Ftoon Kedwan - NLP Application - Natural Language Questions and SQL Using Computational Linguistics-CRC Press (2023)

NLP Application
Applying Natural Language Processing (NLP) concepts to help humans in

their daily life, this book discusses an automatic translation of an unstructured
Natural Language Question (NLQ) into a Structured Query Language (SQL)
statement. Using SQL as a Relational DataBase (RDB) interaction language,
database administrators or general users with little to no SQL querying abilities
are provided with all the knowledge necessary to perform queries on RDBs in
an interactive manner.
Key Features:
• Includes extensive and illustrative examples to simplify the dis-

cussed concepts
• Discusses a novel, and yet simple, approach to NLP
• Introduces a lightweight NLQ into SQL translation approach
through the use of RDB MetaTables as a Hash table
• Extensive literature review and thorough background information
on every tool, concept and technique applied
Providing a unique approach to NLQ into SQL translation, as well as compris-

ing disparate resources on NLP as a whole, this shortform book is of direct use
to administrators and general users of databases.
Ftoon Kedwan is a recognized expert and researcher in the field of Artificial

Intelligence, Software Engineering, Big Data Analysis, Machine Learning,
and Health Informatics. Dr. Kedwan is experienced in industry and academics
and is affiliated with many recognized organizations and universities, such as
Queen’s University, the Royal Military College of Canada (RMC), St. Francis
Xavier University (StFX), the University of Prince Mugrin (UPM), and the
Saudi National Guard Hospital.
NLP Application
Natural Language Questions
and SQL using Computational
Linguistics
Ftoon Kedwan
Cover Image Credit: Shutterstock.com
First edition published 2024

by CRC Press
2385 NW Executive Center Drive, Suite 320, Boca Raton, FL 33431
and by CRC Press

4 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN
CRC Press is an imprint of Taylor & Francis Group, LLC

© 2024 Ftoon Kedwan
Reasonable efforts have been made to publish reliable data and information, but the author and
publisher cannot assume responsibility for the validity of all materials or the consequences of
their use. The authors and publishers have attempted to trace the copyright holders of all material
reproduced in this publication and apologize to copyright holders if permission to publish in this
form has not been obtained. If any copyright material has not been acknowledged please write and
let us know so we may rectify in any future reprint.
Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced,
transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or
hereafter invented, including photocopying, microfilming, and recording, or in any information
storage or retrieval system, without written permission from the publishers.
For permission to photocopy or use material electronically from this work, access www.
copyright.com or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive,
Danvers, MA 01923, 978-750-8400. For works that are not available on CCC please contact
[email protected]
Trademark notice: Product or corporate names may be trademarks or registered trademarks and are
used only for identification and explanation without intent to infringe.
ISBN: 978-1-032-53835-8 (hbk)

ISBN: 978-1-032-53837-2 (pbk)
ISBN: 978-1-003-41389-9 (ebk)
DOI: 10.1201/b23367
Typeset in Times
by SPi Technologies India Pvt Ltd (Straive)
Contents
Preface ix
1 Introduction 1
Basic Research Framework Organization 4
2 Background Study 7
NLQ Input Processing Interface, the NLIDB 7
Interactive Form-Based Interface 7
Keyword-Based Query Interface 8
NLQ-Based Interface 8
Part of Speech (POS) Recognition 9
Linguistic Components Layers 9
Syntactic parser (rule-based) 9
Semantic parser (rule-based) 10
Lexicon 11
Intermediate Language Representation Layer 11
Annotator 11
Disambiguation 12
Matcher/Mapper 13
SQL Template Generator 14
SQL Execution and Result 15
3 Literature Review 17
Related Works 17
NLP 17
ML Algorithms 21
NLQ to SQL Mapping 23
Current Research Work Justification 32
Authoring Interface-Based Systems 33
Enriching the NLQ/SQL Pair 33
Using MLA Algorithms 33
Restricted NLQ Input 33
Lambda Calculus 33
v
vi Contents
Tree Kernels Models 33

Unified Modeling Language (UML) 34
Weighted Links 34
Mapping NLQ Tokens into RDB Lexica 36
NLQ Tokens Extraction 36
RDB Lexica Mapping 36
RDB Lexica Relationships 37
NLP Syntax and Semantics Definition 38
Mapping RDB Lexica into SQL Clauses 39
SQL Clauses Mapping 39
Complexity vs Performance 40
SQL Formation vs SQL Templates 42
Intermediate Representations 44
4 Implementation Plan 47
NLQ Input Interface 47
POS Recognition 48
Disambiguation 54
Matcher/Mapper 55
Mapping NLQ Tokens into RDB Elements 56
Mapping RDB Lexica into SQL Clauses 59
SQL Template Generator 60
SQL Execution and Result 62
5 Implementation User Case Scenario 63

User Case Scenario 68
6 Implementation Testing and Performance Measurements 79

Implementation Environment and System Description 79
DataBase 81
Implementation Testing and Validation 83
Performance Evaluation Measurements 84
7 Implementation Results Discussion 93

Implementation Limitations 94
Mapping Limitations 94
SQL Generation Limitations 95
General Implementation Limitations 96
Contents vii
8 Conclusion and Future Work 99

Conclusion 99
Future Work 100
Appendix 1 103
Appendix 2 105
Appendix 3 107
Appendix 4 109
Appendix 5 111
Appendix 6 113
Appendix 7 115
Appendix 8 117
Appendix 9 119
Glossary 147
References 149
Index 161
Preface
This book applies Natural Language Processing (NLP) concepts to help

humans in their daily life. It discusses an automatic translation of an unstruc-
tured Natural Language Question (NLQ) into a Structured Query Language
(SQL) statement. SQL is used as a Relational DataBase (RDB) interaction
language with special query syntax and a computer-executable artificial lan-
guage. This way, DataBase (DB) administrators or general users with little
or no SQL querying abilities can perform queries on RDBs in an interactive
manner. The Human–Computer Interaction (HCI) happens using users’ NLQs,
which is in English in the proposed research. Users do not need to know any
RDB schema elements or structures such as tables’ names, relationships, for-
mats, attributes, or data types. The RDB schema is a brief description of the
RDB elements’ organization, excluding any RDB values. In this work, a light-
weight NLQ into SQL translation approach is implemented by utilizing an
RDB MetaTable as a Hash table. The main goal is to exploit a manually written
rule-based mapping constraints algorithm. This algorithm maps NLQ tokens’
semantic/syntactic information into RDB elements’ semantic roles (i.e., value,
attribute) in addition to the Wh-Words (e.g., What, Where, How, and Who) of
the actual data and the relationships between the attributes. via pairing and
matching means. The matching RDB elements, called “identified lexica”, are
then mapped into the SQL clauses consistently for SQL generation and execu-
tion. The matching process uses a computational linguistic analysis mapping
algorithm, represented in the MetaTables. This mapping algorithm proved to
be efficient especially with small RDBs with an accuracy of 95% and is about
93% accurate with larger RDBS.
ix
Introduction
1
NLP is a subfield of computer science and engineering under the field of
Artificial Intelligence (AI), as illustrated in Figure 1.
It is developed from the study of language and computational linguistics
[1, 2] and often used to interpret an input Natural Language Question (NLQ)
[3]. NLP’s goal is to analyze and facilitate the interaction between human
language and computing machines. HCI becomes a part of NLP when the
interaction involves the use of natural language. Under NLP, there are several
subareas, including Question Answering Systems (QAS) [4], such as Siri for
iPhones, [5] and summarization tools [6]. Such tools produce a summary of a
long document’s contents or even generate slide presentations. Machine real-
time translation [7], such as Google Translate [8] or BabelFish [9], are among
other examples of NLP subareas. In addition, document classification [10] via
learning models is a famous NLP subarea. It is used to train the classifica-
tion algorithm to identify the category a document should be placed under,
for example, news articles or spam filtering classification. Speech Recognition
Models [11] are another NLP subarea that recognizes spoken language words
which work best only in specific domains.
In the current research, the framework starts with a processing of simple
Online Transactional Processing (OLTP) type of queries. OLTP queries are
simple SELECT, FROM and WHERE statements of the Structured Query
Language (SQL), which is the simplest query form. As in Figure 1, NLP uses
deep learning techniques as part of the AI area. It focusses on computational
linguistics to analyze HCI in terms of the language used for this interactive
communication. Basically, NLP bridges the gap between computers and
humans and facilitates information exchange and retrieval from an adopted
Computer Deep
AI NLP HCI NLIDB
Science Learning
FIGURE 1 Research area breakdown.
DOI: 10.1201/b23367-1 1
2 NLP Application
DataBase (DB), which is, in this case, a Natural Language Interface for
DataBase (NLIDB).
A Relational DataBase (RDB) model [12], which was originally intro-
duced in 1970, is used as the basis of a background data storage and manage-
ment structure. RDB has been chosen as a data storage media for the proposed
research because of the relationships between its entities, including their table
attributes and subfields and their values. These relationships hold significant
information themselves as if they were whole separate entities. In addition,
the information stored on those relationships proved to increase the accuracy
of data retrieval as will be demonstrated later in Chapter 6 on implementation
testing and performance measurements.
RDB’s elements (i.e., Table, Attribute, Relationship, etc.) representation in
Figure 2 describes the relationships between the entity sets to express parts of
Post-Traumatic Stress Disorder (PTSD) RDB semantics. Therefore, an Entity-
Relational Diagram (ERD) [13] was used to demonstrate the RDB structure
because this makes data relationships more visible as follows:
• An entity sets represent a table, i.e., “Patient” table.

• The entity’s features represent specific table’s attributes, i.e.,
“P_Name”.
• Any instance of a specific attribute represents an attribute’s value,
i.e., “Sarah”.
• The directed relationships represent entities’ act or impact on other
entities, i.e., “take”.
In 1976, Chen [13] was the first to graphically model RDB schema entities
using ERD to represent NLQ constructs. In [14], ERD was used to represent
NLQ constructs by analyzing the NLQ constructs’ inter-relationship with the
ERD or even with the Class Diagram Conceptual Schema [15]. However, NLQ
FIGURE 2 ERD for Post-Traumatic Stress Disorder (PTSD) RDB.

1 • Introduction 3
constructs intra-relationships were not studied in the previous works in terms

of mapping the constructs into an RDB query language, such as SQL, which
will be implemented in the current research work. Intra-relationships are the
lexical dependencies between the sentence constructs (i.e., words), such as the
relationship between the verb and its object. This type of relationship supports
the mapping of tokens into lexica.
Furthermore, RDB enables NLQ into SQL mapping using RDB schema
MetaTables, such as in Tables 1 and 2 (see Chapter 2). MetaTables are data
repositories that act as data dictionaries which describe the RDB elements
and the relationships between them using data annotations, span tags and syn-
onyms attachment. In addition to the used RDB MetaTables, the NLQ con-
text and the situation-based linking of knowledge stored in multiple connected
RDB tables all help enhance the accuracy of data retrieval. Under the scope
of the proposed research, the major focus will be on finding an answer to the
question: How can we translate an unstructured full-text NLQ expression to an
SQL statement using RDB schema MetaTables such that it produces accurate
results? This translation mechanism and its framework design, starting from
the NLQ interface and up to identifying the equivalent SQL clauses, aims to
achieve other secondary tasks, such as:
Abbreviation support; so that if a user asks about a patient’s weight using
the abbreviation “kg”, it is recognized as a “kilogram”.
• Support for SQL’s syntactic constructs’ (keywords such as SELECT,

FROM, etc.) synonyms or the absence of them in the NLQ.
• Multiple columns SELECT; to recognize multiple NLQ’s main
nouns or noun phrases.
• Converting operators and numbers (i.e., Equal, Three) into numeri-
cal forms and symbols (i.e., =, 3).
• Deriving tables’ or attributes’ names from the literal values men-
tioned in the NLQ.
• Support for propositional terms (i.e., above, below, between).
• Considering all conditional terms in the NLQ and converting them
into multiple WHERE conditions. This shall apply whether all con-
ditions are applied on the same or different RDB elements.
• Support for aggregate functions (e.g., convert ‘highest’ to ‘MAX
(Hight)’ or ‘youngest’ to ‘MIN(Age)’).
Table 1 is an example of the NLQ MetaTable that breaks down the entered
NLQ into its subsequent tokens. Table 2 is an example of the RDB elements
MetaTable that explains each RDB element in terms of its nature, category,
syntactic role etc. These two tables will be referenced to and elaborated on
frequently throughout this research document.
4 NLP Application
Machine-readable instructions are mandatory to access any type of DB.

This emphasizes the need to find a mapping mechanism between NLQ and
the RDB query languages, such as SQL. Common semantics between NLQ
and artificial languages can be discovered by analyzing the language semantic
and syntax roles. This research work solves the NLQ into SQL translation
problem by manually writing a rule-based mapping algorithm at the word-
processing level for automatic mapping. The manual work is on creating the
rule-based algorithm. After the algorithm is developed, the mapping process
shall be automatic.
The aim of this work is to maintain a simple algorithmic configuration
with high outcome performance. Avoiding the reliance on a huge annotated
corpus or written patterns of translation examples is also very important,
except in the case of simple algorithmic rules. This research idea overcomes
any poor underlying linguistic tools’ performance such as named entity tag-
ger, tokenizer, or dependency parser. This research contribution is a rule-based
algorithm that applies a mapping association between the NLQ’s semantic/
syntactic information, the SQL’s syntactic information and the RDB elements’
semantic roles and it offers an effective translation mechanism to convert
NLQs into SQLs.
BASIC RESEARCH FRAMEWORK

ORGANIZATION
In the implementation of the proposed algorithm, and as illustrated in Figure 3,
user’s NLQ is accepted as an input into a given NLIDB, together with its cor-
responding RDB schema. Next, an NLQ into SQL translation is performed
by an underlying multi-layered NLP framework. Afterwards, the system can
basically understand the NLQ and respond to it with its equivalent SQL query.
FIGURE 3 Basic research organization diagram.

1 • Introduction 5
In the following sections, Chapter 2 is a background study on the research

work pipeline components, which acts as an introduction to the actual research
work implementation plan described in Chapter 4. Chapter 3 is a literature
review of the main research contributions, which are the two mapping algo-
rithms. Chapter 5 is a running example using an implementation user case
scenario to illustrate the role of each component. This chapter also presents
a complete summarization of the algorithm’s processes. Chapter 6 sets out
the algorithm’s implementation testing and validation methods, in addition
to the implementation performance measurements used to illustrate the suc-
cess factors of the proposed research work. Chapter 7 presents the framework
implementation results and discussion and then, finally, Chapter 8 presents the
research conclusion and outlines the scope for future study.
Background
Study 2
For terminology clarification purposes, Figure 4 explains the Natural Lan
guage Query (NLQ) words journey to tokens and then to lexica throughout the
processes executed and summarized on the arrows.
NLQ INPUT PROCESSING

INTERFACE, THE NLIDB
This first step is a point of interaction with the user. The user interface could
be either a simple coded data input interface or a web-based Graphical User
Interface (GUI). GUIs use web design languages such as HTML, CSS, and
PHP. This interface is used to enter a question in an NLQ form. Generally,
NLIDB could be any of the following options:
Interactive Form-Based Interface [16]

Though this looks attractive, it does not always retrieve data variations from
the DB as we can easily define in a formal SQL statement.
FIGURE 4 NLQ words, tokens and lexica terminology clarification.
DOI: 10.1201/b23367-2 7
8 NLP Application
Keyword-Based Query Interface [17]

Extracts keywords from the input NLQ and identifies whether those keywords
are domain-specific or language constructs. Next, the system applies transla
tion rules from the generated knowledge base to use those keywords to form
an SQL query.
NLQ-Based Interface [18]

Unlike previous options, this approach tries to actually understand the NLQ
first before converting it into an SQL statement. To do this, the system applies
lexicon breakdown and NLP syntactic and semantic analysis on the input NLQ.
Lexica is the plural of lexicon, which is the name of a matching RDB element
with a stemmed NLQ token.
The NLQ questions are either:
1) Simple straightforward questions with simple linguistic structures

that have one question word, complete semantic meanings, and no
ambiguous expressions; or
2) Complicated general questions with lots of ambiguous expressions
and fuzzy predicates. The ambiguity could be due to omitting neces
sary words, adding unrelated words, including more than 1 question
word, or the lack thereof.
Furthermore, NLQ questions can be further categorized as questions asking for

a specific entity, quantity, rank of a list of entities or a proportion ratio. They
can also be interrogative questions, being either imperative or declarative in a
negative or affirmative form. The NLP linguistic tools are trained to capture
those question structures and types.
In the proposed research, an NLQ-based interface is chosen as an inter
face to the Natural Language Interface to DataBase (NLIDB) system. This
is because it is the most natural way of communicating with a computing
machine. It is also the most desirable way of communication for decision-
makers, who are the target audience for the proposed research work. In addition,
full semantic understanding of the NLQ by the system is the most essential part
before attempting to translate it to any other query language.
With regard to the adopted NLQ type, only simple straightforward ques
tions, which is the first NLQ question type as explained earlier, are considered.
This is because the main focus and contribution of the proposed research work
is the mapping mechanism, not resolving or disambiguating misstructured or
incomplete NLQs. Yet there is a simple disambiguation module implemented for
2 • Background Study 9
simple NLQ problems. However, addressing the complicated general questions

(i.e., with aggregation or negation) is one of the main suggestion for future study.
PART OF SPEECH (POS) RECOGNITION

To recognize part of speech (POS) for NLQs, the following subsequent layers
must be clarified first.
Linguistic Components Layers

In the proposed system, morphological analysis is performed during the lexical
analysis in order to identify keywords’ properties and break them down into
their basic components. A good example for morphological and lexical analy
sis is WordNet [19]. This traces connections between keywords and separates
the words from their suffixes or affixes (e.g., Client’s to ‘client’ and ‘s’). In
addition, WordNet separates words from their punctuations (a.k.a. non-word
tokens).
Other linguistic components, which are also part of the NLP tools in the
current research framework, include the tokenizer, the POS tagger, the depen
dency tree parser, Named Entity Recognition (NER), and syntactic and seman
tic analyzers such as Link Grammar Parser [20] and Stanford Parser [21].
Those NLP tools are principally used for transforming NLQ tokens into a parse
tree structure to then map them into their appropriate SQL query template.
Discourse Analysis [22] and Pragmatics [23] also fall under the latter NLQ
translation method. SQL syntactic parse trees are usually 100% accurate. The
NLQ parse trees are also accurate but usually introduce some noise affecting
the overall NLQ features representation accuracy [24].
The NLQ grammar is distributed to various linguistic areas, including dis
course, pragmatics and text theory [23]. More specifically, the included lin
guistic areas are Cognitive Designs [25], Descriptive Convenience Modules
[26] or specific techniques such as syntax, semantics, morphology, phonology
and lexica.
Syntactic parser (rule-based)

Syntax is the study of how words are combined to form sentences. It includes
the analysis of POS to show how words are related to one another, and their
aggregations to form larger sentence constructions. The syntax analyser ignores
10 NLP Application
sentences that do not conform to a proper language grammar or rules of how

words are supposed to be combined.
POS tags are needed before sentence parsing. Syntactic parsing analyzes
complex sentences with large amounts of ambiguousness and conditional rela
tionships. It also defines the scope of exclusion and probability statements
(statements on the probability of inclusion [27]).
Under syntactic parsing, there are subtasks that also take place, including:
i. Text Structure Analysis: This is the study of how similarly com
parable words are and how their textual types are constructed to
compose larger textual constructions [28].
ii. Lemmatizer: This module derives the original form (the root word)
of each NLQ word, which happens before tokenization.
iii. Tokenizer: Entered NLQ is tokenized by separating the NLQ tex
tual words into separate tokens. Those tokens will be stored and
passed on to the next module for further analysis.
iv. POS Tagger: This is otherwise known as word category disambigu
ation or grammatical tagging. POS tagging takes into consideration
both the word’s definition and its context in the sentence; to mark
the word in the text or corpus to belong to a particular POS.
Semantic parser (rule-based)

Semantics is the study of the actual meaning of a word such that they relate a
certain language word with its intentional meaning representation. An example
of semantics analysis is relating a syntactic constituent with a predicate [29].
This is because all structures created by the syntactic analyzer have meanings.
In the case when there was no equivalent meaning representation to a certain
syntactic structure, the whole sentence is rejected and ignored by the semantic
parser considering it semantically anomalous.
NLQs are processed based on their accurate semantic properties, indicated
by underlying NLP linguistic tools such as the semantic parser. It is important
to define the context in an NLQ with information further than the equivalent
lexical properties. As such, verbs, subjects, objects and their relationships must
be identified.
Under the semantics analysis and parsing, there are the following definitions:
i. Pragmatics [23]: is studying how the textual context affects expres
sions’ meanings. It explores expression’s implicit meaning using
the sentence’s current structure representation. Next, the expression
is reinterpreted to determine its specific implicit meaning.
ii. Morphology [30]: is studying the units of meanings or functions
(called morphemes) in a certain language. A morpheme can be a
word, affix, prefix, or another word structure.
iii. Phonology [31]: is studying the sound patterns of a certain language

to determine which phones are significant and meaningful (the pho
nemes). Phonology also studies syllables’ structures and the needed
features to describe and interpret the discrete units (segments) in
sentences.
iv. Phonetics [32]: is studying the sounds of spoken language. Phonetics
studies how a phoneme is made and perceived. A phoneme is the
smallest unit of an individual speech sound.
Lexicon
Lexicon studies aim at studying a language vocabulary, that is words and
phrases. In the current research work, RDB lexica are stored in the RDB
MetaTable which is used to map the NLQ words to their formal representa
tions in an RDB (i.e., table names, attribute names, etc.). Lexica are analyzed
by both syntactic and semantic parsers.
Intermediate Language Representation Layer

Those layers use a knowledge base to assist the NLQ with the SQL trans
lation process. A good example for this is the discourse representation
structure. This type of structure converts NLQs into SQLs using an ontology-
based semantic interpreter as an intermediate language representation layer.
Discourse Integration Analysis [22] is the study of information exchange
such as in conversations where sentences meanings may change according
to the preceding/proceeding sentences. An example is the use of pronouns in
sentences such as “He” which will depend on the subject or the actor in the
preceding sentence. In order for the discourse analysis process to be thor
ough, the NLQ tokens must be annotated. Hence, the annotator is essential
in this method.
Annotator
Annotations are metadata that provide additional information on particular
data. Natural language datasets are called corpora (the plural of corpus). When
a corpus is annotated, this is called an annotated corpus. Corpora are used to
train Machine Learning Algorithms (MLA). The annotation types include:
• POS Annotation.
• Phrase Structure Annotation.
• Dependency Structure Annotation.
12 NLP Application
In the current research work, all of the above underlying NLP layers are con
sidered and implemented as part of the POS recognition module, with the
exception of the intermediate language representation layer. The employed
POS recognition Python library is “speech_recognition”. This module is essen
tial to the thorough understanding of the given NLQ for proper and accurate
translation.
DISAMBIGUATION
In the current research work, disambiguation is only required when the mapper
finds more than one matching RDB element to a certain NLQ token. Generally,
a word’s meaning disambiguation has special techniques and algorithms. An
example is the statistical keyword meaning disambiguation process using
N-Gram Vectors [33], where N ranges from 1 to the length of the text. N-Gram
Vectors’ statistics are gathered using a training corpus of English language,
while a customised corpus gather statistics as the system is being used. The
latter corpus requires that the user reviews the presented NLQ interpretation
and makes any necessary changes before submitting it for execution. N-Gram
Vectors are principally used for capturing lexical context. It is considered as
a measure of how likely a given token has a particular meaning in a particular
user NLQ input. This is because every token’s meaning depends on the context
in which it was used. This meaning comparison procedure follows a vector
similarity ranking measure where the higher the meaning rank the closer it is
to the true meaning.
Other words’ meaning disambiguation studies involve computational lin
guistics and statistical language understanding. Such studies focus on word
sense disambiguation within a large corpus. Another method to disambigu
ate a word sense is using word collocations. The word collocations method
measures the meaning likelihood of more than one word when they all exist in
the same NLQ. Iftikhar [34] proposed a solution to the disambiguation prob
lem by parsing domain-specific English NLQs and generate SQL queries by
using the Stanford Parser [21]. This approach is widely used with the Not Only
Structured Query Language (NoSQL) DBs for automatic query and design.
Rajender [35] used a controlled NLIDB interface and recommended SQL
query features to reduce the ambiguity in an NLQ input. This is because the
less ambiguous the domain the more accurate results the system can produce.
What can also help resolve such ambiguity is the weighted relationships or
links [36] between RDB elements. In this method, the relationship’s weight
increases by one each time that particular relationship is used. As such, it is
a given that the bigger the relationship weight is the more often the related
RDB elements are queried. This helps the NLIDB system recommend smarter
options for the user to select from with ordered options, where the topmost
option is the most likely.
In the current research work, NLQ disambiguation is not the main focus.
Thus, a simple NLQ disambiguation module is used by applying Stanford
CoreNLP [21] and “nltk.corpus” Python library [37]. Those tools are solely
used to check for NLQ validity. A syntactic rules checker is also used for any
NLQ grammatical mistakes. The adopted procedure is interactive where an
error message pops up to the user, asking them to rephrase a word or choose
from a few potential spelling corrections. A Naïve Bayes Classifier [38] is
implemented to simply classify user’s response as positive or negative (i.e.,
Yes or No).
MATCHER/MAPPER
The Matcher/Mapper module is considered the most complicated part in NLP
science [14]. Therefore, a keyword-based search has attracted many research
ers as the simplest mapping approach since keywords are explicitly identified.
Researchers used it to improve the information retrieval from DBs and solve
(or avoid) the challenge of understanding NLQ tokens and mapping them into
the underlying DB or schema model [39]. The mapper can be Entity-Attribute-
Value (EAV) Mapper [40], Entity Relational (ER) Mapper [13], or eXtensible
Markup Language (XML) Documents Mapper [41]. NLQs are translated into
SQL queries if the used DB or scheme model is EAV or ER. NLQ can be trans
lated to XML documents only if the used system employs a document-based
data model in the underlying information system. More background informa
tion on this module is given in Chapter 3, the literature review.
In the current research work, the adopted mapper is the EAV mapper in
addition to the RDB relationships. NLQ tokens are mapped into RDB lexica
using the NLQ and RDB MetaTables, Tables 1 and 2, respectively. The match
ing lexica will then be mapped with the SQL clauses. This mapping uses the
proposed rule-based algorithm that is based on the observational assumptions
table discussed later in Table 4 (in Chapter 4).
Table 1 is an example of the NLQ MetaTable that breaks down the entered
NLQ into its subsequent tokens. Table 2 is an example of the RDB elements
MetaTable that explains each RDB element in terms of its nature, category,
syntactic role etc. These two tables will be referenced to and elaborated on
frequently throughout this research document.
14 NLP Application
TABLE 1 NLQ MetaTable

WORDS SYNTACTIC ROLE CATEGORY SYNONYMS
Sarah Instance Value Person, Patient, Physician
Has Verb Relationship Own, Obtain, Have
Physician Noun Attribute Doctor, Provider,
Psychiatrist, Surgeon
TABLE 2 RDB elements MetaTable

SYNTACTIC DATA ENCLOSING
WORDS ROLE CATEGORY TYPE PK/FK SOURCE SYNONYMS
Disease Noun Table Word No RDB: PTSD Illness,
Sickness
Take Verb Relationship Word No Tables: Acquire,
Patient- Absorb, Get,
>Medication Gain
P_ID Noun Attribute Digits Yes Attributes: Patient,
Patient, Identification,
Physician, Number,
Medication, Person
Disease
SQL TEMPLATE GENERATOR

This layer uses input from previous layers in conjunction with the available
MetaTables’ data to translate NLQ into SQL. The complexity of SQL que
ries can be incredibly high because some of them might require aggregation,
nesting or negation with multiple tables selection and joining. Therefore,
logical expressions concatenations might be required to solve potential NLQ
complexity (e.g., negation conjunction, subordinates, superlatives, etc.). Basic
WHERE conditions are combined with AND, OR and NOT operators to fall
onto the nested WHERE clauses where a second SELECT clause is nested in
the WHERE clause.
SQL templates and their complexity classifications determine which SQL
category a particular NLQ falls under. The chosen SQL template’s clauses
are filled with the matching RDB lexica. Furthermore, different translation
systems support different SQL templates. Hence, accepted and recognized

templates must be identified, together with their likelihood ratio and occur
rence frequency.
SQL Statements’ Types:
• Data Query Language (DQL): SELECT

• Data Manipulation Language (DML): INSERT, UPDATE,
DELETE
• Data Definition Language (DDL): CREATE, ALTER, DROP
• Data Control Language (DCL): GRANT, REVOKE
• Transaction Control Language (TCL): BEGIN TRAN, COMMIT
TRAN, ROLLBACK
• Data Administration Commands (DAC): START AUDIT, STOP
AUDIT
In the proposed research work, the focus is on the DQL, which has one com
mand phrase (SELECT) in addition to other supplementary clauses (e.g.,
WHERE, AS, FROM). DQL, though it has one main command phrase, is the
most used phrase among SQL’s other phrases, especially when it comes into
RDB operations. The SELECT keyword is used to send an inquiry to the RDB
seeking a particular piece of information. This could be done via a command
line prompt (i.e., Terminal) or through an Application Program Interface (API).
This research proposes a translation algorithm from NLQ into SQL using the
SELECT command phrase and its supplementary clauses.
SQL EXECUTION AND RESULT

At this stage, the end user must establish a connection with the RDB to send
along generated SQLs. This layer performs all Database Management System
(DBMS) functions. Once the translation process is done and the SQL statement is
generated, the query is executed against the RDB. The query results are integrated
in the form of raw data, that is columns and rows, as NLIDB system output.
The query results may get passed back to the first layer to produce a proper
response to the end user in a natural language form. This is a reverse process to
the original proposed research where SQL is the input and NLQ is the output
for answer generation. This reverse process requires a discourse representa
tion structure. The conversion from SQL into NLQ requires an ontology-based
semantic interpreter. This interpreter is an intermediate language representa
tion layer, which is not used in the original proposed research work.
Literature
Review 3
RELATED WORKS
NLP
LUNAR [42–44] was developed in 1971 to take NLQs about a moon’s rock
sample and present answers from 2 RDBs using Woods’ Procedural Semantics
to reference literature and an Augmented Transition Network (ATN) parser
for chemical data analysis. However, it only handled 78% of the NLQs due to
linguistic limitations as it manages a very narrow and specific domain.
Philips Question Answering Machine (Philiqa) [4] was developed in 1977.
Philiqa separates the syntactic and semantics parsing of NLQ as the semantic
parser is composed of three layers, namely English Formal Language, World
Model Language, and schema DB metadata.
LIFER/LADDER [45] was developed a year later, in 1978, as a DB NLP
system interface to retrieve information regarding US Navy ships. It used a
semantic grammar to parse NLQs. Although LIFER/LADDER supported que-
rying distributed DBs, it could only retrieve data for queries related to 1 single
table, or more than 1 table queries having easy join conditions.
In 1983, ASK [46] was developed as a learning and information manage-
ment system. ASK had the ability to communicate with several external DBs
via the user’s NLQ interface. ASK is considered a learning system due to its
ability to learn new concepts and enhance its performance through user’s inter-
action with the system.
TEAM [16] was developed in 1987 as an NLIDB with high portability and
easy configurability on any DBA system without compatibility issues, which
negatively affected TEAM’s core functionality. NLIDB is an NL-based query
interface which the user can use as a means of interaction with the DB to
DOI: 10.1201/b23367-3 17
18 NLP Application
access or retrieve data using NLQs. This interface tries to understand the NLQ
by parsing and tokenizing it to tokens or lexicons, and then applying syntactic
or semantic analysis to identify terms used in the SQL query formation.
In 1997, [17] a method of conceptual query statement filtration and pro-
cessing interface using NLP was defined to essentially analyze predicates
using full-fledged NL parsing to finally generate a structured query statement.
In 2002, Clinical Data Analytics Language (CliniDAL) [27] initiated
a solution for the mapping problem of keyword-based search using the
similarity-based Top-k algorithm. Top-k algorithm searches for k-records of a
dictionary with a significant similarity matches to a certain NLQ compared to
a predefined similarity threshold. This algorithm was successful with accuracy
of around 84%.
In 2002, a Vietnamese NLIDB interface was developed for the economic
survey of DBs. This proposal also included a WordNet-based NLI to RDBs to
access and query DBs using a user’s NL [18].
In the same year, 2002, DBXplorer employed two preprocessing steps,
PUBLISH which builds symbol tables and associated structures, and SEARCH
which fetches the matching rows from published DBs. Together, PUBLISH
and SEARCH enable Keyword-based search to RDBs [12].
In 2004, PRECISE [47] was developed to use the Semantically Tractable
Sentences concept. The semantic interpretation of the sentences in PRECISE
is done by analyzing language dictionaries and semantic constraints. PRECISE
matches the NLQ tokens with the corresponding DB structures in two stages.
Stage 1, it narrows down possible DB matches to the NLQ tokens using the
Maximum Flow Algorithm (MFA). MFA finds the best single Flow Network
as a directed graph to finally specify one source and one path to increase the
flow’s strength. MFA returns the maximum # of keywords back to the system.
Stage 2 is analyzing the sentence syntactic structure. After that, PRECISE uses
the information returned from both stages to accurately transform the NLQ to
an equivalent SQL query. However, PRECISE has a poor knowledge base as
it only retrieves results that are keyword-based due to NL general complexity
and ambiguity.
In 2005, an NLP question answering system on RDB was defined for NLQ
to SQL query analysis and processing to improve the work on XML processing
for the structural analysis with DB support. Query mapping was used to derive
the DB results [48].
In 2006, NUITS system was implemented as a search algorithm using
DB schema structure and content-level clustering to translate a conditional
keyword-based query and retrieve resulting tuples [49].
The Mayo Clinic information extraction system [50] extracts information
from free text fields (i.e., clinical notes), including named entities (i.e. diseases,
signs, symptoms, procedures, etc.) and their related attributes (i.e. context,
3 • Literature Review 19
status, relatedness to patients, etc.). The system works by implementing simple

NLP tasks such as text tokenization, lemmatization, lexicon verification, and
POS tagging.
English Wizard, developed in 2009, is an English NLQ translation tool
for querying RDB using SQL. English Wizard distinguishing features include
having a graphic user interface (GUI) to issue NLQs or report query results
through client/server applications or other DB reporting tools [51].
GINLIDB, in 2009, employed two types of semantic grammar to support
a wide range of NLQs. The first type is a single lexicon semantic grammar
for lexicon nonterminal words, and the second type forms terminal phrases or
sentences using a composite grammar. In addition, GINLIDB uses ATN for
Syntactic analysis to assure tokens structures’ compatibility with allowable
grammatical structures [52].
In 2010, CHAT-80 system used Prolog to process NLQs into three stages
of representations. First, the system represents NLQ words in their logical con-
stants, and then represents NLQ’s verbs, nouns and adjectives with their prep-
ositions as predicates. For NLQ’s complex phrases or sentences, the system
represents them by forming conjunctions of predicates on the DB. CHAT-80
integrates WordNet as a lexicon and the ontology as the semantic interpreter’s
knowledge base. CHAT-80 is considered a domain-independent and portable
NLIDB. It also uses OWL Ontology to define RDB Entity-Relational model to
increase the accuracy of NLQ sentences [53].
In 2012, an NLI for RDB system called Natural Language Application
Program Interface (LANLI) was developed [54]. LANLI utilized semantic and
syntactic tree generation for query execution defined on a Local Appropriator.
LANLI does the NLQ words matching with the corpus tokens via a match-
ing algorithm before using the query formulating tree. LANLI is an effective
NLIDB retrieval system due to its use of accurate tree formation algorithm for
both the NLQ and the built-in DB, besides the use of knowledge dictionary-
like tables that works at interpreting the knowledge that the NLQs might have.
In 2014, restricted NLQ to SQL translation algorithm was proposed
through CliniDAL [55] focusing on managing the complexity of data extrac-
tion from Entity-Attribute-Value (EAV) design model. The algorithm starts
with the Controller receiving an NLQ with its main parameters’ quantity to
perform advanced search. The NLQ gets parsed, categorized and optimized via
the Query Processor using the query (parse) tree and the information stored in
the data source context model, which is a kind of metadata, to produce a query
or parse tree including entities, comparatives and their categories (i.e. Patient,
Medical, Temporal, etc.). CliniDAL’s grammatical parser recognizes the part
of the Restricted NLQ via a free-text resolution mechanism in addition to the
data analytics post-processing steps for a question deep analysis. The third
step is the Query Translator that identifies the NLQ tokens in the query tree
20 NLP Application
so the Mapper can map them to their internal conceptual representation in the
Clinical Information System (CIS). This mapping process is done using the
similarity-based Top-k algorithm, in addition to embedded NLP tools (e.g.,
tokenization, abbreviation expansion and lemmatization). The mapping results
are stored in a generic context model which feeds the translation process with
necessary information just like an index. As such, the corresponding tables and
fields for the NLQ tokens are extracted from the data source context model to
be fed into the SQL SELECT clause, and the value tables are extracted from
the SELECT clause to generate an SQL FROM clause. The query (parse) tree
leaf nodes represent query constraint categories which reflect conjunction or
disjunction using algebraic computations for clinical attributes and their val-
ues. CliniDAL uses the unique unified term, TERM_ID, and its synonyms as
the internal identifier for software processing purposes, as in composing SQL
statements.
In 2017, the special purpose CliniDAL [42] was introduced to integrate
a concept-based free-text search to its underlying structure of parsing, map-
ping, translation and temporal expression recognition, which are indepen-
dently functioning apart from the CIS, to be able to query both structured
and unstructured schema fields (e.g. patient notes) for a thorough knowledge
retrieval from the CIS. This translation is done using the pivoting approach
by joining fact tables computationally instead of using the DBMS functional-
ities. Aggregation or statistical functions require further post-processing to be
applied and computed.
In 2018, an NLI to RDB called QUEST [56] was developed on top of IBM
Watson UIMA pipeline (McCord) and Cognos. QUEST emphasized focus on
nested queries, rather than simple queries, without restricting the user with
guided NLQs. QUEST workflow consists of two major components, QUEST
Engine Box and Cognos Engine Box. Quest Engine Box includes the schema-
independent rule templates that work by extracting lexicalized rules from the
schema annotation file in the online part of this box, besides the rule-based
semantic parsing module that generates lexicalized rules used in the seman-
tic parsing. Quest Engine Box also includes a semantic parser built on many
Watson NLP components, including English Slot Grammar (ESG) parser,
Predicate Argument Structure (PAS), and Subtree Pattern Matching Framework
(SPMF). This box will finally produce a list of SQL sub-queries that are later
fed into the QUEST other box, the Cognos Engine Box. The latter box focuses
on final SQL statement generation and execution on IBM DB2 server. QUEST
proved to be quite accurate compared to previous similar attempts.
Despite the success of above attempts on NLIDB, token-based [57], form/
template-based [58], menu-based [59] or controlled NL-based search, are simi-
lar but simpler approaches require much more effort as the accuracy of the
translation process depends on the accuracy of the mapping process.
ML Algorithms
NLP tools and ML techniques are among the most advanced practices for
information extraction at present [42]. In respect of ML techniques, they could
be either rule-based or hybrid approaches for features identification and selec-
tion or rules classification processes. A novel supervised learning model was
proposed for i2b2 [60], that integrates rule-based engines and two ML algo-
rithms for medication information extraction (i.e. drug names, dosage, mode,
frequency, duration). This integration proved efficiency during the information
extraction of the drug administration reason from unstructured clinical records
with an F-score of 0.856.
ML algorithms have a wide range of applications in dimensionality reduc-
tion, clustering, classification, and multilinear subspace learning [61, 62]. In
an NLP, ML is used to extract query patterns to improve the response time by
creating links between the NLP input sources and prefetching predicted sets of
SQL templates into a temporary cache memory [63]. ML algorithms are typi-
cally bundled in a library and integrated with query and analytics systems. The
main ML properties include scalability, distributed execution and lightweight
algorithms [62].
NLP and knowledge-based ML algorithms are used in knowledge process-
ing and retrieval since the 1980s [64]. Boyan et al. [65] optimized web search
engines by indexing plain text documents, and then used reinforcement learn-
ing techniques to adjust documents’ rankings by propagating rewards through
a graph. Chen et al. [66] used inductive learning techniques, including sym-
bolic ID3 learning, genetic algorithms, and simulated annealing to enhance
information processing and retrieval and knowledge representation. Similarly,
Hazlehurst et al. [67] used ML in his query engine system to facilitate auto-
matic information retrieval based on query similarity measures through the
development of an Intelligent Query Engine (IQE) system.
Unsupervised learning by probabilistic Latent Semantic Analysis is used
in information retrieval, NLP, and ML [68]. Hofmann used text and linguis-
tic datasets to develop an automated document indexing technique using a
temperature-controlled version of the Expectation Maximization algorithm for
model fitting [68]. Further, Popov et al. introduced the Knowledge and Information
Management framework for automatic annotation, indexing, extraction and
retrieval of documents from RDF repositories based on semantic queries [69].
Rukshan et al. [70] developed a rule-based NL Web Interface for DB
(NLWIDB). They built their rules by teaching the system how to recognise
rules that represent several different tables and attributes in NLWIDB system,
what are the escape words and ignore them, in addition to DB data dictionaries,
Rules for the aggregate function MAX, and rules indicating several different
22 NLP Application
ways to represent an ‘and’ or ‘as well as’ concept, or interval ‘equal’ concept.
Data dictionaries are often used to define the relationship between the attri-
butes to know, for example, which attribute comes first in a comparative opera-
tion, and which falls afterwards in the comparative structure. This NLWIDB is
similar to the current research idea, however, we intend on building a simpler
and more generic algorithm that could be applied on various domains systems.
Our algorithm does not use DB elements as the basis or rules identification, as
in Rukshan et al.’s research; rather, it uses general sentence structure pattern
recognition to form an equivalent SQL statement.
SPARK [71] maps query keywords to ontology resources. The translation
result is a ranked list of queries in an RDF-based Query Language (QL) for-
mat, called SPARQL, created in SPARK using a probabilistic ranking model.
The ontology resources used by SPARK are mapped items represented in a
graph format used to feed SPARQL queries.
Similar to SPARK, PANTO [60] translates the keyword-based queries to
SPARQL, but PANTO can handle complex query keywords (i.e. negation, com-
parative and superlatives). Also, PANTO uses a parse tree, instead of graph rep-
resentation, to represent the intermediate results to generate a SPARQL query.
i2b2 medication extraction challenge [60] proposed a high accuracy infor-
mation extraction of medication concepts from clinical notes using Named
Entity Recognition approaches with pure ML methods or hybrid approaches of
ML and rule-based systems for concept identification and relation extraction.
Keyword++ framework [43] improves NLIDB and addresses NLQs’
incompleteness and imprecision when searching a DB. Keyword++ works by
translating keyword-based queries into SQL via mapping the keywords to their
predicates. The scoring process is done using deferential query pairs.
Keymantic system [72] handles keyword-based queries over RDBs using
schema data types and other intentional knowledge means in addition to web-
based lexical resources or ontologies. Keymantic generates mapping con-
figurations of keywords to their consequent DB terms to determine the best
configuration to be used in the SQL generation
HeidelTime [73, 74] is a temporal tagger that uses a hybrid rule-based and
ML approach for extracting and classifying temporal expressions on clinical
textual reports which also successfully solved the i2b2 NLP challenge.
CliniDAL [27] composes Restricted Natural Language Query (RNLQs)
to extract knowledge from CISs for analytics purposes. CliniDAL’s RNLQ to
SQL mapping and translation algorithms are enhanced by adopting a temporal
analyzer component that employs a two-layer rule-based method to interpret
the temporal expressions of the query, whether they are absolute times or rela-
tive times/events. The Temporal Analyzer automatically finds and maps those
expressions to their corresponding temporal entities of the underlying data ele-
ments of the CIS’s different data design models.
Similar to CliniDAL, TimeText [75] is a temporal system architecture for

extracting, representing and reasoning temporal information in clinical tex-
tual reports (i.e., discharge summaries). However, it only answers very simple
NLQs.
NLQ to SQL Mapping

Giordani and Moschitti [76] designed an NLQ translation system that gener-
ates SQLs based on grammatical relations and matching metadata using NL
linguistic and syntactic dependencies to build potential SELECT and WHERE
clauses, by producing basic expressions and combining them with the conjunc-
tion or negation expressions, and metadata to build FROM clauses that contain
all DB tables that S and W clauses refer to, supported by pairing with highest-
weight meaningful joins, with MySQL framework in the back end. However,
queries that involve less joins and SQLs embedding the most meaningful ref-
erenced tables are preferred. Those clauses are then combined using a smart
algorithm to form the final list of possible SQL statements that have matching
structure and clauses’ components related to the DB metadata, NLQ tokens
and their grammar dependencies, which are mapped with NLQ tokens using a
mapping algorithm. Generated SQLs have a weighting scheme which relies on
how many results are found, to order them based on probability of correctness.
NLQ tokens’ textual relationships are represented by the typed dependency
relations called the Stanford Dependencies Collapsed (SDC). SDC works
by representing the NLQq by its typed SDCq list, which is prepared by first
pruning out the NLQ relations of useless stop/escape words, then stemming/
lemmatizing the remaining NLQ’s grammatical relations to reach the opti-
mized list SDCopt used to build the clauses SELECT S and WHERE W, and
lastly adding the relations’ synonyms to the SDCopt list. After that, an iterative
algorithm q is applied, which adds the modified stems to Π and/or Σ (e.g. sub-
ject or object) categories to search the DB metadata for matching fields with
weighted projection-oriented stems and generate the SQL clauses S, F, W, or
the nested queries, so the answers set A = SELECT S × FROM F × WHERE W
contains all potential SQLs related to q. At the end, the system will select the
single SQL from the A set that maximizes the probability of answering query
Q correctly. This NLIDB system effectiveness and accuracy at selecting the
correct SQL depends on the order of the SQL in the generated list. SQLs on
top of the list (top 10) have 81% correct data retrieval, and 92% on the top 5
SQLs. Nevertheless, authors believe that these accuracies can be improved by
learning a reranker to reorder the top 10 SQLs. Yet, this NLIDB is considered
novel as it is expert-independent since all needed knowledge is already in the
DB metadata stores [76].
24 NLP Application
A similar NLIDB system was designed by [77] with comparable per-

formance despite using different approaches as an expert user who specified
semantic grammars is used to enrich the DB metadata and also implemented
ad hoc rules in a semantic parser [77]. Other similar work is the KRISP [78]
system, which achieved 94% Precision and 78% Recall of correctly retrieved
SQL answers.
Giordani and Moschitti [24] have innovated a novel model design for auto-
matic mapping of NL semantics into SQL-like languages by doing the map-
ping at syntactic level between the two languages. After that, Support Vector
Machines (SVM) ML algorithm is applied on the mapping results to derive the
common high-level semantics to automatically translate NLQs into SQLs. To
do this, syntactic parsers were used to define NLQ and SQL trees through the
ML algorithm using a statistical and shallow model. SVM is used here to build
the training and test sets, where the ML model input is a corpus of questions
and their correct answers. SVM then automatically generates the annotated and
labeled set of all probable correct and incorrect relational Question/Answer
(Q/A) pairs. Those RDB pairs are encoded in SVMs using Kernel Functions to
compute the number of common substructures between two trees and produce
the union of the shallow feature spaces of NLQs and SQLs. Moreover, Kernel
Functions is also a combination of Tree Kernels (e.g. Syntactic Tree Kernel
(STK) and Its Extension with leaf features (STKe)) applied to syntactic trees
and Linear Kernels applied to bag-of-words, and both applied to the syntactic
trees of NLQs and SQLs to train the classifier over those pairs to select the cor-
rect SQLs for an NLQ. Then, map this new NLQ to the set of available SQLs
and rank all available SQLs according to their classifier scores and only use the
higher scores NLQs. Ranking potential SQLs to a given NLQ is done through
an SVM using advanced kernels to generate a set of probable NLQ/SQL pairs
and classify them to correct or incorrect using an automatic categorizer on their
syntactic trees by applying Charniak’s syntactic parser on NLQs and a modifi-
cation of the SQL derivation tree using an ad hoc parser on SQL queries. Then,
the top-ranked pairs are selected according to the automatic categorizer prob-
ability score output. Authors tested the mapping algorithm of NLQs into SQLs
using a standard 10-fold cross-validation, the standard deviation, the learning
curve and the average accuracy of correct SQLs selection for each NLQ. This
approach proved to be able to capture the shared semantics between NLQs and
SQLs. It also proved that the implemented kernel improves the baseline model
(32%) according to the cross-validation experiments by choosing correct SQLs
to a certain NLQ. However, a polynomial kernel (POLY) of 3rd degree on a
bag of words is better than STK because it consists of individual tokens that
does not exist in STK. Overall, kernel methods are reliable in describing rela-
tional problems by means of simple building blocks [24].
Tseng and Chen [79] aim at validating the conceptual data modeling power
in the NLIDB area via extending the Unified Modeling Language (UML) [80,
81] concepts using the extended UML class diagram’s representations to cap-
ture and transform NLQs with fuzzy semantics into the logical form of SQLs
for DB access with the help of a Structured Object Model (SOM) representation
[82] that is applied to transform class diagrams into SQLs for query execution
[50]. This approach maps semantic roles to a class diagram schema [80, 81,
83] and their application concepts, which is one of the UML 9 diagrams used
to demonstrate the relationships (e.g. Generalization and Association) among a
group of classes. Carlson described several constraints to build semantic roles
in English sentences [84].
UML is a standard graphical notation of Object-Oriented (OO) model-
ing and information systems design tool used for requirement analysis and
software design. UML class diagrams are used to model the DB’s static rela-
tionships and static data models (the DB schema) by referring to the DB’s
conceptual schema. SOM methodology is a conceptual data-model-driven
programming tool used to navigate, analyze, and design DB applications and
process DB queries [79].
Authors of [79] aim to explore NLQ constructs’ relationships with the OO
world for the purpose of mapping NLQ constructs that contain vague terms
specified in fuzzy modifiers (i.e. ‘good’ or ‘bad’) into the corresponding class
diagrams through an NLI, to eventually form an SQL statement which, upon
execution, delivers answers and a corresponding degree of vagueness. Authors
focused on the fuzzy set theory [49] because it is a method of representing
vague data with imprecise terms or linguistic variables [85, 86]. Linguistic
variables consist of NL words or sentences (i.e. old, young), excluding num-
bers (i.e. 20 or 30), yet imprecise NLQ terms and concepts can be precisely
modeled using these linguistic variables by specifying natural and simple spec-
ifications and characterizations of imprecise concepts and values.
In [79] real-world objects’ connectivity paths are mapped to SQLs during
the NLQ execution by extracting the class diagram from the NLQ in a form
of a sub-graph/tree (a validation sub-tree) of the SOM diagram that contains
relevant objects connecting the source and the target, that have been identified
by the user earlier in a form of objects and attributes. The source is objects and
their associations that have valued attributes to illustrate the relationship of
objects and attributes of interest, while the object of ultimate destination is the
target. Results are then sent to the connectivity matrix to look for any existing
logical path between the source and the target to eventually map the logical
path to an equivalent QL statement which can be simplified by inner joins.
Schema and membership function represented in class diagram are used to link
each fuzzy modifier with their corresponding fuzzy classes.
26 NLP Application
Isoda [87] discovered that OO-based analysis applications enable intuitive

and natural real-world modelling by identifying the corresponding classes of
those real-world objects.
Moreno and Van [88] proved that NLQs constructs’ conceptual modeling
formalization can be mapped naturally into an OO conceptual model, just like
how Metais [89] mentioned about NLQ and DB conceptual schema, and how
efficient they are at representing the real world’s conceptualization features.
Yager et al. [90] prove that the fuzzy sets (consists of fuzzy terms like
young and rarely) theory provides a linguistic-based application of NLQs
modeling. Furthermore, fuzzy NLQs allow users to describe real-world objects
more intuitively through vague predicates including larger number of tuples.
Based on [79], fuzzy NLQs with linguistic terms and fuzzy terms are more
flexible compared with precise NLQs as they provide more potential answers in
case of no available direct answers. Regarding the linguistic inter-relationship
with DB schema, NLQs are linguistically analyzed to reduce ambiguity and
complexity by using a linked dictionary and predefined grammar-based rules,
while a DB schema acts as a DB conceptual design blueprint.
In the same vein, Owei [91] came up with a concept-based QL that facili-
tates query formulation by means of DB queries’ conceptual abstraction to
exploit the semantic data models and map NLQ’s constructs to their equivalent
specific objects in the real-world DB.
L2S [92] is a hybrid approach to transform NLQs into SQL. It maps NL
vocabularies to SQL using semantic information via underlying tools and uses
bipartite tree-like graph-based processing models for handling and evaluat-
ing the remaining lexicons for the matching stage. This hybrid approach was
designed to help language transformation systems that lack adequate training
data and corpus, specific domain background knowledge and observation anal-
ysis and inefficient employed NLQ linguistic tools (e.g., tokenizer and parser).
Most language transformation approaches or question answering systems
rely on rich annotated training data with employed statistical models or non-
statistical rule-based approaches. Generally speaking, rule/grammar-based
approaches [93] require extensive manual rules defining and customizing in
case of any DB change to maintain accuracy and is used more often in real-life
industrial systems [94].
Statistical models are mostly used in academic research, and they work
by building a training set of correct and incorrect NLQ to SQL pairs [24].
The language transformation then becomes a binary classification task to cor-
rectly map and label each NLQ to its equivalent SQL. In addition, ML Features
extraction is derived from tokens and syntactic trees of correctly mapped and
labeled NLQs into SQLs transformation processes [95].
In addition to statistical and non-statistical models, there are syntactic-
based analysis [59] that is not rule-based, as well as graph-based models [47]
and Prolog-like representation [96], in addition to complex NLQ to SQL trans-

forming operations in frameworks such as the high-level ontology data repre-
sentation with huge amount of data.
L2S functions are spread across three steps [92]. During the first pre-
processing step, NLQ and DB are analyzed, and all ML features are extracted.
This step is supported by three pre-processing components: a linguistic com-
ponent that analyzes the NLQ to generate the list of lexicons and discover any
conditional tokens (i.e. greater/less than). This component has embodied tools
such as Name Entity recognizer (NER) [97] and Coltech-parser in GATE [98]
to extract semantic information that lies within an NLQ. Also, this compo-
nent handles the NLQ tokenization and named entity tagging by using Java
Annotation Patterns Engine (JAPE) grammars. The output of this first step is
an attachment constraint. The second component is the lexicon, used for DB
elements analysis and their attachment constraints. The third component is the
ambiguity Solving, used for NLQ input correction to guarantee an ambiguity-
free NLQ. In this Component, L2S compares all NLQ tokens with DB ele-
ments to find out the ambiguous tokens that match with non or more than two
DB elements through the use of ellipsis method or the highest possibility selec-
tion. The next L2S step is the Matching step which has two main components,
the Semantic Matching and the Graph-Based Matching components. This step
handles the interlingual mapping to produce equivalent SQL elements to the
output of the first step. The third step is generating a complete and accurate
SQL statement. Authors of [92] built a mapping table manually to match the
DB lexicons with the NLQ tokens. L2S transforms NLQ and DB elements into
a tree-like graph, then extracts the SQL from the maximum bipartite match-
ing algorithm result. L2S effectiveness was validated and the results main-
tained high accuracy over different domains. It was also proved that switching
domains requires minimal customization work.
According to [99], NLP is today an active technique of Human–Computer
Interaction, especially in the social media era [99], developed a structural
design method to automatically convert and translate simple DDL and DML
queries with standard join conditions from an NLQ format into SQLs through
NLQ’s semantic extraction and optimized SQL generation. This work also pro-
vides a user-friendly NLI for end users to easily access the social web DB via
any web source. Authors used Java Programming Language and its technical
tools to build the NLIBD system’s front-end and used R-tool as a data collec-
tor to gather data from social web sources. For data storage in the system’s
backend, authors used an oracle DB, the SQL server. They also used a limited
Data Dictionary to store all system-related words. The system would receive
an NLQ, process it, collect data using the R-tool interface, extract the semantic
knowledge from the social web source, and finally generate the associated SQL
statement.
28 NLP Application
In [99], a simple architectural layout framework is proposed. It starts

with the four pre-processing phase modules; Morphological Analysis Module,
Semantic Analysis Module, Mapping Table Module, and the Reports’ Retrieval
Module. In the first module, the Morphological Analysis, NLQ is received and
tokenized into words, which will be then passed on to the extractor to stem them
by identifying their root words. At this stage, unwanted words are removed.
Then, the extractor does the stemming processes using the Porter algorithm,
in addition to maintaining previously tokenized words, a.k.a. the predefined
words, from previous NLQs to compare them with the newly tokenized words
to extract the main keywords, which are passed to the second module. NLQ
tokens’ synonyms are identified from the integrated DB elements’ names (i.e.,
column or table names), which are used to replace the extracted keywords.
The Semantic Analysis Module generates a parse tree from the identified
keywords and passes it on to the third module. The Mapping Table Module
has all potential SQL templates and knows the maximum possibility of each
NLP word, and hence does all the identified words mapping using the mapping
table. The best suitable query is generated and passed on to the fourth and last
module, the Reports’ Retrieval Module, to deliver it to the end user as a report.
The noticeable effort in this work is that every entered NLQ goes through
a Syntactic Rules Checker for any grammatical mistakes. Also, semantic anal-
ysis is used to map NLQ tokens to DB objects. The combination of tokens’
meanings defines the NLQ general meaning, which is used to come up with a
list of potential SQL queries among which the end user has to choose.
Akshay [100] proposes an NLP search system interface for online Appli
cations, search engines and DBs requiring high accuracy and efficiency.
Kaur [101] illustrates the useful usage of Regular Expressions (regexps),
which are generic representations for strings, in NLP phonology and morphol-
ogy, text search and analysis, speech recognition and information extraction.
However, clear collections of regexps in NLQ sentences are not clearly specified.
Avinash [102] uses domain ontology in NLIDBs for NLQ’s semantic
analysis and emphasizes on employing domain and language knowledge at the
semantic level to enhance precision and accuracy.
Kaur and Bali [45] examined an NLQ into SQL conversion interface mod
ule by means of NLQ’s syntactic and semantic analysis, but this module is unable
of processing complete semantic conversion for complex NLQ sentences.
Arati [103] used Probabilistic Context Free Grammar (PCFG) as an
NLDBI system design method for NLP. It works by using NLQ’s syntactic and
semantic knowledge to convert NLQ into an internal representation and then
into SQL via a representation converter. However, finding the right grammar
for optimization is challenging in this system design.
Dshish [104] used NLP for query optimization to translate NLQ into SQL
by means of semantic grammar analysis, while using the LIFER/LADDER
method in the syntax analysis. Since the LIFER/LADDER system only sup-
ports simple SQLs formation, this translation architecture is largely restricted.
For RDBMSs, Gage [105] proposed a method of an AI application in addi-
tion to fuzzy logic applications, phrase recognition and substitution, multilin-
gual solutions and SQL keyword mapping to transform NLQs into a SQLs.
Alessandra [24] used Syntactic Pairing for semantic mapping between
NLQs and SQLs for eventual NLQ translation using SVM algorithm to design
an RDB of syntax trees for NLQs and SQLs pairs, and kernel functions to
encode those pairs.
Gauri [106] also used semantic grammar for NLQ into SQL translation.
In the semantic analysis, the author used the Lexicon to store all grammatical
words, and the post-preprocessor to transform NLQs’ semantic representations
into SQL. However, this architecture can only translate simple NLQs, but not
flexible.
Karande and Patil [51] used grammar and parsing in an NLIDB system for
data selection and extraction by performing simple SQLs (i.e. SQL with a join
operation or few constraints) on a DB. This architecture used an ATN parser to
generate parse trees.
Ott [107] explained the process of SQLs Automatic Generation via an
NLIDBs using an internal intermediate semantic representation language
based on formal logic of the NLQs that is then mapped to SQL+. This approach
is based on First Order Predicate Calculus Logic resembled by DB-Oriented
Logical Form (DBLF), with some SQL operators and functions (e.g., negation,
aggregation, range, and set operator for SQL SELECT).
This approach, called SQL+, aims to solve some of the SQL restrictions
such as handling ordinals via loop operators (e.g., the 6th lowest, the 3rd high-
est). To replace the loop operator, SQL + expressions are entered into a pro-
gramming interface to SQL supplied with a cursor management.
SQL+ strives to save the ultimate power of NLQ by augmenting the SQLs
in a way that each NLQ token is represented and answered by SQL expres-
sions. Experiment results prove that even complex queries can be generated
following three strategies, the Join, the Temporary Relation and the Negation
Strategy, in addition to a mixture of these strategies [107].
For the join strategy, and in the DBLF formula, for each new relation
reference a join operation is built in the SQL FROM clause recursively in a
top-down direction. Universal quantifiers are usually implemented by creating
counters using double-nested constructs as in (NOT EXISTS [sub-SQL]) which
has been used in the TQA system [108]. However, [107] uses the temporary
relation creation strategy instead to handle universal and numeric quantifiers
and to handle ordinals and a mixture of aggregate functions as well. The tem-
porary relations are created using the join-strategy to easily embed them in any
SQL expression. Hence, whenever there is a quantifier, a temporary relation
30 NLP Application
is built for it recursively. For the negation strategy, and in DBLF, negation is
handled by setting the “reverse” marker for yes/no questions if the negation at
the beginning of the sentence, and by using (NOT IN [subquery]) constructs
in case of verb-negation and negated quantifiers in other positions. Both nega-
tion handling methods are doable if the negation occurs in front of a simple
predicate, and in this case, the number and position of negation particles is not
restricted. For the Mixed strategies, any of the previous three strategies can be
mixed arbitrarily as in building temporary relations when aggregate functions
or ordinals occur.
TQA [108] is an NLI that transforms NLQs into SQL directly using
semantic grammar and deep structure grammar to obtain a higher performance
and better ellipses and anaphora handling. However, TQA and similar systems
are almost not transportable and barely adaptable to other DB domains.
PHLIQAl [4], ASK [46] and TEAM [16] adopt the intermediate semantic
representation languages connected with a Conceptual Schema (CS) to provide
an efficient NLQ into SQL transformation tool. CS helps mapping NLQ tokens
to their DB lexicon representations because it stores all NLQ tokens, DB terms
of relations and attributes and their taxonomies, in addition to the DB hierarchy
structure and metadata.
The USL system [109] is adaptable and transportable because it has a
customization device. Yet its intermediate semantic and structure language
is syntax-oriented, and not based on predicate logic. Hence, some semantic
meanings are represented though tree structures forms.
TQA and USL techniques together form the LanguageAccess system
[107], which has a unique SQL generation component that uses the DBLF
and the Conceptual Logical Form (CLF) as two distinct intermediate semantic
representation languages.
LanguageAccess system works through many steps. First, a phrase struc-
ture grammar parses an NLQ to generate parse trees, which are then mapped
to CLF using the CS. Generated CLF formulae are paraphrased in NL and
then presented to the end user for ambiguous tokens interpretations and mean-
ing verification. Once the end user chooses a CLF formula, using the CS, it
gets transformed to DBLF, the source of SQL generation, which is then trans-
formed to SQLs. DBLF considers the DB values internal representations (e.g.
strings, numbers), the temporary relations order, and the generated expressions
delivery mechanism to DBs.
Authors of [7, 110, 111] used NLQs semantic parsing to model algorithms to
map NLQs to SQLs. Similar research work is done by [112] using specific semantic
grammar. Authors of [7, 110, 113] used lambda calculus and applied it on NLQs
meaning representation for the NLQ to SQL mapping process. Furthermore,
[114] used ILP framework’s defined rules and constrains to map NLQs using
their semantic parsing. For the same purpose, [7, 110, 111] followed a time-
consuming and expensive approach by producing NLQ tokens’ meaning
representations manually. Similarly, [112] developed an authoring system
through extensive expertise time and efforts on semantic grammar specifica-
tion. Authors of [7, 110, 113] developed a supervision-extensive system using
lambda-calculus to map NLQs to their corresponding meaning representations.
Similar to KRISP [78], Giordani and Moschitti [115] developed a model using
only Q/A pairs of syntactic trees as the SQL compiler provides the NLQs deriva-
tion tree that is required to translate factoid NLQs into structural RDB SQLs with
generative parsers that are discriminatively reranked using an advanced ML SVM-
ranker based on string tree kernels. The reranker reorders the potential NLQ/SQL
pairs list which has a recall of 94%, recalling the correct answers in this system.
The system in [115] does not depend on NLQ-annotated meaning
resources (e.g. Prolog data, Lambda calculus, MR, or SQLs) or any manual
semantic representations except for some synonym relations that are missing
in WordNet. The first phase is the generation phase where NLQ tokens’ lexi-
cal dependencies and DB metadata-induced lexicon in addition to WordNet
are used, instead of a full NLQ’s semantic interpretation, to build the SQL
clauses (i.e. SELECT, WHERE, FROM, joins, etc.) recursively with the help
of some rules and a heuristic weighting scheme. DB metadata does the rela-
tions disambiguation tasks and includes DB data types, Primary Keys (PKs)
and Foreign Keys (FKs) and other constraints, names of entities, columns and
tables according to domain semantics, and is also called DB catalog usually
stored as INFO_SCHEMA (IS) in a DB. The output of the generation phase is
a ranked potential SQLs list created by the generative parser.
In Dependency Syntactic Parsing is used to extract NLQ tokens’ lexical
relations and dependencies. According to [19], WordNet is efficient at expand-
ing predicate arguments to their meaning interpretations and synonyms;
however, WordNet generalizes the relation arguments but does not guarantee
NLQ’s lack of ambiguity and noise which affects its meaning interpretation
significantly. Therefore, this system generates every possible SQL with all of
its clauses, including ambiguous ones, based on NLQs lexical and grammati-
cal relations dependencies matches, extracted by the Stanford Dependencies
Parser [116], and SQL clauses’ logical and syntactic formulation structures.
The first relation executed on the GEOQUERIES corpus in the [115] algo-
rithm is the FROM clause relation to find the corresponding DB tuples consid-
ering the optional condition in the WHERE clause and then match the results
with the SELECT clause attributes. In case of any empty clauses or nested
queries mismatching, this algorithm will generate no results; otherwise, correct
SQLs are generated among the top three SQLs in 93% of the times using stan-
dard 10-fold cross-validation performance measure. This high accuracy and
32 NLP Application
recall are due to the robust and heuristic weights-based reranker that is built
using SVM-Light-TK6 extending the SVM-Light optimizer [117] by employ-
ing the tree kernels [118, 119] to use the addition STKn + STKs or the multi-
plication STKn × STKs. Default reranker parameters are used such as in the
normalized kernels, λ = 0.4 and cost and trade-off parameters = 1. However,
this approach mandates the existence of possible SQLs in advance as no new
SQLs can be generated by the algorithm, it only verifies if an entered NLQ has
a corresponding SQL to produce a correct answer.
Conceptually similar to [115], Lu et al.’s [120] mapping system does
not depend on NLQ annotation either, but on a generative model and the
(MODELIII+R) which is a discriminative reranking technique. Also, DCS sys-
tem [121] does not depend on a DB annotation either and works as well as a
mapping system enriched with prototype triggers (DCS+). In addition, from Q/A
pairs, SEMRESP employs a semantic parser learner [122] that works best on
annotated logical forms (SQLs). Kwiatkowski et al. [123] developed UBL system
that when trained on SQLs and Q/A pairs, it is able to use restricted lexical items
together with some CCG combinatory rules to learn newly entered NLQ lexicons.
CURRENT RESEARCH WORK

JUSTIFICATION
The main goal proposed in this work is to find a simple but accurate mapping
mechanism between NLQ and machine-readable instructions such as the RDB
query language, SQL. To date, there is no adequate NLQ into SQL translation
mechanism that does not compromise accuracy and precision with complexity
or exaggerated simplicity to an unfunctional level. Such translation mecha-
nisms have numerous rules exceptions and resulted errors when applied on
other RDBs. Thus, the proposed research exploits a simple manually written
rule-based mapping constraints algorithm as a design to a unique and accu-
rate NLQ into SQL translation mechanism. This algorithm maps NLQ tokens
semantic and syntactic information into RDB elements categories (i.e., value,
attribute, etc.) and then into SQL clauses consistently. The algorithm uses
computational linguistics analysis pairing and matching mechanisms through
MetaTables. The study of translating NLQ into SQL-like languages has a long
history starting from the 1971 to date [2–4, 12, 16–18, 35, 39, 40, 41, 46–49,
52, 53, 57, 59, 60, 65–69, 70–73, 99, 100, 124–130].
According to the literature [57, 74], mapping NLQ into SQL occurs using
any of the following approaches:
Authoring Interface-Based Systems [131]

By using semantic grammar specifications designed by extensive expertise
time and efforts to identify and modify RDB elements and concepts (e.g.,
CatchPhrase Authoring tool [131]).
Enriching the NLQ/SQL Pair

By adding extra metadata to the pairs to easily find a semantic interpretation
for NLQ’s ambiguous phrases for the matching problem (e.g., Inductive Logic
Programming (ILP) [75]).
Using MLA Algorithms

By using correct NLQ/SQL pairs’ corpora. A corpus induces semantic gram-
mar parsing to map NLQs into SQLs by training a Support Vector Machine
(SVM) classifier [33] based on string subsequence kernels (i.e., Krisp [132]).
Restricted NLQ Input [35]

By using a keyword-based search structure [128] through a form, template or
a menu-based Natural Language Interface (NLI) [52] to facilitate the mapping
process.
Lambda Calculus [7, 112, 114]

Applied on NLQs meaning representation for the NLQ into SQL mapping
process.
Tree Kernels Models [29, 47, 78, 100, 126]

A Kernel Function [110, 113, 133] is a combination of Tree Kernels [134]
such as the Polynomial Kernel (POLY) [135], Syntactic Tree Kernel (STK)
[33] and its extension with leaf features (STKe) [132]. They can be applied on
NLQs/SQLs pairs syntactic trees, while Linear Kernels [136] are applied on a
“bag-of-words”. They are used to train the classifier over those pairs to select a
correct SQL for a given NLQ.
34 NLP Application
Unified Modeling Language (UML) [34, 118]

A standard graphical notation of Object Oriented (OO) Modeling [137] and an
information system design tool. UML is a combination of Rumbaugh’s Object-
Modeling Technique (OMT) [119], Booch’s OO Analysis and Design [74], and
Jacobson’s Objectory [100]. UML class diagrams are used to model the DB’s
static relationships and static data model (DB schema) by referring to the DB’s
conceptual schema.
Weighted Links [81]

Is a mapping system that works by pairing with the highest weight meaningful
joins between RDB lexica and SQL clauses.
What follows is an explanation of each of the above mapping approaches
and a justification of excluding them by choosing the second approach, enrich-
ing the NLQ/SQL pairs, as the most effective one.
In the current research, the second approach is adopted because it does not
restrict the user to using certain domain-specific keywords [40, 72], as is the
case in the restricted NLQ input approach. This is because the aim in the cur-
rent work is to facilitate the HCI without users’ prior knowledge of the RDB
schema and the system’s infrastructure, underlying NLP linguistic tools or any
specific query language.
In addition, NLQ is the most natural way of communication for humans.
The first approach, authoring interface-based systems, relies heavily on end-
user input through all the interface screens that they have to go through to spec-
ify and modify the used keywords or phrases. Hence, it might seem to the user
that it would have been easier for them if they had any programming knowl-
edge to enter the SQL statement directly without using the Natural Language
Interface (NLI) screens. Hence, the current work only involves end users in
the case of any spelling mistakes or ambiguate phrases. In this regard, in [40,
52, 68, 69, 87], authors described the menu-based or restricted keyword-based
NLQ approaches as methods of mapping. In their paper, they explained how
insignificant restricted NLQ input systems are in terms of accuracy and recall.
Besides, it also has portability problems even with advanced algorithms such
as Similarity-Based Top-k Algorithm [79] that compares the similarity between
dictionary k-records and NLQ tokens. This algorithm achieved an accuracy of
84% only [100], whereas the current research system achieved an accuracy of
as high as 95%.
Most language translation approaches or QAS systems rely on rich anno-
tated training data (corpus) with employed statistical models or non-statistical
rule-based approaches. As such, the third approach that relies on MLA algo-
rithms requires the presence of huge domain-specific (specific keywords used
in the NLQ) NLQ/SQL translation pairs’ corpora. Such a corpus is difficult to
create because it is very time-consuming and a tedious task required by a domain
expert. NLQ/SQL pairs corpus requires hundreds of manually written pairs writ-
ten and examined by a domain expert to train and test the system [70, 76, 138].
Avinash [102] employed a domain-specific ontology for the NLQ’s
semantic analysis. As a result, Avinash’s algorithm would fall under the over-
customization problem, making the system unfunctional on any other domain.
It is also neither transportable nor adaptable to other DB environments, except
with extensive re-customisation. Such domain-specific systems assume the user
is familiar with the DB schema, data and contents. On the other hand, the cur-
rent research work uses simple algorithmic rules and is domain-independent.
Hence, it does not assume prior knowledge of the adopted RDB schema or
require any annotated corpora for training the system. Instead, it uses linguistic
tools to understand and translate the input NLQ. However, the used NLQ/SQL
pairs are only used for algorithm testing and validation purposes. Furthermore,
relying heavily on MLAs proved to be not effective in decreasing the transla-
tion error rates or increasing accuracy [139]. This remains the case even after
supplying the MLA algorithm with a dedicated Error Handling Module [77].
In this regard, the current research work took proactive measures by using
NLP linguistic techniques to make sure the NLQ is fully understood and well
interpreted. This full interpretation happens through the intermediate linguistic
layers and the RDB MetaTable before going any further with the processing;
to avoid potential future errors or jeopardize accuracy. Computational linguis-
tics is used here in the form of linguistics-based mapping constraints using
manually written rule-based algorithms. Those manually written algorithms
are mainly observational assumptions summarised in Table 4 (Chapter 4).
Table 4 specifies RDB schema categories and semantic roles to map the identi-
fied RDB lexica into the SQL clauses and keywords.
Generally speaking, rule/grammar-based approaches [102] require exten
sive manual rules defining and customizing in case of any DB change to
maintain accuracy [140]. However, the rule-based observational algorithm
implemented in the current research work is totally domain-independent and
transportable to any NLQ translation framework. Generally, mapping is a com-
plicated science [14] because low mapping accuracy systems are immediately
abandoned by end users due to the lack of system reliability and trust. Hence,
this research work proposes a cutting-edge translation mechanism using com-
putational linguistics. However, there are several aspects of the proposed
research contribution which will be discussed in reference to the two mapping
algorithms in Figure 8 (Chapter 4).
36 NLP Application
MAPPING NLQ TOKENS

INTO RDB LEXICA
NLQ Tokens Extraction

In the current research work, NLQ tokens extraction and their types identifica-
tion happen through deep computational linguistics processes. The processes
are done via underlying NLP linguistic tools that use an English word seman-
tics dictionary (WordNet), RDB MetaTable (for the mapping algorithm) and
a mapping table for unique values namely, Primary Keys (PKs) and Foreign
Keys (FKs). The adopted linguistic method proved to be more accurate and
effective than other tokens extraction methods such as:
• Morphological and word group analyzers for tokens extraction [35],

• Pattern Matching [141] to identify keywords and their types,
• NER Recognizer [93] alone with the Coltech-parser in GATE [94]
to tokenize and extract NLQ’s semantic information,
• Java Annotation Patterns Engine (JAPE) grammars [88] for NLQ
tokenization and named entity tagging,
• Porter algorithm [97] to extract tokens’ stems,
• Unification-Based Learning (UBL) algorithm [98] which uses res
tricted lexical items and Combinatory Categorial Grammar (CCG)
rules [98] to learn and extract NLQ tokens,
• Dependency Syntactic Parsing [134] to extract tokens and their lexi-
cal relations,
• Dependency-Based Compositional Semantics (DCS) [92] system
enriched with Prototype Triggers [92], or
• Separate value and table extractor interfaces [29], which is a com-
promising approach for not supporting the RDB schema elements’
MetaTables and synonyms such as in the current proposed system.
RDB Lexica Mapping

Even recent studies in this field [142] failed to score high accuracy for the
tokens mapping algorithm or handling complex SQLs, despite using state-of-
the-art tools and techniques. An example of recent works in 2018 is Spider
[142], which does its mapping using a huge human labeled NLQ/SQL pairs
corpus as a training and testing dataset. Such datasets are created using com-
plex and cross-domain semantic parsing and SQL patterns coverage. However,
Spider’s performance surprisingly resulted in a very low matching and map-
ping accuracy. Hence, the current research work is distinct from most of
the previous language translation mechanism efforts because the focus here
gives highest priority to simplicity and accuracy of the algorithm’s matching
outcome.
The current research work employs the NLQ MetaTable (Table 1) to
map NLQ tokens into RDB lexica. The NLQ MetaTable covers NLQ words,
their linguistic or syntactic roles (noun, verb, etc.), matching RDB category
(table, value, etc.), generic data type (words, digits, mixed, etc.), unique as
PK or FK, besides their synonyms and enclosing source (i.e., tables or attri-
butes). MetaTables are used to check for tokens’ existence as a first goal, then
mapping them to their logical role as a relationship, table, attribute or value.
The general-purpose English language ontology (WordNet) are used to sup-
port the MetaTables with words’ synonyms, semantic meanings and lexical
analysis.
The implemented MetaTables fill up the low accuracy gap in language
translation algorithms that do not use any sort of deep DB schema data dic-
tionaries such as [81, 123], or just a limited data dictionary such as [43].
According to [19], WordNet is efficient at expanding NLQ predicate argu-
ments to their meaning interpretations and synonyms. However, WordNet
generalizes the relation arguments and does not guarantee NLQ’s lack of
ambiguity and noise, which significantly affects its meaning interpretation.
Hence, supportive techniques are employed in the current research work such
as the disambiguation module. In addition, to avoid confusion around the
RDB unique values, data profiling [121] is performed on large RDB’s statis-
tics to automatically compile the mapping table of unique values, PKs and
FKs, based on which RDB elements are queried more often. Mapping tables
are manually built for smaller RDBs, while using a data-profiling technique
to build them for larger RDBs. Unique values are stored in the mapping table
by specifying their hosting sources while a hashing function is used to access
them instantly.
RDB Lexica Relationships

NLQ parsing and dependency trees (i.e., Augmented Transition Network
(ATN) Lexical Relations Parser [143]) are used in the current research work as
part of the NLP semantic and syntactic parsing. Those parsers generate gram-
mar parse trees to explain the NLQ tokens’ dependencies and relations [144].
38 NLP Application
Besides, the RDB lexical join conditions are also discovered between any two
words or values. The joint is based on the words’ or values’ connectivity sta-
tus with each other or having common parent node in the dependency tree.
The parsing helps with the NLQ semantics extraction and RDB lexical data
selection. RDB elements relationships are controlled by using only verbs to
represent any connectivity in the RDB schema. The verbs’ parameters (subject
or object) are mapped with the RDB relationship’s corresponding elements:
tables, attributes or values. If the NLQ verb is unidentified or missing, the rela-
tionship between NLQ tokens will be found by analysing the matching RDB
lexica intrarelationships with each other.
There are other methods in the literature that identify lexical dependen-
cies and grammatical relations, such as Stanford Dependencies Parser [145],
Dependency Syntactic Parser [134] and Dependency-Based Compositional
Semantics (DCS) Parser [92]. The current research work used a simple way
of representing RDB elements inter-/intra-relationships. This representation
restricts the RDB schema relationships to be in the form of a verb for easy
mapping between NLQ verbs and RDB relationships.
NLP Syntax and Semantics Definition

The current research discovered common semantics between NLQ and SQL
structures by analyzing both languages’ syntax roles. In the same vein, under-
standing the NLQ, by finding the combination of its tokens’ meanings, is the
most essential part in the language mapping and translation process. Thus,
computational linguistic studies at the words processing level is employed as
opposed to approaches similar to Lambda Calculus or Tree Kernels Models
mentioned above in the fifth and sixth mapping approaches.
The current research framework overcomes any poor underlying linguistic
tools’ performance that are meant to analyse NLQ syntax and semantics. It over-
comes such inadequacies by using the supportive RDB schema knowledge and
semantic data models (MetaTables) and WordNet ontology. Furthermore, NLP
tools, such as NER tagger, tokenizer or dependency parser, are also employed
in addition to the syntactic-based analysis knowledge [52] to generate parse
trees from the identified tokens for proper mapping with the related RDB ele-
ments. Nevertheless, relying solely on NLQ’s syntactic and semantic analysis
for the mapping process is not sufficient and produces substantially low pre-
cision, False Positive Ratio (FPR) and True Negative Ratio (TNR) as indi-
cated in [1, 7, 46, 51, 57, 76, 115, 116, 131]. Such systems include the LIFER/
LADDER method in [45], SVM algorithm, or SVM-Light optimizer [146],
NLQ/SQL syntax trees encoded via Kernel Functions [24] or the Probabilistic
Context Free Grammar (PCFG) method [111] which proved to be challenging

in terms of finding the right grammar for optimization. Hence, the current work
supports the NLQ’s syntactic and semantic grammar analysis with computa-
tional linguistics in the form of RDB and NLQ MetaTables.
MAPPING RDB LEXICA

INTO SQL CLAUSES
SQL Clauses Mapping

While the current work uses computational linguistics mapping constraints to
transform RDB lexica into SQL clauses and keywords, [119] uses the extended
UML class diagrams representations [34, 104, 118]. Those representations are
used to extract fuzzy tokens’ semantic roles, which are imprecise terms or lin-
guistic variables consisting of fuzzy terms like ‘young’, ‘rarely’, ‘good’ or ‘bad’
[60, 83, 86, 103, 117, 137, 141]. Fuzzy tokens’ semantic roles are extracted in
the form of a validation sub-graph or tree of the Self Organizing Maps (SOM)
diagram representation [147]. SOM diagrams transform UML class diagrams
into SQL clauses [148] using the fuzzy set theory [49]. According to [119],
fuzzy NLQs with fuzzy linguistic terms are more flexible compared with pre-
cise NLQs as they provide more potential answers in the case of no available
direct answers. However, though this might provide higher measures of recall,
it is significantly compromising the FPR ratio. Hence, fuzzy NLQs are not
considered in the current research.
In [119], similar approaches to the current work are implemented for RDB
lexica mapping into SQL clauses. The work of [119] uses RDB relationships
to map the lexica into NLQs linguistic semantic roles’ classes as a conceptual
data model. However, since the current work uses supportive NLP linguistic
tools, it is more capable of “understanding” the NLQ statement before translat-
ing it into SQL query, which highly contributes to the increase in the translation
accuracy. Regarding the linguistic inter-relationships within the RDB schema
in the current work, not only WordNet is used, but also a Natural Language
Toolkit (NLTK) and NLP linguistic tools. In addition, a manual rule-based
algorithm is also used to define how NLQ linguistic roles match with the RDB
elements, which does not exist in [119] and which explains the variance in
translation accuracy and precision in comparison.
Regarding the linguistic analysis used in [119], the user has to identify
the fuzzy NLQ source (object), its associations or relationships and the target
40 NLP Application
(attribute) to connect them together for the UML class diagram extraction
phase. The results are derived from the connectivity matrix by searching for
any existing logical path between the source and the target to eventually map
them into an equivalent SQL template. In comparison, and since the current
research work aims for a seemingly natural HCI interaction, the user does not
have to identify any semantic roles in their NLQ. This is because the underly-
ing NLP tools does this for them. Also, the relationships are identified by the
NLQ verbs, so the user is communicating more information in their NLQ using
the current research algorithm compared to the other literature works. Hence,
it is considered more advanced and user-friendly than that in [119]. Also, not
only objects and attributes are extracted from the NLQ; the proposed research
work extracts much lower-level linguistic and semantic roles (i.e., gerunds and
prepositions) which help select the matching RDB lexica with higher accuracy
and precision.
Complexity vs Performance
The current research work is considered significantly simpler than most com-
plex mapping approaches such as [29] as it relies on fewer, but more effective,
underlying NLP linguistic tools and mapping rules. An example of a com-
plex language translation model is the Generative Pre-trained Transformer 3
(GPT-3) [30], introduced in May 2020. GPT-3 is an AI deep learning language
translation model developed by OpenAI [31]. GPT-3 is an enormous artificial
neural networks model with a capacity of 175 billion machine learning param-
eters [32]. Therefore, the performance and quality of GPT-3 language transla-
tion and question-answering models are so high [2]. GPT-3 is used to generate
NLP applications, convert NLQ into SQL, produce human-like text and design
machine learning models.
However, GPT-3 NLP systems of pre-trained language representation
must be trained on text, in-context information and big data (i.e., a DB that
contains all internet contents, a huge library of books and all of Wikipedia) to
make any predictions [31]. Furthermore, for the model training, GPT-3 uses
model parallelism within each matrix multiply to train the incredibly large
GPT-3 models [30]. The model training is executed on Microsoft’s high-band-
width clusters of V100 GPUs.
Training on such advanced computational resources would largely con-
tribute to the excellent performance of GPT-3 models. The biggest weakness
of this model is its extreme complexity, advanced technology requirements and
that it is only efficient once trained because GPT-3 does not have access to the
underlying table schema.
Algorithm simplification is necessary in such language-based applications.

An example of a complicated system is L2S [92] that compares all existing
NLQ tokens with existing DB elements. This approach consumes a lot of time
to run through all DB elements to compare them with the NLQ tokens. L2S
uses NLP tools, tokens’ semantic mapper and graph-based matcher, hence,
simplicity is key in the current work. Other examples of complicated systems
are in [29] and [24] where a hybrid approach is implemented using Bipartite
Tree-Like Graph-Based Processing Model [89], Ellipsis Method [82] and
the Highest Possibility Selection Algorithm [50]. Those approaches require
a domain-specific background knowledge and a thorough training dataset.
Hence, they are not considered in the current research work. The RDB lexica
into SQL clauses mapping algorithms in the literature ranged from simple to
complex methods.
After thorough study and research, it has become clear that the proposed
algorithm in the current research work is the best in terms of performance,
simplicity, usability and adaptability to different framework environments and
RDB domains. Both implemented mappers have access to an embedded lin-
guistic semantic-role frame schema (WordNet and Stanford CoreNLP Toolkit),
MetaTables and the RDB schema and MetaTables. Those resources are essen-
tial for SQL templates formation and generation which is a popular problem
under the NLP era.
The majority of NLQ into SQL mapping processes employ sophisticated
semantic and syntactic analysis procedures on the input NLQ [24, 79, 149–151].
However, those analyses are computationally expensive. Hence, the current
research work employs a lightweight approach for this type of query transla-
tions. In particular, the use of MetaTables which defines the lexicon seman-
tic role (i.e., noun, verb, etc.) and its adjacent SQL slot, prioritising accuracy
above complexity. Complex algorithms, such as the weighted links approach,
compromise accuracy for complexity. An example is in [76] that generates
ordered and weighted SQLs scheme using Weighted Neural Networks [39] and
Stanford Dependencies Collapsed (SDC) [152] as grammatical dependency
relations between NLQ tokens. This system is expensive to implement and
unscalable to bigger RDBs. It also prioritizes SQLs based on probability of
correctness instead of accuracy and precision.
However, what is interesting in [76] is their use of linguistics in their
algorithm where they identify NLQ’s subject or object to search the DB for
matching attributes. This matching uses Weighted Projection-Oriented Stems
[127] to generate the SQL clauses accordingly. Yet the translation accuracy of
this algorithm still falls behind the proposed algorithm in the current research
work. This is because the current work uses further linguistic categories (i.e.,
adjectives, pronouns etc.) in addition to using the verbs to find the attributes’
42 NLP Application
and values’ intra-relationships instead of using a heavy weighted tool such as

the Weighted Projection-Oriented Stems in [76].
Photon [153] is another neural network-based NLQ into SQL translation
approach that translates NLQs by parsing them into executable SQL queries.
Photon is the state-of-the-art NLIDB introduced to the public in June 2020.
It employs several modules as sublayers, such as a deep learning-based neu-
ral semantic parsing module, an NLQ corrector module, a DB engine and a
response generator module. For the neural semantic parsing layer, BERT and
a bi-directional Long-Short Term Memories (LSTM) machine learning algo-
rithms were used to produce hidden representations that match NLQ tokens
with table and attribute names. The NLQ corrector module detects untranslat-
able NLQs by highlighting an ambiguous or a confusion span around the token
and then asks the user to rephrase the NLQ accordingly. Although Photon rep-
resents a state-of-the-art NLIDB and an NLQ translation mechanism, it still
falls under the complex translation models while lighter weight translation
algorithms are sought for. In addition, the Photon model relies on training data-
sets to train its translation algorithm, table value augmentation module, static
SQL correctness checking module and neural translatability detector module.
Furthermore, deep learning and neural networks approaches, generally speak-
ing, tend to act as a black box where it is hard to interpret their predictions and
hard to analyse their performance and evaluation metrics.
An example of a simple mapping algorithm is SAVVY [154] that uses
pattern matching of DB query languages as a mapping algorithm. SAVVY
does not apply any NLQ interpretation modules or parsing elaborations for the
mapping process. Hence, its translation accuracy and overall performance is
highly jeopardized.
SQL Formation vs SQL Templates

The current research work simplifies SQL queries generation by using ready
SQL templates. Yet SQL construction constraints are used in the mapping
algorithm to guarantee accurate SQL template selection. Despite the pres-
ence of flexible SQL templates, some recent works [52, 91, 105, 107, 108,
120] still use other methods to construct their own SQLs from scratch, which
adds an extra computational complexity to the language translation system. An
example is [91], which defined a concept-based query language to facilitate
SQL construction by means of NLQ Conceptual Abstraction [88]. This work
adds an additional unnecessary complex layer on top of the original system
architecture.
Furthermore, [103, 106] used semantic grammar analysis to store all
grammatical words to be used for mapping NLQ’s intermediate semantic
representation into SQLs. Due to this system’s complexity, this architecture

can only translate simple NLQs, but not flexible with nested or cascaded SQLs.
In comparison, the current proposed system does not map whole NLQs into
existing SQLs, but maps NLQ lexica to the SQL clauses and keywords. This is
to enable the translator algorithm to be domain-independent and configurable
on any other environment, without the need of developing a training and test-
ing datasets of NLQ/SQL pairs for every new domain such as in [24].
In [24], a dataset of labeled NLQ/SQL pairs training and testing datas-
ets are generated and classified to correct or incorrect using Kernel Functions
and an SVM classifier. This mapping algorithm is at the syntactic level using
NLQ semantics to build syntactic trees to select SQLs according to their prob-
ability scores. Giordani and Moschitti [24] applies the statistical and shallow
Charniak’s Syntactic Parser [126] to compute the number of shared high-level
semantics and common syntactic substructures between two trees and produce
the union of the shallow feature spaces [24]. Such exclusive domain-specific
systems are highly expensive and their performance is subjective to the accu-
racy and correctness of the employed training and testing datasets, which are
manually written by a human domain expert. As such, the KRISP system [78]
achieved a 78% recall of correctly retrieved SQL answers, while the current
research work achieved a 96% recall on the small RDB (2.5 MB) and a 93%
recall on the large one (200.5 MB) due to the use of a light weighted mapping
algorithm mapped to ready SQL templates.
Another example of recent works that generate their own SQLs is [115],
which used syntactic trees of NLQ/SQL pairs as an SQL compiler to derive
NLQ parsing trees. In [115], NLQ tokens’ lexical dependencies, DB schema
and some synonym relations are used to map DB lexica with the SQL clauses
via a Heuristic Weighting Scheme [41]. Because [115] does not use any NLQ
annotated meaning resources (i.e., Prolog data [155] or Lambda Calculus
[129]) or any other manual semantic interpretation and representation to fully
understand the NLQ, the SQL generator performance was considerably low.
Therefore, authors of [115] applied a reranker [120] to try and boost accuracy
using an advanced Heuristic Weights-Based SVM-Ranker [36] based on String
Tree Kernels [128]. The reranker indeed increased the recall of correct answers
up to 94%, which is still lower than the recall of the proposed research work.
This is because in the current work, RDB lexica MetaTable is used for lexical
relations disambiguation. A mapping table is also used, which includes RDB
data types, PKs and FKs and names of entities (unique values), in addition to
other rule-based mapping constraints. Hence, building an SQL generator is
more complicated in the language translation field and as a result increases
the complexity of the translation algorithm. This is the main reason the current
research work uses SQL templates and puts extra focus on passing accurate
RDB lexia into SQL templates generator for a better performance and output.
44 NLP Application
Neural networks have not been used in the current research work, nor
for any of the mapping mechanisms. The reason why will be clearer with
some recent work examples such as [124, 156]. SEQ2SQL [124] is a deep
Sequence to Sequence Neural Network Algorithm [157] for generating an SQL
from an NLQ semantic parsing tree. SEQ2SQL uses Reinforcement Learning
Algorithm [124] and rewards from in-the-loop query execution to learn an
SQL generation policy. It uses a dataset of 80,654 hand-annotated NLQ/SQL
pairs to generate the SQL conditions which is incompatible with Cross Entropy
Loss Optimization [158] training tasks. This Seq2SQL execution accuracy is
59.4% and the logical form accuracy is 48.3%.
SEQ2SQL does not use any manually written rule-based grammar like
what is implemented in the current research work. In another recent work in
2019 [156], a sequence-to-sequence neural network model has been proved to
be inefficient and unscalable on large RDBs. Moreover, SQLNet [124] is a map-
ping algorithm without the use of a reinforcement learning algorithm. SQLNet
showed small improvements only by training an MLA sequence-to-sequence-
style model to generate SQL queries when order does not matter as a solution
to the “order-matters” problem. Xu et al. [124] used Dependency Graphs [116]
and the Column Attention Mechanism [159] for performance improvement.
Though this work combined most novel techniques, the model has to be fre-
quently and periodically retrained to reflect the latest dataset updates, which
increases the system’s maintenance costs and computational complexity.
The work in [157] overcomes the shortcomings of sequence-to-sequence
models through a Deep-Learning-Based Model [124] for SQL generation
by predicting and generating the SQL directly for any given NLQ. Then, the
model edits the SQL with the Attentive-Copying Mechanism [160], a Recover
Technique [3] and Task-Specific Look-Up Tables [161]. Though this recent
work proved its flexibility and efficiency, the authors had to create their own
NLQ/SQL pairs manually. Besides, they also had to customize the used RDB,
which is a kind of over-customization to the used framework and environment
applied on. Hence, results are highly questionable in terms of generalizability,
applicability and adaptability on other domains. On the other hand, the cur-
rent research work used RDBs that are public sources namely, Zomato and
WikiSQL.
Intermediate Representations
The current research work tries to save every possible information given by the
NLQ tokens so that each of them is used and represented in the SQL clauses
and expressions production. Therefore, multiple NLP tools, MetaTables and

mapping tables (for unique values) are implemented to fully understand the
NLQ and map its tokens to their corresponding RDB elements. Then, the
identified attributes are fed into the SQL SELECT clause, while the tables
are extracted from the SELECT clause to generate an SQL FROM clause, and
the values are used as conditional statements in the WHERE clause. For this
simple mapping purpose, other works in the literature use additional interme-
diate layers to represent NLQ tokens as SQL clauses, which, upon investiga-
tion, turned to be not as effective as the NLP tools, MetaTables and mapping
tables.
An example of the NLQ intermediate semantic representation layers is
using Regular Expressions (regexps) [101] to represent NLP tokens phonol-
ogy and morphology. This representation happens by applying First Order
Predicate Calculus Logic [162] using DB-Oriented Logical Form (DBLF) and
Conceptual Logical Form (CLF) with some SQL operators and functions to
build and generate SQLs [107]. Yet, the use of regular expressions “regexps”
collections in NLQ sentences are not clearly articulated in the literature.
Another example is CliniDAL [27], which used EAV type of DB metadata
and grammatical parse trees to process NLQ tokens to then be mapped to their
internal conceptual representation layer using the Similarity-Based Top-K
Algorithm [138]. More examples of the intermediate layers include PHLIQAl
[4], ASK [46], USL [109], and SEMRESP [7, 129, 111, 122] which defined
intermediate tokens meaning representations manually.
Also, the supervision-extensive system [7, 113, 126] used Lambda
Calculus to map tokens to their corresponding meaning representations. TEAM
[16] adopted intermediate semantic representation layers connected with a DB
conceptual schema. In addition, L2S [92] transforms DB lexica into an inter-
mediate tree-like graph then extracts the SQL from the Maximum Bipartite
Matching Algorithm [163].
All those great efforts unfortunately proved to be not as effective because
of the high complexity and the time-consuming nature of the approaches.
Those deficiencies mandated the introduction of a new system that translates
NLQs into SQLs while maintaining a high simplicity and performance pre-
sented in the current research. The proposed research highlights a new solution
to NLP and language translation problems.
Table 17 (Appendix 9) highlights the main similar works in the literature
with their advantages and disadvantages summarized in comparison with the
proposed work in the current research.
In what follows, a layout of the similar works presented in a chronologi-
cal order. This section is also summarized in Table 17, Appendix 9 (Figure 5).
46 NLP Application
NLQ into SQL mapping Approaches

•Authoring Interface Based Systems
•Enriching the NLQ/SQL Pairs via Inducve Logic Programming
•Using MLA Algorithms
•Restricted NLQ Input
•Lambda Calculus
•Tree Kernels Models
•Unified Modeling Language (UML)
•Weighted Links
NLQ Tokens into RDB Lexica Mapping (NLQ Tokens Extracon)
•Morphological and Word Group Analyzers
•Paern Matching
•Name Enty Recognizer (NER) Alone with Coltech-Parser in GATE
•Java Annotaon Paerns Engine (JAPE) Grammars
•Porter Algorithm
•Unificaon-Based Learning (UBL) Algorithm
•Dependency Syntacc Parsing
•Separate Value and Table Extractor Interfaces
NLQ Tokens into RDB Lexica Mapping (RDB Lexica Mapping)
•Spider System
•WordNet alone
NLQ Tokens into RDB Lexica Mapping (RDB Lexica Relaonships)
•Stanford Dependencies Parser
•Dependency Syntacc Parsing
•Dependency-Based Composional Semancs (DCS) System Enriched with Prototype Triggers
NLQ Tokens into RDBComposional
•Dependency-Based Lexica Mapping (NLP(DCS)
Semancs syntax and Enriched
System semancs)with Prototype Triggers
•Named Enty Tagger
•Dependency Parser
•LIFER/LADDER Method
•NLQ/SQL Syntax Trees Encoded Via Kernel Funcons
•The Probabilisc Context Free Grammar (PCFG) Method
RDB Lexica into SQL Clauses Mapping (SQL clauses mapping)
•The Extended UML Class Diagrams Representaons
•RDB Relaonships and Linguisc Analysis
RDB Lexica into SQL Clauses Mapping (Complexity vs Performance)
•L2S System
•Biparte Tree-Like Graph-Based Processing Model
•Ellipsis Method
•The Highest Possibility Selecon
•Weighted Neural Networks and Stanford Dependencies Collapsed (SDC)
•Paern Matching of SQL
RDB Lexica into SQL Clauses Mapping (SQL Formaon vs SQL Templates)
•NLQ Conceptual Abstracon
•Semanc Grammar Analysis
•Kernel Funcons, SVM Classifier, and the Stascal and Shallow Charniak’s Syntacc Parser
•Heurisc Weighng Scheme
•A Deep Sequence to Sequence Neural Network
•MLA Sequence-To-Sequence-Style Model
•A Deep-Learning-Based Model
RDB Lexica into SQL Clauses Mapping (Intermediate Representaon)
•Regular Expressions (regexps)
•The Similarity-Based Top-K Algorithm
•Lambda-Calculus
•An Intermediate Tree-Like Graph
FIGURE 5 Research ideas with their current existing solutions.

Implementation
Plan 4
The current research framework components start with the NLIDB interface
that the user uses to enter the NLQ sentence. To understand the input NLQ,
the NLQ must go through POS recognition via the underlying NLP tasks, such
as lemmatizing, tokenizing, annotating and rule-based parsing. The following
step, disambiguation, is a conditional step that the NLQ will go through only
in the case that there was a vague POS word (i.e., Is “content” a noun or an
adjective? Is “separate” a verb or an adjective?). After that, the identified NLQ
tokens will be delivered to the matcher/mapper step for mapping the tokens
into the elements of the RDB schema MetaTables and the identified lexica into
the SQL clauses. The matching lexica will be used in the SQL generation step,
which will be executed next.
NLQ INPUT INTERFACE

Before running any script, required dependencies and packages, which are
all open source and available, must be downloaded and imported through the
Python terminal. Then, the user will insert an NLQ into the data input interface.
The user will be returned either the generated SQL results from the MySQL
DB or an error alert. The error alert could be concerning the entered NLQ lin-
guistic issues or an error of an existence of more than 1 match or no match at
all to the NLQ arguments in the RDB MetaTable.
The MetaTables of NLQ and RDB are created by adding span tags to
the RDB elements or the NLQ tokens to attach them with their syntactic and
semantic roles. They are also annotated with their synonyms using the WordNet
ontology functions. The cost of adding the MetaTables data are fractional to the
original size of the RDB itself. For small RDBs, it could add extra 3% on top
of the original RDB size. For larger RDBs, such as WikiSQL, it could add up
to 10% extra storage space. When the RDB changes, the translation processing
DOI: 10.1201/b23367-4 47
48 NLP Application
NLQ Input
Part of Speech Recognion

Lemmazing text using a Tokenizer, Annotator, Semanc or
Syntacc (rule-based) Parser.
Disambiguaon
e.g. Is Adam a paent or a physician?
Matcher/Mapper
Matching NLQ tokens with RDB schema metatables (lexicon)
and the general-purpose English language ontology.
SQL Template Generator
SQL Execuon
Result
FIGURE 6 Detailed research architecture pipeline.
is not affected by the update because every added record will be automatically
annotated by the framework to include necessary annotations and metadata.
The NLQ is inserted through an NLI screen as input data up to 200 charac-
ters with the help of two Python libraries, namely, “server.bot”, which accepts
the input NLQ, and “text_processing.text_nlp”, which initially processes the
NLQ by separating the words and passing them as arguments to the next mod-
ule. The NLI will identify NLQ words as arguments, which will later help pre-
paring them for identifying their semantic and syntactic roles. Figure 6 briefly
summarizes the steps taken to transform an NLQ into an SQL statement. Those
steps will be further clarified throughout this chapter.
POS RECOGNITION
The multilayered translation algorithm framework splits the NLQ into its con-
stituent tokens. Then, these tokens are compared with the RDB MetaTables’
4 • Implementation Plan 49
contents to single out keywords in the NLQ sentence. With the tokens match-
ing schema data, a.k.a. the lexica, the NLQ should be able to be parsed seman-
tically to identify tokens’ semantic-role frames (i.e., noun, verb, etc.) which
helps the translation process. Semantic parsing is done by generating the pars-
ing tree using the Stanford CoreNLP library, with input from the English lan-
guage ontology, WordNet, which feeds the system with NLQ words meanings
(semantics).
The first process performed on the NLQ string is lemmatizing and stem-
ming its words into their broken-down original root forms. This is done by
deleting the words’ inflectional endings and returning them to their base forms,
such as transforming ‘entries’ into ‘entry’. Lemmatizing eases the selection
and mapping process of equivalent RDB elements. It also facilitates the tokens’
syntactic and semantic meaning recognition. Then comes the steps of pars-
ing and tokenizing the words’ stems into tokens according to the predefined
grammatical rules and the built-in syntactic roles. Those syntactic roles will be
mapped to specific RDB elements, for instance, NLQ verbs are mapped with
RDB relationships.
PSEUDOCODE 1 ALGORITHM TO CONSTRUCT AN SQL

QUERY FROM AN NLQ INPUT
Begin
Split NLQ text to individual ordered words and store
into string array A
Delete any escape words from A
Map words in array A with RDB elements E
Replace words in array A by their matching synonyms
and type from E
If there is ambiguate word W in A then
Ask user “What is W?” and match word W with E
End If
If there is a conditional phrase C in A
Replace C with equivalent conditional operator
in O
Attach O to conditioned attribute name as a
suffix and store in A
End If
Do
50 NLP Application
Store attributes and their conditional

operators and tables and relationships for matched
elements E in array R
Generate SQL template matching the number and
type of tokens in R
Construct SQL query using array R tokens and
store it in variable Q
While for each table or attribute or relationship or
conditional operator in array R
Execute generated SQL query
End
For any NLQ translation process, both the parsed tokens and their subse-
quent POS tags must be clearly and accurately identified. This is performed by
an underlying multilayered pipeline which starts with tagging an NLQ POS.
Then, the tokenizer, annotator, semantic and syntactic (rule-based) parsers will
be applied and any punctuation marks will be removed. Part of this step is
omitting the meaningless excess escape words that are predefined in the system
(i.e., a, an, to, of, in, at, are, whose, for, etc.) from the NLQ words group. After
parsing, a parse tree is generated and a dictionary of tokens’ names, syntactic
roles and synonyms are maintained in the NLQ MetaTable. Also, the NLQ’s
subjects, objects, verbs and other linguistic roles are identified. Hence, each
tokenized word is registered into the NLQ MetaTable by the syntactic analyzer.
Tokens are then passed to the semantic analyzer for further processing.
The semantic analyzer employs a word-type identifier using a language
vocabulary dictionary or ontology such as WordNet. The word-type identi-
fier, such as WordNet, identifies what semantic role does a word or a phrase
(i.e., common or proper noun) play in a sentence and what is their role
assigner (the verb). Furthermore, the semantic analyzer is able to identify
conditional or symbolic words and map them with their relative represen-
tation from the language ontology. For example, the phrase “bigger than”
will be replaced by the operator “>”. In other words, the semantic analyz-
er’s entity annotator detects the conditional or symbolic words amongst the
input NLQ entities. Then, it replaces them with their equivalent semantic
types identified previously by the schema annotator. The entities replace-
ment creates a new form of the same NLQ that is easier for the SQL genera-
tor or pattern matcher to detect.
The entity annotator is not the only annotator the NLQ deals with. There
are other annotators the NLQ gets passed through such as the numerical
annotator, date annotator, comparator annotator, etc. In future work, this

step can be further improved to search for the previous annotation results to
check for any stored matching patterns of lexicalized rules. This step shall
help determine the suited SQL template type or its further sub-queries’ divi-
sions faster.
The NLQ gets converted into a stream of tokens and a token ID is pro-
vided to each word of the NLQ. The tokens are classified into their linguistic
categories such as nouns, pronouns, verbs or literal values (string/integer vari-
ables). The algorithm maps the tokens into tables, attributes, values or relation-
ships according to their linguistic categories and semantic roles. The rest of
the acquired information will be used to formulate SQL query clauses (i.e.,
comparative or operational expressions) according to the tagged tokens.
The Python NLP lightweight library (TextBlob) is used as an NLQ POS
recognizer (i.e., “speech_recognition” library). NLQ tagger and lemmatizer
are implemented to facilitate the equivalent RDB elements selection. In addi-
tion, other libraries are also considered including, but not limited to, “string_
punctuation”, “Stanford CoreNLPspellcheck”, “nltk.corpus” (using WordNet),
“textblob.word”, “wordNetLemmatizer”, “nltk.stem”, “sentence_tokenize”,
“word_tokenize”, “nltk.tokenize”, “unicodedata” (for mathematical operations
and symbols), “textt_processing”, “text_nlp” and “server.tokenizer”. Those
libraries’ usage and distribution is explained in Figures 7 and 16 (Chapter 6).
The system checks each NLQ word’s semantic role and adds it to the
registry to be passed on to the next step, as illustrated in PseudoCode 2
(Appendix 1). For example, if the first element in the list (index[0]) is a com-
mon noun, the code would check if the NLQ word is a table. Also, if there
is a corresponding attribute to a value, add the word to the values list, and
so on. The algorithm applies the ‘Maximum Length Algorithm’, illustrated in
PseudoCode 3, to remove tokens from the attributes list if the tokens are also in
the values list. This algorithm enables the system to avoid duplicate use of the
same tokens, which helps in avoiding potential errors and inaccuracy.
PSEUDOCODE 3 MAXIMUM LENGTH ALGORITHM
for values(a, v)
if a ϵ attributes()
remove a from attributes()
end if
end for
52 NLP Application
FIGURE 7 Detailed research organization pipeline (light gray boxes are Python
libraries; dark gray boxes are tasks & gray blobs are passed-on data).
To recognize literal values, the functions “parse_decimal” and “Number

FormatError” are used from the “babel.numbers” library. “parser” and
“WordNetLemmatizer” functions from the “nltk.stem” Python library are used
to insert the RDB elements’ synonyms into the RDB schema automatically.
This happens by adding the synonym and then creating the relationship
‘IS_LIKE’ with the corresponding RDB element as illustrated in PseudoCode 4.
PSEUDOCODE 4 SYNONYMS MATCHING WITH RDB

ELEMENTS
insert_synonyms()
for s ϵ synonyms and e ϵ elements
if s is similar to e and similarity > 0.75 then
merge (s, e) as (s)-[IS_LIKE]->(e)
end if
end for
During the tokenizer and tagger stage, illustrated in PseudoCode 5

(Appendix 2), NLQ POS, semantic analysis and syntactic representation are
expressed and extracted by tagging them. This helps to extract the tokens’ lin-
guistic sub-parts (i.e., adjectives and noun/verb phrases) for accurate mapping
later. The POS tagger will tag each NLQ token to define its syntactic role. The
tokenizer analyzes the NLQ token types discussed in Table 4 and returns the
extracted sentence structure parts (i.e., verbs, noun phrases) mentioned in the
NLQ, together with the tagged version of the lemmatized tokens. In addition,
tokens’ semantic roles under the SQL scope, which is a value, attribute, table
or a relationship, will also be returned.
Not only RDB elements are tagged with their synonyms, but SQL key-
words are also tagged with synonyms and semantic information. This tagging
happens using the semantics dictionary (WordNet) and the “nltk.tokenize”
libraries, namely, “sentence_tokenize” and “word_tokenize”, as illustrated in
PseudoCode 6.
PSEUDOCODE 6 SQL KEYWORDS TAGGING WITH THEIR

SYNONYMS
sql_tagging()
for attributes and tables and conditions
if sql ≠ Ø then
apply semantics_dict[synonyms]
add attributes synonyms to select
add tables synonyms to from
add conditions synonyms to where
end if
end for
return sql_tagging(tags)
54 NLP Application
Other SQL keywords such as aggregate functions (e.g., AVG, SUM, etc.)
or comparison operations (e.g., >, <, =, etc.), defined in the Python “unicode-
data” library, are also tagged with their synonyms for easy and accurate map-
ping, as illustrated in PseudoCode 7 (Appendix 3).
After all of the NLQ and SQL words are tagged with their synonyms, the
algorithm will start the testing module to validate the similarity of RDB Lexica
and NLQ tokens compared with their tagged synonyms. If the “Similarity”
is greater than or equal to 75% (the least-acceptable similarity variance), it
is considered a matching synonym. In this case, lexica or tokens are tagged
with their matching synonyms according to their semantics using WordNet
synonym datasets.
DISAMBIGUATION
NLQ input disambiguation is an intermediate process and is done through con-
textual analysis. When the system cannot make a decision due to some ambi-
guity, it asks the user for further input. This occurs in case of the presence of
more than one match for a particular NLQ token (e.g., “Is Adam a patient or
a physician?”). However, engaging the user is solely for clarifying a certain
ambiguity in the NLQ input by choosing from a list of suggestions of similar
words or synonyms present in the lexica list.
In future work, and as a further disambiguation step to guarantee gener-
ated SQL accuracy, a feedback system could be applied after NLQ analysis.
This feedback system asks the user to confirm the translated NLQ into SQL
query by asking the user “is this the desired SQL?”. However, since we assume
the user’s ignorance of any programming abilities, including SQL, this feed-
back system is not applied in the current research work.
The RDB elements with identical names are carefully managed according
to the NLQ MetaTable (Table 1) and the RDB elements’ MetaTable (Table 2).
Hence, the ambiguity-checking module will eventually have a list of all identi-
cally named elements and their locations in the RDB.
Every entered NLQ goes through a syntactic rules checker for any gram-
matical mistakes. This module checks the NLQ validity or the need for a user
clarification for any ambiguity or spelling mistakes using the Python libraries
“unittest” and “textblob”. The algorithm will proceed to the next step if the
NLQ is valid. Otherwise, the algorithm will look for a clarification or spell-
ing correction response from the user by asking them to choose from a few
potential corrections. Then, the user’s response is classified to either positive
(i.e., Yes) or Negative (i.e., No). This classification happens using the Naïve
Bayes Classifier from the prebuilt Python library “textblob.classifiers” used

as an SQL grammar classifier. If the user’s response is positive, it will use the
corrected form, otherwise, it will use their original NLQ form and work with
it. This step uses the “Stanford CoreNLP” and “nltk.corpus” libraries to check
for the NLQ validity and uses the syntactic rules checker to check for spelling
errors as illustrated in PseudoCode 8 (Appendix 4).
MATCHER/MAPPER
In this phase, synonyms of NLQ tokens are replaced with their equivalent
names from the embedded lexica list. Then, SQL keywords are mapped and
appended with their corresponding RDB lexica. The Matcher/Mapper mod-
ule applies all mapping conditions listed later in Table 4 which covers NLQ
tokens, their associated RDB lexica, SQL clauses, conditional or operational
expressions or mathematical symbols.
This module has access to MetaTables (data dictionaries) of all attributes,
relationships, tables and unique values (Mapping Tables). Both mappers in
Figure 8 can refer to an embedded linguistic semantic-role frame schema, data
or language dictionary, or the underlying RDB schema. This layer uses RDB
schema knowledge (the semantic data models, MetaTables) and related syn-
tactic knowledge to properly map NLQ tokens to the related RDB structure
and contents.
In regard to unique RDB values, and since it’s a storage crisis to store all
RDB values in a RAM or CACHE memory, only unique values and PKs and
FKs will be stored in a mapping table. The unique values’ hosting attributes
and tables will be specified, and a hashing function will be used to access them.
For smaller RDBs (i.e., Zomato), and, as explained in Table 3, the mapping
table is built using the Python dictionary “server.map” that finds associations
Stemmed
• Linguisc Analysis RDB Elements
NLQ Tokens
RDB Matching • Rule-Based Mapping

SQL Clauses
Lexica
FIGURE 8 The two embedded mappers.

56 NLP Application
TABLE 3 Mapping table design options

RDB SIZE EXAMPLE APPLIED TECHNIQUE DATA USED
Small Zomato Python dictionary Associations between NLQ
“server.map” tokens and RDB elements
Large WikiSQL Data profiling RDB elements’ statistics
between NLQ tokens and RDB elements that are often queried together. For
larger RDBs (i.e., WikiSQL), data profiling is performed on RDB elements’
statistics to automatically compile the mapping table. This compilation is
based on which RDB elements are queried more often, and then stored in
the mapping table as a hashing function. The mapping table is expressed as
mapping_table[unique_value] = corresponding_attribute.
Compared to the great value the mapping table adds to the algorithm’s
accuracy, there would not be any significant overhead added by integrating
a mapping table. Ye, the bigger the RDB the bigger the mapping table size,
which affects resources usage in terms of storage capacity.
Mapping NLQ Tokens into RDB Elements

This unit matches NLQ tokens with RDB schema MetaTables (lexica list) to
check for their existence. This unit also checks the general-purpose English
language ontology (WordNet) for NLQ tokens’ synonyms and meanings.
Before discussing the mapping algorithm itself, the RDB schema relationships,
its lexica (tables’ and attributes’ names) and the conditional and operational
expressions must be defined.
First, relationships in the RDB schema will be defined and registered.
Then, they will be matched with the NLQ verb. If the NLQ verb is unidentified
or missing, the relationship between the NLQ tokens will be found through
analysing the lexica intra-relationships with each other as explained step by
step in PseudoCode 9.
PSEUDOCODE 9 NLQ RELATIONSHIPS DEFINITION
/* register relationships */
for attributes in rdbSchema do
for attribute1(lexicon1, attribute1) and
attribute2(lexicon2, attribute2) do
relationships ← relation(attribute1, attribute2)
end for
end for
/* if NLQ has no verbs */
if nlq(verb) = True
check relationships(synonyms)
else
check relation(attributes) in relationships
end if
Now that RDB relationships have been defined and registered, the algo-
rithm is able to retrieve matching RDB lexica, and their hosting attribute or
table. This matching happens in accordance with the matching NLQ lexica and
the relationships built between them. The retrieved data will be then passed
on to the next step to be used in the SQL clauses mapping as explained in
PseudoCode 10.
PSEUDOCODE 10 FINDING RDB LEXICON PARENT

ATTRIBUTE, TABLE AND RELATIONSHIPS
for rdbSchema(lexicon) do
find parent and relationship
return parent(attribute), parent(table),
Relationship(verb)
NLQ tokens are mapped with their internal representation in the RDB
schema via the MetaTables and synonyms, and then mapped to the SQL
clauses. Each input token is mapped with its associated RDB element (lexicon)
category (e.g., value, column, table or relationship).
The mapper translates the NLQ literal conditions and constraints, whether
they are temporal or event-based, into the SQL query clauses such as translat-
ing “Older than 30” to “Age > 30”. The mapper also extracts matches of func-
tion or structure words (i.e., linking words or comparison words) and search
tokens (i.e., Wh-question words) from the annotated NLQ. Function words
could be prepositions (i.e., of, in, between, at), pronouns (i.e., he, they, it),
determiners (i.e., the, a, my, neither), conjunctions (i.e., and, or, when, while),
auxiliary (i.e., is, am, are, have, got) or particles (i.e., as, no, not).
This module checks for the presence of any NLQ conditional, operational
or mathematical expressions (i.e., min, max, avg, etc.) in the NLQ to custom-
ize the WHERE statement accordingly to retrieve only relevant data from the
RDB, as explained in PseudoCode 11.
58 NLP Application
PSEUDOCODE 11 CHECKING NLQ FOR EXPRESSIONS
if nlq(words) ← expr(cond, oper, math)

where_clause = True
adjust where_clause with conditional[] or
operational[] or mathematical[]
else
where_clause = False
end if
TABLE 4 Main rule-based assumptions

SCHEMA
NLQ TOKEN TYPE EXAMPLE CATEGORY SQL SLOT
Instance (Proper Sarah Value WHERE condition
Noun)
Adjective, Adverb, Strongly Attribute SELECT or WHERE
Gerund (if accompanied with a
value) clause
Number (Literal 100 Value WHERE condition
Value)
Common Noun Patient Attribute/Table SELECT/FROM selection
operator clause
Comparative Most Conditional MAX, MIN, AVG, etc.
Expression Values clauses or with WHERE
clause
Comparative Equal Conditional =, >, <, <>, ><, >=, <=,
Operation Values etc. with WHERE clause
Verb has Relationship WHERE condition, JOIN,
AS, or IN
Wh-phrases What Value’s Attribute N/A
Indicator
Prepositions With N/A N/A
Conjunction/ And N/A WHERE condition AND,
Disjunction OR, etc.
During the parsing phase, the NLQ is decomposed into a head-noun,

noun modifiers, verbs that relate semantic roles together, objects and relation-
ships descriptivism and adjectives or adverbs that describe verbs. Hence, NLQ
tokens can be any of the token types in Table 4.
According to those observation-based assumptions summarized in Table 4,
an SQL template can be easily generated. The SQL template generator mainly
needs to know the number of attributes, tables and relationships. In addi-

tion, further information are fed to the SQL generator, such as the AND/OR
clauses (for JOIN clauses), conditional comparative expressions (for WHERE
or AGGREGATE clauses), the conditional comparative operations (for
INTERVAL clauses or controlled values) and numbers and instances (literals)
as values. For example, if the token is a value, then the corresponding attribute
(object) is its column name.
Moreover, synonyms of SQL clauses are also considered. For example,
‘search’, ‘show’, ‘find’, ‘get’ or even the word ‘select’ are all synonyms of the
SQL clause “SELECT”. Similarly, ‘count’, ‘how many’ or ‘how much’ are
synonyms of the “COUNT” statement. Also, ‘where’, ‘who has’ or ‘with’ are
synonyms of the “WHERE” clause.
Since the RDB lexica may not be explicitly used in the NLQ, the matcher/
mapper unit tries to match an NLQ token to an equivalent RDB lexicon by
comparing every token (and its synonyms) to its potential RDB element (or its
synonyms). In addition, NLQ verbs will also be matched with their equivalent
RDB schema relationships to locate where in the RDB schema is this token
being referenced to. If a match is found, the algorithm replaces the token with
the matching lexicon and returns the match in the form (table, attribute, value,
relationship) with each element surrounded by span tags. If the token is found
to be an RDB value, the attribute and subsequent table will be known auto-
matically. This step uses the Python library “server.mapper” as explained in
PseudoCode 12 (Appendix 5).
Dependency trees, derived from the Stanford CoreNLP syntactic trees, are
used to explain the relationships between any two values based on their con-
nectivity status or having common parent node in the dependency tree.
RDBs illustrate relationships based on the data types. As such, values and
attributes existing in the same entity (column) are related, so as attributes’
tables or tables connected with a particular relationship.
In the current research work, verbs will be mapped to associated relation-
ships, and the verbs’ parameters (subject or object noun phrases) into their cor-
responding attributes in the lexica list. Identifying this relationship association
proved to increase the corresponding attributes selection accuracy. Thus, the
tokens list is defined in this phase by the lexical analyzer using the language
ontology WordNet, and eventually replaced by the RDB MetaTable lexica and
passed to the syntactic analyzer.
Mapping RDB Lexica into SQL Clauses

This mapping uses the proposed rule-based algorithm that is based on the
assumptions table (Table 4).
60 NLP Application
The first step is building the main SQL clauses, the SELECT, FROM and
WHERE clauses. The attribute names will be fed into the SELECT clause.
Hence, the SELECT keyword is appended with the table attributes. Attributes
are identified by semantically analysing the Wh-word’s main noun phrase or
head noun (main noun in a noun phrase). The WHERE keyword is mapped
with the attribute-value pairs derived from the NLQ semantics. The FROM
keyword is mapped with all involved tables’ names referenced in the SELECT
and WHERE clauses. If there is more than one table, tables will be joined and
added to the FROM clause. If there is a data retrieval condition, a WHERE
clause will be added, and conditions will be joined as illustrated in PseudoCode
13 (Appendix 6).
In this phase, the key mapping function is mapping SQL clauses and key-
words with the NLQ identified lexica, and then building the SQL query. The
tables list which tables names should be selected from must be identified. The
list of relationships, attributes and values with their associated attributes should
also be identified in the form (attribute, value).
SQL TEMPLATE GENERATOR

SQL formation is done in this stage. SQL components (i.e., tables’ names, lex-
ica hosting sources, attribute-value pairs, data retrieval conditions and relation-
ships) are identified from the input NLQ and arranged in a proper sequence.
The identified NLQ lexica, schema matching elements, and the identified
operators (if any) are then fed into the SQL template generator to generate a
proper SQL statement. SQL templates will be selected based on the numbers
of identified tables, attributes and attribute-value pairs. After that, the system
establishes a connection with the RDB to transfer SQLs to the RDBMS for
execution.
The RDB schema contains unique identifiers (e.g., PKs and FKs) list
stored in a dependency table (the mapping table). This table uniquely identifies
each instance of each attribute, and whether they are connected via a relation-
ship with any other RDB elements. Each attribute’s unique identifier is added
to the SQL query constraints to guarantee that only the particular information
of interest is returned.
The SQL templates list of SELECT statements considers possible SQLs
depending on the NLQ question and desired answer using all input from previ-
ous steps. For NLQs with explicit SELECT parameters, the proposed assump-
tions-based system uses an investigative heuristic procedure to determine what
parameters belong in which SQL slot. The parameters can be used either as
query constraints (e.g., WHERE, IN, etc.) if they already have values, or as
part of the SELECT statement; if they need their values to be retrieved from the
RDB. After that, the input values and necessary operators are used to construct
the query constraints in the proper SQL template. An example of assigning a
suitable operator for every WHERE conditional pair (attributes and values) is
converting the NLQ string “equal” to the SQL keyword “LIKE” or the operator
“=” or converting “smaller or equal” to the operator “<=”.
In this work, only the following SQL main clauses are considered, in addi-
tion to other supplementary clauses (e.g., AS, COUNT, etc.):
• SELECT: identifies desired attributes to be retrieved according to

the NLQ processor.
• FROM: identifies the tables where the SELECT attributes are from,
or where the attribute-value pairs appearing in the WHERE con-
ditional statements are originally from. In case of multiple tables,
relationships between tables are identified using JOIN.
• WHERE: identifies the conditions and criteria that must be applied
on the retrieved data in a form of conjunctions of attributes and their
desired values. If there is more than 1 table, JOIN conditions are
used.
Furthermore, it is important to determine SQL classes which a system can

or cannot generate. All adopted SQL queries are simple, covering SELECT
and WHERE clauses. Therefore, the supported SQL statements are declared in
Figure 9 (Chapter 5).
The chosen SQL template solely relies on the number of NLQ tokens
related to tables, attributes and values. Yet the type of the generated SQL tem-
plate could be nested, aggregated, negated, or basic selection, joining and pro-
jection. Those types are further categorized in Appendix7.
The Python libraries “string.template” and “server.sql_templates” are used
to construct and generate SQL statements in the form: SELECT {attributes}
FROM {table} [, {table}] (WHERE {attribute=value} [and {attribute=value}]).
The algorithm uses default template strings (placeholders) until it receives the
selected lexica from the previous steps, particularly from the Matcher/Mapper
step. This process performs in accordance with Table 4 tokens mapping rules
and as illustrated in PseudoCode 14 (Appendix 8). It is important to note that
all SQL templates use the DISTINCT keyword as per the embedded Maximum
Length Algorithm explained earlier.
The first step in the SQL template generator module is connecting the
SQL templates generator environment to the MySQL server and the MySQL
DB session via the “mysql.connector” function. The SQL template generator
will choose which template should be chosen to generate the query. This SQL
62 NLP Application
template election is based on the required number of involved tables, attributes

and attribute-value pairs derived from the RDB schema using MySQL DB and
the Python libraries “pymysql” and “server.sql”. After selecting the right tem-
plate, the algorithm will return the generated SQL statement with the lexica
inserted appropriately. Then, the query is pushed forward to the MySQL server
for execution on the connected RDB as illustrated in PseudoCode 14.
PSEUDOCODE 14 SQL GENERATION AND EXECUTION
if generated_sql = True then

execute(sql)
else print(Sorry, there were no results for your query!)
end if
SQL EXECUTION AND RESULT

After the generated SQL query execution, data is fetched from the RDB and
displayed to the user as raw data. An example of usage is the following query
line entered into the Python command line interface.
python3 -m nlqsql.main -d zomato/city.sql -j output.json -i

‘What is the average size of restaurants with name Burrito?’
The output would be:
{‘select’: {‘attribute’:‘size’, ‘type’:‘AVG’},

‘from’ : {‘table’:‘restaurant’},
‘where’: {‘conditions’:[{‘attribute’:‘name’,
‘operator’:‘=‘, ‘value’:‘Burrito‘},]},}
And the execution result is:
-----------------
| AVERAGE (*) |
-----------------
| 23 Square Feet |
----------------
Implementation
User Case
Scenario
5
To match an NLQ to a proper SQL template, NLQ text will be analyzed and
tokenized to be matched against the RDB index. The NLQ goes through a full-
text search after it has been tokenized, which is different from the common
keyword search.
Figure 9 shows the directed RDB chart diagram for the Post-Traumatic
Stress Disorder (PTSD) RDB. RDB representation is used here instead of
ERD because the RDB relationships are richer in information than ERDs [13].
Tables 5–8 are the RDB tables. Table 9 is the NLQ MetaTable and Table 10 is
the PTSD RDB elements MetaTable that stores all the entities in the RDB and
their metadata.
All definitions of the RDB entities are stored in Table 10 to describe the
tables and attributes. Using Tables 9 and 10, the current automatic mapping
algorithm can produce considerably accurate mapping results.
RDB keywords related to different tables and attributes are stored together.
Hence, the algorithm is able to map the NLQ tokens to their internal represen-
tation of source attributes and tables in the RDB. To reduce ambiguity, the
relationships between attributes are controlled in the RDB design to be in the
form of verbs only (Figure 10).
Table 10 stores all of the definitions of RDB entities. This describes the
tables, attributes, and unique values. Using Table 9, the current automatic
mapping algorithm can produce more accurate mapping results since it is
able to map the NLQ tokens to their internal representation of source attri-
butes and tables in the RDB. This is because all DB keywords related to
different tables and attributes are stored together. To reduce ambiguity, the
relationships between attributes are controlled in the RDB design to be only
verbs.
DOI: 10.1201/b23367-5 63
64 NLP Application
SQL
SELECT FROM WHERE
One Mulple All Disnct Aggregate One Mulple Disjuncon Cross- Aggregate
Column Columns Columns Select Funcons
Tables
Condion
Nested
Condions Juncon Condion
Operators
Funcons
Count- Equal Sum In

Select Operator Condion
Sum- Not Equal Avg In

Select Operator Condion
Greater-
Than Min In
Avg-Select
Operator Condion
Less-Than Max In
Min-Select
Operator Condion
Max- Like
Select Operator
Between
Operator
FIGURE 9 Included SQL classes.

5 • Implementation User Case Scenario 65
TABLE 5 Medications table

P_ID MED_NAME MED_CODE
43159 VOLTAREN 75MG TABLET M01AB05
31896 TYLENOL and CODEINE TAB N02AA59
32424 ARTHROTEC 50 TABLET M01AB55
37772 AMITRIPTYLINE HCL 10MG TAB N06AA09
42235 NASONEX 50 MCG NASAL SPRAY R01AD09
TABLE 6 Patients table

P_ID P_SEX P_BY P_NAME
43159 Female 1986 Adam
31896 Male 1989 Sarah
32424 Female 1989 Ahmed
37772 Female 1980 Ted
42235 Male 1955 Lin
TABLE 7 Physicians table

PH_ID PH_NAME PH_BY
80702 John 1958
80701 Sally 1977
80702 Tom 1980
80703 Matt 1964
80701 Abby 1982
TABLE 8 Diseases table

P_ID PH_ID DISEASE_NAME
43159 80702 PTSD
31896 80701 Depressive Disorder
32424 80702 Anxiety Depression Disorder
37772 80703 Hypertension
42235 80701 PTSD
66 NLP Application
TABLE 9 NLQ MetaTable

WORDS NATURE CATEGORY SYNONYMS
Sarah Instance Value Person, Patient
Has Verb Relationship Own, Obtain, Have
Physician Noun Attribute Doctor, Provider,
Psychiatric, Surgeon
TABLE 10 Mapping table for unique values

UNIQUE VALUES SOURCE PK/FK SYNONYMS
Adam PTSD. No Person, Patient, ill
Patient.P_Name
43159 PTSD. PK Patient, Identification, Number
Patient.P_ID
John PTSD.Physician. No Person, Physician, employee
Ph_BY
80703 PTSD.Disease. FK Physician, Identification, Number
Ph_ID
AMITRIPTYLINE PTSD. No AMITRIPTYLINE, HCL, 10, Milli
HCL 10MG TAB Medication. Gram, Tblet, Drug, Medication
Med_Name
FIGURE 10 PTSD RDB chart diagram.

The examples discussed in the below use case scenarios are as follows:
Q1: What is Adam’s birth date? (Simple Query)

RA1: ΠP_BY (σP_Name = “Adam” Patients)
SQL1: SELECT P_BY FROM Patients WHERE P_Name = “Adam”;
Q2: Who is the physician that Sarah has? (Nested Query)

RA2: Π Physician.Ph_Name (σPatient.P_Name = “Sarah” Physician ⋈ Patient)
SQL2: SELECT Physician.Ph_Name FROM Physician INNER JOIN Patient
ON Patient.P_ID = Physician.P_ID WHERE Patient.P_Name = “Sarah”;
Q3: What is the most popular illness? (Simple Query)

RA3: Π Disease_Name (σMAX (Disease_Name) Disease)
SQL3: SELECT MAX ( [ ALL | DISTINCT ] Disease_Name) FROM Disease;
Q4: What drug is Ahmed taking? (Nested Query)

RA4: Π Medication.Med_Name (σPatient.P_Name = “Ahmed” Medication ⋈ Patient)
SQL4: SELECT Medication.Med_Name FROM Medication INNER JOIN
Patient
ON Patient.P_ID = Medication.P_ID WHERE Patient.P_Name =
“Ahmed”;
Q5: What medications did John prescribe for his patients? (Cascaded Query)
RA5: Π Medication.Med_Name, Medication.Med_Code (σPhysician.Ph_Name = “John” (Medication ⋈
Patient) ⋈ Physician)
SQL5: SELECT Med_Name, Med_Code FROM Medication WHERE P_ID
IN
(SELECT P_ID FROM Patient WHERE P_ID IN
(SELECT P_ID FROM Physician WHERE Ph_Name = “John”)));
Or
SELECT Medication.Med_Name, Medication.Med_Code FROM
(Medication INNER JOIN Patient
ON Medication.P_ID = Patient.P_ID) INNER JOIN Physician ON
Physician.Ph_ID = Physician.Ph_ID)
WHERE Physician.Ph_Name = “John”;
68 NLP Application
USER CASE SCENARIO

Let us assume there is a physician with the below NLQs, how could the pro-
posed algorithm reach the consequent SQL in order to be executed on the sys-
tem for answers?
Example 1:
Q1: What is Adam’s birth date?

The first step is the NLQ words breakdown process to its separate tokens.
Tokens Breakdown:
• Adam = Instance = Value.

• Birth date = noun phrase = attribute.
The NLQ words or phrases considered as tokens are those that present a
particular meaning. Such tokens will eventually participate in the iden-
tification of the RDB tables, attributes, relationships, operators (MAX,
AVG) or values. This is because any given token may have 1 of 5 possible
matches: a table, an attribute, a value, a relationship or an operator.
After searching the RDB for the instance “Adam”, it was found under
Patients.P_Name. so, the attribute name is found.
The second valuable token is “Birth Date”. Since every RDB ele-
ment (e.g., attribute) has a list of synonyms, BirthDate was matched
with Patients.P_BY. The noun phrase “Birth Date” is also a synonym of
the physician’s birth Year (Ph_BY), hence, the system must determine
the best RDB element match among all possible matches. This is done
using knowledge from other tokens’ processing. As such, since “Adam”
was found under “Patient” table, then the winning Birth Date match is
“P_BY”. Other matches’ determining mechanisms involve technical pro-
cedures such as statistical similarity measures (e.g., N-Grams Vectors’
Comparison Method). The “P_BY” here will be fed to the WHERE
clause. If there are no WHERE clauses, all DB relations and attributes will
be considered to find valid conditions. Some NLQs might not have condi-
tions, meaning there would not be a WHERE clause in the SQL template.
Generally speaking, any tables mentioned in the SELECT or WHERE
clauses should, by default, be included in the FORM clause to avoid any
SQL execution failure. Efficiency of this approach will be evaluated later
using accuracy measures.
So, the acquired information are:
• Table = Patient
• Attribute 1 = P_BY
• Attribute 2 = P_Name
The SQL Template used here is:
SELECT (Attribute1) FROM (Table) WHERE (Attribute2) = (Value);

Now we have all the information we need to execute the query as
follows:
SELECT P_BY FROM Patients WHERE P_Name = “Adam”;

Figure 11 summarizes the steps followed to solve example 1.
The benefit from Figure 12, the tokens breakdown analysis diagrams is to
show the ability to reach the source attribute, table and related RDB from
the NLQ tokens. Finding them helps feeding the SQL template with its
necessary arguments.
Token 1 Token 2
Birth Date
Adam
Type
Type Noune Phrase
Instance
Category
Aribute
Category
Search RDB for Aribute = Birth Date
Value
Birth Date synonyms found under 2 tables
Paent (P_BY) and Physician (Ph_BY)
Search RDB for Value = Adam
Paent is closer to the proper answer as it matches the
Found Under Aribute table of Adam's.
P_Name
Resulted Aribute
P_BY
Source
Source
Paent Table Paent Table
FIGURE 11 Example 1 tokens breakdown analysis.

70
NLP Application
Token 3 Token 4
Token 2
Token 1 has
Pgysician
Sarah
Type
Who Verb Type
Type Noune Phrase
Category
Instance Relaonship
Category
Search RDB for Relaonship = has
Aribute
Category
Type 2 has relaonship found
Value Paent has Disease and Paent has Physician Search RDB for Aribute = Physician
Wh-Word Algorithm stops here to use knowledge from other

Search RDB for Value = Physician table includes a proper noun value
tokens to decide on the winning relaonship aribute (as indicated by "Who") called
Adam aributes. Ph_Name, with a synonym of "Physician"
Found Under Aribute

From 4th token, the aribute of interest is "Physician"
P_Name Resulted Aribute
Category Resulted Relaonship Ph_Name
Paent has Physician
Proper Noun Source
(Person) Value Source Source
Indicator Paent Table PTSD DB Physician Table
FIGURE 12 Example 2 Tokens Breakdown Analysis.

Example 2:
Q2: Who is the physician that Sarah has?

NLQ Tokens Breakdown:
• Physician = Common Noun = Attribute

• Sarah = Instance (Proper Noun) = Value
• Has = Verb = Relationship
Here, relational DB is necessary because the relationship between the

tables “Patient” and “Physician” is important. This relationship will help
identify the table that the instance “Sarah” resides in, and who her physi-
cian is in the Physician table. This is identified by matching the patient
ID in the two tables. All of these processes are computationally expensive
in a non-relational DBMS. The difference here is that there is a critical
piece of information attached to the relationship between the two tables,
which must be a verb, that is translated automatically by the algorithm
to match the verb in the NLQ (or its synonyms). Therefore, the acquired
information is:
• Table 1 = Patient
• Table 2 = Physician
• Attribute 2 = Ph_Name
For Sarah’s identification, it is similar to example 1. This NLQ example

has a verb (has), which is translated automatically to a schema relation-
ship. After a search among the schema relationships matching the verb
“has” (or its synonyms e.g., have, obtain, acquire, etc.), more than 1 match-
ing relationship appeared coming out of the “Patient” table. Hence, we
will use the remaining information we have (Physician) to narrow down
the results. The adjacent physician name to the patient name “Sarah” in
the “Physician” table will be looked up. This is because this algorithm
depends solely on linguistic searching tools, while other matching mecha-
nisms involve technical procedures such as statistical similarity measures
(e.g., N-Grams Vectors’ Comparison Method).
After searching the RDB for the instance “Sarah”, it was found under
“Patients.P_Name”, so the attribute name (Patients) is automatically
found. This attribute name will later be fed to the WHERE clause.
The second valuable token is “Physician”. Since every RDB element (e.g.,
attribute) has a list of synonyms, the table identified here is “Physician”,
and the Attribute is (Ph_Name) with the synonym “Physician Name”.
72 NLP Application
This attribute was chosen because there is only one table containing the
attribute “Physician” as a synonym to the attribute stored in its metadata,
“Ph_Name”.
Although the word Physician also exists in the Ph_BD attribute meta-
data, the WH-Word in the NLQ (Who) refers to a human name instance
(value), not consecutive digits or a number as in the Physician Birth Date
(Ph_BD) attribute values.
Therefore, the acquired information are:
Since we have more than one table, the suitable SQL Template here is:
SELECT (Table2).(Attribute2) FROM (Table2) INNER JOIN (Table1)

ON (Table1).(Attribute3) = (Table2).(Attribute4) WHERE (Table1).
(Attribute1) = (Value);
The “Unique Identifiers List” library has a list of all unique IDs. The
SQL template will mandate this function to look for the appropriate IDs
from both tables to use in filling the SQL template. This will result in
identifying P_ID attribute in both tables.
Now that all needed information are found, the algorithm is ready to
execute the query with proper join clauses between the two tables as
follows:
SELECT Physician.Ph_Name FROM Physician INNER JOIN Patient

ON Patient.P_ID = Physician.P_ID WHERE Patient.P_Name =
“Sarah”;
Example 3:
Q3: What is the most popular illness?

Tokens Breakdown:
• What =Value Indicator

• Most = Comparative Expression = MAX clause
• Illness = Common Noun = Attribute
Based on our main assumptions, any comparative expression will help iden-
tify the SQL comparative clause. In this case it is a MAX, and to calculate
the maximum of any range, we have to know the values within that range.
From the token “Disease”, and after a search around the RDB, we found
only one table called disease with a synonym of illness. Under that table,
there is one attribute containing the word disease, which is Disease_
Name. This concludes all the required information to use the following
SQL template:
SELECT MAX(COUNT(Attribute)) FROM Table;

And with the following acquired information:
• Table = Disease
• Attribute = Disease_Name
Using the identified arguments, we can execute the following SQL

query:
SELECT MAX ( [ ALL | DISTINCT ] Disease_Name) FROM Disease;

Figure 13 summarizes the steps taken to solve Example 3.
Token 3
Token 1 Token 2
Illness
What Most
Type
Common Noun
Category
Type Type Aribute
Comparav Search RDB for Aribute = Illness

Wh-Word
e Expression
Illness synonyms found under Disease Table
Resulted Aribute
Disease_Name
Category Category
Proper Source
SQL MAX
Noun Value
Clause Disease Table
Indicator

74 NLP Application
Example 4:
Q4: What drug is Ahmed taking?

Tokens Breakdown:
• What=Value Indicator
• Ahmed= Instance = Value
• Drug = Common Noun = Attribute
• Taking = Verb = Relationship
Same as previous examples, except that the word “drug” has more than 1
matching. We have 1 table called Medication with a synonym of “Drug”,
but we have 2 attributes under the Medication table with synonyms of
“Drug”, namely Med_Name and Med_Code. Since the NLQ has no fur-
ther tokens to decide which attribute the user is referring to, we will output
both of them.
For complex and nested queries like this example, the mapping and
translation algorithm can be applied recursively.
Following the same steps of previous examples, we reach to the follow-
ing acquired information:
• Table 1 = Medication
• Attribute 1 = Med_Name
• Attribute 2 = Med_Code
• Attribute 3 = P_ID
And SQL query:
SELECT Medication.Med_Name, Medication.Med_Code FROM

Medication INNER JOIN Patient ON Patient.P_ID = Medication.P_ID
WHERE Patient.P_Name = “Ahmed”;
Token 2 Token 3 Token 4
Token 1
Sarah Taking Drug
What Type
Type
Type
Instance Common Noun
Verb
Category Category
5 • Implementation User Case Scenario

Category
Type Value Aribute
Relaonship
Wh-Word Search RDB for Value Search RDB for Aribute = Drug
= Sarah
Search RDB for Relaonship = take
Drug synonyms found under Medicaon Table
Found Under
Aribute Resulted Relaonship
Resulted Matching Aributes
P_Name Paent take Medicaon
Med_Name, Med_Code
Category
Proper Source Source Source

Noun Value
Indicator Paent Table PTSD DB Medicaon Table
75
76 NLP Application
Example 5:
Q5: What medications did John prescribe for his patients?

Tokens Breakdown:
• What =Value Indicator

• John = Instance = Value
• Medications = Common Noun = Attribute
• Prescribe = Verb = Relationship
Using the word “prescribe” will help us identify who is John, which is a
very common name, could be in the patient’s table, as well as the physi-
cians. Searching the RDB, we’ll only find 1 relationship pointing from
physician to patient. Hence, we will look for the word value “John” in the
Physician table, with proper join clauses to the three tables, following the
below SQL template:
SELECT (Table1).(Attribute1) FROM ((Table1) INNER JOIN (Table2)

ON (Table1).(Attribute2) = (Table2).(Attribute2) INNER JOIN (Table3)
ON (Table3).(Attribute3) = (Table3).(Attribute3)
WHERE (Table3).(Attribute4) = (Value);
With the acquired information:
• Table 1 = Medication
• Attribute 1 = Med_Name
• Attribute 2 = P_ID
• Attribute 3 = Ph_ID
We reach the following SQL query:
SELECT Medication.Med_Name FROM (Medication INNER JOIN

Patient
ON Medication.P_ID = Patient.P_ID) INNER JOIN Physician ON
Physician.Ph_ID = Physician.Ph_ID)
WHERE Physician.Ph_Name = “John”;
Token 2 Token 3 Token 4
Token 1
John Prescribe Medicaon
What
Type
Type Type
Instance
Verb Common Noun
Category
Category
5 • Implementation User Case Scenario

Category
Type Value
Relaonship Aribute
Wh-Word Search RDB for

Value = John
Search RDB for Relaonship = Prescribe Medicaon match found under Medicaon Table
Found Under
Aribute Resulted Relaonship Resulted Aribute
P_Name Physician Prescribe Paent Med_Name
Category
Proper Source Source Source

Noun Value
Indicator Paent Table PTSD DB Medicaon Table
77
Implementation
Testing and
Performance
6
Measurements
IMPLEMENTATION
ENVIRONMENT AND
SYSTEM DESCRIPTION
The machine used for this experiment is a MacBook Pro. It was used to run
this experiment with macOS Mojave, version 10.14.2 (18C54). The proces-
sor speed is 2.9 GHz, Intel Core i7 (SATA Physical Interconnect), and 64bit
architecture. The memory is 8 GB of RAM (distributed among two memory
slots, each of which accepts a 1600 MHz memory speed and Double Data
Rate 3 (DDR3) type of memory module), and 750 GB of disk space. The used
MacBook has 1 Processor and 2 Cores, with 256 KB per core.
For the implementation coding and execution, Python 3.7 [164] was
chosen as the programming language due to its clear syntax and popular
NLP libraries for RDB processing tasks. The Integrated Development
Environment (IDE) PyCharm C, Xcode and XQuartz were used to develop
and compile the source codes as they have a Python unit-testing frame-
work that allows for unit-testing automation in consistence with the Python
Software Foundation [130]. The system’s required dependencies include
essential tools and supportive tools. All of the tools are downloaded and
DOI: 10.1201/b23367-6 79
80 NLP Application
installed locally on the experiment machine. The essential tools are declared
in Figure 16, including:
• Python 3.7 [164]: A concise and lightweight programming language

that is compatible with most OS platforms.
• MySQL Community Server 8.0.18 [165]: MySQL RDB backend
server.
• MySQL RDB [166]: The RDB tool used to store and query data.
• NLTK [40]: Provides Python-compatible libraries for NLQ lem-
matizing, tokenizing, tagging, parsing, classifying and semantic
reasoning. It also supports interfaces to over 50 lexical resources in
addition to WordNet corpora.
• TextBlob [167]: A Python library to process NLP tasks such as POS
tagging, classification, noun phrase extraction and sentiment analysis.
• Stanford CoreNLP 3.9.2 [21]: Provides a set of integrated NLP tools
to apply linguistic analysis on any incoming NLQ via a Python-
compatible API. It offers sentences’ structure sentiment analysis
and syntactic and grammatical dependencies analysis. In addition,
FIGURE 16 Framework structure, tools and libraries.

6 • Implementation Testing and Performance Measurements 81
Stanford CoreNLP provides a stemmer, POS tagger, dates and

times normalizer, NER, annotator, parser and bootstrapped pattern
learning. Also, it offers the open information extraction tools such
as extracting relationships between NLQ tokens.
• WordNet [19]: A large English lexical DB that includes nouns,
verbs, adjectives, etc., in addition to the “synsets” library, which is
a grouped set of cognitive synonyms.
The system’s supportive tools include:
• IDE PyCharm C [168]: The Python IDE for code development and
unit testing.
• XQuartz 2.7.11 [169]: A development environment designed for
Apple OS X with supportive libraries and applications.
• Xcode 11 [170]: An application development tool for Apple OSX,
used in this implementation to check codes’ syntactic rightness.
• MySQL Workbench [171]: An SQL development and administra-
tion tool used mainly for visual modeling.
DATABASE
The current implementation uses MySQL DBMS as a backend environment.
The implementation testing uses two RDBs, Zomato RDB [172] for algorithm
testing, and the WikiSQL RDB [173] for algorithm validation. The testing
process using a small RDB confirms the framework’s functionality, while the
framework validation process evaluates the framework’s accuracy, efficiency
and productivity.
Results from both Zomato (small RDB) and WikiSQL (large RDB) will be
compared based on the RDB size. Table 11 compares between the two RDBs
in terms of their number of instances or records, the number of tables and the
public data source where they were published.
Zomato RDB [172], published in 2008, is a small RDB with a size of
2.5MB having 9,552 NLQ and SQL pairs stored in three comma-separated
value (csv) file tables. Zomato RDB is about a restaurant search engine sup-
plied by the public data platform “Kaggle”. Zomato RDB has the schema dem-
onstrated in Figure 17.
The WikiSQL_DEV RDB [174], published in 2017, was chosen because
of its large RDB. It has 200.5 MB of 80,654 manually annotated RDB of NLQ
and SQL pairs in 24,241 tables from Wikipedia. This RDB is used for develop-
ing NLIs for RDBs. Moreover, WikiSQL is considered the largest web-based
82 NLP Application
TABLE 11 Two RDBs comparison

ZOMATO WIKISQL
Size 2.5 MB 200.5 MB
NLQ/SQL Instances 9,552 80,654
Tables 3 24,241
Data Source Kaggle Wikipedia
FIGURE 17 Zomato RDB schema.
realistic hand-annotated semantic parsing RDB [175]. This is because of the

RDB’s large magnitude and variety of logical form examples, tables, columns,
lengths and types of questions, and the length of queries. Hence, it is the ideal
RDB to generalize the implemented mapping algorithm to new and diverse
queries and table schemata. Some examples of WikiSQL_DEV NLQ/SQL
pairs are in Table 12.
TABLE 12 Examples of WikiSQL_DEV NLQ/SQL pairs

# NLQ TABLES SQL
1 How many capital “Country(exonym)”, SELECT COUNT (Capital
cities does Australia “Capital(exonym)”, (endonym)) FROM
have? “Country(endonym)”, 1-1008653-1 WHERE
“Capital(endonym)”, Country(endonym)=
“Official or native Australia
language(s) (alphabet/
script)”
2 What are the “Rd”, “Name”, “Pole SELECT (Name) FROM
races that Johnny Position”, “Fastest 1-10706879-3 WHERE
Rutherford has won? Lap”, “Winning Winning driver=Johnny
driver”, “Winning Rutherford
team”, “Report”
3 What is the number “Player”, “No. (s)”, SELECT(No. (s)) FROM
of the player who “Height in Ft.”, 1-11734041-9 WHERE
went to Southern “Position”, “Years for School/Club Team/
University? Rockets”, “School/ Country=Southern
Club Team/Country” University
4 What is the toll “Name”, “Location”, SELECT (Heavy vehicle (3/4
for heavy vehicles “Light vehicle”, axles)) FROM 1-1211545-2
with 3/4 axles at “Heavy vehicle WHERE Name=Verkeerdevlei
Verkeerdevlei toll (2 axles)”, “Heavy Toll Plaza
plaza? vehicle (3/4 axles)”,
“Heavy vehicle
(5+ axles)”
5 How many millions “No. in series”, SELECT COUNT (U.S.
of U.S. viewers “No. in season”, viewers (millions)) FROM
watched the episode “Title”, “Directed 1-12570759-2 WHERE
"Buzzkill"? by”, “Written Title="Buzzkill"
by”, “Original
air date”, “U.S.
viewers(millions)”
IMPLEMENTATION TESTING
AND VALIDATION
Testing the proposed mapping algorithm happens by running a randomized
shuffling of the NLQ/SQL pairs from the Zomato RDB. This step uses four
library functions namely, “random.shuffle”, “collections.defaultdict”, “tqdm”
and “sql_parse.get_incorrect_sqls”. First, the underlying NLP tools and the
Matcher/Mapper module are tested by feeding the system the NLQ lemma-
tized tokens. Then, the tokens go through the Matcher/Mapper module to
84 NLP Application
match the tokens with their synonyms built into the NLQ MetaTable. After
that, tokens and their synonyms will be mapped to their adjacent RDB values,
attributes, tables or relationships, each based on their syntactic role. To test
the SQL template generator module, a set of RDB lexica will be passed to this
module and the generated SQL will be examined for correctness, accuracy and
other performance metrics discussed in the next section.
PERFORMANCE EVALUATION
MEASUREMENTS
The purpose of the proposed algorithm is generating SQLs from NLQs auto-
matically. It is important to obtain a reliable estimate of performance for this
language translation algorithm. However, the algorithm’s accuracy perfor-
mance may rely on other factors besides the learning algorithm itself. Such
factors might include class distribution, effect (cost) of misclassification and
the size of training and test sets. Therefore, to validate the algorithm’s perfor-
mance and efficiency, more detailed accuracy measures are used to test the
generated SQLs accuracy, precision and recall using:
• False Positive Ratio (FPR = C/(C+D)): the incorrectly classified

queries as positives, but they are actually negatives.
• True Negative Ratio (TNR = D/(C+D)): the correctly classified que-
ries as negatives.
• False Negative Ratio (FNR= B/ (A+B)): the incorrectly classified
queries as negatives, but they are actually positives.
• True Positive Ratio (TPR = A/(A+B)): the correctly classified que-
ries as positives.
Where A = True Positive, B = True Negative, C = False Positive and D = False

Negative.
The classification process here compares the generated SQL by the current
framework against the designated SQL that is originally present in the testing
RDB.
The recall performance measure represents the proportion of positive
case (correct) queries which are correctly generated, Recall(R) = A/(A+B). It
also measures the presentation ratio of all relevant words by the system. The
words here represent the derived lexica that are correctly identified from the
RDB and lead to correct SQL generation. In this case, Recall = number of
relevant words (lexica) retrieved from RDB/number of relevant words (lexica)

not retrieved.
Precision is the proportion of the generated positive case (correct) queries
which are correctly generated and considered as correct SQL constructions,
Precision (P) = A/(A+C). It also measures how efficient the system is in retriev-
ing only relevant words (lexica). In this case, precision is a measure of the abil-
ity of a system to retrieve and present only relevant lexica. Precision = number
of relevant lexica retrieved/total number of lexica retrieved.
Moreover, results’ correctness or accuracy is the proportion of total num-
ber of positive (correct) SQL generations which were correctly generated.
Accuracy = (A+D)/(A+B+C+D). Unordered sets of retrieved queries can be
evaluated by Precision and Recall. For ranked sets, after each query retrieval,
precision should be plotted against recall.
The Receiver Operating Characteristics (ROC) curves [176] will also be
used. ROC is a machine learning graphical plot of the TPR (a.k.a. sensitivity)
against the FPR (a.k.a. 1-specificity). It makes a comparison between the two
translation experiments in the current work, the experiment using the Zomato
RDB and the second experiment using the WikiSQL RDB. This classification
test takes into consideration the generated SQL by the current framework and
the SQL already present in the testing RDB. The ROC curves show where the
two experiment sets would possibly connect. This is because every TPR or
FPR prediction instance is a single point on the ROC space. The bigger the
area under the ROC curve the bigger the benefit of using the associated test.
In other words, predictors’ curves that are closer to the top-left corner provide
better accuracy performance. Depending on the matching accuracy between
the output SQL and the original NLQ input, the proposed algorithm is evalu-
ated and documented.
In the first experiment with the Zomato RDB, 20 iterations (epochs) are
executed on the system where the input NLQs are executed and their equiva-
lent SQLs are generated as output. Then, the implementation resulted with the
following performance metrics declared in Table 13 and Figure 18.
TABLE 13 First experiment confusion matrix with Zomato RDB

F-MEASURE: 94.5 ACTUAL POSITIVE ACTUAL NEGATIVE
Predicted Positive 43 TPR (A) 3 FPR (C)
Type I error
Predicted Negative 2 FNR (B) 49 TNR (D)
Type II error
Accuracy: 94.85% Recall: 0.96 Precision: 0.93
86 NLP Application
FIGURE 18 ROC curve for the first experiment with Zomato RDB.
For the second experiment with the WikiSQL RDB, the implementation
resulted with the following performance metrics declared in Table 14 and
Figure 19.
Compared with other similar research works on WikiSQL, the proposed
work still achieves the highest accuracy measure as illustrated in Table 15 and
Figure 20.
While the aforementioned performance measurements are sufficient to
answer the current research question, the average time translating each query
remains 1.5 minutes. This could be mainly due to the humble computer system
TABLE 14 Second experiment confusion matrix with WikiSQL RDB

F-MEASURE: 92 ACTUAL POSITIVE ACTUAL NEGATIVE
Predicted Positive 42 TPR (A) 4 FPR (C)
Type I Error
Predicted Negative 3 FNR (B) 48 TNR (D)
Type II Error
Accuracy: 92.78% Recall: 0.93 Precision: 0.91
FIGURE 19 ROC curve for the second experiment with WikiSQL RDB.
TABLE 15 NLQ to SQL translation work on WikiSQL RDB

# SOURCE APPROACH ACCURACY (%)
1 Proposed Algorithm Computational Linguistics. 92.78
2 SEQ2SQL [177] MLA through reinforcement 59.4
Learning.
3 TypeSQL [174] NLQ’s token type recognition and 82.6
2 bi-directional LSTM.
4 SQLOVA [178] Table- and context-aware NLQ 89.6
word contextualization and
representations.
5 X-SQL [179] Reinforce schema representation 91.8
with context.
6 WHERE clause Attentional Recurrent Neural 88.6
variants [180] Network (RNN).
7 DialSQL [181] A dialogue-based framework that 69
boosts the performance of existing
algorithms via user interaction.
(Continued)
88 NLP Application
TABLE 15 (Continued) NLQ to SQL translation work on WikiSQL RDB

# SOURCE APPROACH ACCURACY (%)
8 SQLNet [182] A dependency graph, a sequence- 68.3
to-set model and the column
attention mechanism.
9 Question Patterns Question-pattern models Unmeasured
[183] containing dependency graphs.
10 ValueNet [184] A neural model based on an 67
encoder-decoder architecture to
synthesize the SQL query.
NLQ to SQL Translation Work on WikiSQL RDB

100.00%
90.00%
80.00%
70.00%
60.00%
50.00%
40.00%
30.00%
20.00%
10.00%
0.00%
Proposed SEQ2SQL TypeSQL SQLOVA X-SQL WHERE clause DialSQL SQLNet
Work variants
Accuracy
FIGURE 20 WikiSQL works accuracy comparison.
used to run the testing and validation processes. However, this processing time
could be enhanced when analyzing the exact reasons of delay using further
performance analysis. For example, each server executing the NLQ into SQL
translation requests could be examined using the following specific perfor-
mance metrics:
• Residence time, RT = W/C: the system’s resource usage. This rep-

resents the average time queries spend in the server (actual service
time + waiting time).
• Utilization, U = B/T: the average percentage of server’s busy time.
• Throughput, X = C/T: the average percentage at which the server
completes queries’ translation requests.
• Queue length, N = W/T: the average number of queries at the server,
whether executing the translations or waiting for service.
• Mean service time, S = B/C: the average time the server is busy with
queries’ translation processes.
• Area under graph, W = : the total time the server used to translate
all queries.
Where T = Total Period, C = Completed and B = Busy Time.
After running the analysis procedure, and as per Figure 21, it turns out that the
process phase that took the longest time is the matching and mapping phase.
This is to be expected since it does most of the tasks executed by the transla-
tion algorithm. A surprising discovery is the amount of time spent on the query
execution and results retrieval from the MySQL RDB as follows:
• NL Interface = 0.04 secs

• POS Recognition = 0.25 secs
• Disambiguation = 0.06 secs
• Matcher/Mapper = 0.63 secs
• SQL Template Generator = 0.07 secs
• SQL Execution and Results = 0.45 secs
This time consumption breakdown represents the average time taken by each
module to execute a translation task. They were computed after running a
group of translation tasks and calculating the average time consumed by them
combined.
NL Interface
2%
POS Recognion
SQL Execuon and
17%
Results
30%
Disambiguaon
4%
SQL Template
Generator
5%
Matcher/Mapper
42%
FIGURE 21 Translation average time distribution of 1.5 minutes.

90 NLP Application
When an SQL is passed to the DBMS engine for query processing, it is

compiled by the MySQL Community Server compiler for result retrieval. This
process is faster when dealing with small RDBs, but with large RDBs, the que-
ries take much longer to execute. This is mainly because the MySQL RDB is
not designed for big data application requirements.
The additional translation time consumed is proportional to the size of the
RDB since the matching and mapping process will have to examine the whole
RDB for potential matches. For example, with a 2MB RDB, the translation
could take 5 seconds to complete, while it could take up to 1.5 minutes for a
translation process to complete with an RDB of size 1GB.
What also affects the time consumption on the SQL execution and results
phase is not having a query optimizer module. Since the same SQL query can
be written in several ways, query optimization chooses the best way a query
could be syntactically expressed. Not well-formed or optimized queries take
longer to execute, affecting the overall process performance.
Finally, MySQL Community Server does not use proper temporary cach-
ing. Hence, frequently accessed data are not temporarily stored in the cache to
insure faster future accesses. Consequently, the MySQL Server needs to estab-
lish a new service request whenever a new SQL query arrives for execution.
Generally, the server processing time consumed depends on:
• The number of queries.

• The amount of service each query needs.
• The time required for the server to process individual queries.
• The policy used to select the next query from the queue, the Queueing
Model [185] (e.g., the First-Come-First-Served [186] or Priority
Scheduling [187]).
The Queueing Model [188] could be used to enhance the translation perfor-
mance by grouping together the similar SQL types in the queue. However,
queueing models have dependency side effects considering the relationships
between the SQLs in the queue and the corresponding service times for each
SQL execution process.
Those effects could be mitigated by using similar calculations based on
predicted workload intensity [189] and service requirements [190]. The work-
load intensity [189] is a measure of the number of translation requests made in
a given time interval. The service requirements [190] represent the amount of
time each query translation request requires from the server in the processing
system.
If we assume that the system is fast enough to handle the arriving transla-
tion requests, the queries’ translation completion rate (throughput) would equal
the arrival rate. In this ideal case, the implementation environment would be
called “jobs-flow balance” [191] where each query translation duration equals
zero minutes instead of 1.5 minutes, as is the case in the current implementa-
tion environment.
Implementation
Results Discussion 7
In terms of precision, Zomato RDB scored 93% while WikiSQL scored 91%.
Those are the proportion of the correctly generated queries. It also measures
the algorithm’s efficiency in the identification and retrieval of matching RDB
lexica since the retrieval of the wrong lexica would cause lower accuracy mea-
sures due to wrong SQL generations.
From the aforementioned precision and recall measures, the F-measure can
be derived where F-measure = 2PR/ (P + R). The F-measure is the average
performance measure of the matching RDB lexica retrieved as a result of an
accurate matching and mapping process during the NLQ into SQL translation.
Hence, the F-measure basically measures the accuracy of the data retrieved
from an RDB as a result of applying an algorithm. The implementation execu-
tion using Zomato RDB had an F-measure of 94.5%, while WikiSQL had a
92%. In Table 16, a comparison between the two RDB experiments’ perfor-
mance measures is summarized in a confusion matrix. The numbers in the table
represent averages over all runs of both experiments considering all queries in
a run.
Figure 22 illustrates and summarizes all aforementioned performance
metrics measures. With regard to the peaks of accuracy, recall, precision and
F-measure bars in Figure 22, and in addition to the error rates (FPR and FNR)
comparison, it can be concluded that the proposed algorithm is functioning
properly and as needed. Hence, the mapping algorithm does indeed select the
correct RDB elements successfully and map them with the correct SQL clauses
using the novel mapping mechanism. This mapping is based on linguistics
studies of the sentence structure by breaking down the sentence into the words
level and study the words’ inter-/intra-relationships.
Moreover, the area under the ROC curves, called AUC and shown in
Figures 18 and 19, for Zomato RDB is almost 100% while WikiSQL AUC area
is 93%. We conclude from this AUC comparison that the proposed algorithm
shows accurate results for smaller RDBs, but not as much accuracy for RDBs
that cover big data.
DOI: 10.1201/b23367-7 93
94 NLP Application
TABLE 16 Confusion matrix comparison between the two RDB experiments

MEASURE ZOMATO RDB WIKISQL RDB
Accuracy 94.85 92.78
Recall 96 93
TPR 43 42
FPR 3 4
TNR 49 48
FNR 2 3
Precision 93 91
F-Measure 94.5 92
120
100
80
60
40
20
0
Accuracy Recall Precision F-Measure TPR FPR TNR FNR
Zomato DB WikiSQL DB
FIGURE 22 Comparison between the two RDB experiments.
IMPLEMENTATION LIMITATIONS
Mapping Limitations
The reason for the lack of accuracy in bigger RDBs, according to the pro-
posed algorithm’s experiments, is the system’s confusion between the actual
7 • Implementation Results Discussion 95
RDB elements’ names in the RDB MetaTable and the synonyms table as a
whole. Thus, if a field is actually named “Birth_Day”, and another field is
named “BD” but has a synonym of “Birthday”, the system will give prior-
ity to the field named “Birth_Day”, which is the main source of confusion.
However, the adoption of the synonym’s table in NLP is quite immature
and could be improved using appropriate machine learning techniques. Such
techniques include classifying the synonyms and recognizing the actual col-
umn names.
Another cause of inaccuracy in the proposed framework is the mapping
table. When NER or data profiling is used to import RDB’s unique values and
fields and tables’ names, the algorithm will be obstructed from correctly map-
ping an NLQ token to an RDB value. This occurs when the NLQ token is not a
unique value and therefore not included in the mapping table. Though the algo-
rithm is supposed to search for the mentioned NLQ value in the whole RDB, it
still starts with the RDB’s unique values table (the mapping table) to minimize
the searching time. This precedence prioritizes the RDB’s unique values list,
stored in the mapping table, over the entire RDB elements which increases the
chances that the included unique values are mistakenly selected as a matching
lexicon. Yet the value retrieval accuracy would still not be guaranteed since
this depends greatly on how clearly the data clerk have entered the data and
whether it had adequate synonyms attached to it.
SQL Generation Limitations

Other minor SQL generation errors, especially in the WikiSQL RDB, were
due to the system’s inability to grasp a thorough semantic understanding of
the NLQ. An example is the system’s inability to understand that the “king’s
speech” is the same as the “speech of the king”. Other errors were related to
SELECTing the wrong field due to NLQ’s unresolved ambiguity.
In addition, the huge size of the WikiSQL RDB increases the likelihood
of the presence of similarly named fields and tables which might even have
identical labeling and synonyms. To mitigate this problem, RDBs must be pre-
processed to have unique attribute and table names. Based on the proposed
system and experiment results, the smaller the NLQ/SQL training patterns set
the better results the system produces.
Larger sets have higher chances to cause rules conflictions or complica-
tions, while adding more rules would not always solve this conflict or increase
results’ accuracy. Translation failures could also be due to missing SQL clauses
and not being retrieved because of ambiguous NLQs, or possibly due to mis-
matching in nested SQLs.
96 NLP Application
General Implementation Limitations

In the proposed system, the aim is building an automatic NLQ translation into
an understandable language by a machine, that is SQL. However, NLIDBs
require a significant amount of manual work for rule-based constraints, gram-
mar specifications and RDB MetaTables annotation. In addition, RDB enti-
ties must be organized by properly naming tables and fields, converting all
relationships’ names into verbs, defining unique tables and identifying stored
data types. With all said efforts, there is no need for a tailored data dictionary
like most NLIDB question answering systems. This is because the designed
MetaTables already have all knowledge-based facts needed for NLQs transla-
tion. On the other hand, the algorithm does require a significant amount of
manual work in exchange, which can be worth it if the system will not need
heavy future maintenance or customization along the way.
There are other general implementation limitations in this work, such as:
• The use of MetaTables as a Knowledge Base implies that the system

is domain-dependent. This mandates MetaTables reconfiguration
on any other RDB system, on which it is implemented, to be used by
a DBA. The amount of manual work required is preprocessing the
new RDB to have unique attribute and table names. Also, any acro-
nyms or abbreviations must be excluded and changed to their proper
namings as it is hard for the system to annotate acronyms with syn-
onyms. The rest of the reconfiguration process is automatic in that
the system has to process the RDB to be annotated and tagged with
necessary metadata as illustrated in the RDB MetaTable.
• The Zomato RDB mapping table covers unique values of 5 fields and
is about 12 KB, but with a larger DB, the mapping table becomes
a limitation on the system resources (i.e., storage disks). As such,
about 10% extra storage on large RDBS will be required to store all
the metadata present in the RDB MetaTable.
• After executing SQL statements, the results returned are in a less
human-understandable format; that is, columns and rows. Hence,
the result should be manipulated to enable its presentation in an
NLQ format. As an example, Yes/No questions, such as “Is there
inpatients in the ICU Ward 5?”, will return records if the answer is
“Yes” and return no results if the actual result is “No”!
• The coding language used to implement such an algorithm must
possess the capability to connect to an RDBMS and handle NLP
tasks, i.e., Python or Java.
7 • Implementation Results Discussion 97
• Some queries are unanswerable and cannot be converted to SQL

because they are too general, i.e., “Who is alive?” Another rea-
son could be that the NLQ equivalent token in the RDB or its
MetaTables cannot be mapped due to vague terms used in the NLQ
sentence, i.e., “young” instead of “age between 4 to 7” or “heavy”
instead of “weight more than 100 kilograms”. The direct reason for
unanswerable queries is the nonexistence of any matching records
to the query conditions and constraints. In this case, the adoption of
Fuzzy Logic solutions is necessary.
• Limitations regarding SQL language itself. For example, there is
no loop operator or curser management capability in SQL. Also,
expressing a universal quantifier in SQL requires a double-negated
EXIST construct.
• Complicated or nested queries (i.e., aggregate functions) are not
covered in the proposed algorithm as the model is tested on simple
SELECT statements only. The current approach can be expanded
to more complex queries. However, that might require an SQL con-
struction module rather than an SQL template generator.
• The system has no embedded temporary memory (i.e., Cache or
RAM) that stores past NLQs for accelerated frequent data retrieval.
Temporary storage memory could also be used to store recently
retrieved tables temporarily in case the user asks a follow-up ques-
tion. It would be helpful if the temporary memory hosts related data
until the user asks an NLQ of a different subject.
Conclusion
and Future
Work
8
CONCLUSION
It is almost impossible to find a technical application that does not require
some sort of data storage and retrieval. The era of HCI, and the means of inter-
action, that is NLP, is in an ever-growing mode. In summary, the proposed
algorithm aims at solving the language translation gap in the literature with the
proposed mapping algorithm. The algorithm is designed to work on any RDB
schema domain.
The mapping happens in accordance with a manually written rule-based
mapping constraints algorithm. Two mappers are developed, one to map NLQ
tokens into RDB elements, and another to map the identified RDB lexica into
SQL clauses. The algorithm starts by analyzing the input NLQ by executing a
series of NLP tasks. At each analysis stage, the data is further processed to lead
to the formation and generation of an SQL. At the end, the SQL is executed,
and the data are fetched from the RDB and displayed to the user.
The proposed algorithm covers many recent literature works’ shortcom-
ings using the following solutions:
i. Limited HCI interaction with users to assure the most natural way
of communication, that is direct questioning and answering. This
way the user does not need to identify any NLQ tokens semantic
roles.
ii. Does not need an annotated NLQ/SQL pairs corpus for training,
making it domain-independent and adaptable on any environment.
DOI: 10.1201/b23367-8 99
100 NLP Application
iii. Uses simple algorithmic rules based on computational linguistics to

fully understand the input NLQ for best translation accuracy.
iv. Inventing NLQ and RDB MetaTables to increase mapping accuracy
between NLQ tokens, RDB lexica, and then SQL clauses.
v. Maps NLQ verbs with the RDB relationships as the simplest and
most effective way of representing RDB elements relationships.
vi. Overcomes any poor underlying linguistic tools’ performance by
using the supportive MetaTables and WordNet ontology.
vii. Supports NLQ’s syntactic and semantic grammar analysis with com
putational linguistic algorithms (MetaTables constraints). Those con-
straints assist the NLQ tokens mapping to the RDB lexica, without
using heavy-weighted and complicated techniques.
viii. Presents a significantly simpler algorithm as it relies on fewer, but
more effective, linguistic tools and mapping rules without using
intermediate representational layers.
To the best of our knowledge, this proposed research work for NLQ into SQL
mapping and translation presents a novel mechanism. This work bridges the
gap between RDBs and nontechnical DB administrators through a simple
language translation algorithm using strong underlying NLP techniques. This
work enables nontechnical users with no knowledge of RDB semantics to have
the capability to retrieve information from employed RDBs.
The validation of the proposed research experiments and results has
shown promising NLQ into SQL transformation and translation performance.
As such, the smaller RDB performed a 95% accuracy, which is more than the
larger RDB, which scored about 93%. This conclusion is in accordance with
the applied performance metrics and measures such as accuracy, precision,
recall and F-Measure.
However, larger RDBs in this experiment identified clear areas of improve-
ment to enhance their language transformation accuracy to higher than a 93%
accuracy. Another big area of improvement is further simplifying the algo-
rithm coding and testing it on better implementation environment and technical
resources. The aim is to minimise the translation time as it takes an average of
1.5 minutes to return a well-formed SQL, given an NLQ.
FUTURE WORK
Since the research around NLIDBs is only a few years old, there are so many
future work opportunities to expand this work, including but not limited to:
8 • Conclusion and Future Work 101
• Adopting NoSQL queries.

• Scaling up to distributed storage RDBs.
• Employing community detection algorithms.
• Domain independent schema building.
• NLQ ambiguity and uncertainty resolution through fuzzy constraints.
• Dealing with vague and imprecise data through fuzzy RDBs that
store fuzzy attribute values and fuzzy truth values.
• Attaching a well-designed user interface for NLQ input.
• Investigating neural network learning approaches for SQL rank-
ing and classification based on a weighting scheme or an error/
correctness rate.
• Processing NLQs with NLQ modifiers (i.e., almost, nearly, very).
• Processing NLQs expressed in NLQ time-stamped forms with
prepositions (i.e., on, during, since).
• Outputting and transforming query translation and execution results
into XML documents format. XML format is a standard scheme to
store, interchange or exchange semi-structured to structured data.
Appendix 1
PSEUDOCODE 2 NLQ TOKENS LABELING
word = NLQ(tokens)
for rslt = match_label[Table, Attribute, Value,
Relationship]
if rslt[0] then
label token as Table
end if
if rslt[1]
if rslt[2] then
label token as Value
return(rslt[1], rslt[2])
else
label token as Attribute
end if
if rslt[3] then
label token as Relationships
end if
end for
103
Appendix 2
PSEUDOCODE 5 TOKENS’ TYPE DEFINITION
for TextBlob(sentence[]) ← tokens

if token type is noun_phrase then
tag token as noun_phrase
use as lexicon(table_name, attribute)
if token type is string or number then
tag token as Literal_Value
use as lexicon(value)
elif token type is proper_noun then
tag token as proper_noun
elif token type is literal_value then
tag token as literal_value
elif token type is verb then
tag token as verb
use as lexicon(relationship)
elif token type is adverb then
tag token as adverb
use as lexicon(attribute)
elif token type is adjective then
tag token as adjective
use as lexicon(attribute)
elif token type is preposition then
tag token as preposition
elif token type is Wh_question then
tag token as Wh_question
use as lexicon reference
elif token type is conjunction_phrase then
tag token as conjunction_phrase
use as lexicon condition
105
106 Appendix 2
elif token type is disjunction_phrase then

tag token as disjunction_phrase
elif token type is comparative_expression then
tag token as comparative_expression
else token type is operational_expression
tag token as operational_expression
return sentence[tags]
Appendix 3
PSEUDOCODE 7 SYNONYMS TAGGING OF SQL

COMPARATIVE OPERATIONS KEYWORDS
keywords_synonyms()
if keyword is average then
add synonyms[‘average’, ‘avg’]
elif keyword is great then
add synonyms[‘greater’,’gt’,’>’,’larger’,’more
than’, ‘is greater than’]
elif keyword is small then
add synonyms[‘smaller’,’st’,’<‘,’lesser
than’,’less than’, ‘is less than’]
elif keyword is greater_or_equal then
add synonyms[greater or equal’, ‘gt or eq’,
‘>=‘, ‘larger or equal’, ‘more than or equal’]
elif keyword is smaller_or_equal then
add synonyms[‘smaller or equal’, ‘st or eq’,
‘<=‘, ‘lesser than or equal’, ‘less than or equal’]
elif keyword is equal then
add synonyms[‘equal’, ‘eq’, ‘=‘, ‘similar’,
‘same as’, ‘is’]
elif keyword is sum then
add synonyms[‘what is the total’, ‘sum’]
elif keyword is max then
add synonyms[‘what is the maximum’, ‘max’,
‘maximum’]
elif keyword is min then
add synonyms[‘what is the minimum’, ‘min’,
‘minimum’]
elif keyword is count then
add synonyms['how many’, ‘count’]
elif keyword is junction then
add synonyms[‘and’, ‘addition’, ‘add’,
‘junction’]
107
108 Appendix 3
elif keyword is disjunction then

add synonyms[‘or’, ‘either’, ‘disjunction’]
elif keyword is between then
add synonyms[‘among’, ‘between’, ‘range’]
elif keyword is order_by then
add synonyms[‘order by’, ‘order’, ‘organise’]
elif keyword is asc then
add synonyms[‘asc’, ‘ascending’, ‘small to
big’, ‘top to bottom’]
elif keyword is desc then
add synonyms[‘desc’, ‘descending’, ‘big to
small’, ‘bottom up’]
elif keyword is group_by then
add synonyms[‘group by’, ‘group’]
elif keyword is negation then
add synonyms[‘negation’, ‘not’, ‘negative’, ‘is
not’, ‘are not’, ‘does not’]
elif keyword is like then
add synonyms[‘what is the’, ‘like’, ‘similar
to’, ‘same as’]
else keyword is distinct
add synonyms[‘distinct’, ‘unique’]
end if
return keywords_synonyms(tags)
Appendix 4
PSEUDOCODE 8 NLQ SPELLING CHECK FUNCTION
input = nlq(words)
output = correct_nlq(input)
while input ≠ Ø do
if Spellcheck(nlq) = error then
print (‘Sorry, there is an error in your NLQ.’,
‘nlq’)
reset input = user_response(nlq(words))
return input
if ambiguitycheck(input) = true
print out (‘What did you mean by’,
ambiguate(word), ‘?’)
classify user_response()
if user_response = true then
set input ← user_response(clarification)
else
set input = user_response(originalNLQ)
end if
end if
else Spellcheck(nlq) ≠ error
end if
end while
109
Appendix 5
PSEUDOCODE 12 MATCHING NLQ TOKENS TO

EQUIVALENT RDB ELEMENTS
for nlq(token) do
/* mapping tokens with their equivalent lexica or
their synonyms */
if lexica[matching_lexicon(table, attribute,
value, relationship), synonym] ← token then
find (matching_lexicon(table) -
[HAS_ATTRIBUTE]-> matching_lexicon(attribute) -
[HAS_VALUE]-> matching_lexicon(value)) -
[HAS_RELATIONSHIP]->
matching_lexicon(relationship)
Compare token with matching_lexicon and
synonym
spanTag matching_lexicon where
matching_lexicon is similar to token and
similarity > 0.75
return matching_lexicon
elif matching_lexicon > 1 then
print (‘which word did you mean to use?’,
lexicon[0], ‘or’, lexicon[1])
matching_lexicon ← user_response()
return matching_lexicon
else matching_lexicon[] ↚ token
matching_lexicon(table) or
matching_lexicon(attribute) or
matching_lexicon(value) or
matching_lexicon(relationship) = False
return error
end if
/* find the corresponding RDB elements from the
identified ones */
111
112 Appendix 5
if matching_lexicon(value) = True then

matching_lexicon(attribute) ←
current_attribute
elif matching_lexicon(attribute) = True then
matching_lexicon(table) ← current_table
else matching_lexicon(table) = True
matching_lexicon(relationship) = current_
relationship
end if
return matching_lexicon(table, attribute, value,
relationship)
replace token with matching_lexicon
end for
Appendix 6
PSEUDOCODE 13 BUILDING SQL MAIN CLAUSES
for sql_clauses(select, from, where) do
/* define where clause */

if lexicon(attribute, value) = 1 then
include select_clause(attribute, value)
else lexicon(attribute, value) > 1 then
include select_clause(attributes, values)
separated with ‘,’
return select_clause
end if
/* define from clause */

if lexicon(table) = 1 then
include from_clause(table)
else lexicon(table) > 1 then
include from_clause(tables)
join from_clause(tables)
return from_clause
end if
/* define where clause */

where_clause ← condition_type[min, max, avg,
sum, count, distinct]
if condition_type = True then
include where_clause(conditions)
add conditions with ‘and’
return where_clause
end if
return select_clause + from_clause + where_clause

end for
113
Appendix 7
Included SQL Query Types:
• Simple Queries:
– SELECT 1 column in a table (or more) without conditions to
present all data under selected column.
• Nested Queries (Subqueries):
– SELECT 1 column in a table (or more) with WHERE condition\s.
• Cascaded Queries:
– Join 2 or more columns FROM 2 or more tables in the SELECT/
FROM statement without conditions like:
table-name1 JOIN table-name2 ON attribute1(PK of table1) =
attribute2 (attribute in table2 and also FK of table1)
– Join 2 or more columns FROM 2 or more tables in the SELECT/
FROM statement with WHERE conditions. The WHERE clause
is a single condition or a joint of several conditions.
• Negation Queries:
– Using the NOT Operator with SQL syntax to negate a WHERE
condition.
• Simple WHERE Conditions:
– 1 Simple operational condition (=,>, <, etc.)
– 1 Aggregation condition (max, min, etc.)
– 1 Negation Condition (NOT)
• Complex WHERE Conditions:
– 2 or more operational conditions (=,>, <, etc.)
– 2 or more Aggregation conditions (max, min, etc.) concatenated
“=” with a value specified by the end user.
– 2 or more Negation conditions (NOT AND)
– Including subordinates and conjunctions
• Order/group by:
– Asci. (Alphabetical, numeric).
– Desc. (Alphabetical, numeric).
115
Appendix 8
PSEUDOCODE 3 SQL TEMPLATE EXAMPLES
class Templates:
/* zero attributes, one table */
temp100 = Template(‘SELECT DISTINCT * FROM $table’)
/* zero attributes, one table, one attribute-value
pair */
temp101 = Template(‘SELECT DISTINCT * FROM $table
WHERE $attribute='$value'‘)
/* one attribute, one table */
temp110 = Template(‘SELECT DISTINCT $attribute FROM
$table’)
/* one attribute, one table, two attribute-value
pairs (AND) */
temp112 = Template(‘SELECT DISTINCT $attribute FROM
$table WHERE $attribute1='$value1' AND
$attribute2='$value2'‘)
/* two attributes, one table */
temp120 = Template(‘SELECT DISTINCT $attribute1,
$attribute2 FROM $table’)
/* zero attributes, two tables */
temp200 = Template(‘SELECT DISTINCT * FROM $table1
NATURAL JOIN $table2’)
/* zero attributes, two tables, one attribute-value
pair */
NATURAL JOIN $table2 WHERE $attribute='$value'‘)
/* zero attributes, three tables, one attribute-
value pair (AND) */
NATURAL JOIN $table2 NATURAL JOIN $table3 WHERE
$attribute='value'‘)
117
118 Appendix 8
/* zero attributes, three tables, two attribute-

value pairs (AND) */
NATURAL JOIN $table2 NATURAL JOIN $table3 WHERE
$attribute1='$value1' AND $attribute2='$value2'‘)
/* one attribute, three tables, two attribute-value
pairs (AND) */
temp312 = Template(‘SELECT DISTINCT $attribute
FROM $table1 NATURAL JOIN $table2 NATURAL
JOIN $table3 WHERE $attribute1='$value1' AND
$attribute2='$value2'‘)
Appendix 9
119
120
TABLE 17 Literature works comparison
Appendix 9
EXISTING HOW THESIS SYSTEM
# AREA SOLUTIONS ADVANTAGE DISADVANTAGE DIFFERS?
1 NLQ into SQL Authoring • Uses semantic grammar • Relies heavily on end- • Only involves end users
mapping Interface Based specification, which is a user input throughout in the case of any NLQ
Approaches Systems language definition that multiple interface screens words spelling mistakes
provides accurate rules to modify the used or ambiguous phrases.
for linguistic expressions keywords or phrases. • For linguistic expressions
semantic parsing. • Requires extensive semantic parsing,
expertise time and NLP tools are used to
efforts to identify and lemmatize, tokenize,
specify RDB elements define and tag each NLQ
and concepts. token.
2 Enriching • Widely used in MLA • Requires extensive • The rule-based
the NLQ/ problems. manual rules defining observational algorithm
SQL Pairs via • Provides logical and customizing in case implemented is totally
Inductive Logic knowledge and of any DB change to domain-independent and
Programming reasoning. maintain accuracy. portable on any natural
language translation
framework.
• Adds extra metadata
to the NLQ/SQL pairs to
easily find a semantic
interpretation for NLQ’s
ambiguous phrases for
accurate mapping.
3 Using MLA • NLQ/SQL pairs’ corpora • Requires a huge domain • Uses simple
Algorithms induces semantic specific NLQ/SQL algorithmic rules and is
grammar parsing to map translation pairs’ corpora domain independent.
NLQs into their SQLs. that is manually written. • It does not assume
• Used by training a • Data preparation is prior knowledge of the
Support Vector Machine time consuming and a adopted RDB schema or
(SVM) classifier, which tedious task. require any annotated
is an efficient MLA • Requires a domain corpora for training
for high dimensional expert to train and test • NLQ/SQL pairs are only
datasets. the system. used for algorithm
• System is over- testing and validation
customized and purposes.
unfunctional on any • Focus is on
other domain. understanding the NLQ
• Assumes the user is to avoid potential future
familiar with the DB errors or jeopardize
schema, data and accuracy.
contents.
• Relying heavily on
MLAs are not effective
in decreasing the
translation error rates or
Appendix 9
increasing accuracy.
• SVM algorithm needs a
lot of memory space.
• SVM is not scalable to
larger DBs.
(Continued)
121
122
Appendix 9
TABLE 17 (Continued) Literature works comparison
4 Restricted NLQ • Uses a simple keyword- • Restricts the user to • The current work
Input based search structure. using certain domain- facilitates the interaction
• Uses a user-friendly specific keywords. between humans and
form or template based • Insignificant in terms of computers without NLQ
or menu based NLI to accuracy and recall. restrictions.
facilitate the mapping • Has portability problems • Has a limited interaction
process. even with advanced with the user to assure
algorithms such as the most natural way
similarity-based Top-k of communication,
algorithm. direct questioning and
answering, without
needing the user to
identify any NLQ tokens
semantic roles.
• Provides high accuracy
and recall.
• Compatible with any RDB
domain and a translation
environment.
5 Lambda Calculus • Uses NLQs meaning • Has some complicated • Uses a compatible
representation for the language logic. programming language,
mapping process. • Too abstract in many Python, that could
• A simple high-level cases. be translated to any
language model of • Very slow in execution. other language using
computation. • Hard to define rules with grammatical parse trees
its logical expressions. and language compilers.
• The current speed is an
average of 1.5 mins per
query.
6 Tree Kernels • Applies kernel functions • Requires a fully • Does not require an
Models on NLQ/SQL pairs syntactic annotated NLQ/SQL NLQ/SQL pairs corpus to
trees to learn its grammar. pairs corpus. develop.
• Applies linear kernels • Unable to recognise • The employed NLP tools
on a “bag-of-words” structural similarities and carefully understands the
to train the classifier to syntactic relations in an NLQ and recognizes its
select the correct SQLs NLQ. structural similarities and
for a given NLQ. • Has lower performance syntactic relations.
when scaled up to larger • Has an insignificantly
DBs. lower performance with
larger RDBs as well but is
still acceptable.
Appendix 9
(Continued)
123
124
Appendix 9
7 Unified Modeling • Used to model the DB’s • Limited to a few class • MetaTables and mapping
Language (UML) static relationships and diagram concepts tables are used. They
data models. (e.g., classes, accommodate any type
• Refers to the DB’s attributes, associations, and kind of data.
conceptual schema. aggregation and • Only NLQ input is
generalization). required from the user.
• The end user has to • Rule-based algorithm
identify classes and their is compatible with
constituents. all computational
• UML models environments.
visualization
requires compatible
environments.
8 Weighted Links • Uses the highest weight • Compromises accuracy • Accuracy and simplicity
meaningful joins for with complexity. are both the main focus
mapping between NLQ • Requires a huge in the current work.
tokens, RDB lexica and annotated training • No training dataset is
SQL clauses. dataset. needed.
• Computationally
expensive.
9 NLQ Tokens Morphological • Used for tokens • Requires a huge • The English word
into RDB Lexica and Word extraction. annotated training semantics dictionary
Mapping Group Analyzers • Analyses words’ dataset. (WordNet) is used to
(NLQ Tokens morphology. • Mapping accuracy is extract words’ semantic
Extraction) considerably low. information.
10 Pattern Matching • Used to find keywords • Requires a huge • NLQ tokens
types. annotated training extraction and their
• Facilitates learning other dataset. types identification
domains’ features. • Hard to analyse NLQ/ happens through NLP
SQL pairs mismatching computational linguistics
causes. processes, mainly the
Lemmatizer and the
tokenizer.
• The assumption-based
rules make it easy to find
out causes of NLQ/SQL
pairs mismatching.
11 Name Entity • Used to tokenize and • Restricted to only • While NER tagging is
Recognizer extract NLQ’s semantic recognize the NLQ part of the underlying
(NER) Alone information. tokens that already exist NLP tools, the main
with Coltech- in the NER resource. data source is acquired
Parser in GATE • Integrating data from from the NLQ and RDB
Appendix 9
external resources MetaTables.
is computationally
expensive.
(Continued)
125
126 Appendix 9
• Mapping accuracy is • MetaTables are used
considerably low. to check for tokens’
existence as a first goal,
then mapping them
to their logical role as
a relationship, table,
attribute or value.
• WordNet is used to
support the MetaTables
with words’ synonyms,
meanings and Lexical
Analysis.
12 Java Annotation • Used for NLQ • Less expressive than • No source trees are
Patterns tokenization and NER SQL-like languages. required except for the
Engine (JAPE) tagging. • It is memory extensive in NLQ tokens’ relations
Grammars • Generates a tag that it creates a whole analysis step.
probability distribution. structured source tree
• Applies a rich feature for every DB element.
representation.
13 Porter Algorithm • Used to extract tokens’ • Only supports few • The current work is
stems. languages. language independent.
• Does not require • Not a practical approach • Does not mandate the
knowledge structures’ as it requires a huge availability of a huge
reprocessing. memory to process. memory, except for the
• Has a high false positives storage of the RDB and
rate. the MetaTables.
• Hard to implement in
other languages.
14 Unification-Based • Extracts NLQ tokens • Long processing time. • Average processing time
Learning (UBL) using restricted lexical • Complicated nature of of 1.5 mins per query.
Algorithm items and Combinatory stemmer. • The NLP tools easily
Categorial Grammar • Has a high error rate in identify and recognize
(CCG) rules. recognizing NLQ noun tokens’ semantic roles
phrases. and their lexical relations.
• Difficulty in analyzing
tokens relations.
15 Dependency • Used to extract tokens • Potential data loss • NLTK parser parses the
Syntactic Parsing and their lexical during parse tree NLQ tokens according
relations. generation and to the built-in semantic
• Replaces parse trees with expansion. roles that are mapped to
dependency structures. • Eventual error specific RDBs elements.
Appendix 9
• Captures meaningful propagation while • A parse tree is generated
dependency relations applying the greedy and a dictionary of table
directly. parsing. names, attributes and
• Language dependent. tokens are maintained,
and NLQ’s subjects, objects
127
and verbs are identified.
(Continued)
128
Appendix 9
16 Separate Value • A compromising • Requires a big annotated • Supports RDB schema
and Table approach for not training dataset. elements’ MetaTables
Extractor supporting the RDB • Does not provide and synonyms for tokens’
Interfaces schema elements’ information on NLQ semantic information.
MetaTables and tokens’ semantic • No need for a rich
synonyms. relationships. annotated corpus of
• Ideal for complex NLQ/ • Requires a long time to NLQ/SQL pairs for
SQL pairs. process. algorithm training.
• Domain-independent
and configurable on any
working environment.
17 NLQ Tokens into Spider System • Uses a rich corpus • Uses a huge human • Focuses on simplicity and
RDB Lexica created using complex labeled NLQ/SQL corpus accuracy of the algorithm’s
Mapping and cross-domain for training and testing. mapping outcome with
(RDB Lexica semantic parsing and • Mapping accuracy is not highest priority.
Mapping) SQL patterns coverage. significantly high. • Uses NLQ MetaTable to
• An incremental map NLQ tokens into
approach, new RDB lexica.
experiences affect • The implemented
processing. MetaTables fill up the low
accuracy gap in language
translation algorithms
that do not use any sort
of deep DB schema data
dictionaries or just a
limited Data Dictionary.
18 WordNet alone • Efficient at expanding • Generalizes the relation • Supportive techniques are
NLQ predicate arguments and does employed in the current
arguments to their not guarantee NLQ’s research work such as the
meaning interpretations lack of ambiguity and disambiguation module.
and synonyms. noise which significantly • To avoid confusion
• Handles complex NLQs affects its meaning around the RDB unique
without an ontological interpretation. values, data profiling
distinction. is performed on RDB’s
statistics to automatically
compile the mapping
table of unique values,
PKs and FKs.
• Unique values, PKs
and FKs are stored in
a mapping table by
specifying their hosting
attributes and tables
while a hashing function
is used to access them
instantly.
(Continued)
Appendix 9
129
130 Appendix 9
19 NLQ Tokens into Stanford • Can parse any language • It is an outdated • NLQ can be in any
RDB Lexica Dependencies in any free word order. lexicalized parser, which form as long as it has
Mapping Parser • Displays all sentence leads to unnecessary correct spellings and no
(RDB Lexica structure and tokens errors. ambiguous tokens.
Relationships) dependencies. • Sentences must follow
• Uses parsing trees to Chomsky Normal Form
represent syntax and (CNF) style.
semantics.
20 Dependency • Simple and expressive. • Does not show any • The simplest and
Syntactic Parsing • Displays each token in semantic information. most effective way
the NLQ in a high level. • Some parsing trees of representing RDB
are erroneous in that elements relationships
they never lead to the is by restricting the RDB
targeted RDB elements. schema relationships
• Potential false early to be in the form of a
prune out. verb for easy mapping
between NLQ verbs and
RDB relationships.
21 Dependency- • Parsing and learning is • Requires manual • Simple rule-based
Based done using logical forms annotation of logical algorithm that uses the
Compositional (trees). forms and semantic semantic role of a verb to
Semantics (DCS) • Has rich a properties parsing. link RDB lexica with each
System Enriched set of computations, • Complex other.
with Prototype statistics and linguistics. implementation. • No manual annotation is
Triggers needed.
22 NLQ Tokens into Named Entity • Recognizes a wide • Does not remember • Previously tagged
RDB Lexica Tagger range of literal values previously tagged entity entity sets are saved
Mapping (NLP of named or numerical sets. (temporarily in case of
syntax and entity sets. • Supports limited limited storage) in the
semantics) languages. NLQ MetaTable.
• Does not show • Supports the NLQ’s
dependencies between syntactic and semantic
named entity sets. grammar analysis with
computational linguistics
algorithms in the form of
RDB and NLQ MetaTables
to assist tokens mapping
into RDB lexica.
• NLP syntactic and
semantic tools show the
source tables of each
token, which explains
tokens’ relationships.
23 Dependency • Produces parallel • Does not recognise • Considers understanding
Parser syntactic dependency complex language the NLQ, by finding
trees. phenomena. the combination of its
• Constructs dependency tokens’ meanings, is the
Appendix 9
trees directly without most essential part in the
any parse trees mapping and translation
conversions. process.
(Continued)
131
132 Appendix 9
• Employs computational
linguistic studies at the
words processing level is
employed.
• The current research
discovered common
semantics between NLQ
and SQL by analyzing the
language syntax roles.
24 LIFER/LADDER • Uses NLQ syntactic and • Not sufficient and • This research framework
Method semantic analysis alone. produces substantially overcomes any poor
• Simple and easy to low precision, FPR and underlying linguistic
implement. TNR. tools’ performance that
are meant to analyse NLQ
syntax and semantics
by using the supportive
MetaTables and WordNet
ontology. Such NLP tools
include named entity
tagger, tokenizer, or
dependency parser.
25 NLQ/SQL Syntax • Learns multiple NLQ • Limited to available data • RDB schema knowledge,
Trees Encoded syntactic features. resources. the semantic data
Via Kernel • Represents unrestricted • Expensive development models in the form
Functions features of domain- and high time of MetaTables, and
specific knowledge. consumption. syntactic-based analysis
• Does not show knowledge are used to
dependencies between generate parse trees from
named entity sets. the identified tokens
to properly map NLQ
tokens to the related RDB
elements.
26 The Probabilistic • Models NLQ features • Proved to be challenging • The current research
Context Free using production in terms of finding discovered common
Grammar rules with estimated the right grammar for semantics between NLQ
(PCFG) Method probabilities. optimization. and SQL by analyzing the
• Uses overlapping and • Iterative production language syntax roles.
interdependent features rules lead to inherited • Does not require
to build its probability computational annotated datasets.
models. complexity.
• Requires an annotated
training and testing
dataset.
Appendix 9
(Continued)
133
134 Appendix 9
27 RDB Lexica into The Extended • Extracts fuzzy tokens • Has a high False Positive • Uses computational
SQL Clauses UML Class semantic roles in a form Ratio. linguistics mapping
Mapping Diagrams of a validation sub- constraints to transform
(SQL clauses Representations graph or tree of the Self lexica into SQL clauses
mapping) Organizing Maps (SOM) and keywords.
diagram representation • Computational linguistics
that transforms class is used here in the form
diagrams into SQL of linguistics-based
clauses using the fuzzy mapping constraints
set theory. using a manually written
• More flexible than the rule-based algorithm.
MLA approaches. • Those algorithms are
• Provides higher mainly observational
measures of recall. assumptions.
• The MetaTable specifies
RDB schema categories
(value, relationship,
attribute, etc.) to map
the identified RDB lexica
into SQL clauses and
keywords.
• Provides high measures
or accuracy and recall.
28 RDB Relationships • Relationships are used to • The user has to • Current work uses
and Linguistic map the lexica into NLQs identify the source, RDB relationships
Analysis linguistic semantic roles’ its associations or and NLP tools, which
classes as a conceptual relationships and the are more capable of
data model. target in the fuzzy “understanding” the
• The mapping results NLQ to connect them NLQ statement before
are derived from the for UML class diagram translating it to an SQL
connectivity matrix extraction. query.
by searching for any • Extraction is not • This method highly
existing logical path thorough or exhaustive. contributes to the
between the source increase in the translation
(objects) and the target accuracy.
(attributes) to eventually • Regarding the linguistic
map the logical path to inter-relationships within
an equivalent SQL. the RDB schema in the
current work, not only
WordNet is used, but also
an NLTK and NLP tools.
Besides, a manual rule-
based algorithm is also
used to define how NLQ
linguistic roles match with
Appendix 9
the RDB elements, which
explains the variance in
translation accuracy and
precision in comparison.
(Continued)
135
136
Appendix 9
• Assures a seemingly
natural interaction
between the user and
the computer. Hence,
the user does not have
to identify any semantic
roles in their NLQ. The
underlying NLP tools does
this for them.
• The relationships are
identified by the NLQ
verbs, so the user is
communicating more
information in their NLQ
using the current research
algorithm compared to
the other literature works.
Hence, it is considered
more advanced and user
friendly.
• Not only objects and
attributes are extracted
from the NLQ, the
proposed research work
extracts much lower level
linguistic and semantic
roles such as gerunds
and prepositions which
help select the matching
RDB lexica with higher
accuracy and precision.
29 RDB Lexica into L2S System • Compares all existing • A complicated system • Considered significantly
SQL Clauses NLQ tokens with existing that consumes a lot of simpler than most
Mapping DB elements using NLP time to run through complex mapping
(Complexity vs tools, tokens’ semantic all DB elements for approaches as it relies on
Performance) mapper and a graph- comparison with NLQ fewer, but more effective,
based matcher. tokens. underlying linguistic tools
• Computationally and mapping rules.
expensive.
30 Bipartite Tree-Like • Employs sophisticated • A complicated system • The current work is
Graph-Based semantic and syntactic that requires a domain- the best in terms of
Processing analysis of the input specific background performance, simplicity
Model NLQ. knowledge and a and adaptability to
thorough training and different framework
testing dataset. environments and RDB
domains.
• The use of MetaTables to
Appendix 9
define the lexica semantic
roles and their adjacent
SQL slots for better
mapping accuracy.
(Continued)
137
138
Appendix 9
31 Ellipsis Method • Deals with instances • Requires that NLQ be • Employs a lightweight
of ellipsis (less than a explained by the user. approach for query
sentence). • A memory-based translations with no
• Uses a computationally learning method. need to storage spaces
cheap and robust • Produces a few other than storing the
approach. mismatched SQLs. MetaTables and the
mapping tables.
32 The Highest • Automatically discards • Relies heavily on labeled • No need for labeled
Possibility any features that are not training and testing data for developing and
Selection necessary. data, which is expensive executing.
• Memorizes and searches and tedious to create. • Can be generalizable
previous NLQ encounters • Not generalizable across across domains.
to find relatable other domains.
features.
33 Weighted Neural • Generates ordered • Computationally • The translation accuracy
Networks and weighted SQLs expensive. of this algorithm still falls
and Stanford schemata. • Unscalable to bigger behind the proposed
Dependencies • Uses linguistics in their RDBs. algorithm because of the
Collapsed (SDC) algorithm. use of further semantic
roles and linguistic
categories (i.e., adjectives,
pronouns etc.).
• Uses NLQ’s subject • Prioritizes SQLs based on • Uses the verbs to find the
or object to search probability of correctness attributes’ and values’
the DB for matching instead of accuracy and relationships instead of
attributes with weighted precision. using a heavy weighted
projection-oriented tool such as the weighted
stems and generate the projection-oriented stems.
SQL clauses accordingly.
34 Pattern Matching • Used as a mapping • Does not apply any NLQ • Both implemented
of SQL algorithm. interpretation modules mappers have access to
• Easy to develop and or parsing elaborations. an embedded linguistic
execute. • Its translation accuracy semantic-role frame
• Handles mixed data and overall performance schema (WordNet
types of NLQ tokens. is highly jeopardized. and Stanford CoreNLP
Toolkit), data dictionary
(MetaTables) and the RDB
schema. Those resources
are essential for accurate
SQL query formation and
generation.
(Continued)
Appendix 9
139
140
Appendix 9
35 RDB Lexica into NLQ Conceptual • A concept-based query • SQLs are constructed • Simplifies SQL queries
SQL Clauses Abstraction language to facilitate from scratch, which adds generation by using
Mapping (SQL SQL formulation. extra computational ready SQL templates.
Formation vs • Scalable to large complexity to the • SQL construction
SQL Templates) datasets. language translation constraints are used in
system. the mapping algorithm to
• Adds an additional guarantee accurate SQL
unnecessary layer on top template selection.
of the original system • This approach is
architecture. considered as a simple
and accurate method of
generating SQLs.
36 Semantic • Used to store all • Due to this system’s • Accommodates more
Grammar grammatical words to be complexity, this complex SQL types such
Analysis used for mapping NLQ’s architecture can only as nested or cascaded
intermediate semantic translate simple NLQs. SQLs.
representation into SQL • Not flexible with nested
clauses. or cascaded SQLs.
37 Kernel Functions, • Used to classify NLQ/ • Achieves low recall of • No need to develop
SVM Classifier, SQL pairs as correct or correctly retrieved SQL a training and testing
and the incorrect. answers. datasets of NLQ/SQL pairs
Statistical • The mapping algorithm • Requires labeled training for every new domain.
and Shallow is at the syntactic level. and testing datasets of • High recall due to the use
Charniak’s • Uses NLQ semantics to NLQ/SQL pairs. of an accurate mapping
Syntactic Parser build syntactic trees to • Such exclusive domain- algorithm mapped to
select SQLs according to specific systems are ready SQL templates.
their probability scores. highly expensive.
• A parser is applied to • Its performance is
compute the number subjective to the
of shared high-level accuracy and correctness
semantics and common of the training and
syntactic substructures testing datasets, which
between 2 trees to are manually written by
produce the union of a human domain expert.
the shallow feature
spaces.
(Continued)
Appendix 9
141
142 Appendix 9
38 Heuristic • NLQ/SQL pairs syntactic • There is no use of any • Uses simple algorithmic
Weighting trees are used as an SQL NLQ annotated meaning rules based on
Scheme compiler to derive NLQ resources or manual computational linguistics
parsing trees. semantic interpretation to fully understand the
• NLQ tokens’ lexical and representation to input NLQ to ensure
dependencies, DB fully understand the highest translation
schema and some NLQ. accuracy.
synonym relations are • The SQL generator • RDB MetaTable is used
used to map DB lexica performance is for lexical relations
with the SQL clauses. considerably low. disambiguation.
• SQL generator is built • A mapping table is also
from scratch, which used, which includes RDB
adds high complexity to lexica data types, PKs and
the language translation FKs and names of entity
algorithm. sets (unique values),
in addition to other
rule-based mapping
constraints.
• Uses SQL templates and
puts extra focus on passing
accurate RDB lexia into
SQL templates generator
for better performance
and correct output.
39 A Deep Sequence • Generates an SQL from • Incompatible with cross • Uses manually written
to Sequence NLQ semantic parsing. entropy loss optimization rule-based grammar for
Neural Network • Uses reinforcement training tasks. the mapping.
learning and rewards • Requires manually • Produces high measures
from in-the-loop query annotated NLQ/SQL of accuracy and recall.
execution to learn an pairs for generating the • No need for training and
SQL generation policy. SQL conditions. testing labeled data.
• Execution accuracy is as
low as 59.4% and logical
form accuracy is 48.3%.
• Proved to be inefficient
and unscalable on large
RDBs.
40 MLA Sequence- • Employs a mapping • Has very low • Translates NLQs into
To-Sequence- algorithm without performance measures. SQLs while maintaining
Style Model reinforcement learning. • Had to use dependency high simplicity and
• Showed small graphs and the column performance.
improvements to attention mechanism • The system does not
generate SQL queries for performance need to be updated or
when order does not improvement. maintained periodically.
matter. • The model has to
• Solves the “order- be frequently and
Appendix 9
matters” design periodically retrained to
problem. reflect the latest dataset
updates, which increases
the system’s maintenance
costs and computational
143
complexity.
(Continued)
144
Appendix 9
41 A Deep-Learning- • Predicts and generates • NLQ/SQL pairs were • Uses RDBs that are public
Based Model the SQL directly for any manually written for source namely, Zomato
given NLQ. model training and and WikiSQL, only for
• Uses attentive-copying testing. algorithm validation and
mechanism, a recover • The used RDB is testing.
technique and task- specifically customized • Does not need labeled
specific look-up tables, to the used framework data for developing.
to edit the generated and environment applied • The translator algorithm
SQL. on. is domain-independent
• Overcomes the • Highly questionable in and configurable on any
shortcomings of terms of generalizability, other environment.
sequence-to-sequence applicability and
models. adaptability on other
• Proved its flexibility and domains.
efficiency
42 RDB Lexica into Regular • Represents NLP • regexps collections in • Uses NLQ MetaTables
SQL Clauses Expressions tokens phonology and NLQ sentences are not and RDB MetaTables
Mapping (regexps) morphology. clearly articulated in the to increase accuracy of
(Intermediate • Uses NLQ intermediate literature. mappings between NLQ
Representation) semantic representation • Proved to be not as tokens into RDB lexica
layers to represent NLQ effective as the NLP and then into SQL clauses.
lexica as SQL clauses. tools, MetaTables and • Tries to save every
• Tokens representation mapping tables in possible information
happens by applying terms of accuracy and given by the NLQ so
First Order Predicate precision. as each NLQ token is
Calculus Logic resembled used and represented
by DB-Oriented Logical in the SQL clauses and
Form (DBLF) and expressions production.
Conceptual Logical Form • Uses multiple NLP tools,
(CLF) with some SQL MetaTables and mapping
operators and functions tables for unique values
to build and generate to fully understand the
SQLs. NLQ and map its tokens
to their corresponding
RDB elements.
(Continued)
Appendix 9
145
146 Appendix 9
43 The Similarity- • Processes NLQ tokens • Proved to be not • No conceptual
Based Top-K to map them to their effective because of representation is needed
Algorithm internal conceptual the high complexity for the mapping.
representation layer. and time consumption • The identified attributes
• Uses Entity-Attribute- approaches applied. are automatically mapped
Value (EAV) DB into the SQL SELECT
metadata and clause, while the tables
grammatical parse trees. are extracted from the
SELECT clause to generate
an SQL FROM clause, and
the values are used as a
conditional statement in
the WHERE clause.
44 Lambda-Calculus • Uses lambda-calculus • A supervision-extensive • No meaning
to map tokens to their system. representation is needed
corresponding meaning for the mapping.
representations. • Does not require any
human supervision to
properly function.
45 An Intermediate • Transforms DB lexica into • Computationally • No internal graphical
Tree-Like Graph an intermediate tree-like expensive and representation is needed
graph. processing is time for the mapping.
• Extracts the SQL from consuming. • Processing the mapping
the maximum bipartite has an average speed of
matching algorithm. 1.5 mins.
Glossary
AI Artificial Intelligence
API Application Program Interface
CSV Comma-Separated Values
DAC Data Administration Commands
DB DataBase
DBMS Database Management System
DCL Data Control Language
DDL Data Definition Language
DML Data Manipulation Language
DQL Data Query Language
EAV Entity-Attribute-Value
ERD Entity-Relational Diagram
FKs Foreign Keys
FNR False Negative Ratio
FPR False Positive Ratio
IDE Integrated Development Environment
MLA Machine Learning Algorithm
NER Named Entity Recognition
NL Natural Language
NLI Natural Language Interface
NLIDB Natural Language Interface for DataBase
NLP Natural Language Processing
NLQ Natural Language Question
NLTK Natural Language Toolkit
NLTSQLC NL into SQL Convertor
NoSQL Not Only Structured Query Language
OLTP Online Transactional Processing
PKs Primary Keys
POS Part of Speech
PTSD Post-Traumatic Stress Disorder
QAS Question Answering Systems
QL Query Language
RA Relational Algebra
RDB Relational DataBase
147
148 Glossary
RDBMS Relational Database Management Systems

ROC Receiver Operating Characteristics
SQL Structured Query Language
TCL Transaction Control Language
TNR True Negative Ratio
TPR True Positive Ratio
References
[1] Nihalani, N. (2010). An intelligent interface for relational databases. Human–
Computer Interaction, 6, 7.
[2] Natural language processing (NLP), Wikipedia. Last accessed March 29, 2020.
https://en.wikipedia.org/wiki/Natural_language_processing
[3] Kaur, J., Chauhan, B., & Korepal, J. K. (2013). Implementation of query proces-
sor using automata and natural language processing. International Journal of
Scientific and Research Publications, 3(5), 1–5.
[4] Bronnenberg, W. J. H. J., Landsbergen, S., Scha, R., Schoenmakers, W., & van
Utteren, E. (1978). PHLIQA-1, a question-answering system for data-base con-
sultation in natural English. Philips Technical Review, 38, 229–239.
[5] Aron, J. (2011). How innovative is Apple’s new voice assistant, Siri? 24.
[6] Saggion, H. (2006, May). Multilingual multidocument summarization tools and
evaluation. In LREC (pp. 1312–1317).
[7] Wong, Y. W., & Mooney, R. J. (2006, June). Learning for semantic parsing
with statistical machine translation. In Proceedings of the Main Conference on
Human Language Technology Conference of the North American Chapter of
the Association of Computational Linguistics (pp. 439–446). Association for
Computational Linguistics.
[8] Patil, S., & Davies, P. (2014). Use of Google Translate in medical communica-
tion: Evaluation of accuracy. BMJ, 349, g7392.
[9] Eckert, M., Bry, F., Brodt, S., Poppe, O., & Hausmann, S. (2011). A CEP
babelfish: Languages for complex event processing and querying surveyed. In
Reasoning in event-based distributed systems (pp. 47–70). Springer, Berlin,
Heidelberg.
[10] Aggarwal, C. C., & Zhai, C. (2012). A survey of text classification algorithms. In
Mining text data (pp. 163–222). Springer, Boston, MA.
[11] Chorowski, J. K., Bahdanau, D., Serdyuk, D., Cho, K., & Bengio, Y. (2015).
Attention-based models for speech recognition. In Advances in neural infor
mation processing systems (pp. 577–585).
[12] Sanjay Agrwal, S. C. D. (2002). DBXplorer: A system for keyword base search
over relational database. In Proceedings of 18th International Conference on
Data Engineering. IEEE.
[13] Chen, P. P. S. (1976). The entity–relationship model – toward a unified view of
data. ACM Transactions on Database Systems (TODS), 1(1), 9–36.
[14] Tseng, F. S., Chen, A. L., & Yang, W. P. (1992). On mapping natural lan-
guage constructs into relational algebra through ER representation. Data and
Knowledge Engineering, 9(1), 97–118.
149
150 References
[15] Queralt, A., & Teniente, E. (2006, November). Reasoning on UML class dia-
grams with OCL constraints. In International Conference on Conceptual
Modeling (pp. 497–512). Springer, Berlin, Heidelberg.
[16] Grosz, B. J., Appelt, D. E., Martin, P. A., & Pereira, F. C. (1987). TEAM: An
experiment in the design of transportable natural-language interfaces. Artificial
Intelligence, 32(2), 173–243.
[17] Owei, V., Rhee, H. S., & Navathe, S. (1997). Natural language query filtra-
tion in the conceptual query language. In Proceedings of the Thirtieth Hawaii
International Conference on System Sciences (Vol. 3, pp. 539–549). IEEE.
[18] Nguyen, D. T., Hoang, T. D., & Pham, S. B. (2002). A Vietnamese natural lan-
guage interface to database. In 2012 IEEE Sixth International Conference on
Semantic Computing (pp. 130–133). IEEE, China.
[19] Miller, G. A. (1995). WordNet: A lexical database for English. Communications
of the ACM, 38(11), 39–41.
[20] Sleator, D. D., & Temperley, D. (1995). Parsing English with a link grammar.
arXiv preprint cmp-lg/9508004.
[21] Stanford CoreNLP. (2014). Stanford CoreNLP – Natural language software.
Last accessed December 23, 2019. Retrieved from: https://stanfordnlp.github.io/
CoreNLP/
[22] Johnstone, B. (2018). Discourse analysis. John Wiley & Sons.
[23] Leech, G. N. (2016). Principles of pragmatics. Routledge.
[24] Giordani, A., & Moschitti, A. (2009, June). Semantic mapping between natu-
ral language questions and SQL queries via syntactic pairing. In International
Conference on Application of Natural Language to Information Systems
(pp. 207–221). Springer, Berlin, Heidelberg.
[25] Zhang, J., Scardamalia, M., Reeve, R., & Messina, R. (2009). Designs for collec-
tive cognitive responsibility in knowledge-building communities. The Journal of
the Learning Sciences, 18(1), 7–44.
[26] Gallè, F., Mancusi, C., Di Onofrio, V., Visciano, A., Alfano, V., Mastronuzzi, R.,
… & Liguori, G. (2011). Awareness of health risks related to body art practices
among youth in Naples, Italy: A descriptive convenience sample study. BMC
Public Health, 11(1), 625.
[27] Safari, L., & Patrick, J. D. (2019). An enhancement on Clinical Data Analytics
Language (CliniDAL) by integration of free text concept search. Journal of
Intelligent Information Systems, 52(1), 33–55.
[28] Kando, N. (1999, November). Text structure analysis as a tool to make retrieved
documents usable. In Proceedings of the 4th International Workshop on Infor
mation Retrieval with Asian Languages (pp. 126–135).
[29] Zhang, M., Zhang, J., & Su, J. (2006, June). Exploring syntactic features for table
extraction using a convolution tree kernel. In Proceedings of the Main Conference
on Human Language Technology Conference of the North American Chapter
of the Association of Computational Linguistics (pp. 288–295). Association for
[30] Sagar, R. (2020, June 3). OpenAI releases GPT-3, the largest model so far.
Analytics India Magazine. Retrieved October 14, 2020.
[31] Chalmers, David (2020, July 30). “GPT-3 and General Intelligence”. Daily
Nous. Last accessed October 14, 2020. Retrieved from: https://dailynous.
com/2020/07/30/philosophers-gpt-3/#chalmers
References 151
[32] OpenAI, Discovering and enacting the path to safe artificial general intelligence,
2020. Last accessed October 13, 2020. Retrieved from: https://openai.com/
[33] Bybee, J. L. (1985). Morphology: A study of the relation between meaning and
form (Vol. 9). John Benjamins Publishing.
[34] Smith, N. V. (1973). The acquisition of phonology: A case study. Cambridge
University Press.
[35] Stetson, R. H. (2014). Motor phonetics: A study of speech movements in action.
Springer.
[36] Zhang, D., & Lee, W. S. (2003, July). Question classification using support vec-
tor machines. In Proceedings of the 26th Annual International ACM SIGIR
Conference on Research and Development in Information Retrieval (pp. 26–32).
ACM.
[37] Iftikhar, A., Iftikhar, E., & Mehmood, M. K. (2016, August). Domain specific
query generation from natural language text. In 2016 Sixth International
Conference on Innovative Computing Technology (INTECH) (pp. 502–506).
IEEE.
[38] Kumar, R., & Dua, M. (2014, April). Translating controlled natural language
query into SQL query using pattern matching technique. In International
Conference for Convergence for Technology–2014 (pp. 1–5). IEEE.
[39] Boyd-Graber, J., Fellbaum, C., Osherson, D., & Schapire, R. (2006, January).
Adding dense, weighted connections to WordNet. In Proceedings of the Third
International WordNet Conference (pp. 29–36).
[40] NLTK 3.4.5 Documentation. (2019). Natural language toolkit. Last accessed
December 23, 2019. Retrieved from: http://www.nltk.org/
[41] Rish, I. (2001, August). An empirical study of the naive Bayes classifier. In IJCAI
2001 Workshop on Empirical Methods in Artificial Intelligence (Vol. 3, No. 22,
pp. 41–46).
[42] Safari, L., & Patrick, J. D. (2018). Complex analyses on clinical information
systems using restricted natural language querying to resolve time-event depen-
dencies. Journal of Biomedical Informatics, 82, 13–30.
[43] Ganti, V., He, Y., & Xin, D. (2010). Keyword++: A framework to improve
keyword search over entity databases. Proceedings of the VLDB Endowment,
3(1–2), 711–722.
[44] Woods, W. A. (1981). Procedural semantics as a theory of meaning. Bolt Beranek
and Newman Inc.
[45] Kaur, S., & Bali, R. S. (2012). SQL generation and execution from natural lan-
guage processing. International Journal of Computing & Business Research,
2229–6166.
[46] Bhadgale Anil, M., Gavas Sanhita, R., Pati Meghana, M., & Pinki, R. (2013).
Natural language to SQL conversion system. International Journal of Computer
Science Engineering and Information Technology Research (IJCSEITR), 3(2),
161–166, ISSN 2249-6831.
[47] Popescu, A. M., Etzioni, O., & Kautz, H. (2003, January). Towards a theory of
natural language interfaces to databases. In Proceedings of the 8th International
Conference on Intelligent User Interfaces (pp. 149–157). ACM.
[48] Parlikar, A., Shrivastava, N., Khullar, V., & Sanyal, S. (2005). NQML: Natural
query markup language. In 2005 International Conference on Natural Language
Processing and Knowledge Engineering (pp. 184–188). IEEE.
152 References
[49] Peng, Z., Zhang, J., Qin, L., Wang, S., Yu, J. X., & Ding, B. (2006, September).
NUITS: A novel user interface for efficient keyword search over databases. In
Proceedings of the 32nd International Conference on Very Large Data Bases
(pp. 1143–1146). VLDB Endowment.
[50] Feijs, L. M. G. (2000). Natural language and message sequence chart representa-
tion of use cases. Information and Software Technology, 42(9), 633–647.
[51] Karande, N. D., & Patil, G. A. (2009). Natural language database interface
for selection of data using grammar and parsing. World Academy of Science,
Engineering and Technology, 3, 11–26.
[52] El-Mouadib, F. A., Zubi, Z. S., & Almagrous, A. A. (2009). Generic interactive
natural language interface to databases (GINLIDB). International Journal of
Computers, 3(3).
[53] Li, H., & Shi, Y. (2010, February). A wordnet-based natural language interface
to relational databases. In 2010 The 2nd International Conference on Computer
and Automation Engineering (ICCAE) (Vol. 1, pp. 514–518). IEEE.
[54] Enikuomehin, A. O., & Okwufulueze, D. O. (2012). An algorithm for solving
natural language query execution problems on relational databases. International
Journal of Advanced Computer Science and Applications, 3(10), 180–182.
[55] Chen, P. P. S. (1983). English sentence structure and entity-relationship dia-
grams. Information Sciences, 29(2–3), 127–149.
[56] QUEST/a natural language interface to relational databases (2018). In Proceedings
of the Eleventh International Conference on Language Resources and Evaluation
(LREC).
[57] Desai, B., & Stratica, N. (2004). Schema-based natural language semantic map-
ping. In Proceedings of the 9th International Conference on Applications of
Natural Language to Information Systems.
[58] Becker, T. (2002, May). Practical, template–based natural language generation
with tag. In Proceedings of the Sixth International Workshop on Tree Adjoining
Grammar and Related Frameworks (TAG+ 6) (pp. 80–83).
[59] Androutsopoulos, I., Ritchie, G. D., & Thanisch, P. (1995). Natural language
interfaces to databases–an introduction. Natural Language Engineering, 1(1),
29–81.
[60] Patrick, J., & Li, M. (2010). High accuracy information extraction of medica-
tion information from clinical notes: 2009 i2b2 medication extraction challenge.
Journal of the American Medical Informatics Association, 17(5), 524–527.
[61] Chaudhari, P. (2013). Natural language statement to SQL query translator.
International Journal of Computer Applications, 82(5), 18–22.
[62] Gao, K., Mei, G., Piccialli, F., Cuomo, S., Tu, J., & Huo, Z. (2020). Julia lan-
guage in machine learning/algorithms, applications, and open issues. Computer
Science Review, 37, 100254.
[63] Hasan, R., & Gandon, F. (2014, August). A machine learning approach to sparql
query performance prediction. In 2014 IEEE/WIC/ACM International Joint
Conferences on Web Intelligence (WI) and Intelligent Agent Technologies (IAT)
(Vol. 1, pp. 266–273). IEEE.
[64] Wang, H., Ma, C., & Zhou, L. (2009, December). A brief review of machine
learning and its application. In 2009 International Conference on Information
Engineering and Computer Science (pp. 1–4). IEEE.
References 153
[65] Boyan, J., Freitag, D., & Joachims, T. (1996, August). A machine learning archi-
tecture for optimizing web search engines. In AAAI Workshop on Internet Based
Information Systems (pp. 1–8).
[66] Chen, H., Shankaranarayanan, G., She, L., & Iyer, A. (1998). A machine learn-
ing approach to inductive query by examples: An experiment using relevance
feedback, ID3, genetic algorithms, and simulated annealing. Journal of the
American Society for Information Science, 49(8), 693–705.
[67] Hazlehurst, B. L., Burke, S. M., & Nybakken, K. E. (1999). Intelligent query
system for automatically indexing information in a database and automatically
categorizing users. U.S. Patent No. 5,974,412. Washington, DC: U.S. Patent and
Trademark Office. WebMD Inc.
[68] Hofmann, T. (2001). Unsupervised learning by probabilistic latent semantic
analysis. Machine Learning, 42(1–2), 177–196.
[69] Popov, B., Kiryakov, A., Ognyanoff, D., Manov, D., & Kirilov, A. (2004). KIM –
A semantic platform for information extraction and retrieval. Journal of Natural
Language Engineering, 10(3–4), 375–392.
[70] Alexander, R., Rukshan, P., & Mahesan, S. (2013). Natural language web inter-
face for database (NLWIDB). arXiv preprint arXiv:1308.3830.
[71] Zhou, Q., Wang, C., Xiong, M., Wang, H., & Yu, Y. (2007). SPARK: Adapting
keyword query to semantic search. In The semantic web (pp. 694–707). Springer,
Berlin, Heidelberg.
[72] Bergamaschi, S., Domnori, E., Guerra, F., Orsini, M., Lado, R. T., & Velegrakis,
Y. (2010). Keymantic: Semantic keyword-based searching in data integration
systems. Proceedings of the VLDB Endowment, 3(1–2), 1637–1640.
[73] Strötgen, J., & Gertz, M. (2010, July). HeidelTime: High quality rule-based
extraction and normalization of temporal expressions. In Proceedings of the 5th
International Workshop on Semantic Evaluation (pp. 321–324). Association for
[74] Sohn, S., Wagholikar, K. B., Li, D., Jonnalagadda, S. R., Tao, C., Komandur
Elayavilli, R., & Liu, H. (2013). Comprehensive temporal information detection
from clinical text: Medical events, time, and TLINK identification. Journal of
the American Medical Informatics Association, 20(5), 836–842.
[75] Zhou, L., Friedman, C., Parsons, S., & Hripcsak, G. (2005). System architec-
ture for temporal information extraction, representation and reasoning in clinical
narrative reports. In AMIA Annual Symposium Proceedings (Vol. 2005, p. 869).
American Medical Informatics Association.
[76] Giordani, A., & Moschitti, A. (2012, June). Generating SQL queries using natu-
ral language syntactic dependencies and metadata. In International Conference
on Application of Natural Language to Information Systems (pp. 164–170).
Springer, Berlin, Heidelberg.
[77] Giordani, A., & Moschitti, A. (2010, May). Corpora for Automatically Learning
to Map Natural Language Questions into SQL Queries. In LREC.
[78] Kate, R. J., & Mooney, R. J. (2006, July). Using string-kernels for learn-
ing semantic parsers. In Proceedings of the 21st International Conference on
Computational Linguistics and the 44th Annual Meeting of the Association
for Computational Linguistics (pp. 913–920). Association for Computational
Linguistics.
154 References
[79] Tseng, F. S., & Chen, C. L. (2006, September). Extending the UML concepts to
transform natural language queries with fuzzy semantics into SQL. Information
and Software Technology, 48(9), 901–914.
[80] Booch, G. (2005). The unified modeling language user guide. Pearson Education
India.
[81] Oestereich, B. (2002). Developing software with UML: Object-oriented analysis
and design in practice. Pearson Education.
[82] Higa, K., & Owei, V. (1991, January). A data model driven database query tool.
In Proceedings of the Twenty-Fourth Annual Hawaii International Conference
on System Sciences (Vol. 3, pp. 53–62). IEEE.
[83] Muller, R. J. (1999). Database design for smarties: Using UML for data model-
ing. Morgan Kaufmann.
[84] Winston, P. H. (1992). Artificial intelligence. Addison-Wesley Longman
Publishing Co., Inc.
[85] Schmucker, K. J., & Zadeh, L. A. (1984). Fuzzy sets natural language computa-
tions and risk analysis. Rockville Md/ Computer Science Press.
[86] Zadeh, L. A. (1975). The concept of a linguistic variable and its application to
approximate reasoning-I. Information Sciences, 8, 199–249.
[87] Isoda, S. (2001). Object-oriented real-world modeling revisited. Journal of
Systems and Software, 59(2), 153–162.
[88] Moreno, A. M., & Van De Riet, R. P. (1997, June). Justification of the equivalence
between linguistic and conceptual patterns for the object model. In Proceedings
of the International Workshop on Applications of Natural Language to Infor
mation Systems, Vancouver.
[89] Métais, E. (2002). Enhancing information systems management with natu-
ral language processing techniques. Data & Knowledge Engineering, 41(2–3),
247–272.
[90] Yager, R. R., Reformat, M. Z., & To, N. D. (2019). Drawing on the iPad to input
fuzzy sets with an application to linguistic data science. Information Sciences,
479, 277–291.
[91] Owei, V., Navathe, S. B., & Rhee, H. S. (2002). An abbreviated concept-based
query language and its exploratory evaluation. Journal of Systems and Software,
63(1), 45–67.
[92] Hoang, D. T., Le Nguyen, M., & Pham, S. B. (2015, October). L2S: Transforming
natural language questions into SQL queries. In 2015 Seventh International
Conference on Knowledge and Systems Engineering (KSE) (pp. 85–90).
IEEE.
[93] Costa, P. D., Almeida, J. P. A., Pires, L. F., & van Sinderen, M. (2008,
November). Evaluation of a rule-based approach for context-aware services. In
IEEE GLOBECOM 2008 – 2008 IEEE Global Telecommunications Conference
(pp. 1–5). IEEE.
[94] Garcia, K. K., Lumain, M. A., Wong, J. A., Yap, J. G., & Cheng, C. (2008,
November). Natural language database interface for the community based moni-
toring system. In Proceedings of the 22nd Pacific Asia Conference on Language,
Information and Computation (pp. 384–390).
[95] International Conference on Applications of Natural Language to Information
Systems (13th: 2008: London, England). (2008). Natural Language and Information
References 155
Systems 13th International Conference on Applications of Natural Language to

Information Systems, NLDB 2008 London, UK, June 24–27, 2008, Proceedings.
Berlin, Heidelberg: Springer Berlin Heidelberg: Imprint: Springer.
[96] Waltz, D. L. (1978). An English language question answering system for a large
relational database. Communications of the ACM, 21(7), 526–539.
[97] Nguyen, D. B., Hoang, S. H., Pham, S. B., & Nguyen, T. P. (2010, March).
Named entity recognition for Vietnamese. In Asian Conference on Intelligent
Information and Database Systems (pp. 205–214). Springer, Berlin, Heidelberg.
[98] Cunningham, H., Maynard, D., Bontcheva, K., & Tablan, V. (2002, July). GATE:
An architecture for development of robust HLT applications. In Proceedings
of the 40th Annual Meeting on Association for Computational Linguistics
(pp. 168–175). Association for Computational Linguistics.
[99] Sathick, K. J., & Jaya, A. (2015). Natural language to SQL generation for seman-
tic knowledge extraction in social web sources. Indian Journal of Science and
Technology, 8(1), 1–10.
[100] Satav, A. G., Ausekar, A. B., Bihani, R. M., & Shaikh, A. (2014). A proposed
natural language query processing system. International Journal of Science and
Applied Information Technology, 3(2). http://warse.org/pdfs/2014/ijsait01322014.
pdf
[101] Kaur, G. (2014). Usage of regular expressions in NLP. International Journal of
Research in Engineering and Technology IJERT, 3(1), 7.
[102] Agrawal, A. J., & Kakde, O. G. (2013). Semantic analysis of natural language que-
ries using domain ontology for information access from database. International
Journal of Intelligent Systems and Applications, 5(12), 81.
[103] Deshpande, A. K., & Devale, P. R. (2012). Natural language query processing
using probabilistic context free grammar. International Journal of Advances in
Engineering & Technology, 3(2), 568.
[104] Tamrakar, A., & Dubey, D. (2012). Query Optimisation using Natural Language
Processing 1.
[105] Michael, G. (2012). A Survey of Natural Language Processing Techniques for
the Simplification of User Interaction with Relational Database Management
Systems, California Polytechnic State University, San Luis Obispo.
[106] Rao, G., Agarwal, C., Chaudhry, S., Kulkarni, N., & Patil, D. S. (2010). Natural
language query processing using semantic grammar. International Journal on
Computer Science and Engineering, 2(2), 219–223.
[107] Ott, N. (1992). Aspects of the automatic generation of SQL statements in a natu-
ral language query interface. Information Systems, 17(2), 147–159.
[108] Petrick, S. R. (1984). Natural language database query systems. Technical
Report, RC 10508, IBM Thomas J. Watson Research Laboratory.
[109] Lehmann, H., Ott, N., & Zoeppritz, M. (1985). A multilingual interface to data-
bases. IEEE Database Engineering, 8(3), 7–13.
[110] Wong, Y. W., & Mooney, R. (2007, June). Learning synchronous grammars
for semantic parsing with lambda calculus. In Proceedings of the 45th Annual
Meeting of the Association of Computational Linguistics (pp. 960–967).
[111] Ge, R., & Mooney, R. (2005, June). A statistical semantic parser that integrates
syntax and semantics. In Proceedings of the Ninth Conference on Computational
Natural Language Learning (CoNLL – 2005) (pp. 9–16).
156 References
[112] Minock, M., Olofsson, P., & Näslund, A. (2008, June). Towards building robust
natural language interfaces to databases. In International Conference on Appli
cation of Natural Language to Information Systems (pp. 187–198). Springer,
Berlin, Heidelberg.
[113] Zettlemoyer, L. S., & Collins, M. (2012). Learning to map sentences to logi-
cal form: Structured classification with probabilistic categorial grammars. arXiv
preprint arXiv:1207.1420.
[114] Tang, L. R., & Mooney, R. J. (2001, September). Using multiple clause con-
structors in inductive logic programming for semantic parsing. In European
Conference on Machine Learning (pp. 466–477). Springer, Berlin, Heidelberg.
[115] Giordani, A., & Moschitti, A. (2012, December). Translating questions to SQL
queries with generative parsers discriminatively reranked. In Proceedings of
COLING 2012: Posters (pp. 401–410).
[116] De Marneffe, M. C., MacCartney, B., & Manning, C. D. (2006, May). Generating
typed dependency parses from phrase structure parses. In LREC (Vol. 6,
pp. 449–454).
[117] Joachims, T. (1998). Making large-scale SVM learning practical (No. 1998, 28).
Technical Report.
[118] Moschitti, A. (2006, September). Efficient convolution kernels for dependency
and constituent syntactic trees. In European Conference on Machine Learning
(pp. 318–329). Springer, Berlin, Heidelberg.
[119] Giordani, A., & Moschitti, A. (2009, September). Syntactic structural kernels
for natural language interfaces to databases. In Joint European Conference
on Machine Learning and Knowledge Discovery in Databases (pp. 391–406).
Springer, Berlin, Heidelberg.
[120] Lu, W., Ng, H. T., Lee, W. S., & Zettlemoyer, L. (2008, October). A generative
model for parsing natural language to meaning representations. In Proceedings
of the 2008 Conference on Empirical Methods in Natural Language Processing
(pp. 783–792).
[121] Liang, P., Jordan, M. I., & Klein, D. (2013). Learning dependency-based compo-
sitional semantics. Computational Linguistics, 39(2), 389–446.
[122] Clarke, J., Goldwasser, D., Chang, M. W., & Roth, D. (2010, July). Driving
semantic parsing from the world’s response. In Proceedings of the Fourteenth
Conference on Computational Natural Language Learning (pp. 18–27). Asso
ciation for Computational Linguistics.
[123] Kwiatkowski, T., Zettlemoyer, L., Goldwater, S., & Steedman, M. (2010, October).
Inducing probabilistic CCG grammars from logical form with higher-order
unification. In Proceedings of the 2010 Conference on Empirical Methods in
Natural Language Processing (pp. 1223–1233). Association for Computational
Linguistics.
[124] Xu, X., Liu, C., & Song, D. (2017). SQLnet: Generating structured queries
from natural language without reinforcement learning. arXiv preprint arXiv:
1711.04436.
[125] Thompson, B. H., & Thompson, F. B. (1985). ASK is transportable in half a
dozen ways. ACM Transactions on Information Systems, 3(2), 185–203.
References 157
[126] Kudo, T., Suzuki, J., & Isozaki, H. (2005, June). Boosting-based parse reranking
with subtree features. In Proceedings of the 43rd Annual Meeting on Association
for Computational Linguistics (pp. 189–196). Association for Computational
Linguistics.
[127] Toutanova, K., Markova, P., & Manning, C. (2004). The leaf path projection view
of parse trees: Exploring string kernels for HPSG parse selection. In Proceedings
of the 2004 Conference on Empirical Methods in Natural Language Processing
(pp. 166–173).
[128] Kazama, J. I., & Torisawa, K. (2005, October). Speeding up training with tree
kernels for node table labeling. In Proceedings of the Conference on Human
Language Technology and Empirical Methods in Natural Language Processing
(pp. 137–144). Association for Computational Linguistics.
[129] Gaikwad Mahesh, P. (2013). Natural language interface to database. Inter
national Journal of Engineering and Innovative Technology (IJEIT), 2(8), 3–5.
[130] Papalexakis, E., Faloutsos, C., & Sidiropoulos, N. D. (2012). ParCube: Sparse
parallelizable tensor decompositions. In Joint European Conference on Machine
Learning and Knowledge Discovery in Databases (pp. 521–536). Springer,
Berlin, Heidelberg.
[131] Safari, L., & Patrick, J. D. (2014). Restricted natural language based querying of
clinical databases. Journal of Biomedical Informatics, 52, 338–353.
[132] Chandra, Y., & Mihalcea, R. (2006). Natural language interfaces to databases,
University of North Texas (Doctoral dissertation, Thesis (MS)).
[133] Shen, L., Sarkar, A., & Joshi, A. K. (2003, July). Using LTAG based features in parse
reranking. In Proceedings of the 2003 Conference on Empirical Methods in Natural
Language Processing (pp. 89–96). Association for Computational Linguistics.
[134] Collins, M., & Duffy, N. (2002, July). New ranking algorithms for parsing
and tagging: Kernels over discrete structures, and the voted perceptron. In
Proceedings of the 40th Annual Meeting on Association for Computational
Linguistics (pp. 263–270). Association for Computational Linguistics.
[135] Kudo, T., & Matsumoto, Y. (2003, July). Fast methods for kernel-based text anal-
ysis. In Proceedings of the 41st Annual Meeting on Association for Computational
Linguistics-Volume 1 (pp. 24–31). Association for Computational Linguistics.
[136] Cumby, C. M., & Roth, D. (2003). On kernel methods for relational learning.
In Proceedings of the 20th International Conference on Machine Learning
(ICML-03) (pp. 107–114).
[137] Culotta, A., & Sorensen, J. (2004, July). Dependency tree kernels for table
extraction. In Proceedings of the 42nd Annual Meeting on Association for
Computational Linguistics (p. 423). Association for Computational Linguistics.
[138] Ghosal, D., Waghmare, T., Satam, S., & Hajirnis, C. (2016). SQL query for-
mation using natural language processing (NLP). International Journal of
Advanced Research in Computer and Communication Engineering, 5, 3.
[139] Zhang, J., Tang, J., Ma, C., Tong, H., Jing, Y., & Li, J. (2015, August). Panther:
Fast top-k similarity search on large networks. In Proceedings of the 21st ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining
(pp. 1445–1454).
158 References
[140] Ghosh, P. K., Dey, S., & Sengupta, S. (2014). Automatic SQL query formation
from natural language query. International Journal of Computer Applications,
975, 8887.
[141] Choudhary, N., & Gore, S. (2015, September). Impact of intellisense on the
accuracy of natural language interface to database. In 2015 4th International
Conference on Reliability, Infocom Technologies and Optimization (ICRITO)
(Trends and Future Directions) (pp. 1–5). IEEE.
[142] Willett, P. (2006). The Porter stemming algorithm: Then and now. Program.
[143] Yu, T., Zhang, R., Yang, K., Yasunaga, M., Wang, D., Li, Z., … & Zhang, Z.
(2018). Spider: A large-scale human-labeled dataset for complex and cross-
domain semantic parsing and text-to-SQL task. arXiv preprint arXiv:1809.08887.
[144] Nelken, R., & Francez, N. (2000, July). Querying temporal databases using
controlled natural language. In Proceedings of the 18th Conference on Compu
tational Linguistics – Volume 2 (pp. 1076–1080). Association for Computational
Linguistics.
[145] Naumann, F. (2014). Data profiling revisited. ACM SIGMOD Record, 42(4), 40–49.
[146] Singh, G., & Solanki, A. (2016). An algorithm to transform natural language into
SQL queries for relational databases. Selforganizology, 3(3), 100–116.
[147] Zadeh, L. A. (1965). Fuzzy sets. Information and Control, 8(3), 338–353.
[148] Zadeh, L. A. (1978). PRUF – A meaning representation language for natural
languages. International Journal of Man-Machine Studies, 10(4), 395–460.
[149] Dalrymple, M., Shieber, S. M., & Pereira, F. C. (1991). Ellipsis and higher-order
unification. Linguistics and Philosophy, 14(4), 399–452.
[150] Tanaka, H., & Guo, P. (1999). Portfolio selection based on upper and lower expo-
nential possibility distributions. European Journal of Operational Research,
114(1), 115–126.
[151] Kang, I.-S., Bae, J.-H., & Lee, J.-H. Database semantics representation for natu-
ral language access. In Proceedings of the First International Symposium on
Cyber Worlds (CW ’02), ISBN:0-7695-1862-1.
[152] De Marneffe, M. C., & Manning, C. D. (2008). Stanford typed dependencies
manual (pp. 338–345). Technical Report, Stanford University.
[153] Zeng, J., Lin, X. V., Xiong, C., Socher, R., Lyu, M. R., King, I., & Hoi, S. C.
H. (2020). Photon: A robust cross-domain text-to-SQL system. In Proceedings
of the 58th Annual Meeting of the Association for Computational Linguistics:
System Demonstrations (pp. 204–214). ACL.
[154] Poole, D. L., & Mackworth, A. K. (2010). Artificial intelligence: Foundations of
computational agents. Cambridge University Press.
[155] Warren, D. H., Pereira, L. M., & Pereira, F. (1977). Prolog – the language and its
implementation compared with Lisp. ACM SIGPLAN Notices, 12(8), 109–115.
[156] Wang, P., Shi, T., & Reddy, C. K. (2019). A translate-edit model for natural lan-
guage question to SQL query generation on multi-relational healthcare Data.
arXiv preprint arXiv:1908.01839.
[157] Yao, K., & Zweig, G. (2015). Sequence-to-sequence neural net models for graph-
eme-to-phoneme conversion. arXiv preprint arXiv:1506.00196.
[158] Zhang, Z., & Sabuncu, M. (2018). Generalized cross-entropy loss for training
deep neural networks with noisy labels. In Advances in Neural Information
Processing Systems (pp. 8778–8788).
References 159
[159] Seo, M., Kembhavi, A., Farhadi, A., & Hajishirzi, H. (2016). Bidirectional atten-
tion flow for machine comprehension. arXiv preprint arXiv:1611.01603.
[160] Ke, N. R., Goyal, A. G. A. P., Bilaniuk, O., Binas, J., Mozer, M. C., Pal, C., &
Bengio, Y. (2018). Sparse attentive backtracking: Temporal credit assignment
through reminding. In Advances in neural information processing systems
(pp. 7640–7651).
[161] Michaelsen, S. M., Dannenbaum, R., & Levin, M. F. (2006). Task-specific train-
ing with trunk restraint on arm recovery in stroke: Randomized control trial.
Stroke, 37(1), 186–192.
[162] Löb, M. H. (1976). Embedding first order predicate logic in fragments of intu-
itionistic logic. The Journal of Symbolic Logic, 41(4), 705–718.
[163] Yu, B., Lin, X., & Wu, Y. (1991). The tree representation of the graph used in
binary image processing. Information Processing Letters, 37(1), 55–59.
[164] Python. (2018, June). Python 3.7.0 Home Page. Last accessed December 23,
2019. Retrieved from: https://www.python.org/downloads/release/python-370/
[165] MySQL Community Downloads. (2019). MySQL Community Server 8.0.18
Home Page. Last accessed December 23, 2019. Retrieved from: https://dev.
mysql.com/downloads/mysql/
[166] MySQLdb. (2012). Welcome to MySQLdb’s documentation! Last accessed
December 23, 2019. Retrieved from: https://mysqldb.readthedocs.io/en/latest/
[167] TextBlob. (2015). TextBlob: Simplified Text Processing. Last accessed December
23, 2019. Retrieved from: https://textblob.readthedocs.io/en/dev/
[168] JetBrains. (2019). IDE PyCharm C Home Page. Last accessed December 23,
2019. Retrieved from: https://www.jetbrains.com/pycharm/
[169] XQuartz. (2016, October). XQuartz 2.7.11 Home Page. Last accessed December
23, 2019. Retrieved from: https://www.xquartz.org/index.html
[170] Apple Developer. (2019). Xcode 11 Home Page. Last accessed December 23,
2019. Retrieved from: https://developer.apple.com/xcode/
[171] MySQL. (2019). MySQL Workbench Home Page. Last accessed December 23,
2019. Retrieved from: https://www.mysql.com/products/workbench/
[172] Kaggle. (2017). Zomato Restaurants Data. Last accessed December 23,
2019. Retrieved from: https://www.kaggle.com/shrutimehta/zomato-restaurants-
data
[173] GitHub. (2017). WikiSQL RDB – A large annotated semantic parsing corpus
for developing natural language interfaces. Last accessed December 23, 2019.
Retrieved from: https://github.com/salesforce/WikiSQL
[174] Zhong, V., Xiong, C., & Socher, R. (2017). Seq2SQL: Generating structured
queries from natural language using reinforcement learning. arXiv preprint
arXiv:1709.00103.
[175] Zhong, V., Xiong, C., & Socher, R. (2017). Seq2SQL: Generating structured
queries from natural language using reinforcement learning. arXiv preprint
arXiv:1709.00103.
[176] Streiner, D. L., & Cairney, J. (2007). What’s under the ROC? An introduction to
receiver operating characteristics curves. The Canadian Journal of Psychiatry,
52(2), 121–128.
[177] Original implementation code extended from “Couderc, B., & Ferrero, J. (2015,
June). fr2SQL: Interrogation de bases de données en français”.
160 References
[178] Yu, T., Li, Z., Zhang, Z., Zhang, R., & Radev, D. (2018). TypeSQL:
Knowledge-based type-aware neural text-to-SQL generation. arXiv preprint
arXiv:1804.09769.
[179] Hwang, W., Yim, J., Park, S., & Seo, M. (2019). A comprehensive explora-
tion on wikisql with table-aware word contextualization. arXiv preprint arXiv:
1902.01069.
[180] He, P., Mao, Y., Chakrabarti, K., & Chen, W. (2019). X-SQL: Reinforce schema
representation with context. arXiv preprint arXiv:1908.08113.
[181] Yavuz, S., Gur, I., Su, Y., & Yan, X. (2018). What it takes to achieve 100%
condition accuracy on WikiSQL. In Proceedings of the 2018 Conference on
Empirical Methods in Natural Language Processing (pp. 1702–1711).
[182] Gur, I., Yavuz, S., Su, Y., & Yan, X. (2018, July). DialSQL: Dialogue based
structured query generation. In Proceedings of the 56th Annual Meeting of the
Association for Computational Linguistics (Vol. 1: Long Papers pp. 1339–1349).
[183] Zhekova, M., & Totkov, G. (2021). Question patterns for natural language trans-
lation in SQL queries. International Journal on Information Technologies &
Security, 13(2), 43–54.
[184] Brunner, U., & Stockinger, K. (2021, April). Valuenet: A natural language-to-sql
system that learns from database information. In 2021 IEEE 37th International
Conference on Data Engineering (ICDE) (pp. 2177–2182). IEEE.
[185] Xu, X., Liu, C., & Song, D. (2017). SQLnet: Generating structured queries
from natural language without reinforcement learning. arXiv preprint arXiv:
1711.04436.
[186] Talreja, R., & Whitt, W. (2008). Fluid models for overloaded multiclass many-
server queueing systems with first-come, first-served routing. Management
Science, 54(8), 1513–1527.
[187] Tosirisuk, P., & Chandra, J. (1990). Multiple finite source queueing model with
dynamic priority scheduling. Naval Research Logistics (NRL), 37(3), 365–381.
[188] Carbonell, J. R., Ward, J. L., & Senders, J. W. (1968). A queueing model of
visual sampling experimental validation. IEEE Transactions on Man-Machine
Systems, 9(3), 82–87.
[189] Hoi, S. Y., Ismail, N., Ong, L. C., & Kang, J. (2010). Determining nurse staff-
ing needs: The workload intensity measurement system. Journal of Nursing
Management, 18(1), 44–53.
[190] Robinson, W. N. (2003, September). Monitoring web service requirements. In
Proceedings 11th IEEE International Requirements Engineering Conference,
2003 (pp. 65–74). IEEE.
[191] Kim, C., & Kameda, H. (1990). Optimal static load balancing of multi-class
jobs in a distributed computer system. IEICE Transactions (1976–1990), 73(7),
1207–1214.
Index
Pages in italics refer to figures and pages in bold refer to tables.
A Entity-Attribute-Value (EAV), 13, 19, 146

Entity-Relational Diagram (ERD), 2, 63
Agrawal, A. J., 28, 34 Entity Relational (ER) Mapper, 13
ASK, 17 eXtensible Markup Language (XML), 13,
Augmented Transition Network (ATN), 17, 37 18, 101
B
Bali, R. S., 28 F
Booch’s OO Analysis and Design, 34
False Positive Ratio (FPR), 38, 84, 134
Boyan, J., 21
Foreign Keys (FKs), 31, 36
Freitag, D., 21
C
CHAT-80, 19
Chen, C. L., 25 G
Chen, H., 21
Chen, P. P. S., 2 Generative Pre-trained Transformer 3 (GPT-
Clinical Data Analytics Language (CliniDAL), 3), 40
18–20, 22–23, 45 generic interactive natural language interface
Clinical Information System (CIS), 20 to databases (GINLIDB), 19
Coltech-parser, 27, 36, 46, 125 Giordani, A., 23–24, 29, 31, 43
Conceptual Logical Form (CLF), 30, 45, 145 Graphical User Interface (GUI), 7
D
I
DB-Oriented Logical Form (DBLF), 29–30,
45, 145 IDE PyCharm C, 79, 81
DBXplorer, 18 Integrated Development Environment (IDE),
Dependency-Based Compositional Semantics 79
(DCS) Parser, 38 Intelligent Query Engine (IQE) system, 21
Deshpande, A. K., 28 Isoda, S., 26
disambiguation, 8, 10, 12–13, 31, 37, 43, 47, Iyer, A., 21
54–55, 89, 129, 142
J
E
Java Annotation Patterns Engine (JAPE), 27,
English Slot Grammar (ESG), 20 36, 46, 126
English Wizard, 19 Joachims, T., 21
161
162 Index
K Natural Language Interface for DataBase

(NLIDB), 2, 4, 7–8, 12–13, 15,
Karande, N. D., 29 17–20, 22–25, 29, 42, 47, 96, 100
Kaur, G., 28 Natural Language Query (NLQ)
Kaur, S., 28 -based interface, 8–9
Kirilov, A., 21 checking for expressions, 54–55
Kiryakov, A., 21 disambiguation, 54–55
form-based interface, 7
input interface
L
architecture pipeline, 48
Lambda Calculus, 30–31, 33, 38, 43, 45, 46, MetaTables, 47
123, 146 keyword-based query interface, 8
LIFER/LADDER, 17, 28–29, 38, 46, 132 mapping table design, 56
Liu, C., 44 mapping tokens into RDB elements, 56–59
LUNAR, 17 MetaTable, 14, 54, 84
NLP syntax and semantics, 38–39
rule-based assumptions, 58
M into SQL mapping
architectural layout framework, 28
MacBook Pro, 79 authoring interface-based systems, 33
Machine Learning Algorithms (MLA), 11, DB metadata, 23
21–23, 42 Dependency Syntactic Parsing, 31
Manov, D., 21 four pre-processing phase modules, 28
Mapping Table Module, 27–28 fuzzy set theory, 25
Matcher/Mapper module, 13–14, 14, 48, 55, Human–Computer Interaction, 27
83 hybrid approach, 26
Maximum Bipartite Matching Algorithm, 27, Kernel Functions, 24
45, 146 Lambda calculus, 33, 123
Maximum Flow Algorithm (MFA), 18 MLA algorithms, 33
Mayo Clinic information extraction system, MySQL framework in the back end, 23
18 NLIDB system, 24
Métais, E., 26 NLQ input, 33
MetaTables, 3, 11, 13, 14, 32, 35–39, 41, 43, NLQ tokens, 23
45, 47, 48, 50, 54–57, 59, 63, 66, research ideas, 46
84, 95 restricted NLQ input, 122
Michael, G., 29 SQL+, 29
Moreno, A. M., 26 statistical models, 26
Morphological Analysis Module, 28 string tree kernels, 31
Moschitti, A., 23–24, 31, 43 Support Vector Machines, 24
MySQL Community Server 8.0.18, 80 Syntactic Pairing, 29
MySQL RDB, 80 and translation, 100
MySQL Workbench, 81 tree kernels models, 33, 123
UML class diagram’s, 25
unified modeling language, 34, 124
N using MLA algorithms, 120
weighted links, 34–35, 124
Named Entity Recognition (NER), 9, 27, 36, into SQL transformation, 100
38, 46, 81, 95, 125–126 tokens extraction, 36
Natural Language Application Program tokens into RDB lexica mapping
Interface (LANLI), 19 dependency-based compositional
Natural Language Interface (NLI), 18–20, 25, semantics, 130
27, 30, 33–34, 48, 122 dependency parser, 131
Index 163
dependency syntactic parsing, 127, pseudocode

130 construct an SQL query from an NLQ
Java Annotation Patterns Engine, 126 input, 49–50
LIFER/LADDER method, 131 maximum length algorithm, 51
named entity tagger, 131 SQL keywords tagging with their
name entity recognizer, 125 synonyms, 53
pattern matching, 125 synonyms matching with RDB
Porter Aagorithm, 127 elements, 53
probabilistic context free grammar, Python NLP lightweight library, 51
133 RDB MetaTables, 48
spider system, 128 RDB relationships, 49
Stanford dependencies parser, 130 semantic analysis and syntactic
unification-based learning, 127 representation, 53
WordNet, 129 semantic analyzer, 50
two embedded mappers, 55 Post-Traumatic Stress Disorder (PTSD), 2, 2,
Natural Language Toolkit (NLTK), 39, 80, 63, 65, 66, 66, 70, 75, 77
127, 135 PRECISE, 18
NLP techniques, 100 Predicate Argument Structure (PAS), 20
NLQ, see Natural Language Query Primary Keys (PKs), 31, 36
NUITS system, 18 Probabilistic Context Free Grammar (PCFG),
28, 39, 133
pseudocode
O building SQL main clauses, 113
construct an SQL query from an NLQ
Object-Oriented (OO) modeling, 25 input, 49–50
Ognyanoff, D., 21 finding RDB lexicon parent attribute, table
Online Transactional Processing (OLTP), 1 and relationships, 57–57
Ott, N., 29 matching NLQ tokens to equivalent RDB
Owei, V., 26 elements, 111–112
maximum length algorithm, 51
P NLQ relationships definition, 56–57
NLQ spelling check function, 109
PANTO, 22 NLQ tokens labeling, 103
part of speech (POS) SQL generation and execution, 62
intermediate language representation SQL keywords tagging with their
layer, 11 synonyms, 53
annotator, 11–12 SQL template examples, 117–118
linguistic components layers, 9 synonyms matching with RDB elements,
lexicon, 11 53
semantic parser (rule-based), 10–11 synonyms tagging of SQL comparative
syntactic parser (rule-based), 9–10 operations keywords, 107–108
Patil, G. A., 29 tokens’ type definition, 105–106
Philips Question Answering Machine PyCharm C, 79
(Philiqa), 17 Python 3.7, 7, 80
Photon, 42 Python Software Foundation, 79
Popov, B., 21
Porter algorithm, 28, 36, 46, 127
POS recognition
Q
entity annotator, 50–51
NLQ string, 49 QUEST, 20
NLQ translation process, 50 Question Answering Systems (QAS), 1
organization pipeline, 52 Queueing Model, 90
164 Index
R Structured Query Language (SQL), 1

classes, 60, 64
Rao, G., 29 clauses mapping, 39–40
RDB, see Relational DataBase complexity vs performance, 40–42
Receiver Operating Characteristics (ROC) example 1
curves, 85 NLQ words breakdown process, 68
Reformat, M. Z., 26 template, 69
Relational DataBase (RDB), 2 tokens breakdown analysis, 69
comparison, 82 example 2
confusion matrix comparison, 94 NLQ words breakdown process, 71–72
elements, 99 SQL template, 72
F-measure, 93 tokens breakdown analysis, 70
lexica into SQL clauses, 99 example 3
lexica into SQL clauses mapping template, 73
bipartite tree-like graph-based tokens breakdown analysis, 73
processing model, 137 tokens breakdown process, 72–73
deep-learning-based model, 144 example 4
ellipsis method, 138 query, 74
extended UML class diagrams, 134 tokens breakdown analysis, 75
heuristic weighting scheme, 142 tokens breakdown process, 74
L2S system, 137 example 5
Lambda-calculus, 146 template, 76
pattern matching of SQL, 139 tokens breakdown analysis, 77
relationships and linguistic analysis, tokens breakdown process, 76
134 execution and result, 15, 62
semantic grammar analysis, 140 formation vs templates, 42–44
Top-K algorithm, 146 generation limitations, 95
weighted neural networks, 138 limitations, 97
lexica mapping, 36–37 MetaTables, 96
lexica relationships, 37–38 NLQ MetaTable, 66
mapping limitations, 94–95 PTSD RDB chart diagram, 66
MetaTables, 3 query, 67
NLQ into SQL translation, 93 query types, 115
query language, 3, 32 template generator, 14–15, 60–62
schema domain, 99 Subtree Pattern Matching Framework (SPMF),
Reports’ Retrieval Module, 28 20
Restricted Natural Language Query (RNLQs), 22 Support Vector Machines (SVM), 24
Rukshan, P., 21
T
S
TEAM, 17–18
Satav, A. G., 28 TextBlob, 51, 54, 80
Self Organizing Maps (SOM), 39, 134 To, N. D., 26
Semantic Analysis Module, 28 Top-k Algorithm, 18, 20, 34, 45, 46, 122,
Shankaranarayanan, G., 21 146
She, L., 21 True Negative Ratio (TNR), 38, 84
Song, D., 44 Tseng, F. S., 25
SPARK, 22
SQL, see Structured Query Language
U
Stanford CoreNLP 3.9.2, 80–81
Stanford Dependencies Collapsed (SDC), 23 Unified Modeling Language (UML), 25,
Structured Object Model (SOM), 25 34–35, 46, 124
Index 165
V X
Van De Riet, R. P., 26 Xcode, 11, 79, 81
XQuartz, 79, 81
Xu, X., 44
W
Y
weighted links, 34–35
Yager, R. R., 26
WikiSQL, 93
confusion matrix with, 86
NLQ to SQL translation, 87–88 Z
RDB, 81 Zomato RDB, 81, 85, 93
ROC curve for, 87 confusion matrix with, 85
to SQL Translation Work on WikiSQL mapping table, 96
RDB, 88 ROC curve, 86
WordNet, 81 schema, 82

Ftoon Kedwan - NLP Application - Natural Language Questions and SQL Using Computational Linguistics-CRC Press (2023)

Uploaded by

Copyright:

Available Formats

Ftoon Kedwan - NLP Application - Natural Language Questions and SQL Using Computational Linguistics-CRC Press (2023)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Ftoon Kedwan - NLP Application - Natural Language Questions and SQL Using Computational Linguistics-CRC Press (2023)

Uploaded by

Copyright:

Available Formats

NLP Application

Applying Natural Language Processing (NLP) concepts to help humans in

• Includes extensive and illustrative examples to simplify the dis-

Providing a unique approach to NLQ into SQL translation, as well as compris-

Ftoon Kedwan is a recognized expert and researcher in the field of Artificial

First edition published 2024

and by CRC Press

CRC Press is an imprint of Taylor & Francis Group, LLC

ISBN: 978-1-032-53835-8 (hbk)

Tree Kernels Models 33

5 Implementation User Case Scenario 63

6 Implementation Testing and Performance Measurements 79

7 Implementation Results Discussion 93

8 Conclusion and Future Work 99

This book applies Natural Language Processing (NLP) concepts to help

FIGURE 1 Research area breakdown.

• An entity sets represent a table, i.e., “Patient” table.

FIGURE 2 ERD for Post-Traumatic Stress Disorder (PTSD) RDB.

constructs intra-relationships were not studied in the previous works in terms

• Support for SQL’s syntactic constructs’ (keywords such as SELECT,

Machine-readable instructions are mandatory to access any type of DB.

BASIC RESEARCH FRAMEWORK

FIGURE 3 Basic research organization diagram.

In the following sections, Chapter 2 is a background study on the research

NLQ INPUT PROCESSING

Interactive Form-Based Interface [16]

FIGURE 4 NLQ words, tokens and lexica terminology clarification.

Keyword-Based Query Interface [17]

NLQ-Based Interface [18]

1) Simple straightforward questions with simple linguistic structures

Furthermore, NLQ questions can be further categorized as questions asking for

simple NLQ problems. However, addressing the complicated general questions

PART OF SPEECH (POS) RECOGNITION

Linguistic Components Layers

Syntactic parser (rule-based)

sentences that do not conform to a proper language grammar or rules of how

Semantic parser (rule-based)

iii. Phonology [31]: is studying the sound patterns of a certain language

Intermediate Language Representation Layer

TABLE 1 NLQ MetaTable

TABLE 2 RDB elements MetaTable

SQL TEMPLATE GENERATOR

systems support different SQL templates. Hence, accepted and recognized

• Data Query Language (DQL): SELECT

SQL EXECUTION AND RESULT

status, relatedness to patients, etc.). The system works by implementing simple

Similar to CliniDAL, TimeText [75] is a temporal system architecture for

NLQ to SQL Mapping

A similar NLIDB system was designed by [77] with comparable per-

Isoda [87] discovered that OO-based analysis applications enable intuitive

and Prolog-like representation [96], in addition to complex NLQ to SQL trans-

In [99], a simple architectural layout framework is proposed. It starts

CURRENT RESEARCH WORK

Authoring Interface-Based Systems [131]

Enriching the NLQ/SQL Pair

Using MLA Algorithms

Restricted NLQ Input [35]

Lambda Calculus [7, 112, 114]

Tree Kernels Models [29, 47, 78, 100, 126]

Part of Speech Recognion

PSEUDOCODE 1 ALGORITHM TO CONSTRUCT AN SQL

PSEUDOCODE 4 SYNONYMS MATCHING WITH RDB

PSEUDOCODE 6 SQL KEYWORDS TAGGING WITH THEIR

PSEUDOCODE 10 FINDING RDB LEXICON PARENT