See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/313851778

Resume Parser with Natural Language Processing

Thesis · March 2017

DOI: 10.13140/RG.2.2.11709.05607

CITATIONS READS

18 34,625

4 authors:

Satyaki Sanyal Souvik Hazra

KIIT University Infineon Technologies
4 PUBLICATIONS 22 CITATIONS 27 PUBLICATIONS 378 CITATIONS

SEE PROFILE SEE PROFILE

Neelanjan Ghosh Soumyashree Adhikary

1 PUBLICATION 18 CITATIONS
KIIT University
1 PUBLICATION 18 CITATIONS
SEE PROFILE
SEE PROFILE

All content following this page was uploaded by Satyaki Sanyal on 20 February 2017.

The user has requested enhancement of the downloaded file.

ISSN 2321 3361 © 2017 IJESC

Research Article Volume 7 Issue No. 2

Resume Parser with Natural Language Processing

Satyaki Sanyal1 , Souvik Hazra 2 , Sou myashree Adhikary 3 , Neelanjan Ghosh 4
School of Electronics Engineering 1, 3
School of Electrical Engineering 2, 4
KIIT University, India

Abstract:
Parse informat ion fro m a resume using natural language processing, find the keywords, cluster them onto sectors based on their
keywords and lastly show the most relevant resume to the employer based on keyword matching. First, the user uploads a resume
to the web platform. The parser parses all the necessary informat ion fro m the resume and auto fills a form for the user to
proofread. Once the user confirms, the resume is saved into our NoSQL database ready to show itself to the employers. Also, t he
user gets their resume in both JSON fo rmat and pdf.

Keywords: Resu me parser, resume analy zer, text min ing, natural language processing, resume JSON, semantic analysis

I. PROB LEM S TATEMENT came into existence. These agencies required the applicants to
upload their resumes on their websites in particular formats. The
To design a model this can parse informat ion fro m unstructured agencies would then go through the structured data and shortlist
resumes and transform it to a structured JSON format. Also, candidates for the company. This process had a major
present the extracted resumes to the employer based on the job drawback. There were numerous agencies and each had their
description. own unique format. To overcome all the above problems an
intelligent algorithm was required which could parse
II. INTRODUCTION information fro m any unstructured resumes, sort it based on the
clusters and rank it finally. The model uses natural language
Corporate companies and recruitment agencies process processing to understand the resume and then parse the
numerous resumes daily. This is no task for humans. An information fro m it. Once information is parsed it is stored in
automated intelligent system is required which can take out all the database. When the employer posts a job opening, the
the vital informat ion from the unstructured resumes and system ranks the resumes based on keyword matching and
transform all of them to a common structured format which can shows the most relevant ones to the employer.
then be ranked for a specific job position. Parsed information
include name, email address, social profiles, personal websites, IV. PREPROCESSING
years of work experience, work experiences, years of
education, education experiences, publications, certificat ions, Data preprocessing is the first and foremost step of natural
volunteer experiences, keywords and finally the cluster of the language processing. Data preprocessing is a technique of data
resume (ex: co mputer science, human resource, etc.). The mining which transforms raw data into a comprehensible
parsed informat ion is then stored in a database (NoSQL in this format. Data fro m the real wo rld is mostly inadequate,
case) for later use. Unlike other unstructured data (ex: email conflicting and contains innumerable errors. The method of
body, web page contents, etc.), resumes are a bit structured. Data preprocessing has proven to resolve such issues. Data
Information is stored in discrete sets. Each set contains data preprocessing thus further processes the raw data. Data is made
about the person's contact, work experience or education to pass through a series of steps in the time of preprocessing:
details. In spite of this resumes are d ifficu lt to parse. This is
because they vary in types of information, their order, writing Data Cleaning: Processes, like filling in missing values,
style, etc. Moreover, they can be written in various formats. smoothing noisy data or resolving inconsistencies, cleanses the
Some of the common ones include '.txt', '.pdf', '.doc', '.docx', data.
'.odt', '.rtf' etc. To parse the data from different kinds of
resumes effectively and efficiently, the model must not rely on Data Integration: Data consisting of various representations
the order or type of data. are clustered together and the clashes between the data are
taken care of.
III. HISTORY OF HIRING
Data Transformation: Data is distributed, assembled and
The process of hiring has evolved over the period of time. In th e theorized.
first generation hiring model, the companies would advertise
their vacancies on newspapers and television. The applicants Data Reduction: The objective of this step is to present a
would send in their resumes via post and their resumes would be contracted model in a data warehouse.
sorted manually. Once shortlisted, the hiring team would call the
applicants for further rounds of interview. Needless to say, this Data Discretization: In this step, the number of values of an
was a time-consuming procedure. But the industries started uninterrupted characteristic is reduced by division of the range
growing and so did the hiring needs. Hence the companies of intervals of characteristics.
started outsourcing their hiring process. Hiring consultancies

International Journal of Engineering Science and Computing, February 2017 4484 http://ijesc.org/
Tokenization: To kenization is the task of chopping off a taught to identify whether the word is a noun, verb, adjective,
provided character sequence and a detailed document unit. It adverb, etc.
does away with certain characters like punctuation and the
chopped units are further called tokens. It can be illustrated as
follows:

FIGUR E.1. TOKENIZATION

Tokens are usually referred to as terms or words, but

sometimes fabricating a type/token distinction is essential. A
specimen of an array of characters in a document that is
assembled as a helpfu l acceptable unit for processing is called a FIGUR E.3. POS TAGGING
token. Whereas, the group of tokens which consists of same
character sequence is called type. And a type that is added to Chunking: Also known as shallow parsing, chunking is
the dictionary of IR system is called term. We can co mp letely actually the recognition of parts of speech and short phrases.
differentiate a set of index terms fro m tokens. As an examp le We can determine if the words are nouns, verbs, adjectives, etc
we can say, they can be acceptable identifiers in taxono my, but by Parts of Speech tagging, but from this, we cannot get any
mostly in modern IR systems, they have a strong relatio n with clue about the sentence of phrase structure in the sentence. At
tokens in the document. Nevertheless, as a substitute for being times, some mo re informat ion than parts of speech of words are
totally the tokens appearing in the document, they are mostly useful, but the full parse tree that we would get from parsing is
derived fro m them by various processes of normalizat ion. not needed. Named-entity recognition is a citation when
chunking might be preferable. In NER, the goal is finding
Stemming: According to linguistic morphology and retrieval named entities, wh ich mostly has a tendency to be noun
of informat ion, stemming is the mechanism of lowering altered phrases, so in the following sentence we would want to know if
or derived words to their word stem, root or base. The stem 'The angry bear' is there or not: "The angry bear chased the
shouldn't always match the morphological root of the words. It frightened little squirrel" But one wouldn't necessarily care if
is satisfactory that relevant words map to the same stem even the angry bear is the subject of the sentence or not. Chunking is
when the stem is not even a valid root. Algorithms for this also commonly used in tasks like example -based machine
process are being studied in co mputer science fro m the 1960s. natural language understanding, speech generation, and others
Conflat ion is when a number of search engines treat word with as a pre-processing step.
the same stem as synonyms as an example of query expansion.

FIGUR E.2. STEMMING

Lemmatization: In linguistics, lemmatization is a procedure of

organising the altered form of a word, such that their analysis FIGUR E.4. CHUNKING
can take place as a single term, identified by the word's
dictionary form (lemma). In co mputational linguistics, the V. THE MODEL
procedure of concluding the lemma of a word depending upon
its predetermined meaning. It depends on rightly identifying the We can classify the model into three major parts which we will
predetermined part of speech and what a word in a sentence be discussing in depth.
means, as well as in a bigger situation surrounding the A. The uploader
sentences, which can include neighboring sentences, and even B. The parser
an entire document, which contradicts stemming. So, an
algorith m of lemmatizat ion is an open platform fo r research. A. THE UPLOADER

Parts of speech tagging: In corpus linguistics, the procedure The client may be a giant corporate firm who wants to parse
of marking up a text (corpus) which is analogous to a particular and rank their tens and thousands of unstructured resumes, or a
part of speech depending on both its definition and context student trying to beautify his/her unstructured text resume and
(like how it is related to the adjoining wo rds in a paragraph, convert into a beautiful pdf format. In either case, the algorithm
sentence or phrase) is called part-of-speech tagging (POS remains same. First, the client uploads a file. The algorithm
tagging or POST). It is also known as grammat ical tagging or blocks any file having an extension other than '.pdf', '.doc',
word-category disambiguation. School-age children are '.docx', '.odt', ‘.ods’, '.t xt'. After the client has successfully
commonly taught a simp lified version of this when they are uploaded the file, the algorithm takes the file, reads the

International Journal of Engineering Science and Computing, February 2017 4485 http://ijesc.org/
contents and writes the content into a text file before passing on data between the starting and the ending of them, which we call
the data to the parser. as segments. Out of the many exceptions which might occur,
one which is common is that the first segment generally
contains the name as well as the contact information of the
person. Now we program chunkers or Named Entity
Recognizer to extract data from each segment specifically. This
method makes the system efficient and reduces its complexity.
Now if due to some reason the recognizer runs on a wrong
piece of data, the system will produce unexpected results.

II. S YNTACTIC ANALYSIS

FIGUR E.5. THE UPLOADER The syntactic analysis determines the structure of data. The
architecture comprises of a hierarchy of expressions, the
B. THE PARS ER smallest being a basic symbol and the largest being sentences.
We can visualise the architecture as a tree whose nodes
Once uploaded as a text file by the up loader, the parser comes represents the expressions. Values stored in the nodes represent
in play. It parses all the relevant data from the uploaded resume the basic symbols. The root represents the sentence.
including name, emails , contact numbers, social profile links,
personal websites, years of work experience, work experiences, Parse Tree:
years of education, degrees, volunteer experiences,
publications, skills, cluster(s) and languages through natural The parse tree is generated by the parser with syntactic
language processing and without any human interaction. Now, analysis. A parse tree or a parsing tree is an organised,
what exactly is natural language processing? entrenched structure which we use to represent the syntactic
analysis of a string. They categorically reflect the syntax of the
Natural Language Processing: input data, making them noticeable fro m the abstract syntax
trees used in programming.
Natural language processing is a branch of artificial
intelligence and computational linguistics. It can be defined as
the process which is involved in the interaction between a
computer and natural language i.e the language, spoken by
humans. It is directly related to the field of hu man-co mputer
interaction. Now that natural language processing is properly
defined, we will be using the following constraints of NLP to
parse the informat ion fro m the resumes:

I. Lexical Analysis

II. Syntactic Analysis

III. Semantic Analysis

I. LEXICAL ANALYS IS

The pilot stage of the compiler is lexical analysis. The altered

source code is taken from the language preprocessor which FIGUR E.7. PARS E TREE REPRES ENTING S YNTACTIC
writes in the form of sentences. The analyzer removes any ANALYS IS
comments or whitespace from the source code, breaking these
syntaxes into a chain of tokens. II. S EMANTIC ANALYS IS

Semantic analysis can be defined as the study of semantics i.e

the structure and meaning of speech. This process relates
syntactic structure to the level of the writing as a whole from
the levels of clauses, phrases, paragraph and sentences. It
relates to their language-independent meanings. Let's take an
example. Person A has a resume which states he has graduated
fro m the "University of Calcutta" and person B has a resume
which says he has graduated from "Calcutta University".
FIGURE.6. LEXICAL ANALYSIS Essentially they both graduated from the same place. So what
semantic analy zer does is convert "University of Calcutta" to
Considering our case, the resume is discriminated onto various "Calcutta University". In Information Retrieval research, text
segments including contact informat ion, educational classification system is given the utmost focus which bounds
experiences, work experiences and more. We use a database or the decisions to either relevant or non-relevant depending upon
a data dictionary to hold the keywords or headings we find the information need of the user. It is not a hard task to get the
common in most of the resumes. Now when a new resume is user information need.
taken, the parser searches for the keywords and extracts all the

International Journal of Engineering Science and Computing, February 2017 4486 http://ijesc.org/
 THE MODEL

FIGUR E.9.2. WEBS ITE AND EMAIL

FIGUR E.9.3. ADDRESS

FIGUR E.9.4 SOCIAL LINKS

FIGURE.8. THE MO DEL

Now that we understand what Lexical, Syntactic and Semantic

analysis does, let us see how the system works. The data is
given to the system as a raw string. The Lexical analyzer pre-
processes the data and tokenizes them. The Syntactic analyzer FIGUR E.9.5. WORK EXPERIENCE
takes the tokens and finds the structure in it. The parse tree
diagrammat ically represents the syntactic structure in the form
of a tree. The Semantic analyzer studies the structure of the data
to find their language-independent meaning.

VI. EXPECTED OUTCOME

Both employers and candidates will be appeased by our system.

Lots of strain on the head of candidate or employee in Online
Recruit ment System will have been reduced by this online tool. FIGUR E.9.6. EDUCATIONAL EXPERIENCE
The system will parse all the resumes and store them in the
database. Then it will rank them using Artificial Intelligence or
AI and predict which candidate is best suited for the job, thus
making the hiring system authentic. So me screen-shots of the
result of our resume parser are portrayed below. Once the user
confirms the result of our parser the system generates a JSON
resume and stores it in the NoSQL database. FIGUR E.9.7. VOLUNTEER EXPERIENC E

FIGUR E.9.1. BAS IC INFORMATION

FIGUR E.9.8. S KILLS

International Journal of Engineering Science and Computing, February 2017 4487 http://ijesc.org/
VII. JSON OUTPUT "title": "First prize in robotics",
"date": "15-08-2016",
{ "awarder": " VIT Vellore",
"basics": { }],
"name": "Satyaki Sanyal", "publications": [{
"label": "Programmer", "name": "recru it ment predictions with id3 decision tree",
"picture": "", "publisher": "international journal of advanced engineering
"email": "sanyal.satyaki09@g mail.co m", and research development",
"phone": "(+91) 9178449492", "release Date": "22-10-2016",
"website": "http://www.satyakisanyal.com", "website": "http://www.ijaerd.co m",
"summary": "A summary of John Doe...", "summary": "we have tried to solve the problem
"location": { of recru it ment with cognitive
"address": "Acharya Prafulla Chandra Road", computing. we have used decision trees to predict the
"postal Code": "700020", candidates best suited for the job and we have used random
"city": "Kolkata", forest for better pedictions .."
"country": "India", }],
}, "skills": [{
"profiles": [{ "HTM L",
"Twitter": "http://www.twitter.co m/Satsan95", "CSS",
"Github": "www.github.com/Satyaki0924", "Javascript",
"LinkedIn": "http://linkedin.co m/in/satyaki- “Python”,
sanyal-708424b7" “Machine Learn ing”,
}] “Deep Learning”
}, }],
"work": [{ "languages": [{
"company": " Venturesity", "name": " English",
"position": "Intern", "level": "Expert"
"start Date": "01-11-2016", },
"end Date": "01-02-2017", {
"summary": "My job at venturesity was to work on "name": "Hindi",
natural language processing and make online parsing "level": "Expert"
systems", }],
}, "interests": [{
{ “Swimming”,
"company": " Geo metric Ltd.", “Reading books”,
"position": "Intern", “Watching movies”,
"start Date": "01-05-2016", “Coding”
"end Date": "01-07-2016", ]
"summary": "My job at Geo met ric was to work }]
on image recognition and simu late "references": [{
self driv ing with reinforcement learn ing", "name": "John Doe",
}, "reference": "Satyaki was an asset to our
], company.."
"volunteer": [{ }]
"organization": "IBM", }
"position": "Mentor",
"start Date": "05-10-2016", VIII. CONCLUS ION AND FUTUR E WORK
"end Date": "05-10-2016",
"summary": "I was a data analytics mentor for We successfully converted different formats of resumes to text
IBM", and parse relevant information fro m there. We also were able to
}], scrape keywords from different social networking sites
"education": [{ including Stack Overflow, LinkedIn, etc and find the similarity
"institution": "KIIT University", between them with which we could determine the genre of the
"area": "Electronics and Electrical", resume (e.g: Co mputer science, Management, Sales, human
"study Type": "Btech", resource, etc). Future work includes ranking the resume and
"start Date": "2014", analysing information about the candidate from social
"end Date": "2018", networking sites like Face book and Twitter so that we can
}, decide more accurately and authentically whether or not to offer
{ the candidate, a job.
"institution": "Gundecha Education Academy",
"area": "Science", IX. ACKNOWLEDGEMENT
"study Type": "Indian School Cert ificate",
"start Date": "2012", We are grateful to all anonymous reviewers for their valuable
"end Date": "2014", feedback.
}],
"awards": [{

International Journal of Engineering Science and Computing, February 2017 4488 http://ijesc.org/
X. REFER ENCE

[1]. F. Ciravegna, “Adaptive information ext raction fro m text by

rule induction and generalisation,” in Proceedings of the 17th
International Joint Conference on Artificial Intelligence
(IJCAI2001), 2001.

[2]. A. Chandel, P. Nagesh, and S. Sarawagi, “Efficient batch

top-k search for dictionary-based entity recognition,” in
Proceedings of the 22nd IEEE International Conference on Data
Engineering (ICDE), 2006.

[3]. S. Chakrabarti, Min ing the Web: Discovering Knowledge

fro m Hypertext Data. Morgan-Kauffman, 2002

[4]. M. J. Cafarella, D. Downey, S. Soderland, and O. Etzioni,

“KnowItNow: Fast, scalable information extraction from the
web,” in Conference on Human Language Technologies
(HLT/ EMNLP), 2005.

[5]. M. J. Cafarella and O. Et zioni, “A search engine for natural

language applications,” in WWW, pp. 442–452, 2005.

[6]. https://www.ijircce.co m/upload/2016 /april/ 218_ Intellig

ent.pdf

[7]. https://www.tutorialspoint.com/compiler_design/images
/token_passing.jpg

[8]. http://www.n ltk.org/book/tree_images/ch08-tree-6.png

International Journal of Engineering Science and Computing, February 2017 4489 http://ijesc.org/
View publication stats