Natural Language Processing

CS 1462

Vector Space Models, Feature

Extraction , Information Retrieval, and
Text Categorization
Computer Science
3 rd semester - 1444
Dr. Fahman Saeed
[email protected]

1
Some slides borrows from Carl Sable
 If you are following along in the book…

 Part of this topic, on vector space models, is related to

sections 6.3 through 6.5 of the current draft of the textbook
 Information retrieval is mentioned through various
chapters, and is discussed in Section 23.1 of the current
draft (the rest of the chapter is on question answering)
 The naïve Bayes approach to text categorization is the
subject of Chapter 4 of the textbook
 Information retrieval and text categorization (especially the
latter) are highly related to the work we maybe do in course
project.

2
 Information Retrieval

 Information retrieval (IR) is the task of returning, or retrieving,

relevant documents in response to a particular natural language query
 A document in the context of IR can generally pertain to all manner of
media, including text, photographs, audio, and video
 In this course, we are only concerned with the retrieval of text documents
(or at least, documents that include text)
 Depending on the application, these text documents might be web pages,
news articles, technical papers, encyclopedia entries, or smaller units such
as paragraphs or sentences
 IR has been an extremely important application of NLP, some colleges offer
entire courses on IR
 A collection refers to a set of documents being used to satisfy user requests
 A term refers to a lexical item (typically a word or token) that occurs in the
collection
 A query represents the user's information need expressed as a set of terms

3
 Conventional IR

 Many conventional IR systems assume that the

meaning of a document resides solely in the set of
terms (unigrams) that it contains
 Because such systems ignore syntactic (and arguably,
semantic) information, these are often said to use
bag-of-words approaches
 We will see that conventional text categorization
methods also use bag-of-words approaches
 Some modern NLP methods involving word
embeddings also make a bag-of-words assumption
(if the ordering of embeddings is not considered)
4
 Vector Space Models

 In the vector space model of IR, documents and queries are

represented as vectors of features
 Each feature corresponds to one term (typically a word) that occurs in
the collection
 The number of dimensions of each vector is equal to the size of the
vocabulary
 The value of each feature is called the term weight
 The term weight is often a function of the term's frequency, possibly
combined with other factors; we'll discuss this more soon
 Generally, we can represent a document, dj, as a vector, as follows: dj
= (w1,j , w2,j , w3,j , … , wN,j )

5
 Vector Space Models

 Above, N, the number of dimensions of the vector, is equal to the

number of distinct terms that occur in the collection, and wi,j is the
weight that term i is assigned in document j
 N can often be in the tens or hundreds of thousands
 Other text sequences, including queries, can be represented as vectors
in the same manner
 However, most documents and most queries will not contain most of
the possible terms that can occur (and hence, most of the feature
weights will be zero); i.e., the vectors are sparse
 In practice, representations are used which do not require us to store
all the zeros; we'll also discuss this more soon

6
 Word Embedding

 It is a matrix of words with particular values for each

word to identify its meaning and its distance from
the other words.
 If we had five different words it is:
) ‫ كتاب‬, ‫ كلب‬, ‫ تفاحة‬, ‫ رجل‬,‫( الصبر‬
 To establish mathematical relations between them,
we can ask and answer questions for each term, such
as:

7
Basic idea of Embedding

8
 Basic idea of Embedding

 These statistics are used to detect the approximate

meaning of the word in circulation, compare words,
and determine as an example: how close )‫ (تفاح‬is to
)‫ (برتقال‬and how far each is from )‫(الصبر‬.
 To create an embedding word, the model must be
trained to reliably estimate each word's meaning
using millions of words.

9
 Text Vectors

 In this step, we calculate the required words based

on another number of words using embedding word
values for each word. For instance: How can we find
Russia's capital from the US capital?

10
 Text Vectors

By research, Moscow is the closest city to this value, and this applies to
several other countries.

11
 Text Vectors

 X = King – Man + Woman

12
 Measuring Similarity Between Vectors

 The cosine similarity metric is often used in IR to measure the similarity

between two vectors
 Since we don't know the angle between words, how is similarity cosine
calculated?
 Do not forget that if we deal with two words that are far from each other, the
angle will be 90, and the cos will be 0, indicating that they are different
 if we deal with two words that are identical, the angle will be 0, and the cos will
be 1, indicating that they are similar, and any number between them indicates
their similarity.

13
Measuring Similarity Between Vectors

14
 Text Feature Extraction

 Bag-of-words(BoW)
 TF-IDF or ( Term Frequency(TF) — Inverse Dense
Frequency(IDF) )
 One Hot Encoding

15
 Bag-of-words(BoW)

 Bag-of-words(BoW) is a statistical language model

used to analyze text and documents based on word
count.
 The model does not account for word order within a
document.
 BoW can be implemented as a Python dictionary with each
key set to a word and each value set to the number of times
that word appears in a text.

16
 Bag-of-words(BoW) Example

 It was the best of times,

 it was the worst of times,
 it was the age of wisdom,
 it was the age of foolishness
 it was the epoch of belief,
 it was the epoch of incredulity,
 it was the season of Light,
 it was the season of Darkness,
 it was the spring of hope,
 it was the winter of despair

17
 Bag-of-words(BoW) Example

 First, all words are identified, with repetitions removed:

 “it” “was” “the” “best” “of” “times” “worst” “age” “wisdom”
“foolishness”
 To determine each sentence's word presence, we develop a
matrix:

 Each row vector can be used as feature for Ngram or SA

 Its drawbacks are its slowness and lack of focus on word
meanings. 18
 TF-IDF or ( Term Frequency(TF) — Inverse Dense
Frequency(IDF) )

 TF-IDF is a technique which is used to find meaning of sentences

consisting of words and cancels out the incapabilities of Bag of Words
technique which is good for text classification or for helping a machine
read words in numbers.
 TF-IDF gives us a way to associate each word in a document
with a number that represents how relevant each word is in
that document. Such numbers can be then used as features of
machine learning models
 The TF-IDF score of a term in a document measures its importance in
the document, which can be used for various applications, such as
information retrieval, text classification, and clustering.

19
 TF-IDF or ( Term Frequency(TF) — Inverse Dense
Frequency(IDF) )

• We utilize a logarithmic to minimize the value of the greatest

numbers and reduce growth. If the term is repeated twenty times in
a file, it is better, but not twenty times better.
• N, the total number of files, and second, df, the number of files that
mention the desired word.
• If the word is rarely used and appears in a few documents, then the
log of a large number (for example, a million) and has a large value
will be 6 (more important)
• If it appears in all files, it will be logarithm 1, i.e. 0. (less important)
20
 TF-IDF - Example

4 files

TF

21
 TF-IDF - Example

IDF

22
23
 Arabic language encoding

 For Arabic language, there are several text encoding

standards that can be used, such as Unicode, which is the
most commonly used encoding method.
 Unicode works with Arabic script and supports a range of
characters including diacritics and ligatures.
 Another popular encoding standard is the Arabic script
encoding (ASE), which supports a wide range of characters
and ligatures.
 These encoding standards play a crucial role in Arabic NLP
tasks such as text classification, sentiment analysis, and
language modelling.

24
 Arabic language encoding

 an example of encoding the word "‫" أجهزة‬using Arabic

Script Encoding (ASE).
 In ASE, Arabic letters are encoded using a combination
of decimal and hexadecimal values. Here are the decimal
and hexadecimal values of each Arabic letter in the word
":"‫أجهزة‬
 ‫ أ‬: - Decimal 1570, Hexadecimal 0643
 ‫ج‬- Decimal 1580, Hexadecimal 064C-
 ‫ ه‬- Decimal 1607, Hexadecimal 0647
 ‫ ز‬Decimal 1592, Hexadecimal 0632
 ‫ ة‬Decimal 1577, Hexadecimal 0629

25
 Arabic language encoding

 To encode the word in ASE, we convert each decimal

value to a two-digit hexadecimal value and
concatenate them together.
 So, the encoded form of " "‫ أجهزة‬in ASE would be:
"0643064C064706320629".
 To decode the encoded string back to the original
text, we simply split the string into pairs of two digits
and convert them back to decimal values, which then
correspond to the respective Arabic characters.

26
 Other Issues

 As with other NLP applications, we need to decide whether an IR system should apply
stemming (common), apply lemmatization (not common), convert all letters to the same
case, etc.
 We need to be consistent with the text normalization techniques applied to the collection
 Many conventional IR systems also use stop lists, containing stop words, which are just
lists of very common words to exclude from the computation
 This doesn't change results much, since these words tend to have very low IDF values; it
could make the system more efficient, but makes it more difficult to search for phrases
 Another technique used to by some conventional IR systems is to add synonyms of query
terms to the queries to help locate documents that are relevant but don't use overlapping
terms
 The vector space model is not the only way to implement a conventional IR system;
another approach (more common for earlier IR systems) was to use a Bayesian model
 We will consider a Bayesian approach for text categorization, in addition to an approach
using a vector space model, and also a k-nearest neighbors (KNN) approach
 We will later discuss how to evaluate text categorization systems, which uses some of the
same metrics (but text categorization doesn't share all the issues that make evaluation of
IR more difficult)
27
 Ranking Web Pages

 I add: Another issue not discussed in the textbook is ranking algorithms

that are specific to web search (e.g., how Google ranks its results)
 Such ranking could just be based on similarity metrics (e.g., the cosine
metric to compare TF*IDF vectors)
 Historically, early web retrieval search engines did this, but they could
easily be fooled
 Some websites, for example, put invisible, popular search terms in their
pages to fool search engines
 Popular web search engines today treat the web as a directed graph;
intuitively, good pages will tend to be more popular and have more
incoming directed links pointing to them
 The Hyperlinked-Induced Topic Search (HITS) algorithm involves
authorities (pages that have a lot of incoming links from other good pages)
and hubs (pages that point to a lot of authorities)
 Given a query, the algorithm first determines an initial set of web pages
that are relevant to the query, and then uses an iterative approach to
determine authorities and hubs

28
 Text Categorization

 Text categorization (TC) is the automatic labeling of documents based on text

contained in, or associated with, the documents into one or more predefined
categories, a.k.a. classes
 Some sources use the term text classification to refer to this task; however, that
term is also commonly used in a more general sense to include TC, IR, and
sometimes clustering
 Some TC tasks assume that the categories are independent, i.e., binary or Boolean
 Each document can then be assigned to zero, one, or multiple categories; most of
the conventional NLP literature on TC deals with such categories
 Other TC tasks assume mutually exclusive and exhaustive categories; each
document is then assigned to exactly one category (the best fit)
 If a task involves exactly two mutually exclusive and exhaustive categories (e.g.,
Indoor vs. Outdoor), you can alternatively view the task as involving one binary
category
 One special case of TC involves categories that form a hierarchy or taxonomy; a
conventional example was the Yahoo! directory of web pages (which eventually
became defunct)
30
 Applications of Text Categorization

 Some (of the many) applications of text categorization include:

 Classification of e-mail as spam or not spam (a.k.a. ham), i.e., spam filtering
 Classification of news into topical sections (e.g., Google News, Columbia
Newsblaster)
 Websites as pornography or not pornography
 Reviews as positive or negative (a type of sentiment classification, which can also
apply to tweets or news reports, and may involve more fine-grained categories)
 Automatic grading of essays (actually used by ETS)

 One interesting early application of text categorization was

authorship attribution
 Mostellar and Wallace (1964) examined a subset of the Federalist Papers
 There was a historical dispute about whether the author was James Madison or
Alexander Hamilton
 Using techniques that were very different than modern techniques (involving
words such as "of" and "upon"), they definitively concluded that Madison was the
author

31
 Conventional TC Approaches

 As with information retrieval, many conventional TC approaches rely on a bag-of-words

approach to represent documents
 Some approaches use a vector space model
 Some TC tasks that use alternative, task-specific approaches or representations; for example,
authorship attribution typically uses very few words, and essay grading looks at other features
of documents
 Most conventional TC systems use single words (unigrams) as terms; larger N-grams typically
did not improve results
 Different weighting schemes are possible for vector space models; for several TC strategies,
TF*IDF weights were the most common in conventional TC systems relying on vector space
models
 We will cover three conventional TC approaches; namely, Rocchio/TF*IDF, k-nearest
neighbors, and naïve Bayes; only naïve Bayes is discussed in our textbook
 Other conventional approaches that have been applied to TC include decision trees, support
vector machines, and conventional neural networks
 More recently, deep learning approaches involving word embeddings have been applied to TC,
but mostly for tasks involving short sequences of text (e.g., tweets and individual sentences)
 We will learn about deep learning approaches for TC during the seconed part of our course

32
 Text Normalization for TC

 As with IR, you must decide what constitutes a term (i.e.,

what tokenization and other text normalization strategies are
used)
 Some specific issues that TC shares in common with IR
include:
 Whether or not to make the system case sensitive (capitalization could
indicate a proper noun or the start of a sentence)
 Whether or not to use stemming (e.g., Porter stemming) or
lemmatization
 Whether or not to use a stop list
 Whether to apply POS tagging (there are at least two potential ways that
this could be used)
 Most conventional TC approaches that rely on a vector space
model normalize document vectors; some also normalize
category vectors

33
 Machine Learning for TC

 Almost all TC systems use supervised machine learning techniques; the

training data is in the form of labeled documents
 If the categories are binary, you need positive and negative examples for each
category
 If the categories are mutually exclusive and exhaustive, you need examples of each
category
 One advantage of machine learning (ML) for TC is that it is general
 To move to a new set of categories, we only need a new training set
 Sometimes it is possible to create or obtain a training set without manually labeling
any documents, but often this is not the case
 You may need to recruit volunteers to label documents
 Creating a proper training set is often the most expensive part of creating a TC
system, but resources like Amazon Mechanical Turk can make it simpler
 Some issues: How large should the training set be? How should labels be collected?
How many volunteers will label each document? What level of agreement should be
required? Etc.
 Once a TC system is trained, most text categorization approaches can be applied to
new documents one at a time; sometimes this is necessary (e.g., for spam filtering)

34
 K-Nearest Neighbors

 The k-nearest neighbors (KNN) approach is an example of instance-based

learning (a.k.a. example-based learning, memory-based learning, or lazy
learning)
 Training simply consists of recording which training instances belong to each
category
 When a new document arrive, it is compared to those in training set
 Assuming documents are represented as numerical vectors, the "closeness" of
one document to another can be measured according to Euclidean distance:
D(di , dj )= σN
x=1(wx,i − wx,j )
2

 Once all distances are computed, you find the k closest documents from the
training set
 The value of k can be manually coded or chosen based on cross-validation
experiments or a tuning set
 The categories of these k nearest neighbors will be used to select the category
or categories of the new document
 As described, this method is general, not just for text categorization

35
 Choosing a Category with KNN

 If the categories are mutually exclusive and exhaustive, the system

can simply choose the most common category among the k closest
documents
 If the categories are binary, the simplest approach is to choose all
categories that are assigned to over half of the k closest documents
 Better is to weight each of the k documents inversely proportional to
1
σk
i=1 Idi ,c ∗
D d,di +ε
its distance: Sim(d, c) = 1
σk
i=1D d,d +ε
i

 In the formula, Idi ,c is 1 if di belongs to c, 0 and otherwise; ε is a

small constant to avoid division by 0

36
 KNN Applied to TC

 For conventional text categorization, the document vectors would often be

TF*IDF vectors
 The KNN approach has performed very well in the literature for
conventional text categorization,
 For better result, we can use similarity (e.g., cosine similarity) instead of
distance to find and weight the k-nearest neighbors
 Documents can then be weighted proportional to their similarity instead of
inversely proportional to their distance
 Training for KNN is trivial and fast, but that is not too important, because
training only needs to be performed once
 To classify a document, however, it must be compared with every document
in the training set
 Efficiency is a major problem with KNN; it is generally considered to be
computationally intensive
 To combat this, there have been some efforts to find approximate nearest
neighbors more efficiently; this may be significantly faster with only a small
decrease in accuracy

37
 Naïve Bayes

 The naïve Bayes approach to categorization is a probabilistic approach based on

Bayes' theorem (which was previously encountered in many course):
P dc ∗P(c)
cො = argmax P c d = argmax = argmax [P d c ∗ P c ]
c∈C c∈C P(d) c∈C

 The argmax above assumes that categories are mutually exclusive and exhaustive
 We eliminated the denominator, P(d), because it is independent of the categories
 P(c) is just the prior probability of a category
 This can be estimated based on the training set as the frequency of training
documents falling into the category, c
 We still must estimate P(d|c)
 It is typically assumed that d can be represented as a set of features with values
 We start with the "naïve" assumption that the probability distribution for each
feature in a document given its category is not affected by the other features
 Note that naïve Bayes, as described so far, is a general approach, not just for TC

38
 Naïve Bayes Applied to TC

 Naïve Bayes for text categorization is still considered a bag-of-words approach, but it does not
use a vector space model, and it does not rely on TF*IDF word weights (or any word weights)
 The features are the words of the vocabulary, and they are typically considered Boolean features
 In other words, all that matters is whether each word does or does not appear in the document
 The "naïve" assumption for TC is that the probability of seeing a word in a document given its
specific category is not affected by the other words in the document
 Note that this is clearly not true in reality, but Naïve Bayes can still perform well for TC
 The standard technique is to estimate, based on the training set, the probability of seeing each
possible term (or word), t, in each possible category, c; we will call this P෡(t|c)
 This leads to: P෡ d c = ςt∈d P෡(t|c)
 The probability estimates are small, so we use log probabilities instead of probabilities, leading
to the predicted category (assuming categories are mutually exclusive), cො, for a document, d,
being:

cො = argmax log P c + ෍ log P෠ (t|c)

c∈C
t∈d

39
 Implementation Details of Naïve Bayes for TC

 As previously mentioned, P(c) can be estimated based on the training set as

the frequency of training documents falling into the category, c
 P෠ (t|c) is typically estimated as the percentage of training documents of
category c that contain the term t; in other words, this is a maximum
likelihood estimate, which we discussed in previous topics
 P෠ (t|c) = # training documents in category c that contain word t / # training documents in category c
 When looping through the terms in the document to categorize, you can
either loop through distinct terms or all term instances
 Looping through all term instances gives more weight to words that appear
multiple times
 However, a paper by Ken Church shows that the probability of a word
appearing twice in a document is closer to p/2 than p2
 Smoothing is very important to avoid 0 probabilities when using naïve
Bayes (we briefly discussed smoothing techniques during our topic on N-
grams)
 Naïve Bayes can achieve very good results for many sets of for TC with
various categories and other tasks

40
 Evaluating TC Systems

 Evaluating a TC system is generally easier than for IR, because there is typically a labeled test
set
 Overall accuracy is often a reasonable metric for mutually exclusive and exhaustive
categories
 However, this is not reliable for binary categories; most documents will generally not belong to
most categories, so a trivial system that says no to everything might appear to do very well
 This was true of the most common conventional TC corpus, the Reuter's corpus; it includes
9,603 training documents and 3,299 test documents; there are 135 "economic subject
categories"
 Therefore, for each category, it is common to measure precision, recall, and a metric called
F1 (or the more general F-measure)
 These metrics can be used for categorization in general, not just TC
 Consider a confusion matrix (which is a type of contingency table) where rows represent a
system's predictions, and the columns represent the actual categorizations of documents
 The next slide shows a generic example of such a confusion matrix for a single binary category
 We will define all the metrics mentioned above in relation to such a confusion matrix (formulas
are shown on the next page)

41
 TC Evaluation Metrics (per category)

 Consider the following confusion matrix for a single binary category:

Actual=Yes Actual=No
Prediction=Ye A B
s

 We can define: Prediction=No C D

 Overall accuracy = (A + D) / (A + B + C + D); this is the fraction of predictions related to
this category that were correct
 Precision = A / (A + B); of the documents that were predicted to belong to this category, this
is the fraction of them are actually do belong to the category
 Recall = A / (A + C); of the documents in the collection that actually belong to this category,
this is the fraction of them were predicted to belong to the category
 F1 = (2 * Precision * Recall) / (Precision + Recall); this combines precision and recall into a
single metric, in between precision and recall, closer to the lower of the two
 We can define these metrics similarly for mutually exclusive and exhaustive
categories (the confusion matrix would have a row and column for each
category)

42
 TC Evaluation Metrics (per system)

 When there are multiple binary categories, it is useful to compute metrics that
evaluate the system as a whole
 Two ways to combine the category-specific metrics are micro-averaging and
macro-averaging
 To compute the micro-averages for the first three metrics, you combine all the
confusion matrices (for all categories) to obtain global counts for A, B, C, and D
 You then apply the formulas from the previous slide to compute the micro-averaged
overall accuracy, precision, and recall
 The formula for F1 remains the same and is based on the micro-averaged values of
precision and recall
 To compute macro-averages for the first three metrics, you average together the
values of overall accuracy, precision, or recall for the individual categories
 The formula for F1 remains the same and is based on the macro-averaged values of
precision and recall
 Micro-averaging weighs each decision equally; macro-averaging weighs each
category equally

43
 Properties of Categories

 There are many factors to consider that can affect which

methods perform the best for a particular text categorization
task
 First, we will consider some properties of the categories (most
of these are not specific to TC):
 Are they mutually exclusive and exhaustive or independent, binary
categories?
 Are they general or specific categories?
 Related to the last question, do the documents within a category vary a
lot, or do they tend to be very similar to each other?
 Are they topic-based categories? Are they action oriented? Do they
involve sentiment? Are they abstract?
 Are there many categories or few?
 Are some categories significantly more likely than others (i.e., do some
categories contain a relatively high or low percentage of documents)?

44
 Properties of Data

 Next, we will consider some properties of the data (some of these

are specific to TC):
 Are the documents short or long?
 How big is the training set (i.e., how many documents are in it)?
 Is the training set likely to be very representative of the documents to which the
system will be applied later?
 Are the training labels accurate or is there a high degree of error? (This could
depend on how the training data is collected)
 Is there any missing data? (This is not generally an issue with TC, but it is an
issue that some machine learning methods handle better than others)
 Is there additional metadata (text or otherwise) associated with documents?
 Properties of the categories and data may very well affect which ML,
and specifically TC, methods perform well for a given task
 Especially considering the No Free Lunch theorem, we have no
reason to assume that one method will perform the best for all tasks

45
Wishing you a fruitful educatio
nal experience

46