Module 4 (INFORMATION RETRIEVAL)

The huge amount of information stored in electronic form has placed heavy demands on
information retrieval (IR) systems. As a result, information retrieval has become an
important research area.

This chapter focuses on the design of IR systems. It discusses the design features of these
systems and introduces various IR models, such as:

• Classical models: Boolean, probabilistic, and vector space retrieval models

• Non-classical models: Information logic, situation theory, and interaction information
retrieval models
• Alternative models: Clustering, fuzzy logic, and latent semantic indexing (LSI)

A detailed discussion of the vector space model is also included. The final topic in this chapter
is evaluation models used to assess the effectiveness of IR systems.

9.1 INTRODUCTION

Information Retrieval (IR) is about helping people find the information they are
looking for. When someone wants to know something, they type a question or a few
words (a query), usually in normal language. The IR system looks through a lot of stored
information and brings back documents (like articles, web pages, or reports) that seem
related to what the person asked.

History: Work on IR started around the 1960s, mainly focusing on searching text. In the
past, IR didn’t use much help from Natural Language Processing (NLP) (which is the
study of how computers understand human language). But now, because of the internet,
IR and NLP are becoming more connected. Many smart techniques from NLP are now
used in IR, like:

• Latent Semantic Indexing (LSI): Finding hidden relationships between words and
topics.
• Vector Space Retrieval: Treating documents and queries like points in space and finding
which ones are closest.

Important Point:
An IR system does not directly tell the user the answer to their question.
Instead, it shows documents where the user might find the answer.

As Lancaster (1979) said, an IR system just helps the user know where to find the
information — it doesn’t directly teach or explain the answer.

Also, in IR, a document can mean more than just text. It could also be a photo, a video, or
an audio recording.
However, this chapter only talks about text documents.
Module 4 (INFORMATION RETRIEVAL)

• IR systems: Your question might be a bit vague, and the system brings you documents
that might have what you’re looking for.

9.2 DESIGN FEATURES OF INFORMATION RETRIEVAL SYSTEMS

When you use an Information Retrieval (IR) system (like Google), the process
works like this:

1. You (the user) realize you need some information.

2. You think about it and type a query (some words describing what you
want).

3. The IR system takes your query and searches in a big collection of

documents.

4. It then shows you the documents that seem to be related to your query.

5. You look through these relevant documents to find your answer.

Module 4 (INFORMATION RETRIEVAL)

In short:
→ User → Information Need → Query → IR System → Documents → Relevant
Documents

Now, some extra important points explained simply:

• Documents are not searched directly by their full text because it would
be too slow and heavy for the computer.

• Instead, documents are summarized using keywords (important words or

phrases from the text).
Example: An article about "How to grow roses" might have keywords like
"grow", "roses", "plants".

• The job of turning full documents into lists of keywords is called indexing.

• The Inverted Index is a special structure where:

o Each keyword points to the list of documents that have that keyword.

o Example:
"roses" → Document 3, Document 5, Document 10

To make the system faster and smarter:

• Stop word elimination:

Words like "the", "is", "and" are removed because they are not useful for
search.

• Stemming:
Words are cut down to their root form.
Example: running, runs, and runner → run

• Zipf’s Law:
Very common or very rare words are ignored because they don't help much
in finding good documents.
Module 4 (INFORMATION RETRIEVAL)

Not all keywords are equally important!

Some words are more important for understanding what a document is about.
That’s why IR systems give weights to keywords — higher weights for more
important words.

Choosing the right keywords and weights is very important for making the IR
system work well!

9.2.1 Indexing

What is Indexing?

• When there are just a few documents, an IR system can easily read each
document to see if it matches a query.
• But if there are thousands or millions of documents, reading each one takes
too much time.
• So, indexing is done:
It means changing the documents into a simple list of important words
or phrases (called terms).

Why Index?

• The goal is to find good "descriptors" — important words that describe the
document.
• A good descriptor:
o Tells what the document is about.
o Helps separate it from other documents.
Module 4 (INFORMATION RETRIEVAL)

Hans Peter Luhn's Idea (1957-1958)

• He said:
o Look at how often a word appears in a document.
o Words that appear too much (like "the", "and") are not helpful.
o Words that appear only once are also not very helpful.
o Words that appear a middle number of times are the best — they help
describe the document well.

This became a famous idea for finding important words!

How a Document is Represented

Take the sentence:

“Design features of information retrieval systems”

We can break it into single words:

• Design, features, information, retrieval, systems

Or into multi-word phrases (more meaningful groups):

• Design
• Features
• Information retrieval
• Information retrieval systems

How Multi-word Terms are Found

1. N-grams:
Look for word pairs that often come together (like "information retrieval").
2. POS Tagging:
Tag each word with its grammar role (noun, verb, etc.) to find smart word
groups.
3. NLP Tools:
Use computer programs that understand language to find important phrases.
4. Manual selection:
Humans pick good phrases by hand.
Module 4 (INFORMATION RETRIEVAL)

TREC's Phrase Extraction Method

In a big project called TREC, they used a simple rule:

1. Take any two non-common words that are next to each other.
2. Keep only those word pairs that appear in 25 or more documents.

TREC = Text REtrieval Conference

Phrase Normalization

• Different ways of saying the same thing should be treated as the same phrase.
• Example:
o "President Kalam"
o "President of India"
→ Both point to the same person.
• Example:
o "Text categorization"
o "Categorization of text"
o "Text categorizAll can be te"
→ reated as the same thing: text categorize.

Summary:

• Indexing turns documents into important words/phrases.

• This makes searching much faster and better.
• Multi-word phrases and normalization help handle different ways of saying the
same thing.
Module 4 (INFORMATION RETRIEVAL)

9.2.2 Eliminating Stop Words

What is Stop Word Elimination?

• Stop words are very common words like:

"the", "and", "is", "on", "at", "a", "an", etc.

• These words are important for grammar, but not important for finding
topics.

• Because they appear in almost every document, they don't help in

identifying what the document is about.

Why Remove Stop Words?

• By removing stop words, the IR system has fewer words to store and
search.

• This makes the system faster and saves memory.

Examples of Stop Words:

• Articles: a, an, the

• Prepositions: in, on, at

• Helping verbs: is, are, was, were

Problems with Removing Stop Words:

Sometimes, removing stop words can cause problems, because:

• Some stop words are important in certain cases.

Module 4 (INFORMATION RETRIEVAL)

• Example:

o "Vitamin A" → If we remove "A", the meaning is lost.

o Famous phrases like "to be or not to be" → If stop words are removed,
the whole meaning disappears.

We remove stop words to make searching faster, but we must be careful because
sometimes stop words are important for the meaning.
Module 4 (INFORMATION RETRIEVAL)

9.2.3 Stemming

What is Stemming?

• Stemming means cutting words down to their basic root form.

• It removes prefixes and suffixes from words.
• Example:
compute, computing, computes, computer → all become comput

Why Do We Use Stemming?

• So that different forms of the same word are treated as one word.
• This groups similar words together.
• It helps the system find more relevant documents during a search.

Example:

Original sentence:
“Design features of information retrieval systems”

After:

• Stop words are removed (like "of")

• Words are stemmed and made lowercase

Result:
{design, featur, inform, retriev, system}

(Notice: "features" became "featur", "information" became "inform")

Most Famous Stemming Method:

• Porter Stemmer (created by Martin Porter in 1980).

• It's still very popular because it’s simple and works well.
Module 4 (INFORMATION RETRIEVAL)

Problems with Stemming

• Sometimes, stemming makes different words look the same, even when they
mean different things.
• Example:
o computation and computer are stemmed similarly,
o But they talk about different ideas.
• This can cause:
o Higher Recall (you get more documents — even some that are slightly
related).
o Lower Precision (you get some wrong documents — not exactly what
you want).

Stemming helps find more documents by grouping similar words, but sometimes it
may confuse different ideas.

2.4 Zipf's Law

What is Zipf's Law?

• Zipf’s Law says:

If you multiply how often a word appears (frequency) by its rank
(position when ordered by frequency), the result is about the same for
all words.
Module 4 (INFORMATION RETRIEVAL)

In easier words:

• The most common word appears a lot.

• The second most common word appears about half as much.

• The third most common word appears about one-third as much, and so on.

What the graph (Figure 9.2) shows:

• Tall bars on the left: Some words occur very frequently (like "the", "is",
"and").

• Short bars on the right: Many words occur very rarely (like
"encyclopedia", "symphony").

What does this mean for Information Retrieval (IR)?

• Common words (like "the", "is") are not helpful because they are in almost
every document.
They don't tell much about what the document is really about.

• Rare words (that appear only once or twice) are also not very helpful,
because they might not appear in any user’s search query.

• Medium-frequency words are the most useful!

They describe the document’s topic and help find relevant
documents.

Why eliminate stop words?

• High-frequency words (like "the", "and", "is") are filtered out.

• This shrinks the index (saves memory and time) and makes search results
better.
Module 4 (INFORMATION RETRIEVAL)

Zipf’s Law helps us understand that in a big collection of text:

• Few words appear a lot,

• Many words appear rarely,

• Medium-frequency words are the best for searching.

9.3 INFORMATION RETRIEVAL MODELS

An Information Retrieval (IR) model is a framework that defines several aspects

of the retrieval process, including:

• How documents and user queries are represented,

• How the system retrieves relevant documents based on a user's query,
• How the retrieved documents are ranked.

An IR system is composed of:

1. A model for documents

2. A model for queries
3. A matching function that compares queries to documents.

The main goal of an IR model is to retrieve all documents relevant to a user’s

query, forming the core function of an IR system.

Various IR models have been developed, each differing in how documents and
queries are represented and how retrieval is carried out.
Module 4 (INFORMATION RETRIEVAL)

These models are typically classified into three broad categories:

1. Classical Models of IR
2. Non-Classical Models of IR
3. Alternative Models of IR
4. Term Weighting

[Link] IR Model :It is the simplest and easy to implement IR model. This
model is based on mathematical knowledge that was easily recognized and
understood as well. Boolean, Vector and Probabilistic are the three classical IR
models.

Classical Models of IR

The three classical models are:

• Boolean Model
• Vector Space Model
• Probabilistic Model

9.4.1 Boolean Model

Introduced in the 1950s, the Boolean model is the oldest of the three classical
information retrieval (IR) models. It is based on Boolean logic and classical set
theory.

In this model:

• Documents are represented as a set of keywords, usually stored in an

inverted file.

Boolean Model in Information Retrieval (IR) -

1. What is the Boolean Model?

• It's the earliest model for information retrieval (from the 1950s).

• It uses Boolean logic — meaning it works with simple true/false decisions.

Module 4 (INFORMATION RETRIEVAL)

• Documents are described by keywords (like "information", "retrieval",

"query").

• These keywords are stored in a special file called an inverted file, which
tells you which document contains which keywords.

2. How Do Queries Work?

• Users ask questions (queries) using AND, OR, and NOT:

o AND: Document must contain both words.

o OR: Document must contain at least one word.

o NOT: Document must not contain the word.

3. How Are Documents Retrieved?

• The system checks if a document satisfies the full query.

• If it matches, the document is retrieved.

• Important:

o It’s all or nothing — either a document matches completely, or it

doesn’t.

o There’s no ranking of documents based on how relevant they are —

they are all treated equally if they match.

• Documents:

o D₁ = "Information retrieval is concerned with organization, storage,

retrieval, and evaluation of information."

o D₂ = "A user needs to formulate a request in the form of a query."

o D₃ = "The retrieval system responds by retrieving relevant

documents."

• Keywords (Terms):
Module 4 (INFORMATION RETRIEVAL)

o T = {information, retrieval, query}

• Document Representations:

o d₁ = {information, retrieval, query}

o d₂ = {information, query}

o d₃ = {retrieval, query}

• Query Example:

o Query Q = information AND retrieval

Steps to retrieve documents:

1. Find documents that contain information:

→ {d₁, d₂}

2. Find documents that contain retrieval:

→ {d₁, d₃}

3. Take the intersection (common documents) of both sets:

→ {d₁}

Result:

• Only Document D₁ matches both "information" AND "retrieval".

• So, D₁ is retrieved.

Problems with Boolean Model

• No partial match: A document must completely match the query, otherwise

it’s ignored — even if it’s almost a good match.

• No ranking: You don't know which document is better; it just gives you a
list.

• Hard queries: Users must think carefully to create a correct Boolean query
(which is not easy).
Module 4 (INFORMATION RETRIEVAL)

In the Boolean Information Retrieval Model:

• When more than one document fits the query exactly, all matching documents are
retrieved.
• The system does not try to decide which document is better or more relevant — it treats
all matching documents equally.

Advantages (Good Points):

• Simple — easy to understand.

• Efficient — quick to find matching documents.
• Good precision and recall — if the user writes a good query, it works well.

Problems (Bad Points):

1. No partial matching:
o If a document is somewhat relevant but not an exact match, the system ignores it.
o It only sees "full match" or "no match" — no in-between.
2. No ranking:
o The system can't say "this document is better than that one."
o It only cares whether the keywords are present, not how important they are.
3. Hard for users:
o Users usually don't write perfect Boolean expressions like:

Improvements to the Boolean Model:

Because of these problems, people created better models:

• P-norm model (by Salton, 1983):

→ allows more flexibility in matching documents.
Module 4 (INFORMATION RETRIEVAL)

• Fuzzy set model (by Paice, 1984):

→ allows "degrees of matching" (not just yes/no).

The Boolean model is fast and simple, but too strict.

It either finds exact matches or nothing.
It doesn't rank documents or understand "partially correct" matches .

9.4.2 Probabilistic Model

The probabilistic model is a method used in Information Retrieval (IR). Its job is
to find documents that are most likely to be relevant to what the user is searching
for.

It was developed by researchers like Maron and Kuhns (1960) and improved by
Robertson and Sparck Jones (1976).

Main Idea:

When you search something (a query), the system looks at each document and asks:

"What is the chance this document is relevant to the query?"

Then:

• It gives a probability score to each document.

• If that score is high enough, it shows you the document.

What’s Given?

You have:

• A set of documents (D),

• A search query (q),
• A cut-off threshold (α, alpha), which is the minimum acceptable probability.
Module 4 (INFORMATION RETRIEVAL)

Pick

1. It's more likely to be relevant than irrelevant.

2. Its probability of relevance is at least α.

The final set S is the list of documents that meet these two conditions.

How the Model Works (Step-by-Step):

1. The system looks at each document.

2. It calculates two values:
o How likely it is relevant (P(R/d)).
o How likely it is irrelevant (P(I/d)).
3. It compares the values.
4. If the relevance is greater or equal to irrelevance AND meets the threshold, it
keeps the document.

Why Use This Model?

Module 4 (INFORMATION RETRIEVAL)

• It ranks documents from most to least likely to help.

• It allows partial matches, not just exact ones (like Boolean).
• It improves over time if you give it feedback.

Challenges:

• It's hard to set the right threshold (α).

• Sometimes, there aren’t enough relevant documents to guess probabilities
well.
• It assumes words are independent, but in real life, words often appear together

9.4.3 Vector Space Model

The Vector Space Model (VSM) is a method used by search engines to find and rank
documents based on how closely they match a user's search (query).

It was developed by researchers like Luhn (1959), Salton (1968), and others, and has been one
of the most studied models in the field of information retrieval.

Main Idea:

The model treats documents and queries as vectors—like arrows in space. Each term (word)
becomes a dimension in a big imaginary space.

Then, it compares the angle between a query vector and each document vector to see how
similar they are.

How It Works (Step-by-Step):

1. Turn text into numbers (vectors):

• Every document and queryis converted into a vector.

• Each word in the entire collection is treated as a dimension.
• A vector is like a list of numbers. Each number tells you how important a word is in that
document or query.

o .
Module 4 (INFORMATION RETRIEVAL)

TF-IDF = Term Frequency - Inverse Document Frequency

It gives more weight to important words that appear often in a document but not in many other
documents.

For example, let’s say your collection has 3 terms:

“cat”, “dog”, “fish”.

Then, a document that has “cat” and “dog” will look like this:
D1 = [1, 1, 0]
(Query or document with “fish” only would be: [0, 0, 1])

3. Measure Similarity:

• Once both the query and each document are vectors, the system compares them.
• The most common method is cosine similarity.

Cosine similarity = how small the angle is between the query and document vectors.
A smaller angle → more similar → better match.

What Happens in Practice?

1. You search something (e.g., "renewable energy").

2. The query is turned into a vector.
3. Each document is also a vector.
4. The system computes similarity between the query vector and all document vectors.
5. Documents with higher similarity scores are considered more relevant and appear
higher in the search results.

Strengths of VSM:

• Can handle partial matches (even if not all words match).

• Supports ranking based on how closely documents match the query.
• Easy to understand and use in practice.

Limitations:

• Ignores the order of words (e.g., "white house" vs. "house white").
• Assumes terms are independent
• Can be inefficient for very large datasets without optimizations.
Module 4 (INFORMATION RETRIEVAL)

• Each row represents a term.

• Each column represents a document.

• Each cell wij tells you how important term i is in document j.

So this is called a term-document matrix.

This matrix helps the computer:

• Compare documents to queries

• Rank documents based on similarity

• Store all document info in a mathematical way

Example 9.2 Consider the documents and terms in Example 9.1. Let the
weights be assigned based on the frequency of the term within
Module 4 (INFORMATION RETRIEVAL)

the document. Then, the associated vectors will be

You are given a few documents, and for each one, a vector of numbers shows
how often a term (word) appears in that document.

The term-document matrix looks like this:

D1 D2 D3

T1 2 1 0

T2 2 0 1

T3 1 1 1

This means:

• T1 appears 2 times in D1, 1 time in D2, and 0 in D3

• T2 appears 2 times in D1, 0 in D2, and 1 in D3

• T3 appears once in every document

Problem with Raw Values

Using raw frequencies like this causes a problem:

Module 4 (INFORMATION RETRIEVAL)

A document with more words (longer) will look "stronger" or more important just
because of higher numbers. But that doesn’t mean it's more relevant to a query.

For example:

• Document D1: (2, 2, 1)

• Document D2: (1, 0, 1)

Even though D2 might be just as relevant in terms of word patterns, its
vector is smaller because it's a shorter document.

• Solution: Normalization
To solve the problem, we make all document vectors the same length (unit
length = 1).
This is called normalizing the vectors.
Module 4 (INFORMATION RETRIEVAL)
Module 4 (INFORMATION RETRIEVAL)

9.4.4 Term Weighting

Luhn’s Idea (1958)

Luhn said:

• Common words (like "the", "is") are in every document → Not helpful!
• Very rare words (like typos or strange terms) → Not helpful either!
• Mid-frequency words (used sometimes) → Very helpful!

These "middle" words are the best for telling documents apart.

TF-IDF: The Modern Way

TF-IDF = Term Frequency × Inverse Document Frequency

• Term Frequency (TF): How often the word appears in a document.

• Inverse Document Frequency (IDF): How rare the word is across all documents.
Module 4 (INFORMATION RETRIEVAL)

The Formula You Shared:

Where:

• n = total number of documents

• nᵢ = number of documents that contain term i

This is used to calculate IDF (Inverse Document Frequency).

If a term appears in many documents, then:

• nᵢ is large
• Son / nᵢ is small → low importance

If a term appears in only a few documents, then:

• nᵢ is small
• Son / nᵢ is large → high importance

Putting it All Together:

TF-IDF gives high scores to words that appear a lot in one document, but not in many others.

This helps us find the words that make each document unique or special!

Where:

• n is the total number of documents in the collection, and

• ni_ is the number of documents in which term appears.
Module 4 (INFORMATION RETRIEVAL)

IDF gives higher weight to rare words and lower weight to common words.
This helps the system know which words are more useful for telling documents
apart.

If a word is in every document (like “the” or “and”), then:

• It doesn’t help us distinguish between documents.

• So its IDF is zero, making its importance = 0.

Later, researchers combined how often a word appears in a document (TF) and
how rare it is across all documents (IDF).

This gave us the famous TF-IDF formula.

To calculate the importance of a word in a document, we do:

TF-IDF=TF×IDF

So, we:

• Multiply how often the word appears in the document (TF)

• By how rare it is across all documents (IDF using log formula)

TF-IDF combines:
Module 4 (INFORMATION RETRIEVAL)

• Local info → How often the word appears in one document

• Global info → How rare it is in all documents

This gives a balanced way to say how important a word is for searching.

TF-IDF works really well in systems where:

• The set of documents is fixed

• The user can ask different queries

In systems where new documents keep coming (like email spam detection):

• There is no fixed document set

• So we use a training set (older documents) to calculate TF-IDF
• And assume new documents are similar in behavior

Example 9.3

Consider a document represented by three terms:

(tornado, swirl, wind)

The raw term frequencies (tf) in the document are:

• tornado: 4
• swirl: 1
• wind: 1

Now, suppose in a collection of 100 documents:

• The term tornado appears in 15 documents,

• swirl in 20 documents, and
• wind in 40 documents.

We compute the idf for each term using the formula:

Module 4 (INFORMATION RETRIEVAL)

Thus, the tf-idf weighted representation of the document is approximately:

This shows that “tornado” is the most significant term in the document, reflecting both its high
frequency within the document and its relative rarity across the collection.

Figure 9.3 and Table 9.4 to understand the document processing and term weighting process
clearly. The Documents (Figure 9.3)
Module 4 (INFORMATION RETRIEVAL)

Here are 3 small sample documents:

1. Document 1: Vector space model

2. Document 2: Probabilistic retrieval model
3. Document 3: Intelligent techniques in information retrieval

We want to process these so we can compare them and search them effectively.

🔁 Step-by-step Process to Create Vector Representation

Step 1: Tokenization

Tokenization breaks the sentence into individual words (tokens) and makes them lowercase, removing
punctuation.

Example:
Document 1:

• Original: "Vector space model"

• After tokenization: ["vector", "space", "model"]
Module 4 (INFORMATION RETRIEVAL)

Step 2: Stop Word Removal

Common useless words like “in,” “and,” “the,” etc., are removed.
In Document 3:

• Original: "Intelligent techniques in information retrieval"

• After stop word removal: ["intelligent", "techniques", "information",
"retrieval"]

("in" is removed)

Step 3: Stemming

This reduces words to their root form using a process called stemming.
Example:

• "intelligent" → "intellig"
• "techniques" → "technique"
• "retrieval" → "retriev"
• "information" → "inform"
• "probabilistic" → "probabilist"

So now:

• Document 1: ["vector", "space", "model"]

• Document 2: ["probabilist", "retriev", "model"]
• Document 3: ["intellig", "technique", "inform", "retriev"]

Step 4: Term Weighting (Binary in Table 9.4)

In this case, a simple binary weighting is used:

• 1 = if the word is present in the document

• 0 = if the word is not present

Resulting Vector Table (Table 9.4)

Module 4 (INFORMATION RETRIEVAL)

Stemmed Term Doc 1 Doc 2 Doc 3

inform 0 0 1

intellig 0 0 1

model 1 1 0

probabilist 0 1 0

retriev 0 1 1

space 1 0 0

technique 0 0 1

vector 1 0 0

Each document is now represented as a vector of 0s and 1s, like this:

• Doc 1 vector: [0, 0, 1, 0, 0, 1, 0, 1]

• Doc 2 vector: [0, 0, 1, 1, 1, 0, 0, 0]
• Doc 3 vector: [1, 1, 0, 0, 1, 0, 1, 0]

These vectors allow us to:

• Compare documents using math (e.g., cosine similarity)

• Search for relevant documents
• Rank documents by relevance to a query
Module 4 (INFORMATION RETRIEVAL)

9.4.5 Similarity Measures

This explains how similarity between documents and queries can be measured
using a vector space model in information retrieval.

What’s the idea?

• Think of documents and queries as vectors (arrows) in a multi-dimensional

space.
• Each dimension represents a term (word) — in the example, two terms are
used: tᵢ and tⱼ.
• A document or query is placed in this space based on how many times each
term appears (called term frequency).

The example in the figure:

Each item (query or document) is a vector:

• Query q = (2, 1) → 2 occurrences of tᵢ, 1 of tⱼ

• Document d₁ = (2, 0) → 2 of tᵢ, 0 of tⱼ
Module 4 (INFORMATION RETRIEVAL)

• Document d₂ = (1, 0) → 1 of tᵢ, 0 of tⱼ

• Document d₃ = (1, 1) → 1 of both tᵢ and tⱼ

What’s measured?

• The angle between the query and each document vector (θ₁, θ₂, θ₃).
• Smaller angle = more similar → Closer direction means more matching
terms.

Why use this?

• It helps rank documents based on how similar they are to a user’s search
query.
• Instead of just counting matching words, this method uses math (vector
similarity) to compare documents and queries.

That’s what the image is showing — vectors pointing in directions based on words in
documents, and we measure the angle between them to find the closest match.

This is about how to measure the similarity between a query (what you search for) and a
document (what’s in a collection), using mathematical formulas. These formulas compare how
many terms (words) are common and how often they appear.

Important Terms:

• dⱼ: a document
• qₖ: a query
• wⱼᵢ: weight (e.g., frequency) of term i in document j
• wᵢₖ: weight of term i in query k
• m: total number of terms in vocabulary
Module 4 (INFORMATION RETRIEVAL)

1. Inngjer Product (Dot Product)

This is the basic way to compare document and query:

///.,

• Inner product just checks how much the document and query match.
• Dice coefficient says: “Give more score if they match well and are short.”
• Jaccard says: “Score is based on what they have in common vs total words used.”
Module 4 (INFORMATION RETRIEVAL)

Cosine Similarity

The top part (numerator) adds up the matching term weights. The bottom part (denominator)
normalizes them based on how long the vectors are (i.e., how many terms).

• It treats the document and the query as vectors (lists of numbers representing term
importance).
•
• It calculates the cosine of the angle between these two vectors.

• If the angle is 0°, the cosine is 1 → the texts are very similar (they use the same terms).
• If the angle is 90°, the cosine is 0 → the texts are completely different (no shared terms).

9.5 Non-classical models of IR:

These models are different from traditional IR models like Boolean, vector space, or probabilistic
models because they use new ideas to improve search results—especially when classical methods aren’t
good enough.

What Are Non-Classical IR Models?

These models don’t rely on just similarity, keywords, or probabilities like traditional IR.
Instead, they use deeper concepts like logic, situations, and interaction to understand what users really
want when they search.

They don’t follow the traditional way of searching documents by keyword matches or probability.
Instead, they use logic, meaning, and connections.
Module 4 (INFORMATION RETRIEVAL)

1. Information Logic Model

• This model doesn't just look for matching words.

• It tries to understand the meaning of a query and document using logic.
• Imagine you have two sentences:
o x: "It is raining."
o y: "The ground is wet."
It tries to check if y → x is true (i.e., if "the ground is wet" means "it's raining").
• But sometimes, that logic is uncertain — maybe the ground is wet because someone washed
it.
• So we measure uncertainty: how much extra info is needed to believe that one sentence
proves the other.
• This idea comes from a researcher named van Rijsbergen.

2. Situation Theory Model

• It uses something called infons. Think of them as little pieces of information.

• Each infon says if something is true or false.

Example:
Infon = "Adil is serving a dish" → this is marked as true (value = 1).

• The system checks if the situation supports this info (like someone seeing Adil serve food).
• It’s like matching facts to what’s happening.

3. Understanding Document Relevance

• A document is relevant if it supports the meaning of the query (not just the words).
• Even if it doesn't have the exact words, it might still be useful if related words are used.
o Example: Query = “car”, Document = “vehicle” → related terms
• Tools like WordNet help find related meanings (like synonyms, related concepts).
• If we change the document slightly using related terms, it may now be relevant.
Module 4 (INFORMATION RETRIEVAL)

4. Interaction IR Model

• This model sees documents as connected like a network.

• Your query interacts with this network to find results.
• Inspired by quantum mechanics: how observing something changes it.
• Documents and queries are treated like neurons in a brain (neural network).
• When a query is entered:
o New connections are made.
o Existing ones might change.
• Result: The system adapts itself to better understand and find answers.

This model can be implemented using Artificial Neural Networks, which are used in modern AI systems.

9.6 Alternative IR Model

An Alternative IR (Information Retrieval) Model is just a better version of the

classical IR model (which mostly matches keywords).

It uses techniques from other fields like:

• Math
• AI
• Logic
• Linguistics

These help the system understand meaning better and give smarter search results,
not just exact word matches.

Examples of Alternative IR Models

1. Cluster Model

• It groups similar documents together into clusters.

• If your query matches a document in one cluster, the system assumes other
documents in that group are also relevant.
Module 4 (INFORMATION RETRIEVAL)

• It helps search faster and smarter by searching in clusters instead of every

single document.

Example: A cluster might contain documents about “dogs”. So if you search

"puppy care," it searches mostly in the “dogs” group.

2. Fuzzy Model

• “Fuzzy” means not exact — it deals with uncertainty.

• Instead of saying a document is either relevant or not (0 or 1), it gives a
score between 0 and 1.
• It says: "This document is mostly relevant" or "somewhat matches" your
query.

Example: If your query is “smartphones”, a document on “mobile phones” might

be 0.8 relevant (not perfect, but close enough).

3. Latent Semantic Indexing (LSI)

• It looks at hidden meanings behind words.

• LSI doesn’t just look at exact words — it finds concepts and relationships
between words.
• It uses math (matrix factorization) to find patterns in word usage across
documents.

Example: If you search for "doctor", LSI might also show documents with
“physician” or “medical expert” — even if the word "doctor" isn’t there.

Model Simple Meaning

Cluster
Groups similar documents and searches within them.
Model
Module 4 (INFORMATION RETRIEVAL)

Model Simple Meaning

Matches documents by how much they relate to the query, not just
Fuzzy Model
exact match.
LSI Understands hidden meanings and related words in documents

Cluster Model
The Cluster Model is a way to make searching faster and more efficient by grouping similar
documents together.

Why do we need it?

When you search for something in a huge collection of documents, it's slow and wasteful to
check every single document. The idea is to cut down the number of documents we need to
look at.

What’s the basic idea?

This is based on something called the cluster hypothesis, which says:

Documents that are similar to each other are likely to be relevant to the same topic or search
query.

So, instead of comparing your search with every document, we:

1. Group similar documents into clusters.

2. Choose one representative (like a summary or sample) from each group.
3. Compare your search with just those representatives.
4. If a match is found, we only search inside that matching cluster, not the whole
collection.

Why is this helpful?

It saves time. If there are 1,000,000 documents but only 10 clusters, you only check 10
representatives first—then maybe look into just one group more deeply. That’s a big speed-up.

clustering documents based on how similar they are, to make searching faster. Instead of comparing a
search query with all documents, we group similar documents and compare the query with the groups.
Module 4 (INFORMATION RETRIEVAL)

Step-by-step explanation:

1. Start with a set of documents

We have a collection of documents:

D = {d1, d2, ..., dm}

2. Similarity matrix

We calculate how similar each document is to every other document using a similarity matrix.
Each entry e_ij in the matrix tells us how similar document di is to document dj.

3. Forming clusters

We decide on a threshold value T.

If two documents are more similar than T, they are grouped into the same cluster.
We do this for all documents and end up with groups:

C = {C1, C2, ..., Cp}

Each Ck is a cluster of similar documents.

Creating a representative for each cluster

Each cluster has a representative vector (like a summary of the whole cluster), called:

τk = {a1k, a2k, ..., amk}

Each value aik in this vector is calculated like this:

• Look at all documents in the cluster.

• For a specific term ti, average its importance (weight) across all documents in the cluster.
• That average is aik.

Formula:
Module 4 (INFORMATION RETRIEVAL)

aik = (sum of aij for all documents dj in Ci) / |Ci|

Where aij = weight of term ti in document dj.

Searching (Retrieval)

Now when someone searches using a query:

• The query is represented as a vector q.

• We compare q with each cluster representative.

Similarity between query q and cluster Ck is:

sk = sum(ai * qi) for all i

If sk is above a threshold, then we search only inside that cluster.

Why do all this?

• It reduces search time.

• Instead of matching the query with every document, we only check the most promising
cluster(s).
• It still gives accurate results, because similar documents are grouped together.
Module 4 (INFORMATION RETRIEVAL)

Great — let’s walk through the problem in the image step by step using very simple words based only
on that example:

Given:

You are given a matrix called A:

A =
[1 1 0]
[0 1 0]
[1 1 1]
[0 0 1]
[1 1 0]
Module 4 (INFORMATION RETRIEVAL)

Each column is a document:

• d1 = [1, 0, 1, 0, 1]
• d2 = [1, 1, 1, 0, 1]
• d3 = [0, 0, 1, 1, 0]

Each row is a term (a word).

Similarity Matrix (How similar are documents):

This part is already given in the image:

d1 d2 d3
d1 1.0 0.9 0.4
d2 0.9 1.0 0.4
d3 0.4 0.4 1.0

This means:

• d1 and d2 are very similar (0.9)

• d1 and d3 are not very similar (0.4)
• d2 and d3 are also not very similar (0.4)

Step 1: Form clusters (based on similarity ≥ 0.7)

Since d1 and d2 have 0.9 similarity, they are grouped together:

• Cluster C1 = {d1, d2}

d3 is not similar enough to d1 or d2, so it’s in its own group:

• Cluster C2 = {d3}

Step 2: Find cluster representatives

We now create a "summary" for each cluster, called a representative vector:

Module 4 (INFORMATION RETRIEVAL)

🔹 For Cluster C1 = {d1, d2}:

We take the average of their values:

d1 = [1, 0, 1, 0, 1]
d2 = [1, 1, 1, 0, 1]

Average (add and divide by 2):

r1 = [(1+1)/2, (0+1)/2, (1+1)/2, (0+0)/2, (1+1)/2]
= [1, 0.5, 1, 0, 1]

So:
r1 = [1, 0.5, 1, 0, 1]

🔹 For Cluster C2 = {d3}:

Just take d3 as it is (only one document):

r2 = [0, 0, 1, 1, 0]

Step 3: Searching (Retrieval)

When someone searches:

• We convert the query into a vector.

• Compare that query vector with r1 and r2.
• If it's closer to r1, we search only inside Cluster 1 (d1, d2).
• If it's closer to r2, we search only in Cluster 2 (d3).

9.6.2 Fuzzy Model

When you search something (a query), not all documents match 100%. Some match a little,
some a lot. The fuzzy model helps us measure how much each document matches your search
— not just yes or no.
Module 4 (INFORMATION RETRIEVAL)

Imagine This:

Let’s say you search for:

"retrieval"

Now there are 3 documents:

• d₁ = {information, retrieval, query}

• d₂ = {retrieval, query, model}
• d₃ = {information, retrieval}

In a regular model, it would say:

• Does document have the word "retrieval"?

Yes → show it
No → skip it

But in a fuzzy model, we ask:

• How strongly does "retrieval" appear in this document?

It gives a number between 0 and 1:

• 1 means perfect match

• 0 means no match
• 0.3 or 0.5 means partial match

What do those numbers mean?

Each term (word) in a document is assigned a weight.

This weight shows how important that word is in the document.

Example for "retrieval":

Document Weight of "retrieval"

d₁ 1/3 (medium)
d₂ 1/3 (medium)
d₃ 1/3 (medium)

All documents have "retrieval", but none too strongly. So they get 1/3 weight.
Module 4 (INFORMATION RETRIEVAL)

Now, You Search:

"retrieval AND model"

It means:
"Show me documents that talk about both retrieval and model."

We look at weights for both terms and take the minimum value:

• If a document has 1/3 for retrieval and 1/3 for model → keep it
• If one is 0 → skip it

So only d₂ is returned, because it has both "retrieval" and "model".

*9.6.3 Latent Semantic Indexing Model

Latent Semantic Indexing (LSI) is a method used in search engines to improve how information is found.
It tries to understand the meaning behind the words in documents, not just the words themselves.

Why Use LSI?

• Sometimes, different words can mean the same thing (like "car" and "automobile").

• LSI helps match documents to a search query even if they don’t use the exact same words.

• It does this by discovering hidden patterns in how words appear together across many
documents.

Step-by-Step Breakdown of Image 1:

1. Matrix Formation:

• A term-by-document matrix (W) is created.

o Rows = words (terms)

o Columns = documents

o Each cell WijW_{ij}Wij = how important word i is in document j (using term frequency or
other methods).
Module 4 (INFORMATION RETRIEVAL)

2. Use SVD to Break the Matrix:

• SVD (Singular Value Decomposition) breaks this large matrix WWW into three smaller
matrices:

Query Representation in Reduced Space:

Just like documents are reduced to a smaller topic space, the search query is also transformed using the
same concept.
Module 4 (INFORMATION RETRIEVAL)

2. Similarity Matching:

• Once the query and documents are in the same space, we can measure how close they are.

• Common method: Cosine similarity (measures angle between vectors — not just word overlap).

This lets the system:

Find documents that are semantically similar, even if they don’t share exact words with the query.

Why is this powerful?

• Keyword-based search: Only finds documents that have the same words.

• LSI-based search: Can find relevant documents that use different words but talk about similar
topics.

So LSI improves search results by:

Capturing meaning instead of just matching words.
Module 4 (INFORMATION RETRIEVAL)

Term-Document Matrix and Its SVD

Module 4 (INFORMATION RETRIEVAL)
Module 4 (INFORMATION RETRIEVAL)

Result of SVD and Low-Dimensional Representation

This image shows the actual matrices after applying SVD and reducing to 2 dimensions.
Module 4 (INFORMATION RETRIEVAL)

Why is this useful?

• Traditional retrieval looks for exact word matches.

• LSI + SVD finds hidden/semantic relationships:

o Even if two documents don’t share words directly, if they use related words, LSI can still
group or rank them as similar.

In short:

• You start with a term-document matrix.

• Use SVD to break it into components.

• Keep top 2 concepts (dimensions).

• Get a simplified version of your data.

• Helps find meaning-based similarities among documents.

9.7 Evaluation of the IR System

When we use an IR system (like a search engine), we want it to give us useful, correct results. To
evaluate how well it works, we measure several things:
Module 4 (INFORMATION RETRIEVAL)

1. System-Driven vs. User-Centered Models

• System-driven models: Focus on how well the system ranks documents.

• User-centered models: Focus on how happy the users are with the results.

Clevedon's Six Evaluation Criteria

These are six things we look at when evaluating a system:

1. Coverage of the collection: How much of the total information the system includes.

2. Time lag: How fast the system responds to a user's query.

3. Presentation format: How results are shown to the user.

4. User effort: How hard the user has to work to find what they need.

5. Precision: Out of the documents shown, how many are actually relevant?

6. Recall: Out of all relevant documents, how many did the system find?

9.7.1 Relevance
• Relevance means how useful or related a document is to your query.

• It is subjective: Different users may find different documents useful.

• Evaluating "true" relevance is hard because:

o We can't expect humans to check every document.

o Relevance can be partial (not just 0% or 100%).

There are different relevance frameworks, like system-based, psychological, and situational—where the
user's situation, time, and context are considered.

9.7.2 Effectiveness Measures

Effectiveness means how well a system (like a search engine) satisfies a user by returning documents
that are relevant to what they were looking for.

This is measured using two important metrics:

• Precision
Module 4 (INFORMATION RETRIEVAL)

• Recall

Precision and Recall

• Precision = How accurate the system is.

It tells you: Out of all the documents the system gave you, how many were actually relevant?

Example: If the system gives you 10 results, and 7 are relevant, precision = 7/10 = 0.7 (or 70%)

• Recall = How complete the system is.

It tells you: Out of all the relevant documents that exist, how many did the system actually find for
you?

Example: If there are 20 relevant documents in total and the system finds 10, recall = 10/20 = 0.5 (or
50%)
Module 4 (INFORMATION RETRIEVAL)

• To calculate recall, we need to know the total number of relevant documents in the entire
collection. But in real life, this is hard to know unless we use special test datasets like Cranfield
or TREC, which come with pre-labeled data.
Module 4 (INFORMATION RETRIEVAL)
Module 4 (INFORMATION RETRIEVAL)
Module 4 (INFORMATION RETRIEVAL)

IR Test Collections (Table 9.5)

This table shows different datasets used by researchers to test and evaluate Information Retrieval (IR)
systems. Each collection has:

• A number of documents

• A number of queries (questions or search inputs)

Example:

• TREC-1 has the largest dataset with 742,611 documents and 100 queries.

• Cranfield is smaller with 1,400 documents and 225 queries.

These collections help compare how well different IR systems perform.

Trade-off Between Precision and Recall (Fig. 9.14)

• Precision: How accurate the results are (how many of the retrieved documents are relevant).

• Recall: How complete the results are (how many of the total relevant documents were actually
retrieved).

There’s a trade-off:

• If you try to increase recall by retrieving more documents, you may include irrelevant ones,
reducing precision.

• If you only retrieve a few highly relevant documents (high precision), you may miss many other
useful ones (low recall).

Ideal Situation: Retrieve all relevant documents without any junk.

Module 4 (INFORMATION RETRIEVAL)
Module 4 (INFORMATION RETRIEVAL)

Discussion on Precision and Recall

• Many researchers have studied how precision and recall are related.

• Some use mathematical models to understand or improve them.

• A common idea is to use multi-stage retrieval to try to improve both, even if it's hard to achieve
both high precision and high recall at once.

How to Evaluate IR System Performance

• Precision at cut-off: Check how precise the results are after retrieving a certain number of
documents (e.g., top 5, top 10).

• Non-interpolated average precision: A special average that focuses only on the points where
relevant documents appear in the results.

Example (Tables 9.6 & 9.7)

Let’s say we retrieved 10 documents and marked which ones are relevant (with “x”).

Out of the top 5 documents, 3 are relevant:

• → Precision@5 = 3/5 = 0.3

Out of the top 10 documents, 5 are relevant:

• → Precision@10 = 5/10 = 0.5

This helps us measure how effective the IR system is in returning useful information to the user.
Module 4 (INFORMATION RETRIEVAL)
Module 4 (INFORMATION RETRIEVAL)
Module 4 (INFORMATION RETRIEVAL)
Module 4 (INFORMATION RETRIEVAL)
Module 4 (INFORMATION RETRIEVAL)
Module 4 (INFORMATION RETRIEVAL)

WordNet
What Is WordNet?

Think of WordNet as a very smart online dictionary made for computers and researchers.

It helps computers understand:

• What words mean.

• How words are related to other words.

• WordNet = A super dictionary that groups words with the same meanings, explains them with
examples, and links them to other related words.

Who made it?

Created by George A. Miller and his team at Princeton University.

• What does it do?

o Organizes words based on their meanings and relationships.
o Words are grouped into synsets (short for “synonym sets”):
▪ A synset = a group of words that mean the same thing.

What Is Inside WordNet?

WordNet contains three big lists:

1. Words that are nouns (like river, book),

2. Words that are verbs (like run, read),

3. Words that are adjectives/adverbs (like happy, quickly).

Example: The Word "Read"

WordNet shows that:

Module 4 (INFORMATION RETRIEVAL)

• “Read” has 1 noun meaning: like “a good read” (a book or article).

• “Read” has 11 verb meanings! Each one is a little different:

Meaning No. What it means Example sentence

1 Read something printed “Have you read the book?”

2 Say what is written “What does the law say?”

3 Say it out loud “He will read the speech aloud”

4 Read computer data “The computer read the disk”

5 Understand hidden meaning “She read the sky and predicted rain”

... ... ...

So, WordNet explains every meaning of “read” clearly.

Why Is WordNet Important?

It helps machines:

• Understand human language better.

• Tell the correct meaning of words depending on how they're used in a sentence.

Main Idea: Words have different meanings and are connected like a family tree.

Let’s say we take the word “river.”

Figure 12.5 – Hypernym Chain for "River"

This shows how "river" fits in a bigger group of things:

• River is a type of ➡ stream/watercourse

• Which is a type of ➡ body of water

• Which is a type of ➡ thing

• Which is a type of ➡ entity

Module 4 (INFORMATION RETRIEVAL)

What is a river?

WordNet says:

• A river is a large stream.

• A stream is a type of water body.

• A water body is a kind of thing or entity.

➡ So, a river is a part of a big chain of words.

This chain is shown in Figure 12.5

Word Connection Examples:

Here’s how words are connected in WordNet. It shows these in 3 small tables (Figures 12.2, 12.3, and
12.4):

Table 1: Nouns (names of things)

Word type Means… Example

Hypernym Big category "Tree" is a hypernym of "Oak" (Tree is a bigger group)

Hyponym Small type "White oak" is a hyponym of "Oak" (It's a smaller type)

Meronym Part of something "Trunk" is part of "Tree"

Holonym Whole that includes parts "Tree" includes "Trunk"

Antonym Opposite "Victory" is the opposite of "Defeat"

Table 2: Verbs (actions)

Word type Means… Example

Hypernym Big action "Travel" is a big action that includes "Wander"

Troponym Small action "Stroll" is a slower kind of "Walk"

Entail One action needs another If you "Snore", you must be "Sleeping"
Module 4 (INFORMATION RETRIEVAL)

Antonym Opposite "Increase" is the opposite of "Decrease"

Table 3: Opposites

Word Type Example

Adjective heavy ↔ light

Adverb quickly ↔ slowly

WordNet helps us:

• Understand all the meanings of a word

• Find similar or related words

• Find opposites

• Know what group a word belongs to

• WordNet is a lexical database of English words.

• It is freely available at [Link].

• It groups words into sets of synonyms called synsets, and records various semantic (meaning-
based) relations among them.

WordNets for Other Languages

• WordNet has versions for many languages, like:

o EuroWordNet (covers European languages)

o Hindi WordNet (developed by CFILT, IIT Bombay)

• These versions follow similar principles as English WordNet but may also include language-
specific relations (e.g., Hindi-specific grammar or usage).

Hindi WordNet:

• Contains:

o 26,208 synsets

o 56,928 Hindi words

Module 4 (INFORMATION RETRIEVAL)

• Each entry has:

o Synset (group of synonyms)

o Gloss (meaning)

o Ontology (position in the knowledge structure)

Example: WordNet Entries

Figure 12.6: English Word “Laugh”

• Shows different types of laughing under the main concept:

o bray = laugh loudly and harshly

o belly laugh = deep, hearty laugh

o giggle = laugh nervously

o snicker = laugh quietly

o guffaw = laugh loudly and boisterously

o etc.

These are called troponyms (specific types of a more general action).

Figure 12.7: Hindi Word “आकाांक्षा (aakanksha)”

• Means desire, aspiration, ambition.

• The entry shows multiple synonyms in Hindi.

• Example sentence:

o "हर पिता की अिने िुत्र से यह आकाांक्षा रहती है पक वह अिने जीवन में सफल हो।"
→ "Every father has this aspiration that his son should succeed in life."

Figure 12.8: Marathi Word “पाव (pau)”

• Means “step” or “footstep”.

• Shows various Marathi phrases and how the word is used.

Module 4 (INFORMATION RETRIEVAL)

Applications of WordNet

1. Word Sense Disambiguation (WSD)

• Helps in identifying the correct meaning of a word based on context.

• Uses synsets and semantic relationships (like hypernyms and hyponyms) to determine which
sense is most appropriate.

• Important in search engines, chatbots, translation, etc.

2. Automatic Query Expansion

• Helps in improving search accuracy.

• Not only matches exact query words but also includes related words (synonyms).

• Example: searching for “automobile” should also find documents with the word “car”.

FrameNet
What is FrameNet?

• FrameNet is like a big dictionary for computers that helps them understand how words are
used in real-life situations.

• It connects words (like verbs or nouns) to scenarios (called frames).

What is a Frame?

Think of a frame as a situation or event.

Each frame has:

• A main word (called the target word or predicate) — usually a verb like “arrest”, “say”, “send”.

• Roles that describe who is doing what — like "who got arrested", "who said something", "who
received something".

FrameNet helps computers understand the meaning of sentences by breaking them into parts (like
action + who + what). This is useful in many areas.

Imagine a Scene Like in a Movie

Module 4 (INFORMATION RETRIEVAL)

Let’s say there’s a sentence:

"The police arrested the thief."

What’s happening here?

• There’s an action → “arrested”

• There are people in the action → “the police” and “the thief”

FrameNet looks at this like a little movie scene. Every sentence tells a little story!

Frame = The Scene

In this case, the frame is called ARREST — it’s about the situation of someone getting caught by the
police.

FrameNet says:

• Who did the arresting? → Authorities = “The police”

• Who got arrested? → Suspect = “The thief”

• What action happened? → Target word = “arrested”

Applications

FrameNet is used in:

1. Automatic Semantic Parsing

This means:

Letting a computer understand who did what in a sentence.

Module 4 (INFORMATION RETRIEVAL)

Example:

• "The umpire stopped the match."

• "The match stopped due to rain."

In both, the word "match" is involved — but in the first, it's an object; in the second, it's the subject.

FrameNet helps the computer see that "match" plays the same role (theme) in both, even though
grammar is different.

2. Question-Answering Systems

FrameNet helps computers answer questions like:

Question: "Who sent the packet to Khushbu?"

Sentence: "Khushbu received a packet from the examination cell."

The computer can figure out:

• Who sent it? → examination cell

• Who received it? → Khushbu

• What was sent? → a packet

This is possible because of semantic roles like:

• Sender

• Receiver (Recipient)

• Goods (Item sent)

3. Information Retrieval (IR)

FrameNet helps in search engines or databases.

Instead of just searching for exact words, systems can search using meaning.

Example:
If you search “Who nabbed the thief?”, the system might find “The police arrested the thief,” because
FrameNet knows "nab" and "arrest" belong to the same ARREST frame.

4. Text Summarization
Module 4 (INFORMATION RETRIEVAL)

It helps summarize documents by focusing on the main roles and actions, not just keywords.

5. Machine Translation

FrameNet helps in better language translation by understanding the meaning of a sentence.

For example:

• “He gave her a gift” → uses the TRANSFER frame

• In another language, it can translate correctly by matching the same frame.

6. Word Sense Disambiguation

Some words have many meanings.

Example:

• "Bank" (can mean riverbank or financial bank)

FrameNet helps understand the correct meaning based on the frame.

Stemmers
Module 4 (INFORMATION RETRIEVAL)

What is Stemming?

• Stemming means reducing words to their base/root form.

• For example, "running", "runs", and "runner" can all be reduced to "run".

• It helps in search engines and other NLP tasks by grouping similar words.

Popular Stemmers:

1. Porter’s Stemmer – Very common and widely used.

2. Lovins Stemmer

3. Paice/Husk Stemmer

Each stemmer works differently, and the output from the same text can vary.

Section 12.4.1 – Stemmers for European Languages

• Many stemmers exist for languages like English, Russian, French, German, etc.

• A tool called Snowball provides these stemmers.

• Website mentioned: [Link]

Section 12.4.2 – Stemmers for Indian Languages

• Standard stemmers are still developing for Indian languages.

• Researchers like Ramanathan and Rao and Majumder et al. worked on Hindi stemming.

• IIT Bombay's CFILT developed stemmers for Indian languages.

• Website: [Link]

Section 12.4.3 – Applications of Stemming

• Used in search engines to improve search results.

o Example: Search for "astronauts" also finds "astronaut".

• Helps reduce storage space and improves document matching.

• Also useful in:

o Text summarization

o Text categorization
Module 4 (INFORMATION RETRIEVAL)

Part-of-Speech Tagger
What is POS Tagging?

POS tagging means telling what kind of word each word is in a sentence — like whether a word is a
noun, verb, adjective, etc.

For example:
In the sentence "She runs fast."

• "She" = pronoun

• "runs" = verb

• "fast" = adverb

Why is it useful?

POS tagging helps computers understand language better. It's used in:

• Speech synthesis – So computers can speak clearly.

• Machine translation – To translate between languages correctly.

• Information retrieval – To find the right information.

• Information extraction – To pull out facts from texts.

What is the Stanford POS Tagger?

It’s a computer program that does POS tagging.

• It’s very smart because it doesn’t just look at one word.

• It also looks at the words before and after to figure out the correct tag.

• It uses a technique called Maximum Entropy Markov Models (a statistical method that learns
from examples).

• • It chooses the best possible labels for all the words in a sentence.
• • It works fast and accurately and performs better than many other tools.

12.5.3 – TnT Tagger

Module 4 (INFORMATION RETRIEVAL)

• TnT stands for Trigrams’n’Tags.

• It’s another POS tagger that uses a method called Hidden Markov Model (HMM) (again, just a
math trick).

• It is very fast, works well for new words, and is as accurate as more complex methods.

Brill Tagger

• It learns rules like “If a word ends with -ing, it might be a verb.”

• It corrects mistakes step by step.

• It's like learning from experience.

• Works well even if it’s not using probabilities.

CLAWS Tagger

• It uses probabilities + rules (a hybrid system).

• Adapts well to different kinds of texts.

• Accuracy is very good (96–97%).

Module 4 (INFORMATION RETRIEVAL)

Tree-Tagger

• It uses decision trees to make choices.

• Like asking smart "yes/no" questions to find the right tag.

• Works even if there’s not much training data.

• Accuracy is above 96%.

ACOPOS Taggers (A set of 4 taggers)

1. Maximum Entropy Tagger (MET)

o Learns from many features (like nearby words) to choose the best tag.

2. Trigram Tagger (T3)

o Looks at 3-word combinations to predict the tag.

3. Transformation-based Tagger (TBT)

o Fixes mistakes by applying rules one after another.

4. Example-based Tagger (ET)

o Remembers past examples and uses them to tag new sentences.

POS Tagging for Indian Languages

• Harder because we don’t have as much labeled data.

• Languages like Hindi and Marathi are rich and complex.

• Researchers use:

o Small sample of tagged data,

o Then use that to teach a machine (called bootstrapping).

• Work is being done in places like IIT Bombay, CDAC, etc.

POS taggers like helpers that read a sentence and say:

"This is a noun, this is a verb..."

They learn using rules, examples, or math (probabilities), and they are used in things like Google
Translate or voice assistants.
Module 4 (INFORMATION RETRIEVAL)

12.6.1 IR Test Collection

IR means Information Retrieval — like when you search something on Google.

Researchers need data to test how good their search systems are. These test sets have:

• Questions (queries),

• Documents (text to search from),

• Correct answers (so we can check if the system works well).

Example: LETOR

• LETOR means Learning to Rank.

• It is a special test set made by Microsoft.

• It helps check how well a system can rank search results (put best answers on top).

• It contains:

o Questions,

o Documents,

o Features to help learn,

o Correct answers.

12.6.2 Summarization Data

When we build a summarizer (a tool that makes text shorter), we need to test it.

So we need:
Module 4 (INFORMATION RETRIEVAL)

• Long documents,

• Good human-written summaries (called gold summaries).

DUC

• DUC = Document Understanding Conference.

• It gives sets of documents with summaries.

• Used to check how good a summarizer is.

• IR test collections help test search systems.

• Summarization data helps test summary tools.

• These datasets are shared by universities and research groups to help build better tools.

This is a news article from 1988 about Hurricane Gilbert. It comes from a dataset used in
summarization research (DUC 2002). Researchers use these articles to test how well computer systems
can summarize long news into short, clear summaries.

Full Article (What happened)

The article talks about:

• Hurricane Gilbert, which started as a tropical storm and became a hurricane.

• It was heading toward the Dominican Republic with strong winds (75 to 92 mph).

• The Dominican government warned people on the southern coast to be ready.

• 100,000 people lived in the high-risk area.

• The storm caused strong winds, rains, and flooding in Puerto Rico and nearby islands.

• No deaths were reported.

• Another storm, Hurricane Florence, was mentioned but had already become weaker.

Extract (Summary of the article)

Module 4 (INFORMATION RETRIEVAL)

This is a short version of the article. It keeps only the important facts, such as:

• When and where Hurricane Gilbert formed.

• Its speed and position.

• The warnings issued by weather services.

• The potential danger to Puerto Rico, the Virgin Islands, and the Dominican Republic.

Why is this important?

Researchers give both the full article and this short summary to a machine (like a computer program).
Then they ask:

“Can the machine write a summary like this?”

If yes, the summarization system is working well.

Word Sense Disambiguation (WSD)

This section talks about how computers understand the correct meaning of words that have multiple
meanings. For example, the word “bank” can mean a financial institution or the side of a river.

To help with this task, researchers use special text collections:

• SEMCOR: A collection of texts where words are already labeled with their correct meanings
using a tool called WordNet.

• Open Mind Word Expert: A project where regular people help label word meanings online to
build a large dataset.

These resources help computers learn how to correctly understand words in different contexts.

Asian Language Corpora (Collections of Language Data)

This part discusses a project called EMILLE, created in the UK to support natural language processing
(NLP) in South Asian languages like Hindi, Bengali, Urdu, and others.

• The project created written and spoken data in 14 languages.

• The spoken data was recorded from BBC Asia radio programs.
Module 4 (INFORMATION RETRIEVAL)

• They also built translated texts, like government leaflets in English and other South Asian
languages, to help with building translation tools.

• Some of the data was labeled with grammar information (called "annotated").

The EMILLE data is free for research and is useful for developing tools like machine translation
systems and speech processing tools in South Asian languages.