Introduction to Information Retrieval & Models

1 Introduction to Information Retrieval
2 Information Retrieval Models

Mounia Lalmas
Yahoo! Research Barcelona

Information Retrieval Summer School 2011
Bangalore, India

Mounia Lalmas (Yahoo! Research) 20-21 June 2011 1 / 171

Organisation About Myself

About myself

Positions
1999-2008 Lecturer to Professor, Queen Mary University of London
2008-2010 Microsoft Research/RAEng Research Professor, University
of Glasgow
2011- Visiting Principal Scientist, Yahoo! Research Barcelona

Research topics
XML retrieval and evaluation (INEX)
Quantum theory to model interactive information retrieval
Aggregated search
Bridging the digital divide (Eastern Cape, South Africa)
Models and measures of user engagement


Organisation Outline

Outline

1 Introduction
2 Indexing (brief) and TF-IDF
3 Evaluation (brief)

4 Retrieval Models I: Boolean, VSM, BIRM and BM25
5 Retrieval Models II: Probabilities, Language Models, and DFR
6 Retrieval Models III: Query Reformulation and Relevance feedback
7 Retrieval Models IV: PageRank, inference networks, others

8 Conclusions
9 References


Introduction to IR Outline

Outline

Terminology
Retrieval Tasks
A Conceptual Model for IR
Document and Document Representation
Queries
Best-match retrieval
History
Topics in IR
Information Retrieval vs Information Extraction vs Web Search
Important forums (Conferences and Journals)


Introduction to IR Terminology

Terminology

General: Information Retrieval, Information Need, Query, Retrieval Model,
Retrieval Engine, Search Engine, Relevance, Relevance Feedback, Evalua-
tion, Information Seeking, Human-Computer-Interaction, Browsing, Inter-
faces, Ad-hoc Retrieval, Filtering

Related: Document Management, Knowledge Engineering

Expert: term frequency, document frequency, inverse document frequency,
vector-space model, probabilistic model, BM25, DFR, page rank, stemming,
precision, recall



Document/Information/Knowledge

Retrieval/Management/Engineering

document information knowledge
retrieval indexing ranking reasoning
management scanning ﬁltering learning
engineering structuring modelling annotating



Information Need

Example of an information need in the context of the world wide web:

Find all documents (information!) about universities in India that
(1) oﬀer master degrees in Information Retrieval and (2) are reg-
istered with ACM SIGIR. The information (the document!) should
include full curriculum, fees, student campus, e-mail and other con-
tact details.

Formal representation of an information need = Query



Information Retrieval: Informal Deﬁnition

Representation, storage, organisation and access of information
(information items, information objects, documents).

Find relevant (useful) information

Goal of an IR system - RECALL
Retrieve all relevant documents (e.g. legal)
Goal of an IR system - PRECISION
Retrieve the most relevant documents (e.g. web).
Goal of an IR system:
Retrieve as few non-relevant documents as possible.
Retrieve relevant documents before non-relevant documents.



Information Retrieval in Computer Science

Retrieval

s
tem
Sof

e
nc
Sys
tw

ge
O

lli
pe

are

te
ra

In
tin
Sys ory
The

g
tem

Theoretical Computer
Science
g
min Arc
gram hitect
Pro ure
l

Information
ia
fic

se

Sy
Eng
rti

aba

ste
A

m
Dat

in

s
eeri
ng



Information Retrieval / Data Retrieval

Information Retrieval Data Retrieval
Matching vague exact
Model probabilistic deterministic
Query language natural artiﬁcial
Query speciﬁcation incomplete complete
Items wanted relevant all (matching)
Error handling insensitive sensitive


Introduction to IR Retrieval Tasks

Retrieval Tasks

Ad-hoc retrieval (querying) – pull technology
Interactive query formulation
Filtering – push technology
Categorisation
Clustering
Search by browsing (hypertext)
Search by browsing and querying (digital libraries and the web)


Introduction to IR Conceptual Model for IR

A Conceptual Model for IR

Documents Information need

Indexing Formulation

Relevance feedback
Document representation Query

Retrieval function

Retrieved documents


Introduction to IR Documents and Document Representations

Documents

Unit of retrieval
A passage of free text
composed of text, strings of characters from an alphabet
composed of natural language:
newspaper articles, journal paper, dictionary deﬁnition, e-mail
messages
size of documents:
arbitrary, newspaper article vs journal article vs e-mail
Sub-document can also be a unit of retrieval (passage, XML
element, answer to a question)


Introduction to IR Documents and Document Representations

Document Representation

Free-text representation: extracted directly from text, good
performance in broad domains.
Controlled vocabulary representation: most concise
representation, good performance in narrow domains with
limited number of (expert) users.

Full-text representation: most complete representation,
optimal performance, huge resource requirements.
Reduced (partial) content representation: stopwords,
stemming, noun phrases, compression.

Structure representation: chapter, section, paragraph.


Introduction to IR Queries

Queries

Information Need
Simple queries
composed of two or three, perhaps of dozen of keywords
e.g. as in web retrieval
Boolean queries
‘neural network AND speech recognition’
e.g. as in online catalog and patent search
Context queries
proximity search, phrase queries
e.g. neural ftand network distance at most 5 words (XQuery Full Text)


Introduction to IR Best-Match Retrieval

Best-Match Retrieval

Compare the terms in a document and query
Compute “similarity” between each document in the collection and
the query based on the terms they have in common
Sorting the document in order of decreasing similarity with the query
The outputs are a ranked list and displayed to the user – the top ones
are more relevant as judged by the system

Document term descriptors to access text
←→
User term descriptors characterising user needs


Introduction to IR History

History

Manual IR in libraries: manual indexing; manual categorisation
70ies and 80ies: Automatic IR in libraries
90ies: IR on the web and in digital libraries

Success factors: Response time, coverage, interactivity, low (no!) costs,
precision-oriented (you do not “feel” the recall)

precision ≈ correctness, recall ≈ completeness


Introduction to IR Topics in IR

(Some) Topics in IR

1 Retrieval models (ranking function, learning to rank, machine learning)
2 Text processing (“Indexing”): NLP / understanding (language models)
3 Interactivity and users
4 Eﬃciency, compression, MapReduce, Scalability
5 Distributed IR (data fusion, aggregated search, federated search)
6 Multimedia: image, video, sound, speech
7 Evaluation including crowd-sourcing
8 Web retrieval and social media search
9 Cross-lingual IR (FIRE), Structured Data (XML),
10 Digital libraries, Enterprise Search, Legal IR, Patent Search, Genomics IR
(see very nice (old) article in http://www.dlib.org/dlib/november95/11croft.html)


Introduction to IR Information Retrieval vs Information Extraction

Information Retrieval vs Information Extraction

Information Retrieval
Given a set of terms and a set of document terms select only the most
relevant document (precision), and preferably all the relevant ones
(recall)
Information Extraction
Extract from the text what the document means
IR can FIND documents but needs not “understand” them


Introduction to IR Information Retrieval vs Web Search

Information Retrieval vs Web Search

Most people equate information retrieval with web search
Information retrieval is concerned with the ﬁnding of (any kind of)
relevant information


Introduction to IR Information Retrieval Forums

Information Retrieval Forums

Conferences: SIGIR, CIKM, SPIRE, FQAS, BCS-IRSG (ECIR), RIAO,
SAC-IAR, IIIX, EDCL, JCDL, IRF, ICTIR
http://www.sigir.org/events/events-upcoming.html

Journals: TOIS, IP&M, IR, JDOC, JASIST
http://www.sigir.org/resources.html


Indexing and TF-IDF Outline

Outline

Terminology
Generating document representations
Edit Distance and Soundex (e.g. used for spelling correction)
Index term weighting
Inverted ﬁle


Indexing and TF-IDF Terminology

Document views

attribute (data) view
Sailing in author = ’B. Smith’
Greece crdate = ’25.05.98’
ladate = ’30.06.99’
B. Smith
content layout view
view logical view
sailing head
greece title
mediterenean author
fish chapter
sunset section
section
chapter



“Media” types

text
image
graphic
audio (sound, speech, music)
video
animation



Document types

monomedia document: text document, etc.
multimedia document: document containing parts of diﬀerent
media
hypertext document: document with links; referred to as non-linear
document
hypermedia document: multimedia + hypertext
user generated (content) document: blogs, comments, tweets


Indexing and TF-IDF Generating Document Representations

Outline

1 Indexing language
2 Overview
3 Identify words
4 Remove stop-words
5 Detect other specials (e. g. formulae analysis, date detection)
6 Stem words
7 Using a Thesaurus (before or after stemming!?)



Indexing Language

Language used to describe documents and queries
Index terms selected subset of words
Derived from the text or arrived at independently
Keyword searching
Statistical analysis of document based of word occurrence frequency
Automated, eﬃcient and potentially inaccurate
Searching using controlled vocabularies
More accurate results but time consuming if documents manually
indexed



Overview

Documents Structure

Words (Semantic id’s!)

Stop−words

Nouns, proper nouns

Stems

Terms (Index terms)



Identify words

deﬁne word separators: white spaces are word separators
ignore punctuation: ’.’, ’,’, etc. is transposed to white space (word
separator); exception: numbers such as 1.7 or 10,000
deal with hyphen (dash) and underscore: ’-’, ’ ’; exception: formulae
deal with apostrophe and quotes
deal with numbers
deal with strings and compounds (phrases)
transpose to lower case (lower case preferred to upper case)



Remove stop-words

is, a, the, or, and, ...
not?
other?

Stop-word list often deﬁned manually. How would you derive the
stop-word list automatically?

Reduction: between 30 and 50 per cent



Detection of special expressions

Approach 1: Use heuristic rules (regular expressions) for recognising
nouns, proper nouns, credit card numbers, email addresses, phone
numbers, dates, web page locators, etc.

Approach 2: Learn rules

application-dependent!



Stemming approaches

1 dictionary-based: play+ing → play
2 rule-based:
Prefix or suffix (“affix”) removal, language-dependent.
{ sailing, sailor } → sail
Widely used stemmer: Porter stemmer (Snowball)
3 n-gram-based: information → { inf, nfo, for }

For other languages, e.g. Japanese, Chinese, etc very different



Stemming - Plural (in English)

1 If word ends in “ies” but not “eies”, aies”
“ies → “y
2 if word ends in “es” but not “aes, “ees, “oes”
“es → “e”
3 if word ends in “s but not “us” or “ss”
“s” → “”

First applicable rule is the one used



Stemming issues and terminology

under-stemming and over-stemming
over-stemming: { organisation, organ } → org
Polysemous: Several related (homologous) meanings.
Homonym: Several distinct meanings (e. g. bank, left).
Homograph: Same spelling but meaning depends on pronunciation (e.g.
bear).
Homophone: Similar pronunciation but diﬀerent spellings (night and
knight, worth and worse).
Morpheme: The smallest part of a word with a meaning.



Example

“The destruction of the Amazon rain forests”

Case normalisation
Stop word removal (From ﬁxed list)

“destruction amazon rain forests”

Suﬃx removal (stemming).

“destruct amazon rain forest”



Using a Thesaurus

bank: 1. ﬁnance institute; 2. river edge.
sailor: person sailing on boats.
Disambiguation: If bank occurs, then decide whether it has the ﬁnancial or
the river meaning.

Widely known thesaurus: WordNet
http://wordnet.princeton.edu/perl/webwn


Indexing and TF-IDF Edit distance

Edit Distance

What is the value of edit-distance(”Virginia”, ”Vermont”)=?
Scan Virginia, and replace non-matching characters.
Virginia
Verg inia
Verminia
Vermonia
Vernonta
Vermont
edit-distance(virginia, vermont) = 5


Indexing and TF-IDF Soundex

Soundex

1 b,f,p,v
2 c,g,j,k,q,s,x,z
3 d, t
Soundex translation table:
4 l
5 m, n
6 r
What about vowels?
Soundex code: One letter plus 3 digits.
Keep ﬁrst letter, discard vowels, discard repetitions, sequential consonants,
etc.


Indexing and TF-IDF Soundex

Soundex: Examples

Miller M460
Peterson P362
Petersen P362
Auerbach A612
Uhrbach U612
Moskowitz M232
Moskovitz M213


Indexing and TF-IDF Index Term Weighting

Index Term Weighting

Effectiveness of an indexing language:
Exhaustivity
number of different topics indexed
high exhaustivity: high recall and low precision
Specificity
ability of the indexing language to describe topics precisely
high specificity: high precision and low recall



Index Term Weighting

Exhaustivity
related to the number of index terms assigned to a given document
Speciﬁcity
number of documents to which a term is assigned in a collection
related to the distribution of index terms in collection

Index term weighting
index term frequency: occurrence frequency of a term in document
document frequency: number of documents in which a term occurs



Zipf’s law [1949]

Distribution of word frequencies is similar for different texts (natural
language) of significantly large size
f
Frequency of words

r
Words by rank order

Zipf’s law holds even for different languages!



Luhn’s analysis — Hypothesis

[1957]
Hypothesis: Frequency of words is a measurement of word
signiﬁcance.
... a measurement of the power of a word to discriminate documents
by their content ...



Luhn’s analysis — Observation

f Upper cut−off Lower cut−off
Frequency of words

common words

Resolving power

Significant rare words
words
r
Words by rank order



Luhn’s analysis — Explanation

Resolving/Discriminating power of words
Optimal power half way between the cut-oﬀs



tf-idf

weight(t, d) = tf (t, d) × idf (t)
N number of documents in collection
n(t) number of documents in which term t occurs
idf (t) inverse document frequency of term t
occ(t, d) occurrence of term t in document d
tmax term in document d with highest occurrence
tf (t, d) term frequency of t in document d
SMART retrieval system, Salton [1971]



occ(t, d)
tf (t, d) :=
occ(tmax , d)

With lifting factor:

occ(t, d)
tf (t, d) := l + (1 − l)
occ(tmax , d)

occ(t, d)
tf (t, d) := 0.5 + 0.5
occ(tmax , d)

N n(t)
idf (t) := log idf (t) = − log
n(t) N


Indexing and TF-IDF Inverted ﬁle

Inverted ﬁle
TERM IDF DOC TF

Word-oriented mechanism for indexing collections to speed up
searching
Searching:
vocabulary search (query terms)
retrieval of occurrence
manipulation of occurrence



Document vs Inverted Views



What goes in the inverted ﬁle

Boolean retrieval
Just the document number
Ranked Retrieval
Document number and term weight (TF, IDF, TF*IDF, ...)
Proximity operators
Word oﬀsets for each occurrence of the term
Example: t17 (doc1,49) (doc1,70) (doc2,3)



How big is the inverted ﬁle

Very compact for Boolean retrieval
About 10% of the size of the documents
If an aggressive stopword list is used
Not much larger for ranked retrieval
Perhaps 20%
Enormous for proximity operators
Sometimes larger than the documents
But access is fast - you know where to look


Evaluation

Outline

What to evaluate
Test collections
Precision and recall


Evaluation What can we evaluate in IR

What can we evaluate in IR

coverage of the collection: extent to which the system includes
relevant material
this is (was) important in web retrieval (since it was the case that
individual search - Altavista, Lycos, etc) engine covers maybe up to
16% of the web space.
efficiency in terms of speed, memory usage, etc.
time lag (efficiency): average interval between the time a request is
made and the answer is given
presentation of the output, has to do with interface and visualisation
issues.
effort involved by user in obtaining answers to a request
recall of the system: proportion of relevant documents retrieved
precision of the system: proportion of the retrieved documents that
are actually relevant


Evaluation What can we evaluate in IR

System-oriented evaluation

Test collection methodology
Benchmark (data set) upon which eﬀectiveness is measured and
compared
Data that tell us for a given query what are the relevant documents.
Measuring eﬀectiveness has been the most predominant in IR
evaluation:
recall of the system: proportion of relevant documents retrieved
precision of the system: proportion of the retrieved documents that
are actually relevant
Looking at these two aspects is part of what is called
system-oriented evaluation.


Evaluation Test Collections

Test Collections

Compare retrieval performance using a test collection
Document collection, that is the document themselves. The document
collection depends on the task, e.g. evaluating web retrieval requires a
collection of HTML documents.
Queries / requests, which simulate real user information needs.
Relevance judgements, stating for a query the relevant documents.
To compare the performance of two techniques:
each technique used to evaluate test queries
results (set or ranked list) compared using some performance measure
most common measures -precision and recall
Usually use multiple measures to get diﬀerent views of performance
Usually test with multiple collections as performance is collection
dependent


Evaluation Precision and Recall

Effectiveness

We recall that the goal of an IR system is to retrieve as many relevant
documents as possible and as few non-relevant documents as possible.
Evaluating the above consists of a comparative evaluation of technical
performance of IR system(s):
In traditional IR, technical performance means the effectiveness of the
IR system: the ability of the IR system to retrieve relevant documents
and suppress non-relevant documents
Effectiveness is measured by the combination of recall and precision.



Recall / Precision

For a given query, the document collection can be divided into three sets:
the set of retrieved document, the set of relevant documents, and the rest
of the documents.

Note: knowing which documents are relevant comes from the test
collection



Recall / Precision
In the ideal case, the set of retrieved documents is equal to the set of
relevant documents. However, in most cases, the two sets will be diﬀerent.
This diﬀerence is formally measured with precision and recall.

number of relevant documents retrieved
Precision =
number of documents retrieved
Recall =
number of relevant documents


Recall / Precision

Precision =
number of documents retrieved
Recall =
number of relevant documents

The above two measures do not take into account where the relevant
documents are retrieved, this is, at which rank (crucial since the output of
most IR systems is a ranked list of documents).
This is very important because an eﬀective IR system should not only
retrieve as many relevant documents as possible and as few non-relevant
documents as possible, but also it should retrieve relevant documents
before the non-relevant ones.


Recall / Precision

Let us assume that for a given query, the following documents are relevant (10
relevant documents)
{d3, d5, d9, d25, d39, d44, d56, d71, d89, d123}
Now suppose that the following documents are retrieved for that query:
rank doc precision recall rank doc precision recall
1 d123 1/1 1/10 8 d129
2 d84 9 d187
3 d56 2/3 2/10 10 d25 4/10 4/10
4 d6 11 d48
5 d8 12 d250
6 d9 3/6 3/10 13 d113
7 d511 14 d3 5/14 5/10
For each relevant document (in red bold), we calculate the precision value and the
recall value. For example, for d56, we have 3 retrieved documents, and 2 among
them are relevant, so the precision is 2/3. We have 2 of the relevant documents so
far retrieved (the total number of relevant documents being 10), so recall is 2/10.



Recall / Precision

For each query, we obtain pairs of recall and precision values
In our example, we would obtain (1/10, 1/1) (2/10, 2/3) (3/10, 3/6)
(4/10, 4/10) (5/10, 5/14) . . . which are usually expressed in % (10%,
100%) (20%, 66.66%) (30%, 50%) (40%, 40%) (50%, 35.71%) . . .
This can be read for instance: at 20% recall, we have 66.66% precision;
at 50% recall, we have 35.71% precision

The pairs of values are plotted into a graph, which has the following curve



The complete methodology

For each IR system / IR system version

For each query in the test collection
We ﬁrst run the query against the system to obtain a ranked list of
retrieved documents
We use the ranking and relevance judgements to calculate
recall/precision pairs
Then we average recall / precision values across all queries, to obtain
an overall measure of the eﬀectiveness.



Averaging

Recall in % Precision in %
Query 1 Query 2 Average
10 80 60 70
20 80 50 65
30 60 40 50
The same information
40 60 30 45
can be displayed in a
50 40 25 32.5
plot.
60 40 20 30
70 30 15 22.5
80 30 10 20
90 20 5 11.5
100 20 5 11.5



Comparison of systems

We can compare IR systems / system versions. For example, here we see
that at low recall, system 2 is better than system 1, but this changes from
recall value 30%, etc. It is common to calculate an average precision value
across all recall levels, so that to have a single value to compare, so called
Mean average precision (MAP).


Retrieval Models

Outline

Notations - Revision
Components of a retrieval model
Retrieval Models I: Boolean, VSM, BIRM and BM25
Retrieval Models II: Probabilities, Language Models, and DFR
Retrieval Models III: Relevance feedback
Retrieval Models IV: PageRank, inference networks, others


Retrieval Models Notations – Revision

(Mathematical) Notations – Revision

d → q: d implies q as in classical logic
d ∩ q: the intersection of the set d and the set q
|d|: the cardinal of the set d, ie the number of elements in the set d
d ∪ q: the union of the set d and the set q
i=1,n ai = a1 + a2 + . . . + an
i=1,n ai = a1 · a2 · . . . · an


Retrieval Models Components of a retrieval model


D is the set of document representations (called call from now on
documents for simplicity)
Q is the set of information need representations (called from now on
queries)
R(d, q) is a ranking function that
associates a real number, usually between 0 and 1, for a document
d ∈ D and a query q ∈ Q
can be used to deﬁne an ordering for the documents in D with respect
to the query q,
where the ordering is suppose to reﬂect relevance (hopefully).


Retrieval Models Components of a retrieval model


For each retrieval model, we will make explicit the three components:
Document representation d
Query q
Ranking function R(d, q)


Retrieval Models Retrieval Models I: Boolean, VSM, BIRM and BM25

Boolean model

Retrieve documents that make the query true.

1 if d → q
R(d, q) =
0 otherwise

Query (and document): logical combination of index terms
q = (sailing ∧ boats) ∨ (bowskill ∧¬ south coast)
sailing = { d1, d2, d3, d4 }
boats = { d1, d2 }
“Query evaluation” based on inverted ﬁle:
bowskill= { d1, d2, d3 }
south coast= { d1 }
Negation “felt” to be useful — proved to be not eﬀective! Leads to empty
results!
No ranking: either a document is retrieved or not: {d1 , d2 , d3 }



Set-based models — co-ordination level match

Query q: set of index terms and Document d: set of index terms
Ranking based on the cardinality of d ∩ q, ie number of terms
common to the document and the query

R(d, q) = |d ∩ q|

For q = { sailing, boats, bowskill }, the outcome is a (partially
ranked) list of documents
|d ∩ q| documents rank
3 d1, d2 1
2 d3 3
1 d4 4
Note: counter-intuitive results may occur due to document and query
length.



Set-based models — other coeﬃcients

R(d, q) is based on |d ∩ q| but with some normalisation so that to not
obtain counter-intuitive results
2 · |d ∩ q|
Dice: R(d, q) =
|d| + |q|

|d ∩ q|
Jaccard: R(d, q) =
|d ∪ q|

|d ∩ q|
Cosine: R(d, q) =
|d| · |q|



Vector space model — Introduction

Set of n terms {t1 , t2 , . . . , tn } (order is important)
Document represented as a vector: d =< d1 , d2 , . . . , dn >
Query represented as a vector: q =< q1 , q2 , . . . , qn >
di = weight of term ti in document d (eg, based on tf × idf )
qi = weight of term ti in query q (eg, 1 if ti ∈ q; 0 otherwise)
Ranking function, called retrieval status value (often written RSV):

i=1,n di qi
R(d, q) = RSV (d, q) = = cos α
( i=1,n di2 )1/2 ( i=1,n qi2 )1/2

Dimension of the vector is n
Special case: binary vectors (cosine function as given in previous slide)



Vector space model — Graphical interpretation

t2
d2 D

q2 Q

α

d1 q1 t1

Here n = 2, meaning two terms in the collection.



Vector space model — Vector Notation

Document represented as a vector: d =< d1 , d2 , . . . , dn >
Query represented as a vector: q =< q1 , q2 , . . . , qn >
Ranking function (retrieval status value):

i=1,n di qi
R(d, q) = 2 )1/2 (
= cos α
( i=1,n di i=1,n qi2 )1/2

d ·q
R(d, q) = sim(d, q) = cos α =
d2 · q2



Generalised vector space model

R(d, q) := d T · G · q

1 1
G=
0 1

1 1 q1 q1
d1 d2 · · = d1 (d1 + d2 ) ·
0 1 q2 q2

= d1 · q1 + (d1 + d2 ) · q2 = d1 · q1 + d2 · q2 + d1 · q2



Generalised vector-space model

1 1
G= R(d, q) = d1 · q1 + d2 · q2 + d1 · q2
0 1

Relationships between terms are considered, as modelled in G :
Upper triangle: G1,2 = 1 produces (d1 · q2 ). Lower triangle: ???
Here if term t1 in document, and term t2 in query, then consider a
match, but not the other way round.
Example: t1 =dog and t2 =animal



Probabilistic Model

“Given a user query q and a document d, estimate
the probability that the user will ﬁnd d relevant.”

We only discuss the Binary Independence Retrieval Model (BIRM)
based on information related to presence and absence of terms in
relevant and non-relevant documents
information acquired through relevance feedback process:
user stating which of the retrieved documents are relevant /
non-relevant (covered later)



Binary independence retrieval model (BIRM)

A document is described by presence/absence of terms:
d =< x1 , x2 , . . . , xn > with n = number of terms.

1 if document d indexed by ti
xi =
0 otherwise

1- compute for given query q:
P(r |d, q), the probability of d being relevant (r )
P(¬r |d, q), the probability of d not being relevant (¬r )
2- then decide whether document represented by d is relevant to query
q.
The decision is expressed by the Bayes’ decision rule.



BIRM: The Bayes’ decision rule

For each query q deﬁned as a set of terms, we have a set of relevant
documents (binary vectors)
P(r |d, q): probability of judgement being relevant (r ) given
document d and query q
P(¬r |d, q): probability of judgement being not relevant (¬r ) given
document d and query q

if P(r |d, q) > P(¬r |d, q)
retrieve d
Bayesian decision rule: else
do not retrieve d



BIRM: Bayes’ decision rule and retrieval function

Bayes’ decision rule:
“if P(r |d, q) > P(¬r |d, q) then retrieve d; otherwise don’t”

From above decision rule, a retrieval function R(d, q) = g (d, q) is derived:

g (d, q) > C retrieve document represented by d
g (d, q) =
g (d, q) ≤ C do not retrieve document represented by d

for some constant C

We show how g (d, q) is obtained.



BIRM: The Bayes’ decision rule

if P(r |d, q) > P(¬r |d, q)
retrieve d
else
do not retrieve d

The rule says: if P(r |d, q) > P(¬r |d, q) then d is relevant for query
q; otherwise d is not relevant.
To implement this rule, need to compute P(r |d, q) and P(¬r |d, q)
Since these probabilities are with respect to same query q, simplify
the above to P(r |d) and P(¬r |d)
→ We show how to obtain g (d, q) = g (d)



BIRM: Bayes’ theorem

The rule is implemented through the use of Bayes’ theorem

P(d|r ) · P(r ) P(d|¬r ) · P(¬r )
P(r |d) = P(¬r |d) =
P(d) P(d)

P(d): probability of observing d at random, ie probability of d irrespective
of whether it is relevant or not.
P(d|r ): probability of observing d given relevance
P(d|¬r ): probability of observing d given non relevance
P(r ): prior probability of observing a relevant document
P(¬r ): prior probability of observing a non relevant document
Note that from probability theory: P(d) = P(d|r ) · P(r ) + P(d|¬r ) · P(¬r )



BIRM: Independence Assumption

We recall that d =< x1 , x2 , . . . , xn > where xi = 1 or 0.

BIRM assume independence with respect to relevance:

P(d|r ) = P(< x1 , . . . , xn > |r ) = P(xi |r )
i=1,n

BIRM assume independence with respect to non relevance:

P(d|¬r ) = P(< x1 , . . . , xn > |¬r ) = P(xi |¬r )
i=1,n



BIRM: Notations

ai := P(xi = 1|r ): probability that term ti occurs in a relevant
document
1 − ai = P(xi = 0|r ): probability that term ti does not occur in a
relevant document
bi := P(xi = 1|¬r ): probability that term ti occurs in a non-
relevant document
1 − bi = P(xi = 0|¬r ): probability that term ti does not occur in a
non-relevant document

(In literature, you often ﬁnd pi and qi . Leads to confusion with P and q!)



BIRM: Using the notations

P(d|r ) = P(xi |r ) = aixi (1 − ai )1−xi
i=1,n i=1,n

P(d|¬r ) = P(xi |¬r ) = bixi (1 − bi )1−xi
i=1,n i=1,n

Example: Document d =< 0, 1, 1, 0, 0, 1 > and n = 6 (6 terms):
P(< 0, 1, 1, 0, 0, 1 > |r ) = (1 − a1) · a2 · a3 · (1 − a4 ) · (1 − a5 ) · a6
P(< 0, 1, 1, 0, 0, >)|¬r ) = (1 − b1 ) · b2 · b3 · (1 − b4 ) · (1 − b5 ) · b6



BIRM: The way to the retrieval function g (d)

We return now to:
P(d|r ) · P(r )
>1
P(d|¬r ) · P(¬r )

P(r )
For a set of documents, P(¬r ) is constant, so we only have to deal with:

P(d|r )
>1
P(d|¬r )

Using the indepenpendence assumptions, and notations:
xi 1−xi
i=1,n P(xi |r ) i=1,n ai (1 − ai )
= xi 1−xi
>1
i=1,n P(xi |¬r ) i=1,n bi (1 − bi )




From the following:
xi 1−xi
i=1,n ai (1 − ai )
xi 1−xi
>1
i=1,n bi (1 − bi )

We take the log:
xi 1−xi
i=1,n ai (1 − ai )
log xi 1−xi
> log(1) = 0
i=1,n bi (1 − bi )

This gives:

ai (1 − bi ) 1 − ai
xi log + log >0
bi (1 − ai ) 1 − bi
i=1,n i=1,n




From:
ai (1 − bi ) 1 − ai
xi log + log >0
bi (1 − ai ) 1 − bi
i=1,n i=1,n

we obtain:

g (d) = i=1,n ci · xi + C

where
ai (1 − bi )
ci = log
bi (1 − ai )
1 − ai
C= log
1 − bi
i=1,n



BIRM: Why such a g (d)?

ci are weights associated with terms ti , eg discrimination power.
Simple addition:
for ci > 0, term ti occurring in document is a good indication of
relevance
for ci < 0, term ti occuring in document is a good indication of
non-relevance
for ci = 0, term ti occuring in document means nothing
C constant for all documents given the same query:
cut-oﬀ value
Retrieval strategy:
if g (d) ≥ C then retrieve d; otherwise do not retrieve d
or simply rank by g (d) value (ignore C )



BIRM: Estimating ci

RELEVANT NON-RELEVANT
xi = 1 ri ni − ri ni
For each term ti :
xi = 0 R − ri N − ni − R + ri N − ni
R N −R N
ni : number of documents with term ti
ri : number of relevant documents with term ti
R: number of relevant documents
N: number of documents

These data can be extracted after a relevance feedback process: user points out
the relevant documents from a list of retrieved documents.



BIRM: Estimating ci

We recall:
ai (1 − ai ): probability that a relevant document contains (does not
contain) the term ti
bi (1 − bi ): probability that a non relevant document contains (does
not contain) the term ti

ri ni − ri
ai = bi =
R N −R
so
ai (1 − bi ) ri /(R − ri )
ci = log = log
bi (1 − ai ) (ni − ri )/(N − ni − R + ri )



BIRM: Estimating ci - RSJ weights

ri /(R − ri )
ci = log
(ni − ri )/(N − ni − R + ri )

is usually re-written:

(ri + 0.5)/(R − ri + 0.5)
ci = log
(ni − ri + 0.5)/(N − ni − R + ri + 0.5)

0.5 is added to keep the ci value from being inﬁnite when ri and R are
small.
ci is also referred to as term weight in BIRM; also referred to as
Robertson-Spark Jones (RSJ) weights and written w (1) .



BIRM: How does it work in practice?

When no sample is available, R is not known
set ai = 0.5 and bi = ni /N
leads to ci = log(N − ni )/ni which can be viewed as a probabilisitic idf
g (d) thus with idf weights produces initial ranking
Relevance feedback is then applied, and R, ri can be deﬁned, which
has been shown to improve ranking.



BIRM: Example – Using the original ci weights

2 terms t1 and t2 ; d = (x1 , x2 ); 20 documents d1 , . . . d20 ;
the query is made of term t1 and t2

d Rel x1 x2 d Rel x1 x2 d Rel x1 x2
d1 r 1 1 d2 r 1 1 d3 r 1 1
d4 r 1 1 d5 ¬r 1 1 d6 r 1 0
d7 r 1 0 d8 r 1 0 d9 r 1 0
d10 ¬r 1 0 d11 ¬r 1 0 d12 r 0 1
d13 r 0 1 d14 r 0 1 d15 ¬r 0 1
d16 ¬r 0 1 d17 ¬r 0 1 d18 r 0 0
d19 ¬r 0 0 d20 ¬r 0 0

N = 20; R = 12; r1 = 8; r2 = 7; n1 = 11 and n2 = 11



BIRM: Example

a1 = r1 /R = 8/12; a2 = 7/12;
b1 = (n1 − r1 )/(N − R) = (11 − 8)/(20 − 12) = 3/8; b2 = 4/8
Thus: (use ln for the logs)

a1 (1 − b1 )
c1 = log = log 10/3 = 1.20
b1 (1 − a1 )

c2 = log 7/5 = 0.34
Retrieval function: g (D) = 1.20x1 + 0.34x2 + C



BIRM: Example: Result

Retrieval results (here we ignore C ):

Rank Document g (d)
Rank 1 d1 , d2 , d3 , d4 , d5 1.54
Rank 6 d6 , d7 , d8 , d9 , d10 , d11 1.20
Rank 12 d12 , d13 , d14 , d15 , d16 , d17 0.34



BIRM: Summary

Probabilistic model uses probability theory to model the “uncertainty”
in the retrieval process.
Assumptions (here independence assumptions) are made explicit
Term weight (ci ) without relevance information is inverse document
frequency (this can be proven).
Relevance feedback can improve the ranking by giving better
probability estimates of term weights.
No use of within-document term frequencies or document lengths.



Building on the probabilistic model: Okapi weighting

Okapi system is based on the probabilistic model
BIRM does not perform as well as the vector space model
does not use term frequency (tf ) and document length (dl)
hurt performance on long documents

What Okapi does:
add a tf component like in the vector space model
separate document and query length normalization
several tuning constants, which depend on the collection



BM25 (Best-match Okapi weight)

R(d, q) = BM25(d, q) =

(k1 + 1)tf (t, d) (k3 + 1)tf (t, q) avgdl − dl
wt · · + k2 · |q| ·
t∈q
K + tf (t, d) k3 + tf (t, q) avgdl + dl

K = k1 ((1 − b) + (b · dl)/avdl)
wt term weight based on relevance feedback (RSJ - w (1) )
or IDF
tf (t, d), tf (t, q) within term frequencies – document and query
k1 , k2 , k3 , b tuning parameters
dl, avgdl document length and average document length



BM25 – Parameters

(k1 + 1)tf (t, d) (k3 + 1)tf (t, q) avgdl − dl
wt · · + k2 · |q| ·
t∈q
K + tf (t, d) k3 + tf (t, q) avgdl + dl

K = k1 ((1 − b) + (b · dl)/avdl)
k1 : governs the importance of within document frequency tf (t, q)
k2 : compensation factor for the high within document frequency values
in large documents
k3 : governs the importance of within query frequency tf (t, q)
b : relative importance of within document frequency and document
length
The theoretical basis for the Okapi formula is the use of Poisson distributions to
model within document frequency in relevant documents, and in non-relevant
documents
(not discussed here).


BM25 (Best-match Okapi weight)

Experiments show:
k2 =0; k3 large; b closer to 1
Leading for instance to ( with k1 = 1 and b = 0.75):

tf (t, d)
BM25(d, q) = wt ·
t∈q
K + tf (t, d)

K = 0.35 + (0.75 · dl)/avdl
In experiments, Okapi weights give the best performance.
BM25 often used as baseline model in retrieval experiments.



Summary

The vector space model is the most basic one.
The BIRM is one of the important pieces of IR theory.
A ranking based on the probability of relevance is optimal with respect
to a cost function where the costs for reading relevant documents are
low and the costs for reading non-relevant documents are high
(probability ranking principle).
BM25 Okapi model is often the most “eﬀective” model, the model to
“beat” in retrieval experiments.
BM25F (BM25 Field) – take document structure and anchor text into
account


Retrieval Models Retrieval Models II: Probabilities, Language models and DFR

Outline

A recap
Language model (LM)
Divergence from randomness model (DFR)



Boolean Model - Recap

No Rank: A document is judged to be relevant if the terms in the
document satisﬁes the logical expression of the query

A document is represented as a set of keywords (i.e. model of documents)

Queries are Boolean expressions of keywords, connected by AND, OR,
and NOT (i.e. model of queries)



Vector Space Model - Recap

Rank according to the similarity metric (e.g. cosine) between the
query and document.
The smaller the angle between the document and query the more
similar they are believed to be.
Documents are represented by a term vector
Queries are represented by a similar vector
Ad-hoc weightings (term frequency x inverse document frequency )
used
No optimal ranking



Binary Independence Retrieval Model

Rank by the probability of a document being relevant to the query:

P(r |d, q)

Documents are represented by a binary term vector
Absence or presence of terms
We cannot estimate P(r |d, q) directly, so we evoke Bayes’ rule, to
obtain P(d|q, r ), which itself leads to the function g (d, q)
Based on the probability ranking principle, which “ensures” an
optimal ranking



Empirically Based
Probability Ranking Principle
Success measured by
Minimises risk
experimental results
Justiﬁes decision
Ad hoc weighting schemes
Few properties provable Theoretical Framework
Nice theoretical properties, but
Sometimes you want to analyze
performance beneﬁts are unclear
properties of methods
Extensible
Is similar, relevant?



Generative Probabilistic Models

The generative approach A generator which produces events/tokens
with some probability

URN Metaphor a bucket of diﬀerent colour balls (10 red, 5 blue, 3
yellow, 2 white)
What is the probability of drawing a yellow ball? 3/20
What is the probability of drawing (with replacement) a red ball and a
white ball? 10/20 × 2/20 = 1/2 × 1/10
What is the probability of generating the sequence red ball and white
ball (with replacement)?
IR Metaphor: Documents are urns, full of tokens (balls) of (in)
diﬀerent terms (colours)



Generative Models - Language model

A statistical model for generating data

Probability distribution over samples for a given language
M → t1 t2 t3 t4



Generative Probabilistic Models

What is the probability of producing the query from a document?
P(q|d)
Referred to as the query-likelihood
Assumptions:
The probability of a document being relevant is strongly correlated
with the probability of a query given a document, i.e. P(d|r ) is
correlated with P(q|d)
System’s task is to estimate for each of the documents in the
collection, which one is the most likely to generate the query.



Language Models in IR (1998)

Let us assume we point blindly, one at a time, at 3 words in a
document
What is the probability that I, by accident, pointed at the words
“Master”, “computer”, and “Science”?
Compute the probability, and use it to rank the documents.



Types of language models

Unigram Models (Assume word independence)

Bigram Models

There are others . . .



The fundamental problem

Usually we do not know the model M, but have a sample
representative of that model

First estimate a model from a sample
Then compute the observation probability



Example for unigram models

(Urn metaphor)



Example - Ranking documents with unigram models

Rank models (documents) by probability of generating the query
Q:
P( | ) = 4/9 ∗ 2/9 ∗ 4/9 ∗ 3/9 = 96/6561

P( | ) = 3/9 ∗ 3/9 ∗ 3/9 ∗ 3/9 = 81/6561

P( | ) = 2/9 ∗ 3/9 ∗ 2/9 ∗ 4/9 = 48/6561

P( | ) = 2/9 ∗ 5/9 ∗ 2/9 ∗ 2/9 = 40/6561



Standard LM Approach

Assume that query terms are drawn identically and independently
from a document (unigram models)

P(q|d) = P(t|d)n(t,q)
t∈q

(where n(t, q) is the number of term t in query q)
Maximum Likelihood Estimate of P(t|d)
Simply use the number of times the query term occurs in the document
divided by the total number of term occurrences.
Problem: Zero Probability (frequency) Problem



The Zero-frequency Problem

Suppose some event not in our example
Model will assign zero probability to that event
And to any set of events involving the unseen event
Happens frequently with language
It is incorrect to infer zero probabilities
Especially when dealing with incomplete samples



Document Models

Solution:
Infer a language model (θd ) for each document, where P(t|θd ) > 0 for
all t
Then we can estimate P(q|θd )
Standard approach is to use the probability of a term p(t) to smooth
the document model, thus

P(t|θd ) = λP(t|d) + (1 − λ)P(t)



Estimating Document Models

Basic Components
Probability of a term given a document (maximum likelihood estimate)

n(t, d)
P(t|d) =
t n(t , d)

Probability of a term given the collection

d n(t, d)
P(t) =
t d n(t , d )

n(t, d) is the number of times term t occurs in document d



Smoothing

Idea: shift part of probability mass to unseen events
Interpolation with background (General English in our case)
Reﬂects expected frequency of events
Plays role of IDF in LM

λ + (1 − λ)



Estimating Document Models

Example of Smoothing methods
Laplace
n(t, d) + α
P(t|θd ) =
t n(t , d) + α|T |
|T | is the number of term in the vocabulary
Jelinek-Mercer

P(t|θd ) = λ · P(t|d) + (1 − λ) · P(t)

Dirichlet
|d| µ
P(t|θd ) = · P(t|d) + · P(t)
|d| + µ |d| + µ



Language Models - Implementation

We assume the following LM (Jelinek-Mercer smoothing):
n
P(q = t1 , t2 , . . . , tn |d) = ((1 − λ) · P(ti ) + λ · P(ti |d))
i=1

It can be shown that the above leads to:
n
λ · P(ti |d)
P(q = t1 , t2 , . . . , tn |d) ≈ log(1 + )
(1 − λ) · P(ti )
i=1

for ranking purpose (again use log to obtain summation)



Document Priors

Remember P(d|q) = P(q|d)P(d)/P(q) ≈ P(q|d)P(d)
P(d) is typically assumed to be uniform so is usually ignored leading
to P(d|q) ≈ P(q|d)
P(d) provides an interesting avenue for encoding a priori knowledge
about the document
Document length (longer doc → more relevant)
Average Word Length (bigger words → more relevant)
Time of publication (newer doc → more relevant)
Number of web links (more in links → more relevant)
PageRank (more popular → more relevant)



“Language Modelling”

Not just “English”
But also, the language of
author
newspaper
text document
image
structure
...



Summary LM

Approach based on “probability” of relevance (like BIRM) but RSV is
based on P(q|d) (ranking can be done in terms of P(d|q)) and not
P(d|q, r )
Based on the probability that a term occurs in a sequence of terms.
BIRM is based on the probability that term does or does not occur in
a set of (retrieved) documents



Binomial Distribution

N
P(n) = · p n · (1 − p)N−n
n

Imagine you go on a sailing trip on the East Coast of England. Every
second day, there is a beautiful sunset, i.e. p = 1/2. You go sailing for a
week (N = 7). What is your chance to have exactly three (n = 3)
beautiful sunset?
7
· p 3 · (1 − p)7−3 = 0.2734
3



Divergence from Randomness (DFR)

See http://terrier.org/docs/v2.2.1/dfr description.html
Basic idea: ”The more the divergence of the within-document term
frequency from its frequency within the collection, the more divergent
from randomness the term is, meaning the more the information
carried by the term in the document.”

weight(t|d) ∝ − log PM (t ∈ d|collection)

M stands for the type of model of the divergence from randomness
employed to compute the probability.
In the next slide, the binomial distribution (B) is used as the model of
the divergence from randomness.



Binomial Distribution as Randomness Model

TF Term frequency of term t (occurrence of t) in the collection
tf Term frequency of term t in the document d
p Probability to draw a document (p = 1/N, N is number of documents)

TF
− log PB (t ∈ d|collection) = − log · p tf · (1 − p)TF −tf
tf

The probability that
the event (that occurs with probability p) occurs tf times in TF trials
a document occurs tf times in TF trials
a sunny day (which occurs with 1/N) occurs on tf days in a TF days holiday



Binomial Distribution as Randomness Model

TF
− log PB (t ∈ d|collection) = − log · p tf · (1 − p)TF −tf
tf

If N is the number of documents, then TF /N = λ is the average
occurrence of term t.
The above is minimum for tf = λ, meaning that term t has a random
distribution – its distribution does not diverge from randomness – and as
such is not informative.


Retrieval Models Retrieval Models III: Query Reformulation and Relevance feedback

Outline

Query reformulation
Relevance feedback
Local analysis (also called pseudo-relevance feedback)
Global analysis



Query reformulation

No detailed knowledge of collection and retrieval environment
Diﬃcult to formulate queries well designed for retrieval
Need many formulations of queries for good retrieval
First formulation is usually a naive attempt to retrieve relevant
information
The idea is when documents are initially retrieved:
They should be examined for relevance information
Then we can improve the query for retrieving additional relevant
documents
Query reformulation:
Expanding original query with new terms
Reweighing the terms in the (expanded) query



Three approaches

Relevance feedback
Approaches based on feedback from users
E.g. Rocchio, Ide, Binary Independence Model (covered in Retrieval
Models I)
Local analysis (also called pseudo-relevance feedback)
Approaches based on information derived from the set of initially
retrieved documents (local set of documents)
Global analysis
Approaches based on global information derived from the document
collection



Relevance feedback

Relevance feedback can be viewed as an iterative cycle:
1 User presented with a list of retrieved documents
2 User marks those which are relevant (or not relevant)
In practice: top 10-20 ranked documents are examined
Incremental: one document after the other
3 The relevance feedback algorithm selects important terms from
documents assessed relevant by users
4 The relevance feedback algorithm emphasises the importance of these
terms in a new query in the following ways:
Query expansion: add these terms to the query
Term reweighing: modify the term weights in the query
Query expansion + term reweighing
5 The updated query is submitted to the system
6 If the user is satisﬁed with the new set of retrieved documents, then
the relevance feedback process stops, otherwise go to step 2



Relevance feedback - continued

The approaches are (to recap):
Approach 1: Add/Remove/Change query terms
Approach 2: Re-weight query terms
Note: Approach 1 and 2 are not really diﬀerent conceptually. Why?



Relevance Feedback — Vector-Space Model

Reformulation of a query moves the query vector closer to relevant
documents, closer to the optimal query.
C : set of documents (the collection)
R: set of relevant documents in the collection

1 1
qopt = di − di
|R| |C | − |R|
di ∈R di ∈R

Do you know R?



Relevance Feedback — Rochio

You need to use an estimate for R.
Dr : set of relevant and retrieved documents
Dn : set of non-relevant and retrieved documents

1 1
qnext = α · qprev + β · di − γ · di
|Dr | |Dn |
di ∈Dr di ∈Dn

Known as the Rocchio formula
The factors α, β, γ control the eﬀect of previous query, relevant
documents and non-relevant documents on the new query



Relevance Feedback — Rochio

1 1
|Dr | |Dn |
di ∈Dr di ∈Dn

Usually information in relevant documents more important than in
non-relevant documents (γ << β).
Positive relevance feedback (γ = 0) is when we only extract
information from documents assessed relevant.
α emphasises the importance of the original query (qprev ).



Relevance Feedback — Rochio in practice

1 1
qnext = qprev + β · di − γ · di
|Dr | |Dn |
di ∈Dr di ∈Dn

α=1
Terms forming the reformulated query (qprev ) are those:
in the original query,
that appear in more relevant documents that non-relevant documents
that appear in more than half of the relevant documents
Negative weights ignored



Relevance Feedback — Ide

di ∈Dr di ∈Dn

Start with α = β = γ = 1.
The cardinalities of the sets of relevant and non-relevant documents are
not considered.



Issues in relevance feedback

Relevance feedback
Often users are not reliable in making relevance assessments, or do not
make relevance assessments
Implicit relevance feedback by looking at what users access
clicks in web logs (see lecture on web search)
it works well because of the “wisdom of the crowd”
Positive, negative
Partial relevance assessments (e.g very relevant or partially relevant)?
Why is a document relevant?
Interactive query expansion (as opposed to automatic)
Users choose the terms to be added



Local analysis

Examine documents retrieved for query to determine query expansion
with no user assistance
Two strategies are used to add terms to the query:
Local clustering (terms that are synonyms, stemming variations)
Local context analysis (terms close to query terms in text - proximity of
terms in text)
(BUT ALSO session analysis (queries used in same sessions as analyzed
from logs) for query recommendation/suggestion)
Two issues:
Query drift: if top documents are not that relevant, the reformulated
query may not reﬂect the user information need
Computation cost high since must be done at retrieval time (on-line)



Global analysis

Expand query using information from whole set of documents in
collection
No user assistance
Make use of of a global thesaurus that is build based on the
document collection.
Two issues:
Approach to build thesaurus (e.g. term co-occurrence)
Approach to select terms for query expansion (e.g. the top 20 terms
ranked according to IDF value)
(BUT ALSO session analysis (queries used in same sessions as
analyzed from logs) for query recommendation/suggestion)


Retrieval Models Retrieval Models IV: PageRank, inference networks, others

Outline

Other models
Inference network model
Extended boolean model
Fuzzy model
Page rank model
Uncertain inference (logical) model



Inference Network Model

Information Retrieval using inference networks
Greatest strength for IR is in providing a framework permitting
combination of multiple distinct evidence sources to support a
relevance judgement (probability) on a given document.
Bayesian networks
Model causal relationship between events
Infer the belief that an event holds based on observations of other
events
Turtle and Croft Inference Model
Belief model

We only discuss the inference model, which is a simpliﬁed version of a
Bayesian network for IR.
Note that the “combination of evidence is now dealt mostly with machine
learning approaches (features and parameters) = learning to rank.



Bayesian Network – A toy example

& Mounia Lalmas (Yahoo! Research) 20-21 June 2011 147 / 171


Bayesian Network – Links at dependency

Link Matrix
Attached to each node
Give inﬂuences of parents on that node.
Nodes with no parent get a prior probability: e.g., f, d.
Interior node : conditional probability of all combinations of values of
its parents: e.g., n, g, t.
Make use of independence assumptions to make computation
tractable
Nodes not connected by a link: no direct conditioning.
Make use of link matrices (discussed later).



Inference Network

Bayesian Network used to model documents, the document contents,
and the query.
Two sub-networks:
the Document Network (DN) produced during indexing and then static
during retrieval;
the Query Network (QN) produced from the query text during retrieval.


' $

Inference Network
Inference Network

& %


Document network

document nodes:
nodes for each document
retrievable units, i.e. those items we wish to see in the resultant ranking
document concept nodes:
nodes for each concept with the collection
causal link (→) between document node and document concept node:
document content is represented by concept
a conditional probability, or weight, to indicate strength of link
evaluation of a node on the basis of value of the parent nodes and the
conditional probabilities.



Document network - Probabilities

document nodes:
prior document probability
1
usually set to P(d) = n for n the number of documents.
document concept nodes:
usually terms, so ri = ti , or can make use of a thesaurus (for related
terms)
conditional probability P(ti |d) based on within document frequency,
tf × idf , etc.



Query network

represent the submitted query
query concept nodes: framework of nodes that represent the required
concepts
query operator nodes, e.g. #and, #not, etc which permits statistical
operators and statistical approximations of the Boolean operators
user information need node
#and AND the terms
#or OR the terms
#not Negate the term (incoming belief)
#sum Sum of the incoming beliefs
#wsum Weighted sum of the incoming beliefs
#max Maximum of the incoming beliefs



Example



Link Matrix

Link matrix L contains probabilities of the form

¯ ¯ ¯ ¯
P(I |t1 , t2 ) P(I |t1 , t2 ) P(I |t1 , t2 ) P(I |t1 , t2 )
¯|t1 , t2 ) P(¯|t1 , t2 ) P(¯|t1 , t2 ) P(¯|t1 , t2 )
P(I I ¯ I ¯ I ¯ ¯

A special setting of L with wi , i = 1, 2 be query term weights, and
ws = w1 + w2
w1 +w2 w1 w2
ws ws ws 0
w2 w1 w1 +w2
0 ws ws ws )



Probabilistic Inference Network (PIN) Model

Special setting of the link matrix using weights ws = w1 + w2 for two
query terms leads to:
(Here q, is what is called I in this model, as there can be several
queries, or variants, or sub-parts)

1
R(d, q) = P(I |d) = · (w1 · P(t1 |d) + w2 · P(t2 |d))
ws
Complexity linear with number of query terms.
System: INQUERY



Page rank

PR page rank reﬂecting its “authority”, which can be used to rank page
u, v two pages
U set of pages
N(v ) number of outgoing links in page v
PR(v )
PR(u) := d + (1 − d) ·
v →u
N(v )

the summation is over the set of pages v that have a link to page u
damping factor reﬂection the probability that a random (web) surfer
arrives at a page
page rank is used in Google

(See lecture on Web Search)



Fuzzy set model (“vagueness”)

Query term = fuzzy set
Document = has a membership (between 0 and 1) to that set
Fuzzy set theory:
set A (in universe U) whose boundaries are not well deﬁned
(e.g., tall, nice, relevant)
membership function µA : U → [0, 1]



Fuzzy Model

Set-membership function: µA : U → [0; 1] where A is a term and U is
a set of documents.
- sailing = { (0.9, d1), (0.8, d2) }
- boats = { (0.5, d1), (0.8, d2) }
Membership based on tf, idf, or correlation matrix C :
n
- Ci,j := ni +nji,j i,j
−n
- ni,j : Number of documents in which ti and tj occur
- ni : Number of documents in which ti occurs
- µi (d) := 1 − tj ∈d (1 − Ci,j )
The weight of a document in the set of term ti is computed as the
disjunction of all document terms related to term ti .



Fuzzy set model

µi (d) := 1 − (1 − Ci,j )
tj ∈d

Document d belongs to fuzzy set (term) ti if its own terms (the tj s)
are related to ti :
one term tj in document d very related to ti (Cij ∼ 1):
membership of d to fuzzy set ti close to 1 (µi (d) ∼ 1)
no term tj in document d related to ti (all Cij ∼ 0):
membership of d to fuzzy set ti close to 0 (µi (d) ∼ 0)



Fuzzy Model

Retrieval is based on fuzzy operations:
if query is q=“A OR B”: R(d, q) = max(µA (d), µB (d))
if query is q=“A AND B”: R(d, q) = min(µA (d), µB (d))
if query is q=“A AND NOT B”:
R(d, q) = max(µA (d), µB (d)) − min(µA (d), µB (d))



Extended Boolean model

Deal with problems:
t1 AND t2 (too few documents retrieved) → aﬀect recall
t1 OR t2 (too many documents retrieved) → aﬀect precision
Use weights d = (w1 , w2 ) for terms t1 and t2
OR-queries:
2 2
w1 + w2
R(d, q) =
2
AND-queries:

(1 − w1 )2 + (1 − w2 )2
R(d, q) = 1 −
2



Extended Boolean model (Examples)

t1 and t2 in document (w1 = 1 and w2 = 1):
OR-query = 1
AND-query = 1
t1 and t2 not in document (w1 = 0 and w2 = 0):
OR-query = 0
AND-query = 0
t1 in document and√2 not in document (w1 = 1 and w2 = 0):
t
OR-query =1/ 2 = 0.707 < 1
√
AND-query = 1 − 1/ 2 = 0.29 > 0



Uncertain Inference (Logical) Model

Model the notion of relevance as inference.
Document d is relevant to query q ⇐⇒ d → q
The inference will be uncertain, because the IR process is uncertain
(based on estimates)



IR as an inference

document and query: logical formulae d and q
retrieval: search for document which implies the query: d → q

logical view
d = {t1 , t2 , t3 } d = t1 ∧ t2 ∧ t3
q = {t1 , t3 } q = t1 ∧ t3
d →q

Advantage: from term-based retrieval to knowledge-based retrieval
d = square, q = rectangle, and thesaurus: square → rectangle:
document d relevant to query q



IR as an uncertain inference

document and query: logical formulae d and q
d = quadrangle, q = rectangle
document d maybe relevant to query q:
uncertain knowledge is required:
quadrangle, → rectangle with uncertainty 0.3
Retrieval: Estimating the probability that document infer the query:

R(d, q) = P(d → q)



Uncertain Inference Model

What is the difference between P(d → q) and P(q → d)?
View d and q as conjunctions of the terms.
P(d → q) = 1, if d contains all query terms. q is a subset of d; d is
exhaustive.
P(q → d) = 1, if d contains only query terms. d is a subset of q. d is
specific.
Lots of work 1990-2000 (Logic-based IR, FERMI EU project)
Direct application to XML retrieval (exhaustivity and specificity-based
definition relevance at INEX)


Conclusions Recap

What is Information Retrieval
TF-IDF
Precision and Recall
Retrieval Models
Classical Models (Bolean, TF-IDF and Length Normalization, Vector
Space, Probabilistic Model (BIRM))
Alternative Set Theoretic Models (Set-based, Extended Boolean,
Fuzzy Set)
Alternative Algebraic Models (Generalized Vector Space, Latent
Semantic Indexing (LSI), Neural Network)
Alternative Probabilistic Models (BM25, Language Models,
Divergence from Randomness (DFR), Bayesian Network)
Other models (XML, Multimedia, Enterprise, Web, etc)


Acknowledgements

These lectures are based on lecture and presentation slides of the following
persons, who are greatly acknowledged:

Thomas Roelleke, Queen Mary University of London
Arjen de Vries, CWI Amsterdam and University of Delf
Iadh Ounis, University of Glasgow
Gianni Amati, Fondazione Ugo Bordoni, Rome


Acknowledgements Textbook

Introduction (Chapter 1)
IR models (Chapter 3)
Query reformulation and relevance feedback (Chapter 5)
Evaluation - brief (Chapter 4)
Indexing - brief (Chapter 6)
TF-IDF (Chapter 3)


References

Aizawa, A. (2003).
An information-theoretic perspective of tf-idf measures.
Information Processing and Management, 39:45–65.

Amati, G. and Rijsbergen, C. J. (2002).
Term frequency normalization via Pareto distributions.
In Crestani, F., Girolami, M., and Rijsbergen, C. J., editors, 24th BCS-IRSG European Colloquium on IR Research,
Glasgow, Scotland.

Amati, G. and van Rijsbergen, C. J. (2002).
Probabilistic models of information retrieval based on measuring the divergence from randomness.
ACM Transaction on Information Systems (TOIS), 20(4):357–389.

Baeza-Yates, R. and Ribeiro-Neto, B. (2011).
Modern Information Retrieval - the concepts and technology behind search.
Addison Wesley.

Belew, R. K. (2000).
Finding out about.
Cambridge University Press.

Bookstein, A. (1980).
Fuzzy requests: An approach to weighted Boolean searches.
Journal of the American Society for Information Science, 31:240–247.

Brin, S. and Page, L. (1998).
The anatomy of a large-scale hypertextual web search engine.
Computer Networks, 30(1-7):107–117.

Church, K. and Gale, W. (1995).
Inverse document frequency (idf): A measure of deviation from Poisson.
In Proceedings of the Third Workshop on Very Large Corpora, pages 121–130.

Cooper, W. (1991).


References

Some inconsistencies and misnomers in probabilistic IR.
In Bookstein, A., Chiaramella, Y., Salton, G., and Raghavan, V., editors, Proceedings of the Fourteenth Annual
International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 57–61, New York.

Cooper, W. S. (1988).
Getting beyond Boole.
Information Processing and Management, 24(3):243–248.

Cooper, W. S. (1994).
Triennial acm sigir award presentation and paper: The formalism of probability theory in ir: A foundation for an
encumbrance.
In [Croft and van Rijsbergen, 1994], pages 242–248.

Craswell, N., Hawking, D., and Robertson, S. E. (2001).
Effective site finding using link anchor information.
In SIGIR, pages 250–257.

Crestani, F. and van Rijsbergen, C. J. (1995).
Probability kinematics in information retrieval.
In Fox, E., Ingwersen, P., and Fidel, R., editors, Proceedings of the 18th Annual International ACM SIGIR Conference on
Research and Development in Information Retrieval, pages 291–299, New York. ACM.

Croft, B. and Lafferty, J., editors (2003).
Language Modeling for Information Retrieval.
Kluwer.
Croft, W. and Harper, D. (1979).
Using probabilistic models of document retrieval without relevance information.
Journal of Documentation, 35:285–295.

Croft, W. and Turtle, H. (1992).
Retrieval of complex objects.
In Pirotte, A., Delobel, C., and Gottlob, G., editors, Advances in Database Technology — EDBT’92, pages 217–229,
Berlin et al. Springer.


References

Croft, W. B. and van Rijsbergen, C. J., editors (1994).
Proceedings of the Seventeenth Annual International ACM SIGIR Conference on Research and Development in
Information Retrieval, London, et al. Springer-Verlag.

Deerwester, S., Dumais, S., Furnas, G., Landauer, T., and Harshman, R. (1990).
Indexing by latent semantic analysis.
Journal of the American Society for Information Science, 41(6):391–407.

Dumais, S. T., Furnas, G. W.and Landauer, T. K., and Deerwester, S. (1988).
Using latent semantic analysis to improve information retrieval.
pages 281–285.

Fang, H. and Zhai, C. (2005).
An exploration of axiomatic approaches to information retrieval.
In SIGIR ’05: Proceedings of the 28th annual international ACM SIGIR conference on Research and development in
information retrieval, pages 480–487, New York, NY, USA. ACM.

Fuhr, N. (1992).
Probabilistic models in information retrieval.
The Computer Journal, 35(3):243–255.

Grossman, D. A. and Frieder, O. (1998).
Information Retrieval: Algorithms and Heuristics.
Kluwer, Massachusetts.

Grossman, D. A. and Frieder, O. (2004).
Information Retrieval. Algorithms and Heuristics, 2nd ed., volume 15 of The Information Retrieval Series.
Springer.

He, B. and Ounis, I. (2005).
Term frequency normalisation tuning for BM25 and DFR models.
In ECIR, pages 200–214.

Hiemstra, D. (2000).


References

A probabilistic justification for using tf.idf term weighting in information retrieval.
International Journal on Digital Libraries, 3(2):131–139.

Joachims, T. (2000).
Estimating the generalization performance of an svm efficiently.
In [Langley, 2000], pages 431–438.

Kleinberg, J. (1999).
Authoritative sources in a hyperlinked environment.
Journal of ACM, 46.

Klinkenberg, R. and Joachims, T. (2000).
Detecting concept drift with support vector machines.
In [Langley, 2000], pages 487–494.

Lafferty, J. and Zhai, C. (2003).
Probabilistic Relevance Models Based on Document and Query Generation, chapter 1.
In [Croft and Lafferty, 2003].

Langley, P., editor (2000).
Proceedings of the Seventeenth International Conference on Machine Learning (ICML 2000), Stanford University,
Standord, CA, USA, June 29 - July 2, 2000. Morgan Kaufmann.

Lavrenko, V. and Croft, W. B. (2001).
Relevance-based language models.
In SIGIR, pages 120–127.

Luk, R. W. P. (2008).
On event space and rank equivalence between probabilistic retrieval models.
Inf. Retr., 11(6):539–561.

Manning, C. D., Raghavan, P., and Schuetze, H., editors (2008).
Introduction to Information Retrieval.
Cambridge University Press.


References

Margulis, E. (1992).
N-poisson document modelling.
In Belkin, N., Ingwersen, P., and Pejtersen, M., editors, Proceedings of the Fifteenth Annual International ACM SIGIR
Conference on Research and Development in Information Retrieval, pages 177–189, New York.

Maron, M. and Kuhns, J. (1960).
On relevance, probabilistic indexing, and information retrieval.
Journal of the ACM, 7:216–244.

Meghini, C., Sebastiani, F., Straccia, U., and Thanos, C. (1993).
A model of information retrieval based on a terminological logic.
In Korfhage, R., Rasmussen, E., and Willett, P., editors, Proceedings of the Sixteenth Annual International ACM SIGIR
Conference on Research and Development in Information Retrieval, pages 298–308, New York. ACM.

Metzler, D. and Croft, W. B. (2004).
Combining the language model and inference network approaches to retrieval.
Information Processing & Management, 40(5):735–750.

Nie, J. (1992).
Towards a probabilistic modal logic for semantic-based information retrieval.
In Belkin, N., Ingwersen, P., and Pejtersen, M., editors, Proceedings of the Fifteenth Annual International ACM SIGIR
Conference on Research and Development in Information Retrieval, pages 140–151, New York.

Ponte, J. and Croft, W. (1998).
A language modeling approach to information retrieval.
In Croft, W. B., Moﬀat, A., van Rijsbergen, C. J., Wilkinson, R., and Zobel, J., editors, Proceedings of the 21st Annual
International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 275–281, New York.
ACM.
Robertson, S. (2004).
Understanding inverse document frequency: On theoretical arguments for idf.
Journal of Documentation, 60:503–520.

Robertson, S. (2005).
On event spaces and probabilistic models in information retrieval.

References

Information Retrieval Journal, 8(2):319–329.

Robertson, S. and Sparck Jones, K. (1976).
Relevance weighting of search terms.
Journal of the American Society for Information Science, 27:129–146.

Robertson, S. E. and Walker, S. (1994).
Some simple eﬀective approximations to the 2-Poisson model for probabilistic weighted retrieval.
In [Croft and van Rijsbergen, 1994], pages 232–241.

Robertson, S. E., Walker, S., and Hancock-Beaulieu, M. (1995).
Large test collection experiments on an operational interactive system: Okapi at TREC.
Information Processing and Management, 31:345–360.

Rocchio, J. (1966).
Document retrieval systems - optimization and evaluation.
Report ISR-10 to the NSF, Computation Laboratory, Harvard University.

Rocchio, J. (1971).
Relevance feedback in information retrieval.
In [Salton, 1971].

Roelleke, T., Tsikrika, T., and Kazai, G. (2006).
A general matrix framework for modelling information retrieval.
Journal on Information Processing & Management (IP&M), Special Issue on Theory in Information Retrieval, 42(1).

Roelleke, T. and Wang, J. (2006).
A parallel derivation of probabilistic information retrieval models.
In ACM SIGIR, pages 107–114, Seattle, USA.

Roelleke, T. and Wang, J. (2008).
TF-IDF uncovered: A study of theories and probabilities.
In ACM SIGIR, pages 435–442, Singapore.

Roelleke, T., Wu, H., Wang, J., and Azzam, H. (2008).

References

Modelling retrieval models in a probabilistic relational algebra with a new operator: The relational Bayes.
VLDB Journal, 17(1):5–37.

Salton, G., editor (1971).
The SMART Retrieval System - Experiments in Automatic Document Processing.
Prentice Hall, Englewood, Cliﬀs, New Jersey.

Salton, G., Fox, E., and Wu, H. (1983).
Extended Boolean information retrieval.
Communications of the ACM, 26:1022–1036.

Salton, G., Wong, A., and Yang, C. (1975).
A vector space model for automatic indexing.
Communications of the ACM, 18:613–620.

Sebastiani, F. (2002).
Machine learning in automated text categorization.
ACM Comput. Surv., 34(1):1–47.

Turtle, H. and Croft, W. (1991a).
Eﬃcient probabilistic inference for text retrieval.
In Proceedings RIAO 91, pages 644–661, Paris, France.

Turtle, H. and Croft, W. (1991b).
Evaluation of an inference network-based retrieval model.
ACM Transactions on Information Systems, 9(3):187–222.

Turtle, H. and Croft, W. (1992).
A comparison of text retrieval models.
The Computer Journal, 35.

Turtle, H. and Croft, W. B. (1990).
Inference networks for document retrieval.


References

In Vidick, J.-L., editor, Proceedings of the 13th International Conference on Research and Development in Information
Retrieval, pages 1–24, New York. ACM.

van Rijsbergen, C. J. (1979).
Information Retrieval.
Butterworths, London, 2. edition.
http://www.dcs.glasgow.ac.uk/Keith/Preface.html.

A non-classical logic for information retrieval.
The Computer Journal, 29(6):481–485.

Towards an information logic.
In Belkin, N. and van Rijsbergen, C. J., editors, Proceedings of the Twelfth Annual International ACM SIGIR Conference
on Research and Development in Information Retrieval, pages 77–86, New York.

Wong, S. and Yao, Y. (1995).
On modeling information retrieval with probabilistic inference.
ACM Transactions on Information Systems, 13(1):38–68.

Zaragoza, H., Hiemstra, D., and Tipping, M. (2003).
Bayesian extension to the language model for ad hoc information retrieval.
In SIGIR ’03: Proceedings of the 26th annual international ACM SIGIR conference on research and development in
information retrieval, pages 4–9, New York, NY, USA. ACM Press.


Introduction to Information Retrieval & Models

Recommended

More Related Content

What's hot (20)

Similar to Introduction to Information Retrieval & Models (15)

More from Mounia Lalmas-Roelleke (20)

Recently uploaded (20)

Introduction to Information Retrieval & Models