Introduction to Information Retrieval

Introduction to
Information Retrieval
Block B
Lecture 8: Query expansion
and Probabilistic Retrieval
Introduction to Information Retrieval

 Covered material: chapter 9 upto (and including)

9.1.1

6
Introduction to Information Retrieval

Part 1 of this lecture

 Improving results
 For high recall. E.g.,
 searching for aircraft doesn’t match with plane;
 nor thermodynamic with heat

 Options for improving results…

 Global methods
 Query expansion
 Thesauri
 Automatic thesaurus generation
 Local methods
 Relevance feedback
 Pseudo relevance feedback
Introduction to Information Retrieval Sec. 9.1

Relevance Feedback
 Relevance feedback:
 The user issues a (short, simple) query
 The system offers results
 The user marks some results as relevant or non-relevant.
 The system computes a better representation of the
information need based on feedback.

 Relevance feedback can go through one or more

iterations.
Introduction to Information Retrieval

Levels of awareness
Idea: it may be difficult to formulate a good query
when you don’t know the collection well, so iterate

 visceral need (unconscious awareness)

 searcher can recognize some characteristics
 conscious need:
 searcher can judge relevance
 formalized need:
 searcher has implicit or explicit formulation of need;
 if implicit: can judge relevancy of description
 compromised need:
 searcher can compare different solutions.
Introduction to Information Retrieval Sec. 9.1

Relevance feedback
 We will use ad hoc retrieval to refer to regular
retrieval without relevance feedback.

 We now look at examples of relevance feedback that

highlight different aspects.
Introduction to Information Retrieval

Similar pages

snippet

Remember:  Sim( D1 , D1 )  Sim( D1 , Dn ) 

 
AA 
T
   
 Sim( D , D )  Sim( D , D ) 
 n 1 n n 
Introduction to Information Retrieval


description document  A1,1 ,..., A1,m  Column
vectors
 A1,1  A1,m  are the
T
 A1,1 
    standard

A          D1T
A 
A  A   1,m 
 n ,1 n ,m 

description document  An ,1 ,..., An ,m 
T
 An ,1 
 
 D1T       DnT
  A 
   Note that
T  n ,m 
 DT  D  T
 
 
1

 n A      D1 , D2 ,..., Dn
T

 DT 
 n
15
Introduction to Information Retrieval

Finding the most similar documents.

T
D  D  T T
    1 1

A A      
T

 DT   DT 
 n  n
 D1T   D1T D1  D1 Dn 
T
   

    D1 , D2 ,..., Dn       
 DT   T 
 Dn Dn 
T
 n  Dn D1
 Sim( D1 , D1 )  Sim( D1 , Dn ) 
 
    
 Sim( D , D )  Sim( D , D ) 
 n 1 n n 
16
Introduction to Information Retrieval Sec. 9.1.1

Results for Initial Query

Introduction to Information Retrieval Sec. 9.1.1

Results after Relevance Feedback

Introduction to Information Retrieval Sec. 9.1.1

Initial query/results
 Initial query: New space satellite applications

+ 1. 0.539, 08/13/91, NASA Hasn’t Scrapped Imaging Spectrometer

+ 2. 0.533, 07/09/91, NASA Scratches Environment Gear From Satellite Plan
3. 0.528, 04/04/90, Science Panel Backs NASA Satellite Plan, But Urges Launches of Smaller Probes
4. 0.526, 09/09/91, A NASA Satellite Project Accomplishes Incredible Feat: Staying Within Budget
5. 0.525, 07/24/90, Scientist Who Exposed Global Warming Proposes Satellites for Climate Research
6. 0.524, 08/22/90, Report Provides Support for the Critics Of Using Big Satellites to Study Climate
7. 0.516, 04/13/87, Arianespace Receives Satellite Launch Pact From Telesat Canada
+ 8. 0.509, 12/02/87, Telecommunications Tale of Two Companies

 User then marks relevant documents with “+”.

Introduction to Information Retrieval Sec. 9.1.1

Initial query/results
 Initial query: New space satellite applications
 New information from relevant documents:
1. 0.539, 08/13/91, NASA Hasn’t Scrapped Imaging Spectrometer
2. 0.533, 07/09/91, NASA Scratches Environment Gear From Satellite Plan
8. 0.509, 12/02/87, Telecommunications Tale of Two Companies

 But also negative information:

3. 0.528, 04/04/90, Science Panel Backs NASA Satellite Plan, But Urges Launches of Smaller Probes
4. 0.526, 09/09/91, A NASA Satellite Project Accomplishes Incredible Feat: Staying Within Budget
5. 0.525, 07/24/90, Scientist Who Exposed Global Warming Proposes Satellites for Climate Research
6. 0.524, 08/22/90, Report Provides Support for the Critics Of Using Big Satellites to Study Climate
7. 0.516, 04/13/87, Arianespace Receives Satellite Launch Pact From Telesat Canada

Assumptions:
- words in relevant documents are relevancy indicators
- words in nonrelevant documents indicate irrelevancy
Introduction to Information Retrieval Sec. 9.1.1

Expanded query after relevance feedback

 Initial query: New space satellite applications

2.074 new 15.106 space

30.816 satellite 5.660 application
5.991 nasa 5.196 eos
Modified query
4.196 launch 3.972 aster
Expanded query
3.516 instrument 3.446 arianespace
3.004 bundespost 2.806 ss
2.790 rocket 2.053 scientist
2.003 broadcast 1.172 earth
0.836 oil 0.646 measure
Introduction to Information Retrieval Sec. 9.1.1

Results for expanded query

2 1. 0.513, 07/09/91, NASA Scratches Environment Gear From Satellite Plan
1 2. 0.500, 08/13/91, NASA Hasn’t Scrapped Imaging Spectrometer
3. 0.493, 08/07/89, When the Pentagon Launches a Secret Satellite, Space Sleuths Do
Some Spy Work of Their Own
4. 0.493, 07/31/89, NASA Uses ‘Warm’ Superconductors For Fast Circuit
8 5. 0.492, 12/02/87, Telecommunications Tale of Two Companies
6. 0.491, 07/09/91, Soviets May Adapt Parts of SS-20 Missile For Commercial Use
7. 0.490, 07/12/88, Gaping Gap: Pentagon Lags in Race To Match the Soviets In Rocket
Launchers
8. 0.490, 06/14/90, Rescue of Satellite By Space Agency To Cost $90 Million

(Initial query: New space satellite applications)

Introduction to Information Retrieval

HOW DO WE DO IT?

29
Introduction to Information Retrieval Sec. 9.1.1

Key concept: Centroid

 The centroid is the center of mass of a set of points
 Recall that we represent documents as points in a
high-dimensional space

 Definition: Centroid of a set C of documents.

1
 (C )  
| C | dC
d
Introduction to Information Retrieval

Relevancy computation
 Assume query vector is normalized to length 1,

then the relevancy of a document is the total of

supply-demand match:
t
sim(d , q)  d  q   d i  qi
i 1

31 2015/10/19 B1 - Relevance Feedback

Introduction to Information Retrieval

Feedback algorithm
 The algorithm:

Evaluate query q
repeat
Offer k most relevant (yet unseen) documents: T
Ask feedback, splitting T into
set Cr of relevant documents and
set Cnr of nonrelevant documents.
Compute modified query qm
Evaluate modified query qm
until satisfied
32 2015/10/19 B1 - Relevance Feedback
Introduction to Information Retrieval

Optimal query
 Problem:
 given
 set Cr of relevant documents
 set Cnr of irrelevant documents
 find a query q that best generalizes Cr and Cnr
R
Cr Cnr D

 Solution: use bonus-malus strategy

 bonus: similarity with relevant document
 malus: similarity with irrelevant document
Introduction to Information Retrieval

Optimal query
 bonus for q: average similarity with (known) relevant
documents
1 1
bonus (q)  
Cr dCr
Sim(d , q)  
Cr dCr
d  q 

 malus for q: average similarity with (known)

irrelevant documents
malus (q) 
1
 Sim(d , q) 
1
 d  q 
Cnr dCnr Cnr dCnr

 Total score(q) = bonus(q) – malus(q)

34 2015/10/19 B1 - Relevance Feedback
Introduction to Information Retrieval

Optimization problem
 Find query q with length 1, maximizing bonus - malus
score:
score(q) 
1
 d  q   1
 d  q 
Cr dC Cnr dC
r nr

 1   1 
   d   q    d   q
 Cr d Cr   Cnr d C nr 

  (Cr )   (Cnr )   q

 cos   (Cr )   (Cnr ), q  

 Optimal for
 (Cr )   (Cnr )
q cos (φ) maximal when φ = 0
35  (Cr )   (Cnr ) 2015/10/19 B1 - Relevance Feedback
Introduction to Information Retrieval Sec. 9.1.1

Rocchio 1971 Algorithm (SMART)

 Used in practice:

qm    q0     (Cr )     (Cnr )

 qm = modified query vector;

 q0 = original query vector;
 α,β,γ: weights (hand-chosen or set empirically)

 New query moves toward relevant documents and away

from irrelevant documents
Introduction to Information Retrieval Sec. 9.1.1

Relevance Feedback in vector spaces

 Relevance feedback can improve recall and

precision

 Relevance feedback is most useful for increasing

recall in situations where recall is important

 Users can be expected to review results and to take time

to iterate
Introduction to Information Retrieval Sec. 9.1.1

Positive vs Negative Feedback

qm    q0     (Cr )     (Cnr )

 Positive feedback is more valuable than negative

feedback (so, set  < ; e.g.  = 0.25,  = 0.75).

 Many systems only allow positive feedback (=0).

Introduction to Information Retrieval Sec. 9.1.3

Relevance Feedback: Assumptions

 A1: User has sufficient knowledge for initial query.

 A2: Relevance prototypes are “well-behaved”.

 Term distribution in relevant documents will be similar
 Term distribution in non-relevant documents will be
different from those in relevant documents
 Either: All relevant documents are tightly clustered around a
single prototype.
 Or: There are different prototypes, but they have significant
vocabulary overlap.
 Similarities between relevant and irrelevant documents are small
Introduction to Information Retrieval

Relevance Feedback: Problems

 Long queries are inefficient for typical IR engine.
 Long response times for user.
 High cost for retrieval system.
 Partial solution:
 Only reweight certain prominent terms
 Perhaps top 20 by term frequency
 Users are often reluctant to provide explicit
feedback
 It’s often harder to understand why a particular
document was retrieved after applying relevance
feedback
Introduction to Information Retrieval Sec. 9.1.5

Evaluating relevance feedback strategies

 Use q0 and compute precision and recall graph
 Use qm and compute precision recall graph
 Assess on all documents in the collection
 Spectacular improvements, but … it’s cheating!
 Partly due to known relevant documents ranked higher
 Must evaluate with respect to documents not seen by user
 Use documents in residual collection (set of documents minus those
assessed relevant or irrelevant)
 Measures usually then lower than for original query
 But a more realistic evaluation
 Relative performance can be validly compared
 Empirically, one round of relevance feedback is often very useful.
Two rounds is sometimes marginally useful.
Introduction to Information Retrieval

Outline

❶ Probabilistic Approach to Retrieval

❷ Basic Probability Theory

❸ Probability Ranking Principle

❹ Appraisal & Extensions

81
Introduction to Information Retrieval

Bayesian learning theory

• Terms (semantic units) provide an indication for relevance.
– From sample documents get an idea of the kind of terms that occur in
relevant and irrelevant documents.

• Main idea: interchange cause and effect!

– Thus: terms that seem to be (ir)relevant are an indication of
(ir)relevancy.

• Example:
– most students are young
– reversal: a young person is likely to be a student
– being young is an positive indication of being a student
83
Introduction to Information Retrieval

The algorithm
• In the probabilistic approach we use more or less the
same approach as for relevance feedback:
term weights pt and qt get some initial value.

repeat
Evaluate retrieval status value for each document
Offer k most relevant (unseen) documents: T
Ask feedback, splitting T into
set Cr of relevant documents and
set Cnr of nonrelevant documents.
Compute new weights for the terms (pt and qt)
until satisfied
87
Introduction to Information Retrieval

Start point analysis

• Suppose searcher has indicated sets
– R of relevant documents
– S of irrelevant documents

R S D

• Derive term distributions from feedback

probability of observing term t in relevant document (positive evidence)
P t R   p(t )  pt
probability of observing term t in irrelevant document (negative evidence):
P t S   q(t )  qt
88
Introduction to Information Retrieval

Example
a: System and human system engineering
terms: interface, user, system, human, computer,
testing of EPS response, time, EPS, trees, graph, minors, survey
b: A survey of user opinion of computer
system response time
c: The EPS user interface management
Relevant documents:
system
R = {f, g}
d: Human machine interface for ABC
computer applications Nonrelevant documents:
e: Relation of user perceived response S = {a, i}
time to error measurement
f: The generation of random, binary, p(trees) = 1
ordered trees q(system) = 0.5
p(graph) = 0.5 q(human) = 0.5
g: The intersection graph of paths in trees q(EPS) = 0.5
h: Graph minors IV: Widths of trees and q(graph) = 0.5
well-quasi-ordering q(minors) = 0.5
q(survey) = 0.5
i: Graph minors: A survey

89
Introduction to Information Retrieval

The stochastic model

Introduction to Information Retrieval

The stochastic model

• Assume a stochastic variable judgment that can take the
values rel and nrel.
• The searcher judgment considering a document with
representation d:
P (judgment=rel | d) = P (rel | d)

• Positive evidence: P (rel | d)

– the probability that the outcome of the searcher judgment is rel,
when a document with representation d is offered to the searcher.
• Negative evidence: P (nrel | d) = 1 – P (rel | d)
– the probability that the outcome of the searcher judgment is nrel,
when “document d ” is offered to the searcher.
• We will focus on odds: positive against negative evidence.
91
Introduction to Information Retrieval

Odds of being a relevant document

• So the odds of a document like d being judged as relevant are:

– the probability Pr (rel | d) of d being relevant

against
– the probability Pr (nrel | d) of d being nonrelevant

• How do we use the feedback from the searcher to estimate

these probabilities?
d
R S D

92
Introduction to Information Retrieval

Example
terms: interface, user, system, human, computer,
response, time, EPS, trees, graph, minors, survey

Relevant documents: p(trees) = 1 q(system) = 0.5

p(graph) = 0.5 q(human) = 0.5
R = {f, g} q(EPS) = 0.5
Nonrelevant documents: q(graph) = 0.5
q(minors) = 0.5
S = {a, i} q(survey) = 0.5

Consider document h:
Graph minors IV: Widths of trees and well-quasi-ordering

What is the probability of this kind of document

being a relevant document

P (h | rel) = P ({graphs, minors, trees} | rel)

= P (graphs | rel) * P (minors | rel) * P (trees | rel)
* (1 – P (interface | rel)) * (1 – P (user | rel)) * …
93
Introduction to Information Retrieval

Incorporating knowledge
Introduction to Information Retrieval

Incorporating knowledge
• Given (a) document (like) d, what are its odds of being
relevant?
P(rel | d )
Odds(rel | d )  P(A | B) 
P(B | A)P(A)
P(nrel | d ) P(B)

P(d | rel ) P(rel )

P(d )

P(d | nrel ) P(nrel )
P(d ) generality
P(d | rel ) P(rel ) odds
 
P(d | nrel ) P(nrel )
P(d | rel )
~
P(d | nrel )
• This leads us to the probability of documents like d within the
collection of relevant and nonrelevant documents. 95
Introduction to Information Retrieval

Simplification: independence
assumption
Introduction to Information Retrieval

Linked dependence assumption

• The probability of a document corresponds to likeliness of its
specific combination of terms.

• Terms are assumed to be sufficiently independent of each

other: how likely is t in a relevant document

P(d | rel)

t
P(d t | rel ) 1 if t  d
dt  
P(d | nrel) t P(d t | nrel) 0 otherwise

how likely is t in a nonrelevant document

This is called the linked independence assumption

98
Introduction to Information Retrieval

Using our knowledge

Introduction to Information Retrieval

Using our knowledge

• Make the following approximation:

P(d | rel) P(d | R) Model of relevant docs for Q


P(d | nrel) P(d | S ) Model of non-relevant docs for Q

• This can be further elaborated, using the linked independence

assumption:

 td
P(t | R)

 td
(1  P(t | R))
 td
P(t | S )  td
(1  P(t | S ))


 td
pt


td
(1  pt )
 td
qt 
td
(1  qt )
100
Introduction to Information Retrieval

From the high hat

Using a high hat trick leads to a simpler formula:

1 if t  d
 td
pt

 td
(1  pt )
dt
dt  
= 10 ifotherwise
 td
qt  td
(1  qt ) td,
0


t t
p dt


t t
(1  p )1 d t

 t
otherwise
dt 1 d t
pt (1  pt )
q t t
dt
 (1  q )
t t
1 d t
qt dt (1  qt )1 dt

To be simplified
To be simplified even further!

101
Introduction to Information Retrieval

Simplification
• Simplify this expression by taking the logarithm:
 pt (1  pt )1 dt 
dt
log t dt 
1 d t 
 qt (1  qt ) 
 pt 1  pt 
 t  d t log  (1  d t ) log 
 qt 1  qt 
 pt 1  qt  1  pt
 t d t log    t log
 1  pt qt  1  qt
     
document dependent document independent
102
Introduction to Information Retrieval

Retrieval status value

• Isolate document dependent part:
 pt 1  qt 
RSV (d )  t d t log   
 1  pt qt 
 Odds ( pt ) 
 td log  
 Odds (qt ) 
 Odds ( pt ) 
st  log 
 Odds (qt ) 
• Remark: this may be interpreted as the inner vector product
d●s where s is the newly constructed term weight vector!
103
Introduction to Information Retrieval

Retrieval status value

• It will be convenient to introduce:

pRSV (d)  2 RSV (d )

Odds( p t ) 
 t d log 
Odds(q )
t 
2
Odds( pt )

t d Odds(q )
t

 104
Introduction to Information Retrieval

Estimating the distribution

Introduction to Information Retrieval

Robertson-Sparck Jones Model

• Let r: number of documents in R r s
• s: number of documents in S
rt st
• rt: number of documents in R having term t
• st: number of documents in S having term t
rt st
Then:
pt  qt 
r s
• (Robertson & Sparck Jones 76)

rt  0.5 st  0.5
mixture
pt  qt 
r 1 s 1 Stephen Robertson
• Instead of 0.5, alternative adjustments have been proposed

overall rt  nt
nt  rt  nt
st
term pt  N
qt  N
prob r 1 n  r 1 106
Introduction to Information Retrieval

No Relevance Info
(Croft &
• We will assume pi to be a constant (typically 0.5) Harper 79)
• Estimate qi by assuming all documents to be non-relevant

N  rt
pt  constant qt 
rt

Bruce Croft

N  rt  0.5
point-5 formula as extension qt 
rt  1

David Harper
108
Introduction to Information Retrieval

EXERCISES

135
Introduction to Information Retrieval

The exercise
• Consider an excerpt of the cooperators of the Information
Sciences Department some years ago:
– Vaandrager, Jacobs, Mader, Dael, Janssen, Capretta, Sarbo,
Barendregt, Barendsen, Parijs, Basten, Paulussen, Tax, Schering,
Schouten, Achten, Peek, Weert, Fehnker
• Cooperators are characterized by letters in their family name,
using the following letters:
– a, e, n, r, t
• Assume some searcher is interested in those persons having
the letters 'r' and 'e' in their family name. The searcher starts
with the query ‘e'.
• Perform 3 cycles of the feedback algorithm with k=3, using
standard Rocchio (α = β = γ = 1).
Introduction to Information Retrieval

Exercises
2. Comment on the following statement:
Terms neither in positive (R) nor in negative documents (S), are
equally likely to be found in relevant and non-relevant documents

Compare this to the Robertson & Sparck Jones estimation for pt and qt.

3. What is the contribution of terms not in R or S, to the

retrieval status value?

137
Introduction to Information Retrieval

Exercises
4. We do a simulation of probabilistic retrieval
• Our collection consists of cooperators of the Information
Science department (April 21, 1998):
– The characterization of a cooperator is the set of letters in the family
name of this cooperator.
– Our (hidden) information need is the set of cooperators with a
specific letter in their family name.

• The family names of the cooperators:

– Vaandrager, Jacobs, Mader, Dael, Janssen, Capretta, Sarbo, Barendregt, Barendsen, Parijs, Basten,
Paulussen, Tax, Schering, Schouten, Achten, Peek, Weert, Fehnker, Weide, Leijenhorst, Meijer,
Eekelen, Dekkers, Wennink, Verhallen, Serrarens, Messink, Geuvers, Devillers, Wichers Schreur,
Diepen, Lieshout, Pil, Willems-Halders, Giommi, Plasmeijer, Kleijnen, Klop, Smetsers, Snel,
Holwerda, Bommel, Wondergem, Jones, Hoogenhof, Zorner, Bosman, Koster, Low, Arampatzis,
Griffioen, Groningen, Grootjen, Bruijning, Stefanova, Stoelinga, Hubbers, Budde, Huis in 't Veld,
Huisman, Wupper, Kuster, Vytopil 138
Introduction to Information Retrieval

Analysis
• Assume query: m
• Random set: Parijs, Giommi, Serrarens, Jones, Stefanova
– Relevant (R): Giommi
– Nonrelevant (S): Parijs, Serrarens, Jones, Stefanova

• Carry out 2 experiments with different sample size: resp 5 and

10.

139
Introduction to Information Retrieval

END OF PRESENTATION

140