Introduction to Information Retrieval

Relevance Feedback
 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: user feedback on relevance of

docs in initial set of results
 User issues a (short, simple) query
 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.
 Idea: it may be difficult to formulate a good query
when you don’t know the collection well, so iterate
Introduction to Information Retrieval Sec. 9.1.1

Relevance Feedback: Example

 Image search engine

http://nayana.ece.ucsb.edu/imsearch/imsearch.html
Introduction to Information Retrieval Sec. 9.1.1

Results for Initial Query

Introduction to Information Retrieval Sec. 9.1.1

Relevance Feedback
Introduction to Information Retrieval Sec. 9.1.1

Results after Relevance Feedback

Introduction to Information Retrieval

Ad hoc results for query canine

Introduction to Information Retrieval

Ad hoc results for query canine

source: Fernando Diaz
Introduction to Information Retrieval

User feedback: Select what is relevant

source: Fernando Diaz
Introduction to Information Retrieval

Results after relevance feedback

source: Fernando Diaz
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

Expanded query after relevance feedback

 2.074 new 15.106 space
 30.816 satellite 5.660 application
 5.991 nasa 5.196 eos
 4.196 launch 3.972 aster
 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
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
 1 
 (C )  
| C | dC
d

where C is a set of documents.

Introduction to Information Retrieval Sec. 9.1.1

Rocchio Algorithm
 The Rocchio algorithm uses the vector space model
to pick a relevance feedback query
 Rocchio seeks the query qopt that maximizes
    
qopt  arg max [cos( q ,  (Cr ))  cos(q ,  (Cnr ))]

q

 Tries to separate docs marked relevant and non-

relevant  1  1 
qopt 
C d j 
C  d j
r d j Cr nr d j Cr

 Problem: we don’t know the truly relevant docs

Introduction to Information Retrieval Sec. 9.1.1

The Theoretically Best Query

x x
x x
o x x
x x x x
x x
o x
o x x
o x
 o o
x
x

x non-relevant documents
Optimal
query o relevant documents
Introduction to Information Retrieval Sec. 9.1.1

Rocchio 1971 Algorithm (SMART)

 Used in practice:
  1  1 
q m   q0  
Dr  d j  
Dnr  d j
d j Dr d j Dnr

 Dr = set of known relevant doc vectors

 Dnr = set of known irrelevant doc vectors
 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

Subtleties to note
 Tradeoff α vs. β/γ : If we have a lot of judged
documents, we want a higher β/γ.
 Some weights in query vector can go negative
 Negative term weights are ignored (set to 0)
Introduction to Information Retrieval Sec. 9.1.1

Relevance feedback on initial query

Initial x
query x
x
o x
 x x
x x
x x
o x
x o
x o x
o o
x
x x
x
x known non-relevant documents
Revised
query o known relevant documents
Introduction to Information Retrieval Sec. 9.1.1

Relevance Feedback in vector spaces

 We can modify the query based on relevance

feedback and apply standard vector space model.
 Use only the docs that were marked.
 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

 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 Sec. 9.1.3

Violation of A1

 User does not have sufficient initial knowledge.

 Examples:
 Misspellings (Brittany Speers).
 Cross-language information retrieval (hígado).
 Mismatch of searcher’s vocabulary vs. collection
vocabulary
 Cosmonaut/astronaut
Introduction to Information Retrieval Sec. 9.1.3

Violation of A2

 There are several relevance prototypes.

 Examples:
 Burma/Myanmar
 Contradictory government policies
 Pop stars that worked at Burger King
 Often: instances of a general concept
 Good editorial content can address problem
 Report on contradictory government policies
Introduction to Information Retrieval

Relevance Feedback: Problems

 Long queries are inefficient for typical IR engine.
 Long response times for user.
W
 High cost for retrieval system. hy
?
 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

Evaluation of 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)
 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 Sec. 9.1.5

Evaluation of relevance feedback

 Second method – assess only the docs not rated by
the user in the first round
 Could make relevance feedback look worse than it really is
 Can still assess relative performance of algorithms
 Most satisfactory – use two collections each with
their own relevance assessments
 q0 and user feedback from first collection
 qm run on second collection and measured
Introduction to Information Retrieval Sec. 9.1.4

Relevance Feedback on the Web

 Some search engines offer a similar/related pages feature (this is a

trivial form of relevance feedback)
 Google (link-based)
α/β/γ ??
 Altavista
 Stanford WebBase
 But some don’t because it’s hard to explain to average user:
 Alltheweb
 bing
 Yahoo
 Excite initially had true relevance feedback, but abandoned it due
to lack of use.
Introduction to Information Retrieval Sec. 9.1.4

Excite Relevance Feedback

Spink et al. 2000
 Only about 4% of query sessions from a user used
relevance feedback option
 Expressed as “More like this” link next to each result
 But about 70% of users only looked at first page of
results and didn’t pursue things further
 So 4% is about 1/8 of people extending search
 Relevance feedback improved results about 2/3 of
the time
Introduction to Information Retrieval Sec. 9.1.6

Pseudo relevance feedback

 Pseudo-relevance feedback automates the “manual”
part of true relevance feedback.
 Pseudo-relevance algorithm:
 Retrieve a ranked list of hits for the user’s query
 Assume that the top k documents are relevant.
 Do relevance feedback (e.g., Rocchio)
 Works very well on average
 But can go horribly wrong for some queries.
 Several iterations can cause query drift.
Introduction to Information Retrieval Sec. 9.2.2

Query Expansion

 In relevance feedback, users give additional input

(relevant/non-relevant) on documents, which is
used to reweight terms in the documents
 In query expansion, users give additional input
(good/bad search term) on words or phrases
Introduction to Information Retrieval

Query assist

Would you expect such a feature to increase the query

volume at a search engine?
Introduction to Information Retrieval Sec. 9.2.2

How do we augment the user query?

 Manual thesaurus
 E.g. MedLine: physician, syn: doc, doctor, MD, medico
 Can be query rather than just synonyms
 Global Analysis: (static; of all documents in collection)
 Automatically derived thesaurus
 (co-occurrence statistics)
 Refinements based on query log mining
 Common on the web
 Local Analysis: (dynamic)
 Analysis of documents in result set
Introduction to Information Retrieval Sec. 9.2.2

Example of manual thesaurus

Introduction to Information Retrieval Sec. 9.2.2

Thesaurus-based query expansion

 For each term, t, in a query, expand the query with synonyms and
related words of t from the thesaurus
 feline → feline cat
 May weight added terms less than original query terms.
 Generally increases recall
 Widely used in many science/engineering fields
 May significantly decrease precision, particularly with ambiguous
terms.
 “interest rate”  “interest rate fascinate evaluate”
 There is a high cost of manually producing a thesaurus
 And for updating it for scientific changes
Introduction to Information Retrieval Sec. 9.2.3

Automatic Thesaurus Generation

 Attempt to generate a thesaurus automatically by
analyzing the collection of documents
 Fundamental notion: similarity between two words
 Definition 1: Two words are similar if they co-occur with
similar words.
 Definition 2: Two words are similar if they occur in a given
grammatical relation with the same words.
 You can harvest, peel, eat, prepare, etc. apples and pears,
so apples and pears must be similar.
 Co-occurrence based is more robust, grammatical
relations are more accurate.
Introduction to Information Retrieval Sec. 9.2.3

Co-occurrence Thesaurus
 Simplest way to compute one is based on term-term similarities in C = AAT where A is term-document matrix.
 wi,j = (normalized) weight for (ti ,dj)

 For each ti, pick terms with high values in C dj N

ti What does C
contain if A is
a term-doc
incidence
M (0/1) matrix?
Introduction to Information Retrieval Sec. 9.2.3

Automatic Thesaurus Generation

Example
Introduction to Information Retrieval Sec. 9.2.3

Automatic Thesaurus Generation

Discussion
 Quality of associations is usually a problem.
 Term ambiguity may introduce irrelevant statistically
correlated terms.
 “Apple computer”  “Apple red fruit computer”
 Problems:
 False positives: Words deemed similar that are not
 False negatives: Words deemed dissimilar that are similar
 Since terms are highly correlated anyway, expansion
may not retrieve many additional documents.
Introduction to Information Retrieval

Indirect relevance feedback

 On the web, DirectHit introduced a form of indirect
relevance feedback.
 DirectHit ranked documents higher that users look at
more often.
 Clicked on links are assumed likely to be relevant
 Assuming the displayed summaries are good, etc.
 Globally: Not necessarily user or query specific.
 This is the general area of clickstream mining
 Today – handled as part of machine-learned ranking
Introduction to Information Retrieval

Resources
IIR Ch 9
MG Ch. 4.7
MIR Ch. 5.2 – 5.4