Introduction to Information Retrieval

Introduction to
Information Retrieval
CS276
Information Retrieval and Web Search
Christopher Manning and Prabhakar Raghavan
Lecture 8: Evaluation
Introduction to Information Retrieval Sec. 6.2

This lecture
 How do we know if our results are any good?
 Evaluating a search engine
 Benchmarks
 Precision and recall
 Results summaries:
 Making our good results usable to a user

2
Introduction to Information Retrieval

EVALUATING SEARCH ENGINES

Introduction to Information Retrieval Sec. 8.6

Measures for a search engine

 How fast does it index
 Number of documents/hour
 (Average document size)
 How fast does it search
 Latency as a function of index size
 Expressiveness of query language
 Ability to express complex information needs
 Speed on complex queries
 Uncluttered UI
 Is it free?
4
Introduction to Information Retrieval Sec. 8.6

Measures for a search engine

 All of the preceding criteria are measurable: we can
quantify speed/size
 we can make expressiveness precise
 The key measure: user happiness
 What is this?
 Speed of response/size of index are factors
 But blindingly fast, useless answers won’t make a user
happy
 Need a way of quantifying user happiness

5
Introduction to Information Retrieval Sec. 8.6.2

Measuring user happiness

 Issue: who is the user we are trying to make happy?
 Depends on the setting
 Web engine:
 User finds what they want and return to the engine
 Can measure rate of return users
 User completes their task – search as a means, not end
 See Russell http://dmrussell.googlepages.com/JCDL-talk-June-
2007-short.pdf
 eCommerce site: user finds what they want and buy
 Is it the end-user, or the eCommerce site, whose happiness
we measure?
 Measure time to purchase, or fraction of searchers who
become buyers?
6
Introduction to Information Retrieval Sec. 8.6.2

Measuring user happiness

 Enterprise (company/govt/academic): Care about
“user productivity”
 How much time do my users save when looking for
information?
 Many other criteria having to do with breadth of access,
secure access, etc.

7
Introduction to Information Retrieval Sec. 8.1

Happiness: elusive to measure

 Most common proxy: relevance of search results
 But how do you measure relevance?
 We will detail a methodology here, then examine
its issues
 Relevance measurement requires 3 elements:
1. A benchmark document collection
2. A benchmark suite of queries
3. A usually binary assessment of either Relevant or
Nonrelevant for each query and each document
 Some work on more-than-binary, but not the standard
8
Introduction to Information Retrieval Sec. 8.1

Evaluating an IR system
 Note: the information need is translated into a query
 Relevance is assessed relative to the information
need not the query
 E.g., Information need: I'm looking for information on
whether drinking red wine is more effective at
reducing your risk of heart attacks than white wine.
 Query: wine red white heart attack effective
 You evaluate whether the doc addresses the
information need, not whether it has these words

9
Introduction to Information Retrieval Sec. 8.2

Standard relevance benchmarks

 TREC - National Institute of Standards and
Technology (NIST) has run a large IR test bed for
many years
 Reuters and other benchmark doc collections used
 “Retrieval tasks” specified
 sometimes as queries
 Human experts mark, for each query and for each
doc, Relevant or Nonrelevant
 or at least for subset of docs that some system returned
for that query
10
Introduction to Information Retrieval Sec. 8.3

Unranked retrieval evaluation:

Precision and Recall
 Precision: fraction of retrieved docs that are relevant
= P(relevant|retrieved)
 Recall: fraction of relevant docs that are retrieved =
P(retrieved|relevant)

Relevant Nonrelevant
Retrieved tp fp
Not Retrieved fn tn

 Precision P = tp/(tp + fp)

 Recall R = tp/(tp + fn)
11
Introduction to Information Retrieval Sec. 8.3

Should we instead use the accuracy

measure for evaluation?
 Given a query, an engine classifies each doc as
“Relevant” or “Nonrelevant”
 The accuracy of an engine: the fraction of these
classifications that are correct
 (tp + tn) / ( tp + fp + fn + tn)
 Accuracy is a commonly used evaluation measure in
machine learning classification work
 Why is this not a very useful evaluation measure in
IR?

12
Introduction to Information Retrieval Sec. 8.3

Why not just use accuracy?

 How to build a 99.9999% accurate search engine on a
low budget….

Search for:
0 matching results found.

 People doing information retrieval want to find

something and have a certain tolerance for junk.
13
Introduction to Information Retrieval Sec. 8.3

Precision/Recall
 You can get high recall (but low precision) by
retrieving all docs for all queries!
 Recall is a non-decreasing function of the number
of docs retrieved

 In a good system, precision decreases as either the

number of docs retrieved or recall increases
 This is not a theorem, but a result with strong empirical
confirmation

14
Introduction to Information Retrieval Sec. 8.3

Difficulties in using precision/recall

 Should average over large document
collection/query ensembles
 Need human relevance assessments
 People aren’t reliable assessors
 Assessments have to be binary
 Nuanced assessments?
 Heavily skewed by collection/authorship
 Results may not translate from one domain to another

15
Introduction to Information Retrieval Sec. 8.3

A combined measure: F
 Combined measure that assesses precision/recall
tradeoff is F measure (weighted harmonic mean):

(   1) PR
1 2
F 
1
  (1   )
1  PR
2

P R
 People usually use balanced F1 measure
 i.e., with  = 1 or  = ½
 Harmonic mean is a conservative average
 See CJ van Rijsbergen, Information Retrieval
16
Introduction to Information Retrieval Sec. 8.3

F1 and other averages

Combined Measures

100

80 Minimum
Maximum
60
Arithmetic
40 Geometric
Harmonic
20

0
0 20 40 60 80 100
Precision (Recall fixed at 70%)

17
Introduction to Information Retrieval Sec. 8.4

Evaluating ranked results

 Evaluation of ranked results:
 The system can return any number of results
 By taking various numbers of the top returned documents
(levels of recall), the evaluator can produce a precision-
recall curve

18
Introduction to Information Retrieval Sec. 8.4

A precision-recall curve

1.0

0.8
Precision

0.6

0.4

0.2

0.0
0.0 0.2 0.4 0.6 0.8 1.0
Recall
19
Introduction to Information Retrieval Sec. 8.4

Averaging over queries

 A precision-recall graph for one query isn’t a very
sensible thing to look at
 You need to average performance over a whole
bunch of queries.
 But there’s a technical issue:
 Precision-recall calculations place some points on the
graph
 How do you determine a value (interpolate) between the
points?

20
Introduction to Information Retrieval Sec. 8.4

Interpolated precision
 Idea: If locally precision increases with increasing
recall, then you should get to count that…
 So you max of precisions to right of value

21
Introduction to Information Retrieval Sec. 8.4

Evaluation
 Graphs are good, but people want summary measures!
 Precision at fixed retrieval level
 Precision-at-k: Precision of top k results
 Perhaps appropriate for most of web search: all people want are
good matches on the first one or two results pages
 But: averages badly and has an arbitrary parameter of k
 11-point interpolated average precision
 The standard measure in the early TREC competitions: you take
the precision at 11 levels of recall varying from 0 to 1 by tenths of
the documents, using interpolation (the value for 0 is always
interpolated!), and average them
 Evaluates performance at all recall levels

22
Introduction to Information Retrieval Sec. 8.4

Typical (good) 11 point precisions

 SabIR/Cornell 8A1 11pt precision from TREC 8 (1999)
1

0.8

0.6
Precision

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1
Recall
23
Introduction to Information Retrieval Sec. 8.4

Yet more evaluation measures…

 Mean average precision (MAP)
 Average of the precision value obtained for the top k
documents, each time a relevant doc is retrieved
 Avoids interpolation, use of fixed recall levels
 MAP for query collection is arithmetic ave.
 Macro-averaging: each query counts equally
 R-precision
 If have known (though perhaps incomplete) set of relevant
documents of size Rel, then calculate precision of top Rel
docs returned
 Perfect system could score 1.0.
24
Introduction to Information Retrieval Sec. 8.4

Variance
 For a test collection, it is usual that a system does
crummily on some information needs (e.g., MAP =
0.1) and excellently on others (e.g., MAP = 0.7)
 Indeed, it is usually the case that the variance in
performance of the same system across queries is
much greater than the variance of different systems
on the same query.

 That is, there are easy information needs and hard

ones!
25
Introduction to Information Retrieval

CREATING TEST COLLECTIONS

FOR IR EVALUATION
Introduction to Information Retrieval Sec. 8.5

Test Collections

27
Introduction to Information Retrieval Sec. 8.5

From document collections

to test collections
 Still need
 Test queries
 Relevance assessments
 Test queries
 Must be germane to docs available
 Best designed by domain experts
 Random query terms generally not a good idea
 Relevance assessments
 Human judges, time-consuming
 Are human panels perfect?
28
Introduction to Information Retrieval Sec. 8.5

Unit of Evaluation
 We can compute precision, recall, F, and ROC curve
for different units.
 Possible units
 Documents (most common)
 Facts (used in some TREC evaluations)
 Entities (e.g., car companies)
 May produce different results. Why?

29
Introduction to Information Retrieval Sec. 8.5

Kappa measure for inter-judge

(dis)agreement
 Kappa measure
 Agreement measure among judges
 Designed for categorical judgments
 Corrects for chance agreement
 Kappa = [ P(A) – P(E) ] / [ 1 – P(E) ]
 P(A) – proportion of time judges agree
 P(E) – what agreement would be by chance
 Kappa = 0 for chance agreement, 1 for total agreement.

30
Introduction to Information Retrieval Sec. 8.5
P(A)? P(E)?
Kappa Measure: Example
Number of docs Judge 1 Judge 2

300 Relevant Relevant

70 Nonrelevant Nonrelevant

20 Relevant Nonrelevant

10 Nonrelevant Relevant

31
Introduction to Information Retrieval Sec. 8.5

Kappa Example
 P(A) = 370/400 = 0.925
 P(nonrelevant) = (10+20+70+70)/800 = 0.2125
 P(relevant) = (10+20+300+300)/800 = 0.7878
 P(E) = 0.2125^2 + 0.7878^2 = 0.665
 Kappa = (0.925 – 0.665)/(1-0.665) = 0.776

 Kappa > 0.8 = good agreement

 0.67 < Kappa < 0.8 -> “tentative conclusions” (Carletta ’96)
 Depends on purpose of study
 For >2 judges: average pairwise kappas
32
Introduction to Information Retrieval Sec. 8.2

TREC
 TREC Ad Hoc task from first 8 TRECs is standard IR task
 50 detailed information needs a year
 Human evaluation of pooled results returned
 More recently other related things: Web track, HARD
 A TREC query (TREC 5)
<top>
<num> Number: 225
<desc> Description:
What is the main function of the Federal Emergency Management
Agency (FEMA) and the funding level provided to meet emergencies?
Also, what resources are available to FEMA such as people,
equipment, facilities?
</top>
33
Introduction to Information Retrieval Sec. 8.2

Standard relevance benchmarks:

Others
 GOV2
 Another TREC/NIST collection
 25 million web pages
 Largest collection that is easily available
 But still 3 orders of magnitude smaller than what
Google/Yahoo/MSN index
 NTCIR
 East Asian language and cross-language information retrieval
 Cross Language Evaluation Forum (CLEF)
 This evaluation series has concentrated on European languages
and cross-language information retrieval.
 Many others

34
Introduction to Information Retrieval Sec. 8.5

Interjudge Agreement: TREC 3

35
Introduction to Information Retrieval Sec. 8.5

Impact of Inter-judge Agreement

 Impact on absolute performance measure can be significant
(0.32 vs 0.39)
 Little impact on ranking of different systems or relative
performance
 Suppose we want to know if algorithm A is better than
algorithm B
 A standard information retrieval experiment will give us a
reliable answer to this question.

36
Introduction to Information Retrieval Sec. 8.5.1

Critique of pure relevance

 Relevance vs Marginal Relevance
 A document can be redundant even if it is highly relevant
 Duplicates
 The same information from different sources
 Marginal relevance is a better measure of utility for the
user.
 Using facts/entities as evaluation units more directly
measures true relevance.
 But harder to create evaluation set
 See Carbonell reference
37
Introduction to Information Retrieval Sec. 8.6.3

Can we avoid human judgment?

 No
 Makes experimental work hard
 Especially on a large scale
 In some very specific settings, can use proxies
 E.g.: for approximate vector space retrieval, we can
compare the cosine distance closeness of the closest docs
to those found by an approximate retrieval algorithm
 But once we have test collections, we can reuse them
(so long as we don’t overtrain too badly)

38
Introduction to Information Retrieval Sec. 8.6.3

Evaluation at large search engines

 Search engines have test collections of queries and hand-ranked
results
 Recall is difficult to measure on the web
 Search engines often use precision at top k, e.g., k = 10
 . . . or measures that reward you more for getting rank 1 right than
for getting rank 10 right.
 NDCG (Normalized Cumulative Discounted Gain)
 Search engines also use non-relevance-based measures.
 Clickthrough on first result
 Not very reliable if you look at a single clickthrough … but pretty
reliable in the aggregate.
 Studies of user behavior in the lab
 A/B testing
39
Introduction to Information Retrieval Sec. 8.6.3

A/B testing
 Purpose: Test a single innovation
 Prerequisite: You have a large search engine up and running.
 Have most users use old system
 Divert a small proportion of traffic (e.g., 1%) to the new
system that includes the innovation
 Evaluate with an “automatic” measure like clickthrough on
first result
 Now we can directly see if the innovation does improve user
happiness.
 Probably the evaluation methodology that large search
engines trust most
 In principle less powerful than doing a multivariate regression
analysis, but easier to understand

40
Introduction to Information Retrieval Sec. 8.7

RESULTS PRESENTATION

41
Introduction to Information Retrieval Sec. 8.7

Result Summaries
 Having ranked the documents matching a query, we
wish to present a results list
 Most commonly, a list of the document titles plus a
short summary, aka “10 blue links”

42
Introduction to Information Retrieval Sec. 8.7

Summaries
 The title is often automatically extracted from document
metadata. What about the summaries?
 This description is crucial.
 User can identify good/relevant hits based on description.
 Two basic kinds:
 Static
 Dynamic
 A static summary of a document is always the same,
regardless of the query that hit the doc
 A dynamic summary is a query-dependent attempt to explain
why the document was retrieved for the query at hand

43
Introduction to Information Retrieval Sec. 8.7

Static summaries
 In typical systems, the static summary is a subset of
the document
 Simplest heuristic: the first 50 (or so – this can be
varied) words of the document
 Summary cached at indexing time
 More sophisticated: extract from each document a set
of “key” sentences
 Simple NLP heuristics to score each sentence
 Summary is made up of top-scoring sentences.
 Most sophisticated: NLP used to synthesize a summary
 Seldom used in IR; cf. text summarization work

44
Introduction to Information Retrieval Sec. 8.7

Dynamic summaries
 Present one or more “windows” within the document that
contain several of the query terms
 “KWIC” snippets: Keyword in Context presentation

45
Introduction to Information Retrieval Sec. 8.7

Techniques for dynamic summaries

 Find small windows in doc that contain query terms
 Requires fast window lookup in a document cache
 Score each window wrt query
 Use various features such as window width, position in
document, etc.
 Combine features through a scoring function –
methodology to be covered Nov 12th
 Challenges in evaluation: judging summaries
 Easier to do pairwise comparisons rather than binary
relevance assessments

46
Introduction to Information Retrieval

Quicklinks
 For a navigational query such as united airlines
user’s need likely satisfied on www.united.com
 Quicklinks provide navigational cues on that home
page

47
Introduction to Information Retrieval

48
Introduction to Information Retrieval

Alternative results presentations?

 An active area of HCI research
 An alternative: http://www.searchme.com / copies the idea
of Apple’s Cover Flow for search results
 (searchme recently went out of business)

49
Introduction to Information Retrieval

Resources for this lecture

 IIR 8
 MIR Chapter 3
 MG 4.5
 Carbonell and Goldstein 1998. The use of MMR,
diversity-based reranking for reordering documents
and producing summaries. SIGIR 21.

50