1 modern information retrieval chapter 3. evaluation
TRANSCRIPT
11
Modern Information Modern Information RetrievalRetrieval
Chapter 3. EvaluationChapter 3. Evaluation
22
Why System Evaluation?Why System Evaluation?
• There are many retrieval models/ There are many retrieval models/ algorithms/ systems, which one is the best?algorithms/ systems, which one is the best?
• What is the best component for:What is the best component for:– Ranking function (dot-product, cosine, …)Ranking function (dot-product, cosine, …)– Term selection (stopword removal, stemming…)Term selection (stopword removal, stemming…)– Term weighting (TF, TF-IDF,…)Term weighting (TF, TF-IDF,…)
• How far down the ranked list will a user How far down the ranked list will a user need to look to find some/all relevant need to look to find some/all relevant documents?documents?
33
Difficulties in Evaluating IR Difficulties in Evaluating IR SystemsSystems
• Effectiveness is related to the Effectiveness is related to the relevancyrelevancy of of retrieved items.retrieved items.
• Relevancy is not typically binary but continuous.Relevancy is not typically binary but continuous.• Even if relevancy is binary, it can be a difficult Even if relevancy is binary, it can be a difficult
judgment to make.judgment to make.• Relevancy, from a human standpoint, is:Relevancy, from a human standpoint, is:
– Subjective: Depends upon a specific user’s judgment.Subjective: Depends upon a specific user’s judgment.– Situational: Relates to user’s current needs.Situational: Relates to user’s current needs.– Cognitive: Depends on human perception and Cognitive: Depends on human perception and
behavior.behavior.– Dynamic: Changes over time.Dynamic: Changes over time.
44
Human Labeled CorporaHuman Labeled Corpora (Gold Standard) (Gold Standard)• Start with a corpus of documents.Start with a corpus of documents.• Collect a set of queries for this corpus.Collect a set of queries for this corpus.• Have one or more human experts Have one or more human experts
exhaustively label the relevant exhaustively label the relevant documents for each query.documents for each query.
• Typically assumes binary relevance Typically assumes binary relevance judgments.judgments.
• Requires considerable human effort for Requires considerable human effort for large document/query corpora.large document/query corpora.
55
documents relevant of number Total
retrieved documents relevant of Number recall
retrieved documents of number Total
retrieved documents relevant of Number precision
Relevant documents
Retrieved documents
Entire document collection
retrieved & relevant
not retrieved but relevant
retrieved & irrelevant
Not retrieved & irrelevant
retrieved not retrieved
rele
vant
irre
leva
nt
Precision and RecallPrecision and Recall
66
Precision and RecallPrecision and Recall• PrecisionPrecision
– The ability to retrieveThe ability to retrieve top-ranked top-ranked documents that are mostly relevant.documents that are mostly relevant.
• RecallRecall– The ability of the search to find The ability of the search to find allall of of
the relevant items in the corpus.the relevant items in the corpus.
77
Determining Recall is Determining Recall is DifficultDifficult
• Total number of relevant items is Total number of relevant items is sometimes not available:sometimes not available:– Sample across the database and Sample across the database and
perform relevance judgment on these perform relevance judgment on these items.items.
– Apply different retrieval algorithms to Apply different retrieval algorithms to the same database for the same query. the same database for the same query. The aggregate of relevant items is taken The aggregate of relevant items is taken as the total relevant set.as the total relevant set.
88
Trade-off between Recall Trade-off between Recall and Precisionand Precision
10
1
Recall
Pre
cisi
on
The idealReturns relevant documents butmisses many useful ones too
Returns most relevantdocuments but includes lots of junk
99
Computing Recall/Precision Computing Recall/Precision PointsPoints• For a given query, produce the ranked list of For a given query, produce the ranked list of
retrievals.retrievals.
• Adjusting a threshold on this ranked list produces Adjusting a threshold on this ranked list produces different sets of retrieved documents, and different sets of retrieved documents, and therefore different recall/precision measures.therefore different recall/precision measures.
• Mark each document in the ranked list that is Mark each document in the ranked list that is relevant according to the gold standard.relevant according to the gold standard.
• Compute a recall/precision pair for each position Compute a recall/precision pair for each position in the ranked list that contains a relevant in the ranked list that contains a relevant document.document.
1010
R=3/6=0.5; P=3/4=0.75
Computing Recall/Precision Computing Recall/Precision Points: An ExamplePoints: An Example
n doc # relevant
1 588 x2 589 x3 5764 590 x5 9866 592 x7 9848 9889 57810 98511 10312 59113 772 x14 990
Let total # of relevant docs = 6Check each new recall point:
R=1/6=0.167; P=1/1=1
R=2/6=0.333; P=2/2=1
R=5/6=0.833; p=5/13=0.38
R=4/6=0.667; P=4/6=0.667
Missing one relevant document.
Never reach 100% recall
1111
Interpolating a Interpolating a Recall/Precision CurveRecall/Precision Curve
• Interpolate a precision value for each Interpolate a precision value for each standard recall levelstandard recall level::– rrjj {0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, {0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8,
0.9, 1.0}0.9, 1.0}
– rr00 = 0.0, r = 0.0, r11 = 0.1, …, r = 0.1, …, r1010=1.0=1.0
• The interpolated precision at the The interpolated precision at the jj-th -th standard recall level is the maximum standard recall level is the maximum known precision at any recall level known precision at any recall level between the between the jj-th and (-th and (j j + 1)-th level:+ 1)-th level:)(max)(
1
rPrPjj rrr
j
1212
Interpolating a Interpolating a Recall/Precision Curve: An Recall/Precision Curve: An ExampleExample
0.4 0.8
1.0
0.8
0.6
0.4
0.2
0.2 1.00.6 Recall
Prec
isio
n
1313
Average Recall/Precision Average Recall/Precision CurveCurve• Typically average performance over a Typically average performance over a
large large setset of queries. of queries.
• Compute average precision at each Compute average precision at each standard recall level across all queries.standard recall level across all queries.
• Plot average precision/recall curves to Plot average precision/recall curves to evaluate overall system performance evaluate overall system performance on a document/query corpus.on a document/query corpus.
1414
Compare Two or More Compare Two or More SystemsSystems• The curve closest to the upper right-The curve closest to the upper right-
hand corner of the graph indicates hand corner of the graph indicates the best performancethe best performance
0
0.2
0.4
0.6
0.8
1
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Recall
Precision
NoStem Stem
1515
R- PrecisionR- Precision• Precision at the R-th position in the Precision at the R-th position in the
ranking of results for a query that ranking of results for a query that has R relevant documents.has R relevant documents.
n doc # relevant
1 588 x2 589 x3 5764 590 x5 9866 592 x7 9848 9889 57810 98511 10312 59113 772 x14 990
R = # of relevant docs = 6
R-Precision = 4/6 = 0.67
1616
F-MeasureF-Measure
• One measure of performance that takes One measure of performance that takes into account both recall and precision.into account both recall and precision.
• Harmonic mean of recall and precision:Harmonic mean of recall and precision:
• Compared to arithmetic mean, both need Compared to arithmetic mean, both need to be high for harmonic mean to be high.to be high for harmonic mean to be high.
PRRP
PRF 11
22
1717
E Measure (parameterized F E Measure (parameterized F Measure)Measure)
• A variant of F measure that allows A variant of F measure that allows weighting emphasis on precision over weighting emphasis on precision over recall:recall:
• Value of Value of controls trade-off: controls trade-off:– = 1: Equally weight precision and recall (E=F).= 1: Equally weight precision and recall (E=F).– > 1: Weight recall more.> 1: Weight recall more.– < 1: Weight precision more.< 1: Weight precision more.
PRRP
PRE
1
2
2
2
2
)1()1(
1818
Fallout RateFallout Rate
• Problems with both precision and Problems with both precision and recall:recall:– Number of irrelevant documents in the Number of irrelevant documents in the
collection is not taken into account.collection is not taken into account.– Recall is undefined when there is no Recall is undefined when there is no
relevant document in the collection.relevant document in the collection.– Precision is undefined when no Precision is undefined when no
document is retrieved.document is retrieved.
collection the in items tnonrelevan of no. totalretrieved items tnonrelevan of no.
Fallout
1919
Subjective Relevance Subjective Relevance MeasureMeasure• Novelty RatioNovelty Ratio: The proportion of items retrieved : The proportion of items retrieved
and judged relevant by the user and of which and judged relevant by the user and of which they were previously unaware.they were previously unaware.– Ability to findAbility to find new new information on a topic. information on a topic.
• Coverage RatioCoverage Ratio: The proportion of relevant : The proportion of relevant items retrieved out of the total relevant items retrieved out of the total relevant documents documents knownknown to a user prior to the search. to a user prior to the search.– Relevant when the user wants to locate documents Relevant when the user wants to locate documents
which they have seen before (e.g., the budget report which they have seen before (e.g., the budget report for Year 2000).for Year 2000).
2020
Other Factors to ConsiderOther Factors to Consider• User effortUser effort: Work required from the user : Work required from the user
in formulating queries, conducting the in formulating queries, conducting the search, and screening the output.search, and screening the output.
• Response timeResponse time: Time interval between : Time interval between receipt of a user query and the receipt of a user query and the presentation of system responses.presentation of system responses.
• Form of presentationForm of presentation: Influence of search : Influence of search output format on the user’s ability to output format on the user’s ability to utilize the retrieved materials.utilize the retrieved materials.
2121
Experimental Setup for Experimental Setup for BenchmarkingBenchmarking• AAnalyticalnalytical performance evaluation is difficult performance evaluation is difficult
for document retrieval systems because many for document retrieval systems because many characteristics such as relevance, distribution of characteristics such as relevance, distribution of words, etc., are difficult to describe with words, etc., are difficult to describe with mathematical precision.mathematical precision.
• Performance is measured by Performance is measured by benchmarkingbenchmarking. . That is, the retrieval effectiveness of a system is That is, the retrieval effectiveness of a system is evaluated on a evaluated on a given set of documentsgiven set of documents, , queriesqueries, , and and relevance judgmentsrelevance judgments..
• Performance data is valid only for the Performance data is valid only for the environment under which the system is environment under which the system is evaluated. evaluated.
2222
BenchmarksBenchmarks• A benchmark collection contains:A benchmark collection contains:
– A set of standard documents and queries/topics.A set of standard documents and queries/topics.– A list of relevant documents for each query.A list of relevant documents for each query.
• Standard collections for traditional IR:Standard collections for traditional IR:– Smart collection: ftp://ftp.cs.cornell.edu/pub/smartSmart collection: ftp://ftp.cs.cornell.edu/pub/smart– TREC: http://trec.nist.gov/TREC: http://trec.nist.gov/
Standard document collection
Standard queries
Algorithm under test Evaluation
Standard result
Retrieved result
Precision and recall
2323
Benchmarking Benchmarking The The ProblemsProblems
• Performance data is valid only for a Performance data is valid only for a particular benchmark.particular benchmark.
• Building a benchmark corpus is a Building a benchmark corpus is a difficult task.difficult task.
• Benchmark web corpora are just Benchmark web corpora are just starting to be developed.starting to be developed.
• Benchmark foreign-language corpora Benchmark foreign-language corpora are just starting to be developed.are just starting to be developed.
2424
CollectionCollection Number Of Number Of Number Of Number Of Raw Size Raw Size Name Name Documents Documents Queries Queries (Mbytes) (Mbytes) CACM CACM 3,204 3,204 64 64 1.5 1.5 CISICISI 1,460 1,460 112 112 1.3 1.3 CRANCRAN 1,400 1,400 225 225 1.6 1.6 MED MED 1,033 1,033 30 30 1.1 1.1 TIME TIME 425 425 83 83 1.5 1.5
• Different researchers used different test Different researchers used different test collections and evaluation techniques.collections and evaluation techniques.
Early Test CollectionsEarly Test Collections
2525
The TREC Benchmark The TREC Benchmark • TREC: Text REtrieval Conference (http://trec.nist.gov/) Originated from the TIPSTER program sponsored by Defense Advanced Research Projects Agency (DARPA).
• Became an annual conference in 1992, co-sponsored by the National Institute of Standards and Technology (NIST) and DARPA.
• Participants are given parts of a standard set of documents and TOPICS (from which queries have to be derived) in different stages for training and testing.
• Participants submit the P/R values for the final document and query corpus and present their results at the conference.
2626
The TREC Objectives The TREC Objectives • Provide a common ground for comparing different IR techniques.
– Same set of documents and queries, and same evaluation method.• Sharing of resources and experiences in developing the benchmark.
– With major sponsorship from government to develop large benchmark collections.
• Encourage participation from industry and academia.• Development of new evaluation techniques, particularly for new applications.
– Retrieval, routing/filtering, non-English collection, web-based collection, question answering.
2727
TREC AdvantagesTREC Advantages
• Large scale (compared to a few MB in the SMART Large scale (compared to a few MB in the SMART Collection).Collection).
• Relevance judgments provided.Relevance judgments provided.• Under continuous development with support from the Under continuous development with support from the
U.S. Government.U.S. Government.• Wide participation:Wide participation:
– TREC 1: 28 papers 360 pages.TREC 1: 28 papers 360 pages.– TREC 4: 37 papers 560 pages.TREC 4: 37 papers 560 pages.– TREC 7: 61 papers 600 pages. TREC 7: 61 papers 600 pages. – TREC 8: 74 papers.TREC 8: 74 papers.– TREC 2004 had 103 participants from over 20 countries TREC 2004 had 103 participants from over 20 countries
2828
TREC ImpactsTREC ImpactsCornell University TREC Systems
0
0.1
0.2
0.3
0.4
0.5
'92
System
'93
System
'94
System
'95
System
'96
System
'97
System
'98
System
Mea
n A
vera
ge P
reci
sion
TREC-1
TREC-2
TREC-3
TREC-4
TREC-5
TREC-6
TREC-7
2929
TREC TasksTREC Tasks• Ad hocAd hoc: New questions are being asked on : New questions are being asked on
a static set of data. a static set of data. • Routing: Routing: Same questions are being asked, Same questions are being asked,
but new information is being searched. but new information is being searched. (news clipping, library profiling).(news clipping, library profiling).
• New tasks added after TREC 5 - New tasks added after TREC 5 - Interactive, multilingual, natural Interactive, multilingual, natural language, multiple database merging, language, multiple database merging, filtering, very large corpus (20 GB, 7.5 filtering, very large corpus (20 GB, 7.5 million documents), question answering.million documents), question answering.
303019
92
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
TREC TracksTREC TracksRetrieval in a domain
Answers, not docs
Web searching, size
Beyond text
Beyond just English
Human-in-the-loop
Streamed text
Static text Ad Hoc, Robust
Interactive, HARD
X{X,Y,Z} ChineseSpanish
VideoSpeechOCR
TerabyteWebVLC
NoveltyQ&A
FilteringRouting
Genome
3131
TREC-6 Cross Language TREC-6 Cross Language TaskTask• Task: starting in one language, Task: starting in one language,
search for documents written in a search for documents written in a variety of languagesvariety of languages•Document collectionDocument collection
– SDA: French (257 MB), German (331 MB)SDA: French (257 MB), German (331 MB)– Neue Zürcher Zeitung (198 MB, in German)Neue Zürcher Zeitung (198 MB, in German)– AP (759MB, in English)AP (759MB, in English)
•25 topics:25 topics:– Mostly built in English and translated at NIST into Mostly built in English and translated at NIST into
French and German; also translated elsewhere into French and German; also translated elsewhere into Spanish and DutchSpanish and Dutch
•Relevance judgments made at NIST by two tri-Relevance judgments made at NIST by two tri-lingual surrogate users (who built the topics)lingual surrogate users (who built the topics)
3232
TREC-7 Cross Language TREC-7 Cross Language TaskTask
•Document collection:Document collection:– SDA: French (257 MB), German (331 MB), Italian SDA: French (257 MB), German (331 MB), Italian
(194MB)(194MB)– Neue Zürcher Zeitung (198 MB, in German)Neue Zürcher Zeitung (198 MB, in German)– AP (759MB, in English)AP (759MB, in English)
•28 topics (7 each from 4 sites)28 topics (7 each from 4 sites)
•Relevance judgments made Relevance judgments made independently at each site by native independently at each site by native speakersspeakers
English: NI ST German: University BonnFrench: EPFL Lausanne I talian: CNR Pisa
3333
CLIR Evaluation IssuesCLIR Evaluation Issues
• How to ensure that the “translated” topics How to ensure that the “translated” topics represent how an information request would represent how an information request would be made in a given languagebe made in a given language
• How to ensure that there is enough How to ensure that there is enough common understanding of the topic so that common understanding of the topic so that the relevance judgments are consistentthe relevance judgments are consistent
• How to ensure that the sampling techniques How to ensure that the sampling techniques for the relevance judging are complete for the relevance judging are complete enoughenough
3434
CLEF 2000CLEF 2000
• TREC European CLIR task moved to the TREC European CLIR task moved to the new Cross Language Evaluation Forum new Cross Language Evaluation Forum (CLEF) in 2000; number of participants (CLEF) in 2000; number of participants grew from 12 to 20!!grew from 12 to 20!!
• Document collection stayed the same but Document collection stayed the same but the topics were created in 8 languagesthe topics were created in 8 languages
• The “seriousness” of the experiments The “seriousness” of the experiments also took a big jump also took a big jump
3535
CLEF 2004CLEF 2004
• Document collection expanded to include Document collection expanded to include major newspapers in 10 languages!!major newspapers in 10 languages!!
• This means that the CLEF test collection is This means that the CLEF test collection is now a cooperative project across at least 10 now a cooperative project across at least 10 groups in Europegroups in Europe
• Topics translated into 14 languagesTopics translated into 14 languages
• Fifty-five participating groupsFifty-five participating groups
• Six tasks, including question answering and Six tasks, including question answering and image retrieval image retrieval
363619
92
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
TREC TracksTREC TracksRetrieval in a domain
Answers, not docs
Web searching, size
Beyond text
Beyond just English
Human-in-the-loop
Streamed text
Static text Ad Hoc, Robust
Interactive, HARD
X{X,Y,Z} ChineseSpanish
VideoSpeechOCR
TerabyteWebVLC
NoveltyQ&A
FilteringRouting
Genome
3737
Spoken Document RetrievalSpoken Document Retrieval
• Documents: TREC-7 had 100 hours of Documents: TREC-7 had 100 hours of news broadcasts; 550 hours/21,500 news broadcasts; 550 hours/21,500 stories in TREC-8. stories in TREC-8.
• Topics: similar to “standard” TREC Topics: similar to “standard” TREC topics, 23 in TREC-7 and 50 in TREC-8topics, 23 in TREC-7 and 50 in TREC-8
• ““Documents” were available as several Documents” were available as several baseline recognizer outputs (at different baseline recognizer outputs (at different error rates), along with transcriptserror rates), along with transcripts
3838
Video RetrievalVideo Retrieval
• Video retrieval is not just speech retrieval, Video retrieval is not just speech retrieval, even though that is a major component of even though that is a major component of many current systemsmany current systems
• TREC 2001 had a video retrieval track with TREC 2001 had a video retrieval track with 11 hours of video, 2 tasks (shot boundary 11 hours of video, 2 tasks (shot boundary and search), and 12 participantsand search), and 12 participants
• TRECvid 2004 had 70 hours, 4 tasks TRECvid 2004 had 70 hours, 4 tasks (feature extraction and story segmentation (feature extraction and story segmentation added) and 33 participantsadded) and 33 participants
393919
92
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
TREC TracksTREC TracksRetrieval in a domain
Answers, not docs
Web searching, size
Beyond text
Beyond just English
Human-in-the-loop
Streamed text
Static text Ad Hoc, Robust
Interactive, HARD
X{X,Y,Z} ChineseSpanish
VideoSpeechOCR
TerabyteWebVLC
NoveltyQ&A
FilteringRouting
Genome
4040
TREC-8 Factoid QATREC-8 Factoid QA
• Document collection was “regular” TREC-8 Document collection was “regular” TREC-8
• Topics are now questions (198 of them)Topics are now questions (198 of them)– What many calories in a Big Mac?What many calories in a Big Mac?– Where is the Taj Mahal?Where is the Taj Mahal?
• Task is to retrieve answer strings of 50 or Task is to retrieve answer strings of 50 or 250 bytes, not a document list250 bytes, not a document list
• Humans determine correct answers from Humans determine correct answers from what is submittedwhat is submitted
4141
Moving beyond factoid QAMoving beyond factoid QA
• Use of questions from MSN, AskJeeves logs in Use of questions from MSN, AskJeeves logs in 2000 and 20012000 and 2001
• Addition of questions with no answers in 2001 Addition of questions with no answers in 2001 and reduction to 50-byte answersand reduction to 50-byte answers
• Requirement of exact answers in 2002Requirement of exact answers in 2002• Addition of “definition”/who is questions in 2003Addition of “definition”/who is questions in 2003• Expansion of these questions to include exact Expansion of these questions to include exact
and list slots in 2004and list slots in 2004• Addition of events, pilot of relationship Addition of events, pilot of relationship
questions planned for 2005questions planned for 2005
4242
Cross-language QACross-language QA
• CLEF started Cross-language QA in 2004CLEF started Cross-language QA in 2004
• Lots of interest, lots of unique issuesLots of interest, lots of unique issues– Do cultural aspects make this more difficult Do cultural aspects make this more difficult
than cross-language document retrieval??than cross-language document retrieval??– What language should the answers be in??What language should the answers be in??– Are there specific types of questions that Are there specific types of questions that
should be tested??should be tested??– Would it be interesting to test question Would it be interesting to test question
answering in a specific domain??answering in a specific domain??
4343
For more informationFor more information
trec.nist.govtrec.nist.gov for more on TREC for more on TREC
www.clef-campaign.org www.clef-campaign.org for more on for more on CLEFCLEF
research.nii.ac.jp/ntcirresearch.nii.ac.jp/ntcir for more on for more on NTCIR (evaluation in Japanese, NTCIR (evaluation in Japanese, Chinese, and Korean)Chinese, and Korean)
4444
Characteristics of the TREC Characteristics of the TREC Collection Collection
• Both long and short documents (from a few Both long and short documents (from a few hundred to over one thousand unique terms in hundred to over one thousand unique terms in a document).a document).
• Test documents consist of: Test documents consist of: WSJWSJ Wall Street Journal articles (1986-1992) Wall Street Journal articles (1986-1992) 550 M 550 M AP AP Associate Press Newswire (1989) Associate Press Newswire (1989) 514 M514 M ZIFFZIFF Computer Select Disks (Ziff-Davis Publishing) Computer Select Disks (Ziff-Davis Publishing) 493 M 493 M FR FR Federal Register Federal Register 469 M 469 M DOE DOE Abstracts from Department of Energy reports Abstracts from Department of Energy reports 190 M 190 M
4545
More Details on Document More Details on Document CollectionsCollections• Volume 1 (Mar 1994) - Volume 1 (Mar 1994) - Wall Street JournalWall Street Journal (1987, 1988, 1989), (1987, 1988, 1989),
Federal Register Federal Register (1989), (1989), Associated PressAssociated Press (1989), (1989), Department of Department of Energy abstractsEnergy abstracts, and , and Information from the Computer SelectInformation from the Computer Select disks disks (1989, 1990)(1989, 1990)
• Volume 2 (Mar 1994) - Volume 2 (Mar 1994) - Wall Street JournalWall Street Journal (1990, 1991, 1992), the (1990, 1991, 1992), the Federal Register (1988), Associated Press (1988) and Information Federal Register (1988), Associated Press (1988) and Information from the Computer Select disks (1989, 1990)from the Computer Select disks (1989, 1990)
• Volume 3 (Mar 1994) - Volume 3 (Mar 1994) - San Jose Mercury NewsSan Jose Mercury News (1991), the Associated (1991), the Associated Press (1990), Press (1990), U.S. PatentsU.S. Patents (1983-1991), and Information from the (1983-1991), and Information from the Computer Select disks (1991, 1992)Computer Select disks (1991, 1992)
• Volume 4 (May 1996) - Volume 4 (May 1996) - Financial TimesFinancial Times Limited (1991, 1992, 1993, Limited (1991, 1992, 1993, 1994), the 1994), the Congressional Record of the 103rd CongressCongressional Record of the 103rd Congress (1993), and (1993), and the Federal Register (1994). the Federal Register (1994).
• Volume 5 (Apr 1997) - Volume 5 (Apr 1997) - Foreign Broadcast Information ServiceForeign Broadcast Information Service (1996) (1996) and the and the Los Angeles TimesLos Angeles Times (1989, 1990). (1989, 1990).
4646
TREC Disk 4,5TREC Disk 4,5TREC Disk 4 Congressional Record of the 103rd Congress
approx. 30,000 documents approx. 235 MBFederal Register (1994) approx. 55,000 documents approx. 395 MBFinancial Times (1992-1994) approx. 210,000 documents approx. 565 MB
TREC Disk 5 Data provided from the Foreign Broadcast Information Service approx. 130,000 documents approx. 470 MBLos Angeles Times (randomly selected articles from 1989 & 1990) approx. 130,000 document approx. 475 MB
4747
Sample Document (with Sample Document (with SGML)SGML)<<DOC>DOC> <DOCNO><DOCNO> WSJ870324-0001 WSJ870324-0001 </DOCNO></DOCNO> <HL><HL> John Blair Is Near Accord To Sell Unit, Sources Say John Blair Is Near Accord To Sell Unit, Sources Say </HL></HL> <DD><DD> 03/24/87 03/24/87</DD></DD> <SO><SO> WALL STREET JOURNAL (J) WALL STREET JOURNAL (J) </SO></SO><IN><IN> REL TENDER OFFERS, MERGERS, ACQUISITIONS (TNM) REL TENDER OFFERS, MERGERS, ACQUISITIONS (TNM)
MARKETING, ADVERTISING (MKT) TELECOMMUNICATIONS, MARKETING, ADVERTISING (MKT) TELECOMMUNICATIONS, BROADCASTING, TELEPHONE, TELEGRAPH (TEL) BROADCASTING, TELEPHONE, TELEGRAPH (TEL) </IN></IN>
<DATELINE><DATELINE> NEW YORK NEW YORK </DATELINE></DATELINE> <TEXT><TEXT> John Blair & Co. is close to an agreement to sell its TV John Blair & Co. is close to an agreement to sell its TV
station advertising representation operation and program station advertising representation operation and program production unit to an investor group led by James H. Rosenfield, production unit to an investor group led by James H. Rosenfield, a former CBS Inc. executive, industry sources said. Industry a former CBS Inc. executive, industry sources said. Industry sources put the value of the proposed acquisition at more than sources put the value of the proposed acquisition at more than $100 million. ... $100 million. ...
</TEXT> </TEXT> </DOC> </DOC>
4848
Sample Query (with SGML)Sample Query (with SGML)
<<top> top> <head><head> Tipster Topic Description Tipster Topic Description <num><num> Number: 066 Number: 066 <dom><dom> Domain: Science and Technology Domain: Science and Technology <title><title> Topic: Natural Language Processing Topic: Natural Language Processing <desc><desc> Description: Document will identify a type of natural Description: Document will identify a type of natural
language processing technology which is being developed or language processing technology which is being developed or marketed in the U.S. marketed in the U.S.
<narr><narr> Narrative: A relevant document will identify a company or Narrative: A relevant document will identify a company or institution developing or marketing a natural language institution developing or marketing a natural language processing technology, identify the technology, and identify processing technology, identify the technology, and identify one of more features of the company's product.one of more features of the company's product.
<con><con> Concept(s): 1. natural language processing ;2. translation, Concept(s): 1. natural language processing ;2. translation, language, dictionarylanguage, dictionary
<fac><fac> Factor(s): Factor(s): <nat><nat> Nationality: U.S Nationality: U.S.</nat>.</nat></fac></fac> <def><def> Definitions(s): Definitions(s): </top></top>
4949
TREC PropertiesTREC Properties• Both documents and queries contain Both documents and queries contain
many different kinds of information many different kinds of information (fields).(fields).
• Generation of the formal queries Generation of the formal queries (Boolean, Vector Space, etc.) is the (Boolean, Vector Space, etc.) is the responsibility of the system.responsibility of the system.– A system may be very good at querying A system may be very good at querying
and ranking, but if it generates poor and ranking, but if it generates poor queries from the topic, its final P/R queries from the topic, its final P/R would be poor.would be poor.
5050
Two more TREC Document Two more TREC Document ExamplesExamples
5151
Another Example of TREC Another Example of TREC Topic/QueryTopic/Query
5252
Evaluation Evaluation • Summary table statisticsSummary table statistics: Number of topics, : Number of topics,
number of documents retrieved, number of number of documents retrieved, number of relevant documents.relevant documents.
• Recall-precision averageRecall-precision average: Average precision at : Average precision at 11 recall levels (0 to 1 at 0.1 increments).11 recall levels (0 to 1 at 0.1 increments).
• Document level averageDocument level average: Average precision : Average precision when 5, 10, .., 100, … 1000 documents are when 5, 10, .., 100, … 1000 documents are retrieved.retrieved.
• Average precision histogramAverage precision histogram: Difference of the : Difference of the R-precision for each topic and the average R-R-precision for each topic and the average R-precision of all systems for that topic.precision of all systems for that topic.
5353
5454
Cystic Fibrosis (CF) Cystic Fibrosis (CF) CollectionCollection• 1,239 abstracts of medical journal articles 1,239 abstracts of medical journal articles
on CF.on CF.• 100 information requests (queries) in the 100 information requests (queries) in the
form of complete English questions.form of complete English questions.• Relevant documents determined and Relevant documents determined and
rated by 4 separate medical experts on 0-rated by 4 separate medical experts on 0-2 scale:2 scale:– 0: Not relevant.0: Not relevant.– 1: Marginally relevant.1: Marginally relevant.– 2: Highly relevant.2: Highly relevant.
5555
CF Document FieldsCF Document Fields
• MEDLINE access numberMEDLINE access number
• AuthorAuthor
• TitleTitle
• SourceSource
• Major subjectsMajor subjects
• Minor subjectsMinor subjects
• Abstract (or extract)Abstract (or extract)
• References to other documentsReferences to other documents
• Citations to this documentCitations to this document
5656
Sample CF DocumentSample CF DocumentAN 74154352AU Burnell-R-H. Robertson-E-F.TI Cystic fibrosis in a patient with Kartagener syndrome.SO Am-J-Dis-Child. 1974 May. 127(5). P 746-7.MJ CYSTIC-FIBROSIS: co. KARTAGENER-TRIAD: co.MN CASE-REPORT. CHLORIDES: an. HUMAN. INFANT. LUNG: ra. MALE. SITUS-INVERSUS: co, ra. SODIUM: an. SWEAT: an.AB A patient exhibited the features of both Kartagener syndrome and cystic fibrosis. At most, to the authors' knowledge, this represents the third such report of the combination. Cystic fibrosis should be excluded before a diagnosis of Kartagener syndrome is made.RF 001 KARTAGENER M BEITR KLIN TUBERK 83 489 933 002 SCHWARZ V ARCH DIS CHILD 43 695 968 003 MACE JW CLIN PEDIATR 10 285 971 …CT 1 BOCHKOVA DN GENETIKA (SOVIET GENETICS) 11 154 975 2 WOOD RE AM REV RESPIR DIS 113 833 976 3 MOSSBERG B MT SINAI J MED 44 837 977 …
5757
Sample CF QueriesSample CF QueriesQN 00002QU Can one distinguish between the effects of mucus hypersecretion and infection on the submucosal glands of the respiratory tract in CF?NR 00007RD 169 1000 434 1001 454 0100 498 1000 499 1000 592 0002 875 1011
QN 00004QU What is the lipid composition of CF respiratory secretions?NR 00009RD 503 0001 538 0100 539 0100 540 0100 553 0001 604 2222 669 1010 711 2122 876 2222
NR: Number of Relevant documentsRD: Relevant Documents
Ratings code: Four 0-2 ratings, one from each expert
5858
Preprocessing for VSR Preprocessing for VSR ExperimentsExperiments• Separate file for each document with Separate file for each document with
just:just:– AuthorAuthor– TitleTitle– Major and Minor TopicsMajor and Minor Topics– Abstract (Extract)Abstract (Extract)
• Relevance judgment made binary by Relevance judgment made binary by assuming that assuming that allall documents rated 1 or 2 documents rated 1 or 2 by by anyany expert were relevant. expert were relevant.