现代信息检索导论英文课件：08-QueryExpansion

1、An Introduction to IR8th CourseChapter 9:Relevance feedback and query expansionRecap of the last lectureEvaluating a search engineBenchmarksPrecision and recallResults summariesThis lectureImproving resultsFor high recall. E.g., searching for aircraft doesnt match with plane; nor thermodynamic with

2、heatOptions for improving resultsFocus on relevance feedbackThe complete landscapeGlobal methodsQuery expansionThesauriAutomatic thesaurus generationLocal methodsRelevance feedbackPseudo relevance feedbackRelevance FeedbackRelevance feedback: user feedback on relevance of docs in initial set of resu

3、ltsUser issues a (short, simple) queryThe user marks returned documents 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

4、you dont know the collection well, so iterateRelevance Feedback: ExampleImage search engine http:/imsearch/imsearch.htmlResults for Initial QueryRelevance FeedbackResults after Relevance FeedbackRocchio AlgorithmThe Rocchio algorithm incorporates relevance feedback information into the vector space

5、model.Want to maximize sim (Q, Cr) - sim (Q, Cnr)The optimal query vector for separating relevant and non-relevant documents (with cosine sim.):Qopt = optimal query; Cr = set of rel. doc vectors; N = collection sizeUnrealistic: we dont know relevant documents.The Theoretically Best Query xxxxoooOpti

6、mal queryx non-relevant documentso relevant documentsoooxxxxxxxxxxxxxxRocchio 1971 Algorithm (SMART)Used in practice:qm = modified query vector; q0 = original query vector; ,: weights (hand-chosen or set empirically); Dr = set of known relevant doc vectors; Dnr = set of known irrelevant doc vectorsN

7、ew query moves toward relevant documents and away from irrelevant documentsTradeoff vs. / : If we have a lot of judged documents, we want a higher /.Term weight can go negativeNegative term weights are ignored (set to 0)Relevance feedback on initial query xxxxoooRevised queryx known non-relevant doc

8、umentso known relevant documentsoooxxxxxxxxxxxxxxInitial queryRelevance Feedback in vector spacesWe 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 precisionRelevance feedback is most

9、useful for increasing recall in situations where recall is importantUsers can be expected to review results and to take time to iteratePositive vs Negative FeedbackPositive feedback is more valuable than negative feedback (so, set 10,000 dimensions.3 dimensions: If a document is close to many querie

10、s, then some of these queries must be close to each other.Doesnt hold for a high-dimensional space.Probabilistic relevance feedbackRather than reweighting in a vector spaceIf user has told us some relevant and irrelevant documents, then we can proceed to build a classifier, such as a Naive Bayes mod

11、el:P(tk|R) = |Drk| / |Dr|P(tk|NR) = (Nk - |Drk|) / (N - |Dr|)tk = term in document; Drk = known relevant doc containing tk; Nk = total number of docs containing tkMore in upcoming lectures on classificationThis is effectively another way of changing the query term weightsBut note: the above proposal

12、 preserves no memory of the original weightsRelevance Feedback Example: Initial Query and Top 8 ResultsQuery: New space satellite applications+ 1. 0.539, 08/13/91, NASA Hasnt Scrapped Imaging Spectrometer+ 2. 0.533, 07/09/91, NASA Scratches Environment Gear From Satellite Plan 3. 0.528, 04/04/90, Sc

13、ience 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 Pr

14、ovides 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 CompaniesNote: want high recallRelevance Feedback Example: Expanded Query2.074 new 15.106 spa

15、ce30.816 satellite 5.660 application5.991 nasa 5.196 eos4.196 launch 3.972 aster3.516 instrument 3.446 arianespace3.004 bundespost 2.806 ss2.790 rocket 2.053 scientist2.003 broadcast 1.172 earth0.836 oil 0.646 measureTop 8 Results After Relevance Feedback+ 1. 0.513, 07/09/91, NASA Scratches Environm

16、ent Gear From Satellite Plan+ 2. 0.500, 08/13/91, NASA Hasnt 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+ 5. 0.492, 12/02/87, Telecommu

17、nications 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 MillionOther Uses of Rele

18、vance FeedbackFollowing a changing information needMaintaining an information filter (e.g., for a news feed)Active learningDeciding which examples it is most useful to know the class of to reduce annotation costsPseudo Relevance FeedbackAutomatic local analysisPseudo relevance feedback attempts to a

19、utomate the manual part of relevance feedback.Retrieve an initial set of relevant documents.Assume that top m ranked documents are relevant.Do relevance feedbackMostly works (perhaps better than global analysis!)Found to improve performance in TREC ad-hoc taskDanger of query driftPseudo relevance fe

20、edback:Cornell SMART at TREC 4Results show number of relevant documents out of top 100 for 50 queries (so out of 5000)Results contrast two length normalization schemes (L vs. l), and pseudo relevance feedback (PsRF) (done as adding 20 terms)lnc.ltc3210lnc.ltc-PsRF3634Lnu.ltu3709Lnu.ltu-PsRF4350Indir

21、ect relevance feedbackOn 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 relevantAssuming the displayed summaries are good, etc.Globally: Not user or query specific.This is t

22、he general area of click-stream miningEvaluation of relevance feedback strategiesUse q0 and compute precision and recall graphUse qm and compute precision recall graphAssess on all documents in the collectionSpectacular improvements, but its cheating!Partly due to known relevant documents ranked hig

23、herMust evaluate with respect to documents not seen by userUse documents in residual collection (set of documents minus those assessed relevant)Measures usually then lower than for original queryBut a more realistic evaluationRelative performance can be validly comparedEmpirically, one round of rele

24、vance feedback is often very useful. Two rounds is sometimes marginally useful.Relevance Feedback: AssumptionsA1: 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 docu

25、ments will be different from those in relevant documentsEither: 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 smallViolation of A1User

26、does not have sufficient initial knowledge.Examples:Misspellings (Brittany Speers).Cross-language information retrieval (hgado).Mismatch of searchers vocabulary vs. collection vocabularyCosmonaut/astronautViolation of A2There are several relevance prototypes.Examples:Burma/MyanmarContradictory gover

27、nment policiesPop stars that worked at Burger KingOften: instances of a general conceptGood editorial content can address problemReport on contradictory government policiesRelevance Feedback: ProblemsWhy do most search engines not use relevance feedback?Relevance Feedback: ProblemsLong queries are i

28、nefficient for typical IR engine.Long response times for user.High cost for retrieval system.Partial solution:Only reweight certain prominent termsPerhaps top 20 by term frequencyUsers are often reluctant to provide explicit feedbackIts often harder to understand why a particular document was retrie

29、ved after apply relevance feedbackWhy?Relevance Feedback on the Webin 2003: now less major search engines, but same general storySome search engines offer a similar/related pages feature (this is a trivial form of relevance feedback)Google (link-based)AltavistaStanford WebBaseBut some dont because i

30、ts hard to explain to average user:AllthewebmsnYahooExcite initially had true relevance feedback, but abandoned it due to lack of use./ ?Excite Relevance FeedbackSpink et al. 2000Only about 4% of query sessions from a user used relevance feedback optionExpressed as “More like this” link next to each

31、 resultBut about 70% of users only looked at first page of results and didnt pursue things furtherSo 4% is about 1/8 of people extending searchRelevance feedback improved results about 2/3 of the timeRelevance FeedbackSummaryRelevance feedback has been shown to be very effective at improving relevan

32、ce of results.Requires enough judged documents, otherwise its unstable ( 5 recommended)Requires queries for which the set of relevant documents is medium to largeFull relevance feedback is painful for the user.Full relevance feedback is not very efficient in most IR systems.Other types of interactiv

33、e retrieval may improve relevance by as much with less work.The complete landscapeGlobal methodsQuery expansion/reformulationThesauri (or WordNet)Automatic thesaurus generationGlobal indirect relevance feedbackLocal methodsRelevance feedbackPseudo relevance feedbackQuery Reformulation: Vocabulary To

34、ols FeedbackInformation about stop lists, stemming, etc.Numbers of hits on each term or phraseSuggestionsThesaurus Controlled vocabularyBrowse lists of terms in the inverted indexQueryexpansionQuery ExpansionIn relevance feedback, users give additional input (relevant/non-relevant) on documents, whi

35、ch is used to reweight terms in the documentsIn query expansion, users give additional input (good/bad search term) on words or phrases.Query Expansion: ExampleAlso: see , Types of Query ExpansionGlobal Analysis: (static; of all documents in collection)Controlled vocabularyMaintained by editors (e.g

36、., medline)Manual thesaurusE.g. MedLine: physician, syn: doc, doctor, MD, medicoAutomatically derived thesaurus(co-occurrence statistics)Refinements based on query log miningCommon on the webLocal Analysis: (dynamic)Analysis of documents in result setControlled VocabularyThesaurus-based Query Expans

37、ionThis doesnt require user inputFor each term, t, in a query, expand the query with synonyms and related words of t from the thesaurusfeline feline catMay weight added terms less than original query terms.Generally increases recall.Widely used in many science/engineering fieldsMay significantly dec

38、rease precision, particularly with ambiguous terms.“interest rate” “interest rate fascinate evaluate”There is a high cost of manually producing a thesaurusAnd for updating it for scientific changesAutomatic Thesaurus GenerationAttempt to generate a thesaurus automatically by analyzing the collection

39、 of documentsTwo main approachesCo-occurrence based (co-occurring words are more likely to be similar)Shallow analysis of grammatical relationsEntities that are grown, cooked, eaten, and digested are more likely to be food items.Co-occurrence based is more robust, grammatical relations are more accu

40、rate.Why?Co-occurrence ThesaurusSimplest way to compute one is based on term-term similarities in C = AAT where A is term-document matrix.wi,j = (normalized) weighted count (ti , dj)tidjnmWith integercounts whatdo you getfor a booleancooccurrencematrix?Automatic Thesaurus GenerationExampleAutomatic

41、Thesaurus GenerationDiscussionQuality 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 notFalse negatives: Words deemed dissimilar that are similarSince terms are highly correlated anyway, expansion may not retrieve many additional documents.Query Expansion: SummaryQuery expansion is often effective in increasing recall.Not always with general thesauriFairly successful for subject-specific collectionsIn