automatic term mismatch diagnosis for selective query expansion

31
Automatic Term Mismatch Diagnosis for Selective Query Expansion Le Zhao and Jamie Callan Language Technologies Institute School of Computer Science Carnegie Mellon University Pittsburgh, PA @SIGIR 2012, Portland, OR

Upload: aadi

Post on 25-Feb-2016

61 views

Category:

Documents


0 download

DESCRIPTION

Automatic Term Mismatch Diagnosis for Selective Query Expansion. Le Zhao and Jamie Callan Language Technologies Institute School of Computer Science Carnegie Mellon University Pittsburgh, PA @SIGIR 2012, Portland, OR. Main Points. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

Automatic Term Mismatch Diagnosis for

Selective Query ExpansionLe Zhao and Jamie Callan

Language Technologies InstituteSchool of Computer ScienceCarnegie Mellon University

Pittsburgh, PA@SIGIR 2012, Portland, OR

Page 2: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

Main Points

• An important problem – term mismatch & a traditional solution

• New diagnostic intervention approach

• Simulated user studies

• Diagnosis & intervention effectiveness

2

Page 3: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

3

• Average term mismatch rate: 30-40% [Zhao10]

• A common cause of search failure [Harman03, Zhao10]• Frequent user frustration [Feild10]• Here: 50% - 300% gain in retrieval accuracy

Term Mismatch Problem

Relevant docs not returned Web, short queries, stemmed,

inlinks included

Page 4: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

Term Mismatch Problem

Example query (TREC 2006 Legal discovery task):approval of (cigarette company) logos on television watched by children

4

approval logos television watched children

Mismatch 94% 86% 79% 90% 82%

Highest mismatch rate

High mismatch rate for all query terms in this query

Page 5: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

5

The Traditional Solution: BooleanConjunctive Normal Form (CNF) Expansion

Keyword query:approval of logos on television watched by children

Manual CNF (TREC Legal track 2006): (approval OR guideline OR strategy)AND (logos OR promotion OR signage OR brand OR mascot OR marque OR mark)AND (television OR TV OR cable OR network)AND (watched OR view OR viewer)AND (children OR child OR teen OR juvenile OR kid OR adolescent)

– Expressive & compact (1 CNF == 100s alternatives)– Highly effective (this work: 50-300% over base keyword)

Page 6: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

The Potential

• Query: approval logos television watched children

6

50-300% Recall approval 6.49%logos 14.1%television 21.3%watched 10.4%children 18.0%Overall 2.04%

The Potential

? Recall+guideline+strategy == 12.8%+promotion+signage... == 19.7%+tv+cable+network == 22.4%+view+viewer == 19.5%+child+teen+kid... == 19.3% == 8.74%

Page 7: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

CNF Expansion

• Widely used in practice– Librarians [Lancaster68, Harter86]– Lawyers [Lawlor62, Blair85, Baron07]– Search experts [Clarke95, Hearst96, Mitra98]

• Less well studied in research– Users do not create effective free form Boolean queries

([Hearst09] cites many studies).• Question: How to guide user effort in productive directions

– restricting to CNF expansion (to the mismatch problem)– focusing on problem terms when expanding

7WikiQuery [Open Source IR Workshop] Ad

Page 8: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

Main Points

• An important problem – term mismatch & a traditional solution

• New diagnostic intervention approach

• Simulated user studies

• Diagnosis & intervention effectiveness

8

Page 9: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

Diagnostic Intervention

• Goal– Least amount user effort near optimal performance– E.g. expand 2 terms 90% of total improvement

9

approval of logos on television watched by children

approval of logos on television watched by children

High idf (rare) terms

CNF (approval OR guideline OR strategy) AND logos AND television AND (watch OR view OR viewer)AND children

(approval OR guideline OR strategy) AND logosAND (television OR tv OR cable OR network)AND watch AND children

Query: approval of logos on television watched by children

Diagnosis:

Expansion:

Low terms

CNF

Page 10: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

Diagnostic Intervention

10

[ 0.9 (approval logos television watch children) 0.1 (0.4 guideline 0.3 strategy 0.5 view 0.4 viewer)]

[ 0.9 (approval cigar television watch children) 0.1 (0.4 guideline 0.3 strategy 0.5 tv 0.4 cable 0.2 network) ]

Diagnosis:

Expansion query

Bag of wordExpansion: Bag of wordOriginal query

High idf (rare) termsLow termsapproval of logos on television watched by children

Query: approval of logos on television watched by children

• Goal– Least amount user effort near optimal performance– E.g. expand 2 terms 90% of total improvement

approval of logos on television watched by children

Page 11: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

Diagnostic Intervention

• Diagnosis methods– Baseline: rareness (high idf)– High predicted term mismatch P(t | R) [Zhao10]

• Intervention methods– Baseline: bag of word (Relevance Model [Lavrenko01])

• w/ manual expansion terms• w/ automatic expansion terms

– CNF expansion (probabilistic Boolean ranking)• E.g.

11

_

(approval OR guideline OR strategy)ANDP logos ANDP televisionANDP (watch OR view OR viewer)ANDP children

Page 12: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

Main Points

• An important problem – term mismatch & a traditional solution

• New diagnostic intervention approach

• Evaluation: Simulated user studies

• Diagnosis & intervention effectiveness

12

Page 13: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

13

User Keyword queryDiagnosis

system(P(t | R) or idf)

Problem query termsUser expansionExpansion

terms

Query formulation

(CNF or Keyword)

Retrieval engine Evaluation

(child AND cigar)

(child > cigar)(child teen)

(child OR teen) AND cigar

Diagnostic Intervention (We Hope to)

Page 14: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

14

Diagnostic Intervention (We Hope to)

User Keyword queryDiagnosis

system(P(t | R) or idf)

Problem query termsUser expansionExpansion

terms

Query formulation

(CNF or Keyword)

Retrieval engine Evaluation

(child AND cigar)

(child > cigar)

(child OR teen) AND cigar

(child teen)

Page 15: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

15

Expert user Keyword queryDiagnosis

system(P(t | R) or idf)

Problem query termsUser expansionExpansion

terms

Query formulation

(CNF or Keyword)

Retrieval engine Evaluation

Online simulation

Online simulation

We Ended up Using Simulation

(child teen)

(child OR teen) AND cigar

(child OR teen) AND (cigar OR tobacco)

FullCNFOffline

(child AND cigar)

(child > cigar)

Page 16: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

16

Diagnostic Intervention Datasets

• Document sets– TREC 2007 Legal track, 7 million tobacco corp., train on 2006– TREC 4 Ad hoc track, 0.5 million newswire, train on TREC 3

• CNF Queries– TREC 2007 by lawyers, TREC 4 by Univ. Waterloo [Clarke95]– 50 topics each, 2-3 keywords per query

• Relevance Judgments– TREC 2007 sparse, TREC 4 dense

• Evaluation measures– TREC 2007 statAP, TREC 4 MAP

Page 17: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

Main Points

• An important problem – term mismatch & a traditional solution

• New diagnostic intervention approach

• Simulated user studies

• Diagnosis & intervention effectiveness

17

Page 18: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

18

P(t | R) vs. idf diagnosis

Results – Diagnosis

Diagnostic CNF expansion on TREC 4 and 2007

0 1 2 3 4 All0%

10%20%30%40%50%60%70%80%90%

100%

P(t | R) on TREC 2007idf on TREC 2007P(t | R) on TREC 4idf on TREC 4

# query terms selected

Gain in retrieval (MAP)

8%-50%

No Expansion

Full Expansion

Page 19: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

19

Results – Expansion Intervention

CNF vs. bag-of-word expansion

0 1 2 3 4 All0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

CNF on TREC 4Bag of word on TREC 4CNF on TREC 2007Bag of word on TREC 2007

# query terms selected

Retrieval performance (MAP)

P(t | R) guided expansion on TREC 4 and 2007

50% to300%gain

Similar level of gain in top precision

Page 20: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

Main Points

• An important problem – term mismatch & a traditional solution

• New diagnostic intervention approach

• Simulated user studies

• Diagnosis & intervention effectiveness

20

Page 21: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

21

Conclusions

• One of the most effective ways to engage user interactions– CNF queries gain 50-300% over keyword baseline.

• Mismatch diagnosis simple & effective interactions– Automatic diagnosis saves user effort by 33%.

• Expansion in CNF easier and better than in bag of word– Bag of word requires balanced expansion of all terms.

• New research questions:– How to learn from manual CNF queries to improve

automatic CNF expansion– get ordinary users to create effective CNF expansions

(with the help of interfaces or search tools)

Page 22: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

22

Acknowledgements

Chengtao Wen, Grace Hui Yang, Jin Young Kim, Charlie Clarke, SIGIR Reviewers

Helpful discussions & feedback

Charlie Clarke, Gordon Cormack, Ellen Voorhees, NISTAccess to data

NSF grant IIS-1018317Opinions are solely the authors’.

Page 23: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

23

END

Page 24: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

The Potential

• Query: approval logos television watched children

24

logos +promotion +signage +brand All

Mismatch 85.9% 81.1% 80.9% 80.3% 80.3%

Recall 14.1% 18.9% 19.1% 19.7% 19.7%

50-300% Recall Recalllogos 14.1% +promotion+signage... == 19.7%approval 6.49% +guideline+strategy == 12.8%television 21.3% +tv+cable+network == 22.4%watched 10.4% +view+viewer == 19.5%children 18.0% +child+teen+kid... == 19.3%Overall 2.04% == 8.74%

The Potential

?

Page 25: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

Failure Analysis (vs. baseline)

Diagnosis:• 4 topics: wrong P(t | R) prediction, lower MAP

Intervention:• 3 topics: right diagnosis, but lower MAP• 2 of the 3: no manual expansion for the selected term

– Users do not always recognize which terms need help.• 1 of the 3: wrong expansion terms by expert

– “apatite rocks” in nature, not “apatite” chemical– CNF expansion can be difficult w/o looking at retrieval

results.25

Page 26: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

-1 -0.8 -0.6 -0.4 -0.2 0 0.2

-0.1

-0.05

0

0.05

0.1

0.15y - MAP Difference

x - P(t | R) Difference

(P(t | R) better prediction and better MAP)

(P(t | R) better prediction, but lower MAP)

(idf betterprediction, and better MAP)

Failure Analysis -- Comparing diagnosis methods: P(t | R) vs. idf

26

User didn’t expand Wrong

expansion

Legend query query with unexpanded term(s)

Page 27: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

Term Mismatch Diagnosis

• Predicting term recall - P(t | R) [Zhao10]– Query dependent features (model causes of mismatch)

• Synonyms of term t based on query q’s context• How likely these synonyms occur in place of t• Whether t is an abstract term• How rare t occurs in the collection C

– Regression prediction: fi(t, q, C) P(t | R)– Used in term weighting for long queries

• Lower predicted P(t | R) higher likelihood of mismatch t more problematic

27

Page 28: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

28

Online or Offline Study?

• Controlling confounding variables– Quality of expansion terms– User’s prior knowledge of the topic– Interaction effectiveness & effort

• Enrolling many users• Offline simulations can avoid all these and still make

reasonable observations

Page 29: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

29

Simulation Assumptions

• Full expansion to simulate partial expansions• 3 assumptions about user expansion process

– Independent expansion of query terms• A1: same set of expansion terms for a given query term, no

matter which subset of query terms gets expanded• A2: same sequence of expansion terms, no matter …

– A3: Re-constructing keyword query from CNF• Procedure to ensure vocabulary faithful to that of the original

keyword description• Highly effective CNF queries ensure reasonable kw baseline

Page 30: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

30

Results – Level of Expansion

• More expansion per query term, better retrieval• Result of expansion terms being effective• Queries with significant gain in retrieval after expanding

more than 4 terms:– Topic 84, cigarette sales in James Bond movies

Page 31: Automatic Term Mismatch  Diagnosis for  Selective  Query Expansion

31

Expert User Keyword Query Diagnosis system(P(t | R) or idf)

Problem query terms User expansion

Expansion termsQuery formulation(CNF or Keyword)Retrieval engineEvaluation

Online simulation

(child OR youth) AND (cigar OR tobacco)

(child AND cigar)

(child --> youth)(child OR youth) AND cigar

(child > cigar)

Online simulation

Offline Full CNF Query

Idf (rare) 1.22 1.92 1.69 1.87 1.40

Most infrequent