Multilingual Information Retrieval Using English and Chinese Queries

Transcription

1 Multilingual Information Retrieval Using and Chinese Queries Aitao Chen School of Information Management and Systems University of California, Berkeley CLEF 2001 Workshop: 3-4 Sept, 2001, Darmstadt, Germany

2 Outline Overview over what we did at CLEF-2001 German decompounding Chinese topics translation Merging strategies and alternative methods Conclusions

3 Participation in CLEF-2001 Monolingual task (German and Spanish) Bilingual task (Chinese to ) Multilingual task ( and Chinese)

4 Overview of Multilingual Information Retrieval Using Queries Query Documents SYSTRAN and L&H French German French German Italian Italian Spanish Spanish docs French docs German docs Italian docs Spanish docs merger combined ranked list of documents

5 Chinese Overview of Multilingual Information bilingual dict parallel texts search engine Retrieval Using Chinese Queries SYSTRAN and L&H Query French German Italian Documents French German Italian Spanish Spanish docs French docs German docs Italian docs Spanish docs merger combined ranked list of documents

6 German Decompounding Procedure Create a German base dictionary consisting of single words only (compounds are excluded). Decompose a compound into component words found in the German base dictionary. Choose the decomposition with the minimum number of component words. If there are more than one decompositions having the minimum number of component words, choose the decomposition with the highest probability.

7 German Decompounding: Example 1 Compound: filmfestspiele (film festival) 1. Base dictionary film fest fests festspiele piele s 2. Decompositions: 1. film fest s piele 2. film fest spiele 3. film fests piele 4. film festspiele 3. Result: filmfestspiele = file festspiele

8 German Decompounding: Example 2 Compound: hungerstreiks (hunger strike) 1. Base dictionary erst hung hunger hungers hungerst reik reiks s streik streiks 2. Decompositions: log p(d) 1. hung erst reik s hung erst reiks hunger streik s hunger streiks hungerst reik s hungerst reiks Result: hungerstreiks = hunger streiks

9 German Decompounding: Probability of Decomposition C = W 1 W2 W3 W4 p( C) = p( W1 ) p( W2 ) p( W3 ) p( W4 ) p( w) = n tfc( i= 1 tfc( w) w i ) tfc(w) is the number of times word w occurs in a corpus. n is the number of unique words (including compounds) in a corpus.

10 German Decompounding: Failed Cases 1. erdatmosphäre = erde + atmosphäre (earth atmosphere) 2. mittagessenzeit = mittag essen zeit (noon meal time) (mittagessenzeit = mittagessen zeit) lunch time 3. And others

11 German Decompounding and Monolingual Retrieval Performance Test collections -Decompounding -Stemming -Expansion + Decompounding Change CLEF-2001 (49/225K).3673 (1877/2130).4314 (1949/2130) % CLEF-2000 (37/154K).3189 (673/821).4112 (770/821) % TREC-6/7/8 (73/252K).2993 (1907/2626).3368 (2172/2626) % Only component words of compounds are kept in the queries.

12 German Monolingual Retrieval Performance Precision Recall BK2GGA1 (.4050) BK2GGA2 (.3551) bk2gga1* (.4436) Features: +stemming, +decompounding, -expansion

13 Overview of Chinese to Retrieval Chinese topics segmentation stopwords removal Translation resources Term selection & weighting de-segmentation LDC bilingual wordlist term selection Monolingual Chinese words Bilingual dict (parallel texts) term selection term merging & weighting queries (in words) docs Preprocessing Chinese search engine term selection system retrieval results

14 Chinese Topics Preprocessing: De-segmentation

15 Translation Resources: Creation of Bilingual Dictionary From Parallel Texts Parallel texts: Hong Kong news (4/98-4/2001) and FBIS Chinese collection. Document alignment: + LDC wordlist. Paragraph & sentence alignment: adapted from Gale and Church s length-based model. Association measure: Dunning s maximum likelihood ratio statistic.

16 Term Translation Using Search Engine

17 E1 1 E2 1 E3 1 Term Selection, Merging, and E3 1 E4 1 Weighting (1) Top-3 translations of Chinese word C1 from LDC wordlist. Translations are ranked by occurrence frequency in the LA Times collection. (2) Top-2 translations of Chinese word C1 from parallel texts. Translations are ranked by association weight. (1) (2) (5) E1 1 E2 1 E3 2 E4 1 C1 2 E1.20 E2.20 E3.40 E4.20 Original query term frequency of C1 Final term weights for translations of C1 E1.40 E2.40 E3.80 E4.40 (3) (4) (6)

18 Translation Resources Versus Chinese-to- Performance Precision Recall LDC+HKF+YAHOO (.4112) LDC+HKF (.3599) LDC (.2679) HKF (.2675) Mono (.5553)

19 Multilingual Information Retrieval: Merging Strategy docs French docs Italian docs German docs Spanish docs E1 e1 E2 e2 E50 e50 E51 e51 E1000 e1000 F1 f1 F2 f2 F50 f50 F51 f51 F1000 f1000 I1 i1 I2 i2 I50 i50 I51 i51 I1000 i1000 G1 g1 G2 g2 G50 g50 G51 g51 G1000 g1000 S1 s1 S2 s2 S50 s50 S51 s51 S1000 s1000 E1.8*e1 + 1 E2.8*e2 + 1 E50.8*e E51.8*e51 E1000.8*e1000 F1 f1 + 1 F2 f2 + 1 F50 f F51 f51 F1000 f1000 I1 i1 + 1 I2 i2 + 1 I50 i I51 i51 I1000 i1000 G1 g1 + 1 G2 g2 + 1 G50 g G51 g51 G1000 g1000 S1 s1 + 1 S2 s2 + 1 S50 s S51 s51 S1000 s1000 (1) combine lists; (2) sort by adjusted weight; (3) take top 1000 docs

20 Performance of Multilingual Information Retrieval Using Long Queries Query Documents French French SYSTRAN and L&H German German Italian Italian Spanish Spanish docs French docs German docs Italian docs Spanish docs (.5553) (.4776) (.3789) (.3934) (.4703) merger (.3424) combined ranked list of documents

21 Performance of Multilingual Information Retrieval Using Chinese Long Queries Original Query Chinese bilingual dict parallel texts search engine SYSTRAN and L&H (.4122) Query French German Italian Documents French German Italian Spanish Spanish docs French docs German docs Italian docs Spanish docs (.4122) (.2874) (.2619) (.2509) (.2942) merger (.2217) combined ranked list of documents

22 Multilingual Information Retrieval: Alternative Merging Strategy docs French docs Italian docs German docs Spanish docs E1 e1 E2 e2 E50 e50 E51 e51 E1000 e1000 F1 f1 F2 f2 F50 f50 F51 f51 F1000 f1000 I1 i1 I2 i2 I50 i50 I51 i51 I1000 i1000 G1 g1 G2 g2 G50 g50 G51 g51 G1000 g1000 S1 s1 S2 s2 S50 s50 S51 s51 S1000 s1000 E1 e1/e1 E2 e2/e1 E50 e50/e1 E51 e51/e1 E1000 e1000/e1 F1 f1/f1 F2 f2/f1 F50 f50/f1 F51 f51/f1 F1000 f1000/f1 I1 i1/i1 I2 i2/i1 I50 i50/i1 I51 i51/i1 I1000 i1000/i1 G1 g1/g1 G2 g2/g1 G50 g50/g1 G51 g51/g1 G1000 g1000/g1 S1 s1/s1 S2 s2/s1 S50 s50/s1 S51 s51/s1 S1000 s1000/s1 (1) combine lists; (2) sort by adjusted weight; (3) take top 1000 docs

23 Multilingual Information Retrieval: Alternative Method 1 Multilingual Query Multilingual Document Collection translator French German Italian engine French German Italian Spanish Spanish ranked list of docs in multiple languages

24 Multilingual Information Retrieval: Alternative Method 2 Translated documents Query Original documents engine translator translator French German translator Italian translator Spanish ranked list of docs in

25 Multilingual Information Retrieval: Alternative Method 3 Query Documents French French translator German German Italian Italian Spanish Spanish docs French docs German docs Italian docs Spanish docs translator translator translator translator docs docs docs docs docs combined ranked list of documents

26 Performance of Different ML Methods Precision Recall BK2MUEAA1 (.3424) NormalizedMerging (.3286) ML Alternative 1 (.3126) ML Alternative 3 (.3648)

27 Conclusions German decompounding can significantly improve retrieval performance. Keeping only component words in the query works better than keeping both compounds and component words. Chinese search engine is a valuable resource for translating Chinese proper nouns into. Merging documents by adjusted probability of relevance works reasonably well.