Academia Sinica 05 CROSSLANGUAGE IR AND STATISTICAL MT

  • Slides: 62
Download presentation
Academia Sinica 05 CROSS-LANGUAGE IR AND STATISTICAL MT Jian-Yun Nie DIRO, University of Montreal

Academia Sinica 05 CROSS-LANGUAGE IR AND STATISTICAL MT Jian-Yun Nie DIRO, University of Montreal http: //www. iro. umontreal. ca/~nie 1

2 Outline • What are the problems in CLIR? • The approaches proposed in

2 Outline • What are the problems in CLIR? • The approaches proposed in the literature • Their effectiveness • Remaining problems

3 Problem of CLIR • Cross-language IR (CLIR) • Query in a language (e.

3 Problem of CLIR • Cross-language IR (CLIR) • Query in a language (e. g. Chinese) and documents in another language (English) • Multilingual IR (MLIR) • Query in one language and documents in several languages • Where CLIR and MLIR are useful? • Search for international patents • Identify possible competitors or collaborators in other countries • Search for local information that is only in a local language • Multilingual users: avoid the burden to issue several queries • … • In many cases, the translation of retrieved documents into the language of the query is still necessary (goal of machine translation)

History • In 1970 s, first papers on CLIR • TREC-3 (1994) Spanish (monolingual):

History • In 1970 s, first papers on CLIR • TREC-3 (1994) Spanish (monolingual): El Norte Newspaper SP 1 -25 • TREC-4 (1995) Spanish (monolingual): El Norte Newspaper SP 26 -50 • TREC-5 (1996) Spanish (monolingual): El Norte newspaper and Agence France Presse SP 51 -75 Chinese (monolingual): Xinhua News agency, People’s Daily • TREC-6 (1997) CH 1 -28 Chinese (monolingual), The same documents as TREC-6 CH 29 -54 CLIR: English: Associated Press CL 1 -25 French, German: Schweìzerìsche Depeschenagentur (SDA) • TREC-7 (1998) CLIR: English, French, German, Italian (SDA) CL 26 -53 + German: New Zurich Newspaper (NZZ) CLIR (English, French, German, Italian): as in. TREC-7 CL 54 -81 • TREC-9 (2000) English-Chinese: Chinese newswire articles from Hong Kong CH 55 -79 • TREC 2001 English-Arabic: Arabic newswire from Agence France Presse • TREC-8 (1999) • TREC 2002 1 -25 26 -75 4

5 History • NTCIR (Japon, NII) (1999 -) • Asian languages (CJK) + English

5 History • NTCIR (Japon, NII) (1999 -) • Asian languages (CJK) + English • Patent retrieval, blogs, Evaluation methodology, . . • CLEF (Europe) (2000 -) • European languages • Image retrieval, Wikipedia, … • SEWM: Chinese IR (2004 -) • FIRE: IR in Indian languages (2008 -) • Russian, … • Search engines • Yahoo!: 2006, French/German->German/French, English, Spanish, Italian • Google: 2007, Query translation, translation of retrieved documents

6 Problems in CLIR • Translation of query (or documents) so as to compare

6 Problems in CLIR • Translation of query (or documents) so as to compare them • Similarities with MT • Translation • Similar methods can be used • A task different from MT • Short queries (2 -3 words): HD video recording • Flexible syntax: video hd recording, recording hd… • Goal: help find relevant documents, not to make the translated query readable • The "translation" can be by related words (same/related subjet, …) • Less strict translation • Important to weight translation terms • weight = correctness of translation + Utility for IR • E. g. cost for computers Translation-> 计算机成本,计算机开销, 计算机价格, … utility for RI ->计算机成本,计算机开销, 计算机价格, … …

7 Strategies • Translate the query • Translate the documents • The two strategies

7 Strategies • Translate the query • Translate the documents • The two strategies have similar effectiveness • More complex to translate documents • Translate both query and documents into a third language (pivot language) • Less effective than direct translation • (Related) Transitive translation: French->English-> Chinese

Academia Sinica 05 8 How to translate 1. Machine Translation (MT) 2. Bilingual dictionaries,

Academia Sinica 05 8 How to translate 1. Machine Translation (MT) 2. Bilingual dictionaries, thesauri, lexical resources, … 3. Parallel texts: translated texts Parallel texts encompass translation knowledge

Academia Sinica 05 9 Approach 1: Using MT • Seems to be the ideal

Academia Sinica 05 9 Approach 1: Using MT • Seems to be the ideal tool for CLIR and MLIR (if the translation quality is high) Query in F Translation in E MT Documents in E • Typical effectiveness: 80 -100% of the monolingual effectiveness • Problems: • Quality • Availability • Development cost

Academia Sinica 05 10 Problems of MT • Wrong choice of translation word/term •

Academia Sinica 05 10 Problems of MT • Wrong choice of translation word/term • organic food – nouriture organique (biologique) • train skilled personnel - personnel habile de train (ambiguity) • Wrong syntax • human-assisted machine translation - traduction automatique humainaidée • Unknown words • Personal names: Bérégovoy, Beregovoy �小平 Deng Xiaoping, Deng Hsao-ping, Deng Hsiao p'ing • For CLIR: Choose one translation word • E. g. organic – organique • Better to keep all the synonyms (organique, biologique)? – query expansion effect

11 Exemples Systran Google trafic de stupéfiants (correct) 毒品交易 (correct) 毒品贩运 (correct) • 1.

11 Exemples Systran Google trafic de stupéfiants (correct) 毒品交易 (correct) 毒品贩运 (correct) • 1. drug traffic • 2. drug insurance: assurance de drogue (incorrect) d'assurance médicaments (correct) 药物保险 (correct) �物 保� (correct) • 3. drug research: recherche de drogue (incorrect) la recherche sur les drogues (incorrect) 药物研究 (correct) • 4. drug for treatment of Friedreich’s ataxia: drogue pour le traitement de �物 研究 (correct) médicament pour le traitement de l'ataxie de Friedreich (incorrect) l'Ataxie de Friedreich (correct) Friedreich的不整� 的治�的 �物 (correct) �物 治� 弗里德的共�失� (correct) • 5. drug control: commande de drogue (likely incorrect) contrôle des drogues (likely) 药物管制 (likely) �物 管制 (likely) production de drogue (likely) la production de drogues (likely) 药物生产 (likely) �物 生� (likely) • 6. drug production:

Academia Sinica 05 12 Approach 2: Using bilingual dictionaries • Unavailability of high-quality MT

Academia Sinica 05 12 Approach 2: Using bilingual dictionaries • Unavailability of high-quality MT systems for many language pairs • MT systems are often a closed box that is difficult to adapt to IR task • Bilingual dictionary: • A non-expensive alternative • Usually available

Academia Sinica 05 13 Approach 2: Using bilingual dictionaries • General form of dict.

Academia Sinica 05 13 Approach 2: Using bilingual dictionaries • General form of dict. (e. g. Freedict) access: attaque, accéder, intelligence, entrée, accès academic: étudiant, académique branch: filiale, succursale, spécialité, branche data: données, matériau, data • LDC English-Chinese • AIDS/艾滋病/爱滋病/ • data /材料/资料/事实/数据/基准/ • prevention /阻碍/防止/妨碍/预防/预防法/ • problem /问题/难题/疑问/习题/作图题/将军/课题/困难/难/题是/ • structure /构造/构成/结构/组织/化学构造/石理/纹路/构造物/ 建筑物/建造/物/

Academia Sinica 05 14 Basic methods • Use all the translation terms • data

Academia Sinica 05 14 Basic methods • Use all the translation terms • data /材料/资料/事实/数据/基准/ • structure /构造/构成/结构/组织/化学构造/石理/纹路/构造物/ 建筑物/建造/物/ • Introduce noise • Implicitly, the term with more translations is assigned higher importance • Use the first (or most frequent) translation • Limit to the most frequent translation (when frequency is available) • Not always an appropriate choice • General effectiveness: 50 -60% of monolingual IR • Problems of dictionary • Coverage (unknown words, unknown translations) • [Xu and Weischedel 2005] tested the impact of dictionary coverage on CLIR (En-Ch) • The effectiveness increases till 10 000 entries

Academia Sinica 05 15 Translate the query as a whole • Phrase translation [Ballesteros

Academia Sinica 05 15 Translate the query as a whole • Phrase translation [Ballesteros and Croft, 1996, 1997] base de données: pomme de terre: database potato • Translate phrases first • Then the remaining words • Best global translation for the whole query 1. Candidates: For each query word • 2. Determine all the possible translations (through a dictionary) Selection select the set of translation words that produce the highest cohesion

Academia Sinica 05 16 Cohesion • Cohesion ~ frequency of two translation words together

Academia Sinica 05 16 Cohesion • Cohesion ~ frequency of two translation words together E. g. • data: données, matériau, data • access: attaque, accéder, intelligence, entrée, accès (accès, données) 152 * (accéder, données) 31 (données, entrée) 21 (entrée, matériau) 3 … • Freq. from a document collection or from the Web (Grefenstette 99) • (Gao, Nie et al. 2001) (Liu and Jin 2005)(Seo et al. 2005)… • sim: co-occurrence, mutual information, statistical dependence • Dynamic translation: Graph of terms in two languages connected by dictionary translations + random walk • Improved effectiveness (80 -100% of monolingual IR)

Academia Sinica 05 17 Approach 3: using parallel texts • Training a translation model

Academia Sinica 05 17 Approach 3: using parallel texts • Training a translation model (IBM 1) • Principle: • train a statistical translation model from a set of parallel texts: p(tj|si) • Principle: The more sj appears in parallel texts of ti, the higher p(tj|si). • Given a query, use the translation words with the highest probabilities as its translation

Academia Sinica 05 18 Simple utilization • Determine the probability of a word translation

Academia Sinica 05 18 Simple utilization • Determine the probability of a word translation • One should also take into account the discriminant power of the translation (IDF)

19 example Query #3 What measures are being taken to stem international drug traffic?

19 example Query #3 What measures are being taken to stem international drug traffic? médicament=0. 110892 mesure=0. 091091 international=0. 086505 trafic=0. 052353 drogue=0. 041383 découler=0. 024199 circulation=0. 019576 pharmaceutique=0. 018728 pouvoir=0. 013451 • multiple translations, but prendre=0. 012588 ambiguity is kept extérieur=0. 011669 passer=0. 007799 • Unknown word in target demander=0. 007422 language endiguer=0. 006685 nouveau=0. 006016 stupéfiant=0. 005265 produit=0. 004789

20 IBM 1 + dictionnaire • The weight of each translation word in the

20 IBM 1 + dictionnaire • The weight of each translation word in the dictionary is increased (TREC-6) Without default prob. • MAP-mono = 0. 3731 • MAP-LOGOS = 0. 2866 (76. 8%), MAP-Systran = 0. 2763 (74. 1%)

21 Integrating translation in an IR model (Kraaij et al. 2003) • The problem

21 Integrating translation in an IR model (Kraaij et al. 2003) • The problem of CLIR: • Query translation (QT) • Document translation (DT)

22 Results (CLEF 2000 -2002) Translation model (IBM 1) trained on a web collection

22 Results (CLEF 2000 -2002) Translation model (IBM 1) trained on a web collection

Academia Sinica 05 Principle of translation model training • p(tj|si) is estimated from a

Academia Sinica 05 Principle of translation model training • p(tj|si) is estimated from a parallel training corpus, aligned into parallel sentences • IBM models 1, 2, 3, … • Process: • Input = two sets of parallel texts • Sentence alignment A: Sk Tl (bitext) • Initial probability assignment: t(tj|si, A) • Expectation Maximization (EM): t(tj|si , A) • Final result: t(tj|si) = t(tj|si , A) 23

Academia Sinica 05 24 Details on translation model training on a parallel corpus •

Academia Sinica 05 24 Details on translation model training on a parallel corpus • Sentence alignment • Align a sentence in the source language to its translation(s) in the target language • Translation model • Extract translation relationships • Various models (assumptions)

Academia Sinica 05 25 Sentence alignment • Assumption: • The order of sentences in

Academia Sinica 05 25 Sentence alignment • Assumption: • The order of sentences in two parallel texts is similar • A sentence and its translation have similar length (length-based alignment, e. g. Gale & Church) di: distance for different patterns (0 -1, 1 -1, …) • A translation contains some “known” translation words, or cognates (e. g. Simard et al. 93)

Academia Sinica 05 26 Example of aligned sentences (Canadian Hansards) Débat L'intelligence artificielle Artificial

Academia Sinica 05 26 Example of aligned sentences (Canadian Hansards) Débat L'intelligence artificielle Artificial intelligence A Debat Depuis 35 ans, les spécialistes d'intelligence artificielle cherchent à construire des machines pensantes. Attempts to produce thinking machines have met during the past 35 years with a curious mix of progress and failure. Leurs avancées et leurs insuccès alternent curieusement. Two further points are important. Les symboles et les programmes sont des notions purement abstraites. First, symbols and programs are purely abstract notions.

Academia Sinica 05 TM training: Initial probability assignment t(tj|si, A) même un cardinal n’

Academia Sinica 05 TM training: Initial probability assignment t(tj|si, A) même un cardinal n’ est pas à l’ abri des cartels de la drogue. even a cardinal is not safe from drug cartels. 27

Academia Sinica 05 TM training: Application of EM: t(tj|si, A) même un cardinal n’

Academia Sinica 05 TM training: Application of EM: t(tj|si, A) même un cardinal n’ est pas à l’ abri des cartels de la drogue. even a cardinal is not safe from drug cartels. 28

Academia Sinica 05 29 IBM models (Brown et al. ) • IBM 1: does

Academia Sinica 05 29 IBM models (Brown et al. ) • IBM 1: does not consider positional information and sentence length • IBM 2: considers sentence length and word position • IBM 3, 4, 5: fertility in translation • For CLIR, IBM 1 seems to correspond to the current (bag- of-words) approaches to IR.

IBM translation models: principle • Input: bitexts (set of aligned sentences) • Output: transfer

IBM translation models: principle • Input: bitexts (set of aligned sentences) • Output: transfer probability t(f|e) the (le, 0. 18) (la, 0. 15) (de, 0. 12) … minister (ministre, 0. 8) (le, 0. 12), … people (gens, 0. 25) (les, 0. 16) (personnes, 0. 1), … years (ans, 0. 38) (années, 0. 31) (depuis, 0. 12), … • Each pair of (e, f) is a parameter of the model • • In practice: • Limit to only 1 -1 sentence alignments, of length <=40 words • Words of freq. =1 replaced by UNK • Using EM to re-estimate parameters Academia Sinica 05 30

Word alignment for one sentence pair Source sentence in training: e = e 1,

Word alignment for one sentence pair Source sentence in training: e = e 1, …el (+NULL) Target sentence in training: f = f 1, …fm Only consider alignments in which each target word (or position j) is aligned to a source word (of position a j) The set of all the possible word alignments: A(e, f) Academia Sinica 05 31

Academia Sinica 05 32 General formula Prob. that e is translated into a sentence

Academia Sinica 05 32 General formula Prob. that e is translated into a sentence of length m Prob. that j-th target is aligned with aj -th source word Prob. to produce the word fj at position j

Example c’ est traduit automatiquement NULL it is automatically translated a = (1, 2,

Example c’ est traduit automatiquement NULL it is automatically translated a = (1, 2, 4, 3) Academia Sinica 05 33

IBM model 1 • Simplifications Any length generation is equally probable – a constant

IBM model 1 • Simplifications Any length generation is equally probable – a constant Position alignment is uniformly distributed Context-independent word translation • the model becomes (for one sentence alignment a) Academia Sinica 05 34

Example Model 1 c’ est traduit automatiquement NULL it is automatically translated Academia Sinica

Example Model 1 c’ est traduit automatiquement NULL it is automatically translated Academia Sinica 05 a = (1, 2, 4, 3) 35

Academia Sinica 05 36 Sum up all the alignments • Problem: We want to

Academia Sinica 05 36 Sum up all the alignments • Problem: We want to optimize so as to maximize the likelihood of the given sentence alignments • Solution: Using EM

Parameter estimation 1. 2. An initial value for t(f|e) (f, e are words) Compute

Parameter estimation 1. 2. An initial value for t(f|e) (f, e are words) Compute the count of word alignment e-f in the pair of sentences ( ) (E-step) 3. Maximization (M-step) 4. Loop on 2 -3 Count of f in f Count of e in e Academia Sinica 05 37

Academia Sinica 05 38 Utilization of TM in CLIR • Query: a set of

Academia Sinica 05 38 Utilization of TM in CLIR • Query: a set of source words • Each source word: a set of weighted target words • some filtering: stopwords, prob. threshold, number of translations, … • All the target words query “translation” • Query “translation” with monolingual IR

39 Academia Sinica 05 How effective is this approach? (with the Hansard model) F-E

39 Academia Sinica 05 How effective is this approach? (with the Hansard model) F-E (Trec 6) F-E (Trec 7) E-F (trec 6) E-F (Trec 7) Monolingual 0. 2865 0. 3202 0. 3686 0. 2764 Dict. 0. 1707 (59. 0%) 0. 1701 (53. 1%) 0. 2305 (62. 5%) 0. 1352 (48. 9%) Systran 0. 3098 (107. 0%) 0. 3293 (102. 8) 0. 2727 (74. 0%) 0. 2327 (84. 2%) Hansard TM 0. 2166 (74. 8%) 0. 3124 (97. 6%) 0. 2501 (67. 9%) 0. 2587 (93. 6%) Hansard TM+ dict. 0. 2560 (88. 4%) 0. 3245 (101. 3%) 0. 3053 (82. 8%) 0. 2649 (95. 8%)

Academia Sinica 05 40 Problem of parallel texts • Only a few large parallel

Academia Sinica 05 40 Problem of parallel texts • Only a few large parallel corpora • e. g. Canadian Hansards, EU parliament, Hong Kong Hansards, UN documents, … • Many languages are not covered • Is it possible to extract parallel texts from the Web? • STRANDS • PTMiner

Academia Sinica 05 41 An example of “parallel” pages http: //www. iro. umontreal. ca/index.

Academia Sinica 05 41 An example of “parallel” pages http: //www. iro. umontreal. ca/index. html http: //www. iro. umontreal. ca/index-english. html

Academia Sinica 05 42 STRANDS [Resnik 98] • Assumption: If - A Web page

Academia Sinica 05 42 STRANDS [Resnik 98] • Assumption: If - A Web page contains 2 pointers - The anchor text of each pointer identifies a language Then The two pages referenced are “parallel” French text English text

Academia Sinica 05 43 PTMiner (Nie & Chen 1999) • Candidate Site Selection By

Academia Sinica 05 43 PTMiner (Nie & Chen 1999) • Candidate Site Selection By sending queries to Alta. Vista, find the Web sites that may contain parallel text. • File Name Fetching For each site, fetching all the file names that are indexed by search engines. Use host crawler to thoroughly retrieve file names from each site. • Pair Scanning From the file names fetched, scan for pairs that satisfy the common naming rules.

Academia Sinica 05 44 Candidate Sites Searching • Assumption: A candidate site contains at

Academia Sinica 05 44 Candidate Sites Searching • Assumption: A candidate site contains at least one such Web page referencing another language. • Take advantage of existing search engines (Alta. Vista)

Academia Sinica 05 45 File Name Fetching • Initial set of files (seeds) from

Academia Sinica 05 45 File Name Fetching • Initial set of files (seeds) from a candidate site: host: www. info. gov. hk • Breadth-first exploration from the seeds to discover other documents from the sites

Academia Sinica 05 46 Pair Scanning • Naming examples: index. html v. s. index_f.

Academia Sinica 05 46 Pair Scanning • Naming examples: index. html v. s. index_f. html /english/index. html v. s. /french/index. html • General idea: parallel Web pages = Similar URLs at the difference of a tag identifying a language

Academia Sinica 05 47 Further verification of parallelism • Download files (for verification with

Academia Sinica 05 47 Further verification of parallelism • Download files (for verification with document contents) • Compare file lengths • Check file languages (by an automatic language detector – SILC) • Compare HTML structures • (Sentence alignment)

Academia Sinica 05 48 Mining Results (several years ago) • French-English • Exploration of

Academia Sinica 05 48 Mining Results (several years ago) • French-English • Exploration of 30% of 5, 474 candidate sites • 14, 198 pairs of parallel pages • 135 MB French texts and 118 MB English texts • Chinese-English • 196 candidate sites • 14, 820 pairs of parallel pages • 117. 2 M Chinese texts and 136. 5 M English texts • Several other languages I-E, G-E, D-E, …

49 Academia Sinica 05 CLIR results: F-E (Trec 6) F-E (Trec 7) E-F (Trec

49 Academia Sinica 05 CLIR results: F-E (Trec 6) F-E (Trec 7) E-F (Trec 6) E-F (Trec 7) Monolingual 0. 2865 0. 3202 0. 3686 0. 2764 Systran 0. 3098 (107. 0%) 0. 3293 (102. 8) 0. 2727 (74. 0%) 0. 2327 (84. 2%) Hansard TM 0. 2166 (74. 8%) 0. 3124 (97. 6%) 0. 2501 (67. 9%) 0. 2587 (93. 6%) Web TM 0. 2389 (82. 5%) 0. 3146 (98. 3%) 0. 2504 (67. 9%) 0. 2289 (82. 8%) • Web TM comparable to Hansard TM

CLIR Results: C-E • Chinese: People’s Daily, Xinhua news agency • English: AP C-E

CLIR Results: C-E • Chinese: People’s Daily, Xinhua news agency • English: AP C-E E-C Monolingual 0. 3861 0. 3976 Dictionary (EDict) 0. 1530 (39. 6%) 0. 1427 (35. 9%) TM 0. 1654 (42. 84%) 0. 1591 (40. 02%) TM + Dict 0. 2583 (66. 90%) 0. 2232 (56. 14%) • MT system: E-C: 0. 2001 (50. 3%) C-E: (56 - 70%) Academia Sinica 05 50

51 Other methods – using parallel texts for pseudorelevance feedback • Given a query

51 Other methods – using parallel texts for pseudorelevance feedback • Given a query in F • Find relevant documents in the parallel corpus • Extract keywords from their parallel documents, and consider them as a query translation Query F F Rel. doc. F E Corresponding doc. in E Words in E

52 Other methods - LSI • Monolingual LSI : • Create a latent semantic

52 Other methods - LSI • Monolingual LSI : • Create a latent semantic space • Each dimension represents a combination of initial dimensions (terms, documents) • Comparison of document-query in the new space • Bilingual LSI : • Create a latent semantic space for both languages on a parallel corpus • Concatenate two parallel texts together • Convert terms in both languages into the semantic space • Problems: • The dimensions in the latent space are determined to minimise some representational error – may be different from translational error • Coverage of terms by the parallel corpus • Complexity in creating the semantic space • Effectiveness – usually lower than using a translation model

53 Using a comparable corpus • Comparable: News Articles published in two newspapers on

53 Using a comparable corpus • Comparable: News Articles published in two newspapers on the same day • Estimate cross-lingual similarity (less precise than translation) • Similar methods to co-occurrence analysis • Less effective than using a parallel corpus • To be used only when there is no parallel corpus, or the parallel corpus is not large enough

54 Other problems – unknown words • Proper names (‘Pierre Nadeau’ in Chinese? )

54 Other problems – unknown words • Proper names (‘Pierre Nadeau’ in Chinese? ) • New technical terms (‘web surfing’ in Chinese at the beginning of the web? ) • Solutions • Transliteration • Mining the web

55 Transliteration • Translate a name phonetically • Generate the pronounciation of the name

55 Transliteration • Translate a name phonetically • Generate the pronounciation of the name • Transform the sounds into the target language sounds • Generate the characters to represent the sounds

56 Mining the web - 1 • A site is referred to by several

56 Mining the web - 1 • A site is referred to by several pages with different anchor texts in different languages • Anchor texts as parallel texts • Useful for the translation of organizations (故宫博物馆 - National Museum)

57 Mining the web - 2 • Some "monolingual » texts may contain translations

57 Mining the web - 2 • Some "monolingual » texts may contain translations • �在网上最�的�条,就是�个“Barack Obama”,巴拉克·欧巴�(巴 拉克·奥巴�)。 • �就�生了潜��索引(Latent Semantic Indexing) … • templates: • Source-name (Target-name) • Source-name, Target-name • … • May be used to complete the existing dictionary

58 Other improvement measures • Pre- and post-translation expansion • Query expansion before the

58 Other improvement measures • Pre- and post-translation expansion • Query expansion before the translation • Query expansion after the translation • Fuzzy matching • information - información – informazione • ~cognate • Matching n-grams (e. g. 4 -grams) • Transformation using rules (konvektio -> convection) • Combine translations using different tools

59 Current state • Effectiveness of CLIR • Between European languages ~90 -100% monolingual

59 Current state • Effectiveness of CLIR • Between European languages ~90 -100% monolingual • Between European and Asian languages ~ 80 -100% • A usable quality • One always needs translation of the retrieved documents • The need for CLIR is still limited / Tools for CLIR are limited

60 Remaining problems • Current approaches : • CLIR= translation + monolingual IR •

60 Remaining problems • Current approaches : • CLIR= translation + monolingual IR • The resources and tools are usually developed for MT, not for CLIR • Problem of context • window 7 update -> fenêtre 7 mise à jour • Hints to be used: • window 7 (a frequent context) • window – update -> window • Dependent words do not always form a phrase • Take into account more flexible dependencies (even proximity) • How to train a translation model in such a context? • This is not only a problem in CLIR but also in general IR. • See the lecture on dependency models

61 The future • CLIR≠ translation+ monolingual IR • Translation is a step in

61 The future • CLIR≠ translation+ monolingual IR • Translation is a step in CLIR • For IR • Similar to query expansion • Can use similar approaches to query expansion

Academia Sinica 05 62 Summary • High-quality MT usually offers the best solution •

Academia Sinica 05 62 Summary • High-quality MT usually offers the best solution • Well-trained TM based on parallel texts can match or outperform MT (Kraaij et al. 03) • Dictionary • Simple utilization is not good • Complex approaches improve quality • The performance of CLIR usually lower than monolingual IR (between 50% and 100%) • Filtering noisy parallel corpus is useful • Better translation model = better CLIR effectiveness • Consider compound terms in TM • Ongoing work…