A Brief Survey on Cross-language Information Retrieval (CLIR) - Text Retrieval Perspective by Ying Alvarado ( ) CSE 8337 Lecturer : Dr. Margaret Dunham April 26, 2007
2 Outline Introduction Concept Why important Approach CLIR problems Resource Approaches Example Techniques A CLIR application system CLIR effectiveness CLIR future tasks CLIR communities References
3 Cross Language IR Definition: Users enter their query in one language and the system retrieves relevant documents in other languages. For example, a user may pose their query in English but retrieve relevant documents written in French. Example CLIR applications Cross-Language retrieval from texts Cross-Language retrieval from audio and images [1] Wikipedia, [2] Paul Clough, Bridging the language gap: making digital collections available to a multilingual society, presentation, 2005 In this presentation, we focus on text IR only!
4 Monolingual IR: Documents and user requests in the same language Documents (L 1 ) IR system Request (L 1 ) Results (L 1 ) Monolingual vs. Bilingual vs. Multilingual [2] Paul Clough, Bridging the language gap: making digital collections available to a multilingual society, presentation, 2005 Cross-language IR: Documents and user requests are in different languages (bilingual IR) Documents (L 2 ) Cross-language IR (CLIR) system Request (L 1 )Results(L 2 ) Source language Target language
5 Documents (L 3 ) Multilingual IR (MLIR) system Request (L?)Results (L 2, L 3 or L 4 ) Documents (L 2 ) Documents (L 4 ) e.g. the Web Multilingual IR: Documents in collection in different languages, search requests in any language Monolingual vs. Bilingual vs. Multilingual (con.)
6 Why CLIR? [3] Internet World Stats, TOP TEN LANGUAGES IN THE INTERNET % of all Internet Users Internet Users by Language Internet Penetration by Language Internet Growth for Language ( ) 2007 Estimate World Population for the Language English29.5 %328,666, %139.6 %1,143,218,916 Chinese14.3 %159,001, %392.2 %1,351,737,925 Spanish8.0 %88,920, %260.3 %439,284,783 Japanese7.7 %86,300, %83.3 %128,646,345 German5.3 %58,711, %113.2 %96,025,053 French5.0 %55,521, %355.2 %387,820,873 Portuguese3.6 %40,216, %430.8 %234,099,347 Korean3.1 %34,120, %79.2 %74,811,368 Italian2.8 %30,763, %133.1 %59,546,696 Arabic2.6 %28,540, %931.8 %340,548,157 TOP TEN LANGUAGES81.7 %910,762, %181.4 %4,255,739,462 Rest of World Languages18.3 %203,511, %444.5 %2,318,926,955 WORLD TOTAL100.0 %1,114,274, %208.7 %6,574,666,417 Top Ten Languages Used in the Web ( Number of Internet Users by Language ) Mar. 10, 2007
7 Why CLIR? (con.) [4] D.W. Oard, A Survey of Multilingual Text Retrieval. Computer Science Technical Report Series; Vol. CS-TR A collection may contains documents in many different languages, e.g. the Web. It would be impractical to form a query in each language. The documents may be expressed in more than one languages. For example, Technical documents in which English jargon appears intermixed with narrative text in another language. Academic works which cite the titles of documents in different languages. The user is not sufficiently fluent to express a query in a language, but is able to make use of the documents that are identified. The user is monolingual and wants to query in their native language. Because he can judge relevance even if results not translated have access to document translation [2] Paul Clough, Bridging the language gap: making digital collections available to a multilingual society, presentation, 2005
8 Handling non-ASCII character sets Untranslatable search keys (OOV): e.g. compound words, proper names, special terms Multi-word concepts, e.g. phrases and idioms Ambiguity, e.g. Homonymy and polysemy Word Inflections, e.g. plurals and gender CLIR problems [5] Ari Pirkola, et al. Dictionary-Based Cross-Language Information Retrieval_ Problems, Methods, and Research Findings. Information Retrieval, Vol [2] Paul Clough, Bridging the language gap: making digital collections available to a multilingual society, presentation, 2005
9 Ontology Representation of concepts and relationships Thesaurus it more commonly means a listing of words with similar, related, or opposite meanings It does not include the definition of words Bilingual dictionary a list of words together with additional word-specific information. Bilingual controlled vocabulary carefully selected list of words and phrases, which are used to tag units of information (document or work) so that they may be more easily retrieved by a search Corpora The document collection itself Resources for Translation [6] Jimmy Lin, Cross-Language and Multimedia Information Retrieval. Slides for LBSC 796/INFM 718R [4] D.W. Oard, A Survey of Multilingual Text Retrieval. Computer Science Technical Report Series; Vol. CS-TR [1] Wikipedia. Related pages. [7] Metamodel.com. What are the differences between a vocabulary, a taxonomy, a thesaurus, an ontology, and a meta-model?
10 An example of controlled vocabulary [14] Boxes and Arrows, The hierarchical relationships Women’s Pants: BT Pants NT Casual Pants NT Dress Pants NT Sports Pants The equivalence relationship
11 What to translate? Document translation Text translation E.g., translate entire document collection into English → search collection in English Vector translation Query translation E.g., translate English query into Chinese query → search Chinese document collection [6] Jimmy Lin, Cross-Language and Multimedia Information Retrieval. Slides for LBSC 796/INFM 718R. 2006
12 Tradeoffs Document Translation Documents can be translate and stored offline Dependent on high quality automatic machine translation (MT) system Does not easily deal with changing document sets Query Translation Often easier Disambiguation of query terms may be difficult with short queries [6] Jimmy Lin, Cross-Language and Multimedia Information Retrieval. Slides for LBSC 796/INFM 718R [4] D.W. Oard, A Survey of Multilingual Text Retrieval. Computer Science Technical Report Series; Vol. CS-TR
13 Approaches to query translation Knowledge-based: Several aspects of domain knowledge is manually encoded in to a lexicon. Ontology-based (concept driven) Thesaurus-based Dictionary-based Expensive to construct lexicons; Lag behind the common use of terminology. Corpus-based: directly exploit statistical information about term usage in a corpora; automatically construct lexicon. Parallel corpora: document pairs, sentence pairs, term pairs Comparable corpora: document pairs, similar content Unaligned corpora: documents from the same domain, not translations of one another, not linked in any other way [8] Miguel E. Ruiz, CLIR. Slides for school seminars [9] Rada Mihalcea, Information Retrieval and Web Search. Class slides [4] D.W. Oard, A Survey of Multilingual Text Retrieval. Computer Science Technical Report Series; Vol. CS-TR
14 Applying monolingual IR techniques Query expansion Relevance feedback Stemming Latent semantic analysis Parsing Part of speech tagging …… [4] D.W. Oard, A Survey of Multilingual Text Retrieval. Computer Science Technical Report Series; Vol. CS-TR
15 Multilingual Thesauri Three construction techniques Build it from scratch Translate an existing thesaurus Merge monolingual thesauri For example EuroWordNet 7 languages Built from existing lexical resources Has the same structure as Princeton WordNet [8] Miguel E. Ruiz, CLIR. Slides for school seminars [9] Rada Mihalcea, Information Retrieval and Web Search. Class slides. 2007
16 Pseudo-Relevance Feedback Also call Blind feedback Assume that the top n documents in the result set actually are relevant. Enter query terms in French Find top French documents in parallel corpus Construct a query from English translations Perform a monolingual free text search Top ranked French Documents French Text Retrieval System AltaVista French Query Terms English Translations English Web Pages Parallel Corpus [9] Rada Mihalcea, Information Retrieval and Web Search. Class slides. 2007
17 Different level alignment in parallel corpora Document alignment Already exists Collected from existing corpora Examine document external features Examine document internal features Sentence alignment Easily constructed from aligned documents Match pattern of relative sentence lengths Good first step for term alignment Term alignment Using co-occurrence-based translation [9] Rada Mihalcea, Information Retrieval and Web Search. Class slides. 2007
18 Example of term alignment CSE8337 是一门关于信息存储和检索的课程。 CSE8337 is a class about information storage and retrieval.
19 Co-occurrence-based translation Align terms using co-occurrence statistics assumed that the correct translations of query terms tend to co-occur in target language documents How often do a term pair occur in sentence pairs? Weighted by relative position in the sentences Retain term pairs that occur unusually often [9] Rada Mihalcea, Information Retrieval and Web Search. Class slides. 2007
20 Exploiting Unaligned Corpora Example approach: category-based translation Extract a large number of terms from unaligned coprora of the first and second languages Assign a category to each extracted term by accessing monolingual thesauri of the first and second languages Estimate category-to-category translation probabilities Estimate term-to-term translation probabilities using said category-to-category translation probabilities [15] David Hull, Terminology translation for unaligned comparable corpora using category based translation probabilities. United States Patent Filing date: Dec 18, Issue date: Apr 26, 2005
21 In Summary Term-aligned Sentence-aligned Document-aligned Unaligned Parallel Comparable Knowledge-based Corpus-based Controlled Vocabulary Free Text Cross-Language Text Retrieval Query Translation Document Translation Text Translation Vector Translation Ontology-based Dictionary-based Thesaurus-based [8] Miguel E. Ruiz, CLIR. Slides for school seminars. 2001
22 An experimental system [10] Jiang Chen, et al. Automatic construction of parallel English-Chinese corpus for cross-language information retrieval. Proceedings of the sixth conference on Applied natural language processin Automatic construction of parallel English-Chinese corpus for CLIR A parallel text mining system- PTMiner Finds parallel text from web Parallel Text Mining Algorithm 1. Search for candidate sites - Using existing Web search engines, search for the candidate sites that may contain parallel pages; (by using text anchor) 2. File name fetching - For each candidate site, fetch the URLs of Web pages that are indexed by the search engines; 3. Host crawling - Starting from the URLs collected in the previous step, search through each candidate site separately for more URLs; 4. Pair scan - From the obtained URLs of each site, scan for possible parallel pairs; (by analyzing document external features) 5. Download and verifying - Download the parallel pages, determine file size, language and character set, text length, HTML structure, and filter out non- parallel pairs.
23 The workflow of the mining process Sample anchor texts: “english version” [“in english”, ……] Sample document external features: “file-ch.html” vs. “file-en.html” “…/chinese/…/file.html” vs. “…/english/…file.html” Sample document internal features: Character set, HTML structure [10] Jiang Chen, et al. Automatic construction of parallel English-Chinese corpus for cross-language information retrieval. Proceedings of the sixth conference on Applied natural language processin. 2000
24 An alignment example [10] Jiang Chen, et al. Automatic construction of parallel English-Chinese corpus for cross-language information retrieval. Proceedings of the sixth conference on Applied natural language processin. 2000
25 Part of the lexicons t: ture f: false [10] Jiang Chen, et al. Automatic construction of parallel English-Chinese corpus for cross-language information retrieval. Proceedings of the sixth conference on Applied natural language processin Encoding scheme transformation (for Chinese) Sentence level segmentation Chinese word segmentation English expression extraction SILC: language and encoding identification system Other techniques and tools used:
26 Results pairs of texts (lexicon) C-E has a precision of 77% E-C has a precision of 81.5% CLIR results Test corpus: TREC5 and TREC6 Chinese track [10] Jiang Chen, et al. Automatic construction of parallel English-Chinese corpus for cross-language information retrieval. Proceedings of the sixth conference on Applied natural language processin. 2000
27 Does CLIR work? Best systems at TREC-6 (1997): English-French: 49% of highest French monolingual English-German: 64% of highest German monolingual Best systems at CLEF (2002): English-French: 83% of highest French monolingual English-German: 86% of highest German monolingual Best systems at CLEF (2006): English-French: 93.82% of best French monolingual English-Portuguese: 90.91% of best Portuguese monolingual [2]Paul Clough, Bridging the language gap: making digital collections available to a multilingual society, presentation, 2005 [16] Giorgio M. Di Nunzio, CLEF 2006: Ad Hoc Track Overview. 2006
28 Future tasks [11] D.W. Oard, When You Come to a Fork in the Road, Take It: Multiple Futures for CLIR Research. SIGIR 2002 CLIR [12] Fredric Gey, et al, CROSS LANGUAGE INFORMATION RETRIEVAL: A RESEARCH ROADMAP. SIGIR 2002 CLIR Extend study scope: Web pages, medical literature, USENET newsgroup articles, records of legislative and legal proceedings … Lower cost, improve efficiency Pay more attention on indexing-time optimizations to improve query-time efficiency Consider user ’ s perspective Improve the utility of ranked lists Define suitable criteria for the construction of a valid multilingual Web corpus Get resources for resource-poor languages
29 CLIR Communities TREC Cross Language Track currently focuses on the Arabic language, Cross-Language Evaluation Forum (CLEF) – a spinoff from TREC - covering many European languages, NTCIR Asian Language Evaluation (covering Chinese, Japanese and Korean). [12] Fredric Gey, et al, CROSS LANGUAGE INFORMATION RETRIEVAL: A RESEARCH ROADMAP. SIGIR 2002 CLIR
30 In CLEF 2006, eight tracks were offered to evaluate the performance of systems: multilingual document retrieval on news collections (Ad-hoc) cross-language structured scientific data (Domain-specific) interactive cross-language retrieval multiple language question answering cross-language retrieval on image collections cross-language speech retrieval multilingual web retrieval cross-language geographic retrieval. CLEF [13] Carol Peters, Cross-Language Evaluation Forum - CLEF D-Lib Magazine October 2006
31 References [1] Wikipedia, [2] Paul Clough, Bridging the language gap: making digital collections available to a multilingual society, presentation, 2005 [3] Internet World Stats, [4] D.W. Oard, A Survey of Multilingual Text Retrieval. Computer Science Technical Report Series; Vol. CS-TR [6] Jimmy Lin, Cross-Language and Multimedia Information Retrieval. Slides for LBSC 796/INFM 718R [8] Miguel E. Ruiz, CLIR. Slides for school seminars [5] Ari Pirkola, et al. Dictionary_Based Cross-Language Information Retrieval_ Problems, Methods, and Research Findings. Information Retrieval, Vol [7] Metamodel.com. What are the differences between a vocabulary, a taxonomy, a thesaurus, an ontology, and a meta-model? [9] Rada Mihalcea, Information Retrieval and Web Search. Class slides [10] Jiang Chen, et al. Automatic construction of parallel English-Chinese corpus for cross-language information retrieval. Proceedings of the sixth conference on Applied natural language processin [11] D.W. Oard, When You Come to a Fork in the Road, Take It: Multiple Futures for CLIR Research. SIGIR 2002 CLIR [12] Fredric Gey, et al, CROSS LANGUAGE INFORMATION RETRIEVAL: A RESEARCH ROADMAP. SIGIR 2002 CLIR [13] Carol Peters, Cross-Language Evaluation Forum - CLEF D-Lib Magazine October 2006 [14] Boxes and Arrows, [15] David Hull, Terminology translation for unaligned comparable corpora using category based translation probabilities. United States Patent Filing date: Dec 18, Issue date: Apr 26, 2005 [16] Giorgio M. Di Nunzio, CLEF 2006: Ad Hoc Track Overview. 2006
32 Thank you!