1. Multilingual Synchronization focusing on Wikipedia 2011-03-17

2. Introduction Wikipedia: Multilingual encyclopedia – Supports over 270 languages English, German, Spanish, French, Chinese, Arabic, … Allows cross-lingual navigation with inter-language link – Inter-language links: hyperlinks from any page in one Wikipedia language edition to one or more nearly equivalent or exactly equivalent pages in another Wikipedia language editions – Different quantity of data on each languages Wikipedia other language editions often suffer from lack of information compared to the English version – Multilingual stat on Feb. 2011 » English: 3.5 million articles (Most dominant) » French: 1 million articles (3rd) » Korean: 156,290 articles (22nd)

3. Goal of M-Sync Multilingual Synchronization – Synchronizing contents of Wikipedia from multiple different languages Linking among multiple language contents Combining them to synthesis – The various Wikipedia editions from different languages can offer more precise and detailed information based on different intentions/backgrounds/cultures can fill the gap between different languages and to acquire the integrated knowledge

4. Multilingual Resource Synthesis Construction Association Network from multiple language hyperlink structures – based on occurrence data A Web page makes a direct reference to another Web page via a hyperlink Definitions – Association A relation between any two words or concepts with a strength – Association Network Nodes and edges – Node: entity (concept or a named entity) – Edge: link between entities with weight(the strength of association)

5. Motivating Example BaldnessHair

6. Motivating Example BaldnessHair 탈모증털

7. Motivating Example BaldnessHair 탈모증털 Increase Strength

8. Motivating Example BaldnessHair 탈모증털 Hamilton- Norwood_scale

9. Motivating Example BaldnessHair 탈모증털 Hamilton- Norwood_scale Link Prediction or Keyword Recommendation

10. Multilingual Resource Synthesis: Construction of Association Network Hypothesis – X is associated with Y in L 1  X’ should be associated with Y’ in L 2 Where Y’ is a corresponding term to Y in different language – Assumption » Inter-language links are accurate links to connect two pages about the same entity or concept in different languages Where X’ is a translating term to X in different language – X is associated with Y according to its strength synthesis based on respective occ-score

11. Proposed System Workflow Association set for each English word Link Extraction English Documents Association set for each Korean word Link Extraction Korean Documents Association set for each Chinese word Link Extraction Chinese Documents Calculation of Correlations Association for each pair of English and Korean words Selection of highly Associated pairs of words Calculation of Correlations Association for each pair of Korean and Chinese words Calculation of Correlations Bilingual Dictionary

12. Example of Association Set Association set of “Baldness” alopecia areata adult androgenic alopecia 원형 탈모증 남성형 탈모증 머리카락 營養 頭髮 荷爾蒙 心理

13. Preprocessing: Selecting Target Source languages(5) – English, Spanish, French, Chinese, Korean Extracting target pages with a 5-clique by inter-language links – Assumption: Pages founded in all 5 languages are key pages and the target to sync Enforcing consistency of a link path – If a path from X(L 1 ) to X’(L 2 ) founded once, its inverse path (X’, X) is automatically added to the output 13 en:Badminton es:Bádmintonfr:Badminton zh: 羽毛球 ko: 배드민턴 A subset of UN official languages

14. Link types of Wikipedia internal links to other pages in the wiki – Syntax usage: [[Main Page]] external links to other websites interwiki links to other websites registered to the wiki in advance – Unlike internal links, interwiki links do not use page existence detection – Syntax usage: [[wikipedia:Sunflower]] Interlanguage links to other websites registered as other language versions of the wiki

15. Experiment Link Extraction – Input: 40,000 pages in each languages – Output 5,453,959 links en 3,965,279 links fr 3,368,949 links es 2,397,959 links zh 1,347,811 links ko Association Pairs

16. Experiment: respective mode Compute strength of Association: LF-IDF – is motivated by TF-IDF heuristic

17. Experiment: synthesis mode M: Simple average NM: Average with standard deviation Demo – http://nlplab.kaist.ac.kr/~kekeeo/term http://nlplab.kaist.ac.kr/~kekeeo/term

18. Program Running Environment Link extractor – X seconds / 1 page LFIDF – 1 hours 40,000 page * 5 language – 5,453,959 links en – 3,368,949 links es – 3,965,279 links fr – 1,347,811 links ko – 2,397,959 links zh Multilingual LFIDF – 2 hours Translating – X minutes Aggregating links from multiple sources – 1.5 hours Computing multi weight

19. Example of hyperlinks Example links(out-going) of Seattle: – “northwestern United States” – “Washington” – “Lake Washington” – “Michael McGinn (mayor)” – …

20. Extracting correlated-terms from hypertexts (out-going links) AAA BBB CCC DDD AAA BBB CCC DDD AAA BBB CCC DDD AAA BBB CCC DDD AAA BBB CCC DDD Correlated TermSet 123 language1 language2 language3 language4language5 Correlated TermSet 123, AAA Unified TermSet Translating

21. Extracting correlated-terms from hypertexts (in-coming links) Correlated TermSet 123 language1 language2 language3 language4language5 Correlated TermSet 123, AAA Unified TermSet Translating 123 BBB AAA CCC DDD 123 BBB AAA CCC DDD 123 BBB AAA CCC DDD 123 BBB AAA CCC DDD 123 BBB AAA CCC DDD

22. Translating terms Method – Wikipedia dictionary-based We have collected the cross-lingual term pars to build bilingual word pairs – 4 dictionaries are available: EN-KO, ES-KO, FR-KO, ZH-KO Weakness – Lack of vocabularies – Google translation API-base Weakness – Terms(keywords) are too short to solve the word sense disambiguation using MT

23. Computing weights of co-relatedness using multilingual topic synthesis Weight of correlations – Baseline: frequency-based method – However, different Wikipedia has different viewpoints and concerns – We should give different weight of synthesized correlated term sets according to different lingual usages and frequencies Our proposed solution: – analyzing the topical distributions on each languages and – Computing weights of correlated terms by each topics’ interest – Approach » LDA-based topic distribution using links » Align cross-lingual topic clusters Intersection: common topic Difference: unique topic

24. Evaluation What – Comparison with Discovered new correlated terms without topical synthesis – Simple union approach Discovered new correlated terms with topical synthesis – Ranked union using calculating the strength of relatedness How – Public measures Co-occurrence Mutual information Normalized Google distance – Wikipedia oriented measures Comparison with the featured articles Comparison with the temporal manner

25. Comparison with featured articles Featured articles: – are considered to be the best articles in Wikipedia, as determined by Wikipedia's editors AAA BBB CCC DDD EEE FFF GGG Featured article languageX AAA BBB CCC DDD AAA BBB CCC DDD AAA BBB CCC DDD AAA BBB CCC DDD AAA BBB CCC DDD Correlated TermSet 123 language1 language2 language3 language4language5 Correlated TermSet 123, AAA Unified TermSet Translating Correlated TermSet

26. Comparison with temporal manner AAA BBB CCC DDD EEE FFF GGG Article at t n language1 AAA BBB CCC DDD AAA BBB CCC DDD AAA BBB CCC DDD AAA BBB CCC DDD AAA BBB CCC DDD Correlated TermSet 123 in t1 123 in t n language1 language2 language3 language4language5 Correlated TermSet 123, AAA Unified TermSet Translating Correlated TermSet

27. Contributions To support the seed data (seed keywords) to complete articles in a multilingual manner, or to guide users in generating new articles in Wikipedia To find unknown correlated words using various sources