Self-organizing semantic maps and its application to word alignment in Japanese-Chinese parallel

1. Self-organizing semantic maps and its application to word alignment in Japanese-Chinese parallel Advisor ：Dr. Hsu Graduate：Chun Kai Chen Author：Qing Ma, Kyoko Kanzaki, Yujie Zhang, Masaki Murata, Hitoshi Isahara Neural Networks 17 (2004) 1241–1253

2. Outline • Motivation • Objective • Introduction • Self-organizing monolingual semantic maps • Experimental Results • Conclusions • Personal Opinion

3. Motivation • A number of corpus-based statistical approaches have been used to compute word similarity • It is difficult to recognize the relationships between groups or the relationships between words within groups

4. Objective • We need a technique that can map words from a very large lexicon into a small semantic space • A visible representation where words with similar meanings are placed at the same or neighboring points so that the distance between the points represents the semantic similarity in the words • Semantic maps can be automatically constructed with self-organization

5. Introduction • Presents a method of self-organizing monolingual semantic maps for Chinese and Japanese using SOM for specific purpose • To construct semantic maps of nouns from the point of view of the adnominal constituents • Extended to the construction of Japanese–Chinese bilingual semantic maps

6. Self-organizing monolingual semantic maps 。。。。。。 dijis the word similarity • Data coding • Baseline method • Frequency term-weighting method • TFIDF term-weighting method

7. Data coding Word wi can be defined by a set of its co-occurring words as V(wi) is the input to the SOM only reflects the relationships between a pair of words

8. Data coding method • Baseline method • dij：word similarity • ai & aj： are the numbers of co-occurring words of wi and wj • cij：is the number of co-occurring words that both wi and wj have in common • Frequency term-weighting method • TFIDF term-weighting method

9. Table 1 Comparative results for various coding methods and clustering

10. Evaluation methods • Numerical evaluation • precision • recall • F-measure • Intuitive evaluation • our ‘common sense’ • Comparison with other methods • multivariate statistical analyses

11. Experimental Results

13. TFIDF comparison with PCA Fig. 2. Chinese semantic map using principal component analysis Fig. 1. Chinese semantic map based on TFIDF term-weighted coding

14. Table 2 Clustering results with TFIDF term-weighted coding The underlined words are those classified into incorrect areas.

16. Semantic map comparison with PCA Fig. 3. Japanesesemantic map based on the TFIDF term-weighted coding method Fig. 4. Japanese semantic map using principal component analysis

17. Self-organizing bilingual semantic maps(1/2) • When a translation pair of sentences like • Each Japanese word can therefore be automatically aligned to a Chinese word from this map by measuring its distance • If the Chinese word keyi (can) is closest to the Japanese word seta (can), then the Japanese word seta (can) is regarded as being aligned to the Chinese word keyi (can) (Japanese) keiei toppu ga tei seichou jidai teichaku wo jikkan shite iru koto wo ukagawa seta. (Chinese) youci keyi kanchu, zuigao jingyingzhe shengan jingji ren tingliu zai dishu zengzhang shidai. (English) We can see that upper management has realized that the economy is fixed in an eras of slow growth.

18. Self-organizing bilingual semantic maps(2/2) • A small-scale (10 translation pairs) experimental comparison with the baseline method • Comparison with hierarchical clustering and multivariate statistical analysis

19. Data coding(1/2) (Japanese) keiei toppu ga tei seichou jidai teichaku wo jikkan shite iru koto wo ukagawa seta. (Chinese) youci keyi kanchu, zuigao jingyingzhe shengan jingji ren tingliu zai dishu zengzhang shidai. (English) We can see that upper management has realized that the economy is fixed in an eras of slow growth. Ji (i=1,.,m) are Japanese words forming the Japanese sentence Ci (i=1,.,n) are Chinese words forming the translated Chinese sentence

20. is a co-occurring word of Ji is the normalized co-occurrence frequency is a co-occurring word of either or severals of Jj1;.; Jj;ni is the normalized co-occurrence frequency Data coding(2/2)

21. Semantic map comparison with PCA

22. Semantic map comparison with Baseline Table 3 Word alignment result obtained from semantic map Table 4 Baseline word alignment results

23. Conclusions and Future Work • Proposed a method of self-organizing monolingual semantic maps for Japanese and Chinese • Experimental results proved that these maps were generally consistent with our intuition • Comparison demonstrated that the hierarchical clustering technique is inferior to SOM in terms of classifying ability • Furthermore, multivariate statistical analysis such as principal component analysis and factor analysis gave worse results

24. Conclusions and Future Work • An extension to the automaticconstruction of bilingual semantic maps of Japanese and Chinese • Develop an automatic method of transforming both Japanese and Chinese words

25. Personal Opinion • …..