Improving Calculation of Contextual Similarity for Constructing a Bilingual Dictionary via a Third Language

International Joint Conference on Natural Language Processing, pages 1057–1061, Nagoya, Japan, 14-18 October 2013. Improving Calculation of Contextual Similarity for Constructing a Bilingual Dictionary via a Third Language Takashi Tsunakawa Yosuke Yamamoto Hiroyuki Kaji Graduate School of Informatics, Shizuoka University 3-5-1 Johoku, Naka-ku, Hamamatsu, Shizuoka 432-8011, Japan {tuna, kaji}@inf.shizuoka.ac.jp Abstract A novel method is proposed for measuring contextual similarity by “weighted overlap- ping ratio (WOR)” to construct a bilingual dic- tionary of a new language pair from two bilin- gual dictionaries sharing one language. The WOR alleviates the effect of a noisy seed dic- tionary resulting from merger of two bilingual dictionaries via a third language. Combined use of two word-association measures for ex- tracting contexts from corpora is also proposed to compensate their weaknesses. Experiments on constructing a Japanese-Chinese dictionary via English show that the proposed method outperforms the conventional method based on cosine similarity. 1 Introduction With the growth of communication via the Inter- net, people have more chance to access docu- ments written in various languages. The number of Internet users in the Arabic, Russian, and Chi- nese languages have increased at least tenfold times in the last decade. It was sufficient in the past to bridge the gap between English and an- other language by using bilingual resources and services. For directly accessing Web contents written in various languages, however, multilin- gual dictionaries, translation, and information retrieval are required. The present study focuses on a so-called trian- gulation approach for constructing a bilingual dictionary by merging two bilingual dictionaries sharing one language. For example, if Chinese- to-English and English-to-Arabic dictionaries are available, a Chinese-to-Arabic dictionary can be derived by collecting pairs of Chinese and Ara- bic terms (hereafter, “term pairs”) that have common English translations. A serious problem with this approach, however, is how to filter out false term pairs, caused by polysemy of English terms. To validate each term pair as translations, we calculate the context similarity between the terms on the basis of the distributional hypothe- sis (Harris, 1954). In triangulation approach, calculation of the context similarity in different languages is re- quired. Previous studies have calculated context similarities by context vector projection onto an- other language by using a seed bilingual diction- ary. Though this approach is effective, transla- tion perplexity caused by context vector projec- tion may cause negative effects described in Sec- tion 3.3. Instead of projection, our proposed method avoids this problem by using a “weighted overlapping ratio,” which directly maps words in context vectors in different lan- guages. 2 Related work Tanaka and Umemura (1994) proposed a triangu- lation method of constructing a bilingual diction- ary. Their method has been augmented by using semantic classes (Bond et al., 2001) and parts of speech and cognates (Zhang et al., 2005). Several methods of constructing a bilingual dictionary from contextual similarity have been proposed (Rapp, 1995; Kaji and Aizono, 1996; Tanaka and Iwasaki, 1996; Fung and Yee, 1998; Rapp, 1999; Sammer and Soderland, 2007). Most of them are based on context …

Content truncated to 500 words