Acquiring Compound Word Translations both Automatically and Dynamically

Acquiring Compound Word Translations Both Automatically and Dynamically Yujie Zhang Keihanna Human Info-Communications Research Center, National Institute of Information and Communications Technology 3-5 Hikaridai, Seika-cho, Soraku-gun, Kyoto 619-0289, Japan yujie@nict.go.jp Hitoshi Isahara Keihanna Human Info-Communications Research Center, National Institute of Information and Communications Technology 3-5 Hikaridai, Seika-cho, Soraku-gun, Kyoto 619-0289, Japan isahara@nict.go.jp Abstract This paper addresses the problem of compound word translation and proposes the approaches to acquiring translations. The proposed approaches focus on exploring web data and utilizing English translations to link words of the source language and the correspondences in the target language. The paper uses Japanese-Chinese language pairs for the sake of illustration and shows initial experimental results. The proposed method is language-independent and therefore can be applied to other language pairs. 1 Introduction The translation of compound words is an important problem related to the construction of translation dictionaries. On the one hand, the published paper bilingual dictionaries usually collect translations for simple words but not for compound words. On the other hand, new compound words emerge every day. It is reported that most compound words are nominal words. In natural language comprehension, entities or concepts to be recognized are usually described by nouns or compound nouns. Therefore, acquiring translations for compound nouns is extremely important in machine translation and cross-language information retrieval [Cao and Li, 2002; Nakagawa, 2001]. In resolving this problem, we propose a self-learning mechanism that can dynamically and automatically acquire compound words translations from the web and corpora. Here we take Japanese-Chinese language pairs as an instance for illustration and report initial experiment results. The proposed method can also be applied to other language pairs since it is language-independent. 2 Overview of the self-learning mechanism The self-learning mechanism aims at automatically and dynamically expanding the translation dictionary of compound word. It is implemented in two parts: collecting new compound words (collection part) and then acquiring their translations (acquisition part). In the initialization phase, we design a compound word list for the already known words and extract their composition patterns. When the web data or new corpus data are input into the collection part, the data are matched with the extracted compound word composition patterns. The word strings that matched the patterns are then added to the compound word list if they are not already on the list. When new compound words are added to the list, the acquisition part will be activated to acquire possible translation candidates by using the three approaches described in the next section. The most likely translations will be added to the dictionary. In this way, the translation dictionary of compound words can be expanded automatically and dynamically once new data is available. In the case of Japanese-Chinese language pairs, we collected compound word list from the EDR Japanese-English dictionary [NICT, 2002] and obtained the composition patterns by using the Japanese PACLIC 18, December 8th-10th, 2004, Waseda University, Tokyo - 181 - morphological analyzer ChaSen [Matsumoto et al., 2000]. As a result, 178,854 entries were obtained, among which …

Content truncated to 500 words