ATR-SLT System for SENSEVAL-2 Japanese Translation Task

ATR-SLT System for SENSEVAL-2 Japanese Translation Task Tadashi K urnano, Hideki Kashioka and Hideki Tanaka ATR Spoken Language Translation Research Laboratories 2-2-2 Hikaridai Seika-cho Soraku-gun Kyoto 619-0288 JAPAN {tadashi.kumano, hideki.kashioka, hideki.tanaka}@atr;co.jp Abstract We propose a translation selection system based on the vector space model. When each translation candidate of a word is given as a pair of expressions containing the word and its translation, selecting the transla- tion of the word can be considered equivalent to selecting the expression having the most similar context among candidate expressions. The pro- posed method expresses the context information in "context vectors" constructed from content words co-occurring with the target word. Con- text vectors represent detailed information com- posed of lexical attributes( word forms, semantic codes, etc.) and syntactic relations (syntactic dependency, etc.) of the co-occurring words. We tested the proposed method with the SENSEVAL-2 Japanese translation task. Preci- sion/recall was 45.8% to the gold standard m the experiment with the evaluation set. 1 Introduction The SENSEVAL-2 Japanese translation task de- fines a sense of a Japanese word as an English translation. The same Japanese word in differ- ent contexts may have different English trans- lations; therefore, translation ambiguity arises. Translation Memory (henceforth TM) defin- ing word senses were given to the task partic- ipants. Each target word has translation pairs of Japanese and English expressions as word sense candidates1. The target word is marked in the Japanese expression, but the correspond- ing part is unspecified in the English expression. Hence, selecting the most appropriate transla- tion of the target Japanese word in the evalua- tion expression can be considered to be equiv- alent to selecting the expression with the most similar context in the TM. This is equivalent to the word sense disambiguation problem in a single language. 1Each target word has 21.6 pairs on average. 95 Generally, word sense disambiguation uses context information, such as the frequency of words that co-occur with the target word. The context information is learned from the correctly-annotated training corpora. However, no training corpus was given for the task and the given TM had shorter contexts because the TM expressions were rather incomplete. There- fore, instead of learning the co-occurring words with the target word from the training corpora, we extract detailed information from the TM expressions as context information. We utilize the information of co-occurring words with the target word (context words) as shown below. • lexical attributes (word form, part-of- speech, semantic codes on thesaurus, etc.) • syntactic relations to the target word (de- pendency relation, etc.) We employed the vector space model, which is used for text retrieval (Salton and McGill, 1983) to calculate the similarity between the context word information of evaluation expressions and those of the TM. The detailed context informa- tion are expressed as "context vectors." We use cosine values between context vectors as a mea- sure of similarity. In this paper, we will explain first how to con- struct "context vectors," and then show the ac- curacy of the selection …

Content truncated to 500 words