Word Translation Based on Machine Learning Models Using Translation Memory and Corpora

Word Translation Based on Machine Learning Models Using Translation Memory and Corpora Kiyotaka Uchimotot, Satoshi Sekinet, Masaki Muratat, and Hitoshi Isaharat tCommunications Research Laboratory 2-2-2, Hikari-dai, Seika-cho, Soraku-gun, Kyoto, 619-0289 Japan {uchimoto, murata, isahara}©crl.go.jp :j:New York University 715 Broadway, 7th floor New York, NY 10003, USA sekine@cs.nyu.edu Abstract SENSEVAL-2 was held in Spring, 2001. It con- sisted of several tasks in various languages. In this paper, we describe our system used for one of these tasks: the Japanese translation task. With an accuracy of 63.4%, our system was the third best system in the contest among nine sys- tems developed by seven groups. 1 Introduction In the Japanese translation task, the senses of a word were defined in terms of the word's trans- lations. Given an input sentence and a target word in the sentence, our system first estimates the similarity between the input sentence and parallel example sets called "Translation Mem- ory". It then selects an appropriate transla- tion of the target word by using the example set with the highest similarity. The similarity is calculated using dynamic programming and a machine learning model, which assesses the similarity based on the similarity of a string, words to the left and to the right of the target word in the input sentence, content words in the input sentence and their translations, and co-occurrence of content words in bilingual and monolingual corpora in English and Japanese. 2 Japanese Translation Task In general, the definition of word senses depends on the goal of a task. The goal of the Japanese translation task is word selection in translation, where the target language is English. Therefore, word senses are defined as translations (trans- lated words/ phrases). Before the contest, a Japanese-English par- allel phrase/sentence set (Translation Memory, henceforth referred to as TM) was given to the participants as training data. In the TM, for each Japanese headword, there was a set of pairs 155 of a Japanese expression including a headword and an English translation of the expression. We call these pairs examples. Some of the ex- amples are shown in Figure 1. <entry id= "1" headword= ".it,!j" > <sense id="l-1"> <jexpression> ffl;~: .it,!j'T .Q </jexpression> <eexpression>to feel constrained for one's mother</ eexpression> </sense> <sense id= "1-2" > <jexpression> ffl:"'-O).it,!j </jexpression> <eexpression>constraint toward one's mother</ eexpression> <transmemo> UC</transmemo> </sense> <sense id="1-3"> <jexpression> Mk~ ~ .it,!j l-c b G -5 </jexpression> <eexpression>to request to refrain from donation</ eexpression> </sense> </entry> Figure 1: Examples in TM. In the formal test (contest), the participants were given a set of texts each of which was marked by a target word. For each target word, the participants were required to submit either a sense id of the example (the number assigned to each example in the TM), which can be used to translate the target word, or a translation of the target word. In the latter case, a translation of the word itself, a translation of a sequence of words including the target word, or a transla- tion of …

Content truncated to 500 words