A corpus-centered approach to spoken language translation

A Corpus-Centered Approach to Spoken Language Translation Eiichiro SUMITA, Yasuhiro AKIBA, Takao DO!, Andrew FINCH, Kenji IMAMURA, Michael PAUL, Mitsuo SHIMOHATA, and Taro WATANABE ATR Spoken Language Translation Research Laboratories 2-2-2 Hikaridai, Keihanna Science City, Kyoto 619-0288, JAPAN e iichiro. sumitagatr. co jp Abstract This paper reports the latest performance of com- ponents and features of a project named Corpus- Centered Computation (C'3), which targets a trans- lation technology suitable for spoken language translation. C3 places corpora at the center of the technology. Translation knowledge is extracted from corpora by both EBMT and SMT methods, translation quality is gauged by referring to cor- pora, the best translation among multiple-engine outputs is selected based on corpora and the cor- pora themselves are paraphrased or filtered by automated processes. 1 Introduction Our project, named Corpus-Centered Computation (( 3), proposes solutions for efficiently constructing a high-quality translation subsystem for a speech- to-speech translation system. This paper introduces recent progress in C3. Sec- tions 2 and 3 demonstrate a competition between multiple machine translation systems developed in our project, and Sections 4 and 5 explain the fea- tures that differentiate our project from other cor- pus-based projects. 2 Three Corpus-based MT Systems There are two main strategies in corpus-based ma- chine translation: (i) Example-Based Machine Translation (EBMT; Nagao, 1984; Somers, 1999) and (ii) Statistical Machine Translation (SMT; Brown et al., 1993; Knight, 1997; Ney, 2001; Al- shawi et al., 2000). C3 is developing both tech- nologies in parallel and blending them. In this paper, we introduce three different machine trans- lation systems: Di, HPAT, and SAT. The three MT systems are characterized by dif- ferent translation units. D3, HPAT, and SAT use sentences, phrases, and words, respectively. D3 (Sentence-based EBMT): It retrieves the most similar example by DP-matching of the input and example sentences and adjusts the gap between the input and the retrieved example by using dictionar- ies. (Sumita 2001) HPAT (Phrase-based EBMT): Based on phrase- aligned bilingual trees, transfer patterns are gener- ated. According to the patterns, the source phrase structure is obtained and converted to generate tar- get sentences (Imamura 2002) SAT (Word-based SMT): Watanabe et al. (2002b) implemented SAT dealing with Japanese and English on top of a word-based SMT frame- work (Brown et al. 1993). 3 Competition on the Same Corpus 3.1 Resources In our competitive evaluation of the MT systems, we used the BTEC corpus ', which is a collection of Japanese sentences and their English transla- tions typically found in phrasebooks for tourists. The size is about 150 thousand sentence pairs. A quality evaluation was done using a test set consisting of 345 sentences selected randomly from the above corpus, and the remaining sentences were used for learning and verification. For each source sentence in the test set, 16 reference translations were prepared by 5 bilingual translators. 1 BTEC was called BE in the paper (Takezawa et al., 2002). 171 We used bilingual dictionaries and thesauri of about fifty thousand words for the travel domain. 3.2 Evaluation …

Content truncated to 500 words