EBMT, SMT, hybrid and more: ATR spoken language translation system

EBMT, SMT, Hybrid and More: ATR Spoken Language Translation System Eiichiro SUMITA, Yasuhiro AKIBA, Takao DOI, Andrew FINCH, Kenji IMAMURA, Hideo OKUMA, Michael PAUL, Mitsuo SHIMOHATA, Taro WATANABE ATR Spoken Language Translation Research Laboratories 2-2-2 Hikaridai, Keihanna Science City, Kyoto 619-0288, JAPAN eiichiro.sumita@atr.jp Abstract This paper introduces ATR’s project named Corpus-Centered Computation (C3), which aims at developing a translation technology suitable for spoken language translation. C3 places corpora at the center of its technology. Trans- lation knowledge is extracted from corpora, translation qual- ity is gauged by referring to corpora, the best translation among multiple-engine outputs is selected based on corpora, and the corpora themselves are paraphrased or filtered by automated processes to improve the data quality on which translation engines are based. In particular, this paper reports the hybridization archi- tecture of different machine translation systems, our tech- nologies, their performance on the IWSLT04 task, and para- phrasing methods. 1. Introduction There are two main strategies used in corpus-based transla- tion: 1. Example-Based Machine Translation (EBMT) [1]: EBMT uses the corpus directly. EBMT retrieves the translation examples that are best matched to an input expres- sion and then adjusts the examples to obtain the translation. 2. Statistical Machine Translation (SMT) [2]: SMT learns statistical models for translation from cor- pora and dictionaries and then searches for the best transla- tion at run-time according to the statistical models for lan- guage and translation. By using the IWSLT04 task, this paper describes two endeavors that are independent at this moment: (a) a hy- bridization of EBMT and statistical models, and (b) a new approach for SMT, phrase-based HMM. (a) is used in the “unrestricted” Japanese-to-English track (Section 2), and (b) is used in “supplied” Japanese-to-English and Chinese-to- English tracks (Section 3). In addition, paraphrasing tech- nologies, which are not used in the IWSLT04 task but boost translation performance, are also introduced in Section 4. 2. Hybrid MT System (Unrestricted J-to-E Track) No complete translation system has emerged nor is likely to emerge in the foreseeable future. Every approach to trans- lation has its own way of acquiring translation knowledge and using the knowledge. Each system generates its pe- culiar errors in attempting translation. As a result, trans- lation performance differs sentence-by-sentence, system-by- system. There is the possibility of boosting translation per- formance through exploitation of multiple translations gener- ated by different systems. Among several possible architec- tures to integrate multiple translation engines (Section 2.5), we demonstrate the acrchitecture below (Sections from 2.1 to 2.4) as one effective approach. 2.1. A Hybridization: Multiple EBMTs Followed By A Selector Based On SMT Models It is important to integrate “different” types of element ma- chine translation systems in order to boost the overall per- formance by having them compensate each other. We pro- pose an architecture in which multiple EBMT engines work in parallel and their outputs are passed to a post-process that selects the best candidate according to SMT models. Most EBMT systems employ phrases or sentences as the translation unit so that …

Content truncated to 500 words