Applying TDMT to abstracts on science and technology

MT Summit VII Sept.1999 Applying TDMT to Abstracts on Science and Technology Hideki Kashioka Hiroko Ohta Kazutaka Takao Yoshiko Shirokizawa ATR Interpreting Telecommunications Japan Science and Research Laboratories Technology Corporation Abstract In this paper, we discuss applying a trans- lation model, "Transfer Driven Machine Translation" (TDMT), to document ab- stracts on science and technology. TDMT, a machine translation model, was devel- oped by ATR-ITL to deal with dialogues in the travel domain. ATR-ITL has reported that the TDMT system efficiently trans- lates multi-lingual spoken-dialogs. How- ever, little is known about the ability of TDMT to translate written text transla- tions; therefore, we examined TDMT with written text from English to Japanese, es- pecially abstracts on science and technol- ogy produced by the Japan Science and Technology Corporation (JST). The exper- imental results show that TDMT can de- rive written text translation. 1 Introduction ATR-ITL has developed the TDMT (Transfer Driven Machine Translation) system [ Mima et al., 1998] as a prototype system for spoken dialogue translation. The TDMT system provides an efficient technique for handling spoken dialogues by using the following three features: 1. capability for ungrammatical expressions, 2. real-time translation, and 3. ways to smooth the addition of translation knowl- edge. These features are important requirements for trans- lating spoken dialogues. For written text translation, these features, especially the third, are necessities. Text databases attract many people's attention be- cause they are a convenient way to obtain information. Therefore, The Japan Science and Technology Corpo- ration (JST) has produced a Japanese database com- posed of abstracts on science and technology [JST]. If the machine translation system has the ability to translate an English abstract into a Japanese abstract, then the process of producing Japanese abstracts can be carried out in less time than human translation. We examined the ability of EJ translation for written text with TDMT. In the next section of this paper, we give an overview of TDMT. Section 3 describes the characteristics of document abstracts on science and technology. Section 4 presents the points of system modification, and Section 5 describes an experiment for applying TDMT to written text. In Section 6, we state our conclusions. 2 Overview of TDMT TDMT has three modules: morphological analysis, transfer, and sentence generation. The transfer mod- ule is an essential component that applies ''transfer knowledge" to an input sequence that is an output of a morphological analysis module. In TDMT, a trans- fer module is different, from those of other translation systems. Almost all translation systems analyze the structure of an input expression with source language grammar, which is called a parsing phase. After that, they translate the source language structure into the target language structure with translation rules. Fi- nally, they generate the target sentence from the tar- get language structure. Most systems divide the pars- ing part and the structure translation part. TDMT simultaneously processes parsing and structure trans- lation by using the "transfer knowledge." The "transfer knowledge" is required to treat source/ target language expressions at …

Content truncated to 500 words