Machine Translation by Interaction between Paraphraser and Transfer

Machine Translation by Interaction between Paraphraser and Transfer Kazuhide Yamamoto ATR Spoken Language Translation Research Laboratories 2-2-2, Hikaridai, Seika-cho, Soraku-gun, Kyoto, 619-0288 Japan yamamoto@fw.ipsj.or.jp Abstract A machine translation model has been pro- posed where an input is translated through both source-language and target-language para- phrasing processes. We have implemented our prototype model for the Japanese-Chinese lan- guage pair. This paper describes our core idea of translation, where a source language para- phraser and a language transfer cooperates in translation by exchanging information about the source input. 1 Introduction Humans generally have language capability, mostly for their mother tongue and to a lesser extent for foreign languages. This leads us to making the most of our mother language, even in conducting translation. That is, when we translate our language into a foreign one unfamiliar to us, we may try to paraphrase the source input into easier expressions we can translate. In contrast, there is no such machine trans- lation (MT) model so far proposed where the source language module is biased over the bilin- gual language module. All of the MT models are either those where the bilingual processor takes the initiative over the source language analyzer (conventional analyze-transfer-generate model) or integration models of analyzer and transfer, such as example-based or statistical models. Al- though some MT models have a paraphraser (also called a ‘pre-editor’), such as that of Shirai et al. (1993), paraphrasing is performed in these models because it is necessary to prepare for the subsequent bilingual process. In other words, the paraphraser operates as a sub-module for successful transfer. We have proposed a new MT model that is more similar to the human translation pro- cess than other MT systems (Yamamoto et al., 2001). This model, called the Sandglass model, is designed so that the system can gener- ate a translation through source language para- phrasing, even if the system does not have sufficient bilingual knowledge. In this sense, our model design can be considered a non- professional translator’s model. From the engineering point of view, our model has an advantage in language portability; it is easy to construct an MT for a new language, since our model depends only on source lan- guage and thus can reduce dependence on bilin- gual knowledge. Moreover, the better source language paraphraser we make, the easier the implementation of other language MT becomes. Another advantage is task portability, since all of the paraphrasing knowledge, except for lex- ical paraphrasing knowledge, is independent of the task, so we do not need to fit most of the paraphrasing knowledge to the required task. It is also significant that this model’s paraphraser can be employed not only for MT but also for most natural language processing (NLP) appli- cations. This is possible because both the input and output of a paraphraser is the same natural language. We have been building the Sandglass MT system for the Japanese-Chinese, Chinese- Japanese language pairs (Yamamoto et al., 2001; Zhang and Yamamoto, 2002). We have already constructed a …

Content truncated to 500 words