International Workshop on Spoken Language Translation (IWSLT) 2004

Keihanna Science City, Kyoto, Japan
September 30-October 1, 2004

How Long Will We be Able to Ignore Linguistic Knowledge and Their Formalisms?

Jun'ichi Tsujii

Department of Computer Science, University of Tokyo, Japan

The paradigms of MT proposed so far have their own attractions such as SBMT being good for rapid development of MT systems, EBMT for non-compositional translation, etc. However, it is becoming increasingly clear that proper theories of language are also crucial for quality of NLP systems. In this talk, we will argue that grammar in proper linguistic formalisms can improve performances of systems based on ill-conceived grammarCand that it is the time for another paradigm shift in NLP in general and MT in particular.
   Our experience in parsing has show a parser that uses linguistically sound formalisms with substantial knowledge of lexical items can not only supersede the performance of parsers based on arbitrary forms of grammar but also improve adaptability towards specific domain and widen the scope of applicability in actual NLP application systems. Good grammar formalisms also provide better bases for statistical language models. Since MT have to deal with diverse aspects of language, we need to avoid the naive distinction of different MT paradigms and start to pursue possible integration of good ideas in different paradigms.

Bibliographic reference.  Tsujii, Jun'ichi (2004): "How long will we be able to ignore linguistic knowledge and their formalisms?", In IWSLT-2004 (abstract).