The Role of Linguistic Knowledge Resources in MT

In this chapter we examine, largely by reference to Japanese systems, since they are still little known in the West, the issue of what knowledge sources one expects to find in modern MT systems. Later we go on to ask about the role and need of linguistic knowledge in particular. What linguistic knowledge sources are needed to perform MT depends, at least to a limited extent, on which MT method is used. For example, some US projects (Brown et al., 1989 and Chapter 7 above) made use of very large scale statistical information from texts, while Japanese systems do not. Conversely, an experimental MT system at Kyoto University made use of large lists of sample sentences against which a sentence to be translated is matched (Nagao, 1990), whereas no US systems did this until the turn of the century (the CMU system). Most MT systems however, make use of at least some, or possibly all of the following kinds of lexical knowledge sources as distinct from corpora alone (as in Brown et al.):