A Direction of MT Development 1
Sergei Nirenburg · 1993
I would like to discuss just one among the many potentially fruitful directions of development of our field — the introduction of multi-engine system architectures which would facilitate the choice of the best way to translate a source text passage given a set of system resources. 1. Toward Extended MT Architectures Recent years have witnessed a shift in the balance of scientific and technological efforts in the area of machine translation. All the latest methodological novelties in this field are essentially technology-oriented and do not aim at advancing our knowledge about either basic mechanisms of text comprehension and production or computer models simulating such mechanisms. The two most recently popular technological paradigms in machine translation — example-based translation (EBMT) and statistics-based translation (SBMT) — require knowledge about language only as an afterthought. While the representatives of the above paradigms are still at the stage of either building toy systems (e.g., Furuse and Iida, 1992; McLean, 1992, Jones, 1992, Maruyama and Watanabe, 1992) or struggling with the natural constraints of approaches that eschew the study of language as such (e.g., Brown et al., 1990), a number of proposals have come up for some hybridization of MT. In some such approaches, corpus analysis is used for tuning