Introduction

The growth rate of theoretical studies of language structure and use stubbornly remains higher than the improvement rate of large-scale applications. It has been repeatedly proved that large-scale realistic NLP applications carry a prohibitive price tag of large-scale, routine acquisition of knowledge about language and about the world, collected in computational grammars, lexicons and domain models. Strategically, there are several ways of dealing with this problem:

• biting the bullet and going through a massive knowledge acquisition effort, either general-purpose (e.g., the CYC project, Lenat et al., 1990) or domain-specific (e.g., the KBMT-89 project, Goodman and Nirenburg, 1992);

• seeking ways of bringing down the price of knowledge acquisition by studying ways of automatically or semi-automatically extracting relevant information from machine-readable dictionaries (morphological, syntactic and some semantic information, e.g., in the work of Wilks et al., 1990) or text corpora (for instance, collocations, cf. Smadja, 1991); and

• seeking ways of avoiding the need for massive knowledge acquisition by rejecting the entire established NLP paradigm in favor of knowledge-free, linguistics- and AI-independent approaches.

This last option has been energetically promulgated in the important NLP application of machine translation (MT). The two basic ``non-traditional'' approaches to MT are:

• statistical MT, which seeks to carry out translation based on complex cooccurrence and distribution probability calculations over very large aligned bilingual text corpora; and

• a more modest approach, called example-based MT, which is the topic of this paper.