An Amorphous Object Must be Cut by a Blunt Tool

Abstract

I first proposed Example-Based Machine Translation (EBMT) as an artificial intelligence approach to language translation, and named it the “machine translation by analogy principle.” It was presented at a NATO workshop on Artificial and Human Intelligence which was held in France in 1981. In the late 70s I was struggling to write analysis, transfer and generation grammars for machine translation between Japanese and English. Basic grammar rules are not many, but if we want to handle real texts we have to write many additional grammar rules, say, several hundreds. Still the grammar is not complete. There appear expressions which cannot be handled by such a set of grammar rules. They are called extra-grammatical sentences. I realized that there is no person who has written a complete grammar of English, for example. A language is always changing and there exists no notion of “complete” grammar for a language, and therefore we have to introduce the concept of “learning” to a language system. Another problem I was confronted with was the quality of translated sentences. There were varieties of reasons for the poor quality of the translation. One obvious reason is that machine translation systems so far basically depended on the compositionality principle, that is, an input sentence is decomposed into minimum units (words), and these minimum units are translated unit by unit, and then synthesized into a sentential string of a target language referring to the structure of the input sentence of a source language. In such a machine translation system which is called a rule-based machine translation (RBMT) system, improving a grammar is a hard task. We have to check where the real reason for the failure of analysis of a sentence is, and have to change some of the existing rules or add new rules. By doing this, analysis for that particular sentence will become successful, but there may arise bad side effects, such that some sentences which were analyzed successfully before cease to be analyzed well. In RBMT the addition or change of a rule will have a very complex effect on the systematicity or consistency of a grammar as a whole. This process of grammar