Principles and Idiosyncracies in MT Lexicons
Loft Levin, Sergei Nirenburg · 1993
The purpose of an MT lexicon is to facilitate translation. If we adhere to a symbolic (non-statistical) approach then we need a representation of meaning as the basis of translation. This representation is produced by programs typically called semantic and pragmatic interpreters, based on a variety of knowledge sources, such as grammars and lexicons. The grammar rules and lexicon entries vary in generality (productivity) in that they can apply to a very broad class of phenomena, such as all nouns, or to a particular lexical unit, such as however. The size of these rule domains forms, in fact, a range from the former to the latter. It is, naturally, preferable, from the point of view of generality of theory and economy ("capturing generalizations") to prefer the use of few powerful rules applicable to broad domains. However, language analysis shows that this noble pursuit promises but limited success. In a multitude of routine cases it becomes difficult to make generalizations. This leads to the necessity of directly recording information about how to process small classes of phenomena, specifically those that could not be covered by general rules. The trick is to find the watershed between what can be processed on general principles and what is idiosyncratic in language, what we can calculate and what we have to know literally. When an optimum balance on this issue is achieved, a lot of benefits immediately accrue. Thus,