Adaptation and Generalization

Abstract

In Chapter 6 we transformed CHAMP from a bottom-up parser to a least-deviant-first parser by adding error recovery to the system. While error recovery allows CHAMP to accept a larger language than the kernel grammar alone, the average cost of acceptance (and rejection) increases; although we apply constraints at every opportunity, the fact is that search expanded to a non-zero deviation-level is always larger than a grammatical search would be for the same utterance. If we can modify the grammar to recognize a deviant form directly, subsequent encounters with that deviation will not require error recovery at all. Thus, future sentences containing the deviation will succeed at lower deviation-levels, requiring less search. Giving CHAMP the ability to learn new grammatical components transforms the system from a simple least-deviant-first parser to a least-deviant-first adaptiveparser.