ABSTRACT
INTRODUCTION: CONSTRAINED CHAOS This paper explores several ways of using connectionist networks to parse sentences which are syntactically “extragrammatical.” Extragrammatical is not a well-defined term: A sentence is extragrammatical if it is ungrammatical but would still make sense to many native speakers. Labelling a sentence with any of the terms sensible, ungrammatical, or extragrammatical is a judgment by the person who hears the utterance. It is doubtful that any two people would make the same judgment all the time. In fact, even one person is likely to be inconsistent. Suppose, however, that we look at one user of a language. One of the striking things about human language use is that our speaker will attach an interpretation to almost any utterance she hears. Even a nonsense stream of sounds will be heard as phonemes and words. Sentences with “errors” in them yield enormous amounts of information, e.g. they will be broken into phrases even if they contain mostly (or perhaps all) nonsense words. Contrast this to an interpreter of a formal language. It is relatively easy to build a checker that will decide whether or not a string belongs to the language. If it does, it is also easy to extract a parse of the string; however, it is much harder to derive meaningful information about nonmember strings. This is inadequate for natural language. Even if we had a formal language deriving all of the “correct” sentences of some language, it would be neither an adequate model of the human language faculty nor an adequate basis for a computer natural language processor.
