A THEORY OF SCALAR IMPLICATURE (NATURAL LANGUAGES, PRAGMATICS, INFERENCE)

Speakers may convey many sorts of \u27meaning\u27 via an utterance. While each of these contributes to the utterance\u27s overall communi- cative effect, many are not captured by a truth-functional semantics. One class of non-truth-functional, context-dependent meanings, has been identified by Grice Grice 75 as CONVERSATIONAL IMPLICATURES. This thesis presents a formal account of one type of conversational implicature, termed here SCALAR IMPLICATURE, identified from a study of a large corpus of naturally occurring data collected by the author and others from 1982 through 1985. Scalar implicatures rely for their generation and interpretation upon the assumption that cooperative speakers will say as much as they truthfully can that is relevant to a conversational exchange. For example, B\u27s utterance of (1a) (UNFORMATTED TABLE FOLLOWS) (1) A: How was the party last night? a. B: Some people left early. b. Not all people left early. (TABLE ENDS) may convey to A that, as far as B knows, (1b) also holds--even though the truth of (1b) clearly does not follow from the truth of (1a). Scalar implicatures may be distinguished from other conversa- tional implicatures in that their generation and interpretation is dependent upon the identification of some salient relation that orders a concept referred to in an utterance with other concepts. In 1, for example, the salience of an inclusion relation between \u27some people\u27 and \u27all people\u27 in the discourse is prerequisite to B\u27s implicating that (1b)--and to A\u27s understanding that (1b) has in fact been implicated. To illustrate potential applications of the theory presented, a module of a natural-language interface, QUASI, is described. QUASI calculates scalar implicatures that might be licensed by simple direct responses to yes-no questions. Where licenseable implicatures are not consistent with the system\u27s knowledge base, QUASI proposes alternative responses. This system demonstrates how natural language interfaces can use the calculation of implicit meanings to avoid conveying misinformation and to convey desired information more succinctly