<?xml version="1.0" standalone="yes"?>
<Paper uid="H89-2052">
  <Title>Using Structural Constraints for Speech Act Interpretation</Title>
  <Section position="2" start_page="0" end_page="385" type="intro">
    <SectionTitle>
1. INTRODUCTION
</SectionTitle>
    <Paragraph position="0"> A fully functional natural language understanding system will require much more sophisticated interaction between its structural processing (e.g. parsing, semantic interpretation) and its general knowledge and reasoning abilities. In this paper we explore one aspect of this problem, namely the interpretation of intended stmech acts. This is an excellent problem to study to explore this problem since the interpretation of the intended act depends strongly on the syntactic and semantic structure of the utterance, but also is highly influenced by context, and hence general reasoning. Neither the structural constraints nor the reasoning about the situation fully determines the interpretation alone - rather each adds partial information to the overall solution. As a result, a system organization where the parser and semantic interpreter produce single specific interpretations that are accepted or rejected by the contextual processing will be highly inefficient.</Paragraph>
    <Paragraph position="1"> Systems that perform speech act interpretation to date have attempted to avoid this problem by assuming that the relationship between form and meaning is a simple one. Specifically, almost all work on speech act theory makes what is called the literal meaning hypothesis (e.g. Searle 1969, 1975a, Allen &amp; Perrault, 1980, Litman &amp; Allen, 1987). This is the assumption that the syntactic and semantic structure of the sentence uniquely defines some literal meaning independent of the context of the utterance. All indirect meaning can then be derived by reasoning from the literal interpretation. There are many reasons why this assumption is made, but the pnnciple one is to avoid the general issue of how structural processing and general reasoning interact. If there is a context-independent literal meaning, then the structural processing need only produce this interpretation as its output, and it is then used as the starting point for general reasoning. Thus it allows a full separation of the two parts of the interpretation problem.</Paragraph>
    <Paragraph position="2">  In this paper, I will argue that the literal meaning hypothesis cannot be supported. Primarily, this is because it does not allow for any of the subtleties of phrasing that affect the interpretation of the utterance. In addition, it seems to suggest a model that is contrary to intuition about language learning and processing. These problems result in systems that are insensitive to the syntactic form of the sentence. In addition, systems based on these models could not take advantage of the recognizability of common indirect forms since each sentence was analyzed from first principles. Many people, including the authors, have suggested that certain common forms could be compiled for efficiency reasons. These compiled interpretations, however, still operate from the literal meaning as a starting point and thus are insensitive to syntactic and semantic variation.</Paragraph>
    <Paragraph position="3"> If there is no literal meaning, what replaces its role as the connection between structure and inference? We suggest that the result of structural processing is a specification of the space of possible interpretations. This may include specific limits on what the intention could be (e.g. the sentence in question must be a request of some kind), as well as specific interpretations that are favored because of their familiarity and conventionality. This information can then be used to heuristically control the general plan reasoning so that common forms are recognized quickly, and to constrain the possible interpretations that can be suggested from first principles.</Paragraph>
    <Paragraph position="4"> The resulting framework has significant practical advantages: * the common indkect forms are recognized without costly plan reasoning required by earlier models; * general plan reasoning is constrained in what interpretations it can derive, and thus it more accurate; * the framework allows for the integration of intonational and prosodic cues in spoken language.</Paragraph>
    <Paragraph position="5"> The next section provides some background on speech act theory and the literal meaning hypothesis. Following that, the evidence against the literal meaning hypothesis is discussed in Section 3. We then present our new model in the next two sections: Section 4 outlines the structural interpretation rules that define the interpretation space, and Section 5 discusses the role of plan reasoning in the new framework. Section 6 discusses the implementation briefly, and Section 7 discusses some possible extensions.</Paragraph>
  </Section>
class="xml-element"></Paper>