<?xml version="1.0" standalone="yes"?>
<Paper uid="C90-2031">
  <Title>On Trying to Do Things with Words Another plan-based approach to speech act interpretation</Title>
  <Section position="1" start_page="0" end_page="0" type="abstr">
    <SectionTitle>
Abstract
</SectionTitle>
    <Paragraph position="0"> Usnal plan-based approaches to speech act interpretatiott require that perib~zning a speech act implies its succes:s. These approaches are thus useless for describing failing illecut~onary or perlocutionary acts. We propose an t~lternatiw~ plan-based view of speech acts centered around the notion of trying to do - as opposed to actually doing- an action. This approach is contrasted with that of I'errault which aims to overcome similar problems.</Paragraph>
    <Paragraph position="1"> 1. :~ntroduction The., plan-based approach to the analysis of natural langu~bge dialogues, inspired by the work of Austin \[Austin 62\] and Searle \[Searle 75\] \[Searle/Vanderveken 85\], has dominated most of the dialogue-oriented NLU research since the late seventies, cf. \[Allen 79\], \[Allen 84\], \[Cohen 78\], \[CohenfPerrault 791\], \[Perrault/Allen 80\], \[Litman 85\]~ \[Pollack 86\]. It is characterized by the following assumptions: null o Utterances are actions planned and executed by the speaker to satisfy some of her goals.</Paragraph>
    <Paragraph position="2"> * Speech acts can be represented as operators in planning systems such as STRIPS \[Fikes/Nilsson 71\] which relate actions to their preconditions and effects. null $ The speaker's beliefs and intentions can be inferred by observing her utterances.</Paragraph>
    <Paragraph position="3"> Sew~ral variants of describing actions in planning systems have been proposed in the lkerature. We will use the tbllowing c, onventions: an action is defined by a quintuple comprising * an action description, e.g. (pick-up ?x) in the blocks world domain, o a set of preconditions, e.g. (on-table ?x), * the add-llst, a set of propositions which become true once the action has been performed, e.g. (holding ?x), * the delete-list, a set of propositions which are no longer true after the action has been pertbrmed, e.g.</Paragraph>
    <Paragraph position="4"> (on-table ?x), (r) the body, a list of lower-level actions comprising the described action, e.g. the movements of a robot arm :in the blocks world domain.</Paragraph>
    <Paragraph position="5"> 1Thi:~ paper was written by the first author but describes work jointly undertaken with the second.</Paragraph>
    <Paragraph position="6"> The relationship between these elements is interpreted as conditional generation (following \[Goldman 70\], \[Pollack 86\]): the execution of the body actions generates the defined action (and thus its effects as described in the add- and delete-list) iffthe preconditions hold.</Paragraph>
    <Paragraph position="7"> All three acts involved in making an utterance, the locutionary, illocutionary, and perlocutionary acts, may fail: the addressee may not hear the utterance, she may not understand the speaker, or she may not react according to the speaker's intentions.</Paragraph>
    <Paragraph position="8"> But when actions are defined in terms of their effects, a failure to achieve the effects implies a corresponding failure to perlbrm the action. When I try to drive a nail into the wall with a hammer, and fail, then I have not driven a nail into the wall. Similarly, when I utter a declarative sentence, and fail to convince the hearer, I have failed to perform the intended perlocutionary act. When the addressee does not understand me, 1 have not even performed an illocutionary act, and so on. Regardless of the level of description there's always the chance that the action may fail, i.e. that there was no action. Even granted that we could capture practically all of the relevant cases by describing the action on the level of, say, producing sounds, there's no way to relate that level of description to the intentions of the speaker.</Paragraph>
    <Paragraph position="9"> This is a rather unfortunate result, as we take the observed action as the starting point for inferring the speakers beliefs and intentions -- which may well be the same, regardless of the speech act's success. Clearly, an approach facilitating a uniform treatment of succeeding and failing speech acts would be most welcome.</Paragraph>
    <Paragraph position="10"> There are two ways to cope with failing speech acts.</Paragraph>
    <Paragraph position="11"> The first amounts to weakening the inference from the performance of an action to its effects being achieved, i.e. making it defeasible. The other solution is to describe actions in terms not presupposing their successful execution. The former approach was proposed by Perrault \[Perrault 87\] and further refined by Appelt and Konolige \[Appelt/Konolige 88\], while the latter is the one we use.</Paragraph>
  </Section>
class="xml-element"></Paper>