<?xml version="1.0" standalone="yes"?>
<Paper uid="J86-2002">
  <Title>SUMMARIZING NATURAL LANGUAGE DATABASE RESPONSES</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
2 THE NATURE OF SUMMARY RESPONSES
</SectionTitle>
    <Paragraph position="0"> An important convention of human conversation is that no participant monopolize the discourse, ensuring that control can be shared (Joshi, Kaplan, and Lee 1977). For example it is often considered inappropriate for a speaker to respond with a lengthy list of data; a shorter non-enumerative response is, at times, more appealing. Lengthy response sets could be summarized, or defined by a characteristic or attribute. For instance, the question QI: Which employees engage in profit sharing? may be answered by listing the extension of a set containing perhaps, a long list of names, or by the intensional response SI: All vice-presidents.</Paragraph>
    <Paragraph position="1"> Such summary answers avoid unnecessary and distracting details, and more important, they do not mislead the user. As another example, consider the query Q2 given below (from Reiter et al. 1983): Q2: Which department managers earn over $40k per year? $2-1: Abel, Baker, Charles, Doug.</Paragraph>
    <Paragraph position="2"> $2-2: All of them.</Paragraph>
    <Paragraph position="3"> Response $2-1 is what might be expected of an existing system; response $2-2, the summary response, is normally more appropriate if conversational principles and practices are to be adhered to. By enumerating managers who earn over $40k, the first response implies that there are managers who do not earn that much. Such a scalar implieature follows from the cooperative principle in conversation (Grice 1975:45) that requires a speaker to make his/her &amp;quot;conversational contribution&amp;quot;, such as is required, at the stage at which it occurs, by the accepted purpose or direction of the talk exchange in which \[he/she is\] engaged&amp;quot;. There are four maxims that derive from this principle: (i) the maxim of quantity - make the contribution as informative as desired but not more so; (ii) the maxim of quality - do not say what is believed to be false or that for which evidence is lacking; (iii) the maxim of relation - be relevant; (iv) the maxim of manner - avoid obscurity, avoid ambiguity, be brief, be orderly.</Paragraph>
    <Paragraph position="4"> In the case of query Q2 above, the responder would only enumerate positive instances if he/she could not say the more informative All of them. Thus, $2-1 might mislead a user, who would expect the system to respond with $2-2 if it were true. Of course, the Gricean maxims must be viewed as being phrased relative to the responder's perceptions of the user's knowledge. Thus, the responder would have to know exactly what particular knowledge a given user had before being able to decide with certainty which responses are likely to be misleading. The usefulness of modelling the user will be discussed further in section 6.</Paragraph>
    <Paragraph position="5"> In general, summary responses are aimed at meeting the maxims of quantity and manner. As Q2 illustrates, a summary response can be (somewhat paradoxically) more informative than an extensional response. It can also be briefer and less obscure than an extensional response. It is important to make only the relevant summary responses. Our system tries to maintain relevance through the use of a knowledge base tailored to meet the expectations of different classes of users (see section 4.2). Since we don't produce responses not satisfied by the data, the maxim of quality is not changed by the generation of summary responses. It is possible, however, to produce a summary response which itself violates Grice's maxims. For example, assume the query Who passed CMPT 110? were posed to a summary response generation system. Producing the answer All students who got over 50% (if this happened to be true in the current data base) would normally violate the maxims</Paragraph>
  </Section>
class="xml-element"></Paper>