<?xml version="1.0" standalone="yes"?>
<Paper uid="J90-4002">
  <Title>SENTENTIAL SEMANTICS FOR PROPOSITIONAL ATTITUDES</Title>
  <Section position="15" start_page="0" end_page="0" type="concl">
    <SectionTitle>
4.2 CONCLUSIONS AND FUTURE WORK
</SectionTitle>
    <Paragraph position="0"> This paper has presented a new notation for a sentential theory of attitudes, which unlike most existing notations makes it possible to give a compositional semantics for attitude reports. Our notation distinguishes between the de re arguments of an attitude operator and the dummy variables, which stand for unspecified terms that represent the values of the de re arguments. The choice of dummy variables is quite arbitrary--just as the choice of bound variables in first-order logic is arbitrary. This allows us to impose a convention, which says that in fact the dummy variables are equal to the de re arguments. Given this convention, the logical form of a clause is the same whether it stands alone or appears as the argument of an attitude verb.</Paragraph>
    <Paragraph position="1"> This is a simple proposal, and it would be easy to write and implement a grammar that applies the proposal to a few examples. The real question is whether the proposal is robust--whether it can function in a grammar that covers a variety of phenomena. We chose definite clauses and a first-order object language as our semantic formalism. We found a nonobvious interaction between our proposal for de re attitude reports, and two other problems about quantifi-Computational Linguistics Volume 16, Number 4, December 1990 231 Andrew R. Haas Sentential Semantics for Propositional Attitudes cation: the choice of bound variables in a logical form, and the conjunction and disjunction of quantified NPs. We considered two possibilities for choosing the bound variables: assigning a different variable to every NP using a global counter, or requiring each quantifier to bind a variable that is not bound by any quantifier within its scope. The first approach makes it impossible to use our rules for NP conjunction and disjunction, while the second creates implementation problems for the de re argument lists. We resolved the dilemma by picking the second approach, and then rewriting the grammar to solve the implementation problems. Thus we have shown that the proposal for de re attitude reports is not just a plausible notion--it can be made to work in a grammar that is not trivial.</Paragraph>
    <Paragraph position="2"> The grammar handles three kinds of attitude constructions: an attitude verb taking a clause as its object (&amp;quot;John thought he saw Mary&amp;quot;), an attitude verb taking a clause with a gap (&amp;quot;John knows who Mary likes&amp;quot;), and an attitude verb taking a noun phrase as its object (&amp;quot;John wants a Porsche&amp;quot;). The grammar includes de re~de dicto ambiguities, conjunction of NPs, and a very limited treatment of pronouns.</Paragraph>
    <Paragraph position="3"> Another contribution of our work is that it seems to be the first unification grammar that builds logical forms, and at the same time respects the declarative semantics of the notation. We explicitly choose the bound variables of the logical form, instead of using meta-language variables. We also explain the semantics of our representation for quantified NPs: each NP has an infinite set of readings, one for each ordered pair in the extension of the application function. Several authors treat unification grammars with semantics as poor relations of Montague grammar. Pereira and Shieber (1987) propose to &amp;quot;encode&amp;quot; Montague's ideas in unification grammar, while Moore (1989) fears that building logical forms with unification grammar is &amp;quot;unprincipled feature hacking.&amp;quot; We claim that these problems arise not from shortcomings of unification grammar, but from failure to take unification grammar seriously--which means respecting its declarative semantics.</Paragraph>
    <Paragraph position="4"> The most obvious line for future work is to extend the grammar. It would be fairly easy to include complex wh noun phrases, such as &amp;quot;whose cat&amp;quot; or &amp;quot;how many children&amp;quot; (Cooper's grammar handled these). A more difficult problem arises when a gap is the object of an intensional verb---as in &amp;quot;John knows what Mary wants.&amp;quot; The grammar can generate this sentence, but it assigns only a de re reading: unique(x,name(x,john), some(y,thing(y), know(x,\[y\],q(unique(z,name(z,mary), wish(z,\[z,y\],q(have(z,y)))))))).</Paragraph>
    <Paragraph position="5"> This is the only reading because the gap has an empty quantifier store--there is no quantifier available to be applied to the wff &amp;quot;have(z,y).&amp;quot; Yet there are examples in which such sentences do have de dicto readings. For example, consider &amp;quot;What John wants is a Porsche.&amp;quot; Surely this sentence has a de dicto reading--yet the object of&amp;quot;want&amp;quot; is a gap, not a quantified NP. Cooper discusses this problem, but his grammars could not handle it, and neither can ours.</Paragraph>
    <Paragraph position="6"> Hirst and Fawcett (1986) have argued that the ambiguities in attitude reports are more complex than the familiar distinction between de re and de dicto readings. They claim that the sentence &amp;quot;Nadia wants a dog like Ross's&amp;quot; has a reading in which Nadia doesn't want a particular dog (so the qaantifier 3 is inside the scope of the attitude operator), but the description &amp;quot;dog like Ross's&amp;quot; is the speaker's, not Nadia's (so the description is outside the scope of the attitude operator). This reading is certainly different from the usual de re and de dicto readings, in which either the whole logical form of the NP is under the attitude, or none of it is. To represent it, we must be able to separate the logical form of the NP &amp;quot;a dog like Ross's&amp;quot; into two parts (the quantifier 3 and its range restriction), and we must be able to move the range restriction out of the scope of the attitude without moving the quantifier. This will mean that we cannot use the same mechanism for both quantifier scope ambiguities and the ambiguities that arise from attitudes. These extensions appear feasible, but they amount to a major change in the formalism.</Paragraph>
    <Paragraph position="7"> Another possibility for future work is to incorporate the exi,;t!,ng implementation into a question-answering progrant. This requires finding a way to reason about propositional attitudes efficiently without assuming unique standard designators--which means a substantial generalization of the work of Konolige. It also requires us to make much stronger claims about the properties of 'representation' than we have made in this paper. If these problems can be solved, it should be possible to build a natural language question-answering program that can describe the extent of its own knowledge--answering questions like &amp;quot;Do you know the salary of every employee?&amp;quot;</Paragraph>
  </Section>
class="xml-element"></Paper>