<?xml version="1.0" standalone="yes"?>
<Paper uid="C88-2128">
  <Title>A Uniform Architecture for Parsing and Generation</Title>
  <Section position="9" start_page="617" end_page="617" type="concl">
    <SectionTitle>
7 Further Kesearch
</SectionTitle>
    <Paragraph position="0"> Perhaps the most immediate problem raised by the methodology for generation introduced in this paper is the strong requirement of semantic monotonicity, which serves as yet another instance of the straitjacket of compositionality, The semantic-monotonicity constraint allows the goal logical form to be systematically decomposed so that a.</Paragraph>
    <Paragraph position="1"> dynamic-programming generation process can be indexed by the parts of the decomposition (the subformulas), just as the constraint of string concatenation in context-free grammars allows a goal sentence to be systematically decomposed so that a dynamic-programming parsing process can be indexed by the subparts of that decomposition (the 14Such a proof is currently in preparation.</Paragraph>
    <Paragraph position="2">  substrings). And just az the concatenation restriction of context-free grammars can be problematic, so can the restriction of semantic monotonicity. Finding a weaker constraint on grammars that still allows efficient processing is thus an important research objective.</Paragraph>
    <Paragraph position="3"> Even with the semantic-monotonicity constraint, the process of indexing by the highly structured logical forms is not nearly so efficient as indexing by simple integer string positions. Partial match retrieval or similar techniques from the Prolog literature might be useful here. Nothing has been said al~out the importartt problem of guaranteeing that the syntactic and semantic goal properties will actually be realized in the sentence generated. The success criterion for generation described here would require that the logical form for the sentence generated be identical to the goal logical form. However, there is no guarantee that the other properties of the sentence include those of the goal; only compatibility is guaranteed. Researchers at the University of Stuttgart have proposed solutions to this problem based on the type of existential constraint found in lexieal-functional grammar. We expect that their methods might be applicable within th~ presented architecture.</Paragraph>
    <Paragraph position="4"> Finally, on a more pessimistic note, we turn to a widespread problem in all systems for automatic generation of natural language, which Appelt \[1987\] has discussed under the rubric &amp;quot;the problem of logical-form equivalence&amp;quot;. The mapping from logical forms to natural-language expressions is in general many-to-one. For instance, the logical forms red(x) h ball(x) and ball(x) h red(x) might both be realized as the nominal 'red ball'. However, most systems for describing the string-LF relation declaratively do so by assigning a minimal set of logical forms to each string, with each logical form standing proxy for all its logical equivalents. The situation is represented graphically as Figure 1.</Paragraph>
    <Paragraph position="5"> The problem is complicated further in that, strictly speaking, the class of equivalent logical forms from the standpoint of generation is not really closed under logical equivalence. Instead, what is actually required is a finer-grained notion of intentional equivalence, under which, for instance, p and p A (q Y -~q) would not necessarily be inten.. tionally equivalent; they might correspond to different uttera~aces, one about p only, the other about both p and q.</Paragraph>
    <Paragraph position="6"> In such a system, merely using the grammar per se to drive generation (as we do here) allows for the generation of strings from only a subset of the logical-form expressions that have natural-language relata, that is, LF1, LF2, and LF3 in the figure. We might call these the canonical logical forms. Even if the grammar is reversible, the problem remains, because logical equivalence is in general not computable. And even in restricted cases in which it is computable, for instance a logic with a confluence property under which all logically equivalent 61B expressions do have a canonical form, the problem is not solved unless the notion of canonical form implicit in the logic corresponds exactly to that of the natural-language grammar.</Paragraph>
    <Paragraph position="7"> It should be noted that this kind of problem is quite deep. Any,system that :represents meanings in some way (not necessarily with logical expressions) must face a correlate of this problem. Furthermore, although the issue impinges on syntax because it arises in the realm of grammar, it is primarily a semantic problem, as we will shortly see.</Paragraph>
    <Paragraph position="8"> There are two apparent possible approaches to a resolution of this problem. We might use a logic in vchich logical equivalence classes of expressions are all trivial, that is, any two distinct expressions mean something diiferent. In such a logic, there are no artifactual syntactic remnants in the syntax of the logical language. Furthermore, expressions of the logic must be relatable to expressions of the natural language with a reversible grammar. Alternatively, we could use a logic for which canonical forms, corresponding exactly to the natural language graramar's logical forms, do exist.</Paragraph>
    <Paragraph position="9"> The difference between the two approaches is only an apparent one, for in the latter case the equivalence classes of logical forms can be identified as h)gical forms of a new logical language with no artifactual distinctlons. Thus, the second case reduces to the first. The central problem in either case, then~ is discovery of a logical language which exactly and uniquely represents all the meaning distinctions of natural language utterances and no others. This holy grail, of course, is just the goal of knowledge representation for natural language in general; we are unlikely to be able to rely on a full solution soon.</Paragraph>
    <Paragraph position="10"> However, by looking at approximations of this goal, suitably adapted to the particular problems of generation, we can hope to achieve some progress. In the case of approximations, it does not hold that the two methodologies reduce one to another; in this case, we feel that the second approach--designing a logical language that approximates in its canonical forms those needed for grammatical applications-qs more likely to yield good incremental results.</Paragraph>
  </Section>
class="xml-element"></Paper>