<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-1419">
  <Title>TEXTUAL ECONOMY THROUGH CLOSE COUPLING OF SYNTAX AND SEMANTICS</Title>
  <Section position="3" start_page="0" end_page="179" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Th e problem we address is that of producing efficient descriptions of objects, collections, acti0ns, events, etc. (i.e., any generalized individual from a rich ontology for Natural Language such as those described in \[2\] *and advocated in \[9\]). We are interested in a particular kind of efficiency that we call textual economy, which presupposes a view of sentence generation as goal-directed activity that has broad support in Natural Language Generation (NLG) research \[1, 5, 15, 17\]. According to this view, a system has certain communicative intentions that it aims to fulfill in producing a description. For example, the system might have the goal of identifying an individual or action o~ to the hearer, or ensuring that the hearer knows that has property P. Such goals can be satisfied explicitly by assembling appropriate syntactic constituents--for example, * satisfying the goal of identifying an individual using a noun phrase that refers to it or identifying an action using a verb phrase that specifies it. Textual economy refers to satisfying such goals implicitly, by exploiting the hearer's (or reader's) recognition of inferential links to material elsewhere in the Sentence that is there to satisfy independent communicative goals. Such material is therefore overloaded in the sense of \[18\]. l While there are other ways of increasing the efficiency of descriptions (Section 5), our focus is * on the efficiency to be gained by viewing a large part of generation in terms of describing (generalized) individuals.</Paragraph>
    <Paragraph position="1"> Achieving this however places strong requirements on the representation and reasoning used in generating sentences. The representation must support the generator's proceeding incrementally through the syntax and semantics of the sentence as a whole. The reasoning used must enable the generator to assess quickly and reliably at any stage how the hearer will interpret the current sentence, with its (incomplete) syntax and semantics. Only by evaluating the status of such key questions as</Paragraph>
    <Paragraph position="3"> * what (generalized) individuals would the hearer take the sentence to refer to? * what would the sentence invite the hearer to conclude about those individuals? * how can this sentence be modified or extended? can the generator recognize and exploit an opportunity for textual economy.</Paragraph>
    <Paragraph position="4"> These representational and reasoning requirements are met in the SPUD system for sentence planning and realization \[26, 27\]. SPUD draws on earlier work by Appelt \[1\] in building sentences using planning techniques, sPUD plans the syntax and semantics of a sentence by incorporating lexico-grammatical entries into a partial sentence one-by-one and incrementally assessing the answers to the questions given above. In this paper, we describe the intermediate representations that allow SPUD to do so, since these representations have been glossed over in earlier presentations \[26, 27\]. Reasoning in SPUD is performed using a fast modal theorem prover \[24, 25\] to keep track both of what the sentence entails and what the sentence requires in context. By reasoning about the predicated relationships withinclauses and the informational relationships \[16\] between clauses, sPUD is able to generate sentences that exhibit two forms of textual economy: referential interdependency among noun phrases within a single clause, and pragmatic overloading of clauses in instructions \[7\].</Paragraph>
    <Paragraph position="5"> For an informal example of the textual economy to be gained by taking advantage of predicated relationships within clauses, consider the scene pictured in Figure 1 and the goal of getting the hearer to take the rabbit currently in the hat out of the hat it's currently in. Even though there are several rabbits, several hats, and even a rabbit in a bathtub and a flower in a hat, it would be sufficient here to issue the command: (1) Remove the rabbit from the hat.</Paragraph>
    <Paragraph position="6"> It suffices because one of the semantic features of the verb remove--that its object (here, the rabbit) starts out in the source (here, the hat). distinguishes the intended rabbit and hat in Figure 1 from the other ones. Pragmatic overloading \[7\] illustrates how an informational relation between clauses can support textual economy in the clauses that serve as its &amp;quot;arguments&amp;quot;. In \[7\], we focused on describing (complex) actions, showing how a clause interpreted as conveying the goal13 or termination condition r of an action a partially specified in a related clause forms the basis of a constrained inference that provides additional information about a. For example, (2a) Hold the cup under the spigot...</Paragraph>
    <Paragraph position="7"> (2b) ...to fill it with coffee.</Paragraph>
    <Paragraph position="8">  Here, the two clauses (2a) and (2b) are related by purpose---specifically, enablement. The action t~ described in (2a) will enable the actor to achieve the goal/3 described in (2b). While a itself does not specify the orientation of the cup under the spigot, its purpose fl can lead the hearer to an appropriate choice---4o fill a cup with coffee, the cup must be held vertically, with its concavity pointing upwards. As- noted in \[7\], this constraint depends crucially on the purpose for which a is performed. The purpose specified i n (3b) does not constrain cup orientation in the same way: (3a) Hold the cup under the faucet...</Paragraph>
    <Paragraph position="9"> (3b) ...to wash it.</Paragraph>
    <Paragraph position="10"> Examples like (1) and* (2) suggest that the natural locality for sentence planning is in a description of a generalized individual. Even though such descriptions may play out over several clauses (or even sentences), the predications within clauses and the informational relations across clauses of a description give rise to similar textual economies, that merit a similar treatment.</Paragraph>
  </Section>
class="xml-element"></Paper>