<?xml version="1.0" standalone="yes"?>
<Paper uid="N01-1003">
  <Title>SPoT: A Trainable Sentence Planner</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
2 The Sentence Planning Task
</SectionTitle>
    <Paragraph position="0"> The term &amp;quot;sentence planning&amp;quot; comprises many distinct tasks and many ways of organizing these tasks have been proposed in the literature. In general, the role of the sentence planner is to choose abstract linguistic resources (meaning-bearing lexemes, syntactic constructions) for a text plan. In our case, the output of the dialog manager of a spoken dialog system provides the input to our sentence planner in the form of a single spoken dialog text plan for each of the turns. (In contrast, the dialog managers of most dialog systems today simply output completely formed utterances which are passed on to the TTS module.) Each text plan is an unordered set of elementary speech acts encoding all of the system's communicative goals for the current turn, as illustrated in Figure 1. Each elementary speech act is represented as a type (request, implicit confirm, explicit confirm), with type-specific parameters. The sentence planner must decide among alternative abstract linguistic resources for this text plan; surface realizations of some such alternatives are in Figure 2.</Paragraph>
    <Paragraph position="1"> As already mentioned, we divide the sentence planning task into two phases. In the first phase, the sentence-plan-generator (SPG) generates 12-20 possible sentence plans for a given input text plan. Each speech act is assigned a canonical lexico-structural representation (called a DSyntS - Deep Syntactic Structure (Mel'Vcuk, 1988)). The sentence plan is a tree recording how these elementary DSyntS are combined into larger DSyntSs; the DSyntS for the entire input text plan is associated with the root node of the tree. In the second phase, the sentence plan ranker (SPR) ranks sentence plans generated by the SPG, and then selects the top-ranked output as input to the surface realizer, RealPro (Lavoie and Rambow, 1997). The architecture is summarized in Fig-</Paragraph>
  </Section>
class="xml-element"></Paper>