<?xml version="1.0" standalone="yes"?>
<Paper uid="C90-1013">
  <Title>Generation for Dialogue Translation Using Typed Feature Structure Unification</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
2. Introducing Typed Feature
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Structures
2.1 Selecting Appropriate Rules
</SectionTitle>
      <Paragraph position="0"> The basic mechanism of this generator is the top-down application of the grammar rules and construction of the feature structures of the daughter nodes.</Paragraph>
      <Paragraph position="1"> It is important to avoid the derivation of unnecessary phrase structures by selecting appropriate rules to apply in order to increase the efficiency. Consider the following rules</Paragraph>
      <Paragraph position="3"> (2) The construction of the semantic representations given to the mother nodes of these ~ In this rule, =HC*=&gt; link shows that the first element of the right hand symlx~ls becomes the head daughter and the others the complement daughters. =CH=&gt; link is also supplied for complement-head constructions. A symbol with an exclamation mark (!) indicates a predefined template. In this rule, !m stands for tile mother, i.e., the left-hand VP.</Paragraph>
      <Paragraph position="5"> two rules are the same (predicate-argument structure), as can be seen below:  In generation, appropriate grammar rules must be selected using this representation. For this purpose, the difference between these feature structures must be found in the type of the key feature value's type. In this case, the reln (relation) feature plays the key role, and the value &amp;quot;play&amp;quot; must be of the verb type and &amp;quot;in&amp;quot; the prepositional type.</Paragraph>
      <Paragraph position="6"> Typed feature structures formulated by Ai&amp;quot;t-Kaci (1986) are introduced to handle types in feature structures, because, types cannot be handled by ordinary feature structure unification. Using typed feature structures, the following specifications can be attached to the former rules.</Paragraph>
      <Paragraph position="7"> For (1): (&lt;!m sem cent rein&gt; == \[V j) For (2): (&lt;!m sere cent rein&gt; == \[PJ) These specifications work as constraints to the rule application. The first specification shows that the &lt;sere cont reln&gt; feature of the node is unified to the type V (bracketed, shown in bold italics). If the unification succeeds, i.e., the &lt;sem cont reln&gt; feature is under V type in the type hierarchy, this grammar rule can be applied. The selection of appropriate grammar rules is thus accomplished in a declarative way.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.2 Avoiding Termination Problem
</SectionTitle>
      <Paragraph position="0"> There are various ways to utilize the type hierarchy. One example is subclassifying the categories.</Paragraph>
      <Paragraph position="1"> One of the termination problems Shieber et al. (1989) pointed out is in the left-recursive rules. The rule (1) infinitely appends the subcat list to the daughter VP if the grammar is used for generation. This can be solved by restricting the permissible length of the subcat list t?. The maximum length of the subcat list is 2, excluding the subject. This restriction can be represented as follows.</Paragraph>
      <Paragraph position="2"> 1&amp;quot;'\] Though the restriction cannot be applied to languages like Dutch (Shieber et al., 1989), the limitation is irrelevant to our purpose (translation between Japanese and English).</Paragraph>
      <Paragraph position="3"> (:or (&lt;lm lsubcat&gt; == \[list-end\])) (&lt;!m lsubcat rest&gt; == \[list-end\])) However, this restriction forces the rule (2) to be applied twice to all verbs including intransitive verbs. Derivation of the phrase structures with incorrectly extended subcat lists will fail when the terminal is reached.</Paragraph>
      <Paragraph position="4"> This restriction can be solved more effectively using the type hierarchy. If verbs are classified into three subtypes (Monadic, Dyadic and Triadic) by the numbers of their arguments, the restriction in rule (1) can be written as follows.</Paragraph>
      <Paragraph position="5"> (:or ((&lt;Im sem cent rein&gt; == \[dyadic\])</Paragraph>
      <Paragraph position="7"/>
    </Section>
    <Section position="3" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.3 Relating Types and Categories
</SectionTitle>
      <Paragraph position="0"> Another function of the type hierarchy is using the types as a bridge between the semantics in the feature structure and the category in the CFG rules. Categories (nonterminal symbols) are also expressed by, and are closely related to, types. The following lexical entry definition shows that the complement of the verb is VP.</Paragraph>
      <Paragraph position="2"> If type VP is a subtype of XP, \[XP\] and \[VP\] are unified to bear \[VP\] when the lexical entry &amp;quot;must&amp;quot; is unified in rule (1). In D-PATR, such unspecified categories are treated by the system by introducing special symbols X, Y, etc. Typed feature structures serve as a sound foundation for this task.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>