<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1513">
  <Title>Three reasons to adopt TAG-based surface realisation</Title>
  <Section position="7" start_page="100" end_page="101" type="evalu">
    <SectionTitle>
6 Discussion
</SectionTitle>
    <Paragraph position="0"> We have argued that TAG presents several features that makes it particularly amenable to the development of an optimised surface realiser. We now summarise these features and briefly compare TAG with CCG (Combinatory Categorial Grammar) and HPSG (Head Driven Phrase Structure Grammar) based surface realisation.</Paragraph>
    <Section position="1" start_page="100" end_page="101" type="sub_section">
      <SectionTitle>
6.1 Using tree node types
</SectionTitle>
      <Paragraph position="0"> The different types of tree nodes identified by TAG can be used to support polarity filtering whereby substitution nodes can be associated with negative polarities (requirements) and root nodes with positive polarities (resources). As our preliminary experiments show, polarity filtering has a significant impact on the initial search space, on the space used and on CPU times.</Paragraph>
      <Paragraph position="1"> So far, this particular type of global filtering on the initial search space has been used neither in the HPSG (CCFP99; CO05) nor in the CCG (Whi04) approach. Although it could presumably be adapted to fit these grammars, such an adaptation is in essence less straightforward than in TAG. In CCG, the several combination rules mean that a subcategory can function either as a resource or as a requirement depending on the rule that applies. For instance, in the verbal category (S\NP)/NP, the subcategory S\NP functions as a resource when NPs are type raised (it satisfies the requirement of a type raised NP with category S/(S\NP)). However it will need to be further decomposed into a resource and a requirement if they are not. More in general, polarity specification in CCG would need to take into account the several combination rules in addition to the category structure. In HPSG, it is the interaction of lexical categories with lexical and phrasal rules that will need to be taken into consideration.</Paragraph>
    </Section>
    <Section position="2" start_page="101" end_page="101" type="sub_section">
      <SectionTitle>
6.2 Using rule types
</SectionTitle>
      <Paragraph position="0"> The two types of tree combining operations permitted by TAG can be used to structure the surface realisation algorithm. As we've shown, performing all substitutions before allowing for adjunction greatly reduces the exponential impact of intersective modifiers. Moreover, combining such a substitution-before-adjunction strategy with polarity filtering further improves performance.</Paragraph>
      <Paragraph position="1"> In comparison, the HPSG and the CCG approach do not support such a natural structuring of the algorithm and intersective modifiers induce either a pre- or a post-processing.</Paragraph>
      <Paragraph position="2"> In HPSG, intersective modifiers are discarded during the chart generation phase and adjoined into the generated structures at a later stage. This is inelegant in that (i) intersective modifiers are artificially treated separately and (ii) structures sub-ject to adjunction have to be non monotonically recomputed to reflect the impact of the adjunction in that part of the tree dominating the adjunction.</Paragraph>
      <Paragraph position="3"> In CCG, the input logical form is chunked into subtrees each corresponding to a separate generation subproblem to be solved independently.</Paragraph>
      <Paragraph position="4"> Again the approach is ad hoc in that it does not rely on a given grammatical or linguistic property.</Paragraph>
      <Paragraph position="5"> As a result, e.g., negation needs special treatment to avoid incompleteness (if the heuristic applies, negated sentences cannot be generated). Similarly, it is unclear how long distance dependencies involving modifiers (e.g., Which office did you say that Peter work in ?) are handled.</Paragraph>
    </Section>
    <Section position="3" start_page="101" end_page="101" type="sub_section">
      <SectionTitle>
6.3 Using TAG extended domain of locality
</SectionTitle>
      <Paragraph position="0"> TAG extended domain of locality means that empty semantic items need no special treatment.</Paragraph>
      <Paragraph position="1"> In contrast, both the HPSG and the CCG approach resort to ad hoc filtering rules which, based on a scan of the input semantics, add semantically empty items to the chart.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>