<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1037">
  <Title>Guiding a Constraint Dependency Parser with Supertags</Title>
  <Section position="3" start_page="0" end_page="289" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Supertagging is based on the combination of two powerful and influential ideas of natural language processing: On the one hand, parsing is (at least partially) reduced to a decision on the optimal sequence of categories, a problem for which efficient and easily trainable procedures exist. On the other hand, supertagging exploits complex categories, i.e. tree fragments which much better reflect the mutual compatibility between neighbouring lexical items than say part-of-speech tags.</Paragraph>
    <Paragraph position="1"> Bangalore and Joshi (1999) derived the notion of supertag within the framework of Lexicalized Tree-Adjoining Grammars (LTAG) (Schabes and Joshi, 1991). They considered supertagging a process of almost parsing, since all that needs to be done after having a sufficiently reliable sequence of supertags available is to decide on their combination into a spanning tree for the complete sentence. Thus the approach lends itself easily to pre-processing sentences or filtering parsing results with the goal of guiding the parser or reducing its output ambiguity.</Paragraph>
    <Paragraph position="2"> Nasr and Rambow (2004) estimated that perfect supertag information already provides for a parsing accuracy of 98% if a correct supertag assignment were available. Unfortunately, perfectly reliable supertag information cannot be expected; usually this uncertainty is compensated by running the tagger in multi-tagging mode, expecting that the reliability can be increased by not forcing the tagger to take unreliable decisions but instead offering a set of alternatives from which a subsequent processing component can choose.</Paragraph>
    <Paragraph position="3"> A grammar formalism which seems particularly well suited to decompose structural descriptions into lexicalized tree fragments is dependency grammar. It allows us to define supertags on different levels of granularity (White, 2000; Wang and Harper, 2002), thus facilitating a fine grained analysis of how the different aspects of supertag information influence the parsing behaviour. In the following we will use this characteristic to study in more detail the utility of different kinds of supertag information for guiding the parsing process.</Paragraph>
    <Paragraph position="4"> Usually supertags are combined with a parser in a filtering mode, i.e. parsing hypotheses which are not compatible with the supertag predictions are simply discarded. Drawing on the ability of Weighted Constraint Dependency Grammar (WCDG) (Schr&amp;quot;oder et al., 2000) to deal with defeasible constraints, here we try another option for making available supertag information: Using a score to estimate the general reliability of unique supertag decisions, the information can be combined with evidence derived from other constraints of the grammar in a soft manner. It makes possible to rank parsing hypotheses according to their plausibility and allows the parser to even override potentially wrong supertag decisions.</Paragraph>
    <Paragraph position="5"> Starting from a range of possible supertag models, Section 2 explores the reliability with which dependency-based supertags can be determined on  different levels of granularity. Then, Section 3 describes how supertags are integrated into the existing parser for German. The complex nature of supertags as we define them makes it possible to separate the different structural predictions made by a single supertag into components and study their contributions independently (c.f. Section 4). We can show that indeed the parser is robust enough to tolerate supertag errors and that even with a fairly low tagger performance it can profit from the additional, though unreliable information.</Paragraph>
  </Section>
class="xml-element"></Paper>