<?xml version="1.0" standalone="yes"?>
<Paper uid="A92-1022">
  <Title>References</Title>
  <Section position="9" start_page="160" end_page="160" type="concl">
    <SectionTitle>
7 Conclusion
</SectionTitle>
    <Paragraph position="0"> The availability of a substantial file of parses conforming to a relatively well-specified standard affords a number of opportunities for interesting computational linguistic research. One such opportunity is the evaluation of both grammars and parsing strategies, such as we have shown in this paper. We believe that such evaluation will be crucial to the further development of efficient, broad-coverage syntactic analyzers. By restructuring the generated parse trees to conform to those of the standard, and using agreed-upon comparison metrics, it should be possible to make meaningful comparisons across systems, even those using very different strategies.</Paragraph>
    <Paragraph position="1"> We intend to apply our evaluation methods to additional parsing heuristics in the near future. In particular, we noted earlier that selectional constraints are implemented as penalties in our system (preference semantics), and these evaluation methods can assist us in obtaining a proper balance of syntactic and semantic penalties. null By extending our approach, it should be possible to produce more detailed diagnostic information about the parser output. For example, by stripping away some of the structure from both the standard and the system-generated parses, it would be possible to focus on the system's performance on NP bracketing or on S bracketing. Such an approach should also allow for meaningful comparisons with some systems (e.g., some fast deterministic parsers) which generate only more local structures and postpone attachment decisions.</Paragraph>
  </Section>
class="xml-element"></Paper>