<?xml version="1.0" standalone="yes"?>
<Paper uid="H92-1063">
  <Title>A NEW APPROACH TO TEXT UNDERSTANDING</Title>
  <Section position="6" start_page="319" end_page="319" type="concl">
    <SectionTitle>
4. CONCLUSIONS
</SectionTitle>
    <Paragraph position="0"> PLUM has the following key features:  1. Deterministic parsing and semantic  interpretation of all fragments produced. .</Paragraph>
    <Paragraph position="1"> .</Paragraph>
    <Paragraph position="2"> Event-based and template-based knowledge to find relations among entities when syntax/semantics cannot fmd them. Statistical language models at multiple levels.</Paragraph>
    <Paragraph position="3"> These were key to PLUM performing among the top systems evaluated in MUC-3. Because of the focus on producing syntactic and semantic analyses of fragments when no complete analysis was possible, and because of the assumption that discourse processing can fit the fragments together based on required roles in defined event structures, the system can produce answers end-to-end very early on when porting to a new domain, long before domain-specific lexical items and any domain-specific semantic rules are complete.</Paragraph>
    <Paragraph position="4"> That conclusion is supported quantitatively by our experiments. Recall, the percent of information in the text correctly extracted, grew nearly linearly as the lexicon grew in the experiment. Precision, the percent correct information extracted of information output by PLUM, remained flat as a function of lexicon size, also supporting the intuition that the lexicon was declarative and separate from rule-based processing.</Paragraph>
    <Paragraph position="5"> A second conclusion is that deterministic parsing can be supplemented by locally applied, fragment combining rules that use both the syntactic and semantic properties of fragments produced to resolve ambiguity that syntax alone can not resolve in a deterministic parser. The experiment reported here demonstrates that.</Paragraph>
    <Paragraph position="6"> The degree of success obtained by marrying domainindependent, deterministic parsing with partial understanding and statistical techniques has been quite gratifying. The techniques which seemed so incompatible and antithetical in the seventies have proven synergistic.</Paragraph>
  </Section>
class="xml-element"></Paper>