<?xml version="1.0" standalone="yes"?>
<Paper uid="A92-1032">
  <Title>Robust parsing of natural language descriptions</Title>
  <Section position="6" start_page="229" end_page="229" type="concl">
    <SectionTitle>
6 Discussion and conclusions
</SectionTitle>
    <Paragraph position="0"> We would like to make clear that the approach we have presented is seriously limited, by the fact that we are not able to fully exploit syntactic information. This means that complete sentences containing verbs and subordinate clauses cannot be properly handled. However, our experiments show that the approach behaves well in the field of real business descriptions.</Paragraph>
    <Paragraph position="1"> Semantic knowledge plays a fundamental role in our system, as it has to validate the syntactic relationships proposed by the shallow parsing algorithm. Currently, our knowledge base consists of a taxonomy of 3000 concepts, together with 360 semantic constraints for the conceptual relations. Although the current prototype uses a hand-written knowledge base, some techniques for semi-automatic extraction of semantic knowledge from our large corpus have been studied in our project (Velardi et al., 1991). The idea is that the non-ambiguous sentence fragments present in the corpus capture specific word usage patterns, which, via a human-controlled generalization process, may generate the kind of semantic constraints we are looking for to pu t in the knowledge base.</Paragraph>
    <Paragraph position="2"> A full syntactic parsing would be too expensive for large corpora, and it would also fail to consider useful information embedded within ill-formed sentences. The encouraging preliminary results indicate that our approach constitutes a good compromise between syntactic completeness and meaning extraction.</Paragraph>
    <Paragraph position="3"> In conclusion, the shallow parsing algorithm we have described may play a crucial role in order to trigger a bootstrapping process of knowledge acquisition, giving us some chance to overcome, at least in our domain, a major bottleneck in natural language understanding.</Paragraph>
  </Section>
class="xml-element"></Paper>