<?xml version="1.0" standalone="yes"?>
<Paper uid="P03-1055">
  <Title>Deep Syntactic Processing by Combining Shallow Methods</Title>
  <Section position="9" start_page="0" end_page="0" type="concl">
    <SectionTitle>
7 Conclusions
</SectionTitle>
    <Paragraph position="0"> This paper has three main contributions. First, we show that gap+ features, encoding necessary information for antecedent recovery, do not incur any substantial computational overhead.</Paragraph>
    <Paragraph position="1"> Second, the paper demonstrates that a shallow finite-state model can be successful in detecting sites for discontinuity, a task which is generally understood to require deep syntactic and lexical-semantic knowledge. The results show that, at least in English, local clues for discontinuity are abundant.</Paragraph>
    <Paragraph position="2"> This opens up the possibility of employing shallow finite-state methods in novel situations to exploit non-apparent local information.</Paragraph>
    <Paragraph position="3"> Our final contribution, but the one we wish to emphasize the most, is that the combination of two orthogonal shallow models can be successful at solving tasks which are well beyond their individual power. The accent here is on orthogonality - the two models take different sources of information into account. The tagger makes good use of adjacency at the word level, but is unable to handle deeper recursive structures. A context-free grammar is better at finding vertical phrase structure, but cannot exploit linear information when words are separated by phrase boundaries. As a consequence, the finite-state method helps the parser by efficiently and reliably pruning the search-space of the more complicated PCFG model. The benefits are immediate: the parser is not only faster but more accurate in recovering antecedents. The real power of the finite-state model is that it uses information the parser cannot.</Paragraph>
  </Section>
class="xml-element"></Paper>