<?xml version="1.0" standalone="yes"?>
<Paper uid="P03-1014">
  <Title>Integrated Shallow and Deep Parsing: TopP meets HPSG</Title>
  <Section position="13" start_page="82" end_page="82" type="concl">
    <SectionTitle>
BD
BE
</SectionTitle>
    <Paragraph position="0"> ), in particular, thresholded precision of bracket types PT, and tree entropy +E, with comparable speed-up of factor 2.2/2.3 and 2.27/2.23 (2.25 if averaged). Thresholded entropy ET yields slightly lower gains. This could be due to a non-optimal threshold, or the fact that - while precision differentiates bracket types in terms of their confidence, such that only a small number of brackets are weakened - tree entropy as a global measure penalizes all brackets for a sentence on an equal basis, neutralizing positive effects which - as seen in +/A0P - may still contribute useful information.</Paragraph>
    <Paragraph position="1"> Additional use of chunk brackets (row 10) leads to a slight decrease, probably due to lower precision of chunk brackets. Even more, isolated use of chunk information (row 11) does not yield signifi- null cant gains over the baseline (0.89/1.1). Similar results were reported in (Daum et al., 2003) for integration of chunk- and dependency parsing.</Paragraph>
    <Paragraph position="2">  formance gains, with some outliers in the range of length 25-36. 962 sentences (length BQ3, avg. 11.09) took longer parse time as compared to the baseline (with 5% variance margin). For coverage losses, we isolated two factors: while erroneous topological information could lead the parser astray, we also found cases where topological information prevented spurious HPSG parses to surface. This suggests that the integrated system bears the potential of cross-validation of different components.</Paragraph>
    <Paragraph position="3"> 7Conclusion We demonstrated that integration of shallow topological and deep HPSG processing results in significant performance gains, of factor 2.25--at a high level of deep parser efficiency. We show that macrostructural constraints derived from topological parsing improve significantly over chunk-based constraints. Fine-grained prioritisation in terms of confidence weights could further improve the results. Our annotation-based architecture is now easily extended to address robustness issues beyond lexical matters. By extracting spans for clausal fragments from topological parses, in case of deep parsing fail- null (Daum et al., 2003) report a gain of factor 2.76 relative to a non-PoS-guided baseline, which reduces to factor 1.21 relative to a PoS-prioritised baseline, as in our scenario.</Paragraph>
    <Paragraph position="4"> ure the chart can be inspected for spanning analyses for sub-sentential fragments. Further, we can simplify the input sentence, by pruning adjunct subclauses, and trigger reparsing on the pruned input.</Paragraph>
  </Section>
class="xml-element"></Paper>