<?xml version="1.0" standalone="yes"?>
<Paper uid="A00-2022">
  <Title>O~ Proand Retroactive Packing I \] o passive edges \</Title>
  <Section position="8" start_page="167" end_page="168" type="concl">
    <SectionTitle>
6 Conclusions
</SectionTitle>
    <Paragraph position="0"> We have presented novel algorithms for efficient subsumption checking and pro- and retroactive local ambiguity packing with large feature structures, and have provided strong empirical evidence that our approach can be applied beneficially to chart parsing with a large, broad-coverage HPSG of English.</Paragraph>
    <Paragraph position="1"> By comparison to previous work in unification-based parsing we have demonstrated that pro- and retroactive packing are well-suited to achieve optimal packing; furthermore, experimental results obtained with a publicly-available HPSG processing platform confirm that ambiguity packing can greatly reduce average parse complexity for this type of grammars.</Paragraph>
    <Paragraph position="2"> In related work, Miyao (1999) describes an approach to packing in which alternative feature structures are represented as packed, distributed disjunctions of feature structure fragments. Although the approach may have potential, the shifting of complex accounting into the unification algorithm is at variance with the findings of Kiefer et al. (1999), who report large speed-ups from the elimination of disjunction processing during unification. Unfortunately, the reported evaluation measures and lack of discussion of parser control issues are insufficient to allow a precise comparison.</Paragraph>
    <Paragraph position="3"> We intend to develop the approach presented in this paper in several directions. Firstly, we will enhance the unpacking phase to take advantage of the large number of equivalence packings we observe.</Paragraph>
    <Paragraph position="4"> This will significantly reduce the amount of work it needs to do. Secondly, many application contexts and subsequent layers of semantic processing will not require unfolding the entire parse forest; here, we need to define a selective, incremental unpacking procedure. Finally, applications like VerbMobil favour prioritized best-first rather than all-paths parsing. Using slightly more sophisticated accounting in the agenda, we plan to investigate priority  propagation in a best-first variant of our parser.</Paragraph>
  </Section>
class="xml-element"></Paper>