<?xml version="1.0" standalone="yes"?>
<Paper uid="P96-1030">
  <Title>FAST PARSING USING PRUNING AND GRAMMAR SPECIALIZATION</Title>
  <Section position="7" start_page="226" end_page="227" type="concl">
    <SectionTitle>
5 Conclusions and further directions
</SectionTitle>
    <Paragraph position="0"> Table 2 indicates that EBL and pruning each make processing about three times faster; the combination of both gives a factor of about nine. In fact, as the detailed breakdown shows, even this underestimates the effect on the main parsing phase: when both pruning and EBL are operating, processing times for other components (morphology, pruning and preferences) become the dominant ones. As we have so  the four different analysis alternatives, measured on the 500-utterance test set. The entry for a given row and column holds two figures, showing respectively the number of examples where the &amp;quot;row&amp;quot; variant produced a better translation than the &amp;quot;column&amp;quot; variant and the number where it produced a worse one. Thus for example &amp;quot;EBL+/pruning+&amp;quot; was better than &amp;quot;EBL-/pruning-&amp;quot; on 65 examples, and worse on 24.</Paragraph>
    <Paragraph position="1"> far expended little effort on optimizing these phases of processing, it is reasonable to expect substantial further gains to be possible.</Paragraph>
    <Paragraph position="2"> Even more interestingly, Table 3 shows that real system performance, in terms of producing a good translation, is significantly improved by pruning, and is not degraded by grammar specialization. (The slight improvement in coverage with EBL on is not statistically significant). Our interpretation of these results is that the technical loss of grammar coverage due to the specialization and pruning processes is more than counterbalanced by two positive effects.</Paragraph>
    <Paragraph position="3"> Firstly, fewer utterances time out due to slow processing; secondly, the reduced space of possible analyses means that the problem of selecting between different possible analyses of a given utterance becomes easier.</Paragraph>
    <Paragraph position="4"> To sum up, the methods presented here demonstrate that it is possible to use the combined pruning and grammar specialization method to speed up the whole analysis phase by nearly an order of magnitude, without incurring any real penalty in the form of reduced coverage. We find this an exciting and significant result, and are further continuing our research in this area during the coming year. In the last two paragraphs we sketch some ongoing work.</Paragraph>
    <Paragraph position="5"> All the results presented above pertain to English only. The first topic we have been investigating is the application of the methods described here to processing of other languages. Preliminary experiments we have carried out on the Swedish version of the CLE (Gamb~ick and Rayner 1992) have been encouraging; using exactly the same pruning methods and EBL chunking criteria as for English, we obtain comparable speed-ups. The loss of coverage due to grammar specialization also appears comparable, though we have not yet had time to do the work needed to verify this properly. We intend to do so soon, and also to repeat the experiments on the French version of the CLE (Rayner, Carter and Bouillon, 1996).</Paragraph>
    <Paragraph position="6"> The second topic is a more radical departure, and can be viewed as an attempt to make interleaving of parsing and pruning the basic principle underlying the CLE's linguistic analysis process. Exploiting the &amp;quot;stratified&amp;quot; nature of the EBL-specialized grammar, we group the chunked rules by level, and apply them one level at a time, starting at the bottom.</Paragraph>
    <Paragraph position="7"> After each level, constituent pruning is used to eliminate unlikely constituents. The intent is to achieve a trainable robust parsing model, which can return a useful partial analysis when no single global analysis is found. An initial implementation exists, and is currently being tested; preliminary results here are also very positive. We expect to be able to report on this work more fully in the near future.</Paragraph>
  </Section>
class="xml-element"></Paper>