<?xml version="1.0" standalone="yes"?>
<Paper uid="E85-1008">
  <Title>Effective Parsing With Generalised Phrase Structure Grammar</Title>
  <Section position="3" start_page="60" end_page="60" type="concl">
    <SectionTitle>
5. Conclusions
</SectionTitle>
    <Paragraph position="0"> The parser and rule learner described above perform far far better than the parser by itself - on complex cases, the parser may find the correct analysis several hundred times as quickly using learnt rules as it would have done with Just the basic set.</Paragraph>
    <Paragraph position="1"> Experience with the system to date indicates that the introduction of new rules does not slow down the process of selecting relevant rules all that much, partly because the indexing of patterns against initial elements cuts out quite a lot of potentially pointless searching. It is conceivable that when the system has been run on large numbers of examples, the gains introduced by abstracting over long, unusual strings will be outweighed by the extra effort involved in testing for them when they are not relevant. If so, it may be a good idea to put a limit on the length of string for which compound rules should be recorded. There is no indication as yet that this will be necessary.</Paragraph>
    <Paragraph position="2"> It is of interest that the compound rules the system creates are akin to the productions used in Marcus' deterministic parser \[Marcus\] - patterns of descriptions of items which the parser is prepared to react to, combined with packets of simple actions to be taken when a pattern is recognised. There is no suggestion here that the system described above could ever be fully deterministic - there are Just too many possibilities to be explored for this to be likely but it certainly explores fewer dead ends with learnt compound rules than with the initial basic ones.</Paragraph>
  </Section>
class="xml-element"></Paper>