<?xml version="1.0" standalone="yes"?>
<Paper uid="J94-4005">
  <Title>Training and Scaling Preference Functions for Disambiguation Hiyan Alshawi * AT&amp;T Bell Laboratories</Title>
  <Section position="8" start_page="645" end_page="646" type="concl">
    <SectionTitle>
7. Conclusion
</SectionTitle>
    <Paragraph position="0"> We have presented a relatively simple analytic technique for automatically determining a set of scaling factors for preference functions used in semantic disambiguation. The initial scaling factors produced are optimal with respect to a score provided by a  Hiyan Alshawi and David Carter Training and Scaling Preference Functions training procedure and are further improved by comparison with instances of the task they are intended to perform. The experimental results presented indicate that, by using a fairly crude training score measure (comparing only phrase structure trees) with a few thousand training sentences, the method can yield a set of scaling factors that are significantly better than those derived by a labor-intensive hand-tuning effort. We have also confirmed empirically that considerable differences exist between the effectiveness of differently formulated collocation functions for disambiguation. The experiments provide a basis for selecting among different collocational functions and suggest that a collocation function must be evaluated in the context of other functions, rather than on its own, if the correct selection is to be made.</Paragraph>
    <Paragraph position="1"> It should be possible to extend this work fruitfully in several directions, including the following. Training with a measure defined directly on semantic representations is likely to lead to a further reduction in the disambiguation error rate. The method for computing scaling factors described here has more recently been applied to optimizing preference selection for the task of choosing between analyses arising from different word hypotheses in a speech recognition system (Rayner et al. 1994) and is applicable to other problems, such as choosing between possible target representations in a machine translation system. Finally, it would be interesting to combine the work on semantic collocation functions with that on similarity-based clustering (Pereira, Tishby, and Lee 1993; Dagan, Marcus, and Markovitch 1993), with the aim of overcoming the problem of sparse training data. If this is successful, it might make these functions suitable for disambiguation in domains with larger vocabularies than ATIS.</Paragraph>
  </Section>
class="xml-element"></Paper>