<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-2602">
  <Title>Constraint Satisfaction Inference: Non-probabilistic Global Inference for Sequence Labelling</Title>
  <Section position="8" start_page="45" end_page="45" type="concl">
    <SectionTitle>
7 Conclusion
</SectionTitle>
    <Paragraph position="0"> Theclassification and inference approach is apopular and effective framework for performing sequence labelling in tasks where there is strong interaction between output labels. Most existing inference procedures expect a base classifier that makes probabilistic predictions, that is, rather than predicting asingle class label, aconditional probability distribution over the possible classes is computed. The inference procedure presented in this paper is different in the sense that it can be used with any classifier that is able to estimate a confidence score for its (non-probabilistic) predictions. Constraint satisfaction inference builds upon the class trigram method introduced by Van den Bosch and Daelemans (2005), but reinterprets it as a strategy for generating multiple potential output sequences, from which it selects the sequence that has been found to be most optimal according to a weighted constraint satisfaction formulation of the inference process. In a series of experiments involving six sequence labelling task covering several different areas in natural language processing, constraint satisfaction inference has been shown to improve substantially upon the performance achieved by a simpler inference procedure based on majority voting, proposed in the original work on the class trigram method.</Paragraph>
    <Paragraph position="1"> The work presented in this paper shows there is potential for alternative interpretations of the classification and inference framework that do not rely on probabilistic base classifiers. Future work may well be able to further improve the performance of constraint satisfaction inference, for example, by using more optimised constraint weighting schemes. In addition, alternative ways of formulating constraint satisfaction problems from classifier predictions may be explored; not only for sequence labelling, but also for other domains that could benefit from global inference.</Paragraph>
  </Section>
class="xml-element"></Paper>