<?xml version="1.0" standalone="yes"?>
<Paper uid="N04-4009">
  <Title>Competitive Self-Trained Pronoun Interpretation</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> To conclude, a pronoun interpretation system can be trained solely on raw data using a standard set of morphosyntactic features to achieve performance that approaches that of a state-of-the-art supervised system. Although the self-acquired training data is no doubt highly noisy, the resulting model is still accurate enough to perform well at selecting correct antecedents. As a next step, we will take a closer look at the training data acquired to try to ascertain 3TDT segment 14, which is smaller than the others, provided only about 3800 pronouns in the runs corresponding to the last two rows of Table 1. The overall average performance figures are the same to the first decimal place whether or not the results from this segment are included.</Paragraph>
    <Paragraph position="1"> the underlying reasons for this success.</Paragraph>
    <Paragraph position="2"> There are also a number of variants of the algorithm that could be pursued. For instance, whereas our algorithm uses the current model's probabilities in a winner-take-all strategy for positive example selection, these probabilities could instead be used to dictate the likelihood that examples are assigned a positive outcome, or they could be thresholded in various ways to create a more discerning positive outcome assignment mechanism. Such strategies would avoid the current simplification of assigning a positive outcome to exactly one potential antecedent for each pronoun.</Paragraph>
    <Paragraph position="3"> The relative generality of our feature set was appropriate given the size of the data sets used. The availability of very large raw corpora, however, creates the prospect of using self-training with considerably more fine-grained features than is possible in a supervised scenario, due to the relative infrequency with which they would be found in any corpus of a size that could be feasibly annotated manually. It is thus at least conceivable that a self-trained approach, coupled with a large set of features and a large corpus of raw data, could eventually overtake the performance of the best supervised models.</Paragraph>
  </Section>
class="xml-element"></Paper>