<?xml version="1.0" standalone="yes"?>
<Paper uid="C04-1190">
  <Title>Semi-Supervised Training of a Kernel PCA-Based Model for Word Sense Disambiguation</Title>
  <Section position="8" start_page="0" end_page="0" type="evalu">
    <SectionTitle>
6 Results
</SectionTitle>
    <Paragraph position="0"> Table 4 shows that the composite semi-supervised KPCA model improves on the high-performance supervised KPCA model, for both coarse-grained and fined-grained sense distinctions. The supervised KPCA model significantly outperforms a na&amp;quot;ive Bayes model, and a maximum entropy model, which are among the top performing models for WSD. Note that these results are consistent with the larger study of supervised models conducted by Wu et al. (2004). The composite semi-supervised KPCA model outperforms all of the three supervised models, and in particular, it further improves the  accuracy of the supervised KPCA model.</Paragraph>
    <Paragraph position="1"> Overall, with the addition of the semi-supervised model, the accuracy for disambiguating the verbs increases from 57% to 57.4% on the fine-grained task, and from 66.6% to 67.2% on the coarse-grained task.</Paragraph>
    <Paragraph position="2"> In our composite model, the supervised KPCA model predicts senses with high confidence for more than 94% of the test instances. The predictions of the semi-supervised model are used for the remaining 6% of the test instances. Table 5 shows that it is not necessary to use the semi-supervised training model for all the training instances. In fact, when the supervised model is confident, its predictions are significantly more accurate than those of the semi-supervised model alone.</Paragraph>
    <Paragraph position="3"> When the predictions of the supervised KPCA model are not accurate, the semi-supervised KPCA model out-performs the supervised model. This happens when (1) there is no training instance that is very similar to the test instance considered and when (2) in the absence of relevant features to learn from in the small annotated training set, the supervised KPCA model can only predict the most frequent sense for the current target. In these conditions, our experiment results in Table 6 confirm that the semi-supervised KPCA model benefits from the large additional training data, suggesting it is able to learn useful feature conjunctions, which help to give better predictions. null The composite semi-supervised KPCA model therefore chooses the best model depending on the degree of confidence of the supervised model. All the KPCA weights, for both the supervised and the semi-supervised model, have been pre-computed during training, and it is therefore inexpensive to switch from one model to the other at testing time.</Paragraph>
  </Section>
class="xml-element"></Paper>