<?xml version="1.0" standalone="yes"?>
<Paper uid="W00-0709">
  <Title>Overfitting Avoidance for Stochastic Modeling of Attribute-Value Grammars</Title>
  <Section position="8" start_page="52" end_page="52" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> The feature merging strategy described in this paper may be employed to reduce overfitting in Figure 3: For the model trained on 4600 sentences, features containing elements appearing fewer than 500 times are merged. The overfitting in the unmerged model, represented by the solid line, is less drastic due to more extensive training material, but an improvement can still be seen in the curve of the merged model.</Paragraph>
    <Paragraph position="1"> situations where statistical features are built up compositionally from basic elements. As mentioned, the merging strategy bears certain similarities with other methods of overfitting reduction, such as standard feature cutoffs where entire features appearing less than some number of times are ignored (Ratnaparkhi, 1998). Intuitively, it seems that in a sparse data situation, it would be beneficial to retain the general information in features, rather than ignoring rare features entirely. It would be worthwhile to verify this suspicion by comparing the present approach directly with a simple feature cutoff, and furthermore comparing a simple cutoff to one where the low-frequency features were merged according to the present scheme, rather than simply discarded. It is to be expected that a combination of both approaches would be likely to outperform either individual approach. How much improvement may be gained remains to be seen.</Paragraph>
  </Section>
class="xml-element"></Paper>