<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-1010">
  <Title>A Plethora of Methods for Learning English Countability</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Discussion
</SectionTitle>
    <Paragraph position="0"> There have been at least three earlier approaches to the automatic determination of countability: two using semantic cues and one using corpora. Bond and Vatikiotis-Bateson (2002) determine a noun's countability preferences--as defined in a 5-way classification--from its semantic class in the ALT-J/E lexicon, and show that semantics predicts countability 78% of the time. O'Hara et al. (2003) implemented a similar approach using the much larger Cyc ontology and achieved 89.5% accuracy, mapping onto the 2 classes of countable and uncountable. Schwartz (2002) learned noun countabilities by looking at determiner occurrence in singular noun chunks and was able to tag 11.7% of BNC noun tokens as countable and 39.5% as uncountable, achieving a noun type agreement of 88% and 44%, respectively, with the ALT-J/E lexicon. Our results compare favourably with each of these.</Paragraph>
    <Paragraph position="1"> In a separate evaluation, we took the best-performing classifier (Dist(AllCON,SUITE)) and ran it over open data, using best-500 feature selection (Baldwin and Bond, 2003). The output of the classifier was evaluated relative to hand-annotated data, and the level of agreement found to be around 92.4%, which is approximately equivalent to the agreement between COMLEX and ALT-J/E of 93.8%.</Paragraph>
    <Paragraph position="2"> In conclusion, we have presented a plethora of learning techniques for deep lexical acquisition from corpus data, and applied each to the task of classifying English nouns for countability. We specifically compared two feature representations, based on relative feature occurrence and token-level classification, and two basic classifier architectures, using a suite of binary classifiers and a single multi-class classifier. We also analysed the effects of combing the output of multiple pre-processors, and presented a simple feature selection method. Overall, the best results were obtained using a distribution-based suite of binary classifiers combining the output of multiple pre-processors.</Paragraph>
  </Section>
class="xml-element"></Paper>