<?xml version="1.0" standalone="yes"?>
<Paper uid="P93-1023">
  <Title>TOWARDS THE AUTOMATIC IDENTIFICATION OF ADJECTIVAL SCALES: CLUSTERING ADJECTIVES ACCORDING TO MEANING Vasileios Hatzivassiloglou</Title>
  <Section position="10" start_page="179" end_page="180" type="concl">
    <SectionTitle>
6. CONCLUSIONS AND
FUTURE WORK
</SectionTitle>
    <Paragraph position="0"> We have described a system for extracting groups of semantically related adjectives from large text corpora, with a flexible architecture which allows for multiple knowledge sources influencing similarity to  be easily incorporated into the system. Our evaluation reveals that it has significantly high performance levels, comparable to humans, using only a relatively small amount of input data; in addition, it shows the usefulness of negative knowledge, an original feature of our approach. The system's results can be filtered to produce scalar adjectives that are applicable in any given domain. Furthermore, while we have demonstrated the algorithm on adjectives, it can be directly applied to other word classes once sources of linguistic information for judging their similarity have been identified.</Paragraph>
    <Paragraph position="1"> Our immediate plans are to incorporate more similarity modules into stage two of the system and add a training component to stage three so that the relative weights of the various modules can be estimated. We have identified several additional sources of linguistic knowledge which look promising, namely pairs of adjectives separated by connectives and adverb-adjective pairs. We also plan to extend the adjective-noun module to cover adjectives in predicative positions, in addition to our current use of attributive adjectives. These extensions not only will provide us with a better way of exploiting the information in the corpus but may also help us categorize the adjectives as relational or attributive (Levi, 1978); such a categorization may be useful in classifying them as either scalar or non-scalar. For determining whether a group of adjectives is scalar, we also plan to use the gradability of the adjectives as observed in the corpus. In addition, we are exploring tests for determining whether two adjectives are antonymous, essentially in the opposite direction of the work by Justeson and Katz (1991) , and tests for comparing the relative semantic strength of two adjectives.</Paragraph>
    <Paragraph position="2"> Furthermore, we plan to consider alternative evaluation methods and test our system on a much larger set of adjectives. That was not done for the current evaluation because of the difficulty for humans of constructing large models. We are considering an evaluation method which would use a thesaurus to judge similarity, as well as a supplementary method based on mathematical properties of the clustering. Neither of these methods would access any human models. The mathematical method, which uses cluster silhouettes and the silhouette coefficient (Kaufman and Rousseeuw, 1990), can also be used to automatically determine the proper number of clusters, one of the hardest problems in cluster analysis. We also plan a formal study to evaluate the appropriateness of the clustering method used, by computing and evaluating the results when a hierarchical algorithm is employed instead in stage four. Eventually, we plan to evaluate the system's output by using it to augment adjective entries in a lexicon and test the augmented lexicon in an application such as language generation.</Paragraph>
  </Section>
class="xml-element"></Paper>