<?xml version="1.0" standalone="yes"?>
<Paper uid="E95-1014">
  <Title>Corpus-based Method for Automatic Identification of Support Verbs for Nominalizations</Title>
  <Section position="7" start_page="101" end_page="102" type="concl">
    <SectionTitle>
6 Conclusions
</SectionTitle>
    <Paragraph position="0"> Nominalization is a very productive process.</Paragraph>
    <Paragraph position="1"> The proper choice of collocational support verbs for nominalizations in English is a difficult task for language learners given the unpredictability of the semantically emptied verb that fulfills the syntactic role. Given a robust parser and large corpus, the simple technique of extracting the most common verbs for which the nominalized form is the direct object is not always sufficient, since completely deverbal concrete noun uses share the same lexical surface form. Comparing argument/adjunct structures involving the verbal uses of the predicate and using the most common of these structures as filters on the surface forms possibly corresponding to nominalizations captures the linguistic fact that nominalizations retain the syntactic structures of their underlying predicate. When these filters are applied, the most common supporting verb in the corpus for the recognized nominalized patterns seems to correspond to native speakers' intuition of the support verb associated with the nominalization.</Paragraph>
    <Paragraph position="2"> The experiment described here on a 134</Paragraph>
    <Paragraph position="4"> most common main verbs make (116 cases), begin(37), launch(36)</Paragraph>
    <Paragraph position="6"> using the syntactic structure filtering mechanism.</Paragraph>
    <Paragraph position="7"> megabyte corpus of newspaper text from which was extracted evidence for appeal and other predicates shows how this automated procedure can be applied to any verbnominalization pair given a large corpus and a robust parser. Other work in automated support verb discovery using bilingual dictionaries as a source has been reported in Fontenelle (1993).</Paragraph>
    <Paragraph position="8"> It remains to be seen whether these statistical results are more useful to lexicographers than their more traditional tools of key-wordin-context files and T-score measures. Human experiments would be necessary to demonstrate this. Another useful test of the results would be to compare the results given by this technique against machine-readable dictionary-derived data.</Paragraph>
    <Paragraph position="9"> In conclusion, the interest of this technique is its general approach to corpus linguistics as one of multiple passes over the same corpus material, using results of previous passes to filter and refine data extracted on subsequent passes. We believe that this approach, coupled with lexical resources and robust parsers, offers much promise for the future of corpus exploitation.</Paragraph>
  </Section>
class="xml-element"></Paper>