<?xml version="1.0" standalone="yes"?>
<Paper uid="J02-3004">
  <Title>c(c) 2002 Association for Computational Linguistics The Disambiguation of Nominalizations</Title>
  <Section position="9" start_page="385" end_page="386" type="concl">
    <SectionTitle>
8. Conclusions
</SectionTitle>
    <Paragraph position="0"> In this article we presented work on the automatic interpretation of nominalizations (i.e., compounds whose heads are derived from a verb and whose modifiers are interpreted as its arguments). Nominalizations pose a challenge for empirical approaches, as the argument relations between a head and its modifier are not readily available in a corpus, and therefore they have to be somehow retrieved and approximated. Approximating the nominalized head to its corresponding verb and estimating the frequency of verb-noun relations instead of noun-noun relations accounts for only half of the nominalizations attested in the corpus.</Paragraph>
    <Paragraph position="1"> Our experiments revealed that data sparseness can be overcome by taking advantage of smoothing methods and surface contextual information. We have directly compared and contrasted a variety of smoothing approaches proposed in the literature  Computational Linguistics Volume 28, Number 3 and have shown that these methods yield satisfactory results for the demanding task of semantic disambiguation, especially when coupled with contextual information. Our experiments have shown that contextual information that is easily obtainable from a corpus and computationally cheap is good at predicting object relations, whereas the computationally more expensive smoothing variants are better at guessing subject relations. Combination of context with smoothing variants yields better performance over either context or smoothing alone.</Paragraph>
    <Paragraph position="2"> We combined different information sources (i.e., contextual features and smoothing variants) using Ripper. Although a considerable body of previous research has treated several linguistic phenomena as classification tasks, the interpretation of compound nouns has so far been based on the availability of symbolic knowledge. We show that the application of probabilistic learning to the interpretation of compound nouns is novel and promising. Finally, our experiments revealed that information inherent in the corpus can make up for the lack of distributional evidence by taking advantage of smoothing methods that rely simply on verb-argument tuples extracted from a large corpus and surface contextual information without strictly presupposing the existence of annotated data or taxonomic information.</Paragraph>
  </Section>
class="xml-element"></Paper>