<?xml version="1.0" standalone="yes"?>
<Paper uid="W00-0804">
  <Title>Experiments in Word Domain Disambiguation for Parallel Texts</Title>
  <Section position="7" start_page="30" end_page="32" type="evalu">
    <SectionTitle>
5 Results and Discussion
</SectionTitle>
    <Paragraph position="0"> WSD and WDD on the Semcor Brown Corpus. In the first experiment we wanted to verify that, because of the polysemy reduction induced by domain clustering, WDD is a simpler task than  WSD. For the experiment we used a subset of the Semcor corpus. As for WSD we obtained .66 of correct disambiguation with a sense frequency algorithm on polysemous noun words and .80 on all nouns (this last is also reported in the literature, for example in \[Mihalcea and Moldovan, 1999\]). As for WDD, precision has been computed considering the intersection between the word senses belonging to the domain label with the higher score and the sense tag for that word reported in Semcor. Baseline I and baseline 2, described in section 3.1, respectively gave .81 and .82 in precision, with a significant improvement over the WSD baseline, which confirms the initial hypothesis.</Paragraph>
    <Paragraph position="1"> WDD in parallel texts. In this experiment we wanted to test WDD in the context of parallel texts. Table 5 reports the precision and recall (just in case it is not I) scores for six different WDD algorithms applied to parallel English/Italian texts. Numbers refer to polysemous words only.</Paragraph>
    <Paragraph position="2"> Both the baseline algorithms perfbrm quite well: 83% for Italian and 85% for English in case of baseline 1, and 86% for both languages in case of baseline 2 are similar to the results obtained on the SemCor corpus.</Paragraph>
    <Paragraph position="3"> The algorithm which includes word domain frequency (i.e. baseline 2) reaches the highest score in both languages, indicating that the combination of domain word frequency (considered at step 1 of the algorithm) and domain text frequency (considered at step 2) is a good one. In addition, the fact that results are the same for both languages indicates that the method can smooth the coverage differences among the wordnets.</Paragraph>
    <Paragraph position="4"> We expected a better result for the bilingual extensions. The weak mutual strategy, either considering the synset intersection or the domain labels intersection, brings just minor improvements with respect to the baselines; the strong mutual strategy lowers both the precision and the recall. There are several explanations for these results.</Paragraph>
    <Paragraph position="5"> The difference in sense coverage between the two wordnets, about 30%, may affect the quality of the synset intersection: this would also explain the low degree of recall (68% for Italian and 57% for English). This is particularly evident for the strong mutual strategy, where the relative lexical poorness of the Italian synsets can strongly reduce the number of synsets in the intersection.</Paragraph>
    <Paragraph position="6"> Note also that the length of the synset intersection is about 30-40% of the mean synset number for Italian and English news respectively. This means less material which the disambiguation algorithms can take advantage of: relevant sysnsets can be left out of the intersection. For these reasons it is crucial having wordnet resources at the same level of completion to exploit the mutual help hypothesis.</Paragraph>
    <Paragraph position="7"> Furthermore, there may be a significant amount of senses which are &amp;quot;quasi&amp;quot; aligned. This may happen when two parallel senses map into close synsets, but not in the same one (e.g. one is the direct hypernym of the other). This problem could be overcome considering the IS-A relations during the computation of the intersection. In this situation it is also probable that the senses maintain the same domain label. This would explain why the domain intersection behaves better than the synset intersection (from 74%-68% to 77%-91% for the Italian and from 70%-57% to 80%-91% for the English).</Paragraph>
  </Section>
class="xml-element"></Paper>