<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-0710">
  <Title>Aligning WordNet with Additional Lexical Resources</Title>
  <Section position="3" start_page="73" end_page="74" type="metho">
    <SectionTitle>
2 Integrating Different Resources
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="73" end_page="74" type="sub_section">
      <SectionTitle>
2.1 Relations between Resources
</SectionTitle>
      <Paragraph position="0"> The three lexic',d resources we used are the 1987 revision of Roger's Thesaurus (ROGET) (Kirkpatrick, 1987), the Longman Dictionary of Contemporary English (LDOCE) (Procter, 1978) and the Prolog version of WordNet 1.5 (WN) (Miller et al., 1993).</Paragraph>
      <Paragraph position="1"> The linking of LDOCE and WN is in principle quite similar to Knight and Luk's (1994) approach in the PANGLOSS project. But our form of comparison between LDOCE and WN, motivated by the organisation of individual resources in relation to one another, was simpler but as effective. Figure 1 shows how word senses are organised in the three resources and the arrows indicate the direction of mapping.</Paragraph>
      <Paragraph position="2"> In the middle of the figure is the structure of WN, a hierarchy with the nodes formed from the  synsets. If we now look up ROGET for word x2 in synset X, since words expressing every aspect of an idea are grouped together in ROGET, we can expect to find not only words in synset X, but also those in the coordinate synsets (i.e. M and P, with words ml, m2, Pl, P2, etc.) and the superordinate synsets (i.e. C and A, with words cl, c2, etc.) in the same ROGET paragraph. In other words, the thesaurus class to which x2 belongs should roughly include X U M U P U C t3 A. On the other hand, the LDOCE definition corresponding to the sense of synset X (denoted by Dz) is expected to be similar to the textual gloss of synset X (denoted by GI(X)).</Paragraph>
      <Paragraph position="3"> Nevertheless, given that it is not unusual for dictionary definitions to be phrased with synonyms or superordinate terms, we would also expect to find words from X and C, or even A, in the LDOCE definition. That means we believe D~: ~ Gl(X) and Dzn(XUCUA) ~ C/. We did not include coordinate terms (called &amp;quot;siblings&amp;quot; in Knight and Luk (1994)) because we found that while nouns in WN usually have many coordinate terms, the chance of hitting them in LDOCE definitions is hardly high enough to worth the computation effort.</Paragraph>
    </Section>
    <Section position="2" start_page="74" end_page="74" type="sub_section">
      <SectionTitle>
2.2 The Algorithm
</SectionTitle>
      <Paragraph position="0"> Our algorithm defines a mapping chain from LDOCE to ROGET through WN. It is based on shallow processing within the resources themselves, exploiting their inter-relatedness, and does not rely on extensive statistical data (e.g. as suggested in Yarowsky (1992)). Given a word with part of speech, W(p), the core steps are as follows: Step 1: From LDOCE, get the sense definitions Dr, ..., Dt under the entry W(p).</Paragraph>
      <Paragraph position="1"> Step 2: From WN, find all the synsets Sn{wt,w2,...} such that W(p) E Sn. Also collect the corresponding gloss definitions, Gl(Sn), if any, the hypernym synsets Hyp(S,~), and the coordinate synsets Co(S,~).</Paragraph>
      <Paragraph position="2"> Step 3: Compute a similarity score matrix ,4 for the LDOCE senses and the WN synsets. A similarity score A(i, j) is computed for the i th LDOCE sense and the jta WN synset using a weighted sum of the overlaps between the LDOCE sense and the WN synset, hypernyms, and gloss respectively, that is</Paragraph>
      <Paragraph position="4"> For our tests, we just tried setting at = 3, a., = 5 and a3 -- 2 to reveal the relative significance of finding a hypernym, a synonym, and any word in the textual gloss respectively in the dictionary definition.</Paragraph>
      <Paragraph position="5"> A 120. N. cl. cZ ... (in C); ml. m2.... (in M): pl. p?.. B C ... (in P): xl, x2,... (ia X) ~ ~ (~l.c2.... I, Ol(C'3</Paragraph>
      <Paragraph position="7"> Step 5: Compute a similarity score matrix B for the WN synsets and the ROGET classes. A similarity score B(j, k) is computed for the jth WN synset (taking the synset itself, the hypernyms, and the coordinate terms) and the k th ROGET class, according to the following:</Paragraph>
      <Paragraph position="9"> We have set bl = b2 = b3 = 1. Since a ROGET class contains words expressing every aspect of the same idea, it should be equally likely to find synonyms, hypernyms and coordinate terms in common.</Paragraph>
      <Paragraph position="10"> Step 6: For i = 1 to t (i.e. each LDOCE sense), find max(.A(i,j)) from matrix ,4. Then trace from matrix B the jth row and find max(B(j,k)).</Paragraph>
      <Paragraph position="11"> The i th LDOCE sense should finally be mapped to the ROGET class to which Pk belongs.</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="74" end_page="75" type="metho">
    <SectionTitle>
3 Testing.
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="74" end_page="74" type="sub_section">
      <SectionTitle>
3.1 Materials
</SectionTitle>
      <Paragraph position="0"> Three groups of test words (at present we only test on nouns), each containing 12 random samples, were prepared. Words have five or fewer WN senses in the &amp;quot;low polysemy group&amp;quot; (LO), between six to ten in the &amp;quot;medium polysemy group&amp;quot; (MED) , and 11 or more in the &amp;quot;high polysemy group&amp;quot; (HI). Table 1 shows the list of test words with the number of senses they have in the various lexical resources.</Paragraph>
    </Section>
    <Section position="2" start_page="74" end_page="75" type="sub_section">
      <SectionTitle>
3.2 Design and Hypotheses
</SectionTitle>
      <Paragraph position="0"> Our investigation was divided into three parts.</Paragraph>
      <Paragraph position="1"> While the application of the algorithm was in the third part, the first two parts were preliminaries to gather some information about the three resources so as to give a guide for expecting how well the mapping algorithm would work and such information would also help explain the results.</Paragraph>
      <Paragraph position="2"> Part 1: First, we just consider whether the bulk of information, measured by a simple count of the number of senses, for each word captured by different resources is similar in size. Basically if it was not, it would mean that words are treated too differently in terms of sense distinction in different resources for the mapping to succeed.</Paragraph>
      <Paragraph position="3"> A crude way is to look for some linear relationship for the number of senses per word in different resources. If WN is as linguistically valid as other lexical databases, we would expect strong positive correlation with other resources for the &amp;quot;amount&amp;quot; of information.</Paragraph>
      <Paragraph position="4"> Part 2: Second, we look for some quantitative characterisation of the relationship we find for the resources in Part 1. We achieve this by performing some paired-sample t-tests. If the resources do in fact capture similar &amp;quot;amount&amp;quot; of information, we would not expect to find any statistically significant difference for the mean number of senses per word among different resources.</Paragraph>
      <Paragraph position="5"> Part 3: Third, we now apply the mapping algorithm and try to relate the results to the information regarding the resources found in the previous parts. We map LDOCE senses to WN synsets, and WN synsets to ROGET classes (i.e.</Paragraph>
      <Paragraph position="6"> we skip step 6 in this study).</Paragraph>
      <Paragraph position="7"> We first analyse the results by looking at mapping accuracy and failure, and secondly by characterising the relationship between the two pairs of source and target (i.e. LDOCE and WN, WN and ROGET) by means of the Polysemy Factor P. This measure reflects the granularity of the senses in the source resource (S) with respect to those in the target resource (7&amp;quot;), and is defined as follows:</Paragraph>
      <Paragraph position="9"/>
      <Paragraph position="11"> T' is therefore the ratio between the minimum number of target senses covering all mapped source senses and all mapped source senses. In other words, the smaller this number, the more fine-grained are the senses in S with respect to 7&amp;quot;. It tells us whether what is in common between the source and the target is similarly distingnished, and whether any excess information in the source (as found in Part 2) is attributable to extra fineness of sense distinction.</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="75" end_page="75" type="metho">
    <SectionTitle>
4 Results
</SectionTitle>
    <Paragraph position="0"> the number of senses per word in different resources. The upper half of the matrix shows the overall correlations while the lower half shows correlations within individual groups of test words. (An asterisk denotes statistical significance at the .05 level.) Significant positive correlations are found in general, and also for the LO group between WN and the other two resources. Such abstract relations do not necessarily mean simple sameness, and the exact numerical difference is found in Part 2.</Paragraph>
    <Paragraph position="1">  word in the three resources are shown in Table 3. LDOCE has an average of 1.33 senses more than WN for the LO Group, and this difference is statistically significant (t = 3.75, p &lt; .05). Also, a significant difference is found in the HI group between WN and ROGET, with the latter having 3.83 senses fewer on average  analysed the various types of possible mapping errors. They are summarised in Table 4. Incorrectly Mapped and Unmapped-a are both &amp;quot;misses&amp;quot;, whereas Forced Error and Unmappedb are both &amp;quot;false alarms&amp;quot;. These errors are manifested either in a wrong match or no match at all. The performance of the algorithm on the three groups of nouns is shown in Table 5.</Paragraph>
    <Paragraph position="2">  Statistics from one-way ANOVA show that for the mapping between LDOCE and WN, there are significantly more Forced Errors in the HI group than both the LO and MED group (F = 7.58, p &lt; .05).</Paragraph>
    <Paragraph position="3"> For the mapping between WN and ROGET, the MED group has significantly more Incorrectly Mapped than the LO group (F = 3.72, p &lt; .05).</Paragraph>
    <Paragraph position="4"> Also, there are significantly more Forced Errors in the HI group than the LO and MED group  As far as the Polysemy Factor is concerned, it is found to be significantly lower in the HI group than the LO group (F = 3.63, p &lt; .05) for the mapping between WN and ROGET.</Paragraph>
  </Section>
class="xml-element"></Paper>