<?xml version="1.0" standalone="yes"?>
<Paper uid="A00-2011">
  <Title>Word-for-Word Glossing with Contextually Similar Words</Title>
  <Section position="4" start_page="79" end_page="80" type="metho">
    <SectionTitle>
CONTEXT CONTEXTUALLY SIMILAR WORDS OF DUTY
</SectionTitle>
    <Paragraph position="0"> corporate duty fee, function, levy, liability, obligation, personnel, responsibility, rule, staff, tax, training obligation, requirement, responsibility, role fiducia~ duty the responsibility and tax senses of duty, reflecting the fact that the meaning of duty is indeed ambiguous if corporate duty is its sole context. In contrast, the second row in Table 2 clearly indicates the responsibility sense of duty. While previous word sense disambiguation algorithms rely on a lexicon to provide sense inventories of words, the contextually similar words provide a way of distinguishing between different senses of words without committing to any particular sense inventory.</Paragraph>
    <Paragraph position="1">  For example, suppose we wish to translate into French the word duty in the context corporate fiduciary duty. Step 1 retrieves the candidate translations for duty and its WATs from Figure 2. In Step 2, we construct two lists of contextually similar words, one for the dependency context corporate duty and one for the dependency context fiduciary duty, shown in  is obtained by maximizing the group similarities between the lists of contextually similar words and the WATs.</Paragraph>
    <Paragraph position="2"> Using the group similarity measure from Section 5, Table 3 lists the group similarity scores between each list of contextually similar words and each WAT as well as the final combined score for each candidate translation. The combined score for a candidate is the sum of the logs of all group similarity scores involving its WAT. The correct proposed translation for duty in this context is devoir since its WAT received the highest score.</Paragraph>
    <Paragraph position="3">  similar words of duty in corporate duty and fiduciary duty with the WATs for candidate translations devoir and taxe.</Paragraph>
  </Section>
  <Section position="5" start_page="80" end_page="80" type="metho">
    <SectionTitle>
CANDIDATE CANDIDATE
DEVOIR TAXE
</SectionTitle>
    <Paragraph position="0"> the interconnectivity and closeness measures. The interconnectivity in (a) and (b) remains constant while the closeness in (a) is higher than in (b) since there are more zero similarity pairs in (b).</Paragraph>
    <Paragraph position="1"> Input:  A word w to be translated and a set of dependency contexts involving w.</Paragraph>
    <Paragraph position="2"> Retrieve the candidate translations ofw and the corresponding WATs from the bilingual thesaurus.</Paragraph>
    <Paragraph position="3"> Find the contextually similar words of w in each dependency context using the algorithm from Section 3.</Paragraph>
    <Paragraph position="4"> Compute the group similarity (see details in Section 5) between each set of contextually similar words and each WAT; the results are stored in a matrix t, where t\[i,j\] is the group similarity between the ?h list of contextually similar words and thef h WAT.</Paragraph>
    <Paragraph position="5"> Add the logs of the group similarity scores in column oft to obtain a score for each WAT. The candidate translation corresponding to the WAT with the highest score.</Paragraph>
  </Section>
  <Section position="6" start_page="80" end_page="81" type="metho">
    <SectionTitle>
5. Group Similarity
</SectionTitle>
    <Paragraph position="0"> The corpus-based thesaurus contains only the similarities between individual pairs of words. In our algorithm, we require the similarity between groups of words. The group similarity measure we use is proposed by Karypis et al. (1999). It takes as input two groups of elements, Gl and G2, and a similarity matrix, sim, which specifies the similarity between individual elements. GI and G2 are describable by graphs where the vertices are the words and each weighted edge between vertices wl and w2 represents the similarity, sim(wl, w2), between the words wl and Wz.</Paragraph>
    <Paragraph position="1"> Karypis et al. consider both the interconnectivity and the closeness of the groups. The absolute interconnectivity between G t and G 2, AI(G t, G2), is defined as the aggregate similarity between the two groups: x~Gi YEG2 The absolute closeness between G~ and G2, AC(G~, G2), is defined as the average similarity between a pair of elements, one from each group: Ic, lc l</Paragraph>
  </Section>
  <Section position="7" start_page="81" end_page="81" type="metho">
    <SectionTitle>
WORD CANDIDATE ENGLISH SENSE FREQUENCY OF
TRANSLATION OCCURRENCE
</SectionTitle>
    <Paragraph position="0"> account compte bank account, business 245 rapport report, statement 55 duty devoir responsibility, obligation 80 taxe tax 30 race course contest 87 race racial group 23 suit proems lawsuit 281 costume garment 17 check ch6que draft, bank order 105 contr61e evaluation, verification 25 record record unsurpassed statistic/performance 98 enregistremen t recorded data or documentation 12 The difference between the absolute interconnectivity and the absolute closeness is that the latter takes zero similarity pairs into account. In Figure 6, the interconnectivity in (a) and (b) remains constant. However, the closeness in (a) is higher than in (b) since there are more zero similarity pairs in (b).</Paragraph>
    <Paragraph position="1"> Karypis et al. normalized the absolute interconnectivity and closeness by the internal interconnectivity and closeness of the individual groups. The normalized measures are referred to as relative interconnectivity, RI(GI, G2), and relative closeness, RC(GI, G2). The internal interconnectivity and closeness are obtained by first computing a minimal edge bisection of each group. An even-sized partition {G', G&amp;quot;} of a group G is called a minimal edge bisection of G if AI(G', G&amp;quot;) is minimal among all such partitions. The internal interconnectivity of G, II(G), is defined as II(G) = AI(G', G&amp;quot;) and the internal closeness of G, IC(G), as IC(G) = AC(G', G&amp;quot;).</Paragraph>
    <Paragraph position="2"> Minimal edge bisection is performed for all WATs and all sets of contextually similar words. However, the minimal edge bisection problem is NP-complete (Garey and Johnson, 1979).</Paragraph>
    <Paragraph position="3"> Fortunately, state of the art graph partitioning algorithms can approximate these bisections in polynomial time (Goehring and Saad, 1994; Karypis and Kumar, 1999; Kernighan and Lin, 1970). We used the same approximation methods as in (Karypis et al., 1999).</Paragraph>
    <Paragraph position="4"> The similarity between G1 and G2 is then defined as follows: groupSim(G,, G2)= R/(G,, G2)x RC(G,, G 2 ) where</Paragraph>
    <Paragraph position="6"> is the relative closeness.</Paragraph>
  </Section>
  <Section position="8" start_page="81" end_page="81" type="metho">
    <SectionTitle>
6. Experimental Results
</SectionTitle>
    <Paragraph position="0"> The design of our glossing algorithm is applicable to any source/destination language pair as long as a source language parser is available. We considered English-to-French translations in our experiments.</Paragraph>
    <Paragraph position="1"> We experimented with six English nouns that have multiple French translations: account, duty, race, suit, check, and record. Using the 1987  ROM, we extracted a testing corpus 4 consisting of the first 100 to 300 sentences containing the non-idiomatic usage of the six nouns s. Then, we manually tagged each sentence with one of the candidate translations shown in Table 4.</Paragraph>
    <Paragraph position="2"> Each noun in Table 4 translates more frequently to one candidate translation than the other. In fact, always choosing the candidate procbs as the translation for suit yields 94% accuracy. A better measure for evaluating the system's classifications considers both the algorithm's precision and recall on each candidate translation. Table 5 illustrates the precision and recall of our glossing algorithm for each candidate translation. Albeit precision and recall are used to evaluate the quality of the classifications, overall accuracy is sufficient for comparing different approaches with our system. In Section 3, we presented an algorithm for identifying the contextually similar words of a word in a context using a corpus-based thesaurus and a collocation database. Each of the six nouns has similar words in the corpus-based thesaurus. However, in order to find contextually similar words, at least one similar word for each noun must occur in the collocation database in a given context. Thus, the algorithm for constructing contextually similar words is dependent on the coverage of the collocation database. We estimated this coverage by counting the number of times each of the six nouns, in several different contexts, has at least one contextually similar word. The result is shown in Table 6.</Paragraph>
    <Paragraph position="3"> In Section 5, we described a group similarity metric, groupSim, which we use for comparing a WAT with a set of contextually similar words.</Paragraph>
    <Paragraph position="4"> In Figure 7, we compare the translation accuracy of our algorithm using other group similarity metrics. Suppose G~ and (/2 are two groups of words and w is the word that we wish to translate. The metrics used are: I. closest&amp; sum of similarity of the three closest pairs of words from each group.</Paragraph>
    <Paragraph position="5">  by the frequency with which a word in a given context has at least one contextually similar word.</Paragraph>
  </Section>
  <Section position="9" start_page="81" end_page="81" type="metho">
    <SectionTitle>
WORD NUMBER OF COVERAGE
CONTEXTS
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="10" start_page="81" end_page="81" type="metho">
    <SectionTitle>
5. RC:
</SectionTitle>
    <Paragraph position="0"> as defined in Section 5.</Paragraph>
  </Section>
  <Section position="11" start_page="81" end_page="83" type="metho">
    <SectionTitle>
6. RI:
</SectionTitle>
    <Paragraph position="0"> as defined in Section 5.</Paragraph>
    <Paragraph position="1">  In mostFrequent, we include the results obtained if we always choose the translation that occurs most frequently in the testing corpus. We also compared the accuracy of our glossing algorithm with Systran's translation system by feeding the testing sentences into Systran's web interface 6 and manually examining the results. Figure 8 summarizes the overall accuracy obtained by each system and the baseline on the testing corpus. Systran tended to prefer one candidate translation over the other and committed the majority of its errors on the non-preferred senses.</Paragraph>
    <Paragraph position="2"> Consequently, Systran is very accurate if its preferred sense is the frequent sense (as in account and duty) but is very inaccurate if its preferred sense is the infrequent one (as in race, suit, and check).</Paragraph>
    <Paragraph position="3"> 7. Conclusion and Future Work This paper presents a word-for-word glossing algorithm. The gloss of a word is determined by maximizing the similarity between the set of contextually similar words and the different translations of the word in a bilingual thesaurus.</Paragraph>
  </Section>
class="xml-element"></Paper>