<?xml version="1.0" standalone="yes"?>
<Paper uid="W02-1106">
  <Title>Translating Lexical Semantic Relations: The First Step Towards Multilingual Wordnets*</Title>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2. Translation Equivalents and Semantic
Relations
</SectionTitle>
    <Paragraph position="0"> Note that two translation equivalents (TE) in a pair of languages stand in a lexical semantic relation. The most desirable scenario is that when the two TE's are synonymous, such as the English 'apple' to the Mandarin 'ping2guo3'.</Paragraph>
    <Paragraph position="1"> However, since the conceptual space is not segmented identically for all languages, TE's may often stand in other relations to each other.</Paragraph>
    <Paragraph position="2"> For instance, the Mandarin 'zuo1zhi5' is a hypernym for both the English 'desk' and 'table'.</Paragraph>
    <Paragraph position="3"> Suppose we postulate that the LSR's between TE's are exactly identical in nature to the monolingual LSR's described in wordnets. This means that the lexical semantic relation introduced by translation can be combined with monolingual LRS's. Predicting LSR's in a target language based on source language data become a simple logical operation of combining relational functions when the LSR of translation equivalency is defined. This framework is illustrated in Diagram 1.</Paragraph>
    <Paragraph position="5"> CW1 represents our starting Chinese lemma which can be linked to EW1 through the translation LSR i. The linked EW1 can than provide a set of LSR predictions based on the English WN. Assume that we take the LSR x, which is linked to EW2. That LSR prediction is mapped back to Chinese when EW2 is translated to CW2 with a translation LSR ii. In this model, the relation y, between CW1 and CW2 is a functional combination of the three LSR's i, x, and ii.</Paragraph>
    <Paragraph position="6"> However, it is well known that language translation involves more than semantic correspondences. Social and cultural factors also play a role in (human) choices of translation equivalents. It is not the aim of this paper to predict when or how these semantically non-identical translations arise. The aim is to see how much lexical semantic information is inferable across different languages, regardless of translational idiosyncrasies. In this model, the prediction relies crucially on the semantic information provided by the source language (e.g. English) lexical entry as well as the lexical semantic correspondence of a target language (e.g. Chinese) entry. The translation relations of the relational target pairs, although capable of introducing more idiosyncrasies, are not directly involved in the prediction. Hence we make the generalization that any discrepancy introduced at this level does not affect the logical relation of LSR prediction and adopt a working model described in Diagram 2. We only take into consideration those cases where the translation LSR ii is exactly equivalent, i.e., EW2 = CW2.</Paragraph>
    <Paragraph position="7"> This step also allows us to reduce the maximal number of LSR combination in each prediction to two. Thus we are able to better predict the contribution of each mono- or bi-lingual LSR.</Paragraph>
    <Paragraph position="9"> The unknown LSR y = i + x Diagram 2. Translation-mediated LSR Prediction (Reduced Model, currently adopted)</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.1 LRS Inference as Relational Combination
</SectionTitle>
      <Paragraph position="0"> With the semantic contribution of the translation equivalency defined as a (bilingual) LSR, the inference of LSR in the target language wordnet is a simple combination of semantic relations. The default and ideal situation is where the two TE's are synonymous.</Paragraph>
      <Paragraph position="2"> In this case, the translation LSR is an identical relation; the LSR of the source language wordnet can be directly inherited. This is illustrated in Diagram 3.</Paragraph>
      <Paragraph position="3"> However, when the translation has a non-identical semantic relation, such as antonyms and hypernyms, then the LSR predicted is the combination of the bilingual relation and the monolingual relation. In this paper, we will concentrate on Hypernyms and Hyponyms. The choice is made because these two LSR's are transitive relations by definition and allows clear logical predications when combined. The same, with some qualifications, may apply to the Holonym relations.</Paragraph>
      <Paragraph position="4"> Combinations of other LSR's may not yield clear logical entailments. The scenarios involving Hyponymy and Hypernymy will be discussed in section 3.3.</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3. Cross-lingual LSR Inference: A Study
</SectionTitle>
    <Paragraph position="0"> based on English-Chinese Correspondences In this study, we start with a WN-based</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
English-Chinese Translation Equivalents
</SectionTitle>
      <Paragraph position="0"> Database (TEDB)1. Each translation equivalents pair was based on a WN synset. For quality control, we mark each TE pair for its accuracy as well as the translation semantic relation.</Paragraph>
      <Paragraph position="1"> For this study, the 200 most frequently used Chinese words plus 10 adjectives are chosen (since there is no adjective among the top 200 words in Mandarin). Among the 210 input lemmas, 179 lemmas2 find translation equivalents in the TEDB and are mapped to 497 1 The translation equivalence database was hand-crafted by the CKIP WordNet team. For each of the 99642 English synset head words, three appropriate translation equivalents were chosen whenever possible. At the time when this study was carried out, 42606 TE's were proofed and available 2 The input lemmas for which TE's were unable to find are demonstratives or pronouns for nouns, and aspect markers for adverbs English synsets. The occurring distribution is as follows: 84 N's with 195 times; 41 V's with 161 times; 10 Adj's with 47 times; and 47 Adv's with 94 times. 441 distinct English synsets are covered under this process, since some of the TE's are for the same synset. This means that each input Chinese lemma linked to 2.4 English synsets in average. Based on the TEDB and English WN, the 179 mapped input Chinese lemmas expanded to 597 synonyms. And extending from the 441 English synsets, there are 1056 semantically related synsets in WN, which yields 1743 Chinese words with our TEDB.</Paragraph>
      <Paragraph position="2"> 3.1. Evaluation of the Semantics of Translation Six evaluative tags are assigned for the TEDB. Four of them are remarks for future processing. The LSR marked are  'market, securities_industry': the securities markets in the aggregate Table 2 indicates the relations between the synonyms of an input lemma and the same English synset. Recall that our TEDB gives more than one Chinese translation equivalent to one English WN entry. Hence we can hypothesize that the set of Chinese translation equivalents form a synset. It is natural, then, to examine the semantic relations between other synset members and the original WN entry. Table 1 and 2 show a rather marked difference in terms of the correctness of the synonymy relation. This will be further explained later.</Paragraph>
      <Paragraph position="3"> Syn. Incor.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
(Total Subject=597)
</SectionTitle>
      <Paragraph position="0"> From the data above, we observe two generalizations: First, polysemous lemmas have lower possibility of being synonymous to the corresponding English synset. In addition, we also observe that there is a tendency for some groups, i.e., groups with polysemy and with abstract meanings, to match synonymous English synsets. These findings are helpful in our further studies when constructing CWN, as well as in the application of TEDB.</Paragraph>
    </Section>
    <Section position="3" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.2 Cross-lingual LSR predictions with
</SectionTitle>
      <Paragraph position="0"> synonymous translations The next step is to take the set of English LSR's stipulated on a WN synset and transport them to its Chinese translation equivalents. We evaluated the validity of the inferred semantic relations in Chinese. In this study, we concentrated on three better-defined (and more frequently used) semantic relations: antonyms (ANT); hypernyms (HYP); and hyponyms (HPO). Here, we limit our examination to the Chinese lemmas that are both translation equivalents of an English WN entry and are considered to have synonymous semantic relations to that entry. The nominal and verbal statistics are given in Table 3 and Table 4 respectively.</Paragraph>
      <Paragraph position="1"> Syn. Incor.</Paragraph>
    </Section>
    <Section position="4" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Semantic Relations=402)
</SectionTitle>
      <Paragraph position="0"> From the 148 nouns where the English and Chinese translation equivalents are also synonymous, there are 357 pairs of semantic relations that are marked in English WN and are therefore candidates for inferred relations in Chinese. On average, each nominal RC translation equivalent yields 2.41 inferable semantic relations. The precision of the inferred semantic relation is tabulated below.</Paragraph>
    </Section>
    <Section position="5" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Inference (Nouns)
</SectionTitle>
      <Paragraph position="0"> The study here shows that when no additional relational distance is introduced by translation (i.e. the 75.9% of nominal cases when TE's are synonyms), up to 90% precision can be achieved for bilingual LSR inference. And among the semantic relations examined, antonymous relations are the most reliable when transportabled cross-linguistically.</Paragraph>
      <Paragraph position="1"> For the 112 verbs where the English and Chinese TE's are synonymous, there are 155 pairs of semantic relations that are marked in WN and are therefore candidates for inferred relations in Chinese. In contrast to nominal translation equivalents, each pair of verbal TE only yields 1.38 inferable semantic relations. The precision of the inferred semantic relation is tabulated in Table 6.</Paragraph>
    </Section>
    <Section position="6" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Inference (Verbs)
</SectionTitle>
      <Paragraph position="0"> Similar to the results of nouns, antonymous relations appear reliable in the behaviors of verbs as well. As to the other types of relations, the correct rates seem to be slightly lower than nouns.</Paragraph>
      <Paragraph position="1"> The precision for English-to-Chinese semantic relation inference is 80% for verbs.</Paragraph>
      <Paragraph position="2"> The observed discrepancy in terms of semantic relations inference between nouns and verbs deserves in-depth examination. Firstly, the precision of nominal inference is 8.52% higher than verbal inference. Secondly, the contrast may not be attributed to a specific semantic relation. Both nouns and verbs have the same precision pattern for the three semantic relations that we studied. Inference of antonymous relations is highly reliable in both categories (both 100%).</Paragraph>
      <Paragraph position="3"> Hyponymous inference is second, and about 12% higher than hypernymous inference in each category (the difference is 11.64% for nouns and 12.42% for verbs). And, last but not least, the precision gaps between nouns and verbs, when applicable, are similar for different semantic relations (9.55% for hypernyms and 8.77% for hyponyms). All the above facts support the generalization that nominal semantic relations are more reliably inferred cross-linguistically than verbal semantic relations. A plausible explanation of this generalization is the difference in mutability of nominal and verbal meanings, as reported by Ahrens (1999). Ahrens demonstrated with off-line experiments that verb meanings are more mutable than noun meanings.</Paragraph>
      <Paragraph position="4"> She also reported that verb meanings have the tendency to change under coercive contexts. We may assume that making the cross-lingual transfer is a coercive context in terms of meaning identification. Taking the mutability into account, we can predict that since verb meanings are more likely than nouns to change under given coercive conditions, the changes will affect their semantic relations. Hence the precision for semantic relations inference is lower for verbs than for nouns.</Paragraph>
      <Paragraph position="5"> In the above discussion, we observed that the three semantic relations seem to offer clear generalizations with regard to the precision of the inferences, as shown in Table 7.</Paragraph>
      <Paragraph position="6">  Two generalizations emerge from the above data and call for explanation: First, inference of antonymous relations is highly reliable; second, inference of hypernymous relations is more reliable than inference of hyponymous relations. The fact that inference of antonymous relations is highly precise may be due to either of the following facts. Since the number of antonymic relations encoded is relatively few (only 22 all together), they may all be the most prototypical case. In addition, a pair of antonyms by definition differs in only one semantic feature and has the shortest semantic distance between them. In other words, an antonym (of any word) is simply a privileged (near) synonym whose meaning offers contrast at one particular semantic dimension. Since antonymy presupposes synonymous relations, it preserves the premise of our current semantic relation inference.</Paragraph>
      <Paragraph position="7"> The fact that hyponymous relations can be more reliably inferred cross-linguistically than hypernymous relations is somewhat surprising, since they are symmetric semantic relations. That is, if A is a hypernym of B, then B is a hyponym of A. Logically, there does not seem to be any reason for the two relations to have disjoint distributions when transported to another language. However, more careful study of the conceptual nature of the semantic relations yields a plausible explanation.</Paragraph>
      <Paragraph position="8"> We should take note of the two following facts: First, a hyponym link defined on an English word Y presupposes a conceptual class denoted by Y, and stipulates that Z is a kind of Y (see Diagram 4).</Paragraph>
      <Paragraph position="9"> Diagram 4. class vs. member identity (HPO) Second, a hypernym link defined on Y presupposes an identity class X which is NOT explicitly denoted, and stipulates that Y is a kind of X (see Diagram 5). Hence, it is possible that there is another valid conceptual class W in the target language that Y is a member of. And yet W is not equivalent to X.</Paragraph>
      <Paragraph position="10"> Diagram 5. class vs. member identity (HYP) Since our inference is based on the synonymous relation of the Chinese TE to the English word Y, the conceptual foundation of the semantic relation is largely preserved, and the inference has a high precision. The failure of inference can in most cases be attributed to the fact that the intended HYP has no synonymous TE in Chinese.</Paragraph>
      <Paragraph position="11"> To infer a hyponymous relation, however, we need to presuppose the trans-lingual equivalence of the conceptual class defined by HPO. And since our inference only presupposes the synonymous relation of Y and its TE, and says nothing about HPO, the success of inference of the hyponymous relation is than dependent upon an additional semantic condition. Hence that it will have lower precision can be expected.</Paragraph>
      <Paragraph position="12"> To sum up, our preliminary evaluation found that the precision of cross-lingual inference of semantic relation can be higher than 90% if the inference does not require other conceptual/semantic relations other than the synonymy of the translation equivalents. On the other hand, an additional semantic relation, such as the equivalence of the hypernym node in both languages when inferring hyponym relations, seems to bring down the precision rate by about 10%.</Paragraph>
      <Paragraph position="13"> 3.3. When Translation Introduces an additional</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>