File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/99/p99-1028_metho.xml
Size: 25,141 bytes
Last Modified: 2025-10-06 14:15:25
<?xml version="1.0" standalone="yes"?> <Paper uid="P99-1028"> <Title>Resolving Translation Ambiguity and Target Polysemy in Cross-Language Information Retrieval</Title> <Section position="3" start_page="0" end_page="217" type="metho"> <SectionTitle> 2. Effects of Ambiguities </SectionTitle> <Paragraph position="0"> Translation ambiguity and target polysemy are two major problems in CLIR. Translation ambiguity results from the source language, and target polysemy occurs in target language. Take Chinese-English information retrieval (CEIR) and English-Chinese information retrieval (ECIR) as examples. The former uses Chinese queries to retrieve English documents, while the later employs English queries to retrieve Chinese documents. To explore the difficulties in the query translation of different languages, we gather the sense statistics of English and Chinese words. Table 1 shows the degree of word sense ambiguity (in terms of number of senses) in English and in Chinese, respectively. A Chinese thesaurus, i.e., ~~$~k (tong2yi4ci2ci21in2), (Mei, et al., 1982) and an English thesaurus, i.e., Roget's thesaurus, are used to count the statistics of the senses of words. On the average, an English word has 1.687 senses, and a Chinese word has 1.397 senses. If the top 1000 high frequent words are considered, the English words have 3.527 senses, and the bi-character Chinese words only have 1.504 senses. In summary, Chinese word is comparatively unambiguous, so that translation ambiguity is not serious but target polysemy is serious in CEIR. In contrast, an English word is usually ambiguous. The translation disambiguation is important in ECIR. Consider an example in CEIR. The Chinese word &quot;,~,It&quot; (yin2hang2) is unambiguous, but its English translation &quot;bank&quot; has 9 senses (Longman, 1978). When the Chinese word &quot; ,~ 45- &quot; (yin2hang2) is issued, it is translated into the English counterpart &quot;bank&quot; by dictionary lookup without difficulty, and then &quot;bank&quot; is sent to an IR system. The IR system will retrieve documents that contain this word. Because &quot;bank&quot; is not disambiguated, irrelevant documents will be reported. On the contrary, when &quot;bank&quot; is submitted to an ECIR system, we must disambiguate its meaning at first. If we can find that its correct translation is &quot;-~g-#5&quot;&quot; (yin2hang2), the subsequent operation is very simple. That is, &quot;~'~5-&quot; (yin2hang2) is sent into an IR system, and then documents containing &quot;~l~5&quot;&quot; (yin2hang2) will be presented. In this example, translation disambiguation should be done rather than target polysemy resolution.</Paragraph> <Paragraph position="1"> The above examples do not mean translation disambiguation is not required in CEIR. Some Chinese words may have more than one sense.</Paragraph> <Paragraph position="2"> For example, &quot;k-~ ~ &quot; (yun4dong4) has the following meanings (Lai and Lin, 1987): (1) sport, (2) exercise, (3) movement, (4) motion, (5) campaign, and (6) lobby. Each corresponding English word may have more than one sense.</Paragraph> <Paragraph position="3"> For example, &quot;exercise&quot; may mean a question or set of questions to be answered by a pupil for practice; the use of a power or right; and so on.</Paragraph> <Paragraph position="4"> The multiplication effects of translation ambiguity and target polysemy make query translation harder.</Paragraph> <Paragraph position="5"> 3. Translation Ambiguity and Polysemy</Paragraph> <Section position="1" start_page="215" end_page="217" type="sub_section"> <SectionTitle> Resolution Models </SectionTitle> <Paragraph position="0"> In the recent works, Ballesteros and Croft (1998), and Bian and Chen (1998) employ dictionaries and co-occurrence statistics trained from target language documents to deal with translation ambiguity. We will follow our previous work (Bian and Chen, 1998), which combines the dictionary-based and corpus-based approaches for CEIR. A bilingual dictionary provides the translation equivalents of each query term, and the word co-occurrence information trained from a target language text collection is used to disambiguate the translation. This method considers the content around the translation equivalents to decide the best target word. The translation of a query term can be disambiguated using the co-occurrence of the translation equivalents of this term and other terms. We adopt mutual information (Church, et al., 1989) to measure the strength. This disambiguation method performs good translations even when the multi-term phrases are not found in the bilingual dictionary, or the phrases are not identified in the source language. Before discussion, we take Chinese-English information retrieval as an example to explain our methods. Consider the Chinese query &quot;,~I~'~5-&quot; (yin2hang2) to an English collection again. The ambiguity grows from none (source side) to 9 senses (target side) during query translation.</Paragraph> <Paragraph position="1"> How to incorporate the knowledge from source side to target side is an important issue. To avoid the problem of target polysemy in query translation, we have to restrict the use of a target word by augmenting some other words that usually co-occur with it. That is, we have to make a context for the target word. In our method, the contextual information is derived from the source word.</Paragraph> <Paragraph position="2"> We collect the frequently accompanying nouns and verbs for each word in a Chinese corpus.</Paragraph> <Paragraph position="3"> Those words that co-occur with a given word within a window are selected. The word association strength of a word and its accompanying words is measured by mutual information. For each word C in a Chinese query, we augment it with a sequence of Chinese words trained in the above way. Let these words be CW~, CW2, ..., and CWm. Assume the corresponding English translations of C, CW~, CW2, ..., and CWm are E, EW,, EW2, ..., and EWm, respectively. EWe, EW2, ..., and EWm form an augmented translation restriction of E for C. In other words, the list (E, EW1, EW2, ..., EWm) is called an augmented translation result for C.</Paragraph> <Paragraph position="4"> EWe, EWe, ..., and EWm are a pseudo English context produced from Chinese side. Consider the Chinese word &quot;~I~gS&quot;&quot; (yin2hang2). Some strongly co-related Chinese words in ROCLING balanced corpus (Huang, et al., 1995) are: &quot;I!.g.~,&quot; (tie 1 xian4), &quot;~ ~&quot; (ling3chu 1 ), &quot;_-~_. ~&quot; (li3ang2), &quot;~ 1~&quot; (yalhui4), &quot;;~ ~&quot; (hui4dui4), etc. Thus the augmented translation restriction of &quot;bank&quot; is (rebate, show out, Lyons, negotiate, transfer, ...). Unfortunately, the query translation is not so simple. A word C in a query Q may be ambiguous. Besides, the accompanying words CW~ (1 < i < m) trained from Chinese corpus may be translated into more than one English word.</Paragraph> <Paragraph position="5"> An augmented translation restriction may add erroneous patterns when a word in a restriction has more than one sense. Thus we devise several models to discuss the effects of augmented restrictions. Figure 1 shows the different models and the model refinement procedure. A Chinese query may go through translation ambiguity resolution module (left-to-right), target polysemy resolution module (top-down), or both (i.e., these two modules are integrated at the right corner).</Paragraph> <Paragraph position="6"> In the following, we will show how each module is operated independently, and how the two modules are combined.</Paragraph> <Paragraph position="7"> For a Chinese query which is composed of n words C~, C2, ..., Ca, find the corresponding English translation equivalents in a Chinese-English bilingual dictionary. To discuss the propagation errors from translation ambiguity resolution part in the experiments, we consider the following two alternatives: (a) select all (do-nothing) The strategy does nothing on the translation disambiguation. All the English translation equivalents for the n Chinese words are selected, and are submitted to a monolingual information retrieval system.</Paragraph> <Paragraph position="8"> (b) co-occurrence model (Co-Model) We adopt the strategy discussed previously for translation disambiguation (Bian and Chen, 1998). This method considers the content around the English translation equivalents to decide the best target equivalent.</Paragraph> <Paragraph position="9"> For target polysemy resolution part in Figure 1, we also consider two alternatives. In the first alternative (called A model), we augment restrictions to all the words no matter whether they are ambiguous or not. In the second alternative (called U model), we neglect those Cs that have more than one English translation.</Paragraph> <Paragraph position="10"> Assume Co~), C~2) .... , Co~p) (p < n) have only one English translation. The restrictions are augmented to Co~), C~2) ..... Co~p) only. We apply the above corpus-based method to find the restriction for each English word selected by the translation ambiguity resolution model. Recall that the restrictions are derived from Chinese corpus. The accompanying words trained from Chinese corpus may be translated into more than one English word. Here, the translation ambiguity may occur when translating the restrictions. Three alternatives are considered.</Paragraph> <Paragraph position="11"> In U1 (or A1) model, the terms without ambiguity, i.e., Chinese and English words are one-to-one correspondent in a Chinese-English bilingual dictionary, are added. In UT (or AT) model, th/~ terms with the same parts of speech (POSes) are added. That is, POS is used to select English word. In UTT (or ATT) model, we use mutual information to select top 10 accompanying terms of a Chinese query word, and POS is used to obtain the augmented translation restriction.</Paragraph> <Paragraph position="12"> (1) Only one English Translation: ~ Chinese Restriction C o(I), Ca(2) .... , Co(p) {CWotl) Z ..... CWo(l)ml} ' UT Model &quot;J ER-UT \] ~\]~'-~ (2) More than one English Translation: &quot; {CWof2)t{CWa(p) I ..... ..... CWo(2)m.,}C~/a(p) raF~ ..... (POS Tag Matched) &quot;l I / C a(~-i~, C o(p+2) ..... C o{.) ~. UTT Model ~l ER-UTT I (Top 10 & POS Tag Matched)l</Paragraph> <Paragraph position="14"> In the above treatment, a word C~ in a query Q is translated into (Ei, EWil, EWi2 .... , EWimi). Ei is selected by Co-Model, and EWi~, EWi2 .... , EWimi are augmented by different target polysemy resolution models. Intuitively, Ei, EWil, EWi2 .... , EWim~ should have different weights. Ei is assigned a higher weight, and the words EWil, EWi2 ..... Eim~ in the restriction are assigned lower weights. They are determined by the following formula, where n is number of words in Q and mk is the number of words in a restriction for Ek.</Paragraph> <Paragraph position="16"> Thus six new models, i.e., A1W, ATW, ATTW, U1W, UTW and UTTW, are derived. Finally, we apply Co-model again to disambiguate the pseudo contexts and devise six new models (A1WCO, ATWCO, ATTWCO, U1WCO, UTWCO, and UTTWCO). In these six models, only one restriction word will be selected from the words EWil, EWiz, ..., EWim i via disambiguation with other restrictions.</Paragraph> </Section> </Section> <Section position="4" start_page="217" end_page="221" type="metho"> <SectionTitle> 4. Experimental Results </SectionTitle> <Paragraph position="0"> To evaluate the above models, we employ TREC-6 text collection, TREC topics 301-350 (Harman, 1997), and Smart information retrieval system (Salton and Buckley, 1988). The text collection contains 556,077 documents, and is about 2.2G bytes. Because the goal is to evaluate the performance of Chinese-English information retrieval on different models, we translate the 50 English queries into Chinese by human. The topic 332 is considered as an example in the following. The original English version and the human-translated Chinese version are shown. A TREC topic is composed of several fields. Tags <num>, <title>, <des>, and <narr> denote topic number, title, description, and narrative fields. Narrative provides a complete description of document relevance for the assessors. In our experiments, only the fields of title and description are used to generate queries. <top> <num> Number: 332 <title> Income Tax Evasion <desc> Description: This query is looking for investigations that have targeted evaders of U.S. income tax.</Paragraph> <Paragraph position="1"> <narr> Narrative: A relevant document would mention investigations either in the U.S. or abroad of people suspected of evading U.S. income tax laws. Of particular interest are investigations involving revenue from illegal activities, as a strategy to bring known or suspected criminals to justice. Totally, there are 1,017 words (557 distinct words) in the title and description fields of the 50 translated TREC topics. Among these, 401 words have unique translations and 616 words have multiple translation equivalents in our Chinese-English bilingual dictionary. Table 2 shows the degree of word sense ambiguity in English and in Chinese, respectively. On the average, an English query term has 2.976 senses, and a Chinese query term has 1.828 senses only. In our experiments, LOB corpus is employed to train the co-occurrence statistics for translation ambiguity resolution, and ROCLING balanced corpus (Huang, et al., 1995) is employed to train the restrictions for target polysemy resolution. The mutual information tables are trained using a window size 3 for adjacent words.</Paragraph> <Paragraph position="2"> Table 3 shows the query translation of TREC topic 332. For the sake of space, only title field is shown. In Table 3(a), the first two rows list the original English query and the Chinese query. Rows 3 and 4 demonstrate the English translation by select-all model and co-occurrence model by resolving translation ambiguity only. Table 3(b) shows the augmented translation results using different models. Here, both translation ambiguity and target polysemy are resolved.</Paragraph> <Paragraph position="3"> The following lists the selected restrictions in A1 model.</Paragraph> <Paragraph position="4"> i~_~(evasion): ~.~_N (N: poundage), ~/t~_N (N: scot), ~.tkV (V: stay) ?~-(income): I~g~_N (N: quota) ~(tax): i/~_V (N: evasion), I~_N (N:surtax), ~t ~,_N (N: surplus), ,g'~_N (N: sales tax) Augmented translation restrictions (poundage, scot, stay), (quota), and (evasion, surtax, surplus, sales tax) are added to &quot;evasion&quot;, &quot;income&quot;, and &quot;tax&quot;, respectively. From Longman dictionary, we know there are 3 senses, 1 sense, and 2 senses for &quot;evasion&quot;, &quot;income&quot;, and &quot;tax&quot;, respectively. Augmented restrictions are used to deal with target polysemy problem. Compared with A1 model, only &quot;evasion&quot; is augmented with a translation restriction in U1 model. This is because &quot; &quot;~ ~ &quot; (tao21uo4) has only one translation and &quot;?~-&quot; (suo3de2) and &quot;~&quot; (sui4) have more than one translation. Similarly, the augmented translation restrictions are omitted in the other U-models. Now we consider AT model. The Chinese restrictions, which have the matching POSes, are listed below: i~ (evasion): ~_N (N: poundage), ~l~t~0~,_N (N: scot), L~_V (V: stay), ~N (N: droit, duty, geld, tax), li~l~f~ N (N: custom, douane, tariff), /~.~ V (V: avoid, elude, wangle, welch, welsh; N: avoidance, elusion, evasion, evasiveness, miss, runaround, shirk, skulk), i.~)~_V (V: contravene, infract, infringe; N: contravention, infraction, infringement, sin, violation) ~&quot; ~- (income): ~_V (V: impose; N: division), ~.&~,_V (V: assess, put, tax; N: imposition, taxation), ~A~_N (N: Swiss, Switzer), i~_V (V: minus, subtract), I~I\[$~_N (N: quota), I~l ~_N (N: commonwealth, folk, land, nation, nationality, son, subject) (tax): I~h~_N (N: surtax), .~t~g, N (N: surplus), ~'~ _N (N: sales tax), g~V (V: abase, alight, debase, descend), r~_N (N: altitude, loftiness, tallness; ADJ: high; ADV: loftily), ~V (V: comprise, comprize, embrace, encompass), -~V (V: compete, emulate, vie; N: conflict, contention, duel, strife) (a) Resolving Translation Ambiguity Only original English query income tax evasion Chinese translation by human ~ (tao21uo4) ?~- (suo3de2) $~, (sui4) by select all model (evasion), (earning, finance, income, taking), (droit, duty, geld, tax) by co-occurrence model evasion, income, tax (b) Resolving both Translation Ambiguity and Target Polysemy by AI model by UI model by AT model by UT model :by ATT model by UTT model b-y ATWCO model by UTWCO model by ATTWCO model by UTTWCO model (evasion, poundage, scot, stay), (income, quota), (tax, evasion, surtax, surplus, sales tax) (evasion, poundage, scot, stay), (income), (tax) (evasion; poundage; scot; stay; droit, duty, geld, tax; custom, douane, tariff; avoid, elude, wangle, welch, welsh; contravene, infract, infringe), (income; impose; assess, put, tax; Swiss, Switzer; minus subtract; quota; commonwealth, folk, land, nation, nationality, son, subject), (tax; surtax; surplus; sales tax; abase, alight, debase, descend; altitude, loftiness, tallness; comprise, comprize, embrace, encompass; compete, emulate, vie) (evasion; poundage, scot, stay, droit, duty, geld, tax, custom, douane, tariff, avoid, elude, wangle, welch, welsh, contravene, infract, infringe), (income), (tax) (evasion, poundage, scot, stay, droit, duty, geld, tax, custom, douane, tariff), (income), (tax) (evasion, poundage, scot, stay, droit, duty, geld, tax, custom, douane, tariff), (income), (tax) (evasion, tax), (income, land), (tax, surtax) (evasion, poundage), (income), (tax) (evasion, tax), (income), (tax) (evasion, poundage), (income), (tax) Those English words whose POSes are the same as the corresponding Chinese restrictions are selected as augmented translation restriction.</Paragraph> <Paragraph position="5"> For example, the translation of&quot;~&quot;_V (tao2bi4) has two possible POSes, i.e., V and N, so only &quot;avoid&quot;, &quot;elude&quot;, &quot;wangle&quot;, &quot;welch&quot;, and &quot;welsh&quot; are chosen. The other terms are added in the similar way. Recall that we use mutual information to select the top 10 accompanying terms of a Chinese query term in ATT model.</Paragraph> <Paragraph position="6"> The 5 ~ row shows that the augmented translation restrictions for &quot;?)i&quot;~-&quot; (suo3de2) and &quot;~,&quot; (sui4) are removed because their top 10 Chinese accompanying terms do not have English translations of the same POSes. Finally, we consider ATWCO model. The words &quot;tax&quot;, &quot;land&quot;, and &quot;surtax&quot; are selected from the three lists in 3 rd row of Table 3(b) respectively, by using word co-occurrences.</Paragraph> <Paragraph position="7"> Figure 2 shows the number of relevant documents on the top 1000 retrieved documents for Topics 332 and 337. The performances are stable in all of the +weight (W) models and the enhanced CO restriction (WCO) models, even there are different number of words in translation restrictions. Especially, the enhanced CO restriction models add at most one translated restriction word for each query tenn. They can achieve the similar performance to those models that add more translated restriction words.</Paragraph> <Paragraph position="8"> Surprisingly, the augmented translation results may perform better than the monolingual retrieval. Topic 337 in Figure 2 is an example.</Paragraph> <Paragraph position="9"> Table 4 shows the overall performance of 18 different models for 50 topics. Eleven-point average precision on the top 1000 retrieved documents is adopted to measure the performance of all the experiments. The monolingual information retrieval, i.e., the original English queries to English text collection, is regarded as a baseline model. The performance is 0.1459 under the specified environment. The select-all model, i.e., all the translation equivalents are passed without disambiguation, has 0.0652 average precision. About 44.69% of the performance of the monolingual information retrieval is achieved. When co-occurrence model is employed to resolve translation ambiguity, 0.0831 average precision (56.96% of monolingual information retrieval) is reported.</Paragraph> <Paragraph position="10"> Compared to do-nothing model, the performance is 27.45% increase.</Paragraph> <Paragraph position="11"> Now we consider the treatment of translation ambiguity and target polysemy together.</Paragraph> <Paragraph position="12"> Augmented restrictions are formed in A1, AT, ATT, U1, UT and UTT models, however, their performances are worse than Co-model (translation disambiguation only). The major reason is the restrictions may introduce errors. That can be found from the fact that models U 1, UT, and UTT are better than A1, AT, and ATT.</Paragraph> <Paragraph position="13"> Because the translation of restriction from source language (Chinese) to target language (English) has the translation ambiguity problem, the models (U1 and A1) introduce the unambiguous restriction terms and perform better than other models. Controlled augmentation shows higher performance than uncontrolled augmentation.</Paragraph> <Paragraph position="14"> When different weights are assigned to the original English translation and the augmented restrictions, all the models are improved significantly. The performances of A1W, ATW, ATTW, U1W, UTW, and UTTW are about 10.11% addition to the model for translation disambiguation only. Of these models, the performance change from model AT to model ATW is drastic, i.e., from 0.0419 (28.72%) to 0.0913 (62.58%). It tells us the original English translation plays a major role, but the augmented restriction still has a significant effect on the performance.</Paragraph> <Paragraph position="15"> We know that restriction for each English translation presents a pseudo English context.</Paragraph> <Paragraph position="16"> Thus we apply the co-occurrence model again on the pseudo English contexts. The performances are increased a little. These models add at most one translated restriction word for each query term, but their performances are better than those models that adding more translated restriction words. It tells us that a good translated restriction word for each query term is enough for resolving target polysemy problem. U1WCO, which is the best in these experiments, gains 62.92% of monolingual information retrieval, and 40.80% increase to the do-nothing model (selectall). null</Paragraph> </Section> <Section position="5" start_page="221" end_page="221" type="metho"> <SectionTitle> 5. Concluding Remarks </SectionTitle> <Paragraph position="0"> This paper deals with translation ambiguity and target polysemy at the same time. We utilize two monolingual balanced corpora to learn useful statistical data, i.e., word co-occurrence for translation ambiguity resolution, and translation restrictions for target polysemy resolution.</Paragraph> <Paragraph position="1"> Aligned bilingual corpus or special domain corpus is not required in this design. Experiments show that resolving both translation ambiguity and target polysemy gains about 10.11% performance addition to the method for translation disambiguation in cross-language information retrieval. We also analyze the two factors: word sense ambiguity in source language (translation ambiguity), and word sense ambiguity in target language (target polysemy). The statistics of word sense ambiguities have shown that target polysemy resolution is critical in Chinese-English information retrieval.</Paragraph> <Paragraph position="2"> This treatment is very suitable to translate very short query on Web, The queries on Web are 1.5-2 words on the average (Pinkerton, 1994; Fitzpatrick and Dent, 1997). Because the major components of queries are nouns, at least one word of a short query of length 1.5-2 words is noun. Besides, most of the Chinese nouns are unambiguous, so that translation ambiguity is not serious comparatively, but target polysemy is critical in Chinese-English Web retrieval. The translation restrictions, which introduce pseudo contexts, are helpful for target polysemy resolution. The applications of this method to cross-language Internet searching, the applicability of this method to other language pairs, and the effects of human-computer interaction on resolving translation ambiguity and target polysemy will be studied in the future.</Paragraph> </Section> class="xml-element"></Paper>