Case-Based/Example-Based Machine Translation (CBMT/EBMT) has been proposed as a way of over­coming the knowledge acquisition bottleneck in ma­chine translation. This approach is based on the simple concept of translating sentences by analogy with simi­lar cases stored in a set of casesfa case-base) [1, 2, 3, 4].

This approach has two advantages in terms of knowl­edge acquisition. CBMT/EBMT ensures that (1) if the same case as the input exists in the case-base, then the same result will be obtained, and (2) if a similar case exists in the case-base, then a similar result will be ob­tained. In the first instance, which cases are regarded as the same depends on the equality metrics of the sys­tem. In the second instance, which cases are regarded as similar depends on the similarity metrics. Rule de­velopers or users can control the system on the basis of equality and similarity without understanding the global flow of controls.

In applying this idea to practical machine transla­tion systems, there are still two serious problems. One is that CBMT/EBMT requires a great deal of compu­tation because of its inherent need to retrieve a huge number of cases and calculate their similarities to the input. For practical systems, several hundreds of thou­sands of cases must be accessible.

CBMT/EBMT systems should not impose any restrictions on cases to be added to the case-base in an effort to keep the case-base small, since the similarity metrics depends on the frequencies of cases. If cases are restricted, sufficient information to control the rules is not acquired.

The other problem of CBMT/EBMT is the diffi­culty of defining a semantic distance. Though the­sauri are used as bases for semantic distance calcula­tion in CBMT/EBMT, it may be impossible to define a general semantic distance by using thesauri alone. Semantic distances between words are defined accord­ing to which specific words are related to their trans­lations. For example, in translating the word w^t^ ■£>"(eat,feed,. ..), "^t(dog)-^î-Ê^'&"is equivalent to "a dog eats," "4 (cow)-7&î-jiû^£" is equivalent to "a cow feeds," and ( horse)-^>Vft'^<5" is equiva­lent to "a horse feeds." In these cases, "^"(cow) is closer to "$§" (horse) than "^"(dog), because differ­ent words are selected for each translation of £" with "^"(cow) and ll^t"(dog). But in translating "7Ë&,,(run1gallop,. .. ), (dog)-ffî-y&&n is equiva­lent to "a dog runs," "4- (cowJ-^-jÊS" is equivalent to "a cow runs," and "M (horseJ-^-TË-îi" is equivalent to "a horse gallops." In these cases, "4-" (cow) is closer to "Ü"(dog) than ".^"(horse).

If such incomplete semantic distances calculated by thesauri alone are used for CBMT/EBMT, excep­tional cases may be interpreted as general ones(over-generalization). Over-generalization is a major prob­lem in translating idiomatic expressions. For exam­ple, "M (head)-#*-#Jft£>" has two translations: "hurt one's head " or, idiomatically, "be smart." But "Hoß (head)-^î-^îJii'îi" has only one interpretation, "hurt one's head," though the word "ISpÜ" has almost the same meaning as "M".

It is obvious that uW.M-&-tytl2>" can be trans­lated correctly by adding this translation pair into the case-base. The addition, however, cannot prevent the idiomatic expression "UR-ffi-tyilZ)" from being in­terpreted generally. The idiomatic interpretation still may be adopted for "X-^-^Jft^" if X is more simi­lar to the word "jîjl" than the words in the case-base whose pattern is "X-^-^OttS."

Sato [3] and Sumita [4] weigh each slot depending on how much it affects the translation. However, since such weights are calculated only for each slot, the over-generalization that occurs inside of a slot is not re­solved. To avoid over-generalization, we need some mechanism to encapsulate exceptions rather than to adjust the semantic distance.

A case-base, in contrast to a set of rules, has inherent redundancy, because cases are collected without pre­selection. In the simplest case, if the sentence "A" has only one translation equivalent "a," then the single

But if we view the case-base as a collection of sen­tences, the same sentences rarely seem to occur *. Sen­tences can, however, be divided into smaller fragments which are meaningful units for translation according to the some linguistic models, which we call translation patterns.

These fragments are combined for use in translating sentences. Fragments divided on the basis of trans­lation patterns are obviously more effective than sen­tences, because smaller fragments are more likely to match than full sentences.

We generalize such fragments extracted according to each translation pattern, using a thesaurus, by re­placing the words that occur in cases by more general concepts in the thesaurus. The words to be replaced are determined by their frequencies in the case-base. Frequent occurring fragments should be assigned more weight than less frequent by occurring fragments. The frequencies of fragments are used to weigh generalized cases in generalization.

Semantic distances are calculated for each transla­tion pattern as the importances of generalized cases. Only meaningful categories for the translation pattern are stored as generalized cases, except that the most meaningful category is taken as a default. For example,

lThe case-base should contain natural sentences rather than examples which are only the smallest fragments effective for translation. We distinguish CBMT from EBMT in accordance with this viewpoint.

the word u^C"(dog) may be generalized into the con­cept <dog> for translation of "tS| (. "("a dog barks"), whereas it may be replaced by the more general con­cept <animal>, for other translation patterns in which the concept <dog> is not meaningful-While generalizing cases, we can identify exceptional cases as those which cannot be generalized. Once we identify exceptions, then w« can prevent such excep­tions from being interpreted generally.

In this way, cases are generalized according to the translation pattern into generalized cases with con­cepts as the values of their variables.

In addition to generalized cases, rules can be formu­lated according to translation patterns. Generalized cases and manually written rules are assumed to be the same as objects in CBMT. It is valuable to have rules available as well as cases, especially when the case-base contains insufficient cases. If rules are not available, there must be sufficient cases from the time the system is first used. Incremental development of any domain is possible only if general rules are avail­able.

In accordance with these basic ideas, we propose a method of machine translation in which cases are gen­eralized. In our approach, we define linguistic patterns in translation. According to these patterns, the cases in the case-base are divided into smaller fragments and are generalized. Both rules and generalized cases are used to translate sentences.

CBMT is divided into two sub-processes: (1) best matching, to search for the most similar cases in the case-base, and (2) application control, to control the combination of similar cases for translation. Applica­tion control is a general problem in machine transla­tion, whereas best matching is a problem unique to CBMT. If the best matching process returns certainty factors, the system is controlled using these factors on the basis of the some other model such as Watanabe's [5].

In this paper, we concentrate on best matching using a thesaurus.

^Concepts are enclosed between arrowheads (< and >) in this paper.

The XP;s, the set of GLTPCi the TPH.s, and the thesaurus are used in best matching. The values of variables in Va are extracted from the input sentence by applying pattern matching according to the description of TP, . The best-matching process retrieves the most similar case from the set of GLTPC,.

If Mi — 1, words which are equivalent to the word that is a value of the variable in Va are first searched for in the value of the corresponding variable in Ltof GLTPCiS. If none are found, upper concepts re­trieved in either TPTHi or the thesaurus are searched in turn. The GLTPCi which is found first is the shortest-distance GLTPCi (SDGLTPQ). If C. in GLTPCi is not null, then it is also evaluated, whether it is true or false.

In the CBMT approach, the linguistic model, which is a set of translation patterns, is important both for the compaction ratio of a case-base and for similar­ity metrics. If the model is not appropriate, most cases remains ungeneralized, and unnatural cases are retrieved as similar cases to the input. The problems of constructing linguistic models are the same as in rule-based systems.

However, our approach assumes that the linguistic model does not include controls of rules and general­ized cases. Whether or not this assumption is correct, it is very difficult to define controls in such a way that any exceptional cases are encapsulated properly. Our approach provides an engineering solution to these dif­ficulties.

In our approach, the quality of translations depends on the quantity of cases rather than the quality of the thesaurus. Therefore, it is important to explore (semi-)automatic case acquisition from bilingual cor­pora,

To construct a huge case-base is easier than to con­struct a well-defined thesaurus, because cases are con­structed locally without taking account of side-effects. To define an effective thesaurus for translation, every effective category for translation must be included, and every intermediate category that is effective for trans­lation must be included in order to calculate semantic distances properly.

If, on the other hand, thesauri can be developed in­dependently from the case-base, developers or users can select the most appropriate thesaurus for the do­main.

This paper has described a framework for a machine translation using a mixture of rules and cases general­ized by means of a thesaurus, which is much smaller than the case-base itself. Since the importances of rules and generalized cases are calculated in advance by gen­eralization, it is not necessary to calculate them during the best-matching, which is done by exact matching of words or upper concepts in the thesaurus.

I would like to thank Masayuki Morohashi and the members of Japanese Processing Group of Tokyo Re­search Laboratory, IBM Japan, for their valuable sug­gestions and encouragement, and Michael McDonald for his helpful advice on the wording of this paper.