This paper presents a model of machine translation system that uses examples to translate idioms, collocations and other expressions requiring structural changes. The evaluation result of the model is also presented. It is difficult to find an example that fits for source text exactly. If an example has similar context to source text, the phrase of the example in a target language can be used for target text. So, the model searches an example dictionary for an example that is used in similar context, and adapts the example. An idiom or collocation in the target language corresponds generally to more than one expression in the source language. If the expression in an example has the same meaning as that in source text, it is desirable to convert the expression to fit for the source text and use the example. In the model, some words or phrases in examples are written by general expressions. That is, nouns are expressed as words having some attributes, and verbs are expressed as words that require some cases or have some action attributes. Idioms, collocations and other expressions that require structural change in an example, must agree with the expression in the source text but other phrases that can be translated by general rule can be expressed by general expressions. It is better to use an example when input text has the same meaning with the example, even if it has the different expression on the surface. So, the model accepts the difference between input text and an example. The allowable differences are synonym, the conversion between numeral and ordinal, and the expressions that can be converted to equivalent expression. The model refers to the example dictionary in the phases that are provided between conventional phases and uses words and semanitics structures of examples prior to analysis and generation by general rules. The model is evaluated using about 115,000 examples, and the hit rate of examples and effective rate of expression conversion are calculated. The hit rate of examples is 65% in the sentences that require structural change or have collocations. The hit rate is about 1.7 times as much as it used to be by the use of expression conversion.
抄録全体を表示