Towards semantics-oriented machine translation: Using synonym comparisons to interpret natural language.

摘要

This study investigated a novel type of semantic analysis, involving synonym comparisons within a hierarchical category structure, as a way of overcoming the "semantic barrier" problem in machine translation (MT). The premise is that the words in a sensible sentence will logically define themselves simply on their relationships with the other words in the sentence.;The process involves comparisons of every word meaning (and its synonyms) with all other word meanings (and their synonyms). The strongest comparisons, computed using a type of genetic algorithm, are chosen as the "best" interpretation of the sentence. Unlike MT systems that require language "scripts" to help determine the meanings of words, the synonym comparison method permits an unrestricted vocabulary.;The hierarchy of semantic categories in Roget's International Thesaurus, fourth edition was chosen as the medium for synonyms. The synonym comparison method was shown to be interlingual by verifying that all the categories exist in the following "major" language families: Chinese, Germanic, Indic, Japanese, Malayo-Polynesian, Romance, Semitic, and Slavic. These languages are spoken by over 70% of the world's people. The only design limitation is that sentences must be semantically sensible--nonsense sentences or sentences that are completely ambiguous are not allowed.;A demonstration system was developed on a personal computer with English as the source language and German as the target language. The results were generally successful in translating sentences using only synonym comparisons. The interlingual nature of the method implies that this approach will work for any of the language families investigated, and the results lend support to the implication in transformational grammar that sensible, properly constructed sentences contain their own "natural" deep structure meaning.