Generate and repair machine translation.

摘要

We propose Generate and Repair Machine Translation (GRMT), a constraint-based approach to machine translation (MT) that focuses on accurate translation output. The architecture of GRMT was designed to take advantage of, and have advantages over, the three classic strategies (Direct MT, Interlingual MT and Transfer MT), the nonlinguistic information strategies (Example-Based MT and Statistics-Based MT), and the hybrid strategies (Knowledge-Based MT and Shake-and-Bake MT) with respect to several translation aspects: simplicity, accuracy and multilingualism.; GRMT performs the translation by generating a Translation Candidate (TC), verifying the syntax and semantics of the TC, and repairing the TC when required. GRMT comprises three modules: Analysis Lite Machine Translation (ALMT), Translation Candidate Evaluation (TCE), and Repair and Iterate (RI).; In generating the TC, GRMT refines the scope of translation choices of each input word by taking into account the differences between languages in a unique way. In selecting an appropriate word for each input word, GRMT considers the semantic relationship between words. This semantic relationship is based on the Word Association (WordAsso) number. (WordAsso). WordAsso number is assigned to word class. Words are classified according to the meaning of words and their usage. Word classification is designed and used not only in the word selection process but also in the classifier selection process and in semantic representation.; GRMT is highly modular and extendible in the following respects: each component is separated, not only the translation process components (ALMT, TCE, RI), but also in the knowledge-bases, each component can be extended easily to a larger domain. The adding of new languages is also possible since the source language (SL) and the target language (TL) are treated separately. The SL and the TL are connected via the SL-TL dictionary which contains simple information and is manageable.; An English-Thai translation system has been implemented to illustrate the performance of GRMT. The system has been developed and run under SWI-Prolog 3.2.8. The English and Thai grammars have been developed based on the Head-Driven Phrase Structure Grammar (HPSG) and implemented on the Attribute Logic Engine (ALE).; This English-Thai MT system was evaluated and it performs in the way we intended. ALMT generated acceptable translations (grammatically correct, correct word usage and convey the original meaning) for 47 out of the 90 sentences in the test corpus without repair. WE and RI improved 15 sentences using our current HPSG based grammars and lexicons. Twenty-one sentences which contain logical connections are first separated into linguistic units before the repair can be performed due to a current inadequacy in HPSG's semantic representation. However, each linguistic unit was then repaired successfully. Seven sentences faced with the problems of adding linking words and classifiers in Thai also require further research in order to develop ways to repair these sentences.