MACHINE TRANSLATION FOR FINANCIAL NEWS

Abstract

This thesis presents a project on machine translation (MT) leading to a practical system for English-Chinese financial news translation. An MT method for English-Chinese translation of texts written in prose (excluding the literary and ungrammatically colloquial) is proposed. Based on it, a prototype microcomputer-based MT system has been developed. The eventual operational system will be used by a newspaper company, Singapore Press Holdings (SPH), for translating financial news received through telex from overseas news agencies. Final human touch-ups will be required for the system output. The proposed method translates at the sentence level through a process comprising eight major phases. It attempts to achieve the best possible translation in the most efficient way. Therefore, unambiguous words, phrases and collocations are substituted by their unique equivalents prior to any sentence analysis. The source sentence is analyzed only as much as necessary at later phases whenever more information is needed for resolving homographs or rearranging sentence elements to conform to the target language. During the process, a proper and complete target sentence is gradually formed. In order to obtain good-quality output, selection of appropriate equivalents and sentence reordering are always considered in various contexts, from the most specific in meaning to the general, and the largest in size to the smallest. A corpus of financial news articles selected by SPH's news editors has been used to test the prototype. Half of the articles involve many special financial terms, phrases and expressions while the others can be classified as general texts in terms of word or phrase usage and sentence construction. The machine output has been assessed by three groups of evaluators representing the general newspaper readers, SPH and the MT circle respectively. Results of the evaluations reveal that the method is capable of handling a wide variety of sentence patterns and different types of sentence structural differences. Hence, it is general enough for translating various types of texts. Moreover, it has the advantages of being easily extensible to other subject domains and highly adaptable for another language pair. Extension is possible by merely developing additional domain-specific dictionaries without having to alter the translation algorithmn or the existing dictionaries and rules. The adaptability of the method is supported by its almost entire adoption by a system for Malay-English government report translation. The translation quality of the prototype is deemed to be reasonable by human standard and comparable to or better than some commercially available English-Chinese systems. More importantly, it has been considered by SPH as good enough for system adoption. An overall increase in productivity is anticipated. Other benefits to be gained include the consistent translation of proper nouns, special terms, phrases and expressions, and the elimination of errors due to misinterpretation of handwriting. Furthermore, the system can serve as an aid for dictionary lookup and staff training. Therefore, we have made a proposal for integrating the system and other supporting activities to create a better translation environment. Included also in this thesis is a general survey on the MT field.