Combining tools to improve automatic translation

Combining Tools to Improve Automatic Translation Terence Lewis Hook & Hatton Ltd 1 St Giles Terrace Northampton NN1 2BN United Kingdom Hook_Hatton@compuserve.com Abstract This paper takes a practical look at ways of combining language engineering tools to produce more accurate, “more human” automatic translations. Whilst specific products are discussed, the author believes that the methodology could be successfully implemented with different sets of tools. Keywords Analysis, combination, procedure, separation, tools. Introduction This paper takes a practical look at ways of combining language engineering tools to produce more accurate, “more human” automatic translations. The tools involved are machine translation software, a translation memory application and alignment software, but also small tools or utilities written to perform simple yet very important tasks. The MT program discussed is the author's own Dutch-English translation software, which has been entirely rewritten in Java. The translation memory software used is Trados Translator's Workbench and WinAlign. All the utilities were written in Java by the author. However, the paper is concerned with presenting an approach or methodology which could conceivably be implemented with a totally different set of tools. Background Translation buyers will always prefer the “look and feel” of human translations but can be tempted by the considerable cost savings and speed offered by machine translation (with or without some post-editing). The approach presented here provides a way of increasing the “human look and feel” of automatically generated documents. The use of the word "human" with reference to computer-generated translations is, of course, paradoxical. Can the product of a translation program ever look and feel human? Some members of the translation community would answer this question with a resounding "No". And, if we were to confine our discussion to translation engines that provide no more than word replacement and some basic rearrangement, they would broadly be right. At the level of very simple instructions and technical descriptions, the computer and the human translator may sometimes produce an identical translation. But when it comes to the choice of adjectives, the positioning of adverbs, converting the passive voice to the active voice and making some passives active, or the use of an adverb instead of an adverbial phrase, computer and human will invariably part company. While it is possible to hard-code idiomatic translations and write rules for converting the passive to the active voice, human translations of complex sentences are nearly always more readable. The wide availability of relatively low-cost translation memory applications has certainly led some members of the translation community to prematurely engage in wild shenanigans on the freshly filled graves of some poor MT packages. But translation memory software alone cannot meet the rapidly growing demand for "information-level" or draft translations. Translation memories still have to be created and stocked by humans, and yet many documents contain parts that could be adequately drafted by "professional-level" translation software. The realities of the marketplace led to author to experiment with ways of using the best bits of the two key language engineering technologies. Our fundamental principle in …

Content truncated to 500 words