TransBooster: boosting the performance of wide-coverage machine translation systems

EAMT 2005 Conference Proceedings 189 TransBooster: Boosting the Performance of Wide-Coverage Machine Translation Systems Bart Mellebeek, Anna Khasin, Josef Van Genabith and Andy Way National Centre for Language Technology School of Computing Dublin City University, Dublin 9, Ireland. {mellebeek,akhasin,josef,away}@computing.dcu.ie Abstract. We propose the design, implementation and evaluation of a novel and modular approach to boost the translation performance of existing, wide-coverage, freely available machine translation systems based on reliable and fast automatic decomposition of the translation input and corresponding composition of translation output. We provide details of our method, and experimental results compared to the MT systems SYSTRAN and Logomedia. While many avenues for further experimentation remain, to date we fall just behind the baseline systems on the full 800-sentence testset, but in certain cases our method causes the translation quality obtained via the MT systems to improve. 1. Introduction A significant number of freely available, com- mercial, wide-coverage machine translation (MT) systems populate the market, including SDL In- ternational’s Enterprise Translation Server, Re- verso by Softissimo, Logomedia, Promt and perhaps best known BabelFish by AltaVista. Some of these systems are based on 1st gen- eration ‘word-for-word’ direct translation tech- nology, and share a number of characteristics: (i) they are designed to translate wide-coverage, general language material; (ii) they are robust; and (iii) they perform comparatively limited linguistic analysis. These points are interrelated and bear further elaboration. Detailed automatic linguistic analysis of translation input is poten- tially costly (both in terms of processing time and required lingware such as computational grammars etc.), and in the past has often been inversely related to coverage and robustness. In other words, the more detailed the linguistic analysis, the smaller the coverage of the system, and conversely, the wider the coverage, the less detailed the linguistic analysis. This has led commercial, wide-coverage MT systems to con- centrate on ‘linguistics-lite’, robust design prin- ciples. In order to analyse translation input, they often consider only a limited linguistic context. A consequence of this is that existing commer- cial systems are much stronger when translating shorter sentences than they are on longer, more complex input. The reason behind this is sim- ple: the longer the input sentence to be trans- lated, the more likely that the automatic transla- tion system will be led astray by the complexi- ties in the source and target languages. We contend that better performance in terms of output quality can be achieved than these systems can obtain by processing the texts that they are required to translate at any one time into smaller chunks. Consider the example in (1): (1) The chairman, a long-time rival of Bill Gates, likes fast and confidential deals A reasonable translation into German (estab- lished by a human translator) is: (2) Der Vorsitzende, ein langfristiger Ri- vale von Bill Gates, mag schnelle und vertrauliche Abkommen. However, the translation produced by the Ba- belFish MT system is (3): (3) Der Vorsitzende, ein langfristiger Ri- vale von Bill Gates, Gleiche fasten und vertrauliche Abkommen. Mellebeek, et al. 190 EAMT …

Content truncated to 500 words