An efficient phrase-to-phrase alignment model for arbitrarily long phrase and large corpora

EAMT 2005 Conference Proceedings 13 Codified Close Translation as a Standard for MT Lars Ahrenberg Department of Computer and Information Science Linköpings universitet, SE – 58183 Linköping, Sweden lah@ida.liu.se Abstract. This paper proposes that quality criteria are set up for machine translation sys- tems that attempt to capture what can be considered the upper limit for the performance of MT systems. It argues that the best performance MT currently should aim for is close trans- lation, and that attempts should be made to codify what that entails for any given language pair. Some guidelines for the selection of close translations are proposed and their applica- tion on concrete examples drawn from an English-Swedish parallel corpus is discussed. The definition of different quality levels in relation to a high-quality standard is also touched upon. 1. Introduction Machine translation research and development have long since abandoned the goals of "Fully automatic high quality" translation. While this is obviously a realistic decision, one may won- der if anything has come to replace it. Practical utility is perhaps the most obvious answer and experience tells us that MT systems of very dif- ferent quality and coverage may be used to good advantage. Evaluation is another aspect that has come to the fore as a prerogative for advances in ma- chine translation. Coupled with practical utility evaluation tends to focus on factors that are relevant for the purpose at hand and translation quality may not be the most prominent factor then. However, disregarding the price factor, there can be no doubt that systems are generally more useful the larger their coverage is and the higher their accuracy. As regards translation quality, evaluation re- fers to criteria such as fidelity, intelligibility and fluency (White, 2003). These aspects are usually measured by comparisons with refer- ence translations produced by human transla- tors, whether by human judges or automatically (Papineni et al., 2001). While the reference translations are usually good translations, or even "expert translations", their qualities, or the requirements given to the translators, are sel- dom discussed in any detail, at least not pub- licly. Similarly, parallel corpora, such as the Hansards, that are used for training statistical MT systems, are produced by human translators aiming at high quality by human standards. But this quality level may actually be beyond reach for any known system and translation approach developed to date. The question is, then, what output quality we can expect and demand from a MT system, to- day and in the future. This is an important ques- tion both for researchers, potential customers, and the society at large. Researchers would benefit from having challenging though realistic goals to reach. Customers and users would benefit from having standards to compare with. In particular, it would be useful for customers to have access to quality labels that would give rough but reliable information on the linguistic scope and quality of the output from the system (cf. Hutchins, 2000). Finally, language commu- nities would be better off if the …

Content truncated to 500 words