Comparison of generation strategies for interactive machine translation

Comparison of Generation Strategies for Interactive Machine Translation Oliver Bender, Saˇsa Hasan, David Vilar, Richard Zens, and Hermann Ney Lehrstuhl f¨ur Informatik VI, Computer Science Department RWTH Aachen University D-52056 Aachen, Germany {bender,hasan,vilar,zens,ney}@cs.rwth-aachen.de Abstract. Fully automatic translations are far from being perfect. Non-grammatical sentences are often produced by automatic systems and there is even no guarantee that the meaning of the sentence is preserved. Nevertheless, automatic translation systems can be used to help human translators to produce high-quality translations. This is the goal of the TransType2 project, where an interactive translation tool is being developed that suggests, in real time, possible completions for the sentences that the human translator is typing. This leads to a modification of the generation strategy of the translation system, as now we are looking for the best translation of the given source sentence that is compatible with the prefix. In order to remain within the tight response time constraints of such a system, some simplifications have to be done. In this paper, we review possible generation strategies for an interactive statistical machine translation system and analyze what is the loss in performance when strict time constraints have to be met. Experiments are performed on the Spanish-English and German-English Xerox corpora, which consist of the translation of technical manuals, and the results show that the real time generation strategy causes only a small performance degradation. 1 Introduction Although great progress has been made in the field of automatic translation in the last years, the pro- duced translations are far from being perfect. Apart for very limited domains, no current state-of-the- art system can be directly used for real life applica- tions. The produced sentences often contain gram- matical errors and the preservation of the meaning is not even always achieved. Therefore a manual post-processing of the texts has to be done. The concept of interactive machine translation already has a long history, and the first systems ap- peared in the end of the 1960’s. However in most of these systems the user doesn’t have a direct control over the translation process, and most of the user in- teraction is reduced to performing source language disambiguation on demand. The approach we cen- ter on in this work was first suggested by (Foster et al., 1996) and an implementation was carried out in the TransType project (Langlais et al., 2000). In such an environment, human translators inter- act with a translation system that acts as an assis- tance tool and dynamically provides a list of trans- lations that best complete the part of the source sen- tence already translated. Further refinements were presented in the TransType2 project (Schlumberg- erSema S.A. et al., 2001). The work presented in this paper deals with generation strategies for interactive (statistical) ma- chine translation systems. Clearly, the best ap- proach would be to start a new search for every given prefix. However, in these kind of systems, response time is a crucial factor for a human trans- lator, as delays higher than a …

Content truncated to 500 words