Efficient wordgraph for interactive translation prediction

Efficient Wordgraph Pruning for Interactive Translation Prediction Germ´an Sanchis-Trilles PRHLT Centre Univ. Polit´ec. de Valencia 46022 Valencia, Spain gsanchis@dsic.upv.es Daniel Ortiz-Mart´ınez PRHLT Centre Univ. Polit´ec. de Valencia 46022 Valencia, Spain dortiz@dsic.upv.es Francisco Casacuberta PRHLT Centre Univ. Polit´ec. de Valencia 46022 Valencia, Spain fcn@dsic.upv.es Abstract When applying interactive translation pre- diction in real-life scenarios, response time is critical for the users to accept the in- teractive translation prediction system as a potentially useful tool. In this paper, we report on three different strategies for re- ducing the computation time required by a state-of-the-art interactive translation pre- diction system, so that automatic comple- tions are delivered in real time. The best possibility turns out to be to directly prune the wordgraphs derived from the search procedure, achieving real-time response rates without any degradation whatsoever in the quality of the completions provided. 1 Introduction Despite the recent advances in machine transla- tion (MT) technology, MT systems are not able to provide ready-to-use translations in those contexts where translation accuracy is critical, such as med- ical or political applications, or even in contexts where correctness is demanded, such as hardware manuals or news texts. This has given rise to in- creasing research in computer assisted translation (CAT), where the focus is on how to provide a hu- man translator with the best tools available in order to improve the human’s efficiency. To this purpose, several ongoing FP7 projects were approved by the European Comission, some of them still being ac- tive. These projects pursue a very similar purpose, which is to develop a next generation CAT work- bench. c⃝ 2014 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. One of the most innovative research directions regarding CAT tools implies interactive translation prediction (ITP) (Barrachina et al., 2009). Un- der this paradigm, system and human translator interact more closely than in a conventional post- editing setup, and the ITP engine attempts to pro- vide improved completions for the sentence being translated after each one of the interactions of the human translator. Ideally, a constrained decoding, forced to produce the part of the sentence which has already been validated, should be performed before providing every suggestion. However, a full decoding process gives way to an important problem in ITP: the system needs to be able to provide translation completions in real time, since only a small delay in response time could easily lead users to reject the system. For this purpose, a common approximation is to extract a wordgraph off-line, i.e., before the user is actually sitting in front of the CAT tool. Then, during the ITP proce- dure, suggestions are obtained by searching for the best path in such a graph. In the present work we report on different ap- proaches analysed for the purpose of reducing the size of the wordgraph mentioned above when us- ing a state-of-the-art ITP system. Since response time is critical, we studied three different strategies and measured the response …

Content truncated to 500 words