Improved search strategy for interactive predictions in computer-assisted translation

Improved Search Strategy for Interactive Predictions in Computer-Assisted Translation Fatemeh Azadi ft.azadi@aut.ac.ir Department of Computer Engineering, Amirkabir University of Technology, Tehran, Iran Shahram Khadivi1 skhadivi@ebay.com eBay Inc, Aachen, Germany Abstract The statistical machine translation outputs are not error-free and in a high quality yet. So in the cases that we need high quality translations we definitely need the human intervention. An interactive-predictive machine translation is a framework, which enables the collabora- tion of the human and the translation system. Here, we address the problem of searching the best suffix to propose to the user in the phrase-based interactive prediction scenario. By adding the jump operation to the common edit distance based search, we try to overcome the lack of some of the reorderings in the search graph which might be desired by the user. The experiments results shows that this method improves the base method by 1.35% in KSMR2, and if we combine the edit error in the proposed method with the translation scores given by the statistical models to select the offered suffix, we could gain the KSMR im- provement of about 1.63% compared to the base search method. 1. Introduction Although the significant improvements achieved in the field of statistical machine translation, the current models and therefore the systems which can be built from them are still far from perfect. So, in order to achieve good or even acceptable translations, manual post editing is needed. An alternative to this approach is given by the interactive predictive machine transla- tion paradigm (Barrachina et al., 2009). Under this paradigm, translation is considered as an iterative process where the human translator and the computer collaborates to generate the final translation, and in this way both the human and the computer can help each other. In the interactive predictive MT scenario, the system initially offers a translation for a source sentence. From this translation, the user could mark a prefix as correct and begins to type the rest of the target sentence as it’s desirable for him. By typing the single next charac- ter (or maybe the full next word, according to the system settings) the interactive MT system would suggest a new suffix that completes the user’s confirmed prefix with the new character the user has just typed. This procedure continues iteratively until the user accepts the transla- tion to be complete and correct. One of the important issues in the development of interactive predictive MT systems is the strategy to search and offer the best suffix that completes the prefix confirmed by the user. One of the most common ways for phrase-based translation systems is to use the search graph that is obtained once during the translation process, and complete each user’s confirmed pre- 1 This work has been done when the author was with Amirkabir University of Technology. 2 Keystroke and mouse-action ratio Proceedings of MT Summit XV, vol.1: MT Researchers' Track Miami, Oct 30 - Nov 3, 2015 | p. 319 fix by searching the best path which …

Content truncated to 500 words