Preliminary Experiments on Using Users’ Post-Editions to Enhance a SMT System Oracle-based Training for Phrase-based Statistical Machine Translation

Preliminary Experiments on Using Users’ Post-Editions to Enhance a SMT System Marion Potet1, Emmanuelle Esperança-Rodier1, Hervé Blanchon2, Laurent Besacier1 1UJF-Grenoble 1 ; 2UPMF-Grenoble 2 LIG UMR 5217, Grenoble, 38041, France FistName.Name@imag.fr Abstract The aim is to develop an interactive and iterative translation system able to adapt and improve itself rapidly using user’s feedbacks (collected as translation post- editions). We present a preliminary study in which three methods for integrating feedbacks are tested. The results show that, despite the difficulty to measure the system improvement with automatic met- rics (given the small size of our collected post-edition corpus) a small-sized subjec- tive analysis reveals the contribution and prospects of such methods. 1 Introduction New challenges and uses have emerged thanks to the development of the collaborative Web. The Wikipedia encyclopedia, the Google Image La- beler, the Amazon Mechanical Turk platform or Waze are well-known examples. The aim of these collaborative “services” is to solicit input from users to create content, solve problems and enrich database. Thus, in the field of Machine Translation, the same idea has recently appeared in order to overcome the current flaws of MT systems by using the users themselves and en- hance them as time goes on. There are two typical scenarios of the use of MT, namely assimilation (to get the gist) and dissemination. The second scenario is seen as a scenario for professional translators (it requires post-edition and revision). Our long-term goal is to propose post-edition in both scenarios in order for unskilled users to obtain better translations taking advantage of skilled users’ post-editions © 2011 European Association for Machine Translation. that will further improve the MT outputs. To sum up, we argue that MT systems for both assimila- tion and dissemination can “learn” from their users’ knowledge in order to improve themselves overtime through post-edition. For assimilation, post-editors would be either native speakers of the target language with a good knowledge of the source language or native speakers of the source language with a good knowledge of the target language. For dissemination, we aim at “dis- semination for every one”, post-editors would be professionals or have a good knowledge of the target language and MT would be used to im- prove productivity. In this paper, we experimented with collecting post-editions from native speakers of the source language with a good knowledge of the target language and integrating these contributions into the system in order to improve it. We have ini- tially created a statistical machine translation system (Section 3.1). Then, we have collected translations correction (post-editions) made by volunteer annotators (section 3.2). Finally, we propose and experiment with several methods, to use these post-editions to improve our system: adding data to the training corpus of the system (section 4.1); automatically correcting the out- puts of the system (section 4.2); and re-tuning the weights of the log-linear model (Section 4.3). 2 Context 2.1 Related Work Statistical machine translation (MT) techniques regained interest in the 1990s (Brown et al., 1990). Since then, the evolution of the …

Content truncated to 500 words