Online multi-user adaptive statistical machine translation

Online Multi-User Adaptive Statistical Machine Translation Prashant Mathur prashant@fbk.eu FBK Trento, Italy DISI, University of Trento, Italy Mauro Cettolo cettolo@fbk.eu FBK Trento, Italy Marcello Federico federico@fbk.eu FBK Trento, Italy Jos´e G. C. de Souza desouza@fbk.eu FBK Trento, Italy DISI, University of Trento, Italy Abstract In this paper we investigate the problem of adapting a machine translation system to the feed- back provided by multiple post-editors. It is well know that translators might have very dif- ferent post-editing styles and that this variability hinders the application of online learning methods, which indeed assume a homogeneous source of adaptation data. We hence propose multi-task learning to leverage bias information from each single post-editors in order to con- strain the evolution of the SMT system. A new framework for significance testing with sentence level metrics is described which shows that Multi-Task learning approaches outperforms exist- ing online learning approaches, with significant gains of 1.24 and 1.88 TER score over a strong online adaptive baseline, on a test set of post-edits produced by four translators texts and on a popular benchmark with multiple references, respectively. 1 Introduction In a professional localization environment, a document is post-edited by several professional translators with assistance of tools such as translation memory, dictionary, spell checkers etc. To speed up the process, lately localization companies have started using computer assisted translation (CAT) tools with statistical machine translation (SMT) systems in the backend. The role played by the SMT engine is to provide a translation hypothesis that the translator can post edit to produce high quality translations (Federico et al., 2012). In recent works on online adaptation by Mathur et al. (2013) and Denkowski et al. (2014), the SMT is fed with the post edited sentence, allowing the models to adapt to the corrections made by the translators. These kind of systems works well if the document is being post edited by a single translator because models can adapt to the style of that translator. Problems arise when a document is being post edited by a group of translators which is usually the case with big size documents. In fact, if the SMT system adapts to the corrections of all translators together, it will likely mix or overlap stylistic features of the post-editors and thus not learn to mimic well any of them. On the other side, if the system adapts to each individual post-editor, then clearly Al-Onaizan & Simard (Eds.) Proceedings of AMTA 2014, vol. 1: MT Researchers Vancouver, BC © The Authors 152 useful feedback from other post-editors gets wasted. The main motivation for adapting a SMT system in the backend of CAT tool is that the translation improves over time since fewer mistakes are made after learning from post editions. For example, translator A does not post edit phrase Fibre Channel because he thinks that the phrase is a named entity, while B post edits Fibre Channel → Canale a fibre because he does not recognize it as a named entity. In the backend the SMT system …

Content truncated to 500 words