Cache-based Online Adaptation for Machine Translation Enhanced Computer Assisted Translation

The integration of machine translation in the human translation work flow rises intriguing and challenging research issues. One of them, addressed in this work, is how to dynamically adapt phrase-based statistical MT from user post-editing. By casting the problem in the online machine learning paradigm, we propose a cache-based adaptation technique method that dynamically stores target n-gram and phrase-pair features used by the translator. For the sake of adaptation, during decoding not only recency of the features stored in the cache is rewarded but also their occurrence in similar already translated sentences in the document. Our experimental results show the effectiveness of the devised method both on standard benchmarks and on documents post-edited by professional translators through the real use of the MateCat tool.