Research On Interactive Phrase-based Machine Translation

In current society, the communication between countries is becoming more and more frequent. The demands for translation between different languages become ur-gent. Due to the fact that human translation is relatively inefficient and expensive, automatic translation technology (machine translation) has become an important re-search area. After a long period of development, statistical machine translation has been mainstreamed.Despite the rapid development of statistical machine translation, the translation quality of statistical machine translation is still not perfect due to the complexity of nat-ural languages. Users still need to improve translation quality through human-computer interactions. Interactive machine translation is such a technology that improves the quality of machine translation by human-computer interactions. The demand for in-teractive machine translation is very large because machine translation is not perfect nowadays. Users can only perform left-to-right correction when using current inter-active machine translation systems, so the information provided by users is relatively few. It will limit the ability of interactive machine translation system.In this paper, we make improvements to the weaknesses of the current interactive translation systems. Firstly, we introduced a new interactive translation framework and the corresponding constrained decoding algorithm. Users can modify translation errors at any position instead of left-to-right correction using our framework. With human-computer interactive information, the translation system can generate better translations with constrained decoding algorithm. Secondly, to further reduce human interaction, we proposed automatic suggestion models that can offer suggestions for users automat-ically. Thirdly, we extended the interactive translation framework to reduce translation reordering errors. Finally, we used human-computer interactive information to adapt statistical models and improve the ability of translation system itself.The experimental results show that our proposed framework can improve trans-lation quality much faster than traditional framework. The human interactions can be significantly reduced by our framework, too. With the help of our automatic suggestion models, the human interactions can be further reduced. The extension of our framework can reduce the translation reordering errors based on user provided information. With the help of the historical information of human-computer interaction, our model adapta-tion methods can update the translation system and improve the ability of the translation system.