Research On Optimization Technology Of Statistical Machine Translation Based On User Feedback

With the rapid development of Internet, cross-lingual communication becomes more and more frequent, which brings about an increasing demand for real-time translation between different languages. In recent years, Machine Translation(MT) technology has achieved great progress. However, in the domains with high translation quality requirements, the MT output is still not satisfactory. Recently, researchers incorporate more linguistic knowledge into the statistical methods. However, the improvement on the performance of MT systems is still limited.Under such circumstances, researchers began to seek new solutions. A typical one is having users perform post-editing on the output generated by the MT system, until correct translation is achieved. Researchers found that the user feedback during the human-computer interaction process can be used to optimize the MT system and improve the translation performance continuously. However, while integrating knowledge from different users, current approaches treat all the users equally. In fact, different users have different translation experience, so the confidence of the translation knowledge they feedback is also different. It is important to distinguish users while integrating their knowledge to improve the MT system.Aiming at the above problem, we explored the following aspects. We analyzed the factors that influence the users' confidence and introduced the basic features and translation features to construct the user confidence evaluation model. This model can be used to distinguish the translation knowledge from different users, and update the parameters of phrase table in real time. The modified parameters contain forward phrase translation probability, forward lexical translation probability, reverse phrase translation probability and reverse lexical translation probability. Experimental results show that after using the translation knowledge from users' feedback to optimize the MT system, the translation quality was improved. And distinguishing the knowledge from different users' feedback achieves better optimization performance than not distinguishing users.