| dc.description.abstract | Dialogue translation is designed to translate the text of conversations between people of different languages. Most of machine translation tasks use Sentence-level translation models, but such architectures can’t effectively capture contextual relationships, resulting in the inability to express cross-sentence semantics in a correct way. However, a Document-level translation architecture can deal with the above problems.
Document-level translation model can be used for the task of dialogue translation, but how to choose a useful context to help the current sentence to be translated is still an open issue. In addition, there are fewer studies focusing on the influence of historical chat records on translation effects. The historical information can also effectively help the current sentence to be translated, so we think it should be taken into account. Therefore, we propose a new way of context selection and a model adapted to dialogue translation, called United method and Triple-Unified-Transformer, respectively. Our model can better learn the relationship between sentences, so that the current sentence to be translated can get better results.
In experiment 1, we use three different chat translation datasets and auto evaluation metrics to measure the effectiveness of our proposed United method and Triple-Unified-Transformer. The results show that the performance of one of the baseline models after adding the United method can improve the translation effect. On the other hand, the translation effect of Triple-Unified-Transformer achieve good results in specific datasets (BLEU, BLONDE). Furthermore, In experiment 2, we test the United method and Triple-Unified-Transformer are general and can adapt to different chat situations. The results show that Triple-Unified-Transformer performs best in a specific dataset, which means that our model has the ability to apply in different chat situations, and the baseline model with the United method can perform better. | en_US |