Automatic Evaluation Method Of Machine Translation Based On Text Structure

With the continuous development of globalization,especially the construction of the Belt and Road,the trade between countries along the Belt and Road is getting more frequent.Traditional human translation cannot meet the needs of interlanguage translation.In this process,machine translation plays a significant role.In the process of machine translation development,the evaluation of translation systems is accompanied by the continuous development of machine translation.Machine translation human evaluation and machine translation automatic evaluation are the two main types of evaluation techniques of translation systems.Although the early human evaluation is highly accurate,there are still many problems.On the one hand,it takes considerable manpower and time to accurately evaluate the system.On the other hand,human evaluation has the very large subjectivity.On the contrary,the automatic evaluation metric aims to provide fast,cheap and objective numerical measurement of translation quality,which can make up for the deficiency of the human evaluation metric,making it one of the important contents of machine translation development research in recent years.So far,there have been many representative metrics for automatically evaluating the quality of machine translation.The scores of some evaluation metrics are obtained by cumulative average of the data after sentence evaluation in system level.Such evaluation metrics will lead to the neglect of the entire discourse structure.The use of discourse information is very important for the development of machine translation automatic evaluation technology.Discourse-based evaluation metrics have been followed by many researchers,but for various reasons,their development is relatively slow.This paper mainly studies the automatic evaluation method of machine translation based on discourse structure.The main contents include:1.Machine translation automatic evaluation metric based on discourse coherence C-ENTF.In order to better evaluate the results of machine translation at the document level,we propose an improved automatic evaluation metric C-ENTF by integrating discourse coherence information.First,the discourse is represented as a graph by calculating the correlation between each two sentences.Second,we use the sub-graph frequency of the graph to represent the discourse coherence information.Finally,we incorporate the discourse coherence information into the existing automatic evaluation metric ENTF.The experimental results show that the performance of C-ENTF is improved after integrating discourse coherence information.On WMT 2015,C-ENTF performance increased by 0.5 percentage points compared to ENTF.2.Machine translation automatic evaluation metric based on discourse representation structure DRS-ENTF.In order to better evaluate the results of machine translation at the document level,we propose an improved automatic machine translation evaluation metric DRS-ENTF by integrating the discourse representation structure.The original automatic evaluation method,ENTF,used the average of all sentence scores when calculating the system-level scores,and did not actually use the discourse information.DRS-ENTF adds discourse structure information to ENTF.This method can obtain discourse semantic information to some extent to make up for the problem that ENTF does not use discourse information.The experimental results show that compared to ENTF,DRS-ENTF increased by 0.9 percentage points on WMT 2015,which shows that the performance of DRS-ENTF has been improved after incorporating the discourse representation structure.3.Machine translation automatic evaluation metric integrating multiple discourse structures DS-ENTF.The first two methods show that the use of discourse coherence and discourse representation structure can play the certain role in improving the performance of the automatic evaluation metric,and the two methods reflect different aspects of the discourse structure.Therefore,we can simultaneously incorporate discourse coherence and discourse representation structure into the existing automatic evaluation method ENTF.The experimental results show that the performance of DS-ENTF is improved after the integrating of two different types of discourse structures.Compared to C-ENTF and DRS-ENTF,DS-ENTF performance increased by 0.8 and 0.4 percentage points respectively on WMT 2015;DS-ENTF performance increased by 1.3 percentage points compared to ENTF.