Research On Machine Translation Evaluation Metrics Based On Reference Graph

With the rapid development of science and technology,machine translation tech-nology has got more and more attention from human,its theoretical research and prac-tical applications have become a hot topic in the field of machine learning.Compared with the manual translation,the translation efficiency has been improved obviously and the cost is reduced.Therefore,the research of automated translation technology(Machine Translation)has become a very important research direction.With the development of machine translation technology,the research project of the evaluation of machine translation came into being,we need to measure the quality of the output of a machine translation system.Completing the machine translation it-self is not the ultimate goal,we want to know how much the machine translation can help people,in the same time,having a rough assessment of the quality of a machine translation also help to learn a better machine translation system,these two comple-ment each other.For machine translation evaluation,the traditional way is to carry out manual evaluation,but because of its time-consuming,expensive and non-reproduced results,it has been partially replaced by automatic evaluation methods.The current mainstream automatic evaluation methods use the same strategy,they compare system output with one or several human translations(Reference Translation),and think that an system output which is very similar to references is certainly more accurate than those quite different ones.However,due to the diversity of semantic and expression,a source sentence may have many different correct translations(Translation Diversity),while,the reference translation is limited,this will lead to the fact that in some cases,the automatic evaluation methods can not accurately assess the quality of the system output,and also limits the translation skills of the machine translation systems.In this paper,we make improvements to the weakness of current methods.Due to the expensive cost of human translation,we propose automatic methods to explore the information among the limited references.Firstly,we explored the influence of using different evaluation metrics and choosing different reference translations on ma-chine translation systems.Secondly,expanding the existing reference translation.For datasets having several references,we generate a reference graph to explore information from them.For datasets having only one reference,we use paraphrase table to extend it and get a reference graph with more information.Finally,finding the optimal path from the reference graph to improve efficiency,and applying the automatic method to the existing automatic evaluation methods to help enhance the translation capacity of the machine translation system itself.The experimental results show that our proposed method did obtaining more in-formation from limited resource.For the expansion of multiple reference translations,the system output can be more accurately evaluated.For those sets having single ref-erence,it can not only judge more accurately but also reduce the difference between systems using different references.It also effectively improves the machine translation system by applying it to the parameter learning.