机器翻译系统翻译质量的人工评价模型

发布时间：2018-06-18 18:38

本文选题：机器翻译 + 人工评价　；参考：《广东外语外贸大学》2017年硕士论文

【摘要】：机器翻译研究发展迅速已历经数个重要阶段。它从早期的基于规则的方法发展到基于实例的方法再到统计方法。而机器翻译的翻译评价一直是其中一个活跃的领域。该领域有令人鼓舞的成果,但困难依然存在。自动评价标准不够准确,而翻译质量评估的标准也难以适应机器翻译的质量评价。该论文提出了机器翻译的人工评价模型,旨在解决以上的问题。已有的统计机器翻译的评价指标要么是仅仅从词语选择和词语顺序方面评价译文或者是它们的评价标准呈现相互交叠的情况。该模型包含两个方面:评价指标、评价指标的权重。评价指标有三个:信息性、语法性、语言多样性。该模型的构建基于系统思想理论,一个完整的系统应包含组成元素、相互关系、以及组合成后的系统的复杂性。该论文由组成元素对于系统的意义推导出译文评价的信息性指标,由相互关系对于系统的意义推导出译文评价的语法性指标,以及同理可得的语言多样性指标。评价指标的权重反映了每个指标之间的相对重要性的关系。在对模型的有效性进行测试中,本文选取了具有代表性的两个机器翻译系统:谷歌翻译和百度翻译。同时,该测试还在新闻,经济和体育三个领域随机选取了30个汉语句子,作为翻译原文本。得出的译文将由Asiya(开放在线机器翻译评价软件)上面的METEOR进行打分,同时也将由人工评价模型进行打分。该论文分析比较了由两种评价工具所得的分数结果:METEOR和人工评价模型在评价结果上是一致的。进一步的分析得出人工评价模型的分数结果比METEOR的更加稳定,人工评价模型的分数结果的变异系数是6.58%,METEOR的变异系数是29.1%。因此,该人工评价模型能够提高机器翻译质量评价的准确性。同时,该模型因为评价指标本身可以量化的特点,所以能够将该模型应用到机器翻译系统的自动评价标准里,提升评价准确性,帮助不断改善机器翻译的译文质量。
[Abstract]:The rapid development of machine translation research has gone through several important stages. It developed from the early rule-based approach to the case-based method and then to the statistical method. The translation evaluation of machine translation has always been an active field. There have been encouraging results in this area, but difficulties remain. The automatic evaluation criteria are not accurate enough, and the translation quality assessment criteria are difficult to adapt to machine translation quality evaluation. In this paper, a manual evaluation model of machine translation is proposed to solve the above problems. The existing evaluation criteria of statistical machine translation are either evaluated only in terms of word selection and word order or their evaluation criteria overlap with each other. The model includes two aspects: the evaluation index and the weight of the evaluation index. There are three evaluation indicators: informational, grammatical and linguistic diversity. The model is based on the theory of system theory. A complete system should include the elements, the relationship, and the complexity of the combined system. In this paper, the information index of translation evaluation is derived from the meaning of component elements to the system, the grammatical index of translation evaluation and the linguistic diversity index are derived from the meaning of interrelation to the system. The weight of the evaluation index reflects the relative importance of each index. In order to test the validity of the model, this paper selects two representative machine translation systems: Google Translation and Baidu Translation. At the same time, the test also randomly selected 30 Chinese sentences in news, economics and sports as the original text. The resulting translation will be scored by Meter on Asiya (Open online Machine Translation Evaluation Software) and by a manual evaluation model. This paper analyzes and compares the score results obtained by two evaluation tools: Meter and the artificial evaluation model in the evaluation results are consistent. Further analysis showed that the score of artificial evaluation model was more stable than that of Meter. The coefficient of variation of the score of artificial evaluation model was 6.58 and the coefficient of variation of METEOR was 29.1. Therefore, the artificial evaluation model can improve the accuracy of machine translation quality evaluation. At the same time, the model can be applied to the automatic evaluation standard of machine translation system because of the quantifiable characteristics of the evaluation index itself, so it can improve the accuracy of evaluation and help to improve the translation quality of machine translation.
【学位授予单位】：广东外语外贸大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：H085

【参考文献】