【作者】 王青；

【导师】 马萧；

【作者基本信息】 武汉大学 ， 翻译学， 2021， 博士

【摘要】 机器翻译是自然语言处理的范畴,其基本原理是利用计算机将一种语言文本转换成另一种语言文本。机器翻译的问题归根到底是语言的问题,牵涉语言的语法、语义、语境、上下文信息以及情感交流意向等关键性问题,而这些问题能否得到有效的解决则直接关系到机器翻译质量的优劣程度。近几年,随着机器翻译方式从基于规则、基于语料库到基于神经网络的不断拓展和深化,机器翻译的质量取得了突飞猛进的改善。目前,判断机器翻译质量优劣程度的方式主要有两种——人工评价模式和自动评价模式。前者的评价主体为人,以忠实和通顺标准为主要评价依据,而后者以机器为评价行为的主体,通常将相似度和错误率作为评价的参数和指标。这两种评价方式在某些方面对机器翻译结果的检测起作用,不足之处在于其评估理论与评测实践联系不紧密,相互关系不明显。在前人研究的基础上以及基于机器翻译的语言问题意识,本研究采用理论与技术、定性与定量相结合的研究范式,从汉英机器翻译语料着手,尝试构建一个融语言学、翻译学与计算机技术为一体的机器翻译质量评估模式:第一、以语言学知识为评价的参数来源。首先是词汇层面,忠实的程度体现在词语的使用频率和词语的搭配情况;其次是句子层面,忠实的依据在于句子的长度以及语法成分的一致程度;再者是语义层面,忠实的判断依赖于文本的相似度和文本主题的吻合程度;最后是语用层面,忠实的力度主要表现为情感状态和交流意向的统一性。第二、将具有双向意义的翻译忠实标准作为评价的准则和依据。与传统意义的忠实概念不同,翻译忠实标准的双向性是指翻译结果既应该忠实于原文的语法规则、语言意义、语用功能,也应该忠实于译文的语法规则、语言意义、语用功能,即原文的语言成分诸如词汇、短语、句子等要在译文中完完整整地表达出来,而译文中对应原文内容的转换诸如词汇的选择和句子的编排要符合目标语言的语法规则;原文中构成成分所体现出来的各种关系意义要毫无保留地在译文中得到体现,而译文中与原文意义对等的目标语言成分也应该符合译入语的意义表达方式;在原文中体验到的情感状态和交流意向在译文中也应该获得相同的体验,而译文中反映说话者/作者态度和意图的语言表达方式也要适应目标语言文化系统的语言情感倾向。第三、以计算机技术为评价的辅助手段。在该机器翻译质量评估模式中,数字技术对机器翻译语料的作用主要体现在:1)利用科大讯飞开放平台提供的接口和斯坦福大学开发的Stanford Parser进行中文文本以及英文文本的分词、词性标注、命名实体识别、词频统计、句法分析;2)利用互信息算法进行词语搭配分析;3)利用余弦相似度量方法进行文本相似度计算;4)利用隐含狄利克雷分布模型进行主题词分析等。与传统的纯理论与纯技术的机器翻译质量评估模式不同,该评估模式尝试将理论与技术相结合,其中,语言知识和翻译知识是理论依据,而计算机工具是技术手段,三者相辅相成、缺一不可。这项研究既继承了传统机器翻译研究的特点,也顺应了数字技术与翻译研究相结合的趋势,体现了一种新型的现代翻译研究范式。更多还原

【Abstract】 Machine translation is a category of natural language processing.Its basic principle is to use computers to convert one language text into another.The problem of machine translation is ultimately the problem of language,which involves such key issues as grammar,semantics,context,contextual information and emotional exchange intention,and whether these problems can be effectively solved is directly related to the quality of machine translation.In recent years,with the development of machine translation from rule-based,corpus based to neural network-based,the quality of machine translation has been improved rapidly.At present,there are mainly two ways to judge the quality of machine translation——the manual evaluation model and the automatic evaluation model.The evaluation subject of the former is human,and the criterion of faithfulness and smoothness is the main evaluation basis,however,the latter takes the machine as the main body of the evaluation behavior and takes the similarity and the error rate as the evaluation parameters and indexes.In some ways,these two evaluation methods play an important role in the detection of machine translation results.The deficiency is that the evaluation theory is not closely related to the evaluation practice,and the relationship between them is not prominent.Based on the previous researches and the awareness of language problems in machine translation,this research adopts a research paradigm combining theory and technology,qualitative and quantitative,and attempts to construct a model for Chinese-English machine translation quality assessment integrating linguistics,translation theories and computer technologies.From the perspective of languages,language knowledge is the source of evaluation parameters.At the lexical level,the degree of faithfulness is reflected in the frequency and collocation of words;at the sentence level,the degree of faithfulness is based on the length of sentences and the consistency of grammatical components;at the semantic level,the judgement of the degree of faithfulness depends on the similarity of the texts and the consistency of the text subjects;at the pragmatic level,the strength of faithfulness is mainly manifested in the unity of emotional states and communicative intentions.From the perspective of translations,the criterion of faithfulness with bi-directional significance is taken as the theoretical principle of evaluation.Different from the traditional concept of faithfulness,the bi-directional understanding of the criterion of faithfulness in translation refers to that the translation result should be faithful to the grammatical rules,linguistic meanings and pragmatic functions of the original text,as well as those of the target text,that is,the linguistic components of the original text,such as words,phrases and sentences,should be completely expressed in the target text,while the conversion in the translation,such as the choice of vocabulary and the arrangement of sentences,should conform to the grammar rules of the target language;All kinds of relational meanings embodied in the components of the original text should be reflected in the translation without reservation,and the target language components in the translation that are equivalent to the original meaning should also adapt to the ways of meaning expression in the target language;The emotional states and the communicative intentions experienced in the original text should be experienced in the translation as well,and the language expressions that reflect the attitude and the intention of the speaker or author in the translation should also be adapted to the emotional tendency of the target-language cultural system.From the perspective of tools,computer technology is used as the auxiliary means of evaluation.In this model,the value of technology is mainly reflected in the following aspects: 1)The interface provided by IFLYTEK open platform and Stanford Parser developed by Stanford university are used for word segmentation,part of speech tagging,named entity recognition,word frequency statistics,syntax analysis of Chinese and English texts;2)Mutual Information is used for word collocation analysis;3)Cosine Similarity is used for text similarity calculation;4)Latent Dirichlet Allocation is used for the analysis of theme words,etc.Different from the traditional assessment model based on pure theories or pure technologies,this evaluation model attempts to combine theories and technologies,regarding language knowledge and translation knowledge as the theoretical basis and and computer technologies as the practical tools.These three elements complement each other and are indispensable.This research not only inherits the characteristics of traditional machine translation studies,but also conforms to the trend of combining digital technology with translation studies,which reflects a new paradigm of modern translation studies.更多还原