面向社交媒体的多语种文本情感分析方法研究

With the rise of social media, the need for cross-language communication keeps increasing. Currently, the research field of text sentiment analysis focuses more on the study of languages such as Chinese and English, and less on Russian, and even less on the fusion of multiple languages. Russian, as an important language branch of the Eastern Slavic languages, is the most widely used language in the "Belt and Road" center line. Full understanding of the attitude of the people of the Russian-speaking countries towards the "Belt and Road" initiative is conducive to the smooth implementation of relevant policies. In view of the current situation and the intrinsic characteristics of Russian, and based on the construction of Russian semantic resources such as sentiment dictionary and sentiment corpus, this project aims to apply the sentiment analysis method designed for languages with adequate corpus resources into Russian, the language with relatively scarce resources so as to build a multilingual sentiment analysis model and develop a platform for sentiment analysis of multilingual texts, by digging out the multi-dimensional and deep semantic meaning of sentiment lexicon through comprehensive analysis of their phonetic, morphologic and semantic elements, integrating contextual knowledge and common sense contained in the texts, and utilizing Deep Learning and Transfer Learning such as LSTM to realize sentiment classification. The multilingual sentiment analysis will lay the foundation for cross-linguistic research and also serve as a helpful attempt at transferring the sentiment analysis method from languages with adequate resources to languages with scarce resources so as to provide technical support for the smooth implementation of the "Belt and Road" initiative.

随着社交媒体的兴盛，跨语言的交际需求越来越多。但目前文本情感分析领域偏重于中文和英文等语种的研究，俄文等语种的情感分析较少，多语种融合的研究更少。俄语作为东斯拉夫语支的重要语种，是“一带一路”中心线里使用最广泛的语言，充分了解俄语区民众对“一带一路”政策的态度有利于相关政策的顺利实施。针对目前的情况，本项目从俄文自身的特点出发，将资源充足语种的情感分析方法迁移到资源相对匮乏的语种中，在构建俄文情感词典及情感语料库等语义资源基础上，通过音形义结合的方式多维度深层次挖掘情感词汇的语义表示，融合文本中蕴含的语境知识和常识，利用LSTM等深度学习和迁移学习的方法实现情感分类，进而构建一个多语种融合的情感分析模型，研制多语种文本情感分析平台。多语种的文本情感分析将为跨语言的研究奠定基础，也为资源充足语种到资源匮乏语种的迁移方法进行有益尝试，从而为“一带一路”倡议的顺利实施提供技

随着互联网特别是移动互联网在全球的爆发，全球网民数量暴增，大量的用户在各社交媒体平台中产生极为巨大的数据，这些数据通常包含英文、中文和俄文等多种语言，探索多个语种社交媒体文本中蕴含的民意和情感倾向性具有重要的研究价值。目前，情感分析领域中对中文和英文的研究较多，俄语情感分析的研究较少，因此本项目以社交媒体为载体，分析了多种语言文本中蕴含的情感信息，建立情感分类模型，着重研究了目前资源匮乏的俄文情感分析问题。. 在资源建设方面，构建了多语种情感分析平行语料库和多模态情感语料库，语料库建设中建立了多重的质量保障机制以确保数据标注的一致性，标注中制定了完备的标注规范和标注流程，对其他类型情感分析语料库的建设工作具有较大的参考价值。在情感分析模型的构建方面，通过音形义结合的方式，多维度深层次挖掘情感的语义表示，并融入文本中蕴含的语境知识和常识，完成了面向社交媒体的情感分析模型。情感分析模型分为显式和隐式两种类型，提取俄语中词汇级和语句级的情感特征，在此基础上提出一种基于自注意力机制的CNN-BiLSTM混合模型。该模型探索了资源匮乏语种的情感分类方法，所提出的方法可以应用在舆情分析和评论挖掘等领域。. 围绕着课题的研究内容及工作，发表或录用学术论文23篇，其中SCI检索论文和国际会议论文13篇，北大核心期刊论文10篇，获得软件著作权1项，申请专利1项。