经济评论中的经济特征情感分析

A great amount of economic reviews have accumulated on the Web, which provide a new way to understand current economic climate and predict economic trends. .In economic reviews, the economic aspects typically are large in number, varied in frequencies and complex in structure, which makes extracting and identifying economic aspects very difficult. What is more, the opinion words in economic reviews are more flexible in POS (part of speech), and the sentences are usually verbose, which adds the complexity of building opinion dictionary and extracting sentiment analysis unit. Last but not the least, the negative economic aspects and numeric modifier are also obstacles in designing sentiment analysis model..In this project, we investigate sentiment analysis of economic reviews, including the identification and grouping of economic aspects, the building of economic sentiment dictionary, sentiment analysis unit extraction and sentiment analysis of economic aspects. First of all, we plan to propose methods of acquiring correlation constraints and adapting the topic model, which will be used to identify low-frequency economic aspects, and reduce the adverse impact of high-frequency economic aspects on grouping coverage. Second, the economic sentiment dictionary and negative economic aspects dictionary will be built, which are the basics of text sentiment analysis in economic area. Third, in this research we will design a fine-granularity sentiment analysis model based on economic aspects. Especially, we will design a model combining negative economic aspects and opinion words, and a dynamic context-dependent model to translate the numeric modifiers to discrete levels. These models will be used to tackle the unique problems that are not faced in existing product reviews sentiment analysis..Overall, we believe the research has both important scientific significance and practical application value.

Web上经济评论大数据的涌现，为获取当前经济形势、了解经济发展趋势提供了新途径。经济评论中，经济特征数目众多、频数不一、构成形式复杂，增加了经济特征抽取的难度；情感词的词性更丰富且经济评论中多长句，增加了情感词典构建及情感计算单元抽取的复杂度；奇异经济特征和数字修饰等多种修饰成分，更是增加了情感计算模型设计的难度。本项目从经济特征识别与分组、经济域情感词典构建、情感计算单元抽取和经济特征情感计算等方面开展研究。拟提出关联约束的获取及主题模型的改造方法，以识别低频的经济特征词并降低高频全局经济特征词对分组覆盖的不利影响；构建经济域情感词典、奇异评价对象词典，为经济类题裁文本情感挖掘研究打下基础；设计基于经济特征的细粒度文本情感计算模型，尤其是奇异经济特征与情感词结合模型、数字修饰动态程度模型，以解决目前商品评论情感分析中未遇到的新问题。因此，本项目研究具有重要的科学研究意义和应用推广价值。

Web上经济评论大数据的涌现，为获取当前经济形势、了解经济发展趋势提供了新途径。如何从Web经济评论中抽取经济特征词并构建非结构化经济指标体系、抽取满足细粒度文本分析所需的经济特征-情感词对，是经济评论中经济特征情感分析的关键任务。本项目研究的具体内容包括:(1)针对人工构建非结构化经济指标的局限性，以及主题模型在非结构化经济指标挖掘中存在的问题，我们定义了文档的领域隶属度、词语与主题的语义相关度和词语对主题的贡献度，提出了PSP_HDP (combining documents' domain Properties, word Semantics and words' Presences in topics with HDP)主题模型，改进了文档-主题与主题-词语的分配过程，从而提高了经济主题的区分度和辨识度，更有效地挖掘与经济有关的经济主题和经济要素词。(2)针对经济评论中经济特征构成复杂、数目众多，情感词词性丰富，虚指评价对象和隐式评价对象常见等问题，提出了基于浅层语义与语法分析相结合的经济特征-情感词对抽取方法；基于语义和领域知识的虚指评价对象判别和替换方法；基于特殊情感词搭配表、上下文搭配表及频繁搭配表的隐式评价对象识别方法。(3)针对评论中特征和情感词的低频词难以发现及细粒度的特征-情感词对到主题的分配问题，①提出了语义关系约束的主题模型SRC-LDA(Semantic Relation Constrained LDA)，该模型提高了相同主题下主题词分配的关联度和不同主题下主题词分配的区分度，可以更多地发现细粒度特征词、情感词及其之间的语义关联性。②提出了关联约束主题模型AC-LDA (Association Constrained LDA)，该模型通过改善全局特征词和局部特征词的主题区分度，改进了特征词和情感词的主题内聚度，从而提高了低频次级特征词、低频情感词和无特征情感词的提取效率；(4)其他扩展研究。包括中文隐式实体关系抽取、个性化投资方法推荐、合作元素抽取、虚假评论制造者的识别等。这些扩展研究也都是评论文本挖掘的热点问题。