基于情感序列和情感选择的句子多维复合情感识别研究

Sentence emotion recognition is the basis of textual emotion analysis and understanding. According to the characteristic of compound emotion expression in text, this study proposes to recognize sentence emotion from a view of multi-dimensional compound emotions. For the sake of breaking through the problems that the words bear emotion ambiguity and compound emotions are hard to be recognized when depending on emotion lexicons, we first detect and recognize multi-dimensional compound emotions for emotional words in a sentence. By coding of multi-dimensional compound emotions, the links between the basic emotions are established, and the multi-label classification problem is transformed into a single label classification problem, improving the efficiency of the algorithm. On this basis, emotion sequence composed by key emotion elements is proposed to express the emotions of each clause in a sentence, and a sequence-based model combining with word positions is built for clause multi-dimensional compound emotion recognition.In this model, the data sparse problem will be solved by decreasing the number of emotion sequences and by designing the classification algorithm based on subsequence similarity. A strategy based on emotion selection is proposed for the sentences with multiple clauses: the problem of weighting clause features is considered as an optimization problem solving, and genetic algorithm will be used to solve this problem. This research will lay a theory and technology foundation for secondary prevention of mental disorders,treatment, and promotion of human-computer interaction.

句子情感识别是文本情感分析和理解的基础，针对文本情感表达具有复合性的特点，本研究提出从多维复合情感的角度研究句子情感识别问题。为突破依赖情感词典而无法识别具有情感歧义和复合情感词的问题，首先对句子中的情感词进行检测和多维复合情感识别。通过对多维情感进行编码建立基本情感之间的联系，将这一多标签分类问题转化为单标签分类问题，提高算法效率。在此基础上，利用由情感关键元素组成的情感序列来表达句中各分句的情感，并结合词语的位置信息建立序列分类模型识别分句多维复合情感。同时，通过减少情感序列的数量和设计基于子序列相似的分类算法解决该模型可能产生的数据稀疏问题。对具有多个分句的复杂句，提出基于情感选择的策略：将对分句的特征权重计算问题理解为最优化求解问题，并利用遗传算法作为该最优化问题的搜索算法。对这些关键技术的研究将为辅助心理疾病的预防、治疗，和促进人机交互打下理论和技术基础。

面向语言的情感语义理解对提高多模态情感识别的正确率，增强人机交互起着重要作用，句子情感识别是文本情感理解的核心基础问题之一。本项目主要研究句子多维复合情感识别问题，具体包括：（1）句子中的关键情感元素分析及其在句子中表达的多维复合情感识别。（2）基于情感模式的分句多维复合情感识别。（3）分句情感选择。. 本项目取得的重要研究结果主要包括：（1）针对句子情感表达特点，本项目提出基于情感模式和情感选择的句子多维复合情感识别方法。为突破依赖情感词典带来的问题，首先对句子中的情感元素进行多维复合情感识别，通过对多维复合情感进行编码建立基本情感之间的联系，从而将这一多标签分类问题(Multi-Label Classification)转化为单标签分类问题(Single-Label Classification)，提高算法效率。（2）提出利用情感模式来表达句中各分句的情感，并采用基于序列的模型识别分句多维复合情感。同时，通过减少情感模式的数量和设计基于子序列相似的分类算法解决该模型可能产生的数据稀疏问题。（3）对具有多个分句的句子，提出基于情感选择的策略：对分句的特征权重计算问题理解为最优化求解问题，并利用遗传算法GA(genetic algorithm)对该问题求解。. 项目按照原订计划取得了预期的研究成果，项目组成员在相关期刊及会议上共发表24篇论文，其中包括SCI收录期刊论文7篇，EI收录期刊论文8篇，国际会议优秀论文2篇。同时，该项目的相关研究成果获得“中国人工智能学会吴文俊人工智能科学技术创新奖一等奖”。