社交网络中热点事件的负面情绪识别及演化分析研究

Negative emotions of people continue to accumulate in their real life, making the phenomenon of "negative labeling" increasingly serious in social networks. In the process of hot events' formation, fermentation and dissemination, "labeling thinking" spreads negative emotions and distorts social cognition. However, the existing research is difficult to effectively depict and analyze the negative emotions in hot events due to their diversity, randomness, mutual infectivity, and the fact that their mode of transmission is latent, oscillatory, dynamic and deviant. In view of this problem, this project aims to study how to automatically identify negative emotions and to predict their evolution in hot events. The main contents are as follows: (1) In view of the diversity and randomness of negative emotions, the deep reinforcement learning method is used to automatically identify them. (2) In view of the latent, circulatory and dynamic characteristics of negative emotions' propagation, we introduce multi-dimensions characteristics and use the two-order hidden semi-Markov model to study the evolution mechanism. (3) In view of the dynamism and deviation of negative emotions, we use k-means and parallel algorithm to improve the spectral clustering algorithm; clustering analysis is used to depict the evolution trend; and regression equation is used to test the results of clustering analysis. This project is expected to provide a new theoretical basis for the research on the method of automatic recognition, evolution analysis, and evolution trend prediction of negative emotions.

人们在现实生活中的负面情绪不断累积，使得社交网络中“负面标签化”现象日益严重，在热点事件的形成、发酵和传播过程中，“标签化思维”会扩散负面情绪从而扭曲社会认知。然而现有的研究难以对热点事件中的负面情绪进行有效的刻画和分析，因其具有多样性、随机性、相互感染性及其传播方式的潜伏性、循环震荡性、动态性和偏差性等。针对该问题，本项目拟研究如何自动识别热点事件中的负面情绪并预测其演化趋势。主要包括：①针对负面情绪的多样性和随机性，采用深度强化学习方法对其进行自动识别。②针对负面情绪传播的潜伏性、循环震荡性、动态性，融合多维特征并采用二阶隐半马尔可夫模型研究其演化机理。③针对负面情绪的动态性、偏差性，采用k-means、并行算法来改进谱聚类算法，对其演化趋势进行聚类分析，并对聚类分析的结果采用回归方程进行分析。本项目有望在理论上为负面情绪自动识别方法、演化分析方法及演化趋势预测方法提供新的理论依据。

随着电子商务和社交网络的普及，互联网上涌现了大量的各种评论数据，其中很大一部分都蕴含负面情绪。如果能对这些数据进行有效处理与分析，可以很好地了解网民的观点和立场，有助于掌握网络舆情的关键信息。本项目已对与负面情绪识别相关的内容进行了系统地研究，主要包括文本负面情绪识别、对话文本负面情绪分类、方面级情感分类、跨领域的情感分类、跨领域的情绪分类、视觉域自适应等。重点是融合深度学习和宽度学习对负面情绪识别和情绪分类进行研究，主要包括文本的负面情绪识别、对话文本的负面情绪分类、文本的情绪分类。根据实际应用环境，设计并实现相应的原型软件系统，并利用该系统来验证所提出的算法及模型的有效性。本项目研究成果申请专利3项，登记软件著作权3项。发表重要国际期刊论文9篇、重要国内学术会议论文1篇、重要国际学术会议论文5篇。培养青年教师4人。参加有影响的国内外学术会议20多次，累计邀请5位国内外专家举办学术交流报告5次。该项目的研究为热点事件中负面情绪识别提供一种有效的解决方案，也有效地促进了相关方向的研究和发展。