基于多传播信息流博弈的微博谣言自动检测方法研究

Microblog rumors have become one of the hottest social focuses recently. Thus, research on automatic Microblog rumors detection is regarded as the frontier technology of the public opinion analysis. This project applies the public opinion diffusion analysis, the game theory and the deep learning technology as the fundament, so as to study the nature of rumors from the falsity and the negativity factors. In the project, we firstly propose a multi-diffusion information flow based diffusion model which is composed of diffusion links, opinions and emotion. Then this model is used to forecast the diffusion path of topic, and the gain of communicators is then analyzed. Secondly, the deep neural networks based self-taught learning is used to learn the falsity and negativity semantic features from large-scale Microblog data. Finally, a social intelligent game method is studied to simulate the intelligent judging and diffusion process of social members towards the falsity and negativity of topic in multi-diffusion information flow model, and to implement the automatic detection of rumors. The research significances of this project are: 1) A novel approach to an automatic Microblog rumors detection system can be established by using social intelligent game from the perspective of multi-diffusion information flow. 2) The approach develops a feasible way for the research of internet rumors, and has an important academic significance in enriching the public opinion analysis methods and promoting the research of natural language processing technology and sentiment analysis technology. 3) The application of this project shows an important value in the purification of the internet public opinion space and the guidance of social public opinion.

微博谣言是当前社会关注的热点，面向微博的自动谣言检测也是舆情分析技术研究的前沿。本课题从虚假性和负面性两个方面来研究微博谣言，构建微博谣言自动检测系统。首先，本课题通过深度神经网络自学习机制从微博大数据中学习言论虚假性和负面性的语义特征；接着，根据学习到的语义特征，结合社会网络结构、用户模型构建微博言论传播的链接、观点和情感三个层次的多传播信息流模型，预测言论的传播路径，分析言论传播者的传播收益；最后，引入社交智能博弈的方法来模拟多传播信息流模型中社交个体对言论虚假性和负面性的智能判别与传播过程，实现对谣言的自动检测。本课题的研究意义在于从多传播信息流的角度，通过社交博弈来构造微博谣言自动检测系统，这为网络谣言研究开拓了一条可行的途径，对丰富网络舆情分析方法、促进自然语言处理技术和倾向性分析技术的研究都有着重要的学术意义，也对当前网络环境的净化，社会舆情的分析疏导具有非常现实的应用价值。

微博谣言是当前社会关注的热点，面向微博的自动谣言检测也是舆情分析技术研究的前沿。本课题从谣言传播博弈的角度来研究微博谣言，构建微博谣言自动检测系统。三年来主要开展了以下研究内容：首先，本课题提出了基于视觉倾向性话题模型的图像内容倾向性分析方法，该方法通过图像的相关话题内容从而自动挖掘视觉语义中与倾向性相关的视觉语义表示；其次，本课题提出了基于显示情感信号的无监督情感语义学习方法，该方法通过利用数据中与目标相关的无标注信息进行学习，从大量无标注数据中学习与目标空间相近的语义信息空间，并提取鲁棒的语义空间信息表示，最终利用少量标注达到与一定数据规模的有标注数据类似的分类效果。最后，本课题提出了责任均分的谣言传播博弈模型，该方法结合社会网络结构和用户模型构建微博言论传播的链接、观点和情感三个层次的多传播信息流模型，预测言论的传播路径，并通过社交智能博弈的方法来模拟多传播信息流模型中社交个体对言论是否属于谣言的智能判别与传播过程，分析言论传播者的传播收益，实现对谣言的自动检测。本课题的成功实施，不仅仅在语义层面上构建了谣言的负面性及传播博弈的分析方法，并大大改善谣言检测的准确性，同时将核心算法应用在用户倾向性分析和青少年网络信息过滤等方面并取得了良好的效果。