社会网络中非参数化的事件演化分析

Analysis of Non-Parametric Event Evolution in Social Networks

社会网络,如微博和Twitter,已经成为数十亿人关注事件的重要平台。人们不仅关注所发生的事情,更关注事件的演化。因此,监控社交网络中事件的发展是至关重要的。在新闻文章和短文本(社交网络中的文本)中有一些挖掘事件演化的研究。由于短文的形式比新闻文章短,适用于新闻文章的方法不能直接应用于社交网络。一些应用于短文本的方法不考虑语义信息,有些方法无法发现长期跨度事件的演化,特别是中间有间断的事件。鉴于此,我们提出了一种非参数的方法来发现事件演化(故事情节)。首先使用贝叶斯模型测量短文本的语义相关性。其次,使用基于嵌入表示的算法来生成长期和短期事件的故事线。我们进一步使用Dirichlet过程自动学习适当数量的主题。与其他方法相比,三个人工标记数据集的详细实验结果证明了我们方法的有效性。

Social networks, such as Weibo and Twitter, have become important platforms where billions of individuals follow events. People not only concern about what happened, but also pay more attention to how the event gradually progresses. Therefore, it is crucial to monitor the development of events in social networks. There are some research mining event evolutions in news articles and short texts (texts in social networks). Methods designed for news articles cannot be directly applied to social network due to the fact that short texts are more ill-formed and shorter than news articles. Some methods that apply to short texts do not take into account semantic information, and some cannot discover evolution of long-term spanning events, especially intermittent events. In light of this, we propose a non-parametric method to discover event evolution (storylines). Firstly, a bayesian model is used to measure semantic correlation of short texts. Secondly, an embedded representation-based algorithm is used to generate storylines for long-term and short-term events. We further use dirichlet process to automatically learn an appropriate number of topics. In comparison with other methods, detailed experimental results on three manually labeled data sets demonstrate the effectiveness of our method.