舆情场景下基于层次知识的话题推荐方法

A Hierarchical Knowledge Based Topic Recommendation Method in Public Opinion Scenario

随着信息技术的飞速发展,互联网成为了舆情传播的主要载体.各种舆情事件不断涌现,并在网民的参与下广泛传播,由此可能引发强烈的社会反响.因此,如何实现网络舆情事件快速发现与个性化监测需求的精准推送,成为了当前舆情的重点关注内容.对于舆情场景下用户交互信息稀疏导致的兴趣难以刻画的问题,提出了一种基于层次知识的话题推荐模型.模型通过引入层次知识来扩充语义增加话题之间的潜在信息关联,分别对层次知识、话题和用户建模得到对应的嵌入向量表示,再结合多层感知机匹配模型预测用户点击率.实验结果表明,该模型在与多个基线算法的对比中,在F1(the balanced F score)和AUC(the area under curve)指标的平均值上分别提升了6.7%和4.9%.

With the rapid development of information technology,Internet has become the main carrier of public opinion spreading.All kinds of public opinion events come out one after the other,which can be quickly spread on different media in a short time and receive extensive attention and participation from large-scale Internet users.It may also trigger a strong reaction in the Internet space and the real society,and even induce large-scale mass incidents.Therefore,quick monitoring and effective early warning of online public opinion events become more and more important.In response to the growing scale of Internet information and the expanding scope of its dissemination,how to discover online public opinion events and push the precise personalized monitoring information has become the focus of current public opinion applications.Aiming at the problem that the interest is hard to capture in case of the sparsity of user interaction,a topic recommendation model based on HKN(hierarchical knowledge network)is proposed.The model expands the semantics and increases potential information association between topics by using hierarchical knowledge.It models the hierarchical knowledge,topics,and users to obtain corresponding embeddings.With those embeddings,we use a multi-layer perceptron matching model to predict the CTR(click through rate).Experimental results show that the HKN model outperforms multiple baseline algorithms by 6.7%and 4.9%on the average of F1(the balanced F score)and AUC(the area under curve)metrics CTR value respectively.