基于双向分层语义模型的多源新闻评论情绪预测

Multi-Source Emotion Tagging for Online News Comments Using Bi-Directional Hierarchical Semantic Representation Model

随着在线新闻服务的迅猛发展,用户在阅读新闻后可以非常方便地表达自己的主观情绪,有效分析和预测用户的情绪有助于新闻服务提供商为新闻用户提供更好的服务.情绪标注研究已经取得了很多优秀的成果,但仍然存在着一些问题:1)传统的方法将整个文档看作单词流或词袋,不能对句子间的逻辑关系进行建模,在文档中的句子间包含逻辑关系时,这些方法无法适当地表达文档的语义;2)这些方法只用了文档本身的语义,忽略了与该文档相关的其他信息源中信息,而这些信息源对该文档的语义表达也有一定的影响.为了解决这些问题,提出了一种基于多信息源的在线新闻评论双向分层语义表示模型,称为双向分层语义神经网络(bi-directional hierarchical semantic neural network,Bi-HSNN),该模型既捕获句子中词语所表达的情感,又自底向上地学习文档中句子间的逻辑关系,并利用评论、新闻和用户投票等多种信息源对在线新闻评论的情绪进行标注.在真实数据集上的一系列实验,验证了该模型的有效性.

With the rapid growth of news services,users can now actively respond to online news by expressing subjective emotions.Such emotions can help understand the preferences and perspectives of individual users,and thus may facilitate online publishers to provide users with more relevant services.Research on emotion tagging has obtained promising achievement,but there are still some problems：Firstly,traditional methods regard a document as a flow or bag of words,which cannot extract the logical relationship features among sentences appropriately.Therefore,these methods cannot express the semantic of the document properly when there exists logical relationship among the sentences in the document.Secondly,these methods use only the semantic of the document itself,while ignoring the accompanying information sources, which can significantly influence the interpretation of the sentiment contained in documents.In order to solve these problems,this paper proposes a hierarchical semantic representation model of news comments using multiple information sources,called bi-directional hierarchical semantic neural network（Bi-HSNN）,which not only captures the sentiment among words in sentences,but also applies a bottom-up way to learn the logical relationship among sentences in the document.This paper tackles the task of emotion tagging on comments of online news by exploiting multiple information sources including comments,news articles,and the user-generated emotion votes.A series of experiments on real-world datasets have demonstrated the effectiveness of the proposed model.