基于多层关系图模型的中文评价对象与评价词抽取方法

中文评价对象与评价词抽取是文本倾向性分析的重要问题.如何利用评价对象与评价词之间的语法、共现等关系设计模型是提高抽取精度的关键.本文提出了一种基于多层关系图模型的中文评价对象与评价词抽取方法.该方法首先利用词对齐模型抽取评价对象与评价词搭配；然后，考虑评价对象与评价词的依存句法关系、评价对象内部的共现关系和评价词内部的共现关系，建立多层情感关系图，接着利用随机游走方法计算候选评价对象与评价词的置信度；最后，选取置信度高的候选评价对象与评价词作为输出.实验结果表明，与现有的方法相比，本文所提出的方法不仅对评价对象和评价词的抽取精度均有显著提升，而且具有良好的鲁棒性.

A Multi-layer Relation Graph Model for Extracting Opinion Targets and Opinion Words

Mining opinion targets and opinion words is a fundamental task for the Chinese online media to mine opinion and analyze sentiment. The key to enhancing the effectiveness of opinion target and opinion word is to integrate syntactic relations and co-occurrence relations between opinion target and opinion word into the mining model. A novel approach based on a multi-layer relation graph model is proposed to extract opinion targets and opinion words from Chinese social media. First, the word alignment model is employed to extract the candidates of opinion target and opinion word candidates. Second, a multi-layer relation graph is constructed by the syntactic inter-relations between opinion target and opinion word, the co-occurrence intra-relations among opinion targets, and the co-occurrence intra-relations among opinion words. Third, a random-walk algorithm is adopted to calculate the confidence of each opinion target candidate and opinion word candidate. Finally, opinion targets and opinion words are extracted according to their confidence values. Experimental results show that the presented method can not only achieve significant improvement over previous methods, but also have good robustness.