一种面向观点挖掘的多粒度话题情感联合模型

Multi-grain joint model of topic and sentiment for opinion mining

为了提高文本观点挖掘的效率,通过扩展标准话题模型,提出了一种新颖的多粒度话题情感联合模型(MG-TSJ).模型将文本话题区分为全局和局部两类,同时挖掘文本中涉及的多层次话题信息和情感倾向信息.该模型采用非监督的学习方法,解决了现有方法存在的领域依赖问题.通过在测试语料库上进行实验,该模型在文本情感倾向性分类任务中的准确率达到82.6%,具有和监督分类系统相当的性能;挖掘话题集合呈现层次化、语义相关的特点,证明了MG-TSJ模型对观点挖掘是可行的和有效的.

Based on extensions to standard topic modeling methods, a novel multi-grain joint model of topic and sentiment is proposed to improve efficiency of opinion mining. This model extracts sentiment and hierarchy topic from the text simultaneously, which distinguishes between local topics and global topics. The proposed model adopts the unsupervised learning method to address the issue of being domain dependent in existing methods. According to experiments, this model achieves an accuracy of 82. 6% for sentiment classification. It has a performance comparable to that of supervised sentiment classification methods. Moreover, the acquired collection of topics is hierarchy and semantic related. It is proved that the proposed model is feasible and effective for opinion mining.