基于多粒度计算和多准则融合的情感分类  

Sentiment classification based on multigranularity computing and multi-criteria fusion

作　　者：王丙坤[1] 黄永峰[1] 李星[1] 

机构地区：[1]清华大学电子工程系,北京100084

出　　处：《清华大学学报（自然科学版）》2015年第5期497-502,共6页Journal of Tsinghua University(Science and Technology)

基　　金：国家"八六三"高技术项目(2012AA011004);清华大学自主科研计划项目(20111081023)

摘　　要：随着在线用户生成内容的激增,无监督情感分类方法有着广泛应用前景。现有基于情感词的无监督情感分类方法没有考虑句子类型和句间关系对情感分类的影响,分类效果较差;基于自学习的无监督情感分类方法在生成伪标注数据集时,又会引入较多错误。针对上述问题,该文提出了一种基于多粒度计算和多准则融合的无监督情感分类方法。该方法通过多粒度计算,提高现有基于情感词的无监督情感分类精度;同时通过多准则融合来减少伪标注数据错误率。在3个真实中文数据集上的实验结果表明:与现有无监督情感分类方法相比,该方法平均提高了6.5%的分类精度。The large amount of online user-generated content on the Web has created a need for unsupervised sentiment classification methods.Unsupervised sentiment classification methods based on sentiment words do not work well because the complex sentence structures and sentence types are seldom taken into account.Unsupervised sentiment classification methods based on self-learning have many errors when generating pseudo-labelled datasets.These limitations are reduced by the current method based on multi-granularity computing and multi-criteria fusion. The multi-granularity computing improves the accuracy of unsupervised sentiment classification methods based on sentiment words.The multi-criteria fusion reduces the number of errors in the pseudo-labelled data from the self-learning.Tests using three real Chinese review datasets show that the classification accuracy is 6.5% more accurate on average than with existing unsupervised sentiment classification methods.

关 键 词：情感分类 无监督方法 多粒度计算 多准则融合 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]