基于多特征融合的微博主题情感挖掘

摘要

微博情感分析是社会媒体挖掘中的重要任务之一,在恐怖组织识别、个性化推荐、舆情分析等方面具有重要的理论和应用价值.但与传统文本数据不同,微博消息短小而凌乱,包含着大量诸如微博表情符号之类的特有信息,同时微博情感是与其讨论主题是密切相关的.多数现有的微博情感分析方法都没有将微博主题与微博情感进行协同分析,或者在微博主题情感分析过程中没有考虑将用户关系、用户性格情绪等特征数据,从而导致微博情感分析与主题检测的效果难尽人意.为此,提出了一个基于多特征融合的微博主题情感挖掘模型TSMMF(Topic Sentiment Model based on Multi-feature Fusion),该模型将情感表情符号与微博用户性格情绪特征纳入到图模型LDA中实现微博主题与情感的同步推导.实验结果表明,与当前用于短文本情感主题挖掘的最优模型(JST,SLDA与DPLDA)相比较,TSMMF具有更优的微博主题情感检测性能.%Sentiment analysis in microblogging is an important task in mining social media,and has important theoretical and application value in the terrorist organization identification,personalized recommendation,public opinion analysis,etc.However,different from traditional texts,messages in microblogging are short and irregular,and contain multifarious features such as emoticons,update time and etc,and in microblogging sentiment of a message is closely related to its topic.Most existing sentiment mining approaches cannot achieve cooperating analysis of topic and sentiment of messages in microblogging,or do not take factors such as social relations and users emotional personality into consideration,and this may lead to unsatisfactory sentiment classification and topic detection.To address the issues,a probabilistic model,TSMMF (Topic Sentiment Model based on Multi-feature Fusion) is proposed,which introduces emoticons and microbloggers personality into LDA inference framework,models emotion and personality of microbloggers under the guidance of emotional psychology theory,uses social relations among microbloggers to initialize topics of messages,utilizes Gibbs sampling techniques to estimate parameters in the model,and finally achieves synchronized detection of sentiment and topic in microblogging.Extensive experiments show that TSMMF outperforms state-of-the-art unsupervised approaches JST,SLDA and DPLDA significantly in terms of sentiment classification accuracy,and compared to the typical semi-supervised sentiment analysis algorithm SSA-ST,TSMMF performs as well as SSA-ST,but unlike SSA-ST,TSMMF can work without labeled training datasets.And so it has promising performance.