基于词加权LDA算法的无监督情感分类

主题情感混合模型可以有效地提取语料的主题信息和情感倾向。本文针对现有主题/情感分析方法主题间区分度较低的问题提出了一种词加权LDA算法（weighted latent dirichlet allocation algorithm，WLDA），该算法可以实现无监督的主题提取和情感分析。通过计算语料中词汇与情感种子词的距离，在吉布斯采样中对不同词汇赋予不同权重，利用每个主题下的关键词判断主题的情感倾向，进而得到每篇文档的情感分布。这种方法增强了具有情感倾向的词汇在采样过程中的影响，从而改善了主题间的区分性。实验表明，与JST（Joint Sentiment/Topic model）模型相比，WLDA不仅在采样中迭代速度快，也能够更好地实现主题提取和情感分类。
The topic and sentiment unification model can efficiently detect topics and emotions for a given corpus. Faced with the low discriminability of topics in sentiment/topic analysis methods, this paper proposes a novel method, the weighted latent dirichlet allocation algorithm (WLDA), which can acquire sentiments and topics without supervision. The model assigns weights to terms during Gibbs sampling by calculating the distance between seed words and terms, then counts the weights of key words to estimate the sentiment orientation of each topic and obtain the emotional distribution throughout documents. This method enhances the impact of words that convey emotional attitudes and obtains more discriminative topics as a consequence. The experiments show that WLDA, compared with the joint sentiment/topic model (JST), not only has a higher iteration sampling speed, but also gives better results for topic extraction and sentiment classification