中文摘要：

相比于句子级和文档级情感分析，词语级的情感分析呈现出领域性和上下文相关性，难以得到良好的应用．提出一种基于概率图模型的情感分析方法，先通过分析训练语料建立一种具有先验概率的图模型，用于计算语料中词语的情感概率值，再利用信息熵公式将概率值归一化为情感特征值，最后使用该特征值训练SVM分类器对测试语料进行分类．论文在理论上证明该方法能够有效运用于具有“评价对象”和“评价词”二元特性的商品评论的情感判定，且在未表明用户明显态度语句的极性判定中也可以获得良好的效果．实验结果也显示，该方法比传统SVM分类方法在准确率上有明显提高．

英文摘要：

Compared to the sentiment analysis of sentence and document level,analysis of word level can＇t be effectively applied as thefield limited and contextual relevance. The paper proposed a sentiment analysis algorithm using probabilistic graphical models. Firstly,the method analyzed the training corpus to establish a prior probability graph model, which can calculate the sentiment probability val-ue of the words which are extracted from the corpus. Then it normalized the value into the sentiment features with the information en-tropy formula. Finally it trained SVM classifier with the sentiment features to classify the test corpus. In theory, the method can beef-fectively applied in the sentiment analysis of product review with both ＂evaluation objects＂ and ＂evaluation＂, as well as with the im-plicit review. Experiment results showed that this method achieves higher accuracy than the traditional SVM classification.