Topic-Level Sentiment Analysis in Social Networks with Pair-wise User Influence

摘要

Inspired by the principle of Homophily which suggests that opinions are influenced by connection, we introduce relations into sentiment analysis in the context of social networks, which also helps to reduce the content sparsity by utilizing the networked SNS data. We propose a model which utilized textual content and link structure simultaneously to evaluate pair-wise social influence on topic level between users. The framework depicts the topic distribution for each user by LDA based on text information; and model the pair-wise influence between users on topic level by measuring their centralities and interactive weights. The learned influence is then applied into sentiment classification as supplementary features. The experiment results on two datasets show that the model incorporating user relations outperforms the methods which based on textual features only.