Topic-Specific Political Stance Inference in Social Networks With Case Studies

Abstract

Topic-specific political stance inference in social networks (SNs) aims at inferring target users' attitudes toward different target topics. Traditional methods mainly used a language model to classify sentiments from the postings of the SN users. However, people's stances are not always equal to their sentiments. Some others tried to build separate models toward different target topics. In many cases, though SN users talked about the target topics, the information given was limited; or they only expressed attitudes toward some other issues except the target topics. When information is incomplete, the methods that treat the topics independently fail to work, let alone for users who didn't post any of the topics. To solve the above problems, we introduced a political knowledge graph (PKG) to supplement side information for users and topics and proposed a united Knowledge Graph-aware and Social Network-enhanced framework (KGSN) to capture not only the knowledge connections between topics but also the social connections between users. KGSN utilized two levels of graph convolutional networks, the one at the knowledge graph level generating knowledge-aware representations merging knowledge entities for the users and topics respectively, and the one at the social graph level generating social-enhanced representations merging social neighbors for the target users. Beyond that, the respective topic-specific attention mechanisms were leveraged to emphasize special knowledge entities in the knowledge graph and special neighboring users in the social graph. The advantages of KGSN are that: first, it can infer users' attitudes toward more than one topic in one model; second, it can infer users' implicit attitudes toward the target topics through users' explicit attitudes toward the other issues; last but not least, even for users without any postings, KGSN can infer users' implicit attitudes through their social neighbors. Finally, extensive experiments were conducted to demonstrate the superiority of KGSN over state-of-the-art models and case studies were investigated to testify the effectiveness of the model.