Enhancing microblog sentiment analysis through multi-level feature interaction fusion with social relationship guidance

In sentiment analysis of microblogs that involve social relationships, a common approach is to expand the features of target microblogs using microblog relationship networks. However, the current research methodology only relies on individual interaction behaviors on social platforms to construct these networks, disregarding the guiding influence of microblog relationship networks on microblogs. Consequently, this leads to the feature expansion of microblogs while introducing interference among them. To address this problem, this study aims to construct a more precise microblog relationship network by incorporating multiple interactive behaviors from social platforms. This network will serve as a guide for sentiment interactions among microblog texts, thereby mitigating feature interference. Firstly, we utilize various interaction behaviors on social platforms to build a microblog relationship network. We employ a LINE network embedding to represent the microblog relationship network as microblog relationship features. Secondly, we extract word-level and sentence-level features from the microblog text using a BERT pre-training model. The word-level features are combined using convolutional neural networks. Subsequently, the word-level and sentence-level features of microblogs are separately guided for interaction fusion through relational features. An attention network is then constructed to fuse the post-interaction features in a single step. Prior to secondary fusion, the primary fusion features, post-interaction word-level features, and post-interaction sentence-level features are weighted, and the sentiment categories of microblogs are outputted. Finally, we compare the proposed method with the text-only microblog sentiment analysis approach and the sentiment analysis method that incorporates social relationships on two real datasets. The comparison results demonstrate the superiority of our proposed method.