Microblog Sentiment Analysis with Multi-Head Self-Attention Pooling and Multi-Granularity Feature Interaction Fusion

[Objective] This paper tries to efficiently and accurately extract sentiment information from Weibo texts and improve sentiment analysis performance. [Methods] First, we used WoBERT Plus and ALBERT to dynamically encode the character and word-level texts. Then, we extracted key local features with convolution operation. Next, we utilized cross-channel feature fusion and multi-head self-attention pooling operation to extract global semantic information and filter out critical data. Finally, we fused character-level and word-level semantic information using a multi-granularity feature interaction fusion operation and generated the classification results with the Softmax function. [Results] This model’s accuracy and F1 value were 98.51% and 98.53% on the weibo_senti_100k dataset and 80.11% and 75.62% on the SMP2020-EWECT dataset, respectively. Its performance was better than the advanced sentiment analysis models on each dataset. [Limitations] Our model does not include multimodal information such as video, image, and audio for sentiment classification. [Conclusions] The proposed model could effectively accomplish sentiment analysis of Weibo texts. © 2023 Data Analysis and Knowledge Discovery. All rights reserved.