A Novel Microblog Sentiment Classification Method Based on Top-k Pooling

Microblog is a popular social media platform for information sharing and dissemination. Recently, microblog sentiment classification has received a lot of attention in many real-world applications, such as stock price prediction, public opinion monitoring, and crisis management. Existing microblog sentiment classification methods identify the sentiment polarity mainly based on the textual content. As microblog is usually short and contains a lot of noisy, it is very challenge to learn powerful representation only relied on the textual information. In this paper, we argue that the multi-head self-attention, which can capture complicated interactions information between words in sequence, would inevitably to introduce lots of noisy information. Moreover, the user sentimental polarity also plays an important role in judging the sentiment of text. Therefore, we propose a novel microblog sentiment classification method, called Multi-head Self-attention Based on Top-k pooling (MSBT). Specifically, we design a Top-k pooling layer to alleviate the issue caused by the multi-head self-attention network, and then incorporate user historical sentiment tendency in the loop of microblog sentiment classification. Extensive experiments conducted on a real-word dataset MDUHI demonstrates that our proposed approach MSBT can significantly improve the performance of microblog sentiment classification, e.g., the F1 score is 0.98% higher than the optimal benchmark method NPA, reaching 93.39%.