Emotional Polarity Attention Mechanism for Text Sentiment Analysis

Text Sentiment Analysis (TSA) is a classic research topic in the field of Natural Language Processing (NLP) and has important application value for downstream tasks. Previous TSA methods focus on accurately representing emotional words in the semantic space to complete the recognition of text emotions. However, traditional methods are difficult to establish the intrinsic strong connections between emotional words in fine granularity and to accurately distinguish emotional tendencies. The research work of psychologists such as Carroll Ellis Izard, David Krech, and Robert Pluchik pointed out that for humans, our emotions can be compounded from the most basic emotions to construct many types of emotional states. Based on the classic attention mechanism, this paper proposes an Emotional Polarity Attention Mechanism (EPAM), which establishes strong connections between emotional words through emotional compounding theory, to enhance the understanding and representation accuracy of emotional words. We embed EPAM as a network layer into the current mainstream text classification models to build experimental group models and then verify the effectiveness of EPAM. Comparative experiments on four standard text classification datasets confirm that our model is effective and able to mine the intrinsic connections of emotional words.