Mixed emotion extraction analysis and visualisation of social media text
Yuming Li, Johnny Chan, Gabrielle Peko, David Sundaram
Abstract
With the widespread use of social media and accelerated development of artificial intelligence, sentiment analysis is regarded as an important way to help enterprises understand user needs and conduct brand monitoring. It can also assist businesses in making data-driven decisions about product development, marketing strategies, and customer service. However, as social media information continues to grow exponentially, and industry demands increase, sentiment analysis should no longer be limited to fundamental polarity classification of positive, neutral, and negative. Instead, it should move to more precise classification of emotions. Therefore, in this paper, we expand sentiment analysis to analysis of eight different emotions based on Plutchik's wheel of emotions, and define it as a multi-label classification task to identify complex and mixed emotions in text. We achieved an overall precision of 0.7958 for the eight emotions multi-label classification based on the attention-based bidirectional long short-term memory with convolution layer (AC-BiLSTM) model on the SemEval-2018 dataset. In addition, we proposed the introduction of the NRC emotion lexicon and emotion correlation constraints to optimise the emotion classification results. This ultimately increased the overall precision to 0.8228 demonstrating the effectiveness of our approach. Finally, we store and visualise the emotion analysis results in a graph structure, in order to achieve deductibility and traceability of emotions.