Affect level opinion mining of Twitter streams

Abstract

Twitter is a social media platform which is used by millions of users to express their opinions freely. However, it is almost impossible to analyze the opinion manually due to the sheer number of Tweets generated per day. Therefore, automated analysis of emotions in Tweets, which is also known as affect level opinion mining in the literature is crucial. Emotion analysis in this study is performed at two levels: Emotion Category Classification and Emotion Intensity Prediction. One key challenge in identifying emotion categories is the presence of implicit emotions. This study introduces a model that enables reuse of the same deep neural network architecture with different word embeddings for the extraction of different features related to implicit emotion classification. We presented this model at 9𝑡ℎ Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis (WASSA-2018). Our system was ranked among the top ten systems (8𝑡ℎ) amidst constrained corpus usage. Our implicit emotion classifier outperformed the baseline system by more than 8%, achieving a 68.1% macro F1-Score. We solved the emotion intensity task with transfer learning techniques. Among the models used in transferring features were a sentiment classifier, emotion classifier, emoji classifier and emotion intensity predictor. Our transfer learning based intensity predictor outperformed existing best in two out of four emotions. We were able to achieve an average Pearson score of 79.81%. Additionally, we propose a technique to visualize the importance of each word in a tweet to get a better understanding of the model. Finally, we developed a web-platform that utilizes our emotion analysis models to summarize and view the opinion of a group of tweets.