Novel Approaches to Social Emotion Analysis

Abstract

Emotion is a phenomenon that plays an important role in our daily communications. Social emotion refers to the emotion experienced by the reader exposed to a text as opposed to the emotion conveyed from the author's perspective. The work presented in this thesis introduces novel approaches to the field of social emotion analysis. More specifically, the aim is to propose approaches for social emotion prediction as well as to establish cause extraction for social emotion as an information extraction task. The main research question to be answered is “What are the most effective approaches and requirements for predicting social emotion in text and identifying their underlying causes?”. The primary contribution of the thesis is two main approaches for predicting the social emotion: the Comments Aggregation Approach, and the Comments Integration Approach. The Comments Aggregation Approach predicts the social emotion of the post by aggregating the analysed emotion in readers' comments. Two models were developed using this approach. The Lexicon-based model is a basic model that was developed based on an emotion lexicon to work as a proof-of-concept for the Comments Aggregation Approach. Moreover, the Transfer Learning-based model improves the Comments Aggregation Approach by utilising a pre-trained model for the writer's emotion classification. In addition, the Comments Integration Approach predicts social emotion by analysing the combined documents of posts and their corresponding comments; two models were also developed using this approach. The first is the Topic-based model, which employs the machinery of topic models, and the second is the Transformer-based model, which uses the transformer architecture for social emotion prediction. On the other hand, the thesis contributes to the Social Emotion Cause Extraction (SECE) by defining the task and providing the labelled data, as well as providing the evaluation approach and baselines.