Abstract
Understanding user emotions to identify user opinion, sentiment, stance, and preferences has become a hot topic of research in the last few years. Many studies and datasets are designed for user emotion analysis including news websites, blogs, and user tweets. However, there is little exploration of political emotions in the Indian context for multi-label emotion detection. This paper presents a PoliEMO dataset—a novel benchmark corpus of political tweets in a multi-label setup for Indian elections, consisting of over 3512 tweets manually annotated. In this work, 6792 labels were generated for six emotion categories: anger, insult, joy, neutral, sadness, and shameful. Next, PoliEMO dataset is used to understand emotions in a multi-label context using state-of-the-art machine learning algorithms with multi-label classifier (binary relevance (BR), label powerset (LP), classifier chain (CC), and multi-label k-nearest neighbors (MkNN)) and deep learning models like convolutional neural network (CNN), long short-term memory (LSTM), bidirectional long short-term memory (Bi-LSTM), and transfer learning model, i.e., bidirectional encoder representations from transformers (BERT). Experiments and results show Bi-LSTM performs better with micro-averaged F1 score of 0.81, macro-averaged F1 score of 0.78, and accuracy 0.68 as compared to state-of-the-art approaches.