Multi-label Emotion Classification on Social Media Comments using Deep learning

Abstract Social media is an online platform that people use to develop social networks or relationships with others. Every day, millions of people use different social media to express their thoughts, emotions, and experiences. An emotion is a complex psychological event that involves a mixture of reactions occurring in the human body and brain, usually triggered by a mental content(Almeida et al., 2018). Multi-label text emotion classification is the problem that aims to identify all possible emotions from a given text that best represents the author's mental state. Many researches have been done on text emotion classification in English, Arabic, and Chinese language. However, most of them focus on single-label emotion classification which is unable to identify all present emotions in the given instance. To the best of our knowledge, there is no research conducted on multi-label emotion classification for Amharic text. In addition to this, there is no available dataset to conduct multi-label emotion classification research. These reasons motivate us to do research on multi-label Amharic text emotion classification. for this research we collected 18000 datasets from different social platforms YouTube, Facebook, and Twitter. The dataset is annotated by psychologists and other professionals. We use word2vec and one hot encoding to prepare the feature vector. We train and test four deep-learning approaches such as LSTM, BILSTM, CNN, and GRU. We perform the experiment by feeding one hot encoding and word2vec features to these for deep learning models and achieve the best accuracy with one hot vector. We achieve test accuracy of 53.1%, 54.5%, 54%, and 39.7% for LSTM, BILSTM, CNN, and GRU respectively. For the future we conduct this research using a large dataset with transformer models (BRT, ROBERTA, and XLNET) and test the performance of these models on Amharic text multi-label emotion classification.