A Deep Learning Emotion Classification Framework for Low Resource Languages

Abstract

Emotion classification from text is the process of identifying and classifying emotions expressed in textual data. Emotions can be feelings such as anger, joy, suspense, sadness and neutral. Developing a machine learning model to identify emotions in a low-resourced language with a limited set of linguistic resources and annotated corpora is a challenge. This research proposes a Deep Learning Emotion Classification Framework to identify and classify emotions in low-resourced languages such as Hindi. The proposed framework combines a classification model and a low resource optimization technique in a novel way. An annotated corpus of Hindi short stories consisting of 20,304 sentences is used to train the models for predicting five categories of emotions: anger, joy, suspense, sadness, and neutral talk. To resolve the class imbalance in the dataset SMOTE technique is applied. The optimal classification model is selected through experimentation that compares machine learning models and pre-trained models. Machine learning and deep learning models are SVM, Logistic Regression, Random Forest, CNN, BiLSTM, and CNN+BiLSTM. The pre-trained models, mBERT, IndicBERT, and a hybrid model, mBERT+BiLSTM. The models are evaluated based on macro average recall, macro average precision, and macro average F1 score. Results demonstrate that the hybrid model mBERT+BiLSTM out perform other models with a test accuracy of 57%.