Multilabel Emotion Recognition Through Sequence Labeling and Sentence Classification Models using Textual Data

Humans possess limited primary emotions,impacting decision-making, relationship-building, and interactions; understanding these emotions from written text challenges natural language processing. Textual communication lacks visual cues, making it difficult to detect subtle emotions. Multilabel emotion recognition helps machines recognize emotions, especially in text, providing nuanced understanding and insights into human behavior. Research on multilabel emotion detection focuses on sequence labeling and sentence classification. This study capitalizes on this by comparing the two approaches. Bidirectional Long Short-Term Memory (BiLSTM) and Robustly Optimized Bidirectional EncoderRepresentations Pretraining Approach (RoBERTa) models were developed for each approach, respectively. Training and testing are conducted on the Multi-modal Multi-Label Emotion, Intensity, and Sentiment Dialogue Dataset (MEISD), with text labeled with corresponding emotions. Primary emotions recognized by the models are mapped to complex emotions by combining them using Robert Plutchik’s Emotion Wheel. The models are evaluated for their performance on multilabel emotion recognition using the metrics of Precision, Recall, Accuracy, and F1-score. Results obtained have been promising with 93% average accuracy for both models. The distinction between the models is apparent in their confusion matrix with varying outcomes for different emotions. Both offer reliable and competitive performance for sentiment classification tasks.