Leveraging NLP Techniques for Robust Emotion Recognition in Text
O. Bhaskaru, M. Vali, Khaja Vali Syed, Waseem Mohammad · 2025
This study introduces a novel emotion classification framework that harnesses the power of Bidirectional Encoders Representation from Transformer's (BERT)-based contextual embeddings in synergy with a Bidirectional Long Short Term Memory (BiLSTM) network. Traditional models often fall short in capturing the intricate semantic dependencies and emotional subtleties present in textual and emoji-based communication. To address these challenges, our approach leverages BERT's ability to encode deep contextual relationships and BiLSTM's strength in modelling sequential patterns, thereby enhancing emotion recognition accuracy. We assess our method on two well-established benchmark datasets: a large-scale text emotion recognition corpus comprising 40,000 sentences and an emoji-based emotion detection dataset with 108 samples. The textbased dataset enables the model to learn complex linguistic patterns and emotional cues embedded within written language, whereas the emoji-based dataset aids in decoding non-verbal emotional expressions conveyed through symbolic representations. By integrating these diverse modalities, our approach fosters a more comprehensive and adaptive emotion classification system. Empirical evaluations reveal that our model consistently outperforms conventional machine learning and deep learning baselines, achieving superior classification accuracy while significantly reducing misclassification rates. The combination of contextual embeddings and sequential learning offers a remarkable improvement in robustness, making emotion detection more precise and reliable. These results highlight the transformative potential of DL in advancing sentiment analysis and emotion-aware applications, paving the way for enhanced human computer interaction, affective computing, and psychological assessment tools.