Application of BiLSTM CRF Model Based on Hierarchical Attention for Implicit Emotion Recognition in Literary Texts

Abstract

In recent years, emotion recognition in textual narratives has become an important area of affective computing and natural language understanding. Unlike explicit emotional expressions, implicit emotions are concealed within linguistic context, narrative tone, and metaphorical cues, posing significant challenges for traditional machine learning approaches. To address this issue, this study proposes a Hierarchical Attention-based BiLSTM-CRF model for Implicit Emotion Recognition in literary texts. The model integrates a hierarchical attention mechanism to capture both sentence-level and document-level semantic dependencies, while the BiLSTM-CRF architecture effectively encodes sequential context and label inter-dependencies. The proposed model was evaluated on a multi-class emotion dataset containing seven emotion categories such as Joy, Sadness, Anger, Fear, Disgust, Surprise, and Neutral. Experimental results demonstrate that the proposed model yields superior performance compared with conventional deep learning baselines, including HDEL, BERT-BiLSTM-CRF, GT-BiLSTM, and Hierarchical Attention Networks. Specifically, the model achieves an overall Accuracy of 0.92, Precision of 0.91, Recall of 0.90, F1-Score of 0.90, and ROC-AUC of 0.95, indicating enhanced ability to identify subtle emotional implications within complex literary expressions. These findings highlight the effectiveness of combining hierarchical attention with sequence labeling for advanced emotion cognition and provide a valuable framework for future studies on literary emotion analysis and intelligent human–computer interaction.