A Method for Classifying Ancient Poetry Emotions Based on Pre-trained Models and Attention Mechanisms

Abstract

Chinese classical poetry is a treasure of Chinese culture with rich and subtle emotional connotations. Studying the realization of automated emotional analysis of ancient poetry is of great significance for fields such as digital humanities and literary studies. This paper proposes an emotion classification model based on a pre-trained model and combined with the attention mechanism in response to the deficiencies in emotion classification accuracy and multi-label emotion classification in existing emotion classification models of ancient poetry and prose. This model is based on the RoBERTa-wwm-ext-large pre-trained model. An attention mechanism is introduced for the multi-label classification task to capture the emotional keywords in the verses, a weighted binary cross-entropy loss function was adopted to address the severe category imbalance problem in multi-label tasks. Experimental results show that the method achieves an accuracy of 90.80% for the single-label five-classification task on the THU fine-grained sentimental poetry corpus (FSPC) dataset, surpassing the CDBERT model's result of 51.40. On the Chinese Classical Poetry dataset (CCPD), the multi-label 12-class tasks Micro-F1 and Macro-F1 achieved 99.40% and 99.38% respectively, outperforming the CDBERT model by 89.95% and 74.90% respectively. This study validates the effectiveness of pre-trained models combined with attention mechanisms in sentiment analysis of classical poetry and provides new ideas for sentiment analysis research in classical literature.