A Study on the Application of Computer-Assisted Sentiment Analysis Methods of Literary Works in Literary and Cultural Education

Abstract

This paper addresses the need for emotional analysis of literary works by constructing a customized Chinese novel dataset and proposing a progressive learning framework called PL-ERC. This framework generates pseudo-sentence labels through dialogue-level annotation, combining noise density segmentation strategies with LSTM temporal modeling to achieve fine-grained emotional transfer learning. Cross-domain experiments show that the average F1 score reaches 88.87% in a 10-shot task across seven categories of literary works, outperforming the best baseline ConVEx by 2.9%. Ablation experiments reveal that removing the self-training strategy results in the largest F1 drop of 17.72%, while replacing LSTM with RNN reduces F1 by 12.88%. In terms of sentiment classification accuracy, the accuracy rate on the DMSC seven-classification dataset was 97.17%, with an identification rate of over 98% for the anger/happiness categories. Educational evidence shows that the weighted total score of the experimental class applying this method was 94.25 ± 3.92, significantly better than the control class (84.00 ± 7.89, p < 0.01). Students' critical thinking ability scores were 4.67 (control class: 3.51), and aesthetic experience ability scores were 4.72 (control class: 3.74), confirming that technology-enabled education significantly enhances the effectiveness of literary and cultural education.