Decoding Emotion in Ancient Poetry: Leveraging Generative Models for Classical Chinese Sentiment Analysis

Publication type
U
Publication status
In press
Authors
Du, Q, De Langhe, L., Lefever, E., & Hoste, V.
Series
Proceedings of the 15th International Conference on Recent Advances in Natural Language Processing
Publisher
NCOMA Ltd., Shoumen, Bulgaria (Varna, Bulgaria)
Conference
International Conference Recent Advances in Natural Language Processing (Varna, Bulgaria)
Download
(.pdf)
View in Biblio
(externe link)

Abstract

This study explores the use of generative language models for sentiment analysis of classical Chinese poetry, aiming to better understand emotional expression in literary texts. Using the FSPC dataset, we evaluate two models, Qwen-2.5 and LLaMA-3.1, under various prompting strategies. Initial experiments show that base models struggle with task-specific instructions. By applying different instruction tuning strategies with Low-Rank Adaptation (LoRA), we significantly enhance the models’ ability to follow task instructions and capture poetic sentiment, with LLaMA-3.1 achieving the best results (67.10% accuracy, 65.42% macro F1), demonstrate competitive performance against data-intensive, domain-adapted baselines. We further examine the effects of prompt language and multi-task learning, finding that English prompts outperform Chinese ones. These results highlight the promise of instruction-tuned generative models in sentiment analysis of classical Chinese poetry, and underscore the importance of prompt formulation in literary understanding tasks.