A Citation Text Sentiment Recognition Method Based on Dynamic Weight Decay Mechanism

Abstract

This paper systematically investigates deep learning-based text sentiment recognition methods and their application in academic citation analysis. To address the limitations of traditional sentiment dictionaries, which rely on static weight allocation and exhibit insufficient domain adaptability, a dynamic weight decay mechanism is proposed. Building on the SO-PMI algorithm for constructing domain-specific dictionaries, this method incorporates decay factors based on word frequency distribution, sentiment concentration, and contextual position. These factors dynamically adjust sentiment word weights, effectively suppressing interference from high-frequency general terms while enhancing the representation of low-frequency, strongly affective words. Experimental results on citation sentiment classification tasks demonstrate that the proposed method significantly improves the recognition accuracy of neutral and weak sentiment citations. This approach offers a novel pathway for constructing more adaptive and accurate sentiment analysis systems in academic and other specialized domains. Furthermore, the study highlights the potential of integrating domain-specific knowledge and contextual features to refine sentiment analysis models, paving the way for future research on cross-domain and resource-efficient sentiment recognition techniques.