Answers are grounded in this paper and cite their sources.
Citation sentiment analysis (CSA) faces several challenges, including implicit sentiment expression, class imbalance, and the unstable availability of fine-grained structured evidence. Existing methods use structured quadruples to make evaluative information in citation contexts more explicit. However, LLM-generated outputs may still suffer from invalid structures, limited textual alignment, and varying semantic quality, especially for neutral citations with weak evaluative signals. To address these issues, we propose QCSA, an LLM-driven framework for progressive quadruple construction and prompt-augmented learning. For neutral citations with weak structured evidence, QCSA adopts a conditional three-stage construction strategy. It prioritizes valid semantic evidence while improving the completeness of structured prompt inputs. For classification, QCSA uses a shared semantic anchor-guided gated residual fusion mechanism to jointly model the original citation context and quadruple prompts. A weighted multi-backbone ensemble is further used as an optional enhancement to explore complementary information across pretrained language models. Experimental results show that QCSA achieves a Macro-F1 of 87.81% on the Athar dataset, while QCSA-Single (RoBERTa) reaches 96.49% on CSCE. These results indicate competitive performance on both benchmark datasets. Ablation and multi-seed experiments further show that the progressive quadruple construction provides a relatively stable contribution, while gated residual fusion offers a limited supplementary gain.
No takes yet. Share an insight, caveat, or question.
You et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: