ScholarWorks@중앙대학교

초록

Understanding the functional role of citations in scientific literature requires more than analyzing textual content alone. This study explores whether structural positioning specifically, the section in which a citation appears can improve citation intent classification. We introduce a heuristic weighting mechanism that adjusts sentence embeddings based on the citation's location within standardized document sections (e.g., Introduction, Methods, Conclusion). Using a dataset of 818 full-text articles, we evaluate this approach across multiple embedding strategies, including TF-IDF and pretrained transformer models such as BERT, RoBERTa, MiniLM, MPNet, and SciBERT. Each model is trained with and without section-based weighting to quantify the contribution of structural signals. Experimental results show consistent improvements in macro F1 scores, particularly for underrepresented intent classes, without degrading overall performance. These findings suggest that incorporating document structure provides valuable context for enhancing citation intent classification.

키워드