Box-enhanced context fusion for citation intent classification

Citation intent classification is a core research topic in scientometrics and information retrieval. This task analyzes citation contexts and relationships to categorize citation behaviors such as providing support, making comparisons, or offering contextual background. Identifying these behaviors helps reveal the specific role of citations in scholarly communication and clarifies the underlying authorial intent. Existing text-encoding approaches typically formulate citation intent classification as a conventional text classification task. However, this formulation overlooks two critical factors: the influence of citation marker position on classification performance and the implicit hierarchical relationships among texts within the same class.To address these limitations, we propose GLB-Cite, a novel citation intent classification model that fuses local citation marker information with a global context representation to obtain a unified fused representation. This representation is then projected into two parallel spaces: a vector embedding space for semantic discrimination and a box embedding space for modeling overlap between citation contexts and intent classes. By jointly optimizing these two branches, GLB-Cite balances semantic expressiveness and geometric boundary modeling, yielding notable gains especially for intent classes with overlapping semantics and ambiguous decision boundaries.We evaluate the performance of GLB-Cite on two public datasets, SciCite and ACL-ARC, using the macro-F1 score as the primary evaluation metric.Experimental results show that GLB-Cite achieves state-of-the-art performance, with F1 scores of 87.69% on SciCite and 79.79% on ACL-ARC.