ScholarWorks@중앙대학교

초록

This paper proposes a novel synthetic intent-based model for the citation intent classification, called SynIntent. Unlike prior works that treat citation contexts as plain text and rely on deep neural networks (including language models (LMs)) for representation learning, we investigate an unexplored information for this task: latent intent. To exploit these signals, we represent each citation context as a bipartite heterogeneous graph which composed of two node types: sentence nodes and latent intent nodes. Initial node representations are produced by a graph initialization module and subsequently refined through a graph attention network to capture sentence-intent dependencies, yielding a richer and more informative contextual representation. By leveraging latent intent information and sentence–intent relationships, SynIntent provides a novel, deeper and more holistic understanding of the semantic structure underlying citation contexts. The experiments on public datasets demonstrate that SynIntent outperforms the state-of-the-art models in both multiclass and multilabel classification scenarios. Specifically, we achieve a remarkable increase of 4.3% macro-F1 and 4.4% in accuracy on the ACL-ARC dataset. For the multilabel classification, we also observe improvements of 1% for strict metrics and 2% for weak metrics on the MultiCite dataset.

키워드