Scholarly web is a vast network of knowledge connected by citations. However, this system is increasingly compromised by miscitation, where references do not support or even contradict the claims they are cited for. Current miscitation detection methods, which primarily rely on semantic similarity or network anomalies, struggle to capture the nuanced relationship between a citation's context and its place in the wider network. While large language models (LLMs) offer powerful capabilities in semantic reasoning for this task, their deployment is hindered by hallucination risks and high computational costs. In this work, we introduce LLM-Augmented Graph Learning-based Miscitation Detector (LAGMiD), a novel framework that leverages LLMs for deep semantic reasoning over citation graphs and distills this knowledge into graph neural networks (GNNs) for efficient and scalable miscitation detection. Specifically, LAGMiD introduces an evidence-chain reasoning mechanism, which uses chain-of-thought prompting, to perform multi-hop citation tracing and assess semantic fidelity. To reduce LLM inference costs, we design a knowledge distillation method aligning GNN embeddings with intermediate LLM reasoning states. A collaborative learning strategy further routes complex cases to the LLM while optimizing the GNN for structure-based generalization. Experiments on three real-world benchmarks show that LAGMiD achieves state-of-the-art miscitation detection with significantly reduced inference cost.

Abstract Scientific text classification is essential for efficiently organizing and assimilating scientific knowledge. However, existing methods struggle to classify ultra‐short scientific texts due to their limited content and complex hierarchical labeling. To overcome these challenges, we introduce the BERT‐HMCN framework, which combines Bidirectional Encoder Representations from Transformers (BERT) with a Hierarchical Multi‐label Classification Network (HMCN). This framework introduces a novel level‐fixed fine‐tuning strategy that strengthens the connection between text semantics and hierarchical labels, enhancing the representation of ultra‐short texts. We evaluated BERT‐HMCN's performance on a dataset of 75,065 program titles from the National Natural Science Foundation of China. Our results show that BERT‐HMCN outperforms existing models in both overall performance and hierarchical accuracy. We also conducted a comparative analysis with autoregressive large language models (LLMs), illustrating the strengths of each in different contexts. Further analysis confirms the effectiveness and robustness of the BERT‐HMCN framework. We discuss its theoretical contributions and practical applications, underscoring the broader implications of these results in scientific text classification and other related fields.

Graph Neural Networks (GNNs) have gained attention for their ability in capturing node interactions to generate node representations. However, their performances are frequently restricted in real-world directed networks with natural hierarchical structures. Most current GNNs incorporate information from immediate neighbors or within predefined receptive fields, potentially overlooking long-range dependencies inherent in hierarchical structures. They also tend to neglect node adaptability, which varies based on their positions. To address these limitations, we propose a new model called Hierarchy-Aware Adaptive Graph Neural Network (HAGNN) to adaptively capture hierarchical long-range dependencies. Technically, HAGNN creates a hierarchical structure based on directional pair-wise node interactions, revealing underlying hierarchical relationships among nodes. The inferred hierarchy helps to identify certain key nodes, named Source Hubs in our research, which serve as hierarchical contexts for individual nodes. Shortcuts adaptively connect these Source Hubs with distant nodes, enabling efficient message passing for informative long-range interactions. Through comprehensive experiments across multiple datasets, our proposed model outperforms several baseline methods, thus establishing a new state-of-the-art in performance. Further analysis demonstrates the effectiveness of our approach in capturing relevant adaptive hierarchical contexts, leading to improved and explainable node representation.