A Joint Multi-Task Learning Framework for Citation Intent and Citation Evaluation Classification
Yang Zhao, Guangyin ZhangAbstract
Purpose
This study aims to improve fine-grained understanding of citation contexts in scientific literature by jointly modeling two complementary aspects of citation behavior: citation intent and citation evaluation.
Design/methodology/approach
We propose a joint multi-task learning (Joint-MTL) framework that simultaneously models citation intent classification and citation evaluation classification through a shared encoder with task-specific prediction heads. To support the evaluation task, we construct a new manually annotated dataset, CiteEva. The model is trained using an alternating optimization strategy across the two tasks, enabling the shared encoder to learn representations that capture both functional and evaluative characteristics of citation contexts.
Findings
Experiments conducted on the public SciCite dataset and the newly constructed CiteEva dataset show that the joint learning framework consistently outperforms strong single-task baselines and several recent multi-task or feature-based approaches. Additional analysis reveals a moderate representation overlap between the two tasks, indicating that they share meaningful semantic signals while maintaining task-specific distinctions. These findings empirically support the effectiveness of jointly modeling citation intent and citation evaluation.
Research limitations
The current study focuses on two citation-related tasks and evaluates the framework on a limited set of datasets and citation categories. Future work could extend the approach to additional citation analysis tasks and larger-scale corpora.
Practical implications
The proposed approach can support applications such as scientific impact assessment, literature review assistance, and scholarly knowledge mining by enabling more nuanced interpretation of citation roles and evaluative stances.
Originality/value
This study provides empirical evidence for the benefit of jointly modeling citation intent and citation evaluation and introduces CiteEva, a new high-quality dataset for citation evaluation research, contributing to more comprehensive citation context analysis.