EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Label-Aware Evidence Modulation for Multi-Granularity Multi-Label Classification of Scientific Documents

Shuming Zhong, Jingmin Luan, Yao Yu, Zhenhe Ma

IEEE Access · 2026 · 14 · 74688-74708

DOI: 10.1109/access.2026.3692705 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Shuming Zhong und 3 weiteren Autor:innen. Die Arbeit nennt 42 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Natural Science Foundation of China, Industry-Academia-Research Cooperation Project between Hebei-located Universities and Shijiazhuang City.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Multi-granularity multi-label classification of scientific documents is a fundamental task in digital libraries, requiring models to predict both coarse-grained (L1) and fine-grained (L2) labels using limited text like titles and abstracts. However, at the fine-grained level, challenges such as severe label semantic overlap, pronounced long-tail distributions, and sparse supervision often cause existing text-driven models relying on label-query alignment to suffer from “insufficient discriminative evidence” and unstable tail predictions. To address these issues, we propose the Label-aware Evidence Modulation (LEM) framework. Unlike methods depending on external knowledge graphs, LEM induces reusable semantic evidence directly from the internal text representation space. Specifically, LEM integrates an Evidence-induced Cluster and Cross-prototype Graph (ECCG) branch alongside the text backbone. It constructs cluster-level nodes via soft clustering, captures cross-document shared patterns through prototype routing, and generates label-specific evidence via a readout module. Furthermore, an evidence modulation module is designed to adaptively control evidence injection intensity, suppressing irrelevant interference. Experiments on OpenAlex and AAPD datasets demonstrate that LEM consistently outperforms the Label-Query Baseline, improving L2 Macro-F1 by + 0.0119 and + 0.0206, respectively. Ablation studies further validate the contributions of internal evidence induction and adaptive modulation modules to the performance gains.

Förderung

  • National Natural Science Foundation of China
  • Industry-Academia-Research Cooperation Project between Hebei-located Universities and Shijiazhuang City

Zitieren

APA

Zhong, S., Luan, J., Yu, Y., & Ma, Z. (2026). Label-Aware Evidence Modulation for Multi-Granularity Multi-Label Classification of Scientific Documents. IEEE Access, 14, 74688-74708. https://doi.org/10.1109/access.2026.3692705

Vancouver

Zhong S, Luan J, Yu Y, Ma Z. Label-Aware Evidence Modulation for Multi-Granularity Multi-Label Classification of Scientific Documents. IEEE Access. 2026;14:74688-74708. doi:10.1109/access.2026.3692705

BibTeX
@article{Zhong2026,
  author = {Zhong, Shuming and Luan, Jingmin and Yu, Yao and Ma, Zhenhe},
  title = {Label-Aware Evidence Modulation for Multi-Granularity Multi-Label Classification of Scientific Documents},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {74688--74708},
  doi = {10.1109/access.2026.3692705}
}
Kostenlos starten →