SPUR-XMC: Context-Adaptive Spurious Correlation Mitigation for Extreme Multi-Label Classification of Biomedical Literature

Ratri Mukherjee, Shailesh Dahal, Kishlay Jha


Abstract
Extreme multi-label text classification (XMTC) of biomedical literature aims to classify articles into a set of pre-defined labels from a large and highly skewed label distribution space. While recent pretrained language model (PLM)-based XMTC approaches have made substantial advances, they still struggle to generalize to tail labels representing concepts of high biomedical significance (e.g., rare diseases). A key factor contributing to poor tail-label performance is the presence of spurious correlations, where models inadvertently learn non-informative feature-label associations. To address this issue, we propose a novel context-adaptive spurious correlation mitigation technique, SPUR-XMC, which selectively attenuates spurious influences according to the context of the input article. A unique advantage of the proposed framework is that it does not modify existing PLM architectures for XMTC and instead operates as an add-on enhancement module. Extensive experiments across benchmark biomedical literature datasets demonstrate the effectiveness of the proposed approach.
Anthology ID:
2026.findings-emnlp.1597
Volume:
Findings of the Association for Computational Linguistics: EMNLP 2026
Month:
October
Year:
2026
Address:
Budapest, Hungary
Editors:
Hung-Yi Lee, Sunipa Dev, Jessy Li, Desmond Elliott
Venue:
Findings
SIG:
SIGDAT
Publisher:
Association for Computational Linguistics
Note:
Pages:
34025–34045
Language:
URL:
https://aclanthology.org/2026.findings-emnlp.1597/
DOI:
10.18653/2026.findings-emnlp.1597
Bibkey:
Cite (ACL):
Ratri Mukherjee, Shailesh Dahal, and Kishlay Jha. 2026. SPUR-XMC: Context-Adaptive Spurious Correlation Mitigation for Extreme Multi-Label Classification of Biomedical Literature. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 34025–34045, Budapest, Hungary. Association for Computational Linguistics.
Cite (Informal):
SPUR-XMC: Context-Adaptive Spurious Correlation Mitigation for Extreme Multi-Label Classification of Biomedical Literature (Mukherjee et al., Findings 2026)
Copy Citation:
PDF:
https://aclanthology.org/2026.findings-emnlp.1597.pdf