Effective Multi-Task Learning for Biomedical Named Entity Recognition

2025-07-24

AI Summary

Biomedical named entity recognition (BM-NER) faces challenges including terminological complexity, pervasive nested entity structures, and inconsistent annotations across datasets. To address these, we propose SRU-NER: a framework featuring a Slot-based Recurrent Unit (SRU) that explicitly models nested entities; multi-task learning to jointly train on heterogeneous biomedical and general-domain NER datasets; and a novel dynamic loss adjustment mechanism that automatically suppresses supervision signals for unannotated entity types—effectively mitigating negative transfer induced by annotation discrepancies. Crucially, SRU-NER enables unified modeling of both biomedical and general-domain NER. Experimental results demonstrate state-of-the-art or competitive performance across multiple benchmark datasets. Cross-domain evaluation and human assessment further validate its strong generalization capability and practical effectiveness.

Technology Category

• Natural Language Processing: Summarization
• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Knowledge Representation and Reasoning: Nonmonotonic Reasoning

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs

Abstract

Biomedical Named Entity Recognition presents significant challenges due to the complexity of biomedical terminology and inconsistencies in annotation across datasets. This paper introduces SRU-NER (Slot-based Recurrent Unit NER), a novel approach designed to handle nested named entities while integrating multiple datasets through an effective multi-task learning strategy. SRU-NER mitigates annotation gaps by dynamically adjusting loss computation to avoid penalizing predictions of entity types absent in a given dataset. Through extensive experiments, including a cross-corpus evaluation and human assessment of the model's predictions, SRU-NER achieves competitive performance in biomedical and general-domain NER tasks, while improving cross-domain generalization.

Problem

• Handling nested biomedical named entities
• Integrating multiple datasets via multi-task learning
• Improving cross-domain generalization in NER

Innovation

• SRU-NER handles nested named entities
• Dynamic loss adjustment for annotation gaps
• Multi-task learning integrates multiple datasets