DEIB: A Dual-stage Enhanced Information Bottleneck Framework for Joint Entity and Relation Extraction

Joint Entity and Relation Extraction (JERE) aims to simultaneously identify entities and their semantic relations. Despite substantial progress, current JERE models are still hindered by redundant contextual information and noise introduced by external knowledge, which obscure the sparse scientific semantics embedded in biomedical texts, while semantic inconsistencies between entity recognition and relation extraction further weaken the coherence of joint reasoning. To mitigate these limitations, we propose a Dual-stage Enhanced Information Bottleneck (DEIB) framework. DEIB incorporates dual-stage IB modules that progressively compress span-level and knowledge-enhanced representations, thereby filtering task-irrelevant contextual signals and suppressing noise arising from imperfect external knowledge. In addition, a Semantic-Consistent Interaction Module (SCIM) is designed to alleviate cross-subtask semantic inconsistency by reformulating the traditional biaffine scorer into a cross-task reasoning layer that captures high-order dependencies between entity and relation representations. Through ablation studies and case analyses, we confirm that each component of the DEIB framework is critical to improving robustness, as they collectively suppress redundant context and knowledge noise while enhancing cross-subtask semantic consistency. Furthermore, comparative experiments on the BioRelEx and ADE datasets show that DEIB achieves state-of-the-art results, with F1-scores of 93.42% and 79.85% for entity and relation extraction on BioRelEx, and 93.31% and 87.26% on ADE. These experimental results validate that DEIB is accurate and effective.