再创佳绩 近日，实验室博士孙新圆关于文档级生物医学关系抽取的研究成果被人工智能领域国际知名期刊《Engineering Applications of Artificial Intelligence》接收。该期刊专注于人工智能在工程领域的应用研究，是中科院分区TOP期刊，最新影响因子8，被CCF列为C类推荐期刊。本工作提出了一种带有动态记忆机制的多阶段推理框架，通过异构图推理、动态记忆架构以及十字交叉注意力机制，有效解决了复杂文档中长距离实体交互与数据噪声的问题，显著提升了关系抽取的综合推理能力。题目：Multi-Stage Reasoning Framework for Biomedical Document-Level Relation Extraction with Dynamic Memory Mechanism摘要：Document-level Biomedical Relation Extraction identifies and classifies the diverse types of interactions that occur between biomedical entities across specialized literature. Graph-based techniques have significantly advanced the field of relation extraction. However, most existing methods focus on optimizing a single reasoning capability, lacking enhancement in holistic reasoning capabilities; moreover, previous approaches fail to address the noise issues present in training data, particularly within large-scale distantly supervised data. To overcome these limitations, we propose the Multi-Stage Reasoning Framework with Dynamic Memory Mechanism, which integrates contextual and structural representations to enhance relation extraction capabilities for Document-level Biomedical Relation Extraction. Our approach introduces three key innovations: 1) a heterogeneous graph reasoning component that models fine-grained interactions of cross-sentence entities through a co-reference resolution module; 2) a dynamic memory-augmented architecture with trainable memory slots for preserving longrange entity interactions; 3) an iterative interactive reasoning module enabling axis entity pairs interaction via criss-cross attention. Extensive experiments on three public datasets demonstrate the effectiveness of our proposed model. Further experimental results demonstrate that our model significantly outperforms mainstream large language models on Document-level Biomedical Relation Extraction tasks.Our code is available athttps://github.com/SXY09/MSRDM
夜雨聆风
