作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2025, Vol. 51 ›› Issue (1): 106-117. doi: 10.19678/j.issn.1000-3428.0068877

• 人工智能与模式识别 • 上一篇    下一篇

基于证据和图推理的文档级关系抽取方法: 以医学关系为例

周雪阳1,2, 傅启明1,2,*(), 陈建平2,3,4, 陈延明5, 陆悠1,2, 王蕴哲1,2   

  1. 1. 苏州科技大学电子与信息工程学院, 江苏 苏州 215009
    2. 苏州科技大学江苏省建筑智慧节能重点实验室, 江苏 苏州 215009
    3. 苏州科技大学建筑与城市规划学院, 江苏 苏州 215009
    4. 重庆工业大数据创新中心有限公司, 重庆 400707
    5. 苏州大学附属第二医院神经外科, 江苏 苏州 215004
  • 收稿日期:2023-11-20 出版日期:2025-01-15 发布日期:2024-04-25
  • 通讯作者: 傅启明
  • 基金资助:
    国家重点研发计划(2020YFC2006602); 国家自然科学基金(62102278); 国家自然科学基金(62072324); 江苏省高等学校自然科学研究项目(21KJA520005); 江苏省重点研发计划(BE2020026); 江苏省研究生教育教学改革项目; 江苏省研究生科研与实践创新计划项目(KYCX23_3321)

Document-Level Relation Extraction Method Based on Evidence and Graph Inference: A Case Study of Medical Relationships

ZHOU Xueyang1,2, FU Qiming1,2,*(), CHEN Jianping2,3,4, CHEN Yanming5, LU You1,2, WANG Yunzhe1,2   

  1. 1. Electronic and Information Engineering, Suzhou University of Science and Technology, Suzhou 215009, Jiangsu, China
    2. Jiangsu Province Key Laboratory of Intelligent Building Energy Efficiency, Suzhou University of Science and Technology, Suzhou 215009, Jiangsu, China
    3. Architecture and Urban Planning, Suzhou University of Science and Technology, Suzhou 215009, Jiangsu, China
    4. Chongqing Industrial Big Data Innovation Center Co., Ltd., Chongqing 400707, China
    5. Department of Neurosurgery, The Second Affiliated Hospital of Soochow University, Suzhou 215004, Jiangsu, China
  • Received:2023-11-20 Online:2025-01-15 Published:2024-04-25
  • Contact: FU Qiming

摘要:

针对生物医学文献句式冗长、实体密集从而导致关系抽取复杂度高、难度大的问题, 提出一种证据路径增强的图推理框架(EPE-GR)。首先建立一种引入结构化偏差的图注意力机制(B-GAT)增强图推理中信息聚合的指向性, 结合提及级和实体级图建模学习全局交互特征和局部依赖信息; 其次使用启发式搜索聚焦证据句子, 同时构建一种基于掩膜多头注意力(MMHA)机制的路径推理结构, 强化非邻居证据句子之间的相关性并缓解细粒度证据编码带来的复杂度剧增的问题; 最后协同全局、局部和路径推理预测实体之间的语义关系。与已有方法相比, EPE-GR在药物-突变相互作用(DMI)数据集和化学物质诱导疾病(CDR)数据集上都获得了最佳的性能, 前者在二分类和多分类任务的设定下相比次优方法准确率分别提高了5.65和5.13百分点, 后者F1值提高了2.85百分点, 证明所提方法是一个有效的生物医学文档级关系抽取方法且具有较好的泛化能力。此外, 通过进一步的实验表明所提出的关系依赖建模和证据路径推理机制能够有效提升模型推理句间关系的能力。

关键词: 关系抽取, 图推理, 路径推理, 证据增强, 图注意力机制, 多头注意力机制

Abstract:

To address the challenges of complex and difficult relation extraction caused by long sentences and entity density in biomedical literature, this study proposes an Evidence Path Enhanced Graphical Reasoning Framework (EPE-GR). First, a graph attention mechanism that introduces structured bias (B-GAT) is established to enhance the directionality of information aggregation, combined with mention- and entity-level graph modeling to capture global and local features. Second, a heuristic search is used to focus on evidence sentences, and a path inference structure based on a mask multi-head attention mechanism is constructed to strengthen the correlation between non-neighbor evidence sentences and alleviate the complexity surge caused by fine-grained evidence encoding. Finally, global, local, and path reasoning are collaboratively used to predict semantic relations between entities. Compared to existing methods, EPE-GR demonstrates superior performance on Drug-Mutation Interaction (DMI) dataset and Chemical Induced Disease (CDR) dataset. For DMIs, the proposed framework improves accuracy by 5.65 percentage points in binary classification and 5.13 percentage points in multi-classification. For CDRs, the F1 value increases by 2.85 percentage points. These results confirm that EPE-GR is an effective document-level biomedical relationship extraction method with strong generalization ability. Further experiments highlight the effectiveness of the proposed relationship dependency modeling and evidence path inference mechanism in enhancing inter-sentence relation model inference.

Key words: relation extraction, graphical reasoning, path reasoning, evidence enhancing, graph attention mechanism, multi-head attention mechanism