利用自然语言处理技术从生物医学文本中抽取药物治疗、疾病诊断等事件以及事件中涉及的疾病、药物等实体，对于生物医学领域相关学术研究以及各类生物医学应用系统具有重要意义。针对生物医学文本中的缩略词及专业术语难以识别和生物医学语义关系难以嵌入的问题，提出了一种融合外部知识和图卷积神经网络的生物医学信息联合识别模型。图卷积神经网络构建了包含实体和语义关系的异构图，能够迭代地融合本地知识图和外部知识图中的交互信息，根据得到的交互信息来进行生物医学实体对之间关系的抽取任务。预训练编码后利用图卷积神经网络构建本地和外部知识两个知识图，获得两个图中每个节点的特征表示，并且通过注意力实体链接的方法将两个图进行融合与信息迭代，进而抽取其最后一层隐藏层来完成最终的分类识别。其中统一医学语言系统（unified medical language system, UMLS）被用作实体消歧的外部知识库，实体链接器根据注意力权重选择对应实体。通过在MLEE语料库上进行的实验表明，联合任务能够实现事件抽取和触发词、元素识别的综合性能。

The use of natural language processing techniques to extract events such as drug treatment and disease diagnosis from biomedical texts and the entities involved in the events, such as diseases and drugs, is important for academic research and various biomedical application systems in the field of biomedicine. The idea of this paper is a joint biomedical information recognition model that incorporates external knowledge and graph convolutional neural network to address the problems of difficult recognition of abbreviations and specialized terms in biomedical texts and difficult embedding of biomedical semantic relationships. A heterogeneous graph was constructed by the graph convolutional neural network which contains entities and semantic relations, and is able to iteratively fuse the interaction information in the local knowledge graph and the external knowledge graph to perform the task of extracting relations between biomedical entity pairs based on the obtained interaction information. After pre-training and coding, two knowledge graphs were constructed using a graph convolutional neural network to obtain feature representations of each node in the two graphs, and the two graphs were fused and iterated through the attentional entity linking method to extract the last hidden layer for final classification recognition. The UMLS was used as an external knowledge base for entity disambiguation, and the entity linker selects the corresponding entities according to the attention weights. Experiments conducted on the MLEE corpus show that the joint task is able to achieve comprehensive performance in event extraction and trigger word and element recognition.