In order to support the construction of knowledge graphs in heterogeneous biological texts and databases, a deep learning framework combining entity recognition, standardized mapping, relation extraction and graph fusion was proposed. The experimental data are collected from PubMed abstracts, biological terminological databases and structured relational databases, and 182000 entity mentions, 396000 candidate relations and 128000 triplet candidates are obtained after cleaning. In the first stage, the entity coding network based on BioBERT extracted the context semantic features and stably completed the term standardization mapping. In the second stage, the relation classification module identified the semantic types between entities and generated candidate triples. In the third stage, the fusion building module performs node merging, relation alignment and structure writing. Experimental results show that the framework achieves 93.2% entity recognition accuracy, 91.3% standardized accuracy, and 89.8% macro average F1 value of relation classification. The node repetition rate is controlled to 4.1%, the effective write success rate is 88.4%, and the single-hop query response time is 84 ms.
