Abstract: Frequent occurrence of diseases has posed a serious threat to the healthy development of the laying hens industry. Previous studies have focused mainly on the application of artificial intelligence technologies to the intelligent diagnosis of laying hens diseases at present. However, the low accuracy with single data or knowledge cannot fully meet the diagnosis of complex laying hens diseases, particularly for the small number of diagnosable diseases. Graph structure and knowledge graphs can be expected to easily acquire a large amount of data. Furthermore, diagnosing diseases has been limited in the incomplete analysis of correlation information and incomplete knowledge of laying hens using a single text description. This study aims to propose a Bidirectional Encoder Representation from Transformers (BERT) that integrates the knowledge graph and text of typical laying hens diseases using a Transformer based Bidirectional Long Short Term Memory Network (BiLSTM), in order to construct BERT-LHDKG (BERT Laying Hens Disease Knowledge). The diagnostic model was used for intelligent diagnosis of 38 typical diseases in the laying hens, such as Mycoplasma Synoviae, Newcastle Disease, and Infectious Rhinitis. The disease text and knowledge graph data were more comprehensively combined to conduct a comprehensive analysis of the incidence of laying hens. The triple vectors were introduced to represent the knowledge graph. The text feature vector and the triple vector were added inside the BERT model to form a fusion vector. More useful features were extracted for the disease analysis and diagnosis. The performance comparison test showed that the macroprecision of the BERT-LHDKG diagnostic model was 94.27%, the macrorecall was 94.12%, and the macro-F1 was 94.01%. The macroprecision was improved by 10.02, 2.64, 2.18 percentage points, the macrorecall was improved by 10.28, 2.29, 2.29 percentage points, and macro-F1 was improved by 10.66, 2.51, 2.19 percentage points, respectively, compared with deep learning models, such as TextCNN, the BERT model combined with CNN (convolutional neural networks), and the ERNIE model combined with BiLSTM. Better performance was also achieved in the five viral, bacterial, toxic, and metabolic diseases that were prone to occur in laying hens farming. The macro-F1 values of the BERT-LHDKG diagnostic model were 96.43%, 95.57%, 96.72%, and 98.24%, respectively. The ablation experiments showed that the removal of each layer from the BERT-LHDKG structure reduced the diagnostic accuracy of the improved model. Meanwhile, the macroprecision, macrorecall, and macro-F1 of the improved model decreased by 5.14, 4.98, and 5.46 percentage points, respectively, after the removal of the laying hens disease knowledge graph. Therefore, the knowledge graph was integrated to link the entities and relationships in the epidemic texts to the corresponding entities in the knowledge graph. The semantic information of the text was enriched to fully understand the content of the text. Thereby, the accuracy and robustness of the improved model was improved for epidemic diagnosis. In addition, the laying hen disease diagnosis web system was developed on the basis of the BERT-LHDKG diagnostic model, indicating the flexibility of remote diagnosis for laying hen diseases through human-machine dialogue.