In view of the redundancy of knowledge in the field of tobacco grading and the absence of a professional platform for academic retrieving, the knowledge graph of tobacco grading was constructed by collecting multi-source tobacco grading data and combining the top-down method, and an intelligent question and answer system was developed on this basis. The core technologies are as follows. 1) Collecting tobacco leaf grading data through named entity recognition(NER) and relation extraction(RE) to extract triplet information, and import it into the Neo4j platform for storage. 2) For question semantic parsing, the BERT-BiGRU-MHSA-CRF model fused with graph data was used to improve the entity recognition effect of question sentences, and the self-attention mechanism was integrated into the BERT-TextCNN model to parse user hierarchical intent. Then, the cypher query statement was automatically constructed by matching the template and replacing the slot information, and the most accurate answer was retrieved and returned in the Neo4j knowledge base. The results showed that the constructed knowledge graph contains 6 620 entities and more than 14 000 relationships. The harmonic mean F1 of the question entity recognition model BERT-BiGRU-MHSA-CRF was 94.12%, and the F1 of the hierarchical intent recognition model BERT-TextCNN-Attention was 98.77%. In summary, the system can quickly retrieve and accurately answer multiple types of questions related to tobacco grading, which can provide auxiliary functions for graders.