Abstract: The directly-acquired texts often have problems such as weak semantic connection, excessive length, and polysemy. Therefore, this paper proposes a two-stage triplet knowledge extraction method via location-wise based on BERT pre-training. The BERT-Span model was used to achieve entity recognition of tourism through boundary prediction. A relationship extraction model combining positional perception attention and head-tail entity type was constructed based on the character, semantics, location, and entity type characteristics. The experimental results on the Shanxi tourism dataset show that the proposed method is superior to benchmark models in the F1 value.