In order to construct a knowledge graph of Hainan marine intangible cultural heritage, this paper designs the graph database by setting the annotation system and semantic structure of the intangible heritage metadata. For the named entity recognition of the knowledge graph, a model combining long and short-term memory neural networks and Bi-Lstm-CRF with conditional random field is established. The recognized entities are processed based on the Bert pre-training model, and the semantic relations are extracted by combining the cluster search method. Compared with other models, the accuracy rate of the Bs-Spert model for the extraction of semantic relations of the Hainan marine NRM dataset is improved by 0.64%, 0.85% and 0.98% compared with GCNRE, Match Blank and MTDS, respectively. Meanwhile, other metrics of the Bs-Spert model also performed well, with a recall rate of 82.26% and an F1-value of 74.01%. By constructing a knowledge map based on the graph database, we can analyze the current situation of Hainan’s marine NRM more comprehensively and thus propose corresponding countermeasures.