Given that the text paragraphs in the field of wildlife under special state protection are lengthy, the content is fragmented, and the vast majority of the data are unstructured text data. To this end, this paper proposes a knowledge graph construction method for national key protected wildlife based on the lightweight pretrained language model ALBERT, which realizes the construction of the graph while extracting ternary groups from the massive amount of data through a top-down approach. In the self-constructed dataset of wildlife under special state protection, the ALBERT model encodes the features of word vectors, fusing the BiLSTM-CRF and BiLSTM-Attention methods for entity recognition and relation extraction. The experimental results demonstrate that the self-constructed dataset achieves F1 values of 93.72% and 96.04% for entity recognition and relation extraction, respectively. The Neo4j graph database integrates and stores the extracted triples. The constructed knowledge graph can provide an effective information base for subsequent ecological protection research and policy formulation.