As China’s agriculture has developed towards modern agriculture, agricultural finance has played a crucial role in promoting China’s rapid agricultural development. Due to the ineffective interpretation of policy support of agricultural finance, poor agricultural areas cannot benefit from the current policy. The purpose of this paper is to propose a model that combines the ELECTRA model with a bi-directional long- and short-term memory network, BiLSTM and conditional random field (CRF) for identifying agricultural finance policy information. The first step is to convert the plain text of agricultural finance policy into word vectors using ELECTRA’s pre-trained language model. By incorporating zigzag character substitution in the generator, the adaptability of the pre-training target to the annotation task on agricultural finance policy is improved. Next, the BiLSTM neural network is used to obtain the contextual abstract features of the serialised text. Finally, the global optimisation sequence decoding annotation is combined with the CRF to extract the structured agricultural finance policy information. The results indicate that the agricultural finance policy information extraction model is more effective than other models in capturing agricultural finance policy information and has excellent information extraction capabilities.