LSTM is an improved variant of RNN, which has a high efficiency in processing time load data and can effectively solve the common problems of gradient vanishing and exploding in RNN networks. Dictionary extraction has become a hot topic in corpus linguistics research. This article studies a bilingual dictionary extraction algorithm based on recurrent neural networks. Firstly, the recurrent neural network model is studied, including the RNN basic model and the LSTM model; Next, an extraction model based on LSTM was constructed, and the model training process and dictionary extraction process were designed; Then, the model algorithm was studied, including activation functions and optimization functions; Finally, simulation experiments were conducted, and evaluation metrics, experimental environments, and parameter configurations were designed. Four commonly used neural network models, including RNN, GRU, BPTT, and LSTM, were compared and analyzed. The results showed that LSTM has significant advantages in bilingual dictionary extraction.