To understand the user"s query intention is an important task in the field of QA. Traditional user intention recognition mainly uses methods based on template matching or artificial feature set, aiming at its high cost and low scalability, a hybrid neural network intention recognition model based on BERT word embedding and BiGRU-Attention is proposed. The model treats intention recognition as a classification problem. Firstly, the word embedding pre-trained by BERT is used as the input. Secondly, complete context learning is carried out through BiGRU network and the final representation of continuous statements is obtained. Through the attention mechanism, the neural network can be assisted to give more weight to the words that have a direct effect on intention recognition in sentences, making intention recognition more explicit. Finally, user intention recognition is realized through classification. Experimental results show that the proposed method can effectively improve the accuracy of intention recognition accuracy.