The neural network ranking model has been widely used in the ranking task of the information retrieval field. It requires extremely high data quality; however, the information retrieval datasets usually contain a lot of noise, and documents irrelevant to the query cannot be accurately obtained. High-quality negative samples are essential to training a high-performance neural network ranking model. Inspired by the existing doc2query method, we propose a deep and end-to-end model AQGM. This model increases the diversity of queries and enhances the quality of negative samples by learning mismatched query document pairs and generating adversarial queries irrelevant to the documents and similar to the original query. Then, we train a deep ranking model based on BERT with the real samples and the samples generated by the AQGM model. Compared with the baseline model BERT-base, our model improves the MRR index by 0.3% and 3.2% on the MSMARCO and TrecQA datasets, respectively.