In visual question answering (VQA), most existing models suffer from language biases which make models not robust. Recently, many approaches have been proposed to alleviate language biases by generating samples for the VQA task. These methods require the model to distinguish original samples from synthetic samples, to ensure that the model fully understands two modalities of both visual and linguistic information rather than just predicts answers based on language biases. However, these models are still not sensitive enough to changes of key information in questions. To make full use of the key information in questions, we design a self-supervised strategy to make the nouns of questions be focused for enhancing the robustness of VQA models. Its auxiliary training process, predicting answers for synthetic samples generated by masking the last noun in questions, alleviates the negative influence of language biases. Experiments conducted on VQA-CP v2 and VQA v2 datasets show that our method achieves better results than other VQA models.