Abstract: Question matching aims to determine whether the intentions of two different questions are similar. In recent years, with the development of large-scale pre-trained language models, the state-of-the-art performance of question matching has been achieved by these deep models. However, due to the assumption of independent and identical distribution (I.I.D), the performance of these deep models in real scenarios is still limited by the adequacy of training data and the distribution drift between target data and training data. In this article, we propose a novel method for question matching based on the paradigm of gradual machine learning. It extracts diverse semantic features from different perspectives, and then constructs a factor graphs by fusing the extracted feature information to enable gradual learning from easy to hard. In feature modeling, we design and implement two types of features: 1) TF-IDF based keyword features; 2) DNN based deep features. Finally, we have validated the efficacy of the proposed method by a comparative study on the open-sourced benchmark datasets, LCQMC and BQ corpus. Our extensive experiments show that compared to pure deep learning models, our proposed method can effectively improve the accuracy of question matching, and its performance advantage increases with the decrease of labeled training data.