Research and application of matching network

With the rapid development of deep learning and natural language processing, more and more systems have applied deep learning models. However, a large number of data for training is a major bottleneck of deep learning at present. For the postgraduate thesis oral defense system, our model still utilizes the word retrieval method to match teachers and students who have the same research field because of the small amount of data and information. In this paper, we propose a two-stage training framework to improve the system matching correlation which fine-tunes the pre-trained model on specific downstream data and then utilizes contrastive learning and matching network to conduct self-supervised training. At the same time, the framework uses adversarial training to improve the robustness of the model. We evaluate our approach on the dataset of our system, and experiment results demonstrate the effectiveness of our approach.