High school math text similarity studies based on CNN and BiLSTM

Text similarity has been widely studied and applied in information retrieval, question answering system and so on. However, in the specific field of mathematics, there is little research at present. In this paper, aiming at the characteristics of high school mathematics text, such as strong logic, formula doping, including letters and numbers, a symmetrical twin network structure model combining CNN and BiLSTM is proposed. The local mathematical text feature matrix extracted by CNN and the global text feature matrix extracted by BiLSTM are modeled at different levels to obtain the deep level language features of sentences. Then the two similarity matrixes are fused and spliced to maximize the mathematical text features. Finally, the text similarity value is calculated by the matching layer, and then the answer score is obtained. The accuracy of the model is 84.3% when tested on real mathematics test questions. The experimental results show the effectiveness of the model.