Text similarity of autonomous marking system based on deep learning

Abstract

With the practical application of the self-marking system, it can give an objective and fair score, but it can not meet the needs of classroom teaching. It also needs to give students rapid feedback on the vocabulary, sentence, text structure, content relevance, and other dimensions presented in the test paper. The scoring method based on artificial features is the earliest self-marking scoring method, which uses experts to design some scoring features from the language quality, content quality, and text structure of the test paper. It takes the scoring task as a regression or classification task to score or rate the test paper. In this paper, based on the deep learning theory, a method for text similarity detection using a twin network is proposed. Considering the interaction between text pairs, we integrate expressivity pooling based on bidirectional GRU and measure the similarity by distance calculation formula. The experimental data show that the two-way GRU network integrated with expression pooling can obtain the interaction between text pairs so that the extracted features of the text are more comprehensive. The model has a better effect in the study of the similarity of text pairs. This method can reduce the difference in scoring results caused by different subjective consciousness of raters, make the scoring results more objective and persuasive, and improve the accuracy and efficiency of scoring.