Literature Classification Model of Deep Learning Based on BERT-BiLSTM——Taking COVID-19 as an Example

Li, Z.

[Objective] In the beginning of 2020, the outbreak of covid-19 epidemic occurred, and the number of literatures related to it increased rapidly. In order to meet the increasing needs of literature classification, this paper will explore an automated literature classification model. [method] firstly, more than 20000 articles related to covid-19 in CNKI were collected as the marked data set, and the title, keyword and abstract information of the articles were extracted. Then, different combinations of the title, keyword and abstract were used as the input of different features of the model, and the support vector machine (SVM) based literature classification model and BERT-LSTM based literature classification model were trained respectively and the effect of the model was compared. [Results] The accuracy of the feature combination of “title + keyword + abstract” in the BERT-BiLSTM model was 85.79%, which was higher than the accuracy of the feature combination of “title + keyword” and “keyword + abstract”. The accuracy of the benchmark model (SVM) with the combination of “title + keyword + abstract” is 78.79%, so the model based on BERT-BiLSTM can significantly improve the classification effect. [limitation] this paper only classifies four categories, which are classified under R category, so the classification is very few. © 2021, Springer Nature Switzerland AG.