Feature Mining of News Communication Topic Elements Based on BERT Model

Abstract

In order to solve the problems of lack of standardization, fuzzy semantics and sparse features in news topic texts, a feature mining of news communication topic elements based on the BERT model is proposed. In the research, multi-layer fully connected layer feature extraction is performed on the output of the news topic text in the BERT model, and the final extracted text features are purified by the feature projection method to enhance the classification effect.Then the feature projection network is fused in the hidden layer inside the BERT model for feature projection, so as to enhance and purify the classification features through the feature projection of the hidden layer. Experiments are performed on Toutiao, Sohu News, THUC News-L, and THUC News-S datasets. The experimental results show that compared with the baseline BERT method, the two methods have better performance in terms of accuracy and macro-average F1 value, and the highest accuracy is 86. 96%, 86. 17%, 94. 40% and 93.73%, respectively, which verifies the feasibility and effectiveness of the proposed method. It is concluded that the proposed method for news topic text classification combining BERT and FP net is effective and efficient.