ERNIE and Multi-Feature Fusion for News Topic Classification

Abstract

Traditional news topic classification methods suffer from inaccurate text semantics, sparse text features and low classification accuracy. Based on this, this paper proposes a news topic classification method based on Enhanced Language Representation with Informative Entities (ERNIE) and multi-feature fusion. A semantically more accurate representation of text embedding is obtained by ERNIE. In addition, this paper extracts word, context and key sentence based on the news text. The key sentences of the news are obtained through the TextRank algorithm, which enables the model to focus on the content points of the news. Finally, this paper uses the attention mechanism to realize the fusion of multiple features. The proposed method is experimented on BBCNews. The experimental results show that we achieve classification accuracies superior to those of the compared methods, while validating the structural validity of the proposed method. The method in this paper has a positive effect on promoting the research of news topic classification.