WCSE

News Topic Prediction Via Transformer

Abstract—News topic prediction has gotten a lot of attention recently. Existing methods provide recommendation prediction primarily through content-based or collaborative filtering techniques. However, these methods do not take emotional information embedded in news text into account, they performcomparatively poor in the tasks of news classification and recommendation. To fill this gap, we propose EMTransformer, a simple encoder-decoder model that incorporates emotional information to improve news topic prediction. Concretely, we begin with improving A-TFIDF, an aligned TFIDF method, for extracting the keywords from news components in a more precise manner. Then, we construct the emotional dictionary by calculating mutual information degree and information entropy. Finally, we combine the original Transformer with emotional encodings for news topic prediction. We conduct extensive experiments on two real-world datasets, publicly available on People's Daily Online and Xinhua Net. The results demonstrate EM-Transformer outperforms classical baselines on the two data sources, and that emotional encoding increases in model quality with 15.97% and 6.37% in accuracy respectively, as oppose to the original Transformer.

Index Terms—Transformer, TFIDF, topic prediction, emotional encoding

Jibing Gong, Kai Yu, Chaoyuan Huang, Yuting Lin, Chenglong Wang, Jinye Zhao, Shishan Gong, Huanhuan Li
School of Information Science and Engineering, Yanshan University, CHINA

[Download]

Cite: Jibing Gong, Kai Yu, Chaoyuan Huang, Yuting Lin, Chenglong Wang, Jinye Zhao, Shishan Gong, Huanhuan Li, "News Topic Prediction Via Transformer, " Proceedings of 2022 the 12th International Workshop on Computer Science and Engineering (WCSE 2022), pp. 159-169, June 24-27, 2022.