[3Xin4] Poster session 1

[3Xin4-58] News Article Reading Time Estimation by Multimodal Machine Learning

〇Shotaro Ishihara¹, Yasufumi Nakama (1.Nikkei Inc.)

Keywords:Multimodal, Machine learning, Deep learning, Reading time

The page views are widely used for quantitative evaluation of contents such as articles and advertisements, but reading time is a more detailed metric for understanding user preferences. Since we can observe the degree of the reading activities from reading time, it is useful for more sophisticated content recommendation and analysis. This paper emphasizes the importance of reading time in the news media and discusses implementation methods for the prediction. The simplest implementation is based on the hypothesis that reading time correlates with the length of the body text. However, the analysis of actual users of Japanese financial news revealed that reading time did not strongly correlate with the length of the body text. Furthermore, our experiments showed that we can construct more accurate machine learning models by using the information from multiple modalities, such as article headlines and text, thumbnail images, and the user’s past reading time. We also discussed the performance degradation caused by missing modalities for practical use.