Automatic Generation of Soccer Commentary from Play Data Using Rule-based Method

Abstract

Text commentary on sports websites is manually entered by sports data companies. It’s very convenient for those who cannot watch the live game but want to follow it in real time. However, making it requires a lot of work, especially in soccer, where a large number of events take place in a 90-minute period by players. As a labor-saving method, we propose an AI-based automatic text commentary generation system using Play data which has a time series of events made of numerical data. We compared the capability of two methods, an NN based method proposed in a previous study and a newly proposed Rule-based method, by Automatic evaluation and Human evaluation, respectively. The results show that the Rule based method significantly outperforms the NN-based method in terms of content selection, grammatical accuracy, and proper noun output, indicating the effectiveness for Data2Text task.