Chan, Anthony Hing-HungAnthony Hing-HungChanSiu, Wan ChiWan ChiSiuChan, Sin WaiSin WaiChanChan, Cheuk YiuCheuk YiuChanHui, Chun ChuenChun ChuenHui2023-05-252023-05-252022https://repository.sfu.edu.hk/handle/sfu/3809We watch soccer games because of the excitement, the skill of players, the brand name of soccer teams, etc. However, contribution of the commentary is also indispensable. At this stage of our technology development, it is good to have automatic commentary to be provided by our computer. This paper is on the production of soccer commentary automatically making use of hi-tech and deep learning. There are many aspects on the production. However, we just concentrate on the production of key commentaries. We make use of spatial-temporal representation with 2 stages of operations to design our transformer network. This involves Temporal-Grouped Attention that dynamically separates and groups all channels together for the advantage of extracting temporal domain features, Local-Global Mixed Attention that enforces the model to find relationship between local and global feature representations in one attention structure, and Selective Feature Aggregation to intelligently select the final weights between the two attention networks. Our results provides key commentaries in line with other state-of-the-art approaches but with high potential to be further developed to real-time commentary system.enconference proceedings10.1109/TENCON55691.2022.9978078