A fresh look at sports news generation

Sports game summarization, the process of creating news from live commentaries, has a few challenges. Current methods rely on automated data collection, which leads to a lot of inconsistencies. Moreover, the gap between live commentaries and the finished sports news isn’t always bridged effectively, hurting the quality of summaries. To tackle this, a new dataset, K-SportsSum, has been introduced. This dataset stands out for two reasons. First, it includes a vast amount of data, with 7, 854 pairs of commentaries and news articles. To ensure high quality, K-SportsSum uses a manual cleaning process. Second, it closes the knowledge gap by offering a large-scale knowledge corpus. This corpus includes information on 523 sports teams and 14, 724 players.

Alongside the dataset, there's a knowledge-enhanced summarizer. This tool uses both live commentaries and the knowledge corpus to generate sports news. When tested on K-SportsSum and another dataset called SportsSum, the model showed impressive results. It even outperformed previous state-of-the-art models. Human evaluations also confirmed that the generated news was more informative.

Actions