Abstract
Sports game summarization aims at generating sports news from live commentaries. However, existing datasets are all constructed through automated collection and cleaning processes, resulting in a lot of noise. Besides, current works neglect the knowledge gap between live commentaries and sports news, which limits the performance of sports game summarization. In this paper, we introduce K-SportsSum, a new dataset with two characteristics: (1) K-SportsSum collects a large amount of data from massive games. It has 7,854 commentary-news pairs. To improve the quality, K-SportsSum employs a manual cleaning process; (2) Different from existing datasets, to narrow the knowledge gap, K-SportsSum further provides a large-scale knowledge corpus that contains the information of 523 sports teams and 14,724 sports players. Additionally, we also introduce a knowledge-enhanced summarizer that utilizes both live commentaries and the knowledge to generate sports news. Extensive experiments on K-SportsSum and SportsSum datasets show that our model achieves new state-of-the-art performances. Qualitative analysis and human study further verify that our model generates more informative sports news.
Introduction
Generating professional news articles from live sports commentaries is a complex task that goes beyond standard text summarization. Unlike general document summarization, sports reporting involves transforming informal, real-time play-by-play descriptions into structured, formal narratives.
Researchers identify three primary hurdles in this domain:
- The Length Challenge: Live commentaries are extensive, often exceeding the input limits of standard transformer models.
- The Style Gap: Commentaries are colloquial and reactive, while news articles are formal and analytical.
- The Knowledge Gap: Professional news often includes background information—such as player statistics or team history—that is absent from the live commentary.
The K-SportsSum Dataset
To address the limitations of existing datasets, which often contain "noise" like advertisements and irrelevant hyperlinks, we introduce K-SportsSum. This dataset consists of 7,854 manually cleaned commentary-news pairs.
The cleaning process involves removing descriptions of other games, advertisements, and history-related descriptions that do not pertain to the current match. Furthermore, K-SportsSum includes a comprehensive Knowledge Corpus with structured data for 14,724 players and Wikipedia-aligned articles for 523 teams, providing the necessary external context for high-quality summarization.
Methodology: Knowledge Enhanced Summarizer (KES)
We propose a two-step Knowledge Enhanced Summarizer (KES) framework:
- Key Commentary Selection: A RoBERTa-based selector identifies critical match events (e.g., goals, cards) within a sliding window, filtering out filler commentary.
- Knowledge Retrieval: Named Entity Recognition (NER) identifies players and teams, which are then linked to the Knowledge Corpus to retrieve relevant background facts.
- Knowledge-Aware Rewriting: A multilingual T5 (mT5) model generates the final news sentences. We utilize Segment Embeddings to distinguish between entity types and Knowledge Embeddings to inject the retrieved external facts.
Results and Discussion
Quantitative evaluation shows that KES achieves a ROUGE-L score of 47.17 on K-SportsSum, significantly outperforming previous state-of-the-art models. Human evaluations further confirm that KES produces more informative and fluent news. By bridging the knowledge gap, the model can enrich summaries with details like a player's physical attributes or team nicknames, moving closer to professional journalistic standards.
Conclusion
K-SportsSum and the KES model represent a significant step forward in automated sports reporting. By focusing on manual data quality and external knowledge integration, we demonstrate that summarization in specialized domains is as much about "enrichment" as it is about compression. Future work will explore the use of long-context transformers and explicit knowledge graph utilization to further reduce factual hallucinations.
Summary updated