Event-Grounded Football News Generation from Match Videos with Parameter-Efficient Large Language Models

Abstract

Automated football news generation from raw videos requires bridging spatiotemporal perception with factual text composition. This study develops an end-to-end, event-based framework that converts match videos into fact-grounded reports. The framework uses an Inflated Three-Dimensional ConvNet (I3D) backbone with multi-scale temporal context aggregation at 5-, 15-, and 30-second intervals to identify key events, including goals, cards, substitutions, and shots. The recognized events are structured into JavaScript Object Notation (JSON) logs, which serve as a traceable event interface. For report generation, Llama 3-8B is adapted using Low-Rank Adaptation (LoRA), a parameter-efficient fine-tuning (PEFT) strategy that enables the model to acquire football terminology and news logic with limited computational overhead. To ensure editorial safety, the framework incorporates score-state verification, atomic fact validation, and manual review for low-confidence detections. Experiments on SoccerNet-v2 show the I3D backbone achieves a weighted F1 score of 90.60%, while the PEFT-adapted large language model reaches a BERTScore of 0.925. By integrating confidence filtering and factual verification, the Event-to-text F1 score reaches to 0.872, and the unsupported statement rate decreases to 4.8%. These results demonstrate that the systems primary value lies in its explicit event interface and traceable constraint mechanism rather than in developing a novel video backbone. This framework provides an auditable engineering pathway for automated sports journalism, balancing factual grounding with parameter-efficient deployment.