Natural Language Generation for Automated Sports Broadcasting

This thesis addresses the design and implementation of an intelligent system built on a Retrieval-Augmented Generation (RAG) framework coupled with a Large Language Model (LLM), engineered to autonomously generate football match commentary from minimal structured input. The architecture fuses advanced neural language generation with dynamic retrieval of player- and team-specific statistics, enabling the production of detailed and contextually relevant narratives. The approach encompasses the development of an event annotation framework and a user-friendly interface that enables users to select match events and provide relevant contextual details. A thoughtfully engineered prompting strategy, complemented by few-shot examples, directs the LLM to generate coherent and contextually precise commentary while maintaining factual integrity, including information such as goal scorer, assist provider, type of shot, and event timing. Evaluation covers both quantitative metrics-such as accuracy and coverage of events-and qualitative measures, including human judgments of clarity, informativeness, and narrative quality. The results indicate that the RAG-based LLM system achieves an event-level accuracy of 99% and that, based on human evaluations, the generated commentaries were rated at a level comparable to human-written ones. This architecture therefore provides a flexible and scalable solution for automated sports commentary, with potential applications in live broadcasting, online platforms, and as a supportive tool for human commentators.