Live Commenting of Football Games Using Machine Learning and Natural Language Generation

Sammanfattning: In this degree project, automatic live comments of the events in a football game are generated in text format. The events are detected using machine learning, where CNNs are fit using audio recordings and player positions of games. Suitable features are extracted, where several models are fit to detect different types of events. The results indicate that the detections of sound powers and referee whistles are sensitive to the arena, where difficult to determine the event of the referee whistle. However, ongoing attacks are detected accurately. The detected events are commented using natural language generation, where the comments are generated using data-to-text generation. The results indicate that the complexity of the comments is sensitive to the information able to be extracted of the events.