Toward Collaborative Game Commentating Utilizing Pre-Trained Generative Language Models

Abstract

In this paper, we propose a novel task of collaborative game commentating, an artificial intelligence agent capable of collaboratively commentating with a human commentator in Live-Streaming of Esports. To this end, we propose a collaborative game commentating system that employs a pre-trained language model trained using commentaries by professional commentators, along with metadata including title and tags. The conducted experiments show that (1) fine-tuned Text-to-Text Transfer Transformer (T5) model, a state-of-the-art generative language model, could produce more clearer and precise commentary and better recall the words from the reference commentary, as it effectively improves the scores on evaluation metrics that are widely used for concise text generation task after tuning the model. (2) The more information used for the current method fusion of information, the clearer and more precise generated commentary is. However, it performs worse to recall the words from reference commentary.