Learning a game commentary generator with grounded move expressions

Abstract

This paper describes a machine learning-based approach for generating natural language comments on Shogi games. We generate comments by using a discriminative language model trained with a large amount of Shogi game records and comments made by human experts. Central to our method is accurate mapping of move expressions appearing in experts' comments to game states (i.e. positions) of Shogi, because the discriminative language model is trained with textual expressions paired with corresponding Shogi positions. We describe how such mapping can be performed by using evaluation information obtained from a Shogi program. Experimental results show that we can actually generate helpful comments for some positions.