Automated Game Commentary Generator

Hirotaka Kameko^†, Shinsuke Mori^‡, Yoshimasa Tsuruoka^†
†The University of Tokyo, ^‡Kyoto University

Overview

We are aiming to build a system that can automatically generate a commentary for a Shogi game. Shogi (将棋) is a chess-like board game and is also known as Japanese chess. For detail rules, please refer to Wikipedia

The task mainly consists of two problems.

• How to decide what to comment, such as positions, moves, or the result of searching.
• How to comment it.

We propose a machine learning-based approach for generating comments. We generate comments by using a discriminative language model trained with a large amount of game records and comments by human experts.

Besides, comments of Shogi usually consist of textual descriptions and move expressions such as "P-7f" (７六歩) and "Bx8h+" (８八角成). These move expressions are generated not by a language model but by a game tree. We also propose a method that maps these move expressions and game states of Shogi.

Demo

Our system generates comments for Shogi games of professional players in real time.

Selected Results

Move: P-4f

Word Value 腰掛け (reclining) 0.99 指 (action) 0.99 方 (plan) 0.99 銀 (silver) 0.98

Generated: ▲４六歩から腰掛け銀を目指す指し方もあるところだが、 (trans: Black can play P-4f and aim for reclining-silver strategy, but...)

Tools

Publication

Hirotaka Kameko, Shinsuke Mori, Yoshimasa Tsuruoka.
Learning a Game Commentary Generator with Grounded Move Expressions
Proceedings of 2015 IEEE Conference on Computational Intelligence and Games (2015 IEEE CIG), [paper] [slide] (to appear).