Game commentary is an open‑ended generation task that demands multimodal perception, strategic reasoning, and contextual knowledge. Existing research on AI‑generated game commentary (AI‑GGC) is fragmented across different games, modalities, and evaluation protocols, and conventional overlap‑based or holistic metrics fail to capture the functional heterogeneity of commentary.
We introduce GameCommBench, a unified benchmark that spans board games, sports, and esports. Commentary in the benchmark is aligned with heterogeneous game contexts and annotated by commentary type, providing a structured evaluation resource.
Building on this, we propose Type‑Aware Commentary Evaluation (TACE), a framework that categorizes commentary into types such as live observation, narration, and strategic analysis, and assigns dedicated metrics and scoring criteria to each type. Reliability tests and human agreement studies confirm the validity of TACE.
Using TACE to benchmark several representative AI commentators reveals non‑uniform capability profiles: real‑time observation and deep strategic analysis emerge as major bottlenecks, while aspects like emotional rendering and post‑match recap perform comparatively better.
Together, GameCommBench and TACE establish a comparable and interpretable foundation for AI‑GGC evaluation, offering diagnostic insights for future model improvements.
Review