Tree-of-Report: Table-to-Text Generation for Sports Game Reports with Tree-Structured Prompting

Shang-Hsuan Chiang, Tsan-Tsung Yang, Kuang-Da Wang, Wei-Yao Wang, An-Zi Yen, Wen-Chih Peng

Keywords: data-to-text generation, inference methods, automatic evaluation

TL;DR: We propose Tree-of-Report, a novel framework that uses tree-structured prompting to improve table-to-text generation in sports game reports, showing its effectiveness and efficiency across multiple datasets.

Abstract: Generating sports game reports from structured table data is a challenging table-to-text generation task that requires balancing structured data comprehension with narrative storytelling. While model-based approaches demand large training datasets, prompt-based methods with large language models (LLMs) often suffer from hallucination issues due to poor table comprehension. To address these challenges, we propose Tree-of-Report, a novel framework that divides the generation process into three stages: Content Planning, Operation Execution, and Content Generating. Our method decomposes large tables into smaller sub-tables using a hierarchical tree structure, enabling more effective table comprehension. Additionally, it merges and rewrites texts to produce more detailed and coherent long-form outputs. Experimental results on the RotoWire, MLB, and ShuttleSet+ datasets show that Tree-of-Report outperforms existing prompt-based baselines with relatively lower time and cost, demonstrating its advantage in both effectiveness and efficiency. In summary, this work sets a new precedent for prompt-based table-to-text generation in sports game reports.

Archival Status: Non‑archival

Paper Length: Long Paper (up to 8 pages of content)

Submission Number: 276