Tree-of-Text Framework Boosts Sports Report Generation from Tables
A new prompting framework called Tree-of-Text improves table-to-text generation for sports game reports. Proposed on arXiv, it uses a tree structure to guide large language models through three stages: content planning, operation execution, and content generation. The method outperforms existing approaches on ShuttleSet+ and excels in RG and CO metrics on RotoWire-FG, as well as CS and CO on other datasets. It addresses hallucination and weak table comprehension common in prompt-based methods.
Key facts
- Tree-of-Text is a tree-structured prompting framework
- It targets table-to-text generation in the sports domain
- The framework uses a three-stage process: Content Planning, Operation Execution, Content Generation
- It outperforms existing methods on ShuttleSet+ dataset
- Leads in RG and CO metrics on RotoWire-FG
- Excels in CS and CO on other datasets
- Addresses hallucination and weak table comprehension in LLMs
- Published on arXiv with ID 2604.26501
Entities
Institutions
- arXiv