SkillTGR: A New Framework for Operation-wise Table Question Answering
A research paper introduces Skill-augmented Table Graph Reasoning (SkillTGR), a framework for operation-wise Table Question Answering (TableQA). The authors identify that existing methods treat all questions uniformly, obscuring LLMs' poor performance on complex operations like aggregation and arithmetic. They propose a fine-grained question taxonomy and release two datasets, WikiTQ-ow and TabFact-ow, for evaluation. SkillTGR addresses modeling bottlenecks by using a graph structure to preserve table integrity and leveraging reusable patterns across similar operations.
Key facts
- SkillTGR is proposed for operation-wise TableQA.
- Existing methods treat all questions uniformly, hiding LLM struggles with complex operations.
- A fine-grained question taxonomy is introduced.
- Two datasets, WikiTQ-ow and TabFact-ow, are released.
- SkillTGR uses a graph structure to avoid the 'lost-in-the-middle' issue.
- The framework leverages reusable patterns across similar operations.
- The paper is from arXiv with ID 2607.22633.
- The approach aims to improve reasoning on complex table operations.
Entities
Institutions
- arXiv