DIAGRAMS: A Framework for Reasoning-Level Attribution in Diagram QA
The recently introduced DIAGRAMS framework tackles reasoning-level attribution in diagram question answering (QA) by associating each question-answer pair with all relevant visual areas necessary for answer derivation, rather than solely the final answer. This system separates interface logic from specific dataset JSON formats through an internal meta-schema and dataset adapters. It conducts evidence selection conditioned on QA and identifies necessary regions, creating missing QA pairs or potential regions for human review. Testing was conducted on six Diagram QA datasets, revealing that evidence suggested by the model enhanced the efficiency of annotations.
Key facts
- DIAGRAMS is a lightweight, schema-driven review framework for Diagram QA.
- It decouples interface logic from dataset-specific JSON structures.
- The system uses an internal meta-schema and dataset adapters.
- It performs QA-conditioned evidence selection and proposes required regions.
- When QA pairs or candidate regions are missing, it generates them.
- It supports human verification and refinement.
- The framework was tested across six Diagram QA datasets.
- Model-suggested evidence improves annotation efficiency.
Entities
Institutions
- arXiv