GQA-Q2Q: A Large-scale Benchmark for Entity Ambiguity Resolution in Visual Question Answering
Abstract
Vision-language models achieve remarkable results on various visual question-answering benchmarks. However, they struggle with ambiguous questions where the target entity is unclear. To address this issue, it is essential to construct a benchmark dataset that pairs ambiguous questions with entity-grounded clarifying sub-questions. This paper presents GQA-Q2Q, a large-scale dataset for resolving entity ambiguity in visual question answering via entity-grounded clarifying sub-questions. GQA-Q2Q is constructed through a template-assisted human annotation pipeline. The pipeline generates candidate sub-questions from attribute, relation, and position templates and asks human annotators to verify, select, or revise them. As a result, GQA-Q2Q contains 46,554 human-verified clarifying sub-questions with target-instance grounding. This paper also presents a proposed ambiguity-resolving VQA model that leverages the generated sub-questions before producing final answers. The experimental results show that the proposed model outperforms zero-shot and ambiguity-resolving vision-language baselines. The code and dataset are available at https://github.com/gminipark/GQA-Q2Q.git.