Scholarlink

Visual Question Answering (VQA) requires models to understand both images and natural language questions to provide accurate answers. Existing VQA techniques are having language bias issue with predictions mostly depending on linguistic priors rather than visual grounding. By adding an extra Image-Text Matching (ITM) loss and CLS-based contrastive learning to the LXMERT architecture, this research work offers a novel way to reduce language bias. Our approach enhances visual grounding and induces alignment between image-question pairings by directly applying contrastive loss and using ITM signals to reinforce the model. The proposed model demonstrates improved bias reduction performance while maintaining competitive accuracy compared to existing VQA methods. Specifically, it achieves a bias rate of 24.28% on the VQAv2 dataset, outperforming recent baselines like CSS and RUBi during testing on a subset of the VQA v2 dataset of samples, demonstrating both quantitative improvement and qualitative resilience. This approach provides a scalable and interpretable framework that reduces biases in VQA, increasing the reliability and equity of vision-language systems.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Bias Reduction in Visual Question Answering Using CLS-Based Contrastive Learning and ITM Loss

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Pooja Singh,
  • Gourav Bathla

摘要

Visual Question Answering (VQA) requires models to understand both images and natural language questions to provide accurate answers. Existing VQA techniques are having language bias issue with predictions mostly depending on linguistic priors rather than visual grounding. By adding an extra Image-Text Matching (ITM) loss and CLS-based contrastive learning to the LXMERT architecture, this research work offers a novel way to reduce language bias. Our approach enhances visual grounding and induces alignment between image-question pairings by directly applying contrastive loss and using ITM signals to reinforce the model. The proposed model demonstrates improved bias reduction performance while maintaining competitive accuracy compared to existing VQA methods. Specifically, it achieves a bias rate of 24.28% on the VQAv2 dataset, outperforming recent baselines like CSS and RUBi during testing on a subset of the VQA v2 dataset of samples, demonstrating both quantitative improvement and qualitative resilience. This approach provides a scalable and interpretable framework that reduces biases in VQA, increasing the reliability and equity of vision-language systems.