LBF-VQA: Towards Language Bias-Free Visual Question Answering with Multi-Space Collaborative Debiasing

Visual Question Answering (VQA), aimed at improving AI-driven interactions and solving complex visual-linguistic tasks, has increasingly garnered attention as a pivotal research domain in both academic and industrial spheres. Despite progress in VQA, current studies still suffer from the challenge of language bias posed by spurious semantic correlations and minority class collapse, leading to semantic ambiguities and distribution shifts that hinder robust performance across challenging scenarios. To address these challenges, we propose a robust multi-space collaborative debiasing paradigm, termed “LBF-VQA” which systematically leverages multi-space collaborative debiasing strategies to achieve language bias-free VQA, encompassing both Euclidean space debiasing (ESD) and Spherical space debiasing (SSD). By strategically introducing bias-examples and their corresponding counter-examples, the ESD strategy focuses on uncovering hidden prior correlations and the complex interactions between modality and semantics within the Euclidean space. Benefiting from the infinite contrastive and distribution debiasing learning mechanisms, the SSD strategy is devoted to effectively preventing the collapse of minority classes while enhancing the manifold representations of instance de-bias and distribution de-dependence in the Spherical space. Furthermore, we meticulously constructed a specialized medical dataset intentionally embedded with deliberate language bias to comprehensively examine the negative effects of language bias on medical VQA systems. Extensive experiments on multiple general and medical VQA benchmarks consistently verify the effectiveness and generalizability of our LBF-VQA in handling various complex VQA scenarios than state-of-the-art baselines.