Event-Centric Augmentation for Video Question Answering

Video question answering (VideoQA) seeks reliable answers by temporally grounding question-relevant events and modeling their complex semantics. However, the scarcity of question-answer (QA) annotations and the absence of event-span supervision hinder the understanding of complex videos, leading to unreliable predictions. Meanwhile, the inherent supervisory potential within these rich video semantics has been largely overlooked. To address these challenges, we propose a novel plug-and-play, event-centric augmentation framework that strengthens both event understanding and grounding for end-to-end VideoQA models. Specifically, an SQA generation module automatically synthesizes diverse event annotations (Span-Question-Answer triplets) for the sampled event segments, which are then refined by filtering out misaligned or weakly correlated ones via the designed SQA cleaning module. Additionally, we attach pseudo event spans to the original QA pairs and merge them with the synthesized annotations. Benefiting from the augmented event-level spans, we introduce a simple yet effective weakly supervised grounding paradigm that explicitly encourages the model to explore and capture question-critical events, improving the temporal sensitivity of event clues and the trustworthiness of the predicted answer. Extensive experimental results demonstrate that our framework significantly improves both answer accuracy and critical event grounding compared to existing baselines on standard datasets. The source codes are released athttps://github.com/XizeWu/EcA