ViTexQA: A Multi-Frame Temporal Perception Dataset for Video Text Question Answering

Researchers have introduced two new datasets and a model to improve video text understanding and multilingual healthcare AI systems. The datasets address limitations in existing resources for video-text QA and multilingual healthcare environments.

The ViTexQA dataset, presented on arXiv^[1], is a large-scale video-text QA dataset built using a Chain-of-Thought (CoT) annotation pipeline with temporal constraints. It requires cross-frame text fusion to answer questions, enforcing true temporal reliance. FrameThinker, a model for multi-frame temporal reasoning, uses two-stage training with CoT-Guided Supervised Fine-Tuning (SFT) and Temporally-grounded Reinforcement Learning (RL). The method outperforms SOTA baselines on ViTexQA, lifting ROUGE-L by 6.3%^[1]. A separate study on arXiv^[2] introduced a multilingual hematology visual question answering dataset to address the limitation of existing English-centric resources in multilingual healthcare environments. The dataset includes 110K bilingual question-answer pairs for 20K leukemic and normal single-cell images^[2]. The dataset is clinically validated and constructed using morphology-aware annotations from LeukemiaAttri and WBCAtt datasets. It is supported by a domain-specific Urdu hematology dictionary to ensure linguistic consistency and clinical correctness^[2].