Research on multimodal learning, language models, and data-centric AI.
A benchmark and taxonomy for phonemizing noisy user-generated text, paired with a compositional approach for irregular spellings, abbreviations, and code-mixing.
A retrieval-augmented dense video captioning approach that injects supervised saliency signals into both the retrieval and generation stages.
Tackles weakly-supervised dense video captioning through similarity-aware guidance and inter-caption augmentation.
Bridges the text-video modality gap by using generated captions as cross-modal context, enriched through stochastic augmentation during training.
A dense video captioning framework that reweights video frames by saliency and adaptively retrieves relevant captions at inference time.
Refines noisy synthetic image-caption datasets through a one-to-many mapping that re-aligns each image with its best-matching captions.
