Implicit-VidSRL Dataset

• Paper: https://arxiv.org/abs/2505.21068
• Project: https://anilbatra2185.github.io/p/ividsrl/
• Curated by: Anil Batra, Laura Sevilla-Lara, Marcus Rohrbach, Frank Keller
• Language(s) (NLP): English
• License: CC-BY-4.0

Dataset Summary

Implicit-VidSRL is a benchmark for the understanding of procedural steps in instructional videos. The dataset uses semantic role labeling (SRL) to model the semantics of narrated instructional video as simple predicate-argument structures like {verb,what,where/with}. The dataset contains implicit and explicit arguments which can be infered from contextual information in multimodal cooking procedures.

Dataset Details

Dataset Sources

• Youcook2: http://youcook2.eecs.umich.edu/
• Tasty: https://cvml.comp.nus.edu.sg/tasty/

Dataset Fields

Each record in the dataset contains the following fields:

• video_id (str): Video Id from the source dataset.
• dataset (str): The name of the source dataset.
• title (int): The title of the recipe.
• duration (str): Duration of video in seconds.
• timestamps (List[List[float]]): The list of timestamps corresponding to each recipe step in the video. Each timestamp entry contains start and end timestamp.
• sentences (List[str]): The list of steps in the recipe. Each recipe step is natural text in English.
• srl (List[Dict]): The new annotation for each recipe step. Each list item corresponds to the recipe step with key enteries of verb/what/where containing ingredient entities. In addition it also has information about the implicit entities for each step.

Citation

@inproceedings{
    batra2025implicit,
    title={Predicting Implicit Arguments in Procedural Video Instructions},
    author={Anil Batra, Laura Sevilla-Lara, Marcus Rohrbach, Frank Keller},
    booktitle={The 63rd Annual Meeting of the Association for Computational Linguistics},
    year={2025}
}