- Authors: Yuwei Bao, Keunwoo Peter Yu, Yichi Zhang, Shane Storks, Itamar Bar-Yossef, Alexander De La Iglesia, Megan Su, Xiao Lin Zheng, Joyce Chai
- Organization: University of Michigan, Computer Science and Engineering
- Published in: EMNLP 2023, Singapore
- Links: Arxiv, Github, Dataset
WTaG is human-human task guidance dataset with natural language communications, mistake and corrections, and synchronized egocentric videos with transcriptions. WTaG is richly annotated with step detection, user and instructor dialog intention, and mistakes.
- #Videos Length: 10 hours
- #Videos: 56
- #Recipes: 3
- #User: 17
- #Instructor: 3
- #Total Utterances: 4233
- Median Video Length: 10 mins
- Synchronized user and instructor egocentric video + audio transcriptions
- Annotation: Step detection
- Annotation: Human filtered audio transcriptions
- Annotation: User utterrance dialog intention
- Annotation: User mistakes
- Annotations: Instructor utterrance dialog intention (+ detailed)
Despite tremendous advances in AI, it remains a significant challenge to develop interactive task guidance systems that can offer situated, personalized guidance and assist humans in various tasks. These systems need to have a sophisticated understanding of the user as well as the environment, and make timely accurate decisions on when and what to say. To address this issue, we created a new multimodal benchmark dataset, Watch, Talk and Guide (WTaG) based on natural interaction between a human user and a human instructor. We further proposed two tasks: User and Environment Understanding, and Instructor Decision Making. We leveraged several foundation models to study to what extent these models can be quickly adapted to perceptually enabled task guidance. Our quantitative, qualitative, and human evaluation results show that these models can demonstrate fair performances in some cases with no task-specific training, but a fast and reliable adaptation remains a significant challenge. Our benchmark and baselines will provide a stepping stone for future work on situated task guidance.
- User intent prediction: Dialog intent of user’s last utterance, if any (options).
- Step detection: Current step (options).
- Mistake Existence and Mistake Type: Did the user make a mistake at time t (yes/no). If so, what type of mistake (options).
- When to Talk: Should the instructor talk at time t (yes/no).
- Instructor Intent: If yes to 1, instructor’s dialog intention (options).
- Instruction Type: If yes to 1 and intent in 2 is “Instruction”, what type (options).
- Guidance generation: If yes to 1, what to say in natural language.
To evaluate the three methods on WTaG:
python src/pipeline.py [-h] --MTYPE MTYPE --in_path IN_PATH --video_list VIDEO_LIST --out_path
OUT_PATH
optional arguments:
-h, --help show this help message and exit
--MTYPE MTYPE, -t MTYPE
Vision to Language Method Type: ["lanOnly", "blip2", "objDet"]
--in_path IN_PATH, -i IN_PATH
Video input path
--video_list VIDEO_LIST, -l VIDEO_LIST
Path to the file with a list of videos to be processed
--out_path OUT_PATH, -o OUT_PATH
Guidance output path
@inproceedings{bao-etal-2023-foundation,
title = "Can Foundation Models Watch, Talk and Guide You Step by Step to Make a Cake?",
author = "Bao, Yuwei and
Yu, Keunwoo and
Zhang, Yichi and
Storks, Shane and
Bar-Yossef, Itamar and
de la Iglesia, Alex and
Su, Megan and
Zheng, Xiao and
Chai, Joyce",
editor = "Bouamor, Houda and
Pino, Juan and
Bali, Kalika",
booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2023",
month = dec,
year = "2023",
address = "Singapore",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2023.findings-emnlp.824",
doi = "10.18653/v1/2023.findings-emnlp.824",
pages = "12325--12341",
}
Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change. For further questions, please contact yuweibao@umich.edu.