zjuchenlong / youwikihow

YouwikiHow dataset for weakly-supervised article grounding

YouwikiHow

The repo contains the YouwikiHow dataset and original scripts to build this dataset.

Download the Processed Datasets

Dataset layout

YouwikiHow/
|   +-- annotations/
|       +-- wikihow_data.pkl
|       +-- crosstask_test.pkl
|   +-- features/
|       +-- train.csv
|       +-- test.csv
|       +-- train_s3d_features_from_ht100m
|       +-- test_s3d_features_fr16_sz256_nf16

Training Set

The training set wikihow_data.pkl is a dictionary, key is the wikiHow ID, and value are:

• task_full_name: wikiHow task full name.
• task_text: wikihow articles.
  • simplified_headline_list: List of all high-level summary sentences.
  • simplified_article_list: list of all low-level articles.
  • head2sent: The sentence IDs mapping from high-level (key) sentences to low-level (value) sentences.
  • sent2head: The sentence IDs mapping from low-level (key) sentences to high-level (value) sentences.
• task_video: The list of all YouTube videos correspoinding to this wikiHow task.
• video_duration: The total duration of all YouTube videos.

Test Set

The test set crosstask_test.pkl is same format as the training set, with two more keys:

• step2headline: The mannually mapping between the CrossTask steps (in original datasets) and the wikiHow articles (i.e., headlines).
• crosstask_annotations: The ground-truth for evaluation. Keys are the video ids, and Values are all possible ground-truth annotations propagated from CrossTask.

Download Visual Features

OneDrive

• train_s3d_features_from_ht100m
• test_s3d_features_fr16_sz256_nf16

Build from Scratch (Original Datasets)

Download Original Datasets

• wikiHow: Replace the path in dataloader/wikiHow_text.py with corresponding wikiHow dataset path.
• CrossTask: Replace the path in dataloader/CrossTask.py with corresponding CrossTask dataset path.
• HowTo100M: Replace the path in datasetloader/HowTo100M.py with corrsponding HowTo100M dataset path.

YouwikiHow Preprocessing

1. Use manually rules to filter HowTo100M tasks for the training set.

python task_selection.py

2. Save the manually mapped crosstask annotations as the training set

python test_set_crosstask_reader.py

3. Generate S3D features using S3D_Feature_Extractors with csv files in features/train.csv and features/test.csv.

Citations

@inproceedings{chen2022weakly,
  title={Weakly-Supervised Temporal Article Grounding},
  author={Chen, Long and Niu, Yulei and Chen, Brian and Lin, Xudong and Han, Guangxing and Thomas, Christopher and Ayyubi, Hammad and Ji, Heng and Chang, Shih-Fu},
  booktitle={Empirical Methods in Natural Language Processing (EMNLP), 2022},
  year={2022}
}

youwikihow's Issues

The visual training dataset is missing!

The visual training dataset on google drive is missing, and train_s3d_features_from_ht100m cannot be found, could your please share the link again?