In order to download more captions, you need a YouTube API key and pip install the following:
pip install pandas
pip install youtube_transcript_api
pip install google-api-python-clientYou can get the YouTube API key from here, and then use it to populate DEVELOPER_KEY in config.py
To request transcripts from a youtube channel, populate CHANNEL_NAME in config.py, and run python prepare_data.py.
The script will do a search for this channel name and come back with a list of potential channel_ids, suggesting the top-most. Please click on weblink to confirm that you want to proceed with this channel. Note that all data return from calls to the YouTube API, both directly via the googleapiclient and indirectly via youtube_transcript_api are backed up incrementally to reduce the likelihood that you have to re-request the same data and risk hitting your daily request limit (but you still don't want to waste a bunch of time on the wrong channel).
See a colab notebook example of how to scrape a new YouTube channel or update an existing one.
The final result of running prepare_data.py is a panda's DataFrame saved in json that contains transcripts and metadata for any video on that channel that has (in the requested language):
The differences between these two transcripts will be labeled in a way that is non-destructive to the data, and hopefully flexible enough to be repurposed into more useful labels.
For example the tokens that are mutually different between the two transcripts (rather than a one-sided insertion) can be labeled 1 with all other tokens as 0, and then trained on a token-level classification task to recognize such 'errors' in auto-generated transcripts. Subsequently, the 1's on these tokens can be replaced with a <MASK> and trained in a language modeling task, where the 'correct' token for the masked word could come from the 'manually corrected' transcript, rather than the auto-generated transcript. Here the goal is for a language model to learn to fill in a suitable alternative word, when an incorrect word is masked out of a caption.
See the colab notebook checkout_data_and_new_label_creation.ipynb for more details on the dataset and how to initially prepare it for such a training task.
A portion of this dataset has been contributed to Hugging Face's Datasets library for ease of use, with details here.
One example application that was prepared as part of a MLC Research Jam presentation is token classification of the masked out YouTube errors (and their error-type). See a colab notebook example of this token classification model.
Finally, check out slides for the MLC presentation, which also provides an overview for this dataset as well.