SACX-backup

Backup for SACX keyword extraction pipeline

NOTICE SEPARATE LUMI BRANCH FOR RUNNING THIS ON LUMI!!!

What it does?

• Runs multiple multilabel classification models and uses the Integrated Gradients method to explain the results of the classifier.
  • i.e. IG method scores words based on their relevance in the classification
  • We're trying to see which words consistently score high -> keywords
• Aggregates the results over all trained models and produces lists of keywords for each class in the classification.

How to run?

• run train_and_explain.py with different seed values:
  • with Slurm: run sbatch sl-train-explain.bash <insert seed value here> with 10+ different seed values.
  • without Slurm: run python train_and_explain.py --seed=<insert seed value here> with 10+ different seed values.
  • Create folder "explanations" beforehand, unless you define the parameter --explanations=<new path> to point somewhere else.
  • you can modify these parameters on the command line/slurm script or in the file arguments.py.
• run kws.py with the results produced by training-explaining.
• If needed, corpus information can be used. In that case, before kws.py, run count_class_words.py which calculates term and document frequencies.
  • give the result as a parameter --corpus_information to kws.py and a limit for term frequency as --min_word_freq

Multilingual?

• you can give multiple languages as --language='["en", "fr", "zh"]' etc.
• trains with all languages, explains them separately.
• for laguages that do not use white space to separate words, you can use a spacy-parser by defining for example --parse_separately="zh" and --parser_model="zh_core_web_md"
  • in this case, feed the languages in separately, so that only one parser is loaded at a time. I.e. if there is only one language that needs separate parsing, this can be done:

python train_and_explain.py --language='["en", "fr", "jp"]' <other params> --parse_separately="jp" --parser_model="jp_core_news_md"

• ... but if you have, say "zh" and "jp", then they need to be explained separately:

python train_multilabel.py --language='["jp", "zh"]' --save_model=<path to model>
python explain_multilabel.py --trained_model=<path to model> --language='["jp"]' --parse_separately="jp" --parser_model="jp_core_news_md"
python explain_multilabel.py --trained_model=<path to model> --language='["zh"]' --parse_separately="zh" --parser_model="zh_core_web_md"