personalized-nlp

Personalized prediction applied to various subjective natural language processing (NLP) tasks

Download data

To download data, enter personalized_nlp folder and type in:

dvc pull

How to run experiments:

First, preprocess selected dataset with scripts.process_data pipeline, to assigns folds to texts:

python -m scripts.process_data --annotations_df_path .../personalized-nlp/storage/data/unhealthy_conversations/uc_annotations.csv --texts_df_path .../personalized-nlp/storage/data/unhealthy_conversations/uc_texts.csv --annotator_col annotator_id --num_folds 5 --text_col text_id

Then, you can run experiments with:

python -m personalized_nlp.experiments.uhnealthy

How to add DataModule for a new dataset

Copy one of the existing dataset classes (personalized_nlp/datasets/) and modify paths and settings. Next, copy one of the experiments (personalized_nlp/experiments/) and customize the settings.

How to select folding setup

Set the stratify_folds_by argument in datamodule: None for standard train-val-test split, 'users' for users/past-present-future1-future2 split and 'texts' for texts folds.