Skip to content

Latest commit

 

History

History
63 lines (50 loc) · 2.72 KB

README.md

File metadata and controls

63 lines (50 loc) · 2.72 KB
Raw

causal-text

This is an implementation of the paper Causal Effects of Linguistic Properties.

It is a package for computing the causal effects of text. Concretely this means algorithms for quantifying the degree of influence some user-defiend property (E.g. sentiment, respect) has on an outcome (email reply time), while controlling for potential confounds (topic, etc).

Setup

pip install -r requirements.txt

Quickstart

Run the full TextCause algorithm on a simulated dataset.

python main.py --run_cb

Run the full TextCause algorithm on a dataset of your choosing:

python main.py --run_cb --data /path/to/your/data.tsv

Usage Details

  • Prepare your data. This system expects a TSV file, with columns
    • text: string, the text you're studying
    • Y: int, binary outcome of interest
    • C: int, categorical confounder
    • T_proxy: int, your binary treatment indicator, e.g. the output of a classifier or lexicon
    • T_true (optional): int, binary indicator for the "true" (i.e. non-predicted) treatment
  • Run the system.
    • For python main.py --data /path/to/your/data.tsv --no_simulate
    • And if you want to run BERT for text adjustment (i.e. the full TextCause algorithm): python main.py --data /path/to/your/data.tsv --no_simulate --run_cb
    • If you want to run the simulation: python main.py --run_cb
  • Look at your results. When finished, the system will print out all of its hyperparameters and a bunch of different ATE estimates. The estimators are:
    • unadj_T: the unadjusted effect of T
    • ate_T: backdoor-adjusted effect of T
    • unadj_T_proxy: the unadjusted effect of T proxy
    • ate_T_proxy: backdoor-adjusted effect of T proxy
    • ate_matrix: matrix-adjusted effect of T_proxy using the measurement model P(T_hat | T)
    • ate_T_plus_reg: backdoor-adjusted effect of a boosted T
    • ate_T_plus_pu: backdoor-adjusted effect of a boosted T, but where the label improvement comes from a one-class classifier instead of a logistic regression
    • ate_cb_T_proxy: text-adjusted ATE estimate
    • ate_cb_T_plus: the full TextCause algorithm; test adjustment + T boosting.

Citation

Please cite this paper if you make use of the repo:

@inproceedings{pryzant2021causal,
 author = {Pryzant, Reid and Card, Dallas and Jurafsky, Dan and Veitch, Victor and Sridhar, Dhanya},
 booktitle = {Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL)},
 link = {https://nlp.stanford.edu/pubs/pryzant2021causal.pdf},
 title = {Causal Effects of Linguistic Properties},
 url = {https://nlp.stanford.edu/pubs/pryzant2021causal.pdf},
 year = {2021}
}