Skip to content

Code for "Evaluating Explainable AI: Which Algorithmic Explanations Help Users Predict Model Behavior?"

License

Notifications You must be signed in to change notification settings

peterbhase/InterpretableNLP-ACL2020

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

9 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Evaluating Explainable AI

This is the codebase for "Evaluating Explainable AI: Which Algorithmic Explanations Help Users Predict Model Behavior?" (ACL 2020) [pdf]

Repository Structure

|__ text/ --> Directory with all text models and experimental scripts
        |__ data/ --> includes task data and simulation test data for conducting tests
        |__ anchor/ --> Anchor code from original authors
        |__ lime/ --> LIME code from original authors
        |__ saved_models/ --> training reports for task and prototype models
        |__ src/
            |models/ --> code for neural task model and prototype model, including  decision boundary and prototype explanation methods 
            |classes/ --> supporting classes and utility functions for explanations
            |*/ --> directories for supporting classes including network layers and data loaders
        |__ figure_examples.py --> script for generating example explanations used in paper figures
        |__ gather-experiment-data.py --> script for gathering simulation test data
        |__ nearest-neighbors.py --> script for finding nearerest neighbors to prototypes
        |__ run_tagger.py --> script for evaluating classifier accuracy
        |__ requirements.txt --> package requirements for text experiments       
|__ tabular/ -->
        |__ data/ --> includes task data and simulation test data for conducting tests
        |__ anchor/ --> Anchor code from original authors
        |__ saved_models/ --> training reports for task and prototype models
        |__ src/
            |models/ --> code for neural task model and prototype model, including  decision boundary and prototype explanation methods 
            |classes/ --> supporting classes and utility functions for explanations
            |*/ --> directories for supporting classes including network layers and data loaders
        |__ gather-experiment-data.py --> script for gathering simulation test data
        |__ nearest-neighbors.py --> script for finding nearerest neighbors to prototypes
        |__ run_tagger.py --> script for evaluating classifier accuracy
        |__ requirements.txt --> package requirements for tab experiments

results_analysis.Rmd --> R markdown file that computes all empirical/statistical results in paper

Requirements

  • Python 3.6
  • see 'requirements.txt' in each subdirectory for data domain specific requirements

Reproducing Experiments

  1. Set-up: Install the requirements. Run python -m spacy download en_core_web_lg. Download the 840B.300d glove embeddings from https://nlp.stanford.edu/projects/glove/.

  2. Task models: Training both prototype and blackbox models can be done with the main.py scripts in the text and tabular directories. For text data, provide the glove path as --emb-fn <file_path>. See training reports in each saved_model directory for training arguments for hyperparameters to the prototype models. By default, the script trains blackbox models.

  3. Simulation test data: Simulation test data is collected with gather-experiment-data.py in either directory, using trained neural and prototype models.

  4. Statistical results: results_analysis.Rmd computes all empirical/statistical analysis in the paper

Citation

@inproceedings{hase_evaluating_2020,
    title={Evaluating Explainable AI: Which Algorithmic Explanations Help Users Predict Model Behavior?},
    author={Peter Hase and Mohit Bansal},
    booktitle={ACL},
    url={https://arxiv.org/abs/2005.01831},
    year={2020}
}

About

Code for "Evaluating Explainable AI: Which Algorithmic Explanations Help Users Predict Model Behavior?"

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published