Text Counterfactual Explanations

[drobiu]

Hi all!

I'm working on introducing Text Counterfactual Explanations for Language Model Classifier models to CounterfactualExplanations.jl. The method I'm focusing on is Relevance-based Infilling for Textual Counterfactuals (RELITC) Pyhton code, paper. In short, the method generates an explanation for a string (text) classified by an LM Classifier by computing feature attribution per token (score of how much each token contributed to classifying the text to its class), masking the tokens with the highest attribution scores, and filling in the masks using a fine-tuned Conditional Masked Language Model (CMLM).

To have a fully implemented version of RELITC I think we need to have the following (also somewhat tracked in this project):

• RELITC
  • Feature Attributions
    • For now using the Python package transformers-interpret with PythonCall
  • Fine-Tuning for LM Classifier
  • Fine-Tuning for CMLM
  • Masking top-K% tokens
  • Infilling masked tokens
    • Left-to-Right infilling
    • Uncertainty-based infilling
  • Beam search for best K parameter
    • Fluency score
    • Edit score

I'm working on those features in a separate branch, with this PR: #413 where I'm still working in a Jupyter Notebook, but I'm planning to introduce the features to the CE.jl architecture.

The generate_counterfactual(x, target, data, M, generator) can be used in the following way:

• x being the text(s) to explain
• target being the target class for the CE
• data being optional data if fine-tuning of the LM Classifier or CMLM
• M being the LM Classifier to explain
• and generator being the RELITC method
  so the function signature should be usable in this case as well.

Following the CounterfactualExplanations.jl spirit, we can think of interoperability for other CE methods, such as MiCE, which is a predecessor for RELITC so it should be possible to reuse some of the code.

• MiCE