Repo for the ACL 2022 short paper “Counterfactual Explanations for Natural Language Interfaces” by George Tolkachev, Stephen Mell, Steve Zdancewic, and Osbert Bastani.
Paper link: https://arxiv.org/abs/2204.13192
Video presentation: https://www.youtube.com/watch?v=dwl1pWxo3ho
Our contribution is a novel algorithm for computing counterfactual explanations for semantic parsers. We assume the following process:
-
User provides a command, but the NL interface fails to generate the desired result
-
User now provides desired result (in the form of a denotation/trajectory to goal) in addition to command
-
Our algorithm computes an alternative utterance (counterfactual explanation) that:
a. Is as close as possible to the original utterance
b. The semantic parser correctly processes (i.e. user achieves desired result)
The outcome of this process is that the user has a better understanding of how to modify commands to achieve their goals in future interactions with the system.
Here is an illustration of the above process:
