Graph Neural Networks for Optimal and Efficient Generation of Textual Counterfactuals

Abstract

Counterfactual explanations provide reasoning in the form of changes needed to be made in order for a model to make a different decision. When the model in question is a black-box classifier and the input consists of textual data, many counterfactual editors attempt to gain insights about the inner workings of the model by slightly altering the original instances. However most of them are computationally expensive due to the massive space of alternatives one has to search when altering a text. In this thesis, we propose using the recently thriving deep learning models which specifically operate on graph structured data, called Graph Neural Networks (GNN). We present an editor that generates semantically edited inputs, known as counterfactual interventions, which change the model prediction, thus providing a form of counterfactual explanations for the model. The editor utilizes a special graph type knows as a bipartite graph (or bigraph) along with a GNN that we developed so that it simulates the solution to the Rectangular Linear Assignment Problem (RLAP). During our experiments, we showcase the editor’s flexible nature, and discuss multiple trade-offs regarding explainability, minimality and speed. We test our editor on two NLP tasks - binary sentiment classification and topic classification - and show that the generated edits are contrastive, fluent and minimal, while the whole process remains significantly faster than other state-of-the-art counterfactual editors.