Towards Explainable AI: Assessing the Usefulness and Impact of Added Explainability Features in Legal Document Summarization

Abstract

This study tested two different approaches for adding an explainability feature to the implementation of a legal text summarization solution based on a Deep Learning (DL) model. Both approaches aimed to show the reviewers where the summary originated from by highlighting portions of the source text document. The participants had to review summaries generated by the DL model with two different types of text highlights and with no highlights at all. The study found that participants were significantly faster in completing the task with highlights based on attention scores from the DL model, but not with highlights based on a source attribution method, a model-agnostic formula that compares the source text and summary to identify overlapping language. The participants also reported increased trust in the DL model and expressed a preference for the attention highlights over the other type of highlights. This is because the attention highlights had more use cases, for example, the participants were able to use them to enrich the machine-generated summary. The findings of this study provide insights into the benefits and the challenges of selecting suitable mechanisms to provide explainability for DL models in the summarization task.