DeeperMatcher: Using LLMs for Crowd Based Requirements Engineering

Development teams of popular mobile apps can receive thousands of user reviews daily. At the same time, these developers communicate with each other using completely different communication channels, such as their software repository issue tracker. Although accessible and manageable for developers, the nature of the content in development issues differs starkly from what most users write about in their reviews. Issues may lack the steps to reproduce a bug or insights justifying the priority of a new feature request, which might be latently present in user reviews. Moreover, a specific bug might be reported by only a handful of user reviews and might go completely unnoticed for years abandoned in the app store.

We propose an approach to augment software repository issues with pertinent user reviews to bridge the gap between what users write about in their app reviews and the issues managed by developers. Our approach automatically retrieves user reviews with high semantic textual similarity (STS) to the issue content, computed using a state-of-the-art large language model (LLM), and suggests reviews relevant to the developers' grasp on the issue. We conducted a preliminary evaluation using data from the BikeSP project and the Brave open-source browser, with over 9k issues and more than 180k user comments. Our early results testify to the potential of the approach. Besides providing insights into what users ask as a feature or report as a bug, the information found in the pertinent matches points toward a novel and promising way of leveraging user reviews for a concerted evolution of apps.