Minimizing Protoform Redundancy to Enhance Linguistic Summaries in Object Detection

Large language models (LLMs) have gained widespread popularity because of their remarkable capabilities. However, they cannot yet be relied upon to process large amounts of raw data and consistently produce trustworthy results while preserving privacy. In contrast, linguistic summaries can be generated from these data sources using fuzzy linguistic proforms ranked by accuracy and other relevant metrics. Unfortunately, such summaries often yield a large number of redundant statements, making them difficult to parse and thus reducing their overall effectiveness. To address this issue, we introduce a method for processing these statements, represented as a directed acyclic graph that significantly reduces redundancy in the resulting summaries. We take advantage of LLMs for the narrow task of further enhancing and summarizing our already reduced protoforms, presenting a combined approach. To facilitate the application of this technique to object detection, we also propose methods for computing linguistic qualities over heterogeneous sets, such as different types of declarations (e.g., false alarms vs. missed targets). We demonstrate the effectiveness of our approach on both synthetic data and a real person-detection model tested on data generated using the Unreal Engine.