Duplicate bug reports are a frequent issue in software development, consuming significant resources and complicating prioritization. Despite progress in detection systems, there’s a need for better understanding of how different sections of bug reports contribute to detection performance. This work focuses on model explainability by analyzing the influence of bug report sections on automatic duplicate detection, emphasizing reducing false positives and negatives. It aims to identify key sections that reduce computational overhead while considering project-specific differences. We use two type of models: section-based and cross-section-based. Each bug report section or combination of sections is processed through a Siamese transformer network and classified using an MLP. Evaluations on Mozilla projects (i.e., Firefox, Core, Thunderbird) found that the “title” and “description” sections are most effective for duplicate detection, while “steps to reproduce” and “actual results” cause confusion.