Advancing LLM-Based Issue Report Classification with Explained Few-Shot Learning, Intent Extraction, Ensemble, and Summarization

Abstract

Automated issue report classification is increasingly important as software projects face growing volumes of complex reports. This paper investigates LLM-based strategies for classifying long issue reports, comparing intent-oriented and example-driven approaches. The proposed methods combine intent extraction, ensemble voting, explained few-shot learning, and summarization to improve classification quality and consistency. The evaluation compares these approaches against neural and LLM-based baselines using models including GPT-4o, GPT-3.5-turbo, and Qwen 2.5-32B. The results show improvements over the state of the art, highlight differences among model families, and provide a taxonomy of classification challenges that can guide future research and practical adoption.