A Deep Learning Approach to Automated Bug Triaging: Investigating the Impact of Text Vectorization Methods on Effectiveness of CNN-LSTM

Abstract

Manual bug triage is a significant bottleneck in modern software development, leading to costly project delays. While automated systems offer a solution, their performance is fundamentally tied to their ability to comprehend the semantic content of bug reports. The text representation methods that power these systems have evolved significantly, yet a systematic comparison to understand their true impact on triage accuracy has been lacking. This paper conducts a large-scale empirical study to fill this gap, comparing nine text vectorization methods—spanning from classical TF-IDF to state-of-the-art Sentence-Transformers (SBERT)—within a hybrid CNN-LSTM framework. We evaluate the models on six benchmark datasets from projects like Google Chromium and Mozilla. Performance is measured using Top-k accuracy, a standard metric for this recommendation-style task, which assesses whether the correct developer is ranked among the top predictions. Our findings reveal that sentence-level embeddings from SBERT consistently and significantly achieve the highest Top-k accuracy, outperforming all other techniques, including contextual models like BERT. Our SBERT-based model establishes a new state-of-the-art, demonstrating that a holistic, sentence-level semantic understanding is critical for effective bug-to-developer assignment.