Context: Users and developers use bug tracking systems to report errors that occur during the development and testing of software. The manual identification of duplicates is a tedious task especially with software that have large bug repositories. In this context, their automatic detection becomes a necessary task that can help prevent frequently fixing the same bug.
Objective: In this paper, we propose BERT-MLP, a novel pre-trained language model using Bidirectional Encoder Representations from Transformers (BERT) for duplicate bug report detection with the aim of improving the detection rate compared to existing works.
Method: Our approach considers only unstructured data. These are fed into the BERT model in order to learn the contextual relationships between words. The output is fed into a Multi-Layer Perceptron (MLP) classifier, which represents our base duplicate bug report detector.
Results: Our approach was evaluated on three projects: Mozilla Firefox, Eclipse Platform and Thunderbird. It achieved an accuracy of 92.11%, 94.08% and 89.03% respectively for Mozilla, Eclipse and Thunderbird. A comparison with a Dual- Channel Convolutional Neural Network (DC-CNN) model and other pre-trained models, including RoBERTa and Sentence-Bert has been conducted. Results showed that BERT-MLP outper- formed, the second best performing model (DC-CNN) by 12% in accuracy for Eclipse and 9% for both Mozilla and Thunderbird, respectively.