Duplicate detection of multimodal and domain-specific trouble reports when having few samples : An evaluation of models using natural language processing, machine learning, and Siamese networks pre-trained on automatically labeled data

Abstract: Trouble and bug reports are essential in software maintenance and for identifying faults—a challenging and time-consuming task. In cases when the fault and reports are similar or identical to previous and already resolved ones, the effort can be reduced significantly making the prospect of automatically detecting duplicates very compelling. In this work, common methods and techniques in the literature are evaluated and compared on domain-specific and multimodal trouble reports from Ericsson software. The number of samples is few, which is a case not so well-studied in the area. On this basis, both traditional and more recent techniques based on deep learning are considered with the goal of accurately detecting duplicates. Firstly, the more traditional approach based on natural language processing and machine learning is evaluated using different vectorization techniques and similarity measures adapted and customized to the domain-specific trouble reports. The multimodality and many fields of the trouble reports call for a wide range of techniques, including term frequency-inverse document frequency, BM25, and latent semantic analysis. A pipeline processing each data field of the trouble reports independently and automatically weighing the importance of each data field is proposed. The best performing model achieves a recall rate of 89% for a duplicate candidate list size of 10. Further, obtaining knowledge on which types of data are most important for duplicate detection is explored through what is known as Shapley values. Results indicate that utilizing all types of data indeed improve performance, and that date and code parameters are strong indicators. Secondly, a Siamese network based on Transformer-encoders is evaluated on data fields believed to have some underlying representation of the semantic meaning or sequentially important information, which a deep model can capture. To alleviate the issues when having few samples, pre-training through automatic data labeling is studied. Results show an increase in performance compared to not pre-training the Siamese network. However, compared to the more traditional model it performs on par, indicating that traditional models may perform equally well when having few samples besides also being simpler, more robust, and faster.