Learning to Categorize Bug Reports with LSTM Networks

The manual routing of bug reports to specialized expertteams is a time-consuming and expensive process. In this paper,we investigated how this process can be automated by trainingdeep networks and state-of-the-art classifiers from thousands ofreal bug reports from a software company. Different combinationsof the natural language processing methodslemmatization,postagger,bigramandstopword removalwere evaluated in theclassification algorithmsLinear Support Vector Machines(SVMs),multinomial naive Bayes, andLong Short Term Memory(LSTM)networks. For feature processing we used theTerm Frequency-Inverse Document Frequency(TF-IDF) method. Best results wereobtained with a combination of thebigrammethod and linearSVMs. Similar prediction performance values were observedwith LSTM networks that however promise to improve furtherwith larger datasets. The bug triage tool was implemented in amicroservice architecture using docker containers which allowsfor extending individual components and simplifies applicationsto other text classification problems.