Automatic Labeling of SDN Controller Defect Text based on Neural Topic Model

Software defined networking controller defects may cause unexpected failures and reduce network reliability. The historical defect texts provide information on defecting symptoms and fixing strategies, but they are not sufficiently labeled. An approach based on a neural topic model is proposed to automatically assign phrase labels to the defect text. First, five types of information that can be extracted from the defect text are provided to guide the generation of candidate phrases. Second, contextualized embedding representations are extracted by the pre-training BERT Overflow model and input to the contextualized topic model to extract a more coherent topic distribution. Third, the accuracy of the unsupervised assignment of labels is improved by filtering candidate phrases through two-level label filtering. Finally, FastText is applied to train a multi-label classifier to assign labels to the new defect text. The experimental results demonstrate that the proposed approach can effectively assign interpretable labels to the defect text.