Software Bug Prioritization via Hierarchical Attention Networks

Software development and deployment is often infested with bugs, and manually prioritising bug reports using domain knowledge is inaccurate and time-consuming. We propose a Hierarchical Attention-based recursive deep learning approach to automate bug prioritisation. Our method takes data from Linux and Chrome bug databases, generates word embeddings using GloVe, FastText, and Word2Vec, and processes them through a Hierarchical Attention Network (HAN). Results show strong performance on well-structured datasets (Linux) and highlight limitations on noisier data (Chrome).

• ›Hierarchical Attention Network applied to bug report prioritisation, capturing both word and sentence level semantics.
• ›Evaluation of GloVe, FastText, and Word2Vec embeddings on real-world bug databases.
• ›Analysis of model behaviour on well-structured (Linux) vs. unstructured (Chrome) bug reports.
• ›Identification of dataset structure as a key factor in deep learning-based bug classification performance.

Bug reports are tokenised and embedded using three embedding strategies (GloVe, FastText, Word2Vec). A Hierarchical Attention Network encodes the report at both the word level (word attention) and the sentence level (sentence attention), producing a document representation used for priority classification. The model is trained and evaluated on Linux kernel Bugzilla (~16.5k reports, 9 priority classes) and Chromium (~59k reports, 3 type classes).

The research question was whether an emotion score, concatenated to the document vector before the final softmax, improves priority prediction. It does — and by more than any embedding swap.

• ›Adding an emotion score gave the single largest improvement, +0.038 F1 over the word2vec baseline, beating every embedding change.
• ›Among embeddings alone, GloVe performed best.
• ›Hierarchical attention works well on well-structured Linux reports and degrades on unstructured Chromium ones — prompting an architecture change where the RNN blocks were replaced by two GRUs of differing cell sizes plus a parallel GRU path stacked into the softmax layer.
• ›On Chromium the architecture reached 0.7873 accuracy, but the emotion variant and the larger embeddings were never run there: compute was the binding constraint, and that experiment remains open.

This study project builds on the released code for 'Automated labeling of bugs and tickets using attention-based mechanisms in recurrent neural networks' (Lyubinets, Boiko and Nicholas, DSMP 2018). The contributions here are the emotion-score variant, the GRU architecture change, and the comparative embedding study.