Adaptive Title-Description Fusion separately encoded bug-report titles and descriptions and learned their relative contribution. On five-class Mozilla Bugzilla data it achieved Macro-F1 0.771, compared with 0.756 for standard gated fusion and 0.739 for direct concatenation. The result supports adaptive fusion but does not establish cross-project generalisation or real maintenance impact.
Key findings
- ATDF achieved Macro-F1 0.771 versus 0.756 for gated fusion and 0.739 for direct concatenation. The relative contribution of titles and descriptions varied across severity classes and repositories in the current experiment.
Why this matters globally
Automated severity triage could help teams prioritise large defect backlogs. Separately weighting short and long text fields may generalise conceptually to other software-engineering documents.
Thai researcher contribution
Researchers from Mahasarakham University's Faculty of Informatics and Khon Kaen University's Computer Engineering department jointly developed and evaluated the text-classification framework.
Limitations to consider
Evaluation was limited to Mozilla repositories and potentially inconsistent historical labels. Confidence intervals, significance tests, temporal splits, and computational cost are not reported in the abstract. A 0.015 Macro-F1 gain over gated fusion may be modest operationally, and developer or repair-time outcomes were not measured.