Abstract:
Duplication of bug reports poses a significant impact on software development efficiency with
12-25% of bugs report being duplicated on large projects. Manual identification techniques are
both time consuming and prone to error. This study examines whether semantic metadata that
encodes content-level similarities together with simplified deep learning architectures can be
more effective than relying on complex models. We aim to examine existing practices and determine
weaknesses, research and prove semantic content-based metadata to be useful in robust
identification, structure and test various deep learning models to define the best methods to use
to achieve higher levels of performance. Existing studies reveal that machine learning methods
outperform traditional information-retrieval techniques by a significant margin, while deep
learning approaches achieve higher accuracy but often suffer from limited feature diversity. We
collected 535,477 of quality pairs with a 70%-10%-20% split for training, validation, and testing.
Our feature engineering on DistilBERT yielded a mean correlation of 0.1888, surpassing
traditional metadata. We compared 6 architectures, namely, LSTM, CNN, LSTM+Metadata,
Hybrid+Attention, Hybrid+Attention with Metadata, and proposed LSTM+CNN+Metadata,
evaluated through accuracy, precision, recall, F1-score, and AUC-ROC metrics. The proposed
architecture resulted 92.53% F1-score, which is better than complex attention-based models.
Contextual processing is proven to be better, as LSTM performs higher than CNN. This paper
highlights that the quality of features is more critical for model performance than model complexity,
presenting a cost-effective, accurate, and scalable method for automated duplicate bug
detection in real-world applications.
