Scholarlink

<p>In software development, bug reports (BRs) are essential for identifying defects, but the volume of reports in large projects makes manual relatedness analysis slow and error-prone. We study machine-learning approaches for predicting BR relatedness under a file-overlap target, File-Change Similarity (FCS). We compare TF-IDF, frozen sentence-level T5 embeddings (without domain fine-tuning), and a hybrid lexical-semantic representation. Our pipeline covers data retrieval, preprocessing, vectorization, normalization, neural-network training, and evaluation. We evaluated 56 models using various modeling strategies. Analysis reveals that using complete vectors as features is more effective than cosine distance. The hybrid approach shows competitive descriptive performance comparable to TF-IDF alone. Fine-tuning on 14 models tested 168 hyperparameter combinations, with Adam and RMSprop optimizers showing best performance. Key contributions include evaluating T5 and TF-IDF performance for BRs, exploring a hybrid approach, and providing a methodological framework for representation comparison. This research offers suggestions for improving efficiency in development and resource allocation. In the context of frozen T5 embeddings, the findings on T5 performance and the comparison with strong TF-IDF baselines drive future research directions. Since the T5 weights were not specifically trained on the bug report domain (frozen), these results serve as a baseline for future fine-tuning experiments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Lexical and semantic representations for similar bug report detection: A TF-IDF and frozen T5 comparative study

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月25日

查看全文
  • Iann Barbosa,
  • João Brunet,
  • Franklin Ramalho

摘要

In software development, bug reports (BRs) are essential for identifying defects, but the volume of reports in large projects makes manual relatedness analysis slow and error-prone. We study machine-learning approaches for predicting BR relatedness under a file-overlap target, File-Change Similarity (FCS). We compare TF-IDF, frozen sentence-level T5 embeddings (without domain fine-tuning), and a hybrid lexical-semantic representation. Our pipeline covers data retrieval, preprocessing, vectorization, normalization, neural-network training, and evaluation. We evaluated 56 models using various modeling strategies. Analysis reveals that using complete vectors as features is more effective than cosine distance. The hybrid approach shows competitive descriptive performance comparable to TF-IDF alone. Fine-tuning on 14 models tested 168 hyperparameter combinations, with Adam and RMSprop optimizers showing best performance. Key contributions include evaluating T5 and TF-IDF performance for BRs, exploring a hybrid approach, and providing a methodological framework for representation comparison. This research offers suggestions for improving efficiency in development and resource allocation. In the context of frozen T5 embeddings, the findings on T5 performance and the comparison with strong TF-IDF baselines drive future research directions. Since the T5 weights were not specifically trained on the bug report domain (frozen), these results serve as a baseline for future fine-tuning experiments.