Scholarlink

<p>In software development, code review (CR) is a well-established process for ensuring code quality. Modern CR relies strongly on written communication between developers and reviewers. Therefore, the quality of interactions and the presence of unhealthy behaviors can significantly impact the CR process. The criticizing nature of CR comments increases the potential for various harmful and inappropriate language, which can negatively affect the entire CR process. The precise identification and categorization of these behaviors based on their root causes provides insights for designing automated diagnostic models that can detect these behaviors. To achieve this goal, studies have been conducted on unhealthy behaviors in CR comments, each emphasizing different aspects of these behaviors. This research addresses the existing gaps in identifying and diagnosing these behaviors by investigating various forms of unhealthy behavior in written review comments, with the intention of enhancing the overall quality of CR and interpersonal communication. We introduced a new concept, “Counterproductive" behavior, which expands on the well-known concept of toxicity to encompass other related traits. To validate our definition, we analyzed a human-labeled dataset of CR comments using statistical methods. Hypothesis testing was conducted to assess the comprehensiveness of our definition, while additional analyses, including ANOVA and Tukey’s HSD, were performed to investigate relationships between various counterproductive traits in CR communications. Furthermore, we analyzed the extracted relationships psychologically, considering harmful interpersonal behaviors and their broader implications. We also developed a recall-focused model for detecting counterproductive behavior in CR comments, utilizing an ensemble learning technique based on counterproductive traits. Given the nature of the problem, we prioritized recall as the most critical metric and achieved a mean recall of 94% ± 13% across counterproductive traits, which entailed a deliberate recall-precision trade-off, yielding average precision of 79% ± 7% while accepting some risk of false positives.</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

Do not review my code harshly: identifying counterproductive behavior in code review comments

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月2日

查看全文
  • Fatemeh Shaterian,
  • Marzieh Sadri,
  • MohammadAmin Fazli,
  • Jafar Habibi

摘要

In software development, code review (CR) is a well-established process for ensuring code quality. Modern CR relies strongly on written communication between developers and reviewers. Therefore, the quality of interactions and the presence of unhealthy behaviors can significantly impact the CR process. The criticizing nature of CR comments increases the potential for various harmful and inappropriate language, which can negatively affect the entire CR process. The precise identification and categorization of these behaviors based on their root causes provides insights for designing automated diagnostic models that can detect these behaviors. To achieve this goal, studies have been conducted on unhealthy behaviors in CR comments, each emphasizing different aspects of these behaviors. This research addresses the existing gaps in identifying and diagnosing these behaviors by investigating various forms of unhealthy behavior in written review comments, with the intention of enhancing the overall quality of CR and interpersonal communication. We introduced a new concept, “Counterproductive" behavior, which expands on the well-known concept of toxicity to encompass other related traits. To validate our definition, we analyzed a human-labeled dataset of CR comments using statistical methods. Hypothesis testing was conducted to assess the comprehensiveness of our definition, while additional analyses, including ANOVA and Tukey’s HSD, were performed to investigate relationships between various counterproductive traits in CR communications. Furthermore, we analyzed the extracted relationships psychologically, considering harmful interpersonal behaviors and their broader implications. We also developed a recall-focused model for detecting counterproductive behavior in CR comments, utilizing an ensemble learning technique based on counterproductive traits. Given the nature of the problem, we prioritized recall as the most critical metric and achieved a mean recall of 94% ± 13% across counterproductive traits, which entailed a deliberate recall-precision trade-off, yielding average precision of 79% ± 7% while accepting some risk of false positives.