Scholarlink

<p>Social coding platforms like GitHub facilitate collaborative software development through pull requests (PRs), which generate discussions that significantly impact code quality, requirements, and design. Such conversations become a rich source of insights for improving development practices and predicting project outcomes and are subject to several human aspects that have been linked to code quality and PR acceptance. Sentiment analysis is one of the many ways to try to understand these human aspects. However, PR discussions are multifaceted, often involving technical jargon and aspects which limits the utility of general-purpose sentiment analysis tools. This has led to the creation of SE-specific tools, but recent studies have also observed that they demonstrate limited effectiveness. Thus, this study explores the potential of using large language models (LLMs) for this purpose, given their enhanced contextual understanding and ability to process technical language. We evaluated ten LLMs across proprietary and open-source categories, using two complementary datasets: a curated Gold dataset and the PRemo dataset, which captures real-world PR discussions. The models were assessed under zero-shot, few-shot and chain-of-thought prompting techniques on 8,913 messages. In addition, we establish baselines by evaluating fine-tuned transformer-based models. Results show that GPT-4o achieved the highest overall performance across the LLMs, though smaller models, such as Mistral Small and Deepseek-R1 32B delivered competitive results. Transformer-based models achieved excellent performance on the Gold dataset but exhibited degradation on the PRemo dataset. Finally, we conducted a qualitative analysis of misclassified instances, revealing recurring challenges related to technical terminology, sentiment-charged keywords, message length, and contextual ambiguity. These findings suggest that model selection should balance performance requirements against practical constraints, rather than defaulting to the largest available models.</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

Leveraging large language models for sentiment analysis in GitHub pull request discussions

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月9日

查看全文
  • Daniel Coutinho,
  • Breno Braga,
  • Theo Canuto,
  • Juliana Alves Pereira,
  • Wesley K. G. Assunção,
  • Igor Steinmacher,
  • Marco Gerosa,
  • Alessandro Garcia

摘要

Social coding platforms like GitHub facilitate collaborative software development through pull requests (PRs), which generate discussions that significantly impact code quality, requirements, and design. Such conversations become a rich source of insights for improving development practices and predicting project outcomes and are subject to several human aspects that have been linked to code quality and PR acceptance. Sentiment analysis is one of the many ways to try to understand these human aspects. However, PR discussions are multifaceted, often involving technical jargon and aspects which limits the utility of general-purpose sentiment analysis tools. This has led to the creation of SE-specific tools, but recent studies have also observed that they demonstrate limited effectiveness. Thus, this study explores the potential of using large language models (LLMs) for this purpose, given their enhanced contextual understanding and ability to process technical language. We evaluated ten LLMs across proprietary and open-source categories, using two complementary datasets: a curated Gold dataset and the PRemo dataset, which captures real-world PR discussions. The models were assessed under zero-shot, few-shot and chain-of-thought prompting techniques on 8,913 messages. In addition, we establish baselines by evaluating fine-tuned transformer-based models. Results show that GPT-4o achieved the highest overall performance across the LLMs, though smaller models, such as Mistral Small and Deepseek-R1 32B delivered competitive results. Transformer-based models achieved excellent performance on the Gold dataset but exhibited degradation on the PRemo dataset. Finally, we conducted a qualitative analysis of misclassified instances, revealing recurring challenges related to technical terminology, sentiment-charged keywords, message length, and contextual ambiguity. These findings suggest that model selection should balance performance requirements against practical constraints, rather than defaulting to the largest available models.