Scholarlink

This study investigates the role of sentiment analysis in understanding discussions and predicting the scores of answers on Stack Overflow (SO), a vital platform for the software development community. A dual approach, combining machine learning models and expert evaluation, was employed to analyze the correlation between sentiments in discussions and voting outcomes. Embedding techniques such as Word2Vec, GloVe, and BERT were utilized alongside regression models including Random Forest and deep learning models to predict answer scores, revealing minimal performance differences among embeddings. Interestingly, deep learning models demonstrated better predictive performance, while BERT embeddings did not show significant advantages over classical techniques. An expert opinion study involving 13 participants highlighted a stronger alignment with negative sentiments, suggesting greater clarity in their interpretation. These findings highlight the potential of sentiment analysis in understanding user engagement on SO and emphasize the need for fine-tuning domain-specific models for improved results.

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

Exploring Sentiments in Stack Overflow Score and Discussion: A Dual Approach with Machine Learning Models and Expert Evaluation

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Golshid Ranjbaran,
  • Banani Roy,
  • Chanchal K. Roy,
  • Kevin A. Schneider

摘要

This study investigates the role of sentiment analysis in understanding discussions and predicting the scores of answers on Stack Overflow (SO), a vital platform for the software development community. A dual approach, combining machine learning models and expert evaluation, was employed to analyze the correlation between sentiments in discussions and voting outcomes. Embedding techniques such as Word2Vec, GloVe, and BERT were utilized alongside regression models including Random Forest and deep learning models to predict answer scores, revealing minimal performance differences among embeddings. Interestingly, deep learning models demonstrated better predictive performance, while BERT embeddings did not show significant advantages over classical techniques. An expert opinion study involving 13 participants highlighted a stronger alignment with negative sentiments, suggesting greater clarity in their interpretation. These findings highlight the potential of sentiment analysis in understanding user engagement on SO and emphasize the need for fine-tuning domain-specific models for improved results.