Scholarlink

This paper proposes a two-stage method that uses Latent Dirichlet Allocation (LDA) to help organize and review academic articles. The goal is to support researchers when exploring large amounts of literature. In the first stage, LDA is trained on a small set of articles that were selected by reading their titles and abstracts. This allows the model to find common topics and related keywords. In the second stage, the topics are used to classify the full dataset, and the results are checked for consistency. A case study on sustainable supply chain design is used, based on 541 articles from the Web of Science. Three different subsets are tested to see how the selection method affects the model’s performance. The results show that when one person selects the articles, the topics are more consistent and easier to interpret. In contrast, when several people choose the articles, the results vary more and are harder to group. Statistical tests support this difference, confirming that human selection can strongly influence automated topic models. And highlight the usefulness of combining machine learning and human judgment to help with literature reviews.

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

A Two Stage Approach to Supporting Human Judgment in Literature Review Using Latent Dirichlet Allocation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Lucia González,
  • Lorena Espinoza Pérez,
  • Andrea Espinoza Pérez,
  • Óscar C. Vásquez

摘要

This paper proposes a two-stage method that uses Latent Dirichlet Allocation (LDA) to help organize and review academic articles. The goal is to support researchers when exploring large amounts of literature. In the first stage, LDA is trained on a small set of articles that were selected by reading their titles and abstracts. This allows the model to find common topics and related keywords. In the second stage, the topics are used to classify the full dataset, and the results are checked for consistency. A case study on sustainable supply chain design is used, based on 541 articles from the Web of Science. Three different subsets are tested to see how the selection method affects the model’s performance. The results show that when one person selects the articles, the topics are more consistent and easier to interpret. In contrast, when several people choose the articles, the results vary more and are harder to group. Statistical tests support this difference, confirming that human selection can strongly influence automated topic models. And highlight the usefulness of combining machine learning and human judgment to help with literature reviews.