Scholarlink

<p>Interviewing is employed in numerous fields to collect qualitative data, yet analysis is laborious and time-consuming. Recently, large language models (LLMs) have attracted attention as good assistants for qualitative data analysis because they can ingest and aggregate large volumes of text quickly. However, the process by which an LLM summarizes interview content is opaque and the evaluation of LLM-generated results remains highly dependent on the individual researcher. To overcome this problem, we propose a novel method that combines LLMs with network analysis to create a transparent and objectively evaluable workflow. Specifically, we extracted items from transcripts using an LLM, the LLM determined similarities for all pairs to construct a network, we classified it via network clustering, and the LLM generated themes. We show the usefulness of the method through case studies aligned with evolutionary economic geography. As a result, the proposed method makes the determinations of the LLM intuitively legible by introducing a network representation of similarity and its clustering. Because both process and result rest on optimization formulations, anyone can evaluate them.</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

Hybrid approach combining large language model and network science for interview data analysis

  • 期刊论文

  • 发表日期：

    2025年9月1日

查看全文
  • Taiyo Nakatani,
  • Kaito Tsuji,
  • Saya Watanabe,
  • Takaya Ohsato,
  • Hiroaki Nanjo,
  • Takaaki Aoki

摘要

Interviewing is employed in numerous fields to collect qualitative data, yet analysis is laborious and time-consuming. Recently, large language models (LLMs) have attracted attention as good assistants for qualitative data analysis because they can ingest and aggregate large volumes of text quickly. However, the process by which an LLM summarizes interview content is opaque and the evaluation of LLM-generated results remains highly dependent on the individual researcher. To overcome this problem, we propose a novel method that combines LLMs with network analysis to create a transparent and objectively evaluable workflow. Specifically, we extracted items from transcripts using an LLM, the LLM determined similarities for all pairs to construct a network, we classified it via network clustering, and the LLM generated themes. We show the usefulness of the method through case studies aligned with evolutionary economic geography. As a result, the proposed method makes the determinations of the LLM intuitively legible by introducing a network representation of similarity and its clustering. Because both process and result rest on optimization formulations, anyone can evaluate them.