This paper proposes HICode, a novel analysis pipeline leveraging large-scale language models (LLMs) to overcome the limitations of manual labeling or statistical tools (topic modeling) for large-scale text corpora analysis. Inspired by qualitative research methods, HICode consists of a two-step process: inductively generating labels directly from data and hierarchically clustering them to derive new topics. We measure the consistency with human-generated topics across three diverse datasets and validate its robustness through automated and human evaluation. A case study analyzing litigation documents related to the US opioid crisis reveals a pharmaceutical company's aggressive marketing strategy and demonstrates the potential of HICode for deep analysis of large-scale data.
