Searching for Radical Healing: Comparing Neural Topic Modeling and Human Coding in Qualitative Research

• In Event: Qualitative Research and AI

Abstract

The efficiency and accuracy of coding in qualitative research are critical yet often challenging. Although traditional human coding is nuanced and context-sensitive, it is also time-consuming and prone to bias. This study compares the efficiency, reliability, and contextual understanding of BERT-based topic modeling and traditional human coding in qualitative research. Using data from interviews with undergraduate mentors in an afterschool STEM program for Black youth, this study found that BERT significantly reduces coding time but lacks the nuanced understanding of human coders. Reliability measures indicated very low agreement between BERT and human coders for most codes. We discuss the implications of combining BERT’s speed with human insights to enhance large-scale qualitative research and create justice-centered methodological possibilities.

Authors

• Adenike Omolara Adefisayo, Clemson University

• Atefeh Behboudi, Vanderbilt University

• Golnaz Arastoopour Irgens, Vanderbilt University

• Renee Mary Lyons, Clemson University

• Corliss Outley, Clemson University

• Rhondda R. Thomas, Clemson University

• Gail Awan, Urban League of the Upstate

• Eleanor Hatcher, Clemson University