Closing the Interpretive Loop with BERT, Our Neural Topic Modeling Friend

Natural language processing (NLP) holds significant potential for analyzing and interpreting patterns in large qualitative datasets. However, the use of topic modeling to assist quantitative ethnography (QE) researchers in interpreting links among codes within Epistemic Network Analysis (ENA) models has not been explored. This study examines how BERTopic, a neural topic model, supports researchers in interpreting co-occurrences of codes in ENA models. Our findings show that BERTopic provides phrases and representative documents that align with human-developed themes, as well as alternative topics offering new insights into the nuanced connections among codes. We also found that a data pre-processing approach utilizing only the coded lines within a stanza window was critical to the effectiveness of BERTopic in developing interpretable and representative topics. We suggest that integrating neural topic modeling, such as BERTopic, with an Only Coded Lines approach can enhance the interpretative process in QE studies.