TY - JOUR

T1 - Экспертная валидация тематического моделирования (BERTopic) и последующее применение LLM к выделенным темам в качественном исследовании в рамках тематического анализа

AU - Червоткина, Полина Игоревна

PY - 2026/7/3

Y1 - 2026/7/3

N2 - This article explores the application of machine learning and generative language models to the analysis of qualitative interviews. The aim of the study is to develop and validate a mixed methodology combining topic modeling (BERTopic) with expert supervision. The empirical base consisted of 20 transcripts of semi-structured interviews with family members about health care practices. In the first stage, BERTopic, UMAP, and HDBSCAN were used to identify clusters of replicas based on bigrams and trigrams, and then with the addition of unigrams. Hyperparameter selection was then performed. The second stage involved expert validation of the clusters and the construction of a guided model with the addition of expert-generated themes (seed themes). The third stage involved visualizing the relationships between themes through term co-occurrence graphs. In the fourth stage, the results of the previous stages were summarized and grouped using DeepSeek. The results were compared with the encoder's performance on the BERT model topics and interview transcripts.It was found that BERTopic with bigrams and trigrams demonstrates higher topic diversity (0.96) and separability (silhouette coefficient of 0.69) compared to the model including unigrams (topic diversity of 0.84, silhouette coefficient of 0.56). Guided models made it possible to reduce the proportion of noisy utterances, identify topics not detected during the initial model run (e.g., device errors), and improve topic separability metrics (silhouette coefficient of 0.72 for the model with bigrams and trigrams, 0.73 for the model with unigrams). LLM demonstrates limited ability to independently formulate topics not specified in the prompt, but its performance is moderate with researcher coding. The paper proposes methods for reducing hallucinations and improving the quality of thematic generalization in LLM.

AB - This article explores the application of machine learning and generative language models to the analysis of qualitative interviews. The aim of the study is to develop and validate a mixed methodology combining topic modeling (BERTopic) with expert supervision. The empirical base consisted of 20 transcripts of semi-structured interviews with family members about health care practices. In the first stage, BERTopic, UMAP, and HDBSCAN were used to identify clusters of replicas based on bigrams and trigrams, and then with the addition of unigrams. Hyperparameter selection was then performed. The second stage involved expert validation of the clusters and the construction of a guided model with the addition of expert-generated themes (seed themes). The third stage involved visualizing the relationships between themes through term co-occurrence graphs. In the fourth stage, the results of the previous stages were summarized and grouped using DeepSeek. The results were compared with the encoder's performance on the BERT model topics and interview transcripts.It was found that BERTopic with bigrams and trigrams demonstrates higher topic diversity (0.96) and separability (silhouette coefficient of 0.69) compared to the model including unigrams (topic diversity of 0.84, silhouette coefficient of 0.56). Guided models made it possible to reduce the proportion of noisy utterances, identify topics not detected during the initial model run (e.g., device errors), and improve topic separability metrics (silhouette coefficient of 0.72 for the model with bigrams and trigrams, 0.73 for the model with unigrams). LLM demonstrates limited ability to independently formulate topics not specified in the prompt, but its performance is moderate with researcher coding. The paper proposes methods for reducing hallucinations and improving the quality of thematic generalization in LLM.

UR - https://elibrary.ru/XBLIML

UR - https://www.mendeley.com/catalogue/9c713cdb-cad8-37ca-a18a-eb7bd4165694/

U2 - 10.19181/inter.2026.18.2.8

DO - 10.19181/inter.2026.18.2.8

M3 - статья

VL - 18

SP - 91

EP - 125

JO - Интеракция. Интервью. Интерпретация

JF - Интеракция. Интервью. Интерпретация

SN - 2307-2075

IS - 2

ER -