BERTopic Modeling

BERTopic is a topic modeling technique that clusters transformer-based document embeddings, typically with UMAP and HDBSCAN, and describes each cluster with class-based TF-IDF keywords.

BERTopic은 트랜스포머 기반 문서 임베딩을 보통 UMAP과 HDBSCAN으로 군집화하고 클래스 기반 TF-IDF 키워드로 각 군집을 설명하는 토픽모델링 기법이다.

How do I discover the main topics in a large set of short texts?

대량의 짧은 텍스트에서 주요 주제를 어떻게 찾아내나요?

BERTopic embeds each document with a pretrained sentence transformer, reduces dimensions with UMAP, clusters the documents with HDBSCAN, and labels each topic by its most distinctive words using class-based TF-IDF. It fits short texts such as reviews, tweets, or open-ended responses where word-count models such as LDA struggle with sparse vocabulary.

When to use BERTopic Modeling

• Large collections of short texts such as reviews or open-ended answers
• Exploring themes without predefined categories
• Tracking how topic prevalence changes over time or across groups

When to use something else

• Each document should be a mixture of several topics with probabilistic proportions: use LDA or a structural topic model
• Categories are known in advance: use a supervised text classifier

Assumptions and requirements

• Embedding model suited to the language of the texts
• Enough documents for density-based clustering to form stable topics
• Results vary with UMAP and HDBSCAN settings and random seed
• Topics require human review and labeling

What the analysis reports

• Topic assignment for each document, including an outlier topic
• Top keywords per topic (c-TF-IDF)
• Topic sizes and representative documents
• Intertopic distance map or topic-over-time plot
• Coherence or diversity scores

How to report it

BERTopic identified [k] topics among [N] documents; the largest topic ([value]% of documents) was characterized by the keywords [word 1], [word 2], and [word 3].

Related methods

• Latent Dirichlet Allocation (LDA)
• Structural Topic Model
• UMAP Dimensionality Reduction

Run BERTopic Modeling on Quorant