BERTopic
Topic modeling library leveraging transformer embeddings and c-TF-IDF.
About
BERTopic approaches topic modeling with a modular pipeline built on transformer embeddings: sentence-transformers encode documents, UMAP reduces dimensionality, HDBSCAN clusters the results, and a class-based TF-IDF procedure (c-TF-IDF) extracts interpretable keywords for each topic. Every stage can be swapped, which lets one library support more than a dozen variants including guided, supervised, semi-supervised, dynamic topics over time, hierarchical, zero-shot, online or incremental, and multimodal topic modeling over images and text. Topic representations can be refined with LLMs such as OpenAI models, and multilingual embedding models extend coverage to over 50 languages. Interactive visualizations show topic hierarchies, similarity heatmaps, and per-document distributions. Created and maintained by Maarten Grootendorst, released under the MIT license with an accompanying paper, BERTopic is widely used by researchers and data scientists analyzing large document collections. It requires Python 3.10 or newer and installs from PyPI.
Should you use BERTopic?
Pick it when
Pick BERTopic when you need readable topics from thousands of documents such as reviews, tickets, or survey answers, and want options like topics over time, guided seeds, or LLM-written labels without building the pipeline yourself.
Look elsewhere when
Skip it on very large corpora with no GPU, since the default pipeline embeds every document with a transformer before clustering, or when you want each document modeled as a mix of topics, which Gensim's LDA does directly.
Alternatives to BERTopic
- Gensim
Classic LDA and LSA that stream corpora larger than RAM on plain CPUs, but with bag-of-words topics and no embedding semantics or interactive visuals.
- KeyBERT
Same author's lighter tool that extracts key phrases per document, enough when you need tags for each text rather than corpus-wide topics.
- Sentence-Transformers
The embedding layer BERTopic uses; go direct if you only need clustering or similarity search, not topic labels and visualizations.
- scikit-learn
LDA and NMF topic models with no transformer dependency, light and familiar, but no c-TF-IDF labels, dynamic topics or visual tooling.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Category
- Natural Language Processing
- Price
- Free
- Platform
- Local/Desktop
- Difficulty
- Easy (2/5)
- License
- MIT
- Added
- Apr 3, 2026