BERTopic

0.17.4 · active · verified Wed Apr 15

BERTopic is a topic modeling technique that leverages state-of-the-art transformer models (like BERT) and a class-based TF-IDF procedure to create dense clusters, resulting in easily interpretable topics while retaining important words in their descriptions. It is currently at version 0.17.4 and actively maintained with regular updates and feature enhancements.

Warnings

breaking BERTopic models are not guaranteed to be compatible across different versions. When saving and loading models, ensure that the BERTopic version, Python version, and dependency versions (e.g., UMAP, HDBSCAN) are identical to prevent errors or unexpected behavior.
Fix: Pin the versions of `bertopic` and its core dependencies (`umap-learn`, `hdbscan`, `sentence-transformers`) in your `requirements.txt` or `pyproject.toml` to match the environment where the model was trained.
gotcha The default UMAP algorithm used for dimensionality reduction has a stochastic nature, meaning repeated runs with the same data can yield slightly different topic results.
Fix: For reproducible results, initialize UMAP with a `random_state` and pass it to BERTopic: `from umap import UMAP; umap_model = UMAP(random_state=42); topic_model = BERTopic(umap_model=umap_model)`.
gotcha For optimal performance, especially with large documents or when iterating over parameters, it is recommended to pre-calculate embeddings and pass them to BERTopic.
Fix: Calculate embeddings once using `sentence-transformers` (or your preferred embedding model) and then pass them as the `embeddings` argument to `topic_model.fit_transform(docs, embeddings=precomputed_embeddings)`.
gotcha By default, BERTopic initializes with an English-optimized embedding model ('all-MiniLM-L6-v2'). For multilingual datasets, you must explicitly specify the language.
Fix: For multilingual documents, initialize BERTopic with `topic_model = BERTopic(language="multilingual")`. This will load a multilingual model ('paraphrase-multilingual-MiniLM-L12-v2').
deprecated Support for Python 3.8 was officially dropped with recent BERTopic versions, notably around 0.17.x releases.
Fix: Upgrade your Python environment to Python 3.9 or higher. Python 3.10 and 3.11 are explicitly supported, with 3.13 support added in recent patches.

Install

pip install bertopic Base Installation
pip install bertopic[flair,gensim,spacy,use] Install with specific embedding backends
pip install bertopic[vision] Install for topic modeling with images

Imports

BERTopic
```
from bertopic import BERTopic
```

Quickstart

This quickstart example demonstrates how to initialize BERTopic, fit it to a dataset (the 20 newsgroups dataset is commonly used), and retrieve information about the discovered topics. The `fit_transform` method processes the documents, returning topic assignments and probabilities.

from bertopic import BERTopic
from sklearn.datasets import fetch_20newsgroups

# Fetch documents (e.g., 20 newsgroups dataset)
docs = fetch_20newsgroups(subset='all', remove=('headers', 'footers', 'quotes'))['data']

# Initialize and train BERTopic model
topic_model = BERTopic()
topics, probs = topic_model.fit_transform(docs)

# Get information about the frequent topics
topic_info = topic_model.get_topic_info()
print(topic_info.head())

view raw JSON →