BERTopic — Neural Topic Modeling with Transformers

Introduction

BERTopic is a topic modeling technique that combines document embeddings from sentence transformers with a class-based TF-IDF procedure (c-TF-IDF) to generate coherent topic representations. Unlike LDA, it does not assume a fixed number of topics and can discover the optimal topic count automatically through density-based clustering.

What BERTopic Does

• Discovers topics from document collections without specifying the number of topics in advance
• Represents each topic with interpretable keywords derived from c-TF-IDF scores
• Supports dynamic topic modeling to track how topics evolve over time
• Enables hierarchical topic merging for exploring topic granularity
• Generates topic visualizations including intertopic distance maps and bar charts

Architecture Overview

BERTopic's pipeline has four stages. First, documents are embedded into dense vectors using a sentence transformer (or any custom embedding model). Second, UMAP reduces the embedding dimensionality while preserving local structure. Third, HDBSCAN clusters the reduced embeddings into groups, with outliers assigned to a noise topic. Fourth, c-TF-IDF computes per-topic word importance scores to produce human-readable topic labels. Each stage is modular and can be swapped with alternatives (PCA instead of UMAP, k-means instead of HDBSCAN, KeyBERT for representation).

Self-Hosting & Configuration

• Install with pip install bertopic (includes sentence-transformers, umap-learn, hdbscan)
• Configure the embedding model via BERTopic(embedding_model='all-MiniLM-L6-v2') or pass a custom model
• Adjust UMAP and HDBSCAN parameters for cluster granularity and noise handling
• Save and load trained models with topic_model.save() and BERTopic.load()
• Use GPU-accelerated embeddings for faster processing on large corpora

Key Features

• Modular design: swap out embedding, reduction, clustering, and representation components independently
• Dynamic topic modeling tracks topic evolution across timestamps
• Online learning mode for incremental updates without retraining from scratch
• Multi-modal support: combine text with images for visual topic modeling
• LLM-enhanced topic labels using GPT or other language models for more descriptive names

Comparison with Similar Tools

• Gensim LDA — classic probabilistic topic model requiring a fixed topic count; BERTopic discovers topics automatically with richer representations
• Top2Vec — similar embedding-based approach; BERTopic adds c-TF-IDF for more interpretable topics and offers more customization
• Scikit-learn NMF/LDA — lightweight but limited to bag-of-words; BERTopic uses contextual embeddings
• CorEx — information-theoretic topic model supporting anchored topics; BERTopic is more flexible with modern transformers
• LDA2Vec — combines LDA with word2vec; less actively maintained than BERTopic

FAQ

Q: How many documents does BERTopic need to produce good topics? A: Generally a few hundred documents suffice, but quality improves with thousands. Very small corpora (under 100 docs) may not cluster well.

Q: Can I specify the number of topics? A: Yes. Set nr_topics to reduce topics after fitting, or use min_topic_size and HDBSCAN parameters to control granularity during clustering.

Q: Does BERTopic support languages other than English? A: Yes. Use a multilingual embedding model (e.g., paraphrase-multilingual-MiniLM-L12-v2) and BERTopic will discover topics in any language.

Q: Can I update a trained model with new documents? A: Yes. BERTopic supports online learning via partial_fit() and transform() to incorporate new documents without full retraining.

Sources

• https://github.com/MaartenGr/BERTopic
• https://maartengr.github.io/BERTopic/