What it does

Contextualized Topic Models (CTM) are a family of neural topic models that feed BERT-style embeddings into topic modeling. The package offers two main variants: CombinedTM, which mixes contextual embeddings with a traditional bag-of-words representation to produce more coherent topics, and ZeroShotTM, which relies purely on embeddings and can handle missing vocabulary at test time or work cross-lingually when paired with multilingual sentence transformers. There’s also Kitty, a human-in-the-loop classifier for quick document labeling and filtering.

The interesting bit

The architecture is deliberately embedding-agnostic — swap in whatever new sentence transformer HuggingFace releases next week without rewriting the model. The authors also spotted a subtle preprocessing trap: BERT wants raw text, but your bag-of-words wants cleaned text, so the toolkit handles feeding different versions of the same document to each branch.

Key highlights

• Published at ACL and EACL 2021; ~1,300 stars
• Supports any SBERT-compatible embedding model, including multilingual ones
• Zero-shot cross-lingual topic modeling via ZeroShotTM
• Supervised variant (SuperCTM) and a human-in-the-loop classifier (Kitty)
• Multiple Colab tutorials for different use cases
• Preprocessing pipeline included to handle the raw-vs-cleaned text split

Caveats

• Bag-of-words vocabulary should stay under ~2,000 terms for stable training; larger vocabs mean more parameters and worse fitting
• Multilingual models used on English-only data may underperform compared to English-specific embeddings
• CUDA setup is left to you — the docs just point to PyTorch’s instructions

Verdict

Worth a look if you’re still running LDA and wincing at the incoherent topic dumps. Skip it if your vocabulary is massive and unprunable, or if you need a fully unsupervised black-box solution with no preprocessing decisions to make.