MilaNLProc/contextualized-topic-models

# Technical Summary Implements two complementary neural architectures—CombinedTM integrates contextualized embeddings with bag-of-words reconstruction via a VAE-like framework, while ZeroShotTM operates on embeddings alone for cross-lingual and zero-shot capabilities. Leverages Sentence-BERT for flexible embedding generation across any HuggingFace model, with careful preprocessing workflows to manage vocabulary size (≤2000 terms recommended) and balance between preprocessed BoW and raw text for embeddings. Includes Kitty, a human-in-the-loop classifier submodule for interactive document clustering and annotation workflows.

1,266 stars and 1,326 monthly downloads. No commits in the last 6 months. Available on PyPI.