contextualized-topic-models

Why we included this project

Topic modeling usually starts from word counts, but this package takes a different route: it runs documents through contextualized embeddings like BERT and then fits a neural topic model on top, which tends to produce more coherent topics than classic bag-of-words methods. Two model variants handle different cases. CombinedTM mixes the embeddings with the bag of words for better coherence, while ZeroShotTM can label documents containing words it never saw during training, and when trained on multilingual embeddings it works across languages without extra labeled data. The library includes a preprocessing pipeline, visualization helpers, and a human-in-the-loop classifier submodule, plus Colab tutorials that make a first experiment easy. Teams doing text-as-data analysis, content organization, or exploratory NLP research will find a practical starting point, and swapping in different embeddings is straightforward.