KeyBERT

KeyBERT takes a deliberately minimal approach to keyword and keyphrase extraction: it embeds a document with a BERT style model, embeds candidate n-grams from the same text, and ranks the candidates by cosine similarity to the document embedding, so the extracted phrases are the ones most semantically representative of the whole. Maarten Grootendorst's library adds two diversification strategies, Maximal Marginal Relevance and Max Sum Distance, to avoid returning near duplicate phrases. It supports embedding backends including Sentence-Transformers, Flair, spaCy, Gensim, and lightweight Model2Vec models, while the companion KeyLLM module performs keyword extraction with large language models, batching similar documents for efficiency. Useful results take a pip install and about three lines of Python, which is why data scientists reach for it when heavier NLP pipelines are overkill. Released under the MIT license.