KeyBERT Keyword Extraction

KeyBERT is a keyword extraction method that embeds a document and its candidate words or phrases with a BERT-based model and selects the candidates whose embeddings are most similar to the document.

KeyBERT는 BERT 기반 모형으로 문서와 후보 단어·구를 임베딩한 뒤 문서 임베딩과 가장 유사한 후보를 선택하는 키워드 추출 방법이다.

How do I extract the most representative keywords from texts using language model embeddings?

언어모형 임베딩을 이용해 텍스트에서 가장 대표적인 키워드를 어떻게 추출하나요?

KeyBERT generates candidate n-grams, embeds them and the document with a sentence-transformer model, and ranks candidates by cosine similarity to the document embedding. It fits extracting semantically representative keywords from short to medium texts, and maximal marginal relevance can be used to make keywords more diverse.

When to use KeyBERT Keyword Extraction

• Summarizing documents, abstracts or open-ended responses with keywords
• Keyword extraction where semantic meaning matters more than frequency
• Creating labels or tags for a collection of texts

When to use something else

• You need corpus-level themes across documents: use LDA or BERTopic topic modeling
• Simple frequency-based weighting is sufficient: use TF-IDF keyword ranking

Assumptions and requirements

• Text in a language supported by the embedding model
• Candidate n-gram range and stop-word handling specified
• Number of keywords per document chosen in advance
• Diversity setting (MMR or Max Sum) when redundancy is a concern

What the analysis reports

• Top keywords per document
• Similarity score for each keyword
• Aggregated keyword frequencies across the corpus

How to report it

KeyBERT extracted the top [k] keyphrases per document, and the most frequent across the corpus were [keyword 1], [keyword 2] and [keyword 3].

Related methods

• TF-IDF
• BERTopic
• LDA Topic Modeling

Run KeyBERT Keyword Extraction on Quorant