NMF Topic Modeling

NMF topic modeling applies non-negative matrix factorization to a document-term matrix, typically TF-IDF weighted, to decompose it into a small set of topics defined by weighted words and document-topic weights.

NMF 토픽모델링은 보통 TF-IDF로 가중한 문서-단어 행렬에 비음수 행렬분해를 적용하여, 가중 단어로 정의되는 소수의 토픽과 문서-토픽 가중치로 분해하는 방법이다.

How do I find the main themes in a collection of open-ended text responses?

개방형 텍스트 응답 모음에서 주요 주제를 어떻게 찾나요?

NMF topic modeling factorizes a TF-IDF document-term matrix into topics, each described by its highest-weighted words, and gives each document a weight on every topic. It fits short texts such as survey answers or reviews and often produces more distinct topics than LDA on such data.

When to use NMF Topic Modeling

• A text column such as open-ended responses, reviews or abstracts
• Discovering a small number of recurring themes
• Short documents where TF-IDF weighting helps

When to use something else

• A probabilistic model of topic proportions is required; use latent Dirichlet allocation
• Topics based on contextual meaning; use BERTopic

Assumptions and requirements

• Number of topics chosen in advance
• Preprocessing such as stop-word removal and tokenization
• Enough documents to estimate stable topics
• Non-negative input matrix

What the analysis reports

• Top words for each topic
• Document-topic weight matrix
• Topic coherence scores
• Representative documents for each topic

How to report it

NMF identified [k] topics with a mean coherence of [value]; the largest topic, characterized by [words], accounted for [value]% of documents.

Related methods

• Latent Dirichlet Allocation
• BERTopic
• Latent Semantic Analysis

Run NMF Topic Modeling on Quorant