Topic modeling is a text-mining technique that finds the main themes in a large set of documents without labeled training data. Each topic is a group of words that tend to appear together, and each document is described as a mixture of topics. It is a method from natural language processing and is not specific to SEO.
How does it work?
IBM describes most approaches as treating documents as a bag of words: order is ignored, and the model works from how often terms occur. It builds a document-term matrix and groups co-occurring words into topics. A topic is a probability distribution over the vocabulary (Wikipedia), for example high probabilities for “crawl,” “index” and “sitemap”.
Two classic methods are Latent Semantic Analysis (LSA), which uses singular value decomposition, and Latent Dirichlet Allocation (LDA). Blei, Ng and Jordan introduced LDA in the Journal of Machine Learning Research in 2003. Newer approaches use neural networks and language model embeddings.
What is it used for in SEO?
SEOs apply it to a set of pages, such as a site’s articles or competitor content, to see which themes appear and where coverage is thin. This is a use of the method, not a Google ranking feature, and Google has not said it relies on topic models. The result needs interpretation: the algorithm outputs word lists, and a person names the topics.
Example
Running LDA with eight topics over 2,000 help articles might return a topic with the top words “invoice, payment, refund, card, charge.” A reader labels it “Billing.” The number of topics is chosen by the user, and IBM lists that choice as an open difficulty.
Sources
- Latent Dirichlet Allocation – Blei, Ng, Jordan, JMLR 2003jmlr.org
- What is topic modeling? – IBMibm.com
- Topic model – Wikipediaen.wikipedia.org
Last updated: