A Dirichlet distribution describes probability mixtures: nonnegative weights that sum to one. In topic modeling, it represents how strongly a document combines several topics. section 4.5
Latent Dirichlet allocation (LDiA) uses this distribution as a prior for the topics assigned to documents and for the words assigned to topics. Each document becomes a mixture of an arbitrary number of selected topics. section 4.5
The document-generation analogy separates the modeling decisions:
- A Poisson distribution determines how many words the document contains. section 4.5.1
- A Dirichlet distribution determines how many topics to mix into the document. section 4.5.1
- A term-topic matrix supplies word probabilities for the selected topics. section 4.5.1
A document character sheet
Imagine rolling complicated dice to generate a document, like combining dice rolls to create a Dungeons & Dragons character sheet. One roll chooses the document's topic mixture; later steps select topic-appropriate words. section 4.5.1
