Topic Modeling: Documents as Mixtures of Topics

Topic models such as LDA assume every document is built from a small number of hidden themes: a document-specific mixture of topics, and each topic itself a fixed distribution over words. This simulator makes that generative process visible — a topic-proportion vector θ is drawn from a Dirichlet(α) distribution, then word-tokens are sampled one at a time, each first picking a topic from θ and then a word from that topic's word-cloud, flying across the scene to assemble the document in real time. Adjust the number of topics, the concentration α, and the document length to see how peaked or blended real documents can look, and how much a short document's word counts can drift from the θ that generated it.