What you can do

topica is a general topic-modeling toolkit. This section tours what it does. If your goal is a publishable analysis, pair it with Publishing in a journal.

• :material-shape: The models

  Thirteen model families, from LDA through STM, HDP, dynamic and supervised topics, to short-text models, all with one consistent API.

• :material-broom: Preprocessing

  Tokenize, build a Corpus, prune the vocabulary, detect phrases, and split long documents while preserving metadata.

• :material-chart-bell-curve: Covariates & STM

  Relate topics to document metadata: prevalence and content covariates, effect estimation, clustered SEs, GLM links.

• :material-check-decagram: Diagnostics & validation

  Coherence, exclusivity, intrusion tests, stability, alignment, FREX labels, and pyLDAvis, all model-agnostic.

• :material-compare: Distinguishing words

  Fighting Words: which words separate two corpora, with significance.

• :material-message-text: Short text

  Models built for tweets, headlines, and survey answers (PT, GSDMM).

• :material-arrow-right-circle: Held-out inference

  transform new documents onto a fitted model across every model family.

Everything returns NumPy arrays, fits are deterministic for a fixed seed, and the variational models parallelize across cores automatically.