Changelog

Source: NEWS.md

faSTM 0.0.0.9000 (development)

A fast, modern Structural Topic Model for R: an stm-compatible fitting backend on a multithreaded Rust core (topica), plus an effects and tidy-workflow toolkit on top.

Fitting

• stm(): fast STM fit via topica’s Rust fit_ctm, returning a c("faSTM", "STM") object compatible with the stm package’s post-fit functions (labelTopics, plot, findThoughts, sageLabels, toLDAvis). Prevalence and content (SAGE) covariates, spectral / LDA / random / custom initialization, and a num_threads knob. ~6x faster than stm on the poliblog example (and ~13x on search_k), wall-clock to convergence.
• Multiple content covariates: content = ~ a + b crosses covariates into a saturated SAGE content model (one topic-word distribution per level combination), beyond stm’s single content variable.
• SVI kwargs (inference, batch_size, tau, kappa) plumbed through ahead of topica’s STM-SVI landing (topica #231); gated for prevalence/content for now.

Covariate effects (estimateEffect())

• Effects via the method of composition (per-document posterior draws pooled by Rubin’s rules), drawing from each document’s own Laplace covariance, so topic-estimation uncertainty propagates into the standard errors.
• Survey weights (weighted least squares) and cluster-robust standard errors (sandwich vcov with stm’s finite-sample correction).
• Random effects in prevalence: lme4-style (term | group) terms fit per draw and pooled.
• ame() for average marginal effects; effect_estimates() for tidy effect tables; combine for aggregate-topic effects.

Inspection & topic quality

• coherence() with Mimno, NPMI, and C_V measures; semantic_coherence(), exclusivity(), frex_scores(), topic_correlation(), topic_corr_graph().
• label_topics(), sage_labels(), find_thoughts(), topic_terms(), topic_proportions(), content_topics().

Model selection

• search_k(): parallelizes across K (bstewart/stm#262) and selects by held-out likelihood or any coherence measure; select_model(), many_topics(), make_heldout() / eval_heldout().

Tidyverse & out-of-sample

• broom generics: tidy() (beta / frex / gamma), glance(), augment(); predict() for new documents.
• as_corpus() reads prepared text from quanteda / tidytext; from_tidy().

Data

• poliblog: the CMU 2008 political-blog corpus (the stm vignette example).
• congress: 1,679 balanced U.S. House/Senate floor speeches, Congresses 100–111 (party × chamber × time), for the multiple-content-covariate and over-time demos.

Vignettes

• faSTM: the stm vignette, run on faSTM. The stm workflow, fit live.
• Beyond stm: faSTM’s extensions. The tools above, on the congress corpus.