NEWS

sts 1.6 (2026-10-23)

initialization = "stm" (including stmSeed): the sentiment-discourse coefficients kappa_s were always initialized to zero, because the STM topic coefficients were read where the content-covariate interaction coefficients were intended. kappa_t and kappa_s are now initialized from the STM's content-covariate effects (intercept and slope of a document-weighted least-squares fit on the centered levels of initializationVar), which reproduces the STM topic-word distributions exactly for a binary initializationVar. On the Gadarian example this gives a higher final ELBO and faster convergence; fits with initialization = "stm" will differ from sts 1.5. An informative error is now given if stmSeed has no content covariate matching initializationVar.

The package no longer depends on the stm package, which is scheduled for archival on CRAN. Apart from the bug fix above, results are numerically identical to sts 1.5 for the same random seed.
- The spectral (anchor word) initialization used by initialization = "anchor" (the default) and the STM fit used by initialization = "stm" are now implemented internally (R/stmPort.R, src/stmfuns.cpp), adapted from stm 1.3.8 (MIT license; the stm authors are credited in Authors@R and inst/COPYRIGHTS).
- New exported functions textProcessor(), prepDocuments() and make.heldout(), ported from stm, so that raw text can be prepared for sts() and held-out data created without stm.
- New data set gadarian (from stm) used in the examples.
- A prefit STM model can still be supplied through stmSeed.
sts() now gives an informative error for an unknown initialization value.
Updated the reference to Chen and Mankad (2025), Management Science 71(7).
The vignette is now registered through the R.rsp::asis engine.
Added SnowballC (used for stemming in textProcessor()) to Imports and removed stm, knitr and rmarkdown.