sts 1.6
Bug fix
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.
Removal of the stm
dependency
- 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.
sts 1.5