← Back to all sparks
Q

quanteda.textmodels

ANALYTICS
Velocity0.0

R package quanteda.textmodels by quanteda — release notes from GitHub.

Split out of quanteda, then quiet - one new classifier since 2020.

r-packagetext-classificationnlpquantedamaintenance
Current state
quanteda.textmodels holds the scaling and classification models factored out of quanteda's main package. The visible history is thin: a logistic regression classifier and a native C++ rewrite of svmlin in late 2020, an SVM default change in early 2021, and after that only compatibility work. The most recent release fixes a namespace break caused by quanteda 4.1.0 dropping RcppArmadillo.
Where it's heading
The package now moves when its parent or a dependency moves, not on its own schedule. Four of the six most recent releases exist to track changes in quanteda, Matrix, or CRAN policy. The modelling decisions that were made - defaulting textmodel_svm() to the L2-regularized L2-loss dual solver, reducing svmlin to a single algorithm - have not been revisited since.
Prediction
The next release most likely follows another upstream change in quanteda or a Matrix and Rcpp dependency rather than adding a model.

Recent moves

  1. 1y ago

    Namespace fixed after quanteda dropped RcppArmadillo

    The clearest illustration of the dependency relationship: the parent package removed a compiled dependency and this one had to be republished to keep loading.

    View source ↗
  2. 3y ago

    CRAN issues and documentation fixed

    A one-line release note covering CRAN checks and documentation, with no modelling change - two years after the previous substantive work.

    View source ↗
  3. 3y ago

    Compatibility with Matrix 1.4.2

    A sparse-matrix dependency bump absorbed with no user-facing change, part of the run of upstream-driven releases that fills the package's recent history.

    View source ↗
  4. 5y ago

    Duplicate example dfm removed

    Housekeeping from the split: a dataset that already lived in quanteda was dropped from this package rather than shipped twice.

    View source ↗
  5. 5y ago

    SVM default switched to the L2-regularized dual solver

    Changes what textmodel_svm() fits by default and adds a type argument to override it. A real behavioural change for anyone who ran the function without arguments, and the last modelling decision in this feed.

    View source ↗
  6. 5y ago

    Logistic regression classifier added; svmlin rewritten in C++

    ⚡ SPARK

    The high-water mark of the package's own development. Every release after it either tunes what shipped here or repairs it against a moving dependency.

    View source ↗