In underthesea v9.2.9, we replaced the fasttext Python package (a wrapper around Facebook's C++ library) with a pure Rust implementation inside underthesea-core. The result: identical predictions, simpler installation, and one fewer C++ dependency in our stack.
In underthesea v9.2.9, we've completely rewritten the text classification pipeline using our Rust-based TextClassifier. This delivers up to 273x faster inference compared to the previous sklearn-based implementation.
In underthesea v9.2.5, we completed the migration from python-crfsuite to our native Rust implementation underthesea-core. This change resulted in a 20% performance improvement across all CRF-based NLP tasks, plus up to 10x faster training through systematic optimization.