kcws

Why we included this project

Chinese is written without spaces between words, so a segmenter that handles the language properly is usually the first real dependency in a Chinese NLP stack, and kcws fills that spot with BiLSTM+CRF and IDCNN sequence models. The repo carries a trained word2vec vocabulary, a POS-tagging model, and a Bazel-built backend service exposing a segmentation API. The README documents the full pipeline, from preparing annotated corpora and character embeddings through training, freezing the graph, and serving over a web endpoint. A custom dictionary works at decode time: give domain terms weights and the segmenter keeps them intact instead of splitting them. Teams that want to reproduce the reference results or adapt segmentation and POS tagging to their own Chinese text will find the models, training scripts, and runnable service in one place.