Kashgari

Why we included this project

Teams that want a supervised NLP model without hand-writing every training loop can get a long way with Kashgari. It sits on tf.keras and provides ready-made pipelines for named entity recognition, part-of-speech tagging, and text classification, and you can swap plain Word2Vec for pretrained BERT or GPT-2 embeddings without rewriting your data loading or evaluation code. That makes it a practical baseline for sequence labeling, whether you are doing research or running a production service, and the codebase is readable enough to modify when a default architecture does not fit your data. The 2.0 release moved the project to TensorFlow 2, so the current version uses modern TF rather than legacy APIs, and models export in SavedModel format for serving. If you mostly work with Chinese or English text and want a solid baseline model in an afternoon, this removes a lot of the boilerplate.