Kashgari is an open-source transfer learning framework for natural language processing tasks such as text classification and sequence labeling. Built on tf.keras and released under the Apache-2.0 license, it has earned more than 2,300 GitHub stars; the codebase was last updated in 2024 and remains a compact, production-oriented NLP toolkit.
Kashgari’s premise is that most applied NLP work โ named entity recognition, sentiment and intent classification, part-of-speech tagging โ follows the same shape: embed text, encode it, decode labels. The framework wraps that shape in a small, consistent API with pluggable embeddings, including Word2Vec, BERT and GPT-2, so swapping the language representation behind a task takes a line of code. It handles the surrounding chores too: corpus loading, label management, training callbacks and model export, aiming at developers who need a working classifier or tagger in an afternoon rather than a research project.
Kashgari is free and open source under the Apache-2.0 license, with no paid edition.
Kashgari fits developers and data scientists who need dependable, classic NLP pipelines โ entity extraction, text classification โ especially for Chinese and multilingual text, without adopting a heavyweight framework.
Kashgari has been especially popular in the Chinese NLP community, where it is commonly used for Chinese named entity recognition, sentiment analysis and text classification with modest labeled datasets. Its embedding layer supports loading pretrained vectors and transformer checkpoints, and trained models can be saved and reloaded for serving with a few lines of code. For teams maintaining classic TensorFlow NLP pipelines, it remains a pragmatic, well-documented option that favors convention and simplicity over configurability.