Keras TextClassification

Keras TextClassification is an open Python codebase from the Chinese NLP blogger Macropodus (github.com/yongzhuo) that collects dozens of neural architectures for Chinese text classification and adjacent tasks under one training loop. The model zoo spans classic convolutional and recurrent designs — FastText, TextCNN, CharCNN, TextRNN, RCNN, DPCNN — alongside transformer encoders including BERT, XLNet and ALBERT, with CapsuleNet and Transformer-encoder variants in the same tree. Scope reaches beyond single-label classification into multi-label problems, sentence-pair similarity, text regression and short-text summarization. Architectures are assembled from three reusable pieces (embedding layers, graph layers and full models), so you can swap character, word or subword embeddings and load pre-trained weights without rewriting the training script. Two companion packages cover preprocessing that most tutorials skip. macro-correct performs Chinese spelling and punctuation correction with a deliberately small dependency set of pytorch, transformers, numpy and opencc; its macbert4mdcspell_v3 model, per the author's February 2026 write-up, was trained on 20 million mixed samples (including classical Chinese and textbook data) and shows improved F1 with fewer over-corrections. near-synonym generates Chinese antonyms, synonyms and near-synonyms via word2vec plus NLI, or through prompt-based BERT-MLM fine-tuning with beam search. Distribution is deliberately low-tech: source code published through CSDN articles and GitHub rather than a managed product. That fits researchers, graduate students and ML engineers who want to read and modify architectures. If a Chinese pipeline touches classification, spelling correction and augmentation, this covers three stages in one place — teams that need a hosted endpoint or commercial support should look at cloud NLP APIs instead.