Teaching 2014-era CNNs to classify Chinese news

What it does This repo implements two old-school neural approaches—a shallow Conv1D CNN and a two-layer GRU/LSTM RNN—to sort Chinese news into ten categories. It works at the character level, using a fixed-length sequence of 600 character IDs, and trains on a 50,000-sample slice of the THUCNews corpus.

The interesting bit The project is essentially a faithful translation of seminal 2014–2015 English text-classification papers to Chinese, but the real value lies in the boring parts it refuses to hide: vocabulary serialization, fixed-length padding, batch shuffling, and per-class F1 reporting. In an era of one-line transformers pipelines, this is a useful anatomy lesson.

Key highlights

• Character-level models for Chinese, avoiding English-style word tokenization.
• CNN achieves 96.04% test accuracy after just three epochs on the 10-class dataset.
• Includes preprocessing scripts (cnews_loader.py) and helpers to carve the THUCNews subset.
• Emits per-class precision, recall, and a full confusion matrix rather than a single top-line score.
• Uses tf.nn.conv1d instead of the more common 2D convolution workaround for text.

Caveats

• Targets TensorFlow 1.3+ and supports Python 2, so the code uses the graph-and-session APIs of a bygone era.
• The README truncates during RNN training logs and never publishes final RNN test metrics, leaving its performance relative to the CNN unclear.
• You must download the THUCNews subset manually; it is not bundled in the repository.

Verdict Worth a look if you are learning how classical deep-learning text classification works under the hood or need a baseline for Chinese news. If you want a modern, production-ready classifier, this is a museum piece.