lstm-char-cnn-tensorflow

Why we included this project

Yoon Kim's 2016 character-aware neural language model is a good way to see how a CNN over characters, a highway network, and an LSTM language model fit together, and this TensorFlow port keeps that architecture easy to follow. The CLI exposes the main hyperparameters directly, and it trains and evaluates on Penn Treebank out of the box, so you can run it and watch perplexity drop. The author is upfront that the paper's exact results were not reproduced and that the implementation has a known performance issue, so treat it as a study aid rather than something to deploy. If you are learning how subword and character-level representations work, it is a useful companion to the original Lua code.