carpedm20/lstm-char-cnn-tensorflow

Combines character-level and word-level CNN feature extraction with highway networks feeding into an LSTM language model, enabling the network to learn subword patterns while maintaining long-range dependencies. Implements configurable multi-kernel convolutions (kernels 1-7 with 50-200 feature maps) over character embeddings before word representation, allowing flexible hybrid architectures toggled via `--use_char` and `--use_word` flags. Built in TensorFlow with support for Penn Treebank and other datasets, offering both LSTM and LSTMTDNN model variants with full hyperparameter control for embedding dimensions, sequence length, and dropout.

780 stars. No commits in the last 6 months.