Implementation of the character-based language model proposed in the paper Character-Aware Neural Language Models (AAAI 2016) using the Keras neural networks library.
The code is based on the original LUA source code for the Torch library.
The code is compatible with Python 2.7-3.6 and requires Keras >= 2.1.2.
Data should be split into train.txt, valid.txt, and test.txt
Each line of the .txt file should be a sentence. The English Penn Treebank (PTB) data (Tomas Mikolov's pre-processed version with vocab size equal to 10K, widely used by the language modeling community) is given as the default.
You can reproduce the results of the paper as follows
Large character-level model (LSTM-CharCNN-Large in the paper). This is the default: should get ~82 on valid and ~79 on test. Takes ~3.5 hours with Theano (GPU/CuDNN).
python train.py --savefile char-large
Large word-level model (LSTM-Word-Large in the paper). This should get ~89 on valid and ~85 on test.
python train.py --savefile word-large --highway_layers 0 --use_chars 0 --use_words 1
By default train.py will evaluate the model on test data after training using the last epoch's model, and also will be slow due to
the way the data is set up.
Evaluation can be performed via the following script:
python evaluate.py --model cv/char-large --vocabulary data/ptb/vocab.npz --init init.npy --text data/ptb/test.txt --calc
With the --calc option the state of the network is not reset after each sentence, and the mean value of the initial state is saved in the --init file. Using this cross-sentence information helps in the case of the provided PTB data but it is not useful for sentences in random order.
For this later case, we can evaluate the perplexity using a precomputed --init file as the initial state of the LSTM networks at the beginning of each sentence
python evaluate.py --model cv/char-large --vocabulary data/ptb/vocab.npz --init init.npy --text data/ptb/test.txt
MIT