Character-Aware Neural Language Models

Tensorflow implementation of Character-Aware Neural Language Models. The original code of author can be found here.

This implementation contains:

1. Word-level and Character-level Convolutional Neural Network

2. Highway Network

3. Recurrent Neural Network Language Model

The current implementation has a performance issue. See #3.

Prerequisites

• Python 2.7 or Python 3.3+

• Tensorflow

Usage

To train a model with ptb dataset:

$ python main.py --dataset ptb

To test an existing model:

$ python main.py --dataset ptb --forward_only True

To see all training options, run:

$ python main.py --help

which will print

usage: main.py [-h] [--epoch EPOCH] [--word_embed_dim WORD_EMBED_DIM]
              [--char_embed_dim CHAR_EMBED_DIM]
              [--max_word_length MAX_WORD_LENGTH] [--batch_size BATCH_SIZE]
              [--seq_length SEQ_LENGTH] [--learning_rate LEARNING_RATE]
              [--decay DECAY] [--dropout_prob DROPOUT_PROB]
              [--feature_maps FEATURE_MAPS] [--kernels KERNELS]
              [--model MODEL] [--data_dir DATA_DIR] [--dataset DATASET]
              [--checkpoint_dir CHECKPOINT_DIR]
              [--forward_only [FORWARD_ONLY]] [--noforward_only]
              [--use_char [USE_CHAR]] [--nouse_char] [--use_word [USE_WORD]]
              [--nouse_word]

optional arguments:
  -h, --help            show this help message and exit
  --epoch EPOCH         Epoch to train [25]
  --word_embed_dim WORD_EMBED_DIM
                        The dimension of word embedding matrix [650]
  --char_embed_dim CHAR_EMBED_DIM
                        The dimension of char embedding matrix [15]
  --max_word_length MAX_WORD_LENGTH
                        The maximum length of word [65]
  --batch_size BATCH_SIZE
                        The size of batch images [100]
  --seq_length SEQ_LENGTH
                        The # of timesteps to unroll for [35]
  --learning_rate LEARNING_RATE
                        Learning rate [1.0]
  --decay DECAY         Decay of SGD [0.5]
  --dropout_prob DROPOUT_PROB
                        Probability of dropout layer [0.5]
  --feature_maps FEATURE_MAPS
                        The # of feature maps in CNN
                        [50,100,150,200,200,200,200]
  --kernels KERNELS     The width of CNN kernels [1,2,3,4,5,6,7]
  --model MODEL         The type of model to train and test [LSTM, LSTMTDNN]
  --data_dir DATA_DIR   The name of data directory [data]
  --dataset DATASET     The name of dataset [ptb]
  --checkpoint_dir CHECKPOINT_DIR
                        Directory name to save the checkpoints [checkpoint]
  --forward_only [FORWARD_ONLY]
                        True for forward only, False for training [False]
  --noforward_only
  --use_char [USE_CHAR]
                        Use character-level language model [True]
  --nouse_char
  --use_word [USE_WORD]
                        Use word-level language [False]
  --nouse_word

but more options can be found in models/LSTMTDNN and models/TDNN.

Performance

Failed to reproduce the results of paper (2016.02.12). If you are looking for a code that reproduced the paper's result, see https://github.com/mkroutikov/tf-lstm-char-cnn.

The perplexity on the test sets of Penn Treebank (PTB) corpora.

| Name | Character embed | LSTM hidden units | Paper (Y Kim 2016) | This repo. | |:---------------:|:---------------:|:-----------------:|:------------------:|:-----------:| | LSTM-Char-Small | 15 | 100 | 92.3 | in progress | | LSTM-Char-Large | 15 | 150 | 78.9 | in progress |

Author

Taehoon Kim / @carpedm20