keras-bi-lm 0.21.0

Train the Bi-LM model and use it as a feature extraction method

Introduction

The repository contains a class for training a bidirectional language model that extracts features for each position in a sentence.

Install

pip install keras-bi-lm

Usage

Train and save the Bi-LM model

Before using it as a feature extraction method, the language model must be trained on a large corpora.

from keras_bi_lm import BiLM

sentences = [
    ['All', 'work', 'and', 'no', 'play'],
    ['makes', 'Jack', 'a', 'dull', 'boy', '.'],
]
token_dict = {
    '': 0, '<UNK>': 1, '<EOS>': 2,
    'all': 3, 'work': 4, 'and': 5, 'no': 6, 'play': 7,
    'makes': 8, 'a': 9, 'dull': 10, 'boy': 11, '.': 12,
}
token_dict_rev = {v: k for k, v in token_dict.items()}
inputs, outputs = BiLM.get_batch(sentences,
                                 token_dict,
                                 ignore_case=True,
                                 unk_index=token_dict['<UNK>'],
                                 eos_index=token_dict['<EOS>'])

bi_lm = BiLM(token_num=len(token_dict), embedding_dim=10, rnn_units=10)
bi_lm.model.summary()
bi_lm.fit(np.repeat(inputs, 2 ** 12, axis=0),
          [
              np.repeat(outputs[0], 2 ** 12, axis=0),
              np.repeat(outputs[1], 2 ** 12, axis=0),
          ],
          epochs=5)
bi_lm.save_model('bi_lm.h5')

BiLM()

The core class that contains the model to be trained and used. Key parameters:

• token_num: Number of words or characters.

• embedding_dim: The dimension of embedding layer.

• rnn_layer_num: The number of stacked bidirectional RNNs.

• rnn_units: An integer or a list representing the number of units of RNNs in one direction.

• rnn_keep_num: How many layers are used for predicting the probabilities of the next word.

• rnn_type: Type of RNN, ‘gru’ or ‘lstm’.

BiLM.get_batch()

A helper function that converts sentences to batch inputs and outputs for training the model.

• sentences: A list of list of tokens.

• token_dict: The dict that maps a token to an integer. <UNK> and <EOS> should be preserved.

• ignore_case: Whether ignoring the case of the token.

• unk_index: The index for unknown token.

• eos_index: The index for ending of sentence.

Load and use the Bi-LM model

from keras_bi_lm import BiLM

bi_lm = BiLM(model_path='bi_lm.h5')  # or `bi_lm.load_model('bi_lm.h5')`
input_layer, output_layer = bi_lm.get_feature_layers()
model = keras.models.Model(inputs=input_layer, outputs=output_layer)
model.summary()

The output_layer is the time-distributed feature and all the parameters in the layers of the model are not trainable.

Use ELMo-like Weighted Sum of Trained Layers

from keras_bi_lm import BiLM

bi_lm = BiLM(token_num=20000,
             embedding_dim=300,
             rnn_layer_num=3,
             rnn_keep_num=4,
             rnn_units=300,
             rnn_type='lstm')
# ...
# Train the Bi-LM model
# ...

input_layer, output_layer = bi_lm.get_feature_layers(use_weighted_sum=True)
model = keras.models.Model(inputs=input_layer, outputs=output_layer)
model.summary()

When rnn_keep_num is greater than rnn_layer_num, the embedding layer is also used for weighting.

Demo

See demo directory:

cd demo
./get_data.sh
pip install -r requirements.txt
python setiment_analysis.py

Citation

Just cite the paper you’ve seen.