chiragjn / deep-char-cnn-lstm

Deep Character CNN LSTM Encoder with Classification and Similarity Models

In Keras

Overall Idea:

• Convolve over character embeddings with different kernel sizes
• Concat them to get the char-word embedding
• Pass them through a Dense layer with Residual connection
• Optionally concat them with separate word embedding
• Pass sequence of obtained word embeddings through a LSTM encoder
• Train with a constrastive loss function (see References)

Work in Progress

• TODO: Add loading utils
• TODO: Add preprocessing and padding utils
• TODO: Add batching utils
• TODO: Add model training code
• TODO: Add model continue-training code
• TODO: Test Similarity implementation on Quora similar pair dataset
• TODO: Test Classification implementation on Kaggle Toxic internet comments dataset
• TODO: Tune Hyperparameters and try different modifications to architectures
• TODO: Take Hyperparameters using argparse
• TODO: Add tensorboard and tfdbg support

Example Usage:

from model import ClassifierModel, SimilarityModel

classifier = ClassifierModel(vocab_size=10000,
                             charset_size=100,
                             num_classes=5,
                             mode=ClassifierModel.MULTILABEL,
                             char_kernel_sizes=(3,),
                             encoder_hidden_units=128,
                             bidirectional=False)
classifier.compile_model()

similarity_model = SimilarityModel(vocab_size=10000,
                                   charset_size=100,
                                   num_negative_samples=1)
similarity_model.compile_model()

References:

Overall Idea

1. Siamese Recurrent Architectures for Learning Sentence Similarity (2016)

Encoder architecture heavily inspired from

1. Character-Aware Neural Language Models (2015), Kim et. al.
2. dpressel/baseline

Loss function taken from

1. A Latent Semantic Model with Convolutional-Pooling Structure for Information Retrieval (2014)

Other Contrastive Loss functions to try

1. StarSpace: Embed All The Things! (2017) Wu et. al.
2. Comparision of loss functions for deep embedding