Deep Learning and Natural Language Processing (Group 1)
This repository contains a partial implementaion of the model presented in Mousa & Schuller (2017) implemented in Python using Keras and TensorFlow.
To train the models, you need to download the Stanford IMDB dataset, found here, and the pretrained GloVe vectors, found here. Unpack the dataset and the vectors in the repository folder.
Make sure you have all the necesarry libraries to run it, notably Keras, TensorFlow, Nltk, Sklearn and Numpy, but Python will let you know if you're missing something.
Before training run mkdir data and mkdir model to make sure you have the directories to store the data and the models.
To train the cBLSTM language models, open up cBLSTM-Language-Model.py and make sure the top parameters are set correctly. Here you choose some sizes for the network, whether to train the positive or the negative LM, and chose a filename. Then run the file.
To train the BLSTM binary classifier, open up BLSTM-Binary-Classifier.py and set the parameters and filename. Then run the file.
Models are saved along the way in the model directory. Datasets and embedding matrices are saved in the data and can be resued by out commenting their generation in each Python-file.
Open up Combined-classifier.py and set the filenames of the three models to be used. All models are assumed to be in the model folder. Run the file.
All of the data preprocessing is included in the prepros.py-file and a range of utility functions are included in the utils.py, from embedding matrix generation, perplexity calculation, and dataset-splitting and -compiling.