Generative-Language-Models

A library using variational autoencoders to generate sentences. The repo contains two models for doing this: the Bowman model and the Prototype and Edit model. For in-depth information about this work please read thesis.pdf.

Installation

pip install tensorflow
pip install tensorflow-probability
pip install nltk

Data

Both models require a file containing word embeddings. This path must be specified in the config file. Both files also require three files in tsv format: train, test and valid containing the data to train and evaluate the model.

Bowman Model Data

Set of sentences separated by new lines.

Prototype and Edit Model Data

A set of pairs of sentences: a source sentence and a target sentence which are separated with a tab. Each pair of sentences is separated with a new line.

Usage

Training

python -m bowman.main
python -m pe_model.main

Interaction

python -m bowman.interact

There are a number of functions in the interact file than can be used to perform different tasks with the trained model.