Paraphrase generation based on LSTM-VAE by wanzeyu

Overview

Keras implementation for A Deep Generative Framework for Paraphrase Generation.

Resource Used:

1. MSRP paraphrase corpus

2. Fasttext's pretrained vector

Requirements:

1. Keras
2. Numpy

In this project I try to implement novel VAE-LSTM architecture mentioned in A Deep Generative Framework for Paraphrase Generation.

QuickStart

Use python example.py. The code will train on the corpus and print predicted result after every epoch.

By default ,the program will be in training mode and will save model to three files.

The default epoch is 200.

If you want to change the parameters. You have to change according lines.

Resource

• test_source.txt is the original file
• test_target.txt is the paraphrase file
• wiki.simple.vec is fasttext's pretrained vector on simple wiki

Model Architecture

encoder model:

decoder model:

vae overview:

Problems

1. I didn't write a proper test function or use BLEU to evaluate. This has to be done after I found out what's wrong with my implementation.

Progress

• [x] Implement the basic framework of the thesis
• [ ] Write the evaluation code
• [x] Refactor and clean up the messy code

References

• Keras implementation of LSTM Variational Autoencoder
• Toni-Antonova/VAE-Text-Generation
• A Deep Generative Framework for Paraphrase Generation
• Building Autoencoders in Keras
• Generating sentences from a continuous space