Adversarial Learning on the Latent Space for Diverse Dialog Generation

PyTorch implementation of our paper Adversarial Learning on the Latent Space for Diverse Dialog Generation, accepted to COLING 2020, Barcelona, Spain (Online).

Installation

Install allennlp 0.9.0 using conda or pip pip install pyro-ppl

Running Experiments

• VAE allennlp train -s models/dialog_vae --include-package src experiments/vae.jsonnet
• GAN (run after VAE) allennlp train -s models/dialog_gan_mse --include-package src experiments/dialog_gen.jsonnet

Important Code

• experiments
  • vae.jsonnet (Configuration for VAE)
  • dialog_gen.jsonnet (Configuration for GAN with MSE)
• src
  • data
    • dataset.py (DatasetReaders)
    • iterator.py (Special iterator for GAN)
  • modules
    • encoders
      • variational_encoder.py (Variational Encoder for VAE)
    • decoders
      • decoder.py (Decoder base class)
      • variational_decoder.py (Variational Decoder for VAE)
    • metrics.py (LMPPL and NLTKSentenceBLEU)
    • annealer.py (Different Annealing Schedules)
  • models
    • vae.py (Variational Autoencoder, SampleGenerator)
    • dialog_generator.py (Generator Module for GAN)
    • dialog_discriminator.py (Discriminator Module for GAN)
    • dialog_gan.py (Dialog Gan, GanOptimizer and SampleGenerator)

If you find the code useful, please cite:

@inproceedings{khan-etal-2020-adversarial,
title = "Adversarial Learning on the Latent Space for Diverse Dialog Generation",
author = "Khan, Kashif  and
Sahu, Gaurav  and
Balasubramanian, Vikash  and
Mou, Lili  and
Vechtomova, Olga",
booktitle = "Proceedings of the 28th International Conference on Computational Linguistics",
month = dec,
year = "2020",
address = "Barcelona, Spain (Online)",
publisher = "International Committee on Computational Linguistics",
url = "https://www.aclweb.org/anthology/2020.coling-main.441",
pages = "5026--5034",
abstract = "Generating relevant responses in a dialog is challenging, and requires not only proper modeling of context in the conversation, but also being able to generate fluent sentences during inference. In this paper, we propose a two-step framework based on generative adversarial nets for generating conditioned responses. Our model first learns a meaningful representation of sentences by autoencoding, and then learns to map an input query to the response representation, which is in turn decoded as a response sentence. Both quantitative and qualitative evaluations show that our model generates more fluent, relevant, and diverse responses than existing state-of-the-art methods.",
}