A Hybrid Convolutional Variational Autoencoder for Text Generation

[flrngel]

https://arxiv.org/abs/1702.02390

Abstract

• this model uses rnn + feed-forward convolutional architecture
• author claims this model has
  • fast train convergence
  • handles long sequences
  • avoid major difficulties from VAE on textual data

1. Introduction

• author claims this model is
  • first work that applies deconvolution in latent variable generative model of natural text
• this paper discuss
  • optimization difficulties of VAEs for text
    • propose effective ways to address them

2. Related Work

• Techniques to improve VAE training
  • KL-term annealing and input dropout
  • imposing structured sparsity on latent variables

3. Model

• this model respect latent variables that can be able to output sample realistic sentences with latent space

3.1. Variational Autoencoder

• VAE forces the model to map an input to a region of the space

3.2. Deconvolutional Networks

• Deconvolutional layer's goal
  • to perform an inverse convolution operation and increase spatial size of the input while decreasing the number of feature maps.
• Deconvolutional layer's benefit
  • efficient GPU implementation that makes fully parallel
  • feed-forward is typically easier to optimize than using recurrent counterparts

3.3. Hybrid Convolutional-Recurrent VAE (paper model)

• VAE + RNN architecture
• RNN is for consuming deconvolutional decoder to have dependency with previous outputs which is,
  instead of
• encode every detail of a text fragment instead of high level feature (like semantic)

3.4. Optimization Difficulties

• input dropout helped
• add auxiliary reconstruction term computed from last deconvolutional layer,
• finally,
• autoregressive part reuses these features

4. Experiments

4.1. Comparision with LSTM VAE

Historyless decoding

• paper model's historyless decoding was better
• computationally faster (factor 2)

Decoding with history

• Paper checks
  • is historyless decoding generallizes well?
  • how model copes latent variable well
• paper claims their model does not fail on long texts

4.2. Controlling the KL term

Aux cost weight

• using input dropout increases final loss but this is trade off
• note that model finds non-trivial latent vectors when a is large enough

Receptive field

• goal is to study the relationship between KL term values and expressiveness of decoder
  -RNN decoder in LSTM VAE completely ignore information on latent vector
• Aux helps as Figure 6