Improve Diverse Text Generation by Self Labeling Conditional Variational Auto Encoder

TL;DR

This paper introduces SLCVAE, a novel method that explicitly addresses the KL-vanishing problem in CVAEs for diverse text generation by using a self-labeling approach to improve expressiveness and diversity.

Contribution

The paper proposes an explicit objective and a labeling network to enhance CVAE diversity, along with a new dataset for diverse text generation research.

Findings

Significant improvement in generation diversity.

Comparable accuracy to state-of-the-art methods.

Effective mitigation of KL-vanishing problem.

Abstract

Diversity plays a vital role in many text generating applications. In recent years, Conditional Variational Auto Encoders (CVAE) have shown promising performances for this task. However, they often encounter the so called KL-Vanishing problem. Previous works mitigated such problem by heuristic methods such as strengthening the encoder or weakening the decoder while optimizing the CVAE objective function. Nevertheless, the optimizing direction of these methods are implicit and it is hard to find an appropriate degree to which these methods should be applied. In this paper, we propose an explicit optimizing objective to complement the CVAE to directly pull away from KL-vanishing. In fact, this objective term guides the encoder towards the "best encoder" of the decoder to enhance the expressiveness. A labeling network is introduced to estimate the "best encoder". It provides a continuous label in the latent space of CVAE to help build a close connection between latent variables and targets. The whole proposed method is named Self Labeling CVAE~(SLCVAE). To accelerate the research of diverse text generation, we also propose a large native one-to-many dataset. Extensive experiments are conducted on two tasks, which show that our method largely improves the generating diversity while achieving comparable accuracy compared with state-of-art algorithms.

Figures

Full text

Improve Diverse Text Generation by Self Labeled Conditional Variational Auto Encoder

Supplemental Materials