Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Adversarial learning in statistical dialogue systems

Dialektakis Georgios

Full record

---

Summary

In the past few years, the machine learning community has shifted its attention in Generative Adversarial Networks (GANs) and has shown their enormous potential in image, video and audio generation. Nevertheless, they haven't been used widely in the field of Spoken Dialogue Systems (SDS). In this work, we investigate a novel use of GANs in the field of SDS. Drawing intuition from recent related work, we investigate the use of a form of GANs, the Adversarial Autoencoder (AAE), as we want to explore efficient Belief State (BS) space representations through generative adversarial modeling. We review the difficulties that arise when training a GAN and we propose techniques to improve the training process. In particular, we propose the use of the Wasserstein Adversarial Autoencoder (WAAE), which is based on the Wasserstein loss, and we investigate its effectiveness compared to the baseline AAEs. We also examine the efficiency of the Denoising Adversarial Autoencoder (DAAE) in noisy environments. To evaluate our models, we implemented our algorithms in the PyDial toolkit and we performed several experiments employing two Reinforcement Learning (RL) algorithms, GP-SARSA and LSPI. These two algorithms receive the BS representation from the AAE and optimize the dialogue policy. Our experiments confirm the ability of the generative adversarial modeling to robustly represent the BS space, since the proposed method exhibits state-of-the-art performance, particularly in environments with high levels of noise.

Available Files

Services