Learning Robust Dialog Policies in Noisy Environments

TL;DR

This paper presents a method for training robust dialog policies for virtual assistants in noisy environments using a user simulator and deep reinforcement learning, improving efficiency and success rates.

Contribution

It introduces a realistic user simulator for noisy voice interactions and demonstrates that learned policies outperform fixed rules in noisy conditions.

Findings

Simulated dialogs are indistinguishable from human dialogs.

Learned policies achieve similar success rates with fewer turns.

Robust policies perform well in noisy environments.

Abstract

Modern virtual personal assistants provide a convenient interface for completing daily tasks via voice commands. An important consideration for these assistants is the ability to recover from automatic speech recognition (ASR) and natural language understanding (NLU) errors. In this paper, we focus on learning robust dialog policies to recover from these errors. To this end, we develop a user simulator which interacts with the assistant through voice commands in realistic scenarios with noisy audio, and use it to learn dialog policies through deep reinforcement learning. We show that dialogs generated by our simulator are indistinguishable from human generated dialogs, as determined by human evaluators. Furthermore, preliminary experimental results show that the learned policies in noisy environments achieve the same execution success rate with fewer dialog turns compared to fixed rule-based policies.