GenTUS: Simulating User Behaviour and Language in Task-oriented Dialogues with Generative Transformers

TL;DR

GenTUS is a transformer-based user simulator that jointly optimizes user policy and natural language generation, producing more natural, diverse dialogues and generalizing to unseen scenarios, thereby improving training for task-oriented dialogue systems.

Contribution

It introduces a generative transformer-based user simulator that jointly optimizes policy and language, enhancing naturalness and transferability in dialogue training.

Findings

GenTUS generates more natural language than baseline models.

It can transfer to unseen ontologies in a zero-shot manner.

GenTUS's behavior can be shaped with reinforcement learning.

Abstract

User simulators (USs) are commonly used to train task-oriented dialogue systems (DSs) via reinforcement learning. The interactions often take place on semantic level for efficiency, but there is still a gap from semantic actions to natural language, which causes a mismatch between training and deployment environment. Incorporating a natural language generation (NLG) module with USs during training can partly deal with this problem. However, since the policy and NLG of USs are optimised separately, these simulated user utterances may not be natural enough in a given context. In this work, we propose a generative transformer-based user simulator (GenTUS). GenTUS consists of an encoder-decoder structure, which means it can optimise both the user policy and natural language generation jointly. GenTUS generates both semantic actions and natural language utterances, preserving interpretability and enhancing language variation. In addition, by representing the inputs and outputs as word sequences and by using a large pre-trained language model we can achieve generalisability in feature representation. We evaluate GenTUS with automatic metrics and human evaluation. Our results show that GenTUS generates more natural language and is able to transfer to an unseen ontology in a zero-shot fashion. In addition, its behaviour can be further shaped with reinforcement learning opening the door to training specialised user simulators.