Active Dialogue Simulation in Conversational Systems

Anonymous

Keywords: Conversational System, Semantic Parser, GPT-3, Active Learning, Dialogue Generation

TL;DR: Active dialogue simulation in conversational systems using GPT-3.

Abstract: Our goal is to utilize large language models and active learning to replace Wizard-of-Oz (WoZ) collection via crowdsourcing for bootstrapping training data for task-driven semantic parsers. We first demonstrate the utility of utterances generated by GPT-3 when seeded with prior training dialogues, as evaluated by human judges. We then explore two approaches for example selection: maximizing model (parser) uncertainty on generated outputs, and maximizing lexical diversity. We find that large language models can generate useful training data, and that there is a promising direction in active generation to maximize the impact of each such example.