Template Guided Text Generation for Task-Oriented Dialogue

TL;DR

This paper presents a schema-guided and template-based approach for natural language generation in task-oriented dialogue systems, improving coherence, robustness, and sample efficiency across multiple APIs.

Contribution

It introduces a novel combination of schema-guided prompts and template concatenation with language model rewriting for flexible, domain-independent NLG.

Findings

Outperforms strong baselines in automatic and human evaluations

Demonstrates robustness to out-of-domain inputs

Shows improved sample efficiency

Abstract

Virtual assistants such as Google Assistant, Amazon Alexa, and Apple Siri enable users to interact with a large number of services and APIs on the web using natural language. In this work, we investigate two methods for Natural Language Generation (NLG) using a single domain-independent model across a large number of APIs. First, we propose a schema-guided approach which conditions the generation on a schema describing the API in natural language. Our second method investigates the use of a small number of templates, growing linearly in number of slots, to convey the semantics of the API. To generate utterances for an arbitrary slot combination, a few simple templates are first concatenated to give a semantically correct, but possibly incoherent and ungrammatical utterance. A pre-trained language model is subsequently employed to rewrite it into coherent, natural sounding text. Through automatic metrics and human evaluation, we show that our method improves over strong baselines, is robust to out-of-domain inputs and shows improved sample efficiency.