Few-Shot Natural Language Generation by Rewriting Templates

By Mihir Kale, Abhinav Rastogi. 2020

12 citations recorded by the source corpus.

Virtual assistants such as Google Assistant, Alexa and Siri enable users to interact with a large number of services and APIs on the web using natural language. The response generation module converts the actions generated by a policy module into a natural language utterance. Traditionally, template based approaches have been used for response generation in virtual assistants. However, such approaches are not feasible for commercial assistants, which need to support a large number of services. Defining templates for a large number of slot combinations for each of the services supported by large scale assistants becomes tedious. In this work, we propose a template rewriting method for Natural Language Generation (NLG), where the number of templates scales only linearly with the number of slots. A set of simple templates is used to convert actions into utterances, which are 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, while being much more sample efficient.

Subfields: Artificial Intelligence, Natural Language Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Natural Language Processing

Citations in the corpus

The GEM Benchmark: Natural Language Generation, its Evaluation and Metrics (2021, methodology): “We additionally reformat the dialog acts to directly conform to the format described in the paper (Kale and Rastogi, 2020).”
Continual Learning in Task-Oriented Dialogue Systems (2020, background): “In datasets such as SGD, the slot has binary values e.g., yes or no, and thus we exclude those from the count as in Kale and Rastogi (2020).”
Improving Text-to-Text Pre-trained Models for the Graph-to-Text Task (2020, methodology): “Similar with our property lexicalization technique, Kale and Rastogi (2020) explored having natural text templates for each unique slot in dialog acts.”
Getting to Production with Few-shot Natural Language Generation Models (2021, methodology): “Using templates alongside pre-trained language models for NLG has been recently introduced by Kale and Rastogi (2020), where templates for simple input scenarios are concatenated to form a template for a more complicated scenario.”