Efficient Zero-Shot Semantic Parsing with Paraphrasing from Pretrained Language Models

Anonymous

Abstract: Building a domain-specific semantic parser with little or no domain-specific training data remains a challenging task. Previous work has shown that crowdsourced paraphrases of synthetic (grammar-generated) utterances can be used to train semantic parsing models for new domains with good results. We investigate whether semantic parsers for new domains can be built with no additional human effort, obtaining paraphrases of grammar-generated utterances from large neural language models, such as Google's T5 and EleutherAI's GPT-J, as an alternative to crowd-sourcing. While our models trained with automated paraphrases generated by pretrained language models do not outperform supervised models trained with similar amounts of human-generated domain-specific data, they perform well in a zero-shot setting, where no domain-specific data is available for a new domain. Additionally, unlike the current state-of-the-art in zero-shot semantic parsing, our approach does not require the use of large transformer-based language models at inference-time. Using the Overnight dataset, we show that automated paraphrases can be used to train a semantic parsing model that outperforms or is competitive with state-of-the-art-models in the zero-shot setting, while requiring a small fraction of the time and energy costs at inference time.

Paper Type: long