Compositional Task-Oriented Parsing as Abstractive Question Answering

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4418 - 4427 July 10-15, 2022 ©2022 Association for Computational Linguistics Compositional Task-Oriented Parsing as Abstractive Question Answering Wenting Zhao1 ∗and Konstantine Arkoudas2 and Weiqi Sun2 and Claire Cardie1 1Cornell University 2Amazon Alexa AI {wzhao,cardie}@cs.cornell.edu, {arkoudk,weiqisun}@amazon.com Abstract Task-oriented parsing (TOP) aims to convert natural language into machine-readable repre- sentations of specific tasks, such as setting an alarm. A popular approach to TOP is to apply seq2seq models to generate linearized parse trees. A more recent line of work argues that pretrained seq2seq models are better at gener- ating outputs that are themselves natural lan- guage, so they replace linearized parse trees with canonical natural-language paraphrases that can then be easily translated into parse trees, resulting in so-called naturalized parsers. In this work we continue to explore naturalized semantic parsing by presenting a general re- duction of TOP to abstractive question answer- ing that overcomes some limitations of canon- ical paraphrasing. Experimental results show that our QA-based technique outperforms state- of-the-art methods in full-data settings while achieving dramatic improvements in few-shot settings. 1 Introduction Task-oriented parsing (TOP) takes an utterance as input and generates an unambiguous specification of a task that can be executed by machine (Gupta et al., 2018). Traditional approaches to TOP treat the task as an instance of slot filling (Liu and Lane, 2016), first classifying the intent of the utterance as a whole and then tagging tokens with slot la- bels that identify relevant entities (such as numbers, persons, dates or times, organizations, and so on). However, this approach only works for simple ut- terances that have flat rather than compositional semantics. That is, slot-filling approaches cannot produce nested (or “hierarchical”) meaning repre- sentations, such as the one shown in Fig 1, where slots and intents can be composite and contain other slots (or intents) as proper parts. A straightforward way to handle compositional semantics is to formulate TOP as a seq2seq prob- ∗Work done during an internship at Amazon Alexa AI. lem, where the input sequence is the utterance and the output sequence is a linearized represen- tation of a semantic tree (shown in the bottom part of Fig 1). The recent development of high- performing transformer-based pretrained language models (PLMs) (Lewis et al., 2020; Raffel et al., 2020; Brown et al., 2020) that can be fine-tuned on specific tasks (such as a particular TOP dataset) has made this formulation feasible. This approach has achieved state-of-the-art per- formance on a variety of TOP datasets (Rongali et al., 2020; Aghajanyan et al., 2020; Chen et al., 2020). However, the output sequences are typically not expressed in natural language but rather in a mixture of natural language and special symbols (such as [SL:) that were not seen during pretrain- ing and have no meaning to the PLMs. Depending on the task, the output sequences may hardly con- tain any natural language at all (for example, if the outputs …

Content truncated to 500 words