Laziness Is a Virtue When It Comes to Compositionality in Neural Semantic Parsing

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 8434–8448 July 9-14, 2023 ©2023 Association for Computational Linguistics Laziness Is a Virtue When It Comes to Compositionality in Neural Semantic Parsing Maxwell Crouse, Pavan Kapanipathi, Subhajit Chaudhury, Tahira Naseem, Ramon Astudillo, Achille Fokoue, Tim Klinger { maxwell.crouse, subhajit, ramon.astudillo }@ibm.com { kapanipa, tnaseem, achille, tklinger }@us.ibm.com IBM Research Abstract Nearly all general-purpose neural semantic parsers generate logical forms in a strictly top- down autoregressive fashion. Though such sys- tems have achieved impressive results across a variety of datasets and domains, recent works have called into question whether they are ulti- mately limited in their ability to composition- ally generalize. In this work, we approach se- mantic parsing from, quite literally, the oppo- site direction; that is, we introduce a neural semantic parsing generation method that con- structs logical forms from the bottom up, be- ginning from the logical form’s leaves. The system we introduce is lazy in that it incre- mentally builds up a set of potential seman- tic parses, but only expands and processes the most promising candidate parses at each gen- eration step. Such a parsimonious expansion scheme allows the system to maintain an ar- bitrarily large set of parse hypotheses that are never realized and thus incur minimal compu- tational overhead. We evaluate our approach on compositional generalization; specifically, on the challenging CFQ dataset and three Text- to-SQL datasets where we show that our novel, bottom-up semantic parsing technique outper- forms general-purpose semantic parsers while also being competitive with comparable neural parsers that have been designed for each task. 1 Introduction Compositionality is inherent to natural language, with the meaning of complex text or speech un- derstood through the composition of constituent words and phrases (Montague, 1973). For instance, having observed the usage of phrases like "edited by" and "directed by" in isolation, a human would be able to easily understand the question "Was Toy Story edited by and directed by John Lasseter?". The ability to take individual components and com- bine them together in novel ways is known as com- positional generalization, and is a key feature of human intelligence that has been shown to play a significant role in why humans are so efficient at learning (Lake et al., 2017). Compositional generalization has been identified as a major point of weakness in neural methods for semantic parsing (Lake and Baroni, 2018; Hig- gins et al., 2018). Accordingly, this deficiency has been taken up as a challenge by the machine learn- ing community, leading to a slew of methods (Liu et al., 2021; Herzig et al., 2021; Gai et al., 2021) and datasets (Keysers et al., 2019; Kim and Linzen, 2020) targeted towards compositional generaliza- tion. However, while there has been progress in determining factors that allow systems to composi- tionally generalize (Furrer et al., 2020; Oren et al., 2020), there yet remains gaps in our understanding as to why these neural models fail. In this paper, we …

Content truncated to 500 words