Compositional Generalization Requires Compositional Parsers

Anonymous

Abstract: A growing body of research has focused on the task of \textit{compositional generalization}, the ability of a semantic parser to dynamically combine known linguistic elements in novel structures. We analyze the accuracy of different parsers on the recent COGS corpus (Kim and Linzen, 2020). While lexical generalization tasks are solvable by almost all existing models, tasks involving changes to the linguistic structure are hard for even the best sequence-to-sequence models. Structural generalization tasks can be solved with models that have compositionality built in; we present new results confirming this from the AM parser (Groschwitz et al., 2021). We further analyze the role of syntactic generalization in compositional generalization, and we discuss ramifications for the design of both semantic parsers and compositional generalization datasets.