Three Benchmarks for Distributional Approaches to Natural Language Syntax

Human language abilities are far richer than what is represented in the kinds of monolingual corpora that are standardly used to evaluate statistical models of language learning. This article summarizes a series of findings from language acquisition, crosslanguage typology, and language processing, that illustrate the challenges that any serious model of natural language syntax must meet. Even a putative ideal statistical learner of cooccurrences in corpora will struggle to meet the challenges of complexity, crosslanguage consistency, and causality, unless it is able to take advantage of the rich representational primitives motivated by linguistics and psycholinguistics.