An argument from acquisition: Comparing English metrical stress representations by how learnable they are from child-directed speech

Lisa Pearl, Timothy Ho, & Zephyr Detrano

University of California, Irvine

Abstract

One (often implicit) motivation for a theory of knowledge representation (KR) comes from an argument from acquisition, with the idea that language acquisition is straightforward if children’s hypothesis space is defined by the correct KR. Acquisition is then the process of selecting the correct grammar from that hypothesis space, based on language input. To compare KR theories, we establish quantitative acquisition- based metrics that assess learnability from child-directed speech data. We conduct a learnability analysis for three KR theories proposed for metrical phonology and evalu- ate them on English, a language that is notoriously irregular with respect to metrical phonology and therefore non-trivial to learn. We find that all three KR theories have similar learnability potential, but the proposed English grammars within each theory are not the grammars able to account for the most English child-directed speech data. This suggests learnability issues exist for the proposed English grammar in all three theories if a learner is attempting to learn a grammar that accounts for as much acqui- sitional intake data as possible. We discuss ways a learner may still be able to learn the English grammar from English child-directed speech by incorporating (i) additional useful linguistic knowledge about English metrical phonology interactions and (ii) bi- ases to selectively learn from the input. We additionally discuss which aspects of the proposed English grammars are hurting learnability, observing that small changes in parameter values or constraint rankings lead to significantly better learnability results.