Rapid generalization in phonotactic learning

Abstract

Speakers judge novel strings to be better potential words of their language if those strings consist of sound sequences that are attested in the language. These intuitions are often generalized to new sequences that share some properties with attested ones: Participants exposed to an artificial language where all words start with the voiced stops [b] and [d] will prefer words that start with other voiced stops (e.g., [g]) to words that start with vowels or nasals. The current study tracks the evolution of generalization across sounds during the early stages of artificial language learning. In Experiments 1 and 2, participants received varying amounts of exposure to an artificial language. Learners rapidly generalized to new sounds: In fact, following short exposure to the language, attested patterns were not distinguished from unattested patterns that were similar in their phonological properties to the attested ones. Following additional exposure, participants showed an increasing preference for attested sounds, alongside sustained generalization to unattested ones. Finally, Experiment 3 tested whether participants can rapidly generalize to new sounds based on a single type of sound. We discuss the implications of our results for computational models of phonotactic learning.

Inductive learning of locality relations in segmental phonology

“…For example, the Universal Grammar (UG) notion familiar from generative models amounts to a set of domain-specific and (largely) categorical inductive biases, delimiting the space of possible (i.e., learnable) linguistic systems. In recent years, a growing body of research has used artificial grammar learning experiments (Reber, 1967(Reber, , 1989 to probe for inductive biases in a controlled laboratory setting (e.g., Baer-Henney, Kügler, & van de Vijver, 2015;Baer-Henney & van de Vijver, 2012;Carpenter, 2010;Cristia, Mielke, Daland, & Peperkamp, 2013;Finley, 2011Finley, , 2012Finley, , 2015Finley & Badecker, 2009;Gallagher, 2013;Kapatsinski, 2010;Kuo, 2009;Lai, 2015;Linzen & Gallagher, 2017;Moreton, 2008Moreton, , 2012Peperkamp & Dupoux, 2007;Pycha, Nowak, Shin, & Shosted, 2003;White, 2014;Wilson, 2003Wilson, , 2006 for overviews, see Culbertson, 2012;Moreton & Pater, 2012a, 2012b. In the literature on artificial phonology learning, two main types of inductive biases have been proposed, which appear to influence the ease with which a learner is able to acquire sound patterns.…”

There is a simplicity bias when generalising from ambiguous data

“…There are four other issues that we wish to touch upon. First, how do we square our results with those of Gerken (2006) and Linzen & Gallagher (2014, 2017, who argue that their results suggest that learners might be acquiring more complex (or specific) generalisations? As a reviewer points out, it is reasonable to reinterpret Gerken's (2006) results as showing that learners need stimulus variation to infer a more abstract generalisation.…”

“…Furthermore, in each of the above papers the comparisons were over generalisations with different representational primitives. For example, Linzen & Gallagher's (2014, 2017 specific generalisation involved segments, while the simpler one involved features. However, the specific one is only more complex if we assume that segments are not themselves incremental or algorithmic models of learning, but rather computational-level models; this is worth further investigating.…”

“…A second relevant study is that of Linzen & Gallagher (2014, 2017. The authors were interested in the time-course of generalisation, but some of the results in their experiments are relevant to the question of how generalising proceeds when the data is consistent with multiple different generalisations.…”

Vowel but not consonant identity and the very informal English lexicon

“…Recent experimental findings from Linzen & Gallagher (2017) also suggest that total consonantal identity of the Aymara type are very quickly learned in an artificial language learning context. Do these attested examples differ in kind from the consonant identity pattern we did not observe in shitgibbons?…”