Using Rational Models to Interpret the Results of Experiments on Accent Adaptation

Exposure to unfamiliar non-native speech tends to improve comprehension. One hypothesis holds that listeners adapt to non-native-accented speech through distributional learning—by inferring the statistics of the talker's phonetic cues. Models based on this hypothesis provide a good fit to incremental changes after exposure to atypical native speech. These models have, however, not previously been applied to non-native accents, which typically differ from native speech in many dimensions. Motivated by a seeming failure to replicate a well-replicated finding from accent adaptation, we use ideal observers to test whether our results can be understood solely based on the statistics of the relevant cue distributions in the native- and non-native-accented speech. The simple computational model we use for this purpose can be used predictively by other researchers working on similar questions. All code and data are shared.

Distributional learning of recursive structures

Languages differ regarding the depth, structure, and syntactic domains of recursive structures. Even within a single language, some structures allow infinite self-embedding while others are more restricted. For example, English allows infinite free embedding of the prenominal genitive -s, whereas the postnominal genitive of is largely restricted to only one level and to a limited set of items. Therefore, while the ability for recursion is considered as a crucial part of the language faculty, speakers need to learn from experience which specific structures allow free embedding and which do not. One effort to account for the mechanism that underlies this learning process, the distributional learning proposal, suggests that the recursion of a structure (e.g. X1’s-X2) is licensed if the X1 position and the X2 position are productively substitutable in the input. A series of corpus studies have confirmed the availability of such distributional cues in child directed speech. The present study further tests the distributional learning proposal with an artificial language learning experiment. We found that, as predicted, participants exposed to productive input were more likely to accept unattested strings at both one and two-embedding levels than participants exposed to unproductive input. Therefore, our results suggest that speakers can indeed use distributional information at one level to learn whether or not a structure is freely recursive.

Early phonetic learning without phonetic categories: Insights from large-scale simulations on realistic input

Before they even speak, infants become attuned to the sounds of the language(s) they hear, processing native phonetic contrasts more easily than nonnative ones. For example, between 6 to 8 mo and 10 to 12 mo, infants learning American English get better at distinguishing English and [l], as in “rock” vs. “lock,” relative to infants learning Japanese. Influential accounts of this early phonetic learning phenomenon initially proposed that infants group sounds into native vowel- and consonant-like phonetic categories—like and [l] in English—through a statistical clustering mechanism dubbed “distributional learning.” The feasibility of this mechanism for learning phonetic categories has been challenged, however. Here, we demonstrate that a distributional learning algorithm operating on naturalistic speech can predict early phonetic learning, as observed in Japanese and American English infants, suggesting that infants might learn through distributional learning after all. We further show, however, that, contrary to the original distributional learning proposal, our model learns units too brief and too fine-grained acoustically to correspond to phonetic categories. This challenges the influential idea that what infants learn are phonetic categories. More broadly, our work introduces a mechanism-driven approach to the study of early phonetic learning, together with a quantitative modeling framework that can handle realistic input. This allows accounts of early phonetic learning to be linked to concrete, systematic predictions regarding infants’ attunement.

Child-directed speech is statistically optimized for meaning extraction

The way infants manage to extract meaning from the speech stream when learning their first language is a highly complex adaptive behavior. This behavior chiefly relies on the ability to extract information from speech they hear and combine it with the external environment they encounter. However, little is known about the underlying distribution of information in speech that conditions this ability. Here we examine properties of this distribution that support meaning extraction in three different types of speech: child-directed speech, adult conversation, and, as a control, written language. We find that verb meanings in child-directed speech can already be successfully extracted from simple co-occurrences of neighboring words, whereas meaning extraction in the other types of speech fundamentally requires access to more complex structural relations between neighboring words. These results suggest that child-directed speech is ideally shaped for a learner who has not yet mastered the structural complexity of her language and therefore mainly relies on distributional learning mechanisms to develop an understanding of linguistic meanings.

Consistent Unsupervised Estimators for Anchored PCFGs

Learning probabilistic context-free grammars (PCFGs) from strings is a classic problem in computational linguistics since Horning ( 1969 ). Here we present an algorithm based on distributional learning that is a consistent estimator for a large class of PCFGs that satisfy certain natural conditions including being anchored (Stratos et al., 2016 ). We proceed via a reparameterization of (top–down) PCFGs that we call a bottom–up weighted context-free grammar. We show that if the grammar is anchored and satisfies additional restrictions on its ambiguity, then the parameters can be directly related to distributional properties of the anchoring strings; we show the asymptotic correctness of a naive estimator and present some simulations using synthetic data that show that algorithms based on this approach have good finite sample behavior.

What constrains distributional learning in adults?

One of the many remarkable features of human language is it's flexibility: during acquisition, any normally-developing human infant can acquire any human language, and during adulthood, language users quickly and flexibly adapt to a wide range of talker variation. Both language acquisition in infants and adaptation in adults have been hypothesized to be forms of distributional learning, where flexibility is driven by sensitivity to statistical properties of sensory stimuli and the corresponding underlying linguistic structures. Despite the similarities between these forms of linguistic flexibility, there are obvious differences as well, chief among them being that adults have a much harder time acquiring the same unfamiliar languages that they would have picked up naturally during infancy. This suggests that there are strong constraints on distributional learning during adulthood. This paper provides further, direct evidence for these constraints, by showing that American English listeners struggle to learn voice-onset time (VOT) distributions that are atypical of American English. Moreover, computational modeling shows that the pattern of distributional learning (or lack thereof) across different VOT distributions is consistent with Bayesian belief-updating, starting from prior beliefs that are very similar to the VOT distributions produced by a typical talker of American English. Together, this suggests that distributional learning in adults is constrained by prior experience with other talkers, and that distributional learning may be a computational principle of human language that operates throughout the lifespan.