A Role for Chunk Formation in Statistical Learning of Second Language Syntax

Abstract

Humans are remarkably sensitive to the statistical structure of language. However, different mechanisms have been proposed to account for such statistical sensitivities. The present study compared adult learning of syntax and the ability of two models of statistical learning to simulate human performance: Simple Recurrent Networks, which learn by predictive computation, and PARSER, which learns chunks as a byproduct of general principles of associative learning and memory. In the first stage, a semiartificial language paradigm was used to gather human data. In the second stage, a simulation paradigm was then used to compare the patterns of performance of the SRN and PARSER. After the human adults and the computational models were trained on sentences from the semiartificial language with probabilistic syntax, their learning outcomes were compared. Neither model was able to fully reproduce the human data, which may indicate less robust statistical learning effects in adults; however, PARSER was able to simulate more of the adult learning data than the SRN, suggesting a possible role for chunk formation in early phases of adult learning of second language syntax.

Establishing Evidence of Learning in Experiments Employing Artificial Linguistic Systems

“…1), enabling the researchers to argue that chance had been validated as a baseline for their particular experiments. However, in very similar studies, control participants have been found to deviate from chance overall (Rebuschat & Williams, 2012 , Exp. 2) or on specifi c items (Hamrick, 2014a ;Rebuschat & Williams, 2012 ;Rebuschat et al, 2015 ), making it untenable to consider the corresponding experimental groups' above-chance performance alone as evidence of target learning. Extrapolating out to the area of research as a whole, without a significant body of work establishing the precise circumstances under which participants will simply guess and perform at chance in the absence of training/learning, repeated fi ndings of nonchance performance in human controls underscore that it may be invalid, and therefore irresponsible, to employ chance as the sole baseline without direct experiment-specifi c evidence to support its use.…”

“…Considering, moreover, that SLA researchers tend to collect data from rather homogeneous samples (Plonsky, 2015 ), and that homogeneous samples are more likely to have internal correlations manifesting as systematic group-level behavior (Field, 2009 ), it seems reasonable to expect participants to show at least some similar response tendencies independently of training/learning, thereby making chance an inappropriate baseline. Additional examples of this (from Hamrick, 2013Hamrick, , 2014aRebuschat et al, 2015 ) will be presented in the following text in the sections on untrained and trained controls.…”

“…Research employing AGs/ALSs has repeatedly demonstrated that it is not only possible but indeed common for experimental participants to contrive nontarget patterns or internalize other unintended but real (e.g., partial) patterns during training. The development of nontarget "knowledge" sometimes makes no difference in participants' test scores (e.g., Hamrick, 2013 ), but it can sometimes boost (e.g., Hamrick, 2014a ;Knowlton & Squire, 1996 ) or harm their performance, such as when participants devise misguided rules of thumb that they reportedly use to override their intuitions (e.g., Rebuschat et al, 2015 ). To illustrate, consider Hamrick's ( 2013 ) experiments designed to investigate incidental learning of novel syntactic structures.…”

“…One option might be to use predictive multilevel models, which would allow researchers to analyze test results on a trial-by-trial basis. For instance, with Hamrick's ( 2014a ) data, test item number could be used as a predictor to assess whether learning occurred during the test phase and, if so, whether the amount differed across groups. Analyzing further, a researcher could check for interactions between potentially relevant test-item features (e.g., verb fi nality, grammaticality) and group membership, to determine whether there were different bases for the experimental and control groups' performances (Baayen, Davidson, & Bates, 2008 ).…”

Practice and progression in Second Language Research methods

“…In the past, it has been common for researchers to fit a single type of computational model, say, a connectionist neural network, to human data using very specific parameters (e.g, Ellis & Schmidt 1997;Williams & Kuribara 2008). However, in Phillip Hamrick's recent research (Hamrick 2014), he has examined multiple competing computational models with different learning algorithms against adult L2 learning data. He tested these models across a range of parameters, essentially ensuring that the goodness of fit of any model is due to the intrinsic properties of that model, rather than a very limited set of possible parameters within the model.…”