Computational modeling of biased phonological learning

36

2016-11-11

2019-11-10

150000

Computational modeling of biased phonological learning

Computational modeling, Language acquisition, Learning bias, Morphology, Phonology

Linguistics and LanguagesPsycholinguistics

201611159006

Seed Fund for Basic Research for New Staff

2016

Completed

Goal of this project is to create a computational software tool designed to simulate the trajectory of phonological learning based on child production corpora and experimental data. Background Research on phonological learning proposes that infants track the distributional properties of speech sounds in their language from as early as 7 month olds and use this information to perform various phonological tasks – discriminating sounds (Anderson, Morgan, & White, 2003; Maye, Werker, & Gerken, 2002), learning phonotactics (Chambers, Onishi, & Fisher, 2003), segmenting words from string of speech (Saffran, Aslin, & Newport, 1996), and detecting patterns of alternations (White, Peperkamp, & Morgan 2008; White and Sundra 2014). One notable pattern of infants’ phonological learning is that learners are equipped with learning biases, thus some patterns are more readily learnable than others. For example, cognitively simple patterns are easier to learn (Chambers, Onishi, & Fisher, 2011; Cristià & Seidl, 2008; Saffran & Thiessen, 2003), and phonologically less marked patterns are easier to learn (Jusczyk, Smolensky, & Allocco, 2002; See Seidl & Buckley, 2005 for the opposing evidence). Evidence for such biases come from perception experiments, and from a computational modeling approach (Peperkamp, Le Calvez, Nadal, & Dupoux, 2006; White, 2013; Wilson, 2006). While empirical evidence for the learning biases in early stages of phonological learning (i.e., when infants learn phoneme inventories or phonotactic patterns of their languages) has been extensively provided, little is known for the types of learning biases young learners bring in later stages of learning, namely to the task of learning morphophonological alternations (e.g., [khæt-s] ~ [dɔg-z]). Within constraint-based theoretical frameworks of phonology, some learning biases have been proposed based on conceptual reasoning (e.g., see Albright and Do, (in press) for references), but the role of such biases has rarely been empirically tested. Experimental approach is a good way to empirically test hypotheses on learning biases and the PI plans to submit an ECS proposal in October 2016 to test the role of learning biases with school-aged children in order to understand how learners encode morphologically-conditioned alternation patterns into their partially mastered phonological grammar. Additional evidence for learning biases can come from computational approach. By running computational learning simulations, we can embody the hypothesized learning biases in the software code, and let the software compute the exact consequences of the changes of learning. This goal of this project is to implement computational software to test hypotheses on learning biases which within theoretical work of phonology are argued to play a crucial role in learning phonological patterns. Several computational learning models are available in the field of phonology, but all of them are designed to simulate a specific phonological learning task such as learning phoneme inventories or phonotactic patterns (see Wilson 2006 for the overview). An exciting area in understanding learning biases is to examine how a computational learner, which has already mastered parts of phonological grammar, such as sound contrasts or phonotactics, will adjust its grammar to accommodate newly acquired knowledge such as morphophonological alternation patterns of a given language. When the learning trajectory across different phonological learning tasks is simulated, we can understand how the learner incorporates newly acquired linguistic knowledge into its existing grammar, and to see how learning biases facilitate or hinder each step involved in the process of phonological acquisition.