Generalization from newly learned words reveals structural properties of the human reading system.

Abstract

Connectionist accounts of quasiregular domains, such as spelling-sound correspondences in English, represent exception words (e.g., pint) amidst regular words (e.g., mint) via a graded "warping" mechanism. Warping allows the model to extend the dominant pronunciation to nonwords (regularization) with minimal interference (spillover) from the exceptions. We tested for a behavioral marker of warping by investigating the degree to which participants generalized from newly learned made-up words, which ranged from sharing the dominant pronunciation (regulars), a subordinate pronunciation (ambiguous), or a previously non-existent (exception) pronunciation. The new words were learned over two days, and generalization was assessed 48 hours later using nonword neighbors of the new words in a tempo naming task. The frequency of regularization (a measure of generalization) was directly related to degree of warping required to learn the pronunciation of the new word. Simulations using the Plaut et al. (1996) model further support a warping interpretation. Our findings highlight the need to develop theories of representation that are integrally tied to how those representations are learned and generalized.Keywords: quasiregularity; connectionist models; word learning; tempo naming. Warping 3Generalization from newly learned words reveals structural properties of the human reading system Mastery of reading in alphabetic languages, such as English, requires an individual to learn correspondences between strings of letters and their pronunciations. Typically, this mapping is consistent, whereby a string of letters is pronounced in one way across many words (int in mint, hint, print). Although such regularities can simplify reading acquisition, there are also exceptions that are inconsistent with these regularities (e.g., pint) that must also be learned, and these clearly pose a challenge. Nonetheless, most people become proficient readers. Why does learning an exception word like pint not disrupt reading orthographically similar words like mint? This paper investigates how skilled readers represent both regular words and exceptions to the rule, and yet exhibit only minimal interference between these types of items.Considerable computational works has been devoted to understanding how quasiregularity is represented in memory in a way that enables accurate reading of both regulars and exceptions. For example, the DRC model (Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001) has maintained that two qualitatively different mechanisms are required to do so: a rulebased system to deal with regulars and new words, and a memory-based system to accommodate exceptions. Others have argued that it is possible for a single mechanism to represent both regularities and the various degrees of inconsistency characteristic of natural language (e.g., Plaut, Seidenberg, Patterson, & McClelland, 1996). A connectionist, parallel distributed processing (henceforth PDP) network that maps between spelling and sound via an intermediate pool ...

Statistical learning research: A critical review and possible new directions.

“…Aside from the very simplified, extremely limited nature of SL stimuli compared to the noisy input of real-world language (as discussed in our section on A Realistic View of the Learning Environment), several factors affect the relative slowness of language learning. One relates to how new experience is affected by and subsequently impacts on existing patterned regularities learned from prior exposure to language (for related discussion, see Armstrong, Dumay, Kim, & Pitt, 2017; Jones & Macken, 2015). However, another potential factor is the interaction between experience and the gradual maturation of different neural systems.…”

Morpho-phonological regularities influence the dynamics of real-time word recognition: Evidence from artificial language learning

“…These models do not start with a predefined set of internal representations and processing steps, but rather acquire the representations that appear to best capture the task at hand (e.g., Elman & Zipser, 1988). This is clearly seen in recent models of reading (Armstrong et al, 2017;Kim et al, 2013) where even a handful of items are sufficient to warp internal representations linking letters and sounds. During training, these models adopt internal representations that are in some ways a hybrid of rule-based mappings (e.g., O makes the /ɑ/ sound) and clusters of exceptions or item-based learning.…”

The contributions of decoding skill and lexical knowledge to the development of irregular word reading.

“…[Figure 1 about here] Connectionist models based on the triangle model do not incorporate self-teaching in learning at present. Arguably, therefore, they may not apply to how children start to learn orthographic forms at the onset of reading development (Plaut et al, 1996;Seidenberg & McClelland, 1989; see also Kim et al, 2013;Armstrong et al, 2017). Connectionist triangle models vary, but all are governed by the principles of nonlinearity, adaptivity, and distributed representations (Plaut et al, 1996;Seidenberg & McClelland, 1989).…”

“…enberg & McClelland, 1989; see alsoArmstrong et al, 2017;Kim et al, 2013). Connectionist triangle models vary, but all are governed by the principles of nonlinearity, adaptivity, and distributed This document is copyrighted by the American Psychological Association or one of its allied publishers.This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.…”