The work of Susan Gathercole and others on nonword repetition has increased general interest in the relationship between memory and language, and has provided a fertile theoretical framework for researchers to explore how the language system makes use of the phonological loop, a component in Baddeley's (1986) working memory model. Gathercole (2006) integrated a number of findings from a variety of research methodologies and populations to support this theoretical framework. She also discussed how this framework might be applied to increase our understanding of language disorders characterized by word learning problems. Synthesizing the findings from many and diverse areas of study to formulate a coherent theory is a challenge, but this challenge must be undertaken if we hope to significantly advance our understanding of human behavior, including the differences in processing that are found across the life span and across individuals. I admire Gathercole for undertaking this challenge (and the additional challenge of applying the theory to language disorders) in the area of word learning.

I also wholeheartedly agree with Gathercole (2006) regarding the value of using a variety of research methodologies to identify phenomena that are worthy of further exploration, and the strength of experimental manipulation to determine the underlying causal relationships among relevant variables. Future research involving the manipulation of several key variables, such as phonotactic probability and neighborhood density, could significantly advance our understanding of word learning, as well as provide insight into the relationship between word learning and other linguistic processes, such as the production and recognition of spoken words.

Phonotactic probability refers to the frequency with which phonological segments or sequences of segments occur in the language (Vitevitch, Luce, Charles-Luce, & Kemmerer, 1997). Certain segments, such as word initial /s/, and sequences of segments, such as /sˆ/, are more common than other segments, such as word initial /j/, and sequences of segments, such as /ji/. See Vitevitch and Luce (2004) for information on how to obtain phonotactic probability values based on American English pronunciations. Neighborhood density refers to the number of words that sound like a given word; these similar sounding words are called neighbors. One way to estimate the number of neighbors a word has is to add, subtract, or substitute a single phoneme in that word (Landauer & Streeter, 1973; Luce & Pisoni, 1998). For example, adding, subtracting, or substituting a single phoneme in the word cat yields as neighbors the words scat, at, hat, kit, and cab (note, cat has other neighbors, but only a few were listed for illustration). Words that have many neighbors are said to have a dense neighborhood, whereas words that have few neighbors are said to have a sparse neighborhood.

A number of studies have examined the influence of phonotactic probability and neighborhood density on spoken word recognition in adults, and have yielded important insights about the use of lexical and sublexical representations in spoken word recognition (Vitevitch & Luce, 1998, 1999, 2005; see also Pitt & Samuel, 1995). Lexical representations correspond to whole word forms, whereas sublexical representations correspond to parts of words, such as phonological segments or sequences of segments. When lexical representations are used to process spoken stimuli, effects of neighborhood density are observed such that sparse items are responded to more quickly than dense items. When sublexical representations are used to process spoken stimuli, effects of phonotactic probability are observed, such that high probability items are responded to more quickly than low probability items.

In tasks that do not require lexical representations to be accessed for accurate performance, such as auditory shadowing (also called nonword repetition) and same–different matching, sublexical representations can be used to process spoken stimuli. For example, Vitevitch and Luce (1998, 1999, 2005) found that high probability/dense nonwords were responded to more quickly than low probability/sparse nonwords in auditory shadowing and same–different matching tasks, suggesting the use of sublexical representations in processing. Although lexical representations are typically used to process real words (see the results of the real words in Vitevitch & Luce, 1998, 1999, in those same tasks), sublexical representations can be used in certain circumstances to process real words in English, resulting in high probability/dense words being responded to more quickly than low probability/sparse words (Vitevitch, 2003).

In tasks that require lexical representations to be accessed for accurate performance, such as lexical decision, lexical representations are used to process spoken stimuli. As is typical for real words, but interestingly, for nonwords as well, in a lexical decision task, high probability/dense items were responded to more slowly than low probability/sparse items (Vitevitch & Luce, 1999; see also Luce & Pisoni, 1998). The manipulation of phonotactic probability/neighborhood density, lexicality, and task demand in these experiments show that there are two types of representation with different processes operating on them being used to recognize spoken items.

However, spoken word recognition is not the only linguistic process influenced by phonotactic probability and neighborhood density. Storkel (2001, 2003; Storkel & Rogers, 2000) found that these variables also influence word learning in children. After multiple presentations of a novel referent and a novel word form, children in these studies demonstrated in a number of tasks (and even after a delay of 1 week) that they had learned high probability/dense novel words more rapidly than low probability/sparse novel words, suggesting that phonotactic probability influenced the formation of semantic representations and the association between semantic and lexical representations, key components of word learning.

Using a similar task with nonwords varying in phonotactic probability, Storkel (2004) found that normal children were able to capitalize on the phonological similarity of some novel words to known words by learning high probability/dense novel words more rapidly than low probability/sparse novel words, whereas children with phonological delays tended to learn novel words that were dissimilar to other words, suggesting that children with phonological delays may have difficulty processing or resolving among phonologically similar words. Testing children that are normally developing or that have language delays in experiments that manipulate phonotactic probability/neighborhood density may help us better understand the process of word learning as well as the nature of various language disorders.

It is important to note, however, that there is a positive correlation between neighborhood density and phonotactic probability (Vitevitch, Luce, Pisoni, & Auer, 1999). Although most studies compared stimuli that were high in phonotactic probability and had dense neighborhoods to stimuli that were low in phonotactic probability and had sparse neighborhoods (e.g., Vitevitch & Luce, 1998), it is possible to manipulate one variable while controlling the other, or to orthogonally manipulate both variables (e.g., Luce & Large, 2001; Vitevitch, Armbruster, & Chu, 2004). Studies of spoken word recognition and production that have separated the influence of neighborhood density and phonotactic probability have helped us better understand the unique contribution of lexical and sublexical representations to various language processes. By manipulating both of these variables, future studies of word learning might better determine the unique role that each variable plays in learning novel word forms (e.g., Gathercole, Frankish, Pickering, & Peaker, 1999; Roodenrys & Hinton, 2002; Storkel, Armbruster, & Hogan, in press; Thorn & Frankish, 2005). Furthermore, such studies might help us better understand what contributes to poor word learning in children with language impairments, and may lead to improved tools for diagnostics and treatment.

The work of Gathercole and others has done much to advance our understanding of word learning, but there is still much about the process of word learning in normal and clinical populations that we do not understand. As Gathercole (2006) suggests, the experimental manipulation of certain variables can further advance our understanding of the word learning process. Future work that capitalizes on the manipulation of tasks and key variables has the potential to not only increase our understanding of the word learning process, but to greatly improve our models and theories of word learning and related language processes.