Early linguistic experiences are intimately tied to later language learning outcomes (e.g., Chilosi et al., Reference Chilosi, Comparini, Scusa, Orazini, Forli, Cipriani and Berrettini2013; Kaushanskaya & Marian, Reference Kaushanskaya and Marian2009b). The underlying neural and cognitive processes mediating this relationship remain unclear. Pierce, Genesee, Delcenserie, and Morgan (Reference Pierce, Genesee, Delcenserie and Morgan2017) propose that the phonological working memory (PWM) system is the critical component responsible for linking early linguistic experiences to later language development. Their argument arises from research demonstrating that exposure to linguistic input early in life shapes the kinds of phonological representations that are formed about the sounds within one's native language (Kuhl, Reference Kuhl2004; Majerus et al., Reference Majerus, Amand, Boniver, Demanez, Demanez and van der Linden2005; Mody, Schwartz, Gravel, & Ruben, Reference Mody, Schwartz, Gravel and Ruben1999; Nittrouer & Burton, Reference Nittrouer and Burton2005). These phonological representations, which the authors believe to be created and stored using the PWM system (Gathercole & Baddeley, Reference Gathercole and Baddeley1989), are highly predictive of later language outcomes (Bernhardt, Kemp, & Werker, Reference Bernhardt, Kemp and Werker2007; Kuhl, Reference Kuhl2010; Molfese & Molfese, Reference Molfese and Molfese1985; Tsao, Liu, & Kuhl, Reference Tsao, Liu and Kuhl2004). Explaining language development via memory processes has many benefits. For instance, drawing on a large extant literature on memory-related processes allows researchers to make novel predictions about language learning, especially in regard to its connections to other kinds of learning (e.g., Gathercole, Reference Gathercole2006). Identifying PWM as a key driver to language development is a useful insight likely to generate much further research, but it may not go far enough in tying language acquisition to more domain-general aspects of the human cognitive architecture. We believe that a complete account of language will require consideration of additional aspects of the human cognitive architecture working alongside the PWM system.

To illustrate this point, we will discuss a type of learning that is both domain general and widely thought to contribute to the relation between experience, representations, and language outcomes: statistical learning (Saffran, Aslin, & Newport, Reference Saffran, Aslin and Newport1996). Statistical learning refers to the sensitivity to statistical information in the environment and is thought to arise from the way that humans store and access information in memory (Thiessen, Kronstein, & Hufnagle, Reference Thiessen, Kronstein and Hufnagle2013). Because real-world languages can be characterized by a variety of statistical regularities, for example, syllables within a word are more likely to co-occur than syllables spanning a word boundary (Harris, Reference Harris1954; Swingley, Reference Swingley, Hahn and Stoness1999), statistical learning and real language learning are thought to share many of the same underlying mechanisms, a conclusion that has been supported by neuroimaging studies (Conway & Pisoni, Reference Conway and Pisoni2008; Friederici, Steinhauer, & Pfeifer, Reference Friederici, Steinhauer and Pfeifer2002; Petersson, Forkstam, & Ingvar, Reference Petersson, Forkstam and Ingvar2004). Statistical learning, like language itself, clearly involves contributions from working memory processes (e.g., Brady, Konkle, & Alvarez, Reference Brady, Konkle and Alvarez2009; Umemoto, Scolari, Vogel, & Awh, Reference Umemoto, Scolari, Vogel and Awh2010). In addition, though, we suggest that long-term memory and attention is necessary for a complete explanation of both statistical learning and language outcomes (Baker, Olson, & Behrmann, Reference Baker, Olson and Behrmann2004; Schapiro, Gregory, Landau, McCloskey, & Turk-Browne, Reference Schapiro, Gregory, Landau, McCloskey and Turk-Browne2014). We believe that these processes work in tandem with the PWM system to support statistical learning, and by extension language learning.

Because speakers produce utterances that are consistent with the phonology of their native language, the phonological utterances to which infants are exposed are characterized by a statistical structure (e.g., Goldinger, Reference Goldinger1996, Reference Goldinger1998; Hay, Graf Estes, Wang, & Saffran, Reference Hay, Graf Estes, Wang and Saffran2015; Houston & Jusczyk, Reference Houston and Jusczyk2003; Kuhl et al., Reference Kuhl, Stevens, Hayashi, Deguchi, Kiritani and Iverson2006). That is, phonemes that are near prototypical category members are produced more often than phonemes that are in an ambiguous region between two or more phonemic categories (e.g., Vallabha, McClelland, Pons, Werker, & Amano, Reference Vallabha, McClelland, Pons, Werker and Amano2007). This statistical structure means that infants can discover phonemic categories by learning from the distribution of speech sounds in the input (Maye, Werker, & Gerken, Reference Maye, Werker and Gerken2002). Models of this kind of statistical learning demonstrate how infants’ ability to track the frequency and variability of phonological exemplars can arise out of domain-general processes such as similarity-based comparison, decay, integration, and abstraction (e.g., Thiessen & Pavlik, Reference Thiessen and Pavlik2013). Because these processes are supported by the long-term memory system (e.g., Cabeza, Rao, Wagner, Mayer, & Schacter, Reference Cabeza, Rao, Wagner, Mayer and Schacter2001; Collins & Quillian, Reference Collins and Quillian1969; Meyer & Schvaneveldt, Reference Meyer and Schvaneveldt1971; Zeithamova, Maddox, & Schnyer, Reference Zeithamova, Maddox and Schnyer2008), accounts of statistical learning provide an alternate pathway through which refinement of phonological representations is accomplished that does not rely solely on PWM. The involvement of long-term memory processes is supported by research showing a bidirectional relationship between word forms, which are stored in long-term memory (e.g., Buchwald & Rapp, Reference Buchwald and Rapp2009), and phonological representations (Thiessen & Pavlik, Reference Thiessen and Pavlik2016). In addition, a cognitive framework invoking processes with intimate ties to the long-term memory system may be better suited to explain the extended developmental time course of phonological refinement compared to PWM alone (e.g., Marchetto & Bonatti, Reference Marchetto and Bonatti2015; Younger, Hollich, & Furrer, Reference Younger, Hollich and Furrer2004).

Accounts of statistical learning also highlight the importance of attention in supporting the link between prior experience and language outcomes. Given a sequence of elements, infants and adults can detect that some elements of the input are more likely to co-occur than others (e.g., Saffan et al., Reference Saffran, Aslin and Newport1996; Thiessen et al., Reference Thiessen, Kronstein and Hufnagle2013). Many models of statistical learning suggest that this process involves chunking (e.g., Perruchet & Vinter, Reference Perruchet and Vinter1998). That is, due to limited working memory capacity, a fluent stream of input is processed in a piecewise manner. Through processes of activation, decay, and interference, groups of elements that are frequently encountered (e.g., syllables within a word) are strengthened, whereas groups of elements infrequently encountered (e.g., syllables spanning a word boundary) are weakened. Critically, however, these chunks are not formed solely at random. Instead, learners are more likely to detect co-occurrence relations among elements of the input to which they have attended (e.g., Baker et al., Reference Baker, Olson and Behrmann2004; Toro, Sinnett, & Soto-Faraco, Reference Toro, Sinnett and Soto-Faraco2005). The relationship between attention and working memory is a complex one, and these constructs may be partially overlapping (e.g., Awh & Jonides, Reference Awh and Jonides2001; Engle & Kane, Reference Engle and Kane2003). At minimum, though, what we mean by the term attention involves a more active role on the part of the learner than a simple refinement of representations in working memory. We suggest that learners do not simply reflect the statistical structure of the input. Rather, as their knowledge grows, learners use implicitly and explicitly focus on aspects of the input that are likely to be relevant or informative (e.g., Griffiths & Mitchell, Reference Griffiths and Mitchell2008; Kruschke, Reference Kruschke2001). For example, most content words in English are marked with lexical stress on the first syllable (e.g., Cutler & Carter, Reference Cutler and Carter1987). After infants have learned enough English words to detect this regularity, they utilize lexical stress in subsequent word segmentation (e.g., Johnson & Jusczyk, Reference Johnson and Jusczyk2001; Jusczyk, Cutler, & Redanz, Reference Jusczyk, Cutler and Redanz1993; Thiessen & Saffran, Reference Thiessen and Saffran2003). This does not seem to be due to a refinement in the ability to represent lexical stress, as infants who have not yet learned the stress pattern of their native language represent differences in syllabic stress (e.g., Sansavini, Bertoncini, & Giovanelli, Reference Sansavini, Bertoncini and Giovanelli1997). Instead, it appears to be attributable to preferentially attending to stressed syllables as candidate word onsets (e.g., Johnson & Jusczyk, Reference Johnson and Jusczyk2001; Jusczyk et al., Reference Jusczyk, Cutler and Redanz1993; Thiessen & Saffran, Reference Thiessen and Saffran2003).

As we have discussed, there are both empirical and theoretical reasons to favor a broader connection between domain-general cognitive processes and language outcomes; one that involves, in addition to PWM, long-term memory and attentional processes. An additional advantage of such a perspective is that it can speak to the difficulties in developing accurate and fine-tuned phonological representations for languages learned later in life without invoking age-related qualitative change in cognitive mechanisms (Thiessen, Girard, & Erickson, Reference Thiessen, Girard and Erickson2016). This perspective can also explain why late second language learners who have formed appropriate phonological representations may still have difficulties with language learning (e.g., Bradlow, Pisoni, Akahane-Yamada, & Tohkura, Reference Bradlow, Pisoni, Akahane-Yamada and Tohkura1997; Ingvalson, McClelland, & Holt, Reference Ingvalson, McClelland and Holt2011; Pallier, Bosch, & Sebastián-Gallés, Reference Pallier, Bosch and Sebastián-Gallés1997). Statistical learning experiments have demonstrated that linguistic cues are weighted differentially as a function of perceptual characteristics (e.g., salience; cf. Frost, Armstrong, Siegelman, & Christiansen, Reference Frost, Armstrong, Siegelman and Christiansen2015), cognitive ability (Bartolotti, Marian, Schroeder, & Shook, Reference Bartolotti, Marian, Schroeder and Shook2011; Weiss, Gerfen, & Mitchel, Reference Weiss, Gerfen and Mitchel2010), and prior experience (Cutler & Norris, Reference Cutler and Norris1988; Kaushanskaya & Marian, Reference Kaushanskaya and Marian2009a; Thiessen & Saffran, Reference Thiessen and Saffran2003; Toro, Sebastian-Gallés, & Mattys, Reference Toro, Sebastian-Gallés and Mattys2009). As such, the ability to use a linguistic cue is much more nuanced than simply using the most informative cue, and depends on the interplay between prior experience, working memory, long-term memory, and attention. We suggest that a similar perspective will be useful for understanding language development more broadly, and are excited by the possibilities outlined by Pierce et al. (Reference Pierce, Genesee, Delcenserie and Morgan2017) in thinking about language development as deeply interconnected with domain-general cognitive processes.