Lexical–Semantic Cues Induce Sound Pattern Stability in Children With Developmental Language Disorder

Abstract

Purpose

When learning novel word forms, preschoolers with developmental language disorder (DLD; also known as specific language impairment) produce speech targets inaccurately and with a high degree of intraword variability. The aim of the current study is to specify whether and how layering lexical–semantic information onto novel phonological strings would induce increased organization of sound production patterns.

Method

Twenty-one preschoolers with DLD and 21 peers with typical language (ranging in age from 4;1 to 5;11 [years;months]) imitated multiple renditions of novel words, half with (i.e., words) and half without (i.e., nonwords) a linked visual referent. Methods from network science were used to assess the stability and patterning of syllable sequences. Sound accuracy was also measured.

Results

Children with DLD were less accurate and more variable than their typical peers. However, once word forms were associated with a visual referent, network stability, but not accuracy, improved for children with DLD.

Conclusions

Children with DLD showed significant word form deficits as they acquired novel words and nonwords. The inclusion of a meaningful referent resulted in increased sound sequence stability, suggesting that lexical–semantic information provides a bootstrap for phonological organization in children with DLD.

Developmental language disorder (DLD; also known as specific language impairment) is classically identified based on deficits in morphosyntax (Leonard, 2014; Rice et al., 1995), yet children with this disorder often present with weaknesses across multiple aspects of word learning. In this work, we focus on sound pattern learning, especially whether and how the inclusion of a lexical–semantic cue may induce increased stability of a newly acquired word form. Network science approaches (Benham et al., 2018) are used to assess stabilization processes that incorporate lexical–semantic and phonological components of word learning.

In the semantic domain, children with DLD are relatively weak in the depth of their vocabulary knowledge (Gray, 2004, 2006), especially when they define words or draw pictures that incorporate semantic detail (Mainela-Arnold et al., 2010; McGregor et al., 2002, 2013). They also recognize fewer semantic features than their typically developing (TD) peers when fast-mapping features to objects and actions (Alt et al., 2004). Deficits in word recognition are also apparent, with children with DLD demonstrating slower reaction times in lexical decision tasks, suggesting weaker word representations (Edwards & Lahey, 1996). Furthermore, studies of word naming have revealed that children with DLD not only retrieve words less accurately (Gray, 2005) but also differentiate semantic associations among words poorly, which may be attributed to weaker organization of lexical–semantic detail in children with impaired language (Lahey & Edwards, 1999).

Children with DLD show difficulties in mapping new meanings to new forms or in fast mapping (e.g., Rice et al., 1990). However, it is the form, not the meaning, component that is thought by many to be most affected—preschoolers with DLD are similar to their TD peers in their comprehension of the meaning of a novel word (Dollaghan, 1987; Gray, 2004). Children with DLD also retrieve relevant semantic information about referents (Leonard et al., 2019) but are consistently less accurate than their peers when producing novel forms (Dollaghan, 1987; Dollaghan & Campbell, 1998; Heisler et al., 2010; Leonard et al., 2019; McGregor et al., 2013). Furthermore, when producing multiple imitations of the same novel word forms, preschoolers with DLD not only are less accurate but also demonstrate higher intraword variability in their productions, as compared to their peers with typical language development (Benham et al., 2018; Goffman et al., 2007; Heisler et al., 2010).

From a clinical perspective, intraword variability (referred to as inconsistency by Holm et al., 2007) in speech production may be considered an atypical characteristic of speech sound development (Dodd, 2014; Holm et al., 2007). However, other developmental theories suggest that variability may serve an important purpose as an exploratory learning mechanism (Thelen & Smith, 1994). As applied to speech production, intraword stability may have a protracted developmental trajectory in typical learners (Sosa, 2015). Regardless of perspective, variability of phonological form is an important consideration in speech production, one that may be especially implicated in children with DLD as they acquire new words.

When considering classic models of speech production, it is not surprising that form and meaning may be differentially affected. In these models, word production is hierarchically organized, with discrete levels of semantic selection, lexical storage, phonological encoding, and articulatory output (Fromkin, 1971; Garrett, 1988; Levelt et al., 1999). Semantic, lexical, phonological, and articulatory–phonetic representations are thought to be stored relatively independently of each other; specific deficits may occur within each domain. However, there is also evidence of significant interactivity across levels of representation (e.g., Baese-Berk & Goldrick, 2009; Goldrick et al., 2011).

Children, both typical and with DLD, evidence this interactivity across lexical–semantic and articulatory–phonetic levels of processing. Heisler et al. (2010) asked whether the inclusion of lexical information (i.e., a visual referent) would influence speech production processes, including the stability of articulatory movement, as children with DLD and their TD peers produced novel word forms. To assess articulatory variability of lip and jaw movement, they applied the spatiotemporal index (Smith et al., 1995) to repeated productions of novel words, some with and some without referential status assigned. They found that novel strings paired with visual referents were produced with greater articulatory stability than those without visual referents, indicating interactivity between lexical and articulatory levels. Importantly, children with DLD were less accurate and more segmentally variable than their typical peers. Accuracy (i.e., percent consonants correct [PCC]) and segmental variability (i.e., type–token ratio) were not affected by the inclusion of a semantic referent. Thus, articulatory and phonological levels, while interactive, can operate independently. An important consideration that we address in this work is whether another phonological outcome—the stability of sound organization—is similarly affected by the inclusion of a referent.

A central question of this work is whether standard metrics of sound accuracy are sensitive indices of learning at lexical and phonological levels. Phonetic accuracy is frequently used to assess learning at the segmental level. However, shifts in organization are not addressed using this approach—accuracy levels may remain stable while individual segments vary. For example, a child may produce the word “baby” as “daby” and “aby.” The initial [b] is in error in both cases, but the child's instability is not incorporated in this analysis. Benham et al. (2018) proposed that metrics from network science would be well equipped to characterize changes in the organization of sound and syllable sequences in nonword productions in children with DLD. Application of these metrics revealed that children with DLD not only produced novel forms less accurately but also showed more variable syllable sequences. A critical question is whether the organization of syllable sequences can be altered by linking novel word forms to lexical–semantic information.

In this work, we focus on stability—and its counterpart “variability”—as children acquire new word forms, either with or without meaning. This approach is grounded in developmental findings showing that an increase in variability indexes a transition in the organization of a behavior. A classic example is in the transition from bimanual to unimanual and back to bimanual reaching (Corbetta & Bojczyk, 2002; Thelen & Smith, 1994). Infants first reach bimanually and then develop stable unimanual reaching patterns. However, at the onset of walking, there is a return to the less mature and less stable bimanual reaching pattern. These fluctuations in the structure and stability of motor patterns are thought to reflect reorganization as task demands change.

As applied to speech sound development, some have suggested that changes in variability signal important transitions in the organization of a child's productive phonology, especially as children move from holistic, word-based phonological units of representation to more fine-grained, segmental units (Sosa & Stoel-Gammon, 2006). Others have emphasized the role of the word in structuring a child's phonology (Vihman et al., 2009; Vihman & Velleman, 1989). For example, in a seminal study of sound and word production in children over the first 2 years of life, Ferguson and Farwell (1975) described sound production variability (e.g., saying “baby” as both [bidi] and [bɑbi]) as a typical process in early language development. They observed that children were often more accurate in their productions early on; then, as children acquired new words and sound patterns, there was often a period of instability and regression to less accurate forms (i.e., a “U-shaped” pattern of development). Importantly, they proposed that the observed instabilities in young children's speech were related to learning sounds on a word-by-word basis. They hypothesized that this would explain why some phonemes are produced correctly only in specific contexts (e.g., a [b] produced correctly in “ball” but not in “bottle”). In their view, the word plays a central role in structuring a child's early productive phonology.

It is our contention that these broader principles of behavioral organization, notably variability, stability, and reorganization, will index learning in children with typical and atypical language development. In the following sections, we highlight the potential contributions of network science to capture the organizational properties of a child's productive phonology and why this approach to development may be especially powerful when applied to children with DLD. We emphasize that novel network science methods will provide new insights for theoretical accounts and models of phonological learning, especially organization and stabilization, in children with typical and atypical language development. In particular, we investigate how children organize their sound pattern networks when a lexical level of processing is invoked via the inclusion of a semantic referent.

Indexing Sound Organization Using Network Science Approaches

Network science is a robust tool that language researchers are beginning to use to quantify complexity across multiple domains of word processing (Vitevitch, 2008, 2014), for instance, in lexical–phonological processing (e.g., Benham et al., 2018; Chan & Vitevitch, 2010; Vitevitch & Luce, 2016) and lexical–semantic processing (e.g., Beckage et al., 2011; Peters & Borovsky, 2019). As applied to language systems, a network consists of primary units (i.e., nodes or vertices), which, at the lexical and phonological levels, may be individual words, syllables, segments, features, and so forth. By assessing how the nodes of a network interact with each other via the links (i.e., edges) between them, organizational and structural patterns emerge. This approach is promising for capturing differences in linguistic structure in clinical populations, such as in children with language delays and impairments. For instance, Beckage et al. (2011) used network science to reveal the semantic structure of typically developing toddlers as compared to late-talking toddlers. Using children's words from the MacArthur–Bates Communicative Development Inventories (Fenson et al., 2007), they constructed semantic association networks based on words toddlers produced. Networks of typically developing toddlers were characterized by high clustering and hubs, suggesting systematic organization as new lexical entries were acquired; late talkers exhibited fewer local hubs with reduced interconnectivity between clusters. These findings reveal distinct processes by which these two groups of toddlers acquire words, with late talkers more readily acquiring semantically isolated, rather than highly related, words.

Similar deficits in network organization have been shown in the semantic structure of 8- to 10-year-old children with DLD. For example, Brooks et al. (2017) found that the semantic networks of children with DLD included associations that, compared to those of typical peers, were tightly interconnected, with less differentiation among semantic clusters. The authors concluded that these poorly differentiated semantic clusters may lead to inefficiencies in lexical processing and difficulties in word finding.

Less work has been conducted in sound pattern organization. An important question in children's phonological development relates to the sequential organization of sound units, such as segments or syllables. Network science approaches provide a tool for assessing the organization of syllable sequences. As a conceptual example of this approach, in the word “baby,” there are two syllables, namely, /beɪ/ and /bi/. A child who encounters the word “baby” for the first time might produce the word in the following ways (syllables marked by the placement of “.”): [beɪ.bi] produced 6 times, [bɑ.di] produced twice, and [bi.di] produced twice. One could construct a network for this production set in which each syllable is a node in the network (i.e., a total of four nodes for [beɪ], [bi], [bɑ], and [di]). The two adjacent syllables in each production would be connected by a line, or an edge, and a thicker line would indicate a more frequent pattern of production, the edge weight. There would be 10 edges overall for the 10 productions, but only three distinct edge patterns representing the three different productions. Note that a thicker line (or greater edge weight) would connect the syllables [beɪ] and [bi], as this was the child's most frequent production pattern and happens to be the correct production. In a separate analysis, we could characterize this child's phonetic accuracy of consonants produced as 16 consonants correct out of 20, with an accuracy score of 80% for this production set.

To simulate this process in a child's first encounter with novel words, Benham et al. (2018) examined preschoolers with DLD and their typical age-matched peers and constructed networks of syllables produced in imitations of two-syllable nonwords. Children with DLD had significantly more syllable candidates (or different nodes) in their networks, more links between syllables, and fewer stable connection patterns. This resulted in less organized and more variable productions of sound sequences. The authors proposed that production deficits in children with DLD may relate to difficulty with the sequential arrangement of syllable sequences, consistent with a deficit in sequential pattern learning (Goffman & Gerken, 2020; Hsu & Bishop, 2014; Vuolo et al., 2017). These results motivated this work, in which we ask whether production networks may show increased stability (or variability) when meaning is linked to the phonological form. That is, we ask whether the inclusion of a lexical–semantic cue may facilitate the organization of phonological production networks.

This Study

The objective of this study is to evaluate whether and how the inclusion of a visual referent, that is, a lexical–semantic cue, facilitates the acquisition of a stable and accurate word form as children with DLD and their TD peers acquire novel words. We ask children with DLD and children with typical language development to produce multiple nonwords. In one condition, following multiple productions, a visual referent is introduced. In a control condition, while number of exposures is identical, there is no referent. The children then produce the novel words again. Thus, the children engage in equal degrees of exposure and practice, but in one condition, a referent is introduced, and in the other, it is not.

In our first question, we examine how children, both TD and with DLD, map form to meaning in the referential condition. Because we hypothesize that form deficits are a core component of DLD, we predict that these children will not show difficulties in comprehending those novel word–referent pairs they have learned. The second question is concerned with how phonological sequences are organized as word forms attain referential status. We use network science as a core analysis to determine whether and how syllable sequences are restructured with the inclusion of a referential cue. One possibility is that, as evidenced in the literature on early typical development, additional referential information may induce a greater degree of production variability, consistent with lexical reorganization (Ferguson & Farwell, 1975; Sosa & Stoel-Gammon, 2006). Conversely, the additional referential cue may solidify the lexical–phonological link, thereby decreasing variability; children with DLD may use referential information to strengthen the stability of their novel word form representations (Heisler et al., 2010).

Finally, in our third question, we determine whether there are qualitative differences in the phonological constituents of a word once a referent has been associated—specifically whether the inclusion of a referent substantively alters the forms children use. To detect reorganization or stabilization, we ask whether the inclusion of a referent promotes the selection of a new set of syllables and syllable sequences. We hypothesize that a referent may promote the selection of different sound elements, consistent with the lexical restructuring hypothesis (e.g., Metsala & Walley, 1998). This hypothesis predicts that, as the lexicon expands, there are shifts in phonological organization that move away from representations stored as holistic chunks to representations comprised of individual segments. In this way, the emerging lexicon promotes the selection of new speech sounds. This measure provides critical qualitative information about the nature of phonological stabilization. For instance, it may be that children stabilize on a preferred number of syllables and syllable connections; however, a qualitative approach (addressed by the third question) is necessary to determine if the substance of the syllables and syllable connections is the same between phases of the study.

Method

Participants

Forty-two children participated, ranging in age from 4;1 to 5;11 (years;months). Twenty-one participants met exclusionary and inclusionary criteria for DLD as outlined by Leonard (2014). These included 10 girls and 11 boys with a mean age of 5;0 (SD = 0.43 years). Twenty-one participants had typical language development and served as a control group, which included 12 girls and 9 boys with a mean age of 4;11 (SD = 0.52 years). Because of the phonological focus of this study, all children were monolingual English speakers. Data were collected at Purdue University and analyzed at The University of Texas at Dallas; approvals were obtained from the institutional review board of both universities. Informed consent was obtained prior to participating in the study.

To be included in the study, all children were required to pass a bilateral hearing screening with 20 dB HL pure tones presented at 500, 1000, 2000, and 4000 Hz. Children were reported to have no history of neurological or neurodevelopmental disorder and received a score of minimal to no symptoms of autism spectrum disorder on the Childhood Autism Rating Scale–Second Edition (Schopler et al., 2010). Children showed a standard score of 85 or better on an assessment of nonverbal reasoning abilities, that is, the Columbia Mental Maturity Scale–Third Edition (Burgemeister et al., 1972).

The Structured Photographic Expressive Language Test–Preschool: Second Edition (Dawson et al., 2005) or the Structured Photographic Expressive Language Test–Third Edition (Dawson et al., 2003) were used as the qualifying language measures using a standard score cutoff of 87 for the Structured Photographic Expressive Language Test–Preschool: Second Edition, which has shown to have high sensitivity (96%) and specificity (95%) for DLD in this age group (Greenslade et al., 2009). For the Structured Photographic Expressive Language Test–Third Edition, we used a cutoff of 85 (1 SD below the mean). It has been reported that, while sensitivity and specificity are not included in the manual, this cutoff provides a sensitivity value of 71.9% and a specificity value of 100% (Perona et al., 2005). We collected additional speech (Consonant Inventory on the Bankson–Bernthal Test of Phonology [Bankson & Bernthal, 1990]) and vocabulary (Peabody Picture Vocabulary Test–Fourth Edition [Dunn & Dunn, 2007], Expressive Vocabulary Test–Second Edition [Williams, 2007]) measures on which children with DLD have been reported to vary (e.g., Leonard, 2014). Diagnostic measures are summarized in Table 1.

Table 1.

Mean standard scores and standard deviations for behavioral assessments by group.

Assessment	DLD			TD
	M	SD	Range	M	SD	Range
CMMS-3	101.90	7.78	91–125	116.67	9.06	103–140
CARS 2	16.31	1.35	15–20	15.12	0.27	15–16
SPELT-P2/SPELT-3	75.38	10.01	61–87	111.29	11.07	90–130
BBTOP-CI	71.19	8.65	65–90	100.10	8.51	86–117
PPVT-4	101.33	12.59	80–120	116.86	10.93	97–141
EVT-2	98.72 a	12.13	75–119	115.24	11.56	93–137

Note. DLD = children with developmental language disorder; TD = children with typical language development; CMMS-3 = Columbia Mental Maturity Scale–Third Edition; CARS 2 = Childhood Autism Rating Scale–Second Edition; SPELT-P2 = Structured Photographic Expressive Language Test–Preschool: Second Edition; SPELT-3 = Structured Photographic Expressive Language Test–Third Edition; BBTOP-CI = Bankson-Bernthal Test of Phonology–Consonant Inventory; PPVT-4 = Peabody Picture Vocabulary Test–Fourth Edition; EVT-2 = Expressive Vocabulary Test–Second Edition.

^a Three children in the DLD group did not complete the EVT-2, so this value reflects the mean of 18 participants.

Study Design

The present analyses were conducted on a subset of data from a larger study on the role of richness of semantic content in word form learning. In this larger study, children learned six novel words over the course of 3 days (Gladfelter & Goffman, 2017). The sessions occurred at least 24 hr, but no more than a week, apart to ensure that participants slept and consolidation could occur. In this study, we focused on the second session when the semantic cue condition incorporated a visual referent to map form to meaning. Thus, we included four of the six novel words. During Day 1, children produced all four of these words with no semantic content included (i.e., these were nonwords). Productions from Day 1 are not included in the present analysis; however, it is important to note that Day 1 consisted of the same amount of practice with nonwords across conditions, so referential and nonword conditions were controlled for practice. On Day 2, the focus of this study, two of these four words were assigned a visual referent (i.e., these two now had meaning) and two were not. Because the core question was whether there were shifts in children's phonological organization when first linking a referent to a novel word form, the focus of the present analysis is on word pairs that, on Day 2 of the study, are associated with a referential cue and control words that remain nonwords.

Stimuli

Children repeated four from a possible six trochaic CVCCVC (“C” refers to a consonant; “V” refers to a vowel) words, paired in the learning conditions as follows: /f^ʃpəm/ and /p^vgəb/, /m^fpəm/ and /b^pkəv/, and /p^btəm/ and /f^spəb/. These were counterbalanced in the larger study; hence, word pairs were evenly distributed across conditions, and assignment of word pairs to condition was randomized across all participants. In this study, children produced four of the six words, counterbalanced across conditions. Stimuli were constructed such that syllables came from low-density neighborhoods (Storkel & Hoover, 2010), with the neighborhood density of each syllable ranging from 0 to 15 neighbors (M = 6.3, SD = 4.3). Medial consonant clusters were of low phonotactic probability (Vitevitch & Luce, 2004); the positional biphone frequency of the medial cluster ranged from 0 to 0.0081 (M = 0.0014, SD = 0.0032). These aspects of the words were controlled because low neighborhood density and low phonotactic probability have shown an advantage for learning (Gladfelter & Goffman, 2013; Heisler & Goffman, 2016; Storkel, 2001). Words were additionally constrained by an initial, a medial, and a final labial consonant to facilitate articulatory kinematic analysis (not reported here), which requires the detection of lip opening and closing motion. Stimuli were recorded in a sound booth by a female talker in a child-directed voice as well as digitized and equalized at 70 dB using Praat (Boersma & Weenink, 2012).

Procedure

Participants were seated approximately 8 ft in front of a 76.2-cm Dell monitor, which was used to present PowerPoint slides containing visual images and audio playback of the stimuli. Children looked at referential or nonreferential images (see below for details) in a PowerPoint display and heard and imitated novel words. In each condition (nonreferential or referential), the two words were quasi-randomly ordered, with no more than two productions in a row of the same word. Productions were recorded with a high-quality audio signal using a Marantz CD recorder and a Shure BETA 87 microphone. Video was recorded with a Panasonic DVD camcorder. Children engaged in three phases of the study for each of the nonreferential and referential conditions, summarized in Table 2 and described below.

Table 2.

Schematic of experimental conditions.

Condition	Day 1	Day 2
Control	No referent	No referent
“Nonreferential”
	Pretest: Abstract images Exposure: Abstract images Posttest: Abstract images	Pretest: Abstract images Exposure: Abstract images Posttest: Abstract images
Lexical manipulation	No referent	Lexical referent
“Referential”
	Pretest: Abstract images Exposure: Abstract images Posttest: Abstract images	Pretest: Abstract images Exposure: Referential images Posttest: Abstract images

Note. The present analysis only includes data from Day 2.

In Phase 1, the pretest phase, children were presented with a word pair randomly associated with varying abstract and colorful images. Children repeated the words in direct imitation 12 times each. Participants were assigned a word pair for the nonreferential condition and a different word pair for the referential condition.

Following the pretest phase, children engaged in Phase 2, the exposure phase. This phase was perceptual only—children did not produce any of the word forms. It was only in this phase that the nonreferential and referential conditions differed. In the exposure phase for the nonreferential condition, children watched and listened 7 times as each word was randomly paired with varying abstract and colorful images, as in the pretest phase. In the referential condition, these word forms were paired with a referent—a drawing of a novel object—also 7 times each. Two additional novel objects were presented on the same screen as the target (see Figure 1); these served as foils in the comprehension probes that followed. An arrow was used to indicate the object referent to which the novel word corresponded. Because the novel word being named was paired with the same referent on each exposure, mutual exclusivity also supported learning. Phase 3, the posttest phase, was the same across both conditions, in which children repeated their assigned novel words 12 times each, again paired with the varying abstract images as in the pretest phase.

Figure 1.

Two object referents (left) assigned to novel phonological strings in the exposure phase as well as two foils (right).

For the novel words that were assigned referents, comprehension probes followed the test phase. Children were presented with a recording of a target word and an array of four objects—the two target objects and the two foils. They were instructed to listen to the word and point to the picture of its corresponding referent. For the next target word, children were presented with a new array of the same four pictures, but in a different spatial arrangement. Children selected the picture that matched the target word. Responses were recorded by the experimenter, and no feedback on accuracy was provided.

Analyses

Comprehension Probes

The first question related to whether preschoolers with typical and atypical language mapped a word form to a lexical referent. To verify that children were creating this association, the child's selection of a referent from the field of four pictures was coded on a 3-point scale (0, 1, and 2). A score of 0 corresponded to a no response or the selection of a foil, a score of 1 corresponded to the selection of the lexical competitor (i.e., the other lexical target), and a score of 2 corresponded to the selection of the correct lexical target. In this way, the responses for the lexical targets were weighted more heavily than the selection of a foil or no response. Responses for the two probes were averaged for a single value on the 3-point scale for statistical comparison.

Phonetic Transcription

The primary analyses were derived from broad phonetic transcription of the children's productions. Graduate speech-language pathology research assistants, with training in phonetics and child speech, transcribed each word the children produced. Productions that were disfluent or contained yawning, whispering, laughter, sighing, or long pauses between syllables were excluded from analysis. Interrater reliability was calculated between two transcribers using a sample of 20% of all productions; these were evenly distributed between the two experimental groups. An item-by-item comparison between the two coders showed 93% reliability. This was deemed a sufficiently high level of reliability, and transcriptions from the primary transcriber were used for the analyses. For the network analysis, transcriptions were converted to Klattese (Klatt, 1987), a computer-readable version of the International Phonetic Alphabet.

Phonological Outcomes: Segmental Accuracy

The next question pertained to how accurately children imitated word forms, in both the referential and control nonreferential conditions. Accuracy was computed for each production based on the number of consonants correct out of the total number of target consonants and multiplied by 100 to yield a percentage or PCC (Shriberg & Kwiatkowski, 1982). The number of correct consonants was summed for both words for each phase (pretest and posttest) such that, for statistical comparison, each child has a single value for PCC in the pretest phase and a single value for PCC in the posttest phase. For this analysis, PCC was calculated for the first 10 productions of each word that met the criteria for an acceptable production, as described above. Therefore, each cell represents the composite value of the number of correct consonants out of 80 target consonants (four consonants per word, 10 productions each of two words). Omissions and substitutions of target consonants were both scored as zero; additions were not considered errors. If a child switched the order of two target phonemes in the output (e.g., the child said /m^pfəm/ for the target /m^fpəm/), we counted both the “f” and the “p” as in error; this production would be scored as having two, out of four, consonants correct.

Phonological Outcomes: Network Analysis

The primary objective of this work was to examine whether and how children organize and stabilize their phonological systems as they produce novel word forms in referential or nonreferential contexts. We used approaches from network science to characterize the variable paths children take in arranging phonological elements—in this analysis, the syllable. To assess the relationship between syllable forms in the pre- and posttest conditions, we followed the network analysis procedures outlined in Benham et al. (2018). To summarize, each production was divided into two CVC syllables. Each syllable was considered a “node” in the network. A link or an “edge” connected the two nodes if they were produced in succession. For example, the two nodes “m^f” and “pəm” would be connected by an edge if the child produced the word “m^fpəm.” Therefore, edges can be considered a word production consisting of two linked syllable nodes, and the number of distinct edges refers to the number of distinct syllable connections. If the child omitted a medial consonant as in a CVCVC form, we considered the CV to be the first syllable and the CVC to be the second syllable (i.e., Maximal Onset Principle). The frequency of each production was also considered such that a stronger weight between two nodes represented a more frequent production pattern, depicted graphically as a thicker line between two nodes. Note that this analysis did not compare the child's productions to a target; instead, the focus was on the variability of a child's production set and whether the child stabilized on a preferred production pattern regardless of accuracy.

Co-occurrence lists of syllables for the pre- and posttest conditions were generated and loaded into NodeXL (Smith et al., 2010), an open-source network software program. Since we were interested in the number of syllable variants and the different connection patterns between them, we used this program to calculate the number of nodes and edges, respectively. We were also interested in the average edge weight for each production set (i.e., pretest and posttest). Recall that this experiment had a controlled number of productions, and the number of total edges for each child would be the same (i.e., 20 productions for pre- and posttest conditions). For this reason, we instead used the number of distinct edges in our analyses. To calculate the average edge weight, the total number of productions, 20, was divided by the number of distinct edges. A higher edge weight indicates a more frequent and stable production pattern. Note that, for this analysis, one network was generated for the pretest phase and another for the posttest phase for each child, which included both target words. Therefore, each child had one value for each measure (number of nodes, number of edges, and edge weight) for each of the two phases (pretest and posttest).

Proportion of Production Similarity in Referential Versus Nonreferential Conditions

The analyses described thus far are quantitative in nature and emphasize the number of nodes and edges produced in referential and nonreferential conditions. It is also important to qualitatively assess which nodes and edges children produce as a word form is assigned referential status. For example, it is possible that, in the posttest phase, children will converge on a subset of specific nodes or edges produced in the pretest phase. Alternatively, it is also possible that the introduction of a referential association will promote the selection of an entirely new set of word forms. Therefore, in this analysis, we assess the qualitative features of the novel forms over time and in the transition from nonreferential to referential learning via determining the proportion of production similarity from the pretest to the posttest phase. In this metric, we calculated the proportion of distinct syllables (i.e., nodes) and word forms (i.e., edges) that the child produced in both the pre- and posttest conditions out of the total number of distinct nodes or edges in the pre- and posttest conditions. We use the term “distinct” to refer to a nonweighted value, meaning if a child produced the word /m^fpəm/ 10 times, this word would only be counted as one distinct form. This provides an estimate of how well children retained word forms and their constituent syllables across the pre- and posttest phases in referential and nonreferential conditions. A higher proportion suggests that more syllables (nodes) and syllable connections (edges) were retained from the pretest condition in the posttest condition. A lower proportion suggests that children produced new forms in the posttest phase that were not produced in the pretest phase.

Statistical Analyses

In Question 1, we asked whether children with DLD perform more poorly than their typical peers in the comprehension probes. A one-way analysis of variance (ANOVA) with group (DLD and TD) as the between-subjects factor was used to detect differences in performance on the comprehension probes in the referential condition. For Question 2, we assess phonological performance as word forms transition from nonreferential to referential or, in our control condition, remain nonreferential. We used a mixed-model ANOVA with group (DLD and TD) as the between-subjects factor and with condition (nonreferential and referential) and phase (pretest and posttest) as the within-subject factors (Question 2a). Because we are especially interested in learning, we conducted an additional analysis focused on difference scores from pre- to posttest phases (Question 2b). We used a mixed-model ANOVA with group (DLD and TD) as the between-subjects factor and with the difference scores of the dependent variables (posttest minus pretest score) as the within-subject factor. Significant interactions were followed up by analyzing the referential and nonreferential conditions separately. For the third and final question, we assess the proportion of production similarity for nodes and edges in a mixed-model ANOVA using group as the between-subjects factor and condition as the within-subject factor. A Type I error rate of below .05 was considered significant for statistical tests. The data that support the findings of this study are openly available in Open Science Framework (https://osf.io/bpt2z/).

Results

Question 1: Comprehension Probes in Referential Condition

Performance on comprehension probes in the referential condition was assessed following the posttest phase. A one-way ANOVA revealed no effect for group, F(1, 40) = 0.44, p = .51, MSE = 0.21, η_p ² = .01; DLD and TD groups similarly link form and referent.

Question 2a: Performance on Phonological Outcomes in Nonreferential and Referential Conditions

The second question is concerned with phonological outcomes, including phonetic accuracy (i.e., PCC) and network science measures (number of nodes, number of edges, and edge weight), as children learn words with and without referential status. Means and standard deviations for all dependent variables are reported in Table 3. Main effects are always reported, along with the Group × Condition interaction. Additional interactions not directly related to the core hypotheses are reported in Table 4.

Table 3.

Means and standard deviations (in parentheses) of all variables on Day 2 of the study.

Dependent variable	Nonreferential				Referential
	Pretest		Posttest		Pretest		Posttest
	DLD	TD	DLD	TD	DLD	TD	DLD	TD
PCC	0.69 (0.20)	0.88 (0.10)	0.68 (0.21)	0.90 (0.08)	0.63 (0.18)	0.81 (0.14)	0.65 (0.20)	0.82 (0.16)
Nodes	8.14 (2.69)	5.86 (1.71)	7.38 (2.01)	5.24 (1.26)	9.29 (3.18)	6.57 (2.79)	6.81 (1.75)	5.71 (1.59)
Edges	6.33 (2.74)	3.81 (1.83)	5.52 (2.18)	3.19 (1.17)	7.52 (2.75)	4.86 (3.04)	5.10 (1.84)	4.05 (2.01)
Edge weight	4.01 (2.37)	6.29 (2.53)	4.08 (1.35)	7.11 (2.49)	3.08 (1.37)	5.59 (2.88)	4.53 (1.91)	5.96 (2.57)

Note. Ranges for the proportion of production similarity are given in the text and shown in Figure 5. DLD = children with developmental language disorder; TD = children with typical language development; PCC = percent consonants correct.

Table 4.

Additional statistical comparisons for analyses in Question 2a.

Measure	F(1, 40)	p	MSE	ηp 2
PCC
Phase × Group	0.70	.41	0.001	.02
Condition × Phase	0.71	.41	0.001	.02
Condition × Phase × Group	3.73	.06	0.01	.09
Nodes
Phase × Group	3.61	.07	8.15	.08
Condition × Phase	5.48	.02	10.01	.12
Condition × Phase × Group	3.13	.08	5.72	.07
Edges
Phase × Group	3.96	.05	8.60	.09
Condition × Phase	5.77	.02	8.60	.13
Condition × Phase × Group	3.60	.07	5.36	.08
Edge weight
Phase × Group	0.22	.64	0.29	.01
Condition × Phase	1.43	.24	2.25	.03
Condition × Phase × Group	5.56	.02	8.76	.12

Note. MSE = mean square error; PCC = percent consonants correct.

PCC

Children with DLD demonstrated significantly lower phonetic accuracy than their typical peers, F(1, 40) = 17.64, p < .001, MSE = 1.54, η_p ² = .31. There was also a significant condition effect, F(1, 40) = 8.69, p = .01, MSE = 0.15, η_p ² = .18, with a lower PCC in the referential condition. There was no significant Condition × Group interaction, F(1, 40) = 0.51, p = .48, MSE = 0.01, η_p ² = .01, and no significant effect of phase (pretest vs. posttest), F(1, 40) = 1.29, p = .26, MSE = 0.003, η_p ² = .03. Overall, children with DLD were weaker than children with typical language development in phonetic accuracy.

Nodes

There was a significant group effect, F(1, 40) = 15.84, p < .001, MSE = 178.15, η_p ² = .28, with children with DLD producing more nodes (i.e., syllable candidates) than their typical peers. There was no significant effect of condition, F(1, 40) = 1.90, p = .18, MSE = 8.15, η_p ² = .05, or Condition × Group interaction, F(1, 40) = 0.23, p = .63, MSE = 1.01, η_p ² = .01. There was a significant effect of phase, F(1, 40) = 25.85, p < .001, MSE = 58.34, η_p ² = .39, with fewer nodes in the posttest phase. Main effects reveal that children with DLD produced more nodes than their TD peers. The presence of fewer nodes in the posttest phase is consistent with phonological stabilization.

Edges

There was a significant group effect, F(1, 40) = 14.89, p < .001, MSE = 192.86, η_p ² = .27, with children with DLD producing more edges (i.e., syllable connections) than their typical peers. There was a significant condition effect, F(1, 40) = 4.66, p = .04, MSE = 18.67, η_p ² = .10, with more edges in the referential condition. There was no significant Condition × Group interaction, F(1, 40) = 0.86, p = .36, MSE = 3.43, η_p ² = .02. There was a significant effect of phase, F(1, 40) = 26.36, p < .001, MSE = 57.17, η_p ² = .40, with fewer edges in the posttest phase. Children with DLD produced more edges than their TD peers. Fewer edges in the posttest phase suggest phonological stabilization.

Edge Weight

As with edges, edge weight also showed a significant effect of group, F(1, 40) = 15.76, p < .001, MSE = 224.69, η_p ² = .28, with children with DLD demonstrating a lower edge weight than their TD peers. That is, children with DLD produced a greater number of different productions for a specific word target. There was also a significant condition effect, F(1, 40) = 4.74, p = .04, MSE = 14.40, η_p ² = .11, with a higher edge weight in the nonreferential condition. This was true for both groups, since there was no significant Condition × Group interaction, F(1, 40) = 1.61, p = .21, MSE = 4.88, η_p ² = .04. Children in both groups stabilized their productions over the course of learning, with a significant effect of phase, F(1, 40) = 14.43, p < .001, MSE = 19.40, η_p ² = .27, with a higher edge weight in the posttest phase.

To summarize, children with DLD were consistently weaker than their TD peers in their phonetic accuracy and in their phonological network organization. In their networks, they produced more nodes (or syllable candidates), more edges (or syllable connections), and a lower edge weight (or less frequent syllable connections) than their typical peers. There were significant condition effects for outcome variables: PCC decreased in the referential condition, and network analyses revealed a significant increase in the number of edges and a decrease in edge weight in the referential condition, suggesting shifts in phonological learning. Of course, a critical test of our hypotheses regarding the influence of a referent on phonological learning is to directly assess change within each individual child. We use difference scores across pre- and posttest phases to accomplish this goal in Question 2b.

Question 2b: Evidence of Phonological Learning in Nonreferential and Referential Conditions

PCC

There was no significant group effect, F(1, 40) = 0.70, p = .41, MSE = 0.003, η_p ² = .02; no significant condition effect, F(1, 40) = 0.71, p = .41, MSE = 0.003, η_p ² = .02; and a trending Group × Condition interaction, F(1, 40) = 3.73, p = .06, MSE = 0.01, η_p ² = .09, as illustrated in Figure 2. Overall, there was little evidence of learning for either group.

Figure 2.

Difference scores for percent consonants correct (PCC) in the nonreferential and referential conditions. Error bars denote standard error. DLD = children with developmental language disorder; TD = children with typical language development.

Nodes

There was no effect of group, F(1, 40) = 3.61, p = .07, MSE = 16.30, η_p ² = .08, although there was a trend, with children with DLD showing marginally larger difference scores for nodes compared with their TD peers. There was a significant condition effect, F(1, 40) = 5.48, p = .02, MSE = 20.01, η_p ² = .12, with more change in the referential than the nonreferential condition, and a trending interaction, F(1, 40) = 3.13, p = .08, MSE = 11.44, η_p ² = .07. On the basis of visual inspection of Figure 3 as well as the trending statistical effects, we followed up the learning effects by examining each condition separately. There was no significant group effect of difference scores for the nonreferential condition, F(1, 40) = 0.07, p = .80, MSE = 0.21, η_p ² = .002. However, in the referential condition, children with DLD demonstrated a greater reduction in the number of nodes than their TD peers, F(1, 40) = 5.62, p = .02, MSE = 27.52, η_p ² = .12. In the referential condition, 86% (18/21) of children with DLD and 52% (11/21) of children with typical language development showed fewer nodes in the posttest condition than in the pretest condition.

Figure 3.

Difference score for nodes (top), edges (center), and edge weight (bottom) in nonreferential (control) and referential conditions. Error bars denote standard error. DLD = children with developmental language disorder; TD = children with typical language development.

Edges

There was a trending effect of group, F(1, 40) = 3.96, p = .05, MSE = 17.19, η_p ² = .09, where children with DLD had marginally larger difference scores for edges overall. The referential condition showed a greater degree of change than the nonreferential condition, F(1, 40) = 5.77, p = .02, MSE = 17.19, η_p ² = .13. Children with DLD trended to demonstrating more change in the referential condition than their TD peers, F(1, 40) = 3.60, p = .07, MSE = 10.71, η_p ² = .08.

In our follow-up analyses of each condition separately, there was no significant group effect of difference scores for the nonreferential condition, F(1, 40) = 0.11, p = .74, MSE = 0.38, η_p ² = .003. In the referential condition, children with DLD demonstrated a greater reduction in the number of edges than their TD peers, F(1, 40) = 7.13, p = .01, MSE = 27.52, η_p ² = .15. This was supported by the observation that, in the referential condition, 90% (19/21) of children with DLD and 48% (10/21) of children with typical language development had fewer edges in the posttest condition than in the pretest condition.

Edge Weight

There was no significant group effect, F(1, 40) = 0.22, p = .64, MSE = 0.58, η_p ² = .01, or condition effect, F(1, 40) = 1.43, p = .24, MSE = 4.50, η_p ² = .03, but there was a significant Group × Condition interaction, F(1, 40) = 5.56, p = .02, MSE = 17.51, η_p ² = .12. Children with DLD, but not those with typical language development, showed increased edge weight (indicating convergence to more stable word forms) from the pretest to the posttest phase in the referential condition.

In our follow-up analyses, there was no significant group effect of difference scores for the nonreferential condition, F(1, 40) = 1.46, p = .24, MSE = 5.85, η_p ² = .04. However, there was evidence that children with DLD stabilized on preferred forms in the referential condition, F(1, 40) = 6.74, p = .01, MSE = 12.24, η_p ² = .14. In the referential condition, 90% (19/21) of children with DLD and 48% (10/21) of children with typical language development had a higher edge weight in the posttest condition than in the pretest condition. Graphic depictions of the networks for the pre- and posttest phases of the referential condition are presented in Figure 4.

Figure 4.

Network plots for the referential condition, collapsed by group for pre- and posttest conditions. These are for visualization only and represent 420 productions in each plot. DLD = children with developmental language disorder; TD = children with typical language development.

To summarize, children with DLD converged on more stable sound pattern networks when new word forms were assigned a referent. There was little evidence that children became more accurate over the course of learning. However, when assessing phonological network organization, children with DLD produced fewer nodes and edges and increased edge weight, but only once words were associated with a visual referent. The inclusion of a visual referent promoted the stability of these phonological components, but only for children with DLD and not their TD peers. We next address how the inclusion of a referent influences the quality of the nodes and edges selected from the pretest to the posttest phase.

Question 3: Proportion of Production Similarity Between Pretest and Posttest in Referential and Nonreferential Conditions

In Question 3, we use a proportion metric to determine whether there are qualitative differences in the phonological forms between pre- and posttest conditions once a lexical referent has been associated, as compared to the control condition in which there is never a referent assigned. We calculated the number of shared syllables (i.e., nodes) and syllable connections (i.e., edges) between the pre- and posttest phases, divided by the total number of distinct nodes or edges in the pre- and posttest phases.

Nodes

Means and standard errors are shown in Figure 5. Children with DLD did not differ from their TD peers in the proportion of shared nodes in the pretest compared with the posttest conditions, F(1, 40) = 3.88, p = .06, MSE = 0.08, η_p ² = .09. There was also no significant effect of referential versus nonreferential condition, F(1, 40) = 1.04, p = .32, MSE = 0.02, η_p ² = .03, and no Group × Condition interaction, F(1, 40) = 0.13, p = .72, MSE = 0.002, η_p ² = .003. As shown in Figure 5, there was a high degree of overlap of nodes produced from the pretest to the posttest phase, ranging from 82% to 91%.

Figure 5.

Proportion of production similarity between pre- and posttest phases for nodes (top) and edges (bottom). Error bars denote standard error. DLD = children with developmental language disorder; TD = children with typical language development.

Edges

There was a significant group effect, F(1, 40) = 6.63, p = .01, MSE = 0.27, η_p ² = .14. Children with DLD retained fewer of the same edges between pre- and posttest phases than those with typical language development. There was no significant effect of referential versus nonreferential condition, F(1, 40) = 0.03, p = .87, MSE = 0.001, η_p ² = .001, and no significant Group × Condition interaction, F(1, 40) = 1.30, p = .26, MSE = 0.05, η_p ² = .03. The means and standard error are shown in Figure 5. Notably, children with DLD showed an overlap of 68% (SD = 20%) in the nonreferential condition and 73% (SD = 19%) in the referential condition. Children with typical language development showed increased overlap across pre- and posttest phases, with 84% (SD = 22%) in the nonreferential condition and 79% (SD = 19%) in the referential condition.

To summarize, the inclusion of a visual referent, while increasing the stability of phonological networks in children with DLD, had no influence on the consistency of the constituents of the networks (i.e., the nodes) or the interconnectivity of these nodes (i.e., the edges). Once children, typical or with DLD, selected a set of nodes, they retained those through the posttest, even after the perceptual learning phase. Children with DLD did show less overlap of edges than their TD peers. This reduced overlap was observed in both referential and nonreferential conditions. In summary, while the stability of networks was conditioned by referential status in children with DLD, the actual substance of the nodes and edges was not.

Discussion

Grammatical weaknesses have long been a central focus of research in children with DLD. However, phonological aspects of novel word form learning are also clearly implicated (e.g., Coady & Evans, 2008; Dollaghan & Campbell, 1998; Graf Estes et al., 2007), which we contend form an important component of the deficit profile in DLD. The present findings support an impaired phonological system in children with DLD and contribute to theoretical accounts of phonological learning, suggesting interconnectivity between lexical–semantic and phonological levels of word learning (Heisler & Goffman, 2016; Heisler et al., 2010).

In this study, consistent with previous work, we found that children with DLD showed increased speech sound errors as compared to their typical peers. In addition, as shown via network science approaches, these children evidenced more disorganized phonological systems, with a greater inventory of syllables (i.e., nodes) and increased syllable connections (i.e., edges) for a specific word target. Reduced edge weight showed that children with DLD used less stable connection patterns than their TD peers. Similar to Benham et al. (2018), the organization of syllable sequences is clearly implicated in DLD. One possibility is that weaknesses in phonological organization relate to other sequence learning and statistical learning deficits. Children with DLD are weak at learning the patterns governing sequenced elements, even in the manual domain (Hsu & Bishop, 2014; Lum et al., 2014; Tomblin et al., 2007; Ullman & Pierpont, 2005). These domain-general sequential learning difficulties may also affect the organization of syllable co-occurrence patterns (Benham et al., 2018; Evans et al., 2009). Predictions from the procedural deficit hypothesis (Ullman et al., 2020; Ullman & Pierpont, 2005) suggest that phonological learning, which relies heavily on the sequential ordering of speech sounds, will be affected in children with DLD.

The major goal of the current study was to determine whether, during the earliest phases of word learning, we could modify phonological organization in children with DLD by strengthening the link between lexical–semantic and phonological levels of representation. Our results indicated that, only for children with DLD, once words were associated with a referent, phonological forms demonstrated increased organization and stability. Specifically, increasing the amount of lexical–semantic detail scaffolded phonological organization—but not accuracy—for children with DLD. The qualitative selection of nodes and edges also did not differ following a referent.

Classic models of speech production are hierarchically organized, with semantic, lexical, phonological, and articulatory levels processed separately (Fromkin, 1971; Garrett, 1988; Levelt et al., 1999). In our study, the opportunity to practice and hear the novel form without associating it with a meaningful referent (the nonreferential condition) did not improve phonological outcomes. Rather, linking phonological forms to meaning activated increased sound organization for children with DLD. Enhancement of meaning has been reported to facilitate both semantic and phonological outcomes for learners with DLD, suggesting some degree of interactivity across these processing levels. For example, Gray (2005) found that, in children with DLD, semantic cues strengthened the form–referent link, whereas phonological cues maintained the form representations to promote retrieval. These results suggest that lexical–semantic and phonological levels are not entirely independent and that manipulation of input at one level may positively influence learning at a different level. In both typical and atypical learning, at least during particular developmental periods, strengths in one domain may bootstrap weaknesses in others (e.g., Gleitman et al., 2005).

In our findings, associating a meaningful object referent strengthened phonological organization. This result is also consistent with Ullman and Pullman's (2015) declarative memory compensation hypothesis, in which it is posited that children with DLD rely on the declarative system to compensate for procedural deficits. Applied to the present work, one possibility is that weaknesses in the sequential arrangement of sound elements in children with DLD are improved by engaging the declarative system to link form to meaning. The procedural deficit hypothesis (Ullman et al., 2020; Ullman & Pullman, 2015) also speaks to the use of declarative memory functions as a scaffold or an intervention approach. Indeed, the present data provide empirical support that the addition of a referent helps children stabilize their productions. Ullman and Pullman propose that the declarative memory system will compensate for functions subserved by the procedural system.

The present findings reveal that some aspects of sound pattern learning are affected by linking a referent to a novel word form. Phonological organization and stability increased for children with DLD following the referential exposure, yet phonetic accuracy did not. Many previous studies have shown that children with DLD, especially in the preschool years, often show a co-occurring deficit in speech sound production (Alt et al., 2004; Benham et al., 2018; Deevy et al., 2010). This was true in our sample as well, with 90% (19/21) of children with DLD performing below expected levels on a standardized single-word articulation assessment. However, it was important that speech sound errors, as measured using PCC, did not show learning effects, either with practice (in the nonreferential condition) or as a function of inclusion of meaning (in the referential condition). Conversely, phonological organization did show learning effects, evidenced in increased network stability following the inclusion of a referent. Thus, it did not appear that phonetic accuracy influenced sound pattern organization.

A central question that emerges from the present findings is whether an increase in phonological stability is a positive outcome. As previously stated, processes of variability, reorganization, and stability are thought to index the onset of behavioral transitions (Thelen & Smith, 1994; Vihman et al., 2009; Vihman & Velleman, 1989). However, in this study, the inclusion of a referent did not induce more variability but rather strengthened the stability of phonological forms. Crucially, stability did not promote phonetic accuracy—children with DLD stabilized on incorrect targets. It is possible that a shift toward a more stable production system inhibits children with DLD from acquiring accurate speech targets through flexible exploration (Thelen & Smith, 1994). It is also possible that children with DLD weigh the production of a stable word form more heavily than that of an accurate form. This is consistent with prior work that demonstrates how the linking of a referential cue to a phonological form facilitates the stability of articulatory movements, with no improvement in phonetic accuracy (Heisler et al., 2010). It is important in future studies to assess more protracted word learning trajectories to specify transitions in stability and flexibility of the speech production system over time.

A missing link is to determine whether this novel approach to phonological learning—that is, the organization of phonological constituents—captures a deficit specific to DLD or whether these findings are developmental in nature. For instance, it may be that toddlers, who are notoriously variable in their form production (e.g., Ferguson & Farwell, 1975; Sosa, 2015; Sosa & Stoel-Gammon, 2012), also demonstrate difficulty with the organization of novel sound sequences, suggesting a protracted time-course of phonological organization. Alternatively, it may be that the sequential organization of phonological elements captures a unique characteristic of sound production in DLD. Given the finding that children with DLD benefit from semantic cues in their sound pattern learning, it will be important to consider whether very young children also rely on semantic cues to stabilize their production of novel phonological forms.

In the early stages of word learning, the processes by which children learn words and sounds are intricately interconnected, and the interactivity between lexical–semantic and phonological systems manifests in predictable ways (Ferguson & Farwell, 1975; Oller et al., 1976; Schwartz & Leonard, 1982; Stoel-Gammon, 2011). Over time, the processes by which TD children acquire new words become less reliant on articulatory and phonetic constraints and are integrated within a web of more complex semantic and syntactic relationships (Borovsky et al., 2016; Gleitman et al., 2005; Golinkoff et al., 1994). It appears that children with DLD use lexical–semantic information to stabilize phonological form. Future studies that assess whether this is a remnant of an earlier period of word acquisition in which lexical–semantic information plays a key role in strengthening phonological representations will be especially fruitful for delineating the trajectory of word learning skills in children with typical and atypical language development.

Other Applications of Network Science

The focus of this study was on the organization of syllable sequences as a function of referential status in children with DLD. Network science approaches provided insights not afforded by standard measures of accuracy. We suggest that network science has broader applications to other developmental and clinical issues. As an example applied to phonological factors, manipulations of phonotactic probability (i.e., the likelihood that a sequence of phonemes occurs in a given language) have been shown to affect the learnability of novel word forms, typically measured by phonetic accuracy, in both children with typical language (Heisler & Goffman, 2016; Storkel, 2001) and children with weaker language (Gray & Brinkley, 2011; Munson et al., 2005; Storkel & Hoover, 2011). However, the extent to which phonotactic manipulations relate to phonological sequence learning has not yet been explored (but see Chan & Vitevitch, 2010; Siew & Vitevitch, 2020; Vitevitch, 2008, 2014, for discussions on other network-driven methods for assessing phonological and lexical organization). This is just one example of how network science approaches may extend to other questions related to language processing and production.

We have shown that a network approach reveals deficits in phonological sequence learning in children with DLD. Such an approach may also be informative as applied to children with speech sound disorders. For example, investigators have proposed that subtypes of speech sound disorder may be classified by inconsistency in children's error patterns (Dodd, 2014; Holm et al., 2007; but see Iuzzini-Seigel et al., 2017). A network science approach could contribute to the characterization of different profiles of phonological production patterns that support subtypes of children with speech sound disorders.

Network science approaches may also be applied to sound pattern organization in multilingual children. For instance, research on phonological development in Spanish–English bilingual children has demonstrated that, while children appear to generally organize each language's phonology separately, there is some evidence of interactivity and transfer across phonetic inventories of the two languages (Fabiano-Smith & Barlow, 2010; Fabiano-Smith & Goldstein, 2010). Assessing how sequential sound patterns are organized in bilingual language acquisition using a network science approach has the potential to inform theories of bilingual and multilingual phonological development and to provide a new index of atypical patterns of organization. These are just examples of how the application of network science provides a new avenue for analysis of sound pattern learning across linguistic domains, developmental periods, and speech and language disorders.

Funding Statement

This research was supported by the National Institute on Deafness and Other Communication Disorders Grants R01 DC04826 and R01 DC016813, awarded to Lisa Goffman, as well as F31 DC017904, awarded to Sara Benham. Portions of this work were funded by Grant T32 DC00030 at Purdue University, awarded to Laurence Leonard.