A Real-time Mechanism Underlying Lexical Deficits in Developmental Language Disorder: Between-Word Inhibition

Abstract

Eight to 11% of children have a clinical disorder in oral language (Developmental Language Disorder, DLD). Language deficits in DLD can affect all levels of language and persist through adulthood. Word-level processing may be critical as words link phonology, orthography, syntax and semantics. Thus, a lexical deficit could cascade throughout language. Cognitively, word recognition is a competition process: as the input (e.g., lizard) unfolds, multiple candidates (liver, wizard) compete for recognition. Children with DLD do not fully resolve this competition, but it is unclear what cognitive mechanisms underlie this. We examined lexical inhibition—the ability of more active words to suppress competitors—in 79 adolescents with and without DLD. Participants heard words (e.g. net) in which the onset was manipulated to briefly favor a competitor (neck). This was predicted to inhibit the target, slowing recognition. Word recognition was measured using a task in which participants heard the stimulus, and clicked on a picture of the item from an array of competitors, while eye-movements were monitored as a measure of how strongly the participant was committed to that interpretation over time. TD listeners showed evidence of inhibition with greater interference for stimuli that briefly activated a competitor word. DLD listeners did not. This suggests deficits in DLD may stem from a failure to engage lexical inhibition. This in turn could have ripple effects throughout the language system. This supports theoretical approaches to DLD that emphasize lexical-level deficits; and deficits in real-time processing.

Introduction

Language is fundamental for social interaction, education and culture. Thus, understanding how this complex behavior develops is of paramount importance. The introduction to the vast majority of papers on language acquisition start by explaining the enormous complexity of the problem of acquiring a language due to classic problems like a lack of negative evidence, referential ambiguity or the lack of acoustic invariance. Despite this, most of these introductions end by arguing that for most children it develops effortlessly – a seemingly compelling mystery.

Yet, language development is not guaranteed. Developmental Language Disorder (DLD) is a common developmental disorder in which children exhibit poor oral language despite the lack of an obvious cause such as a hearing impairment, neurological disorder or gross developmental disorder. While prior clinical definitions excluded children with poor non-verbal cognitive abilities, the DSM-5 has dropped this criterion as these are not entirely separable constructs and there is little evidence for a qualitatively distinct profile of language in children with DLD with and without a cognitive deficit (see, Norbury et al., 2016, for a discussion). Importantly, by this definition, DLD affects 8-11% of children (Norbury et al., 2016; Pennington & Bishop, 2009; Tomblin et al., 1997), a number that far exceeds that of other pervasive developmental or cognitive disorders. This extraordinarlily high incidence suggests that many children do not fully solve the problems of language acquisition.

DLD has important developmental ramifications. It is highly stable: children do not “grow out” of these deficits (Bornstein, Hahn, & Putnick, 2016; Conti-Ramsden, St. Clair, Pickles, & Durkin, 2012). Over half of children with DLD also show a clinical reading deficit (Pennington & Bishop, 2009), and DLD predicts poor academic and life outcomes (Clegg, Hollis, Mawhood, & Rutter, 2005; Conti-Ramsden, Durkin, Simkin, & Knox, 2009). Thus, at a practical level, understanding DLD is crucial for improving developmental outcomes in multiple domains. At a theoretical level, it is necessary for understanding the range of developmental trajectories for language, and their impact on multiple domains of cognition.

While traditional approaches to DLD have emphasized syntax, DLD affects every level of language. In fact, deficits in even low-level skills like phonological processing persist through adolescence and adulthood (Clegg et al., 2005; Snowling, Bishop, Stothard, Chipchase, & Kaplan, 2006). In seeking to bridge these lower and higher levels of language, lexical (or word-level) processes are promising (Nation, 2014), as words are a hub in the language system linking speech perception, production, reading, morphology, syntax and semantics. Thus, a lexical-level impairment could cascade throughout levels of language. Indeed, DLD has been linked to a variety of deficits in word learning and lexical processing (Dollaghan, 1998; Mainela-Arnold, Evans, & Coady, 2008; McGregor, Oleson, Bahnsen, & Duff, 2013; McMurray, Munson, & Tomblin, 2014; McMurray, Samelson, Lee, & Tomblin, 2010; Montgomery, 2002).

DLD is often described as a learning deficit (e.g., Ullman & Pierpont, 2005). However, children with DLD show deficits recognizing even highly familiar words (Dollaghan, 1998; Mainela-Arnold et al., 2008; McMurray et al., 2010). This raises the possibility that the core lexical deficit in DLD is not necessarily learning, but is better characterized as a deficit in real-time processes (which could in principle cascade to affect learning). In DLD, differences in real-time word recognition processes could shape the input to learning, or they could make it difficult for children to access their lexical knowledge in the moment. In typical listeners, the cognitive mechanisms of real-time word recognition are well understood; this theoretical basis could offer a useful platform for understanding mechanisms that may be impaired in DLD.

Word recognition is characterized by competition (Dahan & Magnuson, 2006, for a review). As speech unfolds, the partial input is consistent with many words. Rather than waiting to hear the entire word, these candidates are immediately activated. For example, after hearing li-, listeners activate lizard, lips, and listen. As more input arrives, activation is updated and these words compete until one remains. A range of competitors are implicitly considered, including rhymes (e.g., after hearing lizard, wizard, Connine, Blasko, & Titone, 1993), anadromes (e.g., after cat, tack, Toscano, Anderson, & McMurray, 2013), and embedded words (e.g., the bone in trombone, Luce & Cluff, 1998). Thus, lexical competition is a flexible process reflecting listeners’ partial consideration of a range of plausible matches to the input.

A number of studies have characterized the timecourse of lexical competition in children with DLD using the visual world paradigm (VWP; Tanenhaus, Spivey-Knowlton, Eberhard, & Sedivy, 1995). In the VWP, listeners hear a word and match it to a visual referent from a display containing the target (e.g., lizard) and competitors (wizard, liver). The timecourse of fixations to each referent reflects how strongly its label is considered. These studies reveal a characteristic pattern in DLD. McMurray et al. (2010) examined adolescents with DLD and found that the earliest fixations to the target and competitors were similar between adolescents with and without DLD; however, by the end of processing, adolescents with DLD did not fully commit to a single candidate. At asymptote, they fixated the target less and continued fixating competitors at levels greater than baseline unrelated objects (see also McMurray et al., 2014). This mirrors studies with gated stimuli (Dollaghan, 1998; Mainela-Arnold et al., 2008). In these studies, children with DLD and TD show similar responses after hearing early portions of the stimuli, but children with DLD vacillate between interpretations at later gates. This is also seen in an Event Related Potentials (ERP) study by Helenius, Parviainen, Paetau, and Salmelin (2009) who showed a reduced ERP response to repeated words (less detection of the repetition); this was taken as evidence that lexical activation decays more rapidly in DLD.

In sum, this work offers clear evidence for specific deficits in spoken word recognition. VWP, gating and ERP studies all suggest there are not deficits in early lexical access, but there is difficulty in fully resolving competition and maintaining activation over time. Moreover, follow up work suggests that these specific deficits do not likely derive from perceptual problems. Lexical level deficits are not increased when fine-grained acoustic cues are manipulated (McMurray et al., 2014; Montgomery, 2002; Stark & Montgomery, 1995).

This work suggests listeners with DLD do not fully resolve lexical competition. However, it is not clear why they are unable to do so. The ability to resolve competition derives from a number of processes that could be affected in DLD (McClelland & Elman, 1986). For example, if bottom up processes mapping the input to lexical candidates are weaker in DLD, activation for words may build more slowly. Alternatively, listeners with DLD may have difficulty maintaining activation for the correct word (a deficit in lexical “decay”; Helenius et al., 2009; McMurray et al., 2010). Finally, listeners with DLD may be impaired in lexical inhibition (Dahan, Magnuson, Tanenhaus, & Hogan, 2001; Luce & Pisoni, 1998), by which more active words suppress competitors. This is unrelated to controlled forms of inhibition (executive function), but is seen as a marker of whether words are fully integrated in the lexicon (McMurray, Kapnoula, & Gaskell, 2016). Inhibition could help listeners eliminate consideration of incorrect candidates. If inhibition were mistuned in DLD, this could account for their impairments in word recognition.

This study examined lexical inhibition in adolescents with and without DLD. Our goal was not to characterize the “cause” of DLD, and as we discuss later, a deficit in inhibition could be both a product of poor language development and a cause of it. Rather our goal is to more precisely characterize the computational differences at a given point in development that underlie the deficit in word recognition that is commonly observed in DLD.

In order to isolate the effect of inhibition, we used a VWP task developed by Dahan et al. (2001) for typical adults. The vowel of a syllable contains subtle acoustic cues arising from anticipatory coarticulation that predict subsequent phonemes. We created stimuli in which coarticulation favored a competing word. For example, in this condition (the WordSplice condition), the ne in net was spliced from neck to create ne_ckt. This boost to competitor activation was expected to then inhibit the target, delay in its recognition. This was compared to an intact form in which the vowel was taken from another exemplar of the same word (ne_tt, the MatchSplice condition). However, the WordSplice, ne_ckt, is not a good exemplar of the target word; this could slow recognition for reasons unrelated to inhibition, since the input is a less-good match to listeners’ internal templates. Thus, in a third condition the vowel portion came from a nonword (e.g., nep). This provided a comparably poor exemplar of net (ne_pt), which does not activate a lexical competitor (NWsplice) condition. We thus, estimate the precise contribution of lexical inhibition by comparing the dynamics of lexical activation between the WordSplice and NWsplice conditions. Lexical activation should be delayed in both condition, but any enhanced delay in the WordSplice condition would indicate that inhibition was operating. We compared this estimate between DLD and TD children to determine if children with DLD show differences in lexical inhibition.

The dynamics of activating the target word (net) was assessed in a VWP task in which participants heard one of the spliced forms in the presence of four objects: the target (net), a cohort competitor that overlapped in a single phoneme (c.f., Dahan et al., 2001), and two unrelated items. The competing form, neck, was not present on the screen, so any interference can only derive from implicit activation of the competitor and its inhibitory effect on the target.

We assessed this in adolescents with and without DLD. We sampled adolescents for two reasons. First, our prior work on the dynamics of spoken word recognition in DLD has focused on this group. This was originally for convenience, since these children were on the tail of a longitudinal study (McMurray et al., 2014; McMurray et al., 2010). However, the profile of lexical competiion more broadly is well characterized at this age making this a useful place to start. Second, emerging evidence suggests that spoken word recognition dynamics continue to develop during childhood and least up to adolescence (McMurray, Danelz, Rigler, & Seedorff, 2018; Rigler et al., 2015). Indeed preliminary work using the subphonemic paradigm with younger TD children suggests that lexical inhibition may not be adult like until about 12 years of age (Blomquist & McMurray, in preparation). Thus, older adolescents are a sample in which word recognition is somewhat more stable than earlier in development, potentially revealing larger differences between TD and DLD children. Finally, for this paradigm it was important that participants knew all the words (particularly the non-displayed words). Studying older children was one way to ensure this was the case.

We frame predictions here in terms of the lower level differences in the computational properties of the word recognition system (in the general discussion we relate these findings to broader theories of DLD). In particular we offer four predictions.

First, our primary marker of inhibition is the difference in fixations between the WordSplice and NWSplice. Thus, if children with DLD show a deficit in inhibition, this will appear as reduced difference or even no difference between these conditions in adolescents with DLD.

Second, an alternative hypothesis is that word recognition deficits in DLD derive from weaker bottom-up driven activation growth (the ability to activate words from the input). This hypothesis predicts DLD adolescents will show slower activation for the target in all conditions. Moreover, while inhibition (the difference between WordSplice and NWSplice) should be operative, its effects may be observed later in the timecourse of processing (since inhibition cannot have a strong effect until the word becomes somwaht active).

Third, word recognition is built on perceptual and phonological processes; consequently a deficit in auditory perception or phonological processing could cascade to a lexical deficit. Such deficits have been proposed but are fairly controversial (Tallal & Piercy, 1973a; Vandewalle, Boets, Ghesquiere, & Zink, 2012) (though see Bishop, Adams, Nation, & Rosen, 2005; Rosen, 2003). Moreover, the idea that such deficits may underlie deficits in word recognition given prior evidence that word recognition in children with DLD do not show differentially sensitive to fine-grained acoustic or phonetic differences in word recognition tasks (McMurray et al., 2014; Montgomery, 2002; Stark & Montgomery, 1995). We can capture one aspect of such sensitivity by comparing the MatchSplice and the NWSplice conditions (which reflects only a bottom up acoustic mismatch). If a perceptual deficit is the source of the differnces in DLD, then children with DLD should should be differently sensitive to coarticulatory mismatch than TD listeners.

Fourth, in our prior work, we hypothesized that word recognition deficits may be driven by decay (McMurray et al., 2010). This was primarily supported by computational modeling showing which manipulated decay to achieve a better fit to VWP data than other parameters (including inhibition) (and see Helenius et al., 2009, for evidence from ERPs). However, there are uncertainties in the precise dynamics of TRACE, which were developed long before we had tools like the VWP with which to validate them. It is not clear how enhanced decay may play out in the context of the subphonemic mismatch paradigm. However, if no inhibitory deficit is found, this may support an account based on decay.

Finally, we note that word recognition deficits in DLD may derive from differences in more than one processing parameter. Thus, our work is intended to isolate a potential deficit in inhibition, not necessarily to rule out all other such deficits.

Methods

Participants

Seventy-nine adolescents (M= 16;5; range = 14;5 – 19;11; 46 female) participated. Each received $20/hour compensation. All participants (or their parents, if minors) underwent informed consent following a protocol approved by the University of Iowa IRB.

Recruiting and Sample Size.

This was a convenience sample, which presented some difficulty in obtaining. We could not rely on provider referrals (e.g., referres from a speech language pathologist, SLP) for two reasons. First, few adolescents with DLD receive SLP services. Second, provider-referred samples may not be well matched to control participants, since the factors that lead an adolescent to seek SLP services are hard to quantify (see Bishop & Hayiou-Thomas, 2008, for an example of this difficulty). Instead, we used a large group screening, and targeted recruitment based on scores obtained on screening materials.

To this end, we developed two screening tasks that could be delivered orally to groups of students in their high schools. This was conducted in Cedar Rapids (approximately 30 miles from the University of Iowa) which has greater socio-economic and demographic diversity. From these scores, we identified children as potentially DLD or TD, and drove them to Iowa City in groups of 4-6 students. These then received a full language and non-verbal IQ assessment, and were immediately tested in several studies. These studies included several VWP studies as well as structural imaging. Not all children did all studies (based on time, as well as the power identified in advance for each study). As formal assessments scores were not known when a participant’s schedule was created, we had to rely only on screening data to determine the sample for this study, resulting in some imbalances between TD and DLD groups.

Most participants (N=71) were recruited via this screening procedure. Participants completed two screenings (see Supplement S1 for more detail). First, there was a 30 item color-and-shape token task that required participants to match complex sentences to displays of 8 colored shapes. Second, there was a 25-item agent-action vocabulary task. Our work has shown that together, these are highly predictive of language (r= .58) and non-verbal IQ (r=.60). Instructions were delivered orally to groups of students who responded on paper worksheets. We recruited participants with average performance across both tasks less than 20% correct (potentially DLD) or greater than 50% correct (potentially TD). Eight additional participants were recruited as potentially TD using college entrance test scores (SAT Critical Reading scores >= 600 or composite ACT >= 25). They did not undergo the screening procedures, but did undergo full language and cognitive testing.

We targeted 80 participants, under the assumption that we could obtain 30 with DLD using the procedures described below. This led to a minimum detectable effect (for an unpaired t-test) of d = .65 for between group comparisons (e.g., DLD vs. TD in Figure 1A), assuming α=.05, and β=.8, and an MDE of d=.529 and d=.404 for within group comparisons (e.g., Match vs. WordSplice) within the DLD and TD groups respectively. Data were collected in a single wave of recruitment and were not analyzed prior to the completion of the final sample.

Figure 1:

A) Composite language and nonverbal cognitive scores for each participant. B) Distribution of composite language and nonverbal IQ scores.

Language and cognitive assessments.

After screening, participants recruited into the study underwent a language and non-verbal assessment by a trained assessor. Vocabulary was assessed with the Peabody Picture Vocabulary Test Revised Version 4 (Dunn & Dunn, 2007), and the Expressive Vocabulary Test Version 2 (Williams, 1997). Higher level language ability was assessed with the Clinical Evaluation of Language Fundamentals Version 4 (Semel, Wiig, & Secord, 2006): Recalling Sentences, and Understanding Spoken Paragraphs. These four measures were averaged to compute a composite language measure. Nonverbal cognitive ability was assessed using the Weschler Abbreviated Scale of Intelligence Version 2 (Wechsler & Hsiao-Pin, 2011), the Block Design (WASI-BD) and Matrix Reasoning (WASI-MR) subtests. These scores were converted to standard scores and averaged to create a composite non-verbal measure.

Distribution of Participant Ability.

Participants had an average language composite of 93.84 (SD=13.08) and an average non-verbal composite of 93.10 (SD=14.27). For ease of analysis (and for clinical relevance), participants were designated DLD if their composite standard score on these assessments was less than or equal to −1 SD. This resulted in 22 teenagers with DLD (M=16;4; years range 14;7 - 18;7; 15 female) and 57 TD (Typically Developing) teenagers (M=16;5; range 14;5-19;11; 31 female; Figure 1).

As expected, language and nonverbal ability were correlated (r=.63, p<.0001; Figure 1). Participants in the TD group had a mean composite non-verbal score of 96.7 (SD=13.5), and an average language score of 100.2 (SD=8.7). This contrasts with the DLD group where non-verbal scores averaged 83.7 (SD=12.0) and language averaged 77.32 (SD=6.1). This relationship between language and cognition was expected as it had been observed previously in Tomblin’s longitudinal project on DLD, and has since been dropped from the exclusionary criteria for DLD in the DSM-5. Given this, attempts to dissociate language and IQ (either statistically, or via sampling) would have resulted in a sample that is not representative of the population of children who struggle (or succeed) with language (Dennis et al., 2009). Moreover, prior work has found little or no influence of non-verbal intelligence on lexical processing (McMurray et al., 2014; McMurray et al., 2010). Thus, we focused our analysis on language composite scores.

Design and Items

This study used the subphonemic mismatch paradigm (Dahan et al., 2001). Target words consisted of 28 CVC English words (adapted from Kapnoula & McMurray, 2016a), with the constraint that they had one lexical competitor that differed only in final place of articulation (e.g., net / neck), and another that also differed only in place of articulation that did not form a word (e.g., nep). For each target word, three variants were constructed by splicing different onsets (e.g., ne) onto the same coda (t). This left subtle coarticulatory cues in the onsets that would bias the percept toward the correct word or a competing word. In the MatchSplice condition, the onset was taken from the base word and coarticulation matched the coda (ne_tt). In the WordSplice condition, the onset came from the competing word (ne_ckt). In the nonword splice or NWsplice condition, the onset came from a nonword (ne_pt).

Target words were paired with three additional words to use as foils in the VWP. These included one word that shared its initial phoneme with the target word (nurse) and two unrelated words (duck, book). Words in a given set always appeared together across trials. Each of these four words was the auditory stimulus across trials. Each word in a set was played one time in each splice condition (e.g., net was heard once in the MatchSplice, WordSplice, and NWsplice conditions). The three foil items were presented in the MatchSplice and two distinct NWsplice conditions. This led to 28 sets x 3 conditions x 4 words / set = 336 trials.

Four of the 28 experimental items had accuracy below 95% (hub, pit, steak, zip) for MatchSplice trials. As there were no visible competitors on the screen, accuracy should have been quite high. This raised concerns about the auditory or visual stimuli, or the suitability of these words for this population. Thus, these items were removed from all analyses.

Stimuli.

Auditory Stimuli.

Auditory stimuli were constructed from natural recordings of a male talker. For each word (net) and the to-be spliced forms (neck, nep), he produced several exemplars in a neutral carrier sentence (The next word will be…). The most coarticulated exemplar was selected by a group of phonetically trained listeners who listened to the onsets (e.g., ne…) and selected the one that led to the most robust percept of the following consonant. These recordings were then excised from the sentences for use in the experiment. Onsets were separated from their final consonants at the zero-crossing nearest the onset of the closure. In the MatchSplice condition, the onset and coda came from the different exemplars of the same word. We then appended 100 msec of silence to the onset of each stimulus, and the stimuli were amplitude normalized in Praat to 70 dB. The average duration of the onset portions of the experimental words (onset to closure) was 303 msec, and the average duration of the entire stimulus was 534 msec (not including the silent period).

Visual Stimuli.

Visual stimuli consisted of 112 pictures constructed using a standard lab protocol. For each, 5-10 pictures were downloaded from a clipart database, and reviewed by a focus group of graduate and undergraduate students. The group selected the image that was most prototypical for that stimulus. It was then edited for consistency with other visual stimuli. Finally, it was verified by an independent lab member with extensive VWP experience.

Procedure

The experiment was run using Experiment Builder (SR Research, Oakville, On, Canada). After undergoing informed consent, participants were seated in front of a computer in a sound attenuated room. The computer used a 17” (5:4) monitor operating at 1280 × 1024 resolution. The eye-tracker was then calibrated. During the first phase, participants were familiarized to the visual stimuli. They saw each picture with its printed name, and advanced through the pictures by pressing the space bar.

Next, participants began the experimental phase. On each trial participants saw four pictures with one picture near each corner of the screen, and a blue dot in the center (see Figure 2). Pictures were 200 × 200 pixels, separated by 780 pixels horizontally and 524 pixels vertically. After 500 msec, the dot in the center of the screen changed from blue to red, indicating that the participants could use a mouse to click the dot to hear the auditory stimulus. After hearing the auditory stimulus, participants clicked on the picture that best represented the auditory word. The relative positions of the four pictures were randomized on each trial.

Figure 2:

Sample display of a single trial. Here, the target was the experimental item mug (which could be cross-spliced with the non-displayed competitor, mud); the cohort competitor was mitt, and the unrelated items were spark and truck.

For most participants, this experiment was conducted as part of a longer testing procedure in which participants performed one or more eye-tracking experiments, standardized assessments, and structural MRI (for a different study). Each experiment was conducted as a wholly separate block of trials with a break in between them; the order of procedures varied among subjects and was largely counterbalanced.

Eye-tracking Methods

Eye movements were monitored at 500 Hz with an SR Research Eyelink 1000 eye-tracker in chinrest mode. When possible both pupil and corneal reflection were used for tracking eye position. A standard 9 point calibration was used, with a drift correction every 24 trials. The fixation record was automatically parsed by the Eyelink control software into saccades, fixations and blinks with the default psychophysical parameters. As in prior studies (McMurray et al., 2010), a saccade and the subsequent fixation were combined into a single unit, a “look” which began at the onset of the saccade, ended at the end of the fixation¹. In assigning looks to regions of interest, 100 pixels were added to each edge of the images. This did not result in any overlap.

These looks were collapsed into a continuous proportion of fixating each competitor on the screen over time. This consists of the number of trials on which a given candidate was fixated divided by all trials (looks to nothing and blinks are included in the denominator). This will be referred to as the proportion of fixations.

We examined the quality of the eye-tracks by comparing the calibration quality between the two groups. Using the average of the nine points, we found no difference (TD: M=0.434, SD=.152; DLD: M=.442, SD=.17; T(77)=.19, p=.85). Similarly using the maximum deviation, there was no difference (TD: M=.941, SD=.321; DLD: M=1.069, SD=.392; T(77)=1.49, p=.14).

Statistical Analysis

The proportion of fixations over time were analyzed with the bootstrapped difference of timeseries (Oleson, Cavanaugh, McMurray, & Brown, 2017) implemented in the BDOTS R package. In this approach, a 4-parameter logistic function is fit to data for each subject/condition. BDOTs then uses parametric bootstrapping to estimate confidence intervals around these functions (1000 iterations) and conducts an appropriate t-test (paired or unpaired) at each time using these confidence intervals. Next, the autocorrelation among test statistics is used to identify a new alpha level which controls family-wise error. Significant regions are then identified. BDOTs is currently only operationalized for t-tests (not an F statistic), so we conducted pairwise comparisons between groups (within splice condition) and among conditions (within group). We approximate interactions by computing the differences between splice conditions and comparing across groups, and we follow that with a traditional mixed effects model to establish the interactions. . Recommended reporting procedures for BDOTs are to report the adjusted alpha, observed autocorrelation (ρ), and the quality of the individual fits (Seedorff, Oleson, & McMurray, in press).

Results

We first examined accuracy of the mouse-click results on MatchSplice trials, on which participants heard the most natural form of the word. Participants responded correctly on 98.78% of trials, and all participants but one had over 90% accuracy (that one was better than 85%). While DLD participants as a whole (M=97%, SD=4%) were slightly less accurate than TD (M= 99%, SD = %2; F(1,77)=17.67, η_p=0.19, p < .0001), both groups performed exceedingly accurately. In subsequent analyses, eye-movements were only analyzed for correct trials.

Our primary analyses asked whether the time course of fixations differed as a function of language status and experimental conditoin. To determine this, we computed the proportion of trials on which the subject was fixating that target at each 4 msec interval. This was done for each subject in each condition. We then used BDOTs (Oleson et al., 2017) to identify time regions at which pairs curves differed significantly as a function of language group or condition (data and R scripts are publicly available at https://osf.io/kfwxu/). Details of BDOTs analyses are shown in Table 1.

Table 1:

Results of BDOTS analysis. Shown is the corrected alpha (α*) and autocorrelation among the t-statistics (ρ) used to compute α*; the number of fits of varying degrees of quality, and times over which each contrast was significant. Lexical inhibition (bold effects) is indicated when the WordSplice (W) condition was lower than the NWSplice (NW) or secondarily, the MatchSplice condition. Times for significant regions are relative to trial onset, and should be adjusted by −300 msec to reflect 100 msec of silence before the stimulus, and 200 msec to plan and launch an eye-movement.

Analysis	Comparison	Statistics	Fit Quality		Significant Regions
			Level	N
Figure 3a	TD vs. DLD: MatchSplice	ɑ* = 0.0014 ρ= 0.9994	R2>=0.95 R2>=0.8 R2<0.8 Dropped	78 1 0 0	TD > DLD 592-2500msec
Figure 3b	TD: Match vs. NW	ɑ* = 0.0017 ρ= 0.9993	R2>=0.95 R2>=0.8 R2<0.8 Dropped	113 1 0 0	NW > Match 204-612 msec Match > NW 780-2500 msec
	TD: Word vs. NW	ɑ* = 0.0015 ρ= 0.990	R2>=0.95 R2>=0.8 R2<0.8 Dropped	111 3 0 0	NW > Word 528-1276 msec
Figure 3c	DLD: Match vs. NW	ɑ* = 0.0019 ρ = 0.9984	R2>=0.95 R2>=0.8 R2<0.8 Dropped	41 2 0 1	Match > NW 696-2500 msec
	DLD: Word vs. NW	ɑ* = 0.0022 ρ = 0.9989	R2>=0.95 R2>=0.8 R2<0.8 Dropped	40 3 0 1	⊘
Figure 4a	TD and DLD: Match vs. NW	ɑ* = 0.0013 ρ = 0.9982	R2>=0.95 R2>=0.8 R2<0.8 Dropped	153 3 0 1	DLD > TD 464-648 msec
Figure 4b	TD and DLD: NW vs. Word	ɑ* = 0.0015 ρ = 0.9987	R2>=0.95 R2>=0.8 R2<0.8 Dropped	149 6 1 1	TD > DLD 548-1148 msec
Figure 4c	TD and DLD: Match vs. Word	ɑ* = 0.0019 ρ= 0.9992	R2>=0.95 R2>=0.8 R2<0.8 Dropped	15 5 0 0	TD > DLD 688-1116 msec
Figure 5	TD: Match vs. NW	ɑ* = 0.0005 ρ = 0.9988	R2>=0.95 R2>=0.8 R2<0.8 Dropped	43 1 0 0	NW > Match 428-552 msec Match > NW 820-2500 msec
	TD: Word vs. NW	ɑ* = 0.0005 ρ = 0.9987	R2>=0.95 R2>=0.8 R2<0.8 Dropped	42 2 0 0	NW > Word 524-1184 msec

Our first analysis replicated the failure to fully commit to the target shown by prior work (McMurray et al., 2014; McMurray et al., 2010). This examined only the MatchSplice trials, as the closest match to prior work which used unmodified words. Figure 3A shows a similar profile of target looking to prior studies. Both groups showed similar looking early, but displayed a significant deviation starting at 592 msec, and which was maintained throughout the rest of trial, particularly at asymptotic portions of the curves. It takes 200 msec to plan and launch an eye-movement and there was 100 msec of silence preceding the word. As a result, these effects arose at roughly 300 msec into the word (roughly after the closure) and persisted throughout the timecourse of processing, with the largest differences at roughly 1000 msec (700 msec into processing) and later. This is consistent with prior work that suggest the effect of language ability is not immediate (e.g. at word onset), but primarily affects the later portion of the timecourse of processing (Dollaghan, 1998; McMurray et al., 2010). Additional analyses reported in online Supplement S2 asked whether DLD and TD children differed in the immediacy of lexical access. These showed no differences, confirming that word recognition deficits in DLD are not a matter of delayed or slowed lexical access, but rather reflect differences later in the timecourse of processing (Dollaghan, 1998; Mainela-Arnold et al., 2008; McMurray et al., 2010).

Figure 3.

A) Looks to target for MatchSplice trials for TD and DLD participants. Average fixations are smoothed with a 12 msec triangular windows, and error bars computed with BDOTs. Trial onset is at 0 msec (100 msec before stimulus onset). Target fixations were significantly lower in DLD than TD children from 592 – 2000 msec (Table 1). B) Looks to the target as a function of time during MatchSplice (M), WordSplice (W) and NWsplice (N) trials for TD participants. Target fixations in the WordSplice (W) condition were significantly lower than the NWSplice (N) condition from 528 -1276 msec (indicating the effect of lexical inhibition during this window; see Table 1 for other contrasts). C) Looks to the target as a function of time and stimulus condition for DLD participants. At no point was the WordSplice condition lower than the NWSplice (though there was an effect of coarticulation).

Next, we examined the TD listeners to replicate Dahan et al. (2001) and document effects of both splicing and inhibition. We first isolated the effect of coarticulatory mismatch as the contrast of the NWsplice condition with the MatchSplice condition in the TD group (Figure 3B). TD participants showed fewer target looks in the NWsplice condition relative to the MatchSplice, from 780 msec through the end of the trial. This indicates sensitivity to the splicing manipulation², starting at 480 msec into processing (just before the end of the word). The fact that this effect was not immediate suggests the effect of mismatching acoustic input may have acted to destabilize lexical activation (rather than delay its initial growth). Importantly, an effect of lexical inhibition was revealed in the significant difference in between the WordSplice condition with the NWsplice from 528 and 1275 msec. Given the 300 msec delay (oculomotor delay + stimulus onset), this effect began 228 msec into processing, before the end of the the final stop consonant. This means that TD listeners were showing evidence of activating the competing word at about the time when any acoustically biasing information would have been available. The difference between the WordSplice and NWSplice conditions persisted out to about 975 msec in processing – well after the end of the word (at 534 msec) suggesting a long-lived inhibition effect.

DLD participants showed a markedly different pattern (Figure 3C). DLD listeners also showed similar reductions in target fixations due to the coarticulatory mismatch (the NWsplice vs. MatchSplice contrast, which was significant from 696-2500 msec). However, the NWsplice and WordSplice conditions did not differ significantly at any time.

We next sought to determine if the pattern of response to each contrast differed significantly between groups (approximating an interaction). For this, for each participant, we computed the difference between stimulus conditions over time. This difference curve captures the amount of interference observed in a particular condition. Here, a positive curve indicates the presence of interference, while a negative curve indicates facilitation. These were again compared with BDOTs (Table 1; Figure 4A). Panel A shows the expected interference due to coarticulatory mismatch. As expected both DLD and TD adolescents showed similar degrees of interference due to coarticulation, with only a brief difference between 464 and 648 msec. However, when we computed the effect of lexical inhibition (NWsplice – WordSplice; Figure 4B), we saw a startling reversal, with DLD adolescents showing a negative effect (facilitation) that differed from the inhibitory effect shown by the TD children from 548 to 1148 msec. In processing time this translates to an effect that began shortly before closure (248 msec) and ended about 200 msec after word onset (848 msec). We were concerned that this measure may reflect noise in the NWsplice condition, so we also computed lexical inhibition as the MatchSplice – WordSplice (which includes both the lexical effect and the sensitivity to coarticulatory mismatch; Figure 4C). This showed a difference from 688 to 1116 msec, with again DLD listeners showing less evidence of inhibition.

Figure 4.

Difference curves for fixations to target as a function of stimulus condition. A) Effect of coarticulatory mismatch: NWsplice (N) trials subtracted from looks to target for MatchSplice (M) trials for TD and DLD participants. These did not differ between groups except for a short period at 496-552 msec. B) Effect of lexical inhibition, over and above coarticulatory mismatch: WordSplice trials subtracted from N trials. This showed a significant group difference between 512 and 1780 msec with greater inhibition in the TD group. C) Effect of lexical inhibition, including coarticulatory mismatch: W trials subtracted from M trials. These differed from 696-1048 msec with greater inhibition in TD adolescents.

Items effects and Interactions.

Current versions of BDOTs have two limitations relevant to the present study. First, because there is rarely sufficient data to fit curves by subject and item simultaneously, it is not possible to capture item-level variance. This is less an issue here because the within subject variable is also within-item (Raaijmakers, Schrijnemakers, & Gremmen, 1999), and because the primary variable of interest is between-subject. Second, BDOTs cannot directly compute interaction terms, but must do so indirectly via a differences of differences (as in Figure 4).

To confirm that the effects were robust against both concerns, we replicated our primary (planned) analyses with a mixed effects model using a traditional area under the curve approach as the DV. This of course, has the added researcher d.f. of which time window to use (a primary reason we planned on a BDOTs approach). Here, we overcome this by using a time window identified as significant in the BDOTs analysis. We thus computed the average target fixations between 500 and 1300 msec. This window was chosen based on the WordSplice vs. NWSplice analyses: these conditions differed from 528 to 1276 msec; and the difference of difference analyses suggested group differences between these conditions in a similar window (688 to 1186).

A linear mixed model was constructed with group (TD=−.5, DLD=+.5, and then centered), and condition as DVs. Condition was coded as two contrast codes: MatchSplice (.67) vs. other splice (−.33, Word and NW) and NWSplice (+.5) vs. WordSplice (−.5). Random effects structure was selected via an examination of the full model-space, using AIC to select the best fitting model (Seedorff, Oleson, & McMurray, submitted). This led to a random slope of both condition codes on subject and item. Significance of fixed effects was evaluated using the Satterthwaite approximation of d.f.

This model found a significant effect of language status (B=.0656, SE=.032, t(77.1)=2.075, p=.0413) with fewer fixations in the TD than DLD group. Neither condition contrast was significant as main effects (Match vs. Other: B=.0289, SE=.023, t(68.5)=1.26, p=.21; NW vs. WordSplice: B=−.0182, SE=.024, t(85.1)=−.76, p=.45). Moreover, while language group did not interact with the Match vs. Other contrast (B=.0116, SE=.021, t(1195.0)=0.57, p=.57), it did interact with the NWSplice vs. WordSplice contrast (B=.0475, SE=.024, t(2955.8)=2.02, p=.0434). Thus, this confirms the differential response of the DLD children to the manipulation of inhibition.

Effect of Non-verbal skills.

Given the expected correlation between language and non-verbal cognitive skills, we asked if this deficit was specific to language abilities or if it also reflected some non-verbal component. To this effect, we conducted an exploratory analysis in which we identified the subset of 22 TD children (henceforth, the TD/M group) that were best matched on overall non-verbal skills. This was done by selecting the TD participant whose non-verbal skills were the single closest match to each DLD participant. If the lack of evidence for lexical inhibition in the DLD group derives from their poor non-verbal IQ, we should observe no evidence for inhibition in this group.

This group was closely matched to the DLD adolescents in non-verbal cognitive ability (M_DLD=83.7, SD=12.0; M_TD/M=83.6, SD=6.1; T(42)=.02, p=.98), but they preserved the difference in language (M_DLD=77.3, SD=6.1; M_TD/M=96.4, SD=7.2; T(42)=9.4, p<.001). Further, the TD/M group did not differ from the DLD children in age (T(24)=1.0, p=.29).

As Figure 5 shows, there was a continued difference between the WordSplice and NWsplice conditions. This was confirmed by a BDOTS analysis which revealed a significant difference between 524 and 1184 msec. This suggests that the effect of DLD on inhibition was driven by their poor language skills, not their poor cognitive abilities, since the latter alone was not sufficient to drive a difference in lexical inhibition in TD children.

Figure 5:

Looks to the target as a function of time and condition in the TD/M group who was matched to the DLD group on non-verbal cognitive abilities.

Sample Size and Power.

A final concern is the imbalance in the sample sizes between the TD (N=57) and DLD (N=22) groups. What if the apparent null effect in the DLD group is simply due to the fact that this group had a reduced sample size? This is unlikely for two reasons. First, our findings are also supported by a significant difference of differences between groups (Figure 3), and by a significant interaction in a traditional mixed model). Second, the prior analysis used a TD sample that was matched on both size and IQ to the DLD group. This analysis still found the effect.

However, the broader question remains as to whether we would have had the power to detect a significant difference between the WordSplice and NWSplice conditions in the TD group if we used the DLD sized sample of 22? If we did not have the power, this may call into question the inability to detect an effect of inhibition in the DLD group.

To address this, we conducted a resampling analysis. On each run of this simulation, we randomly selected 22 of the 57 children in the TD group. These were then subjected to a BDOTs analysis comparing fixations to the target in the NWSplice condition to those in the WordSplice condition. This was done 500 times to determine the power of detecting an effect of inhibition with only 22 participants. We found that in 419 / 500 runs (83.8%) there was a significant difference between the NWsplice and WordSplice conditions in the predicted direction (more fixations to the target in the NWSplice condition). Moreover, the average duration of this significant time window was 639 msec and it started on average at 597 msec – about what was detected in the primary analysis. This suggests that we would have had ample power to detect an effect in the targeted range in a design with equal samples.

Discussion

We examined the timecourse of spoken word recognition in adolescents who had DLD or who were TD. Specifically, we asked whether their documented deficits in spoken word recognition (Dollaghan, 1998; Helenius et al., 2009; Mainela-Arnold et al., 2008; McMurray et al., 2010; Montgomery, 2002; Stark & Montgomery, 1995) derive from differences in local inhibition between lexical items (i.e., the degree to which a strongly active word can suppress competitors). Our analyses first showed that adolescents with DLD showed a reduction in asymptotic target fixations relative to TD adolescents, even in the MatchSplice condition. This is consistent with prior studies (McMurray et al., 2014; McMurray et al., 2010) (and see, Dollaghan, 1998; Helenius et al., 2009; Mainela-Arnold et al., 2008). This is typically interpreted as an inability to fully commit to the target (or suppress competitors). Indeed this is likely to be the primary functional deficit in word recognition that is associated with DLD, an inability to fully resolve competition (or in competition resolution).

Our study offers a mechanistic account of these deficits in word recognition, by revealing one mechanism underlying such differences: differences in lexical inhibition. In TD adolescents (Figure 3B), when coarticulation at word onset favored another word (the WordSplice condition), interference was seen over and above the simple effect of coarticulatory mismatch (the NWSplice condition). In contrast, for adolescents with DLD (Figure 3C), the WordSplice condition did not differ from the NWsplice condtion, offering no evidence for functional lexical inhibition in this population. Moreover our follow-up analysis showed evidence of lexical inhibition in a sub-set of the TD adolescents who had preserved language but with poor non-verbal skills (equal to the DLD group). This offers clear evidence linking lexical inhibition to specifically language ability.

We also note that this is highly specific deficit. Adolescents with DLD did not show a different response when the stimulus reflected a purely acoustic mismatch with the target word (the NWsplice condition). This was somewhat remarkable for two reasons. First, as we describe there are longstanding debates over whether DLD may derive from a phonological or auditory deficit (Bishop et al., 2005; Corriveau, Pasquini, & Goswami, 2007; Rosen, 2003; Tallal & Piercy, 1973b)—these results suggest fairly precise auditory encoding. Second, they also suggest that the deficit we observe here is highly specific to lexical level interactions.

Limitations

Before discussing the implications of these findings for DLD, we note a few limitations of our study. First, we examined adolescents, and it is not clear whether or not these findings generalize to younger children with DLD. As we describe, addressing this question is important for determining whether a deficit in lexical inhibition developmentally “causes” DLD, or whether poor language development in general (DLD) leads to differences in inhibition (or both).

A second concern is whether the DLD children were activating the competitor words (e.g., neck when the stimulus was ne_ckt) at all. This was not something we could measure since the presence of referent for the competing word on the screen would have created an additional confounding difference between the WordSplice and NWSplice conditions. However, there is little reason to think that they would not have shown early competitor activation. Prior VWP studies (McMurray et al., 2014; McMurray et al., 2010) suggest that early competitor fixations do not differ in children with DLD, and gating work confirms something similar (Dollaghan, 1998; Mainela-Arnold et al., 2008). Moreover, competitor activation is typically construed as a consequence of the fact that listeners activate words immediately; since they do not wait for the entire word to begin lexical access, multiple items (the target and competitors) are active simply because the early portions of the stimulus (e.g., ne…) are consistent with many words. In this regard, we showed that DLD children and TD do not differ in the immediacy of word recognition (see Supplement S2), again, suggesting that the competing word would have been similarly active. Thus, while we cannot conclusively rule out this hypothesis, there is little evidence in favor of it.

Third, it is important to confirm these findings with other measures to rule out the possibility that the absence of lexical inhibition in adolescents with DLD reflects a task-specific response rather than a general pattern of language processing. In this regard, real-time measures of brain-function like EEG and MEG are promising and have already revealed deficits in the timecourse of lexical processing in DLD (c.f., Helenius et al., 2009; Malins et al., 2013). However, the mapping between typical components studied in these paradigms (e.g., the N400, the N1) and the kind of lexical competition processes studied here is indirect, and as such there is currently no direct index of lexical inhibition using such measures (though we do not think such a measure is implausible). This raises the need for further methodological work.

Components of Competition

In computational models of spoken word recognition (Hannagan, Magnuson, & Grainger, 2013; McClelland & Elman, 1986; Norris, 1994), the process of resolving competition derives from several computational properties of the system: lexical inhibition clearly plays a role, but resolution is also dependent on the strength of bottom-up activation flow, the decay or maintenance of lexical activation, and other processes. Could differences in these other pathways give rise to what looks like an impairment in inhibition?

For example, if learners are slower to activate a competing word, it won’t be sufficiently active to inhibit the target. However, this might predict children with DLD would be delayed early in processing, or would show an effect of inhibition later in processing. Neither were observed.

Moreover, we found no evidence of a perceptual deficit in our data. Children with DLD were just as sensitive to fine-grained coarticulatory differences that were not predicted to alter lexical activation. Indeed for a brief period (Figure 4A) they appeared to be more sensitive (or more rapidly sensitive). This adds to the evidence that deficits in lexical processing may be just that, and that they do not derive from perceptual factors (McMurray et al., 2014; Montgomery, 2002; Stark & Montgomery, 1995). That is not to discount the possibility of an auditory deficit. While this is still a matter of debate (Bishop et al., 2005; Rosen, 2003), large scale studies show some evidence for general issues in auditory processing (Grube, Kumar, Cooper, Turton, & Griffiths, 2012; Vandewalle et al., 2012). We add then that it is unlikely that deficits in word recognition are the specific consequence of an auditory deficit, rather auditory deficits may exert their influence via other aspects of language.

The possibility of a deficit in lexical decay is more challenging to explain. In McMurray et al. (2010), we proposed that increased decay might prevent the target word from fully activating (e.g., it is “leaky”), and this would lead it to inhibit competitors less, resulting in greater activation for competitors. This was supported by TRACE modeling showing a strong fit to the data. It is not clear if such a deficit could lead to differences in the response to the WordSplice condition, though it would clearly depend on a complex interaction of factors.

While we cannot rule out such differences, no matter which pathway is impaired, one can look at the results from this paradigm at a more functional level. That is, it asks if a briefly activated competing word (neck in the context of ne_ckt) can become active rapidly enough to suppress activation for the target in real-time. Achieving such rapid interactions may require many computational pathways (inhibition, decay, etc.) to be tuned appropriately. Our data suggest that functionally speaking, the lexical network in adolescents with DLD does not process words in a way that allows inhibition to emerge in real-time. As a result, these listeners may not be able to activate and suppress competitors rapidly enough to take advantage of this adaptive component of word recognition.

How can a small deficit in inhibition relate to language deficits as a whole?

While our study was narrowly designed to identify a difference in lexical inhibition, such a deficit could be impactful throughout levels of language processing, or it could be a marker of a broader deficit.

In our experimental preparation, inhibition manifested as greater interference. This was due to a specific experimental manipulation, however, and should not be construed as a claim that inhibition is deleterious for word recognition. Inhibition occurs automatically and is pervasive during even normal language processing. For example, inhibition is the dominant explanation for effects of neighborhood and cohort density on word recognition (Luce & Pisoni, 1998; Magnuson, Dixon, Tanenhaus, & Aslin, 2007), effects that can be seen in a variety of task with unmanipulated words. In fact, in everyday language processing, lexical Inhibition is thought to be functionally useful, helping listeners resolve brief periods of competition more quickly and more fully (McClelland & Elman, 1986), deal with embedded words (Gow & Gordon, 1995), and generally achieve more stable representations of the input. Thus, while our paradigm isolated inhibition in a rather unecological way, it was meant to measure something that is quite normal and useful in everyday language.

Deficits in lexical level inhibition could have cascading consequences throughout language. Work in sentence level and syntactic processing suggests that higher levels of language (e.g., semantics, syntax) do not wait for lexical competition to resolve (Apfelbaum, Blumstein, & McMurray, 2011; Levy, Bicknell, Slattery, & Rayner, 2009; Zwitserlood, 1989). Thus, if lexical level processes are not converging on a single candidate, this could have deleterious downstream effects. In this context, good inhibition may help suppress activation for lexical competitors’; this in turn would block the ability of these (irrelevant) words to activate new (irrelevant) semantic and syntactic interpretations. Such a cascade can explain how a small deficit in lexical inhibition could lead to deficits throughout language. If word recognition does not settle on a single interpretation of a word, this could create noise throughout the system. This could help account for “downstream” deficits in sentence processing (Borovsky, Burns, Elman, & Evans, 2013; Nation, Marshall, & Altmann, 2003) as the inability to resolve competition for each word leads to the accumulation of possibilities in sentence processing (see Levy et al., 2009, for evidence of this kind of cascade in typical listeners). To the extent that competition at one level must resolve on a single candidate (e.g., a word) for the next level, the lack of inhibition creates higher entropy (noisier) representations for subsequent processing.

Such a deficit could also impair language learning. Apfelbaum and McMurray (2017) showed that while learning new word/object mappings, learners link available meanings to both the target word and to briefly active competitors. In children with reduced inhibition, this effect could be enhanced, leading developing learners to form spurious associations between a referent and many competing words. Indeed, computational models of phoneme and word learning (McMurray, Aslin, & Toscano, 2009; McMurray, Horst, & Samuelson, 2012) suggest that local inhibition may be essential for unsupervised statistical learning.

However, a deficit in local forms of inhibition like lexical inhibition may not be specific to word recognition. Competition may underlie many aspects of language including syntactic and semantic processing as well as speech production (MacDonald, Pearlmutter, & Seidenberg, 1994; Rapp & Goldrick, 2000). Thus, the deficit in lexical inhibition observed here could also be a marker of broader deficits in lateral inhibition throughout language. Supporting this, deficits in resolving competition have also been noted at the semantic level in DLD (Norbury, 2005). Lateral inhibition is also critical in visual word recognition (Castles, Davis, Cavalot, & Forster, 2007; C. J. Davis & Lupker, 2006), potentially explaining links between DLD and reading disability. Moreover, there is increasing interest in local inhibition as a process embedded in many perceptual and cognitive processes (as opposed to a centralized cognitive controller: Eisenreich, Akaishi, & Hayden, 2017). Thus, a deficit in local inhibition could account for slowing in many processes in DLD (Miller, Kail, Leonard, & Tomblin, 2001), as inhibition is a key part of models of decision making and response selection (Usher & McClelland, 2001).

This argument (and our data) should not be construed as offering evidence that a deficit in inhibition precedes and causes DLD. Our work suggests that a concurrent deficit in inhibition could cascade to concurrent deficits in language processing and learning. It is equally likely that poor language development may be the developmental cause of an inhibitory deficit. For example, a slower growing vocabulary (McGregor et al., 2013; Rice & Hoffman, 2015) could place less pressure on the lexical system to develop inhibitory connections. Potentially supporting this, the efficiency of lexical competition appears to be developing develop through adolescence (McMurray et al., 2018; Rigler et al., 2015; Sekerina & Brooks, 2007), and a recent study from our lab using a similar paradigm (Blomquist & McMurray, in preparation), suggests lexical inhibition may not be robust until about 12. That suggests that lexical inhibition may be developmentally intertwined with broader aspects of language development. Longitudinal work will help characterize how developmental changes in language capacities lead to changes in word recognition and inhibition, and how these may cascade to drive developmental change in other language capacities. While these developmental questions await an answer, our work clearly shows that a key mechanistic factor driving the word recognition deficits in adolescents with DLD is a lack of functional inhibition.

Theories of DLD and the Source of the Inhibitory Deficit

While our study clearly identifies a deficit in lexical inhibition, it is unclear what developmental factors give rise to differences in lexical inhibition and how this might relate to broader thinking on DLD. Our study was not designed to test these broader ideas (which would probably require more than one study), so this discussion cannot be conclusive.

Vocabulary.

As we’ve mentioned vocabulary growth is one possibility: identifying words in the face of increasing numbers of competitors may place demands on the system to develop stronger inhibition. As we describe shortly, we’ve recently documented exactly this kind of plasticity in the laboratory (Kapnoula & McMurray, 2016a). In this way, the lack of inhibition in DLD may be a consequence of their smaller vocabularies. However, at the same time, lateral inhibition may be essential for learning, particularly under high degrees of uncertainty (McMurray et al., 2012); thus, this could also be a causal mechanism in their poor vocabulary.

Executive Function.

A second possibility is domain general executive function, deficits of which have been linked to DLD (Henry, Messer, & Nash, 2012). Executive function is not a single skill, but a constellation of functions (Barkley, 2012; Miyake et al., 2000). Of these, inhibitory control is likely the function most relevant for managing competition among words. This domain general, controlled form of inhibition can be distinguished from the more local form of inhibition we studied here (c.f., Eisenreich et al., 2017). In contrast to general models of EF, the local form of inhibition studied here is implicit, automatic and embedded in a specific system,.in this case, the lexical system (see Munakata et al., 2011, for a discussion of how these forms of inhibition relate).

It is unclear if domain general cognitive control plays a role in resolving lexical competition in optimal listening conditions (e.g. in quiet). This argument cannot be ruled out by the present data. However, a priori, the argument that EF plays such a role is tenuous for several reasons. First, all models of spoken word recognition conceptualize inhibition as local among words (lateral inhibition) (Hannagan et al., 2013; McClelland & Elman, 1986; Norris, 1994). That is, inhibition is localized to the lexicon, not a top down operation originating from a separate central executive control system. Thus, if EF is involved, this would require a rethinking of substantial work in psycholinguistics. Second, there is no empirical evidence linking top-down inhibitory control to lexical competition. The only finding that comes close is a recent study by Zhang and Samuel (2018) that links domain general capacity or resource limits (but not inhibitory control) to competition resolution. However, Zhang and Samuel found that capacity primarily influences late regions in the time course of processing, whereas we found differences starting at 284 msec post-stimulus onset. Finally, one might have predicted that if domain-general EF were involved in lexical inhibition, we would have found that lexical inhibition in our task is more strongly linked to non-verbal cognitive skills than language. However, the IQ matched sample (the TD/M group, Figure 4) shows that the deficit in lexical inhibition is clearly rooted in language. Thus, there is little evidence to suggest that a deficit in executive function may underlie the word recognition deficits observed in DLD. However, at the same time, the conceptual overlap between the two forms of inhibition (domain-general and local) clearly warrants studies that explicitly examine the link between these constructs. Here, a comparison between standard EF tasks (e.g., Flanker, Stroop, Simon) and sub-phonemic mismatch may be an ideal platform in which to ask this question.

Learning.

Third, there has been increasing interest in the idea that learning may be impaired in DLD. This has been linked to specific learning systems such as the procedural (Lee, Nopoulos, & Tomblin, 2013; Ullman & Pierpont, 2005) as well as to more functional forms of learning such as word learning or sequence learning (Hsu & Bishop, 2014) that may not be restricted to single neurobiological system. At this point substantial evidence exists for deficits in a variety of learning processes (see Obeid, Brooks, Powers, Gillespie-Lynch, & Lum, 2016, for a meta-analysis of a number of forms of learning).

An important theory in this vein is that DLD derives from a procedural learning deficit (Hedenius et al., 2011; Tomblin, Mainela-Arnold, & Zhang, 2007; Ullman, 2004). It is not clear whether this can account for our data. One could argue that procedural learning is necessary to establish the lexical network, and thus our deficit could be a product of poor learning. However, countering this, most versions of the procedural learning account (Ullman, 2004) suggest that lexical development is largely declarative, not procedural. This is a view that is consistent with work suggesting sleep-based consolidation is necessary for establishing inhibitory connections among words (M. H. Davis & Gaskell, 2009; Dumay & Gaskell, 2007). However, we note that which learning systems are involved in word learning is still an unresolved question (Fernandes, Kolinsky, & Ventura, 2009; Kapnoula & McMurray, 2016b; Kapnoula, Packard, Gupta, & McMurray, 2015). Moreover, and children with DLD may also show deficits in learning that are not procedural (Hsu & Bishop, 2014). Thus, without a clearer sense of a) which learning systems are impaired or not, and b) which learning systems underlie the formation of inhibitory connections in the lexicon, it is not clear that a learning based account would predict our findings a priori.

One intriguing possibility however, is that these inhibitory deficits are not the consequence of learning deficits, but they are the source: an inability to resolve competition in the moment makes it more difficult to learn language. Many accounts of learning neglect the role of robustly encoding or representing the input prior to linking it to other representations. Consequently, a deficit in real-time processing (e.g., from an inhibitory deficit) may simply provide poorer fodder for learning such that even an intact learning system may show poor outcomes. In fact, both empirical and computational work suggests that robust suppression of competitors in the moment may be important for establishing more robust mappings during learning (Apfelbaum & McMurray, 2017; McMurray et al., 2012; Zhao, Packard, McMurray, & Gupta, in press). This prompts a need to consider the relationship between real-time and learning-time processes in development and in language impairment.

Neural processes.

Finally, the differences we observe here in lexical inhibition are likely mediated by neural differences. Spoken word recognition involves a broad cortical network including temporal and frontal areas, as well as white matter tracts like the arcuate and uncinate fasciculi that connect them (Gow, 2012; Hickok & Poeppel, 2007). Structural changes throughout this network could mediate these differences (Leonard, Eckert, Given, Virginia, & Eden, 2006; Verhoeven et al., 2012). To the extent that lexical inhibition is local to the lexical system, we suspect the largest neural changes will be found within regions of these networks. Here, structural or functional differences in inhibitory interneurons may account for these effects (Markram et al., 2004). These interneurons form the GABAergic system that provides inhibitory functions in the cortex. This complex system may offer a number of sources for variation in inhibitory function.

Plasticity.

Ultimately, however, such differences may reflect developmental differences, not cause them. Thus, it is crucial to investigate avenues of plasticity in lexical inhibition. This was investigated by a recent study in our laboratory (Kapnoula & McMurray, 2016a). TD adults were trained on a common set of tasks and items under conditions which either stressed the need to resolve differences among highly similar words (e.g. distinguishing neck and net was required for responding correctly in the task), or did not. After 30 minutes of training, the group exposed to the higher competition version of training were better able to resolve competition (tested in the subphonemic mismatch paradigm used here). This suggests that the cognitive and neural underpinnings of lexical competition may be responsive to the right kind of experience. Further, it opens avenues for remediation of this deficit that may be core to a widespread developmental disorders.

Conclusion.

Our data suggest a specific deficit in real-time processing of familiar words, and they isolate this deficit to local inhibition among words. This offers mechanistic explanation for concurrent deficits in spoken word recognition, and with the potential to explain some concurrent deficits in other areas of language. Developmentally, it is not clear whether an inhibitory deficit is a cause of DLD, or a consequence of poor development. However, the computational specificity (and importance) of inhibition provides a useful target for further studies examining development.

The deficit in lexical inhibition we have observed here is not clearly predicted nor explainable by any extant accounts of DLD. We are not proposing it as a new unifying account, merely as a specific computational mechanism that appears to be awry at this age and which may underlie concrete deficits in word recognition also observed at this age. However, a deficit in inhibition has the potential to explain a wide number of studies showing lexical level deficits in DLD (Dollaghan, 1998; Mainela-Arnold et al., 2008; McMurray et al., 2014; McMurray et al., 2010; Montgomery, 2002; Stark & Montgomery, 1995), and they may cascade to affect processing at multiple levels. This promotes a different view of the core deficits in DLD that highlights two things that are deemphasized in most thinking on this disorder. First, DLD disorder may be characterized by underlying differences at the lexical (Nation, 2014), rather than phonological or grammatical level. Such a lexical deficit could explain the pervasive deficits observed in DLD, given the fact that words link phonology, syntax, semantics and orthography Second, we suggest that the lexical deficit may be one of real-time lexical processing, not necessarily learning or representation. Our work here identifies a specific aspect of processing, local inhibition among candidate words. Given plasticity in this aspect of word recognition (Kapnoula & McMurray, 2016a), this suggests that understanding the real-time dynamics of language processing may offer unique insight into the causes and treatments of developmental disorders like DLD.