Authors: Justin B. Kueser, Arielle Borovsky, Patricia Deevy, Mine Muezzinoglu, Claney Outzen, Laurence B. Leonard
Categories: Language
Source: Journal of Speech, Language, and Hearing Research : JSLHR
Children with developmental language disorder (DLD) tend to interpret noncanonical sentences like passives using event probability (EP) information regardless of structure (e.g., by interpreting “The dog was chased by the squirrel” as “The dog chased the squirrel”). Verbs are a major source of EP information in adults and children with typical development (TD), who know that “chase” implies an unequal relationship among participants. Individuals with DLD have poor verb knowledge and verb-based sentence processing. Yet, they also appear to rely more on EP information than their peers. This paradox raises two (a) How do children with DLD use verb-based EP information alongside other information in online passive sentence interpretation? (b) How does verb vocabulary knowledge support EP information use?
We created novel EP biases by showing animations of agents with consistent action tendencies (e.g., clumsy vs. helpful actions). We then used eye tracking to examine how this EP information was used during online passive sentence processing. Participants were 4- to 5-year-old children with DLD (n = 20) and same-age peers with TD (n = 20).
In Experiment 1, children with DLD quickly integrated verb-based EP information with morphosyntax close to the verb but failed to do so with distant morphosyntax. In Experiment 2, the quality of children's sentence-specific verb vocabulary knowledge was positively associated with the use of EP information in both groups.
Depending on the morphosyntactic context, children with DLD and TD used EP information differently, but verb vocabulary knowledge aided its use.
https://doi.org/10.23641/asha.25491805
Children with developmental language disorder (DLD) have a persistent difficulty with language use and understanding that cannot be attributed to hearing loss, gross neurological impairment, or low nonverbal intelligence (Bishop et al., 2016) and are frequently encountered in the general population (comprising 7% of school-age children; Tomblin et al., 1997). Compared to children with typical development (TD), children with DLD have poorer sentence comprehension due to their reduced response to morphosyntactic information and word order cues (Montgomery et al., 2017) and reduced working memory capacity, speed of processing, and attention (Leonard et al., 2007; Montgomery & Evans, 2009; Montgomery et al., 2018), among other factors (Leonard, 2014). One area of strength, however, appears to lie in how children with DLD leverage event knowledge to interpret sentences; they demonstrate better comprehension of sentences describing probable compared to improbable events (compare, e.g., “The dog chases the squirrel” to “The squirrel chases the dog”), sometimes approaching the performance of their peers with TD even with complex sentences like passives (Gowie & Powers, 1972, 1978; Hutson, 1975; Hutson & Powers, 1974; Powers, 1973; Precious & Conti-Ramsden, 1988; van der Lely & Dewart, 1986).
Like children with DLD, adults and children with TD experience an advantage for processing probable over improbable language (e.g., Chapman & Kohn, 1978; Ferreira, 2003; Hagoort et al., 2004; Matsuki et al., 2011; Milburn et al., 2016; Troyer & Kutas, 2020; Warren et al., 2008). Knowledge of event probability (EP) may aid unfolding and incremental language comprehension (Venhuizen et al., 2019), serve as a default “good enough” interpretation if deeper interpretive processes fail (Ferreira & Patson, 2007), or override other noise and uncertainty in language input or representations (Hahn et al., 2022). EP-driven strategies are so robust that they can even lead to misinterpretation, particularly in sentences with complex semantics or syntax (e.g., Ferreira, 2003; Paape et al., 2020; Wason & Reich, 1979). For example, even adults sometimes interpret simple passive sentences like, “The dog was chased by the squirrel,” as “The dog chased the squirrel,” because the event described is so improbable (Ferreira, 2003).
Comprehenders rapidly compute and use EP information based on a range of event-related conceptual knowledge (McRae & Matsuki, 2009). Such knowledge is closely linked with verb meaning, including selectional restrictions on verbs' arguments and detailed information about how different participants in events may act (as agents) and be acted upon (as patients). For adults and children with TD, verb meaning and associated event knowledge support prediction of upcoming sentence material (Altmann & Kamide, 1999; Borovsky et al., 2012; Fernald et al., 2008; for a recent review on prediction in language comprehension, see Ryskin & Nieuwland, 2023). During incremental sentence interpretation, this verb-based EP knowledge is integrated with multiple levels of linguistic structure, including individual words (Hare et al., 2009), sentence-internal semantics (Altmann & Kamide, 1999; Fernald et al., 2008; Hagoort et al., 2004), and discourse (Delogu et al., 2018; Metusalem et al., 2012). Morphosyntax also plays a key role in these processes; for example, individuals make more accurate and quicker use of EP information in simple active sentences compared to more complex sentences with noncanonical word orders (Ferreira, 2003). Incremental integration of background EP knowledge with linguistic information produces detailed situation models—mental models reflecting comprehenders' understanding of participants in events and their relations—that are important for robust language comprehension (Johnson-Laird, 1995; Van Dijk & Kintsch, 1983; Zwaan & Radvansky, 1998).
Intriguingly, while individuals with DLD can leverage EP information, they show key differences in other skills related to verb knowledge and incremental integration of linguistic information compared to those with TD. For example, individuals with DLD have less detailed knowledge of verbs (Alt et al., 2004), slower verb learning (Kan & Windsor, 2010), smaller verb vocabularies (Conti-Ramsden & Jones, 1997; Kelly, 1997; Rice & Bode, 1993; Watkins et al., 1993), and differences in using verb and thematic role knowledge in online (i.e., in-the-moment) incremental sentence processing (Andreu et al., 2013; Borovsky et al., 2013; Nation et al., 2003). Children with DLD also have difficulties integrating word-level lexical–semantic information into sentence representations (Pijnacker et al., 2017; Pizzioli & Schelstraete, 2013) and integrating prior sentence material with later material across long distances (Purdy et al., 2014). Yet, despite weaknesses in the processes underlying the use of EP information, children with DLD still show advantages for interpreting probable over improbable sentences.
Children with DLD may use EP information differently than children with TD, perhaps preferring alternative sources of this information or deploying it in different ways. In this study, we ask how children with DLD derive an advantage from EP information in language comprehension despite poor verb knowledge and poor online integration of semantic and morphosyntactic information within sentences. We review three possibilities below.
The first possibility is that children with DLD use EP information primarily in offline (i.e., after completion of the sentence) comprehension rather than by integrating it with sentence-internal information during online processing (post sentence completion only hypothesis). Most studies showing that children with DLD use EP information in sentence interpretation have used offline comprehension pointing or enactment tasks (Gowie & Powers, 1972, 1978; Hutson, 1975; Hutson & Powers, 1974; Powers, 1973; Precious & Conti-Ramsden, 1988; van der Lely & Dewart, 1986). In such tasks, children with DLD may show superficial success in interpreting a sentence like, “The squirrel was chased by the dog,” by simply attending to the content words within the sentence, ignoring word order and morphosyntax, and selecting the most likely interpretation from an array of choices, or by enacting a probable event given an array of toys. Thus, children with DLD may show little advantage from EP information in online comprehension but superficial increases in offline comprehension accuracy.
A second possibility is that when children with DLD interpret sentences that describe probable events, their online processing looks similar to that of children with TD (similar processing for probable hypothesis). Some recent studies examining sentence processing in children with DLD have used uniformly improbable sentences for exactly this reason. For example, a line of studies by Montgomery and colleagues (Montgomery et al., 2017, 2018, 2019) examined how the syntactic and cognitive abilities of children with DLD influenced their sentence processing. In order to neutralize the potentially beneficial effect of semantics, sentences described improbable events (e.g., “The tree touches the mouse”). In the absence of semantic cues, children with DLD were found to have less robust sentence processing. In contrast, for sentences describing probable events, children with DLD may be better able to direct their more limited cognitive resources to process otherwise challenging language material. Such an advantage would be seen both in faster online processing and in more accurate offline comprehension of sentences matching children's EP biases.
A third, intermediate possibility lies between the other two—children with DLD may show advantages in processing for some sentences and not for others (similar processing depending on demands hypothesis). For example, advantages might primarily lie in offline comprehension if EP information is based on verbs about which a child has shallow knowledge or if integrating sentence-internal material with EP biases requires dealing with difficult language (e.g., long-distance dependencies, complex syntax, or verb morphology). If verb knowledge is strong or if linguistic demands are low, however, children with DLD may use EP information much like their peers with TD.
To distinguish among these possibilities, we focus on two core questions about the online incremental processing of passive sentences in English by children with DLD: First, how is EP information integrated with other sentence material? Second, how does verb knowledge shape EP use?
Passive sentences have an ideal structure to answer these questions. In passive sentences, the order of participants conflicts with the usual agent–action–patient word-order biases of English (compare “The dog chased the squirrel” to “The squirrel was chased by the dog”), indicating that verb knowledge (and associated EP information) can contribute to predictions of the sentence-final agent. These sentences are also developmentally Though children with TD eventually reach adult levels of proficiency in understanding these sentences by age 11–12 years, children with DLD may never do so (Dick et al., 2004).
There are two sources of EP information relevant to passives. One, animacy, is already well understood. Animacy of the patient differentiates the two major types of passive nonreversible and reversible passives. Nonreversible passive sentences have an inanimate patient (e.g., “ball” in “The ball was chased by the dog”), and reversible passive sentences have an animate patient (e.g., “cat” in “The cat was chased by the dog”). Reversible passives are so-called because they can plausibly support a reversed interpretation where the patient is misinterpreted as the agent (e.g., “The cat chased the dog”), whereas for nonreversible passives, such an interpretation is often implausible (e.g., “The ball chased the dog”). Children with TD and DLD often interpret passive sentences as if they were active sentences, a bias that is even stronger for children with DLD than for children with TD (Evans, 2002; Friedmann & Novogrodsky, 2007; Montgomery & Evans, 2009; Montgomery et al., 2017; van der Lely, 1996). This bias arises because children with TD and with DLD use a first-animate-noun-as-agent interpretation strategy where agenthood is assigned to the first animate noun in a sentence, and children with DLD apply this strategy more inflexibly than children with TD (Bishop et al., 2000; Evans, 2002; Evans & MacWhinney, 1999; van der Lely, 1996; van der Lely & Harris, 1990).
While agency is an important source of EP information for passives, in this study, we focus on verb-based EP information, which comes from verb meaning and its match with potential agents' meanings and is less well understood though potentially impactful across a wide variety of sentence types. Comprehenders have knowledge not only about the types of patients that verbs tend to select (e.g., “chase” selecting “squirrel” or other animals) but also about the types of agents that tend to participate in different actions (e.g., knowing that dogs, not mice, tend to chase squirrels; Matsuki et al., 2011; McRae et al., 1997). In passive sentences, the verb precedes the agent. (In a sentence like, “The squirrel was chased by the dog,” the verb, “chased,” appears before the agent, “dog.”) Therefore, in incremental interpretation of passive sentences, prediction of an upcoming agent is entailed by verb-based EP information to the extent that the agent's prior participation in events is consistent with the verb's meaning. Examining this verb-based EP information in passive sentences offers an opportunity to investigate how children with DLD integrate EP information with other sentence-internal information and to learn how verb vocabulary knowledge supports this process.
Passive sentences are ideal for assessing how children with DLD integrate verb-based EP information with morphosyntactic cues. Importantly, these morphosyntactic cues in passives vary in their distance to the verb, including the verb-proximal past participle (e.g., “chased”) and the verb-distant by phrase. These morphosyntactic cues can reinforce verb-based EP information or compete with it depending on the sentence's ultimate match with the verb-based EP information and the material already heard in the sentence. For sentences that match verb-based EP biases, after hearing the past participle –ed morpheme (e.g., “The squirrel was chased . . . [not chasing]”), it should be relatively easy to predict the upcoming agent (e.g., “dog”) and even more so after hearing the by phrase, which would provide more evidence that the sentence is in passive and not active voice. Yet, this process may be difficult for children with DLD, who have reduced response to passive-sentence morphology (Leonard et al., 2006) and, in particular, difficulty integrating information across long distances (Purdy et al., 2014), which may include links between the verb's meaning and the by phrase. Therefore, we may find that children with DLD fail to show increased use of verb-based EP information when the sentence-internal morphosyntactic information supports its use, particularly across long distances.
Passive sentences are also ideal for testing the influence of verb knowledge on EP use, again because the meaning of a verb constrains predictions about the upcoming agent. Children with DLD, by virtue of their poorer verb knowledge (Alt et al., 2004; Conti-Ramsden & Jones, 1997; Kan & Windsor, 2010; Kelly, 1997; Rice & Bode, 1993; Watkins et al., 1993), may have delayed and less accurate prediction of agents based on EP information after hearing the verb in passive sentences. On the other hand, if children with DLD have strong knowledge of a verb's meaning, they may be able to devote more resources to overcoming difficulties with passive sentence morphology or integration across long distances.
Each hypothesis detailed above predicts unique patterns of performance in comprehension passive sentences. Hypothesis 1 (“post sentence completion only”) suggests that children with DLD will integrate EP biases with sentence-internal verb and noun semantics to produce correct interpretations but will do so more slowly than children with TD, resulting in different online processing but equivalent success in comprehension for probable sentences. Hypothesis 2 (“similar processing for probable”) suggests that integration of EP biases with sentence-internal information will be eased in online processing for probable sentences, resulting in similar processing for probable sentences across groups and equivalent success in comprehension; in contrast, groups will still differ for improbable sentences. Hypothesis 3 (“similar processing depending on demands”) suggests that syntactic and semantic information may interact in online processing. For parts of the sentence with relatively limited linguistic demands (as, e.g., in integrating morphosyntax and semantics close to the verb), children with DLD may look like their peers with TD during online processing for probable sentences. For parts of the sentence with more complexity, divergences between groups may be observed.
In Experiment 1, we ask how children with TD and DLD use verb-based EP information in conjunction with sentence-internal information like morphosyntax in online passive sentence interpretation using eye tracking. We recruited 4- to 5-year-old children with DLD and same-age peers with TD to serve as participants. We created novel EP biases based on the overlap between attributes of a verb and a potential agent by showing scenes setting up strong associations between agents and particular kinds of actions. The children then heard nonreversible and reversible passive sentences that matched or did not match these EP biases. For example, one agent was shown doing helpful actions while another was clumsy. Based on these attributes, when presented with a sentence like, “The car was bumped accidentally by the . . . ,” children are likely to predict the clumsy over the helpful agent. We used novel EP biases rather than the natural affordances of certain agents (e.g., “judge–convict,” “waiter–serve”) for two reasons. First, the available data about agent–verb relationships are limited, primarily consisting of adult ratings for highly specific and emotionally valent agents and verbs that are unlikely to be known by children (e.g., “terrorize,” “evaluate,” “freshman”; McRae et al., 1997). Second, by directly associating agents with sets of verbs, we reduced the likelihood that agents would bring additional associations that could confound our results.
We recorded children's eye gaze to images representing potential sentence interpretations as they heard the sentences and asked children to point to their final interpretation as a measure of offline comprehension. We explored how children with DLD might demonstrate slowdowns and difficulties in responding to and integrating verb-based EP information with other sentential cues relative to children with TD (see Table 1 for a summary of the predictions). We focus on two sentential windows in the main text—one window from the onset of the past participle –ed ending to the onset of the by phrase (e.g., “The car was bumped accidentally by the man”) and another window from the onset of the by phrase to the onset of the agent (e.g., “The car was bumped accidentally by the man”)—to test how the distance of morphological cues from the verb influences integration with verb-based EP information.
In Experiment 2, we consider how children's prior verb vocabulary knowledge impacts these processes given evidence for verb knowledge deficits in children with DLD. We engaged the children in a word association task to measure the quality of their vocabulary knowledge of the specific verbs used in the sentences in Experiment 1 and examined how knowledge quality affected children's use of verb-based EP information on a sentence-by-sentence basis. We expected that the quality of children's sentence-specific vocabulary knowledge would lead to a greater processing advantage from EP information.
This study was conducted in accordance with the policies of the institutional review board (IRB) at Purdue University (IRB Protocol 1904021980). Participants' caregivers provided written consent, and participants provided verbal assent before each session.
Participants were 20 children with DLD (nine girls, 11 boys) and 20 children with TD (nine girls, 11 boys; see Table 2). The groups were matched on age, MDLD = 57.7 months, SDDLD = 6.63; MTD = 57.5 months, SDTD = 5.66; t(37.08) = 0.10, p = .919, d = 0.03. Children this age in both groups are not likely to demonstrate floor or ceiling performance on the comprehension of reversible or nonreversible passive sentences (van der Lely & Harris, 1990). The study was conducted in American English. Participants were users of Midwestern American English typical of Indiana (N = 40). Participants were White (n = 38), Asian (n = 1), and of multiple races (n = 1), with three also being Hispanic.
The children with DLD were required to score at or below 87 on the Structured Photographic Expressive Language Test–Primary, Second Edition (SPELT-P2; Dawson et al., 2005), the empirically derived cutoff score for high sensitivity and specificity for this test (Greenslade et al., 2009). Children with TD were required to score above 87 on the SPELT-P2 and have no parental concerns about language development. The children with DLD also demonstrated minimal or no symptoms of autism spectrum disorder on the Childhood Autism Rating Scale–Second Edition (Schopler & Van Bourgondien, 2010). The average number of years of maternal education was 16.56 (SD = 2.12) for the group with TD and 16.41 (SD = 2.50) for the group with DLD.
Children were required to pass a pure-tone hearing screening; have no history of neurological disorders; and score within the average range on the Kaufman Assessment Battery for Children–Second Edition (Kaufman & Kaufman, 2018), a measure of nonverbal intelligence. Children were additionally administered the Peabody Picture Vocabulary Test–Fourth Edition (PPVT-4; Dunn & Dunn, 2007) as a supplemental measure of receptive vocabulary knowledge. Three additional children with DLD and one additional child with TD attempted but opted not to complete the eye-tracking task.
Each block of the eye-tracking study consisted of two an agent attribute assignment part and a sentence prediction part (see Figure 1 for a visualization of the study's design).
Figure 1. Schematic of the overall design for the eye-tracking task. The examples correspond to a “clumsy” agent. In the agent attribute learning part, children would hear sentences about and see the man doing clumsy activities. In the agent test part, children would be asked what they learned about the man. In the sentence prediction part, each child heard reversible and nonreversible passives that demonstrated agent attribute–verb match and mismatch. T
agentTactionrefers to images with the correct target agent doing the correct target action. NTagentTactionrefers to images with the incorrect agent doing the correct target action. TagentNTactionrefers to images with the correct agent doing the incorrect action. NTagentNTactionrefers to images with the incorrect agent doing the incorrect action.
During the agent attribute learning part, we set up associations between agents and clusters of different verbs to establish EP biases for the sentence prediction part (see Table 3). In each block, we showed two agents doing five different actions; one agent's actions came from one semantically coherent category (e.g., clumsy slipping, tipping, tripping), while the other agent's actions came from an unrelated, highly contrastive semantically coherent category (e.g., affectionate cuddling, tickling, petting). The agents in each pair were highly distinguishable, and the assignment of agent to verb group was counterbalanced across children.
For each block of the study, we identified seven semantically coherent verbs in each of the two verb groups (see Table 3). Verbs were included if they were able to appear in transitive sentences. All of the verbs chosen were in the bottom 25th percentile of adult age-of-acquisition ratings in the Kuperman et al. (2012) data set, indicating that these verbs were judged to be relatively early learned. We used WordNet, VerbNet, and English Verb Classes and Alternations to identify verbs from similar semantic classes that were potentially reversible, allowing for human agents and patients (Levin, 1993; Princeton University, 2010; Schuler, 2005).
We verified that these verbs patterned together semantically by using a large corpus of English language material to create a vector representation of each verb's meaning using Word2Vec (Mikolov et al., 2013). We measured semantic distance between the vector representations of each pair of verb groups to ensure that the meanings of these verbs clustered into two separate groups (Kogan et al., 2006). This procedure demonstrated that, within a block, verbs in each of the semantically coherent groups patterned together and that the verbs did not pattern with the other group (see Supplemental Material S1 for details).
We created 10-s videos of each agent–verb combination using Vyond, an online animation software (Vyond, 2019). The action in each video was described using active past tense sentences (e.g., “The man knocked over a chair. He knocked over a chair”). The sentences were spoken by an American English speaker at a child-directed pace and were normalized for amplitude using Praat (Boersma & Weenink, 2019).
The children first saw a still image of an agent and heard a sentence taking the form, “Look at this [man/woman]. [He/she] is so [attribute]. Let's find out what [he/she] did today.” The children then saw the five action videos associated with an agent played two times in sequence. The children then saw the still image of the agent and heard a sentence taking the form, “Wow! That [man/woman] is really [attribute].” An example video can be found on the Open Science Framework (https://osf.io/b89q5/; see Clumsy_Man.mp4).
The children saw the stimuli associated with the two agents two times in an interleaved fashion. The stimuli for each agent were also seen in the middle of the sentence prediction part of the task to ensure that the EP biases had not weakened over time. This presentation was abbreviated; action videos were each played once rather than twice. The first agent exposure lasted approximately 4 min, and the second exposure lasted 2 min.
We measured learning of the association between the agents and their attribute labels using two agent test trials after the initial agent attribute learning part. Children pointed to the agent associated with each attribute (e.g., “Which one is clumsy? Point to the one who is clumsy”). Children in both groups demonstrated high accuracy, MDLD = 92.9%, MTD = 98.7%. Data were excluded for a child's blocks if either of these trials was incorrectly answered, resulting in the exclusion of 10 of 160 blocks across 10 participants (nDLD = 9, nTD = 1).
The sentence prediction part followed the agent attribute learning part. There were two types of sentence prediction tests and fillers. Half of the test sentences were nonreversible passives (e.g., “The car was bumped accidentally by the man”), and half were reversible passives (e.g., “The woman was bumped accidentally by the man”). In addition, half of the sentences demonstrated a match (and half, a mismatch) between the verb and its associated agent. For example, if the man participated in clumsy actions in the agent attribute learning part, half of the passive sentences demonstrated an agent match (e.g., “The car was bumped accidentally by the man”) and half demonstrated an agent mismatch (e.g., “The car was bumped accidentally by the woman”). Two semantically related verbs able to appear felicitously with inanimate and animate patients were chosen for each verb group for the test trials (see Table 3). We included an intercept for item (including the test verb) in the random effects structure for the statistical models, which served to statistically control for the influence of variation across test verbs.
The passive sentences followed the pattern, “The [patient] was [verb stem] + [past participle ending] [adverbial material] by the [man/woman]” (e.g., “The car was bumped accidentally by the man”). We embedded neutral adverbial material (e.g., “accidentally”) between the past participle and the by phrase to increase the time available for measuring anticipatory looks in response to the past participle. These sentences were audio-recorded using the same speaker of American English as in the agent attribute learning part. All of the individual parts of the sentences that were potentially important for their interpretation were standardized in duration in Praat. These durations were as “the”: 173.00 ms, “[patient] + was”: 740.30 ms, “[verb stem]”: 325.60 ms, “[part participle ending] + [adverbial material]”: 1,084.57 ms, “by the”: 308.32 ms, and “[agent]”: 568.20 ms. The entire sentence lasted 3,200 ms. We refer to these time windows as the “determiner,” “patient + auxiliary,” “verb stem,” “past participle,” “by phrase,” and “agent” windows, respectively. The parts of the sentences were standardized in length not through the addition of extra pauses but by slightly elongating (or contracting) their duration, allowing for control over the points at which informative information was provided to the listener. The first author ensured that the stimuli sounded natural after the duration manipulation. Finally, the sentences were normalized in amplitude.
In order to measure children's online interpretation of these sentences, we used the visual world eye-tracking paradigm (e.g., Huettig et al., 2011). We presented four images on-screen during each test trial (see Figures 1 and 2). One image was the Target Agent, Target Action (TagentTaction) image, showing the correct interpretation of the sentence with the target agent doing the target action. Another image was the Nontarget Agent, Target Action (NTagentTaction) image, which showed the nontarget agent doing the target action. The Target Agent, Nontarget Action (TagentNTaction) image showed the target agent participating in a nontarget action. Finally, the Nontarget Agent, Nontarget Action (NTagentNTaction) image showed the nontarget agent participating in the nontarget action. Nontarget actions were not in the same semantic category as either of the agents' semantically coherent verb categories. Nontarget actions appeared with equal frequency with each of the test verbs. Examples of visual stimuli corresponding to the “clumsy” agent appear in Figure 2.
Figure 2. Examples of sets of T
agentTaction, NTagentTaction, TagentNTaction, and NTagentNTaction(from top-left clockwise within each array) images for the “clumsy” agent across passive sentence types and match conditions.
Filler trials consisted of active sentences in the past progressive tense that described events seen in the agent attribute learning part (e.g., “The woman was petting the dog”). These trials served two purposes. First, they ensured that children were unable to predict whether a sentence was in active or passive voice until hearing the disambiguating morphological content on the verb (i.e., −ing or the past participle –ed ending). Second, they reinforced the EP biases set up in the agent attribute learning part by showing the agents participating in these same actions. These active sentences were audio-recorded, their parts standardized in duration, and their amplitude normalized as with the passive sentences. The four images appearing on-screen were evenly divided between actions from the two agents' agent attribute learning parts.
In each block of the study, there were 16 passive sentences (four reversible matching sentences, four reversible mismatching sentences, four nonreversible matching sentences, and four nonreversible mismatching sentences). Each test verb was used four times, once in each of the reversible/nonreversible and matching/mismatching sentences. There were 12 active distractor sentences (six for each agent), pseudo-randomly distributed among the passive sentences so that no more than two consecutive trials came from the same verb category, no more than three consecutive trials had the same target agent, and no verb appeared in consecutive trials. There were eight counterbalance orders controlling the following assignment of agent to attribute, first agent exposed within a block, trial order (reversed in half the counterbalances), and whether blocks were presented on the first or second day of testing. Across participants, different passive sentence types were seen first equally often across the counterbalance orders. Each block of the sentence prediction part took approximately 5 min to complete.
Children's fixations to each of the four images in the 64 passive trials were recorded at 500 Hz using an EyeLink 1000 Plus eye-tracking system, and a 24-in. screen was positioned 580–620 mm away from the children. The system was calibrated using a five-point routine, and a central fixation image was presented before every test trial to calibrate the machine in case of drift or movement. We used the default fixation thresholds for the EyeLink machine, which filters out fixations shorter than 2 ms that could be due to noise, error in the eye-tracking routines, or unsystematic looking patterns. An experimenter sat next to the child and provided noncontingent praise. Trials were excluded from analysis if there was more than 80% track loss (e.g., eye blinks or head turns away from screen) across the sentence. This resulted in the exclusion of 3% of trials (81 out of 2,400 completed trials). Children's pointing responses were recorded using a button box operated by the experimenter. Two blocks of the study were presented in each session and were separated by a short 5-min break. The two sessions of the study were separated by approximately a week. Participant assessment testing and other tasks for Experiment 2 were completed before the eye-tracking study. Each session lasted approximately 40 min.
Bayesian mixed-effects models were used to analyze the statistical patterns in our data. For readers unfamiliar with Bayesian statistics, we highlight three differences from frequentist statistics that bear on the interpretation of the model output (Kruschke & Liddell, 2018). First, the meaning of a confidence interval is different. For a Bayesian model, there is a 95% probability that the true value for a parameter lies in the 95% credible interval (CI). Second, p values, when used, take on a different meaning. We use pd values (“p direction”), which are the proportion of posterior samples below or above zero for point estimates above or below zero, respectively. In our tables, we denote with asterisks pd values greater than .95 and mainly focus on their interpretation given the number of interactions we could potentially consider. However, we note that effects with pd values below .95 should not be considered “insignificant” or absent. Finally, when considering post hoc contrasts, we present medians, 95% CIs, and the proportion of samples within the region of practical equivalence (ROPE; Kruschke, 2018; Makowski, Ben-Shachar, Chen, & Lüdecke, 2019). The ROPE for a model corresponds to the a priori range of negligible effect sizes. If fewer than 2.5% of the samples are within the ROPE (denoted pROPE), there is strong evidence that an effect size is larger than negligible. A low pd value can co-occur with a low pROPE if, for example, the posterior samples' distribution is widely distributed around 0; pd and pROPE should each be interpreted considering the other. In our analysis, we first use pd to focus on effects likely to be present, and, for effects that are likely, we then use pROPE to determine whether the effects are meaningfully large.
Factors were sum-coded to allow for the interpretation of lower order effects in the model output. We used R Version 4.2.1 (R Development Core Team, 2008), emmeans Version 1.8.2 (Lenth, 2016), tidybayes Version 3.0.2 (Kay, 2021), and bayestestR Version 0.13.0 (Makowski, Ben-Shachar, & Lüdecke, 2019) to perform post hoc testing and brms Version 2.18.1 (Bürkner, 2018) to fit the statistical models.
Offline comprehension accuracy, measured using children's postsentence image pointing, was analyzed using Bayesian mixed-effects logistic regression in separate models for each choice. We included random intercepts for participant and item and random slopes for sentence type, group, and match condition. Independent variables were group (DLD vs. TD), passive sentence type (nonreversible vs. reversible), condition (verb-based EP match vs. mismatch), and these variables' interactions. There were 2,319 observations for these models. As priors on the fixed effects, we used a weakly informative normal distribution centered on 0 with an SD of 0.5. The Bayesian model was fit using a Hamiltonian Monte Carlo procedure, which adaptively samples the posterior distribution of the model coefficients across a number of chains (McElreath, 2020). To do this, the procedure tunes the simulation in a “warmup” phase and then produces posterior samples in a “postwarmup” phase. We collected 2,000 warmup samples and 4,500 postwarmup samples across four chains, resulting in 10,000 postwarmup samples for the model. We chose to have 10,000 postwarmup samples so that we would have adequate precision for 95% CIs. We defined the ROPE for post hoc contrasts based on the range recommended in Kruschke (2018) for a negligible effect size in logistic regression models. This ROPE range for the offline comprehension accuracy model is −0.055 to 0.055 on the log odds scale, or, for example, the difference between 50% and 52.75% accuracy.
For the eye-tracking data, we examined how the children's looks to the TagentTaction image varied as the sentence was heard. Further, we included only sentences that were correctly interpreted (as judged by pointing accuracy). Importantly, there was evidence that the children's correct responses were not correct only by chance but were instead reflective of true interpretive All of the children had a proportion of correct responses greater than .25 across the passive trials, the chance level for the pointing task (given that there were four options).
The 500-Hz raw samples were binned into intervals of 50 ms, producing a proportion of TagentTaction looking every 50 ms during the sentence, thereby reducing the effects of noise and unsystematic looking. This binning procedure has been used in other work (e.g., Borovsky et al., 2013). We used linear splines to analyze the binned data (Cho et al., 2022; Yamashiro et al., 2019). In this approach, connected line segments (i.e., splines) are used to model how TagentTaction looking changes, potentially abruptly, in response to events in the sentence. Events are modeled as “knots” in a line segment where changes of slope may occur. As potentially informative information occurred at the beginning of each time window, knots were positioned at the beginning of the determiner, patient + auxiliary, verb stem, past participle, by phrase, and agent time windows. Segments for each time window were created following Yamashiro et al. (2019). These segments were entered into the model as separate terms. Interactions of the factors of interest with these line segments model how the slope of TagentTaction looking differs with respect to these factors within each time window.
We used a Bayesian linear mixed-effects model with an autoregressive error term and crossed random effects. Logit-adjusted proportion of TagentTaction looking within the 50-ms bins was the dependent variable. Group, passive sentence type, match condition, and their interactions were independent variables along with interactions with the spline time terms for each window. The time terms were scaled so that a 1-unit change corresponded to 100 ms. Random intercepts for participant and item were included. Match condition and sentence type were included as random slopes on the participant random intercept, and group was included as a random slope on the item random intercept. As looking across adjacent time points within a trial is likely to be correlated, an AR(1) autoregressive covariance structure of the residuals was used as a function of time within individual trials. The estimate for autocorrelation ρ was .95, 95% CI [0.95, 0.95]. We used a weakly informative normal distribution with a mean of 0 and an SD of 0.5 as a prior for the betas. All other priors were the uninformative defaults for brms. There were 2,000 warmup samples and 10,000 postwarmup samples across eight chains. For this model, there were 97,990 observations distributed across 40 children and 64 items. The ROPE used for this model was −0.0055 to 0.0055 as the spline terms were scaled such that 1 unit represented 100 ms. This range, in logit units, is equivalent to an increase from 50% TagentTaction looking to 52.75% TagentTaction looking (a change of 0.055 on the logit scale) over the course of 1 s.
Figure 3 shows the children's offline comprehension choices presented across groups, sentence types, and match conditions. In the main text, we focus on the logistic mixed-effects model for offline comprehension pointing to the TagentTaction choice. The main findings from the models for the NTagentTaction, TagentNTaction, and NTagentNTaction choices are summarized below, but the full analyses are presented in Supplemental Material S1 for brevity.
Figure 3. Pointing accuracy across groups, sentence types, and match conditions. T
agentTactionrefers to images with the correct target agent doing the correct target action. NTagentTactionrefers to images with the incorrect agent doing the correct target action. TagentNTactionrefers to images with the correct agent doing the incorrect action. NTagentNTactionrefers to images with the incorrect agent doing the incorrect action. Error bars are standard errors. DLD = developmental language disorder; TD = typical development.
The mixed-effects logistic regression model for the TagentTaction (correct) choice showed strong effects of passive sentence type and group (see Table 4). Specifically, across groups and match conditions, the proportion of trials where children pointed to the correct image in nonreversible passives, estimated marginal median (EMM) = .80, 95% CI [0.69, 0.88], was greater than the proportion for reversible passives, EMM = .62, 95% CI [0.50, 0.72], odds ratio (OR) = 2.86, pROPE = 0, pd = 1. In addition, across passive sentence types and match conditions, the group with TD, EMM = .86, 95% CI [0.76, 0.92], exhibited a greater proportion of correct trials than the group with DLD, EMM = .56, 95% CI [0.41, 0.70], OR = 5.16, pROPE = 0, pd = 1.
The effect of match condition was weaker than passivesentence type and group. The difference in accuracy between the match and mismatch trials, EMMMatch-Mismatch = .02, was not clearly outside the ROPE, pROPE = .19, pd = .87, suggesting that match condition had limited effect on correct comprehension accuracy. The interactions between the predictors were also not reliably different from zero, pd < .79, suggesting that these interactions had little effect on the likelihood of a correct pointing response.
In summary, the children with TD were more likely to select the correct image than the children with DLD. Across groups, nonreversible passives were more likely to be correctly interpreted than reversible passives.
The full results for models for the other choices are presented in Supplemental Material S1. Across these models, we found that when the children made errors of interpretation, they were more likely to choose the NTagentTaction and NTagentNTaction images for reversible passives and to choose the TagentNTaction image for nonreversible passives. This pattern fits with children's expected use of the first-animate-noun-as-agent strategy for reversible passives. This strategy appeared to be somewhat weaker for children with DLD, however, given that they chose the TagentNTaction image for reversible passives and the NTagentTaction image for nonreversible passives at increased rates compared to children with TD. There was also evidence that children with DLD were more likely to choose the TagentNTaction image for passives in the match condition than in the mismatch condition. While children with DLD correctly identified the target agent, they seemed to fail to remember the correct verb, suggesting potential trade-offs in the use of EP information and memory. Finally, we note that the absolute differences in proportions for the TagentNTaction and NTagentNTaction models were generally small, suggesting that differences among groups and conditions for these choices may reflect subtle processing factors.
Proportions of TagentTaction looking across groups, sentence types, and match conditions are presented in Figure 4. These data represent only the trials that were correctly interpreted during the offline comprehension pointing task. The following sections present the results for the mixed-effects spline model; each table shows the model coefficients, credible intervals, and pd values for the terms that interact with the past participle (see Table 5) and by phrase (see Table 8) splines. Results for the nonspline main effects, determiner, patient + auxiliary, verb stem, and agent windows are provided in Supplemental Material S1 for brevity. Although these tables are split by window, the results come from the single model that incorporates all of these windows together.
Figure 4. T
agentTaction(correct) looking across groups by match and sentence types for sentences that were ultimately correctly interpreted in the offline comprehension pointing task. Error bars are standard errors. Vertical lines represent the divisions between the determiner, patient + auxiliary, verb stem, past participle, by phrase, and agent windows. Time is in milliseconds post sentence onset. Example sentences correspond to the case where the man is clumsy and the woman is affectionate. DLD = developmental language disorder; TD = typical development.
The past participle spline and its interactions with the other predictors are displayed in Table 5. The main effect of the past participle spline was strong; across groups and conditions, the proportion of TagentTaction looking increased from EMM = .15, 95% CI [0.13, 0.18], to EMM = .27, 95% CI [0.24, 0.31], a reliably positive and nonnegligible growth rate, pd = 1, pROPE = 0, during the 1,085-ms time window. The past participle spline had strong interactions with sentence type, Sentence Type × Group, and Sentence Type × Match Condition × Group.
The interaction with sentence type was strong, pd = 1, pROPE = .002. At the beginning of the past participle window, reversible passives had higher proportions of TagentTaction looking, EMM = .17, 95% CI [0.14, 0.21], than nonreversible passives, EMM = .14, 95% CI [0.11, 0.17]. This initial difference was compounded across the window, with higher growth in TagentTaction looking for reversible passives, b = 1.05, 95% CI [0.76, 1.32], pd = 1, pROPE = 0, than for nonreversible passives, b = 0.37, 95% CI [0.12, 0.61], pd = 1, pROPE = .005. (Note that a 1-unit change in b, on the adjusted logit scale, reflects the difference between, e.g., 50% target looking and 73.1% target looking over the course of 100 ms.) This difference likely reflects that for reversible passives, the past participle verb ending (which is at the beginning of the past participle window) is highly informative when combined with the sentence-initial noun in the patient + auxiliary window and the verb in the verb stem window. In contrast, for nonreversible passives, the sentence is still ambiguous because, in all of the images, the sentence-initial noun is shown as the patient of the action.
Sentence type and group also interacted together with the past participle spline. Figure 5 shows proportions of TagentTaction at the beginning and end of the past participle window, and Table 6 reports group differences in growth across sentence types. At the beginning of the past participle window, the groups likely differed in the proportion of TagentTaction looking for reversible passives, pd = .93, pROPE = .004, but not for nonreversible passives, pd = .65, pROPE = .015. Over the course of the past participle window, the group with DLD demonstrated slightly higher rates of growth in TagentTaction looking for reversible passives than the group with TD, though this difference was not reliable. In contrast, the group with TD demonstrated reliably higher rates of growth in TagentTaction looking for nonreversible passives than the group with DLD. In general, this interaction suggests that for nonreversible passives, the group with TD was responding to the last informative cue (i.e., the verb) more quickly than the group with DLD. For reversible passives, there was some weak evidence for catch-up growth in TagentTaction looking by the group with DLD.
Figure 5. Proportion of T
agentTactionlooking at the start and end of the past participle window for the Sentence Type × Group interaction. Error bars are 95% highest density credible intervals. DLD = developmental language disorder; TD = typical development.
Last, there was evidence for an interaction between sentence type, match condition, and group in the past participle window, pd = .99, pROPE = .034. To contextualize this interaction, the groups' proportions of TagentTaction looking at the beginning and end of the past participle window are presented in the time course plot in Figure 4, and differences in growth across the window are shown in Table 7. At the beginning of the window, there was strong evidence of differences between groups only for the reversible sentences in the match condition; there, children with TD had higher TagentTaction looking than children with DLD, pd = .98, pROPE = .001. The two groups demonstrated different rates of growth in TagentTaction looking across the past participle window for sentences in the match condition but more similar rates for sentences in the mismatch condition. The group with TD had a numerically higher rate of growth in TagentTaction looking than the group with DLD for nonreversible sentences in the mismatch condition and for reversible sentences in the mismatch condition, but these differences were not large. The group differences for the sentences in the match condition were larger and more reliable. The group with TD had higher growth in TagentTaction looking than the group with DLD for nonreversible sentences in the match condition. In contrast, the group with TD had lower growth in TagentTaction looking than the group with DLD for nonreversible sentences in the match condition.
In summary, this interaction suggests that children in the two groups responded differently to EP information. For sentences in the mismatch condition, group differences were small but still favored the group with TD. This may reflect greater exploration of the images by the group with TD compared to the group with DLD, even those not consistent with the EP information. For sentences in the match condition, group differences were robust. For nonreversible sentences in the match condition, children in both groups entered the window with similar rates of TagentTaction looking. The group with TD then demonstrated integration of EP biases with the verb and past participle to quickly increase TagentTaction looking, and the group with DLD did not, showing near-zero rate of growth across the window. For reversible sentences in the match condition, the children with TD had higher rates of TagentTaction looking than children with DLD at the beginning of the window. This initial difference for this sentence type going into the past participle window potentially reflects increased prediction based on EP information by the group with TD before linguistic information conditions it in earlier windows, an interpretation supported by differences in growth in the patient + auxiliary time window (see Supplemental Material S1). During these sentences, however, the children with DLD demonstrated higher growth in TagentTaction looking than the children with TD, clearly catching up with their peers in TagentTaction looking by the end of the past participle window, demonstrating success in integrating EP biases with verb-proximal morphosyntactic information.
The by phrase spline and its interactions with the other predictors are displayed in Table 8. The main effect of the by phrase spline was positive but not unambiguous; across groups and conditions, there was weak evidence that the proportion of TagentTaction looking increased from EMM = .27, 95% CI [0.24, 0.31], to EMM = .29, 95% CI [0.26, 0.33], pd = .94, pROPE = .001, during the 308-ms time window. However, the by phrase spline had strong interactions with match condition and Match Condition × Group.
The interaction between match condition and the by phrase spline indicated that growth in TagentTaction looking was higher for sentences in the mismatch condition, b = 0.07, 95% CI [0.01, 0.13], pd = .99, pROPE = .011, than for sentences in the match condition, b = 0.00, 95% CI [−0.06, 0.06], pd = .54, pROPE = .147.
Additionally, there was a strong interaction between match and group in the window, pd = 1, pROPE = 0. Figure 6 displays group differences in TagentTaction looking at the beginning and end of the by phrase window across match conditions. At the beginning of the window, the group with TD had higher proportions of TagentTaction looking for sentences in the mismatch condition than the group with DLD, pd = 1, pROPE = 0. In contrast, the differences at the beginning of the window between groups for the sentences in the match condition were not large or reliable, pd = .79, pROPE = .011. At the end of the window, these differences reversed within conditions, with the group with TD having greater TagentTaction looking for the match condition and with no reliable differences for the mismatch condition.
Figure 6. Proportion of T
agentTactionlooking at the start and end of the by phrase window for the Match Condition × Group interaction. Error bars are 95% highest density credible intervals. DLD = developmental language disorder; TD = typical development.
The overall interaction was best characterized as a difference between match conditions within group rather than group differences within match conditions. That is, over the course of the by phrase window, the group with TD had higher growth in TagentTaction looking for sentences in the match condition than for sentences in the mismatch condition, bMatch–Mismatch = 0.13, 95% CI [0.03, 0.23], pd = .99, pROPE = .005. In contrast, the group with DLD had higher growth in TagentTaction looking for sentences in the mismatch condition than for sentences in the match condition, bMatch–Mismatch = −27, 95% CI [−0.40, −0.14], pd = 1, pROPE = 0. The overall interaction was driven more by the group with DLD than by the group with TD; the magnitude of difference in growth in TagentTaction looking between match and mismatch sentences was larger for the group with DLD compared to the group with TD, b|TD[mismatch–match]|–|DLD[mismatch–match]| = −0.14, 95% CI [−0.30, 0.03], pd = .95, pROPE = .013.
In summary, this interaction suggests that for sentences in the mismatch condition, the group with DLD demonstrated catch-up growth, having started the window with less TagentTaction looking than the group with TD. For sentences in the match condition, the two groups started the window with similar levels of TagentTaction looking but quickly diverged—the children with DLD demonstrated rapid decreases in TagentTaction looking because they appeared to have difficulty integrating EP information with the verb-distant cue to passivity offered by the word “by” at the beginning of the window.
The goal of Experiment 1 was to determine how the use of verb-based EP information in offline and online passive sentence comprehension differed between children with DLD and TD. We put forward three alternative hypotheses about how children with DLD would use EP first, that children with DLD would primarily use EP information in offline comprehension and not online comprehension; second, that for sentences matching EP biases, children with DLD would show online processing similar to that of children with TD; and third, that children with DLD would show variable facilitation from EP information depending on the semantic and morphosyntactic content within the sentence.
The first hypothesis was not supported by the offline comprehension data as we did not find large effects of verb-based EP information on offline comprehension accuracy. The other major EP-based information was animacy, and here we replicated the robust advantage for nonreversible over nonreversible passive sentences observed in other studies. This difference from prior work for verb-based EP information (e.g., van der Lely & Dewart, 1986) is likely because our EP biases were created by assigning attributes to agents through clusters of similar verbs. Instead of relying on general world knowledge, the children learned EP biases during the experiment and used them a short time later. This potentially reduced the biases' effect because they may not have been as robust as the general world knowledge on which prior work was based. Nevertheless, there were differences between groups in rates of distractor choices across match conditions, suggesting that the short-term EP biases did affect processing, albeit more subtly than observed in pointing and enactment studies.
Instead, the most telling data came from the online processing eye-tracking task. These data did not support the second hypothesis—that the children with DLD would use EP information like their peers with TD in sentences in the match condition. In fact, the largest group differences were seen in sentences in the match condition. Consequently, there was much more support for the third hypothesis—that the children with DLD would use EP information in online processing to the extent that they could integrate it with sentence-internal information. Group differences in EP integration were most strongly observed in reversible passives in the match condition, where the groups showed differences in integrating EP information upon encountering the verb-proximal past participle and verb-distant by phrase windows. Children with TD demonstrated greater growth in TagentTaction looking through the verb stem window—likely as a result of quick application of EP biases given the images on-screen (even before the linguistic stimuli conditioned it)—and started the past participle window with greater overall TagentTaction looking. Upon hearing the past participle, however, the rate of TagentTaction looking for the children with DLD outpaced the rate for the children with TD, producing catch-up growth to the children with TD. For sentences in the mismatch condition, group differences in growth in TagentTaction looking were weaker but favored the group with TD. Overall, this finding indicates that children with DLD demonstrated success in integrating verb-based EP information with local morphosyntax to produce robust growth in interpretation.
In contrast, the group with DLD could not integrate EP information with the more distant morphosyntax in the by phrase window. Here, for passives in the match condition, the children with DLD demonstrated rapid decreases in TagentTaction looking growth, whereas for passives in the mismatch condition, the children with DLD had modest increases in TagentTaction looking growth, catching up to the children with TD. This pattern suggests that the children with DLD had difficulty integrating EP biases with the verb-distant by phrase morphosyntax relative to the children with TD. In response to verb-proximal morphosyntactic material, the children with DLD successfully integrated their EP biases to shape their online processing. Ultimately, however, when confronted with more distant confirmatory evidence in the by phrase, the children with DLD appeared to reevaluate their initial EP-based predictions rather than reinforce them, a pattern consistent with the hypothesis that children with DLD incorporate EP information best when the linguistic demands are low.
There are other potential explanations for the patterns observed in the by phrase window. One possibility is that children with DLD demonstrated decreased attention to the TagentTaction image over time, which would be consistent with their overall weaker sustained attention abilities (Ebert & Kohnert, 2011). However, this possibility seems improbable given that children in both groups demonstrated increases in TagentTaction looking across the long past participle window (where we might most expect difficulties with sustained attention to affect looking) but sharp decreases across the short by phrase window. Another possibility, suggested by a reviewer, is that the word “by” could have been interpreted by children with DLD as referencing a location (rather than as a signal of passivity). We believe that this possibility is also unlikely as there was not a similar drop in TagentTaction looking for the sentences in the mismatch condition, which would be expected if “by” were being misinterpreted by the group with DLD. Overall, because the looking patterns for the group with DLD differ so dramatically across match conditions, the difference between conditions (i.e., the semantic match between agent attributes and the sentences' verbs) must be responsible.
We also saw evidence in the offline comprehension data that the children with DLD had difficulty integrating EP and morphosyntactic information. The interaction between match condition and group for the TagentNTaction choice (i.e., an image showing an unmentioned action with the target agent) showed that the children with DLD were more likely to choose that option for passives in the match condition than in the mismatch condition, whereas the children with TD showed the opposite pattern. This was not a predicted difference. It is, however, consistent with the integration difficulties observed in the online processing data for the group with DLD, as we explain below.
In reversible passive sentence comprehension, both the first and second nouns compete with each other for agent and patient role assignments (Ferreira, 2003; Montgomery et al., 2017). Other work on active sentence interpretation indicates that while predicting the patient in a sentence like, “Toby arrests the . . . ,” adults look at both “police” and “crook,” indicating that both agent- and patient-associated information is activated in response to verbs (Kukona et al., 2011). Together, these studies suggest that prediction of the agent in passive sentence interpretation requires the inhibition of the patient (see Montgomery et al., 2017, p. 2614, for a similar argument applied to children with DLD). Adding EP biases to this task requires considering yet another piece of information related to the verb, and for children with DLD, these demands may be too much. When the children with DLD heard the verb in the passives in the match condition, they seemed to extract information about the potential agent involved (on the basis of EP information) but sometimes failed to retain the meaning of the specific verb. Thus, after hearing a sentence like, “The woman was bumped by the man,” these children identified “man” as the probable agent but had more difficulty retaining the specific verb “bump.” The higher activation of the noun “man” relative to the inhibited noun “woman” carried through to offline comprehension, but the specific verb was lost, resulting in higher rates of TagentNTaction choices (e.g., an image showing the man filming the woman). Given that children with DLD have weak representations of verbs, when challenged with retaining both the agent and the verb in working memory—another area of weakness in these children (Montgomery & Evans, 2009)—it is likely that if one of these elements were to be forgotten, it would be the verb.
Other patterns in the offline comprehension data suggested weaknesses based on vocabulary knowledge. For nonreversible passives, the children with DLD chose the NTagentTaction choice more often than the children with TD; in fact, the difference between groups was more pronounced for this passive type than for reversible passives. This pattern for nonreversible passives was also unexpected. For these passives, the NTagentTaction choice corresponded to a scene in which an unmentioned actor was the agent; for example, in the sentence, “The car was bumped by the man,” the NTagentTaction choice would correspond to a scene in which a woman was bumping a car. Even a simple vocabulary-based strategy involving “adding up” the words encountered over the course of a sentence without respect to their order should be highly effective for nonreversible passives; of the four interpretive options provided to children, only one would be consistent with all of the individual words in a nonreversible passive. However, children with DLD failed to consider these informative vocabulary cues to the same degree as children with TD. Additionally, compared with the children with TD, the children with DLD were more likely to choose the two distractor images. This kind of error could only be due to difficulty with either recognizing individual vocabulary words (e.g., the verb) or integrating across several vocabulary words (e.g., the verb and the agent).
The idea that deficits in vocabulary knowledge—and verb vocabulary knowledge in particular—contributed to the patterns we observed is consistent with an exploratory model comparison reported in Supplemental Material S1 (see “Exploratory Individual Differences Models”) that showed that vocabulary knowledge had a stronger effect than morphosyntactic knowledge in the past participle window and that they had an equal effect in the by phrase window. This idea also gains support from work showing that children with DLD have poor verb vocabulary learning and knowledge (Alt et al., 2004; Conti-Ramsden & Jones, 1997; Kan & Windsor, 2010; Kelly, 1997; Rice & Bode, 1993; Watkins et al., 1993). It is also consistent with how we set up the EP biases, which were created by showing agents participating in closely semantically related actions. One reason why the children with DLD may have demonstrated differences in their use of EP information is that they were less able to relate the actions seen during the learning phase to the semantically associated actions seen during the passive sentences. If a child's knowledge of the test and agent exposure verbs is shallow, then it is likely that the child will also struggle to make the connection between agents' attributes and their later participation in related actions. There may also be difficulty in retaining that information in the face of heightened processing demands, as when, for instance, interpreting passive verb morphology. We explore this possibility in Experiment 2, asking how the children's verb vocabulary knowledge affected their overall use of EP information.
In Experiment 1, we found that the children with DLD used verb-based EP information in different ways from children with TD when interpreting passive sentences. Two factors that seemed likely to produce this difference were reduced sensitivity to morphosyntactic information and poorer verb vocabulary knowledge in the group with DLD. In Experiment 2, we examine how the second factor—the quality of children's knowledge of the verbs—impacts their use of EP information.
We measured children's semantic processing with a repeated word association task previously used with children with DLD (Sheng & McGregor, 2010). The task involved asking children to produce semantic associates of the verbs used in the passive sentences. We expected that the degree to which a child used EP information to process sentences in the match condition would be related to their verb-specific semantic processing. Children with more developed networks of verb semantic associations should be able to respond more quickly to verb-based EP biases to facilitate passive sentence processing. We additionally controlled for overall receptive vocabulary knowledge (as measured by the PPVT-4 score), making this analysis a stringent test of the effect of children's verb-specific semantic knowledge.
In the semantic association task, the children repeatedly generated word associations (e.g., “eat”–“food,” “eat”–“drink,” “eat”–“swallow”) for each of the verbs used in the sentence prediction part of Experiment 1. This task was administered 1 week before Experiment 1. The children were presented with each verb used in the test sentences three times consecutively in progressive form (e.g., “lifting”). On each exposure, the children were asked to produce one semantically related word that was different from prior responses for that word. The children's responses were audio-recorded and transcribed for later scoring.
Before the test trials, experimenters provided models with two words and the children practiced with four words. The model words were “mom” and “birthday.” These words were each played three times, and the experimenters provided a semantically related response (e.g., “cake,” “party,” “present”). The children then practiced on the words “grass,” “digging,” “moon,” and “cutting.” Although the actual test items were verbs, we included nouns in the models and practice items because we found in pilot work that these were effective at conveying to the children that semantically related responses were desired. The first two practice words were played only one time each. Then, the children were told, “Every time you hear a word, I want you to say the first thing you think of. Say something new and different every time.” The final two words were played two times each.
During the task and practice trials, children were provided noncontingent feedback unless they produced multi-word responses or repeated responses within a verb, in which cases they were encouraged to “say one thing” or to “say something different.” Multi-word responses consisting of an article, a quantifier, or a possessive and a noun were treated as single-word responses. The children were given 10 s to provide a response; if one was not forthcoming, we moved onto the next trial. There were 48 trials on this task, which took approximately 10 min to complete. Verbs were presented in random order, but repetitions of a verb appeared consecutively. The children were randomly assigned to one of two counterbalance orders.
The children's responses were scored using a procedure based on the work of Sheng and McGregor (2010). Each response was coded as having a semantic relationship, a phonological relationship, or no relationship with the target verb. Semantic relationships included categorical (“run”–“exercise”), functional (“wear”–“hat”), descriptive (“run”–“fast”), thematic (“walk”–“leash”), causal (“lick”–“wet”), part-whole (“dog”–“tail”), or phrase completions (“give”–“back”). Phonological relationships included alliterations (“carry”–“carrot”) and rhymes (“broom”–“zoom”). Responses with no semantic or phonological relationship, including responses of “I don't know,” were coded as having no relationship. Verbs with any responses consisting of repetitions or inflections of the target verb were excluded from further analysis (n = 16 trials, 1%). Some children produced multi-word phrases despite our instructions. In an effort to include these responses (n = 535 trials, 26%), we subjected each word in the response to the coding scheme. If a word in the response had a semantic relationship with the target verb, we considered the response to be semantically related.
Coding was completed by two experimenters independently. The coders agreed on 97% of responses. The first author recoded responses in cases of disagreement. Across the verbs, 27% of the associates provided by the children with DLD were semantically related to the target verb, 3% were phonologically related responses, and the remaining 70% were unrelated. These values were 44%, 2%, and 54%, respectively, for the children with TD. Correlations between PPVT-4 and SPELT-P2 scores and proportion of semantically related responses were r = .30, t(38) = 1.91, p = .064, and r = .37, t(38) = 2.45, p = .019, respectively.
We used the measure of semantic relatedness to predict the proportion of target looking across the postverb windows in the passive sentences in the match condition (N = 1,153 trials). We chose windows after the verb was completed because, at that point, prediction based on the verb-based EP information was possible and strong effects of match condition were observed in these windows in Experiment 1. The proportion of target looking was calculated by taking the average of the looks to the target across the past participle, by phrase, and agent windows for each participant and trial in the match condition. This proportion was adjusted using the same logit transformation as in Experiment 1.
There were two predictors in the linear mixed-effects model. The first predictor was the semantic association score, each child's proportion of semantically related responses for the specific test verb appearing in the passive sentence. We also included group (TD/DLD) as a factor and its interaction with the semantic association score. In addition, we included PPVT-4 score as a covariate, random intercepts for item and participant, a random slope for the semantic association score on the participant intercept, and a random slope for group on the item intercept. Priors for the fixed effects were normal distributions with a mean of 0 and an SD of 0.5. There were 2,000 warmup samples and 4,500 postwarmup samples collected for each of four chains, resulting in 10,000 postwarmup samples for the entire model. The ROPE range was −0.055 to 0.055.
Figure 7 shows the relationship between the semantic association scores and the proportion of TagentTaction looking. The results of the model are presented in Table 9. There was a strong effect of group consistent with the results from Experiment 1, pd = 1, pROPE = 0. The group with TD had higher proportions of TagentTaction looking, EMM = .28, 95% CI [0.22, 0.34], than the group with DLD, EMM = .14, 95% CI [0.11, 0.18]. There was also a strong positive effect of semantic association score for both groups, pd = .99, pROPE = .003. Trials with a semantic association score of 1 (corresponding to a child having provided a semantic associate on all three trials for a verb) had higher proportions of TagentTaction looking for the associated trials, EMM = .25, 95% CI [0.19, 0.32], than trials with a semantic association score of 0, EMM = .18, 95% CI [0.15, 0.22]. (The ends of the range of semantic association score were selected to illustrate this continuous effect.) Last, we note that the effect of PPVT-4 score was not reliably associated with TagentTaction looking, pd = .86, pROPE = 1. In summary, we found strong evidence for a relationship between the quality of children's semantic knowledge of the test verbs and their use of EP information during online sentence processing and little evidence of a relationship with overall receptive vocabulary.
Figure 7. Relationship between event probability cue use and semantic association score. Ranges indicate 95% credible intervals of T
agentTaction(correct) looking proportion across semantic association scores. DLD = developmental language disorder; TD = typical development.
This study examined children's comprehension of passive sentences, asking how children with DLD differed from their peers with TD in their use of EP information and how verb vocabulary supported this process. In Experiment 1, we found that compared to the children with TD, the children with DLD demonstrated differences in the integration of EP information with morphosyntactic information in online processing. The children with DLD were more likely to integrate EP information early in the sentence with verb-proximal morphosyntactic information but had difficulty integrating EP information with verb-distant morphosyntactic information later in the sentence. In other words, for the children with DLD, when linguistic demands were lower, EP information was used; when linguistic demands were higher, EP information was used less and may even have become detrimental to interpretation.
The patterns in Experiment 1 suggest that one key weakness of the language processing of children with DLD is the coordination between event and linguistic knowledge. Such an idea would be consistent with the developmental patterns we see in these children as they grow older, where lower level morphosyntactic difficulties give way to higher level difficulties with inferential skills, reading comprehension, and discourse, implying ongoing difficulties with event knowledge. Future work should investigate how sensitive children with DLD are to properties of event knowledge, particularly in different contexts and under different task demands. Knowledge displayed in one context may be more difficult to access in another for these children. More generally, the idea of understanding how event knowledge structures relate to linguistic structures across development would be a rich starting point for understanding the interplay between conceptual knowledge, event processing, and language learning in children with DLD. Some work shows that individuals with aphasia are affected more by event knowledge than by linguistic knowledge like co-occurrence and argument structure, suggesting that impairments in language may be compensated for by increased reliance on event knowledge (Dresang et al., 2021; Hayes et al., 2016), a trade-off that could be beneficial to exploit for individuals with DLD if event knowledge can be improved.
In Experiment 2, we found that children in both groups used EP information more when they had better sentence-specific verb vocabulary knowledge event, though the children with DLD had overall poorer verb knowledge on the semantic association task than the children with TD. These patterns support the idea that the children deployed EP information at least in part based on the semantic overlap between the verbs in the agent attribute learning phase and the verbs in the passive sentences. Children with better sentence-specific verb vocabulary knowledge were better able to make the semantic connection and use it to predictively increase TagentTaction looking in the passives in the match condition.
One immediate application of this result is that children's sentence processing can be supported by improving their vocabulary knowledge. Deeper understanding of verbs seems especially important, including their associations with different agents, patients, and events (McRae et al., 1997). Teaching children the meanings of verbs and highlighting their relationships to other event participants may be an important way to improve passive sentence comprehension. One intervention technique designed for adults with aphasia, the Verb Network Strengthening Technique, targets these relationships and has been shown to improve sentence comprehension in these patients (Edmonds, 2016). Other work in adults with aphasia shows similar relations between verb knowledge and verb processing in sentences (Colvin et al., 2019). Extensions of this technique to children with DLD may be fruitful. Importantly, however, the weak association between PPVT-4 scores and TagentTaction looking suggests that targeting overall receptive vocabulary knowledge is not enough—instead, clinicians should focus on the specific words to be processed.
The results provide a foundation for understanding the contribution of EP information and vocabulary knowledge to the interpretation of sentences by children with DLD but there remain many questions to answer. First, though we saw evidence for overall differences in how children responded to the cues in passive sentences, we did not examine the dynamics of their responses across the experiment (see, e.g., Evans, 2002). Future work could address this by examining the evolution of children's processing biases from one sentence to the next. Second, in our online processing analysis, we assumed that TagentTaction looking was associated with a propensity to choose the TagentTaction image in the offline comprehension task. This assumption is reasonable, but future work should explicitly examine how robustly looking patterns at different points in the sentence predict ultimate comprehension choices. Finally, there are several possibilities for clinical applications of the current work that require more investigation to confirm. Dissociations in children's performance in the offline pointing and online processing tasks suggest that reliance on receptive picture-pointing tasks to assess sentence comprehension may be relatively insensitive to children's underlying linguistic and cognitive difficulties. One such difficulty was the clear negative effects of the by phrase on processing, suggesting that focusing on this feature in treatment may help comprehension. Additionally, new clinical interventions for passives could be developed that exploit informative sentence cues to build up flexibility in interpretation. Expectations could be gradually violated, forcing an increased reliance on morphosyntactic cues; pauses could be inserted after the past participle morphological marker on the verb to allow for greater processing time and to reduce subsequent working memory demands; and vocabulary items could be tailored to the child's knowledge. Our data hint that the use of such techniques might be successful, but further investigation will be needed.
This study has several limitations that future work could address. First, our eye-tracking task presented only four scenes per sentence. This means that our claims about the effect of EP information in children with DLD apply to a domain consisting of restricted interpretive options, but in the real world, interpretive options are not so constrained. Examining the use of EP biases by these children in a more open task would clarify the extent to which these biases shape language processing and learning more generally in the real world. A second limitation is that the short-term nature of our task meant that the EP biases exerted influence mainly in online processing. Extending this design to biases based on the natural affordances between verbs and agents (if enough can be identified that are accessible to young children) may help bring out the expected differences in offline comprehension. Relatedly, our measure of learning of the EP biases was not comprehensive. Future studies should assess learning more robustly, perhaps with a productive task like recalling the attribute. Additionally, our measure of verb-specific vocabulary knowledge was relatively simple, though it proved more informative than the overall receptive vocabulary test. Tapping into specific aspects of knowledge (e.g., associations between verbs and agents or patients for instance) might provide more information about which aspects of vocabulary knowledge are most important for sentence processing. Last, our sample was relatively homogeneous in race, dialect used, and maternal education, which may have been a strength in terms of controlling the effects of background and dialect but was also a weakness in terms of generalizability to other groups.
This study provides evidence that children with DLD demonstrate differences from children with TD in their use of EP information in passive sentence interpretation. At the same time, strong verb vocabulary knowledge can support the use of EP information for children in both groups. When EP information must be integrated with morphosyntactic information near to the verb, children with DLD demonstrate similar (or even faster catch-up) processing, but when this morphosyntactic information is far from the verb, children with DLD demonstrate confusion and slower processing. In summary, children with DLD show their best use of EP information when linguistic demands are low and verb knowledge is high.
Data, analysis scripts, and example stimuli are available at https://osf.io/b89q5/.
Justin B. Kueser was supported by National Institute on Deafness and Other Communication Disorders (NIDCD) Grants F31 DC018435 and R01 DC018593. Arielle Borovsky was supported by NIDCD Grant R01 DC018593. Laurence B. Leonard and Patricia Deevy were supported by NIDCD Grant R01 DC014708.
Justin B. Kueser was supported by National Institute on Deafness and Other Communication Disorders (NIDCD) Grants F31 DC018435 and R01 DC018593. Arielle Borovsky was supported by NIDCD Grant R01 DC018593. Laurence B. Leonard and Patricia Deevy were supported by NIDCD Grant R01 DC014708.
Data, analysis scripts, and example stimuli are available at https://osf.io/b89q5/.