Developmental changes in how children generalize from their experience to support predictive linguistic processing

Abstract

Prediction is posited to support fluent comprehension of speech—but how and when do young listeners, who encounter unfamiliar and novel events with high frequency, learn to deploy predictive processing strategies in these unfamiliar circumstances? The current work used a discourse-based event teaching paradigm to explore how English-speaking school-aged children (aged 5;0–8;11 [years;months]; N = 92) generalize from their (experimentally controlled) experience to generate real-time linguistic predictions about novel events during an eye-tracked sentence recognition task. The findings reveal developmental differences in how the initial structure of event exposure supports generalization. Specifically, real-time extension was supported by viewing multiple instances of events involving varied agents in the younger children (5–6 years), whereas older children (7–8 years) extended when they experienced repetition of events with identical agents. The findings support accounts of predictive processing suggesting that learners generate predictions in a variety of less predictable circumstances and suggest practical directions to support early learning and language processing skills.

Introduction

A central question in language development concerns how children use their limited experience to fluently and rapidly interpret speech. Prediction—a skill where language users preactivate potential elements of upcoming speech—is put forward as one means by which young learners “keep up with” the rapidly unfurling speech stream (Pickering & Gambi, 2018). Children (from at least 2–3 years of age if not younger) leverage a variety of semantic, morphosyntactic, and structural cues in the speech stream to support dynamic and predictive comprehension of speech (Deevy, Leonard, & Marchman, 2017; Gambi, Gorrie, Pickering, & Rabagliati, 2018; Huang & Snedeker, 2013; Lew-Williams & Fernald, 2007; Lukyanenko & Fisher, 2016; Mani & Huettig, 2012; Reuter, Dalawella, & Lew-Williams, 2021). Much of this prior work, however, measures predictive processing in contexts that are highly similar to children’s prior linguistic experiences. When considering that linguistic creativity is a hallmark feature of language (Hockett, 1960), the familiarity of stimuli in these studies might not, in fact, reflect the typical experience of language processing in everyday contexts, particularly for young children who are simultaneously learning about language and its accompanying referents. How and when might children deploy predictive mechanisms in less familiar or new circumstances? The current work sought to gain traction on this fundamental question by focusing on how young learners generalize from their prior event experiences to generate predictions for new events in real-time language processing.

Links between predictive processing and language learning

Predictive processing and language learning are posited to support each other in a number of ways such as by helping a learner recognize novelty and drive learning via error-based learning mechanisms and by helping listeners to “free up” processing capacity for learning subsequent novel items (Chang, Dell, & Bock, 2006; Fernald, Zangl, Portillo, & Marchman, 2008; Rabagliati, Gambi, & Pickering, 2016; Reuter, Borovsky, & Lew-Williams, 2019). Numerous studies have supported these hypothesized connections by identifying positive associations between language skills (e.g., vocabulary) and predictive processing in adults and children (Borovsky, Elman, & Fernald, 2012; Gambi, Jindal, Sharpe, Pickering, & Rabagliati, 2021; Hintz, Meyer, & Huettig, 2017; Kukona et al., 2016; Mani & Huettig, 2012; Rommers, Meyer, & Huettig, 2015). Both mechanisms are likely to work together as learners leverage their prior knowledge while interpreting speech in new situations. For example, to generate a prediction about upcoming language (irrespective of situational familiarity), learners must have sufficient linguistic skills (e.g., vocabulary knowledge) to activate and identify the referents of ongoing speech, and linguistic predictions, in turn, may support further language learning. The current work focused on the first part of this predict–learn cycle by identifying the learning conditions that help young learners to generalize beyond their existing linguistic/event knowledge to support predictive processing in previously unheard contexts. Specifically, we explored the conditions under which learners might generate predictions based on the semantic category of an agent that has previously carried out a similar event.

Prediction in new contexts

Typically, many of the paradigms that measure predictive processing in children rely on stimuli that probe highly familiar events. One highly replicated paradigm, which explores verb-mediated predictive priming, finds that adults and children as young as 2 years will fixate toward an object that is associated with a particular verb before that word is mentioned (e.g., listeners fixate more toward COOKIE vs. BOOK while hearing “the baby will eat the …”; Altmann & Kamide, 1999; Fernald, Zangl, et al., 2008; Mani & Huettig, 2012).

Young learners, who are simultaneously learning about the world and language, must frequently interpret language consisting of familiar words that describe new events. For example, what might young children predict when interpreting a sentence like “the duck will eat the …” where children might be able to identify a duck, and might be familiar with eating events, but not specifically what ducks eat? In this case, children may be able to generalize from their knowledge of the diets of similar animals to generate a number of plausible predictions. Despite the high ecological validity of this situation, we do not yet know how children generalize from their existing knowledge during real-time language interpretation in unfamiliar circumstances.

There are some clues suggesting that children can modify and update their predictive processing depending on the context and learning environment and therefore may build on their prior knowledge to generate predictions in new circumstances. For example, 4- and 5-year-olds can generate predictions for nouns following informative verbs when linguistic stimuli are varied in their predictability during a study (Reuter et al., 2021). French-speaking 3- and 4-year-olds can update their expectations for a syntactic category based on a few exposures to a construction (“la petite …”) when it is followed by either a noun or a verb (Havron, Babineau, Fiévet, de Carvalho, & Christophe, 2021; Havron, de Carvalho, Fiévet, & Christophe, 2019). In addition, 4- and 5-year-olds can rapidly update event information (e.g., tracking who has entered or left a scene) through narrative discourse to generate real-time inferences about the likely participant (among multiple possible referents) in an action (Yuile & Fisher, 2021). Adults and children are also able to use information about novel events described through discourse to support real-time prediction (Amato & MacDonald, 2010; Borovsky, Sweeney, Elman, & Fernald, 2014). For example, one prior study exposed children to brief illustrated narratives describing novel events that were unlikely to have been encountered in the children’s own experience (e.g., a monkey that rides a bus) and then found, in a subsequent visual world sentence recognition task, that children as young as 5 years (but not younger) could use this new experience to generate predictions for (previously unpredictable) items based on these (novel) events such as by predictively fixating toward an image of a BUS when hearing “the monkey rides the …” (Borovsky et al., 2014). Together, these studies suggest that children from at least early school age (~5 years) are able to retain new information in their linguistic environment to update their linguistic predictions accordingly. However, no study has yet identified whether or how young learners generalize from these experiences to predict upcoming language in similar but novel circumstances. More concretely, to take the “monkey riding a bus” example from above—would children, upon learning about this event, also generalize from this example that other similar creatures (e.g., gorillas) are potential public transit riders and deploy this knowledge in real-time prediction? Although this specific example may seem fanciful, the central question hits at a fundamental issue regarding how learners might be able to deploy cognitive mechanisms such as generalization to support their real-time interpretation of creative and novel linguistic input. One possibility is that learners may limit predictive processing in contexts that are highly similar to their prior experience and instead may rely on more integrative processing in less familiar contexts. Alternatively, learners may extend from their experience under the right conditions, which might suggest that anticipation does not exist at the level of precise lexical forms (e.g., Ferreira & Chantavarin, 2018).

One study in adults suggests that it is possible for learners to generalize from novel events to support prediction (Borovsky, 2017). This study found that a single experience with a novel event (e.g., “monkey riding a bus,” as in the above example) was not sufficient to support generalization-based prediction in adults. However, reinforcing this initial event experience through a single additional repetition was sufficient to support a generalized prediction to a novel but similar situation.

Thus, whereas adults generalize from relatively limited prior experience to predictively interpret language describing a novel situation, it is unknown whether and to what extent young children may be able to do so. Nevertheless, there are indications from prior memory research that the ability to learn about events and generalize from them is developmentally within the capabilities of children in the age range targeted in this study (5–8 years). For example, preschool-aged children (3–5 years) are able to recall novel and complex events after only a single experience, can learn from events involving variable participants, and can generalize from these experiences (Bauer & Fivush, 1992; Fivush, Kuebli, & Clubb, 1992). In addition, children in the targeted ages incorporate recently learned event information in real-time predictive processing, although their fluency in this skill is improving rapidly in this age range (Borovsky et al., 2014).

The current work asked whether and how children generalize from prior event experience to generate real-time linguistic predictions in two experiments. Both experiments used an event-teaching paradigm that had been developed in a prior study with 3- to 10-year-olds (Borovsky et al, 2014) and used materials from a prior study with adults that were simultaneously developed to be accessible to young children (Borovsky, 2017). The first experiment sought to replicate and extend prior developmental work. Children were initially exposed to novel events through brief illustrated narratives, and the experimental task was designed to determine whether children can simply generate predictions in language describing this newly learned information and whether this knowledge obtained from this single experience may generalize to support real-time linguistic prediction during a similar, but not identical, event. It was expected that the results of the first condition would mirror that from a prior study (Borovsky et al., 2014), where school-aged children (5–10 years) were able to generate real-time predictions for a recently learned event. Based on adult patterns, it was also expected that this single experience would not be sufficient to support generalization during prediction to a similar novel circumstance (Borovsky, 2017). This predicted failure to generalize after a single experience makes intuitive sense: Learners of all ages are unlikely to generate “promiscuous” over-generalizations from single unusual experiences.

Instead, it was anticipated that young learners, like adults, would require additional experience with an event. Therefore, Experiment 2 sought to measure how the structure of repeated encounters with identical and related events might support extension during linguistic prediction. Specifically, the experiment was designed to contrast whether and how variability versus repetition in experience supports generalization in prediction by exploring real-time extension in two conditions that create variability or repetition in the initial event exposure. Importantly, several studies in other domains of language acquisition provide alternating support for repetition and variability in linguistic extension.

Variability during learning is posited to generalization in a variety of language tasks, including in generalization to novel structures (Gerken & Bollt, 2008; Gómez, 2002; Gómez & Maye, 2005; Wonnacott, Boyd, Thomson, & Goldberg, 2012), and in extensions of novel actions (Childers & Paik, 2009; Childers, Paik, Flores, Lai, & Dolan, 2017; Forbes & Farrar, 1995).. Alternatively, the repetition account is supported by findings in the language development literature, which finds that generalization of argument structures is best supported under learning conditions that include a frequent or prototypical exemplar (Goldberg, Casenhiser, & Sethuraman, 2004). Similarly, numerous novel verb extension and event learning studies suggest that novel event/verb learning is less robust under highly variable conditions or conditions that are more distant to the original exposure context (Haryu, Imai, & Okada, 2011; Maguire, Hirsh-Pasek, Golinkoff, & Brandone, 2008) or when agents are not familiar in a novel event (Kersten & Smith, 2003). In addition, a prior study with adults that was similar in structure to the current experiment also found evidence that event repetition (and not event variability) also supported extension during real-time prediction (Borovsky, 2017).

To test whether variability supports generalization in this task, in one experimental condition (the Multiple Story condition), learners were presented with a story that describes events involving multiple (semantically similar) agents completing an action with a coordinated object (e.g., two insects—an ant and a butterfly—wear sunglasses, whereas two woodland creatures—a rabbit and a hamster—wear hats). Then, the effect of this initial experience was measured in how learners generate predicted fixations toward images depicting potential sentence-final objects in sentence stems that included an agent from one of the event categories—for example, by measuring looks toward SUNGLASSES versus HAT while children listen to a sentence containing another insect agent such as “the ladybug wears the ….”

To test whether repetition supports generalization during predictive processing, learners were exposed to novel events in a second experimental condition (the Repeated Story condition), where they viewed the same novel event two times (e.g., they were exposed twice to a brief illustrated narrative about a butterfly wearing sunglasses and a rabbit wearing a hat) before assessing whether this type of experience supports generalization during predictive processing.

This work strategically recruited children aged 5–8 years to explore whether and how learning mechanisms that support extension during prediction may change across a period of development in which children are able to quickly deploy recently acquired event information in sentence processing (as verified by prior work; Borovsky et al., 2014) but may still be developing facility in this skill. Thus, this age range was identified to explore the possibility that children change in performance across ages. One possibility is that younger children (5–6 years) may benefit more from variability than from repetition as compared with older children (7–8 years). This outcome would be consistent with the idea that younger children, who have less experience, may tend to form more specific memories (Barr & Brito, 2013), and therefore require a wider “breadth” of learning examples to help support extension. Such a result would also be consistent with findings that younger school-aged children may have difficulty in revising or updating their representations in discourse and sentence processing compared with school-aged children who are just a few years older (Yazbec, Kaschak, & Borovsky, 2019). Alternatively, it is possible that repetition may better support extension in younger versus older children. This account would be consistent with the idea that younger children, as compared with older children, generally require more encounters or repetitions to reinforce encoding in memory (see Bauer & Fivush, 2013, for a review) and evidence in the language acquisition literature that repetition, or similarity across events, may support relational mappings that support extension, particularly in younger children (as reviewed above; Gentner & Markman, 1997). Therefore, contrasting how variability and repetition in experience influence extension across development in Experiment 2 has the potential to illuminate how the learning mechanisms that support predictive processing in language change across a key point when these skills are still developing.

Experiment 1

Method

Participants

Families with children who were at least 5 years old and under 9 years old (i.e., 5- to 8-year-olds), and who met self-reported inclusionary criteria for the study, were invited to participate. The inclusionary criteria were as follows: English as the primary native language, normal hearing and vision (or corrected-to-normal vision), and no history of diagnosis or treatment for cognitive, speech, language, or attentional disorders. These criteria were verified via parental report. A total of 48 children meeting these criteria agreed to participate in this study, although 2 children failed to provide eye-tracking data and 1 child exhibited very low accuracy/attention to the task, leaving 45 remaining participants who contributed to the final analysis (M_age = 6.9 years, SD = 1.2, range = 5;2–8;11 [years; months]; 27 girls and 18 boys). Following previous sentence processing studies in our group, children were further grouped into 2-year age intervals for some analyses (n_5–6group = 24 [13 girls], M_age = 5.9 - years, SD = 0.61; n_7–8group = 21 (14 girls), M_age = 8.03 years, SD = 0.49). Caregivers provided verbal and written consent to participate in the study, and children provided verbal assent. The study was reviewed and approved by the university ethics board.

Materials

All study materials were identical to those used in Borovsky (2017), which was tested in adult participants. The experiment consisted of two interleaved phases: (1) story presentation and (2) sentence processing task. Materials for each phase are described separately.

Story phase materials.

The narrative structure for vignettes in Experiment 1 established four novel event relationships by introducing two separate agents, who performed the same two actions on two different objects (Fig. 1, Story A). Children listened to four of these stories across the study.

Fig. 1.

Sample story from Experiments 1 and 2. Story A illustrates an example from Experiment 1. In Experiment 2, participants either saw Story A repeated twice (Repeated Story condition) or saw Story A and Story B a single time (Multiple Story condition). Images appeared one at a time and were accompanied by spoken narration (listed above the images). Counterbalanced versions ensured that all possible combinations of agents, actions, and objects occurred.

Narrative structure.

Because the primary goal of the study was to determine how children leverage prior learning experiences to support real-time sentence processing, eight stories were constructed that conveyed novel fictional events using visual, auditory, and linguistic stimuli that were appropriate and accessible to even preschool-aged children, so that the materials would be accessible to even the youngest children (5-year-olds) in our target age range. These studies were designed following a structure outlined in Borovsky et al., (2014), where the novel events depicted in the story narrative followed a consistent structure involving two agents who performed the same two actions on two different thematic objects (and the pairing among event participants was counterbalanced, as described in greater detail below). The standard narrative event sequence is illustrated in Fig. 1, and all relations between event participants (agents, actions, and objects) are listed in Appendix A in the online supplementary material.

As depicted in Fig. 1, Story A, each story conveyed four novel events, each consisting of a single agent, action, and object (e.g., squirrel–holds–teddy bear, bee–turns on–lamp). The narrative design sought to present event relations that were likely to be novel to young children. Therefore, the narratives consisted of low-frequency events that could plausibly occur in a cartoon context but that were unlikely to have been encountered in children’s own experience. As an additional measure to ensure stimulus novelty, the pairings of the items in the events were counterbalanced across versions of the experiment such that all pairings of agents, actions, and objects appeared across all versions of the study with equal frequency.

In addition, the structure of the narration was controlled to help learners establish the connections among the event constituents, but it avoided using language that was identical to that to be used in the subsequent sentence recognition task. For example, in each vignette, after a particular agent was introduced, the narrator described the subsequent events using a pronoun (e.g., “He’s holding the teddy bear”) rather than referring to the agent (not “Squirrel is holding the teddy bear”). This additional control served to support comparisons between real-time recognition in conditions that involved either repetition or generalization from the initial event by ensuring that participants’ performance on the sentence recognition was not biased by having previously heard the same acoustic stimulus during the story presentation phase. Because the counterbalancing also served to ensure that all combinations of event constituents were equally likely to occur across versions, this aspect of the narration also ensured that each event relation was conveyed by an identical acoustic signal (i.e., across versions, children heard the same recording. “He’s holding the teddy bear,” irrespective of the agent who was connected to that event).

Story agents.

Because the primary goal of the study involved measuring how children would generalize from their experiences with events carried out by specific agents to events carried out either by these identical agents or by (taxonomically) similar agents, the agents in each constructed story narrative were carefully selected with several goals in mind. First, care was taken to ensure that the contrasting story agents who were associated with distinct objects belonged to distinct superordinate taxonomic subcategories (i.e., different animal subcategories). These taxonomic categories were chosen to be familiar to young children. To achieve this end, eight animal subcategories that commonly appear in early children’s literature and books were selected, with four each from mammalian and non-mammalian categories. Mammalian categories were woodland creatures, zoo animals, pets, and farm animals. Non-mammalian categories were bugs, birds, reptiles, and sea creatures. To keep the agents and events in the stories maximally distinct, agents in each story were selected from one mammalian category and one non-mammalian category. In addition, to verify that items would be familiar to young children, age-of-acquisition (AoA) norms and corpora of early child-directed speech were consulted. Agents had an age-appropriate AoA according to adult rating norms (M_AoA = 4.8 years, SD = 1.9, range = 2.8–8.2; Kuperman, Stadthagen-Gonzalez, & Brysbaert, 2012), and all were reported to appear in child-directed speech from 3 to 5 years of age, according to the ChildFreq database (Bååth, 2010).

Story narration.

All spoken narration in the story was prerecorded by a female native English speaker using child-directed prosody at 44.1 kHz on a mono channel. Mean intensity of all narrations was normalized to 70 dB.

The spoken narrations accompanied visual depictions of the events in the story, and examples can be viewed in Fig. 1. Note that these verbal narrations never included all elements of the event within a single utterance. For example, in the “squirrel holds teddy bear” combination, the narrator initially introduces the agent, “There’s a squirrel!,” and then describes the event without directly naming the agent again, “He is holding the teddy bear!” This element of the narrative design sought to ensure that the recognition and extension of subsequent sentences did not rely on simple repetition of the same acoustic stimulus from one part of the study to another. Furthermore, this control allowed for the same acoustic stimulus “he is holding the teddy bear” to be used to convey the same critical aspect of the event irrespective of the agent with which it was paired in that particular version.

Story images.

Colorful cartoon images accompanied the spoken story narrations and were 400 × 400 pixels in size (on a 1280 × 1024-pixel monitor).

Sentence recognition phase.

After listening to each story, participants completed 4 trials of an eye-tracked visual world sentence recognition task to assess their real-time ability to recognize and extend the events of the prior story. The structure of the task is depicted in Fig. 2, which illustrates that for each trial participants initially viewed four images that consisted of the thematic images from the prior story on the screen.

Fig. 2.

Example of an eye-tracked image array and possible accompanying sentences. In Experiment 1, participants heard sentences from the Repeat and Generalize conditions. In Experiment 2, only spoken sentences from the Generalize condition were heard. Art1/2, Article 1/2.

Participants then heard sentences that included one of these objects, and their eye movements toward these images were measured as they listened to these sentences. Participants completed a total of 4 trials after each story, yielding a total of 16 trials across the four blocks of the study, with 8 trials in each of two sentence conditions. The subject of each SVO (subject–verb–object) sentence in this task corresponded to one of two conditions: either a Match condition, where the sentence agent was identical to one that appeared in the prior story (e.g., “The bee turns on the lamp”), or a Generalize condition, where the sentence agent was a member of the same animal subcategory as one of the agents in the original story (e.g., “The fly turns on the lamp,” where fly is a member of the same BUG subcategory as one of the original story agents, bee).

Sentence recognition stimuli.

The images in this phase depicted the four thematic agents from the prior story phase (see Fig. 2) and consisted of each image placed on a white background in a 400 × 400-pixel image. This array of four objects was accompanied by a spoken sentence that included a thematic object that named one of the four images. Each image corresponded to one of four conditions, depending on the spoken sentence that accompanied the image: target, agent-related, action-related, or unrelated. For example, in Fig. 2, if the spoken sentence was “The squirrel turns on the flashlight,” then the target item would be FLASHLIGHT, whereas the agent-related item (which is the item that appeared with the squirrel agent in the prior story) would be TEDDY BEAR and the action-related item (or the item that “turns on”) would be LAMP. The rationale for including both agent-related and action-related distractors was so that it would be possible to determine whether predictive processing for the thematic object was driven by participants’ knowledge of the event (which included both the agent and the action) versus simply being driven by participants’ association of the objects that either the agent interacted with (verified by the agent-related cue) or driven by the verb alone (verified by the action-related cue). The unrelated item did not associate with either the agent or the action of the sentence. Fig. 2 illustrates how this image arrangement worked across sentence conditions using the example story in Fig. 1 as a reference. This image counterbalancing arrangement ensured that, for each subject, all images were equally likely to appear in all target and distractor image roles.

Spoken sentences were recorded by the same speaker as for the story narration and were edited in Praat (Boersma & Weenink, 2012) to a mean standard intensity of 70 dB. To ensure that sentences were aligned in duration across conditions, and to facilitate interpretation and alignment of eye movements with sentence stimuli, the timing of each word onset in each sentence was also aligned in Praat. Thus, each sentence that accompanied the sentence recognition task was identical in duration (2567 ms) and intensity. The timings of the words in each sentence were also aligned. Critically, the timing of the verb (where participants would start to have access to information necessary to make an anticipatory eye movement) began at 1059 ms after sentence onset, and the sentence-final object onset began at 2037 ms after sentence onset. Therefore, the critical, anticipatory window spanned a duration of 978 ms total, starting from 1059 ms (verb onset) and ending at 2037 ms (object onset) after sentence onset.

Procedure

After an initial warm-up phase in the laboratory playroom, participants were invited to take part in the experimental task by moving into an adjacent room, where they were directed to sit in a stationary armchair in front of a 17-inch screen and to wear child-sized earphones. The screen was mounted on a moveable arm that could be adjusted individually to the size of the child. The arm mount included an attached remote eye-tracking camera (EyeLink 1000 Plus eye tracker, 500 Hz sample acquisition; SR Research, Ottawa, Canada), and the position of the screen was adjusted so that the position of the eye-tracking camera was maintained approximately 600 ± 20 mm from the child. The child’s head position and distance were automatically tracked by affixing a small standard sticker to the child’s forehead. The experimenter informed children that they would initially hear a short story and that after each story they would be asked to listen to other sentences and to point to images that go with those sentences. Participants completed a single practice trial and were given feedback. Feedback on this practice included instructions (if needed) to minimize pointing until the sentence was completed as well as acknowledgment if participants responded correctly and clarification if they did not. Participants were then invited to ask questions. Next, the eye tracker was calibrated using a standard 5-point calibration routine before the experimental task began.

Each block of the study consisted of a story phase followed by a sentence recognition task phase. The story phase included engaging cartoon illustrations accompanied by child-directed spoken narration and served to establish novel event relations among the agents, actions, and objects. The narration started by initially presenting an isolated image or “page” in the story in silence for 2000 ms before the spoken narration proceeded. Once the narration was completed, a small square-shaped cursor appeared on the screen to indicate that the narration was ready to proceed. To ensure that the child attended to the narration, the experimenter controlled the advancement of the story such that a new frame in the story appeared on the screen only when the child was attentive.

At the completion of the story, the participant began the eye-tracked sentence recognition task. Each trial proceeded as follows. First, the centrally located image appeared centrally on the screen to check that the child was attending to the image and that the calibration remained valid before the onset of the trial. If the calibration check indicated a need for recalibration, then the experiment was paused to complete a new calibration routine (<30 s; rarely needed). Next, four objects that had been mentioned in the preceding story phase appeared on the screen in silence for 2000 ms before the onset of the spoken sentence. The child’s eye movements toward the object images were recorded over the course of the spoken sentence. Eye movements to the four objects were recorded across the course of the sentence at a sampling rate of 500 Hz and were binned into 50-ms samples for analysis. At the completion of the sentence, a small square-shaped cursor reappeared at the center of the screen. After this cursor prompt, the child was directed (if needed) to point toward the image that was mentioned in the sentence, and the experimenter recorded the child’s selected image. Participants were offered a longer break at the midpoint of the experiment, and the entire task lasted approximately 10 min.

Results and discussion

Picture selection accuracy

In each sentence recognition trial, children were asked to point toward the picture that fit the sentence. This response served as a behavioral indicator as to whether the child was attentive and could comprehend that trial. Trials where the child selected the incorrect image (i.e., the one that did not match the object of the sentence) were removed from further eye movement analysis. Aside from 1 child who provided only a single correct response and therefore was removed from subsequent analysis, overall accuracy was very high, with a mean accuracy of 98.5% and a standard deviation of 3.8%; the lowest accuracy rate by any single child was 81.3% (3 of 16 trials incorrect). Together, these findings indicate that children understood the language in the task and attended appropriately to the task throughout the duration of the experiment.

Eye movement analysis

Data cleaning.

Trials with incorrect behavioral responses or excessive track loss were removed from analysis so as to ensure analyses comprised trials where children were attentive and able to comprehend the stimuli. Before data cleaning, there were 720 trials in the initial dataset (360 in the Repeated Story condition and 360 in the Generalize condition). Of these, 11 trials (1.5% of the dataset) were removed where children selected the incorrect picture, leaving 709 trials (356 in the Repeated Story condition trials and 353 in the Generalize condition trials).

Next, following procedures in several prior studies, including a prior study using the same materials and task structure (Borovsky, 2017), trials were removed where there was greater than 80% loss of samples during the two anticipatory windows (verb window and article Window separately) due to track loss, blinks, or children’s inattention to the screen. In the verb window, this criterion identified 31 trials (4.4%) to be removed, leaving 678 trials (338 in the Match condition and 340 in the Generalize condition). In the article window, this criterion identified 51 trials (7.1%) for removal, leaving 658 trials (329 in the Match condition and 329 in the Generalize condition).

Time course visualization.

Fig. 3 illustrates gaze behavior on each task condition and visualizes performance across the younger (5–6 years) and older (7–8 years) age groups on the task. The plots illustrate that in the Match condition, both younger and older children began to uniquely fixate toward the target object before it was spoken. Notably, this pattern replicates behavior noted in a similar task (Borovsky et al., 2014) and indicates that children were successfully able to deploy information about the recently learned event to predict the behavior of the same agent in a sentence processing task. However, in the Generalize condition, children of both ages appeared to have difficulty in using the information from the prior story to make a predictive generalization about the behavior of a similar agent. These visualized patterns are next explored in statistical analyses, where predictive fixations toward the target object are measured as the verb and article are spoken separately.

Fig. 3.

Time course of looks to the interest areas in Experiment 1, broken down by Match Agent and Generalize Agent conditions and by older and younger age groups. Error ribbons represent 95% confidence intervals. yrs, years; Art, Article.

Time window analyses.

The goals of the next analyses were to determine (a) whether in each condition children fixated toward the target predictively (i.e., the target label was spoken) and (b) whether and how the degree of predictive target fixations varied by experimental condition (Match vs. Generalize) and age (as a continuous variable in months). Predictive fixation in any condition was inferred if, on average, looks to the target exceeded looks to the other competitors while the verb and/or article were spoken. The analytic code and results that contributed to these analyses are also posted in the supplementary material.

In the first target versus competitor analyses, mean looks to the target during the verb and article windows were calculated separately. The verb window represents the earliest period of time at which the information necessary to generate an appropriate fixation starts to emerge, whereas the article window represents the time just after the offset of the verb, when this information is fully available, and during which it would be possible to launch robust predictive fixations toward the target before it is spoken (for a similar word-defined time window analysis, see Kamide, Altmann, & Haywood, 2003). Therefore, strong evidence for predictive fixation to the target would emerge in one of two conditions: (a) early prediction, where the target exceeds distractor fixations in both the verb and article windows, and (b) standard prediction: where the target exceeds distractor fixations in the article window only. This comparison was carried out by first calculating each participant’s mean fixations proportion toward each interest area (target, agent-related, action-related, and unrelated) in each condition (Match and Generalize). Participants’ fixations in each age group’s (older and younger) fixations to the target were compared with each distractor separately (vs. the agent-related, action-related, and unrelated distractors). Note that because of the counterbalancing scheme, where each image appeared equally across each interest area region, this comparison represents a balanced measure of looks to the target looking to all images versus fixations toward the same images in each distractor condition. One-way t tests were used to determine whether mean proportion looks to the target exceeded the other three distractor conditions in each time window, with the critical alpha-level threshold adjusted for multiple comparisons via the Bonferroni method, and results are illustrated in Fig. 4 (see also Supplemental Table S1). These comparisons indicated that, both age groups fixated predictively toward the target in the Match condition only. Moreover, younger children’s fixation patterns exhibited prediction in the standard (article-only) window, whereas older children exhibited early prediction in both the verb and article windows. Children’s fixation patterns did not support either early or standard prediction in the Generalize condition in either age group.

Fig. 4.

Mean proportion looking to target and distractor image areas of interest (AOIs) in the verb and article time windows across age groups and conditions. Unfilled circles represent mean looks for single participants, and solid dots signify group means. Lines indicate Bonferroni-corrected significance levels comparing looks to target versus distractors. yrs, years; AgentRel, agent-related; ActionRel, action-related; ns, p > .05, *p < .05, **p < .01, ****p < .0001.

The second analyses sought to explore whether predictive looks to the target varied across condition and age using a linear mixed-effects regression analysis. Here, raw proportions to the target in the preceding verb and article windows (i.e., before the target object was spoken) were separately transformed using a logit approach, where the proportion looks to the target (Prop) are log-transformed as: log [Prop / (1 – Prop)]. Because values of 0 and 1 are undefined in this formula, a small, standard value (defined as half the smallest value in the dataset) was then added to 0 values and subtracted from data points equaling 1 (Barr, 2008; Dink & Ferguson, 2015). Then, the categorical values were contrast-coded with Match condition = .50 and Generalize condition = −.50, and continuous values of age in months were centered and normalized. Initial models including random effects of participants and items failed to converge. Further inspection of the random effects indicated that Items contributed very little variance in both the verb and article time windows, so this term was dropped in both windows, and final models included random effects of participants. The lmer formula for this reported analysis was Logit ~ Age * Condition + (1 | Participant).

Results of this analysis are illustrated in Table 1 for logit fixations to the target in the verb and article time windows separately. In both time windows, a significant effect of age reflected that older children showed more robust (predictive) looks toward the target than younger children. An effect of condition in the article time window reflected that looks to the target in the Match condition exceeded those in the Generalize condition.

Table 1.

Outcomes of linear mixed-effects modeling from Experiment 1.

	Verb logit			Article logit
Predictor	Estimate	Confidence interval	p	Estimate	Confidence interval	p
(Intercept)	−2.25	−2.54 to −1.95	<.001	−1.20	−1.56 to −0.84	<.001
Condition	0.40	−0.14 to 0.95	.150	0.94	0.26–1.63	.007
Age	0.35	0.05–0.65	.025	0.42	0.06–0.79	.028
Condition × Age	−0.03	−0.58 to 0.51	.906	−0.13	−0.82 to 0.56	.704
Random effects
σ 2	13.11			20.32
τ 00 subject	0.17			0.16
n subject	45			45
Observations	678			658
Marginal R2 / conditional R2	.012 / .024			.019 / .027

Note. The p values in bold are significant.

Together, the findings of the statistical analyses indicated three main patterns. First, children were able to remember and use information from the stories to support predictive processing for information that was available in the earlier story. That is, before the object of the sentence was spoken (the target), children began to already generate fixations toward this sentence-final object as the preceding verb and article were spoken. This result, which replicates patterns in this age in prior work (Borovsky et al., 2014), indicates that even 5-year-olds can engage in predictive processing with relatively new information without the need for repeated practice or long-term consolidation.

Second, there was evidence of developmental change in the sample, with anticipatory patterns becoming more robust as children aged. Here, the older age group demonstrated early prediction in the Repeated Story condition by uniquely fixating toward the target before it was mentioned during an earlier period in the sentence when the preceding verb and article were spoken. In contrast, the younger children’s fixation patterns yielded significant predictive target preference only in the (later but still predictive) article window. Together, this change in the speed and magnitude of the predictive processing effect with age is likely to reflect both ongoing improvements in the efficiency of linguistic processing, which is a skill that improves even into adolescence (Rigler et al., 2015), and improvements in memory skills during the early school years (Bauer & Fivush, 2013).

Finally, despite the robust learning and prediction in the Match condition, children, like adults in a prior study (Borovsky, 2017), were not able to use this newly learned information to support generalization during predictive processing. These findings also fit within prior evidence in verb learning, which suggests that even young children seem to extend verb meanings to new events, but only when given multiple (three) different training examples (for 2.5-year-olds; Childers, 2011) or multiple repetitions of the same example (for 5-year-olds; Imai, Haryu, & Okada, 2005). In our study, this lack of generalization during a single event experience may in fact serve to optimize processing by reducing the possibility that a single odd experience might lead to a spurious shift in how learners interprets other situations. Instead, generalized predictive processing may be more appropriate when learners can build on repeated or similar experience. Thus, children may need to either reinforce the initial event encoding (through repeated experience with the same event; Imai et al., 2005) or enrich the encoding (with additional relevant but not identical experience; Childers, 2011) to support subsequent generalization during real-time predictive processing of events carried out by similar agents to a prior experience.

Therefore, the next experiment sought to identify the learning conditions where additional input might best support extension from prior experience during predictive processing by either (a) adding an addition exposure to the same initial event information (Repeated Story condition) or (b) adding exposure about a new but similar event (Multiple Story condition).

Experiment 2

Children in Experiment 1 did not show robust prediction in a case where they needed to generalize from a single learning experience, which suggested that these young learners could benefit from additional learning context. Experiment 2 sought to explore the learning conditions that support generalization during real-time predictive processing by providing additional event experience that repeated or reinforced a novel event representation (Repeated Story condition) or by exposing the learning to a similar event that was carried out by a new similar agent from the same animal subcategory (Multiple Story condition).

If children generalize in the Repeated Story condition (as measured by evidence of anticipatory processing in an eye-tracked auditory sentence comprehension task in this condition), then this finding would suggest that generalization is supported by providing learning conditions that include more frequent prototypical exemplars, consistent with exemplar language learning accounts (e.g., Goldberg et al., 2004). On the other hand, variability accounts (e.g., Gómez, 2002) of language learning would be supported if children generalize in the Multiple Story condition.

Method

Participants

Families with children who were at least 5 years old and under 9 years old (i.e., 5- to 8-year-olds) who had not participated in Experiment 1, and who met the same self-reported inclusionary criteria as in Experiment 1, were invited to participate. A total of 52 families of children meeting these criteria agreed to participate in this study, but 5 of these children were not included in the analysis (2 children opted not to complete the eye-tracking task that day, there was an equipment error that caused a failure to play sounds for another child, and 2 other children exhibited very low accuracy/attention to the task), leaving 47 remaining participants who contributed to the final analysis (M_age = 7;0, SD = 14.2 months, range = 5;0–8;11; 22 girls and 25 boys). As in Experiment 1, children were further grouped into 2-year age intervals for some analyses (n_5–6gourp = 26 [13 girls], M_age = 6.1 years, SD = 0.56; n_7–8group = 21 [10 girls], M_age = 8.1 years, SD = 0.57). Caregivers provided verbal and written consent to participate in the study, and children provided verbal assent. The protocol was reviewed and approved by the university ethics review board.

Materials

Details that are unique to Experiment 2 and that were not reported in Experiment 1 are described below.

Stories.

Children heard four vignettes in Experiment 2 divided equally between two study conditions (Repeated Story and Multiple Story conditions) that sought to contrast whether repeated or varied experience would support generalization. In the Repeated Story condition, children simply heard the four novel event relations that were originally established in Experiment 1 an additional time (Fig. 1, Story A repeated twice). In the Multiple Story condition, the events that appeared in Experiment 1 where then followed by the description of similar events carried out by agents from the same two animal subcategories who performed identical actions to their corresponding subcategory member (Fig. 1, Stories A and B; Multiple Story condition).

Sentence recognition task stimuli.

In Experiment 2, participants only saw test sentences in the Generalize condition (see Fig. 2), where the subject agent was a different member of the same animal subcategory as one of the agents in the prior stories (e.g., “The fly turns on the lamp,” where fly is a member of the same BUG subcategory as one of the original story agents, bee or spider).

Results and discussion

Picture selection accuracy

As in Experiment 1, children’s attention and comprehension were confirmed by asking children to select (by pointing) the image that was an appropriate match for each sentence recognition trial. Incorrect trials were not included in eye movement data analyses. Overall accuracy was very high, with a mean accuracy of 97.14% (SD = 6.0), and the lowest accuracy rate by any single child was 68.8% (5 of 16 trials incorrect). As in Experiment 1, these responses indicated that children remained attentive and understood the language in the task.

Eye movement analysis

Data cleaning.

The initial dataset comprised 780 trials (390 in the Repeated Story condition and 390 in the Multiple Story condition). As in Experiment 1, trials were removed that contained incorrect behavioral responses or excessive track loss. Of the 780 trials, 34 (2.86% of the dataset) were removed where children selected the incorrect picture, leaving 746 trials (376 in the Repeated Story condition and 370 in the Multiple Story condition). Applying the same track loss criterion as in Experiment 1 (>80% missing samples in the verb or article window), an additional 41 trials (5.5%) were removed from further analysis from the verb window, leaving 705 trials (357 in the Repeated Story condition and 348 in the Multiple Story condition). In the article window, 50 trials (6.7%) were removed, leaving 696 trials in the final analysis (352 in the Repeated Story condition and 344 in the Multiple Story condition).

Time course visualization.

Fig. 5 illustrates gaze behavior on each task condition across the younger (5–6 years) and older (7–8 years) age groups. The plots suggest that the two age groups responded differently to the event exposure conditions. Specifically, whereas the older group shows robust predictive looking in the Repeated Story condition and not the Multiple Story condition, the younger group appears to show the opposite pattern. The statistical analyses below provide further insight into these visualized patterns in the data.

Fig. 5.

Time course of looks to the interest areas over the course of the sentence recognition task in Experiment 2, faceted by the Repeated Story and Multiple Story conditions and by the older and younger age groups. Error ribbons represent 95% confidence intervals. Dashed vertical lines indicate word onsets at each position in the sentence. yrs, years; Art, Article.

Time window analyses.

Following the analytic procedures as in Experiment 1, statistical analyses focused on exploring (a) whether in each condition children fixated toward the target predictively (i.e., the target label was spoken) and (b) whether and how the degree of predictive target fixations varied by experimental condition (Repeated Story vs. Multiple Story) and age (in months). The analytic code and reports that contributed to the results below are also posted in the supplementary material.

In the first analyses, mean looks to the target during the verb and article windows were calculated separately, and predictive processing considered either early or standard prediction according to whether target fixations exceeded distractor looking starting from the verb window and continuing through the article window (early prediction) or just during the article window (standard prediction). Shapiro–Wilk tests indicated that the distribution of fixations in the article window was not normally distributed. Therefore, as in Experiment 1, each participant’s mean fixation toward each region of interest in each condition was calculated separately. Then, each age group’s fixations to the target were compared with each distractor separately (vs. the agent-related, action-related, and unrelated targets) using either one-way t tests (verb window) or Wilcoxon tests (article window) with the multiple-comparisons critical alpha-level threshold adjusted via the Bonferroni method. These comparisons are illustrated in Fig. 6 (and in Supplemental Table S2). The statistical comparisons indicated that age groups varied in their predictive patterns across conditions. First, neither age group exhibited early prediction during the verb window. However, there were developmental differences in predictive patterns in the standard predictive window (i.e., during the article). Whereas the younger group generated robust predictive fixations in the Multiple Story condition only, the older group did so in the Repeated Story condition only (Fig. 6).

Fig. 6.

Experiment 2 mean proportion looking to target and distractor image areas of interest (AOIs) in the verb and article time windows across age groups and conditions. Unfilled circles represent mean looks for single participants, and solid dots signify group means. Lines indicate Bonferroni-corrected significance levels comparing looks to target versus distractors. yrs, years; RepeatStory, Repeated Story; MultStory, Multiple Story; AgentRel, agent-related; ActionRel, action-related; ns, p > .05, *p < .05, **p < .01, ***p < .001, ****p < .0001.

The second analysis sought to determine whether children exhibited differential performance in each predictive window as a function of age and condition. Using the same linear mixed-effects modeling approach as in Experiment 1, age was modeled as a continuous variable (age in months, centered and normalized) and condition was contrast-coded with Repeated Story = .50 and Multiple Story = −.50. Fixations to the target were transformed to logit as in the previous analysis. Initial models included both items and participants as random effects. In the article time window, this model converged, and so the model is reported for this analysis as Logit ~ Age * Condition + (1 | Participant) + (1 | Item). In the verb window, this model failed to converge, and closer inspection of the random-effects structure indicated that the participant term accounted for less variance, so this term was dropped from the model, which then did converge. The final model in the verb window was Logit ~ Age * Condition + (1 | Item).

The results of these models are reported in Table 2 and indicate a significant interaction of condition by age during the article time window. Fig. 7 illustrates the model interaction effects of condition and age. This figure suggests that this interaction effect was driven by different responses to the learning conditions as a function of age. Namely, both older and younger children benefited from additional experience with the narrated events, but in different ways. Younger children, who have less experience and knowledge, benefited from variable experience to support extension, whereas predictive extension in older children was supported by additional opportunity to encode the event through repetition. The broader implications of these patterns are addressed in the General Discussion below.

Table 2.

Outcomes of linear mixed-effects modeling from Experiment 2.

	Verb logit			Article logit
Predictor	Estimate	Confidence interval	p	Estimate	Confidence interval	p
(Intercept)	−2.45	−2.74 to −2.16	<.001	−1.45	−1.83 to −1.08	<.001
Condition	0.23	−0.29 to 0.74	.384	−0.04	−0.69 to 0.61	.898
Age	0.06	−0.20 to 0.31	.669	0.14	−0.22 to 0.51	.450
Condition × Age	0.30	−0.22 to 0.82	.257	1.07	0.42–1.72	.001
Random effects
σ 2	12.05			19.01
τ 00 item	0.27			0.10
τ 00 subject	–			0.36
n subject	47			47
Observations	705			696
Marginal R2 / conditional R2	.003 / .025			.016 / .039

Note. The p values in bold are significant.

Fig. 7.

Illustration of Condition × Age interaction model effect in Experiment 2. Age values are centered and normalized, with smaller values indicating younger ages.

General discussion

This work sought to explore how children build on prior experience to deploy predictive mechanisms in new situations. Although it is now clear that children use predictive processing in a variety of linguistic tasks, much prior work had focused, by design, on situations that were highly familiar to young learners. This prior focus on prediction in familiar situations represents an important and necessary first step in our understanding of how predictive mechanisms support language processing during childhood. However, language about unfamiliar situations is a common circumstance for young learners. Therefore, this work sought to gain traction on this fundamental issue in two studies that used a controlled learning paradigm to explore how children’s exposure to related events might support predictive processing in new circumstances that share some elements of the initial learning context. There are several key findings from this work.

First, irrespective of age, a single experience with a new event was not sufficient to support generalization, as demonstrated by the lack of prediction for the Generalize condition in Experiment 1. This finding also mirrors patterns in a similar study with adults, suggesting that this pattern remains consistent at least from school age (if not earlier). Although at first glance this lack of generalization from a single experience may seem suboptimal, it may in fact represent an appropriately conservative pattern of learning. That is, this pattern suggests that learners do not generate inappropriately broad connections between events based on a single new experience. Instead, learners require additional evidence before extracting patterns that apply to other related events.

Experiment 2 built on the first study by exploring whether and how additional experience with narrated events might allow learners to predict in related circumstances. Crucially, children learned about events in narrated contexts that provided additional repetition of a single event with a single agent or, alternatively, learned in variable event contexts with multiple agents taking part in the same event. Thus, the second study explored how simple event repetition (with the Repeated Story condition) as compared with event variability (in the Multiple Story condition) would support subsequent predictive generalization. This experimental manipulation revealed that school-aged learners can generalize from their prior experience during real-time language comprehension but that there are developmental differences in what kinds of experience support real-time prediction. Namely, whereas younger learners’ (5- and 6-year-olds’) predictive extensions benefit from exposure to related but varied events, older learners (7- and 8-year-olds) benefited from repeated experience with a single event.

At first glance, these developmental patterns in how learning context supported generalization may seem surprising, and there are several reasons that might help to explain these differences between the older and younger children in the study. Overall, these differences are likely to reflect maturation in the breadth and depth of children’s knowledge that occurs over this period as children gain experience with the world, develop fluency in literacy skills, and prepare to launch into self-guided learning during reading (Chall, 1996). This idea is also consistent with evidence that children’s understanding of category structure becomes more differentiated in the age range of 4–8 years (Unger, Fisher, Nugent, Ventura, & MacLellan, 2016; Vales, Stevens, & Fisher, 2020). For example, when older (7–8 years) and younger (4–5 years) children take part in a task that uses spatial arrangement of pictures to indicate similarity among objects, older children more closely arrange items within a category domain that shares more features (e.g., overlapping taxonomic and functional context) than those that share only taxonomic cues, whereas younger children show within-domain differentiation to a lesser extent (Vales et al., 2020). This finding suggests that older children are able to make finer distinctions between items within a category than younger children. This difference in the semantic development of children in the same age range as those in the current work provides one explanation for why 7- and 8-year-olds in this work (like adults) did not show generalization in the variability condition—because they may have been more sensitive to the differences among individuals in each agent category compared with the younger group. Care was taken to select the agents in the work so that they would be relatively familiar to even the youngest children in this work (as described in the Method sections), and in prior work these stimuli were normed in adults to verify the category memberships of the agents in the studies. However, this category knowledge was not verified at the participant level in this task. It remains an interesting question for future work to explore how individual differences in learners’ own connections between the variable event conditions might support subsequent prediction.

These developmental patterns suggest that developmental accounts of the drivers of predictive linguistic processing could be enriched by accounting for the structure and content of learners’ knowledge and semantic memory. Specifically, although there is some indication that language processing can be affected by the semantic structure of children’s lexicon (Borovsky, Ellis, Evans, & Elman, 2016; Borovsky & Peters, 2019; Vales & Fisher, 2019). Other research also points to a need to tease apart how maturational and experiential changes in memory skills and children’s own knowledge support real-time linguistic processing and vice versa. For instance, generating prediction error during processing seems to facilitate encoding and retrieval of known and novel words (Gambi, Pickering, & Rabagliati, 2021; Haeuser & Kray, 2021; Reuter et al., 2019). Similarly, memory structures may also influence subsequent knowledge generation and extension, and research in children suggests that there are developmental differences in generating new knowledge from existing information (Bauer, Blue, Xu, & Esposito, 2016; Bauer & Souci, 2010). It is also possible to speculate that, based on earlier noted developmental differences in the specificity of memory formation (with younger children forming more specific memories than older children; Barr & Brito, 2013), variability among multiple agents might help younger children to recognize similarities between related experiences, whereas reinforcement through repetition may support learning and extension in older children who may be better able to draw analogies of their own without prompting. The current study connects these distinct domains of inquiry to suggest that developmental differences in the structure of children’s experience and memory can drive how children continue to learn about events and their real-time comprehension of language, which continues to mature throughout childhood (Bauer & Fivush, 2013; Creel, 2019; Rigler et al., 2015). In sum, these studies suggest a fruitful pathway forward to advance theoretical accounts of language processing by using controlled learning paradigms to systematically measure how the experience and memory of the learner can affect language processing skills across development.

From an applied perspective, the developmental differences noted here also provide practical insights into the conditions that might best support children as they interpret language surrounding new events. For younger children, who are still learning about the world, it may help to facilitate comprehension of a new topic or event by pointing to multiple related examples. These findings are consistent with work suggesting that far generalization of events during novel verb learning can be facilitated by variable learning contexts (e.g., Childers & Paik, 2009; Childers et al., 2017; Forbes & Farrar, 1995). On the other hand, with older children, who are able to self-generate and reinforce robust connections among early words and events (Bauer et al., 2016), it may be simply enough to reinforce new situations through repeated reminders of prior related examples. This pattern in older children might also explain why, in some cases, variable learning environments are supportive for novel event learning and extension (e.g., Haryu et al., 2011; Kersten & Smith, 2003; Maguire et al., 2008). However, a limitation of the current work is that the task does not measure whether the observed differences in prediction have consequences for learning. Future work is needed to explore how developmental differences in learning environments that support subsequent predictive processing might enhance or inhibit subsequent learning.

Finally, these findings also inform theories of how predictive mechanisms support everyday language processing. Specifically, the findings illustrate that even young learners leverage limited experience to engage in predictive processing for new but related events. At a minimum, this observation suggests that the contents of predictive processing might not be narrow in scope. Rather, the findings are more consistent with accounts positing that prediction varies according to the context, knowledge, and learning skills of the listener, as is consistent with pluralistic accounts of prediction and other recent accounts suggesting that linguistic predictions are driven by multiple factors and graded (Kuperberg & Jaeger, 2016; Mani, Daum, & Huettig, 2016).

In sum, language enables the communication of endless imaginative ideas. This integral creativity does not divorce linguistic processing from the fundamental mechanics of cognition, nor does it create insurmountable challenges for predictive language processing even in young learners, who have limited experience from which to generate predictions. Rather, when considering multiple cognitive mechanisms, which change across development, we can see how fluent interpretation of the creative possibilities of language are possible even in young learners by leveraging our prodigious capabilities for learning, generalization, and prediction.