An eye-tracking study of selective trust development in children with and without autism spectrum disorder

Highlights

• We investigated whether children with ASD can selectively learn from speakers.

• Children with ASD, children with DLD and TD children participated in this study

• Speakers differed on previously demonstrated accuracy and on potential reliability.

• Children in all groups preferred word-mappings offered by previously accurate speakers.

• Eye-tracking data revealed that they avoided inaccurate speakers’ testimony.

• Only TD children displayed preference for mappings offered by more reliable speakers.

Abstract

The purpose of this study was to explore whether children with autism display selectivity in social learning. We investigated the processing of word mappings provided by speakers who differed on previously demonstrated accuracy and on potential degree of reliability in three groups of children (children with autism spectrum disorder, children with developmental language disorder, and typically developing children) aged 4–9 years. In Task 1, one speaker consistently misnamed familiar objects and the second speaker consistently gave correct names. In Task 2, both speakers provided correct information but differed on how they could achieve this accuracy. We analyzed how the speakers’ profiles influenced children’s decisions to rely on them in order to learn novel words. We also examined how children attended to the speakers’ testimony by tracking their eye movements and comparing children’ gaze distribution across speakers’ faces and objects of their choice. Results show that children rely on associative trait attribution heuristics to selectively learn from accurate speakers. In Task 1, children in all groups preferred the novel object selected by accurate speakers and directly avoided information provided by previously inaccurate speakers, as revealed by the eye-tracking data. In Task 2, where more sophisticated reasoning about speakers’ reliability was required, only children in the typically developing group performed above chance. Nonverbal intelligence score emerged as a predictor of children’s preference for more reliable informational sources. In addition, children with autism exhibited reduced attention to speakers’ faces compared with children in the comparison groups.

Introduction

Most situations in which children acquire new words are highly ambiguous. A great number of objects or properties can potentially correspond to an unknown word, especially because the intended referent is almost never presented in isolation. Therefore, it is relatively unsurprising that children take into account various social cues to determine the meanings of new words. For instance, tracking a parent’s attention focus is a powerful social cue to determine the reference of her or his utterances (Akhtar and Tomasello, 2000, Baldwin, 1995).

Autism spectrum disorder (ASD) is characterized by impaired social functioning and communication as well as by stereotyped and repetitive behaviors (American Psychiatric Association, 2013). Children with ASD are known to display difficulty in adequately tracking and understanding the communicative intentions of people around them (Baron-Cohen et al., 1985, Leekam and Perner, 1991, Yirmiya et al., 1998). It has been suggested that because the ability to map words onto meanings requires attending to social cues, such as the speaker’s gaze, social deficits may affect word learning in ASD (Baron-Cohen et al., 1997, Preissler and Carey, 2005). According to this view, lexical deficits and delays that are often attested in children with ASD could stem from a limited capacity to follow and understand social cues. In line with this hypothesis, unlike their typically developing (TD) peers, children with ASD show a tendency to map words onto referents located within their own attentional focus rather than on those attended by the speaker (Parish-Morris et al., 2007, Preissler and Carey, 2005). However, these conclusions have been challenged by a series of more recent eye-tracking studies in which children and adults with ASD proved to be sensitive to the direction of the experimenter’s gaze in word-mapping tasks (Bean Ellawadi and McGregor, 2016, Franken et al., 2010, Luyster and Lord, 2009, McGregor et al., 2013, Norbury et al., 2010). Thus, it appears that children with ASD do take speakers into account—to a certain extent, at least—when they acquire and process language.

A promising way to gain better insights into the role that social understanding plays in vocabulary acquisition is to investigate whether social cues in word-learning situations are treated by children in ASD in a flexible way. A child’s lexical repertoire is drawn from multiple sources, some of which can be misleading. Speakers may mispronounce some words because they are not fully fluent in the language they speak to the child, others may lack the relevant expertise, and still others may try to deliberately deceive the child or make jokes. This variation in reliability creates pressure to be selective about which information sources are indeed trustworthy. A beneficial strategy for a child who acquires language, thus, would be to rely on the cues offered by reliable sources but to disregard misleading informants. A large body of research on selective learning in TD infants and preschoolers provides evidence for such flexible use of social information (for reviews, see Mills, 2013, and Sobel & Kushnir, 2013). Even very young children frequently modulate their word-learning strategies according to previously demonstrated behaviors of informants; they selectively choose to learn words from speakers who described themselves as knowledgeable (Sabbagh & Baldwin, 2001) or who demonstrated accurate lexical knowledge (Koenig, Clément, & Harris, 2004). Selective learning has been classically studied using a paradigm by Koenig et al. (2004), where children witness (directly or through video recording) two speakers who provide labels for familiar referents, with one speaker systematically mislabeling the objects (e.g., calling a spoon a cup) and another speaker providing the correct labels. Next, children are presented with an unfamiliar object for which both speakers provide conflicting labels; crucially, children appear to display a robust preference for the label used by the accurate speaker. Multiple subsequent studies demonstrated that early selective trust extends beyond the language domain to other learning situations in which an adult’s testimony serves as a major source of information (Birch et al., 2008, Harris et al., 2018).

Although evidence for the preference for reliable informants in childhood is ubiquitous, explanations for these results in typical development are not yet firmly established and the extent to which they reflect genuine understanding of speakers’ trustworthiness is still a matter of controversy (Birch et al., 2017, Poulin-Dubois and Brosseau-Liard, 2016). One possible explanation, in line with sociopragmatic accounts of language learning, would be that, from early developmental stages, children detect and rely on speakers’ mental states. Under this line of thought, preschoolers use social information in a flexible way by making inferences about speakers’ epistemic states and choosing to trust adults who had demonstrated that they may possess the relevant evidence (Koenig and Harris, 2007, Sobel and Kushnir, 2013).

An alternative—and conflicting—explanation, however, is that selective learning is rooted in a less sophisticated associative mechanism. According to this line of thought, children attribute surface traits to speakers based on their observations of these speakers’ previous behavior. It is quite exceptional to witness speakers mislabeling objects around them. Because such behavior may be perceived as highly anomalous, it could be that children simply bypass information coming from inaccurate speakers. Thus, reliable speakers would be held generally knowledgeable—rather than situationally informed—because they do not exhibit the odd behavior characteristic of inaccurate speakers (Lucas & Lewis, 2010).

Whereas sociopragmatic explanations entail that children genuinely process speakers’ knowledge states, associative responses rest on nothing more than simple global rejection of the unconventional behavior demonstrated by inaccurate speakers. Children’s preference for the information associated with accurate speakers would not arise from rational inference about their expertise but rather is simply induced by a rejection of inaccurate speakers. Thus, even though a preference for accurate speakers may misleadingly appear as a rational choice, it may simply result from a surface trait attribution that leads to the rejection of inaccurate speakers. In other situations, where children erroneously rely on irrelevant attributes, the superficial trait attribution process may result in irrational trusting behavior. For instance, preschoolers were found to selectively learn words from attractive, nice, and strong speakers rather than from unattractive, mean, and weaker ones even when the former were wrong (Bascandziev and Harris, 2014, Fusaro et al., 2011, Landrum et al., 2013).

To assess the potential implication of mental state understanding in selective learning, one could examine correlations between children’s performance on standard theory of mind tasks and tasks on speakers’ reliability. However, currently available results are inconclusive, with some researchers reporting moderate correlations (Brosseau-Liard et al., 2015, DiYanni et al., 2012) and others finding no correlation at all (Brosseau-Liard et al., 2018, Pasquini et al., 2007). To investigate the impact of speakers’ trustworthiness on language acquisition in ASD, a clinical group with robustly documented difficulties in perspective taking and mentalizing may provide more direct insight into the nature of selective learning.

Recall that there is ample evidence that individuals with ASD have limited abilities in tracking the mental states of others during interaction. If selective learning is rooted in mentalizing abilities, as held by the proponents of sociopragmatic theories, then one should expect children with ASD to fail to display selective learning based on their misunderstanding of speakers’ epistemic statuses. However, if children’s learning strategies in tasks where they prefer learning from previously accurate speakers rely on a more general associative mechanism, then children with ASD could prove to be as sensitive as TD children to speakers’ previous accuracy.

It is also possible that in some situations of social learning, TD children spontaneously engage in mental state reasoning but also use less sophisticated mechanisms in other situations. A dual-processing account of selective trust proposes that children simultaneously use two kinds of underlying cognitive process: one unsophisticated and fast mechanism that is driven by trait attribution and one slow process that requires situational inference-making strategies (Hermes, Behne, & Rakoczy, 2018). Importantly, these two mechanisms may coexist and operate in parallel as determined by situational context; it is also likely that a more sophisticated process will appear later during child development.

One way to determine whether a nuanced understanding of epistemic reasoning is available to young children is to expose them to speakers who demonstrate the same degree of accuracy on the surface but differ in the way they achieve this accuracy. For example, Nurmsoo and Robinson (2009) found that preschoolers fail to take into account the difference in information access between two speakers. In this study, children showed no preference for speakers who had excusable reason for erring (blindfolded speakers) over speakers who had no obvious excuse for being inaccurate. However, in another study (Kondrad & Jaswal, 2012), 4- and 5-year-old children were found to overlook semantic errors that were closer to being correct (e.g., the mislabeling of a comb as a brush rather than as a thunderstorm) when the speaker’s errors could be excused. Likewise, Einav and Robinson (2011) reported that by 4 years of age children are sensitive to the conditions under which accuracy is achieved. In this study, children were presented with two speakers, both of whom correctly named familiar objects; however, whereas one speaker needed help to do so, the other one reached correct labels without being helped. The authors found that 4-year-olds, but not 3-year-olds, displayed a preference for the novel labels that were offered by the latter speaker. Age was also found to predict children’s ability to prioritize the accuracy of a speaker over the speaker’s confidence in learning situations (Brosseau-Liard, Cassels, & Birch, 2014). These results suggest that the mechanisms involved in selective learning may be situation dependent and evolve across development, with more sophisticated inference-based mechanisms appearing only later.

Nuanced social selective learning may also remain challenging for younger preschoolers because in such tasks children would need to make inferences at two parallel levels (Landrum, Eaves, & Shafto, 2015). Learning accurate information from others requires drawing inferences both about speakers’ state of knowledge and about the information they provide. In other words, learners must apply their assumptions about the speakers’ epistemic state to make inferences about object–referent pairs. One could argue that failure to learn words selectively by younger children in more complex contexts does not follow so much from a difficulty to evaluate speakers’ epistemic statuses as from a difficulty to apply this information while reasoning about which of two available referents should be associated with a novel word.

Independent of the reason why sophisticated processing of social cues is not entirely in place in younger TD children, one should expect it to be unavailable to children with ASD. A number of studies on language processing reported significant impairments in context-based inferences in ASD (Bodner et al., 2015, Dennis et al., 2001, Happé, 1994). In addition to difficulties in drawing inferences about intentions or mental states (Scott & Baron-Cohen, 1996), some studies have also reported impaired processing of causal links in nonsocial events in autism (Mason, Williams, Kana, Minshew, & Just, 2008). Difficulties in drawing context-based inferences should prevent children with ASD from taking into account nuanced differences in speakers’ behavior and associating it with various degrees of trustworthiness.

To get more insight into the mechanisms involved in social selective learning, it is important to use methods that would not only determine whether children discriminate between reliable and less reliable speakers but also reveal, through a dynamic exploration of social cue processing, how children reach their decisions. To that end, behavioral data on children’s performance should be complemented by visual exploration patterns. It is widely acknowledged that visually exploring people’s faces provides powerful social cues. For instance, one study found that 14-month-old infants more readily followed the gaze of a reliable experimenter versus an unreliable one (Chow, Poulin-Dubois, & Lewis, 2008). Therefore, eye tracking constitutes a promising tool to study selective trust.

Eye tracking is especially relevant for our research question because numerous studies reveal that individuals with ASD attend less than their TD peers to social stimuli, especially when there is competition between social and nonsocial information (Kikuchi et al., 2009, Moore et al., 2012, Riby and Hancock, 2009). Furthermore, even when overall distribution of attention to social and nonsocial stimuli is within the typical range, the time course of gaze shifts to important social cues may still significantly differ between individuals with and without ASD (Fletcher-Watson et al., 2009, Kikuchi et al., 2009).

That said, in studies that explored visual attention in referential tasks, no significant group differences were found in the amount of attention allocated to social (eyes and mouth of a speaker) versus nonsocial (target object and distractor) areas of interest between children with ASD and TD children (Norbury et al., 2010, Tenenbaum et al., 2014). However, the gaze exploration of social and nonsocial information in a more complex word-learning situation—when two speakers produce conflicting information—has not yet been studied. Furthermore, according to a recent meta-analysis, high social content (corresponding to a higher number of people involved in an observed interaction) has been found to be a strong predictor of diminished social attention in autism (Chita-Tegmark, 2016). One could expect that children with ASD would be less attracted by socially relevant information in complex social scenes and, for this reason, would experience difficulties in forming expectations about speakers’ competence.

Another factor that can help to elucidate the relationships between sociopragmatic processes and selective learning in autism is the choice of control groups. The most widespread method in research on ASD is to recruit comparison groups of TD individuals matched on chronological age and/or cognitive abilities—in spite of the fact that children with ASD often exhibit a significant delay in language acquisition. Children with developmental language disorder (DLD),¹ who usually have vocabulary deficits comparable to those observed in children with ASD, arguably constitute a more adequate comparison group for studying word learning in autism. Moreover, such a comparison may help to elucidate the role of attention to social information in selective learning given that children with DLD have been found to privilege socially relevant information in the same way as their TD peers (Hanley et al., 2014). At the same time, children with DLD may exhibit important sociocommunicative impairments usually associated with poor language skills (Botting and Conti-Ramsden, 2008, St Clair et al., 2011). Given that both children with ASD and children with DLD share similar communication difficulties but are likely to differ in their processing of social cues, comparing these populations may increase our understanding of the mechanisms that support selectivity in social learning.

The current study was designed to explore the role of social-pragmatic reasoning in selective learning tasks in both typical and atypical development. Tasks on selective trust have been extensively implemented with TD children and infants, but to the best of our knowledge no study has yet addressed this issue in children with ASD.

Our aim here was to investigate whether children with ASD can selectively use social cues for word learning. We also considered two alternative theoretical accounts of social selective learning: one that explains early sensitivity to information sources, in typical development, by a precocious capacity to reason about knowledge states of speakers and the other that explains this sensitivity by a general surface trait attribution mechanism. One way of adjudicating between these two positions is to investigate selective social learning in children who show impoverished social-pragmatic processing. The past 25 years of research yielded ample and robust evidence that individuals with ASD present such a clinical profile. Given the relatively uncontroversial presence of difficulty in reasoning about other people’s mental states in ASD, two antagonistic predictions may be drawn from the two accounts of selective learning just evoked. First, consistent with the first (inference-based) account, children with ASD should not be able to use social cues in a flexible way in word-learning situations; thus, they should display no preference for word–object mappings offered by a previously accurate speaker and would perform at chance in “classic” tasks on selective trust. Second, consistent with the second (associative surface trait attribution) account, children with ASD may display selective social learning, suggesting that it does not depend on the kind of pragmatic and mentalizing skills whose deficits are otherwise attested in this population.

As seen above, a conservative interpretation of selective learning is that mental state reasoning is not required in all word-learning situations, so that young children may avoid learning from inaccurate speakers through a less sophisticated associative mechanism. Therefore, we explored children’s performance in two different word-learning situations. In a first task, we used a set of short videos based on a classic selective trust experiment with two speakers, one of whom applies familiar nouns to correct objects and the other of whom does so systematically to incorrect objects. In the second task, we used a scenario similar to that of Einav and Robinson (2011), which requires reasoning about speakers’ previous behavior. In this scenario, both speakers give correct answers, but the reliability of the speakers can be assessed through the way these accurate responses were achieved; in familiarization trials, one of the speakers is systematically assisted by a third party, whereas the other speaker always makes her lexical choices by herself. In this scenario, to evaluate the degree of speakers’ potential accuracy, children must build a model of speakers’ epistemic status based on the interaction that these speakers previously had with the third party. Therefore, sensitivity to speakers’ competences in such a task cannot be explained by the surface trait attribution mechanism. Moreover, an above chance performance in this task would strongly suggest a capacity of genuine social-cognitive reasoning. We predicted, therefore, that children with ASD, but not TD children, would fail to display selective learning in this complex situation.

We also conducted an exploratory study of how children allocate attention during learning. We reasoned that children might differentiate between speakers before they make their lexical decisions by diverting their attention away from the demonstration made by the unreliable speaker. To this end, we compared gaze fixation patterns associated with conflicting mappings made by each speaker during test trials in two tasks. We predicted that if children with ASD fail to differentiate between the two speakers in the tasks, then their patterns of gaze distribution should differ from the gaze allocation in the DLD and TD groups. In addition to examining whether children’s gaze patterns are predicted by the previously demonstrated reliability of speakers, this design provides an excellent opportunity to explore whether attention to social information—the speaker’s face—during a dynamic learning scenario is reduced in the ASD group as compared with the DLD and TD groups.

In sum, if selective social learning requires understanding mental states, then it should be impaired in children with ASD but not (or at least not to the same extent) in children with DLD and in TD children. However, one should not expect children with ASD to fail to discriminate between speakers based on their previous accuracy if such discrimination is associated with surface trait attribution. To the best of our knowledge, no study has attempted to analyze distribution of attention to faces of speakers producing conflicting information in word-learning situations. We fill this gap and reason that the patterns of attention allocation should help to understand the final lexical choice made by children in different groups. Children’s attention to the information provided by speakers may vary as a function of speakers’ previous accuracy; children should attend more to the novel object–word pairings associated with a previously accurate speaker. We also expected that the pattern of attention distribution would be associated with group performance; if children with ASD fail to use the speaker’s epistemic status in learning new words, then this failure would be associated with an atypical attention distribution pattern.