No evidence for high inflexible precision of prediction errors in autism during lexical processing

Research has shown that information processing differences associated with autism could impact on language and literacy development. This study tested an approach to autistic cognition that suggests learning occurs via prediction errors, and autistic people have very precise and inflexible predictions that result in more sensitivity to meaningless signal errors than non‐autistic readers. We used this theoretical background to investigate whether differences in prediction coding influence how orthographic (Experiment 1) and semantic information (Experiment 2) is processed by autistic readers. Experiment 1 used a lexical decision task to test whether letter position information was processed less flexibly by autistic than non‐autistic readers. Three types of letter strings: words, transposed letter and substituted letters nonwords were presented. Experiment 2 used a semantic relatedness task to test whether autistic readers processed words with high and low semantic diversity differently to non‐autistic readers. Results showed similar transposed letter and semantic diversity effects for all readers; indicating that orthographic and semantic information are processed similarly by autistic and non‐autistic readers; and therefore, differences in prediction coding were not evident for these lexical processing tasks.


INTRODUCTION
Autism Spectrum Disorder (ASD) is a developmental condition diagnostically characterized by restricted and repetitive patterns of behavior and atypical social interaction and communication (APA, 2013).This condition influences academic attainment and is often associated with literacy difficulties (Jones et al., 2009).At a group level, the reading profile of the autism community across a wide range of ages reflects a pattern of age-appropriate decoding skills coupled with reading comprehension challenges (Brown et al., 2013;Huemer & Mann, 2010).Within this general profile however, there is a wide range of individual variability in skill for different components of reading (McIntyre, Solari, Gonzales, et al., 2017;Nation et al., 2006).In this study, we aim to examine the factors that may underpin reading differences associated with autism.
There are several components that influence reading comprehension development in the population overall, such as oral language proficiency and word decoding (Daneman & Carpenter, 1980;Hoover & Gough, 1990).In the autism community however, these factors alone do not fully explain reading skill variability.Research has suggested that there are autism specific information processing differences that additionally contribute to shaping language and literacy development (e.g., McIntyre, Solari, Gonzales, et al., 2017;McIntyre, Solari, Grimm, et al., 2017;Ricketts et al., 2013).There are many theories that specify the details of these information processing differences that may be associated with autism, such as the Weak Central Coherence Theory (Frith & Happé, 1994;Happé & Frith, 2006); the Executive Function Theory (Hill, 2004;Hughes et al., 1994;Hughes & Russell, 1993;Pennington & Ozonoff, 1996); the Theory of Mind Theory (Baron-Cohen, 1989;Yirmiya et al., 1998); and the Theory of Complex Information Processing (Minshew & Goldstein, 1998).However, these theories are under-specified in the context of language processing and cannot explain the processing differences associated with reading.
A newer approach to understanding cognition in autism (e.g., Brock, 2012;Cannon et al., 2021;Lawson et al., 2014;Palmer et al., 2017;Pellicano & Burr, 2012;van de Cruys et al., 2014) proposes a well-specified mechanistic explanation of autistic cognition and provides scope to predict processing differences associated with reading.This approach is an extension of accounts of human cognition whereby mental experiences (e.g., perception, decision making) are thought to be determined by an interaction between predictions generated based on prior knowledge and incoming sensory signals (e.g., Gregory, 1980).According to this approach, learning occurs via prediction errors, which happen when incoming sensory signals contradict predictions (or expectations).The information generated by prediction errors update memory/knowledge and is used for future predictions.For example, when hearing a new pronunciation of a known word due to a regional accent, the new information needs to be included by updating the expected prior knowledge for that word.
Prediction errors, however, are not always informative.For example, in noisy/uncertain environments, the system needs flexibility to distinguish between error signals that reflect meaningful changes and errors signals that are meaningless and should be ignored.For example, if one encounters a written word that is very similar to a known word, but with a slight variation in spelling (e.g., color instead of color), this could a) reflect an unknown spelling variation of that word (e.g., the difference between USA and UK spelling conventions), resulting in learning a new word spelling and updating this information in memory; or b) be a meaningless case (e.g., a typo) that should be ignored, and so, our knowledge and expectations about the word will remain the same (color).
The most recent and compelling version of this approach posit that the ability to distinguish between meaningful and meaningless error signals is the key difference between autistic and non-autistic cognition.In situations where non-autistic people might perceive error signals as meaningless, autistic people perceive these as meaningful because they make very precise predictions and are less flexible to accept meaningless errors; this is known as High Inflexible Precision of Prediction Errors in Autism (HIPPEA; see van de Cruys et al., 2014 for a review).The impact of HIPPEA on behavior and cognition will vary depending on the expected volatility of uncertainty of a situation.For example, it is expected that autistic people will perform extremely well in circumstances where their precise predictions with very little flexibility are appropriate and match the stimulus in question (e.g., rote learning, error detection), but they will have difficulties and challenges when the stimulus in question is volatile and some generalization is required (e.g., social scenarios).Therefore, very precise predictions in these volatile circumstances would lead to many violations (prediction errors).Autistic people may categorize these violations as special cases and cause new (over) learning to occur (as opposed to discounting prediction errors as meaningless).
In the current study, we tested HIPPEA's theoretical framework in the context of lexical identification, which is the 'backbone' of reading comprehension (e.g., Perfetti & Stafura, 2014;Rayner et al., 2001).Lexical identification consists of identifying the letters that occur in a specific order within a string (orthography) and mapping these letters to their associated sounds (phonology) to retrieve the meaning (semantics) of that word.The faster lexical identification occurs, the more efficient and effortless reading becomes (e.g., Perfetti, 2007).We will examine whether orthographic or semantic information are processed differently by autistic compared to non-autistic readers, given the potential differences in prediction error coding suggested by HIPPEA.

ORTHOGRAPHY: LETTER IDENTITY AND POSITION INFORMATION
Orthographic representations are typically considered to be stable and concrete (e.g., a single correct spelling for each word).Letter identity and position information is essential to distinguish between very similar words (e.g., "calm" and "clam").Recently, some models of isolated word recognition assume that letter identity and position information are encoded independently via a flexible mechanism (e.g., Davis, 2010;Gomez et al., 2008;Grainger & Ziegler, 2011;Whitney, 2001).This flexible letter position encoding mechanism is supported by empirical evidence based on the transposed letter effect (see Frost, 2012 for a review).For example, in lexical decision tasks (where participants must decide if a letter string is a real word or not), transposed letter nonwords (jugde) are incorrectly identified as words (judge) more often and have longer response times compared to substituted letter nonwords (junpe) (e.g., Acha & Perea, 2008;Perea & Lupker, 2003, 2004).This effect has been taken as evidence for flexible letter position encoding in non-autistic readers.
In this context, we will test the HIPPEA approach to investigate whether autistic readers encode letter identity and position information in a similar way as non-autistic readers.In the context of prediction coding, a transposed letter nonword would reflect some meaningless signal error that non-autistic readers ignore, so that the existing orthographic representation can be activated, and the base word identified.Following HIPPEA, autistic readers should show a higher sensitivity to meaningless signal errors, and so, letter-position encoding should be less flexible.We therefore predict no differences between transposed and substituted letter nonword processing for autistic readers.In Experiment 1, we use a lexical decision task to test this.Three types of letter strings will be presented: words, transposed letter (TL) nonwords and substituted letters (SL) nonwords.First, we predict that all readers will be faster and more accurate to identify words than to reject nonwords; this is known as the lexicality effect (Reicher, 1969;Taft, 1986;Wheeler, 1970).Second, we predict a TL effect such that TL nonwords will have longer response times and more errors (incorrectly identifying these as words) relative to SL nonwords.This effect however should be larger for nonautistic than autistic people because autistic people will have very precise expectations on the spelling on the correct word.

SEMANTIC INFORMATION: WORD MEANING AND SEMANTIC DIVERSITY
In contrast to orthographic representations, semantic representations are often considered to be abstract, with words in the English language having a degree of polysemy (see Rodd, 2020).The extent with which the meaning of a word may vary has been conceptualized as semantic diversity.Semantically diverse words can appear in many different contexts with subtle variations in the 'sense' of meaning (e.g., chance), whereas words with low semantic diversity (e.g., carrot) will appear in a very restricted range of contexts and topics (Hoffman et al., 2013).Hoffman and Woollams (2015) studied the impact of semantic diversity on lexical identification in non-autistic readers using both lexical decision and semantic relatedness tasks.They found that high semantic diversity words facilitated word identification during a lexical decision task, but readers needed more time to decide that two words were semantically related when they were high compared to low semantic diversity.These findings suggest that the exposure to diverse linguistic contexts can enhance the quality of semantic representations (Nation, 2017), resulting in faster lexical processing, but adding noise to its semantic representation.According to HIPPEA, however, the accumulation of diverse experiences for a single concept should negatively alter the learning process for autistic readers and difficulties should be experienced in ambiguous situations (van de Cruys et al., 2014).
Therefore, we will test HIPPEA's approach to investigate whether semantically diverse information is processed similarly by autistic and non-autistic readers.Non-autistic readers would add flexibility, nuance and richness to an existing semantic representation when encountering a known word in new context (Nation, 2017).Autistic readers, however, should create a separate semantic representation for each newly experienced instance of a semantically diverse word.In Experiment 2, we use a semantic relatedness task to test this.High and low semantic diverse words were paired with semantic related (scarf-shawl) or unrelated (ocean-shawl) words.Readers will have to decide if the two words are related in meaning or not.First, we predict that high semantic diversity words will have longer response times and more errors than low semantic diversity words.This effect should be larger for non-autistic than autistic readers, because, according to HIPPEA, autistic readers would have individual memory traces for each experience of a semantically diverse word, resulting in less interference from semantic diversity.

METHOD
All materials, data, and analysis scripts for this study can be viewed on the Open Science Framework (OSF): https://osf.io/avfyg/

PARTICIPANTS
We aimed to recruit a minimum of 40 autistic and 40 non-autistic participants, which along with the stimuli set size, were selected to be in line with Brysbaert and Stevens' (2018) recommendations for sufficient power in word recognition experiments (at least 1600 observations per condition) and guidance on the minimum number of participants required to have 0.9 power in each of our experiments (Westfall et al., 2014).Autistic participants were recruited via the Autistica Discover Network and non-autistic people were recruited from the local community.Participants who completed the study were given the option to enter a prize draw for a £50 voucher.
All autistic participants reported that they had written evidence of a formal, clinical diagnosis of an autism spectrum disorder (27 Asperger's syndrome, 23 autism spectrum disorder/condition) assigned by qualified professionals (predominantly clinical psychologists).Many were also able to report that standardized diagnostic assessments were used in their diagnosis (e.g., the Autism Diagnostic Observation Schedule, the Autism Diagnostic Interview, the Adult Asperger Assessment, and the Diagnostic Interview for Social and Communication Disorders).On average, the autistic group reported receiving their diagnosis in adulthood (M = 41.66 years, SD = 14.72,Range = 12-67 years).The benefit of working with an adult sample is that reduces the variability in language skill associated with language onset.

MATERIALS AND PROCEDURE
Participants completed all tasks online in the following order: demographic questions; lexical decision task; Autism Spectrum Quotient 10 ( Allison et al., 2012); semantic relatedness task; reading efficiency assessment, and International Cognitive Ability Resource À16 (Condon & Revelle, 2014;Dworak et al., 2021), with the exception that the order of the lexical decision and semantic-relatedness tasks were counterbalanced.Both word tasks were developed using PsychoPy (Peirce et al., 2019.1) and hosted on Pavlovia (pavlovia.org).All other tasks were developed and presented via Qualtrics (Qualtrics, Provo, UT).Further details of each task are provided below.In total, the entire study took participants approximately 1 h to complete.

Design
A mixed 2 (group: autistic vs. non-autistic) Â 3 (type: TL non-word vs. SL non-word vs. word) design was adopted, with group as a between participant factor and word type as a within participant factor.Note that this is an unbalance design, in which only nonwords contained the letter manipulation (either transposed or substituted).
Stimuli 160 base words were selected from the English Lexicon project (Balota et al., 2007).Each word was a high frequency (Zipf, a logarithmic standardized frequency scale, M = 4.12, SD = 0.49, Range = 3.50-5.56)monomorphemic noun between five and seven letters in length (M = 6.19,SD = 0.69), with few orthographic neighbors (the number of words that share all the letters in the same position except one, M = 0.61, SD = 0.73, Range = 0-2).The second and third letter of each word was manipulated to create a TL nonword and a SL nonword condition.For TL-nonwords, the position of the second and third letter of each word was reversed.For SL-nonwords, the second and third letters were replaced with letters that had similar visual properties (e.g., ascenders with ascenders, descenders with descenders).Bigram frequency, the number of times that a pair of letters in a specific position occurs, did not differ between TL-nonwords (M = 543.10,SD = 914.47)and SL-nonwords (M = 539.24,SD = 1414.83);t (272.11)= 0.03, p = 0.977 (Baayen et al., 1995).Stimuli were counterbalanced across four lists using a Latin square design, that assured each list contained a single version of each stimulus in the form of 80 words and 80 non-words (40 TL, 40 SL).Each participant viewed a single list.

Procedure
Participants were instructed to identify as quickly and accurately as possible, whether a target was a real English word ("Yes" response) or not ("No" response) using a key press.Each trial consisted of a fixation cross (500 ms) followed by a target word (maximum display time 2000 ms).Trials were presented in a pseudorandomised order that assured no condition was viewed more than 3 times consecutively.Prior to completing the 160 experimental trials, participants completed 8 practice trials with feedback on response accuracy.Participants were offered a break halfway through the experimental trials.

Stimuli
Word pairs were adopted and adapted from Hoffman and Woollam's (Hoffman & Woollams, 2015) semantic relatedness judgment task (Experiment 2).160 target words were selected, half of which were of high semD (M = 1.84,SD = 0.09), and half of which had low semD (M = 1.39,SD = 0.17) (Balota et al., 2007).Target words were each paired with two words; one that was related in meaning (e.g., sprint-jog) and one that was unrelated in meaning (e.g., coffin-jog).Word pairs were counterbalanced across two lists using a Latin square design that assured each list contained all 160 target words either in their related (80) or unrelated pair (80).Each participant viewed a single list.

Procedure
Participants were instructed to identify as quickly and accurately as possible, whether the meaning of two words was related or unrelated using a key response.Each trial consisted of a fixation cross (500 ms), followed by the first word in a pair (1000 ms), and then the target word (maximum display time 3000 ms).Trials were presented in random order and participants made a response upon viewing the target (second) word.Prior to completing the 160 experimental trials, participants completed 8 practice trials with feedback on response accuracy.Participants were offered a break halfway through experimental trials.

COGNITIVE ASSESSMENTS
Autism spectrum quotient 10 (AQ-10) The AQ-10 ( Allison et al., 2012) is a 10-item scale designed to measure self-reported levels of autistic traits.The scale involves participants indicating the extent that 10 statements relating to attention to detail, attention switching, communication, imagination, and social interactions, apply to them using a 4-point Likert scale (strongly agree-strongly disagree).Scores range from 0 to 10, with a cut off score of 6 or above being considered indicative of potentially clinical levels of autistic traits, with good internal consistency (Cronbach's α = 0.85), sensitivity (0.88), and specificity (0.91; Allison et al., 2012).

Reading efficiency assessment
Reading skill was assessed by asking participants to read four passages from literary texts (Alice and Wonderland, The Portrait of Dorian Gray, The Time Machine, and the Princess of Mars) with an average length of 524 words (SD = 13).Reading time for each passage was recorded.Following each passage, participants responded to 6 multiple choice comprehension questions with 4 answer choices that probed understanding of detailed aspects of each passage.Effective reading speed scores are calculated by multiplying the proportion of questions participants answered correctly by reading speed, with higher scores being indicative of higher reading skill (e.g., Jackson & McClelland, 1979;Patching & Jordan, 2005a, 2005b).

International cognitive ability resource (ICAR-16)
General cognitive ability was measured using the 16-item ICAR sample test (Condon & Revelle, 2014;Dworak et al., 2021).This task requires participants to respond to four multiple choice verbal reasoning, letter-number sequences, matrix reasoning, and three-dimensional rotation problems.Scores range from 0 to 16, with higher scores being indicative of higher cognitive ability, with prior research reporting good internal consistency (Cronbach's α = 0.81; Condon & Revelle, 2014).Performance on the ICAR-16 has been demonstrated to be strongly associated with full scale IQ as estimated by standardized assessments and education-based measures of achievement (Condon & Revelle, 2014;Young & Keith, 2020).For the current study, items were presented in a set random order and a maximum of 1 min and 30 s was provided for participants to respond to each item.

DATA ANALYSES
For both word processing tasks, linear mixed effect models were adopted, using the lme4 package (Bates et al., 2015) to analyze accuracy (binomial) and RT (linear) data.For each model, participant group and the stimuli condition/s were included as categorical fixed effects, with sliding contrasts specified using the MASS package (Venables & Ripley, 2002).Age was included in each model as a continuous (centred) fixed effect, to act as a co-variate and account for the age difference between the participant groups.For these fixed effects, p-values were calculated using Satterthwaites approximations via the lmerTest package (Kuznetsova et al., 2017).Initially, the full random structure with crossed random effects for both participants and items was included for each model (Barr et al., 2013).However, this approach resulted in models failing to converge.Therefore, for each model, parameters in the random structure were reduced one by one, beginning with interaction effects, until convergence occurred.For each model, we also compared whether a better fit was achieved with the age covariate specified as a simple main effect or as an interactive effect.For all models, no difference was found, and therefore, models including age as a simple main effect, which were more parsimonious, are reported below.The final models specified for each measure can be viewed in the Tables below and the steps taken to reach these models can be viewed in the analysis script within the OSF repository.For brevity, only reliable effects are reported in the text.

RESULTS
All analyses were conducted in R version 1.2.1578 (R Core Team, 2019).

Cognitive assessments
Autistic participants scored significantly higher on the AQ-10 relative to non-autistic participants; t (87.06) = 8.60, p < 0.001, but did not differ in their reading comprehension accuracy; t (85.78) = 0.19, p = 0.849, reading efficiency; t (58.57) = 0.88, p = 0.385, or ICAR-16 scores; t (87.84) = 1.13, p = 0.261.This indicates that the autistic and non-autistic participants had comparable reading skill and general cognitive ability, yet participants with a diagnosis of autism reported higher levels of autistic traits relative to non-autistic participants.For descriptive statistics of each cognitive measure, please see Table 1.
EXPERIMENT 1: LEXICAL DECISION TASK Lexicality effect (word vs. nonword) The first step of the lexical decision analysis was focused on the lexicality effect (word vs. non-words) and as such, the data for non-words was collapsed across TL and SL nonword conditions.For accuracy, a significant word effect was found, with accuracy rates being higher for words relative to non-words.This was qualified by a group Â word type interaction, that indicated the autistic participants had higher accuracy for words relative to non-autistic participants.For reaction times (correct responses), there was also a word effect, with responses being significantly faster for words relative to non-words.An effect of age was detected, with response times increasing with participant age.Descriptive statistics and full model output can be viewed in Tables 2 and 3 respectively.
Transposed letter effect (TL-nonword vs. SLnonword) The second step of the lexical decision analysis was restricted to non-words and examined the transposed letter effect (TL-nonword vs. SL-nonword).For accuracy, a transposed letter effect was observed, with participants being less accurate at detecting TL nonwords relative to SL nonwords.For reaction time, there was also evidence of a transposed letter effect, with responses being faster for SL nonwords relative to TL nonwords.An effect of age was detected, with response times increasing with participant age.Model output and descriptive statistics can be viewed in Tables 2 and 3 respectively.

Experiment 2: Semantic relatedness task
For accuracy, a significant effect of relatedness and semD was detected.These effects reflect participants having higher accuracy in response to unrelated word pairs relative to related word pairs, and higher accuracy for low semD words relative to high semD words.The main effect of relatedness was qualified by a group Â T A B L E 1 Mean (M), standard deviation (SD), and range for AQ-10, comprehension accuracy, reading efficiency and ICAR16 scores across autistic and non-autistic groups.Note: Three non-autistic participants had very short reading times, that resulted in extremely high efficiency scores (>18,000).When these participants are removed from the dataset, there is still no difference between autistic and non-autistic participant group reading efficiency; t (81.77) = 0.59, p = 0.559, non-autistic M = 6997, SD = 3111, Range = 2758-15,686.

Group
T A B L E 2 Means (standard deviations) for each condition in Experiment 1.

Accuracy (proportion correct)
Reaction times (ms) relatedness interaction, that indicated autistic participants were less accurate for related word pairs, relative to non-autistic participants.A similar semD and relatedness effect was found for reaction times (correct responses), with participants being faster to respond to low semantic diversity words relative to high semantic diversity words, and faster to respond to related word pairs relative to unrelated word pairs.See Tables 4 and 5 for model output and descriptive statistics respectively.

DISCUSSION
This research aimed to examine whether orthographic and semantic information is processed similarly in autistic compared to non-autistic readers during lexical processing.We adopted the HIPPEA approach of autism and tested whether autistic readers have less flexible letter position encoding (Experiment 1) or store variances in the meaning of words as separate lexical entries (Experiment 2) compared to non-autistic readers.We found no evidence for either of these predictions within our data.We summarize the findings and implications below.
Experiment 1: Orthographic processing Both autistic and non-autistic participants demonstrated a typical lexicality effect in accuracy and reaction times, reflecting words being classified more accurately and quickly than nonwords.This replicates previous research (Reicher, 1969;Taft, 1986;Wheeler, 1970) and extends it to autistic readers.In addition, an unpredicted effect was an interaction that indicated autistic people were more accurate at classifying words relative to non-autistic people.We speculate that this finding could be interpreted as a hint towards autistic people having more precise lexical representations overall; but also, as evidence towards the idea that word-like stimuli activate existing representations to a greater extent in non-autistic than autistic readers (see Knowland et al., 2022 for a similar finding during word learning in children).This finding, however, should be taken cautiously as we did not have any measurement of vocabulary knowledge, or other higher-order processes that could explain this effect.Contrary to our predictions, autistic and non-autistic readers showed a transposed letter effect in accuracy and reaction times: TL non-words were classified less accurately and more slowly than SL non-words.This replicates previous research with non-autistic samples (e.g., Acha & Perea, 2008;Perea & Lupker, 2003, 2004) and suggests that autistic readers also encode letter position information flexibly.Overall, these findings add to the literature that has used different methods and manipulations and indicate orthographic processing does not differ between autistic and non-autistic people (e.g., Micai et al., 2019;Saldaña et al., 2009;Speirs et al., 2011).

Experiment 2: Semantic processing
Both autistic and non-autistic participants demonstrated a typical semantic diversity effect in accuracy and RTs.These effects reflected participants being more accurate and faster to respond to low semD word pairs, relative to high semD word pairs.This replicates previous research with non-autistic samples (e.g., Hoffman & Woollams, 2015) and demonstrates that autistic people are similarly sensitive to semD, and experience interference in accessing and assessing word meaning when encountering words that appear in a variety of contexts.
HIPPEA indicated that when working with uncertain stimuli/environments, which we interpreted to be the case for semantic information that is often variable, that autistic people may store new experiences as new representations, as opposed to adding nuance and generalization to existing representations.If this was the case for lexical-semantic representations, we predicted a reduced or absent semD effect for autistic people.In contrast to this prediction, similar interference from high semD words were observed, which provides no evidence for differences in how semantic representations are stored for high semD words between autistic and non-autistic people.We can therefore assume that both groups have similarly organized lexicons, with a single representation for high semD words, that contains information on the variety of contexts these might appear within and how this might alter the sense of the meaning.
Although contrary to the predictions generated by HIPPEA, this data adds further evidence to the literature that suggests autistic and non-autistic people are similarly sensitive to linguistic context and how this might influence word meaning (e.g., Brock et al., 2008;Brock et al., 2017;Hahn et al., 2015;Hala et al., 2007;Henderson et al., 2011;Norbury, 2005), but using a more subtle manipulation of word meaning ambiguity.There was an interaction between group and relatedness, that indicated autistic people were slightly less accurate when making judgments about related word pairs relative to unrelated word pairs.This does not pertain to any of our hypotheses, but we can speculate that this may perhaps be indicative of autistic people having more stringent criteria for what word meanings are judged to be related due to more precise lexical representations or maybe as a consequence of their differences in vocabulary (e.g., see Bishop, 1993 for a review).However, as we mentioned earlier, we should be cautious in this interpretation because we did not have any measurement of vocabulary or higher-processing knowledge to explain these group differences.

Implications and limitations
The autistic and non-autistic people who took part in these experiments performed extremely similarly, suggesting that differences in prediction error coding (HIPPEA approach) are not associated with how autistic readers process orthographic and semantic information during lexical processing.There could be several reasons why we have not detected lexical processing differences as predicted by the HIPPEA approach.
First, in this study we have used artificial tasks were a single word or word pairs were presented in isolation.Readers needed to either make a lexical or a judgmental decision relating to these words.It could be that psychological processes involved in these tasks are not the same as those involved in more natural reading tasks and that the expectation of the task itself influenced prediction errors (e.g., an expectation to encounter nonwords).
Second, orthographic, and semantic information is processed at very early stages during lexical processing.
Word identification is very rapid and is extremely efficient in skilled adult readers.It is therefore possible that any differences predicted by HIPPEA might not be evident for adult, skilled autistic readers recruited in this study.Note that no differences were detected between reading skill of our autistic and non-autistic groups; and autistic readers were diagnosed in their adulthood, representing a subset of the autism community It could be, therefore, that HIPPEA for lexical processing may still apply to developing readers, or people who experience reading challenges or a delay in language abilities.Third, it is also possible that our semantic information, although varied and uncertain at times, is concrete enough that the processing of this information will not result in any differences being evident between autistic and non-autistic people.For example, although word meaning and contexts vary, there is still some systematicity and limit to this.It is also worth highlighting that we manipulated semantic diversity, which reflects the number of contexts a word may appear in.Although this is related to polysemy, it is not a direct reflection of it, hence it could have weakened the possibility of observing any differences in the autism community.
In summary, although autistic and non-autistic readers did not show any differences in the direction predicted by the HIPPEA approach, we observed some differences between them; specifically, autistic readers were more accurate to identify words and, slightly less accurate to decide whether two words were semantically related or not than non-autistic readers.More research is needed to understand whether these differences can be explained by HIPPEA or if they could be the result of a different way of processing, storing and/or retrieving semantic information.

CONCLUSION
This research demonstrates that autistic adults show typical transposed letter and semantic diversity effects in isolated word recognition tasks, suggesting that the mechanisms to encode letter position information and process semantically diverse words are similar to nonautistic readers.More research is needed to investigate whether reading differences frequently associated with autism (e.g., Brown et al., 2013;Huemer & Mann, 2010) can be related to differences in prediction coding using different materials, tasks and a more representative sample of the autism community.
Mixed effects model parameters for each accuracy and reaction times in Experiment 1. Mixed effects model parameters for accuracy and reaction times in Experiment 2.
T A B L E 3