In this paper, I review studies that have used a number of different methodologies to investigate comprehension of spoken and written language in autism. Difficulties understanding the meaning of spoken and written language are thought to be characteristic of autism. However, language comprehension skills vary widely between individuals and would appear to be closely linked to other core language skills. Moreover, traditional tasks used to assess language comprehension may prove difficult for people with autism for a number of reasons and thus underestimate true levels of comprehension.
Language comprehension in autism
Impaired communication is one of the defining features of autism, but there is enormous variation in the nature and degree of communication difficulties across the autism spectrum. At present, there is a broad consensus that the use of language in social situations (i.e., pragmatics) is affected in all individuals with an autism spectrum disorder (ASD), but there is huge individual variation in the ability to process speech sounds (phonology); to combine words to make grammatical sentences (syntax); and to mark words to indicate the tense of verbs or plurals of nouns (morphology).
The focus of this paper is semantic processing; that is the knowledge and appreciation of word and sentence meaning. It has long been recognised that many people with autism have difficulty in processing semantic information. Indeed, impairments of semantic processing lie at the heart of Uta Frith’s influential ‘weak central coherence’ account of autism (Frith, 1989), which states that autism is characterised by “the inability to draw together information so as to derive coherent and meaningful ideas”. Weak central coherence remains, however, a somewhat nebulous concept and the precise nature of any semantic impairment is still unspecified. Moreover, recent studies have indicated that semantic processing difficulties may be related to an individual’s level of general language ability rather than being a generic feature of autism.
One of the major difficulties in making sense of the wealth of relevant data lies in the fact that a wide range of tasks have been used across studies and performance on these different tasks can be affected by various extraneous factors. The aim of this paper, therefore, is to review the available evidence from studies of semantic processing in autism across a number of different paradigms, considering the conclusions that can be drawn as well as any methodological concerns and confounds. I will also describe a recent study of language comprehension in autism that I conducted with my colleagues, Courtenay Norbury, Shiri Einav, and Kate Nation, in which we used eye-tracking as a means of avoiding some of the concerns affecting earlier studies (Brock et al., 2008). Finally, I will consider the implications for assessment of language in individuals with autism, for theories of autism, and directions for future research.
The idea that individuals with autism may have some form of semantic impairment was first given serious consideration by Hermelin and O’Connor (1967), who gave an immediate memory test to a group of children with autism and a control group of non-autistic children with severe intellectual delay. They found that children in the control group were considerably better at recalling meaningful sentences such as “The fish swims in the pond” than they were at recalling strings of unrelated word such as “By is go tree stroke lets”. Children with autism, however, failed to show a significant advantage for meaningful material, suggesting that they were not attending to the meanings of the words.
A number of subsequent studies have adopted a similar approach, comparing memory for lists of related words (e.g., animals) with memory for lists of unrelated words. However, results have been somewhat mixed. In some studies, individuals with autism have failed to show any benefit for recalling semantically related words (Bowler et al., 1996; Tager-Flusberg 1991), while others have demonstrated entirely normal semantic facilitation (e.g., Fyffe & Prior, 1978; Lopez & Leekam 2003; Ramondo & Milech 1984). It has also been suggested that, where group differences do exist, they may reflect differences in strategies adopted to aid task completion rather than actual deficits in semantic processing or an insensitivity to semantic associations (see Gaigg et al., 2008; Schwartz, 1981).
A common means of investigating semantic associations is via a priming task. Semantic priming occurs when processing of a stimulus is made easier or quicker when it preceded by a semantically related stimulus. For example, people react quicker to the word “dog” when it follows “cat” than when it follows “hat”. Priming effects rely on the ability to make associations between words, which in turn depend on first accessing the meaning of those words. If individuals with autism fail to access meaning or fail to make the semantic association then priming effects should be reduced.
Most published studies have in fact reported normal semantic priming effects. Lopez and Leekam (2003) found that, just like typically developing control children, those with autism read words more rapidly if they were preceded by an appropriate context word. For instance, they were quicker to read the word “jug” if it followed the word “kitchen” rather than the word “garden”. Similar results were reported by Kamio & Toichi (2000; Toichi & Kamio, 2001), who used a fragment-completion task. Participants were asked to read aloud a word with some of the characters missing. Individuals with and without autism were more likely to provide the correct response if the word-fragment was preceded by a semantically related prime word. More recently, Kamio et al. (2007) utilised a lexical decision task in which participants had to decide whether written words were real or made-up. Control children were quicker to respond to real words when they were preceded by a semantically related word, but this effect was not significant in the autism group. This contradicts other findings, but it is important to note that there were only eleven participants in the autism group and there was huge individual variation in their reaction times. Moreover, the authors were only able to use a non-parametric statistical test, which has relatively little statistical power to detect a priming effect.
In sum, the balance of evidence suggests that people with autism show normal semantic priming and therefore do access word meanings. Having said that, all of the above-mentioned studies have tested individuals with relatively high verbal IQs and none have looked at individual variation in performance. It remains theoretically possible, therefore, that some people with autism do show reduced semantic priming.
Another well-researched theme in early as well as contemporary accounts of autism has been the contrast between relatively good text-decoding skills (i.e., reading aloud) and poor reading comprehension. In his seminal description of autism, Kanner (1943) noted that “reading skill is acquired quickly, but the children read monotonously, and a story… is experienced in unrelated portions rather than its coherent totality”. Subsequent studies have broadly supported these observations. Of particular note are studies using the Neale Analysis of Reading Ability – a standardised test in which participants are scored for the accuracy in reading a short story aloud and then their ability to answer questions about the story. Age equivalent (mental age) scores for children with autism are typically lower for reading comprehension than for reading accuracy (Frith & Snowling, 1983; Lockyer & Rutter, 1969; Nation et al., 2006; Rutter & Bartak, 1973). Indeed, Nation et al (2006) found that comprehension was poorer than accuracy for virtually all their sample of children with autism. Notably, those with the lowest reading comprehension scores also had the poorest receptive vocabulary scores, indicating a link between reading comprehension and more general language abilities.
Jolliffe and Baron-Cohen (1999) investigated the ability to make so-called ‘bridging inferences’ about events that are implied but not explicitly stated in the text. They asked participants to read pairs of sentences such as “George left his bathwater running. George cleared up the mess in the bathroom.” They were then given a multiple choice question in which participants had to choose the sentence that best completed the story (i.e., “The bath had overflowed”). High-functioning adults with autism or Asperger syndrome were less likely to choose the correct answer than non-autistic adults, suggesting that they had difficulty integrating the two components of the story to make the inference.
The concern with all of these reading comprehension tests, however, is that the tasks are each quite demanding in a number of ways. Potentially, an individual with autism could have perfectly good understanding of the text, but have difficulty answering the questions because they don’t understand what they have to do, have difficulty formulating an answer, or have forgotten the correct response by the time they get to the question. These issues are highlighted by a recent study in which Saldaña and Frith (2007) used an indirect measure of the ability to make bridging inferences. Participants were asked to read short stories followed by a question. The critical measure was not the answer to the question but the time it took to read the question. Reading times were shorter when the question was primed by an inference that could be derived from the story. For example, children were quicker to read the question “Can rocks be large?” when it followed the story “The Indians pushed the rocks off the cliff onto the cowboys. The cowboys were badly injured”. The story implies that the rocks were large enough to cause serious injury, thereby providing the answer to the question. Contrary to predictions, children with and without autism showed similar reductions in reading times, indicating that both groups had made the appropriate inferences. Notably, the children with autism still performed poorly on a standardized reading comprehension test, emphasising the point that poor performance on traditional reading comprehension tests may underestimate the true level of comprehension.
The weak central coherence account emphasises difficulties in deriving ‘contextualised meaning’. The most widely cited evidence for linguistic weak central coherence comes from studies looking at the use of context to derive the appropriate meanings of homographs. These are words with different meanings that are written the same, so the correct meaning can only be derived by considering the sentence context. Crucially, the homographs used in studies of autism also have different pronunciations associated with their two meanings. For example, the word “tear” is pronounced differently in the sentences “In her eye was a tear” and “In her dress was a tear”. Thus, the extent to which somebody correctly reads the homographs aloud indicates their ability to access the contextually appropriate meaning. Importantly, the reader’s attention is not explicitly drawn to the ambiguity so, as with Saldana and Frith’s inferencing task described above, homograph-reading arguably taps into natural ongoing reading comprehension processes in a way that more conventional tests do not. Frith and Snowling (1983) found that children with autism tended to produce the most common pronunciation of the homographs, regardless of contextual cues. By contrast, typically developing children and children with dyslexia performed much better, indicating that they had made use of the contextual information.
The basic finding that, on average, people with autism make more errors than control participants on the homographs task has been replicated on numerous occasions, both with adults and with children. It is important to note, however, that many people with autism appear to have absolutely no trouble with this task. Three reports (Happé, 1997; Jolliffe & Baron-Cohen 1999; Lopez & Leekam, 2003) with a combined total of 65 participants with autism or Asperger syndrome provide sufficient information to allow individual scores to be derived. The critical condition is when the correct pronunciation of the homograph is the less common one and so the participant has to rely on the preceding sentence context to arrive at the appropriate response (i.e., if they give the most common pronunciation they will be scored incorrect). Across these three studies, 40% of participants with ASD gave the correct response on every single trial and a further 29% made only one error. These are, admittedly, considerably lower proportions than observed for control participants, but the point stands that consistent reports of group differences should not be mistaken for consistent performance at the individual level.
Of further concern is that, while the homographs task avoids some of the pitfalls associated with conventional reading comprehension tests, performance still relies on familiarity with both meanings of the word and the contextual information as well as an understanding that words can have multiple meanings (see Lopez & Leekam, 2003). Snowling and Frith (1986) attempted to address these issues by giving participants training; familiarizing them with both meanings of the homographs, and then re-testing them. In contrast to other studies mentioned above, they found no differences between the performance of children with and without autism. Unfortunately, they only reported the average of performance before and after training, so it is unclear what effect training actually had for each group. Nevertheless, it is interesting to note that children with lower language scores performed relatively poorly, regardless of their autism diagnosis, again highlighting the importance of general language abilities for explaining task performance.
The majority of studies reviewed thus far have, in one way or another, looked at comprehension of written material. However, the assumption is that comprehension difficulties should equally affect spoken material. Two recent studies have, therefore, investigated comprehension of spoken sentences containing homophones – words that sound the same but mean different things. As with homographs, correct interpretation relies on consideration of the context in which they occur, so results should in theory parallel those of homograph-reading studies.
In the first of these studies, Hoy et al. (2004) asked participants to listen to sentences containing a homophone followed by a second sentence that provided disambiguating information. For example “The boy wanted to go to the beech. He wanted to climb in its branches.” They were then asked which of four pictures went with the story – options included pictures related to both meanings of the homophone (e.g., a tree and a seaside picture). Children with autism performed worse than typically developing control children on this task. However, their language scores were also considerably lower, so it is unclear whether their poor performance was a function of their autism diagnosis or their language difficulties.
This issue was addressed in a subsequent study by Norbury (2005) using a similar homophone-comprehension task. Participants heard a single sentence containing a homophone (e.g., “John stole from the bank”). They then saw a single picture corresponding to one of the meanings of the homophone and had to decide whether it matched the sentence. Children with autism performed just as well as non-autistic children with similar language abilities. However, when participants were divided up according to their language ability, it was found that, regardless of their autism diagnosis, those with poorer language skills made more errors when the sentence context was biased towards the less common meaning of the homophone. For example, on hearing “John fished from the bank”, they were less likely to say that it matched a picture of a river and more likely to say that it matched a picture of some money.
These findings suggest that difficulties in processing semantic context are not ubiquitous in autism, as widely argued, but are instead related to an individual’s level of language ability. Having said that, the homophones task faces many of the concerns that were raised against the homographs task. Performance depends not only on the ability to use contextual information provided in the sentence but also knowledge of the different meanings of the homophones and the world-knowledge required to determine whether a particular interpretation of a homophone is sensible. Potentially, any of these factors could contribute to poor performance among children with language difficulties. Moreover, the picture-matching task lacks some of the subtlety of the homographs task, asking participants directly to consider the meaning of the homophones after they have heard the whole sentence. Perhaps an indirect measure of ongoing spoken language comprehension would reveal subtle differences in semantic processing.
This brings us to our eye-tracking study of language comprehension (Brock et al., 2008). Studies have shown that spoken language can have a strong influence on eye-movements. People tend to look at objects that match the words they are hearing; they also look at objects that sound similar to those words or are semantically related. This happens quickly and automatically, so eye-movements can provide important insights into ongoing language comprehension without requiring participants to give answers to direct questions.
In our study, we simply asked participants to listen to spoken sentences while looking at a computer screen on which there were various different objects. All they had to do was press a button if and when they heard a word in the sentence that corresponded to one of the objects on the screen. This task was fairly trivial and most participants scored close to 100% correct. Our real interest was in their patterns of eye-movements as they completed the task. We tested 24 adolescents with autism who had a wide range of language skills. Our control group was made up of 24 typically developing children and non-autistic children with language impairment so the two groups were matched overall for language ability. We could then look at a number of different factors that affected participants’ eye-movements and determine whether they were related to their autism diagnosis or their language score.
First, we investigated whether participants looked at objects that were mentioned in the speech; for example, if they heard “Sam chose the hamster”, did they look at the hamster on the screen more than any other object? Reassuringly, they did, and there was no evidence that adolescents with autism were any different to non-autistic adolescents, indicating that all the participants were processing the meaning of individual words. We also investigated whether participants could anticipate objects that were predicted by the speech but hadn’t actually been heard yet. For example, if they heard “Joe stroked the hamster”, did they wait until after the word “hamster” to look at the hamster or did they look at it immediately after the word “stroked” (because it was the only strokeable object on the screen)? Such anticipatory eye-movements would indicate that participants were making semantic associations between the sentence verbs and the objects to which they were likely to refer and were aware of the semantic properties of the objects (e.g., that hamsters are strokeable items). Again, we found a strong anticipatory effect but failed to find any differences between our two groups or any association between patterns of eye-movements and language ability.
The most interesting results came from our consideration of ‘phonological competitors’. On certain trials, we replaced the target object with a different object that shared the same onset; for example, the hamster was replaced by a hammer. Participants showed a strong ‘competitor effect’. So on hearing “Sam chose the hamster”, they looked briefly at the hammer and then looked away once they had heard sufficient information to realise that the word wasn’t “hammer” after all. Again, we found no differences between adolescents with and without autism and no association with language ability.
Crucially, we then looked at the effect of sentence context on this competitor effect. In a previous study with university undergraduates as participants, we found that the competitor effect was wiped out if the competitor didn’t fit in with the sentence context. Participants would look at the hammer if they heard “Joe chose the hamster” but not if they heard “Joe stroked the hamster”. This effect of context on eye-movements is broadly analogous to the effect of context on the interpretation of ambiguous words (i.e., homographs and homophones). Spoken words unfold over time and halfway through the word “hamster”, it is ambiguous – it could be “hamster”, “hammer”, “hamburger”, or indeed any other word beginning with “ham”. It is this temporary ambiguity that ‘tricks’ people into looking at the competitor object. The results from previous studies using the homographs task predicted that adolescents with autism would be unable to use sentence context to overcome this ambiguity, so would look at the competitor object regardless of context. On the other hand, the results from Norbury’s (2005) homophones study predicted that the effect of context would be related to language ability rather than autism diagnosis. As it happened, the eye-tracking data were entirely consistent with Norbury’s findings. Overall, adolescents with autism did not differ from non-autistic adolescents, but in both groups there was a strong relationship with language ability – those with the poorest language scores tended to look at the competitor even when it was contextually inappropriate.
Having criticized various aspects of the methodology of earlier studies, it is only fair to subject our eye-tracking study to similar scrutiny. Our main concern was that the reduced effect of sentence context on the eye-movements of adolescents with language impairment may have arisen because these participants were simply less familiar with the words used in the sentences, were less motivated, or were paying less attention. However, these factors would have affected eye-movements across all the conditions. We only found an effect of language in the one condition that required participants to use the sentence context to resolve the (temporary) ambiguity. Given this, our results further undermine the weak central coherence account, but suggest that difficulties in processing semantic context on-line are related to language impairment rather than autism.
A number of key issues arise from this brief review. The first is that traditional behavioural tests of comprehension may underestimate the true capabilities of individuals with autism. People with autism often struggle when they are asked to answer questions after the event about a complex passage of prose that they have just heard or read. However, when the test demands are stripped back, indirect measures of comprehension such as semantic priming or eye-tracking can reveal a surprising degree of understanding, often in line with expectations based on general verbal abilities. Having said that, it is important to acknowledge that, at this stage, studies using such techniques have been quite limited in scope and future studies may reveal language comprehension difficulties that cannot be easily explained away. It is also important to distinguish between the kinds of ‘extraneous’ difficulties that only compromise experimental task performance and processing difficulties (such as general memory difficulties) that may also impact upon everyday communication.
The review highlights the potential of techniques such as eye-tracking for assessing language comprehension in autism (cf. Edelson et al., 2008). At present, eye-tracking technology is probably too expensive to be used routinely for assessments in most clinical or educational settings, but costs are falling rapidly and this may be a significant avenue for future developments. In the meantime, this review demonstrates the need for careful consideration of confounding factors that could detrimentally affect performance when trying to evaluate the comprehension skills of an individual with autism.
A second overarching theme coming out of this review is the heterogeneity within the autism spectrum. Whenever we look at individual differences rather than just group averages, it becomes clear that some people with ASD have difficulties on semantic processing tasks whereas others do not. Evidence from homophone-processing and eye-tracking tasks suggests that this heterogeneity in semantic processing is related to more general verbal abilities, particularly when considering the use of semantic context. In the same way that some individuals with ASD have perfectly adequate phonological and grammatical skills, so, it appears, some individuals have age-appropriate semantic skills. However, the fact that semantic and other language skills are not universally impaired in autism does not make them any less important. Clearly, many people with autism do have problems understanding the meaning of language and it is imperative to understand why they do and how best to ameliorate their difficulties.
This revised perspective on semantic impairment in autism also has important implications for theoretical accounts of autism. In particular, recent findings are clearly at odds with the weak central coherence account, according to which, difficulties in extracting meaning and processing words in context are characteristic of autism (Frith, 1989). One resolution would be to acknowledge the possible existence of subgroups of ASD with completely different underlying causes. Thus weak central coherence and semantic processing deficits may only be a characteristic of a subgroup of people on the autistic spectrum. An alternative possibility is to abandon the notion of a central cognitive mechanism and instead think in terms of neural or genetic mechanisms that act as risk factors for ‘symptoms’ associated with weak central coherence. To illustrate in more concrete terms, Brock et al. (2002) hypothesized that aspects of autism attributed to weak central coherence may reflect abnormal connections between different regions of the brain. However, genes influencing the development of neural connectivity are likely to act in a probabilistic manner – they may or may not impact upon the functioning of any particular brain system (Geschwind & Levitt, 2007) so would not always lead to impaired semantic processing or indeed any other given symptom.
These considerations also shed some light on the hypothetical relationship between autism and specific language impairment (SLI); a developmental disorder in which language acquisition is severely delayed in the absence of autism or any other obvious cause. Overlaps between SLI and autism are well-documented in relation to phonology, syntax and morphology, leading some researchers to conclude that autistic individuals who also have language difficulties effectively have both autism and SLI (e.g., Roberts et al., 2004), The evidence reviewed here also points to overlap in terms of semantic processing; in both Norbury’s homophones task and our eye-tracking study, a lack of sensitivity to semantic context was demonstrated by individuals with language impairment regardless of their autism diagnosis. That being said, Williams et al. (2008) have recently argued that similarities between autism and SLI are probably more superficial than real, so it remains to be seen whether the overlap in semantic processing holds up to further detailed investigation.
Bearing these arguments in mind, an important avenue for future research is to consider the brain mechanisms involved in semantic processing in autism. There is now a growing body of evidence from studies of autism using magnetic resonance imaging, event-related potentials and, latterly, magnetoencephalography, testifying to atypical brain responses during semantic processing tasks (e.g., Braeutigam et al., 2008; Harris et al., 2006; Ring et al., 2007). A detailed discussion of these studies is beyond the scope of this review, suffice to say that many of the criticisms raised here against behavioural-cognitive studies may also apply to these brain-imaging studies. Typically, studies consider group averages rather than individual variation and groups are often mismatched in terms of language ability, muddying any attempt to interpret group differences. Furthermore, the tasks used in these studies introduce various confounding demands that could lead to differences in brain activity unrelated to actual semantic processing.
Clearly, there is considerable work to be done before we fully understand the nature of semantic processing impairments in autism at either the cognitive or the brain level. Future research addressing the links between cognitive and neural mechanisms is likely to generate important insights into the underlying causes of communication impairment in autism. Crucially, however, the studies reviewed in this paper highlight the need to acknowledge the heterogeneity that exists within autism as well as potential overlaps with other groups of non-autistic individuals who also have communication difficulties.
Bowler, D. et al. (1997). Neuropsychologia, 35, 65–70.
Braeutigam, S. et al. (2008). European Journal of Neuroscience, 27, 1026-36
Brock, J. et al. (2002). Development and Psychopathology, 14, 209-224.
Brock, J. et al. (2008). Cognition, 108, 896-904.
Edelson, L. et al. (2008). http://imfar.confex.com/imfar/2008/techprogram/P1285.HTM
Friedmann, N. & Novogrodsky, R. (2007). Brain and Language, 101, 50-63.
Frith, U. (1989). Autism: Explaining the enigma. Oxford: Blackwell.
Frith, U., & Snowling, M. (1983). Journal of Developmental Psychology, 1, 329–342.
Fyffe, C., & Prior, M. (1978). British Journal of Psychology, 69, 393–402.
Gaigg, S. et al. (2008). Neuropsychologia, 46, 983-92.
Happé, F. (1997). British Journal of Developmental Psychology, 15, 1–12.
Happé, F. & Frith, U. (2006). Journal of Autism and Developmental Disorders, 36, 5–25.
Harris, G. et al. (2006). Brain and Cognition, 61, 54-68.
Hermelin, B. & O’Connor, N. (1967). British Journal of Psychology, 58, 213–218.
Hoy, J. A., et al. (2004). Autism, 8, 267-81.
Jarrold, C., & Brock, J. (2004). Journal of Autism and Developmental Disorders, 34, 81-86.
Jolliffe, T., & Baron-Cohen, S. (1999). Cognition, 71, 149–185.
Kamio, Y. et al. (2007). Journal of Autism and Developmental Disorders, 37, 1116-22.
Kamio, Y., & Toichi, M. (2000). Journal of Child Psychology and Psychiatry, 41, 859–867.
Kanner, L. (1943). Nervous Child, 35, 217-250.
López, B., & Leekam, S. R. (2003). Journal of Child Psychology & Psychiatry, 44, 285–300.
Nation, K. et al. (2006). Journal of Autism and Developmental Disorders, 36, 911-919.
Noens I. L., & van Berckelaer-Onnes, I. A. (2005). Journal of Communication Disorders, 38, 123-141.
Norbury, C. F. (2005) Journal of Experimental Child Psychology. 90, 142-171.
Ramondo, N., & Milech, D. (1984). British Journal of Psychology, 75, 95–103.
Ring, H. et al. (2007). Journal of Autism and Developmental Disorders, 37, 281-290.
Roberts, J. A., et al. (2004). Applied Psycholinguistics, 25, 429 – 448.
Sabisch, B., et al. (2006). Neuroreport, 17, 1511-1514.
Saldaña, D., & Frith, U. (2007). Journal of Experimental Child Psychology, 96, 310-319.
Snowling, M. J., & Frith, U. (1986). Journal of Experimental Child Psychology, 42, 392–415.
Schwartz, S. (1981). Applied Psycholinguistics, 2, 25-31.
Tager-Flusberg, H. (1991). British Journal of Developmental Psychology, 9, 417–430.
Toichi, M. & Kamio, Y. (2001). Journal of Autism and Developmental Disorders 31, 483–90.
Williams, D. et al. (2008). Psychological Bulletin, 134, 944-63.