Human existence

Bookmark and Share

The google word theory of autism

Search engines such as Google make keyword searches context sensitive with the aid of the information contained in higher-order associations

The human brain also uses this type of information to make infant and child learning sensitive to context

Autism occurs when the processing of this type of information is developmentally impaired

Why important

The scientific paper arguing this theory Autism, Context/Noncontext Information Processing, and Atypical Development has just been published in Autism Research and Treatment.

first-order associations higher-order associations
  First-order associations   Higher-order associations

Theory in a nutshell

“Street” and “Road” very rarely appear close together in webpages. As synonyms they are strongly associated—not by being in the same sentences but by having the same neighboring associations with other words. These higher-order or indirect association play an important role in the context indexing for keywords by search engines including by Google.

The information extracted from higher-order associations underlies the human brain’s capacity to process context—a capacity that is important to learning by infants and children. If this was to be impaired, it would result in context insensitivity across a wide variety of cognitions.

An inability to use context has been identified as a trait of those with autism since the condition was first identified in 1943 by Leo Kanner. But the nature of this context defect has been ill-defined.

The field of computer/cognitive science called Latent Semantic Analysis (LSA) has changed this as it shows that the processing of context requires the extraction of the information contained in higher-order associations. Originally LSA was called Latent Semantic Indexing but changed its name when it was realized that its ideas could model cogntiive development including how children learn the ability to pick up the meaning of unfamiliar words from the context provided by familiar ones (see Landauer and Dumais, Psychological Review, A solution to Plato's problem: The latent semantic analysis theory of acquisition, induction, and representation of knowledge.

LSA at present leaves one question unexplored. How would cognition develop if children were born with an impaired ability to process higher-order associations?

This disability would impair them in learning language. But not importantly not only in this. LSA researchers note higher-order associations exist not just between words but potentially between any entity processed by the brain. Thus an impairment in using the information extracted from higher-order associations could be expected to be important not just for learning unfamiliar words but the development of the use of context in diverse aspects of everyday cognition such as sociability, sensation, motor control and stress regulation.

The extraction of information from higher-order associations is computationally much more complex than that of first-order associations. As a result, much of it is likely to be processed in specialized neural areas that are separate to those used in everyday cognition. Impairment to these specialized processes could therefore be expected to impair the development of a variety of apparently unrelated cognitive faculties

What will happen to a child born with impaired higher-order specialist processes? Intelligent cognition would still be possible since the child could still develop using the information extracted from first-order ones. But its cognition would be atypcial as it would be restricted to using the information extracted only from first-order associations. As a result, the child would think, feel, perceive and act with informaton derived from direct associations in the place of where other children would use the context information gained from higher-order associations. In consequence, such children would grow-up experiencing themselves, other people and the world a manner that was literal, unadaptive and insensitive to context.

This atypicality development would affect many apparently unrelated aspects of development such as those responsible for social interaction, stress regulation, sensation, motor coordination and language

Key parts of the theory

Hiigher-order associations–what are they?

What computationally is context?

What is the relation between autism and context impairment?

What is LSA and what are its origins in search indexing?

How does the information extracted from higher-order associations revolutionize how scientists understand meaning?

How does LSA explain the language impairment of autism?

What is the role of context in everyday cognition?

How does this theory explain the need for sameness and impaired stress regulation in autism?

How does this theory explain the lack of sociability in autism?

How does this theory explain impaired theory of mind in autism?

How does this theory explain islets of preserved cogition in autism?

How does this thery explain sensory and motor impairments in autism?

How this theory link autism to impairment in the brain?

How does this theory explain the differing development of autism?

How might the Google theory of autism be researched?

Charles's Sorting Task-an overlooked means of treating autism

Some nonGoogle theory observations elsewhere on this site on autism

Methodological problems in autism research

New directions for researching autism

 

Higher-order associations–what are they?

illustrates higher- and first-order associations for words in terms of synonyms and next to this their equivalents for nonword entitiesHigher-order associations are the associations that exist between entities through indirect associations mediated via other entities.

The idea of higher-order associations can be grasped by the example of synonyms. Consider “big”, "large" and “vast”, they rarely appear directly together. People normally do not say things like “the big and vast stone broke the wagon”. If one of the words is used the second is a repetition, and so redundant. Nonetheless, the occurrence of one word contains much information about the occurrence of the other. If “big” appears frequently with another word, say, “stone”, so predictably will “vast”; if “big” does not appear with this word, then neither will “vast”. What links “big” and “vast” is not a direct association but having the same kind of direct associations to other words. After all, it is often a matter of indifference as to whether “big” or “vast” gets picked by context for use in any particular sentence to indicate large size. Such higher-order associations exist not just between “big” and “vast”, but also (with varying strength) between every other possible pair of words, and so occur in regard to every one of the tens of thousands of words used in spoken or written vocabulary. These shared associations create at a higher level, a strong but indirect link between “big” and “vast”.

Higher and even more indirect levels of co-occurrence word association also exist. Some of the words with which “big” associate, co-associate with it more or less strongly in the presence of yet further words. Thus, if the word “big” appears in a sentence with “stone”, then it is likely that “heavy” will also appear.

The existence and importance of higher-order associations has only been recently recognized with the development of computer technology, and the use algorithms that can extract the information they contain. As a result, there is a disparity between work upon cognition and first-order associations and that upon higher-order ones. A large philosophical literature exists based upon associations and their relationship with cognition, originating in the work of people such as the eighteenth century philosopher David Hume. This deals with cognition as something that arises from experienced direct associations. In contrast, the only work so far upon cognition and higher-order associations is that done by those that have explored its extraction in the field known as latent semantic analysis.

What computationally is context?

Etymologically the word “context” originates from the Latin words contex and contextus which refer to when things “weave together”, “interweave” and “join”. This sense of “weave together” was generalized from physical entities, like textile threads, to include the coherence links between different word meanings in written composition by Roman rhetoricians such as Quintilianus. Biblical scholars later treated context as the “co-text” between the parts of a text that occur before and after words.

The idea of context has in the twentieth century become further generalized—with context being extended to refer to the “weaving together” of psychological and psychological related entities as studied in cognitive science, Artificial Intelligence, computer interface engineering, pragmatics, linguistics, social theory, neurocognitive and social neuroscience. In psychology, context has been suggested to conceptually relate to the idea of “field” in cognitive styles, and also the background in which perceptual entities are interrelated together in gestalt theory. This idea of central coherence links to the processes by which cognition coheres different parts into larger wholes.

Information stages in cogntion

  circumstance situation context
actual reality perception awareness apprehension
  circumstance situation context

Cognition can be considered to be the information processing stage that follows the sensory impressions made by perception. Prior to cognition, sense organs and the initial analysis of their input by sensory cortices separates out information about the world into co-occurring entities. These entities are arranged in their input, by this perceptual separation, in regard to circumstances.

Such sensory circumstances about entities, however, only provide very limited information about the world. Enriching this perception is the task of cognition. Information exists that can be added to the experience of entities provided by perception. This information exists because entities in sensory impressions do not co-occur together at random. This creates a hidden source of information about them that can be extracted. Individuals are engaged in perception and actions for about 5,000 waking hours each year and so experience a constant stream of new associations in a rich variety of episodes. This provides the developing brains with tens of thousands of different and association rich episodes of experience. This when they gradually accumulated provide considerable information with which to create cognition.

Cognition enhances the experience of entities in two ways. First, by creating an awareness of the situation of perceived entities, and second by creating an apprehension of them in their context. Essentially, the difference between situation and context is that the awareness of an entity’s situation concerns its properties and how they link it in an episode with other entities. The apprehension of an entity’s context, in contrast, concerns how the other entities in the episode and their properties cohere together and interweave in regard to that entity.

Perception and circumstance

circumstance

To illustrate, starting with perception. Imagine being in a dinning room and seeing or feeling an apple in a fruit bowl. This sensory impression tells us that the apple is located within the bowl. This informs us about the circumstances of its location relative to other entities—above, below, adjacent or far away. This information comes with the processes that create the sensory input and their circumstance together.

Awareness and situation

awareness

Shift to having a cognitive awareness of that apple: it is now experienced in terms of its properties such as its organicness, edibility, heaviness and inanimateness. This makes it different as an entity to that of a small helium balloon or a resting bird in a bowl, even if these might share similar circumstances in the bowl to the apple, since unlike these the apple will not drift or fly off. On the other hand we experience the apple, unlike a balloon or bird, as an entity that might soon decay or get eaten. These properties apply to apples generally wherever they are perceived to be located, whether in a room or an orchard, a fruit bowl or on a kitchen table. They are circumstance independent. These properties are, however, still experienced as important and create its particular situation in the dinning room—its heaviness lets us know that unless moved by some force, the apple will remain in the bowl; its organicness that it may, if left, decay or be picked up and eaten.

Apprehension and context

context

Now consider the cognitive apprehension of the context of that apple. When so apprehended, the apple is experienced cognitively in terms of its interconnectedness. The apple is given a role by its surroudning entities and environment such that the apple could be substituted, such that the context they give the apple in the bowl would be preserved by putting another fruit, another type of food (such as a candy bar), or even a wax replica of an apple. Which context relevant properties of the apple—and therefore which kind of entity could be substituted—depends upon what else is in the dinning room, and why it is there. These might fit in with it having the role to provide food for a guest, or it might be to provide ornament as part of the room’s décor.

In other words, awareness concerns what is useful to know about an entity that is independent of its circumstances with other entities but which in its situation forms part of its relationship with them—like spokes from a wheel hub. Apprehension concerns the opposite: knowing what is salient about an entity which is dependent upon its circumstances—those ways in which it is contextually interweaved with links that fit it in as part of an episode with the other entities—spokes that radiate from them to it.

Another way of understanding these differences is that perception is like a photo segmented into its raw entities. It contains only informaiton about the physical relationships between the entities it contains. Awareness is adding labels to those entities so that the properties of the various entities is known. This provides them with a situation. Apprehension is like doing a google on those entities so information of their relevance to each other is added. This provides information about how they might be changed for other ones without changing their context.

What is the relation between autism and context impairment?

The processing of context is an aspect of cognition that has been proposed to be impaired in autism. Such a context impairment is the basis of the weak “central coherence” approach: according to this, normal cognition depends upon a “built-in propensity to form coherence over as wide a range of stimuli as possible, and to generalize over as wide a range of contexts as possible”. Central coherence, moreover, provides, “the everyday tendency to process incoming information in its context – that is, pulling information together for higher-level meaning”. Those with autism have weak central coherence, and due to this: “Normal children everywhere do well when they understand and take account of context. This is not the case in autism”.

The inability to process context in those with autism has also been observed more generally, for example Hugh Morgan: “in autism the prerequisites for creativity are not present. The adult with autism cannot extend the known, or bring together understandings to create new ones, because the known remains confined to the specific context in which it was learnt”. Susan Dodd has observed: “A child with autism who is told to 'pull up your socks' or to 'shake a leg' is unlikely to understand that these phrases have multiple meanings depending on context. The child is likely to interpret the phrases literally, regardless of context.”. An inability to experience things normally in context was also implied in the first description of autism by Kanner: “A situation, a performance, a sentence is not regarded as complete if it is not made up of exactly the same elements that were present at the time the child was first confronted with it. If the slightest ingredient is altered or removed, the total situation is no longer the same and therefore is not accepted as such ..”

Supporting this general link between autism and context, research has found deficits with those with autism in the processing of context for social cues, face processing, memory, perceptual groupings, distinguishing essential from variable aspects in event schemas, and also in the skills of generalizing from structured settings to more naturalistic ones which may contain unpredictable and context-dependent interpretational elements. This impaired contextual processing in autism has also been proposed to affect the production and comprehension of language in regard to homophone pronunciation, lexical ambiguity, sentence understanding, semantic incongruity detection, prosody production and comprehension, metaphor and metonymy appreciation, irony, discourse continuity, and communication pragmatics. Repetitive behaviors in autism have been described as originating in contextual impairments, as such behavior is “an assemblage of behaviors defined by their topographical similarity across contexts, inappropriateness, and repetition”. Individuals with autism, in addition, show superior performance on tasks that require ignoring irrelevant context.

What is LSA and what are its origins in search indexing?

Latent semantic analysis (LSA) can be viewed as an area of science with two parts. First, it offers a mathematical account of the presence and extraction (by the singular value decomposition of a matrix using various transformations into a multidimensional vector space) of the information latent in the higher-order associations contained in past word usage. Second, LSA provides a successfully tested simulation of the use of such information in regard to the many aspects of language such as word sorting and category judgments, estimations of passage coherence, and the quality and quantity of knowledge contained in student psychology essays. In particular, LSA shows that children possess sufficient information (contained in the higher-order association of known words) to guess what unknown words might mean given the context information provided by surrounding words. This has allowed LSA to model how children learn the meaning of unknown words, an aspect of child development that prior to LSA was notable in being difficult to explain.

LSA mathematically uses latent class analysis. So far it remains the only model using this mathematics that has been widely and successfully tested against real data. Because of this, though the principles it uses are more general, the focus here is upon LSA. However, it should be appreciated that LSA and the field associated with it of computational semantics is undergoing rapid development. Moreover, LSA has its limits, notably that it does not extract all potentially useful information. For example, simplifications in its implementation results in it ignoring much of the information available to word learning such as syntax, word order, and real world associations. Nor do the episodes it processes correspond exactly to those of sentences. Implementations that incorporate some of this information are being developed such as those that involve generating “probabilistic topics”.

Historically, LSA originated in computer scientists seeking an automatic means to retrieve documents by keywords. (For this reason it was originally called latent semantic indexing, LSI.) Computer scientists faced the problem that the link between keywords and the words in sought-after-documents depended upon their context. If a searcher types in, for instance, the keywords, “film” and “Marilyn Monroe”, they seek to retrieve not only documents that mention “film”, but also related synonyms found in the same context (such as “movie”, “Hollywood” and “motion picture”). Further, they want to retrieve only those documents that contain the word “film” that fit in with the context of “Marilyn Monroe”, and not ones containing “film” when it means “thin coating”. Computationally, the context sensitivity needed to identify such synonyms and homograph meanings cannot be reduced to the information contained in its direct co-occurrence associations with adjacent words. Computer scientists to overcome this developed methods to determine the synonyms and homographs of keywords by extracting from large corpuses of texts the higher-order (or indirect) associations that words have with each other. This information had not been previously investigated, as previous work upon the information in texts (due to limits upon computer power) was confined only to the extraction of first-order or direct co-occurrence associations. Using information extracted from higher-order associations, LSA applied to word searches has been able to efficiently detect synonyms and homographs needed for keyword document retrieval programs. Although commercial confidentiality prevents the open publication of the implementation of contemporary search engines such as Google, web newsletters such as this one report they incorporate, in addition to PageRanking, the use of LSA-derived context extracting techniques.

How does the information extracted from higher-order associations revolutionize how scientists understand meaning?

LSA provides an empirically tested account of the phenomena of context and what might be called “meaningfulness”. There are two issues here: meaningfulness as a measurable “behavioral” aspect of words, and the ability of LSA to match the performance of human individuals on this measure.

Context and meaning link through “behavioral properties”. These concern not theories about the nature of meaning but objective behavior as shown in human judgments about word meaning. These properties contrast with the “sense” properties traditionally investigated by philosophers that concern intuitions about semantic reference. Behavioral properties concerning the comprehension of meaningfulness affects the objective behavior of human judgments about words in terms of their similarity, their intersubstitutability in sentence contexts, and their predictability in certain sentence contexts.
         
The similar-dissimilarity property.

All words have various degrees of similarity and distance of meaning. “Vast” and “big” are regarded as having a great deal of semantic similarity and closeness, “vast” and “small”, a smaller amount, while “vast” and “kiss” are very distant, for example. What might underlie this property of semantic similarity is uncertain: the resemblance or lack of meaning between words cannot be inferred from their visual or sound identity nor their close association with other words (the level of first-order associations). For example nothing about the perceptual nature of “vast” and “big” – its letters and phonemes – provides information that they are synonyms. Nor is this information provided by their immediate associations with surrounding words in the sentences containing them. In spite of this, humans have an immediate apprehension of which word meanings are similar and which are not.

The synonymy intersubstitutability context property.

Meaning” has the property that words that have roughly the same “meaning” – synonyms – are contextually intersubstitutable. The more similar their meaning, the more alike the contexts in which they appear. Consider “vast” and its synonyms, “big”, “large”, or “vast”, each of these three words can be substituted in most sentences without significant change of sentence meaning. “The big rock broke the wagon”, for example, means roughly the same as “The large rock broke the wagon”, and even, “The vast rock broke the wagon”. Like the distance-similarity property, while the human brain can understand intuitively which words can be swapped, philosophical or other analysis finds it difficult to specify the nature of the processes involved.

The predictability usage property.

Linked to context is the property that words do not appear at random in normally encountered sentences. Expectations and constraints exist–word usage patterns–about which individual words tend to appear with which other ones. This can be experimentally shown with cloze sentences in which single words are omitted and individuals face multi-choice options to pick the missing word. If human cognition is presented with the incomplete sentence, “The ___ stone broke the wagon”, and asked to predict the most likely missing word out of “big”, “forgetful”, and “sweet”, the word picked would be “big”. “The big stone broke the wagon” has an expected pattern of usage, that is absent – even though they are readily comprehendible – in the sentences, “the forgetful stone broke the wagon”, and “the sweet stone broke the wagon” (understandable, for example, if they were to appear in a “fairy story”). Indeed, even without suggested word choices, if single words are cut from writing, human subjects can guess with the remaining before and after words (depending upon sentence position) up to half of them. Further, if the next word in a piece of ordinary writing is covered, and an individual has not read ahead, they will be able to guess it in about one in four times. The ability to fill in missing words in sentences is used in the testing of the comprehension progress of language learners, and the assessment of language competence, for example, in TOEFL (Test of English as a Foreign Language) certification.

Semantic space and LSA

Judgments about semantic closeness cannot be made directly upon the information extracted from higher-order associations. The information this contains about words therefore needs to be changed into a form that can be compared with human judgment behavior. This is done in LSA by converting the information that associates words into a multidimensional vector space. In this space, synonyms occupy the same locality (due to their intersubstitutability in any context), while the different meanings possessed by heteronyms (homographs and homophones) are widely separated (they produce different meanings when put in different contexts). More generally, the closer the meaning of words (in terms of cosine similarity), in this space, the closer mathematically are their vectors, and vice versa. This provides a means to assess semantic closeness created by LSA since such measures can be matched against judgments made by human subjects.

In this multidimensional space, vector locations not only are given to words but also to complete and incomplete sentences. A word can be synonymous, after all, not only with another word, but a group of words – for example, the definition – “being of extreme size” is synonymous with “vast”. Sentences and other groups of words are given a vector location by mathematically adding together all the vectors of their individual words. This provides a new vector location that is a kind of mathematical “center” of their individual vectors. This turns out to be useful as a means of gaining information about unknown words. This is because the words that contextually surround the unknown word due to their coherence contained in their higher-order association information provides information also in regard to the unknown word.

LSA and the behavioral properties of meaning

LSA successfully models the behavioral properties of meaning, such as similarity and the distance judgments, and does so with a performance that matches that of human subjects. It also models successfully the ability to fill in words in TOEFL. For example, applicants to US colleges with English as a second language, if given a word and four possible synonyms will get 64.5% correct; the LSA model, 64.%.

As a consequence, LSA provides strong support to suggest that the higher-order associations which it extracts play a key role in the brain’s generation of the context that underlies the meaning of words (as reflected in the brain’s making of similarity and dissimilarity judgments). This would be an unlikely finding if the higher-order associations extracted by LSA were a mere epiphenomenon to the brain’s processing of meaning.

Landauer and Dumais further find that the ability to infer unknown words from the context of known ones, in the high-dimensional space generated from higher-order associations, plays a key role in language development. A key theoretical problem in child development is explaining the success of children in learning new words. Children learn on average 10 to 15 new word meanings each day but only one of these words can be accounted for by direct instruction. The other nine to 14 word meanings need to be picked up from another source of information. Landauer and Dumais have found that when children encounter an unfamiliar word, the context of its surrounding known words contains sufficient information to enable them to guess its likely meaning. The words children already know together with the information associated with them from their higher-order associations allows them to create a semantic space from the words near the unknown one. These words each have a vector location in this space that can then be added together to create a vector of the context of the unknown word. This vector turns out to be sufficiently near to the meaning of the unknown word to provide a source of information for determining what that might be. The ability to learn new words expands with vocabulary and language experience. LSA explains why: the more words an individual knows, the better the multidimensional semantic space they create, and so the easier it is to use the information surrounding words to learn new ones.

LSA shows that most of the information needed to make judgments about meaning and learn new words comes from higher-order associations not first-order ones. It finds that unlike first-order ones, higher-order associations tend to provide only relatively weak information about word usage. However, higher-order associations are by many orders of magnitude much more numerous than first-order ones, and so (when added up) in total contain much more information. As a result, as Landauer and Dumais note, “About three quarters of LSA’s word knowledge [when tested] is the result of indirect induction, the effect of exposure to text not containing words in the tests”.

How does LSA explain the language impairment of autism?

From Leo Kanner to the Diagnostic and Statistical Manual of Mental Disorders. Fourth edition, language and meaning problems has been central to the diagnosis of autism. They include mutism, echolalia, language acquisition delay, pedanticism, and atypicality in the understanding and use of word meaning. The latter will be of concern here. Kanner noted that, “the autistic child has his own private, original, individualized references, the semantics of which are transferable only to the extent to which any listener can, through his own efforts, trace the source”, and that their word meanings are “rooted in concrete, specific, personal experiences of the child who used them”. Similar comments have been made by Uta Frith and Francesca Happé. They note that children with autism, “may use single words in a simple, associative way, so that “Apple” always means, “Give me apple”. The single words acquired are often esoteric (e.g. “Beethoven”) and not like the first words of a normally-developing pre-schooler. Neologisms (e.g. “bawcet” for bossy), or familiar words with special meanings “yes” meaning “carry me on your shoulders”), also reflect the very concrete context of word and object”. These comments suggest that individuals with autism when they hear unfamiliar words are restricted in their attempts to understand what they might mean to the use of their immediate (and often misleading) associations in the physical world. This, indeed, was described by Kanner in regard to a child called Paul G who said “Peter eater” when ever he saw anything resembling a saucepan. According to his mother, when Paul was two years old, while busy in the kitchen, she was reciting to him the nursery rhyme about “Peter, Peter, pumpkin eater”, when she dropped a saucepan. Ever since Paul has understood “Peter eater” atypically to mean “saucepan”. He had made an association between “Peter eater” and the most salient event at the time of its been said – the dropped saucepan.

This bias in those with autism for using direct co-occurrence associations would suggest a particular kind of failure in word learning. First, that they fail to employ the context-based cognition processes by which nonautistic individuals acquire word meaning. Second, that in its absence, they rely upon the accidental nature of word and event co-occurrences to guess what words might mean. Third, that when they have done this, they are impaired in using context to appreciate the inappropriateness with which they have understood a word. This has two results:

The miscomprehension of homographs by those with autism also provides evidence of a deficiency in the processing of context as modeled by LSA. Four sets of experiments involving those with autism that can read have found evidence that they ignore sentence context when asked to pronounce a homograph. Unlike those without autism, they tend to pronounce the most common occurrence of homographs, rather than the one that fits its sentence context. Thus, they will tend to pronounce the word “tear” in the sentence, “The lady had a TEAR on her dress” in the sense of a cry droplet rather than of a rip or cut. This suggests that they might not access the context used by nonautistic people in word comprehension. LSA has shown that the context needed to disambiguate the different meanings of homographs is extracted from higher-order associations.

LSA has also shown that context information is used by people to judge the degree of coherence present in texts. Thus if LSA-like processes were damaged in autism, the perception and use of such coherence would be expected to be atypical. Evidence suggests that this is the case: individuals with autism tend to fail to comprehend the normal coherence that allows inferences between sentences. In further support of this, as noted in DSM-IV, those with autism seem unable to create continuity between sentences, that they fail to continue the topic of conversations, and, regardless of meaning, repeat words or phrases.

What is the role of context in everyday cognition?

The three things are identified by LSA as important to language—the information contained in higher-order associations, context, and intersubsitutability—are not specific to language. Like words, entities identified in nonlinguistic domains can also have cognitive attributes akin to “synonymy” and “heteronymy”. A car, a donkey, and a sedan-chair may look very different, but they provide similar intersubstitutable means of travel when on a holiday to get from one place to another. They – at least if our desire is holiday transport – can substitute for each other (much as the words, “vast”, “big” or “being of extensive size” can in the same sentence substitute for each other). Here, however, what they share is not the same meaning but the same functionality as required by our needs and our cognition of those needs. “The tourist went in a car to the shops”, or “The tourist went on a donkey to the shops”, or “The tourist went in a sedan-chair to the shops” while different as physical activities are similar in terms of their being solutions to a holiday need to travel. (And, reflecting this, the sentences describing them mean roughly the same in that context). Likewise, in terms of function, an entity can be “heteronymous” in regard to different cognitive relevancies: a car in one context can be a means of transport, in another a place to escape rain, and in yet a further one, a status symbol. It would thus seem plausible to infer that the multidimensional space LSA has found for words and meaning has its parallels for entities in other cognitive domains.

The problem of identification of unknown entities from context is not confined to words. A cognitive problem faced by many faculties is identifying something – a hidden object, an unfamiliar response, or an unidentified aspect of a situation – given the information provided by surrounding known events and entities. This problem is encountered particularly by the faculties that provide humans with a sense of security, social interaction and the apprehension of mental states. Such faculties develop using the information contained in the tens of thousands of episodes encountered in daily experience across the many years in which cognition matures. Thus, such faculties have available the higher-order association information contained across such episodes to generate a multidimensional space that can be used to identify unknowns from their surrounding context.

Context information processing also plays a central and crucial role in behavioral adaptability. This depends upon several skills.

This raises the possibility that the extraction of context information from higher-order associations and its following consolidation into cognition might be widely exploited by faculties other than language in their own development and function. From this it follows that impairments to such processes might produce impairments in these nonlinguistic functions. How closely would these resemble the symptoms that characterize autism?

How does this theory explain the need for sameness and impaired stress regulation in autism?

Two main traits were seen by Leo Kanner to be present in individuals with autism: an “anxiously obsessive desire for the preservation of sameness”, and an “extreme autistic aloneness”. The American Psychiatric Association’s DSM-IV criteria for Autistic Disorder likewise gives its two essential features as “a markedly restricted repertoire of activity and interests”, and “the presence of markedly abnormal or impaired development in social interaction and communication”.

Harm perception atypicalness

One instinct built into humans is a concern to avoid what might cause us harm. Harm avoidance depends upon the cognitive capacity to anticipate and then activate specially evolved neural circuits in the limbic system, and the hypothalamic-pituitary-adrenal (HPA) axis so that the physiological reactions of fright, fear and ”flee or fight” that protect individuals from harm are induced. The normality of these harm protection reactions therefore critically depends upon the integrity of the cognitive processes that detect threats of danger. If people falsely expect harm, they may panic even when they are safe (and so waste their energies or unnecessarily limit their actions), or if they fail to foresee harm, they may not panic when this might enable them to avoid later suffering and injury. Thus, experiences such as panic and anxiety are intimately linked to how successfully the brain’s information processing can predict and identify contexts of actual threat. This makes the efficacy of acquiring such cognitive processes central to the everyday ability to live without suffering constant and undue distress.

Context and harm awareness.

Information used to anticipate danger can depend upon first-order cues or incorporated context. Anticipatory information provided by first-order clues concern such attributes as the appearance of a threat (the shape of a snake) or the unfamiliarity of a situation. However these danger cues are not always trustworthy. For example, the shape of a snake – a first-order perceptual cue – can warn an individual of a dangerous animal, but many entities from a distance look like snakes (large worms, water hoses, plastic snake imitations), and many snakes are not poisonous, (and poisonous ones are not dangerous if dead or in glass tanks). Familiarity is also an unreliable clue: many familiar places and things can be dangerous – for example, a house on a beach does not change in appearance even though a forecast hurricane requires its immediate evacuation. Further, strange places can be safer than familiar ones (the sports center into which people seek refuge from an advancing hurricane). The brain thus is advantaged if it can use context to extend and refine its sense of security so it minimizes false positives (anticipations that are false), and false negatives (failure to anticipate dangers that are real). Unimpaired and experienced cognition has the information to do this since it can incorporate the information contained in the many tens of thousands of varied and diverse situations that have happened either to ourselves, or that which has been witnessed to happen to others, in which harmful or nonharmful events occur or not. Thus, cognition has the opportunity to extract higher-order associations for the contextual anticipation of danger by means of a multidimensional space in which situations can be found to be similar or dissimilar in regard to their risk and security. Such a multidimensional space, as with words, would allow cognitive processes to use the context of the known to identify the unknown (in this case whether a new situation is likely to be safe or dangerous).

The limbic and the hypothalamic-pituitary-adrenal (HPA) axis can use such information to supplement unfamiliarity and other unreliable co-occurrence cues. Thus, as a result of this, the brain can feel safe even when unfamiliarity and first-order cues suggest it should feel danger (for example, when it first travels in an aircraft), provided that context reassures a person that there is no threat. The brain, however, will still remain sensitive to dangers if this context changes: for example, if the warning lights above the passenger seats light up and the captain announces that the passengers must prepare for an emergency landing. These clues while they are not in themselves unfamiliar events (at other occasions the lights come on and the captain speaks), however, they create a new context (passengers do not think the captain is fooling or mistaken) that warns that them that face actual risk.

This analysis of the role of context to anticipate risk and safety raises the question of what happens when it is impaired. In this situation, the limbic system and the hypothalamic-pituitary-adrenal (HPA) axis will have no option but to be limited in its threat activation to those provided by first-order perceptual clues. As a result, individuals with an impairment in extracting context, will be hypersensitive to changes in what is recognizable and well-known. Those with autism are diagnostically characterized by panic and distress over small and irrelevant alterations in their routines and circumstances. As the DSM-IV (American Psychiatric Association 1994) criteria explain, “they may insist on sameness and show resistance to or distress over trivial changes (e.g. a younger child may have a catastrophic reaction to a minor change in the environment such as a new set of curtains or a change in place at the dinner table)”.

How does this explain explain lack of sociability in autism?

Humans are a social primate. Due to this, the human brain generates emotions that bind individuals with kin, friends, and other individuals. This interpersonal involvement with each other is moreover highly contextual. For example, when a person meets a friend, how they interact will happen within various contexts such as their past relationship (there might be past obligations to repay); their relationship to others (they might know a mutual friend whose impressions they value); what they feel (if their friend is sad, they try and cheer them up); social recognition (their friend wants to be valued as a special person), and social mores, rules, manners, and morals (there are contextually defined boundaries upon behavior). All these concerns enrich social interaction with multiple levels of context; nothing analogous occurs in the human relationship with physical objects.

Social psychology, moreover, in its attempts to understand social existence gives a central position to phenomena such as ‘self’, social roles, sense of personal identity, self-presentation, pragmatics, and awareness of in-group and out-group boundaries. These depend for their functioning within social psychology upon the ability of people to understand or experience social events and entities as happening within an interpretative context. Social psychologists, however, do not provide an information processing framework in which to understand at a basic information processing level such phenomena. Processes such as “self”, “personal identity” and “social group perception”, as a result, from a computational perspective, are nebulous lacking an underlying cognitive mechanisms that could support their generation or perception by the brain. But it is striking that social psychologists describe such phenomena as being carried out by processes that are akin to those discussed above in being interpretative, sensitive to context and concerned with ‘meaning’.

Sociability, moreover, builds upon context like the sentences of language by building upon earlier ones and contextually inter-chaining them. In a text, this is reflected in the fact that when a sentence follows another, it expands and develops the meaning of the earlier sentence. Because of this, when the sentences of a text are scrambled up, they do not make sense (their meaning depends upon the sequential flow of context). Likewise, sociability involves an interdependent flow of context: imagine the episodes of a social situation mixed up; they would cease to be recognizable as a social interaction. One action at one moment means one thing, at another moment, something very different. Due to this context dependence, if one element is missed in a social interaction, it can be guessed from its surrounding ones, much as in a cloze sentence, an omitted word can be conjectured from the context provided by surrounding ones. While physical entities and the human relation with them have some aspects of context, they lack this rich embeddedness in context that shapes social behavior and social interpretation.

Every human brain encounters from its earliest years tens of thousands of episodes of social interaction. Thus, it has considerable opportunity to extract social context information from the higher-order associations hidden in them and generate a multidimensional “social intercourse” space. Such a multidimensional space contains context information which can be used to generate social phenomena that get identified as “self”, social roles and personal identity.

Parsimony suggests that it should be assumed (unless good reasons exist to think otherwise), that this information processing system has an important role in enabling and enhancing sociability. This is because such information is available (people are constantly engage in social interactions), that such information is known to be used by one faculty already (language), and that there exists a need for such a information processing system (humans live in communities, and in their affects, they are social primates). If this information processing does not exist, a problem arises why humans have a skill (sociability), (1) that has an information processing similarity to that used in another domain (language), (2) that has the information needed to be processed like it, and yet (3) uses some different and distinct process. Ockham in his razor requires that entities are not multiplied beyond necessity. Thus, even though precise models are lacking at present, it is reasonable to suggest, at least provisionally, that the cognitive processes under sociability depend upon LSA-like computations.

Social emotions and context.

Social psychology links emotions to these above discussed context aspects of sociability and self. For example, consider this description of the human need for social acknowledgement by William James:

We have an innate propensity to get ourselves noticed, and noticed favorably, by our kind. No more fiendish punishment could be devised, were such a thing physically possible, than that one should be turned loose in society and remain absolutely unnoticed by all the members thereof. If no one turned round when we entered, answered when we spoke or minded what we did, but if every person we met “cut us dead,” and acted as if we were non-existing things, a kind of rage and impotent despair would ere long well up in us.

People not only exist in a social context, but as James suggests here that emotions make humans concerned that others experience them as part of the context of their social world.

Atypical social interests. 

The above observations argue that sociability and a sense of self would not normally arise without the ability to extract higher-order associations from interpersonal interactions. First, individuals would not be able to acquire the context sensitivity needed to engage in social behaviors such as greeting and responding to greetings. Second, individuals would not be able to interchain social interactions so that each social interaction followed on appropriately from earlier ones. Third, since roles, self and a sense of social presence derive from the multidimensional “social” space generated from such interactions, together with emotions (people care much to be noticed), individuals will not be motivated to interact. As a result, people lacking the ability to extract and consolidate higher-order association will still have emotions, but such emotions will enable them to relate only through perceptual-based types of cognition: their affect will link to specific behaviors, and resemble their behavior to objects. They will, moreover, lack the ability to experience the many emotions that arise out of context such as the “rage and impotent despair” that James mentions above. Indeed, rather than such social engagement emotions, they will be content if they are socially ignored. They will show poor or inappropriately developed capacity for social roles, personal identity, self-presentation, and social boundaries. In consequence, such individuals will seem to others socially remote, passive, odd and lacking in friends. This description characterizes the core deficits of the social impairment of those with autism. As put by the DSM-IV, in the autistic disorder:

Younger individuals may have little or no interest in establishing friendships. Older individuals may have an interest in friendship but lack understanding of the conventions of social interaction. There may be a lack of spontaneous seeking to share enjoyment. interests, or achievements with other people (e.g. not showing, bringing or pointing out objects they find interesting) (American Psychiatric Association 1994). Further, that “Lack of social or emotional reciprocity may be present (e.g.: not actively participating in simple social play or games, preferring solitary activities, or involving others only as tools or mechanical aids)” (American Psychiatric Association 1994).

How does this theory explain impaired theory of mind in autism?

One context in a social situation is created by human goals, desires, beliefs and knowledge. In spite of these mental processes being hidden, individuals can predict how they will shape the behavior of other individuals. For example, an individual can predict if a child believes a toy is in a box, that they will act in one way, and that if they believe it has been moved, that they will act in a different one. Skill in anticipating how mental states shape behavior is called theory-of-mind.

Mental states, however, cannot be directly inferred from first-order co-occurrence associations of the entities making up the episodes in which they occur. This is because they are not only hidden but rarely correlate to the immediate aspects of a situation. In information processing terms, mental state inference involves inferring the presence of unknown elements from a context of known ones – a cognitive task parallel to episodic inference one of the identification of the meaning of unknown words from the known ones that surround them. Here, however, the unknown element guessed from observed behavior and its situation is an aspect of mental activity.

Such a skill in mind reading will depend upon an ability of the brain to create a multidimensional space in which to locate the observed elements of behavior that can be used to indirectly infer such hidden mental states. The higher-order associations needed to create this space exist since all episodes of behavior arise from people having varied intents, beliefs and desires, and that, over many years, the brain encounters tens of thousands of situations molded by mental phenomena. The higher-order associations contained in them allow for the creation of a multidimensional “mind” space in which episodes in which people do things can be redescribed in terms of contexts and hidden intentional states. From this space, the known elements of a behavioral situation can be used to provide contextual information about the unperceivable ones of intents, beliefs and desires. (Though, as noted below, another source of information in creating this mental state space are mentalistic language terms for intent and mind: both sources of information have a role in developing theory-of-mind competence).

Due to the ability to guess unknown elements from the context of known ones (where these depend upon higher-order associations), it could be expected that impairments in extracting such co-occurrence associations would impair the cognitive ability to infer mental states. This fits in with the atypical deficits of autism. Awareness that others have desires, beliefs and intents (theory-of-mind) is impaired in those with autism. Some individuals with autism pass tasks that test theory-of-mind. But they are different: ordinary people cannot explain in detail the answers they give to such tasks (much as they cannot when they use context for identifying homographs that fit sentence context), in contrast, those with autism that succeed on these tasks provide articulate justifications. This suggests that they have tackled them as puzzles, and extrapolate their answers from superficially similar examples not as a consolidated part of their cognition. Such “example-based” puzzle solving is of limited use in replacing those based upon context: “mind-aware” people with autism fail harder tasks that test more difficult forms of mind awareness, moreover, they fail to use such skills in real life.

How does this theory explain islets of preserved cogition of autism?

A problem may require either the use of first-order co-occurrence associations or the context provided by extracted higher-order associations. But which kind of information processing is needed may not be obvious from the initial perception of a task: habits and superficial details could therefore cause a problem to be tackled inappropriately making it harder and more difficult to solve by misleading cognition to process task irrelevant information. The block design, and the embedded figures task, for example, embed a problem in information that invites a contextual approach but which in fact is irrelevant and distracting to its solution. Another instance is the learning of lists of meaningless words – since words are habitually heard as parts of sentence contexts, an individual will tend to spontaneously treat them as having potential meaning as a group – even though the memorizing of them is easier if they are treated as lists of unconnected elements. In consequence, such tasks will be more straightforwardly done if a deficit exists that suppresses the extraction or consolidation of higher-order information when faced with a problem. (Assuming, of course, that first-order information processing is intact.) If those with autism fail to process context information, they should solve such tasks better than people without autism. This indeed is the case. Kanner noted that people with autism could memorize lists of items better than ordinary people, and people with autism have been reported to be superior to normal people upon tasks such as the block design, and the embedded ones.

In the absence of higher-order associations, the development of diverse cognitions might be maladaptive since they might use, however, in the place of context the noncontextual information provided by available first-order co-occurrence associations. This atypical development need not always be inferior to that normally acquired. It is possible that in the absence of context that cognition might become highly sophisticated in doing certain tasks (albeit with limited usefulness). This even might produce in some individuals novel cognitive skills that show superiorities. Consistent with this, a few individuals with autism show unusual abilities in such activities as drawing, music, calculation and memory. Their exceptionality, seems to derive from replacing a sensitivity to context with elaborate and particularly expert use of noncontextual information.

How does this thery explain sensory and motor impairments in autism?

Some apparently noncontextual aspects of autism might hide the processing of context. This might be the case for sensory gating and motor coordination problems.

Many hundreds of sensory and biomechanical interactions exist (by the nature of the articulate human body being possessed of diverse and separated sensory organs) that bind together what the brain experiences, and governs what the body can perform. These would appear in many cases to require for their efficient processing sensory gating, anticipatory motor adjustment and other modulation that depends upon the detection of an appropriate context. Contractions of agonist muscles, for example, must be initiated contextually to occur with time appropriate relaxation, not just of antagonist muscles but also muscles in parts of the muscular that are distant, (and then only in certain movements contexts). Many movements substitute for each other depending upon context (for example, underhand and overhand throwing). Sensation is actively gated in a similar way dependent upon context: for example, input from joint and skin receptors is temporally “damped” so people do not to “feel” their own movements, and hear their own words when they speak. Touch input also is contextually heightened, (for example, when an individual’s finger tip contacts an object).

Such context gating and adjustment would be optimal if it used the extraction of the higher-order associations gained from the analysis of tens of thousands of past episodes of movement and sensation in which motor outputs and sensory inputs interact together. That such information would enable effective sensory-motor function could explain why motor skills take so long to perfect, for example, the ability to grab objects continue to develop in sophistication into the first decade of life. Due to this dependence of motor and sensory gating upon context, any deficits in extracting the context interactions between motor outputs, sensory input and the surrounding sensory-motor circumstances could produce impaired sensory gating and motor abnormalities. Those with autism have motor problems and sensory disturbances. This raises the possibility that general impairments in context processing could underlie these aspects of autism even though upon first impressions they do not seem be context dependent. Supporting this, Happé and Frith observe that “perceptual abnormalities such as hypersensitivity, clinically/anecdotally reported but little studied in research to date, may relate to context-free processing as expectations and context-based interpretation are known to modulate experience of sensory stimuli”.

How this theory link autism to impairments in the brain?

The information contained in higher-order associations is more difficult to extract than that in first-order associations. To extract the information contained in the first-order association between two entities requires relatively uncompliated computation since it involves just those episodes that contain them. Moreover, it is processed as a normal part of cognitive activity. As a result, it can be acquired directly by the cognitions utilizing such information from the experience of such entities in past episodes. But to find the information contained in the higher-order associations requires processing at a global level that encompases all entities that have occurred in all episodes as a totality. This is something that is not normally done as part of cognition.

Another difference is that the extraction of information from higher-order associations concerns very weak bits of information that accumulate into useful information due to the very large numbers of associations containing these slight bits of information in past episodes. Further this weak information is scattered across the brain. Different parts of the brain, for example, process the “noun” and “verb” qualities of entities, and also the concepts of tools and animals. In consequence, the information in the higher-order associations will be neurally spread and hidden. This will need a process that can extract it irrespective of where it is neurologically located. These computational factors of global, weak, and neurally dispersed information require that the extraction of the information from higher-order associations is specialized by the brain since only this will allow it to be effectively and efficiently processed.

There is also the factor that many cognitive faculties develop in regard to overlapping sets of experiences. For example, the ability to contextually apprehend security and danger derives from past experience in terms of their safety and danger. These episodes are also ones that contain context information needed for apprehending context in language and social interactions. Thus for distinct functions, the same or overlapping episodes will need to be processed to extract context information from higher-order associations. This creates a computational redundancy that can be avoided if one process extracts this information, and then it is transferred for separate consolidation into different cognitions.

Such separation and specialization is consistent with how the brain is internally organized. Information processing in the brain occurs in white matter circuits through which information is transferred, modified and consolidated. This consolidation/transfer process occurs for the cerebral cortex in regard to modification of its information by white matter traits to and from the cerebellum, hippocampus and basal ganglia. Splitting information processing in this way allows areas to specialize in different types of computations.

Consistent with this, a consolidation processing defect has already been suggested to underlie autism in regard to the hippocampus and cerebral cortex and cognitive representations of the environment acquired from experiences. In addition, impairments in white matter integrity have been identified in those with autism that impair the connectivity needed for such modification and consolidation.

The separation in the brain between the processes that extract higher-order information and the cognitions that use this information has important implications for the etiology of its possible impairment. This is because the impairment of such separated and specialized processes could occur without impairment occurring directly to the processes that use extracted contextual information. Their impairments, however, would still have a knock-on impact upon the processes underlying everyday cognition. This is because these otherwise intact processes would be disrupted by the consequences of them not developing with the information extracted from higher-order associations.

This would greatly alter these processes.

First, many cognitions would not be able to carry out the functions that required the information extracted from higher-order associations.

Second, while many normal cognitions would be impaired by the lack of information from higher-order association, some which use only first-order information would not be affected creating an uneven cognitive development.

Third, some cognitions might compensate for missing information with that extracted from first-order associations functioning creating forms of cognition not found in unaffected individuals.

Fourth, in rare cases the lack of information from higher-order associations will result in the processes acquiring greater competence – compared to that present in neurotypical individuals.

Thus, following an impairment to the processes extracting the information from higher-order associations, many faculties would be unimpaired (those not using this information), some impaired (those using it and for which the information from first-order associations cannot replace), some atypical (those compensating the information from higher-order associations from first-order ones), and some superior (those cognitions for which the information from higher-order associations impairs). This would parallel the symptom profile of autism where many cognition competences are preserved though many are impaired—for example, having reading comprehension is impaired but not the ability to read, spell, and do computational tasks and mechanical reasoning that is preserved compared to social reasoning.

The extraction and the use of information can happen on different occasions. Search engines such as Google in their early years crawled the web to create an entirely new index every few weeks, though the resulting index so created became searchable within a fraction of a second. Likewise once higher-order associations are extracted and converted into a multidimensional vector space, this process does not necessarily need to be repeated as the information extracted will be in most cases usable for new entities. Neurobiologically, the extraction of higher-order context is likely to be continuously updated from the constant stream of episodes that arise in everyday life. But, if for any reason this process is interrupted or stopped, the brain will not lose the context information that it has already acquired, employed and consolidated and so made part of mature cognition. As a result, any defect later in life in the processes extracting higher-order context need not necessarily produce noticeable cognitive defects since the cognition of such individuals would still possess functioning context information from the period when such processes had been intact. Impairment will thus be critical only for young individuals before they have acquired a reserve of such extracted context information. As a result, this would limit – as in autism – the effects of impairments upon such processes to when they happen early but not to when they are impaired later in life.

How does this theory explain the differing development of autism?

Autism takes many forms and any candidate theory of autism must be consistent with such diversity. The extraction of information from higher-order associations is part of a circuit with several stages. The circuit contains the acquisition of experience in terms of episodes containing associations, its transfer through white matter tracts to specialized processes that extract the information in regard to its higher-order associations, the transfer of this information by white matter tracts to cognition, and finally its consolidation as context into cognition. Each of these might be separately compromised with different consequences. For example, impairment to the extraction process would result in widespread impairment, while impairment to white matter tracts might limit impairment to only those cognitions for which they provide a specific information transfer.

Differences could also occur within the specialized computational process. For example, varying the size of the window of association elements from which such information is extracted, and the number dimensions into which such extracted information is rerepresented might produce different kinds of atypicalness. These offer diverse ways in which impairment might create different forms of contextual information deficiencies.

Another layer of variables on top of this would be created by the different environmental and educational support that ameliorates or increases their impact on such neurocognitive impairments. The neurobiological factors that cause impairment upon neurocognitive processes, in addition, could independently of this be expected also have their own separate effects upon cognitive integrity (such as causing general mental retardation and epilepsy). Thus, no reason exists to assume (except in broad aspects) that in different individuals the behavioral consequences of an impairment to processing higher-order associations, and so context, will in detail be alike.

How might the Google theory of autism be researched?

Several areas exist by this theory might be researched. One is computational lesioning. This has already occurred for other developmental conditions, such as dyslexia, Williams syndrome, brain injury, and specific language impairment. If the general approach of this theory is correct, then lesioning LSA computer simulations should produce atypical word comprehension and learning. Further study could be made in those with autism and other forms of atypical cognition based upon such lesioned LSA models. Models of autism and context have been developed—LSA suggests ways in which this research could be expanded with more specific notions of context and its impairment.

Atypicalness in autism also suggest areas in which the LSA could model uninvestigated aspects of language. For example, the multidimensional space created by LSA represents, in a manner, conceptual knowledge – an aspect that has resulted in LSA models being successfully used to automatically mark the content of student essays in regard to their knowledge of psychology textbooks. The acquisition of theory-of-mind skills is known to link to language, and the social exposure to mental state words. This suggests it is likely to be extracted by LSA from the usage of mentalistic words. Theory-of-mind skills have not as yet been modeled by computer simulation, nor has the LSA model been tested as to whether it can simulate them. According to the approach proposed here, the lack of theory-of-mind skills in those with autism will derive, like their other impairments, at least in part, from a defect in extracting and using higher-order associations contained in mentalistic words. Given the dependence of theory-of-mind on language, and its proposed dependence upon higher-order context, this should be modelable by LSA.

Charles' Sorting Task—an overlooked means of investigating and treating autism

Another important area is developing tasks for research that can better identify the processes that depend upon the information extracted from higher-order associations. Charles’ Sorting Task is relevant here as it directly measures the capacity to comprehend synonyms. In this task, cards are prepared with sentences omitting words which may or may not be synonyms. An individual is asked to go through the cards guessing the missing word and sorting them out into several “word” piles, one for each type of missing word. The discriminability with which words omitted are sorted into different piles is a measure of their semantic closeness, and an individual’s capacity to extract and utilize such context information. (The Charles' Sorting Task is described in Charles, W.G., The categorization of sentential contexts. Journal of Psycholinguistic Research, 1988. 17(5): p. 403-411 and Charles, W.G. and G.A. Miller, Contexts of antonymous adjectives. Applied Psycholinguistics, 1989. 10(03): p. 357-375.)

The Charles Sorting Task has so far not been used by autism researchers. The approach given here would suggest performance in it in those with autism would be considerably impaired both in tasks using words, and versions in which in the place of sentences, cards describe episodes are made up of nonlinguistic entities related to life experiences, in particular, social interactions.

A further possibility is that such cards could form the basis of a method for enhancing better adaptive life and interpersonal skills in those with autism. Several kinds of therapeutic task can be imagined using cards re prepared with sentences or cartoon illustrations describing situations some of which are context similar and others which are not. This could be done in several ways both to exercise what remaining higher-order abilities exist and to reduce any overcompensation upon first-order ones.