We have no difficulty picking “rat” as the odd one out from the set “goat – deer – rat”. This ready access to semantic structure in our memories supports many essential cognitive capabilities. It allows us to be guided in our current understanding and behavior by prior knowledge and experience. For example, if we learned that a noolbenger looked much like a rodent, then we would rapidly infer that—unlike goats and deer—it probably is not typically consumed by humans.
This basic semantic knowledge is disrupted in Alzheimer’s disease, a degenerative disorder of the brain that affects about 6% of people 65 years and older. The effects of Alzheimer’s disease on a person’s ability to grasp the meaning of words, and the progressive memory loss it entails, were brought to the silver screen in Still Alice, which earned Julianne Moore an Academy Award for Best Actress.
Here she explains “what it’s like” to lose one’s language skills to her daughter:
Neuroscientists have sought to understand the nature of the disruption of knowledge in Alzheimer’s disease by a variety of means. A popular technique involves the use of multidimensional scaling (MDS), which constructs a spatial “map” of semantic space from a person’s responses to the triadic comparison task—that is, multiple trials on which the person has to select the odd one out from a set of three stimuli (“goat – deer – rat”, “zebra – antelope – cat”, and so on).
A recent article in the Psychonomic Society’s journal Behavior Research Methods provides a new perspective on the modeling of the semantic spaces of people with various degrees of cognitive dysfunction. Researchers Michael Lee, Melinea Abramyan, and William Shankle obtained data from nearly 3,000 patients who were tested extensively for cognitive disorders, and who all completed a triadic comparison task involving animals.
Lee and colleagues were particularly interested in how the extent of memory impairment relates to the structure of a person’s semantic knowledge. They therefore divided the sample into 8 groups that differed in their ability to recall the 9 animals that had earlier been presented for the triadic comparison task. The best-performing group recalled all 9 animals without error, the next-best group recalled 8, and so on.
To obtain a map of each group’s semantic structure of the animal category, Lee and colleagues developed a new Bayesian procedure to construct the map from the triadic judgments. Unlike existing approaches, this new procedure yields not only the most likely structure of the space, but also indicates how much confidence one can place in it in light of the likely error in the data.
The main result is shown in the figure below, which shows the semantic representations inferred for each of groups I–VIII:
Recall that Group I exhibited perfect recall, whereas Group VIII performed worst on the free-recall task. The figure therefore provides a picture of the deterioration of people’s semantic space as their memory worsens owing to Alzheimer’s disease or other cognitive disorders.
The black dots in each panel describe the most likely location, in semantic space, of the animals in question. The gray dots provide a measure of uncertainty surrounding that best estimate of location.
Several observations can be made. First, as the severity of memory impairment increases (from Group I through Group VIII), the uncertainty surrounding the location of the stimuli also increases. Second, as severity of impairment increases, the meaningful clustering of the semantic space is eroded. Thus, whereas Group I puts simians and apes (monkeys, gorillas, and chimps) into one cluster, with pets (cats and dogs) in another and assorted small furry things (rat, rabbit, chipmunk, and beaver) in yet another, this differentiation is entirely lost in Group VIII.
Intriguingly, this progression from clear structure to an undifferentiated representation is not uniform: The simian-ape cluster retains its identity for all but the last two groups, whereas the differentiation between the remaining clusters has already almost disappeared for Group V.
Lee and colleagues went on to model performance at the level of individual patients. To do so, they had to assume that all persons shared a common underlying mental representation of the animal category, but that they differed in their ability to access that representation. The assumption of a common representation is justified by work in cultural anthropology which has presented evidence of the constancy of MDS representations of animals.
Individual variation was captured by a parameter that determined how people chose their responses in the triadic comparison task. The details need not concern us here, except to note that when this parameter tends towards zero, responding becomes random. Conversely, when this parameter tends towards positive infinity, responding becomes completely deterministic, which in turn implies that people’s responses adhere entirely to the underlying MDS representation of the category.
The figure below shows the estimated values of that parameter for all patients in the sample broken down by group. In each panel, the gray histogram is for the overall sample of nearly 3,000 individuals, and the black histogram characterizes the participants in that particular group (Note that the group order here is reversed, running from poorest (VIII) at the top to best (I) at the bottom):
It is clear that people whose memory functions well are predominantly responding “deterministically”—that is, their responses conform to the underlying canonical MDS structure of the category. People with a severe memory impairment, by contrast, predominantly perform in a manner that is not much different from random—their parameter estimates cluster around the lower end and often tend towards zero.
Perhaps this parameter captures Alice’s inability to reach the words that she is trying to say, and that she remembers are out there somewhere.
Lee and colleagues suggest that their modeling work yielded insights into the semantic disruption associated with Alzheimer’s disease and other cognitive dysfunctions that were not available by inspection of the data alone. Given that the individual parameter values were estimated on the basis of only 12 responses to triads (involving only 9 animals), those estimates could be obtained even in a primary care setting. The potential for model-based clinical data analysis and diagnosis is promising indeed.