In 1988, a three-year-old child is led into a brightly colored testing room in a psychology department in Bloomington, Indiana. A small toy is brought out and put onto a table in front of the child. The toy was wooden, blue, about two inches square, and U-shaped. "This is a dax." The researchers picked a word that was easy to pronounce, but was definitely something the children had never heard before. Then, seven other toys were brought out. Some of them were the same shape as the DAX, but larger, or smaller. Some were the same size, but were made of cloth or sponge instead of wood. Some were the same size and texture as the original DAX, but different in shape. One by one, the researcher picked up the toys, and asked, "is this a dax?" Sometimes, the researcher would hold up two of the new toys at the same time, and ask, "which of these is a dax?" ("Dax" or "dax" is used to indicate the label, while DAX is used to indicate the object.)
The experiment, conducted by psychologist Barbara Landau of Colombia University (now at Johns Hopkins), with Linda B. Smith and Susan S. Jones of Indiana University, was one of the first experiments that indicated which features of objects young children pay attention to when learning their names. The children (and, in later experiments, adults) generalized the shape of the DAX, rather than its size or texture. Any toy that was roughly U-shaped could be considered a DAX. While there is tremendous controversy among developmental psycholinguists over where the shape bias comes from, it is thought that the bias explains the primary way through which human children learn to apply labels to entire classes of objects (for example, generalizing the class of objects known as cups from learning about a particular cup). And children do this quickly and efficiently. They can learn the label for an object after just one casual instance thanks to a process called fast-mapping. Starting in the second year of life, the typical English-speaking child can add approximately ten new words each day to her lexicon until she reaches an average vocabulary of sixty thousand words by the time she graduates high school.
Humans aren't the only ones who can learn the labels for objects with such ease: dogs do it too. Rico is a border collie who knows the labels for some two hundred different objects. When Rico was asked to fetch a word that he had never heard before, he automatically inferred that the new word belonged with the object he had never seen before. Chaser, another border collie, had a vocabulary of nearly one thousand words, and could categorize objects into higher-order categories, such as "toys."
Do dogs show the same shape bias as human children?
Now, Emile van der Zee and colleagues from the University of Lincoln in the United Kingdom, set out to determine whether dogs and humans generalize words in the same way. That dogs learn words nearly as efficiently as human children is indeed impressive, but, van der Zee reasons, there is one important aspect of word learning in humans that hasn't yet been demonstrated in dogs: the shape bias. He writes, "Although these findings seem to suggest qualitative similarity in word comprehension in the dog and in humans, the presence of a well-established feature characterizing the quality of human word comprehension has so far not been investigated in the dog: a bias to link the meaning of words referring to objects to object shape."
To see whether dogs classify objects as humans do, van der Zee returned to the original 1988 experiment conducted by Landau, Smith, and Jones. Only this time, the experiment participant was a five year old border collie named Gable. Gable's owners proudly asserted that their dog knew the labels for fifty-four different objects. To make sure that Gable wasn't simply a canine Clever Hans, the researchers knew that their first experiment had to establish that the dog really did possess a large vocabulary. He correctly retrieved forty-three out of fifty-four objects, an impressive eighty percent accuracy rate. He only selected an incorrect item on five occasions. On the remaining six trials, he did not retrieve an item at all, and instead hesitated, whined, or appeared to wait for further instructions.
With that aside the researchers set about replicating the original DAX experiment, with a few modifications. The DAX that they used was bigger than the original 1988 version, so that it wouldn't provide a choking hazard. Also, rather than using sponge as one of the alternative textures, as this too would have been a choking hazard, they used a variety of cloth textures.
After being taught the relationship between the word "dax" and the object DAX, Gable was familiarized with each of the six alternative objects. He was presented with ten pairs from within the set of seven objects, and simply asked to retrieve the "dax." Unlike the human children and adults, who generalized "dax" to objects of similar shape as the original DAX, Gable generalized "dax" to objects of similar size in each of the ten trials he was given. That is, he ignored shape and texture, relying instead on size to determine which objects could be considered DAXes. Rather than displaying a shape bias, Gable displayed a size bias.
To rule out the possibility that Gable's size bias could be explained simply by an innate preference for smaller items instead of larger items, the researchers conducted a third experiment that pit shape directly against size (removing texture as a possible distraction). They also made the original object larger. If Gable's performance could be explained by a preference for small objects, then he should prefer the small objects in this experiment as well, despite having been trained on a larger object. This time, the original object was L-shaped, and was called a "gnark."
As in the previous experiment, Gable generalized the word "gnark" to the other medium-sized objects, not to the other GNARK-shaped objects. That is, the size bias was confirmed in this experiment, and since he reliably chose medium-sized objects over the small-sized ones, the bias can't be explained simply by a preference for small objects.
The long tail of the DAX
After the third experiment, the researchers packed up and went home, taking the DAX and GNARK with them. But four months later, they returned and gave Gable and his owners the DAX to play with. Thirty-nine days after that, the researchers returned one last time. Would Gable still show the size bias for the DAX, after a 5-week-long familiarization phase, as he had just ten minutes after learning the word-object association in the first place? They repeated their experiment just as before.
In six out of six trials, Gable correctly picked the DAX out of sets of his other objects, showing that after thirty-nine days, he had firmly re-established the label-object association in his mind. When it came to the critical experiment, however, things were different. Rather than showing a size bias, he actually showed a texture bias!
van der Zee reasons that long-term word-object learning processes might lead to qualitatively different sorts of mental processes for object processing than shorter-term learning processes. More work, clearly, is required, to see whether the patterns observed in Gable can be seen in other dogs, and to more carefully characterize the reasons that the features used by shorter-term associations would be different by those used in longer-term associations. "Gable’s word generalization and word knowledge development are thus qualitatively different from those found in humans," he says, suggesting that the way that humans build their skills for "word comprehension may be distinctive compared to the domestic dog."
The differences between the way that humans and dogs associate words with objects may come down to the way that each species perceptually experiences objects in the first place. Humans rely mainly on vision for identifying objects, making object shape the most salient feature available to the mental machinery responsible for creating those associations. So when we learn the name for a new object, our cognitive systems identify the most obvious visual features of the object and they form an association. Dogs, however, may rely more on olfactory and tactile features, rather than visual ones. Since all objects in these experiments smelled exactly the same (by design), the most salient features available to Gable may have come from manipulating them with his mouth. His cognitive systems, then, may have identified features such as the size of his bite, or the texture of the object felt against his tongue, as the most salient, and used those to create word-object associations.
If van der Zee is right, then some of the structural properties of language may be constrained by the basic perceptual biology of the human mind and brain. What would language look like for a species that relied less on vision for identifying objects and more on sound or touch?
van der Zee E, Zulch H, Mills D (2012) Word Generalization by a Dog (Canis familiaris): Is Shape Important? PLoS ONE 7(11): e49382. doi:10.1371/ journal.pone.0049382
Landau B., Smith L.B. & Jones S.S. (1988). The importance of shape in early lexical learning, Cognitive Development, 3 (3) 299-321. DOI: 10.1016/0885-2014(88)90014-7
Photo of Gable and toys copyright Sally Smith. Other images, including header, adapted from van der Zee et al. (2012).