Rogers & McClelland (R&M) set the ambitious goal of accounting for a wide array of experimental findings in semantic cognition. The ability of their simple distributed connectionist model to account for such a range of phenomena is impressive. However, the authors do not consider crosslinguistic naming data at all in their book, Semantic Cognition (Rogers & McClelland Reference Rogers and McClelland2004). We believe that these data pose serious problems for the model in its current form.
One of the critical properties of the R&M model is its sensitivity to “coherent covariation” of features, for example, “has wings,” “has feathers,” “can fly” are features that coherently covary in birds. Sensitivity to higher-order inter- and intra-category features enables their model to exhibit some of the core characteristics of semantic cognition. For their model to be psychologically plausible, it is important to be able to demonstrate that there are indeed higher-order correlations in the world and that semantic categories straightforwardly map these correlations. In the examples that R&M consider, it seems likely that this could be so – plant and animal categories, for instance, appear to be highly constrained in the ways that the authors outline (see also Malt Reference Malt1995). But for many other categories it is not clear that coherent covariation alone will account for the facts.
Consider event categories as an example. Recent cross-cultural research suggests that there may be coherent covariation among features of events, as there are for object categories. But despite this, the specific instantiation of categories found in different cultures vary substantially. Majid and colleagues (Majid et al. Reference Majid, Bowerman, van Staden and Boster2007a; forthcoming), for example, analyzed naming data collected from 28 typologically, genetically, and geographically diverse languages. Using correspondence analysis, Majid et al. (Reference Majid, Bowerman, van Staden and Boster2007a; forthcoming) found that there were a small number of dimensions that accounted for the semantic categories of cutting and breaking across languages. The first, and most important of these dimensions, was a continuous one that distinguished events where the location of separation in the object was predictable from those where the location was unpredictable (roughly corresponding to “cut” and “break” events). But despite the fact that all languages loaded very highly on this, and the other, dimensions, individual categories from specific languages were very different.
To illustrate this, compare the sheer number of categories used for this domain from two of the languages of the sample – Tzeltal, a language spoken in the Highlands of Mexico, and Yélî Dnye, a language spoken on Rossel Island, an isolated island in Papua New Guinea. Tzeltal speakers used more than 50 different verbs to describe the 60-odd videoclips (Brown Reference Brown2007), while Yélî Dnye speakers used only three (Levinson Reference Levinson2007). Or consider a more detailed contrast – Dutch, Swedish, and Mandarin have a specific category used for events of “cutting-with-scissors,” whereas most other languages lump these events with other events of predictable separation.
How can it be that languages correlate very highly on the dimensional structure but vary so much in the specific categories that they have? One way to reconcile these findings is that the dimensions uncovered by the correspondence analysis are exactly those exhibiting coherent covariation among features. For cutting and breaking events in the real world, there is a tight correlation between the kind of instrument used, the object it is used on, the manner in which it is used, and the end state of the object. Sharp instruments, such as knives, are typically used with rigid but pliable objects, in a deliberate manner to achieve a clean separation. Across the board, languages are sensitive to these regularities, just as R&M propose. But R&M under-emphasize – and perhaps under-appreciate – that different languages come to different conventionalized solutions about how to make reference to these constellations.
For example, English speakers have a hierarchical system of verbs that they can call upon when deciding how to name an event. The same event could be labeled cut or slice, dice, chop, or break and snap, smash, and so on. Swedish speakers, on the other hand, do not have this option open to them. In Swedish, cutting and breaking verbs appear to be organized in a flat structure. There are no general superordinate verbs like cut and break. As a consequence, Swedish speakers are more consistent namers for these events since there are fewer options for how to label a particular event in their language. English speakers are less consistent because there are equally good alternatives available (Majid et al. Reference Majid, Gullberg, van Staden and Bowerman2007b).
This kind of variation in naming highlights the complex interplay of world, concept, and word. Convergent evidence from domains as diverse as event representation (such as the data mentioned above), color (e.g., Kay & Regier Reference Kay and Regier2003), and artifacts (e.g., Malt et al. Reference Malt, Sloman, Gennari, Shi and Wang1999) suggest that names may indeed be constrained by perceptual attributes, but there is still much variation between languages. It is not clear how R&M's model will be able to account for this sort of variation.
R&M hold an ambivalent position about the role that language plays in categorization. On the one hand, they acknowledge that “experience with spoken language may play some role in concept acquisition prior to the infant's ability to produce speech” (Semantic Cognition, p. 145). On the other hand, they stress that in their model “structure arises from the pattern of covariation of properties of objects, and does not depend on the explicit labeling of the objects” (p. 69). If categories are being created independently of linguistic experience, then how do R&M account for how children come to acquire the particular linguistic system of their community? If language is being used from the earliest phases of acquisition, then R&M need to be more explicit about how labels interact with “coherent covariation.”
As we said at the outset, the range of phenomena that R&M tackle are impressive. In their own words, they wish to account for semantic tasks “that require a person to produce or verify semantic information about an object, a depiction of an object, or a set of objects indicated verbally (e.g., by a word)” (Semantic Cognition, p. 2). We have our doubts that a single model can really do it all, given that there is evidence that the representations underlying naming and object recognition are distinct. For example, Malt et al. (Reference Malt, Sloman, Gennari, Shi and Wang1999) have demonstrated that object naming and nonlinguistic sorting are only weakly correlated, suggesting that they are at least partially independent.
To summarize, R&M need to illustrate how their model can be appended to accommodate these findings or acknowledge whether another mechanism is needed.
