Commentary on Michael Arbib

Word counts:

            abstract 60

            main text 1456

            references 193

            entire text (total) 1709

Title: The “Complex First” Paradox: What’s Needed to Lexicalize Neurally Distributed Meanings?

Author name: Markus Werning

Mailing address:

Department of Philosophy

Heinrich-Heine University

Düsseldorf, D-40225

Germany

Phone: +49-211-81-11473

Email: werning@phil.uni-duesseldorf.de

URL: www.phil.uni-duesseldorf.de/thphil/werning

Abstract:

From studies on language acquisition and typology it can be inferred that there is an ontogenetic and phylogenetic priority for the lexicalization of complex concepts, whose neural correlates are more widely distributed in the cortex than those of primitive concepts. It is argued that this poses a problem for the explanation of how an appropriately distributive syntax-semantics interface could evolve.

Main Text:

One of Arbib’s main claims is that, once the human brain had biologically evolved to support the features of a protolanguage (LR1 – LR7), the additional requirements for the possession of a real language (LA1 – LA4) were a matter merely of social development and required no further change in the biological hardware. Among these “soft” features of language, Arbib (with some caution) claims, are symbolization and compositionality, i.e., the fact that ”symbols become words in the modern sense, interchangeable and composable in the expression of meaning” (p. 13). I would like to highlight a problem that concerns the neural syntax-semantics interface and directly address the issues of symbolization and compositionality. It isn’t a problem particular to Arbib’s approach, but a problem for any neurobiological approach towards semantics. I will, however, ask whether the story Arbib tells on the biological evolution of protolanguage provides a framework for an adequate solution.

When one is concerned with the search for the neural correlates of linguistic meanings that are lexical, i.e. verbalized by syntactically unstructured expressions, one is confronted by what I would like to call the “complex first” paradox: The paradox arises from the fact that substance concepts are more frequently lexicalized across languages and their lexical expressions are ontogenetically and (probably) phylogenetically earlier acquired than is the case for attribute concepts. This is not what one would expect from the fact that, according to widely held views (e.g. Barsalou, 1999), the prototypical (first-level) attribute concepts are semantically primitive and the prototypical substance concepts are semantically complex. According to those views, the substance concept [mango], e.g., is made up of the vector of attribute concepts <orange, oval, big, soft, sweet, edible, ...>. This implies that the neural correlates of substance concepts should be widely distributed in the cortex, whereas those of attribute concepts should be relatively local. At first glance the effort for a syntax-semantics interface to address attribute concepts should be less than that for addressing substance concepts. From this point of view, it is surprising that the latter nevertheless are lexicalized first.

Substance concepts serve to re-identify things over time despite of their contingent changes of attributes and so allow us to gather, store and update information in a systematic and enduring way (Millikan, 1998). They are typically expressed by concrete nouns — in English, e.g., by names of individuals like mama, names of basic kinds like mouse and names of stuffs like milk. The great mass of children’s earliest words are concrete nouns. During the so-called naming explosion, when children around 18 months of age first systematically organize their concepts by means of a lexicon, they preponderantly pair substance concepts with concrete nouns, whereas the assignment of adjectives (red, vertical) and abstract nouns (redness, verticality ) to the concepts they express, viz. attribute concepts, comes much later (Ingram, 1989). Some languages even don’t have adjectives or just a closed set of them (Dixon, 1999), while the class of concrete nouns is (arguably) universal and always open.

If one shares the widely held assumption that the understanding of a word must finally result in the (synchronized) activation of neuronal assemblies in the sensomotoric cortices (for review see Pulvermüller, 1999) and if one identifies those assemblies with the concept the word expresses (Werning, 2003, 2004), the priority for the lexicalization of substance concepts calls for an explanation. For, then the neural correlates of prototypical attribute concepts would be constrained to local feature maps as they occur for color concepts in V4 or for concepts of orientation in area 17 of V1. The correlates of substance concepts, in contrast, would be widely distributed over the cortex. We know that the understanding of concrete nouns for tools like hammer, e.g., results in an activity distributed over the premotor and the visual cortex (Martin, Wiggs, Ungerleider, & Haxby, 1996). The assertion that words for substance concepts arouse more widely distributed activity in sensomotoric cortices than words for attribute concepts is, furthermore, supported by a study of Rappelsberger, Weiss, and Schack (2000). They compared the temporal coherence of EEG traces across different regions of cortex that result from the processing of concrete and abstract nouns. For concrete nouns they measured much higher and more widely distributed temporal coherence, than they did for abstract nouns.

Does Arbib’s account provide a framework to explain the ”complex first” paradox? Arbib indeed sympathizes with the view that the first syntactically unstructured signs were semantically complex. Their semantic values, he thinks, were in fact even more complex than those of concrete nouns and represented entire situations. His approach might also explain why there is a stronger evolutionary pressure to lexicalize concepts as complex as those of substances than to lexicalize the less complex attribute concepts. It evidently is rather economic to lexicalize concepts for often recurring, highly specific entities of great survival value. Telling someone that there are mangos nearby is not only shorter, but also more exact than to tell someone that there are orange, oval, big, soft, sweet, edible things around.

What Arbib does not answer though is the following question: How could a mechanism evolve that enables certain regions of cortex that are involved in representing a word (phonologically, as part of a syntactically more complex expression, etc.) to address those regions of the sensomotoric cortices that represent the word’s meaning, i.e., the concept it expresses. Given that semantically complex words are evolutionary prior such an interface towards semantics (it’s sometimes attributed to Wernicke’s area) must have evolved at an early stage in the evolution of language and it must have had strong distributive capacities from the beginning.

It seems to me that Arbib does not at all deal with the problem of addressing meanings. For him the phenomenon of meaning, i.e., the fact that certain signs (words, gestures) stand for certain other objects, events or properties, apparently is not to be explained neurally by a mechanism of addressing, through which neuronal representations of words (or gestures) in Broca’s, Wernicke’s and related areas lead to an activation of the neuronal representations of objects, events and properties in the sensomotoric cortices. His view, rather, seems to be that gestures that were originally neurally represented in F5 (the proposed primate homolog of Broca’s area) and related regions for the purpose of imitation became conventionalized such that those neural representations ceased to be representations of those gestures alone, but of situations that were by convention linked to these gestures. Later on those “meaningful” gestures were substituted by “meaningful” vocalizations, which finally became words and sentences with full-blown semantic values. In Arbib’s framework, still, meanings would remain situated in F5/Broca and closely related areas. There would be no need to propagate activation from regions involved in the processing of the syntax and phonology of expressions to sensomotoric regions that host the concepts semantically expressed by them. But this can’t be true, I suppose. For, meaning comprises empirical content. The capacity to understand the meaning of the word mango must be inherently linked to the capacity to perceptually recognize mangos. F5/Broca, however, is not at all involved in the recognition of mangos. What’s involved here are the sensomotoric cortices. These must be addressed somehow when the word mango is understood.

In the literature on mirror neurons it is often suggested that, since mirror neurons (in F5 and other regions) are sensitive to certain action-object frames, say the grasping of a mango, the neurons represent the grasping of a mango and can hence be identified with the neural correlate of the (unsaturated) concept [grasp(X, mango)]. But this identification can’t be true, either. For, the concept [grasp(X, mango)] is semantically composed from the concepts [grasp] and [mango] plus some specific syntactic structure. The complex concept hence contains the concept [mango] as its constituent. The neural correlate of this constituent concept, however, can’t be located in F5 (etc.) for the reasons mentioned above. Given that conceptual representations are compositional, the neural correlate of the composed concept [grasp(X, mango)], consequently, can’t be located exhaustively – i.e., with all constituents – in F5, either. What strikes me as a better hypothesis about the representational function of those mirror neurons is that they represent the complex syntactic structure needed to generate [grasp(X, mango)], something like [grasp(X, Y)] or even [Z(X, Y)]. The fact that a single mirror neuron prefers mangos over bananas as the Y-argument might be a contingent fact without any representational function, just like a redness neuron in V4 still is a representation of redness rather than redness-on-the-upper-left, although it is sensitive only to red objects on the upper left corner of the visual field. The alternative hypothesis might then, too, explain why F5 is indeed the primate homolog of Broca’s area, which, as we all know, has traditionally been thought of as a syntax processing unit.

References:

Barsalou, L. W. (1999). Perceptual symbol systems. Behavioral and Brain Sciences, 22, 577-660.

Dixon, R. M.W. (1999). Adjectives. In K. Brown, J. Miller, & R. E. Asher (Eds.), Concise encyclopedia of grammatical categories. Amsterdam: Elsevier.

Ingram, D. (1989). First language acquisition; method, description and explanation. Cambridge: Cambridge University Press.

Martin, A., Wiggs, C. L., Ungerleider, L., & Haxby, J. V. (1996). Neural correlates of category-specific knowledge. Nature, 379, 649-52.

Millikan, R. G. (1998). A common structure for concepts of individuals, stuffs and real kinds: More Mama, more milk, and more mouse. Behavioral and Brain Sciences, 21, 55-100.

Pulvermüller, F. (1999). Words in the brain’s language. Behavioral and Brain Sciences, 22, 253-279.

Rappelsberger, P., Weiss, S., & Schack, B. (2000). Coherence and phase relations between EEG traces recorded from different locations. In R. Müller (Ed.), Time and the brain (p. 297-330). Harwood Academic Publishers.

Werning, M. (2003). Synchrony and composition: Toward a cognitive architecture between classicism and connectionism. In B. Löwe, W. Malzkorn, & T. Raesch (Eds.), Applications of mathematical logic in philosophy and linguistics (p. 261-278). Dordrecht: Kluwer.

Werning, M. (2004). The temporal dimension of thought: Cortical foundations of predicative representation. Synthese. (Forthcoming)