Note that the above summary of the central claim makes no mention of language. Clearly, non-human animals do not think in language, and much human thought is also not in a public human language, such as English or Chinese. A recurrent theme in many commentaries (Anderson & Oates, Politzer, Jones) was the association of `predicate' with something essentially linguistic. The predicates discussed in the target article are not linguistic; they are pre-linguistic. Some commentators (Dessalles & Ghadakpour, Politzer) usefully distinguished between two senses of `predicate', a linguistic sense and a non-linguistic sense; these commentaries are thus helpful in clarifying what the target article was (not) about. Jones also noted the different senses of `predicate', and constructively suggested ways to bridge the gap between theories designed to account for the different phenomena which have been labelled `predicate'. Other commentators made no such distinction, and wrongly assumed that the `predicates' discussed in the target article are identified or closely associated with linguistic categories.
Language is used to express thoughts. As formal logical schemes were developed largely with human languages as a model, there clearly is a correlation between such schemes and human languages. This makes the over-readiness to identify predicates as linguistic entities understandable. All that the central claim of the target article requires is that it be possible to spell out the nature of the mapping between representations in a logical scheme with the predicate-argument asymmetry at its heart and the grammatical structure evident in languages. The relationship between the surface structure of languages and the representation of thought is so complex and indirect that it will often mask any direct correlation between the surface linguistic structure and the dorsal/ventral separation.
The target article concentrated exclusively on the brain activity involved in perception. It said nothing about how perceived events are stored in memory, except to imply that the predicate-argument asymmetry is presumably not lost, though perhaps somewhat transformed, in this process. When people speak, they mainly express propositions dredged up from memory or generated creatively by recombining elements from different memories. Only very rarely does a human utterance describe what the speaker is perceiving at the very moment of the utterance. Several commentaries implicitly discuss the nature of the representations at the beginning of the utterance production process or at the end of the utterance interpretation process. Such representations are variously called in the literature `semantic representations', `intentional/conceptual representations', `Logical Form' and so forth. Without exception, all proposals for such representations make use of predicate-argument structure. The ubiquity of predicate-argument structure in such representations calls for an explanation, and I attempted to provide one.
Occam's razor dictates that we should try to find theories which minimize the gaps between the boxes in Figure 1. I doubt whether any of the boxes can be eliminated -- each seems necessary on independent grounds. Commentaries by Cowie and Jones, in their very different ways, respectively philosophical and psychological, address the task of bridging the gaps between the boxes in Figure 1.
Cowie asks `What are the processes of derivation, abstraction, inference, binding, (what to call them?) by which these initial [PREDICATE(x)] representations are manipulated or transformed into the kinds of thoughts expressed by our propositional attitude ascriptions?' Dead right -- but (speaking of the dead) I do not think that this empiricist research programme has been so dead since the 1950s as she supposes. `Empiricism died a death, and Rationalism -- the idea that there's stuff in our minds that doesn't come from the senses (I guess it must be innate) -- emerged as the dominant theoretical orientation throughout the cognitive sciences.' This is too dramatic: a winner-take-all Empiricism versus Rationalism scenario is as implausible as the simplistic Nature versus Nurture `debate'. In my proposal, the actual distinction between (perceptual) predicates and arguments is innate, and doesn't come from the senses. The neuroscientists' discovery of the dorsal/ventral separation simultaneously gives the empiricist philosophical programme of relating sense-data to higher-level representations a boost and a cautionary admonition. The innate mechanisms formalizable as PREDICATE(x) are constitutive of our sense-data, rather than `coming from the senses'.
Jones also well appreciates the nature of the gaps between the boxes in Figure 1. His own psychological research has concentrated on the middle and right-hand boxes, those more involved with language, and he has developed a theory which distinguishes between predicates according to their `ease of predication'. He asks `Are the predicate-argument and ease-of-predication approaches compatible? Having been cantilevered out from opposite sides of the cognitive landscape -- respectively from perception/action and from knowledge/language -- it seems reasonable to hope that these two stretches of predicational bridge can be made to join up in the middle.' I agree, and Jones' analysis of the nature of the problems involved seems acute and accurate.
I have appealed to the idea that predicate-argument structure is somehow transformed, but not beyond recognition, in the translation from perception to memory and thence to underlying linguistic structure. The target article emphasized the logical asymmetry between predicates and individual terms. Several commentators, some apparently quite casually, and some with more deliberation, use formulae which disregard the basic asymmetry, putting predicate terms inside the brackets reserved for argument slots. Examples are `bite(man,dog)' (Dominey) and `herd(deer)' (Lu & Franceschetti).
Another example, at first sight rather confusing, is Gillett's ` ... in the PREDICATE(x) type thought «that frog is bright orange», «that frog» focuses on and tracks an object, and «bright orange» links a feature of the frog to other stimulus arrays instancing that colour.' Here, Gillett has used the predicate term frog as an argument of the other predicate orange. This only coheres in my terms if the proposed representation is not at the perceptual level, where, I hold, representations of the form PREDICATE(x) apply. At the perceptual level, I would represent the experience Gillett describes as FROG(that) & ORANGE(that) (innocuously using `that' for `x'). What is intriguing and challenging is that representations at a level closer to linguistic form preserve a certain amount of the original asymmetry, but allow the replacement of the argument variables by elements otherwise corresponding to predicates. A theory predicting which predicates can thus `migrate' to the argument slots inevitably raises the issue of the origins of the grammatical noun category, since essentially only nouny predicates have this privilege. You can't say That bright orange is frog, at least in English. Noun/verb issues recur in these commentaries and are taken up again several times later in this reply.
I sympathize with Bickerton's warning that we don't want to get back to unjustifiably baroque underlying structures for simple sentences. But research has moved on since the heyday of generative semantics, and the target article was concerned with issues other than simply deriving sentences. In particular, the target article focused on perception. As Figure 1 is intended to make clear, there are gaps to be bridged between the representation formed on perceiving an event, the representation of the remembered event in memory, and the structure underlying a sentence used to describe the event. Now, on seeing Floyd breaking some glass, an observer's immediate representation might resemble [BREAK(e) & FLOYD(x) & AGENT(x) & GLASS(y) & PATIENT(y)] (always with the caveat that these are non-linguistic predicates). When this event is transferred to long-term memory, it is possible that this representation is transformed into something closer to the linguistic form of Floyd broke the glass. It is to be hoped that future research, from both ends of the subject, linguistic and neuroscientific, can shed some light on this question.
Knott also raises the matter of the degree to which the syntactic structures of sentences preserve some characteristics of the raw PREDICATE(x) structure. He argues plausibly that such semantically empty grammatical elements as the there in There's a cup on the table play a similar attention-directing role to the `x' variable in PREDICATE(x). Indeed the etymology in this case is highly suggestive of attention being drawn to a location. This is a topic which deserves a lot more thought, following the idea that when people talk to each other, they are often attempting to recreate experiences in their hearers. Recalling Gillett's example, if, in the absence of the frog, you want to report your FROG(x) & ORANGE(x) experience to me, you use the grammatical subject slot as a surrogate for a deictic act of pointing, filling in with a word best calculated to make me think of the right thing. Just why the best term for the purpose is frog and not orange again raises the noun/verb issue.
Knott's commentary attempts to draw closer parallels than I had envisaged between the structure of perceptual events and grammatical structure. I had relegated the grammatical definite/indefinite distinction to a discourse function, only relevant for communication between people, and playing no part in the solipsistic representation of perceived scenes and events. Knott's idea that there might be some correlate of a definiteness marker in solipsistic representations, indicating something like `entity already known about' is worth following up. The appropriate psychological question would seem to be whether there is any detectable difference between attention directed to a new object for the first time and attention re-directed to an object very recently attended to in the current scene. For example, the kind of attention I might give to a plate, knife, fork and spoon unexpectedly laid out on a sidewalk might differ from the kind of attention I would pay to these same objects while manipulating them at the dinner table. Brinck's distinction, in his commentary, between objects of attention as either causes, attracting attention, or effects, that is products of focal attention, may be closely related to Knott's idea. (See below for discussion of Brinck.)
Bridging gaps involves taking advantages of similarities and accounting independently for differences. If two structures have everything in common, there is no gap to bridge -- they are in fact one and the same structure. Bridgeman states that the central claim of the target article `misses the mark because there is a tight logical relationship between subject and predicate; but information in the two visual streams can be independent and even contradictory'. Note first that `subject and predicate' immediately suggests specifically linguistic structure; the target article acknowledged the differences between perceptual processes and linguistic structure, but claimed that a similar skeletal shape, the asymmetric interaction of elements (processes or terms) of different types, is evident in both. If an archeologist claims that a wooden medieval house was built on the stone foundations of a Roman villa, he may do so on the basis of significant coincidences in the shapes of the separately identified relics. He excavates past the medieval wood to the Roman stone and finds the same basic layout. The archeologist's conclusions do not `miss the mark' because there are obvious differences between the two structures -- different materials, at different levels. Metaphorically, the target article was an excavation from the surface structures of languages (dug through pretty quickly), through the logical structures often called `semantic representations' down to a deep level of mechanisms of visual perception. As with the archeological example, we find a similar layout, in different materials, at the different levels.
I was expecting a more sympathetic hearing from Bridgeman, and his commentary seems to have missed some parts of the target article. Bridgeman emphasizes the possibility of contradiction between information in the dorsal and ventral streams. In fact, the target article, in its discussion of blindsight, also acknowledged this; it is one of the characteristics of the perceptual level lacking a clear parallel at the linguistic level. Bridgeman was critical of the use of the terms `what' and `where' for the two pathways. This usage, however inappropriate, is now commonplace, and the target article stuck with it, while explicitly noting that the terms `motor-oriented' and `cognitive' suggested by Bridgeman himself (Bridgeman et al. 1979) are preferable. This terminological point was picked up by MacNeilage & Davis. Bridgeman's critique brackets the target article with an earlier paper by Landau and Jackendoff. In fact, the target article explicitly distanced itself from L&J's paper, and identified more closely with Bridgeman's own commentary on that paper, emphasizing, in Bridgeman's words, `a level that differentiates linguistic from non-linguistic coding'. In this connection, Piattelli-Palmarini & Harley also, in mentioning the `chasm between closed class and open class lexical items’ missed the important distance between the target article and L&J's paper, which it explicitly criticized. Bridgeman also attributes the development of logic to `linguists', and rather generally conflates logic and grammar. He should hear how linguists gossip about logicians -- a remark of Bickerton's, quoted below, is an example.
Woll's commentary also raises, in a different way, the gap between perception of events, memory for events, and processing of sentences describing events. Woll wonders why, given my model, the parietal lobes are not involved in linguistic comprehension tasks, especially those which require spatial representational resources. This was briefly addressed in the last paragraph of Section 4 of the target article, but some further comment will be helpful, I hope. My claim is about perception -- the basic event of having the attention drawn to some object, and then making a judgement about it. The target article said very little about the further processes by which such PREDICATE(x) experiences are transferred to memory, perhaps later to be resurrected when talking about such experiences. Clearly the parietal regions need to be busy all the time guiding the attention to objects in the here-and-now. If they were still also active in linguistic production and perception of sentences describing scenes experienced (much) earlier, there could be dysfunctional interference with the second-by-second running of the bodyparts involved in attention shifts (saccades, hand and head movements).
The following analogy may be useful. Consider a robot programmed to roam the world, taking and storing digital photographs of kinds of things that its programmers have determined. Such a machine would have mechanisms for directing its lens, zooming, focussing, and adjusting exposure for light conditions. Having directed, focussed, and adjusted exposure, then `Click!' -- the photo is taken, and downloaded (say as a JPEG or GIF) to its memory. If the machine as also programmed to search its data-bank of photos and to provide descriptive summaries of what they contain, there is no reason to suppose that the lens-directing, zooming, focussing and exposure-adjusting mechanisms will be involved in this latter task. No analogy is perfect, and probably the brain is not organized in quite the cleanly modular way of this machine: we may not be surprised to find some small residual parietal activity in parsing sentences, especially those about concrete objects in physical space.
Dessalles & Ghadakpour address the difference between `non-linguistic predicate' (their `R-predication') and `linguistic predicate' (their `C-predication'). The target article, while not dwelling on the evolutionary step(s) from one to the other, may have given the impression that a single step, the development of a labelling capacity, is all that is required. Dessalles & Ghadakpour's point about the human capacity for negating and contrasting predicates is important. An animal may be able to represent the concept APPLE, but can it mentally apply a negation operator to get NOT-APPLE? Dessalles & Ghadakpour assume that only humans are capable of this, and we are surely vastly better at it than non-humans. But there is some evidence that even Alex the parrot can approximate a kind of contrast or negation (Pepperberg, 1999). Dessalles & Ghadakpour claim a qualitative difference between R-predication and C-predication, but also concede that object recognition and categorization may be a likely prerequisite of human predication. On the possibility of continuity in the evolution of linguistic predicates from non-linguistic predicates, see Gillett's commentary, discussed briefly later in this reply.
Dessalles & Ghadakpour point out that computer simulations such as Batali's (2002) and Kirby's (2000) have predicate-argument structure built in, and this enables the simulated populations of agents to evolve language-like systems. Animals haven't evolved language, Dessalles & Ghadakpour argue, so they can't have predicate-argument structure built in. But an important factor is omitted from their argument here. Batali's and Kirby's simulations did indeed have predicate-argument structure built in to the agents (and Batali's PREDICATE(x) semantic representations pre-dated my use of them). But the agents also had symbolic ability, the capacity to acquire arbitrary labels for their inner concepts. Both this symbolic capacity and the built-in predicate-argument structure are necessary to get such language-like systems to evolve. So my reply to this point of Dessalles & Ghadakpour's is that what is missing in animals is not predicate-argument structure, as they claim, but the rampant symbolizing capacity of humans.
While conforming to the `predicate-as-essentially-linguistic' tendency, Anderson & Oates make a different point, which prompts a useful clarification. Their title is `Pre-linguistic agents will form only ego-centric representations'. Predicates, as Anderson & Oates understand the term (their `genuinely objective predicates') are shared by all members of a community, whereas prelinguistic categorizations are (or may be) idiosyncratic and subjective, and thus `egocentric'. I agree. Anderson & Oates argue that the (possible) egocentricity of prelinguistic categorizations, as delivered by the ventral stream, undermines the crucial difference that I rely on between the two information streams. The target article stressed the deictic nature of the variable x argument in PREDICATE(x). It is important to distinguish between `deictic' and `egocentric'. Anderson & Oates are right to say that the agents I envisage are `functionally solipsistic', in that social interaction played no part in my story. But these creatures have a history, and a memory of past events, and a future, and some capacity to plan for it. Animals' mental representations of their pasts and their futures may or may not be functionally solipsistic, but their existence certainly extends beyond the here-and-now. `Deictic' implies `pointing to what is here now'. When a creature attends to some object in the here-and-now and categorizes it in such-and-such a way, the category is one that it has carried around in its head ever since it acquired it (or was born with it), and will probably carry around until its death. The application of a prelinguistic predicate to an object combines ephemeral deictic attention with lasting (if possibly egocentric) mental categories. Prelinguistic predicates may be egocentric, but they are not deictic.
In fact, it is surely likely that prelinguistic creatures of the same species have significantly similar predicates, due to shared adaptive heredity. The work of Luc Steels (1997), to which Anderson & Oates refer, starts with agents which have shared constraints on the categorical distinctions they make. The existence of pre-linguistic predicates facilitates (is a preadaptation for) the later emergence, aided by language, of what Anderson & Oates call `genuinely objective predicates' and what Gillett calls `true concepts'. I share the view, clearly expressed by Gillett, that `` ... convergence in categorization with other competent language users occurs by conversational correction within a co-linguistic human group. By noticing this fact, we can, without denying the continuity between human thought and that of higher animals, bring out a point of difference which increases the power of human epistemic activity and in which language plays a central role.''
Politzer also distinguishes two senses of `predicate' and `predication', but his dichotomy differs from that of Dessalles & Ghadakpour. Politzer is centrally concerned with predicates as realized in language. ` ... in its modern, logical sense, a predicate (henceforth `predicateL') is a function from a singular term to a sentence expressing a proposition about the object to which the singular term refers'. This is somewhat close to my sense in the formula PREDICATE(x). But note that in writing about sentences, and therefore about linguistic predicates, Politzer is implicitly granting the transition from mental prelinguistic predicates to (this type of) linguistic predicates, a transition that others found problematic or at least in need of an account. The definition of predicateL, Politzer comments, is at odds with `Aristotelian sentences' (`A-sentences') which, as the target article noted, do not require the subject and the grammatical predicate (`predicateG') to belong to different kinds of terms. Examples: The philosopher is a man and The man is a philosopher.
The subjects of modern human sentences, Politzer argues, can be interpreted either intensionally or extensionally. As a piece of purely synchronic description, this may be adequate. But Politzer's suggested account of the origin of A-sentences is circular, resting implicitly on an unstated distinction between nouns and adjectives. He writes `From the inception of categorization, the A-sentence predication could start to develop, taking generic individual objects as its subject: ``x is a predator'' (with temporal anteriority) and ``x is yellow'' are conflated into ``the predator is yellow''. Indeed, it would be uneconomical to formulate, e.g. the x is S and P when the first predication is temporally or cognitively already established, hence the shorter formulation the S is P.' Politzer's appeal to `temporal anteriority' is crucial to his argument and quite unjustified. What is to guarantee that the judgement that something is a predator precedes the judgement that something is yellow? One would in fact tend to expect the opposite. Without this ad hoc appeal to anteriority, there is no way to avoid the yellow is predator. Politzer seems implicitly to have accorded `anteriority' to predicates corresponding to nouns, but gives no reason why only noun-predicates are `established first'. Likewise the psychological experiment on syllogisms to which Politzer appeals reflects an unwillingness in subjects to concoct sentences in which nouns and adjectives are in the wrong grammatical positions. Clearly, these arguments don't explain how a logically uniform class of predicates came to be mapped onto grammatically contrasting syntactic categories -- nouns versus others; the Aristotle problem is still unsolved. Finally, on a conciliatory note, Politzer may be right that certain kinds of predicate have `anteriority', although he himself gives no account of what such anteriority is. Perhaps something like Jones' ease of predication (see above) does in fact distinguish noun-predicates from others. None of this critique of Politzer's commentary affects the central claim of the target article, correlating PREDICATE(x) with the ventral/dorsal separation, a claim which Politzer finds plausible.
Anastasio comments on the inappropriateness of discrete, categorical logical representations, like PREDICATE(x), and contrasts formal logic with probability. I agree with Anastasio's main point, that a properly fine-grained account of brain activity has to be in terms of probabilities. Anastasio gives the example of what happens in the ventral stream when someone recognizes an object as satisfying the predicate APPLE, emphasizing that the relevant cortical neurons are not two-state elements, but show graded responses to their inputs. This is absolutely right, as is Arbib's similar comment that neural schemas are likelihood distributions, rather than simply returning 1 or 0. As the target article stated, the brain is vastly more complex and subtle than any formal scheme invented by a logician.
But we should not, and Anastasio does not, throw out the predicate-argument baby with the discrete, categorical bathwater. Anastasio suggests replacing a predicate-argument formula such as APPLE(x) with P(X=APPLE|S), representing the probability that an object X, eliciting sensory input S, is an apple. This models the activity of a neuron in the ventral stream. The predicate-argument distinction is still present in Anastasio's formula, in the form of the two terms APPLE and X, merely hedged around with a probability factor. (I would prefer to replace Anastasio's `equals' sign with something less suggestive of identity and conveying the relation of satisfaction between a predicate and its argument.) And Anastasio has no quarrel with the target article's correlation of predicate and argument with the ventral and dorsal streams. Logic and probability are not incompatible. Probabilistic, or `fuzzy' logics have been developed (see, for example, Zadeh and Kacprzyk, 1992). ``
The target article appealed to a broad class of semantic theories such as Montague Grammar, event semantics and discourse representation theory, which have made plausible attempts to relate quite abstract, and often typically flat, semantic representations to the surface grammar of languages. It is just these semantic theories which Dominey characterizes as `developing a theoretical basis for mapping logic to language and the meanings that can be expressed in language'. But this misrepresents the goals of these theories. They aim to map the structures found in language onto a type of well-specified external world (a `model'). They attempt to provide an account of what natural language sentences are about. Even Frege, in his way, was trying to do just this. Likewise, Bickerton attributes to me `an assumption that language and cognition must be based on logic' and an `insistence on approaching language from a logical point of view'. Carstairs-McCarthy, very similarly, claims I assume `that there must be some stage of linguistic evolution at which (proto)syntax behaved in a fashion that reflected more closely than it does now the way in which predicate-argument structure works in logic' (emphasis added). I do not assume this a priori appropriateness of logic. The work of logicians is to be respected for its rigour, and appreciated when we can discover in logic something that is useful in an account of cognition and language. I plead guilty to cherry-picking what is useful from logic. Hence my rejection of individual constants (as defined in logic) in favour of individual variables, which, I claim, can be correlated quite nicely with the information from the dorsal stream. A logical notation is not an a priori Procrustean starting point, with which semanticists wrestle to make language map onto it. This is just why Cowie sees the target article as vindicating the predicate-argument asymmetry in logic, rather than the other way around. Nowadays the logical notations that semanticists use are theoretical constructs designed with the goal of relating sentences to the world they describe (and in some cases to further functions of sentences as well). They are driven by language -- not by the structures of isolated sentences, but by the complex semantic interrelationships between whole classes of diversely structured sentences.
Dominey's proposal is to ditch the work of formal semanticists because of the quantity of effort expended. He rejects a flat structure bite(e), man(x), dog(y), agent(x), patient(y) as a representation of A man bites a dog, on the grounds that it is arbitrary, unordered and less informative than bite(man,dog). The flat structure is not `arbitrary'; on the contrary it is designed to account economically for a whole range of facts about the sentence in question, facts such as the following. The sentence is a paraphrase of others such as A dog is bitten by a man, It is a dog that the man bites, What bites the dog is a man. Of the situation described by the sentence, the question can be asked `Who (or what) acted deliberately?', eliciting the Agent information separately, and `Who (or what) was affected by the act?', eliciting the Patient information separately. The semantic structure of a sentence is not a matter that can be determined solely by looking at it in isolation. A proposed representation such as bite(man, dog) needs to be backed up by thoroughly reasoned comparison with the alternatives, taking into account a worked-out theory of what in the world the terms denote. How, for example, will Dominey treat the term man in dealing with a sentence such as Peter is a man? Will man be assigned the same denotation for both sentences, and if so what is it?
Obviously, at one end of the theoretical derivation of an individual sentence, there has to be a structure that closely resembles the sentence itself. But once we undertake to relate whole sets of sentences to each other, to the eventualities they describe, and to the psychological mechanisms that process both the sentences and the eventualities, things get more complicated. In terms of the box diagram in Figure 1, Dominey's proposed representations are closer to the right-hand linguistic end. Dominey appears to accept the case for relating neural architecture to PREDICATE(x) structure, but rejects the possibility of mapping PREDICATE(x) structures onto language. But this leaves a gap unbridged.
An analogy might help us out of our difficulty here. Getting (digital) computers to work involves converting everything down to a binary code consisting of nothing but 1s and 0s. Fortunately for users, intermediate levels of representation, such as machine code, assembly language, and high level languages, keep us from toiling with 1s and 0s. Anything can be coded in 1s and 0s, and reliably retrieved. With computers, the effort is necessary, because the basic electronics deals best with 1s and 0s. A theory of the input-output mappings of a particular programmed computer might not need to get down to the 1s and 0s, but if the mappings are at all complicated, some quite abstract underlying representations, no doubt resembling the computers' program(s), would be required. The PREDICATE(x) formula is somewhere in the middle between the firings of individual neurons (analogous to 1s and 0s) and our representations of the inputs and outputs that humans map onto each other, namely sentences and perceptions of objects and events. Just as anything, of whatever dimensionality, can be coded into a linear one-dimensional representation, even with the minimal {1, 0} alphabet, anything can be coded into flat structures of conjoined PREDICATE(x) elements. The question of whether this is the most elegant solution is open, but cannot be easily dismissed.
Dominey writes that `Prelingual infants appear to represent collisions in terms of the properties of the ``collider'' and their influence on the ``collidee''. This supports the hypothesis that contact is represented by a 2(or greater)-place predicate.' Just the contrary, surely. The target article made a brief case that 2-place predications can always be reduced to conjunctions of 1-place predications assigning properties differentially to the participants. And here are the prelingual infants conforming to that conjecture.
The question of whether (perception of) all eventualities can be represented by conjunctions of 1-place predications, or whether some 2-place predications are necessary is also raised by Werning and Arbib, but in a different way from Dominey. Like me, and unlike Dominey, both Werning and Arbib use flat conjunctions of predications with individual variables as their arguments, as in bite(e), man(x), dog(y), agent(x), patient(y). Arbib writes `I don't think this works. We need to replace agent(x) by agent(x, e) to indicate in which event x plays the stipulated role.' For Werning, too, the predicates representing participant roles, such as Agent and Patient, are 2-place, for the same reason. This is common in neo-Davidsonian event semantics. I am interested in pursuing the possibility that such 2-place predications can be eliminated by bracketing together all the predications relating to a single event, and restricting the application of such predicates as agent and patient to the local environments delimited by the brackets. The box notations of Discourse Representation Theory (Kamp, 1981; Kamp & Reyle 1993), or something like them, could perhaps be used for this role-scope-delimiting purpose.
Lu & Franceschetti, following Talmy (2000), point to the need for the components Figure, Motion, Path and Ground in semantic representations. This seems right. The target article discussed some similar examples to theirs, in Section 1.2, but without using Talmy's terminology. Lu & Franceschetti's discussion is quite informal, but there is no obvious reason why, in a more formal representation, these semantic components could not be expressed as 1-place predicates applying to the various objects in the situations and events described.
Werning clearly accepts the standard event-semantic arguments for the kind of flat-style semantic representations, with variables for arguments, that I assume. Thus we tend in the same direction in how far we distance semantic representations from the surface grammar of languages. But Werning makes one concession that I do not make to language-like structure. His semantic representations distinguish between two kinds of variable occupying the argument slots, namely object variables and event variables, and these are used to explain the universal linguistic distinction between nouns and verbs. Werning, like me, is committed to finding neural correlates for semantic representations. He, unlike me, associates predicates over event variables, giving rise to linguistic verbs, with the ventral stream, and predicates over object variables, giving rise to linguistic nouns, with the dorsal stream. I will discuss the important issue of brain areas involved in detection of motion in another section. But note here that Werning's account loses the opportunity to explain the predicate-argument asymmetry. If the SLUMP(e) of his example, paraphraseable as `There was a slumping event', is completely hosted by the dorsal stream, what accounts for the predicate-argument structure that Werning assigns to it? Likewise, if RED(x) is entirely hosted by the ventral stream, what accounts for the predicate-argument structure Werning assumes it to have? I am not claiming that my correlation of the ventral/dorsal separation must be the only possible explanation for the predicate-argument asymmetry, but it is at least one proposed explanation. Werning's account leaves unanswered the question of how, in each stream separately but in parallel, the same basic predicate-argument asymmetry arises.
Can humans reliably reidentify unique individuals? As Carstairs-McCarthy’s example of his possible confusion of Jim Hurford with Tim Hurford shows, we cannot. We are in no better position than the tern chicks, as both we and non-human animals are at the mercy of our senses for reidentifying particulars.
Do humans have concepts of unique individuals? I agree with Carstairs-McCarthy that we do. He is surely right in stating that a Matsigenka who happens to have two relatives glossed as `patrilineal same-sex cousin' will have concepts of these relatives as separate unique individuals. However, I submit that evidence for the human concept of a unique individual rests only on our human command of the meaning of unique and its translations in other languages, and our grammatical intuitions. Since we understand the meaning of the word unique, we can hardly deny that we have the corresponding concept. The only other evidence that humans have concepts of unique individuals comes from our grammatical intuitions. We intuit such facts as that Mary and herself are `co-referential' in Mary admires herself.
There is a curious tension between the answers to the two questions posed above, parallel with a tension between ideal linguistic competence and actual linguistic performance. In a previous paper (Hurford, 1999) I suggested a process by which certain perceived individuals, by virtue of constant presentation of complex combinations of properties salient and important for us, become `cognized individuals', attributed with uniqueness in our belief system. In the terms of Figure 1, certain perceptions get transferred to memory with such frequency and significance for us, that the `whichness' of the objects concerned is preserved. In the words of that paper (reidentifiable) `individuals are abstractions'. Further, by our constant use of language which presupposes and bolsters our beliefs in unique individuals, we may construct and reinforce mental representations containing a certain class of term which, we believe, is used to reidentify unique individuals.
Do non-human animals have concepts of unique individuals? The concept of uniqueness rests on a command of negation. A full command of uniqueness cannot exist without a command of negation and contrast. Dessalles & Ghadakpour argue that non-human animals lack a command of negation and contrast, and I agree that such concepts are likely to be less developed in non-human animals. There is no aspect of animals' behaviour forcing us to attribute concepts of unique individuals to them, as opposed to a more parsimonious attribution of sufficiently specific conjunctions of properties. The signing chimpanzees that Carstairs-McCarthy mentions use signs glossed as Roger and Washoe. But that does not immediately tell us that these animals had a mental conception of Roger, for instance, as unique in a way that the denotata of tree, cage, and house did not. (I keep quiet about Washoe, as there is no space to get into the thorny issue of animals' conceptions of themselves as unique or otherwise.) The question is the extent to which repeated exposure to and familiarity with certain things can give rise, in animals, to a special class of concepts of cognized individuals, each member of which is credited with uniqueness. I surmise that such concepts came late in human evolution -- maybe before something approaching full human language (and gossip), maybe not. We certainly didn't need them to stand for the forerunners of arguments in predicate-argument structures, because the availability of perceived individuals, as opposed to cognized individuals, was there from long before humans evolved. Bickerton's assertion that `they [animals] have a clear concept of a specific individual' is not backed by evidence. We may fondly believe that our pet cats treat us as special individuals until we see them sidle purring around the next house-guest who feeds them.
Do languages have proper names? Yes, obviously, most do, and some, like Matsigenka, don’t. How do we decide whether a language has proper names? The standard linguistic method for answering such a question appeals to the distribution of expressions in wellformed sentences. Bickerton's hypothetical language has two sentences: Last night the wind knocked the hut over and Knocked the hut over seduced your wife last night. Bickerton claims that in the second sentence, but not the first, knocked the hut over is a proper name. This is not, of course, enough data to go on, but we get the idea. I suppose that on similar grounds, he might conclude that the teacher is a `proper name' in The teacher should know better whereas the same expression is not a proper name in Derek is the teacher. But of course, in both sentences, the teacher belongs to the same syntactic category, NP, which is not equivalent to `proper name'. (The copula is here is not crucial, as many languages, e.g. Arabic, Russian, express predication without such a connecting particle.) Bickerton might disagree: `What determines whether something is a proper name is not its internal structure but how it is used.' This might reflect the assumption that `proper name' is a semantic, rather than syntactic, category, so in this case `proper name' means `expression used to refer to an individual'. Even in languages without proper names, it is (unsurprisingly) possible to refer to individuals. The question is whether the precursors of language had available a special category dedicated to this purpose.
(There are several odd misconceptions in Carstairs-McCarthy's commentary. His abstract states that I claim that the priority of empty variables in predicate-argument structure `had an effect on visual perception'. In fact, of course, I claimed just the opposite. He also implies that `proper names are complex to handle in FOPL'. Semanticists attempt to account for the mapping between a large body of natural language (typically English) sentences and models of a world that they describe; what they find hard to handle is dealing with linguistic proper names as straightforward equivalents of individual constants in FOPL.)
The other commenting linguists were Piattelli-Palmarini & Harley. They also discussed individuals, but in connection with a different question. They wonder why the clauses found in natural languages are typically restricted to a maximum of three or four principal participants. The target article (Section 4) attributed this directly to nonlinguistic limitations of short-term memory, as surveyed by Cowan (2001). This explanation is not addressed by Piattelli-Palmarini & Harley, who prefer an explanation in terms of language-internal constraints. But such an appeal to language-internal constraints leaves unanswered the question of where such constraints came from. I have a preference for reductionist explanations; Piattelli-Palmarini & Harley prefer a `translationist' project. I would like to know whether research programmes of this `translationist' stripe can ever explain phenomena of emergence or evolution.
Anderson & Oates join with Werning in rejecting my claim that the origins of perceptual predicates lie solely in the ventral stream. Anderson & Oates suggest that the dorsal pathway could produce representations to underlie predicates like REACHABLE(x), and the ventral pathway could produce representations to underlie predicates like RED(x). This has the same disadvantage as Werning's proposal, noted at the end of Section 5, namely that it does not provide any explanation for why the information coming through these separate streams should have the same predicate-argument format. The blindsight patient mentioned by Anderson & Oates could indeed reach accurately, a feat accomplished by his dorsal stream, but the property of reachability never got transferred upstream to mechanisms involved in reporting on events. The central claim of the target article is that only properties delivered by the ventral stream provide the predicates used in representations which, through memory, can become the basis for linguistic representations.
Werning argues that properties in the general super-category of motion are detected by the dorsal stream. Note first that there is, as Woll's commentary mentions, little evidence that dorsal stream parietal systems are activated in sentence processing, even when space is referred to in spoken language. Thus, if we envisige the diagram in Figure 1 as a kind of (perception > memory > linguistic representation) production line, there is no evidence that any dorsal stream involvement is preserved at the stage of linguistic representations.
Perception of motion and mental representation of motion properties are at present probably the most problematic area for the central claim of the target article, and clearly more research, and perhaps some revision of the central claim, is necessary. But it is becoming clear that `motion' should not be treated as a single category. I mention below a few recent studies that suggest that at least some processing of motion takes place in the ventral stream. Beintema and Lappe (2002:5661) report that `some patients with lesions to motion processing areas in the dorsal stream are severely impaired in image motion perception but can easily perceive biological motion.' Zhou et al. (2003:417) report that `Long-range AM [apparent motion] activated the anterior-temporal lobe in the visual ventral pathway, and the response varied according to form stability. The results suggest that long-range AM is associated with neural systems for form perception.' Vaina et al. (2001) report `whereas face (and form) stimuli activate primarily the ventral system and motion stimuli primarily the dorsal system, recognition of biological motion stimuli may activate both systems as well as their confluence in STS'.
Brinck makes a valuable distinction, which I largely neglected, between stimulus-driven attention and goal-driven attention. As I understand Brinck's terminology, the process he calls `indexing' only applies in stimulus-driven attention. `Not any object will be indexed, but only those that are salient enough to impinge on the subject. Indexing is caused by some property of the object, although that property will not be encoded.' I agree. Section 2.2 in the target article discussed `natural attention-drawing properties', as opposed to other kinds of properties. Brinck challenges this idea: `I do not see the need to introduce ``natural attention-drawing properties'' to account for attention-attraction'. This seems inconsistent with the quotation above about indexing being caused by some property of the object. In his penultimate paragraph, Brinck writes that attention is attracted by sudden and unexpected changes in the subject's immediate environment. If such a sudden and unexpected change is to the whole environment, like the sudden darkness due to a total eclipse, or a bright light suddenly illuminating the whole of a previously dark room, then there is no single object to which attention is drawn. But if the change is more locally constrained, almost certainly it will be a change in a property of some object, as seen from the subject's position. For example, a leaf may flutter or a door may open (I am happy with modes of movement being properties), or as the subject turns her head, redness appears, interpretable as some red object changing its position relative to the subject. Red is generally a more attention-drawing colour than brown (which helps to account for the well-known hierarchy of Basic Color Terms in languages.) The target article cited evidence that young children pay more attention to shape than to other properties of objects. It was largely stimulus-driven attention that was assumed in the target article, and I think the difference between red and brown makes the point. Some properties of objects grab attention faster and more effectively than others, and some properties of objects (such as their weight) hardly grab attention at all.
Turning now to goal-driven attention, it is only here that, as I understand Brinck's terminology, one can speak of `targets of attention'. `Goal-driven attention works top-down, in anticipation of some well-defined item. The subject is searching for a particular object.' The target of attention is, then, the defining property of the sought-for object(s). So indexing is bottom-up, stimulus-driven, while having a target of attention happens in top-down, goal-driven search. Given this, Brinck is correct in saying that indexed objects can never be targets of attention. It follows from these definitions. To say otherwise would be like saying, contradictorily, `I'm looking for the thing that just immediately caught my attention.'
The target article should have made the distinction between stimulus-driven and goal-driven attention. It was essentially about stimulus-driven attention. With that limitation, the arguments in the target article are not undermined by Brinck's commentary. I suggest, furthermore, that stimulus-driven attention is the evolutionarily more primitive form of attention, thus rooting the neural basis of predicate-argument structure firmly in what MacNeilage & Davis, after Darwin, call `lowly origins'.
I have much sympathy with the position of these writers that the evolutionary roots of language are to be found in action. `In the beginning was the deed, not the word', as Goethe's Faust insisted. The target article was mainly concerned with demonstrating a present-day correlation between semantic structure and neural organization. That this neural organization is shared by higher mammals does indicate `lowly origins', but I did not dwell on the evolutionary history of this organization (though it would be fascinating). At one point, I told a brief merely figurative story, repeated by Indurkhya, of the growth of predicate-argument structure from earlier forms of behaviour which were holistic, and did not exhibit anything resembling the dichotomy between predicate and argument. I was once a phonetician, but it is too late for me to catch up with the likes of MacNeilage & Davis and theorize about the origins of speech. And if Indurkhya thinks that my story sped past the interesting bits too fast, he should write his own story.
Indurkhya raises the matter of holistic one-word utterances, as made by children and our ancestors at some stage. Only some such utterances support Indurkhya's view of an action-based system in which no division like that between subject and predicate can be made. If a speaker routinely grunts (like a tennis player) when performing a certain action, then certainly we may see the grunt as in some sense intrinsic to the action. But when a child says `Daddy!', as opposed to `Mummy', although the utterance is a single word, there are nevertheless distinguishable acts of referring to a particular person and assigning it a certain mental category. The target article noted briefly, near the end, that holistic utterances could nevertheless express predicate-argument meanings.
Arbib makes what could seem to be an odd point about the distinction between neural processes and descriptions of those processes. Obviously, for any X, `description of X' is not the same as X. The word electron is not an electron. I agree with Arbib that the formula PREDICATE(x) is not itself a neural process. Who could think otherwise? Perhaps the issue is whether some neural process or configuration described by a scientist's predicate is itself a `representation' available to the animal concerned. I use `representation' in the sense that if an animal can reliably distinguish a certain class of stimuli from others, the neural configurations that enable it to do so constitute a representation of that class of stimuli, for which we humans may or may not happen to have a word, such as red or leopard. In this sense, the representation is available to the animal. I don't make the distinction between representations and `their supporting neuronal states and processes' made by Piattelli-Palmarini & Harley. For them, `representations are descriptions accessed internally by the subject'.
It is an empirical matter what uses the animal can put its representations to. A frog can use its prey-representation for catching prey, but it can't attach a symbolic label like prey or insect to its prey concept, for communicating about prey. Humans can describe their representations in a public code, most animals can't. When a frog jumps at a particular stimulus, it would seem to be internally accessing (or perhaps just using, or even being used by) some configuration in its brain. Perhaps Piattelli-Palmarini & Harley's point is that whatever is accessed internally in this case is not a `description'. I do not claim, of course, that if we open up a brain we will find representations somehow written down in the same kind of public symbols that we humans use to talk about things (any more than we will find a little homunculus looking at a screen). Piattelli-Palmarini & Harley write `Neural states or processes as such have no semantics. They co-vary nomologically and causally with events in the world'. My view is that the causal co-variance of neural states or processes with events in the world is necessary but not sufficient basis of semantics. Smoke is causally co-variant with fire. To a first approximation, semantics, as the term is conventionally used, is restricted to the domain of conventional or non-natural meaning, in Grice's terms.
I agree with Arbib that animals' representations are modulated in complex ways by whatever else is happening in the brain and in the world outside; so such representations are indeed `likelihood distributions over a multi-dimensional parameter space'. Since Wittgenstein, most semanticists have believed that the meanings of words are also likelihood distributions over a multi-dimensional parameter space, so the use of an expression like `non-linguistic predicate' should not be too objectionable.
Haack, S., 1994 ``Deviant Logics''. In The Encyclopedia of Language and Linguistics, edited by R.E.Asher and J.M.Y.Simpson, Pergamon Press, Oxford. pp.891-896.
Kamp, Hans, 1981 ``A theory of truth and semantic representation''. In J.Groenendijk, T.Janssen and M.Stokhof (eds) Formal Methods in the Study of Language: Proceedings of the Third Amsterdam Colloquium. Mathematical Centre Tracts, 277-322. Amsterdam. Reprinted in J.Groenendijk, T.M.V..Janssen and M.Stokhof (eds) (1984) Truth, Interpretation and Information, GRASS 2, Dordrecht, Foris.
Kamp, Hans, and U.Reyle 1993 From Discourse to Logic, Dordrecht, Kluwer.
Pepperberg, Irene, 1999 `Can a parrot respond to absence of information?' In Pepperberg, I. The Alex Studies, Harvard University Press, Cambridge, MA.
Talmy, Leonard, 2000 Toward a cognitive Semantics, MIT Press, Cambridge, MA.
Vaina, L.M., Solomon, J., Chowdhury, S., Sinha, P., and Belliveau, J.W. 2001 `Functional neuroanatomy of biological motion perception in humans', Proceedings of the National Academy of Sciences of the United States of America, 98 (20):11656-11661.
Zadeh, Lotfi A., and Kacprzyk, Janusz, (Eds) 1992 Fuzzy Logic for the Management of Uncertainty, Wiley, New York.
Zhuo, Y., Zhou, T.G., Rao, H.Y., Wang, J.J., Meng, M., Chen, M., Zhou, C., and Chen, L., 2003 `Contributions of the visual ventral pathway to long-range apparent motion', Science, 299 (5605):417-420.