Origins of Grammar
Francis, 25-26 July 2002 -- 2,600 words

In conversations with Per Aage Brandt on 25 July 2002, we developed several new ideas on the origins of grammar, as below, with subsequent elaborations.

Introduction

In terms of the evolution of language, we propose that the problem that had to be solved was to find ways of communicating meanings that were already present and elaborated in human cognition (see The Phoneme Revolution).

Per-Aage brought up the possibility that what is captured by phonemic strings is the denotative meaning, while the connotative meaning remains in the analogical component of the communicative act (we'll need to consider this carefully before we use it). We may envisage an initial situation where the connotative meaning is initally very large -- the only denotation is a few nouns and maybe some verbs and an adjective or two. The task of each speaker is then to enrich the language by bringing some aspect of the connotative meaning into the denotative system. This work would have involved the generation of grammar.

In brief, we propose that what we call grammar is based on an agent-centered understanding of event structure that is tens if not hundreds of millions of years old. (Cite the work of Srini Narayan on grammar and motor routines -- he grounds it in the motor cortex, which is a great place for us to begin too.)

Linguistic capacities in non-human primates

The communicational capacities of other primates is extensive, but qualitatively different from human languages.

1. Elementary symbolic communication is present already in vervets (Seyfarth & Cheney), and likely generally in monkeys (ask Susan Perry & Joe Mason about capuchins). Symbolic communication involves the learned use of arbitrary cues fixed by convention to communicate and understand meanings that exist in the form of mental simulations.
2. The vocabulary of the vervets consists largely in alarm calls. These are typically repeated single-token utterances.
3. Apes in captivity have manifested a limited capacity for creating symbolic cues by combining existing cues.
4. The vocabulary of trained apes is comparable to the vocabulary of a xx-month-old human infant. It is focused on a particular set of communicative needs appropriate for interacting with a caregiver -- that is to say, on objects and services that the individual wants.
5. Chimpanzee communication in captivity may include some elementary grammar -- that is to say, some method of conveying the desired role of an agent. We should look into this -- it may involve simple pairing, as in "tickle Koko" or "eat Koko," where the distinction between subject and object is either left entirely implicit or signaled outside of the conventional signs. Someone has very likely worked on this. I wonder if anyone has tried to teach them case morphemes?
6. More generally, non-human primates are capable of utilizing symbolic cue/meaning pairs, typically to indicate reference to external agents or objects that are feared or desired, by themselves or by other memebers of the troop.
7. It should be noted that the trained chimpanzees are in an ecologically unrealistic environment where the animals have in effect been infantilized. This has the effect of biasing the content of chimpanzee language to requests.
8. We propose to generalize and suggest that the origins of language may be sought in the need to communicate what is feared or desired with respect to external agents or objects.
9. The vocalizations used by vervets and the visual symbols used by trained chimpanzees are arbitrary cues that must be learned, but they provide no method for generating new cues. While it is in principle possible to combine them, the combinatorial space is limited, largely unsuited for introducing new meanings, and apparently rarely utilized.

In an attempt to reconstruct the evolutionary history of language, we adopt the default assumption that the communicative performance of chimpanzees is representative of the communicative capacities of our common ancestors. The task we propose to ourselves is therefore to add in the smallest possible number of additional cognitive capacities to arrive at a theory of language that is as parsimonious as possible.

We propose that the most immediate limitation facing our primate ancestors was the difficulty posed in generating new symbols, or new cues in cue/meaning pairs. The difficulty is not inherently one of a lack of raw material: any combination of arbitrary or analogical gestures, facial expressions, body movements, eye movements, and vocalizations can be utilized as a symbol. Such symbols, however, would suffer from major drawbacks.

1. First of all, it would be difficult to distinguish the denotative meaning from the connotative mening. This is a version of the frame problem raised by Quine and others. The result would be degraded replication.
2. Secondly, they would be hard to remember. Even the speaker who first generated the symbol might not be able to reproduce it accurately on the next occasion. Such a failure of standardization would again lead to degraded replication.
3. Third, the interlocutor would have an even harder time remembering it. In consequence, the symbol would have a hard time spreading, as speakers fail to replicate it.

It's worth noting that these obstacles to replication are structurally eliminated in the lingustic symbols successfully taught to chimpanzees. The researchers very elegantly bypass these difficulties by defining a fixed repertoire of visual tokens.

1. These tokens are purely denotational, instantly resolving a major part of the frame problem.
2. The generation of symbols is left entirely to humans, who fix them in physical form, likely produced and perhaps reproduced mechanically. This removing the problem of imperfect replications by the inventor.
3. While it may be difficult to remember them, a great deal of time and resources are devoted to this task, to a degree that natural circumstances would not be capable of reproducing.
4. The fact that the symbols are located on a grid may also render the task of learning them easier. This lends a degree of systematicity to the symbols that would be extremely difficult to reproduce in nature.
5. Finally, the symbols don't have to be replicated at all -- they merely have to be recalled. Thus, the central problem of language -- that of symbol replication -- has been effectively removed.

In brief, the symbol boards used in chimp language solve precisely the problems that under natural circumstances would have presented enormous obstacles to the development of an effective language. It is our contention that our ancestors in fact were faced with these obstacles.

The success of symbol boards in teaching chimpanzees to talk is thus an admirable achievement -- but the linguistic system as such is above all an admirable human achivement. In our view, such symbol boards present a partial solution to precisely what is difficult about language. However, even they do not provide a sufficiently sophistociated solution.

The primary limitation of the symbol-board solution is of course that it is entirely unrealistic in a natural setting. The problem of delimiting denotation from connotation must be solved without such aids -- and how is this solution possible? Equally, memorability and reproducibility must be resolved without the use of symbol boards, and such boards provide little or no clues to how this may have been accomplished.

A final and perhaps less conclusive limitation of the symbol-board solution is that it is unlikely that the symbols are helpful in thinking. More generally, it is unlikely that the mining of the multimodal space of possible communicative signals for the generation of symbolic cues would aid thinking, in the sense of competing successfully with the mind's existing representational formats. If symbols are not used in thinking, they would be sooner forgotten.

This point provides an interesting middle way between a Whorfian and a Pinkerian perspective: it may be a requirement for a fully-fledged linguistic system that it actually aids thinking -- if not, it would tend to be forgotten.

The limitations to this solution also suggest where the biological differences between chimpanzees and humans lie, pointing the way to the adaptive changes human cognition had to undergo before a fully fledged language was possible. We suggest it indicates the evolution of a biologically based cognitive system of standardized and limited syllabary, with combinatorial properties for generating memorable symbols.

Note that this is a very different argument against the proposition that chimpazees can speak than the usual argument. Where the usual argument focuses on the claim that chimpanzee communication lacks grammar, this argument focuses on the claim that chimpanzees lack a syllabary.

We do need to address the question of chimpanzee grammar. (Cite the work of the researcher who sees various aspects of grammar in gorilla courtship interactions.) On the basis of the argument that what chimpanzees lack is a syllabary, we might expect a straightforward prediction that chimpanzees can communicate grammatically. However, even though there can be no doubt that chimpanzees understand events, we have to raise the question of whether they run mental simulations of events. These are different claims: the ability to parse events is not the same as the ability to run event simulations. The broader model we use here is that communication relies on mental simulations.

Thus, we predict that chimpanzee grammar is constrained by their imaginative capacities, or their capacities to run mental simulations of events. To the extent that chimpanzees are able to simulate events -- for instance, to imagine a desirable event in the future -- we would predict that they are capable of communicating the grammatical role of the agents in such events. Specifically, we predict that chimpanzees can be trained to distinguish between the nominative and the dative, or the giver and the recipient of a gift (see details below). The way to test this would be to teach the chimpanzee a particular morphological tag that indicates role; it could be physically attached to the existing symbol, for instance with a magnet.

The main argument here is that a syllabary is going to be required before you get grammaticalization of language on a significant scale; see The Phoneme Revolution.

Origins of the parts of speech

The evidence suggests that language starts with the names of concrete objects, such as the names of animals.

Chimpanzees are also able to understand verbs, such as tickle and eat, and adjectives such as sweet, hungry and tired.

We may need to provide a principled argument for "parts of speech" -- why nouns, verbs, and adjectives?

Predication and binding

We discussed neural binding as the pre-existing correlate to predication. This involves not only perception but is carried in memory. Our ancestors, in order to create the future and the past, would need sophisticated forms of neural binding in memory. These capacities would be useful without language and the simplest assumption is that they precede language.

Event structure

The main source of grammar, we propose, is our pre-existing capacity for parsing and understanding events. This includes the identification of agents, roles, actions, movements, places, and times (fill this out).

Prototypical social situations

We propose that grammar was developed by formalizing prototypical cases of recurring human relations. Once formalized, the grammatical pattern for this specific type of relation could then be applied through analogical reasoning to other situations, including situations that do not involve human relations. We propose that the primary generative source of prototypical situations is human relations rather than other domains, and that the resulting grammatical structures were only subsequently applied in other areas.

The reason we believe social relations provide the majority of prototypical situations for grammar is that language initially emerged purely as a tool for communication. When you want to make an axe, you don't need to tell the rock what you want; when you want to marry a girl, you need to speak to speak to people about it.

Active and Passive

The active voice is prototypically used to convey the actions of an agent. The prototypical use of the passive voice is likely to communicate an attack or an injury. There are languages where the passive voice cannot be used to convey compliments or positive assessments (details).

Nominative

The prototypical nominative case is designed to identify the agent of an action, as if by naming him.

Accusative

The prototypical accusative case is designed to identify the agent that is the target of someone's action, typically identifying the interlocutor.

Dative

The prototypical dative case is designed to identify the recipient of a gift. It is built to communicate the recurring human relation of giving an object to another -- an action that would have been important to track.

Locative

The prototypical locative is designed to identify the time and place at which some event took place.

The general argument here is that the cases primarily solve the problem of conveying the particular social role adopted by the various agents involved. We suggest that a morphological solution to this problem is a particularly elegant one. It allows speakers to utilize the lexicals they are already familiar with, only modifying them to convey the appropriate social relation.

Embedding

A primary source of embedding is to convey the correct or appropriate status of representations -- in broad terms, to assist the listener in the task of source monitoring. Thus, "Jane thinks that" and "Peter knew that" are ways of tagging simulations and asserting their scope (see Tooby & Cosmides, "Consider the Source").

Word order and morphology

There is some evidence that early languages have relied more on morphology than on word order.

On the one hand, one could argue that there are elements akin to word order in the mind's native format of mental simulations. The main idea here is that mental simulations are agent-based, and that we might expect a word order where there is a tendency for the agent to be mentioned first. Possibly, but less convincingly, there is some reason to believe that the actions or other characteristics of the agent are so closely tied to the agent that they can be expected to be mentioned before other elements of the situation

This argument does not extend very far. That is to say, there are no obvious reasons why the object of the action, or the recipient of the action, or the place and time during which the action took place, should be mentioned in any particular order. In our view, the default solution to the complex problem of roles should be handled by morphology rather than by word order (syntax).

Of course, word adjacency could be used to indicate predication. But if the morphology is in place, it would be more economical to just let the morphological tags be used for predicates too. Once you have morphological tags, the constraints of word order become unnecessary.

What may be predicted to skew language towards syntax and away from morphology is writing. Because writing presents an unambiguous and replicable, external, and visible order, it should have the effect of off-loading the work of morphology onto syntax. It would be interesting to see if there is a single language that developed in the absence of writing that relies primarily upon syntax for its grammar.

The role of the scholar and the poet

We discussed how the task of grammaticizing cognition may have taken a long time, and account for the delay in the emergence of symbolic thought. The central idea is that all of the species-defining cognitive adaptations are in place by around 130,000 years ago, and the puzzle we need to explain is why it took tens of thousands of years before we see any record of major symbolic activity.

In fact, it should be mentioned that even cave paintings and paleolithic statuettes emphasize the analogical rather than the symbolic imagination. It is not until the neolithic that the symbolic imagination clearly begins to dominate in cultural production. While paleolithic art is exquisitely figurative, neolithic art is largely symbolic and appears to rely heavily on verbal narrative. It is worth keeping this contrast in mind -- the paleolithic cultural explosion remains focused on the externalization of analogical thought, or visual imagery. It shows clear signs of having been analytically parsed and accessed introspectively in ways that suggest a fully modern mind. However, the evidence for symbolic thought is not strong, at least at first blush.

The paintings and sculptures do, however, show evidence of a kind of grammar -- that is to say, they are examples of the externalization and communalization of the imagination. It's just that this is not done primarily (or even at all) with the help of a formal system. Instead, the grammar of the paintings suggest an analytical decomposition of visual imagery that comment on or are attuned to how the mind works.

In the case of grammar, we could suggest that we have a similar process taking place. That is to say, grammar is the cave paintings of the mind -- an analytical decomposition of events expressed in symbols. There is a kind of disjunction or transverse shear here, for it is precisely this analytical decomposition of events that we see in neolithic art. This art is produced not only by agricultural communities; all of the art of the new world, for instance, is neolithic, including the art of the Chumash, who did not farm. On the face of it, this suggests that symbolic thought did not assume a dominant role in civilization until the holocene, or the period following the last ice age, some ten thousand years ago. It is above all these post-glacial cultures that are overtly symbolic.

The paleolithic art that remains is clearly focused on exploring the analogical imagination -- an exploration that is systematic but not formalized. This art is characterized not by symbols but by stunning representations of what the mind perceives. They suggest, in other words, a desire and an ability to externalize and make communal the contents of individual minds. In addition, they may show evidence of an explicit or implicit understanding of the epidemiology of representations -- that is to say, of how the human mind works, how it thinks, what it finds striking, and how it learns. The paintings and statuettes exaggerate those features of the referent objects that make them salient or attractive to the human mind. That is to say, they represent objects in ways that make the representations somehow more typical of the category than any actual member -- the Chauvet rhino is a better example of the category rhino than a real rhino. This may be going too far, but the idea would be that the artist has captured something about the mental prototype and externalized it, in effect painting thought.

So this is the main idea of the paleolithic art: it is painted thought. But it is painted analogical thought, painted imagery, not painted symbols. Of course you wouldn't in any sense expect painting to keep up with the development of language -- visual art has specialized and in some cases regressive functions, keeping alive modes of cognition that are being displaced. While art may be both exploratory and hankering towards the past, bringing in modes of cognition that have been ignored, the argument that it represents a different stream of cultural evolution may still hold.

The claim, then, is that the grammatization of cognition was a prolonged process that not only took tens of thousands of years, but that is still on-going. Poets may tap into the rich soup of analogical thinking, or out of this soup they may extract some invariant feature and lexicalize or grammaticalize it.

The visual imagination -- or more generally, the mind's native format of running simulations -- remains basic to cognition. The earliest writing systems clearly aim to access this analogical imagination: the marks symbolize concepts, not words. Such writing systems of course have the advantage that they can be used across linguistic communities, which may have been an advantage already for the Sumerians, who traded extensively. The main advantage, however, may simply have been that these symbols tapped directly into the mind's preferred ancient representational system.

The difficulty here, then, is that we find conflicting strands of evidence. It is clearly not the case that humanity simply moves from an analogical into a symbolic sphere. Early writing systems, which date back only a few thousand years to around 4,000 BC, show a preference for tapping directly into the analogical imagination. The argument may be made that these formalized writing systems are continuous with the symbolism of neolithic art. This suggests a line of continuity from the neolithic revolution to early civilizations that is hardly surprising. The cultural evidence here suggests that symbolic systems have acquired a central significance, but they are poorly grammaticalized. We could look in greater detail into early ideographic systems -- do they show any evidence of grammar? Or are they symbolic systems that exist in parallel with language, but without grammatization? This would be very interesting to find out more about.

With respect to language, then, the argument may have to be that it follows a trajectory that is not duplicated by the visual arts -- and perhaps for good reasons. The basic argument is that grammatization does not require new cognitive machinery; it is simply a certain kind of melody that is played on the register of the mind. It is useful for certain occasions, while other forms of symbolic thought are useful for other occasions.

One, that opens up for the possibility that paleolithic art is symbolic, but symbolic in a manner characteristically different from neolithic art -- not at first blush a very convincing argument.

Two, language may have been subject to a different set of cultural pressures than the visual arts. This is almost certainly correct. The argument is that vocal communication, grounded in a formalized system of phonemes, represents the cutting edge of human culture, far in advance of visual imagery, which is trivial in comparison. Language is much older and would have been subject to intense pressures of natural selection (for an analysis, consider the new framework under phonemes, end of scroll).

What we still lack is a detailed account, with some striking examples, of how grammar may have emerged.

Finally -- and this doesn't really have to be deal with until the third article -- we need to consider how the evolution of language may have triggered the emergence of symbolic thought. This argument actually still holds; what doesn't hold as clearly is that paleolithic art is evidence of symbolic thought -- at least symbolic thought that is in any intimate way connected with language. I do think one could argue that there is a kind of visual grammar in the paintings, but the work is really not a very good fit -- there is analytical decomposition, introspection, high-level understanding of human nature, but not grammar, not events at all in fact, unless we imagine that the paintings are props for larger-scale communal simulations. If this is the argument -- as it was in my CASBS talks in the fall -- then it is only loosely tied to an argument about the origins of grammar..

Predictions

This model generates a series of interesting predictions. For instance, it suggests that children should only be able to learn a grammatical form when they already understand the prototypical social relation that this grammatical form is constructed to convey. Moreover, children are predicted to show a preference for utilizing new grammatical forms in prototypical situations, and these prototypical situations are predicted to be social relations. Once they have mastered a grammatical form in a prototypical social relation, they should slowly be able to begin to utilize the same grammatical form in non-prototypical social relations, or in relations that are not social -- say, their relations with inanimate objects. However, this is predicted to happen slowly and perhaps over several years.

Relation to existing theories

This model is presented as a tentatively coherent perspective on the origins of grammar. We are not committed to the truth of this model, but believe that it is plausible. We see the purpose of our proposal to formulate a research framework.

In our view, this is an evolutionary model more consistent with the adaptationist project than Pinker's proposal of a language instinct. At the same time, we agree with Pinker that the evidence is convincing that certain aspects of language are innate; specifically, the proclivity and capacity to construct a fixed set of phonemes. This is used to generate the combinatorial space for generating names.

The model is in many respects compatible with Lakoff's view of language. In our view, however, a distinction should be made between the capacity to utilize a set of fixed phonemes, which is clearly an evolved adaptation to language, and the various features of grammar, which may not be evolved. While we do not adopt a firm stance on the question of whether aspects of grammar are innate, we find the arguments for innateness less convincing that the argument that grammar is derived from other, innate cognitive structures. We also differ from Lakoff's position in attaching more significance to evolved non-linguistc structures and less significance to ontogenetic learning. These two processes, however, are not in principle distinct -- that is to say, ontogenetic learning is not an alternative to evolutionary adaptations, but a complement to it. It is by virtue of evolved adaptations that we are capable of ontogenetic learning, and this learning is in many cases relatively tightly structured by evolved capacities and proclivities (we'll need to tidy up this argument).

We are not entirely averse to the idea that aspects of grammar are innate. However, in our view the purpose of grammar is to convey information that is already present in the mind in an appropriately structured form -- that is to say, we accept the common-sense or naive view that we have something in mind and then find a way to communicate it. In this view, grammar represents a way of encoding relations, just as lexicals are a way to convey entities, actions, and qualities that are already fully represented in the mind. If this view is largely accurate, then there is no need for natural selection to act on grammatical aptitude -- grammar simply reproduces skills that have already been selected for.

It should be clear from this discussion that we do not argue that language creates our sense of reality. We do not dispute Pinker's claim that the categories conveyed by lexicals are generated by evolved cognitive mechanisms. In fact, we take this argument further: not only are categories based on "prior art" -- so is grammar itself. The biological substrate of language -- adaptations that makes language proper possible -- are largely focused on the phonemic system. These are not simply physiological changes; most crucially, they are cognitive adaptations for constraining denotation to a digital and combinatorial space of phonemes.

Conclusion

Now, the point about this new generative view of language is that we can imagine an infinite series of possible grammars. However, only those grammars that usefully and faithfully reproduce the logic of the preexisting cognitive structures of the speakers will spread. What limits the variability of grammar, then, is simply what Sperber calls the epidemiology of representations. Our model is in this sense a cultural constructivist model of grammar, constrained in its spread by pre-existing evolved modes of thinking.

Another way of saying this is that in this view, language is integrated into cognition according to best practice in software design: the principle of reusing code. Our central argument hinges on this: there is simply no need to evolve a "language module", since the function of language is to communicate what the mind already is doing. In this sense, the Chomskian project is fundamentally misconceived. What you do need to evolve is a low-level communicational protocol, which is designed to provide speakers with a highly effective tool for communicating what they are already thinking. On top of this protocol -- which we argue consists of a highly sophisticated system for digitizing communicative signals -- you can improvise a grammar to convey the thoughts you already have. The basic task such a grammar needs to achieve is to provide a set of morphological endings that indicate role in a well-understood and non-linguistic event-structure.

We moderate this claim by allowing that phonemic strings may play a significant role in thinking. Specifically, we propose that the phonemization of cues in cue/meaning pairs and the grammatization of significant aspects of cognition makes symbolic thought possible. This is a monumental innovation with massive consequences for human life. We see no evidence, however, that the monumental achievements of symbolic thought have in any way been subject to the structuring culling of natural selection, or (more specifically) that they in any sense rely on cognitive mechanisms that have been constructed by natural selection specifically to produce grammar.

In contrast, there is ample evidence to suggest that the ongoing achievements of symbolic thought are made possible by tapping into pre-existing non-linguistic cognitive functions. Our view of grammar, then, is generative, but the generativity comes not from a dedicated language organ but from cognitive mechanisms that were built for other purposes, combined with the language-specific generativity of phonemes.

Bibliography

Briscoe, Ted (1998). Grammatical Acquisition: Coevolution of Language and the Language Acquisition Device. Full text (pdf, external). See his "Theoretical background" for citations. See also his 1999 and 2000 articles.

.