AT THE University of Edinburgh, Simon Kirby and his colleagues are running a world-record-breaking game of Chinese whispers. At any one time, hundreds of players may be passing on literally millions of whispers. A single game can last for many generations, with old players dying and being replaced by new ones. And these games are repeated, with slight variations in rules, many thousands of times. A somewhat whimsical research project? 鈥淔ar from it,鈥 says Kirby. 鈥淲hat we are actually doing is uncovering the origins of the most fundamental property that makes us human 鈥 language.鈥
Of course Kirby鈥檚 鈥減layers鈥 are not real people. They are artificial agents, composed of hundreds of lines of computer code. His 鈥済ame鈥 is a complex computer simulation. And the 鈥渨hispers鈥 are phrases in an artificial language, mutating through time as they are repeatedly spoken, heard and spoken again by different agents. Just as in a real game of Chinese whispers, the interesting thing is to see how utterances change along the way. And that is central to Kirby鈥檚 unusual take on the evolution of language, because he and his team choose to focus on what is spoken, rather than on the speaker. They are working within a new linguistic paradigm, one which considers language as an organism evolved to fit a unique ecological niche 鈥 the human brain.
In this view, language is a parasite, unable to live without us. Although we don鈥檛 need the 鈥渓anguage bug鈥 we give it houseroom in our brains because it allows us to do clever and useful things that we couldn鈥檛 otherwise do. Morten Christiansen, who pioneered similar research in his lab at Cornell University, New York, uses more technical terms to describe the relationship. Language, he says, is a 鈥渘on-obligate mutualistic endo-symbiont鈥. That puts it in the same category as microbes that live in our guts and earn 鈥渇ree鈥 food and board in return for helping us process otherwise indigestible B vitamins.
Advertisement
Orthodox ideas of language evolution emphasise how natural selection shaped our ancestors to improve their talent for complex communication. Noam Chomsky from the Massachusetts Institute of Technology famously pointed out that infants learn language quickly and reliably from sparse and chaotic input. For him and many linguists, this 鈥減overty of the stimulus鈥, as it is known, is evidence that much of our language ability is innate, directly encoded in our genome, and takes the form of a neurologically hard-wired universal grammar. Chomsky鈥檚 colleague Steven Pinker argues that the ability to communicate effectively would have given early humans a 鈥渇itness鈥 advantage. Natural selection favoured genetic mutations that improved our language faculty, so they spread through the hominid gene pool. The legacy is that we all have brains adapted for speech.
鈥淏ut what if our brains are not so specifically designed for language?鈥 asks Kirby. 鈥淲hat if we appear to be biologically adapted to language because language has culturally adapted to us?鈥 Christiansen points out that language changes far faster than biology. Languages as different as Danish and Hindi have evolved in less than 5000 years from a common Proto-Indo-European ancestor. Yet it took between 100,000 and 200,000 years for modern humans to evolve from archaic Homo sapiens. 鈥淥f course language confers selective advantages on our species,鈥 says Christiansen. But, he argues, humans can survive without language, whereas the language bug has to adapt and evolve to live with us, otherwise it will become extinct. 鈥淟anguages that are hard for humans to learn simply die out, or, more likely, never even come into existence,鈥 he says.
It may seem like a minor shift in perspective, but the implications of this way of thinking are radical. For a start, Christiansen questions the need to invoke a Chomskian universal grammar at all. Instead, he argues, language has adapted to plug into more general cognitive processing capacities that were already part of our ancestors鈥 brains before language came along. Among these, Christiansen is focusing on 鈥渟equential learning鈥 鈥 the ability to encode and represent the order of the discrete elements in a sequence. This ability is not unique to humans: mountain gorillas, for example, use it in the complicated preparation of certain spiky plant foods, where a sequence of tasks is required to remove the edible part. The idea that sequential learning underlies language has been given a boost recently by studies of people who have a medical condition known as aphasia, in which language impairment is associated with a breakdown of more general, non-linguistic sequencing abilities, such as the ability to copy a sequence of hand movements.
Christiansen鈥檚 hypothesis is simple: if neural networks designed for sequential learning 鈥 but without any built-in linguistic knowledge 鈥 can learn languages as we do, then there is little need to invoke a universal grammar. In one experiment, he exposed his agents to 32 different artificial grammars, each of which possessed different rules for word-order in sentences. He found the agents could learn the languages. What鈥檚 more, there was a correlation between how easily they learned an artificial grammar and how often a similar grammar occurs among the 6000 languages humans use in the real world. Further evidence that there are parallels between ourselves and agents with no innate linguistic knowledge came when Christiansen selected two of his artificial grammars and trained separate groups of people to use them. The group learning the grammar with the rare word-order had qualitatively similar problems to the neural networks. 鈥淭his is clearly an important result,鈥 he says.
Pinker is not so impressed. 鈥淣o one has ever doubted that languages are adapted to the human brain,鈥 he comments. 鈥淪ince language wasn鈥檛 handed down to us by a committee of Martians, how else could it be?鈥 But we still need innate language abilities. 鈥淲e still need to explain why children learn languages 鈥 real human languages, not simplified artificial ones 鈥 leading to a cycle of evolution over the generations, whereas monkeys don鈥檛, even when exposed to the same input.鈥 (see 鈥淭alking heads鈥). Kirby agrees. 鈥淯ltimately we need to understand why humans are unique in this respect, and why human languages are the way they are,鈥 he says. We are beginning to see that satisfactory answers to these questions must take into account the complex interactions between human evolution and linguistic evolution.鈥
To tackle these issues Kirby鈥檚 team use what鈥檚 known as an 鈥渋terated learning model鈥 in which each new generation of speakers learns by observing and copying the behaviour of the previous one. They have found that the most successful languages to emerge in this simulated world have important similarities to the languages we speak in the real world. Their models also suggest that a major force shaping language evolution is the bottleneck that occurs when infants learn to speak on the basis of the limited input they get from their carers. In other words, the 鈥減overty of the stimulus鈥 isn鈥檛 necessarily a problem that has to be solved by invoking innate language skills, but might actually explain why language is as it is.
Whereas Christiansen designs complex artificial languages for his simulations, Kirby allows his agents to generate their own. In his version of Chinese whispers they begin with 鈥渞andom鈥 languages, with each agent choosing chance symbols for each meaning it wants to express. The agents talk about an imaginary world inhabited by a small group of people. So, for example, one agent prompted to express the meaning 鈥淢ary admires John鈥 might produce the alphabet symbol-string 鈥渓dg鈥. Another might use the strings 鈥済j鈥 to say 鈥淢ary admires Gavin鈥. A third could come up with 鈥渪kq鈥 for 鈥淢ary loves John鈥. These utterances form the input for newborn, language-naive agents. To reflect the observed learning capacities of real children, these baby eavesdroppers can occasionally 鈥渕ind-read鈥 鈥 intuitively pick up the meanings behind individual symbols 鈥 and also try to infer what linguistic system lies behind the utterances.
In the early stages the language is 鈥渉olistic鈥, so the whole of each symbol-string (鈥済j鈥, say) corresponds to the whole of its meaning (鈥淢ary admires Gavin鈥). But such a language is highly unstable, with pairings between symbols and their meanings being broken and remade. Eventually, as generations of agents die and are replaced by new ones, this unstable holistic language evolves into a so-called 鈥渃ompositional鈥 one, in which individual parts of the symbol-string start to correspond to separate elements of its meaning. So, bits like 鈥渢ej鈥 and 鈥渕鈥 might become consistently linked to the meanings 鈥淢ary鈥 and 鈥渁dmires鈥. And after 1000 generations, the agents say things like 鈥済jhftejm鈥 for 鈥淢ary admires John鈥; 鈥済jhftejwp鈥 for 鈥淢ary loves John鈥 and 鈥済jqpftejm鈥 for 鈥淢ary admires Gavin鈥. Such compositionality is a fundamental feature of human language structure.
Kirby鈥檚 model mirrors at least one independently developed theory of language evolution. Alison Wray of Cardiff University, editor of the recent book The Transition to Language (Oxford University Press, 2002), proposes that human language began as a holistic communication system of unique sound strings for complete messages, and only later got broken down into words that had to be strung together with grammar. The traditional explanation for this transition is that our ancestors evolved as natural selection favoured individuals who were better communicators. But in Kirby鈥檚 model speakers don鈥檛 evolve. All agents are born identical throughout the entire simulation, and an agent鈥檚 survival is unrelated to its ability to communicate. Despite this, the language still evolves.
鈥淲hat happens is that at some point in the initial randomness, an apparent regularity shows up,鈥 explains Kirby. Parts of the language start to follow general rules, which makes them easier to learn, so they persist for longer. In evolutionary terms, these bits of language are better survivors because they are better adapted to be learned by the agents. Gradually, this process produces more and larger generalisations, until the whole language follows regular rules (see Diagram). By now the language has become very useful to its users: they can learn it from few examples and communicate meanings they never encountered while learning. 鈥淭he key point,鈥 says Kirby, 鈥渋s that although language turns out to be optimal for communication, this isn鈥檛 what drives its evolution.鈥
More intriguing still, Kirby鈥檚 models indicate that language only takes the leap from holistic to compositional and generalisable under certain circumstances. When agents are exposed to every possible meaning in their world, holistic language is stable and doesn鈥檛 evolve further. But most of Kirby鈥檚 simulations deliberately expose agents to sparse and chaotic input of language, mimicking the real- world circumstances in which infants learn to speak. His radical conclusion is that language did not develop despite the 鈥減overty of the stimulus鈥, but because of it.
For Christiansen, this also explains the so-called 鈥渃ritical period鈥 in childhood, after which people find it increasingly difficult to learn language. 鈥淭o reap the fullest adaptive advantage from their language symbiont, humans need to acquire it as early in life as possible,鈥 he says. 鈥淪o languages have evolved to be learnable primarily by children.鈥 And this has implications for what languages can and cannot be. 鈥淓asily learned linguistic forms will establish themselves early, and thus may pre-empt more complex and potentially more communicative forms,鈥 he says. 鈥淟anguage forms not learnable by children will disappear.鈥 Even if such forms might be of more use to adult speakers.
In a further counter-intuitive twist, Christiansen argues that children鈥檚 limited cognitive abilities in areas such as perception and memory may actually help them learn to speak. Recent evidence from both computer simulations and experiments with human adults learning artificial languages support this idea. 鈥淐hildren鈥檚 perception and memory limitations force them to focus on the basic 鈥榖uilding blocks鈥 for further language learning,鈥 says Christiansen. 鈥淚n contrast, the superior processing abilities of adults prevent them picking up these blocks directly. They have to find them using more complex computations, making language learning more difficult.鈥
Other language experts are taking these ideas seriously. They have particular resonance for proponents of the theory of memes, in which units of culture are seen to evolve in much the same way as genes. Daniel Dennett from Tufts University in Boston points out: 鈥淩ichard Dawkins and others, including me, have been arguing for some years now that memes are symbionts, and words 鈥 and hence language 鈥 are a prime example of memes.鈥 Nevertheless, Dennett is uneasy about extrapolating from computer simulations to the real world. 鈥淚t is important to say, firmly and clearly, just how abstract, how unlike human agents, these software agents are,鈥 he says. Wray agrees that simulations can oversimplify and magnify errors. 鈥淚t鈥檚 certainly possible to get a lot of nonsense out if you don鈥檛 know what you鈥檙e doing,鈥 she says. 鈥淔ortunately, Kirby and his team do.鈥
Even if this work doesn鈥檛 dislodge universal grammar from its central role in language evolution, it adds another dimension to the story. Kirby鈥檚 team is already working on simulations where both the agents and the language evolve, to see how this complex interaction plays out. And there is no doubt that their approach has a future. 鈥淚t鈥檚 definitely the way forward,鈥 says Wray. Kirby鈥檚 colleague, Jim Hurford, agrees. A decade ago, in his book The Language Instinct, Pinker described Hurford as 鈥渢he world鈥檚 only computational evolutionary linguist鈥. 鈥淭here are now several hundreds of us,鈥 says Hurford, 鈥渁nd Kirby is easily one of the most prominent.鈥
Talking heads
Why are we the only animals that speak? Some other primates use holistic language, so why has this never evolved into a complex, compositional language like ours? Morten Christiansen believes it helps to see language as a parasite that has evolved to live in the human brain. The question then becomes: what makes this habitat uniquely desirable for language?
Comparative studies reviewed by Christiansen suggest that humans and other primates share certain fundamental cognitive abilities. Non-human apes, for example, can learn fixed sequences such as the strings of sounds that make up words. They can also chop up the continuous sounds of spoken phrases in appropriate ways. But there are also big differences.
Our brains are much better at learning hierarchical structures than those of other animals. 鈥淭his is crucial for the syntactic processing of language,鈥 says Christiansen, because grammar requires the ability to build up a series of phrases into a meaningful sentence. We are also particularly good at recognising and taking an interest in what other people are attending to, a skill known as 鈥渏oint attention鈥. And we have an aptitude for 鈥渕indreading鈥, attributing separate viewpoints and motivations to other individuals. Christiansen suspects that the ability to speak requires a combination of many such skills.
Even so, there鈥檚 some evidence that the language bug can colonise new territory. Bonobos, most notably Kanzi, have learned to communicate with a lexigram-based artificial language 鈥 a simple compositional system. And Kanzi has tried to communicate with other bonobos, chimps and even a dog, using the lexigrams. 鈥淚t鈥檚 at least conceivable that a group of language-trained bonobos could maintain their language ability in the wild, transmitting it both to non-trained peers and to offspring,鈥 says Christiansen.
- Language Evolution, edited by Morten Christiansen and Simon Kirby, Oxford University Press (2003)