Hauser, Yang, Berwick, Tattersall, Ryan, Watumull, Chomsky and Lewontin have a co-authored article on The Mystery of Language Evolution. It’s a review of current directions in the field with the basic message that we don’t yet understand enough for empirical evidence from animal studies, archaeology, palaeontology, genetics or modelling to inform theories of language evolution. Here I summarise the paper and offer some criticisms.
The core language phenotype of interest, according to the authors, is discrete infinity as exemplified in recursive operations found in combinatorial phonology and hierarchical syntax. The authors argue that the methods of evolutionary biology cannot yet be adequately applied to the evolution of this phenotype.
The paper begins with an illustration of the methods of evolutionary biology in a case where this kind of inference is possible. Túngara frogs (pictured above) have a very simple communication system (males croak to attract females), and we know a lot about the mechanisms underlying production and perception and how it links to fitness. However, the obvious adaptive hypothesis (perception adapted after production) was proven wrong by comparison with living sister species (they had similar perception, but not production capacities, so production adapted to perception). This method is hard to apply to language evolution, because we don’t have a good idea of the mechanisms involved and we have no sister-species to compare ourselves to.
Specifically, the authors focus on 4 domains of inquiry, which they claim cannot contribute to theories of language evolution.
The first domain is animal studies. Birdsong is not a suitable comparison with human language because it is a finite system (unlike discrete infinity), in one sensory channel (although why this is a problem is not clear) and has no compositional semantics. Primates are not adequate species for comparison because they have no vocal learning ability, they have no phonology and they have poor voluntary control over their production. Attempts to explore the linguistic abilities of animals in general are limited because they have a limited or innate lexicon, they show no evidence of having abstract concepts, they don’t combine signals and they have no parallels to grammatical classes or agreement. The most advanced examples of animal communication come from extensive reinforcement training. The conclusion is that “it is not possible to empirically support a continuity thesis whereby a nonhuman animal form served as a precursor to the modern human form”.
Paleontology and archaeology
If living animal species aren’t a good fit, what about our hominid ancestors? The next section argues that we have no evidence for cognitive abilities of Neandertals. Skull fossils, brain size and the burial of the dead can’t tell us anything about cognitive capacities, and “it is abundantly clear that the manufacture of even quite complex stone tools is not necessarily associated with modern cognition” (no citations are provided, so I’m not sure what their argument is).
Confusingly, the authors go on to argue that fossils and a lack of symbolic material culture provide evidence that Neandertals did not speak modern languages. As Dan Dediu pointed out in his EvoLang talk, there are existing societies of humans that have very little material culture, but quite modern language, so a lack of an archaeological trace does not imply a lack of symbolic thought. It’s clear that much more evidence is needed before we can know more about the capacities of Neandertals, which makes the point of this section confusing.
The authors argue that, despite considerable advances in our understanding of genetics, the gap between how geneotypes lead to phenotypes is too large, especially because of many language-related genes are regulatory. For example, we can’t tell if FOXP2 affects the perceptual systems or the computational capacity. Furthermore, human changes in FOXP2 most likely evolved in response to changes in diet, so FOXP2 is did not evolve ‘for’ language (this is hardly a new point, as Simon Fisher often emphasizes it). This suggests that ultimate goal of the authors to describe what discrete infinity evolved ‘for’, but this contradicts the typical claim that langauge was not adaptive and exapted, so apart from stating that molecular biology is complicated, I’m not sure what the point of this section is.
Surveying 4 computational models, the authors conclude that the main problem with modeling is that they presuppose the existence of a recursive ability, either by building it in to mathematical models or by using human participants in artificial learning experiments. This means that models can’t explain why this ability evolved in the first place. In addition, many aren’t verified by comparing the results with empirical data.
The authors also take issue with the idea that fitness is tied to communicative success in some models (their citation of Baronchelli et al.’s model is a bit misleading because that model links domain-general rapid learning to fitness). See below for some points on this section.
The article ends on a more up-beat note, with a look ahead to the future. They point to the following directions that could align research on discrete infinity with evolutionary biology to a greater extent. In the case of animal studies, an ability to measure spontaneous processing in animals in a way that avoided training would be a step ahead. Fossils could still be useful if casts of skulls could be better linked with brain structures (which seems like a tall order). The discovery of symbolic artefacts for Neandertals would help push back the date of the language capacity. Finally, “modelling work must focus on the computations and representations of the core competence for language, recognize the distinction between these internal processes and their potential externalization in communication, and lay out models that can be empirically tested in our own and other species.”
However, the overall conclusion is that the four fields mentioned above, in their current state, cannot contribute to theories of language evolution.
The Mystery of Language Evolution
While I agree that more empirical evidence is a good thing, and it would be nice to have another species to do comparative work on, I see this paper as problematic. First, it is quite disheartening – as a modeller trying to interface with the field of language evolution, my initial impression after reading this is that I should give up. More seriously, the dogmatic emphasis on discrete infinity as the primary or only language phenotype worth studying is extremely limiting. The concept of discrete infinity itself has been criticized by many (e.g. Hurford), and statements like “no matter how far apart [arguments] are from each other, the [hierarchical] association remains” seem absurd when looking at language from a performance perspective. There are plenty of other factors that have been suggested as important aspects of the language phenotype, such as capacity for massive storage (Hurford, at this year’s EvoLang), joint attention, theory of mind, power and kinship, social structure, to name a few.
The article wilfully resists interfacing with these concepts and the wider field. There are some entirely unsubstantiated claims, and statements like “The distribution of the structural properties of language, such as word order and agreement, do not seem to follow any cultural or historical patterns”, ignore empirical work on this topic (e.g. Dunn et al.). The absence of all the authors from the recent Evolution of Language conference is notable, where they might have been able to interact with current research and learn of new discoveries that could address some of their problems, such as evidence of vocal control in Gorillas from Marcus Perlman.
The title of the article – ‘the mystery of language evolution’ – is revealing. By using the language of religious dogma, they are claiming that language evolution is a ‘mystery’: a sacred domain that can only be interpreted and revealed by the anointed few. The billboard of famous names contributes to this impression. If the intention really is to encourage research in this field, this is a dangerous move.
The dismissal of the contributions from the various fields is also unfair. Consider their criticism of modeling. They argue that
“If communication has played a significant role in the evolution of language, its force should be observable in the process of language transmission. The history of language change provides the only testable case for the predictions of this communication optimization thesis, and the evidence points in the opposite direction.”
And go on to cite examples of language change that is apparently independent from the constraints of efficient communication (e.g. vowel mergers, redundant morphology). Leaving aside the studies that do find support for adaptation to efficiency of communication (e.g. Maurits et al., Piantadosi et al.), and specifically for loss of phoneme contrasts (e.g. Wedel, Jackson & Kaplan, see here), and the suggestion that communicative efficiency is best served by robustness, not simplicity, this is still a strange claim, because it’s not clear that diachronic differences in modern languages are qualitatively the same as a move from proto-language to modern language.
Futhermore, the authors claim that “Under the adaptationist assumption in language evolution modeling, languages that facilitate more efficient communication are more successful in transmission to the next generation.” This misses a key part of some models (e.g. Smith, Tamariz & Kirby, 2013) that see learnability, as well as expressivity as properties that emerge from the process of cultural transmission. Modelling work has shown that cognitive biases can be amplified into linguistic universals (Kirby, Dowman & Griffiths, 2007; Thompson, Smith & Kirby, 2012), going completely against the article’s claim that “From these formal systems it is possible to deduce linguistic universals as consequences”. My own modeling work (with Bill Thompson and Kenny Smith, presented at EvoLang) showed that, regardless of the mapping of communicative success to reproduction, understanding the cultural transmission process is still required to make inferences about cognitive processes from observations of the distribution of linguistic features. The model actually helps us formulate, evaluate and generate predictions from theories of language evolution.
Furthermore, it’s not as if modellers are unaware of the limitations of their models. While Hauser et al. suggest that modeling is weak because it either relies on unrealistic computational agents or humans with pre-existing knowledge, standard methodologies from cognitive science actually see these contrasts as beneficial if one integrates the results into a robust argument (Irvine, Roberts & Kirby, 2013). A forthcoming paper by Tao Gong, Lan Shuai and Manghan Zhang reviews evolutionary models and sees many successes as well as problems. The dozen responses to the paper by leading experts demonstrates researchers are engaging seriously with the problems.
It’s true that modelling would benefit from integration with empirical observations, just as molecular biology would benefit from a greater knowledge of transcription factors. This much is fairly obvious to the researchers in the field, so what is the intended message of the paper? Unlike the congregational call of the túngara frog, it’s difficult not to interpret it as “stay off our turf”.
Below are some minor thoughts I had while going through the paper.
The authors claim that phonological assimilation is an automatic, emergent phenomenon, but then that they are represented as rules which “can be described as familiar IF-THEN statements in computational systems” which “children unconsciously and spontaneously follow”. But surely phonological rules are gerneralisations of patterns in form, rather than rules that are represented in the cognition of a speaker?
In the discussion of hierarchical structure, I don’t understand how “the correct semantic interpretation can be derived if we consider the logical sequence in which structures are built”. Surely in an infinitely digital system, this process is going to get complicated without relying on semantics in context at all?
The authors claim that “Chimpanzee hearing is basically identical to ours” (no citations), going against the findings that chimpanzee tuning to speech has qualitative differences to humans, and that Neandertal tuning is much more similar to humans than to chimpanzees (Martinez et al., 2004, 2013).
“A leading proposal in evolutionary modeling is to identify language fitness with communicative success (Baronchelli et al., 2012; Nowak and Komarova, 2001): individuals who communicate with and learn languages more successfully have greater reproductive success.” – This seems to confound the idea of the fitness of a language and the fitness of an individual.
The authors argue that “evolutionary analyses demand a clear specification of the target phenotype,”, but then go on to suggest they have no such specification for their own target:
“to the extent that our account of the language phenotype is diffuse (some general system of cognition rather than a precisely delimited and narrow computational module), the genotype-phenotype mapping will be correspondingly more challenging to address.”
The claim that “The structural constraints of language, … derived from a few familiar languages, have proven remarkably successful in the description and explanation of linguistic diversity” ignores a massive literature on linguistic relativity. I also find the distinction between “familiar [European] languages” and “unfamiliar” Australian languages (section 2) distasteful.
Marc D. Hauser, Charles Yang, Robert C. Berwick, Ian Tattersall, Michael Ryan, Jeffrey Watumull, Noam Chomsky, & Richard Lewontin (2014). The mystery of language evolution Frontiers in Psychology