Research

At the most general level, my work concerns what makes us human and how we evolved to be the way we are, and in particular the origins, evolution and properties of language. While my research concerns mainly linguistics and the language sciences, it also touches on other disciplines including psychology, cognitive neuroscience, biology, anthropology, computer science and statistics. I briefly summarize below some of my main research directions.

Language in the context of human evolution

Language should be studied in the wider context of human evolution, in continuous interaction with the environment and our biology. Essentially modern language and speech are old, dating back half a million years ago and preceding the emergence of anatomically modern humans.

One of my main lines of research concerns the intimate interaction between culture (including language) and biology as a fundamental force driving our evolution. I have been interested for a very long time in human evolution as it provides the background for understand language, its origins, evolution and present-day characteristics. Steve Levinson and myself reviewed the relevant literature spanning palaeoanthropology, archaeology, genetics, anthropology, cognitive neuroscience and the language sciences, andwe concluded that Neanderthals and Denisovans belonged to the same species as ourselves, had very similar (but not identical!) cognitive and linguistic capacities with us, and that they used essentially modern languages. Thus, modern language is old, going back at least half a million years ago to the last common ancestor we shared with the Neanderthals. This paper received a lot of media attention (e.g., Süddeutsche Zeitung, Discovery News) and it still is in the top most ranking papers. I think that cultural evolution and its interactions with biological evolution are the major drivers of language origins, evolution and diversity. As such, I have been interested for a long time in the ongoing developments in evolutionary theory, especially concerning phenotypic plasticity, evo-devo, niche construction and gene-culture co-evolution, and in their relevance for understanding language and culture. A fascinating example is represented by the co-evolution of the recessive mutations causing congenital non-syndromic deafness and village sign languages.

Figure-language-evolution
Two (simplified) models of language evolution. The “received view” on the right posits that language appeared suddenly, essentially fully-formed and only in our species (modern humans), while the “alternative” on the left proposes that language is old, going back to the last common ancestor we shared with the Neanderthals and Denisovans, and evolved gradually. The two models are mirrored to facilitate comparison. The top scale is in million years before present (mya). Gray represents some form of proto-language, while yellow represents modern language. Human lineages from top to bottom: Homo erectus (gray), Neanderthals (red), Denisova (magenta) and modern humans (blue, suggesting the spread across the world, also showing the archaic admixture outside Africa as thin red lines). Please note that the underlying human evolution has been massively simplified for display purposes, and does not show, for instance, the multiple admixture events between various human lineages, including between modern humans outside Africa and Neanderthals and Denisovans.

The structural stability of language

The stability of the structural features of language is a complex outcome, reflecting universal tendencies, macro-regional processes and language-specific factors. There is probably no single “correct” way to define and measure stability, as various approaches capture its different aspects.

Stability is extensively studied in evolutionary biology, but in linguistics there is less agreement, and there are multiple definitions of structural (or typological) stability. I studied stability from a phylogenetic perspective using databases of typological features and Bayesian methods (I developed one specifically for this purpose), and I found that there is agreement across databases, language families, classifications and methods, pointing to a universal tendency for some features to be more stable than others. However, this agreement is not absolute and there is a geographical component, with neighboring families being more similar at the continental scale, suggesting that stability may conserve information about very ancient language relatedness and contact, going back 10-15 thousand years ago. Moreover, I showed that, just like words, structural features evolve in punctuational bursts of accelerated change when languages split. But a more fundamental question concerns the meaning of structural stability, and I think that psychometrics may provide the appropriate framework and conceptual apparatus. I proposed that structural stability is a latent variable that cannot be directly measured, but must be inferred from various observable variables that load on it to various degrees in the way that IQ can only be estimated from extended batteries of diverse tests. This is supported by a comparison of seven different quantitative estimates, and goes against there being a single best definition but rather that each captures an important aspect of structural stability.

Connecting biological and linguistic diversities

Variation between individuals and groups is ubiquitous, structured, motivated by our evolutionary history, and is one of the most positive features of humanity. This biological variation may influence language change, explaining some observed patterns of cross-linguistic universals and diversity, especially (but not exclusively) at the level of the vocal tract, phonetics and phonology.

I am probably best known for my work on the connections between inter-individual differences and cross-linguistic variation. Using statistical methods, Prof. D. Robert Ladd and myself  proposed that the population frequency of the “derived alleles” of two brain growth and development-related genes (ASPM and MCPH1) predicts whether the population speaks a tone language or not. We suggested that the derived alleles influence the learning, production, processing or perception of tone, producing a bias against it (bias that is very weak at the individual level, as every normal child can learn any human language to native proficiency irrespective of their genetic origins) but that can be amplified by the repeated transmission of language in a population in which many learners are biased in such a way. This paper generated a lot of attention from the general public (e.g, The Times Online, New Scientist, Scientific American, Science, National Geographic and Wissenschaft.de) and the scientific community and helped promote research into the influence of non-linguistic factors on language change and diversity. Following this proposal, I showed, using computer simulations, that such weak biases can be amplified in the right social settings, but this amplification is highly non-linear and dependent on the details of the bias and of social structure. Using phylogenetic methods, I have shown that linguistic tone is relatively stable, consistent with the anchoring of tone in the much slower changing (compared to culture) genetic landscape of a population. With many collaborators, I have explored the nature of such a genetic bias affecting tone, looking at individual differences in the perception of missing fundamental tones, the segmentation of a stream of syllables using tone cues, and the learning an artificial tone language using fMRI. Thus, while the jury is still out concerning this association (and there are competing accounts, including climate), the idea that weak biases rooted in our biology might affect language seems to be on the right track.

Figure-biasing
Weak bias amplification through the repeated use and acquisition of language. A highly simplified depiction of how a weak bias (e.g., acting on tone), that is present with different frequencies in two populations (red and blue), might become amplified across generations, affecting language change and “pushing” the languages of the two groups towards two different states that (non-linearly) reflect the variation in bias. This would help explain some of the observed diversity, but if the bias is shared by all humans, it results instead in universal tendencies.

The Vocal Tract (VT) is one of the best places to look for such biases because the genetic influences are more direct, we understand biomechanics, acoustics, and phonetics, we can measure reliability and accurately its anatomy, physiology and phonetic output, and there are large cross-linguistic databases allowing statistical inferences. Between 2012 and 2017, I was awarded a VIDI grant by the Netherlands Organisation for Scientific Research (NWO), Genetic biases in language and speech (or G[3]bils), hosted by the Language and Genetics Department at the Max Planck Institute for Psycholinguistics in Nijmegen. There, we built a model of the hard palate using Bézier curves that can fit existing variation with high accuracy and can generate naturally-looking hard palate shapes, which we integrated in a geometric model of the vocal tract VocalTractLab. This allowed us to build computer agents with a realistic vocal tract, the anatomy of which we accurately control, and which are can produce speech sounds. Using standard machine learning techniques (neural networks and genetic algorithms), we train such agents to learn a given set of speech sounds, and we study the influence of their anatomy on how well they reproduce these sounds in isolation or as part of iterated transmission chains.

Figure-VTL-agents
Studying the influence of Vocal Tract anatomy on vowels across generations using simulated agents. Panel (A): we adapted the VocalTractLab computer model to allow the detailed specification of the shape of the hard palate (yellow ellipse) using Bézier curves and the position of the larynx (red ellipse). Panel (B): we implemented a learning algorithm using generic machine learning techniques (a neural network with synaptic weights optimized by a genetic algorithm) that can be trained to reproduce a given set of vowels. Panel (C): the Iterated Learning Model – a simplified model of language learning and transmission across generations – which amplifies the weak biases due to anatomical variation in our agents’ vocal tract. Such models allow the study of language change when fully controlling the anatomy of vocal tract structures.

We introduced a complementary approach by building realistic biomechanical models implemented in ArtiSynth, that allow us to quantify the influence of anatomical variation on the muscle effort required to produce speech sounds, helping to clarify arguments related to the economy of effort. Using it, we showed that the production of a generalized dental/post-alveolar click is favored, in terms of lower muscle effort and better acoustics, by a reduced alveolar ridge prominence (ARP), providing a possible explanatory mechanism for the distribution of click languages. This biomechanical model is part of a larger project inspired by an older observation that the prototypical “Khoisan” click-language speakers of southern Africa tend to have a reduced ARP, prompting the suggestion that this anatomy might somehow help with click production. Through a large literature review, we found preliminary support for a statistical, metrical difference in the ARP size between “Khoisan” and other populations. Finally, by collaborating with the Donders Institute and the UMC Radboud, we collected a large multi-ethnic sample (ArtiVarK) of phonetic recordings of native and non-native speech sounds (including clicks) in the phonetics lab and during MRI scans, plus detailed information on the anatomy of the vocal tract from 3D intra-oral optical scans. We are currently working on analyzing these data, data that goes beyond the articulation of clicks.

Figure-clicks.jpg
Biomechanical model of the influence of alveolar ridge on click production. Panel (A) shows a comparison of alveolar ridge size between a click language speaker and a white Canadian of European descent. Panel (B) represents the distribution of click languages in Africa (with number of clicks as given by PHOIBLE). Panels (C) and (D) show the three alveolar ridge size conditions and a frame from the simulated articulation of a lingual click using our biomechanical model in ArtiSynth. Panel (E) displays the main results showing that a smaller alveolar ridge is associated, as predicted, with a better volume change (a proxy for the acoustics) and with less overall muscle effort, especially during the constriction and rarefaction phases. See here for details (and video).

An important question concerns the unfolding through historical time of such anatomical influences on speech. We took a first step by investigating, in collaboration with the Laboratory for Human Osteoarchaeology at University of Leiden, the acquisition of information about the hard structures of the vocal tract, using 3D optical scanning, in two post-medieval Dutch populations (Alkmaar, 1484-1574, and Middenbeemster 1829-1866), comparing the jaw and the hard palate.

The genetic foundations of normal variation in speech and language

The genetic foundations of speech and language can be investigated at many levels if we define carefully the aspects of interest and their measurement. The genetics and development of the vocal tract are essential for understanding studying phonetics and phonology.

 

What are the genetic foundations of normal language and speech? To answer this question,  I investigated (together with many collaborators) various aspects of speech and language, including the individual differences during the learning of complex foreign phonology, the changes in brain connectivity (using DTI) associated with the acquisition of artificial and natural languages, and the changes in brain activity (using fMRI) following the learning of an artificial tone language, but also syntactic priming, iconic/ideophonic sensitivity, the learning of Korean consonant contrasts, and color perception and the color lexicon. We are currently working on using anthropological landmarks and semi-landmarks on the vocal tract to estimate the heritability of vocal tract structures and to uncover the genetic foundations of normal vocal tract variation.