Investigating the origins of cross-cultural variation in kinship terminology with artificial language learning

Every society has a set of words for talking about family. This set of kin terminology is often smaller than the number of family relationships, or kin types, that it refers to: the same term can denote multiple types, like grandmother denoting both your father's mother and your mother's mother. Kinship systems - how kin terminology maps onto kin types - vary across languages and affect how we understand family members to be grouped together or distinguished from one another. For instance, in Indonesian, siblings are distinguished terminologically by relative age, while English speakers distinguish their siblings by gender.



Theoretically, there are over 10 billion possible ways to map kin terms to kin types. However, only a handful of these mappings occur in the world's languages: there is no language that uses one single word for 'mother', 'father', and 'sibling', for instance. This pattern of constrained variation suggests that some kinship systems are more likely to survive social and linguistic evolution, perhaps due to advantages in their learnability or communicative function.



Language persists through a cycle of learning and use, placing contrasting requirements on language to be sufficiently expressive, by encoding distinctions between concepts with distinct words, but also simple enough to be learned, by reducing the number of distinct words overall. Kemp and Regier (2012) found that kinship systems across the world's languages optimally trade-off these requirements: though each system varies in simplicity and expressivity, they are maximally simple given their expressivity and maximally expressive given their simplicity. Additionally, being closely linked to cultural conventions about family structures, kin terminology is shaped by social pressures as well as linguistic ones. By combining insights from experimental language evolution and linguistic anthropology, my research will investigate how these linguistic pressures shape kinship terminology, and to what extent social structures can influence kinship terms within the constraints of those pressures, giving rise to the typology of kinship terms we observe cross-linguistically.



We can study typological universals using experimental methods like artificial language learning, which replicate the processes of learning and use by which language evolves. Such studies have shown that the demands of these processes can affect which linguistic properties tend to arise and persist, and can therefore shed light on the pressures which caused constrained variation in kinship systems. For instance, Smith et al (2020) have shown that simpler artificial kinship systems are learned more accurately than complex ones, and that errors in learning tended to reduce the number of distinct kin terms, indicating that learners may have a bias for simplicity in kinship systems.



The proposed research will expand upon this artificial language approach by more accurately simulating the conditions under which kin terms develop, including learning biases, expressivity biases, and the effect of social ideology surrounding family roles. I will foreground the existing cross-cultural variation, testing whether participants' learning or communication behaviour differs for attested versus non-attested organisations of kinship systems. I will use KinBank, a database of kinship terminology for over 1000 societies, to inform the design of these artificial kinship systems. Across a series of experiments and computer simulations of language evolution, I will test whether agents learn attested kinship systems more accurately than unattested ones, and whether they are more successful in communicative tasks with attested versus unattested systems. If we see a higher success rate with attested systems, this would suggest that language evolution favours kinship systems designed for ease of learning and efficient communication.