An evolutionary exploration of paradigm syncretism in the world's languages

Abstract

This thesis explores whether the patterning of cells within paradigmatic language structures can be used to diagnose relationships between languages and whether the expected stability of these systems allows for analysis at greater time depth. In doing so, it answers three important questions - are there cross linguistic paradigmatic structures that lend themselves to phylogenetic analysis, can we effectively model these paradigmatic structures for this type of analysis and do these structures contain sufficient phylogenetic signal to be able to study them computationally.

These questions are approached using pronoun paradigms, a system that is expected to be found cross-linguistically and for which there is significant documentary record. Whilst the lexical content of these paradigms has been studied extensively, the patterning of the forms within the paradigms has not. The thesis examines various methodologies for modelling paradigms. Models built by comparing full and partial syncretism between cells are explored before demonstrating that embeddings derived from Large Language Models (LLM) provide the most promising method of capturing the relationships between the cells of a paradigm.

The thesis explores the methods and their utility through three case studies. Firstly, the well-known and well documented Indo-European language family is explored. It is shown that using LLM embeddings as the input for phylogenetic techniques can replicate results produced using traditional historical linguistic techniques. The techniques are then extended to the Torres Strait region where we answer questions about the relationships between languages from Southern New Guinea, Northern Australia, and the Torres Strait. In the final case study, the data from the first two case studies is bought together with Dravidian language data to test whether there is sufficient evidence to link Dravidian languages with either Indo-European or Australian languages. The case studies highlight several issues with the use of LLM embeddings including the impact that linguistic traditions have on the results of the analysis and the impact of LLM design choices on the success of the analysis. Despite the problems that arose, LLM embeddings are shown to provide a dense representation of the paradigms being analysed and it is shown that using these embeddings provide us with an opportunity to explore the relationships between languages in a structured and mathematical way. Show more