A computational approach to linguistic relativity with empirical experimentation

This thesis investigates linguistic relativity (LR), a hypothesis proposing that language influences thought, through a computational framework utilizing large language models (LLMs). To empirically test this proposition, the study prompt engineers LLMs as proxies simulating bilingual speakers, each representing a collective linguistic identity corresponding to one of 24 distinct native language groups, whose members are characterized by varying levels of English proficiency. Semantic embedding vectors are generated by eliciting adjective associations for nouns from various LLM configurations, thereby forming language specific conceptual representations. These embeddings are used to calculate semantic distances through measures of cosine similarity, whereby all of these language's conceptual representations are mapped to form a semantic field. Further experiments are conducted on these representational bilingual proxies to include both replications of existing studies in LR literature and novel cognitive tasks across a variety of domains, with results analyzed for systematic variations in task performance. The findings indicate that while native speakers of diverse languages employ distinct lexical choices to articulate conceptualizations of identical entities even in a common medium language (English), experimental data did not show statistically significant variations at the cognitive level of thought. Consequently, it is concluded that these cross-linguistic distinctions remain confined to the linguistic surface instead of extending to the deeper cognitive level of thought. Overall, by bridging the fields of linguistics and artificial intelligence, this study contributes to the ongoing investigation of LR within cognitive science studies by presenting a innovative and scalable methodological approach that addresses traditional limitations.