Kelime çağrışımları: Yapay Zeka ve insan kelime çağrışım yapılarının karşılaştırılması

This thesis explores the similarities and differences between human word associations and those generated by large language models (LLMs) through a systematic comparison of responses in a Turkish word association task. Using human participants and selected LLMs, this study examines key metrics, including response diversity, associative strength, sensitivity to linguistic features such as morphology and concreteness, and structural alignment within associative networks. The methodology involves a continuous word association experiment conducted with 381 human participants in addition to simulations using state-of-the-art LLMs, including GPT-4o mini, Trendyol LLM v1.8, and Cosmos LLaMA. Comparative analyses employ techniques such as cosine similarity, Jaccard similarity, and graph-based metrics to evaluate structural and associative parallels. Results indicate notable differences in the associative patterns of humans and LLMs, particularly in response diversity and sensitivity to morphological and semantic properties. While LLMs exhibit strengths in replicating certain associative structures, they face challenges in mimicking the nuanced cognitive processes observed in humans. The findings underscore the potential of word association tasks as a lens for understanding human cognition and evaluating artificial intelligence systems. This research bridges traditional psycholinguistic methodologies with contemporary computational approaches, advancing our understanding of semantic memory and highlighting the evolving potential and limitations of LLMs in approximating human cognition.