Fine-tuned Large Language Models predict human word associations and improve performance across lexical and semantic processing tasks.
Abstract
Recent studies have shown that LLMs can predict various aspects of semantic cognition when prompted with instructions that closely mimic those of humans. This study examines to what degree LLMs can predict human word associations, asking two previously underexplored questions: (1)~Does fine-tuning LLMs on a limited amount of human data increase prediction quality? And (2)~Does a more human-like association prediction also facilitate predicting other behaviors such as lexical processing and similarity judgments? Our findings suggest that the answer to both questions is yes. We first show that word associations can be predicted more accurately after fine-tuning. This was especially evident for the strongest associates, as the median rank of the predicted first associate was 1, a substantial improvement over previous approaches. Fine-tuning also enhances reflection of human response biases (e.g., favoring short, frequent responses). This close alignment with human associations further benefited the prediction of lexical processing and similarity judgments. Our results have implications on using LLMs fine-tuned on word associations in studies of a wide range of cognitive behavior, and on capturing factors that drive diversity in human associations.