Investigating the category effect in freely associated word responses and predicting free word associations using language models

The process of free association is key to human meaning-making, influencing both shared and individual understanding. Useful in many aspects, free association word resources are costly to make and attempts in computational modelling of free associations has yielded limited results only. Designed in two separate steps, current study first investigates whether sensory categories in the brain play a predictive role in the process of freely associating words. For this purpose, the data previously collected by De Deyne et al. (2018), namely Small World of Words (SWOW), was categorized by the aid of the pretrained Word2Vec model and analyzed. Secondly, the study tests whether a computational language model is able to capture the general regularities of freely associated human responses after being fine-tuned with relevant data. For this purpose, two language models, namely Tiny Llama and Google Gemma 3 were fine-tuned with SWOW data and the models were tested. Results show that there is a significant relation between the sensory areas in the brain and free word association responses and it is possible to say that fine-tuning significantly improves semantic understanding in free association responses in both models. However, values also highlight the limitations of both the base and fine-tuned models for handling such an open-ended task.