Measuring children's associative divergent ability through drawings and semantic similarity

Currently, the main method for measuring children's DT is the "Torrance Test of Creative Thinking" (TTCT). However, this method is based on a scale-based relative measurement, which not only has high measurement costs but also yields results that are not actual numerical values, making it impossible to participate in data/pattern mining and other operations.

In this regard, this paper proposes a new approach to measuring children's DT: based on a drawing completion task, using a large language model (LLM) image recognition method to convert drawing information into semantic information, and then using NLP algorithms to calculate semantic similarity (SS), the DT information of 827 children's drawings was measured. The results show that: ① in terms of divergence direction, the theme of "nature" is the most prevalent, with an overall power-law distribution, and the overall SS is concentrated in the range of 0.5-0.75; ② the SS of gradient tasks is significantly more convergent than that of the main task, but there is significant variation in SS across different gradient tasks, indicating that DT is greatly influenced by priming items; ③there is little gender difference in DT, but significant age differences exist, and it is clearly influenced by learned knowledge; ④ after controlling for priming items, the differences in SS are inconsistent, indicating that children's DT does not end after the priming item is recognized as a certain object, and some categories persist throughout the process of completing the drawing.