Constructing a Norm for Children’s Scientific Drawing: Distribution Features Based on Semantic Similarity of Large Language Models

Abstract

Using children’s drawings to understand children’s concept learning has been proven to be an effective method, but there are still two major problems in previous research: 1. The drawings heavily relies on the task, so the ecological validity of the conclusions is low; 2. The subjective interpretation of drawings is inevitable. To address these problems, this study uses the Large Language Model (LLM) to identify the drawing contents of 1096 children’s scientific drawings (covering 6 scientific concepts), and uses the word2vec algorithm to calculate their semantic similarity. The study explores whether there are consistent drawing representations among children with the same theme, and attempts to establish a norm for children’s scientific drawings, providing a baseline reference for following children’s drawing research. The results showed that there were significant differences in the consistency of children’s representations of different concepts, and there was a possibility of consistency bias, that is, the appearance of consistency representations misled LLM. At the same time, linear regression tests were used to analyze the relevant factors that affect children’s representation. The results show that sample size and teaching strategies can affect the accuracy of LLM’s image recognition, while the degree of conceptual abstraction may affect the consistency of representation.