Automated Graphic Divergent Thinking Assessment: A Multimodal Machine Learning Approach

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Educational Psychology

Background:

• Automated assessment of divergent thinking (DT) is crucial for evaluating creativity.
• Image-based DT tests offer rich, multimodal data but pose scoring challenges.
• Current methods often lack objectivity and efficiency in creativity assessment.

Purpose of the Study:

• To develop and validate a multimodal deep learning model for automated scoring of image-based divergent thinking tests.
• To integrate visual and semantic features for enhanced objectivity and efficiency in creativity assessment.
• To evaluate the model's performance against human expert scores.

Main Methods:

• A multimodal deep learning model was developed using ResNet50 for image features and GloVe for text embeddings.
• A fully connected neural network fused these features, trained with Mean Squared Error (MSE) loss and Adam optimization.
• The model was trained and validated on 708 responses from Chinese high school students.

Main Results:

• The model demonstrated strong alignment with human scores on the training set (Pearson r = 0.810).
• Validation on unseen data showed good generalization capacity (r = 0.561).
• Participant-level analysis achieved a correlation of 0.602 with total human scores.

Conclusions:

• Multimodal deep learning effectively captures key dimensions of divergent thinking (novelty, fluency, flexibility).
• This approach significantly reduces subjectivity and streamlines the assessment process for creativity.
• The findings support the use of visual-textual feature fusion for robust, cost-effective cognitive evaluation.