Multilingual semantic distance: Automatic verbal creativity assessment in many languages

Creativity research commonly involves recruiting human raters to judge the originality of responses to divergent thinking tasks, such as the alternate uses task (AUT). These manual scoring practices have benefited the field, but they also have limitations, including labor-intensiveness and subjectivity, which can adversely impact the reliability and validity of assessments. To address these challenges, researchers are increasingly employing automatic scoring approaches, such as distributional models of semantic distance. However, semantic distance has primarily been studied in English-speaking samples, with very little research in the many other languages of the world. In a multilab study (N = 6,522 participants), we aimed to validate semantic distance on the AUT in 12 languages: Arabic, Chinese, Dutch, English, Farsi, French, German, Hebrew, Italian, Polish, Russian, and Spanish. We gathered AUT responses and human creativity ratings (N = 107,672 responses), as well as criterion measures for validation (e.g., creative achievement). We compared two deep learning-based semantic models—multilingual bidirectional encoder representations from transformers and cross-lingual language model RoBERTa—to compute semantic distance and validate this automated metric with human ratings and criterion measures. We found that the top-performing model for each language correlated positively with human creativity ratings, with correlations ranging from medium to large across languages. Regarding criterion validity, semantic distance showed small-to-moderate effect sizes (comparable to human ratings) for openness, creative behavior/achievement, and creative self-concept. We provide open access to our multilingual dataset for future algorithmic development, along with Python code to compute semantic distance in 12 languages.