Predicting Human Similarity Judgments Using Large Language Models