The psychological sciences have been limited by the lack of conceptual clarity and a common taxonomy of psychological constructs and measures. We show that item embeddings (i.e., representations of psychometric items in a vector space obtained from natural language processing methods) can help deal with this problem by quantifying both empirical and conceptual overlap between measures and providing a common conceptual representation. Specifically, we analyse 459 psychological measures consisting of more than 4000 items to show that item embeddings can predict observed empirical correlations between measures, identify jingle-jangle fallacies, and suggest a novel taxonomy of personality constructs. All in all, our work suggests that item embeddings offer a powerful tool to address the incommensurability problem in the psychological sciences.