Abstract
Every month, scores of new measures and constructs enter the behavioral science literature. Researchers struggle to keep an overview of their subfields, let alone the wider field, with an estimated number of allegedly different constructs now exceeding 30,000. Proposing constructs is easy, finding redundancies and overlaps is not. The effort of finding redundancies using participant responses grows quadratically with the number of measures. This cost motivates the use of language models, which can simplify and speed up the process without requiring new data collection. Earlier work employed latent semantic analysis (LSA) to make use of the textual nature of questionnaires in the social and behavioral sciences. However, LSA treats sentences as bags of words and ignores word order, syntax, and negation, all of which shape how respondents interpret items. Recent advances demonstrate that transformer-based sentence embedding models offer a more viable solution for predicting empirical relatedness between survey items. In the presented work, we use a fine-tuned language model, the SurveyBot3000, to introduce the Synthometricon, an open-access web application for systematic detection of conceptual overlap. The application indexes over 570,000 items across more than 97,000 scales from approximately 28,000 instruments in the APA PsycTests database and other sources.