Distributional semantic representations predict high-level human judgment in seven diverse behavioral domains

AbstractThe complex judgments we make about the innumerable objects in the world are made on the basis of our representation of those objects. Thus a model of judgment should specify (a) our representation of the many objects in the world, and (b) how we use this knowledge for making judgments. Here we show that word embeddings, vector representations for words derived from statistics of word use in corpora, proxy this knowledge, and that accurate models of judgment can be trained by regressing human judgment ratings (e.g., femininity of traits) directly on word embeddings. This method achieves higher out-of-sample accuracy than a vector similarity-based baseline and compares favorably to human inter-rater reliability. Word embeddings can also identify the concepts most associated with observed judgments, and can thus shed light on the psychological substrates of judgment. Overall, we provide new methods and insights for predicting and understanding high-level human judgment.

Return to previous page