This Paper is “Un-reject-able-ish”: Learning and Generalization of Novel Compositional Meanings

The ability to generalize previously learned knowledge to novel situations is fundamental for adaptive behavior. When encountering the novel word “un-reject-able-ish” for the first time, one can swiftly infer its meaning by generalizing from the known elements and integrating them based on abstract structural rules, such as the sequential arrangement of word parts. How do we generate such novel, compositional meaning? To address this question, we developed a behavioral paradigm to quantify structural inference for the creation of word meaning. Participants were taught compositional pseudo-words from an artificial language and subsequently tested with a new set of pseudo-words from the same language. Across three behavioral experiments, we demonstrate that participants can efficiently learn and apply structural rules to infer novel, compositional meanings spontaneously. Moreover, our findings reveal two distinct behavioral patterns during this compositional generalization process. Some individuals employ a rule-based “building” strategy that takes into account sequential order rules, while others adopt a “mixing” strategy that combines individual elements’ meanings but places less emphasis on the structural rules. We further demonstrated that this individual variability in reliance on the rule-based building strategy for meaning composition is partially explained by individual differences in stable trait factors, including general nonverbal intelligence and vocabulary. By investigating the mechanisms of compositional generalization within the realm of language, our research bridges the gap between linguistic compositionality and other domain-general cognitive processes, such as structural inference, visual composition and decision making.