Testing semantic compositionality in baboons (Papio papio) through relearning and generalization.

This study investigates whether baboons are capable of semantic compositionality, specifically, whether they can apply compositional rules to new situations (generalization). In language, semantic compositionality is linked to productivity, the generalization of a rule to new combinations. Across four experiments, baboons were trained to match visual stimuli based on either shape or color depending on symbolic cues. Experiments 1-3 tested generalization under different task complexities but consistently failed to show evidence that baboons understood or applied the matching rules beyond memorized combinations. Only in Experiment 4, which used a relearning paradigm rather than generalization, did baboons show improved performance when the rule remained consistent across phases. Four hypotheses were explored to explain the lack of generalization: an iconicity-novelty bias, the possibility that compositionality is present, but that training was not sufficient for generalization, rote memorization of cue-sample pairs, and a difference between implicit and explicit learning. The findings do not allow us to discriminate between these hypotheses.