Abstract rule generalization for composing novel meaning recruits the frontoparietal control network

Abstract The ability to generalize previously learned knowledge to novel situations is crucial for adaptive behavior, representing a form of cognitive flexibility that is particularly relevant in language. Humans excel at combining linguistic building blocks to infer the meanings of novel compositional words, such as “un-reject-able-ish”. How do we accomplish this? While recent research on compositional generalization in relational memory, action planning and vision strongly implicates a medial prefrontal-hippocampal network (Baram et al., 2021; Barron et al., 2020; Schwartenbeck et al., 2023), it remains unclear whether the same network supports compositional inference in language. To this end, we trained participants on an artificial language in which the meanings of compositional words could be derived from known stems and unknown affixes, using abstract affixation rules (e.g., “good-kla” which means “bad”, where “-kla” reverses the meaning of the stem word “good”). According to these rules, word meaning depended on the sequential relation between the stem and the affix (i.e., pre- vs. post-stem). During fMRI, participants performed a semantic priming task, with novel compositional words as either sequential order congruent (e.g., “white-kla”) or incongruent primes (e.g., “kla-white”), and synonyms of the composed meaning as targets (“black”). Our results show that the compositional process engages a broad temporoparietal network, including the hippocampus, while the composed meanings are represented in left-lateralized language areas. Notably, newly composed meanings were decodable already at the time of the prime. Finally, we found that the composition process recruits abstract structure rule representations in a lateral frontoparietal network, rather than the predicted medial frontotemporal network, perhaps because abstract rules in language are more readily formatted as production rules rather than as relational map structures. Competing Interest Statement The authors have declared no competing interest. Data and Code Availability All the research data (e.g., data files, analysis scripts) associated with the current paper are shared in the Donders repository (https://data.donders.ru.nl/).