Decoding the Neural Architecture of Compositional Meaning Generalization
Poster No:
914
Submission Type:
Abstract Submission
Authors:
Xiaochen Zheng1, Mona Garvert2, Hanneke den Ouden1, Lisa Horstman3, David Richter4, Roshan Cools3
Institutions:
1Donders Institute for Brain, Cognition and Behaviour, Nijmegen, Netherlands, 2Julius-Maximilians-Universität Würzburg, Würzburg, Germany, 3Radboud University Medical Center, Nijmegen, Netherlands, 4Vrije Universiteit Amsterdam, Amsterdam, Netherlands
First Author:
Co-Author(s):
Introduction:
The ability to generalize previously learned knowledge to novel situations is essential for adaptive behavior. We are very good at combining "building blocks" for inferring the meaning of novel compositional words. For example, when encountering the word "un-reject-able-ish" for the first time, one can easily infer its meaning by integrating knowledge of its constituent parts based on abstract structural rules (such as the sequential order of word parts). In this study, we investigate the neural mechanisms of the ability to infer novel compositional word meanings. Specifically, we aimed to assess whether compositional inference in language recruits a medial prefrontal-hippocampal network that is also recruited by compositional action planning, compositional vision and constructive relational memory (Baram et al., 2021; Barron et al., 2020; Schwartenbeck et al., 2023).
Methods:
In a pre-registered fMRI study (aspredicted.org/mk5i2.pdf), we taught 43 participants the meanings of artificial compositional words consisting of known stems ("good") and novel affixes ("kla") (Figure 1A). The meaning of the compositional words depended on the position of the novel affix ("goodkla = bad", "kladog = puppy"). We then asked them to infer the meaning of novel compositional words ("richkla =?", "klarich =?") that were either congruent or incongruent with the established rule ("klarich" is incongruent because a small version of "rich" does not exist). In the scanner, participants performed a semantic priming task in which the novel words served as either congruent ("richkla") or incongruent primes ("klarich") and their synonyms ("poor") served as targets. After the scanning session, they were asked to indicate (i) whether these novel words held any meaning and, (ii) if so, what they meant.
Analyses of fMRI data focused on the inferential computation as a function of univariate repetition suppression (Barron et al., 2016) and multivariate representational similarity (RSA, Kriegeskorte et al., 2018).
Analyses of fMRI data focused on the inferential computation as a function of univariate repetition suppression (Barron et al., 2016) and multivariate representational similarity (RSA, Kriegeskorte et al., 2018).
·Figure 1. Experimental paradigm and behavioral results.
Results:
Our results demonstrated that participants were able to generate novel compositional meanings on the fly, successfully inferring meanings of congruent versus incongruent words (Figure 1B). Univariate analysis of congruent versus incongruent prime-related fMRI activity revealed a broad frontal-parietal network, including the hippocampus, a brain area commonly associated with the generalization process of structural relationships (Figure 2A, top panel). Analysis at target words revealed greater repetition suppression when primed with congruent than incongruent words in the left inferior frontal gyrus, which suggests that novel meanings are presented in this linguistic "building" hub (Figure 2A, bottom panel). Furthermore, multivariate RSA revealed that, excitingly, this representation was already decodable at the time of prime (Figure 2B, "meaning representation"). Intriguingly, RSA also revealed representations of the newly derived abstract rules in a bilateral frontoparietal network (Figure 2B, "rule representation"), also commonly implicated in the representation of task state spaces and abstract rules in working memory for goal-directed action planning.
·Figure 2. Univariate and multivariate representational similarity analysis of fMRI data.
Conclusions:
The compositional nature of language enables us to freely combine morphemes into words, words into sentences, and to convey an infinite array of thoughts and ideas. Using an artificial language learning paradigm, we show that participants are able to generalize recently learned structural rules to infer novel meanings on the fly. This compositional process in language engages a domain-general control network, while newly inferred meanings are represented in more language-specific regions (Hagoort, 2005). Our results demonstrate that our ability to generate novel compositional meaning representations for language recruits abstract rule representations in a frontoparietal network that also represents abstract task representations for generative action selection (Vaidya & Badre, 2022).
Higher Cognitive Functions:
Executive Function, Cognitive Control and Decision Making 1
Reasoning and Problem Solving 2
Language:
Language Comprehension and Semantics
Learning and Memory:
Learning and Memory Other
Modeling and Analysis Methods:
Multivariate Approaches
Keywords:
Cognition
FUNCTIONAL MRI
Language
Learning
Memory
Multivariate
Pre-registration
Univariate
1|2Indicates the priority used for review
Provide references using author date format
Barron, H. C., Garvert, M. M., & Behrens, T. E. (2016). Repetition suppression: a means to index neural representations using BOLD?. Philosophical Transactions of the Royal Society B: Biological Sciences, 371(1705), 20150355.
Barron, H. C., Reeve, H. M., Koolschijn, R. S., Perestenko, P. V., Shpektor, A., Nili, H., ... & Dupret, D. (2020). Neuronal computation underlying inferential reasoning in humans and mice. Cell, 183(1), 228-243.
Hagoort, P. (2005). On Broca, brain, and binding: a new framework. Trends in Cognitive Sciences, 9(9), 416-423.
Kriegeskorte, N., Mur, M., & Bandettini, P. A. (2008). Representational similarity analysis-connecting the branches of systems neuroscience. Frontiers in Systems Neuroscience, 4.
Schwartenbeck, P., Baram, A., Liu, Y., Mark, S., Muller, T., Dolan, R., ... & Behrens, T. (2023). Generative replay underlies compositional inference in the hippocampal-prefrontal circuit. Cell, 186(22), 4885-4897.
Vaidya, A. R., & Badre, D. (2022). Abstract task representations for inference and control. Trends in Cognitive Sciences, 26(6), 484-498.