Reinforcement Learning and Semantic Selection: The Role of the Basal Ganglia in Language

Abstract

The role of the subcortical basal ganglia in language is currently not agreed upon. Research relating these structures to language processes is diverse, and existing theories fail to account for the breadth of findings. To address this, I proposed taking a core basal ganglia neurocomputation, that of reinforcement learning, to explain its involvement in linguistic processes. Lexico-semantics was then used as the level of linguistic processing to test this basal ganglia process. My work consisted of three projects investigating the mechanisms supporting semantic selection: two experiments and one computational simulation using the ACT-R cognitive architecture. Behavioral indices of basal ganglia dual-path learning, or reinforcement learning from positive and negative feedback, showed that semantic selection in the absence of context varies in line with the bias in the estimated value of a specific meaning. The ACT-R computational cognitive model provides a mechanistic explanation of, and causal evidence for, the influence of a reinforcement learning-based action selection system in semantic selection. A self-paced reading sentence task provides further insights into how sentence context and the temporal dynamics of semantic selection process affect basal ganglia-mediated ambiguity resolution. Results from the three projects converge to show that individual differences in learning from both positive and negative feedback through basal ganglia competitive dynamics relate to variability in semantic selection processes. Findings are discussed in light of a general reinforcement learning neurocomputational approach to understanding basal ganglia involvement in language, more generally.