The Roads Not Taken: The Neural Dynamics of Lexical Alternatives During Free Language Production

Human language allows multiple ways to express the same thought, raising the possibility that several lexical candidates are activated in parallel before a single word is selected for articulation. While their co-activation is assumed by computational models of word production, until recently, there has been no practical way to estimate the identity of these candidates in unconstrained speech, nor to directly probe their representation in neural activity. Here, we combined high-density electrocorticography recorded during real-world conversation with context-sensitive large language model (LLM) to characterize, for the first time, how these lexical candidates are represented in the brain. We focus on the candidates which remain unsaid, referred to as the "alternatives", by only using samples where the LLM predictions do not include the produced word. We provide evidence for three computational principles of the neural dynamics of lexical alternatives: (1) multiple lexical alternatives to each word in conversation are distinctly represented in neural activity prior to articulation; (2) the set of lexical alternatives is rapidly updated from one word to the next; and (3) alternatives to future words are represented along with alternatives to the upcoming word. Together, these findings reveal that spontaneous speech production is supported by a broad, dynamically evolving and forward-looking set of lexical alternatives, providing a neural account of how the brain prepares not only the words that are spoken, but also those that remain unsaid.