Search Abstracts | Symposia | Slide Sessions | Poster Sessions
From structure to sequence: a Multinomial Processing Tree model of syntactic encoding
Poster Session F, Friday, October 2, 2:45 - 4:45 pm, Wangari Maathai
Jeremy Yeaton1,2, Grant Walker1, Danielle Fahey3, William Matchin4, Julius Fridriksson4, Gregory Hickok1; 1University of California, Irvine, 2University of California, Davis, 3University of Alabama, 4University of South Carolina
Two-stage models of grammatical encoding posit that sentence production unfolds sequentially: first, a hierarchical syntactic structure is constructed encoding the relational organization of the message; second, this structure is linearized into a sequence suitable for articulation. While widely accepted in psycho- and neurolinguistics, such models have never been formally implemented computationally, leaving open the question of whether individual differences in stage-specific processing abilities can be quantified and mapped onto distinct neural substrates. Here we address this gap by introducing a Multinomial Processing Tree (MPT) model that operationalizes the two-stage framework as a generative process and fits it to behavioral data from individuals with aphasia. As input data, we used discourse samples from 79 individuals with chronic post-stroke aphasia retelling the Cinderella story, annotated for syntactic error type. Each utterance was classified into one of five categories: Grammatical, Hierarchical error (structural anomalies including subcategorization violations, exchange errors, and ungrammatical insertions), Linearization error (omission of multiple functional morphemes), Mixed error (features of both), or Omission (single functional morpheme omission, indeterminate as to stage). These categories formed the leaf nodes of the MPT model, which included three latent parameters: a, the probability of successfully constructing the hierarchical representation of an utterance; b, the probability of successfully linearizing a hierarchical representation; and c, the probability that a given error presents as an overt structural error rather than a morpheme omission. Model parameters were estimated using Bayesian Gibbs sampling implemented in JAGS, yielding a posterior distribution of credible parameter values for each participant. We compared model fit against a simple multinomial model using the Deviance Information Criterion (DIC) and validated parameter estimates against observed error rates using Pearson correlations. Lesion-symptom mapping was performed using SVR-LSM on high-resolution MRI data, correcting for lesion volume and thresholding at p<.005 based on 10,000 permutations. The MPT model recovered an average of 93.6% of variance in observed response type rates across error categories, though the simpler multinomial model achieved a lower DIC (1,154 vs. 1,284), driven largely by the MPT's poorer fit to Mixed errors. MPT ability estimates correlated significantly with observed error rates for their corresponding error types: a with hierarchical error rate, b with linearization error rate, and c with omission error rate (all p<.01, Bonferroni-corrected). Lesion mapping revealed that a was associated with damage to the posterior middle temporal gyrus and angular gyrus, while b was associated with damage to the posterior frontal lobe including the middle frontal gyrus and inferior frontal sulcus. Parameter c was associated with damage to the superior frontal sulcus. These findings provide the first computational instantiation of a two-stage model of sentence production and demonstrate that stage-specific ability estimates derived from naturalistic discourse data localize to theoretically predicted neural regions. The convergence of model parameters with established lesion correlates of agrammatism and paragrammatism supports a neurocomputational dissociation between hierarchical and linearization stages of grammatical encoding, and illustrates the potential of MPT modeling as a tool for individual-level characterization of syntactic deficits in aphasia.
Topic Areas: Computational Approaches, Language Production