Poster A32, Thursday, August 16, 10:15 am – 12:00 pm, Room 2000AB

Dissociating prediction and constituent-structure during sentence-structure building

Murielle Fabre1, Shohini Bhattasali1, John Hale1, Christophe Pallier2;1Cornell University, 2INSERM-CEA Cognitive Neuro-imaging Unit

Introduction: Sentence processing is more than decoding linear strings of words. Hierarchical relations between words which impact language comprehension are formalized through tree-like structures, and their complexity has been consistently shown to correlate with activity in core brain areas in the language network (Ben-Shachar et al. 2004; Shetreet and Friedman 2014; Pallier et al. 2011). Along with such structural complexity, different computational parsing strategies can be used to investigate the neural correlates of syntactic structure-building. Modelling how sentence structure can be parsed can reveal different components of sentence processing. Thus, the interest in comparing the fMRI activation patterns to Bottom-Up (BU) and Top-Down (TD) parsing strategies (Fig.1) lies in decomposing the cognitive process of sentence-structure building into different sub-processes. BU can instantiate constituent-structure building as it builds and collects sub-parses towards the end of the phrase/sentence. The rules of a grammar are applied at each incoming word, as seen through the parser action counts (Fig.2). Alternatively, TD better approximates expectation-driven structural processing, as rules are applied predictively, in advance of each word, thus assigning higher scores at the beginning of sentences. Methods: Participants (n=51, 32 female) listened to The Little Prince’s audiobook for 1 hour and 38 minutes. Participants' comprehension was confirmed through multiple-choice questions (90% accuracy, SD = 3.7%). Functional scans were acquired using multi-echo planar imaging sequence (ME-EPI) (TR=2000ms; TE's=12.8, 27.5, 43ms; FA=77 degrees; FOV=240.0mm X 240.0mm; 2X image acceleration; 33 axial slices, voxel-size 3.75 x 3.75 x 3.8mm). Preprocessing was done with AFNI16 and ME-ICA v3.2 (Kundu et al., 2011). The number of parser actions required, word-by-word, to build the correct syntactic tree as determined by the Stanford parser was computed according to two parsing strategies, described above and illustrated by a sentence from the auditory stimuli (Fig.2). Along with these syntactic structure building regressors, we entered four regressors of non-interest into the GLM analysis (SPM12): word-offset, word frequency, pitch, intensity. The regressors were not orthogonalized. The whole-brain main effects were FWE-corrected (T-score>5.3). Results: Regression analyses localized the activation patterns for BU and TD to different areas in the brain. BU showed bilateral clusters. The peak activation was observed in right TP within a main cluster extending to STG through MTG, while ATL involvement was bilateral. Increased activation of LIFG and RIFG stretching over Pars Orbitalis and Triangularis and extending to the anterior Insula and Putamen was observed together with the clusters reported in Fig. 3. For TD, two bilateral clusters were observed along STG extending from its posterior portion to MTG, reaching TP in the right hemisphere (Fig.3). Conclusion: Predictive syntactic processes modeled by TD evoke a pattern of activation that is spatially-dissociable from compositional structure-building modeled by BU. This result replicates findings about surprisal (Willems et al. 2015). Consistent with previous work (Nelson et al. 2017; Brennan et al. 2016), these findings underline that different parts of the language network functionally contribute to different dimension of sentence-structure building during natural language comprehension.

Topic Area: Grammar: Syntax