Project

Self-monitoring of language in productionElin Runnqvist(1), Chotiga Pattamadilok(1), Kristof Strijkers(1),

Pascal Belin(2), F.-Xavier Alario(3), Mireille Bonnard(4)

(1) LPL, (2) INT, (3) LPC, (4) INS

AbstractLa Banque centrale a organisé un mode d’intervention permettant de comploter… de compléter ledispositif » (François Hollande, 2012).

Just like the president of France, once in a while speakers commit “slips of the tongue”. When considering the anatomical and cognitive complexity of language production, involving the use and orchestration of several organs, body parts, mental representations and levels of processing, a pressing question is how is it possible that errors do not occur more often?

Through our recent research we examined the possibility that speech errors may be avoided similarly to how motor actions are controlled (e.g., Pickering & Garrod, 2013). Motor control is thought to involve internal modeling of upcoming actions, predicting their sensory consequences before they occur (i.e., efference copying, e.g., Wolpert et al., 1995) and suppressing their expected sensory response (i.e., reafference cancellation), leading to a larger activity in the relevant sensorial brain areas whenever there is a mismatch between what is expected and what actually occurs. Using low-frequency rTMS (Runnqvist et al., 2016), we found evidence for a causal implication in language production monitoring of the right cerebellum, a neural structure crucial for internal modeling (e.g., Blakemore et al., 2001). In another study (Runnqvist et al., in prep.), we observed that correctly produced words in conditions of high monitoring load resulted in more negative EEG amplitudes around 100 ms after the onset of the vocal response, a finding that can be related to the N100 component indicating auditory reafference cancellation as a consequence of internal modeling (e.g., Heinks-Maldonado et al., 2005).

Other studies obtained evidence that speech errors may be detected in a similar way as we manage irrelevant but potentially conflicting information when driving a car (e.g., Nozari et al., 2010). Both types of activity would be achieved through conflict monitoring involving the anterior cingulate cortex (ACC) and the pre supplementary motor area (pre-SMA). For example, Riès et al. (2011) observed that speech errors, just as errors in other cognitive domains, elicited more pronounced error-related negativities (ERN), an electrophysiological component with a source localized to the ACC. In an fMRI study, Gauvin et al. (2016) showed that speech errors involved increased activation of both ACC and pre-SMA.

The goal of the fMRI experiment proposed here is to advance in this novel line of research examining the possibility of shared processes between language production monitoring, motor control and conflict monitoring by investigating different aspects of error monitoring (prevention and detection) and levels of processing more or less distant to speech motor processes (lexical and phonetic variables).

Publications-

28 février 2020

Fiche-résumé contribution CREx

SLIP (en cours)

Self Monitoting of language in production Investigateurs: Elin Runnqvist (LPL), Chotiga Pattamadilok (LPL), Kristof Strijkers (LPL), Pascal Belin (INT), F.-Xavier Alario (LPC) & Mireille Bonnard (INS).

Durée : depuis mai 2017

Contribution : Aide à la passation, au prétraitement et à l’analyse des données.

Objectif : Etudier la contribution respective des processus impliqués dans le monitoring de la production du langage, le contrôle moteur et le monitoring des conflits en recherchant des aspects différents du monitoring des erreurs (prévention et détection) et des niveaux de traitement plus ou moins distants aux processus articulatoires (variables lexicales et phonétiques).

▮▮ Paradigme – Lecture à voix haute d’une paire de mots avec amorçage induisant en erreur

28 février 2020

Tâche SLIP (adaptation de Baars et al., 1975)

Cette tâche provoque expérimentalement des lapsus de la langue (« Slips of the tongue »). Elle consiste à présenter des paires de mots, les unes après les autres sur un écran, et à demander au sujet de lire à haute voix la paire indicée (celle qui est suivie « ? ») avant l’apparition d’un autre indice (« !!! » ; voir Figure 1 en Annexe). Les paires utilisées pour l’amorçage sont construites pour induire des erreurs (e.g. pour l’amorçage : « balle mouche », « bain mets », « berne mille », pour la cible : « mec bise » construite pour provoquer l’erreur « bec mise »).

▮▮ Passation – Passation effectuée auprès de 24 adultes de langue maternelle française (en cours)

▮▮ Prétraitement – Prétraitement des données sur SPM12

▮▮ Analyse – Analyse univariée sur le cerveau entier, Analyse par régions d’intérêt et connectivité

▮▮ Diffusion – Article soumis en novembre 2020 (Cerebral Cortex Communications ♠). An involvement of the cerebellum in internal and external speech error monitoring".

Annexe

Short description of Experimental Protocol

SLIP Task during fMRI

Specifically, we will use the SLIP task (e.g., Baars et al., 1975), in which word-pairs are read silently, and on a prompt (?), speakers are asked to produce the last pair they have seen before the appearance of aprompt (!!!, see Figure 1). The pairs immediately preceding the target are intended to prime errors (e.g. “balle mouche”, “bain mets”, “berne mille”, for the target “mec bise” priming the error “bec mise”). Keeping the phonological error-inducing priming constant, monitoring load will be varied by manipulating variables known to affect the likelihood of committing an error: lexicality of the primed error (lex for words and nonlex for non-words) and phonetic distance of word onsets (3 levels of distance D, D1 to D3).

28 février 2020

Figure 1. Schematic illustration of experimental procedure and design in SLIP task

Data acquisition and preprocessing

Data were collected on a 3-Tesla Siemens Prisma Scanner (Siemens, Erlangen, Germany) at the Marseille MRI center (Centre IRM-INT@CERIMED, UMR7289 CNRS & AMU) using a 64-channel head coil. Functional images (EPI sequence, 54 slices per volume, multi-band accelerator factor 3, repetition time= 1.224 s, spatial resolution= 2.5 x 2.5 x 2.5 mm, echo time=30 ms, flip angle=65°) covering the whole brain were acquired during the SLIP task. Whole brain anatomical MRI data were acquired using high-resolution structural T1-weighted image (MPRAGE sequence, repetition time= 2.4 s, spatial resolution= 0.8 x 0.8 x 0.8mm, echo time=2.28 ms, flip angle= 8°) in the sagittal plane. Prior to functional imaging, Fieldmap image (Dual echo Gradient-echo acquisition, repetition time= 7.06 s, spatial resolution= 2.5 x 2.5 x 2.5 mm, echo time=59 ms, flip angle= 90°) was also acquired.

The fMRI data were pre-processed and analyzed using Statistical Parametric Mapping software (SPM12, http://www.fil.ion.ucl.ac.uk/spm/software/spm12/) on matlab R2018b (Mathworks Inc., Natick, MA). The anatomical scan was spatially normalized to the avg152 T1-weighted brain template defined by the Montreal Neurological Institute using the default parameters (nonlinear transformation). The fieldmap images were used during the realign and unwarp procedure for distortion and motion correction. Functional volumes were spatially realigned and normalized (using the combination of deformation field, coregistred structural and sliced functional images) and smoothed with an isotropic Gaussian kernel (FWHM=5mm). The Artifact Detection Tools (ART implemented in CONN toolbox (www.nitrc.org/projects/conn, RRID:SCR_009550) was used to define the noninterest regressors related to head movements and functional data outliers (see next section). Automatical ART-based identification of outlier scans used a 97th percentiles superior to normative samples in the definition of the outlier thresholds (global-signal z-threshold of 5 and subjectmotion threshold of 0.9 mm). Four participants were excluded from all analyses, three because of excessive head movements during the acquisition and one

28 février 2020

because of a misunderstanding of the task (reading of all words even non-targets ones). A total of 24 subjects (18 women and 6 men, 23.83 +/- 3.20 years old on average) are retained for all analyses in this study.

Univariate Analysis on Whole BrainFor each subject, a general linear models (GLM) was generated from the SLIP task. The GLM included, for each of the six runs, 7 regressors modelling a combinaison of lexicalicality monitoring load (lex for lexical and nonlex for nonlexical items), phonetic monitoring load (Phon_1, Phon_2 vs Phon_3 as a function of phonetic features distance) and an error responses factor (all types of errors except no responses): lex_Phon1_CR, lex_Phon2_CR, lex_Phon3_CR, nonlex_Phon1_CR, nonlex_Phon2_CR, nonlex_Phon3_CR, Resp_ER (CR for correct response and ER for error). In the GLM, the noninterest regressors were also included using ART text file per subject (each file described outliers scans from global signal and head movements from ART). Regressors were convolved with the canonical hemodynamic response function (HRF), and the default SPM autoregressive model AR(1) was applied. Functional data were filtered with a 128s highpass filter. Statistical parametric maps for each experimental factor and each participant (beta maps) were calculated at the first level and then entered in a second-level one sample t-test analysis of variance (random effects analysis or RFX). All statistical comparisons were performed with a voxelwise threshold of p< .001 and a cluster extent threshold of 5 voxels.

Results on whole brain Analysis for Phonetic monitoring load (n= 24 subjects)In order to characterize the brain areas recruited during phonetic monitoring load, the contrast ‘Contrast ‘PhonLoad_3 minus PhonLoad was studied on correct responses (see the statistical t-maps on Figure 1).

More activation in ‘Phonetic Load 3’ compared to ‘Phonetic Load 1’ condition was revealed in the left posterior cingulate gyrus and bilateral precentral gyri (voxelwise threshold= p< .001 uncorrected). However, these activations were no longer significant when a Family Wise Error (FWE) or False Discovery Rate (FDR) correction is applied to overcome the problem of multiple comparisons.

RFX. Results on Lexical monitoring load (n= 24 subjects)In order to characterize the brain areas recruited during lexical monitoring load, the contrast ‘‘lexical minus nonlexical items’ was studied on correct responses (see the statistical t-maps on Figure 1). Compared to the non lexical monitoring load, the lexical one demonstrated more activation in a left cerebellum region (VI) with a voxelwise threshold of p< .001 and with a clusterwise threshold of p< .05 FWE and FDR-correction. Other significant activations were revealed at voxelwise threshold of p< .001: they were located in cerebellar, right precuneus and bilateral inferior temporal regions.

28 février 2020

Figure 1. Contrast ‘Lexical minus non Lexical monitoring load’.Statistical t-maps are overlaid on MNI cortex slices (5 axial slices and 1 sagittal slice par line) using a voxelwise threshold of p< .001 and an extent threshold of 5 voxels. With this threshold, the significant activations were in various cerebellar regions, in a right precuneus and bilateral inferior temporal regions. When a correction for multiple comparisons was applied, only the cerebellar region located in the left cerebellum VI survived (see slice -28).

Results on Anatomical ROI Analysis

Spherical ROI with a MNI coordinates-centre and a 10mm-radius were created using the MarsBar SPM toolbox (Brett, Anton, Valbregue, & Poline, 2002).

28 février 2020

For a given ROI mask and on the basis on unsmoothed functional images, we extracted each subject’s percent signal change (see http://marsbar.sourceforge.net/faq.html#how-do-i-extract-percent-signal-change-from-my-design-using-batch) using ‘mean’ calculation across voxels. From each contrast (‘Phonetic monitoring load’, ‘Lexical monitoring load’ and ‘Error-related monitoring’), we obtained a vector of 24 percent signal changes (1 per subject) per ROI (n=11). For each ROI, we performed permutation tests (from Laurens R Krol, see https://github.com/lrkrol/permutationTest) to compare the distribution of the percent signal changes to the null hypothesis (normal distribution). Two-tailed statistical tests were conducted using 2000 permutations and 2 types of multiple comparisons (False Discovery Rate, FDR: Benjamini, & Hochberg,1995 and Bonferroni).

Results on Lexical monitoring load

Functional Connectivity AnalysisTask-related functional connectivity was assessed using the CONN toolbox (Whitfield-Gabrieli and Nieto-Castanon, 2012- version 19c). To characterize the connectivity between all pairs of ROIs, we choose a ROI-to-ROI connectivity metrics using percent signal changes as units and bivariate correlations as computation.

28 février 2020

As input images, we imported the individual functional unsmoothed images, the same GLM specification of previous analyses with interest regressors (lex_Phon1_CR, lex_Phon2_CR, lex_Phon3_CR, nonlex_Phon1_CR, nonlex_Phon2_CR, nonlex_Phon3_CR and Resp_ER) and non interest ART regressors (outliers and movement), the individual anatomical images (T1), the individual tissues (white matter, gray matter and CSF) normalized in the MNI space and the 11 anatomical ROIs (ACC_L ACC_R, Pre-SMA_L, Pre-SMA_R, RCB1_R, RCB2_R, SMC_R; SMC_L, SPT_L, pSTG_L, pSTG_R) previously defined in the MNI space.

In CONN toolbox, the denoising step combines the linear regression of potential confounding effects in the BOLD signal, and temporal band-pass filtering. We defined as confounding effects : (1) an anatomical component-based noise correction procedure (aCompCor) that included noise components from cerebral white matter and cerebrospinal areas (Behzadi et al. 2007), (2) the estimated outliers and subject-motion parameters from ART, (3) the first-order linear session and task effects (lex_Phon1_CR, lex_Phon2_CR, lex_Phon3_CR, nonlex_Phon1_CR, nonlex_Phon2_CR, nonlex_Phon3_CR and Resp_ER effects) as suggested by Whitfield-Gabrieli and Nieto-Castanon (Whitfield-Gabrieli and Nieto-Castanon, 2012). Adding task effects as confounding effects allowed limitation of false positives in correlation measures quantifying functional connectivity. For temporal band-pass filtering (linear detrending), we included a high-pass filtering of 0.01 Hz but no low-pass filtering to accept all higher frequencies.

For each participant and for each condition., BOLD timeseries (percent signal changes) were averaged across all voxels of a given ROI. Bivariate correlation coefficients (Pearson’s correlations) between each pair of ROIs were computed and Fisher-transformed, resulting in a 11×11 ROI-to-ROI correlation (RRC) matrix for each participant and for each condition (see for more details, https://web.conn-toolbox.org/fmri-methods/connectivitymeasures/roi-to-roi#h.p_D3yU5EfNjsh9). At CONN’s second-level analysis step, we added user-defined second-level model to define (as in the previous analysis) the following interest contrasts : the contrast ‘Phonetic monitoring load’ (PhonLoad_3 minus PhonLoad 1 on correct responses), the contrast ‘Lexical monitoring load’ (Lexical minus non Lexical on correct responses) and the contrast ‘Error-related monitoring’(Error Responses minus Correct Responses).

Results on Phonetic monitoring load

RRC results on the Phonetic monitoring load The RRC matrix of t-values is plotted (t-Values > 2) at the top of the figure. At the bottom of the figure, only the t-Values > 2 are shown on the sagittal view of a MNI cerebral mesh. The node size represents the degree of connection (the sum of connections from a seed region to all other ROIs). The connections between nodes represent the pairwise t-values >2.

28 février 2020

Results on Lexicality monitoring load

RRC results on Lexical monitoring load The RRC matrix of t-values is plotted (t-Values > 2) at the left of the figure. On the right of the figure, only the t-Values > 2 are shown on the sagittal view of a MNI cerebral mesh. The node size represents the degree of connection (the sum of connections from a seed region to all other ROIs). The connections between nodes represent the pairwise t-values >2.

References

Behzadi, Y., Restom, K., Liau, J., & Liu, T. T. (2007). A component based noise correction method (CompCor) for BOLD and perfusion based fMRI. Neuroimage, 37(1), 90-101Benjamini, Y., & Hochberg, Y. (1995). Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal Statistical Society, Series B, 57, 289–300.

Brett, M., Anton, J. L., Valbregue, R., & Poline, J. B. (2002). Region of interest analysis using an SPM toolbox. Presented at the 8th International Conference on Functional Mapping of the Human Brain, June 2-6, 2002, Sendai, Japan. Available on CD-Rom in NeuroImage, 16(2).Delorme, A. & Makeig, S. EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysis. J Neurosci Methods134, 9–21 (2004).Esterman E, Tamber-Rosenau BJ, Chiu YC & Yantis S (2010). Avoiding non-independence in fMRI data analysis: Leave one subject out. Neuroimage. 2010 April 1; 50(2): 572–576. doi:10.1016/j.neuroimage.2009.10.092.

Fedorenko E, Behr MK & KanwisherN (2011). Functional specificity for high-level linguistic processing in the human brain PNAS 108 (39) 16428-16433; https://doi.org/10.1073/pnas.1112937108Aglieri V, Chaminade T, Takerkart S & Belin P (2018). Functional connectivity within the voice perception network and its behavioural relevance. Neuroimage (183).356-365Whitfield-Gabrieli , Nieto-Castanon, A (2012). Conn: a functional connectivity toolbox for correlated and anticorrelated brain networks. Brain Connect., 2 125-141 https://doi.org/10.1089/brain.2012.0073

28 février 2020