Distributional factors in language processing: evidence from parametric and naturalistic functional MRI

Abstract

Language is a distinctive human ability that supports our daily life interactions. A deep understanding of the brain mechanisms behind language processing is fundamental to create physiological models and find possible alterations derived from specific pathologies. In the last decades, the general technological advancement and the development of more precise and less invasive investigation techniques have increased dramatically our knowledge of the neural correlates of language processing. However, the great variety of human languages, the possibility to communicate across multiple and different channels and the uncertainty about the actual role of some linguistic features leave several open questions with a concrete possibility of neuroscientific innovation. For instance, the features of a highly inflected language like Italian can provide interesting research questions and additional insights on how our brain processes linguistic information. The main aim of this work is to explore in greater details the influence of the linguistic distributional factors on the neural correlates of both language production and comprehension by exploiting the richness of the Italian language. Therefore, three functional magnetic resonance imaging (fMRI) experiments were performed by using both classical parametric and novel naturalistic frameworks. Two fast event-related fMRI experiments investigated the influence of the language distributional factors on the neural correlates of the inflectional process, whereas a third experiment was dedicated to providing additional insights on the linguistic prediction mechanism during natural language comprehension by modelling the neural response with two statistical language models. The first experiment addresses the influence of the inflectional classes (i.e. the conjugations) and their distributional features (e.g. the size, the productivity, and the ortho-phonological consistency) on the generation of the past participle of the Italian verbs by analyzing both neural and behavioral data. The study reports significant effects of the conjugations on the cognitive operations and, for the first time, differential cortical activations in the left middle frontal gyrus, left the supplementary motor area and left anterior cingulate cortex for verbs from different conjugations supporting the hypothesis that the neural correlates of the verb inflection are influenced by specific properties of the inflectional classes. The second experiment explores the influence of the noun inflectional classes and their properties (e.g. the consistency, number and cumulative frequency of members) on the nominal inflection by analyzing the neural data of a group of participants involved in overt inflection task from the singular to plural and vice versa. The study reports an extensive bilateral cortical network involving the cingulate cortex, frontal and temporal areas, and the cerebellum, revealing that the neural activations are modulated by specific distributional features of the noun inflectional paradigm. The third experiment investigates the neural correlates of the linguistic prediction underlying the natural language processing during narrative listening. The interest, therefore, shifts from word (verb or noun) production to structured text understanding. This is done by fitting the fMRI data with models that encode the probabilistic features of the language via the estimation of the so-called surprisal, that is a measure that quantifies the unexpectedness of a word given the previous ones. Two stochastic language models were estimated on a large written Italian corpus to obtain two versions of surprisal: a lexical-only version, based only on the lexical information of the chosen stimulus and a novel semantics-weighted model that integrates both lexical and semantic features. Our study reports better prediction accuracy and better fitting of the fMRI data for the semantics-weighted model. The two models produced both overlapping and distinct activations: while lexical-only surprisal activated secondary auditory areas in the superior temporal gyri and the cerebellum, semantics-weighted surprisal additionally activated the left inferior frontal gyrus. The results support the usefulness of the surprisal models to describe the linguistic prediction and suggest that the proper inclusion of semantics information in the surprisal model may increase its the sensitivity in higher-order language-related areas, with possible implications for future naturalistic fMRI studies of language under normal and (clinically or pharmacologically) modified conditions. Besides the investigation of the influence of the language distributional factors on the neural correlates of both language production and comprehension, an additional aim of this work is the proposal of an innovative procedure in the broader field of the fMRI-neurofeedback (fMRI-NF). In general, the fMRI-NF has been successfully applied in several cognitive domains and it is a procedure based on the possibility to self-modulate the neural signal of a brain region to explore and induce behavioral changes. The proposed method integrates the representational similarity analysis (RSA) and the fMRI-NF framework to train the subjects in modulating their mental state rather than simply regulating the neural signal of a region. The method has been tested in a pilot experiment at 7 Tesla where the subject was asked to imagine concrete objects. The results show that the presented approach is feasible suggesting further investigations and future applications in several domains, including representational and distributional aspects of language processing. [edited by Author]