Concepts are built from percepts: Neural mechanisms supporting sub-lexical to lexico-semantic processing

(cc) 2024 Abraham Sánchez Sánchez (cc by-nc-nd 4.0) Concepts are built from percepts Neural mechanisms supporting sub-lexical to lexico-semantic processing Abraham Sánchez Supervised by Manuel Carreiras & Pedro M. Paz-Alonso 2024 Concepts are built from percepts Neural mechanisms supporting sub-lexical to lexico-semantic processing Abraham Sánchez Supervised by Manuel Carreiras & Pedro M. Paz-Alonso 2024 I ACKNOWLEDGEMENTS Most questions and propositions of the philosophers result from the fact that we do not understand the logic of our language. Ludwig Wittgenstein First of all, many thanks to my supervisors, Manolo and Kepa, for supporting me along the way, for providing valuable feedback and counsel, for teaching me all I needed to know for my PhD, and for giving me all the resources to have a strong profile in the future. I have felt very lucky for some years now. I was lucky to be given the chance to do my PhD at the BCBL, and to learn from incredible professionals. I was also lucky to be surrounded by amazing people that grew to be like a second family. I have had their professional and personal support all these four years, no matter what the problem was, or whether they knew how to solve it or not. And I feel lucky, above all things, to have enjoyed as I have during the path, both developing the works of this thesis, but also in living everyday in such a fun, constructive and rich environment. I remember my first day at the BCBL very well. I did not join in easy times, as I started right in the middle of the COVID-19 pandemic. I remember Chiara coming to my desk (with her mask on, of course) to tell me that the few of us that were working from the centre would go to have lunch at 1pm. I remember slowly meeting the other predocs. Christoforos’ puns, the loud Vicente, and my unfortunate comments about Valladolid to Laura. I remember going on those fun workouts with Giorgio, Piermatteo and Jose, and also to meet at Giorgio’s with all of them and Irene, Inés, Alberto, Asier… to laugh and have as much fun as we could until the curfew pushed us back to our homes. And the “family” coffees at Bizi, of course, including the rich (and sometimes bizarre) conversations with Patxi. So thank you, for all these moments, to all of you, Chiara, Alberto, Jose, Piermatteo, Giorgio, Vicente, Laura Fernández, Christoforos, Inés, Irene, Hana, Patxi, Maca, Edith, Marta, Laura de Frutos, Coco, Dani López, Asier, David Carcedo, Lucía, Eneko, Giulia, Marco, Ihintza… and the rest of the BCBL community I had the luck to share a coffee with. I carry these moments with me. Deep thanks to my housemates, Hana and (later) Tomas, for being like sister and brother during the last years, and for making it possible to live as if at home. And special thanks to Tomas, also for always listening to my consults and giving the greatest advice when I needed it. Most special thanks to María. Simply thank you with all my heart. You have been my pillar, and the deepest well of love, fun and inspiration. Te quiero con cordura. II My deepest gratitude to my parents, my brothers and sister. Without them, this thesis would not have been possible. You gave me the little you had, and always believed in my ability. And finally, I would like to make what, in my humble opinion, is a necessary conceptual exercise. I have put my everything into scientific research. But not with passion, or devotion. These heavenly words pertain to the worlds of religion and belief. I have worked with dedication and professionalism. As all professionals at the BCBL do. The words we use matter. Lets try to understand the logic of our language. CONTENTS ACKNOWLEDGEMENTS ...................................................................................................... I LIST OF ABBREVIATIONS ................................................................................................. VI ABSTRACT .......................................................................................................................... 1 RESUMEN EN CASTELLANO ............................................................................................. 2 GENERAL INTRODUCTION ................................................................................................ 5 Background and Motivation ....................................................................................... 5 Objectives and Thesis Structure ................................................................................ 7 CHAPTER I. INTEGRATION OF LINGUISTIC PERCEPTUAL INFORMATION ................... 9 1.1. LANGUAGE VISUAL PERCEPTION ........................................................................ 9 1.1.1. Visual Pathway ............................................................................................... 9 1.1.2. Beyond V1: Encoding of Complex Visual Features ........................................ 13 1.1.3. Ventral Occipitotemporal Cortex and the Putative Visual Word Form Area .... 16 1.2. INTEGRATION OF AUDITORY LANGUAGE PROCESSING ................................. 19 1.3. LANGUAGE PROCESSING NETWORKS ............................................................. 23 1.3.1. Dorsal Route.................................................................................................. 27 1.3.2. Ventral Route ................................................................................................. 27 CHAPTER II. THE NEUROBIOLOGY OF SEMANTIC REPRESENTATIONS ................... 29 2.1. MODELS FOCUSED ON THE PROCESS ............................................................. 29 2.1.1. Memory, Unification and Control (Hagoort, 2005, 2013) ................................ 29 2.1.2. The Cognitive Control of Semantic Memory (Badre & Wagner, 2007) ........... 32 2.2. MODELS FOR SEMANTIC REPRESENTATIONS ................................................. 34 2.2.1. Distributed versus Distributed-Plus-Hub Perspective (Patterson et al, 2007) . 35 2.2.2. Embodied Abstraction (Binder et al, 2009; Binder & Desai 2011) .................. 37 2.2.3. Controlled Semantic Cognition (CSC) (Ralph et al., 2017) ............................. 39 CHAPTER III. THE STUDY OF NEURAL LEXICAL REPRESENTATIONS ....................... 42 3.1. Psycholinguistic Properties as a Window to Lexical Representations ..................... 43 3.1.1. Word Concreteness and Imageability ............................................................ 43 3.1.2. Word Frequency and Familiarity .................................................................... 45 3.1.3. Phonological Properties ................................................................................. 48 3.2. Novel Measures: Naturalistic Language Processing and Word Vectors .................. 49 3.3. Novel fMRI Approaches: Representational Similarity Analysis (RSA) ..................... 51 CHAPTER IV. THE ROLE OF READING DEMANDS AND WORD FREQUENCY IN THE ACCESS TO LEXICAL UNITS ........................................................................................... 55 4.1. RATIONALE ........................................................................................................... 55 4.2. METHODS ............................................................................................................. 57 4.2.1. Participants .................................................................................................... 57 4.2.2 Materials and Procedure ................................................................................. 57 4.2.3. fMRI Data Acquisition .................................................................................... 58 4.2.4. fMRI Data Analyses ....................................................................................... 58 4.2.5. Functional Connectivity Analyses .................................................................. 60 4.3. RESULTS ............................................................................................................... 60 4.3.1. Behavioural Performance .............................................................................. 60 4.3.2. Whole-brain results ........................................................................................ 61 4.3.3. ROI analysis .................................................................................................. 62 Ventral Network ................................................................................................. 63 Dorsal Network .................................................................................................. 64 4.3.4. Functional Connectivity Analysis .................................................................... 65 4.4. DISCUSSION ......................................................................................................... 66 4.4.1. Word frequency ............................................................................................. 67 4.4.2. Reading demands ......................................................................................... 67 4.4.3. The WFE is modulated by reading demands in anterior IFG .......................... 68 4.4.4. Task-related functional connectivity ............................................................... 69 4.4.5. Limitations ..................................................................................................... 69 4.5. CONCLUSIONS ..................................................................................................... 70 CHAPTER V. NEURAL REPRESENTATIONS OF LEXICO-SEMANTIC KNOWLEDGE: SIMILARITY OF SUB-LEXICAL AND LEXICAL MODELS WITH MULTIVARIATE BRAIN RESPONSES ..................................................................................................................... 71 5.1. RATIONALE ........................................................................................................... 71 5.2. METHODS ............................................................................................................. 72 5.2.1. Participants .................................................................................................... 72 5.2.2. Stimuli and Materials ..................................................................................... 72 5.2.3. Procedure ...................................................................................................... 74 5.2.4. MRI Data Acquisition and Preprocessing ....................................................... 75 5.2.5. Univariate analyses ....................................................................................... 76 5.2.6. RSA searchlight ............................................................................................. 77 5.2.7. ROI-based RSA ............................................................................................. 79 5.3. RESULTS ............................................................................................................... 79 5.3.1. Behavioural Performance .............................................................................. 79 5.3.2. Univariate results ........................................................................................... 81 5.3.3. RSA results ................................................................................................... 83 RSA Searchlight Results.................................................................................... 83 ROI-based RSA Results .......................................................................................... 85 5.4. DISCUSSION ......................................................................................................... 88 5.4.1. Associations Between Word Properties ......................................................... 88 5.4.2. Anterior-to-Posterior dissociation in the left IFG ............................................. 89 5.4.3. Involvement of the left anterior vOTC in lexico-semantic processing ............ 90 5.4.4. Linguistic Properties vs Word Vectors in Semantic Hubs ............................... 91 5.5. CONCLUSIONS........................................................................................................... 92 CHAPTER VI. BEHAVIOURAL CORRELATES OF LEXICO-SEMANTIC REPRESENTATIONS ......................................................................................................... 94 6.1. RATIONALE ........................................................................................................... 94 6.2. METHODS ............................................................................................................. 96 6.2.1. Participants .................................................................................................... 96 6.2.2. Tasks and Materials ....................................................................................... 97 6.2.3. Task Performance Comparison ..................................................................... 98 6.2.4. Drift-Diffusion Models and Analyses .............................................................. 99 6.2.5. Association with Brain Representations ....................................................... 100 6.3. RESULTS ............................................................................................................. 101 6.3.1. Performance in Both Tasks .......................................................................... 101 6.3.2. Drift-Diffusion Results .................................................................................. 103 6.3.3. Correlations between DDMs and Brain Representations ............................. 105 6.4. DISCUSSION ....................................................................................................... 107 6.4.1. Psycholinguistic Properties and Decision-Making ........................................ 107 6.4.2. Association between Drift Rate and Brain Representations ......................... 108 6.5. CONCLUSIONS ................................................................................................... 111 GENERAL DISCUSSION ................................................................................................. 112 Functional Dissociations in the IFG and vOTC and their Dynamic Nature ............. 112 Semantic Representations in Semantic Hubs: Psycholinguistics and Natural Language Processing (NLP) Combined ................................................................. 114 Limitations and Future Directions .......................................................................... 115 Conclusions ........................................................................................................... 116 REFERENCES ................................................................................................................. 118 VI LIST OF ABBREVIATIONS ACC: anterior cingulate cortex AF: arcuate fasciculus AG: angular gyrus AIC: Akaike information criterion ATL: anterior temporal lobe BA: Brodmann area BF: Bayes Factor CSC: controlled semantic cognition model DCT: dual-coding theory DDM: drift-diffusion model dlPFC: dorsolateral prefrontal cortex dmPFC: dorsomedial prefrontal cortex DRC: Dual-route cascaded model EEG: electroencephalography FFG: fusiform gyrus FDR: false discovery rate fMRI: functional magnetic resonance imaging FWE: family-wise error FWHM: full width at half-maximum GLM: general linear model HRF: hemodynamic response function IC: inferior colliculus IFG: inferior frontal gyrus IFOF: inferior fronto-occipital fasciculus ILF: inferior longitudinal fasciculus IOG: inferior occipital gyrus IPL: inferior parietal lobule IPS: intraparietal sulcus ITG: inferior temporal gyrus ITI: inter-trial interval LGN: lateral geniculate nucleus LMM: linear mixed model LO/hOclp4: lateral occipital cortex LRT: Likelihood Ratio test M: magnocellular 6 OF 150 More specifically, I chose this fragment because it touches some of the central aspects and ideas that are the object of study of the current thesis. This work is mainly concerned with the cognitive and neural systems that allow us to form and sustain concepts (the chunks of information about the world around us), and to access them during reading. It pertains to the whole continuum of mechanisms that make it possible to make sense of the abstract symbols and sounds that constitute language, to ultimately form meaningful ideas and connect them to our environment. As illustrated in the Frankenstein’s fragment, such concepts are encapsulated in lexical units, or words. As concepts reflect entities in our world that vary in a wide range of features, so do words, as linguistic references to such concepts. Frankenstein’s creature first acquires those words that have a clear connection to the perceptual world immediately around him, like the milk the cottagers drink, the bread they eat, or the wood they use to make a fire. Other words are still puzzling to him, since he cannot perceive the entity they refer to. One can hardly see goodness or unhappiness, if not linked to very specific and, to a great extent, tacit indicators of such emotional states. It is through repetition, the iterated connection between the perceptual world and the abstract references to it, that Frankenstein’s creature can at least partially grasp such figurative ideas. This highlights a critical point of the current thesis: the importance of the connection between the perceptual input and the abstractions that can be created from it. While the observation of the perceptual world around us allows humans to relatively easily and naturally acquire language, reading requires years of active instruction, and even after that, many struggle to consolidate such a complicated ability (Stein, 2022; Yeatman & White, 2021). Language acquisition entails the link of its sounds to the entities in our environment, while reading requires to link abstract symbols to discrete sounds, to whole-word sounds, and then to link those sounds to meaning. There are many possible pieces that can break in this chain of complex processes. Therefore, it is not surprising that the cognitive and brain mechanisms of language and reading have received a great deal of attention both in the neurobiology of language in general, and in the neurobiology of reading and the associated reading difficulties. Readers must “deal” with two “perceptual worlds”, our natural environment, formed by the objects and perceptual entities around us, and the abstract world of language, formed by its visual symbols and auditory components. Interestingly, the neural mechanisms that act as an interface between these two worlds are still poorly understood. How does our brain transition from the linguistic perceptual world to the conceptual world of ideas and references to our environment? What are the neural underpinnings that sustain the formation of abstractions and semantic knowledge associated with words and concepts? These are questions that are still wide open. 7 OF 150 Objectives and Thesis Structure The general objective of the current thesis is to offer a comprehensive perspective of the neural mechanisms that support the access and use of conceptual information during reading. On the one hand, this encompassess those processes related to deciphering words even before they are processed as a whole. This, in turn, includes the segmentation of the basic units that constitute a word (syllables), and linking them to their sound basic units (phonemes). This is what I will refer to when I use the expression sub-lexical processing. These sublexical processes are mostly (although not exclusively) developed from brain areas at the bottom of the hierarchy, in perceptual regions, and hence are mostly covered by what are termed bottom-up mechanisms. On the other hand, I will cover those processes related to the perception of words as a whole, and will often refer to these as lexical units (as words are the entries of our lexicon). And going a step further, I will often allude to lexico-semantic processes to refer to the cognitive mechanisms that allow for the connection between the whole-word and the concept it represents (i.e., its meaning). Previous knowledge and expectations involve connections from brain areas at the top of the hierarchy (or high level areas) to perceptual and association regions, and hence they are often referred to as topdown mechanisms. Although this thesis is mostly concerned with the second group of processes, they cannot be understood without taking sublexical processing into account. The specific objectives of the thesis are: 1) to investigate the influence of previous knowledge and expectations (here referred to as top-down influences) in how words are accessed at the neural level during reading; 2) to explore the influence of word features that reflect critical psycholinguistic properties on the neural representations of conceptual knowledge; and 3) to assess different models that reflect a variety of lexical properties, from sublexical to lexico-semantic, and analyse their representation at the neural level. The thesis will contain six chapters. The first three will be theoretical chapters that aim to respond to the general objective, by offering a critical review of the available knowledge on lexical representations and the neural mechanisms that support them. This will go from how the brain deciphers visual linguistic input, to how it represents complex abstract information, and how we can explore this through the use of functional magnetic resonance imaging (fMRI). Chapter 1 will cover the neural underpinnings of the visual integration of linguistic information. This will include how visual inputs are decoded by the brain, and how the brain makes it possible to ultimately recognise words. At the end of the chapter, I will cover how these neural processes are integrated into larger brain language networks. Chapter 2 will be dedicated to review the most influential models that tried to explain how semantic information is represented in the brain, and what are the neural mechanisms that support the access and manipulation of this information. And Chapter 3 will tackle the methods we have to study neural 8 OF 150 semantic representations. It will almost exclusively refer to fMRI investigations, and the manipulations that can be made to explore lexico-semantic processing at the neural level. The last three chapters are empirical, and try to address the specific objectives of the thesis, while also contributing to the general framework. In Chapter 4, I present a fMRI study in which I explored the influence of top-down reading demand (perceptual demand versus semantic demand) and word frequency as critical variables that affect the neural responses to words as they are read. Chapter 5 describes an investigation that used novel fMRI analytical approaches to explore neural representations associated with sublexical to lexico-semantic processing. And finally, Chapter 6 will try to complement the study described in Chapter 5 by critically analysing the behavioural correlates associated with the neural representations presented in the previous chapter. 9 OF 150 CHAPTER I. INTEGRATION OF LINGUISTIC PERCEPTUAL INFORMATION 1.1. LANGUAGE VISUAL PERCEPTION 1.1.1. Visual Pathway From a strict neurophysiological point of view, the very first step for the human brain to decipher visual linguistic information is to capture the physical stimuli that constitute the linguistic input (Stein, 2022). These physical stimuli are then transformed into electrical impulses that our brain interprets in order to form the perceptual experience that allows reading. This first-level step is possible thanks to the retina and the visual pathway (Kiley & Usrey, 2016; Stein, 2022). The retina contains photoreceptors that capture light photons, and relay this stimulation to the retinal ganglion cells in the form of electrochemical signals (Sartucci & Porciatti, 2024) that are carried by the optic nerve (Celesia, 2005). After partially crossing its fibres to the contralateral part at the optic chiasm, the optic nerve forms the optic tract. Some of its fibres reach the superior colliculi to control ocular reflexes, and the pulvinar nucleus of the thalamus, where minimal visual processing takes place (Celesia, 2005; Kahle et al., 2022). The majority of the optic tract fibres, however, run laterally, conveying the low-level visual information to the lateral geniculate nucleus (LGN) in the thalamus. From there, the fibres constitute the optic radiation, which reaches the striate cortex in the primary visual cortex, also called V1 area (Celesia, 2005; Kiley & Usrey, 2016), as well as the secondary visual cortex or V2 to a much lesser extent (Alvarez et al., 2015; Arrigo et al., 2016) V1 follows a retinotopic organisation, where each quadrant of the contralateral visual hemifield is represented in a specific location of the striate cortex. This organisation starts in the retina and is kept along the visual pathway, through the optic nerve, tract and radiation (Kahle et al., 2022). In this retinotopic organisation, the fovea, a retinal area with the highest visual sharpness (given to its tissue composition), is overrepresented in V1 as compared to the periphery of the visual field (Celesia, 2005). Furthermore, the neighbouring cells in V1 represent adjacent visual receptive fields to one another (Kiley & Usrey, 2016). Figure 1.1 shows the arrangement of optic fibres, the organisation of the visual quadrants along the visual pathway and their representation in V1. Figure 1.1. A) Arrangement of the optic fibres; B) Position of retinal quadrants along the visual 10 OF 150 pathway, including V1 (bottom right). Taken from Kahle et al (2022). The main input to V1 comes from the LGN, which is formed by six layers of simple cells that already process some basic visual features. The two predominant types of cells in the LGN are the magnocellular (M) and parvocellular (P) cells (Celesia, 2005). While P cells respond to low temporal frequency and high spatial frequency, and have thinner axons and smaller receptive fields, M cells respond to high temporal frequency and low spatial frequency, and have larger receptive fields and thicker axons. This is why P cells are sensitive to chromatic information and visual acuity, whereas M cells are sensitive to fast motion information (Celesia, 2005; Kiley & Usrey, 2016). The receptive fields of all LGN cells are organised in what is called a centre-surround (or ON and OFF) structure that consists in a centre that responds to increment or decrement of light, and an annular portion that surrounds it, which responds in the opposite way to the central portion (Celesia, 2005; Kiley & Usrey, 2016). The V1 area is formed by simple cells and complex cells. Simple cells in V1 also have On and Off receptive fields, but are elongated and capture orientation of visual stimuli, with a preference for a specific orientation (Celesia, 2005; Kiley & Usrey, 2016). Simple cells in V1 receive their input from multiple LGN simple cells. Complex cells in V1 receive their input from multiple simple cells that share the same orientation preference, but differ in their spatial arrangement (Kiley & Usrey, 2016). Figure 1.2 represents the structure of LGN and V1 simple cells, and the connections between them. 11 OF 150 Figure 1.2. Representation of LGN and V1 simple cells. The figure also represents the feedforward model in which V1 simple cells are fed by the input from multiple LGN cells. Taken from Kiley & Usrery (2016). Horizontally, the V1 area is organised in six main layers. The layer 4 contains the majority of the simple cells. M and P cells terminate in different sublayers of layer 4. Layers 5 and 6 send feedback to the LGN and other subcortical regions. The layer 4 projects its fibres to layers 2 and 3, which send their projections to extrastriate areas. Vertically, the V1 area follows a columnar organisation, in which the complex cells are arranged in bundles that contain the orientation information from one of the respective hemifields (left or right), which is why they are also called ocular dominance columns (Kahle et al., 2022; Kiley & Usrey, 2016). The combinations of two adjacent columns from the ipsilateral and contralateral eyes are called hypercolumns, and they represent the receptive field of the corresponding eye (Kahle et al., 2022). This results in two inputs that represent the same information in two different ways (left columns and right columns). This phenomenon is called binocular disparity, and is critical for depth perception (Kiley & Usrey, 2016). Between the ocular dominance columns lay the blobs, columns that are not sensitive to orientation, but that have high cytochrome oxidase content, and are therefore sensitive to colour (Kahle et al., 2022). Blobs are located in layer 2 of V1, and receive their input directly from LGN cells. Figure 1.3 shows the horizontal layers and vertical columnar organisation of V1. 12 OF 150 Figure 3. A) Horizontal organisation of V1 into 6 different layers. B) Vertical organisation of V1 into ocular dominance columns. 1-6. Horizontal layers; 7. ; 8. ; 9. A hypercolumn; 10. Blobs. Taken from Kahle et al. (2022). Thanks to the properties of the visual pathway described thus far, the primary visual cortex and its inputs allow to encode basic visual features, such as orientation, ocular dominance, edges and shape, and stimulus size (Kiley & Usrey, 2016). It is in V1 where the perception of all these features takes place (Celesia, 2005), and only minimal processing of specific features of the visual stimuli occur earlier in the pathway (Celesia, 2005; Kiley & Usrey, 2016). However, more complex characteristics of the visual stimuli are processed later in the visual pathway, in extrastriate areas (Kiley & Usrey, 2016). 13 OF 150 1.1.2. Beyond V1: Encoding of Complex Visual Features So far, we have covered how the human visual pathway, up to the primary visual cortex, converts physical stimuli into electrical impulses that constitute the basis for visual perception of low-level features. However, we have not answered the question of how visual information is categorised and recognised by the human brain, a critical step in the visual integration of linguistic input and, therefore, in reading. Although the primary visual cortex is necessary for the recognition of visual input, this task is actually carried out by a set of areas that reside outside V1, especially in the ventral occipitotemporal cortex (vOTC) (Grill-Spector & Weiner, 2014; Weiner et al., 2014). The striate cortex is connected ventrally to the vOTC, but before reaching the vOTC, V1 sends parallel projections to areas V2-V5, also known as the secondary visual cortex. These areas, as V1, are retinotopically organised, and are especially relevant in processing additional visual features like angle, motion or texture (Furlan & Smith, 2016; Grill-Spector & Weiner, 2014; Okazawa et al., 2016; Zhong & Wang, 2021). Areas V2-V5 process in parallel several different features, while exhibiting each of them a certain preference for a concrete visual characteristic. For instance, it has been shown that V2 encodes contour (edges and corner detection) of visual stimuli (R. Chen et al., 2017; Roe & Ts’o, 2015; Zhong & Wang, 2021), while being necessary for visual awareness (SalminenVaparanta et al., 2012). On the other hand, V3 and V5/MT are especially relevant for motion processing (Furlan & Smith, 2016; Jeschke et al., 2023; Silvanto et al., 2005), although area V2 also contributes to early motion encoding (Furlan & Smith, 2016). Finally, area V4 is often recognised as the main secondary visual region for colour perception (Bouvier & Engel, 2006; Brouwer & Heeger, 2009; Desimone et al., 1985; Pasupathy et al., 2020), although colour encoding is not limited to V4 (Bouvier & Engel, 2006; Brouwer & Heeger, 2009). Furthermore, V4, along with V2, also encodes other visual features like texture (Okazawa et al., 2016) or size (Tanaka & Fujita, 2015), which is why V4 is believed to participate in early visual object recognition (Pasupathy et al., 2020). The categorisation and recognition of visual objects (which includes linguistic units such as words), however, is possible thanks to the function of the vOTC. It has long been reported that lesions in the vOTC in primates lead to difficulties in discriminating visual objects, without impairing the ability to perceive those objects’ spatial configuration (Pohl, 1973). These findings led to the accepted notion that the visual pathway is divided into two main streams (Goodale & Milner, 1992; Mishkin et al., 1983): a dorsal stream that runs predominantly through the occipitoparietal cortex, and a ventral stream that runs through the occipitotemporal cortex. The dorsal stream is also called the where or how pathway, since it supports spatial and motor visual perception. The ventral stream allows the recognition of object identity by establishing a bridge between visual input and information stored in memory (see Figure 1.4), 14 OF 150 and hence has been named the what pathway (C. B. Martin & Barense, 2023). It is important to note that these two streams are not absolutely independent, and that they dynamically interplay with one another. Figure 1.4. Representation of the ventral visual pathway and its function in the primate brain. Taken from Martin & Barense (2023). Although both visual streams are necessary for language processing, the ventral stream is pivotal for reading, as it is principally responsible for word recognition and the integration of visual and language processing (Yeatman & White, 2021). More specifically, it is the vOTC where visual features are integrated to ultimately construct the perception of visual objects (Grill-Spector & Weiner, 2014). The vOTC displays a high level of cytoarchitectonic and functional specialisation, and hence its subregions selectively represent different object categories (Downing et al., 2006; Grill-Spector & Weiner, 2014; Weiner et al., 2014). For instance, some few categories that are considered “ecologically-valid” (i.e., they have been relevant for the evolution of the human species) are commonly clustered together in vOTC subregions: faces are clustered around the midand posterior fusiform gyrus (FFG), and the inferior occipital gyrus (IOG) (Grill-Spector et al., 2017); places usually cluster around the parahippocampal place area (PPA), in the medial portion of the vOTC (Epstein & Baker, 2019; Nasr et al., 2011); body parts tend to be represented in occipitotemporal sulcus (OTS) (Downing et al., 2006; Grill-Spector & Weiner, 2014; Peelen & Downing, 2005; Ritchie et al., 15 OF 150 2021); objects representations are usually also clustered around the OTS, partially overlapping with body clusters, and extending posteriorly (Grill-Spector & Weiner, 2014; Ritchie et al., 2021); and words and symbols are commonly represented in the OTS, partially overlapping with object and body clusters, but extending ventrally (Grill-Spector & Weiner, 2014), within the so-called visual word form area (VWFA) (Dehaene et al., 2002). Figure 1.5 illustrates the functional specialisation of the vOTC. Figure 1.5. Functional clusters of vOTC subregions and the categories that are represented within them. 1) inferior occipital gyrus (IOG); 2) posterior fusiform face-selective region (pFus), or face fusiform area-1 (FFA-1); 3); mid-fusiform face-selective region (mFus) or FFA-2; 4) occipitotemporal sulcus (OTS); 5) parahippocampal place area (PPA); 6) visual word form area (VWFA). MFS: midfusiform sulcus. Taken from Grill-Spector & Weiner (2014). The bottom-line message is that the vOTC is a fundamental brain region for the recognition of visual input, and that it shows high specificity as to what type of stimuli are preferentially represented in its subdivisions. The function of the vOTC allows us to distinguish one kind of stimulus from the others, so that we can respond according to our needs in a particular situation. Taken to reading, this entails being able to recognise visual linguistic input (the bottom-up process) such as words. But in order to be able to recognise words as whole, 22 OF 150 2020). The influence of auditory processing of words over the vOTC is best illustrated by a number of studies that demonstrated automatic top-down activation of the vOTC during speech processing (Cao et al., 2010; Conant et al., 2020; Dziȩgiel-Fivet et al., 2023; Ludersdorfer et al., 2016; Pattamadilok et al., 2019; Planton et al., 2019; Yoncheva et al., 2010; for a review, see Dȩbska et al., 2023). However, there is no clear consensus about the nature of this top-down influence. The proponents of the VWFA link the vOTC response to auditory stimuli to co-activations of the orthographic whole-word code (Dehaene & Cohen, 2011). Another possibility is that vOTC activation reflects phonological processing of smaller orthographic units (e.g., phonemes) (Pattamadilok et al., 2019). This view is supported by studies showing an effect of phonological factors, like syllable structure, over vOTC activation (Conant et al., 2020). A third explanation considers the coexistence of phonological, orthographic and semantic clusters of neurons within the vOTC (Dȩbska et al., 2023). This is a promising account, that aligns with evidence suggesting a division of labour within the vOTC (Lerma-Usabiaga et al., 2018; Sebastian et al., 2014; White et al., 2019; Zemmoura et al., 2015), and that is being increasingly explored with multivariate approaches that can delve into these fine distinctions (Fischer-Baum et al., 2017; X. Wang et al., 2018). We will come back to this matter in Chapter 3. We cannot conclude the section without clarifying a fact that has been pointed out above: that other areas out of the secondary auditory cortex also display tight connectivity with vOTC regions (Lerma-Usabiaga et al., 2018). In fact, the functional coupling with the vOTC changes as a function of literacy, something that goes in line with the interactive account of the vOTC by Price & Devlin (2011). Interestingly, the connectivity between the STG and the vOTC decreases with increasing literacy, while connectivity with parietal areas and the frontoparietal network increases with increasing literacy (López-Barroso et al., 2020; Moulton et al., 2019). This reminds us that language processing, and specifically, reading, requires coordination between several different cognitive processes, and thus, connectivity between several different brain areas. The parietal cortex participates in the mapping of phonological sensory representations of language with its motor articulation, and it is therefore expected to increase its participation with increasing literacy (Hickok & Poeppel, 2007). Aspects like this illustrate how important it is to understand the brain networks involved in language processing, and not just the regional activation associated with it. This is why the next section will be dedicated to language processing networks. 1.3. LANGUAGE PROCESSING NETWORKS Before and during the early 2000s, computational models coming from neuropsychology and linguistics were a popular method for attempting to explain how different 23 OF 150 cognitive processes interact to produce a certain behaviour, especially in the context of brain damage (Protopapas et al., 2016). In short, these models are composed of input units (the perceptual correlates of any experience), hidden units (the cognitive processes that transform the perceptual input) and output units (the behavioural production resulting from the interaction between the former two), components that interact with each other in mathematically expressed relations (Coltheart et al., 2001). Taken to reading, several different computational models tried to mathematically express how the different sub-lexical and lexical components interact to produce visual word recognition and reading aloud. There are multiple examples of computational models of reading (Coltheart et al., 2001; Grainger & Jacobs, 1996; McClelland & Rumelhart, 1981; Plaut et al., 1996), but, far from focusing on these models, I will briefly describe one of them that has received a great deal of attention, and that served as the basis for one of the most popular frameworks for the neurobiology of reading: the Dual-route cascaded model (DRC) (Coltheart et al., 2001). For clarity, I present here, in Figure 1.7, a graphical representation of the DRC model. 24 OF 150 Figure 1.7. A graphical representation of the DRC model, taken from Coltheart et al. (2001). As the name indicates, the DRC model understands that there are two main streams of processing printed words: 1) a grapheme-phoneme correspondence route in charge of converting letters into phonemes (or a phonological route); and 2) a lexical route that processes words as a whole, both orthographically (i.e., the phonological rules that constitute the word entry in the lexicon) and semantically (i.e., the meaning of that word entry in the lexicon). Importantly, Coltheart et al. conceived that these components interact in a cascaded fashion. This is, instead of being activated sequentially after reaching a given threshold, once one process (e.g., input letter units) starts, it can send and receive excitatory or inhibitory connections to/from the ones connected to it. These features made the DRC model an attractive option to be tested with neuroimaging tools in order to investigate the processing routes for language. In fact, the DRC, and other similar dual-route models, can be considered the computational starting point of one of the most popular frameworks in the neurobiology of language: the division of labour 25 OF 150 between the dorsal and ventral routes (Jobard et al., 2003). This conception was based on the traditional idea that sensory processing of language requires the interface between 1) a motor-articulatory system that allows for the integration of auditory input into the motor articulation required to ultimately produce speech; and 2) a conceptual system that allows us to comprehend language (Hickok & Poeppel, 2004). Drawing on evidence in the neurobiology of vision, Hickok and Poeppel (2000, 2004) proposed a similar division of labour for language processing: a dorsal route, critical for auditory-motor integration of language; and a ventral route that is involved in mapping the language units onto meaning. Or put in other words, while the dorsal route takes care of phonological processing, the ventral route is mainly in charge of lexico-semantic processing (Oliver et al., 2017; Sandak et al., 2004). The dorsal route is mainly composed of the middle and posterior STG, the inferior parietal lobule (IPL), premotor cortex (PMC), and the posterior part of the inferior frontal gyrus (IFG), which includes the pars opercularis (Brodmann Area (BA) 44); the ventral route is predominantly formed by the inferior temporal gyrus (ITG), including the vOTC, the anterior temporal lobe (ATL) and the anterior part of the IFG, which includes the pars triangularis (BA 45) and pars orbitalis (BA 47) (Friederici, 2011; Hickok & Poeppel, 2007; Oliver et al., 2017; Sandak et al., 2004; Saur et al., 2008). It is usually assumed that both routes display a certain degree of left lateralisation (Friederici, 2011; Friederici & Gierhan, 2013). However, it has been proposed that while the dorsal route is usually more clearly left-lateralised, the ventral route is more bilaterally distributed (Hickok & Poeppel, 2007). In any case, both hemispheres show a similar structural connectivity pattern between the above-mentioned areas, although each of them can serve distinct cognitive functions during language processing (Dick et al., 2014). Importantly, we currently know that these routes are not isolated from each other, but they can interact with one another, and form redundant connections between the areas that constitute them (LópezBarroso & De Diego-Balaguer, 2017) In general, this view serves us as a useful framework to understand how reading recruits several brain mechanisms in unison, and thus rely on large-scale brain networks. Nevertheless, this traditional view can perhaps be minimally tinged to match the most recent evidence, part of which we already introduced in previous sections. The first nuance that recent evidence can add is the knowledge that other large-scale brain networks can contribute to language processing (Hagoort, 2019; Sierpowska et al., 2022). Language usage recruits both its basic units (e.g. phonemes, words, sentences), and a number of operations that can be made over these units (e.g., recognising, retrieving, or linking a word with stored information and its context) (Deniz et al., 2023; Hagoort, 2019). The consequence of this is that multiple processes can be involved even when processing the single units of language (Roger, Rodrigues De Almeida, et al., 2022). In terms of structural connectivity, it is believed that this feature facilitates stronger and more widespread white matter tracts between the 26 OF 150 human temporal cortex and other areas that act as association hubs, like the parietal or frontal cortices (Braunsdorf et al., 2021; Sierpowska et al., 2022). This connects to the second nuance to the dorsal/ventral framework. Growing evidence from diffusion weighted imaging (DWI) studies pointed to several different tracts contributing to either the dorsal or the ventral route for language (Catani et al., 2005; Friederici, 2009; Saur et al., 2008). This group of evidence led to most language scientist to talk about dorsal “pathways” and ventral “pathways” (Dick et al., 2014; Friederici, 2011; Friederici & Gierhan, 2013). I illustrate this in Figure 1.8, which offers a schematic view of the multiple dorsal and ventral language pathways. Figure 1.8. A schematic view of the dorsal and ventral pathways. FC: Frontal Cortex; OC: Occipital Cortex; TC: Temporal Cortex. Taken from Friederici & Gierhan (2013). And the last nuance, partly illustrated in some of the details of regional activation presented thus far, concerns the specifics of the functional properties of the regions composing both pathways. There is considerable agreement on the idea of multifunctionality of brain regions: the same region can serve multiple purposes, and several different regions can contribute to the same function (Bullmore & Sporns, 2009). This principle can be applied to the language networks, and as an example, we can take back the idea of the vOTC serving both lexico-semantic and sub-lexical processes (Lerma-Usabiaga et al., 2018; White et al., 2019). For all the reasons exposed above, in the sub-sections below, I will describe in further detail the connections and functions covered by each of the two main language networks, and the regions that constitute them. 27 OF 150 1.3.1. Dorsal Route In section 1.2., we described how the left STG supports auditory perception of linguistic stimuli, and more specifically, phonological discrimination of language sounds. Two main white matter fibre bundles connect the STG mainly to the dorsal PMC and partly to the the dorsal IFG (pars opercularis): the superior longitudinal fasciculus (SLF) and the arcuate fasciculus (AF). Focusing on the first dorsal pathway, the STG is connected indirectly, via the IPL, to the dorsal PMC by the SLF (Dick et al., 2014; Friederici & Gierhan, 2013). The IPL is known to be important for actively retrieving and keeping relevant information online (Cabeza et al., 2011; Sestieri et al., 2017), which in the case of language processing, entails phonological working memory (Friederici & Gierhan, 2013). The dorsal PMC is responsible for the motor articulation of speech (Hickok & Poeppel, 2007). As a consequence, lesions to any portion of the SLF, or any of the regions that form this connection, are known to produce deficits in the repetition of speech (Friederici & Gierhan, 2013; Hickok & Poeppel, 2007). A second dorsal pathway connects the left posterior STG to the pars opercularis of the IFG through the AF (and part of the SLF) (Catani et al., 2005; Thiebaut De Schotten et al., 2012). The lack of a clear consensus about the composition of the fibres forming this pathway has sometimes led researchers to name the tract SLF/AF (Dick et al., 2014; Friederici & Gierhan, 2013). The left pars opercularis (BA 44), which lies next to the PMC, has been proposed to take part in providing top-down predictions about the linguistic input (Friederici & Gierhan, 2013). Recently, the posterior part of BA 44, lying next to the PMC, has been linked to action observation and imitation, while the anterior BA 44, connected to BA 45, is relevant for the understanding of complex language rules and sequences processing (Friederici, 2023; Hamzei et al., 2016; Papitto et al., 2020). For simplification, the connection between the left STG and left posterior IFG (also via left PMC) is believed to be involved in complex syntactic processing (Friederici & Gierhan, 2013). We can summarise the role of the dorsal pathways for language by saying that they are in charge of processing the phonological aspects of words, retrieving and keeping in memory these sound units, and generating predictions about the input that will help both in articulating speech and in recognising written and spoken phonological units. 1.3.2. Ventral Route Given that the present work is mainly concerned with lexico-semantic access, it is not daring to say that we will mostly (albeit not exclusively) focus on the regions constituting these ventral pathways along this thesis. For this reason, here I will simply describe the connections of the areas forming the ventral pathways, as well as the overall function of the ventral routes. In the following chapters, I will address the diverse functional specialisation of the different regions involved in these pathways. 28 OF 150 Two main white matter tracts are known to connect the areas around the vOTC, and occipital cortex overall, to the ATL and the anterior part of the IFG (BA 45 and BA 47): the inferior longitudinal fasciculus (ILF) and the inferior fronto-occipital fasciculus (IFOF). A third tract, the uncinate fasciculus, connects the lateral and orbital frontal cortex, BA 47 and the anterior cingulate cortex (ACC) with the ATL, amygdala and parahippocampal cortex (Dick et al., 2014; Thiebaut De Schotten et al., 2012). Additionally, and outside the typical dorsal/ventral framework, the posterior vOTC has been shown to display significant connectivity with the ventral parietal lobe, presumably through the posterior AF, and the inferior longitudinal fasciculus (ILF) (Lerma-Usabiaga et al., 2018; López-Barroso et al., 2020; Moulton et al., 2019). Interestingly, this pattern has been shown to be associated with increasing reading experience (López-Barroso et al., 2020; Moulton et al., 2019). The vOTC has already been described as an area that is crucial for the visual recognition of lexical units, and that receives top-down influences from phonological areas. The ATL is known to be involved in semantic processing during complex sentence reading, and to represent semantic knowledge overall (Dick et al., 2014; Friederici & Gierhan, 2013). Finally, the anterior IFG (BA 45 and BA 47) is associated with semantic categorisation and deliberate access to semantics (Badre & Wagner, 2007; Friederici & Gierhan, 2013), while also participating in syntax processing (Friederici, 2011; Friederici & Gierhan, 2013). We will delve into the functions of ATL and the IFG in Chapter 2. In sum, the ventral pathways take on the task of categorising the linguistic input into discrete lexical units, link it to previously stored knowledge, and access it in a deliberate manner to interact with the world around us. 29 OF 150 CHAPTER II. THE NEUROBIOLOGY OF SEMANTIC REPRESENTATIONS The dorsal/ventral view constitutes a rather useful framework in understanding brain activation associated with specific subprocesses during language processing. Nevertheless, as it was originally tailored to explain the neuroanatomy of speech perception (Hickok & Poeppel, 2004), there are a number of aspects that the dorsal/ventral view leaves out. When it comes to comprehending how the brain creates and accesses semantic representations, several other models are especially helpful. Furthermore, language is not a unitary process, and the regions constituting the networks described thus far are not exclusively recruited for speech perception. Cognitive processes like memory are key to understanding how the human brain builds, represents, and accesses conceptual knowledge. In this chapter, we will focus on the most relevant models that aim at explaining the neurobiology of semantic representations, both through language and memory. For clarity, I will divide them into 1) models that are focused on the processes supporting the access and use of semantic knowledge, and 2) those focused on the stored semantic representations, as well as integrative models that combine aspects from both views. 2.1. MODELS FOCUSED ON THE PROCESS In this section, we will review two theoretical models that are especially concerned with the cognitive processes that allow the access, manipulation and control of the previously acquired linguistic information: the Memory, Unification and Control (MUC) model, proposed by Hagoort (2005) and the cognitive control of semantic memory view, put forward by Badre and Wagner (2007). While the scope of the MUC model, proposed by Hagoort includes syntactic, phonological and semantic processes, the model proposed by Badre and Wagner is mainly concerned with the access to semantic memory. 2.1.1. Memory, Unification and Control (Hagoort, 2005, 2013) Until well into the 21st century, the neurobiology of language was dominated by the Broca-Wernicke model (Hagoort, 2013). As briefly introduced in the previous section, this view understood that the main two neurocognitive components of language, production and perception, were mainly based on the activity of the left IFG and the left posterior STG, respectively (Hickok & Poeppel, 2007). But like most of the models in the early to mid 2000s that aimed at explaining how language is processed at the neural level, Hagoort (2005, 2013) built on the idea that there is a further division of labour outside the classical Broca-Wernicke framework. The starting point of Hagoort’s model is the identified need to explain both what are the discrete cognitive skills that language makes use of (cognitive architecture), and how these skills are supported by brain function (neural architecture). His MUC model tried to give 30 OF 150 response to these two essential needs, and it entails that versatile networks of regions from both hemispheres (although predominantly left) take on different components of language processing that dynamically participate both in speech production and comprehension. In particular, he defines three main components: Memory, Unification and Control (hence the name, MUC). Figure 2.1 illustrates the brain distribution and connections of the components in the MUC model. Figure 2.1. A) Distribution of the components of the MUC model at the neural level: Memory (yellow), Unification (blue) and Control (pink). Numbers indicate Brodmann areas. B) Schematic illustration of the connectivity in the left hemisphere language network. Red circles represent brain areas: fusiform gyrus (fg), temporal gyri (tg), angular gyrus (ag), temporal pole (tp), pars orbitalis (or), pars triangularis (tr) and pars opercularis (op). Grey arrows depict white matter tracts: inferior longitudinal fasciculus (ILF), uncinate fasciculus (UF), extreme capsule (EC) and arcuate fasciculus (AF). Green lines represent interfaces with sensory motor systems: visual cortex (vc), auditory cortex (vc) and motor cortex (mc). Adapted from Hagoort (2013). The Memory component comprises all linguistic knowledge that was acquired and consolidated in neocortical structures. This includes information about phonology of words, knowledge about their syntactic structures (like grammatical gender or word class), and conceptual lexico-semantic information. In the MUC model, phonology information is held in the central to posterior STG. In turn, the semantic information is distributed along several different areas, mainly around the MTG and ITG (Hagoort, 2005), but also posteriorly, in the AG (Hagoort, 2013). The retrieval of syntactic information is associated with the left posterior STG, although for the manipulation of this kind of information, the Unification and Control components and their respective neural pathways become especially relevant. Simple retrieval of stored linguistic information is not sufficient to cover the integrative use of lexico-semantic information and its connection with both different pieces of semantic knowledge and the context in which they appear. In this sense, the Unification component refers to the combination of the items stored in memory in novel ways in order to produce assembled meaningful structures. Unification was originally described through computational 31 OF 150 models at the syntactic level (Vosse & Kempen, 2000). At the sentence level, input words have associated syntactic structures (retrieved from memory) that can be ambiguous. As words are accessed, their syntactic structures are selected and subject to lateral inhibition processes that resolve this ambiguity. This whole mechanism is mainly associated with the activity of the mid to posterior left IFG (BA 45, 44) (Hagoort, 2005; Petersson, 2004). In the MUC model, Hagoort extends the unification component to the semantic and phonological spheres. Semantic Unification would entail the processes that guide the selection of the appropriate meaning among all potential acceptations of a word, or the link between the word and the preceding context. Phonological Unification refers to the selection of relevant segments within an intonation to guide the categorisation of the phonological units forming words and sentences. Although semantic and phonological Unification have been less consistently studied at the neural level, the MUC model proposes that semantic Unification is linked to the anterior left IFG (BA 45, 47), while phonological Unification is associated with the posterior left IFG (BA 44) and adjacent posterior areas including the supplementary motor area (SMA, BA 6) (Fedorenko et al., 2012; Hagoort, 2005, 2013). Figure 2.2 depicts the distribution of Unification processes in the left IFG. Figure 2.2. Distribution of the semantic, phonological and syntactic Unification components in the left IFG. Taken from Hagoort (2013). And finally, the Control component refers to the use of executive control to put the units of language in relation to action and context. When processing language, like, for instance, during reading, our brain needs to extract the essential information that allows us to decipher the visual input into meaningful information. In this task, our previous knowledge and 38 OF 150 model, these convergence areas include other regions outside the ATL, like the AG and SMG, the MTG and ITG, and the anterior vOTC. Another relevant difference is the importance conceded to these convergence areas. While Binder and Desai recognise the role of convergence zones in representing semantic abstract information, they defend a model in which sensory-motor information can give rise to multimodal representations without the express need of an amodal hub, which is why they termed their view embodied abstraction. Under this perspective, conceptual representations are formed by different levels of abstraction from sensory, motor and affective information. The access to each level of abstraction is determined by several factors, like frequency or familiarity with the information retrieved, or the demands of the task during which the information is used. Abstract representations are sufficient in highly familiar contexts, while in novel contexts sensory, motor and affective representations are additionally required. Thus, they describe four levels in the neuroanatomy of these representations (see Figure 2.5): 1) sensory, motor and affective areas (e.g., precentral gyrus or posterior STG); 2) abstract convergence zones (e.g., AG or ITG); 3) control areas that direct the selection of information according to the goals (e.g. dmPFC and IFG); and 4) areas that act as an interface between semantic and episodic information (e.g. precuneus and PCC). Importantly, these modules are flexible and highly interactive throughout the levels of abstraction. Figure 2.5. Embodied abstraction neuroanatomical model. Sensory, motor and affective regions (yellow areas) constitute the perceptual input to abstract convergence zones in parietal and temporal cortices (red areas). Control regions, like the IFG and dmPFC select the information directed to the goal at hand. Interface areas (in green), like the precuneus and PCC, connect semantic memory to 39 OF 150 the episodic experience, through their connections with the MTL. Taken from Binder & Desai (2011). 2.2.3. Controlled Semantic Cognition (CSC) (Ralph et al., 2017) Almost a decade after the distributed-plus-hub theory was put forward, their proponents integrated the most recent evidence regarding the function of the ATL and its functional segregations, as well as other related structures, in a revised review of the neurocomputational underpinnings of semantic cognition (Ralph et al., 2017). They mainly put together two lines of research to propose a double mechanism for semantic cognition: 1) a representation system, dedicated to the abstraction of perceptual input of diverse nature and their interrelations (Lambon Ralph et al., 2010); and 2) a control system that modulates the representations to adapt to the context and its demands (Badre et al., 2005). With this view, referred to as controlled semantic cognition (CSC), the authors reconcile some of the most relevant semantic research focused on the process, and that focused on the semantic representations per se. The representation system has already been described in the model proposed by Patterson and colleagues (see subsection 2.2.1). However, this updated review included some interesting nuances to the ATL function. First, the ATL role as an amodal hub that integrates different sources of information seems to be more specifically centred around the ventrolateral ATL (including ITG and medial FFG) (Abel et al., 2015; Shimotake et al., 2015). Second, the 40 OF 150 ATL can be further structurally and functionally segregated into subregions that convey different sorts of information. From neuroanatomical research, we know that the ATL is composed of graded neuronal ensembles (Ding et al., 2009) that receive structural connections from different areas, with certain degree of preference. For instance, the uncinate fasciculus connects the orbitofrontal cortex and anterior IFG most strongly to the temporopolar cortex, while the extreme capsule and middle longitudinal fasciculus connect the MTG and IPL with the superior part of the ATL. In turn, the ILF connects the visual cortex and the vOTC to the ventral ventromedial ATL (Binney et al., 2012). Importantly, this structural connectivity is believed to be reflected in a functional specialisation of the ATL, in the form of a gradient: the ventrolateral ATL represents abstract categories that do not depend on the input modality or stimulus category (Visser et al., 2012); the medial ATL is less cross-modal, and shows a preference for picture-based materials and concrete concepts (Clarke & Tyler, 2015; Hoffman et al., 2015); the superior ATL (anterior STG) shows a preference for auditory stimuli, spoken language and abstract concepts (Hoffman et al., 2015); and the polar ATL (its most anterior portion) is preferentially recruited by social concepts (Olson et al., 2013). Figure 2.6 illustrates this graded structural and functional segregation of the ATL. Figure 2.6. A) Computational model of the graded function of the ATL. B) Neuroanatomical framework of the functional specialisation of the ATL and its input connections. Taken from Ralph et al. (2017). The second component of the CSC, the control system, is an homologous of the already described set of mechanisms for the cognitive control of semantic knowledge (see section 2.1), with some added particularities. First, the areas that play an important role in executive control include areas posterior to the IFG, like the pre-SMA, and temporo-parietal regions like the posterior MTG and STG, and the IPL. The differences in the implication of 41 OF 150 frontal as opposed to temporo-parietal areas are subtle: lesions to the PFC usually lead to perseverant responses to a higher extent than temporo-parietal lesions do. This led the authors to propose that frontal areas might have a more extensive implication in inhibitory mechanisms than temporo-parietal regions (Ralph et al., 2017). The CSC model unifies many of the theories proposed thus far in the light of the updated evidence. While I present it in a section dedicated to the models that try to explain semantic representations at the neural level (partly because it is a re-elaboration of the distributed-plus-hub model), the CSC model focuses on both the representational aspects and on the cognitive processes that modulate such representations. This makes it a rather integrative model that is well worth describing. Some critical inconsistencies arise after evaluating the most influential models for the representation, access and manipulation of semantic information. We have already pointed out the discrepancies between embodied models and models based on amodal abstractions. While the former see conceptual information as relying on a widely distributed network of sensory-specific brain regions, the latter understand that this type of knowledge depends both on sensory and motor information and their abstract generalisations. To date, it is still unclear to what extent semantic information relies on sensorimotor representations or generalised abstractions, and how these representations vary as a function of the context (Ralph et al., 2017). In a similar vein, whether there are different neural mechanisms for the representation of meaning, with or without language processing, is still an open question (Fedorenko et al., 2024). In this sense, language requires a dynamic interaction between functionally separate systems (e.g., attention, inhibition, motor planning, etc.). But such systems are not completely independent, and often overlap. This makes it hard to tease apart the circumstances under which these networks are recruited, and in which cognitive processes they actually participate. Perspectives that conceive the spatial distribution of language networks as a gradient in which some areas participate dynamically in different cognitive functions (e.g., Lerma-Usabiaga et al., 2018; Ralph et al., 2017) do not challenge the idea that such functional specialisation exists (Fedorenko et al., 2024). Nevertheless, this is still a point of conflict in many specific research topics, including semantic representations (Fedorenko et al., 2024; Kanwisher, 2010). 42 OF 150 CHAPTER III. THE STUDY OF NEURAL LEXICAL REPRESENTATIONS Virtually all the theoretical models exposed thus far emerge from two types of evidence: 1) neuropsychological studies, like the ones that gave rise to the VWFA theory, or the studies on ATL lesions in semantic dementia; and 2) neuroimaging studies using a wide variety of techniques, of which fMRI has been perhaps the most popular, due to its potential to manipulate a wide variety of conditions and contexts, and capture the associated brain responses with high spatial resolution. Although lesional studies have been, and still are, essential for understanding brain function, neuroimaging studies clearly constitute the most flexible way of approaching the analysis of cognitive functioning at the neural level. The possibility of collecting indicators of brain activity during the engagement in diverse tasks and contexts, with a whole possible world of manipulations, is an ideal setting for exploring the neural underpinnings of a given cognitive process. Taken to lexico-semantic representations, we have a great example of this capacity in the meta-analysis described in subsection 2.2.2: by putting together 120 studies that used fMRI to measure brain activation during similar variations of semantic tasks, we could get a very good approximation of the regions that are likely to be involved during semantic processing. An additional point that we can extract from this is the necessity of having good spatial resolution, especially when it comes to exploring neural representations. Throughout the present work, we have alluded to fine-grained spatial divisions in several different brain areas, something that fMRI makes possible. For this reason, we will focus on fMRI, the technique employed in the works presented in this thesis. Many of the conclusions drawn in the theoretical models described in the previous chapter were possible due to the manipulation of specific psycholinguistic features captured in language. For instance, the comparison between abstract and concrete concepts in fMRI tasks is one of the most important contributors to the evidence supporting practically all theoretical frameworks presented (e.g., Binder et al., 2009; Hoffman et al., 2015). Additionally, some of the most recent findings about the functional specialisation of the vOTC were possible due to the use of manipulations that affected lexical features like the length of a word or properties that affect its readability (e.g, Lerma-Usabiaga et al., 2018; White et al., 2019). These possibilities have been considerably improved with recent advances in the field of computational language modelling and with the popularisation of novel analytical approaches applied in fMRI data analysis. The works presented in the current thesis harnessed the manipulation of psycholinguistic properties, partially making use of computational language models and novel fMRI multivariate analytical approaches. For these reasons, in this chapter I will describe the most relevant psycholinguistic properties associated with lexico-semantic access, followed by naturalistic language processing models, and a 43 OF 150 specific form of multivariate analysis approach for fMRI data called representational similarity analysis (RSA). 3.1. Psycholinguistic Properties as a Window to Lexical Representations Whether perceptibly or unconsciously, the chunks of information that we use in our everyday lives reflect an extensive variety of properties that affect how that specific part of semantic knowledge is processed. These chunks of information, that we can refer to as concepts, are captured in what we have been naming lexical units: words. Words can vary, for instance, in how often they appear on media (lexical frequency), or how familiar to us we feel that the content that the word refers to (word familiarity). They can refer to concrete entities that we can perceive or feel in the world around us, or be abstract and not have any evident physical analogue in the real world (word concreteness). This, in turn, will affect how easily we can create a mental visualisation or sensation of the entity referred to by the word (word imageability). They can allude to living beings or inanimate entities, man made objects (word animacy), and the referred entities can be stronger examples of a given category (e.g., dog as a mammal) than others (e.g. bat, also as a mammal) (taxonomic hierarchy). All these features are examples of semantically-related variables that reflect meaningful properties of the conceptual entities. Importantly, manipulating the degree to which words vary in these continua exerts a clear influence over the brain activation patterns generated by accessing the concepts referred to. This has been used to inform about the implication of different brain networks in distinct cognitive processes. As examples, the neural effects of manipulating animacy (Coggan & Tong, 2023; Grill-Spector et al., 2017; Grill-Spector & Weiner, 2014; Jozwik et al., 2022) and taxonomic hierarchy (Ritchie et al., 2021), often studied in the context of visual categorisation of pictures, have greatly contributed to our knowledge about the functional specialisation of the vOTC (see Thorat et al., 2019, and subsection 1.1.2). 3.1.1. Word Concreteness and Imageability As mentioned earlier, a great deal of what we know about how the brain represents conceptual information comes from the analysis of the concreteness continuum. This continuum is often defined subjectively, by asking a large sample of participants of a given language to rate an extensive set of words from 1 (absolutely not perceptible/abstract) to 7 (absolutely perceptible/concrete) (e.g., Duchon et al., 2013). The reader might be wondering: how can such a specific feature of language tell us so much about such a complex neurocognitive ability? A potential answer to this is captured by the dual-coding theory (DCT) (Paivio, 2010, 2014). This view understands that representations can be acquired and held either through perceptual encoding or via verbal encoding. Concrete concepts follow a double 44 OF 150 encoding process, since the referred entities can be perceived through our senses, and verbalised through language. In contrast, abstract concepts are mostly acquired through language, thus counting on weaker support from perceptual experience. The DCT can explain the well reported behavioural effect of concreteness (i.e., abstract concepts being harder to process and more effortful to learn) (Feyereisen et al., 1988; Mkrtychian et al., 2019; Palmer et al., 2013; Schwanenflugel et al., 1988). The fact that abstract words are usually learned later in life, and are on average less familiar than concrete words is also in line with the DCT (Striem-Amit et al., 2018). And most importantly, the neuroimaging findings over the last 20 years are to a great extent consistent and in line with the DCT. Concrete concepts mainly recruit distributed modality-specific sensory and motor areas like the precuneus, PCC, FFG and parahippocampal gyrus (Hoffman et al., 2015; Vignali et al., 2023; J. Wang et al., 2010), but also multimodal and language areas to some extent (Binder et al., 2009; Hoffman et al., 2015; Vignali et al., 2023). In contrast, abstract words are associated with responses in multimodal areas like the ATL (Hoffman et al., 2015; Vignali et al., 2023; J. Wang et al., 2010), the left IFG (Binder et al., 2009; J. Wang et al., 2010) and left posterior MTG and STG (Binder et al., 2009; J. Wang et al., 2010). Figure 3.1 shows the neural distribution of the activations associated with concrete vs. abstract words, taken from the meta-analyses evaluating this comparison by Wang et al (2010) (but see also Binder et al 2009). Figure 3.1. Activation likelihood of abstract and concrete words, taken from the meta-analysis by Wang et al (2010). 45 OF 150 However, there is no absolute agreement in the interpretation of the neural effects of abstract as opposed to concrete conceptual processing. The more embodied perspectives argue that many abstract concepts are grounded in situations, mental states, events or relations between objects, the same way that concrete concepts can reflect abstract features (Borghi & Binkofski, 2014). In addition, abstract concepts are often associated with higher degrees of affective content. Some studies demonstrated that words that can be grounded in emotions are learned earlier in life, and benefit from this emotional valence grounding, while concrete words do not (Ponari et al., 2018). Moreover, some of the neural effects associated with abstract word processing can be partially explained by the influence of affective valence. For instance, activations observed in the rostral ACC could be related with the processing of hedonic value captured by abstract words (Vigliocco et al., 2014). Similarly, affective valence has been found to be well represented in the superior ATL only for abstract words processing (Meersmans et al., 2020). A factor that is tightly related with word concreteness is imageability (a measure that is also obtained from subjective ratings in normative studies). In fact, previous studies have reported considerably high correlations between these two factors (Westbury et al., 2013). Nonetheless, the extent to which a word can evoke a tangible visualisation/sensation can add some information about semantic representations. For instance, a word (e.g., eternal) can be judged as clearly abstract, but as partially imageable (Westbury et al., 2013), perhaps owing to its capacity to evoke bodily sensations and mental imagery. This can provide additional details about the effects observed in fMRI studies. For example, although fMRI evidence found when comparing lowversus high-imageability words overlap considerably with concreteness neural effects (i.e., activation of modality-specific sensory areas like the FFG when processing imageable concepts) (Bedny & Thompson-Schill, 2006; Lewis & Poeppel, 2014), additional effects, like the involvement of the hippocampus when processing easily imageable concepts have been reported (Caplan & Madan, 2016; Klaver et al., 2005). Throughout the rest of the thesis, while I consider the mentioned high collinearity with word concreteness, I assume that word imageability can still contribute to a better understanding of semantic neural representations. 3.1.2. Word Frequency and Familiarity On a different vein, another factor that has contributed enormously to our understanding of the cognitive and neural mechanisms involved in the access and representation of lexical knowledge is word frequency. Our brains can take advantage of repeated exposure to words in different ways. Upon encountering a new word, we face the 46 OF 150 laborious task of deciphering its fundamental orthographic and phonological components. Through repeated exposure, these novel words can become consolidated, incorporating their phonological, morphological, and syntactic characteristics (Hagoort, 2013). They are further connected to other words and concepts, and can become semanticised into an amodal semantic network (see section 2.2; Memetova et al., 2024). This progression streamlines the reading process, reducing the need for exhaustive perceptual analysis of the word, and enhancing reading efficiency (Desai et al., 2020). We can connect this process to the distinction between the dorsal (orthographic-phonological) and ventral (lexico-semantic) language networks explained in section 1.3: reading unknown words require the involvement of the dorsal phonological network, while words that are well known to us may use the ventral lexico-semantic route to process the word as a whole, without the need to decipher its orthographic and phonological components. In practice, the word frequency effect (WFE) is a well-documented phenomenon in the neuroscience of reading. It refers to the fact that low-frequency words pose greater processing challenges compared to high-frequency ones. Extensive behavioural research consistently demonstrated faster responses to high-frequency words as opposed to low-frequency words across a variety of tasks such as word naming, lexical decision, and semantic decision tasks (Brysbaert et al., 2018). At the neural level, several fMRI studies have analysed the WFE. In those studies, the WFE is defined as heightened regional brain activation for low-frequency words, suggesting more demanding cognitive processing. Notably, the IFG emerges as a key neural locus for this effect, although other studies have found a WFE in the vOTC, and less consistently in other areas, like SMA or ACC. Table 3.1 provides a synthesis of previous fMRI investigations on the WFE, outlining their key findings, methodological parameters, and sample sizes. Table 3.1 Previous fMRI studies examining the WFE, main findings, reading tasks used and sample sizes. Studies Regions showing the WFE Reading Tasks Sample sizes (N) Chee et al. (2002) Left IFG (BA 44) Silent reading vs. semantic judgements 16 Fiebach et al. (2002) Left IFG (BA 44, 45) Lexical Decision Task 12 Chee et al. (2003) Left ACC (BA 32), IFG (BA 44, 45), ITC (BA 37) Semantic judgement + 24h Recognition 16 47 OF 150 Kuo et al. (2003) Left Precentral, SMA, IFG (BA 44), vOTC Covert naming + Recognition 28 Joubert et al. (2004) Left IFG (BA 45, 47) Silent reading 10 Kronbichler et al. (2004) Left IFG (BA 45, 47), vOTC (mid) Silent reading 13 Carreiras et al. (2006) Left IFG (BA 44) Lexical Decision vs Reading aloud 16 Graves et al. (2007) Left IFG (BA 45/47), vOTC (post), pSTG Picture naming (overt) 59 Hauk et al. (2008) Bil. IFG (BA 45, 47), vOTC (ant) Silent reading 21 Bruno et al. (2008) Left Precentral, IFG (BA 44, 45), vOTC, pSTG Phonological Lexical Decision Task 28 Carreiras et al. (2009) Bil. ACC/ IFG (BA 45, 47), Precuneus, SMA Lexical Decision Task 20 Schuster et al. (2016) Left IFG (BA 45), vOTC Silent Sentence Reading 56 Rundle et al. (2018) Left ITC (BA 37); vOTC Silent reading + semantic catch trial 19 These findings carry significant implications for the understanding of the neural networks involved in processing information from phonological to semantic. However, the interpretation of these results is diverse. In short, while some authors attribute the effect to the search and access to phonological information (Carreiras et al., 2009; Fiebach et al., 2002), others interpret it as an indicator of more effortful search of semantic information (Chee et al., 2002). In truth, the phonological and semantic interpretations of the WFE are not completely incompatible. Processing a word that appears often on media can be easier both because the access to its orthographic (syllables) and phonological (phonemes) components is enhanced after repeated exposure to it. But as I indicated before, a proficient reader often reads known 54 OF 150 and the conceptual matrix (Kriegeskorte, 2008; Popal et al., 2019). Figure 3.4 represents a schematic application of RSA. In the last few years, RSA has been successfully applied to explore the neural representations of a variety of dimensions in the field of semantics, lexical information and mnemonic representations (Carota et al., 2017, 2021a; Liuzzi et al., 2023; Meersmans et al., 2020, 2022; Viganò et al., 2021; Y. Wang et al., 2023; Yacoby et al., 2021). In Chapter 5, we will cover an empirical application of RSA to disentangle the neural representations of, and dynamic relations between, the most important psycholinguistic variables, as well as NLP models. Figure 3.4. Schematic application of RSA. A) RDM expression of voxel pattern responses to stimuli of different nature (here, houses and faces). B) Multiple possibilities for the comparison between the RDM and an empirical model. The empirical model can be built from a wide range of features, from computational, to data from other techniques (e.g. EEG or cell recordings), and even other species. Adapted from Kriegeskorte et al (2008). 55 OF 150 CHAPTER IV. THE ROLE OF READING DEMANDS AND WORD FREQUENCY IN THE ACCESS TO LEXICAL UNITS 4.1. RATIONALE As we described in section 3.1.2, research on the WFE has typically involved diverse methodological settings. The fMRI methodological procedures used in previous studies have varied greatly, including differences in protocols for multiple comparison corrections or even the absence of such corrections for statistical significance thresholds. This diversity has influenced the observed effects of frequency in the brain regions described in section 3.1.2, making the WFE results difficult to interpret. The inconsistencies in the location of the neural activations found has contributed to a certain degree of diversity in the interpretation of these effects. In some of these studies, the WFE extends to the anterior part of the IFG (i.e., pars orbitalis and pars triangularis, BA 47 and 45, respectively), while in others it is associated with the posterior IFG (i.e., pars opercularis, BA 44), or both (see Table 3.1). In consequence, while some researchers associate the observed WFE in the IFG with phonological processing or retrieval during lexical search (Carreiras et al., 2009; Fiebach et al., 2002), others suggest it hinges on deliberate access to semantic information (Chee et al., 2002). As briefly introduced in the previous chapter, both interpretations are not completely incompatible. Here we propose that, if the WFE relies on phonological processes, then it would be mainly (although not exclusively) observable in regions along the dorsal reading network (i.e., IFG pars opercularis, STG, and/or IPC) involved in mapping visual percepts onto the phonological structure of the language. In contrast, if the WFE relies on lexico-semantic processing, we predict that the WFE will mainly (but not exclusively) rely on the engagement of regions along the ventral reading network (i.e., IFG pars triangularis, IFG pars orbitalis, and/or the vOTC), involved in mapping orthographic-lexical stimuli to words as a whole (Oliver et al., 2017; Sandak et al., 2004). Indeed, the studies defending the semantic interpretation have typically found the WFE in regions of the ventral reading network, whereas the studies defending the phonological interpretation have typically found the WFE to occur in regions along the dorsal reading network (see Table 3.1). The relationship between phenomena such as the WFE and the functional specialisation of the vOTC has also sparked debate. The reader will remember from section 1.1.3 that this region contains the putative VWFA, which would be, according to its proponents, expressly dedicated to the visual recognition of word forms. Conversely, we can also bring back the idea of the vOTC responding to lexical and semantic properties (especially its anterior portion), and even to stimuli other than visual words. The fact that the vOTC has been found to respond to word frequency is part of the evidence against the VWFA perspective (Kronbichler et al., 2004; Kuo et al., 2003; Schuster et al., 2016). Nevertheless, while a number 56 OF 150 of studies have found a WFE in the vOTC (Bruno et al., 2008b; Graves et al., 2007; Hauk et al., 2008; Kronbichler et al., 2004; Rundle et al., 2018; Schuster et al., 2016), others have failed to find such an effect (Carreiras et al., 2006, 2009; Chee et al., 2002, 2003; Fiebach et al., 2002; Joubert et al., 2004), and when found, the locations of the effect in the vOTC are somewhat inconsistent. These facts render the WFE in the vOTC challenging to interpret. On the other hand, it has been indicated that the engagement of both the IFG and the vOTC could be modulated by top-down processes, such as those imposed by reading demands (Price & Devlin, 2011; Rundle et al., 2018; Yang & Zevin, 2014). Previous fMRI studies have used a variety of conditions and tasks, ranging from silent reading to lexical decision or semantic decision tasks (see Table 3.1). Some authors have proposed that, in the case of the IFG, lexical decisions (Rundle et al., 2018) or naming efforts (Vogel et al., 2013) might be amplifying the effects found in this region. In the case of the vOTC, the prolonged exposure time, combined with the demands of tasks like semantic judgement or lexical decision tasks, could make this region to be more engaged during the processing of lowfrequency words (Schuster et al., 2016). Again, these contrasting interpretations are especially controversial in regard to the main theoretical accounts of the vOTC. Finally, and adding to the mentioned limitations, most previous fMRI studies on the WFE have concentrated on regional activation, neglecting to examine the functional connectivity patterns that might underlie this effect. Given these inconsistencies, our primary objective was to examine the WFE in the activation profiles of regions within the ventral and dorsal reading networks using complementary analytical approaches. We also aimed to investigate the potential interaction between reading demands and word frequency in these functional patterns. To achieve this, we employed two versions of a single-word reading task: a perceptual task (low reading demand) and a semantic task (high reading demand). This allowed us to assess the influence of reading demands imposed by these two types of tasks on the WFE. Additionally, we aimed to determine whether there were distinguishable functional connectivity profiles among ventral and dorsal reading regions that respond to word frequency and reading demands. Consistent with previous evidence and theories regarding the division of labour between these two networks (e.g., Pugh et al., 2001), we expected to observe the WFE in regions of the lexico-semantic ventral reading network, such as the IFG and the vOTC, characterised by higher activation in these areas for low-frequency words compared to high-frequency words. Furthermore, we predicted this effect would be stronger in the semantic task than in the perceptual task and that these findings would be replicated across different analytical approaches. In terms of functional connectivity, we anticipated that the WFE would be associated with stronger functional connectivity within regions along the 57 OF 150 ventral reading network. The methods and results discussed in the following sections were published elsewhere as of October of 2023 (Sánchez et al., 2023) 4.2. METHODS 4.2.1. Participants The total sample of the study consisted of 54 right-handed native Spanish-speaking participants, of 29.3 years of age on average (SD = 6.88 years; 30 females). All participants had normal or corrected-to-normal vision, and no known history of neurological or psychiatric illness. Of the initial 57 participants, one participant was excluded due to excessive head motion during scanning (see fMRI data analysis section below), and two participants were excluded due to the absence of recorded responses to the catch (i.e., Go) trials in the fMRI tasks. Language proficiency was assessed using both objective and subjective measures. As an objective measure, we used an adapted Spanish version of the Boston Naming Test (de Bruin et al., 2017). As a subjective measure, participants filled in a language proficiency selfrated questionnaire, in which they evaluated their own proficiency, as well as language exposure. All participants gave written informed consent in compliance with the ethical regulations established by the BCBL Ethics Committee and the guidelines of the Helsinki Declaration. All participants received monetary compensation for their participation. 4.2.2 Materials and Procedure The experimental design consisted of two single-word reading Go/No-Go tasks, one perceptual (low reading demand) and the other semantic (high reading demand). In both tasks, all participants were visually presented with character strings in their native language (i.e., Spanish) that could be words or nonwords. In the perceptual task, participants were asked to press a button any time they saw a coloured letter within a string. In the semantic task, participants had to press a button any time they read a word referring to an animal (instead of strings containing a coloured letter). All stimuli were presented for 1.5 seconds on the centre of the screen. The tasks were divided in different functional runs that were counterbalanced between participants. Each task included a total of 80 words, of which 40 were high frequency words and 40 were low frequency words, and 80 nonwords. Thus, two sets of words and corresponding nonwords were developed and their use in either the perceptual task (i.e., low reading demand) and the semantic task (i.e., high reading demand) was counterbalanced between subjects. Word frequency is objectively defined by measuring the number of appearances of a given word, per million words, on a large sample of text sources (Brysbaert et al., 2018). Its most commonly used measure is the logarithmic transformation of the frequency per million 58 OF 150 words, the Zipf scale, which ranges from 1 (very low frequency) to 7 (very high frequency) (Van Heuven et al., 2014). In both sets, low-frequency words were nouns with a Zipf measure lower than 4, and high-frequency words were nouns above this cutoff. All word measures were obtained from EsPal (Duchon et al., 2013), and the two sets of words were matched on frequency, length (i.e., 5-8 characters) and number of orthographic neighbours. Nonword strings were included as stimuli in the experimental design to address other research questions not relevant for the present work. To reduce the potential reading demands imposed by nonwords as much as possible, they were designed so that they were legal, legible strings. Furthermore, the two sets of nonwords were also matched in length. Additionally, we included 13% of Go trials (i.e., either words with a coloured letter or animal words) as catch trials for each of the two reading tasks. Nonwords and Go trials were modelled as regressors of interest but not analysed. The stimuli used for the perceptual and semantic reading tasks were also counterbalanced between subjects. 4.2.3. fMRI Data Acquisition Whole-brain fMRI data were obtained on a 3-T Siemens TRIO whole-body MRI scanner (Siemens Medical Solutions) at the Basque Center on Cognition, Brain and Language (BCBL), using a 32-channel whole-head coil. The area between the participants’ heads and the coil was padded with foam in order to reduce head movement, and the participants were asked to stay as still as possible. Snuggly fitting headphones (MR Confon) were used to dampen background scanner noise and to allow communication between participants and experimenters. The functional images were acquired using a gradient-echo echo-planar pulse sequence with the following parameters: time repetition (TR) = 2000 ms, time echo (TE) = 25 ms, 35 contiguous 3-mm axial slices, 0-mm inter slice gap, flip angle= 90º, field of view = 218 mm, 64 x 64 matrix. The first four volumes of each scan were discarded to allow T1equilibration effects. The order of the conditions of the study within each run, as well as the inter-trial intervals of variable duration, were determined with an algorithm designed to maximise the efficiency of the recovery of the blood oxygen level-dependent response: Optseq II (Dale, 1999). High-resolution T1-weighted anatomical images were also acquired with the following acquisition parameters: TR = 2300 ms, TE = 2.97 ms, flip angle = 9º, Field of view = 256 mm, 176 volumes per run, voxel size = 1 cubic mm. 4.2.4. fMRI Data Analyses Standard SPM12 (Wellcome Department of Cognitive Neurology, London, UK) preprocessing routines and analysis methods were employed. Images were corrected for 59 OF 150 differences in timing of slice acquisition and realigned to the first volume by means of rigidbody motion transformation. Motion parameters were extracted from this process and were used, after a partial smoothing of 4-mm full width at half-maximum (FWHM) isotropic Gaussian kernel, to inform additional motion correction algorithms implemented by the Artifact Repair toolbox (ArtRepair; Stanford Psychiatric Neuroimaging Laboratory), intended to repair outlier volumes with sudden scan-to-scan motion exceeding 0.5 mm and volumes whose signal fluctuations in global intensity was > 1.3 % SD away from the mean. The correction of these outlier volumes was performed via linear interpolation between the nearest non-outlier time points (Mazaika et al., 2009). Data from 1 subject requiring more than 15% of their volumes to be repaired was discarded. For the final sample of participants, the average percentage of repaired volumes was 1.8% (SD = 2.8%). After volume repair, functional volumes were coregistered to the T1 images using 12-parameter affine transformation and spatially normalised to the Montreal Neurological Institute (MNI) space by applying nonlinear transforms estimated by deforming the MNI template to each individual’s structural volume. During normalisation, the volumes were sampled to 3-mm cubic voxels. Functional volumes were then smoothed with a 7-mm FWHM isotropic Gaussian kernel. Due to the quadratic relation between separate smoothing operations, the total smoothing applied to the functional data was approximately equivalent to smoothing with an 8-mm FWHM Gaussian kernel. Finally, time series were temporally filtered to eliminate contamination from slow frequency drift (high-pass filter with a cutoff period of 128 s). Statistical analyses were performed on individual participant data using the general linear model (GLM). fMRI time series data were modelled by a series of impulses convolved with a canonical hemodynamic response function (HRF). The motion parameters for translation (i.e., x, y, and z) and rotation (i.e., yaw, pitch, and roll) were included as covariates of non-interest in the GLM. Each trial was modelled as an event, time-locked to the onset of the presentation of each character string. The resulting functions were used as covariates in a GLM, along with a basic set of cosine functions that high-pass filtered the data. SPM12 FAST was used for temporal autocorrelation modelling in this GLM due to its optimal performance in terms of removing residual autocorrelated noise in first-level analyses (Olszowy et al., 2019). The least-squares parameter estimates of the height of the best-fitting canonical HRF for each study condition were used in pairwise contrasts. Regarding such analyses, whole-brain contrasts were computed by performing one-sample t-tests on the contrast images. Region-of-interest (ROI) analyses were carried out by using the MARSBAR toolbox for SPM12 (Brett et al., 2002). Six left-lateralised regions along the reading network were functionally identified using two different procedures: I) group level and II) individual-subject 60 OF 150 level. The group ROI identification procedure identified active voxels obtained from the whole brain contrast Words > Null across all participants, cluster Family-wise error (FWE) corrected, p < 0.001 voxel extent. The regions identified included pars orbitalis (centre of mass: -37, 27, -8; mm3= 1656), pars triangularis (centre of mass: -46, 27, 14; mm3= 12704), pars opercularis (centre of mass: -48, 11, 20; mm3= 5856), MTG (centre of mass: -46, -59, -2; mm3= 1008), IPC (centre of mass: -30, -53, 46; mm3= 2200) and vOTC (centre of mass: -43, -59, -17; mm3= 6296). As for the individual ROIs procedure, the same six regions were localised at the individual-subject level. To this end, 5 mm radius spheres were created by selecting the local maxima in each individual subject’s Words > Null contrast (cluster FWE corrected, p < 0.001 voxel extent) that fall within the anatomical mask of the six above-mentioned ROIs. For those subjects that had no voxels over the threshold that fell within the anatomical mask, the closest local maxima that allowed a sphere to be built falling within the mask was selected. The selection of the local maxima for individual ROIs in all participants were systematically checked by two authors (A.S. and P.M.P-A). 4.2.5. Functional Connectivity Analyses We used the beta-series correlation method to compute functional connectivity analyses (Rissman et al., 2004) by using custom Matlab scripts for SPM12. As in ROI analyses, functional connectivity analyses were performed on both group and individual ROIs. The occurrence of each event was modelled with the canonical HRF, which allowed for the extraction of the parameter estimates (i.e., beta correlations) associated with each condition in every voxel. Following this, pairwise connectivity between the 6 left-lateralised ROIs described above was conducted. After Bonferroni’s correction, a value of r > .355 was considered to show a significant functional connectivity between nodes. Further contrasts (i.e., t-tests) on the beta correlations associated with low versus high Frequency, and perceptual versus semantic Task, were carried out after Fisher’s Z transforms (Fisher, 1921) of the beta Pearson’s r correlations values to make the null hypothesis sampling distribution approach that of the normal distribution. 4.3. RESULTS 4.3.1. Behavioural Performance All participants showed a high response rate to Go trials overall, with an average accuracy of 99.4% (SD = 0.02%) for the perceptual task, and 95.4% (SD = 0.05%) for the semantic task. This indicates that participants were focused on the instructions given by the experimenter and performing the task. As expected, accuracy was slightly, but significantly higher for the perceptual task (t = 4.880; p < 0.001). Likewise, on average, response times 61 OF 150 were significantly faster for the perceptual than for the semantic task (t = -8.697; p < 0.001; perceptual M = 560 ms, SD = 90 ms; semantic M = 760 ms, SD = 140 ms). 4.3.2. Whole-brain results When contrasting all trials including words against baseline, the averaged activation map for words across all subjects (see Figure 4.1A) included regions in the occipital cortex, such as the lingual gyrus and the cuneus, the vOTC, IPC, MTG, the middle and superior frontal gyrus, the precentral and postcentral gyri, and the different subregions within the IFG (pars orbitalis, pars triangularis and pars opercularis). Figure 4.1. Whole-brain contrasts and low-high frequency simple contrasts. A) Results of the Words versus Null contrast across all subjects; B) Low > High frequency contrast in the perceptual (low reading demand) task (in red), and the semantic (high reading demand) task (in green). p < 0.05 FWE corrected clusterwise (p < 0.001 uncorrected voxel-extent threshold). Additionally, we computed the whole-brain Low > High frequency contrasts in both the perceptual and the semantic tasks separately. This contrast reflects the brain distribution of the WFE in either task (Figure 4.1B). In the perceptual task, no voxels survived the threshold 62 OF 150 of p < 0.05 (FWE corrected clusterwise, with a p < 0.001 uncorrected voxel-extent threshold). On the other hand, the semantic task showed a WFE exceeding the established threshold in the whole left IFG, with global maximae located in the anterior IFG (pars triangularis and pars orbitalis). 4.3.3. ROI analysis For each of the selected ROIs, a 2x2 ANOVA with Frequency and Task as factors, and percent signal change (PSC) as the dependent measure, was carried out. These analyses were followed by simple post-hoc pairwise t-tests for planned comparisons, for which Bayes factors (BF) are reported below. Results from group ROIs are reported first, followed by individual ROIs results. Figure 4.2 shows an overview of the results regarding the WFE by Task, across all group ROIs. An additional 2x2x2 ANOVA, with ROI type (group versus individual ROI) x Frequency x Task was carried out to determine any possible effects arising from the type of ROIs used (group vs individual). Aside from a significant ROI type x Task interaction in IFG pars triangularis (F = 8.078, p = 0.004), we found no significant main or interactive effects of the factor ROI type and, therefore, only results from group ROIs are shown in Figure 4.2. Table 4.1 shows a summary of the results from the group and individual ROI ANOVAs. Table 4.1. Statistical results from the group and individual ROI ANOVAs. Main effects of Frequency and Task, as well as their interaction are reported. Asterisks denote a statistically significant effect. ROI names Group ROI Individual ROIs Frequency Task Interaction Frequency Task Interaction Orbitalis F = 10.347 p = 0.002* 𝜂2 = 0.163 F = 7.222 p = 0.009* 𝜂2 = 0.119 F = 5.595 p = 0.021* 𝜂2 = 0.095 F = 7.058 p = 0.010* 𝜂2 = 0.121 F = 5.870 p = 0.018* 𝜂2 = 0.103 F = 3.211 p = 0.079 𝜂2 = 0.059 Triangularis F = 6.043 p = 0.017* 𝜂2 = 0.102 F = 11.946 p = 0.001* 𝜂2 = 0.183 F = 4.812 p = 0.017* 𝜂2 = 0.083 F = 4.710 p = 0.034* 𝜂2 = 0.083 F = 43.873 p < 0.001* 𝜂2 = 0.457 F = 2.785 p = 0.101 𝜂2 = 0.050 Opercularis F = 5.384 p = 0.024* 𝜂2 = 0.092 F = 9.345 p = 0.003* 𝜂2 = 0.149 F = 1.957 p = 0.16 𝜂2 = 0.035 F = 8.035 p = 0.006* 𝜂2 = 0.133 F = 20.109 p < 0.001* 𝜂2 = 0.278 F = 0.835 p = 0.364 𝜂2 = 0.015 IPC F = 0.596 p = 0.443 𝜂2 = 0.011 F = 4.425 p = 0.040* 𝜂2 = 0.077 F = 0.309 p = 0.580 𝜂2 = 0.005 F = 0.012 p = 0.715 𝜂2 < 0.001 F = 4.022 p = 0.050* 𝜂2 = 0.071 F = 0.134 p = 0.715 𝜂2 = 0.002 MTG/STG F = 0.367 p = 0.547 𝜂2 = 0.006 F = 2.169 p = 0.146 𝜂2 = 0.039 F = 0.008 p = 0.928 𝜂2 < 0.001 F = 0.077 p = 0.781 𝜂2 = 0.001 F = 0.972 p = 0.328 𝜂2 = 0.018 F = 0.023 p = 0.879 𝜂2 < 0.001 63 OF 150 vOTC F = 0.013 p = 0.909 𝜂2 < 0.001 F = 7.331 p = 0.009* 𝜂2 = 0.121 F = 0.725 p = 0.398 𝜂2 = 0.013 F = 0.060 p = 0.807 𝜂2 = 0.001 F = 9.841 p = 0.002* 𝜂2 = 0.159 F = 1.704 p = 0.197 𝜂2 = 0.031 Ventral Network Pars orbitalis. Following the group ROI approach, a main effect of Frequency was found for this region. We also found a significant main effect of Task. These main effects were qualified by a significant Task x Frequency interaction. Simple-effect post-hoc analyses revealed that this interaction was due to low-frequency words showing stronger regional activation than high-frequency words in the semantic (t = -3.907, p < 0.001, BF = 91.709), but not in the perceptual task (t = -0.687, p = 0.495, BF = 0.186). Results from the group ROI of the pars orbitalis were replicated with individual ROIs, although the Task x Frequency interaction resulted marginally significant, possibly due to differences in signal intensities derived from both approaches. Figure 4.2. ROI analyses. A) Group ROIs employed, obtained from the Words-Null contrast. B) Results from the group ROI analyses. ROIs are represented in the X axis, whereas the difference of the parameter estimates between regional activation of low versus high frequency words is depicted in the Y axis. Boxes with straight lines represent the perceptual task, whereas boxes with dotted lines depict the semantic task. Red asterisks at the bottom part indicate that the region showed a significant main effect of FrequencyTask (*p < .05, BF > 1; **p < .01, BF > 5; ***p < .001, BF > 10). Blue asterisks at the bottom part indicate that the region showed a significant main effect of TaskFrequency (*p < .05, BF > 1; **p < .01, BF > 5; ***p < .001, BF > 10). Black asterisks over the boxes indicate that the region showed a significant Task x Frequency interaction, here depicted as a significant difference 70 OF 150 4.5. CONCLUSIONS By applying different analytical approaches to a large dataset of 54 individuals, here we offer robust evidence that the activity in the left pars opercularis is modulated by word frequency, possibly reflecting phonological processing during lexical search. The activation of the pars opercularis and IPC are also modulated by reading demands, possibly reflecting stronger phonological processing during semantic word reading. The same reading demand effect was observed in the vOTC, but in the absence of any effect of word frequency, which seems to support the notion that this area is influenced by top-down processes. This has potential implications for theories about the role of the vOTC in pre-lexical and lexical processes. Finally, the WFE is modulated by reading demands in pars orbitalis and pars triangularis, since the WFE was only present under semantic reading demands. This is interpreted as an indicator of the role of the anterior IFG in controlled access to semantics. These effects, occurring at the regional activation level, seem to underline the role of the ventral reading network in the access to lexico-semantic information. 71 OF 150 CHAPTER V. NEURAL REPRESENTATIONS OF LEXICO-SEMANTIC KNOWLEDGE: SIMILARITY OF SUB-LEXICAL AND LEXICAL MODELS WITH MULTIVARIATE BRAIN RESPONSES 5.1. RATIONALE As reflected in Chapter 3, interpreting the effects of psycholinguistic properties like word concreteness, familiarity or frequency at the brain level can be troublesome, given the complex interrelation between such properties, and their interaction with expectations imposed by the task. While we tried to tackle the latter factor in the previous chapter, the study was agnostic to complex interactions between word properties, as acknowledged in section 4.4.5. The investigation presented in this chapter was originally designed to overcome the limitations of the study described in Chapter 4. It also attempted to improve our understanding of neural representations associated with phonological processing as compared to those linked to semantic processing. Finally, the current study also tried to offer new insights into the nature of the brain representations elicited by NLPs, thus addressing some of their conceptual limitations as mentioned in section 3.2. More specifically, in Chapter 3 I alluded to the differences in affective valence between abstract and concrete words, as well as their contextual representation. I also mentioned how abstract words are, on average, learned later in life and less familiar than concrete words. The high collinearity between concreteness and imageability, or between frequency and familiarity has also been pointed out. At the neural level, the well documented neural effects of semantic variables like word frequency, concreteness or familiarity, mostly come from investigations that addressed their study in isolation (with some exceptions mentioned in Chapter 3). But in truth, the information about conceptual neural representations that these variables can provide when studied together, in a controlled fashion, has been rarely explored. And most importantly, the implications of univariate designs that mainly focus on one or two of the above-mentioned factors, for both lexico-semantic and sublexical processing, are considerably limited. In the present study, our general objectives were to disentangle the brain representations of several different language properties, from sublexical (i.e., bigram and biphone frequency, phonological neighbours, orthographic distance) to semantic (i.e., concreteness, frequency, familiarity), and to investigate whether these properties are better represented together or in isolation in brain areas critical for language processing. Additionally, we had the specific objective of improving our understanding of the linguistic properties that are reflected in naturalistic language models (i.e. word vectors). To answer the general and specific objectives, we carefully selected a large pool of words that heterogeneously varied in sublexical and lexico-semantic properties, and tested them with fMRI in a simple word reading task. This allowed us to estimate several different models based on diverse combinations of 72 OF 150 all of the above-mentioned language features, and to compare them, in multivariate similarity analyses, with brain activation patterns elicited by reading these words. Comparing models that are intentionally “biased” to represent different combinations of either semantic, phonological or both kinds of features at the same time, has the potential to reveal how the brain regions that are key for language processing represent lexico-semantic knowledge. For instance, comparing phonological and semantic models in the IFG will reveal to what extent the different subregions of the IFG and adjacent areas participate in sublexical, lexicosemantic or both processes. Similarly, contrasting semantic versus phonological models in the vOTC would reveal whether indeed the anterior vOTC is implicated in semantic processing to some degree, while its posterior subdivision shows a preference for sublexical information. Specifically, here we expect 1) that the left IFG will show a dissociation in which the anterior IFG activation is better predicted by semantic models, whereas in its posterior subdivisions, both semantic and phonological models will yield comparable correlation values with brain activation patterns (Sánchez et al., 2023); 2) that the left vOTC will show a dissociation from anterior to posterior, where the anterior but not the posterior vOTC shows sensitivity to semantically-related models, in line with previous recent evidence (Lerma-Usabiaga et al., 2018); and 3) that areas that act as “semantic hubs”, such as the anterior IFG, ATL or the STG, will show especially good brain-model similarity when using models built from semantic features, including word vectors (Binder et al., 2009; Hoffman et al., 2015; J. Wang et al., 2010). 5.2. METHODS 5.2.1. Participants A total of 30 Spanish-speaking, right-handed participants (23 females) aged between 19 and 40 years old (average = 28.5 ± 6.933 years) took part in the study. All participants spoke Spanish as their first language, had normal or corrected-to-normal vision, and had no history of reported neurological or psychiatric disorders. Of the initial 32 participants, two of them were discarded due to excessive head motion. All participants received monetary compensation for their voluntary participation and gave their informed consent to take part in the study, in compliance with the regulations established by the BCBL Ethics Committee and the guidelines of the Helsinki Declaration. 5.2.2. Stimuli and Materials The stimuli employed included a total of 960 spanish words. Half of these words were tested in the MRI, and the other half were tested outside the scanner (see below). All words were nouns extracted from EsPal (Duchon et al., 2013), ranging from 4 to 10 letters in length, 73 OF 150 and including concreteness, familiarity and imageability subjective ratings. These ratings were available in EsPal, and come from normative data described elsewhere (Duchon et al., 2013). They are subjective ratings about the word, ranging from 1 to 7, where 1 means completely abstract (in the case of concreteness), completely not familiar (in the case of familiarity), or referring to an object that is completely impossible to imagine (in the case of imageability). Additionally, all words counted on objective measures of frequency of occurrence, bigram frequency, biphone frequency, number of phonological neighbours, and orthographic Levenshtein distance (OLD20) (see Yarkoni et al., 2008). Each subset of 480 words was split into 240 abstract and 240 concrete words. Each word group was also subsequently divided in 4 conditions: a) 60 words with low familiarity and low frequency; b) 60 words with low familiarity and high frequency; c) 60 words with high familiarity and low frequency; and d) 60 words with high familiarity and high frequency. The cutoff points for each of the variables (concreteness, familiarity and frequency) were based on their approximate median values. Thus, abstract concepts include words with a rating of 4.5/7 or lower, and concrete concepts include words above this value. Likewise, the cutoff value for familiarity was set to 4.5/7. For word frequency we used a Zipf (Brysbaert et al., 2018) scaled value of 3.5 as a cutoff. Although discrete groups were used for general contrasts and to simplify the univariate analyses, we counted on a continuum for all three factors of interest, thus allowing us to perform RSA and hierarchical regression analyses. The two subsets of words were matched in all variables of interest, as well as in word length and phonological neighbours. Figure 5.1 shows the distribution of the subset of words tested in the fMRI task along the continuum of the variables of interest. The behavioural subset, not shown here (but see Figure 6.1), shows a similar pattern. A total of 560 pronounceable non-words, built from the spanish words by using Wuggy (Keuleers & Brysbaert, 2010), were also employed. Of the total, 80 non-words were used in the MRI, and the remaining 480 non-words were used in the behavioural tests. Both subsets were matched in length and phonological neighbours. 74 OF 150 Figure 5.1. Distribution of the 480 words used in the fMRI task along the continuum of concreteness, familiarity and frequency (zipf). Each dot represents a word. The eight colours represent each of the discrete categories created based on the median cutoffs. All materials explained above were tested in two lexical decision tasks, one inside the scanner and another one outside the scanner. The lexical decision task allowed us to have participants reading and processing words, while permitting the registration of reaction times (RTs) and avoiding excessive overload due to complex decision making. The two tasks differed in the proportion of words and pseudowords and in the inter-trial intervals (ITI), but were otherwise identical. Since we needed a high number of observations, and RSA requires a sufficient spacing between trials (over 6 seconds), in the lexical decision task inside the scanner we used 14 % of non-word trials (80 items), for the sake of time. The “proper” lexical decision task outside the scanner (with 50% of non-words and 50% of words), acted as a control for any potential task effects derived from the differential proportion of words versus pseudowords in the lexical decision task inside the scanner. 5.2.3. Procedure Firstly, high resolution T1 images were acquired. Next, during the functional MRI BOLD sequence, participants performed the lexical decision task: they were told to pay attention to 75 OF 150 and read a series of letter strings that might or might not form real words in Spanish. They were instructed to decide whether the word exists or not by pressing the corresponding button. Button assignment was counterbalanced for all participants, and the order of presentation of the items was unique for every two participants (i.e. a total of 15 predefined counterbalanced orders) to ensure that no effects of order influence pattern similarity (Mumford et al., 2014). The stimuli were presented on the centre of the screen for 1 second, followed by a variable ITI of at least 6 seconds. This allowed the haemodynamic response to return to baseline, which allowed us to accurately model each trial separately. The task was divided into 6 identical functional runs of 11:40 minutes each. Immediately after the MRI session, participants performed the “proper” lexical decision task with the remaining 480 words and 480 non-words outside the scanner. The task was the same, except for the proportion of non-words, and the ITI, which was shorter (1 second) in the “proper” lexical decision task (longer ITIs were only required by the fMRI task). The order of the stimuli was also counterbalanced by subject. 5.2.4. MRI Data Acquisition and Preprocessing Whole-brain Images were acquired on a 3-T SIEMENS’s Magnetom Prisma-fit scanner, with 64-channel head coil, at the Basque Center on Cognition Brain and Language (BCBL). Firstly, high-resolution T1-weighted anatomical images were obtained with the following acquisition parameters: TR = 2530 ms, TE = 2.36 ms, flip angle = 7º, Field of view = 256 mm, 176 volumes per run, voxel size = 1 cubic mm. After that, 6 functional runs were acquired. Each fMRI run consisted of a multiband gradient-echo echo-planar imaging sequence with the following parameters: TR = 1000 ms, TE = 35 ms, flip angle = 56º Field of view = 210 mm, 690 volumes per run, voxel size = 2.4 cubic mm, acceleration factor = 5. The first 6 volumes of each run were removed to ensure T1-equilibration effects. The order of the trials in each run, as well as the inter-trial intervals of variable duration, were determined with an optimised algorithm designed to maximise the efficiency of the recovery of BOLD response: Optseq II (Dale, 1999). All images were preprocessed by using custom scripts based on AFNI (Cox, 1996). The T1-weighted image was skull-stripeed and co-registered to the functional images by means of linear affine transformations. Although slice timing correction was not compulsory due to simultaneous acquisition of multiple slices with multiband sequence and a short repetition time (i.e. 1000 ms), the remaining volumes were corrected for potential differences in timing of slice acquisition and realigned to the minimum outlier volume by means of 12 parameter rigid-body motion transformation. For univariate analyses, all images were 76 OF 150 normalised to the MNI152 standard space (2009Lin) by means of non-linear transformations, at a resolution of 2 cubic mm. The resulting images were smoothed with a 4-mm full width at half-maximum (FWHM) isotropic Gaussian kernel, and finally, scaled to get a mean voxel signal of 100. Regarding multivariate analyses, both RSA searchlight and RSA based on regions of interest (ROIs) were carried out. All multivariate analyses were performed in individual subject space, on unsmoothed, unscaled images obtained prior to the normalisation to the MNI space. 5.2.5. Univariate analyses Statistical analyses were performed on individual subject data using the general linear model (GLM), and then submitted to group level analyses. fMRI time series data were modelled by a series of impulses convolved with a canonical hemodynamic response function (HRF). The motion parameters for translation (i.e., x, y, and z) and rotation (i.e., yaw, pitch, and roll), along with a basic set of cosine functions that high-pass filtered the data, were included as covariates of non-interest in the GLM. Each of the 8 categorical conditions, composed of 60 trials, was modelled as an event, time-locked to the onsets of the presentation of each word that belonged to their corresponding category. The least-squares parameter estimates of the height of the best-fitting canonical gamma HRF for each categorical condition were used in pairwise contrasts. Whole-brain main effects and interactions of the variables of interest were explored by performing a 2x2x2 within-subject ANOVA (Concreteness x Familiarity x Frequency). Simple effects were explored by performing pairwise contrasts, computed as one-sample t-tests on the contrast images. Given the nested nature of the categorical conditions (see Stimuli and Materials), the main contrast images (Concreteness, Familiarity and Frequency) were obtained by merging together the corresponding categorical conditions. Thus, the concreteness contrast was obtained by performing a one-sample t-test between the activations coming from the merged 240 abstract words versus the activation of the merged 240 concrete words. Likewise, the familiarity contrast was obtained by performing a one-sample t-test between the activations of the merged 240 low-familiarity words versus the activation of the merged high-familiarity words. And finally, the frequency contrast was obtained from comparing the merged 240 low-frequency words against the merged 240 highfrequency words. After applying a false discovery rate (FDR) cluster-level correction of p < 0.01 to the uncorrected voxel-extent threshold of p < 0.001, a cluster was deemed to be significantly active if it exceeded 14 voxels. 5.2.6. RSA searchlight In order to explore the whole-brain representations of the different lexical features that are objects of interest in the present study, a searchlight-based RSA was carried out. For this, 77 OF 150 we firstly defined 6 different models based on the dissimilarity between each pair of words in their key features: a) word concreteness; b) word familiarity; c) word frequency (zipf); d) semantic features (the combination of a, b and c, plus imageability; e) phonological features (a combination of bigram frequency, biphone frequency, number of phonological neighbours, number of letters, and ; and f) word vectors (Word2Vec) recovered from several different Spanish sources (Almeida & Bilbao, 2018; Bilbao-Jayo & Almeida, 2018). To ensure that the models were expressed in as comparable measures as possible, we employed euclidean distances in the case of unique variables (concreteness, familiarity and frequency), mahalanobis distances in the case of composite phonological and semantic models (given their potential covariability) and cosine distances in the case of word vectors (given the potential influence of the size of the estimated vector). All measures were then normalised to produce a range between 0 (the items are the same) and 1 (the items are completely different). Figure 5.2 shows a graphical representation of the models (representational dissimilarity matrices, RDMs), along with a correlation plot displaying the similarity between each pair of models. Of note, the models represent the words ordered by concreteness for visualisation and consistency purposes. Figure 5.2. Graphical representation of the models employed and pairwise correlation between them. A) Representation matrix of the “simple” models. From top to bottom: Concreteness, Familiarity, Frequency. Each row/column represents a different word (trial). B) Representation matrix of the “combinatorial” models. From top to bottom: Semantic, Phonological, Word2Vec. C) Pairwise correlation matrix between the models. Of note, A and B display dissimilarities matrices, with 1 meaning absolute dissimilarity, and 0 meaning absolute similarity. In turn, C displays correlation, and therefore, 1 means perfect correlation, while 0 means absolute absence of correlation. 78 OF 150 After the preprocessing of the images, beta values for each trial (each word) were estimated with the GLM. A total of 36 parameters corresponded to polynomial terms that estimated changes in signal due to drift (6 terms per run). Another 6 regressors were motion parameters (3 rotational, 3 translational). Each word was estimated as a separate regressor by convolving the onset of the stimulus with a canonical gamma HRF. The resulting images containing the beta values were then masked to include brain voxels only. The searchlight itself was performed with custom scripts based on Python (https://github.com/AbrahamSV/simplyRSA). All possible 5 mm spheres containing at least 50% of voxels falling within the brain mask were predefined to increase computing efficiency. For each sphere, the brain similarity matrix was computed as the cosine distance between each pair of trial-specific vectorised patterns. The brain similarity matrix was then compared to each of the 6 described models by means of ranks correlation (Spearman’s rho). The result of each sphere (i.e. each centre of mass coordinates) was then read back into the brain mask, and normalised to the MNI152 template for averaging, resulting in the whole-brain representation of the similarity with each model. Since we were interested in investigating the whole-brain differences between phonological, semantic, and naturalistic models (Word2Vec), and given the collinearity between the three variables of interest with the semantic model (see Figure 5.2), only the three combinatorial models were analysed in the searchlight. The contribution of the three variables of interest was explored in ROI-based analyses. 5.2.7. ROI-based RSA In order to address the main objectives and to test the proposed hypotheses, we conducted RSA in brain areas that are key for conceptual representations associated with the variables of interest, according to the previous literature mentioned above (Binder et al., 2009; Hoffman et al., 2015; J. Wang et al., 2010). A total of 7 left-lateralised ROIs were anatomically defined by using the HCP atlas available in AFNI (Glasser et al 2016): IFG pars orbitalis, IFG pars triangularis, IFG pars opercularis, ATL, posterior STG, anterior vOTC (fusiform FG4) and posterior vOTC (fusiform FG2). As in the searchlight RSA, the brain similarity matrix was obtained for each ROI by computing the cosine similarity between each pair of trial-specific vectorised patterns. The brain similarity matrix was compared with each of the 6 models, and submitted to a bootstrap analysis of 100.000 iterations, which yielded the employed 95%CI. Each iteration was computed as the Spearman rank’s correlation between the brain RDM and a random permutation of the Word2Vec model. The resulting averaged correlation between each of the ROIs and each model was then compared to the 95th percentile of the bootstrap sample, by means of Pearson and Filon’s Z, adapted from the cocor R package (Diedenhofen 79 OF 150 & Musch, 2015). Although we were mainly interested in analysing the differences between the combinatorial models, in the ROI analyses we included the simple models in order to investigate the contribution of each of the variables of interest to the semantic model. 5.3. RESULTS 5.3.1. Behavioural Performance On average, the global accuracy for all items in the fMRI task was 95.031% ± 4.229, and the average RT for all items was 0.677 sec ± 0.084. In order to evaluate the potential relations between the three variables of interest and RTs, we performed Pearson’s correlation tests over all correct responses by each participant. Confidence intervals (CIs) for the r values are reported. Across the 30 subjects, the average correlation between Concreteness and RT was r = -0.052 ± 0.045, with a 95%CI = [-0.069, -0.036], and an averaged p = 0.309 ± 0.255 . The averaged correlation between Familiarity and RT was r = -0.284 ± 0.083, with a 95% CI = [-0.315, -0.253], and an averaged p = 0.001 ± 0.005. Finally, the average correlation between Frequency and RT was r = -0.246 ± 0.059, with a 95%CI = [-0.269, -0.224], and an averaged p = 0.001 ± 0.007. We also assessed the differences in the lexical effect between all the 8 categorical bins described above, by calculating T values and effect sizes for the contrast non-word’s RT minus each bin’s RT. Non-words tend to show slower RTs than words overall. A significant difference between non-words and any of the 8 bins would mean that words in such bin are easier to access than non-words (i.e., lexical effect). In this sense, the higher the effect size associated with the contrast, the greater the lexical effect. In turn, if the contrast does not show a significant difference, or if the effect size associated is considerably low, it can be considered that the words in the bin are as easy to access as non-words (i.e., no lexical effect). Below, we report averaged false discovery rate (FDR)-corrected p values (i.e. q-values) and averaged Cohen’s d values. The results demonstrated that all bins showed a lexical effect. However, the bins that combined highly familiar and highly frequent words showed the greatest effect sizes, with slight differences between abstract and concrete words. In contrast, bins that contained lower-familiarity and lower-frequency words showed moderate to high effect sizes, thus indicating a less pronounced lexical effect for words in these bins (see Table 5.1 below). Table 5.1. Lexical effects resulting from contrasting each of the categorical bins defined against nonwords. 86 OF 150 Figure 5.5. Results from the bootstrap analysis in the semantic hub ROIs for each of the 6 models. Rho values are shown in the X axis. FDR-corrected q values are reported for significant above-percentile correlations. Single asterisks represent correlations significantly above the 95h percentile of the bootstrap not surviving the FDR correction. Double asterisks represent significant above-percentile correlations surviving the FDR correction. The IFG pars orbitalis ROI did not show significant above-percentile correlations with any of the combinatorial models. When looking at the simple models, only the Frequency model showed a significant above-percentile correlation, which did not survive the FDR correction (Z = 1.663, q = 0.336). The IFG pars triangularis ROI displayed a significant above-percentile correlation with the Semantic model only, but it failed to reach significance after FDR correction (Z = 2.369, q = 0.093). Regarding the simple models, we found a significant correlation with Frequency (Z = 3.789, q = 0.001). In the posterior IFG pars opercularis, the correlations were significantly above the fixed percentile for the Semantic model, although this correlation did not survive the FDR correction 87 OF 150 (Z = 2.052 , q = 0.168). However, this ROI showed a significant correlation surviving the FDR corrections with the Frequency model (Z = 5.264, q < 0.001) The ATL ROI showed no significant correlations with any of the combinatorial models, or any of the simple models. Similarly, the STG ROI did not show any significant correlation with any of the combinatorial or simple models. Figure 5.6 shows the main contrasts against the 95th percentile of the bootstrap distribution for each of the vOTC ROIs (anterior FG4, posterior FG2). While the posterior vOTC did not show any significant correlation with any of the combinatorial or simple models, the anterior vOTC displayed a significant correlation with the Frequency model that became only marginally significant after FDR corrections (Z = 2.654, q = 0.055 ). Figure 5.6. Results from the bootstrap analysis in the vOTC ROIs for each of the 6 models. As in Figure 5.5, X axis represents Rho values. Single asterisks represent correlations significantly above the 95h percentile that did not survive the FDR correction. 88 OF 150 5.4. DISCUSSION In the present study, our goal was to investigate the neural representational space of several linguistic properties that are known to be associated with the access to lexico-semantic knowledge. We further sought to disentangle the interrelations between these variables, along with naturalistic language models. The main findings are discussed below. 5.4.1. Associations Between Word Properties One of our general objectives was to explore how several psycholinguistic variables, from phonology to semantics, are related in terms of brain activation. There is collinearity between word concreteness and variables like affective valence (Meersmans et al., 2020; Vigliocco et al., 2014), frequency of occurrence or subjective familiarity (Striem-Amit et al., 2018). And yet, neuroimaging studies very often investigate these variables in isolation, while ignoring the most common collinear variables. Here, we found that some of the commonly observed neural effects of word concreteness can be qualified by relevant interactions with word familiarity. Our results show that this is the case for the right parahippocampal gyrus, extending to the hippocampus, where highly familiar words showed greater activation than less familiar words, only for words on the lower end of the concreteness continuum. Being the first gate to the medial temporal lobe from the visual inputs, the parahippocampal gyrus is known to be involved in learning and semantic memory, and specifically, in the binding of items to their context (C. B. Martin & Barense, 2023). Some studies have found higher responses in the parahippocampal gyrus to more familiar visual objects (C. B. Martin et al., 2013), and to concrete versus abstract words (Binder et al., 2009; Hoffman et al., 2015). Our data replicated these results, and extended them, by showing that high familiarity evokes greater parahippocampal responses also in the linguistic domain. This effect of word familiarity was especially evident in abstract words. A potential interpretation of this interaction is related with the above-mentioned role of the parahippocampal gyrus in contextual binding. In this sense, the retrieval of familiar information could be more effortful for abstract words, given that they are known to appear in a wider variety of contexts, which makes them more ambiguous and harder to access (Hoffman et al., 2013; Schwanenflugel et al., 1988) In addition, we found a significant interaction between word frequency and familiarity. This is perhaps less surprising, given the known collinearity between these two variables (Tanaka-Ishii & Terada, 2011). Nevertheless, to our knowledge, this is the first study to further explore the interaction of these two variables at the neural level, with potential conceptual implications. The question raises then as to whether these two properties tap into the same cognitive process. Our data suggest that, albeit to a high extent collinear, the effects of familiarity and frequency add up to produce stronger lexical effects. Our behavioural data show 89 OF 150 that words became easiest to access when they are associated with both higher frequency of occurrence in media (objective frequency), and with higher rates of subjective familiarity (subjective frequency). Aligning to this result, our fMRI data show that frequency and familiarity add up to produce stronger effects in areas of the ventral language network (i.e. IFG pars orbitalis and pars triangularis), also including the left parahippocampal gyrus. While the effects of frequency have been associated to both phonological access (Carreiras et al., 2009; Fiebach et al., 2002), and semantic access (Chee et al., 2002, 2003) during word processing, the cognitive effects of familiarity have been almost exclusively associated to the ease of access to semantic knowledge (Neveu & Kaushanskaya, 2023; Shinozuka et al., 2021). Thus, given the role of the ventral reading network in active retrieval of lexico-semantic information (word meaning), it seems that the added effects of frequency and familiarity are mainly linked to semantic access, without contesting the idea that word frequency is also associated with phonological access. 5.4.2. Anterior-to-Posterior dissociation in the left IFG In line with previous findings by our group (Sánchez et al., 2023), we expected to find evidence for an anterior-to-posterior functional dissociation in the left IFG, where its anterior subdivisions would show higher similarity with semantically-related models, whereas the posterior portions of the IFG would show equivalent similarities with both semantic and phonological models. Our results partially support this notion. From the RSA searchlight, we could demonstrate that a semantic model built from psycholinguistic properties showed above threshold similarities in the whole left IFG, with the highest correlations observed in the pars triangularis. Similarly, a naturalistic language model, believed to represent fine-grained semantic relations, showed above threshold similarities in the bilateral pars orbitalis only. However, while the posterior IFG (pars opercularis) displayed above-threshold correlations with the semantic model, we did not observe significant correlations with a phonological model in this area. In turn the phonological model showed significant correlations in more posterior areas, in SMA (BA 6), potentially underlining its role in motor planning and phonological search (Carreiras et al., 2006, 2009). A possibility then is that the division of labour between anterior and posterior IFG expands beyond this area. Under this interpretation the anterior IFG is strongly associated with semantic lexical access (although not exclusively), as illustrated by its high similarity with both a psycholinguistic semantic model, and a naturalistic language model. As we move to the posterior IFG, we still observe, to a certain degree, involvement in semantic processing, although this involvement becomes less pronounced. And going further, the areas immediately posterior/dorsal to the IFG seem to be almost exclusively involved in motor planning associated with phonological processing (Hagoort, 2005, 2013). 90 OF 150 This view could in part explain the results from the ROI-based RSA. We found that only the word frequency model showed significant correlations with the brain activation pattern in the pars triangularis, and especially, in pars opercularis. Given the above-mentioned hybrid nature of word frequency (putatively conveying both semantic processing and phonological search), the fact that the pars opercularis is the area where the frequency model showed the highest significant correlations, illustrate the dynamic role of this area in both phonological and semantic processes. Nevertheless, the ROI-based analyses failed to replicate some of the findings from the searchlight (i.e., we did not find a significant correlation with semantic models in the anterior IFG). Two factors could account for this inconsistency. Firstly, the bootstrap analysis applied in the ROIs is considerably more stringent than the whole-brain searchlight, by establishing a higher threshold for a contrast to be considered significant, after the due corrections. And secondly, although we believe that the employed lexical decision task is optimal for the objectives of the study (as explained above), it might have not maximised semantic processing, for which top-down modulations have been demonstrated to be especially relevant (Sánchez et al., 2023). Future studies should investigate the relation between multivariate representations of psycholinguistic properties in the IFG and varying task demands. But notwithstanding these limitations, our results pinpoint the functional dynamics of the IFG and associated areas, and their role in the different cognitive processes that facilitate lexical access. 5.4.3. Involvement of the left anterior vOTC in lexico-semantic processing We expected to find a functional dissociation between the left anterior and posterior vOTC, where its anterior, but not posterior, subdivision would show sensitivity to semantic models. Our results corroborated this hypothesis. We found significant correlations with the semantic model around the anterior vOTC. Although with a smaller cluster size, we also found a correlation with the Word2Vec model in this area. Furthermore, we found a negative correlation with the phonological model in this anterior vOTC area. This could be indicating that phonological processing might not be as relevant in the anterior vOTC. In turn, the phonological model showed high correlations with primary and secondary visual areas, including V3 and V4. As in the IFG, a potential interpretation is that the expected gradient could extend posteriorly, where secondary visual areas would be sensitive to lower-level features of visual linguistic input. Areas V3 and V4 are known for their role in early visual processing of low-level visual features (Furlan & Smith, 2016; Pasupathy et al., 2020). The area V3 has been associated with the perception of motion related with speech recognition (Jeschke et al., 2023), and Area V4 has often been considered the earliest step for the categorisation of visual input (Okazawa et al., 2016; Pasupathy et al., 2020). Additionally, 91 OF 150 adjacent areas have been associated with spatial and action-related processing, given their co-activation with the parietal and, precisely, with the SMA (Malikovic et al., 2016). Perhaps the most compelling piece of evidence supporting this view is that the anterior vOTC showed significant (albeit marginal) correlations with word frequency, a semantically related variable, while the posterior vOTC did not. This is in line with previous recent evidence indicating a dissociation between the anterior and posterior vOTC (Lerma-Usabiaga et al., 2018; Price & Devlin, 2011; White et al., 2019). Accordingly, we would have expected that, in turn, the posterior vOTC showed certain sensitivity to phonological models. However, as it happened with the IFG, we did not find a significant correlation with the phonological model in the more posterior vOTC subdivision, and thus the results should be taken carefully. 5.4.4. Linguistic Properties vs Word Vectors in Semantic Hubs Our last specific objective was to improve our understanding of the neural effects captured by naturalistic language models, and their relation with psycholinguistic properties. In this sense, it is typically assumed that these models represent, to a high extent, semantic relations between words (Abnar et al., 2018; Grave et al., 2019; Liuzzi et al., 2023). In our study, this was corroborated by the significant correlation between a semantic model built from psycholinguistic variables and a word vector model. At the neural level, we hypothesised that brain areas that are highly involved in abstract processing and retrieval of conceptual information (sometimes referred to as semantic hubs) (Patterson et al., 2007), would show especially good brain similarity with the semantic and the Word2Vec models. This hypothesis was confirmed by the RSA searchlight. We found significant correlations in areas like the anterior IFG, the STG, or the ATL, with some differences between the semantic model and the Word2Vec model. The correlations with the semantic model were more extensive, especially in the IFG and the STG, while not being found in the ATL. In turn the Word2Vec model showed less widespread similarities in the left anterior IFG, inferior ATL and parahippocampal gyrus. This could be indicating that the Word2Vec model might be more specific to semantically related processes (see Carota et al., 2017, 2021) than the semantic model built from psycholinguistic variables. For example, the STG, an area associated both with phonological and semantic processing (Friederici, 2012; Lau et al., 2008; MacGregor et al., 2012), showed high similarities with both the phonological and the semantic model, but not with the Word2Vec model. In addition, correlations with the left inferior ATL were only found with the Word2Vec model. It should be pointed out that the phonological model yielded significant correlations in the superior ATL. A relatively recent review (see section 2.2.3) pointed out that the ATL displays a graded functional specialisation, and thus information of diverse nature is represented along the ATL (Ralph et al., 2017). In 92 OF 150 this sense, the superior ATL is connected with the auditory cortex and hence it is not surprising that phonological representations recruit this area (Zhang et al., 2024). In contrast, the similarities with the Word2Vec model were found in the inferior ATL (inferior temporal/anterior FFG), an area that might convey information of diverse nature (Ralph et al., 2017). Although these results seem promising in the attempt to disentangle the neural effects associated with naturalistic language models, it should be noted that the ROI analyses failed to replicate the findings. Once more, we can refer to the two factors explained above to account for this. Firstly, because we used a lexical decision task, semantic top-down influences are not maximised. And secondly, and adding to this, the ROI contrasts were more stringent, and thus, a subtle semantic effect arising from a lexical decision task, might have been overlooked when following ROI analyses. 5.5. CONCLUSIONS The present study offers new insights into the brain representations of lexico-semantic information. We found that some of the most commonly reported semantic effects, like the left parahippocampal responses to concreteness, might be qualified by familiarity effects. Furthermore, we found considerable support to the view that the left IFG and adjacent posterior dorsal areas (BA 6, SMA) show a functional dissociation, where anterior areas of the IFG are especially relevant for the access to lexico-semantic knowledge, while the posterior IFG participates to a lesser extent in semantic processes, and the SMA is involved in accessing and representing phonological lexical information. A somewhat different pattern was found in the vOTC, extending to the adjacent posterior areas of the primary and secondary visual cortex: while the anterior vOTC exhibited sensitivity to semantically-related properties during word reading, the posterior vOTC did not, and posterior adjacent areas V3 and V4 showed high similarity with a model representing phonological properties. Finally, we found that both naturalistic language models and semantic psycholinguistic models yield significant similarities in brain networks that hold semantic representations. But while the semantic psycholinguistic model showed a more widespread pattern, including the whole IFG and STG, the word vectors model displayed more specific activations in semantic hubs. Our results pinpoint the specific cognitive processes, from phonological to semantic, involved in the access to words in brain networks that are crucial for reading and semantic memory. 93 OF 150 CHAPTER VI. BEHAVIOURAL CORRELATES OF LEXICO-SEMANTIC REPRESENTATIONS 6.1. RATIONALE As it has been described in previous chapters, the neural effects of psycholinguistic variables observed in multiple fMRI studies, in different brain areas, may in part depend on top-down influences imposed by the demands of the task. In many instances, like in the study described in Chapter 4 (also see Meersmans et al., 2022; Soto et al., 2020), the attentional requirements of the task can be intentionally manipulated to explore the different cognitive settings under which the conditions studied elicit a given brain response. However, in other circumstances, most of these factors may play a significant role, without any explicit control over them, either because they are unknown to us, or because it is costful to include such manipulations and/or controls (as was the latter case of the study presented in Chapter 5, due to the time constraints). If we focus only on the behavioural aspects of a cognitive task, decision making itself is influenced by a considerably extensive variety of factors. Some of them offer little information about the cognitive processes of interest, and are unavoidable, like noise or purely motor response planning; others might be variables of interest, and can be harnessed to better understand a given behaviour, like individual differences and the manipulations that alter the cognitive processes induced by the task itself. Whilst most statistical approaches have the potential to assess the influence of known and controlled variables of no interest, it is usually hard to tease apart those unknown noise factors that influence a given decision. This has been the starting point for, and main reason to use an approach designed to better account for decision making processes, the drift diffusion model (DDM) (Ratcliff, 1978; Ratcliff et al., 2016; Ratcliff & McKoon, 2008). This group of computational models try to infer latent cognitive processes from the observable behaviour (usually represented by trial-to-trial RTs) during decision making in cognitive tasks (Myers et al., 2022). Essentially, the DDM simulates the time it takes to reach a decision (or a boundary), in one way or the other, when a choice needs to be made. In its most common application, four parameters are defined that represent different aspects of the evidence accumulation process in order to make a decision: 1) the boundary separation (a) indicates the distance from one boundary (response A) to the other (response B), with greater boundary separation indicating a higher necessary amount of evidence required to reach a decision; 2) the starting point (z) represents the initial state in the decision making process, and can be closer to either boundary if the conditions favour one or other response; 3) the drift, which gives name to the approach, represents the speed at which evidence is accumulated to reach a boundary, with larger positive values indicating faster evidence accumulation towards the upper boundary (response A), and larger negative values 94 OF 150 indicating faster evidence accumulation towards the lower boundary (response B); and 4) the non-decision time represents the time it takes for the nervous system to give a motor response, once the decision boundary is reached (Myers et al., 2022; Shinn et al., 2020). Importantly, all these parameters try to build a simulation that best fits the RTs observed in a given task in a trial-to-trial manner. It is common to define different drifts that represent the different ability of the conditions in the study to affect decision time (Myers et al., 2022). Other more sophisticated models include dynamically changing parameters in order to improve the predictions, and thus the fit of the model (Shinn et al., 2020). We can harness this type of computational modelling to predict the variables that contribute the most to the decision making process. Additionally, by customising the model parameters, we can explore how complex the process is, and to what extent RTs are “inflated” by this complexity in the decision making (e.g., Mueller et al., 2017). This can be especially relevant when it comes to accounting for the brain activation patterns that are obtained under certain cognitive manipulations, and for estimating how influenced they are by top-down influences not directly related with the cognitive process studied. Moreover, DDMs offer the possibility to compare the correlations between the drifts yielded by the models and brain activity patterns. This can reveal whether the ability of any set of variables to influence the decision making process is indeed associated with the patterns of brain activation observed in response to those variables. This is why I decided to employ DDMs to further analyse the performance in the lexical decision tasks described in Chapter 5, both inside and outside the scanner (referred to as fMRI and Behavioural respectively from this point on). Since manipulating the task (as in Chapter 4) was not possible in the study described in Chapter 5, part of the cognitive processes associated with the neural representations observed could not be directly accounted for. Specifically, as described above, a limitation of the previous study was that it was unclear whether the lexical decision task employed could be limiting the extent of semantic processing, or on the contrary, boosting sublexical processing by making participants read the words in a superficial and automated way. DDMs offer a direct way of comparing the ability of different combinations of parameters to better predict the observed RTs. It also has the potential to account for non-decision time in a trial-wise manner. This could also allow us to examine the correlation between the “pure” lexical decision time (not influenced by effects of no interest in the responses given), and the neural effects observed in Chapter 5. The general objective of this chapter was to examine the degree to which sublexical (bigram frequency, number of letters, and orthographic distance) and/or lexico-semantic (word frequency, familiarity, and concreteness) properties, or any combination of these variables, are capable of influencing the decision making process in lexical decision tasks. Additionally, this chapter has the specific objective of analysing the link between fine-grained behavioural 95 OF 150 measures and the neural effects observed. In particular, we sought to examine the association between the model-brain similarities from Chapter 5 on the one hand, and the ability of different semantic and sublexical psycholinguistic variables to influence decision making on the other. To give an answer to the general objective, we will compare the fit of DDMs representing different combinations of the mentioned variables of interest. The specific objective will be addressed by obtaining the best-fitting drift values of the semantic and sublexical DDMs in order to estimate their association with brain similarities obtained from the RSA analysis, which embody the neural representations associated with semantic or sublexical processes described in the previous chapter. We expect that a DDM built from lexico-semantic properties (word frequency, familiarity and concreteness) will be the best DDM at predicting RTs after accounting for model complexity, and that word frequency will be the most significant contributor to the drift rate. This result would indicate that, even in a lexical decision task, lexical processing of semantic properties emerges, as our results from previous chapters seem to indicate. In addition, we hypothesised that, if the neural effects observed in Chapter 5 are explained by lexical decision making, then we should expect significant correlations between the drift values from the DDMs and brain-model similarities. We expect the highest correlations between the best-fitting semantic drifts and the neural representations in those areas that showed significant similarities with the semantic and word vector models (anterior IFG, anterior vOTC and ATL). At the same time, if overall RTs are associated with RSA model similarities, we would expect equally high correlations between model similarities and overall RTs in those areas that are more heavily influenced by complex decision-making, like the posterior IFG, and the vOTC. In contrast, if RSA similarities are unbiased by raw RTs, then we should not expect any significant correlations between this variable and similarity values. 6.2. METHODS 6.2.1. Participants The participants analysed here were the same 30 described in the previous chapter (see section 5.2.1). All participants spoke Spanish as their first language and had no reported history of neurological or psychiatric illnesses. All of them received monetary compensation for their voluntary participation and gave their informed consent to take part in the study, in compliance with the regulations established by the BCBL Ethics Committee and the guidelines of the Helsinki Declaration. 102 OF 150 Figure 6.3. A) Distribution of RTs (in seconds) across subjects in response to words in the Behavioural (left) and fMRI (right) tasks for correct (in blue) and incorrect (in red) responses. B) Distribution of RTs (in seconds) across subjects in response to nonwords in the Behavioural (left) and fMRI (right) tasks for correct (in blue) and incorrect (in red) responses. 6.3.2. Drift-Diffusion Results Subsequently, fit values of the 7 DDMs built were checked for differences between the different models in the fMRI lexical decision task. The model that showed the best fit, after controlling for the number of parameters included, was the Semantic DDM (AIC=-579.315), followed by the Familiarity DDM (AIC=-565.096) and the Frequency DDM (AIC=-549.906). The Concreteness DDM (AIC=-491.970) still performed better than the two Random DDMs, but considerably worse than the Semantic, Familiarity and Frequency DDMS. LRT tests further indicated that the Semantic DDM performed better than the Frequency (p<.001), Familiarity (p=.058) and Concreteness (p<.001) DDMs. Lastly, the Sublexical DDM (AIC=-444.727) did not perform better than the Random 3 DDM (AIC=-464.929). As for nonwords, the overall fit of the model was considerably lower (mean AIC=- 15.227). It should be considered that the drift in this model was kept fixed for all trials, as 103 OF 150 opposed to the rest of the models, whose drifts dynamically varied with the variables of interest. Figure 6.4 shows the AIC values, averaged across subjects, of each of the 7 word models tested and the nonwords model. Figure 6.4. A) Fit of the DDMs in the fMRI lexical decision task, based on the transformation of NLL values to AIC. Lower values indicate better fit. Notice that the Nonwords model is estimated differently (fixed drift) than the rest of the models (dynamically changing drift). B) Distribution of the best-fitting drifts from the Semantic DDM across subjects. C) Distribution of the best-fitting drifts from the Sublexical DDM. The best-fitting drifts of the Semantic DDM were also analysed in order to further examine the contributions of each of the variables to the decision making process. Consistent with the analysis of the AIC values, the highest contributor to the drift of the Semantic DDM was Familiarity (mean drift=1.413), followed by Frequency (mean drift=1.144). Concreteness showed drifts that were close to zero (mean drift=0.023), indicating that this variable had little influence on the overall drift of the model. Additionally, the best-fitting drifts of the Sublexical DDM were explored. Although the fit of this model was not better than the Random 3 DDM, the drifts can provide additional information about the contribution of the variables included in the model. Number of letters drift values were considerably high (mean drift=1.257), being the only variable in the model 104 OF 150 whose drifts significantly differed from zero. Of note, Orthographic distance (OLD20) drifts showed high variability across subjects. 6.3.3. Correlations between DDMs and Brain Representations We systematically inspected the correlations between each of the drifts obtained from the DDM analyses in the fMRI task (see Figure 6.4[B,C]) and the ROI-based RSA similarities yielded by the semantic, phonological and word2vec models. Of all seven ROIs, only the IFG pars orbitalis, the ATL and FG2 yielded significant correlations surviving the FDR correction. Within these areas, we only observed associations between some of the semantic DDM drifts and the semantic RSA model. The results are described below, and represented in Figure 6.5. Figure 6.5. Significant RSA-DDM fMRI drift correlations surviving the FDR correction for an alpha of .05. Each dot represents a subject. In IFG pars orbitalis, the Familiarity drift was positively associated with the semantic RSA model similarities (r=0.563, q=.013), while the Concreteness drift was negatively 105 OF 150 correlated with the same semantic RSA model (r=-0.589, q=.013). The hierarchical linear regression analysis in this ROI indicated that a model with Familiarity drift alone explained 25.3% of the variance (R2=0.253, F=9.162, p=.005), with Familiarity drift significantly predicting RSA similarity (𝛽=.009, p=.005, CI=[.003,.015]). Adding Frequency drift to the model helped explain up to 30.2 % of the variance (R2=0.302, F=5.615, p=.009), with Familiarity still being a significant predictor of RSA similarity (𝛽=.007, p=.036), but with Frequency not being a significant predictor (𝛽=-.003, p=.192). Finally, including Concreteness drift in the model summed up to 32.4% of the variance explained (R2=324, F=3.997, p=.018). However, in this model, none of the drifts were significant predictors of RSA similarity (Familiarity: 𝛽=.005, p=.203; Frequency: 𝛽=-.001., p=.644; Concreteness: 𝛽=-.005, p=.370). This was not likely due to multicollinearity, given the considerably low VIF values associated with the three drifts (Familiarity=1.787, Frequency=2.014, Concreteness=2.931). The followup ANOVA indicated that the above-mentioned increases in the variance explained did not reach significance after the additions of Frequency drift (F=1.786, p=0.192) or Concreteness drift (F=0.833, p=0.370) to the model. In the ATL, the Concreteness drift was negatively correlated with the semantic RSA model similarities (r=-0.582, q=0.017). The regression model with Concreteness drift alone explained 29.6% of the variance (R2=0.296, F=11.37, p=.002), with Concreteness drift as a significant predictor of RSA similarity (𝛽=-.009, p=.002, CI=[-.015, -.004]). Adding Familiarity drift to the model did not contribute to the explained variance (R2=0.297, F=5.489, p=.01). Finally, after adding Frequency drift to the model, the variance explained increased to 31.6% (R2=0.316, F=3.855, p=.02), but Concreteness drift was the only significant predictor of the RSA similarity (Concreteness: 𝛽=-.012, p=.018, CI=[-.023, -.002]; Familiarity: 𝛽<-.001, p=.828, CI=[-.008, .006]; Frequency: 𝛽=.002, p=.408, CI=[-.004, .009]). The follow-up ANOVA indicated that the observed increase in the variance explained after adding Frequency drift as a third predictor was not significant (F=0.709, p=.407). As for the FG2 ROI, Concreteness drift was also negatively correlated with the semantic RSA model similarity (r=-0.549, q=.037). In the subsequent hierarchical linear regression analysis, a model with Concreteness drift alone explained 28.1% of the variance (R2=0.281, F=10.54, p=.003), with Concreteness drift as a significant predictor of RSA similarity (𝛽=-.012, p=.003, CI=[-.021, -.005]). After adding Familiarity drift, the model explained 30.1% of the variance (R2=0.301, F=5.607, p=.009), although with Concreteness as the only marginally significant predictor of RSA similarity (Concreteness: 𝛽=-.010, p=.076, CI=[-.021, .001]; Familiarity: 𝛽=.004, p=.389, CI=[-.005, .014]). Finally, adding Frequency drift to the model did not increase the variance explained by the model (R2=0.303, F=3.625, 106 OF 150 p=.026). The follow-up ANOVA revealed that the increase in the variance explained after adding Familiarity drift was not statistically significant (F=0.74, p= .397). We also explored the associations between raw RTs and RSA model similarities. In this sense, although we observed significant or marginally significant negative correlations in the mid/posterior IFG (pars opercularis: r=-0.334, p=.075; pars triangularis: r=-0.37, p=.047), these correlations did not survive the FDR correction (pars opercularis: q=.264; pars triangularis: q=.264). No additional effects were found in any of the ROIs, for any of the RSA models. 6.4. DISCUSSION In this work, we aimed to quantify the degree to which different psycholinguistic properties are able to influence the decision-making process in two different lexical decision tasks, under different timing and environmental conditions, and with a fixed order. We addressed this by employing Drift-Diffusion Model (DDM) analyses. This approach additionally served the specific aim of assessing the link between the ability of lexico-semantic and sublexical properties to influence decision-making (represented by the drifts of the DDMs), and semantic and phonological brain representations found in Chapter 5. The main findings are discussed below. 6.4.1. Psycholinguistic Properties and Decision-Making In our previous study, we could not ensure that having participants simply decide whether a given string of letters forms a word that exists in their language or not could effectively trigger lexico-semantic processing (i.e., the processing of words as a whole and their meaning). Because we were aware of this potential limitation, we sought to examine the ability of sublexical vs. lexico-semantic properties to influence behaviour during lexical decisions. As we expected, we found that lexico-semantic, more so than sublexical properties, drove most of the evidence accumulation rate towards reaching a decision in two different lexical decision tasks. A DDM based on word-by-word Familiarity, Frequency and Concreteness was the best model to account for evidence accumulation rate, even after controlling for model complexity. In contrast, a DDM built from sublexical variables like the number of letters, orthographic distance and bigram frequency, did not perform better than a DDM built from three randomly generated variables. The most relevant contributor to the decision rate was word Familiarity, followed by word Frequency, variables that are inevitably associated with lexical processing of meaning (Chee et al., 2002, 2003; Neveu & Kaushanskaya, 2023; Shinozuka et al., 2021). In this sense, these two variables sum up to contribute to deciding that a string is a real word. Although intuitively one would expect that 107 OF 150 this type of task should be performed without the need of semantic processing, it is apparent from our results that lexico-semantic properties are triggered automatically in typical readers, even in a seemingly simple linguistic task. It has been typically assumed that lexical access might not be strictly required to perform a lexical decision task, and that it rather relies on automatic visual word recognition (Coltheart, 2004; Coltheart et al., 2022). It should be noted that our results are not taken as a direct indication that lexical access is a unitary process. Just as word Frequency is likely to tap into both phonological access (Carreiras et al., 2009; Fiebach et al., 2002), and semantic access (Chee et al., 2002, 2003), it is possible that both word Familiarity and word length affect lexical decisions simultaneously. In the present work, within a sublexical model built from phonological and orthographic features, a variable like the number of letters constituting the word was likely to influence the evidence accumulation rate during lexical decision, given the observed large drift values associated with it. Although we divided sublexical and lexicosemantic variables into two separate models, it is possible that both kinds of variables converge to influence the ease with which a lexical decision is reached. This interpretation would be in line with previous recent findings indicating that both sublexical and lexicosemantic properties influence the spatiotemporal neural dynamics of lexical processing (Woolnough et al., 2021). The thorough exploration of this complex spatiotemporal interaction, however, was beyond the scope of this work, and suffice it to say that lexico-semantic properties such as word Familiarity and Frequency, were the most important determinants of the observed lexical decision times, followed by a sublexical variable like word length. Our findings were replicated in two different versions of the same task (inside the MRI vs. outside the scanner), which varied in the timing conditions and even in the proportion of nonword stimuli, and each tested in a different set of words. Hence, the influence of lexicosemantic processing during lexical decision-making became apparent in our results. 6.4.2. Association between Drift Rate and Brain Representations Interpreting the associations between neural effects and the observed behaviour is oftentimes a challenging endeavour. In Chapter 5, we described how brain areas such as the anterior IFG, the ATL or the anterior vOTC, are recruited for semantic representations, as illustrated by their representational similarity with psycholinguistic semantic and word vector RSA models. In this work, we aimed to go a step further and inspect whether these representations were associated with semantic drifts, which embody the ability of lexicosemantic properties to influence decision making. If such associations are greater than any potential link between raw RTs and the neural representations, then such finding would further support the interpretation of these neural effects in terms of semantic brain representations. We hypothesised that semantic and word vector representations would show high correlations 108 OF 150 with semantic drifts in the anterior IFG, ATL and anterior vOTC, while raw RTs would also show significant correlations in the posterior IFG and the vOTC. Our results partially supported these expectations. In the anterior IFG, Familiarity drift arose as a good predictor of brain similarities with a semantic model built from psycholinguistic properties, but not with a model built from word vectors. Word Familiarity is a subjective measure of the frequency with which individuals are exposed to conceptual information. It has been associated with semantic access, word learning and memory (Neveu & Kaushanskaya, 2023), and it has been found to elicit IFG responses that could be linked to the level of internalisation of lexical content (Shinozuka et al., 2021). In this sense, our data seem to indicate that it is these characteristics that mainly drive semantic representations in an area that is tightly associated with access to semantics such as the anterior IFG. Nevertheless, it should be noted that this finding was not replicated when using a word vector model, and hence it should be taken cautiously. Although Frequency and Concreteness drifts slightly improved the predictions of semantic similarities in the anterior IFG, the contribution of these two variables was limited. On the one hand, as indicated above, because word Frequency taps into sublexical and lexico-semantic processes, it is not surprising that its contribution to predict semantic brain representations was modest. On the other hand, in our study, word Concreteness seemed to interact with word Familiarity at the neural level (see section 5.4.1), but it did not seem to have a relevant impact on decision making in our tasks. This fact could explain why it did not arise as a good predictor of semantic representations in the anterior IFG. Potentially, this might be indicating that when including other factors that better capitalise on the frequency of exposure to information, a key feature in the modulation of the anterior IFG function (as shown in Chapter 4), word Concreteness becomes less relevant for lexical access. A finding that is more challenging to interpret was the negative correlation between Concreteness drift and semantic brain representations in the anterior IFG, ATL and the posterior vOTC. Drift can adopt negative values, meaning, in this scenario, that the higher the specific drift variable (i.e. Concreteness), the faster the response towards the lower end (i.e., faster evidence accumulation to the “nonword” response). In other words, our study included participants for whom highly concrete words were harder to recognise as words. In many cases, this kind of profile was associated with higher semantic similarities in the mentioned ROIs. A possible explanation is that in these cases, Concreteness is not really affecting semantic processing, but instead reflecting sublexical processing associated with specific features of abstract words. For instance, many abstract words share terminations like -ción (- tion in English) or -dad (-ity in English), which might have helped in recognising these strings as words. If this was the case, then it is plausible that in those participants with negative Concreteness drifts, the semantic RSA model was performing better because Concreteness 109 OF 150 did not compete with the other two semantic variables, thus leading to a clearer, less variable effect over neural representations. This interpretation would be consistent with our observation that, when trying to predict semantic similarities with Concreteness as the starting variable in the hierarchy of a linear regression model, the similarity baseline (represented by the constant term) was greatly different from 0. Moreover, this possibility also aligns with the abovementioned view of lexical access as a dynamic process, in which lexico-semantic and sublexical mechanisms interplay, each taking over under specific circumstances (e.g., Soto et al., 2020). Finally, we expected to observe significant correlations between raw RTs and model similarities in the posterior IFG and the vOTC, owing to the influence of decision-making not related to lexical processing (i.e., noise). Although this was the case in the posterior IFG, such correlations did not survive the multiple comparisons corrections. It is possible that brain representations obtained from multivariate analyses do not exactly reflect factors that affect BOLD signal intensity per se (Haynes & Rees, 2006), such as decision-making in complex scenarios. In turn, RSA is better suited for exploring complex conceptual structures (Frisby et al., 2023), while being less sensitive to single factors (Lewis-Peacock & Norman, 2014). Nevertheless, our findings seem to indicate that the brain representations found in Chapter 5 were unlikely to be due to noise attributed to complex decision-making. Some limitations should be mentioned. Firstly, the timing and hardware differences between the fMRI and the behavioural tasks did not allow a balanced, comprehensive comparison between the results of both tasks. As a consequence, all such comparisons were merely descriptive, and lack the potential to draw any kind of inference. Secondly, although we explored different combinations of psycholinguistic properties into a sublexical and a lexico-semantic model, hybrid models that incorporated both types of variables were not further explored. From our results, it seems that lexical access is not a unitary process, and although it is heavily influenced by lexico-semantic properties in typical readers, some sublexical word properties, like word length, seem to interplay in this mechanism. Future studies should further explore this interpretation. And lastly, we decided to employ the ROIs used in Chapter 5 for a clear interpretation of the effects. However, because these ROIs do not necessarily overlap with the supra-threshold similarities found in the whole-brain searchlight, it is possible that some of the associations between brain representations and the drifts obtained in this work were overlooked. Similarly, some of the effects that spanned outside these ROIs, like those found in the parahippocampal gyrus, were not explored. Nevertheless, our decision of employing the pre-built ROIs was based on our aims and hypotheses, and these questions may be addressed in future analyses. 110 OF 150 6.5. CONCLUSIONS We found robust evidence that lexico-semantic processing can be triggered during a seemingly simple lexical decision task. In fact, lexico-semantic properties, and especially, word Familiarity and word Frequency, arose as the most evident contributors to decision making during lexical decisions. However, other sublexical variables, like word length, seemed to interplay with this lexico-semantic processing to influence lexical access in our study. At the neural level, the ability of word Familiarity to influence lexical access was found to be tightly associated with semantic representations in the anterior IFG. Although other semantic properties, such as word Concreteness, seem to drive part of the semantic representations in areas like the ATL or the vOTC, these associations were possibly due to spurious interactions with sublexical processing, and hence should be further investigated in future studies. In sum, our results highlight the dynamic nature of sublexical and semantic lexical processing mechanisms, and the role of the anterior IFG in semantic access during reading. 111 OF 150 GENERAL DISCUSSION The general objective of this thesis was to give a comprehensive view of the neural underpinnings of the access and use of conceptual information during reading. Specifically, the thesis aimed to improve our understanding of how, from perceptual linguistic information, our brain enables the access to lexical conceptual information, and how it uses it to adapt to the task at hand. To this end, I investigated sublexical to lexico-semantic measurable word properties and their associated brain activation patterns and representations. We can highlight two main contributions. Firstly, the identification of functional dissociations within (and beyond) brain areas that are key in the access to lexico-semantic information, like the IFG (extending to motor planning areas such as the SMA) and vOTC (extending to the IOG), which are dynamically recruited by sublexical and lexico-semantic processes during lexical search. And secondly, this thesis underscores the importance of using psycholinguistic analyses and natural language processing in combination to disentangle neural representations of semantic information in convergence areas like the IFG, ATL and the vOTC. These key findings are discussed in the context of the most recent evidence available. Functional Dissociations in the IFG and vOTC and their Dynamic Nature The idea that the different subdivisions of the left IFG are recruited differentially for specific language and memory processes was put forward nearly two decades ago. In reading, the activity of the anterior IFG (pars orbitalis) has been repeatedly associated with access to meaning and to retrieval of information that is not readily available, as when we process unfamiliar words (Badre & Wagner, 2005; Hagoort, 2005). The areas posterior to the anterior IFG have been associated with the selection among the competing information that is available (Badre & Wagner, 2007), as it is the case of syntactic processing associated with the middle IFG (pars triangularis), or phonological selection and motor planning implicating the posterior IFG (pars opercularis) (Hagoort, 2013). More recently, different views have been proposed that support a rather interactive (and not so hermetic) framework, in which different systems (like declarative memory and language comprehension) interplay (Roger, Banjac, et al., 2022), and rely on overlapping large-scale networks (Fedorenko et al., 2024; Roger, Banjac, et al., 2022). This includes the IFG, a complex structure in which diverse functions, from declarative memory to language production, or action perception, convey (Friederici, 2023). In this sense, several key networks (salience network, SAL; frontoparietal network, FPN; and default mode network, DMN) converge in the IFG, and with this convergence, it is plausible that functions that pertain to phonology and articulation, syntactic and semantic processing, as well as monitoring and control of input and output linguistic processing also convey (see Roger, Banjac, et al., 2022 for a detailed review). 118 OF 150 Hierarchical Organization of the Prefrontal Cortex. Journal of Cognitive Neuroscience, 19(12), 2082–2099. https://doi.org/10.1162/jocn.2007.19.12.2082 Badre, D., Poldrack, R. A., Paré-Blagoev, E. J., Insler, R. Z., & Wagner, A. D. (2005). Dissociable Controlled Retrieval and Generalized Selection Mechanisms in Ventrolateral Prefrontal Cortex. Neuron, 47(6), 907–918. https://doi.org/10.1016/j.neuron.2005.07.023 Badre, D., & Wagner, A. D. (2005). Frontal Lobe Mechanisms that Resolve Proactive Interference. Cerebral Cortex, 15(12), 2003–2012. https://doi.org/10.1093/cercor/bhi075 Badre, D., & Wagner, A. D. (2006). Computational and neurobiological mechanisms underlying cognitive flexibility. Proceedings of the National Academy of Sciences, 103(18), 7186–7191. https://doi.org/10.1073/pnas.0509550103 Badre, D., & Wagner, A. D. (2007). Left ventrolateral prefrontal cortex and the cognitive control of memory. Neuropsychologia, 45(13), 2883–2901. https://doi.org/10.1016/j.neuropsychologia.2007.06.015 Banaszkiewicz, A., Matuszewski, J., Bola, Ł., Szczepanik, M., Kossowski, B., Rutkowski, P., Szwed, M., Emmorey, K., Jednoróg, K., & Marchewka, A. (2021). Multimodal imaging of brain reorganization in hearing late learners of sign language. Human Brain Mapping, 42(2), 384–397. https://doi.org/10.1002/hbm.25229 Barnhart, A. S., & Goldinger, S. D. (2015). Orthographic and phonological neighborhood effects in handwritten word perception. Psychonomic Bulletin & Review, 22(6), 1739– 1745. https://doi.org/10.3758/s13423-015-0846-z Bašnáková, J., Weber, K., Petersson, K. M., Van Berkum, J., & Hagoort, P. (2014). Beyond the Language Given: The Neural Correlates of Inferring Speaker Meaning. Cerebral Cortex, 24(10), 2572–2578. https://doi.org/10.1093/cercor/bht112 Beauchamp, M. S., Argall, B. D., Bodurka, J., Duyn, J. H., & Martin, A. (2004). Unraveling multisensory integration: Patchy organization within human STS multisensory cortex. Nature Neuroscience, 7(11), 1190–1192. https://doi.org/10.1038/nn1333 119 OF 150 Bedny, M., & Thompson-Schill, S. L. (2006). Neuroanatomically separable effects of imageability and grammatical class during single-word comprehension. Brain and Language, 98(2), 127–139. https://doi.org/10.1016/j.bandl.2006.04.008 Benner, J., Reinhardt, J., Christiner, M., Wengenroth, M., Stippich, C., Schneider, P., & Blatow, M. (2023). Temporal hierarchy of cortical responses reflects core-beltparabelt organization of auditory cortex in musicians. Cerebral Cortex, 33(11), 7044– 7060. https://doi.org/10.1093/cercor/bhad020 Bilbao-Jayo, A., & Almeida, A. (2018). Automatic political discourse analysis with multi-scale convolutional neural networks and contextual data. International Journal of Distributed Sensor Networks, 14(11), 155014771881182. https://doi.org/10.1177/1550147718811827 Bilecen, D., Seifritz, E., Scheffler, K., Henning, J., & Schulte, A.-C. (2002). Amplitopicity of the Human Auditory Cortex: An fMRI Study. NeuroImage, 17(2), 710–718. https://doi.org/10.1006/nimg.2002.1133 Binder, J. R., & Desai, R. H. (2011). The neurobiology of semantic memory. Trends in Cognitive Sciences, 15(11), 527–536. https://doi.org/10.1016/j.tics.2011.10.001 Binder, J. R., Desai, R. H., Graves, W. W., & Conant, L. L. (2009). Where Is the Semantic System? A Critical Review and Meta-Analysis of 120 Functional Neuroimaging Studies. Cerebral Cortex, 19(12), 2767–2796. https://doi.org/10.1093/cercor/bhp055 Binney, R. J., Parker, G. J. M., & Lambon Ralph, M. A. (2012). Convergent Connectivity and Graded Specialization in the Rostral Human Temporal Lobe as Revealed by Diffusion-Weighted Imaging Probabilistic Tractography. Journal of Cognitive Neuroscience, 24(10), 1998–2014. https://doi.org/10.1162/jocn_a_00263 Borghi, A. M., & Binkofski, F. (2014). Words as Social Tools: An Embodied View on Abstract Concepts. Springer New York. https://doi.org/10.1007/978-1-4614-9539-0 Bouvier, S. E., & Engel, S. A. (2006). Behavioral Deficits and Cortical Damage Loci in Cerebral Achromatopsia. Cerebral Cortex, 16(2), 183–191. https://doi.org/10.1093/cercor/bhi096 120 OF 150 Braunsdorf, M., Blazquez Freches, G., Roumazeilles, L., Eichert, N., Schurz, M., Uithol, S., Bryant, K. L., & Mars, R. B. (2021). Does the temporal cortex make us human? A review of structural and functional diversity of the primate temporal lobe. Neuroscience & Biobehavioral Reviews, 131, 400–410. https://doi.org/10.1016/j.neubiorev.2021.08.032 Brett, M., Anton, J.-L., Valabregue, R., & Poline, J.-B. (2002). Region of interest analysis using an SPM toolbox. 1. Brouwer, G. J., & Heeger, D. J. (2009). Decoding and Reconstructing Color from Responses in Human Visual Cortex. The Journal of Neuroscience, 29(44), 13992–14003. https://doi.org/10.1523/JNEUROSCI.3577-09.2009 Bruno, J. L., Zumberge, A., Manis, F. R., Lu, Z.-L., & Goldman, J. G. (2008a). Sensitivity to orthographic familiarity in the occipito-temporal region. NeuroImage, 39(4), 1988– 2001. https://doi.org/10.1016/j.neuroimage.2007.10.044 Bruno, J. L., Zumberge, A., Manis, F. R., Lu, Z.-L., & Goldman, J. G. (2008b). Sensitivity to orthographic familiarity in the occipito-temporal region. NeuroImage, 39(4), 1988– 2001. https://doi.org/10.1016/j.neuroimage.2007.10.044 Brysbaert, M., Mandera, P., & Keuleers, E. (2018). The Word Frequency Effect in Word Processing: An Updated Review. Current Directions in Psychological Science, 27(1), 45–50. https://doi.org/10.1177/0963721417727521 Bullmore, E., & Sporns, O. (2009). Complex brain networks: Graph theoretical analysis of structural and functional systems. Nature Reviews Neuroscience, 10(3), 186–198. https://doi.org/10.1038/nrn2575 Cabeza, R., Mazuz, Y. S., Stokes, J., Kragel, J. E., Woldorff, M. G., Ciaramelli, E., Olson, I. R., & Moscovitch, M. (2011). Overlapping Parietal Activity in Memory and Perception: Evidence for the Attention to Memory Model. Journal of Cognitive Neuroscience, 23(11), 3209–3217. https://doi.org/10.1162/jocn_a_00065 Cao, F., Khalid, K., Zaveri, R., Bolger, D. J., Bitan, T., & Booth, J. R. (2010). Neural correlates of priming effects in children during spoken word processing with 121 OF 150 orthographic demands. Brain and Language, 114(2), 80–89. https://doi.org/10.1016/j.bandl.2009.07.005 Caplan, J. B., & Madan, C. R. (2016). Word Imageability Enhances Association-memory by Increasing Hippocampal Engagement. Journal of Cognitive Neuroscience, 28(10), 1522–1538. https://doi.org/10.1162/jocn_a_00992 Carota, F., Kriegeskorte, N., Nili, H., & Pulvermüller, F. (2017). Representational Similarity Mapping of Distributional Semantics in Left Inferior Frontal, Middle Temporal, and Motor Cortex. Cerebral Cortex, 27(1), 294–309. https://doi.org/10.1093/cercor/bhw379 Carota, F., Nili, H., Pulvermüller, F., & Kriegeskorte, N. (2021a). Distinct fronto-temporal substrates of distributional and taxonomic similarity among words: Evidence from RSA of BOLD signals. NeuroImage, 224, 117408. https://doi.org/10.1016/j.neuroimage.2020.117408 Carota, F., Nili, H., Pulvermüller, F., & Kriegeskorte, N. (2021b). Distinct fronto-temporal substrates of distributional and taxonomic similarity among words: Evidence from RSA of BOLD signals. NeuroImage, 224, 117408. https://doi.org/10.1016/j.neuroimage.2020.117408 Carreiras, M., Armstrong, B. C., Perea, M., & Frost, R. (2014). The what, when, where, and how of visual word recognition. Trends in Cognitive Sciences, 18(2), 90–98. https://doi.org/10.1016/j.tics.2013.11.005 Carreiras, M., Mechelli, A., & Price, C. J. (2006). Effect of word and syllable frequency on activation during lexical decision and reading aloud. Human Brain Mapping, 27(12), 963–972. https://doi.org/10.1002/hbm.20236 Carreiras, M., Riba, J., Vergara, M., Heldmann, M., & Münte, T. F. (2009). Syllable congruency and word frequency effects on brain activation. Human Brain Mapping, 30(9), 3079–3088. https://doi.org/10.1002/hbm.20730 Catani, M., Jones, D. K., & Ffytche, D. H. (2005). Perisylvian language networks of the human brain. Annals of Neurology, 57(1), 8–16. https://doi.org/10.1002/ana.20319 122 OF 150 Celesia, G. G. (2005). Chapter 2 Anatomy and physiology of the visual pathways and cortex. In G. G. Celesia (Ed.), Handbook of Clinical Neurophysiology (Vol. 5, pp. 45–63). Elsevier. https://doi.org/10.1016/S1567-4231(09)70199-6 Chao, L. L., Weisberg, J., & Martin, A. (2002). Experience-dependent Modulation of Category-related Cortical Activity. Cerebral Cortex, 12(5), 545–551. https://doi.org/10.1093/cercor/12.5.545 Chee, M. W. L., Hon, N. H. H., Caplan, D., Lee, H. L., & Goh, J. (2002). Frequency of Concrete Words Modulates Prefrontal Activation during Semantic Judgments. NeuroImage, 16(1), 259–268. https://doi.org/10.1006/nimg.2002.1061 Chee, M. W. L., Westphal, C., Goh, J., Graham, S., & Song, A. W. (2003). Word frequency and subsequent memory effects studied using event-related fMRI. NeuroImage, 20(2), 1042–1051. https://doi.org/10.1016/S1053-8119(03)00335-5 Chen, R., Wang, F., Liang, H., & Li, W. (2017). Synergistic Processing of Visual Contours across Cortical Layers in V1 and V2. Neuron, 96(6), 1388-1402.e4. https://doi.org/10.1016/j.neuron.2017.11.004 Chen, Y., Davis, M. H., Pulvermüller, F., & Hauk, O. (2013). Task modulation of brain responses in visual word recognition as studied using EEG/MEG and fMRI. Frontiers in Human Neuroscience, 7. https://doi.org/10.3389/fnhum.2013.00376 Chen, Y., Davis, M. H., Pulvermüller, F., & Hauk, O. (2015). Early Visual Word Processing Is Flexible: Evidence from Spatiotemporal Brain Dynamics. Journal of Cognitive Neuroscience, 27(9), 1738–1751. https://doi.org/10.1162/jocn_a_00815 Chiarello, C., Vaden, K. I., & Eckert, M. A. (2018). Orthographic influence on spoken word identification: Behavioral and fMRI evidence. Neuropsychologia, 111, 103–111. https://doi.org/10.1016/j.neuropsychologia.2018.01.032 Clarke, A., & Tyler, L. K. (2015). Understanding What We See: How We Derive Meaning From Vision. Trends in Cognitive Sciences, 19(11), 677–687. https://doi.org/10.1016/j.tics.2015.08.008 Coggan, D. D., & Tong, F. (2023). Spikiness and animacy as potential organizing principles 123 OF 150 of human ventral visual cortex. Cerebral Cortex, 33(13), 8194–8217. https://doi.org/10.1093/cercor/bhad108 Cohen, L. (2003). Visual Word Recognition in the Left and Right Hemispheres: Anatomical and Functional Correlates of Peripheral Alexias. Cerebral Cortex, 13(12), 1313– 1333. https://doi.org/10.1093/cercor/bhg079 Cohen, L., & Dehaene, S. (2004). Specialization within the ventral stream: The case for the visual word form area. NeuroImage, 22(1), 466–476. https://doi.org/10.1016/j.neuroimage.2003.12.049 Cohen, L., Dehaene, S., Naccache, L., Lehéricy, S., Dehaene-Lambertz, G., Hénaff, M.-A., & Michel, F. (2000). The visual word form area: Spatial and temporal characterization of an initial stage of reading in normal subjects and posterior split-brain patients. Brain, 123(2), 291–307. https://doi.org/10.1093/brain/123.2.291 Cohen, L., Jobert, A., Le Bihan, D., & Dehaene, S. (2004). Distinct unimodal and multimodal regions for word processing in the left temporal cortex. NeuroImage, 23(4), 1256– 1270. https://doi.org/10.1016/j.neuroimage.2004.07.052 Cohen, L., Lehéricy, S., Chochon, F., Lemer, C., Rivaud, S., & Dehaene, S. (2002). Language‐specific tuning of visual cortex? Functional properties of the Visual Word Form Area. Brain, 125(5), 1054–1069. https://doi.org/10.1093/brain/awf094 Coltheart, M. (2004). Are There Lexicons? The Quarterly Journal of Experimental Psychology Section A, 57(7), 1153–1172. https://doi.org/10.1080/02724980443000007 Coltheart, M., Davelaar, E., Jonasson, J. T., & Besner, D. (2022). Access to the Internal Lexicon. In S. Dornič, Attention and Performance VI (1st ed., pp. 535–555). Routledge. https://doi.org/10.4324/9781003309734-29 Coltheart, M., Rastle, K., Perry, C., Langdon, R., & Ziegler, J. (2001). DRC: A dual route cascaded model of visual word recognition and reading aloud. Psychological Review, 108(1), 204–256. https://doi.org/10.1037/0033-295X.108.1.204 Combrisson, E., & Jerbi, K. (2015). Exceeding chance level by chance: The caveat of 124 OF 150 theoretical chance levels in brain signal classification and statistical assessment of decoding accuracy. Journal of Neuroscience Methods, 250, 126–136. https://doi.org/10.1016/j.jneumeth.2015.01.010 Conant, L. L., Liebenthal, E., Desai, A., Seidenberg, M. S., & Binder, J. R. (2020). Differential activation of the visual word form area during auditory phoneme perception in youth with dyslexia. Neuropsychologia, 146, 107543. https://doi.org/10.1016/j.neuropsychologia.2020.107543 Cox, R. W. (1996). AFNI: Software for Analysis and Visualization of Functional Magnetic Resonance Neuroimages. Computers and Biomedical Research, 29(3), 162–173. https://doi.org/10.1006/cbmr.1996.0014 Dale, A. M. (1999). Optimal experimental design for event‐related fMRI. Human Brain Mapping, 8, 109–114. Damasio, A. R., & Damasio, H. (1983). The anatomic basis of pure alexia. Neurology, 33(12), 1573–1573. https://doi.org/10.1212/WNL.33.12.1573 Dampuré, J., Ros, C., Rouet, J., & Vibert, N. (2012). How Word Familiarity Facilitates Visual Search for Verbal Material. Applied Cognitive Psychology, 26(2), 271–288. https://doi.org/10.1002/acp.1821 de Bruin, A., Carreiras, M., & Duñabeitia, J. A. (2017). The BEST Dataset of Language Proficiency. Frontiers in Psychology, 8. https://doi.org/10.3389/fpsyg.2017.00522 Dȩbska, A., Wójcik, M., Chyl, K., Dziȩgiel-Fivet, G., & Jednoróg, K. (2023). Beyond the Visual Word Form Area – a cognitive characterization of the left ventral occipitotemporal cortex. Frontiers in Human Neuroscience, 17, 1199366. https://doi.org/10.3389/fnhum.2023.1199366 Dehaene, S., & Cohen, L. (2007). Cultural Recycling of Cortical Maps. Neuron, 56(2), 384– 398. https://doi.org/10.1016/j.neuron.2007.10.004 Dehaene, S., & Cohen, L. (2011). The unique role of the visual word form area in reading. Trends in Cognitive Sciences, 15(6), 254–262. https://doi.org/10.1016/j.tics.2011.04.003 125 OF 150 Dehaene, S., Le Clec’H, G., Poline, J.-B., Le Bihan, D., & Cohen, L. (2002). The visual word form area: A prelexical representation of visual words in the fusiform gyrus: Neuroreport, 13(3), 321–325. https://doi.org/10.1097/00001756-200203040-00015 Dehaene, S., Naccache, L., Cohen, L., Bihan, D. L., Mangin, J.-F., Poline, J.-B., & Rivière, D. (2001). Cerebral mechanisms of word masking and unconscious repetition priming. Nature Neuroscience, 4(7), 752–758. https://doi.org/10.1038/89551 Déjerine, J. (1892). Contribution a l’étude anatomoclinique et clinique des dif ferentes varietes de cecite verbal. Compte Rendus Hebdomadaires Séances et Mémoires de La Société de Biologie, 4, 61–90. Deniz, F., Tseng, C., Wehbe, L., Dupré La Tour, T., & Gallant, J. L. (2023). Semantic Representations during Language Comprehension Are Affected by Context. The Journal of Neuroscience, 43(17), 3144–3158. https://doi.org/10.1523/JNEUROSCI.2459-21.2023 Desimone, R., Schein, S. J., Moran, J., & Ungerleider, L. G. (1985). Contour, color and shape analysis beyond the striate cortex. Vision Research, 25(3), 441–452. https://doi.org/10.1016/0042-6989(85)90069-0 Devlin, J. T., Jamison, H. L., Gonnerman, L. M., & Matthews, P. M. (2006). The Role of the Posterior Fusiform Gyrus in Reading. Journal of Cognitive Neuroscience, 18(6), 911– 922. https://doi.org/10.1162/jocn.2006.18.6.911 DeWitt, I., & Rauschecker, J. P. (2012). Phoneme and word recognition in the auditory ventral stream. Proceedings of the National Academy of Sciences, 109(8). https://doi.org/10.1073/pnas.1113427109 Diaz, M. T., Karimi, H., Troutman, S. B. W., Gertel, V. H., Cosgrove, A. L., & Zhang, H. (2021). Neural sensitivity to phonological characteristics is stable across the lifespan. NeuroImage, 225, 117511. https://doi.org/10.1016/j.neuroimage.2020.117511 Dick, A. S., Bernal, B., & Tremblay, P. (2014). The Language Connectome: New Pathways, New Concepts. The Neuroscientist, 20(5), 453–467. https://doi.org/10.1177/1073858413513502 126 OF 150 Diedenhofen, B., & Musch, J. (2015). cocor: A Comprehensive Solution for the Statistical Comparison of Correlations. PLOS ONE, 10(4), e0121945. https://doi.org/10.1371/journal.pone.0121945 Ding, S., Van Hoesen, G. W., Cassell, M. D., & Poremba, A. (2009). Parcellation of human temporal polar cortex: A combined analysis of multiple cytoarchitectonic, chemoarchitectonic, and pathological markers. Journal of Comparative Neurology, 514(6), 595–623. https://doi.org/10.1002/cne.22053 Dobbins, I. G., & Wagner, A. D. (2005). Domain-general and Domain-sensitive Prefrontal Mechanisms for Recollecting Events and Detecting Novelty. Cerebral Cortex, 15(11), 1768–1778. https://doi.org/10.1093/cercor/bhi054 Downing, P. E., Chan, A. W.-Y., Peelen, M. V., Dodds, C. M., & Kanwisher, N. (2006). Domain Specificity in Visual Cortex. Cerebral Cortex, 16(10), 1453–1461. https://doi.org/10.1093/cercor/bhj086 Duchon, A., Perea, M., Sebastián-Gallés, N., Martí, A., & Carreiras, M. (2013). EsPal: Onestop shopping for Spanish word properties. Behavior Research Methods, 45(4), 1246–1258. https://doi.org/10.3758/s13428-013-0326-1 Dziȩgiel-Fivet, G., Beck, J., & Jednoróg, K. (2023). The role of the left ventral occipitotemporal cortex in speech processing—The influence of visual deprivation. Frontiers in Human Neuroscience, 17, 1228808. https://doi.org/10.3389/fnhum.2023.1228808 Epstein, R. A., & Baker, C. I. (2019). Scene Perception in the Human Brain. Annual Review of Vision Science, 5(1), 373–397. https://doi.org/10.1146/annurev-vision-091718014809 Farrell, S., & Lewandowsky, S. (Eds.). (2018). Model Comparison. In Computational Modeling of Cognition and Behavior (pp. 241–272). Cambridge University Press; Cambridge Core. https://doi.org/10.1017/CBO9781316272503.011 Fedorenko, E., Duncan, J., & Kanwisher, N. (2012). Language-Selective and DomainGeneral Regions Lie Side by Side within Broca’s Area. Current Biology, 22(21), 127 OF 150 2059–2062. https://doi.org/10.1016/j.cub.2012.09.011 Fedorenko, E., Ivanova, A. A., & Regev, T. I. (2024). The language network as a natural kind within the broader landscape of the human brain. Nature Reviews Neuroscience, 25(5), 289–312. https://doi.org/10.1038/s41583-024-00802-4 Fernandes, M. A., Moscovitch, M., Ziegler, M., & Grady, C. (2005). Brain regions associated with successful and unsuccessful retrieval of verbal episodic memory as revealed by divided attention. Neuropsychologia, 43(8), 1115–1127. https://doi.org/10.1016/j.neuropsychologia.2004.11.026 Feyereisen, P., Van der Borght, F., & Seron, X. (1988). The operativity effect in naming: A re-analysis. Neuropsychologia, 26(3), 401–415. https://doi.org/10.1016/00283932(88)90094-2 Fiebach, C. J., Friederici, A. D., Müller, K., & Cramon, D. Y. von. (2002). fMRI Evidence for Dual Routes to the Mental Lexicon in Visual Word Recognition. Journal of Cognitive Neuroscience, 14(1), 11–23. https://doi.org/10.1162/089892902317205285 Fischer-Baum, S., Bruggemann, D., Gallego, I. F., Li, D. S. P., & Tamez, E. R. (2017). Decoding levels of representation in reading: A representational similarity approach. Cortex, 90, 88–102. https://doi.org/10.1016/j.cortex.2017.02.017 Fisher, R. A. (1921). On the probable error of a coefficient of correlation deduced from a small sample. Metron, 1, 3–32. Frankland, S. M., & Greene, J. D. (2015). An architecture for encoding sentence meaning in left mid-superior temporal cortex. Proceedings of the National Academy of Sciences, 112(37), 11732–11737. https://doi.org/10.1073/pnas.1421236112 Friederici, A. D. (2009). Pathways to language: Fiber tracts in the human brain. Trends in Cognitive Sciences, 13(4), 175–181. https://doi.org/10.1016/j.tics.2009.01.001 Friederici, A. D. (2011). The Brain Basis of Language Processing: From Structure to Function. Physiological Reviews, 91(4), 1357–1392. https://doi.org/10.1152/physrev.00006.2011 Friederici, A. D. (2012). The cortical language circuit: From auditory perception to sentence 134 OF 150 Sciences, 107(6), 2717–2722. https://doi.org/10.1073/pnas.0907307107 Lau, E. F., Phillips, C., & Poeppel, D. (2008). A cortical network for semantics: (De)constructing the N400. Nature Reviews Neuroscience, 9(12), 920–933. https://doi.org/10.1038/nrn2532 Lerma-Usabiaga, G., Carreiras, M., & Paz-Alonso, P. M. (2018). Converging evidence for functional and structural segregation within the left ventral occipitotemporal cortex in reading. Proceedings of the National Academy of Sciences, 115(42), E9981–E9990. https://doi.org/10.1073/pnas.1803003115 Lewis, G., & Poeppel, D. (2014). The role of visual representations during the lexical access of spoken words. Brain and Language, 134, 1–10. https://doi.org/10.1016/j.bandl.2014.03.008 Lewis-Peacock, J. A., & Norman, K. A. (2014). Multivoxel Pattern Analysis of Functional MRI Data. In M. S. Gazzaniga & G. R. Mangun (Eds.), The Cognitive Neurosciences (5th ed.). The MIT Press. https://doi.org/10.7551/mitpress/9504.003.0099 Lim, S. W. H. (2016). The Influence of Orthographic Neighborhood Density and Word Frequency on Visual Word Recognition: Insights from RT Distributional Analyses. Frontiers in Psychology, 7. https://doi.org/10.3389/fpsyg.2016.00401 Liuzzi, A. G., Meersmans, K., Peeters, R., De Deyne, S., Dupont, P., & Vandenberghe, R. (2024). Semantic representations in inferior frontal and lateral temporal cortex during picture naming, reading, and repetition. Human Brain Mapping, 45(2), e26603. https://doi.org/10.1002/hbm.26603 Liuzzi, A. G., Meersmans, K., Storms, G., De Deyne, S., Dupont, P., & Vandenberghe, R. (2023). Independency of Coding for Affective Similarities and for Word Cooccurrences in Temporal Perisylvian Neocortex. Neurobiology of Language, 4(2), 257–279. https://doi.org/10.1162/nol_a_00095 López-Barroso, D., & De Diego-Balaguer, R. (2017). Language Learning Variability within the Dorsal and Ventral Streams as a Cue for Compensatory Mechanisms in Aphasia Recovery. Frontiers in Human Neuroscience, 11, 476. 135 OF 150 https://doi.org/10.3389/fnhum.2017.00476 López-Barroso, D., Thiebaut De Schotten, M., Morais, J., Kolinsky, R., Braga, L. W., Guerreiro-Tauil, A., Dehaene, S., & Cohen, L. (2020). Impact of literacy on the functional connectivity of vision and language related networks. NeuroImage, 213, 116722. https://doi.org/10.1016/j.neuroimage.2020.116722 Ludersdorfer, P., Wimmer, H., Richlan, F., Schurz, M., Hutzler, F., & Kronbichler, M. (2016). Left ventral occipitotemporal activation during orthographic and semantic processing of auditory words. NeuroImage, 124, 834–842. https://doi.org/10.1016/j.neuroimage.2015.09.039 MacGregor, L. J., Pulvermüller, F., Van Casteren, M., & Shtyrov, Y. (2012). Ultra-rapid access to words in the brain. Nature Communications, 3(1), 711. https://doi.org/10.1038/ncomms1715 Malcolm, G. L., Groen, I. I. A., & Baker, C. I. (2016). Making Sense of Real-World Scenes. Trends in Cognitive Sciences, 20(11), 843–856. https://doi.org/10.1016/j.tics.2016.09.003 Malikovic, A., Amunts, K., Schleicher, A., Mohlberg, H., Kujovic, M., Palomero-Gallagher, N., Eickhoff, S. B., & Zilles, K. (2016). Cytoarchitecture of the human lateral occipital cortex: Mapping of two extrastriate areas hOc4la and hOc4lp. Brain Structure and Function, 221(4), 1877–1897. https://doi.org/10.1007/s00429-015-1009-8 Martin, A. (2007). The Representation of Object Concepts in the Brain. Annual Review of Psychology, 58(1), 25–45. https://doi.org/10.1146/annurev.psych.57.102904.190143 Martin, A., Haxby, J. V., Lalonde, F. M., Wiggs, C. L., & Ungerleider, L. G. (1995). Discrete Cortical Regions Associated with Knowledge of Color and Knowledge of Action. Science, 270(5233), 102–105. https://doi.org/10.1126/science.270.5233.102 Martin, C. B., & Barense, M. D. (2023). Perception and Memory in the Ventral Visual Stream and Medial Temporal Lobe. Annual Review of Vision Science, 9(1), 409–434. https://doi.org/10.1146/annurev-vision-120222-014200 Martin, C. B., McLean, D. A., O’Neil, E. B., & Kohler, S. (2013). Distinct Familiarity-Based 136 OF 150 Response Patterns for Faces and Buildings in Perirhinal and Parahippocampal Cortex. Journal of Neuroscience, 33(26), 10915–10923. https://doi.org/10.1523/JNEUROSCI.0126-13.2013 Mazaika, P. K., Hoeft, F., Glover, G. H., & Reiss, A. L. (2009). Methods and Software for fMRI Analysis for Clinical Subjects. Human Brain Mapping Conference, Florence, Italy. McCandliss, B. D., Cohen, L., & Dehaene, S. (2003). The visual word form area: Expertise for reading in the fusiform gyrus. Trends in Cognitive Sciences, 7(7), 293–299. https://doi.org/10.1016/S1364-6613(03)00134-7 McClelland, J. L., & Rogers, T. T. (2003). The parallel distributed processing approach to semantic cognition. Nature Reviews Neuroscience, 4(4), 310–322. https://doi.org/10.1038/nrn1076 McClelland, J. L., & Rumelhart, D. E. (1981). An interactive activation model of context effects in letter perception: I. An account of basic findings. Psychological Review, 88(5), 375–407. https://doi.org/10.1037/0033-295X.88.5.375 Meersmans, K., Bruffaerts, R., Jamoulle, T., Liuzzi, A. G., De Deyne, S., Storms, G., Dupont, P., & Vandenberghe, R. (2020). Representation of associative and affective semantic similarity of abstract words in the lateral temporal perisylvian language regions. NeuroImage, 217, 116892. https://doi.org/10.1016/j.neuroimage.2020.116892 Meersmans, K., Storms, G., De Deyne, S., Bruffaerts, R., Dupont, P., & Vandenberghe, R. (2022). Orienting to different dimensions of word meaning alters the representation of word meaning in early processing regions. Cerebral Cortex, 32(15), 3302–3317. Scopus. https://doi.org/10.1093/cercor/bhab416 Memetova, K. S., Knyazeva, V. M., Stankevich, L. N., Malanchuk, I. G., & Aleksandrov, A. A. (2024). Functional Connectivity of Brain Regions Resulting from Learning Unfamiliar Words: Word Frequency Effect. In A. V. Samsonovich & T. Liu (Eds.), Biologically Inspired Cognitive Architectures 2023 (Vol. 1130, pp. 595–603). Springer Nature Switzerland. https://doi.org/10.1007/978-3-031-50381-8_63 137 OF 150 Mikolov, T., Sutskever, I., Chen, K., Corrado, G., & Dean, J. (2013). Distributed Representations of Words and Phrases and their Compositionality. https://doi.org/10.48550/ARXIV.1310.4546 Mishkin, M., Ungerleider, L. G., & Macko, K. A. (1983). Object vision and spatial vision: Two cortical pathways. Trends in Neurosciences, 6, 414–417. https://doi.org/10.1016/0166-2236(83)90190-X Mitchell, T. M., Shinkareva, S. V., Carlson, A., Chang, K.-M., Malave, V. L., Mason, R. A., & Just, M. A. (2008). Predicting Human Brain Activity Associated with the Meanings of Nouns. Science, 320(5880), 1191–1195. https://doi.org/10.1126/science.1152876 Mkrtychian, N., Blagovechtchenski, E., Kurmakaeva, D., Gnedykh, D., Kostromina, S., & Shtyrov, Y. (2019). Concrete vs. Abstract Semantics: From Mental Representations to Functional Brain Mapping. Frontiers in Human Neuroscience, 13, 267. https://doi.org/10.3389/fnhum.2019.00267 Moerel, M., De Martino, F., & Formisano, E. (2012). Processing of Natural Sounds in Human Auditory Cortex: Tonotopy, Spectral Tuning, and Relation to Voice Sensitivity. The Journal of Neuroscience, 32(41), 14205–14216. https://doi.org/10.1523/JNEUROSCI.1388-12.2012 Moore, C. J., & Price, C. J. (1999). Three Distinct Ventral Occipitotemporal Regions for Reading and Object Naming. NeuroImage, 10(2), 181–192. https://doi.org/10.1006/nimg.1999.0450 Morales, J., Lau, H., & Fleming, S. M. (2018). Domain-General and Domain-Specific Patterns of Activity Supporting Metacognition in Human Prefrontal Cortex. The Journal of Neuroscience, 38(14), 3534–3546. https://doi.org/10.1523/JNEUROSCI.2360-17.2018 Moulton, E., Bouhali, F., Monzalvo, K., Poupon, C., Zhang, H., Dehaene, S., DehaeneLambertz, G., & Dubois, J. (2019). Connectivity between the visual word form area and the parietal lobe improves after the first year of reading instruction: A longitudinal MRI study in children. Brain Structure and Function, 224(4), 1519–1536. 138 OF 150 https://doi.org/10.1007/s00429-019-01855-3 Mueller, C. J., White, C. N., & Kuchinke, L. (2017). Electrophysiological correlates of the drift diffusion model in visual word recognition. Human Brain Mapping, 38(11), 5616– 5627. https://doi.org/10.1002/hbm.23753 Mumford, J. A., Davis, T., & Poldrack, R. A. (2014). The impact of study design on pattern estimation for single-trial multivariate pattern analysis. NeuroImage, 103, 130–138. https://doi.org/10.1016/j.neuroimage.2014.09.026 Myers, C. E., Interian, A., & Moustafa, A. A. (2022). A practical introduction to using the drift diffusion model of decision-making in cognitive psychology, neuroscience, and health sciences. Frontiers in Psychology, 13, 1039172. https://doi.org/10.3389/fpsyg.2022.1039172 Nasr, S., Liu, N., Devaney, K. J., Yue, X., Rajimehr, R., Ungerleider, L. G., & Tootell, R. B. H. (2011). Scene-Selective Cortical Regions in Human and Nonhuman Primates. The Journal of Neuroscience, 31(39), 13771–13785. https://doi.org/10.1523/JNEUROSCI.2792-11.2011 Neveu, A., & Kaushanskaya, M. (2023). Paired-associate versus cross-situational: How do verbal working memory and word familiarity affect word learning? Memory & Cognition, 51(7), 1670–1682. https://doi.org/10.3758/s13421-023-01421-7 Nordt, M., Gomez, J., Natu, V., Jeska, B., Barnett, M., & Grill-Spector, K. (2019). Learning to Read Increases the Informativeness of Distributed Ventral Temporal Responses. Cerebral Cortex, 29(7), 3124–3139. https://doi.org/10.1093/cercor/bhy178 Nordt, M., Gomez, J., Natu, V. S., Rezai, A. A., Finzi, D., Kular, H., & Grill-Spector, K. (2021). Cortical recycling in high-level visual cortex during childhood development. Nature Human Behaviour, 5(12), 1686–1697. https://doi.org/10.1038/s41562-02101141-5 Nordt, M., Gomez, J., Natu, V. S., Rezai, A. A., Finzi, D., Kular, H., & Grill-Spector, K. (2023). Longitudinal development of category representations in ventral temporal cortex predicts word and face recognition. Nature Communications, 14(1), 8010. 139 OF 150 https://doi.org/10.1038/s41467-023-43146-w Norman, K. A., Polyn, S. M., Detre, G. J., & Haxby, J. V. (2006). Beyond mind-reading: Multivoxel pattern analysis of fMRI data. Trends in Cognitive Sciences, 10(9), 424–430. https://doi.org/10.1016/j.tics.2006.07.005 OECD. (2023). PISA 2022 Results (Volume I): The State of Learning and Equity in Education. OECD. https://doi.org/10.1787/53f23881-en Okazawa, G., Tajima, S., & Komatsu, H. (2016). Gradual Development of Visual TextureSelective Properties Between Macaque Areas V2 and V4. Cerebral Cortex, 4867– 4880. https://doi.org/10.1093/cercor/bhw282 Oliver, M., Carreiras, M., & Paz-Alonso, P. M. (2017). Functional Dynamics of Dorsal and Ventral Reading Networks in Bilinguals. Cerebral Cortex, 27(12), 5431–5443. https://doi.org/10.1093/cercor/bhw310 Olson, I. R., McCoy, D., Klobusicky, E., & Ross, L. A. (2013). Social cognition and the anterior temporal lobes: A review and theoretical framework. Social Cognitive and Affective Neuroscience, 8(2), 123–133. https://doi.org/10.1093/scan/nss119 Paivio, A. (2010). Dual coding theory and the mental lexicon. The Mental Lexicon, 5(2), 205– 230. https://doi.org/10.1075/ml.5.2.04pai Paivio, A. (2014). Intelligence, dual coding theory, and the brain. Intelligence, 47, 141–158. https://doi.org/10.1016/j.intell.2014.09.002 Palmer, S. D., MacGregor, L. J., & Havelka, J. (2013). Concreteness effects in singlemeaning, multi-meaning and newly acquired words. Brain Research, 1538, 135–150. https://doi.org/10.1016/j.brainres.2013.09.015 Papitto, G., Friederici, A. D., & Zaccarella, E. (2020). The topographical organization of motor processing: An ALE meta-analysis on six action domains and the relevance of Broca’s region. NeuroImage, 206, 116321. https://doi.org/10.1016/j.neuroimage.2019.116321 Pasupathy, A., Popovkina, D. V., & Kim, T. (2020). Visual Functions of Primate Area V4. Annual Review of Vision Science, 6(1), 363–385. https://doi.org/10.1146/annurev- 140 OF 150 vision-030320-041306 Pattamadilok, C., Planton, S., & Bonnard, M. (2019). Spoken language coding neurons in the Visual Word Form Area: Evidence from a TMS adaptation paradigm. NeuroImage, 186, 278–285. https://doi.org/10.1016/j.neuroimage.2018.11.014 Patterson, K., Nestor, P. J., & Rogers, T. T. (2007). Where do you know what you know? The representation of semantic knowledge in the human brain. Nature Reviews Neuroscience, 8(12), 976–987. https://doi.org/10.1038/nrn2277 Paz-Alonso, P. M., Oliver, M., Lerma-Usabiaga, G., Caballero-Gaudes, C., Quiñones, I., Suárez-Coalla, P., Duñabeitia, J. A., Cuetos, F., & Carreiras, M. (2018). Neural correlates of phonological, orthographic and semantic reading processing in dyslexia. NeuroImage: Clinical, 20, 433–447. https://doi.org/10.1016/j.nicl.2018.08.018 Peelen, M. V., & Downing, P. E. (2005). Selectivity for the Human Body in the Fusiform Gyrus. Journal of Neurophysiology, 93(1), 603–608. https://doi.org/10.1152/jn.00513.2004 Pereira, F., Botvinick, M., & Detre, G. (2013). Using Wikipedia to learn semantic feature representations of concrete concepts in neuroimaging experiments. Artificial Intelligence, 194, 240–252. https://doi.org/10.1016/j.artint.2012.06.005 Petersson, K. (2004). Artificial syntactic violations activate Broca?s region. Cognitive Science, 28(3), 383–407. https://doi.org/10.1016/j.cogsci.2003.12.003 Planton, S., Chanoine, V., Sein, J., Anton, J.-L., Nazarian, B., Pallier, C., & Pattamadilok, C. (2019). Top-down activation of the visuo-orthographic system during spoken sentence processing. NeuroImage, 202, 116135. https://doi.org/10.1016/j.neuroimage.2019.116135 Plaut, D. C., McClelland, J. L., Seidenberg, M. S., & Patterson, K. (1996). Understanding normal and impaired word reading: Computational principles in quasi-regular domains. Psychological Review, 103(1), 56–115. https://doi.org/10.1037/0033295X.103.1.56 Pohl, W. (1973). Dissociation of spatial discrimination deficits following frontal and parietal 141 OF 150 lesions in monkeys. Journal of Comparative and Physiological Psychology, 82(2), 227–239. https://doi.org/10.1037/h0033922 Ponari, M., Norbury, C. F., & Vigliocco, G. (2018). Acquisition of abstract concepts is influenced by emotional valence. Developmental Science, 21(2), e12549. https://doi.org/10.1111/desc.12549 Popal, H., Wang, Y., & Olson, I. R. (2019). A Guide to Representational Similarity Analysis for Social Neuroscience. Social Cognitive and Affective Neuroscience, 14(11), 1243– 1253. https://doi.org/10.1093/scan/nsz099 Price, C. J., & Devlin, J. T. (2003). The myth of the visual word form area. NeuroImage, 19(3), 473–481. https://doi.org/10.1016/S1053-8119(03)00084-3 Price, C. J., & Devlin, J. T. (2011). The Interactive Account of ventral occipitotemporal contributions to reading. Trends in Cognitive Sciences, 15(6), 246–253. https://doi.org/10.1016/j.tics.2011.04.001 Protopapas, A., Orfanidou, E., Taylor, J. S. H., Karavasilis, E., Kapnoula, E. C., Panagiotaropoulou, G., Velonakis, G., Poulou, L. S., Smyrnis, N., & Kelekis, D. (2016). Evaluating cognitive models of visual word recognition using fMRI: Effects of lexical and sublexical variables. NeuroImage, 128, 328–341. https://doi.org/10.1016/j.neuroimage.2016.01.013 Pugh, K. R., Mencl, W. E., Jenner, A. R., Katz, L., Frost, S. J., Lee, J. R., Shaywitz, S. E., & Shaywitz, B. A. (2001). Neurobiological studies of reading and reading disability. Journal of Communication Disorders, 34(6), 479–492. https://doi.org/10.1016/S00219924(01)00060-0 Ralph, M. A. L., Jefferies, E., Patterson, K., & Rogers, T. T. (2017). The neural and computational bases of semantic cognition. Nature Reviews Neuroscience, 18(1), 42–55. https://doi.org/10.1038/nrn.2016.150 Ratcliff, R. (1978). A theory of memory retrieval. Psychological Review, 85(2), 59–108. https://doi.org/10.1037/0033-295X.85.2.59 Ratcliff, R., & McKoon, G. (2008). The Diffusion Decision Model: Theory and Data for Two- 142 OF 150 Choice Decision Tasks. Neural Computation, 20(4), 873–922. https://doi.org/10.1162/neco.2008.12-06-420 Ratcliff, R., Smith, P. L., Brown, S. D., & McKoon, G. (2016). Diffusion Decision Model: Current Issues and History. Trends in Cognitive Sciences, 20(4), 260–281. https://doi.org/10.1016/j.tics.2016.01.007 Rauschecker, J. P. (1998). Parallel Processing in the Auditory Cortex of Primates. Audiology and Neurotology, 3(2–3), 86–103. https://doi.org/10.1159/000013784 Richardson, F. M., Seghier, M. L., Leff, A. P., Thomas, M. S. C., & Price, C. J. (2011). Multiple Routes from Occipital to Temporal Cortices during Reading. Journal of Neuroscience, 31(22), 8239–8247. https://doi.org/10.1523/JNEUROSCI.651910.2011 Riesenhuber, M., & Glezer, L. S. (2017). Evidence for rapid localist plasticity in the ventral visual stream: The example of words. Language, Cognition and Neuroscience, 32(3), 286–294. https://doi.org/10.1080/23273798.2016.1210178 Rissman, J., Gazzaley, A., & D’Esposito, M. (2004). Measuring functional connectivity during distinct stages of a cognitive task. NeuroImage, 23(2), 752–763. https://doi.org/10.1016/j.neuroimage.2004.06.035 Ritchie, J. B., Zeman, A. A., Bosmans, J., Sun, S., Verhaegen, K., & Op De Beeck, H. P. (2021). Untangling the Animacy Organization of Occipitotemporal Cortex. The Journal of Neuroscience, 41(33), 7103–7119. https://doi.org/10.1523/JNEUROSCI.2628-20.2021 Roe, A. W., & Ts’o, D. Y. (2015). Specificity of V1–V2 orientation networks in the primate visual cortex. Cortex, 72, 168–178. https://doi.org/10.1016/j.cortex.2015.07.007 Roger, E., Banjac, S., Thiebaut de Schotten, M., & Baciu, M. (2022). Missing links: The functional unification of language and memory (L∪M). Neuroscience & Biobehavioral Reviews, 133, 104489. https://doi.org/10.1016/j.neubiorev.2021.12.012 Roger, E., Rodrigues De Almeida, L., Loevenbruck, H., Perrone-Bertolotti, M., Cousin, E., Schwartz, J. L., Perrier, P., Dohen, M., Vilain, A., Baraduc, P., Achard, S., & Baciu, 143 OF 150 M. (2022). Unraveling the functional attributes of the language connectome: Crucial subnetworks, flexibility and variability. NeuroImage, 263. https://doi.org/10.1016/j.neuroimage.2022.119672 Rosenberg, M. D., Finn, E. S., Constable, R. T., & Chun, M. M. (2015). Predicting momentto-moment attentional state. NeuroImage, 114, 249–256. https://doi.org/10.1016/j.neuroimage.2015.03.032 Rueckl, J. G., Paz-Alonso, P. M., Molfese, P. J., Kuo, W.-J., Bick, A., Frost, S. J., Hancock, R., Wu, D. H., Mencl, W. E., Duñabeitia, J. A., Lee, J.-R., Oliver, M., Zevin, J. D., Hoeft, F., Carreiras, M., Tzeng, O. J. L., Pugh, K. R., & Frost, R. (2015). Universal brain signature of proficient reading: Evidence from four contrasting languages. Proceedings of the National Academy of Sciences, 112(50), 15510–15515. https://doi.org/10.1073/pnas.1509321112 Rundle, M. M., Coch, D., Connolly, A. C., & Granger, R. H. (2018). Dissociating frequency and animacy effects in visual word processing: An fMRI study. Brain and Language, 183, 54–63. https://doi.org/10.1016/j.bandl.2018.05.005 Saenz, M., & Langers, D. R. M. (2014). Tonotopic mapping of human auditory cortex. Hearing Research, 307, 42–52. https://doi.org/10.1016/j.heares.2013.07.016 Salminen-Vaparanta, N., Koivisto, M., Noreika, V., Vanni, S., & Revonsuo, A. (2012). Neuronavigated transcranial magnetic stimulation suggests that area V2 is necessary for visual awareness. Neuropsychologia, 50(7), 1621–1627. https://doi.org/10.1016/j.neuropsychologia.2012.03.015 Sánchez, A., Carreiras, M., & Paz-Alonso, P. M. (2023). Word frequency and reading demands modulate brain activation in the inferior frontal gyrus. Scientific Reports, 13(1), 17217. https://doi.org/10.1038/s41598-023-44420-z Sandak, R., Mencl, W. E., Frost, S. J., & Pugh, K. R. (2004). The Neurobiological Basis of Skilled and Impaired Reading: Recent Findings and New Directions. Scientific Studies of Reading, 8(3), 273–292. https://doi.org/10.1207/s1532799xssr0803_6 Sapountzis, P., Schluppeck, D., Bowtell, R., & Peirce, J. W. (2010). A comparison of fMRI