Cognitive Network Science symposium

Classic studies in spatial cognition demonstrate that landmarks are cognitive reference points that can create a "warping" effect on people's perception of distances between two locations (e.g., Rosch, 1975). The goal of the current study was to investigate if words that could be considered as "landmark" nodes in a language network of phonological word forms might also show similar effects on people's judgments of phonological distances. We considered two possible operationalizations of landmark words based on the network science literature: one based on closeness centrality, and the other based on identification of keyplayer nodes that led to the maximal fragmentation of the network if removed. Participants listened to English word pairs which contained either landmarks or foils with similar lexical properties to the landmarks, paired with a random word in the network. Preliminary analyses indicated that word pairs containing a keyplayer node were considered to be less similar sounding than the corresponding foils, and there was no effect of closeness centrality. We suggest that landmark nodes in the phonological lexicon may behave as “super-transmitters” in the network, with greater dissipation of activation corresponding to lower probabilities of similarity judgments.

Traditional network models for human memory organization represent conceptual associations using pairwise links. However, conceptual assocations can naturally be depicted as hyperlinks involving more than two concepts. Using word associations data, we introduced two studies where pairwise networks and hypergraphs were compared. In the first study, we quantitatively investigate whether there is any benefit in using the hypergraph model over a pairwise network in predicting (i) test-based age of acquisition norms in children up to age 9 years and (ii) normative learning in toddlers up to age 30 months. In the second study, we build hypergraphs from word associations and use evaluation methods from machine learning features to predict concept concreteness. Our studies reveal that hypergraphs contain richer information and better predict word concreteness in adults, and age of acquisition trends in toddlers. The shift from pairwise networks to hypergraphs represents a significant advancement in memory modeling, offering deeper insights into cognitive phenomena and developmental processes.

High-level cognition, such as creativity and intelligence, involve the ability to efficiently search through memory. Thus, can the way a person searches their memory be indicative of their high-level cognitive abilities? I will present a series of studies where we computationally and empirically examine how performance in a semantic fluency task operationalizes mental navigation through a multidimensional representation of their mental lexicon – a cognitive multiplex network. I will show how such analysis can be used to accurately predict individual differences in creativity, intelligence, and the personality trait Openness to Experience. Critically, this work highlights a common search behavior that emphasizes a more “peripheral” search in relation to heightened cognitive abilities. Furthermore, this work highlights the advantages of studying the mental lexicon as a multidimensional construct.

A psychometric scale reports experiences in terms of items/sentences rated by individuals. We investigate whether psychometric item ratings reflect semantic/syntactic associations between concepts in items. To this aim, we introduce semantic loadings as a semantic counterpart of psychometric factors, i.e. clusters of items obtained by correlations between item ratings. Semantic loadings quantify how clusters of semantically related concepts, as expressed in the texts of items, are allocated across psychometric factors as identified by ratings. As a case study, we focus on 39775 individual responses to the Depression Anxiety and Stress Scale (DASS) with 42 items on a 4-point Likert scale. To identify communities of semantically related concepts, we exploit the cognitive network framework of Textual Forma Mentis Networks (TFMNs), which reconstruct semantic/syntactic links encoded in the texts of items (e.g. "feel" and "sad" in the item "I usually feel sad"). To identify factors we compare eigenvector-based exploratory analysis with Graph Exploratory Analysis (EGA), which can both cluster items (and their texts) according to user ratings. We find that EGA is better at reconstructing the psychological organisation of DASS along the dimensions of anxiety, stress and depression. Following dual coding theory and the Deep Lexical Hypothesis, we posit that the act of reading items activates interconnected concepts and this influences user ratings and their expressed psychological constructs. Our results show a quantitative match: TFMN-based semantic loadings can identify specific aspects of emotional dysregulation, emotional exhaustion, physical distress and tension states of EGA-based psychometric factors, in non-random ways (up to p<0.001). We discuss our results in view of relevant mental distress literature, psychometric scale designing and links with episodic and semantic memories.

Cognitive network science is a growing field at the fringe of mathematical psychology, graph theory and computational network science. A cognitive network is a representation of associative knowledge, where nodes represent atomistic cognitive units (e.g. concepts, word forms, graphemes, etc.) linked together by one or multiple types of cognitive associations (e.g. memory recall patterns, co-occurrences in words, phonological similarities, etc.). Vastly disseminated in psychology in the 60s/70s thanks to the works of Collins, Quillian and Loftus, cognitive networks have been recently re-discovered with the advent of novel computational tools from graph theory but also thanks to the increasing availability of large-scale cognitive datasets, enabling novel methodological frameworks for investigating the interplay between the structure of associative knowledge and a variety of psychological phenomena. This symposium, proposed to MathPsych/ICCM 2024 VIRTUAL, focuses on the exploration of cognitive networks as both representational and computational models for understanding cognitive phenomena. The representational capacity of cognitive networks is highlighted through their ability to act as proxies of the complex organization of associative knowledge. In the last 20 years, several studies, including Steyvers and Tenenbaum (Cog. Sci. 2005), Vitevitch (ASLHA, 2008), Hills et al. (Psych. Sci. 2009) and Kenett et al. (PNAS, 2018) have accumulated strong scientific evidence that associative structure, as incompletely captured by cognitive networks, can explain behavioral data ranging from lexical processing to language acquisition, from word confusability to semantic memory functioning. The recent advent of multiplex lexical networks (Stella et al., Sci.Rep. 2017), encoding simultaneously several types of conceptual associations, opens the way to quantitatively exploring how the interplay between different aspects of associative knowledge (e.g. semantic and phonological representations) can affect the above cognitive phenomena. As remarked in the recent review of the field by Siew et al. (Compl. 2019), cognitive networks are also powerful computational tools. Network provide a substrate for simulating models like spreading activation or mental search via random walks and simple diffusion. Thus, cognitive networks can provide unprecedented ways to: (i) test multiple experimental scenarios by changing simulation parameters or network structure, e.g. simulating spreading activation across multiple priming conditions, (Siew, Beh. Res. Met. 2019); (ii) open the way to automatic frameworks capturing, predicting and understanding psychological phenomena relative to the presence of specific constructs or personality traits, e.g. capturing openness to experience via network exploration (Samuel et al., J. Pers. Res. 2023). These interpretable computational approaches are importantly driven by decades of psychological theories and promote not only computational psychology but also our understanding of psychological and cognitive phenomena. The CNS symposium will provide a crucial opportunity to engage with the latest applications of graph theoretical investigations of cognitive networks, elucidating how the structure of cognitive networks can reveal important characteristics of cognitive processing, such as information flow efficiency, robustness, and the role of hub nodes in cognitive architectures. By integrating empirical findings with theoretical models, the symposium aims to advance our comprehension of cognitive networks in mathematical psychology and cognitive data science. Participants will engage with the latest research methodologies and computational tools used in the analysis of cognitive networks, fostering a deeper scientific understanding of how these models can be applied to investigate cognitive phenomena. This interdisciplinary workshop is designed to facilitate rigorous academic dialogue, encouraging collaboration across fields to refine and expand the application of cognitive networks in understanding the complexities of human cognition.

There are two contrasting views of aging. One sees aging as a process of cognitive decline, a natural consequence of biological aging. The other sees aging as a process of lifelong learning. Of these two views, one is based on an underlying process of atrophy, about which we understand little. The other is based on enrichment. Enrichment is nothing more than learning, and we understand learning processes well. Moreover, there is clear evidence for it: older adults show conspicuous improvements in vocabulary across the lifespan as well as in many other knowledge-related domains. But can enrichment explain recent findings of a fracturing of the semantic lexicon in late life? Or an age-related reduction in similarity ratings? In this talk, I will investigate how understanding the nature of structural changes in knowledge induced by learning across the lifespan can explain these and other changes. In addition, I will show how this account provides important insights into how we understand and measure structure and process in cognitive representations.