Cognitive network science for understanding online social cognitions: A brief review

Social media are digitalising massive amounts of users' cognitions in terms of timelines and emotional content. Such Big Data opens unprecedented opportunities for investigating cognitive phenomena like perception, personality and information diffusion but requires suitable interpretable frameworks. Since social media data come from users' minds, worthy candidates for this challenge are cognitive networks, models of cognition giving structure to mental conceptual associations. This work outlines how cognitive network science can open new, quantitative ways for understanding cognition through online media, like: (i) reconstructing how users semantically and emotionally frame events with contextual knowledge unavailable to machine learning, (ii) investigating conceptual salience/prominence through knowledge structure in social discourse; (iii) studying users' personality traits like openness-to-experience, curiosity, and creativity through language in posts; (iv) bridging cognitive/emotional content and social dynamics via multilayer networks comparing the mindsets of influencers and followers. These advancements combine cognitive-, network- and computer science to understand cognitive mechanisms in both digital and real-world settings but come with limitations concerning representativeness, individual variability and data integration. Such aspects are discussed along the ethical implications of manipulating socio-cognitive data. In the future, reading cognitions through networks and social media can expose cognitive biases amplified by online platforms and relevantly inform policy making, education and markets about massive, complex cognitive trends.


INTRODUCTION
A key component of cognition lies in the ability to express ideas through language. Over the centuries, concepts and emotions were retrieved from the human mind, encapsulated in words and then diffused through written and oral media. Only in the last few decades this process accelerated drastically.
Online social media, mimicking friendship circles, revolutionised people's ways to speak their minds, structuring their stances, knowledge and perceptions through social discourse, timelines and posts (v) profile also online users in terms of their digital footprint based on the language they used online, e.g. assess personality traits and information seeking patterns (Hills 2019).
This complex landscape must be investigated by giving structure to knowledge in social discourse, a task achievable by reconstructing conceptual associations between ideas. Cognitive network science This perspective article reviews the most relevant recent results on cognitive networks and links them to social media in order to obtain novel insights on cognition. Cognitive networks provide quantitative readings of people's minds through their language, thus leading to next-generation algorithms and interpretable models capable of using social media data for grasping mechanisms like attention, This article will identify key potential developments of cognitive network science in relation with the above five cognitive dimensions of social media. Particular attention will be devoted also to the possibility of using multilayer networks for encapsulating within the same framework both social ties and cognitive relationships. Data limitations and ethical monitoring will be discussed in view of relevant research.

Cognitive Networks as Models of the Human Mind
The knowledge humans use for producing social media messages is mostly linguistic and resides in the so-called mental lexicon, a cognitive system apt at acquiring, storing, processing and producing conceptual knowledge (Vitevitch 2019). Despite its name, it is no "common" dictionary but rather represents a highly dynamical and structured system of conceptually interconnected ideas, whose access and testing cannot thus be mediated in a lab. Whereas experimenters could physically manipulate a human brain in the lab in order to test its connections, the mental lexicon remains a cognitive construct, with fascinating influence over the mechanisms of information processing (cf.
Vitevitch 2019, . For this reason, the structure of knowledge in the human mind has to be investigated in other ways, like through cognitive tasks stimulating the mental lexicon (e.g. people writing about a topic) or through representations of conceptual knowledge in the mental lexicon. Cognitive networks can combine both these types of indirect access, as they can be built either through cognitive tasks or as representations of specific semantic, orthographic, phonological, syntactic or even visual aspects of knowledge in the mental lexicon ).
Conceptual networks were first proposed as models of knowledge in the human mind by Quillian (1967), in a hierarchical organisation of concepts interconnected when sharing semantic features.
Network distance on such a structure could account for the time it took for participants to rate the validity of simple statements but not other patterns related to meaning negation. Given the scarcity of datasets at that time, cognitive networks were rapidly forgotten

Information on social media: Language is key but multifaceted
Reading sentences, much like this one, activates the mental lexicon and its cognitive structure of interconnected meanings and emotions (Vitevitch 2019). Analogously, the mental lexicon is used by social media users whenever they express their knowledge through online messages. In this way, language represents a powerful bridge between the online world, the way humans express themselves and the way their minds are organised.
A first issue for investigating social discourse is represented by the massive amounts of available data (Krippendorf, 2018). For instance, only on Twitter, online users produce over 6,000 tweets in one second (cf. Brandwatch.com). This deluge of information contains various types of content: written language, hyperlinks, emojis, pictures and videos (see also Figure 1). Even if one discarded pictures and videos, whose automatic processing remains an open challenge, a rich linguistic information would remain in the form of words, hashtags, social jargon and emojis. All these elements provide key semantic and emotional cues used by social users to express themselves. In particular, even though hashtags and emojis are not a language, e.g. they do not satisfy grammatical rules, they can still

Cognitive networks, social media data and the need for interpretability
To achieve next-generation tools suitable for processing and interpreting knowledge, it is fundamental Recently, Kenett and colleagues (2017) showed that semantic network distance outmatched latent semantic analysis in predicting human judgements about semantic relatedness, with the advantage of semantic distance being interpretable through spreading activation (Quillian, 1967).

Cognitive networks and the "text as data" approach
Humans do not explicitly see conceptual associations when reading sentences and yet they are aware of the syntactic and semantic links combining information units, e.g. words, and conferring meaning and emotions to a given sentence (  were capable of detecting key concepts in both annotated short texts and social media data, providing ways of capturing semantic prominence in texts without word frequency.
The above results open promising directions for further using networks of concepts when investigating social cognitions in online platforms, which is briefly reviewed in the following in terms of exploring cognitive phenomena related to salience, perception and biases.

Studying semantic salience beyond frequency counts
Salience is the state of being prominent in a given context. Semantic salience or prominence characterises concepts that are key for individuals to understand a greater extent of knowledge and its influence over cognitive representations is a crucial research direction (cf. Vivas et al. 2020).
Social discourse can provide data useful for understanding those features influencing semantic prominence and ultimately determining key aspects of social discourse.
A key feature for identifying semantic prominence of individual concepts through social media is frequency, i.e. counting how many times a given word occurs. This metric influences a variety of linguistic tasks (cf. Vitevitch 2019) but it is based on concepts in isolation. A word might occur more or less frequently but also always within the same context or not. Frequency alone would not be able to capture these differences. For this reason, n-grams were introduced in the literature as frequency counts of words together with other n-1 contextually related words (Damashek, 1995). n-grams can account for contextual information but also require prior knowledge on how to select words. Keeping n fixed to 2 but considering pairwise associations among more words leads to co-occurrence networks. Further filtering out non-syntactic links leads to syntactic networks. In this way, syntactic networks are an extension of mere frequency counts (1-grams), considering how conceptual relationships are structured across different contexts or semantic areas, i.e. clusters of words with analogous semantic features (Citraro et al. 2020).
Representing social discourse as a syntactic/semantic network can be a quantitative way of estimating concept salience/prominence in social media data beyond frequency (see also Figure 2).
Prominence would be defined in terms of the semantic relatedness of a concept across contexts, with more prominent concepts being closer to and more well connected than others. Stella (2020b) built on previous results of semantic relatedness/closeness being captured by semantic network distance

Give structure to online semantic frames and emotional profiles
If the network structure can be used for achieving models explaining semantic salience in terms of large-scale conceptual associations, can networks and social media data assist also with the study of perception? A quantitative model for studying perception through social media data should be related to understanding how social media users perceive events through their online discussions. Cognitive networks provide a convenient and powerful way for reconstructing quantitatively the semantic patterns framing ideas on social media discours thanks to the theory of frame semantics (cf. Fillmore, 2001). As also highlighted in Figure 1, meaning is cast upon individual words by conceptual associates referring to them. Semantic frame theory posits that the idea and features of a given entity (say "pandemic") can be reconstructed by considering the other words referring to it syntactically and semantically in a given text (e.g. "people", "infection", "hospitals", "populations", "large") (Baker et al. Investigating perception could then be recast into the problem of analysing network neighbourhoods of semantic/syntactic associates around specific topics of social discourse. Pioneering results towards this direction already show that semantic frames can contain key information about the way social users discuss and perceive their experience and beliefs. Stella (2020) found that online discourse framed the idea of the "gender gap", usually highly criticised, in a context rich with positive jargon, evoking emotions of joy and trust and celebrating women's success in science. Figure   Reconstructing semantic frames from syntactic/semantic dependencies in text is an important task also for unearthing how concepts are structured and promoted by fake news and misinformation channels in the "dark side" of information flow (Hills 2019). In their investigation of conspiratorial theories and the 5G on Twitter, Ahmed and colleagues (2020) found that websites promoting fake news were largely reshared by users framing negative stances towards conspiratorial theories.
Despite criticising fake content, these users ultimately promoted fake news through online discussion. mechanisms. An analogous procedure, considering only conceptual revisiting of the same concepts and semantic areas in tweets was introduced by Monakhov (2020) for detecting trolls, i.e. users inflaming social discourse with abrasive language and influencing even massive voting events or public perceptions (Broniatowski et al. 2018). Monakhov (2020) showed that trolls tend to behave like Lydon-Staley and colleagues' "hunters", revisiting only a limited span of conceptual associations, a feature that was used for detecting trolling with an accuracy of 91% but using only 50 tweets per user.

Multilayer networks as socio-cognitive models
Most of the above works focused only on one type of conceptual associations despite the mental lexicon accounting for multiple types or layers of conceptual relationships, e.g. syntactic, semantic,