An overview of existing ontologies related to education can be found in Stancin et al. (2020). Ontologies about education can be categorized into four main categories: curriculum modeling and management, describing learning domains, describing e-learning services, and describing learning data. Some ontologies cover multiple categories at the same time. Of particular interest for this paper are ontologies belonging to the final category. Ontologies describing learning data mainly focus on modeling learners and their related data. LifeOn is a “ubiquitous lifelong learner model ontology” (with a highlight on learner personality) for adaptive learning systems (Nurjanah, 2018). LifeOn represents static (special needs, preferences, profile) and dynamic (learning history and study plan) attributes of learners with academic and non-academic scopes. In Akharraz et al. (2018), an ontology for reasoning about learners and their needs is presented. The goal of the ontology is to help learners realize intended learning goals by allowing exchanging the learners’ profiles between the components of the e-learning system. Han (2018) presents an interdisciplinary ontology-based model aiming to improve students’ cognitive abilities. The model serves as a guide for building teaching plans by helping college educators understand the learning situation of the individual students. An ontology for modeling learning preferences and predicting learning styles is presented in Rami et al. (2018). Grivokostopoulou et al. (2019) present an ontology used to analyze student’s learning and performance within the context of a tutoring system for Artificial Intelligence. The system aims at supporting both students and teachers. Zhao and Guo (2019) present how big data can be leveraged to provide personalized distance learning while avoiding information resource overload. They propose an ontology to model user interests based on user information and preferences. The learning management system presented in Bouquet and Molinari (2016) represents heterogeneous information through semantic knowledge bases to ensure expressivity and reusability. In Rani et al. (2016), an ontology is used to improve data representation and learner classification within a learning management system. However, learner profiling is only done based on demographic information. A multimodal ontology-based school care system targeted toward students with special educational needs is presented in Hafidh et al. (2019). The presented application combines multiple factors such are education, health, and social interaction. The e-learning framework presented in Sarwar et al. (2019) aims at profiling and categorizing learners to provide content recommendations and adaptations. However, the ontology is only utilized to represent the learning content and not for the learner profiling and recommendation process. In Saleena and Srivatsa (2015) and Kaur et al. (2015), e-learning systems integrating multiple ontologies for modeling learner profiles, with the former being more oriented toward learner evaluation and the latter focusing on the learner’s personality.

Several approaches have been proposed for using ontologies when representing and modeling the complex interactions between human behavior, personality, and emotions. The Mental Functioning Ontology (MF) is a general ontology of aspect of mental functioning Hastings et al. (2012). It represents mental processes (including cognition and general traits). The ontology complies and is aligned with the Ontology for General Medical Science (OGMS) (Ceusters and Smith, 2015) and the Basic Formal Ontology (BFO) (Arp et al., 2015) which represents concepts as continuant (e.g., humans, cognitive representations, and dispositions) or occurrent (e.g., processes like behavior) entities. MF is also partially aligned with the Cognitive Paradigm Ontology (Turner and Laird, 2012) and the Cognitive Atlas (Poldrack et al., 2011). The ontology has two extension modules the Mental Disease Ontology (MFOMD) (Ceusters and Smith, 2010; Hastings et al., 2012) and the Emotion Ontology (MFOEM) (Hastings et al., 2011b, a). The former represents mental disorders in alignment with the Disease Ontology. The latter represents emotions and all related affective phenomena. MFOEM models bearers of affective phenomena, different emotion types, moods, as well as their different building blocks, dimensions, physiological and bodily representation markers (e.g., facial expressions). It also represents the relation between affective phenomena and human behavior. MF and its modules support alignment with self-reporting or articulation of emotional states and responses (Hastings et al., 2014; Larsen and Hastings, 2018). EmOCA, an emotion ontology, can be used to reason about philia and phobia based on emotion expression in a context-aware manner (Berthelon and Sander, 2013). EmotionsOnto is another emotions ontology for developing affective applications and detecting emotions (López Gil et al., 2014). In García-Vélez et al. (2018), an ontology of psychological user profiles (mainly personality traits and facets) is presented. A web application for detecting personality using linguistic feature analysis based on ontologies of personality and other techniques is presented in Sewwandi et al. (2017). An ontology for insider threat risk detection and mitigation through individual (personality, affect, ideology, and other similar attributes) and organizational socio-technical factors is presented in Greitzer et al. (2016). The HeLiS ontology (Bailoni et al., 2016; Dragoni et al., 2018) models the concepts representing the food and physical activity domains. Through modeling detailed food properties and physical activity properties, HeLiS supports the construction of intelligent interfaces for domain experts to support a healthy lifestyle and is extended to representing behavior change in this aspect (Dragoni and Tamma, 2019). Another ontology for representing the psychological barriers preventing behavior change and thus enabling overcoming them is presented in Alfaifi et al. (2018). The ontology does that from the perspective of patient behavior in relation to the necessary behavior change interventions for type 2 diabetes.

The above-discussed ontologies each cover one specific domain or factor of human experience and behavior. Crucially, the represented domains and human factors are based on one specific defining theory, e.g., the two-factor theory of emotions or appraisal theory of emotions, in Berthelon and Sander (2013) and Hastings et al. (2014), respectively. However, human behavior and the psychological factors that influence it are much more complex and not just dependent on one single factor. Therefore, ontologies need to have the flexibility to include and combine different psychologically relevant attributes to construe a formal description of the many factors and their inter-relationship and their relevance for behavior.

In line with this, an integrated ontology of human character and its interaction with behavior and situation is necessary to enable the automated description of human behavior in a situation-sensitive manner, considering overt and covert human factors [for an overview, see Herbert et al. (2020), Herbert (2020)].

To accomplish this aim, computer scientists and psychologists need to work together to build a psychology-driven ontology.

Character Computing is a psychologically driven framework that considers an individual’s state and trait markers, represented as character [for an overview, see El Bolock (2020), Herbert (2020), and Herbert et al. (2020)] to predict human behavior in a specific situation [for more research, see El Bolock et al. (2020a)]. The interaction between these components is referred to as the Character – Behavior – Situation (CBS) triad [see El Bolock (2020); Herbert (2020), and Figure 1). As outlined in Herbert (2020), multiple psychological theories define human behavior and its underlying processes; and affecting factors. The inter-disciplinary field of Character Computing aims to validate, enrich, and possibly update our current understanding of human behavior in various contexts, e.g., education, well-being, health, and human computer interaction. This is done through a joint approach focused on experimental validation and computational, heuristic modeling, as shown in Figure 1 and outlined in Herbert et al. (2020). To provide a holistic definition of human behavior, we need to include overt and covert human factors, i.e., cognitive, affective, and motivational state, personality traits, subjective experiences, and socio-cultural embedding, e.g., Herbert et al. (2020). Following an interactionist psychological approach, human factors interacting with the situation to produce variability in behavior are defined as character within Character Computing. This approach accounts for the interactions between behavior, cognition, and emotions, aligning Character Computing with Artificial Intelligence and agent-based, embodied approaches. This differentiates the use of the term character from its common use in psychology, representing biologically based individual differences (temperament) or latent person variables defined by personality traits (Herbert, 2020). As behavior reactions feedback to an individual’s current state, the behavior needs to be dynamically modeled, as enabled by the CBS triad. A psychologically driven computation approach to Character Computing enables the extraction of behavior patterns and indicative features from data to predict certain human behaviors. By defining human behavior as an artifact sensitive to intra- and inter-individual differences (including language and culture) and resulting from the interactions between character and situation, we enable developing human-centered technology solutions.

Computational ontologies are semantic frameworks that structure and represent knowledge of a specific domain as a formal model. Similar to ontologies in Psychology, computational ontology models consist of domain-relevant concepts, the relationships between them, and the conditions constraining them. Computational ontologies have the advantage of providing a unified representation of a specific domain of knowledge that is shareable among humans and computer systems alike. Automated logic-based reasoning enables the inference of new or implicit knowledge from the explicitly defined domain. Ontologies have successfully been used to model mental processing, human emotions, behavior, and behavior change. Ontologies have also been often used for representing academic knowledge and as the underlying knowledge bases of education applications.

Building on existing ventures, including those mentioned above [e.g., Larsen and Hastings (2018), Maimone et al. (2018), Dragoni et al. (2020), and Hastings et al. (2020)], in this manuscript, we present how an ontology is used to identify and structure university student behavior data during the first COVID-19 lockdown (Herbert et al., 2021a).

As described in more detail in the section ‘‘Methods,’’ the already developed ontology, CCOnto¹, represents knowledge about behavior, situation, and character [see El Bolock et al. (2021, 2020c)]. It is, therefore, an ideal framework that can serve as a common ground for psychoeducational research and provide further insights into the involved concepts to be used as a basis for psychology- and technology-based solutions. A set of existing tools can be used to enable psychologists to browse the relevant ontology parts (Lohmann et al., 2016; Dudáš et al., 2018; Ivanova et al., 2019; El Bolock et al., 2020e; Florrence, 2021) and to apply rules to the knowledge represented by the ontology (Heyvaert et al., 2018; Coelho et al., 2019; El Bolock et al., 2020d; Pittl and Fill, 2020).

The data is modeled to categorize the essential variables defining and affecting the emotional well-being of learners. This enables pinpointing underlying causes for individual students’ different emotional states, predicting other students’ emotional responses, and conveying these predictions and understanding to caregivers (educators and family members). The ontology can be coupled with machine learning to explain black-box output predictions of machine learning models.

CCOnto includes all relevant aspects of character- and behavior-based phenomena in an embodied contextual sense represented by the CBS Triad [first introduced in El Bolock (2020)]. This includes trait markers (e.g., personality and socio-cultural embeddings) and state markers (e.g., affect and well-being) that form the individual character, the models representing them, and the different theories on how they impact behavior in specific situations. The recommended best practice ontology development features include modularity, ontology reuse, continuous collaboration, and agile development based on rapid iterative prototyping (Kotis et al., 2020).

CCOnto is modular, allowing the extension of specific concepts depending on the different use-cases, as shown in Figure 2. Our process of providing a model for the Character Computing framework is thus composed of three interconnected parts. The first, taxonomization, aims to collect all involved concepts and processes into a unified vocabulary while aligning them with existing taxonomies and categorizing ontologies. The second, contextualization, aims to extend the taxonomy regarding different contexts (domains) to represent behavior from an embodied perspective as rule-based functions of interactions of character states and traits with situation determinants. The third application aims to ground behavior, trait, state, and situation variables concerning domain-specific applications to harness the extracted contextual behavior concepts, as outlined in El Bolock et al. (2020b). CCOnto builds on existing efforts of ontologies representing mental functioning, emotions, context-awareness, health, and other related concepts. Specifically, the core part of CCOnto, CCOnto-Core, draws on the Mental Functioning Ontology (Smith and Ceusters, 2010), two of its branches: the Emotions Ontology (Hastings et al., 2014, 2011a,2011b), the Mental Disease Ontology (Hastings et al., 2012 and Larsen and Hastings, 2018), and the Emotions for Context Awareness Ontology (Berthelon and Sander, 2013). The relevant concepts are updated to fit CCOnto and combined with our own concepts. CCOnto-Domain integrates (one or more) existing domain ontologies covering the needed extension domains of interest, e.g., Dragoni et al. (2018). Finally, CCOnto-Apps can build domain-specific applications that use CCOnto-Domain for reasoning about concepts within the ontology based on specific use-cases. For example, we would use the proposed approach to assist educators by indicating the anxiety threat levels of different students based on their behavior for early interference.

The ontology is developed following the agile METHONTOLOGY ontology development methodology (Fernández-López et al., 1997). The methodology consists of five iterative steps whose intermediate output is validated and evaluated by the knowledge engineers and the domain experts. The five steps are (1) specification of the ontology purpose and domain, (2) conceptualization of the domain knowledge and reusing existing ontologies, (3) formalization of the conceptual model, (4) implementation of the formal model into a machine-readable format, and (5) maintaining, documenting and publishing the ontology. The development steps are described in detail in El Bolock et al. (2021). The ontology is written in the Web Ontology Language OWL 2.0 (World Wide Web Consortium, 2012) and implemented using the Protégé tool 5.2.1 (Musen, 2015), enabling rapid prototyping.

A schematic overview of the main entities and relations included in CCOnto is shown in Figure 3, where the entities belonging to different existing ontologies are denoted in unique colors.

Table 1 gives an overview of the main class hierarchy. Only a representative sample of the classes and subclasses relevant to the scope of this paper is included. For example, the focus is on behavior and behavior change during the pandemic. The character, behavior, and situation concepts are formalized as a hierarchy of classes related through parent-child relationships.

Character consists of three main subclass clusters, namely characteristic adaptation, mental emotional functioning, and personality. Characteristic adaptations can vary over time, such as appearance or goals. Mental emotional functioning includes all affect-related concepts incorporated from the emotions ontology (Smith and Ceusters, 2010 and Hastings et al., 2012) and readjusted to fit CCOnto. Personality consists of the five FFM traits, trait anxiety, and BIS/BAS traits (Pickering and Corr, 2008). Behavior is represented as scores of specific behavior-related tasks. We distinguish singular, habitual, and learning behavior and belonging tasks. The situation is thus represented as a specific task, environment, or social and learning situations. The coordinate represents continuous, categorical emotions and is reused from EmOCA (Berthelon and Sander, 2013), alongside the stimulus and impact (of a personality trait) classes.

As described under the sections “Methods” and “Aim of the Present Study,” in this manuscript, we apply the CCOnto ontology CCOnto to the specific question of behavior change among university students during COVID-19. Figure 4 gives an overview of the different usage scenarios of the CCOnto ontology. Notably, the added value and purpose of an ontology is to test the consistency of different hypotheses with respect to the knowledge representation model and the dataset. Inconsistencies indicate one of 4 options to be investigated by the knowledge engineers and researchers: (1) incorrect hypothesis, (2) incorrect or inconsistent data, (3) inconsistent representation of the underlying concepts with respect to the dataset, or (4) incorrect underlying theories. The latter would be of particular benefit when comparing two representation theories of different situation, state, or trait markers, e.g., two theories of emotions. Due to being as theory-agnostic as possible, the ontology supports comparing between different theories. This would help indicate whether some theories are more relevant in some cases as opposed to others.

To highlight the basic usage scenario presented in Figure 4, we used the ontology to test the initial hypotheses about mental and behavior change among university students conceptualized and described in Herbert et al. (2021a). The output of the ontology is aligned with the previously conducted statistical analysis in Herbert et al. (2021a). In this manuscript, we present how we evaluate the ontology usage in the context of categorically structuring information about COVID-19 related behavior change. As proof of concept, we rely on the concept and design of the already conducted psychological online survey (Herbert et al., 2021a) conducted during the first COVID-19 pandemic to assess behavior change across psychological domains among university students studying in Egypt or Germany. The survey items assess different psychological aspects of behavior related to teaching and learning, social behavior, mental health, and well-being [please see Herbert et al. (2021a)]. The results of this psychological study showed several interesting findings confirming that university students across cultures and countries have been seriously affected by the COVID-19 pandemic (Herbert et al., 2021a).

Demographics are one primary identifier of human beings. Gender, age, marital status, countries of origin, and residence all shape and identify behavior. For example, a married university student would require different inference rules to determine procrastination. An international student has different representation requirements than a student visiting a university in his/her hometown, especially in the case of a lockdown.

A person’s socio-cultural embeddings dictate how they perceive certain things and, in turn, how they react toward them [e.g., Banks (1988) and Park (2002)]. We denote this as culture, which refers to ethnicity, religion, and nationality for modeling purposes. We do not differentiate between subtypes of culture but rather consider it as one concept. This means that a single individual can have many values for culture representing their ethnicity, religion, and nationality.

Behavior is defined as any “internally coordinated response(s) (actions or inactions) of humans (animals) (individuals or groups) to internal or external stimuli, via a mechanism that involves nervous system activity.” within the Gene Ontology (Ashburner et al., 2000; Carbon et al., 2021). We focus on a more abstracted notion of behavior that does not necessarily focus on the lower-level behavioral functions unless they directly serve to infer higher-level ones, such as physiological responses to emotions and affective representations related to threat perception during the pandemic.

All behavior-related relations in Table 2 can be used to represent information about a single occurrence of behavior, recurring habits, and behavior change with respect to the pandemic. Also, most of them can be regarded as stand-alone well-being-related behaviors. Thus, depending on each input instance, the information is mapped to one or many of the below indicators simultaneously.

Even with recent theories advocating situated-ness (Barsalou, 2019), embodiment (Niedenthal et al., 2005), and environmental psychology (Russell and Ward, 1982), the situation is not often included as a potent representative of behavior. As advocated by Character Computing (El Bolock, 2020), however, the situation is a central concept representing behavior and thus the third member of the CBS triad modeled by CCOnto. We define a situation as any external factors and conditions of an individual. The situation is to be represented by different (non-mutually exclusive) aspects.

Of particular importance to the purposes of this paper are the different learning patterns and behaviors. Learning factors can be interpreted as and are composed of entities belonging to different categories. Representing the variables related to learning and educational behavior thus benefit the most from the ontology-based categorization.

The proposed approach of categorizing data (specifically psychoeducational behavior data) using ontologies has many implications for the domain of psycho-education building on the outlined use-case. Using the ontology representations, we showed how learning is simultaneously represented by state markers, behaviors, and situation factors. This representation mode sheds new insights into understanding learning from a holistic psychological perspective. Learning behavior depends not only on educational factors or learner’s personality but also on learner states, situation, and other behavior. Given that psychoeducational data is currently often multimodal or implicitly included in other data sources using ontologies similar to CCOnto are integral to consolidating results through knowledge-based representation as outlined above.

Two notable implications of the proposed ontology are enabling, (1) knowledge inference and dataset completion in relation to human factors and (2) combining the high predictive power of machine and deep learning with the high interpretative power of ontologies.