Image schemas have been recognized as fundamental cognitive structures that precede language and serve as foundational elements for conceptualization. Extensive research has investigated their involvement in metaphor construction, analogy formation, and conceptual blending. This paper initiates the project of analysing their role in the representation of concepts provided by Prototype Theory. To tackle this problem, we analyse the Leuven Concept Database which is collecting (common sense) information over the features exhibited by 15 concepts. Specifically, our focus lies in investigating the image schematic content embedded within these features and exploring how this content varies across diferent conceptual domains.
In Knowledge Representation (KR), concepts are often assumed as purely extensional, namely as the set of entities they identify. Experimental findings in Cognitive and Experimental Psychology have long since demonstrated that concepts have much more subtle semantics than what is expressible by simply assuming a purely extensional view of concepts.
Diferent, alternative, cognitive theories of concepts have been proposed in the literature to
provide a working definition of what concepts are and how they are represented in the human
mind. Among these theories, the Prototype theory is one of the most formally developed [
One of its central notions is then the one of feature. While the notion of concept is well
discussed in the literature, both philosophical and psychological [
Notwithstanding this, Prototype Theory suggests representing concepts as a list of features, meant as the characteristics, attributes, or properties usually found across the instances of the concepts. For instance, a Pet may be described as an animal, furry, having an owner: all of these descriptions may be considered features of the concept Pet. Moreover, according to the Prototype Theory, these features have diferent ‘importance’ in the description of the concept. The importance is often expressed through ‘weights’ (i.e. numbers) these features are associated with. For example, the concept Pet could associate a higher importance weight with the feature “having fur” and a lower weight with “having scales”. Nevertheless, the list of features can include (intuitively) contradictory features, allowing in this way the representation of the variability of the category (let us consider again the iguana pet example). Similarly, this explains many of the so-called typicality efect : exemplars sharing more features with the prototype would be more typical, and exemplars sharing a few numbers of features would instead be atypical.
While the Prototype view is widely accepted as a prominent theory for representing concepts, it is not without its share of criticisms, also because of the naive use of the notion of feature.
According to Lakof [ 13, p. 133] “[...] the properties that are relevant for the characterization of human categories are not objectively existing properties that are ‘out there’ in the world. Rather they are ‘interactional properties,’ what we understand as properties by virtue of our interactive functioning in our environment.” Moreover, (ibid. p.135) “Prototype efects are real, but superficial. They may arise from a variety of sources. It is important not to confuse prototype efects with the structure of the category as given by cognitive models.”
Proponents of the Prototype theory argue, contra this kind of critique, that it is important to
distinguish between the theory itself and the operational methodology employed to support it
[
However, according to Hampton: “It is not supposed for example that the mind contains a list
of attributes in the verbal form in which they are generated. Clearly meaning has to be grounded
[...] in experiential sub-symbolic levels of cognition, so it is unhelpful for a psychological model
to give the meaning of one word simply in terms of others unless there is a primitive base of
terms that are defined non-verbally.”[
This is nothing but another variant of the Symbol Grounding Problem, which in AI refers to the issue of connecting symbols, such as words or abstract representations, with their corresponding meanings or referents in the real world. In the context of prototypes’ representation, the problem lies in bridging the gap between the symbolic representations (reified in the list of features which describe the concepts) and the actual sensory experiences or physical entities they are meant to represent. Hampton’s reference to “experiential sub-symbolic levels of cognition” points here to embodied experience as a lower-level source of the representation which is presupposed by Prototype Theory (and mainstream cognitive theories of concepts in general).
Image Schemas refer to recurring patterns of perceptual and embodied experiences that shape
our understanding of the world. As such, they are often been pointed to as a possible way of
solving the Symbol Grounding Problem [
To start tackling this question, in this paper we analyse the content of the Leuven Concept
Database [
While there exist works aiming at automatically extracting Image Schemas from texts [
The Leuven Concept Database (LCD) is a large dataset that links sets of features to concepts (or category labels) and exemplars (or lexical entries). It was compiled by the ConCat group at the University of Leuven between 2004 and 2008. This database consists of 15 categories (with a total of 420 associated exemplars) and 500 associated features. Specifically, it covers the animal domain: birds, fish, insects, mammals, reptiles, and amphibians. It also includes information on the artefact domain: musical instruments, tools, vehicles, clothing, kitchen utensils, and weapons. Additionally, there are categories for fruits and vegetables, as well as activities, such as professions and sports. The data collection involved the participation of at least a thousand students.
Although the material was collected in Dutch, an English translation is provided to facilitate further experimental and modelling approaches.
The studies conducted at the University of Leuven contribute to the ongoing debate between
the Prototype Theory [
The participants’ responses to the feature generation task were manually compiled and adjusted with minimal stemming into 15 distinct Excel tables (one for each category). The frequency of feature production was recorded as an indirect measure of their importance. Moreover, participants were directly asked to rate the importance of each feature in defining the concept it was associated with, on a rating scale ranging from +3 (very important feature) to -3 (very unimportant feature). The ratings assigned to each feature by the participants of the experiments were collected in each table of the database, leading to a representation of concepts very close to the one endorsed by proponents of the Prototype Theory.
In previous work [21], we presented an in-depth case study of the Leuven concept database (LCD) to transform the common-sense knowledge therein into a format that could be efectively utilised for practical applications. More specifically, the goal of the paper was to translate the LCD into OWL (the most widespread language for authoring ontologies [22]) and to use the ontologies built upon it in a concrete computational implementation for concept combination [23, 24]. To achieve this level of formalisation, we explored a hybrid approach that involved analyzing the syntactic structure of the LCD entries along with conducting semantic and ontological analyses. More specifically, we employed notions from Foundational Ontology (FO): foundational or upper ontologies establish the precise definitions of highly general terms, such as object, event, property, quality, relation, process, etc. FO define the thus fundamental categories that are applicable across multiple domains and are inherently embedded within common sense knowledge. These high-level, ontological distinctions, were used, in our case study, both to disambiguate the meaning of some of the features described in the LCD and to inform some of the formalisation choices in the process of translation into OWL.
In the above-mentioned paper [21], we individuated 7 modes of disambiguation which could guide our translation of the LCD: the distinction between Rigid and Non-Rigid properties [25], capturing the diference between properties which are always inherently true of one class, and properties which may change over time; Mereology, namely the theory of part-whole relations; Constitution, capturing the features expressing information on material substances; Quality, helping the formalisation of quality spaces such as size, shape, colour, etc; Action and Ability, ascribing abilities to agents; Functionality, often employed in the artefact domain to capture objects’ purposes and afordances; and finally, we took into account other facets of core knowledge that are grounded in cognition, such as spatiotemporal relations. Among them, we also discussed the Image Schema CONTAINMENT, but we did not carry out a thorough analysis of the other Image Schemas involved in the content of the Database.
In this paper, we aim to go a step further and analyse the features contained in the LCD by studying the image-schematic information they codify. As we already carried out an ontological analysis of the LCD, this will also allow for some consideration of the relationship between Image Schemas and ontological distinctions previously identified.
Image schemas are primitive, cognitive structures that arise from our interactions with the
world, shaping our understanding of reality since childhood. Image schemas are derived from
our sensory and motor interactions and provide the building blocks for abstract reasoning and
conceptual knowledge. These recurring patterns of perception and embodied experiences are
thus foundational elements for conceptualization. Despite a long line of research trying to
give a comprehensive classification, there is still no common agreement on the list of Image
Schemas [26]. In this paper, we manually analyse the content of the LCD by considering the
list of image schemas proposed by Hedblom [
As mentioned above, the LCD contains 500 features, distributed among 15 concepts.
Nevertheless, some features generated during the experiments posed challenges in their interpretation within the context of defining the concept. Some of the features were either blatantly false when applied to the entire class or connected to a semantic context completely unrelated to the experiment’s intended purpose (e.g., a Fish was described as ”a constellation”). Furthermore, certain features captured language and societal biases, such as a Profession being perceived as ”diferent for men and women,” and a Kitchen Utensil being labelled as ”especially used by women,” while a Tool was considered ”primarily used by men.”
As previously mentioned, all generated features underwent an evaluation to assess their relevance to the respective class (see again Section 2.1). Following what was done in [21], and inspired by the concept of salience in Prototype theory, to exclude possibly controversial features, the means of subjects’ judgments were calculated. Entries falling strictly below the threshold of 0 were excluded. This evaluation procedure resulted in the exclusion of 103 features, (accounting for approximately 20% of the total features).
As a result, we retained a total of 397 features that spanned 15 diferent categories. We proceeded to manually annotate all of these features, identifying and documenting any Image Schemas from the previously mentioned list we were able to identify as hidden in the expression of the feature. Figure 3.1 shows an example for the concept Bird.
Out of these 397 features, we successfully categorized 228 features based on image-schematic information. In other words, approximately 40% of the features defied classification within the chosen Image Schemas.
This percentage is however not equally distributed across the diferent entries of the database. As already mentioned, we can distinguish between 4 macro-areas, or domains, in the database: the Animal domain (including Bird, Mammal, Fish, Insect, and Reptile, for a total of 141 features); the Artefact domain (including Vehicle, Kitchen Utensil, Clothing, Tool, Musical Instrument, Weapon, for a total of 155 features); the Plant domain (including Vegetable, Fruit, for a total 53 features); Activity domain (including Profession, Sport, for a total of 48 features).
In the case of the Animal domain, we were able to categorise into image-schematic terms 72% of the features. For the Artefact domain, the percentage is lower, and it goes to 54%, and, similarly in the Plant domain is 58%. The most dificult domain resulted to be that of Activities, where the percentage of classified features drop to 27%.
We can now consider how the selected Image Schemas are distributed across the Database. Clearly, as complex features are involved in the description of the concepts, often the classification did not satisfy a unique Image Schema, but rather a collection (or profile) of Image Schemas [27]. We will present the percentages for each Image Schema individually in the following. 1. CONTAINMENT represents the connection between the interior and exterior of an entity, along with the boundary that separates them [28]. In our analysis we counted as CONTAINMENT also its dynamic counterparts IN and OUT. Based on our analysis, we found that approximately 14% of the features in the Database can be categorized as involving a CONTAINMENT relationship. Breaking it down by domains, there were 24 out of 141 features in the Animal domain (e.g. Birds ‘eat worms’, Fishes ‘contain omega 3’), 16 out of 155 in the Artefact domain (e.g. Kitchen Utensils are ‘used to store food’, Clothes are ‘found in a clothing cupboard’, Weapons ‘contain bullets’), 12 out of 53 in the Plant domain (e.g. Fruit ’contains seeds’), and 4 out of 48 in the Activity domain (e.g.
Sport ’is practised indoor’) that were classified based on this primitive.
2. SOURCE_PATH_GOAL pertains to the movement from a source and directed towards a
goal. Around 11% of the features in the database were classified as implying it, mostly in
the Artefact domain: 30 features out of 155 involved SOURCE_PATH_GOAL (many of
them in the Vehicle category: e.g. with a Vehicle ‘you can transport things’, or a Vehicle
‘can be used to get from one place to another’). In the other domains, the number of
features captured by this Image Schema drops: only 4 out of 141 features in the Animal
domain (e.g. Reptiles ‘descend from dinosaurs); 5 out of 53 in the Plant domain (Fruit
is ‘used for making fruit salad’); 7 out of 48 in the Activities domain (mostly taking a
metaphorical reading), resulting in being the most frequent Image Schema identified in
this domain (e.g. Sport ‘makes you sweat’, a Profession ‘earns you money’).
3. CYCLE is the Image Schema of recurring patterns (such as seasons or daily cycles). For
this reason, it is maybe not surprising that it was found to be especially frequent in the
Plant domain (13 out of 53 features, in cases such as Fruit ‘is seasonal’ and ’rots after
a while’, or Vegetables ‘grow in the garden’). Globally, around 9% of the features were
classified as involving a CYCLE.
4. CONTACT in its easiest variant refers to physical contact between two objects, but may
also involve abstract contact among entities [
We did not find incarnations of VERTICALITY, NEAR_FAR, and ATTRACTION in the Leuven Concept Database, although a number of features may imply some of them (see below).
The analysis presented above has shown that the role of Image Schemas varies across the diferent domains. For instance, we have noticed how their role is predominant in describing the Animal domain, whereas most of the features describing the Activities escaped a classification in terms of image schematic information. There exist diferent possible explanations for this. First, the Activities domain (hence the concepts of Profession and Sport) is arguably more complex to describe than the Animal one, as it is more dificult to instantiate at the basic level the involved categories, and this leads to the use of more complex, or at least more inventive, features in describing them. Identifying the Image Schemas involved in such descriptions may thus require a more in-depth metaphorical analysis of the involved terminology. Second, and aside from these practical dificulties, this efect may also reflect the diferent amount of background, contextual and extra-bodily knowledge which is employed when describing concepts such as Sports and Professions (e.g. a Profession ‘requires a certain education’, or a Sport ‘requires motivation’).
In other cases, the features employed to describe the concepts imply the reference to embodied experience but are of dificult classification with the list of image schema identified for the task. Examples are features referring to sounds (Vehicles ‘produce noise’), colour (Vegetables ‘are green’), or even emotions (Sport ‘is fun’, or Mosquitos ‘make an extremely irritating sound’). Although, in many cases, there is not an obvious direct connection to primitive spatial structures, there is however space for a less straightforward reading of some of the selected Image Schemas, which may allow for an interpretation of some of the above-mentioned features [26]. VERTICALITY, for instance, can allow a metaphorical interpretation of emotions: good emotions are UP in the scale, while bad emotions drag you DOWN [30].
Another interesting point put in evidence by our analysis is the possibility to identify negation of Image Schemas among the features of a concept. An example is the negation of the Image Schema SELF-MOVEMENT in cases such as Vehicles ‘are operated’ or Mammals ‘can’t fly’ (poor bats). The possibility of considering negation of Image Schemas paves the way for interesting lines of research, although may sound controversial, as the negation is not easy to identify as a conceptual primitive.
In conclusion, it is noteworthy to establish a connection between the current study and the ontological analysis conducted in prior research (Righetti, 2022). Several Image Schemas chosen for our investigation align closely with the ontological “modes of disambiguation” proposed in the same work (Righetti, 2022). For instance, the PART_WHOLE schema is easily associated with the mode of ‘Mereology’, providing a clear and evident link. Additionally, the relationship between the ‘Functionality’ mode and the SOURCE_PATH_GOAL schema might be less apparent, but they often run parallel in our analysis of the Leuven Concept Database. A diferent story is about the mode ‘Quality’, encompassing colour, taste and sounds in our ontological analysis, for which was often much more dificult (if not impossible) to identify a corresponding Image Schema.
In conclusion, this paper aimed to address the question of whether the features contained in the Leuven Concept Database (LCD) are associated with image-schematic notions. The purpose was to start tackling the question of whether and in what way we can ground the notion of feature, as used in Prototype Theory, into bodily experience and conceptual primitives such as Image Schemas. One immediate application is to have a better handle on which kinds of features are essential in formal approaches to giving prototype definitions, as e.g. [ 31], and specifically, what is the contribution of image-schematic features in such definitions.
According to our analysis, around 60% of the features gathered in the database can be described in terms of image-schematic information. Moreover, a number of additional features which escaped our classification was still traceable to bodily experiences (e.g. emotions, sound experiences, etc).
Although there is empirical evidence of high agreement when people have to identify image schematic information within text or images [32], clearly this is only a preliminary analysis which will require further study and validation. Consider the feature: Fruit ‘grows on trees’. How many Image Schemas does it imply? How many would agree on the fact that the growing of an apple would imply the SCALING schema? Possibly many. How many would agree on the fact that growing implies the SELF_MOVEMENT schema? Possibly less. The features contained in the LCD can be subject to various interpretations, and the level of strictness in these interpretations may yield slightly diferent outcomes wrt our analysis. Conducting a more thorough analysis that addresses these concerns is an avenue for future research. [19] S. De Giorgis, A. Gangemi, D. Gromann, Imageschemanet: A framester graph for embodied commonsense knowledge, Semantic Web (????) 1–25. [20] R. M. Nosofsky, Similarity, frequency, and category representations., Journal of Experimental Psychology: learning, memory, and cognition 14 (1988) 54. [21] G. Righetti, O. Kutz, D. Porello, N. Troquard, A game of essence and serendipity: Superb owls vs. cooking-woodpeckers, in: M. M. Hedblom, A. A. Kantosalo, R. Confalonieri, O. Kutz, T. Veale (Eds.), Proceedings of the 13th International Conference on Computational Creativity, Bozen-Bolzano, Italy, June 27 - July 1, 2022, Association for Computational Creativity (ACC), 2022, pp. 300–309. URL: http://computationalcreativity.net/iccc22/papers/ ICCC-2022_paper_88.pdf. [22] G. Antoniou, F. v. Harmelen, Web ontology language: Owl, Handbook on ontologies (2009) 91–110. [23] G. Righetti, D. Porello, N. Troquard, O. Kutz, M. Hedblom, P. Galliani, Asymmetric hybrids: Dialogues for computational concept combination, in: Formal Ontology in Information Systems: Proceedings of the 12th International Conference (FOIS 2021), IOS Press, 2021. [24] G. Righetti, D. Porello, N. Troquard, O. Kutz, M. M. Hedblom, P. Galliani, Asymmetric hybrids: Dialogues for computational concept combination (extended abstract), in: L. D. Raedt (Ed.), Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI 2022, Vienna, Austria, 23-29 July 2022, ijcai.org, 2022, pp. 5329–5333.
URL: https://doi.org/10.24963/ijcai.2022/745. doi:1 0 . 2 4 9 6 3 / i j c a i . 2 0 2 2 / 7 4 5 . [25] N. Guarino, C. A. Welty, An overview of ontoclean, Handbook on ontologies (2009) 201–220. [26] J. M. Mandler, C. P. Cánovas, On defining image schemas, Language and cognition 6 (2014) 510–532. [27] M. M. Hedblom, O. Kutz, R. Peñaloza, G. Guizzardi, Image schema combinations and complex events, KI-Künstliche Intelligenz 33 (2019) 279–291. [28] M. M. Hedblom, D. Gromann, O. Kutz, In, out and through: Formalising some dynamic aspects of the image schema containment, in: Proceedings of the 33rd Annual ACM Symposium on Applied Computing, 2018, pp. 918–925. [29] F. Santibáñez, The object image-schema and other dependent schemas, Atlantis (2002) 183–201. [30] M. M. Hedblom, O. Kutz, Hidden Meaning: Using Image Schema Violations to Expose Hidden Semantical Structures in Metaphors, in: Proceedings of the Joint Ontology Workshops (JOWO 2019), Episode V: The Styrian Autumn of Ontology, volume 2518, CEUR-WS, 2019. [31] G. Sacco, L. Bozzato, O. Kutz, Introducing Weighted Prototypes in Description Logics for Defeasible Reasoning, in: Proceedings of the 38th Italian Conference on Computational Logic (CILC 2023), CEUR-WS, 2023. [32] M. M. Hedblom, O. Kutz, Conceptual puzzle pieces: An image schema experiment on object conceptualisation, in: Modeling and Using Context: 11th International and Interdisciplinary Conference, CONTEXT 2019, Trento, Italy, November 20–22, 2019, Proceedings 11, Springer, 2019, pp. 98–111.