ResearchPDF Available

Abstract and Figures

Image schemas are recognised as a fundamental ingredient in human cognition and creative thought. They have been studied extensively in areas such as cognitive linguistics. With the goal of exploring their potential role in computational creative systems, we here study the viability of the idea to formalise image schemas as a set of interlinked theories. We discuss in particular a selection of image schemas related to the notion of ‘path’, and show how they can be mapped to a formalised family of microtheories reflecting the different aspects of path following. Finally, we illustrate the potential of this approach in the area of concept invention, namely by providing several examples illustrating in detail in what way formalised image schema families support the computational modelling of conceptual blending.
Content may be subject to copyright.
Journal of Artificial General Intelligence 6(1) 21-54, 2015 Submitted 2015-08-16
DOI: 10.1515/jagi-2015-0003 Accepted 2015-11-26
Choosing the Right Path:
Image Schema Theory as a Foundation for Concept Invention
Department for Knowledge and Language Engineering (IWS)
Otto-von-Guericke University of Magdeburg, Germany
Research Centre for Knowledge and Data (KRDB)
Free University of Bozen-Bolzano, Italy
Department for Knowledge and Language Engineering (IWS)
Otto-von-Guericke University of Magdeburg, Germany
Editor: Tarek R. Besold, Kai-Uwe K ¨
uhnberger, Tony Veale
Image schemas are recognised as a fundamental ingredient in human cognition and creative
thought. They have been studied extensively in areas such as cognitive linguistics. With the goal
of exploring their potential role in computational creative systems, we here study the viability of
the idea to formalise image schemas as a set of interlinked theories. We discuss in particular a
selection of image schemas related to the notion of ‘path’, and show how they can be mapped to
a formalised family of microtheories reflecting the different aspects of path following. Finally,
we illustrate the potential of this approach in the area of concept invention, namely by providing
several examples illustrating in detail in what way formalised image schema families support the
computational modelling of conceptual blending.
Keywords: image schemas, grounded cognition, computational creativity, concept invention,
conceptual blending
1. Introduction
The cognitive processes underlying concept invention are still largely unexplored ground, although
promising theories have been developed within the last few decades. Two of the most influential
directions are the theory of grounded cognition (Barsalou, 2008) and the embodied mind theory
(Lakoff and Johnson, 1999; Gallese and Lakoff, 2005). Both propose that human cognition is
grounded in our bodily experience with the environment. In the case of grounded cognition, mental
representations are thought to be derived from the embodied experience that is used to structure
concepts, including even the most abstract ones. Embodied cognition takes this one step further by
arguing against mental representations and by claiming that concepts are (or may be identified with)
the neural activation of embodied experiences.
Following the reasoning of how embodied experiences shape our cognition, a more specific
theory of concept formation and language understanding was introduced under the name of image
schemas. It is a theory that focuses on basic spatial cognition. The theory was originally developed
. This paper is a revised and expanded version of (Hedblom, Kutz, and Neuhaus, 2015).
Published by De Gruyter Open under the Creative Commons Attribution 3.0 License.
in the late 80s by Lakoff (1987) and Johnson (1987), and was quickly taken up by other researchers
in the area (e.g. Mandler (1992)).
According to Johnson (1987) “an image schema is a recurring dynamic pattern of our
perceptual interaction and motor programs that gives coherence and structure to our experience.
The ‘image schema’ is thought to be the abstracted spatial pattern from repeated sensorimotor
experience. These mental structures offer a foundation and a way to ground other cognitive
phenomena, such as language capacity, understanding, and reasoning. They offer a connection
between the bodily experienced relationships of physical objects in time and space with the internal
conceptual world of an agent. In language, they can be seen as the conceptual building blocks
for metaphoric and abstract thought. Some of the most commonly mentioned examples of image
Another important approach towards an understanding of concept invention is the theory of
conceptual blending, introduced by Fauconnier and Turner (1998), which developed further the idea
of bisociation introduced by the psychologist Koestler (1964). Conceptual blending proposes that
novel concepts arise from a selective combination of previously known information (see Section 5.1
for a more thorough introduction).
The primary goal of our research is to develop a computational system for concept invention by
combining a formal representation of image schemas with the framework of conceptual blending.
One of the major obstacles for implementing such a system is that image schemas are, typically, not
crisply defined in the literature, but rather presented as mouldable concepts. Their adaptability
is indeed part of the explanatory success of image schemas. Hence, to realise our goal of a
computational system using image schemas in conceptual blending, we need to develop a formal
representation of image schemas that captures their inherent complexity.
In this paper we suggest that image schemas should be considered as members of tightly
connected image schema families, where the connecting relation is based on the notion of family
resemblance. In particular, each of the image schemas covers a particular conceptual-cognitive
scenario within the scope of the schema family. An image schema family may be formally
represented as a set (i.e. a family) of interlinked theories.
To illustrate our approach, we will use the image schema of MOVE ME NT ALONG PATH,
analyse its use in natural language, and sketch a representation of MOVE ME NT ALONG PATH in
first-order logic, which focusses on the different branching points of microtheories involved in the
family. Afterwards, we show how the image schemas may be used for conceptual blending.
The remainder of the paper is structured as follows: In Section 2, a more detailed account of
image schemas is presented by discussing their internal structure and their role in language. In
Section 3, the PATH-following is discussed in more detail. In Section 4, formal approaches to
image schemas are discussed and we introduce our idea of how to gather image schemas in families
of theories by using PATH-following as a proof of concept. This is done by representing PATH-
following as a DOL-graph and as a FOL-axiomatisation. In Section 5, we investigate the role image
schemas can play in concept invention within the framework of conceptual blending theory. Finally,
in Sections 6 and 7, a discussion and a short conclusion are provided.
1. This image schema is also referred to as S OURCE PATH GOAL schema. For reasons that will become more obvious
in Section 4.2, we use MOVEM EN T ALO NG PATH as the most generic term for this image schema.
2. Image schemas
2.1 Image schemas and embodiment
The theory of image schemas stems from the theory of embodied/grounded cognition. It is a theory
that emphasises the role of bodily experiences as a source for cognitive capacities. It has become
increasingly supported by findings in cognitive linguistics and neuroscience (e.g. Tettamanti et al.
(2005); Gallese and Lakoff (2005); Feldman and Narayanan (2004); Wilson and Gibbs (2007);
Louwerse and Jeuniaux (2010)).
The theory offers an interesting view of cognition for approaches to artificial intelligence
as it provides a more direct route to computational cognition than traditional, more hard-coded
approaches. I.e., artificial agents are encouraged to learn, and artificial cognitive structures are
populated by their ‘experiences’, similar to the learning process observed in human children (see
(Chrisley, 2003) for an introduction to embodied artificial intelligence and (Vernon, 2014) for an
overview of such cognitive architectures).
Building on grounded cognition, image schemas are thought to be the mental representations
extracted from bodily experiences, and more specifically, experiences that can be described using
basic spatial relationships. Image schemas are therefore mental abstractions of learnt spatial
relationships (e.g. CONTAINMENT or SUPPO RT).
In the early stages of cognitive development, these image schemas are formed and ‘fine-tuned’
as the experiences with a particular spatial relationship are increased and extended to different
situations (Mandler, 2008; Rohrer, 2005). Due to this fine-tuning, it appears prominent that image
schemas consist of different ‘parts’ (Mandler, 2004). These ‘parts’ can either be removed or added
while still capturing the same basic image schema, generating what can be described as an image
schema ‘family’. Mandler and Pag´
an C´
anovas (2014) refer to these ‘parts’ as spatial primitives; the
fundamental spatial building blocks. As this view is essential to our approach for formalising image
schemas we will regularly return to the notion of spatial primitives in later sections.
The explanatory value of image schemas lies in their function as abstracted spatial relationships.
It is believed that they contain vital information for the understanding of concepts and their
conceptual neighbourhoods. For example, to properly understand what a ‘cup’ is, an infant needs
to learn that cup is kind of CONTAINER, and, thus, other objects may stand in CON TAI NE R-typical
spatial relationships to a cup. (E.g., an object can be in the cup.) Respectively, the notion of
a ‘table’ needs to be connected to a SUPPO RT image schema since otherwise the child will not
understand that objects remain on tables after being placed on them. This way, image schemas
map affordances to objects (in the sense of Affordance Theory (Gibson, 1977)) and can be used to
explain increasingly more complicated concepts. This is done through information transfer and can
be observed in natural language, for example in conceptual metaphors.
2.2 Information transfer and conceptual metaphor
Image schemas are a abstracted spatial relationships that are associated with affordances. Their
cognitive benefit is that they may map affordances to objects, which an agent has not encountered
yet. For example, if the image schema of SUPP ORT has been learnt through perceptual exposure
of ‘plates on tables’, an infant can infer that table-like objects such as ‘desks’ also have the
SUP PO RT image schema and can SUP PO RT objects such as ‘books’ as well. As the environment
becomes increasingly complex for the infant, this information transfer becomes a fundamental part
of cognition.
Information transfer in language is often done through conceptual metaphors. Information is
moved from one known source domain to an unknown target domain. Conceptual metaphors can
be further specialised to image-schematic metaphors. These are the metaphors that do not transfer
general conceptual knowledge from one domain to another but the skeletal structure of the image
schema (K¨
ovecses, 2010). Here the source domain in the analogical transfer is stripped down to
the image schema skeleton which is mapped to the target domain and there fleshed out with local
domain information. In sections 3 and 5 we will discuss several examples of image-schematic
metaphors (and similes), which are the result of conceptual blending.
As language develops and the individual is exposed to increasingly more abstract concepts
than those found in early infancy, image schemas can be used to ground the novel concepts in
already comprehended concepts. Embodied experiences and image schemas are often used in
natural language to explain abstract concepts. For example, in a social hierarchy people can be
either ‘above’ or ‘below’ us, expressions learned from embodied experiences of the image schema
ABOV E, which itself derives from experiencing the human body’s vertical axis. In natural language,
metaphors such as ‘falling from grace’ or ‘the rise to power’ use the same image schema to represent
status and success.
One of the most well-known examples of an attempt at grounding abstract concepts in image
schemas is the work of Lakoff and N´
nez (2000). In Where Mathematics Comes From, they
defend the view that image schemas lay the foundation for abstract concepts in mathematics.
They explain how the notions of addition and subtraction can be traced from back and forward
MOVE MENT ALONG PATH, and extend the reasoning to more abstract constructions such as
complex numbers. While being influential work, it has also received heavy criticism, in particular
targeted at vague terminology and methodology (e.g. Goldin (2001); Schiralli and Sinclair (2003))
and mathematical errors (e.g. Voorhees (2004)).
The image schema CONTAINMENT is commonly described as the sum of the interrelationships
of an inside, an outside and a boundary (Lakoff, 1987). An abstract example of the image schema
CONTAINMENT is the conceptual metaphor “to be in love”. Obviously, there is no spatial region
for the emotional state of love in the same sense as there is for a physical container such as a cup.
Yet, we use the spatial language to talk about the phenomenon of love: e.g. we ‘fall in love’ or ‘fall
out of love’.
There is a clear connection between CONTAINMENT and prepositions such as ‘in’, ‘into’ and
‘out of’. Bennett and Cialone (2014) investigated the CONTAINMENT relationship by searching text
corpora for words similar to containment, e.g. ‘surrounding’ and ‘enclosing’. The authors’ method
distinguished eight different kinds of CONTAINMENT.
Prepositions in combination with verbs often do appear to be the key words that help identify
image schemas in language (Johanson and Papafragou, 2014). Below we will discuss natural
language and conceptual metaphors of the PATH-following image schema family.
2.3 Image schemas and their structure
In this section we, look more closely at how image schemas are thought to be structured. The
first pertinent distinction is that image schemas can be both static and dynamic. For example,
CONTAINMENT can either describe the situation in which the cup already contains coffee, or
alternatively the situation in which the coffee is poured from a source: a kettle, to a goal: a cup,
defined as an INand OUT schema. The static image schemas can in turn be differentiated under three
different categories: orientational (e.g. ABOVE), topological (e.g. CON TACT), and force-dynamic
(e.g. SU PP ORT) (Lakoff and N´
nez, 2000).
A more dynamic way to understand the INand OUT schema is to view them as combinations of
the two image schemas CONTAINMENT and MOV EM ENT ALONG PAT H, building on the idea that
image schemas can be combined with one another to generate more specific and complex image
schemas (Kuhn, 2007; Walton and Worboys, 2009; Mandler and Pag´
an C´
anovas, 2014; Oakley,
2010). Another example is how MOVE ME NT AL ONG PATH easily can be connected with the image
schema LINK resulting in the higher level image schema LINKED PAT H: The image schema concept
that encompasses linked behaviour on two, or more, joint paths. This “Gestalt” grouping of image
schemas means that there must be a distinction between the most perceptually primitive image
schemas and the more complex image schemas.
In language, this corresponds to the observation that combinations of image schemas are
suitable to describe more complex concepts. To illustrate this, Kuhn (2007) suggested that
‘transportation’ can be understood as a combination of the image schemas of SUPPO RT and
MOVE MENT ALONG PATH, and Mandler (2004) suggested that ‘marriage’ can be viewed as a
One proposal to hierarchically structure the range from simple to more complex and dynamic
image schemas is the approach presented in (Mandler and Pag´
an C´
anovas, 2014), which builds
on empirical data from studies on cognitive development. In their work, the umbrella term
‘image schemas’ is divided into three different levels: spatial primitives2(the conceptual building
blocks build from spatial information), image schemas3(simple spatial stories), and conceptual
integrations (image schemas combined with a non-spatial element such as force or emotion).4
One major advantage of partitioning image schemas into spatial primitives, spatial schemas, and
conceptual integration is that it enables a more fine-grained analysis of connections between image
schemas. We believe that the change from one spatial schema to another can be accomplished by
providing or detailing more spatial information, e.g., by adding additional spatial primitives. In
Section 4.2, we will demonstrate this idea by using PATH-following as a proof of concept. In the
next section PATH-following will be introduced in more detail.
3. The image schema family PATH-following
3.1 Introducing PATH-following
In this section, we will explain how image schemas, like MOVEM EN T ALONG PATH, are members
of image schema families. For this purpose we introduce the PATH-following image schema family
2. The notion of spatial primitives is not novel. Research on such semantic building blocks can be found in the linguistic
literature. E.g. the work on spatial semantics by Veale and Keane (1992) and the more general work on semantic
primes that covers more than the spatial and temporal aspects found in spatial primitives (Wierzbicka, 1996).
3. When referring to this concept we will use the term spatial schemas to avoid ambiguity between the wider notion of
‘image schema’ and its narrower sense introduced by Mandler and Pag´
an C´
anovas (2014).
4. For the purposes of this paper, only spatial primitives and spatial schemas will be further discussed. In principle,
our approach is general enough to allow for heterogeneity, also on the logical level. Therefore one may also
include conceptual integrations involving non-spatial elements in our image schema families, cf. the discussion in
Sections 4.2 and 4.3.
and illustrate how the family is organised hierarchically from general to more specific versions of
PAT H-following by the addition of spatial information and primitives.
MOVE MENT ALONG PATH is one of the first image schemas to be acquired in early infancy as
children are immediately exposed to movement from a range of objects. This, in combination
with the neurological priority to process moving objects over static objects, suggests that the
image schema is either innate or learnt at a very early stage in cognitive development (Rohrer,
2005). However, in order to understand how the PATH-following family is fine-tuned and in
‘more completion’ internally structured, experiments with children have provided some insights
on distinguishing how the different spatial schemas may develop.
Firstly, already at an early age children pay more attention to moving objects than resting
objects. Trivial as it may seem, it requires children to detect the spatial primitive OBJECT (or
THI NG) and the spatial schema MOVEME NT OFOBJECT.5Secondly, children tend to remember
the PAT H of the movement of the object. The PATH is a spatial primitive, which is different from
the movement and the moving object.6.
In addition to these two basic spatial primitives and as the child becomes more and more familiar
with PAT H-following, image schemas that contain more spatial information are learned. This means
that in more advanced stages, image schemas may include beyond MOV EM ENT OFOBJECT and
the spatial PATH itself also the spatial primitive END PAT H, and later also a START PATH (Mandler
and Pag´
an C´
anovas, 2014). Already at five months infants can distinguish PATH-following that has
an EN D PATH (the image schema PATH GOAL) from the initial PATH, while the S TART PATH is
less interesting until the end of the first year of life. This is further supported by linguistic analyses
in which an END PATH is initially more interesting than a START PATH (Johanson and Papafragou,
Table 1 summarises the spatial primitives that may be involved in image schemas of the PATH-
following family.7
Spatial primitive Description
OBJECT an object
PAT H the path the object moves along
START PATH the initial location
END PATH the final location
Table 1: Spatial primitives of the PAT H-following family
A more specified example of the PATH-following family is presented by Lakoff and N´
(2000). In accordance with other linguistic literature on image schemas they are focussed on the
SOURCE PAT H GOAL schema, see Figure 1. Here, the object, called trajector, moves from a source
to a goal. END PAT H and START PAT H are not identical to the SOURCE and GOA L found in the
5. OBJECT is understood here in a very wide sense that includes not only solid material objects but entities like waves
on a pond or shadows. Mandler and Pag ´
an C´
anovas (2014) also discuss MOV E as a spatial primitive of its own.
We consider MOVE ME NT OFOBJECT to be a spatial schema, since movement necessarily involves a temporal
dimension and, further, it always involves at least one spatial primitive, since any movement, necessarily, involves at
least one OBJECT that moves.
6. This spatial primitive is not to be confused with the image schema family PATH-following.
7. Table 1 is based on Mandler and Pag´
an C´
anovas (2014), but includes some changes. In particular, as mentioned
above, we do not consider MOV E to be a spatial primitive.
SOURCE PAT H GOAL schema. In S OURCE PATH GOAL, a direction and a purpose are implied in
the image schema, which changes the conceptual nature of the movement. Lakoff and N ´
nez (2000)
make the distinction of ‘elements’, or roles, that to some extent correspond to the spatial primitives
discussed above, but additional distinctions are added. The elements8are listed in Table 2. Most
importantly, they make the clear distinction between end location and goal, as they distinguish
between ‘path’, the actual trajectory of a movement, and ‘route’, the expected movement.
Figure 1: The SOURCE PATH GOA L schema as illustrated by Lakoff and N´
nez (2000).
The distinction, made by Lakoff and N´
nez (2000), between the expected movement and the
actual movement is primarily interesting for a description of how new image schemas relate to
actual events and how new image schemas are learned. Consider, for example, a situation where
a child observes the movement of a billiard ball and is surprised that the ball stops because it is
blocked by another billiard ball. In this case, a given instance of the MOVEMENT ALO NG PATH
spatial schema formed the expectations of the child, which were disappointed by the actual physical
movement, because the expected END PAT H (the goal) does not correspond to the actual END PATH
(end location). Given a repeated exposure to similar events, the child may develop the new spatial
schema BLOCKAGE. After learning BLOCKAGE, the child will no longer be surprised by blocked
movement since the expected END PATH (the goal) will correspond to the actual END PATH (end
location). While the terminological distinction between expected trajectory and actual trajectory is
useful, these do not necessarily need to constitute two different spatial primitives. Indeed, spatial
primitives are parts of image schemas and, thus, always parts of conceptualisations, and not parts of
actual events.
While the notions of path-following of Mandler and Pag´
an C´
anovas (2014) and Lakoff and
nez (2000) coincide widely, there are differences in terminology and definitions. In this paper
we follow primarily the former.
8. We altered the terms to better match the terminology in (Mandler and Pag´
an C´
anovas, 2014), but no change in content
was made.
Element Description
trajector The object
source The initial location
goal The intended end location
route A pre-realised route from source to goal
path The trajectory of motion
position The position of the trajector at a given time
direction The direction of the trajector at a given time
end location End location, may not correspond to the goal location
Table 2: Elements of PATH according to Lakoff and N´
nez (2000)
3.2 Concepts that involve PATH-following
As briefly demonstrated with CONTAINMENT and ABOV E, image schemas can be used as a source
for grounding abstract concepts in already comprehended concrete concepts. In this section we
consider examples for concepts, which involve members of the PATH-following family.
The most straightforward examples of concepts that involve PATH-following are concepts that
are about the spatial relationship of movement between different points. Prepositions such as
from, to, across and through all indicate a kind of PATH-following9. This also includes key
verbs that describe movement, e.g. coming and going. Another example, here for the spatial
schema SOURCE PATH GOA L, is Going from Berlin to Prague. Note that in many cases we do
not provide information about START PAT H and EN D PATH of a movement; e.g. leaving Berlin
and travelling to Berlin are examples for the spatial schemas SOURCE PATH and PATH GOAL,
respectively. Meandering is an example for a concept that realises MOVE ME NT AL ONG PATH,
which involves a PATH but no START PATH or EN D PATH. In contrast, no discernable PATH
is involved in roaming the city, which is an example for MOVEMEN T OFOBJECT. – These
examples illustrate that spatial schemas may be ordered hierarchically with respect to their
content: SOURCE PATH GOA L contains more spatial primitives and more information than, for
example, MOV EM EN T ALONG PAT H, which is the root of the PATH-following family. And
Beyond concepts that involve movement, PATH-following plays an important role in many
abstract concepts and conceptual metaphors.
The concept of “going for a joy ride” realises the spatial schema SOURCE PATH, since it has
a START PATH and a PATH but no EN D PATH. Similarly, the expression “running for president”
describes the process of trying to get elected as president metaphorically as a PATH GOAL. In this
metaphor the PATH consists of the various stages of the process (e.g. announcing a candidacy and
being nominated by a party) with the inauguration as END PATH.
Another metaphor “life is a journey”, studied by Ahrens and Say (1999), makes an analogical
mapping between the passing of time in life, to the passing of spatial regions on a journey. As in
the example mentioned above, where the concept of “being in love” acquired information from the
CONTAINMENT schema, this metaphor gains information from the spatial primitives connected
to the image schema SOURCE PATH GOAL. Here, the most important spatial primitives are
9. Some prepositions include other image schemas at the same time. E.g. ‘through’ involves apart from PATH also some
notion of CONTAINMENT.
START PATH and EN D PATH – in this metaphor they are mapped to the moments of birth and
death, as well as the PATH itself, illustrating how “life goes on” in a successive motion without
A different perspective on life and death is expressed in the metaphorical expression “the circle
of life”. Implied is that life leads to death, but also that death gives rise to life, completing a cyclic
movement – the image schema MOV EM ENT INLOOP S. This image schema can be considered as a
version of PATH-following, in which START PAT H and EN D PATH coincide at the same ‘location’.
These examples illustrate a general pattern, namely that many conceptual metaphors involving
PAT Hs are about processes, and different events during such processes are treated metaphorically as
locations on a path. This leads to a conceptualisation of the abstract concept of time, which we will
further investigate in the next section.
Expression Level in hierarchy
Concrete: Roaming the city MOV EMENT OFOBJECT
Leaving Berlin SOURCE PATH
Travelling to Berlin PAT H GOAL
Going from Prague to Berlin SOURCE PATH GOAL
Abstract: Going for a joy ride SOURCE PATH
Running for president PAT H GOAL
Life is a journey SOURCE PATH GOA L
The circle of life MOV EMENT INLOOP S
Table 3: Summary of the mentioned expressions and their level in the PATH-following hierarchy
3.3 Time and processes as PATH
The conceptualisation of time has been investigated by Boroditsky (2000). Here we follow suit
by looking at how members of the PAT H-following image schema family are widely used as a
conceptual metaphors for time. We consider several examples and discuss the role of PATH-
following image schemas for the conceptualisation of processes in general.
One popular way to conceptualise time is as MOVEME NT AL ON G PATH. Often, time is
conceptualised as having a beginning, a START PAT H; this may be the Big Bang or the moment
of creation in a religious context. Depending on the cosmological preferences, time may also be
conceptualised to have an end, an END PAT H: the Big Rip or an apocalypse.
Other religious traditions embrace the notion of a ‘Wheel of Time’, that is time as a cyclic
repetition of different aeons. The underlying image schema involves a MOVEME NT INLO OP S.
The same image schema is used in the conceptualisation of time within calendars: the seasons are
a continuous cycle where any winter is followed by a new spring. Similarly, the hours of the day
are represented on analogue clocks as 12 marks on a cycle, and the passing of time is visualised as
MOVE MENT INLOOPS of the handles of the clock.
The conceptualisation of time, in itself, is an interesting example for the usage of image
schemas. However, the real significance is that these image schemas can be seen as providing the
conceptual skeletal structure for our understanding of processes. Assume we want to understand a
complex process, e.g. the demographic development of a country, the acceleration of a falling object,
or the economic situation of a country. In these situations we often use two-dimensional coordinate
systems where the vertical axis represents the property in question (e.g. population, speed, GDP,
respectively) and the horizontal axis represents time. These coordinate systems are so useful and so
widely applicable because we can conceptualise arbitrary processes as MOVE ME NT AL ONG PATH,
where the paths represent some important dimension or aspect of the process.
The importance of PATH-following image schemas for the conceptualisation of processes can
be illustrated by considering similes. If we pick from Table 4 randomly a target domain Xfrom the
first column and a source domain Yfrom the second column, the resulting simile Xis like Ywill
be sensible. (Of course, depending on the choice of Xand Ythe simile may be more or less witty.)
Note that the target domains have little or nothing in common. Thus, at least on first glance, one
would not expect that one can compare them meaningfully to one and the same source domain.
Target Domain Source Domain
Watching the football game the swinging of a pendulum
Their marriage a marathon
The story escaping a maze
This piece of music a sail boat during a hurricane
Bob’s career a roller coaster ride
Her thoughts a Prussian military parade
Democracy in Italy stroll in the park
Table 4: PATH similes: <target>is like <source>.
The similes work because all of the concepts in the second column involve physical MOVE-
ME NT AL ONG PATH, which have some pertinent characteristics. These characteristics may
concern the shape of the path itself (e.g. the path of a roller coaster involves many ups and downs
and tight curves, the path out of a maze involves many turns, the path of a pendulum is regular and
between two points), the way the movement is performed (e.g. the movement of a sail boat during a
storm is erratic and involuntary, a stroll in the park is done leisurely), and the effects the movement
may have (e.g. running a marathon is exhausting, a Prussian military parade may be perceived as
threatening). In each of the similes we use some of the pertinent characteristics from the source
domain to describe the process from the target domain. For example, in the simile ‘Bob’s career is
like a Prussian military parade’ we conceptualise the career as a path along time (with career-related
events like promotions as the sites on the path) and transfer characteristics from the movement of a
Prussian military parade on this path. Thus, one way to read the simile is that Bob moves through the
stages of his career in a exceptionally predictable fashion. The example illustrates how the similes
work: first, we conceptualise the process in the target domain as MOVE MENT ALO NG PATH, where
the events of the process are ordered by time, and then we transfer some pertinent characteristics of
the MOV EMENT AL ON G PATH of the source domain to the target domain. This pattern is not just
applicable to the concepts in Table 4. As we discussed above, any process can be conceptualised as
MOVE MENT ALONG PATH, thus, any process could be added as target domain in Table 4. Further,
any concept that involves interesting physical movement along some path could be added as source
domain. Hence, the use of the image schema MOVE ME NT ALONG PATH enables the mechanical
generation of similes for processes.
Similes are a particular form of concept generation in which two domains are combined. This
phenomenon is strongly connected to conceptual blending that we will discuss further in Section 5.1.
To summarise, in this section we have introduced the image schema MOVE MENT ALO NG PATH.
We have seen that it is widely used in natural language and plays an important role in our
understanding of time and processes. The examples show that the notion of PATH-following, at
its core, is about movement along some trajectory. However, there are important differences both
with respect to the spatial primitives that are involved and with respect to the shape of the PATHs.
In the next section, we consider how images schemas can be represented in formal languages. One
particular concern is to represent image schemas in a way that adequately captures the variety and
flexibility of image schemas.
4. Formalising image schemas as graphs of theories
4.1 Previous work on formalising image schemas
Image schema research has had great impact in the cognitive sciences and in particular in cognitive
linguistics. However, within computational cognitive systems, and artificial intelligence in general,
it has not yet been explored to its full potential.
Looking at how image schemas can be computationally acquired, there are studies that attempt
to model early cognitive development and learn from perceptual input. The connectionist model
proposed by Regier (1996) learns to linguistically classify visual stimuli in accordance with the
spatial terms of various natural languages. Similarly, the Dev E-R system by Aguilar and P´
y P´
erez (2015) is a computer model that simulates the first sensorimotor stages in cognitive
development. Their system learns to distinguish and fine-tune visual clues such as nuances of colour,
as well as different sizes of objects and directions of movement. Both approaches demonstrate how
an artificial agent can develop cognitive abilities and language development from perceptual input.
Another study using perceptual input to simulate the development of image schemas was made
by Nayak and Mukerjee (2012). They fed video material of OBJECTs moving INand OUT of boxes
into an unsupervised statistical model in order to capture the dynamic aspects of the CONTAINMENT
schema. From this, the system learned how to categorise different CONTAINMENT contexts and
could in combination with a linguistic corpus generate simple CONTAINMENT-related language
These are examples of systems that learn image schemas and visual relationships from
perceptual input. More commonly, work on formalising image schemas is done when the image
schemas are already identified. Prominent work in this field is the work by Kuhn (2002, 2007). He
argues that image schemas capture abstractions in order to model affordances. Working top-down
rather than bottom-up as above, he uses WordNet to define noun words and connects them to spatial
categorisations related to image schemas based on affordance-related aspects of meaning.
Walton and Worboys (2009) build further on Kuhn’s work by visualising and formalising
the connections between different image schemas using bigraphs. By visually representing the
topological and ‘physical’ image schemas relevant in built environments, they demonstrate how
more complex dynamic image schemas such as BLOCKAGE could be generated using sequences of
bigraph reaction rules on top of simpler static image schemas.
St. Amant et al. (2006) present what they call the Image Schema Language, ISL. In their paper,
they provide a set of diagrams that illustrate how combinations of image schemas can lead to more
complex image schemas, and provide some real life examples.
Brugman and Lakoff (1988) discuss how image schema transformations form networks that
capture the relationships in polysemous words, in particular the preposition ‘over’ is investigated.
This relates to our own approach of how to formalise and formally represent image schemas.
Namely to use the hierarchical structure of image schemas demonstrated previously to represent
image schemas as families of theories.
4.2 Image Schema Families as Graphs of Theories
In the previous sections, we argued for image schemas to be members of families, which are
partially ordered by generality. In the following section, we will describe and visualise an approach
to represent the connections between image schemas, belonging to the same family. In order to
discuss the problem of how more complex image schemas can be constructed through a combination
of different image schemas (e.g. LINKED PATH, MOVEMENT INLOO PS), we will discuss the
possible interconnection these families of theories allow. Formally, we can represent the idea as
a graph10 of theories in DOL, the Distributed Ontology, Modeling and Specification Language
(Mossakowski et al., 2015).
This choice is motivated primarily by two general features of DOL: (1) the heterogeneous
approach, which allows for a variety of image schematic formalisations without being limited to
a single logic, and (2) the focus on linking and modularity. Therefore, DOL provides a rich toolkit
to further formally develop the idea of image schema families in a variety of directions.
In more detail, DOL aims at providing a unified metalanguage for handling the diversity of
ontology, modelling, and specification languages, for which it uses the umbrella term ‘OMS’. In
particular, DOL includes syntactic constructs for:
1. “as-is” use of OMS formulated (as a logical theory) in a specific ontology, modelling or
specification language,
2. defining new OMS by modifying and combining existing OMS (which are possibly written
in different languages), and
3. mappings between OMS, resulting in networks of OMS.
DOL is equipped with an abstract model-theoretic semantics.11 The theoretical underpinnings of
the DOL language have been described in detail in (Kutz, Mossakowski, and L¨
ucke, 2010) and
(Mossakowski, Lange, and Kutz, 2012), whilst a full description of the language can be found in
(Mossakowski et al., 2015) or (in a more condensed form) in (Mossakowski et al., 2013).
Building on similar ideas to those underlying the first-order ontology repository COLORE12
uninger et al., 2012), we propose to capture image schemas as interrelated families of
(heterogeneous) theories. Similar ideas for structuring common sense notions have also been
applied to various notions of time (Van Benthem, 1983; Allen and Hayes, 1985). This general
approach also covers the introduction of non-spatial elements such as ‘force’ as a basic ingredient of
image schemas, as for instance argued for by G¨
ardenfors (2007) and constitute the core of Mandler
and Pag´
an C´
anovas (2014)’s conceptual integrations mentioned above.
10. These graphs are diagrams in the sense of category theory.
11. The final DOL specification was submitted as a standard to the Object Management Group (OMG) in late 2015
12. See
S = G
Path: the image schema family of moving along paths and in loops
add End_Path
with additional
distinguished point
add End_Pathadd Start_Path
add Focal_Point
add Start_Path
S = G
extending an image
schema axiomatically
extending by new spatial
primitives and axioms
add Landmark
Source_Path Path_Goal
add Path
Figure 2: A portion of the family of image schemas related to path following shown as DOL graph.
In Figure 2, some of the first basic stages of the image schema family PATH-following are
presented. Ranging from Mandler’s general definition presented above, of object movement in any
trajectory, to more complex constructions.
The particular image schema family sketched is organised primarily via adding new spatial
primitives to the participating image schemas and/or by refining an image schema’s properties
(extending the axiomatisation). In general, different sets of criteria may be used depending, for ex-
ample, on the context of usage, thereby putting particular image schemas (say, REVOLVE AROU ND)
into a variety of families. Apart from a selection of spatial primitives, other dimensions might be
deemed relevant for defining a particular family, such as their role in the developmental process.
One way MOVEME NT AL ON G PATH can be specialised is as the image schema of MOVE -
ME NT INLO OPS. Note that this change does not involve adding a new spatial primitive, but
just an additional characteristic of the path. The resulting image schema can be further refined
by adding the spatial information of a focal point, which the path revolves around – this leads to
the notion of orbiting, or, by continuously moving the orbiting path away from the focal point, to
create the concept of spirals. Alternatively, we may change MOV EM ENT ALONG PAT H by adding
distinguished points; e.g. the START PATH, the target END PATH, or both.
The MOV EMENT INLOOP S image schema may be further specialised by identifying (the
location of) the START PATH and the EN D PATH. In this case, the path is closed in the sense
that any object which follows the path will end up at the location at where it started its movement.
The difference between a closed path and a looping path is that the closed path has a start and an
end (e.g. a race on a circular track), while the looping path has neither (like an orbit). It is possible
to further refine the schema by adding more designated points (i.e. ‘landmarks’) or other related
spatial primitives.
We will now show how the theories of image schemas and the various branching points in the
graph can be characterised formally.
4.3 Axiomatisation of Path-Following
In this section, we present an axiomatisation of the image schemas represented in Figure 2. The
focus of our axiomatisations is to capture the important differences of the branching points of the
PAT H-following family, not an exhaustive axiomatisation. For the sake of brevity, we will present
only selected axioms in this section. A more complete axiomatisation is available at an Ontohub
Our axiomatisation approach is inspired by semantics in the neo-Davidsonian tradition (David-
son, 1967; Parson, 1990). We consider image schemas as a type of event (in generality quite similar
to the view defended in (Clausner and Croft, 1999) to view image schemas as a kind of ‘domain’)
and consider spatial primitives as thematic roles of these events. Thus, if a given image schema is
enriched by adding a new spatial primitive, this is typically represented by adding a new entity (e.g.
site) and a new relation (e.g. has start path) that determines the thematic role of the new entity in
the event. As representation language we use ISO/IEC 24707 Common Logic. Common Logic is a
standardised language for first-order logic knowledge representation, which supports some limited
form of higher-order quantification and sequence variables (Menzel, 2011).
For the axiomatisation of the image schemas in the PATH-following family we assume an
image schema MOVE ME NT AL ONG PATH as the root of the family. MOVE ME NT ALONG PATH is
derived from a more general notion, namely MOV EM EN T OFOBJECT. This is movement of some
kind that involves only one spatial primitive, namely an OBJECT. This object plays the role of the
trajector within the context of the MOV E. This can be formalised in Common Logic as follows:
(forall (m)
(MovementOfObject m)
(exists (o)
(Movement m)
(Object o)
(has_trajector m o)))))
No additional information about what kind of object is moving and how it is moving is
The schema MOVE ME NT AL ONG PATH is the result of adding a new spatial primitive to
MOVE MENT OFOBJECT, which plays the role of a PATH.
(forall (m)
(MovementAlongPath m)
(exists (p)
(MovementOfObject m)
(Path p)
(has_path m p)))))
Under a PATH we understand a collection of two or more sites, which are connected by successor
relationships. Each of these sites have (relative to the path) at most one successor site. The transitive
closure of the successor relation defines a before relationship (relative to the path); and for any two
different sites x, y of a given path, either xis before yor yis before x(relative to the path).15 This
axiomatisation provides a representation of a quite abstract notion of MOVEME NT AL ON G PATH.
It needs to be sufficiently abstract, since it serves as the root node for the PAT H-following family. All
other image schemas in the family are derived from this root by adding additional spatial primitives
and/or additional axioms.
Given this notion of PATH, we can axiomatise the relationship between the PATH and the
OBJECT, which characterises a MOVE ME NT AL ONG PATH. During the movement, the moving
object needs to pass through all sites of the path in a temporal order, which matches the before-
relationship between the sites:
(forall (p o m s1 s2)
(i f
(MovementAlongPath m)
(has_path m p)
(has_trajector m o)
(before s1 s2 p))
(exists (t1 t2)
14. From an ontological perspective, MOV EM EN T OFOBJECT can be seen as a kind of process (or occurrent). Thus,
any adequate axiomatisation of MOV EM EN T OFOBJECT needs to represent change over time in some form. To
keep things simple, we here just quantify over time points. We assume that time points are ordered by an earlier
relationship. Further, we use two other relationships to connect time points to processes: (has start m t)
means The movement mstarts at time point tand (during t m) means Time point tlies within the interval
during which movement mhappens.
15. The before-relationship is not a total order, since antisymmetry is not postulated.
(Timepoint t1) (Timepoint t2)
(during t1 m) (during t2 m)
(located_at o s1 t1) (located_at o s2 t2)
(earlier t1 t2)))))
The image schema SOURCE PATH is the result of adding the spatial primitive START PATH
to MOV EMENT ALON G PATH. We represent this with the has starts path relationship. The
START PATH of a PATH is a site on the path that is before any other site of the path:
(forall (m)
(SourcePathMovement m)
(exists (s)
(MovementAlongPath m)
(has_start_path m s)))))
(forall (m s1 s2 p)
(i f
(SourcePathMovement m)
(Site s1)
(Site s2)
(no t (= s1 s2))
(has_path m p)
(has_start_path m s1)
(part_of s2 p))
(before s1 s2 p)))
What distinguishes SOURCE PATH from other movements is the following: at the start of a
SOURCE PAT H movement the object that moves is located at the START PATH:
(forall (mstpo)
(i f
(SourcePathMovement m)
(has_start m t)
(has_trajector m o)
(has_start_path m s))
(located_at o s t))))
Analogously, we can define PATH GOAL as a MOV EM ENT ALON G PATH with an EN D PATH.
A SOURCE PATH GOA L is a movement, which includes both landmarks of START PATH and
END PATH. Thus, SOURCE PATH GOA L can be defined as the union (of the axioms) of
CLO SE D PATH MOVE ME NT is a special case of SOURCE PATH GOA L, where the location of
the START PATH and the END PATH of the PATH coincide.
(forall (m s g)
(i f
(has_start_path m s)
(has_end_path m g))
(ClosedPathMovement m)
(SourcePathGoalMovement m)
(= (location_of s) (location_of g))))))
SOURCE PAT H VIA GOAL is a different way to refine SOURCE PATH GOAL . In this case an
additional designated site is added, which lies between the START PATH and the END PATH of the
16. In DOL, this is done by using the keyword ‘and’, which amounts to taking the model-theoretic intersection of the
model classes of the theories of SOURCE PATH and PATH GOA L.
(( forall (m)
(SourcePathViaGoalMovement m)
(exists (s p)
(SourcePathGoalMovement m)
(has_path m p)
(Site s)
(part_of s p)
(no t (has_start_path m s))
(no t (has_end_path m s))))))
Both CL OS ED PATH MOVE ME NT and S OURCE PATH VIA GOAL can be combined in the obvious
A completely different branch of the movement image schema family does not involve either
START PATH or EN D PATH, but the PATH consists of a loop of sites. One way to represent this is
by requiring that the before-relationship is reflexive (with respect to the path of the movement):
(forall (m)
(MovementInLoops m)
(MovementAlongPath m)
(forall (p s)
(i f
(has_path m p)
(Site s)
(part_of s p))
(before s s p))))))
The difference between MOV EMENT INLOOP S and CL OSED PATH MOVE MENT is that in the
latter case both START PATH and END PAT H are present, they just spatially coincide. Hence, the
movement is over when the object meets the target. In contrast, MOVE MENT INLOO PS entails that
the moving object is located at the same location more than once.
REVO LVI NG MOV EMENT is a subtype of MOVEME NT INLO OP S. To define it, we need to
consider two additional factors: the shape of the path is elliptical, and there is a focal point, which
the movement revolves around. The focal point itself is a site, but it is typically the location of an
object. A detailed axiomatisation of this image schema is beyond the scope of this paper, we just
provide an initial sketch:
(forall (m)
(RevolvingMovement m)
(MovementInLoops m)
(exists (p s)
(has_path m p)
(Eliptical (shape p))
(Site s)
(has_focal_point p s))))))
5. Image Schemas in Computational Conceptual Blending
In this section, we will illustrate how formalised families of image schemas, as just sketched above,
can help in the computational modelling of concept invention, an area at the heart of AGI. More
precisely, we will here focus on the highly influential framework of conceptual blending (Fauconnier
and Turner, 2003; Turner, 2007; Lakoff and N´
nez, 2000), and illustrate the foundational role that
a formal theory of image schemas plays in its computational realisation.
5.1 A crash course on conceptual blending
Introduced by Fauconnier and Turner (1998), conceptual blending has been employed very
successfully to understand the process of concept invention, studied particularly within cognitive
psychology and linguistics. The theory argues that at the heart of novel concept creation lies a
combination process involving already existing knowledge and understood concepts. By merging
two, or more, conceptual spaces, a blended conceptual space results. This blend contains
information from both input spaces and has emergent properties due to its own unique composition.
The classic example is the blend of a ‘houseboat’, containing merged information from the input
spaces ‘house’ and ‘boat’.17
One of the central aspects of blending is the the way in which ‘common structure’ between the
input concepts is understood to steer the creation of the new concept. The ‘merging’ of the input
spaces is moderated by this common structure, represented as the generic space, or as it is called in
formal approaches, the base ontology (see Figure 3).18 The common structure of the input spaces is
understood to play a vital role in rendering the newly constructed concept meaningful, as it ensures
that the blended space also contains the structure found in the generic space.
base morphisms
O1 O2
Base Ontology
Input 1 Input 2
blendoid morphisms
Figure 3: The basic integration network
for blending: concepts in the
base ontology are first refined to
concepts in the input ontologies
and then selectively blended
into the blendoid.
However, despite this influential research, within computational creativity and AI in general,
relatively little effort has been devoted to fully formalise these ideas and to make them amenable to
computational techniques, but see (Schorlemmer et al., 2014; Kutz et al., 2014a) for overviews.
Unlike other combination techniques, blending aims at creatively generating (new) concepts on
the basis of input theories whose domains are thematically distinct but whose specifications share
some structural similarity.
17. This and the related blend of ‘boathouse’ were fully formalised in (Kutz et al., 2014b).
18. In the limit case, the shared structure might be trivial, and a concept such as ‘red pencil’ might be understood as a
blend too, by simply imposing properties from one input space onto another.
Kutz et al. (2014a) describe in detail the basic formalisation of conceptual blending, as sketched
by the late Joseph Goguen and discuss some of its variations (Goguen and Harrell, 2010). Moreover,
it is illustrated how the Distributed Ontology Language DOL can be used to declaratively specify
blending diagrams of various shapes, and how the workflow and creative act of generating and
evaluating a new, blended concept can be managed and computationally supported within Ontohub,
aDOL-enabled theory repository with support for a large number of logical languages and formal
linking constructs, see (Kutz et al., 2014b; Mossakowski, Kutz, and Codescu, 2014). The reasoning
engine managing heterogeneous theories and computationally supporting the Ontohub repository
is the Heterogeneous Tool Set Hets (Mossakowski, Maeder, and L ¨
uttich, 2007). The graph for the
structured theory illustrating the PATH-following family and automatically being generated by the
Hets system from its formal specification is shown in Figure 4.
Figure 3 illustrates the basic, classical case of an ontological blending diagram. The lower part
of the diagram shows the generic space (tertium), i.e. the common generalisation of the two input
spaces, which is connected to these via total (theory) morphisms, the base morphisms. The newly
invented concept is at the top of this diagram, and is computed from the base diagram via a colimit.
More precisely, any consistent subset of the colimit of the base diagram may be seen as a newly
invented concept, a blendoid.19 Note that, in general, ontological blending can deal with more than
one base and two input ontologies, and in particular, the sets of input and base nodes need not
exhaust the nodes participating in a base diagram.
5.2 Using image schemas in computational blending
One problem for conceptual blending, and related work on analogy engines (e.g. structure mapping
(Gentner, 1983; Forbus, Falkenhainer, and Gentner, 1989) and heuristic-driven theory projection
(HDPT) (Schmidt et al., 2014)) is the generation of a ‘sensible’ blend. In a completely automatised
system, there is currently no simple way to distinguish the blendoids that a human would consider
meaningful from those that lack cognitive value. This problem grows exponentially in relation to
the size of the input spaces. The larger the input spaces, the more combinations can be generated
resulting in a multitude of possible blendoids, most of which will make little sense if evaluated
by humans. In real life scenarios, the amount of information in the input spaces can be vast,
complicating things for successful concept invention tremendously when looked at as a formal,
combinatorial problem.
A proposal to explain the ease with which humans perform blending is given via the ideas of
packing and unpacking, as well as compression and expanding of conceptual spaces, as outlined
by Turner (2014). These terms aim to capture how we mentally carry around ideas as compressed
‘idea packages’ that we can ‘unpack’ and utilise in different contexts on the fly. The process of
packing and unpacking ideas is important for the contextualised usage of conceptual blends in
various situations. Generally, the idea of optimality principles in blending theory is meant to account
for an evaluation of the quality and appropriateness of the resulting blends (Fauconnier and Turner,
2003). However, there is currently no general formal proposal how such optimality principles could
be implemented computationally, apart from some work on turning such principles into metrics for
rather lightweight formal languages (Pereira and Cardoso, 2003).
19. A technically more precise definition of this notion is given in Kutz et al. (2012). Note also that our usage of the term
‘blendoid’ does not coincide with the (non-primary) blendoids defined in Goguen and Harrell (2010).
is:pathGoalMovement.clif is:sourcePathMovement.clif
sourcePathMovementpathGoalMovement movementInLoops
sourcePathGoalMovement revolvingMovement
sourcePathViaGoalMovement closedPathMovement
Figure 4: The PAT H-following family displayed as a structured DOL theory in Ontohub/Hets.
Hedblom, Kutz, and Neuhaus (2014) suggested that instead of relying on purely syntactic
approaches, image schemas in their role as conceptual building blocks could be used to guide
the computational blending process. The principle idea here is that employing image schemas in
the construction of generic spaces will not only result in a significant reduction of the number of
generated blends, but will moreover filter out many of those blends human evaluators would deem
meaningless. A related and complementary approach is (Veale, Feyaerts, and Forceville, 2013),
where the problem of constraining the search space was addressed by suggesting that blending is
performed in a task-specific context. Here, selecting a task-specific context in a blending scenario
means to simultaneously work forward from the input spaces and backward from the desired
elements of the blend space.
In this line of thinking, one way to use image schemas in blending is to identify them as the
prime ingredient for the construction of a generic space. When performing the search for common
structure in the different input spaces, the search could be guided by mapping (parts of) the content
of the input spaces to nodes in a library of formally represented image schemas. As image schemas
hold semantic value in the form of spatial relationships, the blendoids would be based on the same
content. In theory, this is similar to classic structure mapping that preserves relationships, but as
image schemas model e.g. affordances (Kuhn, 2007), a blendoid will inherit such information as
I1 I2
Base Image Schema
Input concept 1 Input concept 2
Weakening: moving upwards in a family
I1* I2*
weakend I2 theory and / or
move up image schema hierarchy
weaken I1 theory and /or
move up image schema hierarchy
I1 I2
Base Image Schema
Input concept 1 Input concept 2
I1* I2*
weakend I2+ theory
weaken I1+ theory
move down (specialise) I1 image schema move down (specialise) I2 image schema
Strenghening: moving downwards in a family
Figure 5a Figure 5b
Figure 5: Blending using common image schemas: strenghening vs. weakening.
Figure 5 shows the two basic ways of using image schemas within the conceptual blending
workflow. In both cases, the image schematic content takes priority over other information the input
concepts might contain. On the left, following the core model of blending described above, we first
identify different spatial structure within the same image schema family in the input concepts, and
then generalise to the most specific common version within the image schema family to identify a
generic space, using our pre-determined graph of spatial schemas (i.e. we compute the least upper
bound in the lattice). The second case, shown on the right, illustrates the situation where we first
want to specialise or complete the (description of the) spatial schemas found in the input concepts,
before performing a generalisation step and to identify the generic space. This means moving down
in the graph of the image schema family. Of course, also a mix of these two basic approaches is
reasonable, i.e. where one input spatial schema is specialised within a family whilst the other is
generalised in order to identify a generic space based on image-schematic content. Examples for
both cases are described below in Sections 5.3 and 5.4.
5.3 The PATH-following family at work
To study how image schematic content can be used more concretely within conceptual blending,
we will now look at a number of examples. In this section, we will illustrate how moving up and
down within the image schema family of path-following opens up a space of blending possibilities,
infused with the respective semantics of the (versions of) the image schema. In the next section, we
will then discuss in more formal detail how these ideas work on a logic-based level.
As outlined in Section 3.3, processes in general can be easily combined with a variety of more
specific PAT H-following schemas. More specifically, we can explore the basic idea how to combine
the input space of ‘thinking process’, which involves only an underspecified kind of ‘movement
of thoughts’, with a second input space that carries a clearly defined path-following image schema.
This leads intuitively to a number of more or less well known phrases that can be analysed as blends,
including: ‘train of thought’, ‘line of reasoning’, ‘derailment’, ‘flow of arguments’, or ‘stream of
consciousness’, amongst others. Indeed, a central point we want to make in this section is that these
blends work well and appear natural because of the effectiveness of the following heuristics:20 (i)
given two input spaces I1and I2, search for the strongest version Gof some image schema that is
common to both, according to the organisation of a particular image schema family F; (ii) use Gas
generic space; and (iii) use again Fto identify the stronger version of G, say G0, inherent in one of
the two inputs, and use the semantic content of G0to steer the overall selection of axioms for the
blended concept.
To illustrate this process informally, let us briefly consider the concepts of ‘stream of
consciousness’, ‘train of thought’, and ‘line of reasoning’21.
On first inspection, the spatial schema of movement related to ‘thinking’ might be identified as
MOVE MENT OFOBJECT, i.e. without necessarily identifying following a PATH at all. Indeed, in
Figure 2, MOVE ME NT OFOBJECT is marked as an ‘entry point’ to the path-following family.
The stream of consciousness may be seen as an unguided flow of thoughts, in which topics
merge into each other without any defined steps, but rather in a continuous manner. It lacks a clear
START PATH and has no guided movement towards a particular END PATH. It resembles the more
basic forms of PATH-following that, according to Mandler and Pag´
an C´
anovas (2014), is simply
movement in any trajectory.
Atrain of thought22 can be conceptualised in various ways. It differs from a stream of
consciousness by having a more clear direction, often with an intended EN D PATH. It is possible to
say that one “lost their train of thought”, or that “it was hijacked” or how “it reversed its course”. The
‘train’ may be understood as a chain-like spatial object (in which case ‘losing the train’ decodes to
‘disconnecting the chain’) or more plainly as a locomotive. In the Pixar film ‘Inside Out’ (2015), the
‘Train of Thought’ is an actual train that travels the mind of the fictional character Riley Anderson,
and delivers daydreams, facts, opinions, and memories.
Aline of reasoning might be seen as a strengthening of this blend, where the path imposed is
linear. Although a ‘line’, mathematically speaking, has no beginning or end, the way this expression
is normally understood is as a discrete succession of arguments (following logical rules) leading to
an insight (or truth). This blend might therefore be analysed to correspond to S OURCE PATH GOAL
in (Lakoff and N´
nez, 2000), in which there is a clear direction and trajectory of the ‘thought’
In order to understand how blending can result in these concepts, and how image schemas are
involved, let us have a closer look at the input spaces and their relationship to the PATH-following
image schemas. Relevant input spaces include line (perhaps analysed as ‘discrete interval’),
stream/river, train/locomotive, and, as secondary input space, ‘thinking process’.
‘Thinking’ as an input space is difficult to visualise. However, when ‘thinking’ is understood
as a process it can be easily combined with various PAT H-following notions (see Section 3.2
above). As thoughts (in the form of OBJECT) are moved around, the simplest form of thinking
20. By ‘heuristics’ we mean a method that imposes rules on how to select a base (i.e. introduces a preference order
on possible generic spaces) and, moreover, rules to decide which axioms to push into the blend. I.e., without any
heuristics we are left to perform a randomised axiom selection, followed by an evaluation of the resulting blended
21. The examples presented here are chosen to illustrate the basic ideas how to employ families of image schemas in
blending. It is not intended to capture fully the meaning of these terms as they are used in the psychological or
linguistic literature, or indeed the subtle meaning they might carry in natural language.
22. The expression ‘train of thoughts’ appears to have been first used by Thomas Hobbes in his Leviathan (1651): “By
‘consequence of thoughts’ or ‘TRAIN of thoughts’ I mean the occurrence of thoughts, one at a time, in a sequence;
we call this ‘mental discourse’, to distinguish it from discourse in words.
is MOV EMENT OFOBJECT. There is no S TART PATH nor an END PATH. Intuitively, it does not
appear to have any particular PATH (in the sense of a spatial primitive).
A stream is characterised by a continuous flow along a PATH. Whilst a START PATH and
END PATH can be part of a stream-like concept, like in the fleshed out concept of a river with a
source and mouth, they do not constitute an essential part of the concept of stream.
For a train (understood as ‘locomotive’), the concepts of a START PATH and END PAT H has
a much higher significance. The affordances found in trains are primarily those concerning going
from one place to another. A train ride can also be seen as a discrete movement in the sense that for
most train rides, there are more stops than the final destination. This results in a discrete form of the
spatial schema SOURCE PATH GOAL.
When blending such forms of movement with the thinking process, what happens is that the
unspecified form of movement found in ‘thinking process’ is specialised to the PATH-following
characteristics found in the second input space. The result is the conceptual metaphors for the differ-
ent modes of thinking listed above, where the generic space contains just MOV EM ENT OFOBJECT,
and the blended concepts inherit the more complex PATH-following from ‘train’, ‘stream’, or ‘line’.
Path: specialisation (and generalisation) of image schemas in the path family
add End_Path
add End_Pathadd Start_Path
add Start_Path
add Landmark
add Path
add End_Path
add End_Pathadd Start_Path
add Start_Path
add Path
add Discreteness
add Continuity
Specialising arbitrary movement inherent to
'thinking' to discrete Source-Path-Goal
respectively to continous path-following
Figure 6: How ‘thinking’ transforms into ‘train of thought’ respectively ‘stream of consciousness’.
In more detail, Figure 6 shows two specialisations of the basic spatial schema of MOVE -
ME NT OFOBJECT. The first, shown on the left, specialises to a discrete version of the
schema SOURCE PATH GOA L with a designated element and discrete movement, supporting
the ‘train of thought’ blend. The second, shown on the right, specialises to a continuous
version of MOVEMENT AL ON G PATH, where an axiom for gapless movement is added to the
ontology story =
%% A story is defined as a telling of a plot.
(forall (x)
(Story x)
(exists (y) (and (Plot y) (tells x y)))))
%% A plot consists of some events
(forall (x)
(i f (Plot x)
(exists (y) (and (Event y)(part_of y x)))))
%% Every event in a plot is causally connected to at least some other event in the plot.
(forall (x y)
(i f (and (Plot x) (Event y)(part_of y x))
(exists (z) (and (Event z) (part_of z x) (o r (causes y z x) (causes z y x))))))
%% Some stories have a protagonist
(exists (x y) (and (Story x) (has_protagonist x y) (Character)))
Figure 7: Ontology story in DOL and Common Logic.
MOVE MENT ALONG PATH spatial schema to support the ‘flow of consciousness’ blend. As a third
possibility, in ‘line of reasoning’, we would impose additionally a linear (and perhaps discrete) path
onto ‘thinking’.
5.4 Blending in depth
In Section 3.3, we discussed how PATH spatial schemas support similes, where some process (e.g. a
story) is compared to some other concept (e.g. a roller coaster ride). In the same way as the elements
in the PATH schema family may be used for creating similes, they can be used for conceptual
blending. In this case, the PATH spatial schemas play the role of the generic space and may also be
used to strengthen the input spaces. In this subsection, we discuss the process in detail and show
some of the relevant axioms. The example illustrates the blending pattern from Figure 5b.
The input spaces for our blending process are Story and Roller Coaster Ride. They are formally
represented in Figures 7 and 8 as axiomatisations in Common Logic with a DOL wrapper.23 Both
axiomatisations are quite weak. Stories are defined as a telling of a plot, and they may involve a
protagonist. A plot consists of some causally connected events. A roller coaster ride is a scary
amusement ride that follows either some steel or wooden track. The track is fast-paced and consists
at least of a start, a thrill element, and an end. Note that the track is an instantiation of a PATH image
schema, more specifically a SOURCE PATH GOAL.24
Because SOURCE PATH GOA L is embedded in the roller coaster concept, it is natural to use
it as the base space in our blend. (We reuse its axiomatisation from Section 4.3.) However,
SOURCE PAT H GOAL is not present in the space Story, thus we need to strengthen the concept
by adding the image schematic content from SOURCE PATH GOAL to Story. This can be defined in
DOL with the help of a signature map:
23. The first and the last line of each axiomatisation are DOL expressions. Their only purpose is to label these ontologies.
24. To save space we omitted many of the axioms that realise the SOURCE PATH GOAL schema.
ontology linearStory = { sourcePathGoalMovement with
SourcePathGoalMovement |-> LinearStory ,
Path |-> Plot ,
Site |-> Event ,
has_trajector |-> has_protagonist ,
Object |-> Character ,
has_path |-> has_plot ,
has_start_path |-> starts_with ,
has_end_path |-> ends_with ,
successor_of |-> causes } and story
The resulting concept is a Linear Story, that is a story, where the protagonist participates in a linear
succession of events with a clear start and ending.
While strengthening adds new information to an input space, weakening removes some
information. This may be necessary, because the blend may otherwise be logically inconsistent.
But even if logical consistency is not an issue, one of the input spaces may contain information that
is not desirable or irrelevant for a specific blend. For example, the axiomatisation in Figure 8
provides information about the material of roller coasters, which may be removed completely.
Further, the first axiom defines that roller coasters are scary amusement rides. This could be replaced
by an axiom that keeps the information that roller coasters are scary, but omits the connection to
amusement rides. All of these changes can be expressed in DOL with the help of filtering operations
and extensions.
In this example it is not necessary to weaken the Story input space. On the other hand, since the
SOURCE PAT H GOAL image schema is already realised in the input space Roller Coaster, there is
no strengthening necessary. Thus, in the case of this example I1+and I1in Figure 5b are identical;
so are I2+and I2.
ontology rollerCoasterRide =
% A roller coaster ride is a kind of scary amusement ride.
(forall (x)
(iff (RollerCoasterRide x)
(and (AmusementRide x) (Scary x))))
% A roller coaster ride follows some track & involves some Person as participant.
(forall (x)
(iff (RollerCoasterRide x)
(exists (y z)
(and (RollerCoasterTrack y)(follows x y)(Person z)(has_participant x z)))))
% A roller coaster track is fast-paced. It starts with the start, has at
% least one thrill element as part & ends with the end.
(forall (x)
(i f (RollerCoasterTrack x)
(exists (y1 y2 y3)
(and (fast_paced x) (starts_with x y1) (Start y1) (part_of y2 x)
(ThrillElement y2) (ends_with x y3) (End y3)))))
% Roller coaster tracks are either made from wood or steel.
(forall (x)
(i f (RollerCoasterTrack x)
(or (material_of x steel) (material_of x wood))))
Figure 8: Ontology rollerCoasterRide in DOL and Common Logic.
The blended concept is the result of merging both of the weakened Linear Story I1and
weakened Roller Coaster I2. To achieve this is DOL we need to define the interpretations from
the base image schema SOURCE PATH GOAL to I1and I2. The following is the definition for
weakened Roller Coaster.
interpretation base2rollerCoaster: base t o rollerCoasterRideWeakened =
Path |-> RollerCoasterTrack ,
Site |-> RollerCoasterElement ,
SourcePathGoalMovement |-> RollerCoasterRide ,
has_trajector |-> has_participant ,
Object |-> Person ,
has_path |-> follows ,
has_start_path |-> starts_with ,
has_end_path |-> ends_with
Analogously we define another interpretation base2story.25
By combining the two interpretations base2story and base2rollerCoaster we get a
new concept: a Thriller. In DOL the blended concept can be defined as follows:26
ontology blend = combine base2rollerCoaster, base2plot with Story |-> Thriller
Figure 9 provides an overview over the whole blending diagram.
Roller Coaster
strengthened Roller Coaster
weakened Roller Coaster
weakened Linear Story
Linear Story
Figure 9: Blending Thriller with the input spaces Story and Roller Coaster.
The newly defined concept, Thriller, inherits aspects from the input spaces Story and Roller
Coaster as well as from the SOURCE PATH GOAL spatial schema. In particular, thrillers are
25. The interpretation from the base space to I1reuses exactly the same signature map as in the strengthening process.
26. The with Story |-> Thriller part of the definition just renames “Story” into “Thriller” to make the
axiomatisation easier to read. The content of the concept is not affected.
scary, they have a fast-paced linear plot, which involves thrill elements, and a protagonist which
participates in the events of the plot. (Figure 10 shows some of the axioms of the blended theory.)
%% A thriller is scary.
(forall (x) ( i f (Thriller x) (Scary x)))
%% A thriller has a plot and a character, who is its protagonist
(forall (x)
(Thriller x)
(exists (y z)
(and (Plot y) (has_plot x y)(Character z)(has_protagonist x z)))))
%% The plot (of a thriller) is fast-paced and involves a start, at least one thrill
%% element and an end.
(forall (x)
(i f (Plot x)
(exists (y1 y2 y3)
(and (fast_paced x) (starts_with x y1) (Start y1) (part_of y2 x)
(ThrillElement y2) (ends_with x y3) (End y3)))))
%% If x and y are two different events in a plot p,
%% then either x is before y (in p) or the other way round.
(forall (p x y)
(i f
(and (Plot p) (Event y) (Event x) (part_of y p)(part_of x p) (not (= y x)))
(or (before x y p) (before y x p))))
Figure 10: Some axioms of the blended concept Thriller
6. Discussion
One of the hardest problems yet to be solved in artificial general intelligence is the generation
of concepts and their grounding in the environment, commonly known as the symbol grounding
problem. The difficulty lies not only in establishing a relationship between objects in the real world
and symbolic as well as mental representations, but also in the problem of defining ‘meaning’ itself.
This paper rests on the basic ideas of grounded and embodied cognition, in which physical
experiences are thought of as the primary source that gives meaning to concepts. Indeed, some
studies in linguistic neuroscience (e.g. Gallese and Lakoff (2005); Tettamanti et al. (2005)) indicate
that the cortical regions of the sensorimotor cortex are activated also in word comprehension tasks.
If bodily experiences are a primary source in constituting the meaning of concepts, the symbol
grounding problem can be meaningfully approached within this general framework.
We proposed in this paper an approach to computational concept invention in which image
schemas, understood as embodied conceptual building blocks, were utilised in conceptual blending,
the suggested cognitive machinery behind concept invention. To successfully investigate and
evaluate the fruitfulness of this idea, a more comprehensive formalisation of image schemas is
needed. Formalising image schemas has been a rather recent undertaking in artificial intelligence
research as a means to aid computational concept invention and common-sense reasoning (Kuhn,
2002; Walton and Worboys, 2009; Morgenstern, 2001; Goguen and Harrell, 2010; Kutz et al.,
So far, the research on image schemas has illuminated the inherent complexity of the
formalisation problem due to their abstract cognitive nature. At the same time, the incoherent
scholarly terminology and corresponding key definitions make it challenging to find stable ground
for further research. In an attempt to bridge research from several research strands, the main part
of this article was devoted to introducing the idea of how to formally structure image schemas as
families of theories.
Our work differs from the approaches discussed in Section 4.1 by considering the structure of
entire image schema families, using PATH as a proof of concept. While other approaches tend to
look at the interconnections between particular image schemas, we have followed the psychological
research of Mandler and Pag´
an C´
anovas (2014) to analyse formally the PATH-following image
schema family concentrating on the involved spatial primitives. It is our belief that this will allow
for a more fine-tuned and specialised use of image schemas in computational systems.
The most basic image schemas develop early and become more specialised with experience
(Rohrer, 2005). Currently there is no comprehensive and agreed upon list of these most basic image
schemas, although the general consensus is that complex image schemas result from combining
elements taken from various, more simple, image schemas (Oakley, 2010).
It is therefore likely that there are a limited number of core image schemas that, from their
most basic form, can be formally fleshed out into a structure similar to the illustrated PATH
image schema graph. Finding the intersections and combinations of these basic families, thus,
would mean identifying the source of the cognitive machinery behind the development of complex
image schemas. These would be the spatial schemas that contain similar and overlapping spatial
information and share certain spatial primitives. In our graph, MOVEMEN T INLOO PS could be
considered to be in the intersection of the image schema families PATH-following and CYCLE.
MOVE MENT OFOBJECT marks the entry point to the PATH-family, yet lacking itself the spatial
primitive of ‘path’. Here, the DOL language provides some of the tools to make such an
interconnection of families formally feasible, and gives a handle on a formal rendering of the notion
of construal (image schema transformation) discussed by Clausner and Croft (1999).
A second problem is the temporal nature of image schemas. Since image schemas are not only
static but also capture change over time, any axiomatisation thereof needs to address the non-trivial
problem of formally representing time. One motivation for the use of non-classical logics is the
claim that these are cognitively and linguistically more adequate than classical logics involving
variables and direct quantification over objects (Kamp, 1979; Blackburn, de Rijke, and Venema,
Moreover, the cognitive adequateness of particular formalisms has been studied in detail (e.g.
Knauff, Rauh, and Renz (1997)). In this spirit, a large variety of temporal logics has been
proposed to model various temporal aspects of natural language (Prior, 1967; Van Benthem, 1983).
Similarly, qualitative spatial logics have been designed to capture more adequately the way humans
conceptualise and reason about space (Cohn and Renz, 2007).
7. Conclusion
We have here presented an approach in which image schemas are treated as interconnected theories
in a lattice (ordered by theory interpretation). This was motivated by image schematic structure
found in language and the cognitive development of spatial primitives and image schemas. The main
insights, we claim, support the hypotheses that the spatial primitives and their assumed properties
distinguish not only the different usages in natural language and various cognitive stages, but can
be systematically seen as and mapped to branching points in the lattice of image schema theories.
The benefits of this approach lie not only in the provided structuring of image schemas, but
also in how formal systems may use them. By using image schemas in conceptual blending, it
is our belief that computational concept invention has taken a step in the right direction. Image
schemas provide a cognitively very plausible foundation for the idea of a generic space found in the
theory of conceptual blending. In analogy engines, or (formal) approaches to conceptual blending
(Turner, 2014; Kutz et al., 2014a), the presented graph of image schemas can provide a method
for theory weakening and strengthening based on the involved image schemas, employing basic
ideas of amalgams (Onta˜
on and Plaza, 2010). This approach is therefore substantially different
from the more syntactic-driven methods used by the Structure Mapping Engine (SME) (Gentner,
1983; Forbus, Falkenhainer, and Gentner, 1989) or Heuristic-Driven Theory Projection (HDTP)
(Schwering et al., 2009; Schmidt et al., 2014).
Future work will focus on extending the presented formalisation approach to other basic image
schema families. This will include studying their interconnections, formal methods for their
combination to construct complex schemas, as well as algorithmic approaches for detecting image
schemas within given input concepts. Conversely, we hope that the systematic study of the formal
interconnections between image schema families will have unifying value also for image schema
research within the cognitive sciences, and provide some of the still missing systematicity to the
field. Finally, we belief that the semantic and cognitive grounding of the idea of a generic space
in the notion of image schema has great potential for computational realisations of conceptual
Acknowledgments. We thank the reviewers for constructive and valuable feedback. We would
also like to thank John Bateman, Tarek R. Besold, Emilios Cambouropoulos, Tony Veale, and
Mihailo Antovi´
c for valuable input and interesting discussions on topics related to this paper.
The project COINVENT acknowledges the financial support of the Future and Emerging
Technologies (FET) programme within the Seventh Framework Programme for Research of the
European Commission, under FET-Open Grant number: 611553.
Aguilar, W., and P´
erez y P´
erez, R. 2015. Dev E-R: A computational model of early cognitive
development as a creative process. Cognitive Systems Research 33:17–41.
Ahrens, K., and Say, A. L. 1999. Mapping Image Schemas and Traslating Metaphors. In
Proceedings of Pacific Asia Conference on Language, Information and Computation, 1–8.
Allen, J., and Hayes, P. 1985. A Common-Sense Theory of Time. In Proceedings of the 9th
International Joint Conference on Artificial Intelligence (IJCAI-85), 528–531.
Barsalou, L. W. 2008. Grounded cognition. Annual review of psychology 59:617–645.
Bennett, B., and Cialone, C. 2014. Corpus Guided Sense Cluster Analysis: a methodology for
ontology development (with examples from the spatial domain). In Garbacz, P., and Kutz, O.,
eds., 8th International Conference on Formal Ontology in Information Systems (FOIS), volume
267 of Frontiers in Artificial Intelligence and Applications, 213–226. IOS Press.
Blackburn, P.; de Rijke, M.; and Venema, Y. 2001. Modal Logic. Cambridge Tracts in Theoretical
Computer Science. Cambridge University Press.
Boroditsky, L. 2000. Metaphoric structuring: Understanding time through spatial metaphors.
Cognition 75(1):1–28.
Brugman, C., and Lakoff, G. 1988. Cognitive Topology and Lexical Networks. In Stephen Small,
Gary Cottrell, M. T., ed., Lexical ambiguity resolution. 477–508.
Chrisley, R. 2003. Embodied artificial intelligence. Artificial Intelligence 149:131–150.
Clausner, T. C., and Croft, W. 1999. Domains and image schemas. Cognitive Linguistics 10(1):1–
Cohn, A. G., and Renz, J. 2007. Qualitative Spatial Representation and Reasoning. In van
Harmelen et al., F., ed., Handbook of Knowledge Representation. Oxford: Elsevier. 551–596.
Davidson, D. 1967. The logical form of action sentences. In Rescher, N., ed., The logic of decision
and action. 81–94.
Fauconnier, G., and Turner, M. 1998. Conceptual Integration Networks. Cognitive Science
Fauconnier, G., and Turner, M. 2003. The Way We Think: Conceptual Blending and the Mind’s
Hidden Complexities. Basic Books.
Feldman, J., and Narayanan, S. 2004. Embodied meaning in a neural theory of language. Brain
and Language 89(2):385–392.
Forbus, K.; Falkenhainer, B.; and Gentner, D. 1989. The structure-mapping engine. Artificial
Intelligence 41:1–63.
Gallese, V., and Lakoff, G. 2005. The Brain’s concepts: the role of the Sensory-motor system in
conceptual knowledge. Cognitive neuropsychology 22(3):455–79.
ardenfors, P. 2007. Embodiment in Cognition and Culture, volume 71 of Advances in
Consciousness Research. John Benjamins Publishing Company. chapter Cognitive semantics
and image schemas with embodied forces, 57–76.
Gentner, D. 1983. Structure mapping: A theoretical framework for analogy. Cognitive Science
Gibson, J. J. 1977. The theory of affordances, in Perceiving, Acting, and Knowing. Towards an
Ecological Psychology. In Shaw, R., and Bransford, J., eds., Perceiving, Acting, and Knowing:
Toward an Ecological Psychology. Hillsdale: NJ: Lawrence Erlbaum. 67–82.
Goguen, J. A., and Harrell, D. F. 2010. Style: A Computational and Conceptual Blending-Based
Approach. In Argamon, S., and Dubnov, S., eds., The Structure of Style: Algorithmic Approaches
to Understanding Manner and Meaning. Berlin: Springer. 147–170.
Goldin, G. A. 2001. Counting on the metaphorical. Nature 413(6851):18–19.
uninger, M.; Hahmann, T.; Hashemi, A.; Ong, D.; and Ozgovde, A. 2012. Modular First-Order
Ontologies Via Repositories. Applied Ontology 7(2):169–209.
Hedblom, M. M.; Kutz, O.; and Neuhaus, F. 2014. On the cognitive and logical role of
image schemas in computational conceptual blending. In Proceedings of the 2nd International
Workshop on Artificial Intelligence and Cognition (AIC-2014), Torino, Italy, November 26th–
27th, volume Volume 1315 of CEUR-WS.
Hedblom, M. M.; Kutz, O.; and Neuhaus, F. 2015. Image Schemas as Families of Theories.
In Besold, T. R.; K¨
uhnberger, K.-U.; Schorlemmer, M.; and Smaill, A., eds., Proceedings of
the Workshop “Computational Creativity, Concept Invention, and General Intelligence” 2015,
volume 2 of Publications of the Institute of Cognitive Science, 19–33. Institute of Cognitive
Johanson, M., and Papafragou, A. 2014. What does children’s spatial language reveal about spatial
concepts? Evidence from the use of containment expressions. Cognitive science 38(5):881–910.
Johnson, M. 1987. The Body in the Mind. The Bodily Basis of Meaning, Imagination, and
Reasoning. The University of Chicago Press.
Kamp, H. 1979. Instants, events and temporal discourse. In B¨
auerle, R.; Schwarze, C.; and von
Stechow, A., eds., Semantics from Different Points of View. Springer, Berlin. 376–417.
Knauff, M.; Rauh, R.; and Renz, J. 1997. A cognitive assessment of topological spatial relations:
Results from an empirical investigation. In Hirtle, S. C., and Frank, A. U., eds., Spatial
Information Theory: A Theoretical Basis for GIS, volume 1329 of Lecture Notes in Computer
Science. Springer. 193–206.
Koestler, A. 1964. The Act of Creation. Macmillan.
ovecses, Z. 2010. Metaphor:A Practical Introduction. Oxford University Press, USA.
Kuhn, W. 2002. Modeling the Semantics of Geographic Categories through Conceptual Integration.
In Proceedings of GIScience 2002, 108–118. Springer.
Kuhn, W. 2007. An Image-Schematic Account of Spatial Categories. In Winter, S.; Duckham, M.;
Kulik, L.; and Kuipers, B., eds., Spatial Information Theory, volume 4736 of Lecture Notes in
Computer Science. Springer. 152–168.
Kutz, O.; Mossakowski, T.; Hois, J.; Bhatt, M.; and Bateman, J. 2012. Ontological Blending in
DOL. In Besold, T.; K¨
uhnberger, K.-U.; Schorlemmer, M.; and Smaill, A., eds., Computational
Creativity, Concept Invention, and General Intelligence, Proceedings of the 1st International
Workshop C3GI@ECAI, volume 01-2012. Montpellier, France: Publications of the Institute of
Cognitive Science, Osnabr¨
Kutz, O.; Bateman, J.; Neuhaus, F.; Mossakowski, T.; and Bhatt, M. 2014a. E pluribus unum:
Formalisation, Use-Cases, and Computational Support for Conceptual Blending. In Besold, T. R.;
Schorlemmer, M.; and Smaill, A., eds., Computational Creativity Research: Towards Creative
Machines, Thinking Machines. Atlantis/Springer.
Kutz, O.; Neuhaus, F.; Mossakowski, T.; and Codescu, M. 2014b. Blending in the Hub—Towards a
collaborative concept invention platform. In Proceedings of the 5th International Conference on
Computational Creativity.
Kutz, O.; Mossakowski, T.; and L¨
ucke, D. 2010. Carnap, Goguen, and the Hyperontologies: Logical
Pluralism and Heterogeneous Structuring in Ontology Design. Logica Universalis 4(2):255–333.
Special Issue on ‘Is Logic Universal?’.
Lakoff, G., and Johnson, M. 1999. Philosophy in the Flesh. Basic Books.
Lakoff, G., and N´
nez, R. 2000. Where Mathematics Comes from: How the Embodied Mind Brings
Mathematics Into Being. New York: Basic Books.
Lakoff, G. 1987. Women, Fire, and Dangerous Things. What Categories Reveal about the Mind.
The University of Chicago Press.
Louwerse, M. M., and Jeuniaux, P. 2010. The linguistic and embodied nature of conceptual
processing. Cognition 114(1):96–104.
Mandler, J. M., and Pag´
an C´
anovas, C. 2014. On defining image schemas. Language and Cognition
Mandler, J. M. 1992. How to build a baby: II. Conceptual primitives. Psychological review
Mandler, J. M. 2004. The Foundations of Mind : Origins of Conceptual Thought: Origins of
Conceptual Though. New York: Oxford University Press.
Mandler, J. M. 2008. On the Birth and Growth of Concepts. Philosophical Psychology 21(2):207–
Menzel, C. 2011. Knowledge representation, the World Wide Web, and the evolution of logic.
Synthese 182:269–295.
Morgenstern, L. 2001. Mid-Sized Axiomatizations of Commonsense Problems: A Case Study in
Egg Cracking. Studia Logica 67:333–384.
Mossakowski, T.; Kutz, O.; Codescu, M.; and Lange, C. 2013. The Distributed Ontology, Modeling
and Specification Language. In Del Vescovo, C.; Hahmann, T.; Pearce, D.; and Walther, D.,
eds., Proceedings of the 7th International Workshop on Modular Ontologies (WoMO-13), volume
1081. CEUR-WS.
Mossakowski, T.; Codescu, M.; Neuhaus, F.; and Kutz, O. 2015. The Road to Universal
Logic–Festschrift for 50th birthday of Jean-Yves Beziau, Volume II. Studies in Universal Logic.
auser. chapter The distributed ontology, modelling and specification language - DOL.
Mossakowski, T.; Kutz, O.; and Codescu, M. 2014. Ontohub: A semantic repository for
heterogeneous ontologies. In Proc. of the Theory Day in Computer Science (DACS-2014).
Satellite workshop of ICTAC-2014.
Mossakowski, T.; Lange, C.; and Kutz, O. 2012. Three Semantics for the Core of the Distributed
Ontology Language. In Gr¨
uninger, M., ed., 7th International Conference on Formal Ontology in
Information Systems (FOIS), Frontiers in Artificial Intelligence and Applications. IOS Press.
Mossakowski, T.; Maeder, C.; and L ¨
uttich, K. 2007. The Heterogeneous Tool Set. In Grumberg,
O., and Huth, M., eds., Tools and Algorithms for the Construction and Analysis of Systems.
13th International Conference, TACAS 2007, Held as Part of the Joint European Conferences
on Theory and Practice of Software, ETAPS 2007 Braga, Portugal, March 24 - April 1, 2007.
Proceedings, volume 4424 of Lecture Notes in Computer Science, 519–522. Springer.
Nayak, S., and Mukerjee, A. 2012. Concretizing the image schema: How semantics guides the
bootstrapping of syntax. In 2012 IEEE International Conference on Development and Learning
and Epigenetic Robotics, ICDL 2012.
Oakley, T. 2010. Image Schema. In Geeraerts, D., and Cuyckens, H., eds., The Oxford Handbook
of Cognitive Linguistics. Oxford: Oxford University Press. 214–235.
on, S., and Plaza, E. 2010. Amalgams: A formal approach for combining multiple case
solutions. In Case-Based Reasoning. Research and Development. Springer. 257–271.
Parson, T. 1990. Events in the Semantics of English: A Study in Subatomic Semantics. MIT Press.
Pereira, F. C., and Cardoso, A. 2003. Optimality Principles for Conceptual Blending: A First
Computational Approach. AISB Journal 1(4).
Prior, A. N. 1967. Past, Present and Future. Oxford: Oxford University Press.
Regier, T. 1996. The Human Semantic Potential: Spatial Language and Constrained
Connectionism. The MIT Press.
Rohrer, T. 2005. Image Schemata in the Brain. In Hampe, B., and Grady, J. E., eds., From perception
to meaning: Image schemas in cognitive linguistics, volume 29 of Cognitive Linguistics Research.
Walter de Gruyter. 165–196.
Schiralli, M., and Sinclair, N. 2003. A Constructive Response to ‘Where Mathematics Comes
From’. Educational Studies in Mathematics 52:79–91.
Schmidt, M.; Krumnack, U.; Gust, H.; and Khnberger, K.-U. 2014. Heuristic-Driven Theory
Projection: An Overview. In Prade, H., and Richard, G., eds., Computational Approaches to
Analogical Reasoning: Current Trends, Computational Intelligence 548. Springer-Verlag.
Schorlemmer, M.; Smaill, A.; K ¨
uhnberger, K.-U.; Kutz, O.; Colton, S.; Cambouropoulos, E.;
and Pease, A. 2014. COINVENT: Towards a Computational Concept Invention Theory. In
Proceedings of the 5th International Conference on Computational Creativity.
Schwering, A.; Krumnack, U.; K ¨
uhnberger, K.-U.; and Gust, H. 2009. Syntactic Principles of
Heuristic-Driven Theory Projection. Cognitive Systems Research 10(3):251–269.
St. Amant, R.; Morrison, C. T.; Chang, Y.-H.; Cohen, P. R.; and Beal, C. 2006. An image schema
language. In International Conference on Cognitive Modeling (ICCM), 292–297.
Tettamanti, M.; Buccino, G.; Saccuman, M. C.; Gallese, V.; Danna, M.; Scifo, P.; Fazio, F.;
Rizzolatti, G.; and Perani, D. 2005. Listening to action-related sentences activates fronto-parietal
motor circuits. Journal of Cognitive Neuroscience 273–281.
Turner, M. 2007. The Way We Imagine. In Roth, I., ed., Imaginative Minds - Proceedings of the
British Academy. Oxford: OUP. 213–236.
Turner, M. 2014. The Origin of Ideas: Blending, Creativity, and the Human Spark. Oxford
University Press.
Van Benthem, J. F. A. K. 1983. The Logic of Time. Dordrecht, Holland: D. Reidel Publishing
Veale, T., and Keane, M. T. 1992. Conceptual Scaffolding: a Spatially Founded Meaning
Representation for Metaphor Comprehension. Computational Intelligence 8(3):494–519.
Veale, T.; Feyaerts, K.; and Forceville, C. 2013. E Unis Pluribum: Using Mental Agility to
Achieve Creative Duality in Word, Image and Sound. In Creativity and the Agile Mind: A Multi-
Disciplinary Study of a Multi-Faceted Phenomenon (Applications of Cognitive Linguistics). 37–
Vernon, D. 2014. Artificial Cognitive Systems: A Primer. MIT Press.
Voorhees, B. 2004. Embodied Mathematics: Comments on Lakoff and N´
nez. Journal of
Consciousness Studies 11(9):83–88.
Walton, L., and Worboys, M. 2009. An Algebraic Approach to Image Schemas for Geographic
Space. In Proceedings of the 9th International Conference on Spatial Information Theory
(COSIT), 357–370.
Wierzbicka, A. 1996. Semantics : Primes and Universals: Primes and Universals. Oxford
University Press, UK.
Wilson, N. L., and Gibbs, R. W. 2007. Real and imagined body movement primes metaphor
comprehension. Cognitive science 31(4):721–731.
... They may be employed to establish rigorous definitions that capture part of the meaning of natural language expressions. It is generally believed that analyzing natural language leads to a greater understanding of image schemas [6, 15]. Research on image schemas is performed in several disciplines; cognitive linguis- tics [5], developmental psychology [14] and more formal areas (e.g. ...
... In their investigation they were able to identify eight different types of CONTAINMENT. Hedblom et al. [6] showed how the image schema PATH-following represents a family of theories rather than an individual schema and demonstrated how this can be used to ground abstract concepts. With examples, they demonstrated how image schemas capture the information skeleton in (some) linguistic metaphors. ...
... Aiming to take the above mentioned aspects of image-schematic structures, components and combinatorial possibilities into account, Hedblom et al. [6] took a closer look at what in the literature is called the SOURCE PATH GOAL schema [12] . They presented a hierarchical structure, the PATH-following family (see Figure 1), that grew more specialised based on the addition of spatial primitives found in developmental psychology [15]. ...
Conference Paper
Full-text available
Image schemas provide preverbal conceptual structures and are suggested to be the conceptual building blocks from which cognitive phenomena such as language and reasoning are constructed. 'Motion along a path' is one of the first image schemas infants remember, making PATH-following one of the earliest cog-nitive building blocks. We are interested in the importance of this developmentally relevant image schema in abstract adult language. For this purpose, we propose a semi-automated method to extract image-schematic structures related to PATH-following from a multilingual financial terminology. Two major assumptions are that a linguistic mapping of image schemas facilitates the understanding of complex concepts and is persistent across languages. Our results show that complex textual representations can be made simpler to understand by extracting the underlying image schemas and that they are persistent across languages. Another result includes the identification of novel specifications of predefined image-schematic structures.
... Thereby, we focused on the identification of image schemas, and analysis of linguistic expressions. A description of image schemas in a large body of science education and philosophy literature [41,44,[59][60][61][62] thereby functioned as source for their identification in the text. To identify these, we focused on semantical structures such as prepositions, postpositions, verbs, bodypart-metaphors, cases and morphemes. ...
Full-text available
By means of their pivotal role in the outbreak of a variety of diseases, such as, recently, COVID-19, the molecular aspects of cell membrane function have gained considerable attention from researchers in recent decades. The resulting information explosion and the growing interdisciplinary character of cell biology seems, however, to not be represented in science classrooms. Hence, there appears to be a gap between what is scientifically known and what is actually taught in classrooms. Framed by the model of educational reconstruction (MER), the aim of our study is therefore to identify scientific core ideas of cell membrane biology from an educational point of view. This is achieved by conducting qualitative content analysis of relevant cell biology literature. By using Conceptual Metaphor as a theory of understanding, we additionally illuminate the experiential grounding of scientific conceptions. Our results propose that cell membrane biology can be structured into three core ideas, comprising compartmentalisation, physical and chemical properties, and multicellular coordination interrelated by evolution as a key aspect. Our results show that scientists conceive these ideas metaphorically. Embodied part-whole relations seem, for example, to lay the grounds for their understanding of biological function. The outcomes of the study may inform future cell membrane teaching.
... They distinguished no less than eight different CONTAINMENT schemas from natural language (see Figure 3). Similarly, Hedblom et al. visualised image schemas as a series of theories of increasing complexity [7] (see Figure 3). They aimed to explain how image schemas become more fine-tuned in cognitive development as a result of a complex environment [17, 22], by using the PATH-following image schema as evidence. ...
Conference Paper
Full-text available
Embodied cognition has emphasised the importance of sensorimotor processes for concept formation. According to this theory, image schemas are suggested to be a cognitive representation of certain basic spatial relationships. At the same time, the shape of objects is essential for categorisation and understanding. While information on both shapes and image schemas are primarily obtained through visual perception, their cognitive nature differs. Shape is an attribute of an object, whereas an image schema is a basic spatial relationship between (several) objects in a particular environment and context. We here discuss interdependencies between these two modes of perception with regard to their roles in concept formation.
... In order to implement this idea, a sufficiently comprehensive library of formalised image schema theories has to be made available for access by a blending engine. The first such library for the case of 'path following' is developed in [26]. ...
In cognitive science, image schemas are identified as fundamental patterns of cognition. They are schematic prelinguistic conceptualisations of events and serve as conceptual building blocks for concepts. This paper proposes that image schemas can play an important role in computational concept invention, namely within the computational realisation of conceptual blending. We propose to build a library of formalised image schemas, and illustrate how they can guide the search for a base space in the concept invention work flow. Their schematic nature is captured by the idea of organising image schemas into families. Formally, they are represented as heterogeneous, interlinked theories.
Full-text available
The notion of affordance remains elusive, notwithstanding its importance for the representation of agency, cognition, and behaviors. This paper lays down a foundation for an ontology of affordances by elaborating the idea of “core affordance” which would serve as a common ground for explaining existing diverse conceptions of affordances and their interrelationships. For this purpose, it analyzes M. T. Turvey’s dispositional theory of affordances in light of a formal ontology of dispositions. Consequently, two kinds of so-called “core affordances” are proposed: specific and general ones. Inspired directly by Turvey’s original account, a specific core affordance is intimately connected to a specific agent, as it is reciprocal with a counterpart effectivity (which is a disposition) of this agent within the agent-environment system. On the opposite, a general core affordance does not depend on individual agents; rather, its realization involves an action by an instance of a determinate class of agents. The utility of such core affordances is illustrated by examining how they can be leveraged to formalize other major accounts of affordances. Additionally, it is briefly outlined how core affordances can be employed to analyze three notions that are closely allied with affordances: the environment, image schemas, and intentions.
Full-text available
Image schemas were introduced as mental generalisations learned from the sensorimotor experiences in infancy that in adulthood shape language formation and conceptualisations. So far, little empirical research has been devoted to investigate to which degree image schemas are involved in object conceptualisation more concretely. To address this, this experimental study investigates the relationship between abstract image schemas and their involvement in conceptualisations of common, everyday objects. The experimental set-up asks participants to describe objects using abstract representations of image schemas. The results from the study support the claim that image-schematic thinking is prevalent in the conceptualisation of objects, thus providing empirical evidence for the idea that image schemas can serve as conceptual building blocks for the meaning of objects.
Full-text available
Formal knowledge representation struggles to represent the dynamic changes within complex events in a cognitively plausible way. Image schemas, on the other hand, are spatiotemporal relationships used in cognitive science as building blocks to conceptualise objects and events on a high level of abstraction. In this paper, we explore this modelling gap by looking at how image schemas can capture the skeletal information of events and describe segmentation cuts essential for conceptualising dynamic changes. The main contribution of the paper is the introduction of a more systematic approach for the combination of image schemas with one another in order to capture the conceptual representation of complex concepts and events. To reach this goal we use the image schema logic ISL, and, based on foundational research in cognitive linguistics and developmental psychology, we motivate three different methods for the formal combination of image schemas: merge, collection, and structured combination. These methods are then used for formal event segmentation where the changes in image-schematic state generate the points of separation into individual scenes. The paper concludes with a demonstration of our methodology and an ontological analysis of the classic commonsense reasoning problem of ‘cracking an egg.’
Conference Paper
Full-text available
Cognitive linguistics introduced image schemas as a missing link between embodied experiences and high-level conceptualisation in language and metaphorical thinking. They are described as the abstract spatio-temporal relationships that function as conceptual building blocks for everyday concepts and events. Although there is increasing interest in the area of cognitively motivated artificial intelligence, where image schemas are suggested to be a core piece in the puzzle to model human-level conceptualisation and reasoning, so far rather few formal logical approaches can be found in the literature, in particular regarding attention to the dynamic aspects of image schemas. A fundamental problem here is that the typical mainstream approaches in contemporary KR do not map well to various scenarios found in image schema modelling. In this paper, we introduce a spatio-temporal logic for ‘directed movement of objects’, with the aim to model formally image schematic events such as Blockage, Caused_Movement and ‘bouncing’.
This volume collects Davidson's seminal contributions to the philosophy of mind and the philosophy of action. Its overarching thesis is that the ordinary concept of causality we employ to render physical processes intelligible should also be employed in describing and explaining human action. In the first of three subsections into which the papers are thematically organized, Davidson uses causality to give novel analyses of acting for a reason, of intending, weakness of will, and freedom of will. The second section provides the formal and ontological framework for those analyses. In particular, the logical form and attending ontology of action sentences and causal statements is explored. To uphold the analyses, Davidson urges us to accept the existence of non‐recurrent particulars, events, along with that of persons and other objects. The final section employs this ontology of events to provide an anti‐reductionist answer to the mind/matter debate that Davidson labels ‘anomalous monism’. Events enter causal relations regardless of how we describe them but can, for the sake of different explanatory purposes, be subsumed under mutually irreducible descriptions, claims Davidson. Events qualify as mental if caused and rationalized by reasons, but can be so described only if we subsume them under considerations that are not amenable to codification into strict laws. We abandon those considerations, collectively labelled the ‘constitutive ideal of rationality’, if we want to explain the physical occurrence of those very same events; in which case we have to describe them as governed by strict laws. The impossibility of intertranslating the two idioms by means of psychophysical laws blocks any analytically reductive relation between them. The mental and the physical would thus disintegrate were it not for causality, which is operative in both realms through a shared ontology of events.
This chapter provides a concise overview of Heuristic-Driven Theory Projection (HDTP), a powerful framework for computing analogies. The chapter attempts to illuminate HDTP from several different perspectives. On the one hand, the syntactic basis of HDTP is formally specified, in particular, restricted higher-order anti-unification together with a complexity measure is described as the core process to compute a generalization given two input domains (source and target). On the other hand, the substitution-governed alignment and mapping process is analyzed together with the transfer of knowledge from source to target in order to induce hypotheses on the target domain. Additionally, this chapter presents some core ideas concerning the semantics of HDTP as well as the algorithm that computes analogies given two input domains. Finally, some further remarks describe the different (but important) roles heuristics play in this framework.