Analogy
Charles Tijus
Abstract Among cognitive processes, although literally based on logically false
propositions, analogical and metaphorical reasoning are the most used and useful
thinking for communicating, understanding, discovering, problem-solving and
learning. The topic of this chapter about analogy and metaphor, is to address the
kind of computation linking a target category to a source category that belongs to
another domain that might be able to support reasoning properties based on the
fallacy of the falsity of propositions, on imperfection, imprecision and approximation, gradualness, vagueness, fuzziness, uncertainty and implicit plausibility of likeness. Because notable advances in the computation of analogies are from Bernadette
Bouchon-Meunier’s work with her team: the fuzzy logic computation of analogical
reasoning and schemes, we examine how such modeling of the hu-man computation
of analogies can be used in turn to model the machine computa-tion of analogies.
1 Introducing Analogy as One of Two Main Ways
of Thinking
Cognitive processes of both human and machine are for understanding and for
decision-making. These processes are cognitive investigations about external physical things, including people and oneself; about who they are, what they are (structure
and functioning), about the events and actions in which they participate, as well as
ascribing intention to other agents [1, 2]. Although a cognitive system requires information from the external physical world, investigations about external things are
firstly based on what we already know about them: on their internal representation, including the kind of things they are and the categories to which they belong.
This inference-making about what things are and about how they behave are made
for interaction, for decision-making about our judgments about them, about actions
C. Tijus (B)
CHArt, University Paris 8, 2, rue de la Liberté 93526 St Denis Cedex 02, Paris, France
e-mail: tijus@univ-paris8.fr
© The Editor(s) (if applicable) and The Author(s), under exclusive license
to Springer Nature Switzerland AG 2021
M.-J. Lesot and C. Marsala (eds.), Fuzzy Approaches for Soft Computing and Approximate
Reasoning: Theories and Applications, Studies in Fuzziness and Soft Computing 394,
https://doi.org/10.1007/978-3-030-54341-9_6
61
62
C. Tijus
we might perform on them or with them. Classically, there are four kinds of inferences of these cognitive processes—deduction, induction, analogy and metaphor [3,
4]—that we propose to be reduced to two main kinds: deduction and induction as
one kind and analogy and metaphor as another kind.
Inference-making of the first kind is literal, which means that the target domain
of investigation is domain specific constrained: restricted to the categories of this
domain, restricted to the intension and extension of these categories and to the relations among these categories. In that case, things under investigation are supposedly
known; as in computer programming for which a literal value is a value given explicitly in the source code of the algorithm for a given variable. Such literal inferencemaking classically stands for both deduction and induction: either from category to
exemplar (i.e. deduction: what is known about a category can be attributed to subordinate categories and finally to exemplars of these categories): Cats meow, Maine
Coon is a kind of cat, Felix is a Maine Coon, Felix should meow. It stands also from
exemplars to category (i.e. induction: what is known about exemplars of a category
could be attributed to this category): Felix is a cat and meows, Tom is also a cat and
meows: cats might meow. In that case, the top-down (from category to exemplars) or
bottom-up (from exemplars to category) processes are about things known to belong
to the same given category or, for more complex thinking and reasoning, to the same
specific domain.
Contrary to deduction and induction that provide “true” sentences, Inferencemaking of the second kind is unliteral (i.e. not literal; not literally comparable).
Things under the cognitive investigation are cross-domain constrained: they are of
different categories, usually of exclusive domains: inference-making is from singular
to singular, from an exemplar of a given category to exemplar of another category.
Such unliteral inference-making is found with analogical propositions based on
comparison (e.g. the nucleus of the atom is the sun of the solar system) but also with
metaphorical propositions [5] based on attributive categorization (e.g. John is a cat):
John, a troubadour, sings to beg for a little love. Felix, a cat, meows to beg for food.
John’s voice sounds like meowing: propositions such as “John is meowing” or “John
is a cat” are analogies. Such associations, like “Juliet is my sun” (i.e. she brings
me joy and light) in from Romeo’s diary of Shakespeare’s “Romeo and Juliet,” are
based on some fuzzy resemblance, on analogy or metaphor (upon, according to/so
to speak). According to Plato, analogies are founded on a reasoning based on “an
argument from the similarity of things in some ways inferring their similarity in
others” and on a computation based on “partial agreement, likeness or proportion
between things”. Although such analogically-based reasoning provides sentences
that do not have evidential support and are logically “false”, this kind of reasoning
appears to be the most prominent kind of human of thinking [6] and maybe the one
that machines will be using in the very near future for understanding, thinking and
communicating:
• For analogical reasoning that is based on comparison: [A: the nucleus] is to [B:
the Atom] what [C: the sun] is to [C: the solar system]. If one knows the relation
Analogy
63
between the sun and the solar system, s/he can infer the relation between the
nucleus and the atom.
• For metaphorical reasoning that is based on attributive categorization [B: singing]
is to [A: John] what [D: meowing] is to [C: Felix]. If one knows that Felix is
meowing to beg, the conclusion is that John is begging. Thus, Felix is the source
of the analogy, while John is the target of the analogy.
Note that to produce an analogy or a metaphor, the computation is from-targetto-source. For example, to produce “Juliet is my sun” from Juliet Shakespeare had
to find a likeness source. In order to understand the likeness of the source, the cognitive investigation of the listener is a from-source-to-target computation. Thus, to
understand “Juliet is my sun” from “sun,” one must find the likeness to attribute to
Juliet.
Note also that because analogy is based on comparison (x is equivalent to), source
and target can permute. For example, “the nucleus of the atom is the sun of the solar
system” is equivalent to “the sun of the solar system is the nucleus of the atom.”
Conversely, because metaphor is based on implication (x is a kind of y), source and
target cannot permute. “John is a cat,” the reverse is not true: “Cats are not as John”
[C].
As a matter of fact, there is a very challenging scientific and technological issue
to discover and model the what and how of the thinking processes that are able to
produce and understand analogies. The kind of computation linking a target category to a source category that belongs to another domain might be able to support
reasoning properties such as imperfection, imprecision and approximation, gradualness, vagueness, fuzziness, uncertainty and implicit plausibility of likeness. Notable
advances in the computation of analogies are from Bernadette Bouchon-Meunier’s
work with her team: the fuzzy logic computation of analogical reasoning and schemes
[7–16].
2 The Necessity of Fuzzy Logic Computation of Analogical
Reasoning and Schemes
The first main advance of Bernadette Bouchon-Meunier (BBM) about analogical
reasoning is due to the hint of using the main principle of fuzzy logic: i.e. the gradualness of membership function. So, as crisp sets representing precise and certain
descriptions of objects might be regarded as particular cases of fuzzy sets [10],
tautology as identity (the sun is the sun), analogy as comparison (the nucleus of the
atom is the sun of the solar system) and antilogy as metaphor (Juliet is the sun of
Romeo) are fuzzy sets of special kinds. Identity is computed as a particular case
of analogy. Analogy is computed as a particular case of metaphor, with transitivity,
asymmetry and irreflexivity relations.
The first main advance of BBM on analogical reasoning is due to her insight that
the main principle of fuzzy logic might be the central core of analogical thinking
64
C. Tijus
and reasoning. There is a graduality of membership function that can be used both
for “John is a man” and “John is a cat”; certainty (John is a man) being a special
case of uncertainty (John is a cat). So, as crisp sets that represent precise and certain
descriptions of objects, they are to be regarded as particular cases of fuzzy sets [10].
Tautology as identity (the sun is the sun), analogy as comparison (the nucleus of the
atom is as the sun of the solar system) and antilogy as metaphorical contradiction
(Juliet is the sun although Juliet is not the sun) are all fuzzy sets of special kinds.
Identity is computed as a particular case of analogy. Analogy is computed as a
particular case of metaphor, with transitivity, asymmetry and irreflexivity relations.
The second advance of BBM’s team about analogical and metaphorical reasoning
is embedded by the first one. This advance relates to tautology that appears to be
very useful in daily life activities,—as well as antilogy (In London, even when it’s
not raining, it’s raining!) -, to provide useful information. However, according to
classic logic, tautologies such as “the sun is the sun” and “Paris is Paris” are per
se uninformative. Similarly, according to the Grice’s maxims of pragmatics [17],
although they respect quality (truthful, supported by evidence) and manner (avoiding
obscurity and ambiguity), tautologies violate two other maxims: quantity (to be
as informative as possible) and relation (be pertinent). Here again with the BBM
approach, tautologies and antilogies can be seen as special cases of graduality of
certainty-uncertainty, in contradt to a full or null membership. As will be seen in the
next section, the same categorization-based cognitive process appears to be a good
candidate for the computing all of these forms of metaphorical reasoning [18].
The third advance is that analogical reasoning and metaphorical reasoning are
cross-domains: a target object T (Tenor), is investigated from the point of view of
a source that is used as a V (vehicle) for transmission of meanings. Having similes
among the same category (tautology), among different categories of the same domain
(analogy), among different categories across domains, or among inter-domains categories (metaphor), T and V can be computed according to the evaluation of their
“closeness” through fuzzy modifiers in order to measure their similarity [16]. The
advance is that closeness of two objects (i.e. moon and sun) as a semantic distance
(very close, close, far, very far) can be described through fuzzy sets that can manage
approximation. Closeness approximation depends on to T and V role (the moon-sun
closeness in “the moon is a sun” being of a different value in “the sun is a moon”),
according to context (the moon-sun closeness in “tonight the moon is a sun” being of
a different value in “today the moon is a sun”), and motive (the moon-sun closeness
in “in your drawing, the moon is a sun” being of a different value in “in the sky, the
moon is a sun”).
Moving on, we go further to develop (i) what analogy is and what analogy is
not, (ii) analogical reasoning as being metaphorical reasoning, (iii) the powerful
use of approximation and imprecision by the brain using analogies and metaphors,
(iv) the categorical human resolution of analogies and metaphors through fuzzy
inference-making and finally (v) models of solving analogies and metaphors; a
section on artificial cognition could mimic human cognition for producing and
understanding analogical and metaphorical thinking.
Analogy
65
3 What Analogy is and What Analogy is Not
Analogy might play an important role in epistemology, history of art, scientific
discoveries and innovation, but also in the methods of doing art, science and techniques. There are many historical narratives about the emergence of new ideas, of
discoveries and of problem solving. Analogy is generally described in the form “A
is to B as C is to D” (A:B::C:D); where the source is some kind of substitute for
the target for thinking and reasoning [19]. For instance, a well-known narrative is
about how Archimedes found a solution to know whether the crown of King Hiero
of Athens was really made out of pure gold, or if it was contaminated with cheap
silver. After a long day of worrying, he decided to relax with a warm bath. When
he entered the tub, he noticed the water level rising. This was something he knew,
but now he suddenly realized that the water displacement was proportionate to the
volume of the immersed part of his body. Then he put a weight of gold equal to the
crown in a bowl filled with water. Next, the gold was removed and replaced by the
crown. Silver would increase the bulk of the crown and cause the bowl to overflow.
Thus, his body was a cognitive substitute for the crown to solve the problem.
In everyday life, analogies are used and solved either for symbolic representation of things with verbalization, or when acting with physical things, or both. For
instance, a verbal analogical problem solving in the form of “The railway is to the
train as the sea is to the boat” (railway: train:: sea: boat) while the corresponding
physical analogical problem solving is having the engineer in the restless train.
Problem solving of verbal scholar analogies such as “The railway is to the train
as the sea is to boat” (railway: train:: sea: boat). Problems have three terms, to solve
for the fourth (e.g. A: train:: sea: boat) or have one set of two terms and their relation
(railway: train:: C: D) to find the one that has the simile relation among other sets of
objects (wheels: car; sea: boat; passengers: bus).
Note first that, in contrast to the real world problems, these scholar problems have
simile relations that are academic (e.g. synonym, antonym, part-to-whole, category/type, object-to-function, performer to related action, cause and effect, degree of
intensity, and symbol and representation) and that the relation is given. For instance, in
a real world problem, an engineer is asked, to correct and assure passengers comfort
because “this train is a boat” making the relation under investigation implicit. It
could be in that case that the train behaves as a boat because the train’s rails are to the
passengers what the rough seas would be for the passengers of a boat. However,
the relation could include other elements (e.g. due to the wind, due to the mountain
shape as waves, and due to the wheels). Thus, the analogy is not in the form “A is to
B as C is to D” but rather in the form “A is like C”.
Secondly, as a matter of facts, not a single relationship but rather many are candidates for concluding the analogical solution. In addition, the conclusive solution can
be made of a, set of relations with their interactions.
Thirdly, analogy is supposedly done based on literal similarity comparisons [13,
20, 21]. However, between two natural knowledge domain, A and B, it is hard for
individuals to evaluate the similarity between A and B. It is hard to evaluate the size
66
C. Tijus
of the intersection (A ∩ B) of their feature sets as well as the alignment of these
features: which feature in A matches a given feature in B. Such correspondence may
also be at a given level of the decomposition tree of attributes but not in others (e.g.
while red in A and blue in B do not match at the value level; they match as being a
color). Such correspondence may also be at one of the dimensions of the domain
description (e.g. dimension of surface, of structure, of function, of procedure, of
dynamic behaviour).
Fourthly, based on the comparison of A and C, the analogy in the form “A is
to B as C is to D” can be permuted as “C is to D as A is to B.” We reasoned that
such inter-domain comparisons require precise structural alignment and mapping
[16] that must be hard to find. The equivalence between “A is to B” and “C is to
B” should be rare, making analogical relations much more oriented from target to
source than from source to target. This also means that analogy is from complex to
simple. From this point of view, analogy is simplexity [22], given that meanings in
everyday life language provide instructions to build understandable points of view
[23]. Thus, many analogies appear to be assertions that have a metaphorical form,
which is to say, not reversible: if New York is a big apple, a big apple is not New
York.
This metaphor asymmetry is found with classical analogical problem solving, for
instance between two isomorphic sub-domains of algebra and physics [24]. When
students, who learned one of these subtopics and are familiar with that source, are
presented with a target problem based on the unfamiliar but analogous domain (as in
metaphor) the source-to-target transfer is asymmetric. Students who had learned
arithmetic (source) were very likely to transfer to physics (target). In contrast,
students who had learned physics (source) almost never exhibited transfer to the
isomorphic algebra problem (target). If physics is recognized as algebra, algebra is
not recognized as physics.
For studying the underlying cognitive processes of analogy resolution, an alternative of using already known natural domains is to build up unknown experimental
isomorphic micro domains. The building up of experimental domains assert the
analogical match of the two A and B descriptions, while minimizing the noise of
complement sets of A’s features that are not in B (A–B) and the reverse (B–A).
The prominent Cognitive science work of Herbert Simon and collaborators [25],
as well as followers [26], is based on using domain-free puzzles such as the Tower
of Hanoi (TOH) and the build up of term-to-term isomorphs. Thus, there is full
similarity between two A and B TOH isomorphs that are done that way. The same
space problem, the same transitions from state to state, each state in A is having
its analog in B and the same minimum number of moves for reaching the goal
state. Thus, the two TOH isomorphs have the same deep structure, but a different
surface appearance. A persistent experimental result in the literature is that much
exploration is often involved in solving some of the isomorphs, whose problem spaces
are identical, but are packaged differently according to their surface properties. For
instance, the classical 3-disks-TOH (e.g. A) is made of disks of permanent size
that change place while a possible isomorphic problem (e.g. B) is made of disks
of permanent place that change size. Although 7 moves are enough to solve both
Analogy
67
problems, some of these are easy to solve with 11 moves on average (A) while others
require up to 120 moves (B). Some isomorphs (as B) of the same problem space take
16 times as long to solve as other isomorphs. The difficulty varies by a factor of 16,
depending on the surface characteristics.
Studies of analogical transfer between isomorphs of these well defined puzzles
[27, 28] show that solving a particular problem does not help solving an isomorphic
one, except when conditions of transfer are based on generalization and internalization. On the contrary, from externalization according to surface properties and
context, learning cannot be transferred when the problem content provides information that can be perceived and used without being explicitly interpreted and
formulated.
For instance, there are problems for which the solution requires discovering that
each winning state is made of an even number. What are the conditions for transferring this solution to another isomorphic target problem? How much of the “even
number” concept can be generalized to the same problem but have different surface
properties? Here again, it is found that the analogical gain of solving isomorphic
problems is asymmetric.
Let’s examine two problems. The first one is about numbers. The second one
is about tokens. When the problem that is solved first is solely with numbers, the
discovery that number 4 is a winning state can be learned and used as a source
for the isomorphic target problem with a number of tokens because four tokens is
recognized as number 4, it is a winning state. Conversely, when the source problem
is the problem with tokens, it is not helpful to solve the analogous target problem
solely with numbers. The state number 4 in the target problem is not recognized as
equivalent to four tokens in the source problem.
Thus, because learning is based on generalization and deep structure, a problem
that favors generalization and internalization, such as the problem with solely
numbers, is a profitable source while a problem based on externalization through
its surface properties, such as the problem with tokens, will impair transfer; except
for other problems of same surface resemblance [28]. As physics is recognized as
algebra, four tokens is recognized as number 4. As algebra is not recognized as
physics, number 4 is not recognized as the equivalent of four tokens (one can
externally manipulate four tokens without explicitly counting them).
In summary, as noticed by BBM [9], suppose that the price of a house is to be
determined. This can be done according to several criteria such as its size, its state,
or its location. But because each criterion cannot be independently evaluated, we
need references to evaluate the co-occurrences or correlation of attribute values. The
needs of references could be satisfied with a large amount of data about home prices
and values. A more simple and appropriate reasoning to evaluate the particular target
house is to look for a known particular house that can be used as an intra-domain
source for this case-based reasoning: “this house is a large villa: let’s compare it with
a large villa we know”. The question under investigation is therefore the resemblance
of source and target that might be based on comparing their corresponding features.
Thus, the target house H1 (T) can be compared to the source house H2 (S) on
properties that can be aligned and matched [21]. Within this intra domain comparison
68
C. Tijus
of houses, most of the features, and therefore criteria, can be found in the (T ∩ S)
intersection of their feature sets. It could be that some features (e.g. a swimming
pool) belong to the (T-not S) set while others (e.g. a tennis court) belong to the (not
T-S) set and that alignment might be difficult, but most of the reasoning is among
the (T ∩ S) set of features that can be compared. In addition, target house H1 might
serve as a source for house H2 and vice-versa.
Unlike case-based reasoning, analogy is inter-domains. Someone could be
searching for the “Rolls-Royce of houses” meaning a house that would be among
houses what a Rolls-Royce is among cars. Someone else could have said “I found a
house that is a Rolls Royce.” Since the two domains are different, the resemblance
relations are fuzzy and the features and criteria are hard to be aligned and matched.
Contrary to case-based reasoning, if a Rolls-Royce can serve as a source for a large
villa, a large villa will hardly serve as a source for a car. The main reason (unlike
case-based reasoning) the source is not a particular thing, but a category of things
that serves to make the target inherit the properties of the category. If a Rolls Royce is
luxury, opulence, very expensive, distinctive, solid, well known, then the house that
is in the category of Rolls Royce will have these features. It will inherit these properties as being an attributive category just as a particular living being will inherit the
properties of being a mammal [5]. In summary, we argue that analogy and metaphor
are two faces of the same coin.
4 Models of Solving Analogies and Metaphors for Fuzzy
Inference Making
There are many cognitive models of analogy. See [29] for a review. As for other kinds
of cognitive processes, there are two different approaches of modeling and simulating
the resolution of analogies. The first approach is based on a bottom-up decentralized
process such as the model named AMBR that stands for “Associative Memory-Based
Reasoning.” It goes from local to global, blending episodic-contextual and long
term memories. For solving analogies, memory and reasoning are highly integrated
in Neural Nets and high level features are built starting from the local level. There
is an initial distribution of activation resulting from previously solved problems as
sources that can prime the relevant features of the target. The functioning of the
model can be seen as a collection of basic units that are domain specific cognitive
agents that collaborate according to the declarative (what) and procedural (how/why)
knowledge they encapsulate. Thus, as agents (e.g. “railway” and “sea”) that entail
what (support of train or boat) and how/why (for moving) collaborate to produce
analogies such as “The railway is to the train as the sea is to the boat.” Similarly,
“Juliet” and “sun” entail what (important) and how/why (for living) collaborate to
produce metaphors such as “Juliet is my sun.”
Analogy
69
The second approach is somewhat based on a top-down centralized process such
the one used in the model named SMT that stands for “Structure Mapping Theory”
[21, 30, 31]. In SMT, the similarity between the target and the source is evaluated by computing commonalities and differences. The former provides generalization, abstraction and schemas while the latter provides alignable differences, having
“some expression in the base and some corresponding but different expression in the
target.” The computation of commonalities and differences strengthens the structural
alignment of features that guide the analogical process.
These two approaches of analogy are based on the similarity computation of T and
S; the comparison being symmetrical and reversible: T can be compared to S and S
can be compared to T. When two objects, situations or domains are comparable, either
one or the other can serve as a source or as target. However, this is a particular case
of analogy. Most source-target analogies are oriented: a target can imply a source,
while the reverse is not true.
We assert that there is a clear distinction between physical world objects and the
categories humans use to represent them, in order to think, talk and communicate
about them [32]. Categories as sources for understanding can be used literally as
in deduction and induction, but also unliterally in analogies and metaphors in a
fuzzy way. As [33], we maintain that in both cases, categorizing a target as a source
type yields unseen features. If someone reads that in a fictitious country “Xs are
birds,” then one can infer not only that “Xs are animals,” but also that “Xs fly.”
As semantic relations, analogies and metaphors are based on categorization. They
activate a category and its attached features. They also activate the super-ordinate
categories and their respective attached features.
Our categorization approach of analogies is the one that has already been proposed
for metaphors understanding [5, 18, 34] in which a source (e.g. sun) is a cognitive
vehicle to transfer meanings to the target topic (e.g. Juliet) with the notion that
vehicles in metaphors are attributive categories.
In the past, most of the cognitive models of metaphor understanding have adopted
the approach according to which metaphor is an implicit comparison: understanding
a metaphor “X (topic) is Y (vehicle)” consists in converting it into a simile “X (the
topic) is like Y (the vehicle)”. This comparison-based model of metaphor understanding is a mechanism of property matching. This is the reason these models are
confronted with the problem of measuring the similarity of properties as well as with
the problem of calculating the distance between properties, which makes a simile
literal or unliteral—metaphoric.
More recently, an alternative attributive categorization based approach is the
Glucksberg’s class inclusion model: a metaphoric statement of the type “X is Y”
is solved by looking for the category, represented by the term Y, which furnishes
source properties that are potentially relevant for the target topic X.
The general hypothesis is that metaphor understanding consists of including the
topic in the category of the source-vehicle and attributing to it the properties of
that category that are compatible with what is already known about the topic. We
assume that interpretation is constructed on-line and that knowledge about the topic
intervenes at an early stage in processing by constraining the selection of features.
70
C. Tijus
BBM [10] noticed that crisp sets that represent precise and certain descriptions
of objects might be regarded as particular cases of fuzzy sets. Similarly, we argue
that deduction, induction, abduction, induction, analogy and metaphor are particular
cases of metaphor, from the less fuzzy-certain, to the more fuzzy uncertain, analogy
a particular case of metaphor. According to BBM [11], gradual reasoning can be
obtained by using linguistic modifiers such as in [35, 36], the link between gradual
reasoning and analogical reasoning corresponding to the utilization of a relationship
between variations of X and variations of Y expressed in gradual knowledge to infer
a value of Y from a given value of X. Thus BBM and collaborators introduced a
general framework that represents analogy, on the basis of a link between variables
and measures of comparison between values of variables. This analogical scheme is
a common description of several forms of reasoning used in fuzzy control or in the
management of knowledge-based systems, such as deductive reasoning, inductive
reasoning or prototypical reasoning and gradual reasoning.
A general model for the simulation of those modes of inference-making is a
model based on a fuzzy semantic network [37]. Making hierarchies of categories
of the semantic network with Galois Lattices [38] allows partonomy, which is the
decomposition of an object into its physical parts (the what), and meronomy, which is
the decomposition of a category description into its cognitive parts (the how and why
as well as conceptual features). Unlike classical Galois Lattices, the inheritance of
properties and membership link (e.g. “is a” for category; “is a kind of” for subordinate
category to super-ordinate category), can be interval-valued for fuzzy inclusion all
along the path from instances to subordinate categories, then to the highest general
super-ordinate categories. Another fuzzy measure is the extent a given feature can
possibly be the attached feature of a given category. Within the lineage of categories,
it is important for the structure of a category to distinguish among levels of categorization [39]: among subordinate (siamese), basic level (cat) and super-ordinate (animal)
categories. A distinction that can be made with the partonomy and meronomy decomposition of descriptions; allowing gradual evaluation of concreteness of categories
as well as of the domain of comparisons.
These fuzzy and categorical approaches of analogy differ from those that are based
on similarity computed from features comparisons. These are two different cognitive
approaches since psychological studies and in cognitive science show that similarity
does not match categorization [20]. For instance, the similarity score (Russia, Cuba) is
7 and (Cuba, Jamaica) is 8, similarity (Russia, Jamaica) should be around 7, but the
similarity score is 1! People categorize things differently from the way they evaluate
things to be very similar: similar objects can belong to different categories while
dissimilar objects can belong to the same category.
BBM [11] proposed a fuzzy prototype-based reasoning for making and solving
analogies. A fuzzy prototype of a category enables one to generate typicality and
the set of relevant objects and therefore can be used for matching source to target,
as Tverski’s proposal. The degree of typicality depends on both the resemblance
to other objects in the same category and on the dissimilarity to objects in other
categories. Thus, the analogical question at hand is: “does the target gradually satisfy
Analogy
71
the prototype of the source category?” These are solutions for the different modes of
reasoning, including analogy and metaphor.
5 Discussion
For cognitive purposes, objects are psychologically grouped in categories. Once a
category exists, it has an extension that includes all the instances of the category (even
innumerable) and an intension that includes the properties (even innumerable) shared
by the objects (e.g. unseen). Also for cognitive purposes, categories entail categories
as well as forms of categorical hierarchies. The importance of categorization can be
noticed from the following points of view:
• Reasoning: As categorization corresponds often to an abductive process. When
putting an object in an existing category, we provide it the “rule” or “set” of
properties of the category; and because the “is-a-kind-of” relation entails modus
ponens (Socrate is mortal because Socrate is a kind of person that is a kind of
mortal), but not modus tollens (which is not based on categories; i.e. things that
would be a non-person that are non-mortal);
• Comprehension by inference: Since two things, or two categories, are put together,
common properties of the super-ordinate category, act as a filter, indicating the
“what-is-about” in terms of structure, functionality and usability. For instance, a
piano and a guitar are put in “music and band playing music,” while a piano and
a fridge are put in “large heavy objects” and “how to carry large heavy objects”;
and
• Comprehension of the world structure: Since a category in a hierarchy of categories factorize different kinds of properties and provide the causal links between
procedure-function and structure because the “how-to-use” the object as well as
the “in-order-to” will be based on other features of the object, such as structural
properties. For instance, notice that in folk taxonomy “to have wings” and “flying”
are properties of birds.
Analogies and metaphors are usual modes of thinking and reasoning although
based on false categorization: “electricity is like water,” or “this lawyer is really a
shark.” Thus, there is a powerful use of approximation and imprecision by the brain
using analogies and metaphors through fuzzy inference-making.
In their prominent paper on the fuzzy approach to analogical reasoning, BBM and
Valverde [12] address the problem of the representation of resemblances involved
in analogical reasoning and use fuzzy relations to compare situations. As fuzziness
entails the diverse forms of reasoning, from true literal sentences to false unliteral
sentences, it is a powerful computation mode for human thinking and reasoning
that is mainly metaphor and analogy-based reasoning. In addition, analogical and
metaphorical sentences often include modifiers that BBM and Marsala put as the
core of interpretable Fuzzy Systems [16]. For instance, when describing electricity,
a common analogy is a water tank, where charge stands for the water amount, voltage
72
C. Tijus
for the water pressure, and current for the water flow. They are said to be equivalent,
but electricity is not used to explain water flow. There are even situations where the
water analogy is rather misleading. Electricity is like water but they cannot be mixed.
Water is largely used to produce electricity, not the reverse. One might think that a
metaphor such as “this is truly a gem” means a true literal sentence although this is
metaphorical “image-based language” that strengthens the metaphor.
According to BBM [12], we use fuzzy relations to compare situations that can be
used to model a natural analogy: resemblance relations can be used to define a kind
of analogical scheme compatible with approximate reasoning in fuzzy logic, with
measures of satisfiability, resemblance and inclusion. These fuzzy relations can be
regarded as measures of a categorization process devoted to analogy and metaphor
with the purpose of transmitting knowledge from the source to the target.
References
1. Zibetti, E., Tijus, C.: Understanding actions: contextual dimensions and heuristics. In International and Interdisciplinary Conference on Modeling and Using Context, pp. 542–555. Springer,
Berlin, Heidelberg (2005)
2. Hard, B.M., Meyer, M., Baldwin, D.: Attention reorganizes as structure is detected in dynamic
action. Memory & Cognition 47(1), 17–32 (2018)
3. Picard, J.: Les trois modes du raisonnement analogique. Revue Philosophique 104, 242–282
(1927)
4. Goblot, E.: Traité de logique. A. Colin (1920)
5. Glucksberg, S., McGlone, M.S., Manfredi, D.: Property attribution in metaphor comprehension.
J. Mem. Lang. 36(1), 50–67 (1997)
6. Lakoff, G., Johnson, M.: Metaphors we live by. University of Chicago press (2008)
7. Bouchon-Meunier, B., Ramdani, M., Valverde, L.: Fuzzy logic, inductive and analogical
reasoning. In: International Workshop on Fuzzy Logic in Artificial Intelligence (pp. 38–50)
Springer, Berlin, Heidelberg (1993)
8. Bouchon-Meunier, B., Valverde, L.: Analogy relations and inference. In: Second IEEE
International Conference on Fuzzy Systems, pp. 1140–1144 (1993)
9. Bouchon-Meunier, B., Valverde, L.: A resemblance approach to analogical reasoning functions.
In: International Workshop on Fuzzy Logic in Artificial Intelligence, pp. 266–272 Springer,
Berlin, Heidelberg (1995)
10. Bouchon-Meunier, B., Rifqi, M., Bothorel, S.: Towards general measures of comparison of
objects. Fuzzy Sets Syst. 84(2), 143–153 (1996)
11. Bouchon-Meunier, B., Delechamp, J., Marsala, C., Rifqi, M.: Several forms of fuzzy analogical
reasoning. In: Proceedings of the Sixth IEEE International Conference on Fuzzy Systems, vol.
1, pp. 45–50 (1997)
12. Bouchon-Meunier, B., Valverde, L.: A fuzzy approach to analogical reasoning. Soft. Comput.
3(3), 141–147 (1999)
13. Bouchon-Meunier, B., Delechamp, J., Marsala, C., Rifqi, M.: Analogy as a basis of various
forms of approximate reasoning. In: Uncertainty in Intelligent and Information Systems, pp. 70–
79 (2000)
14. Bouchon-Meunier, B.: Une approche floue du raisonnement par analogie. In Tijus, C. (ed.)
Métaphores et Analogies. Collection Traité de Sciences Cognitives, Hermes (2003)
15. Bouchon-Meunier, B., Mesiar, R., Marsala, C., Rifqi, M.: Compositional rule of inference as
an analogical scheme. Fuzzy Sets and Syst. 138(1), 53–65 (2003)
Analogy
73
16. Bouchon-Meunier, B., Marsala, C.: Fuzzy modifiers at the core of interpretable fuzzy systems.
In: Fifty Years of Fuzzy Logic and its Applications, pp. 51–63. Springer, Cham (2015)
17. Grice, H.P.: Logic and Conversation, 41–58 (1975)
18. Glucksberg, S.: The psycholinguistics of metaphor. Trends in Cogn Sci 7(2), 92–96 (2003)
19. Gombrich, E.H.: Mediations on a hobby horse. In: Meditations on a Hobby Horse and Other
Essays on the Theory of Art. L.L. Whyte, London (1963)
20. Tversky, A.: Features of similarity. Psychol. Rev. 84, 327–352 (1977)
21. Gentner, D., Markman, A.B.: Structure mapping in analogy and similarity. Am. Psychol. 52(1),
45–56 (1997)
22. Berthoz, A.: Simplexity: Simplifying Principles for a Complex World. Yale University Press,
USA (2012)
23. Raccah, P.Y.: Linguistic argumentation as a shortcut for the empirical study of argumentative strategies. In: Reflections on Theoretical Issues in Argumentation Theory, pp. 279–293.
Springer, Cham (2015)
24. Bassok, M., Holyoak, K.J.: Interdomain transfer between isomorphic topics in algebra and
physics. J. Exp. Psychol. Learn. Mem. Cogn. 15(1), 153–166 (1999)
25. Kotovsky, K., Hayes, J.R., Simon, H.A.: Why are some problems hard? Evidence from Tower
of Hanoi. Cogn. Psychol. 17(2), 248–294 (1985)
26. Megalakaki, O., Tijus, C., Baiche, R., Poitrenaud, S.: The effect of semantics on problem
solving is to reduce relational complexity. Think. Reason. 18(2), 159–182 (2012)
27. Zhang, J.: The nature of external representations in problem solving. Cogn. Sci. 21(2), 179–217
(1997)
28. Nguyen-Xuan, A., Tijus, C.: Rules discovery: transfer and generalization. In: IEEE International Conference on Research, Innovation and Vision for the Future, RIVF, pp. 9–16
(2008)
29. Kokinov, B., French, R.M.: Computational models of analogy making. Encycloped. Cogn. Sci.
1, 113–118 (2003)
30. Forbus, K.D., Ferguson, R.W., Lovett, A., Gentner, D.: Extending SME to handle large-scale
cognitive modeling. Cogn. Sci. 41(5), 1152–1201 (2017)
31. Lovett, A., Forbus, K.: Modeling visual problem solving as analogical reasoning. Psychol. Rev.
124(1), 60–90 (2017)
32. Tijus, C., Poitrenaud, S., Chene, D.: Similarity and categorization: taxonomic and meronomic
parts of similes, In: Proceedings of the 6th European Congress on System Sciences, vol. 38
(2005)
33. Anderson, J.R.: The adaptive nature of human categorization. Psychol. Rev. 98(3), 409–429
(1991)
34. Pudelko, B., Hamilton, E., Legros, D., Tijus, C.: How context contributes to metaphor understanding. In: International and Interdisciplinary Conference on Modeling and Using Context,
pp. 511–514. Springer, Berlin, Heidelberg (1999)