A Dual Process Architecture for Ontology-based Systems

Antonio Lieto

1,3

, Andrea Minieri

, Alberto Piana

, Daniele P. Radicioni

and Marcello Frixione

Dipartimento di Informatica, Universit`a di Torino, Torino, Italy

DAFIST, Universit`a di Genova, Genova, Italy

ICAR-CNR, Palermo, Italy

Keywords:

Knowledge Representation, Formal Ontologies, Conceptual Spaces, Common Sense Reasoning, Dual Process

Theory, Prototypical Reasoning.

Abstract:

In this work we present an ontology-based system equipped with a hybrid architecture for the representation of

conceptual information. The proposed system aims at extending the representational and reasoning capabilities

of classical ontology-based systems towards more realistic and cognitively grounded scenarios, such as those

envisioned by the prototype theory. The resulting system attempts to reconcile the heterogeneous approach to

the concepts in Cognitive Science and the dual process theories of reasoning and rationality. The system has

been experimentally assessed in a conceptual categorization task where common sense linguistic descriptions

were given in input, and the corresponding target concepts had to be identiﬁed. The results show that the

proposed solution substantially improves on the representational and reasoning “conceptual” capabilities of

standard ontology-based systems.

1 INTRODUCTION

One of the main open problems in the ﬁeld of on-

tology engineering is that formal ontologies do not

allow –for technical convenience– neither the repre-

sentation of concepts in prototypical terms nor forms

of approximate, non monotonic, conceptual reason-

ing. Conversely, in Cognitive Science evidences exist

in favor of prototypical concepts, and typicality-based

conceptual reasoning has been widely investigated in

the ﬁelds of the studies regarding the human cogni-

tion. In this ﬁeld the early work of Rosch (Rosch,

1975) showed that ordinary concepts do not obey the

classical theory (stating that concepts can be deﬁned

in terms of sets of necessary and sufﬁcient condi-

tions). Rather, they exhibit prototypical traits: e.g.,

some members of a category are considered better in-

stances than other ones; more central instances share

certain typical features –such as the ability of ﬂy-

ing for birds– that, in general, cannot be thought of

as necessary nor sufﬁcient conditions. These results

inﬂuenced pioneering KR research, where some ef-

forts were invested in trying to take into account the

suggestions coming from Cognitive Psychology: ar-

tiﬁcial systems were designed –e.g., frames (Min-

sky, 1975) and semantic networks– to represent and

to conduct reasoning on concepts in “non classical”,

prototypical terms.

However, these systems were later sacriﬁced in fa-

vor of a class of formalisms stemmed from structured

inheritance semantic networks and based in a more

rigorous semantics: the ﬁrst system in this line of

research was the KL-ONE system (Brachmann and

Schmolze, 1985). These formalisms are known to-

day as description logics (DLs) (Nardi and Brachman,

2003). In this setting, the representation of prototyp-

ical information (and therefore the possibility of per-

forming non monotonic reasoning) is not allowed,

since the formalisms in this class are primarily in-

tended for deductive, logical inference. However, un-

der a historical perspective, the choice of preferring

classical systems based on a well deﬁned –Tarskian-

like– semantics left unsolved the problem of repre-

senting concepts in prototypical terms. Although in

the ﬁeld of logic oriented KR various fuzzy and non-

monotonic extensions of DL formalisms have been

designed to deal with some aspects of “non-classical”

concepts (Straccia, 2011; Giordano et al., 2013),

nonetheless various theoretical and practical prob-

lems remain unsolved (Frixione and Lieto, 2010).

In this paper a conceptual architecture is presented

This is the case, for example, of exceptions to the in-

heritance mechanism.

Lieto A., Minieri A., Piana A., P. Radicioni D. and Frixione M..

A Dual Process Architecture for Ontology-based Systems.

DOI: 10.5220/0005070800480055

In Proceedings of the International Conference on Knowledge Engineering and Ontology Development (KEOD-2014), pages 48-55

ISBN: 978-989-758-049-9

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

that, embedded in a larger knowledge-based system,

aims at extending the representational and reasoning

capabilities available to traditional ontology-based

frameworks.

The paper is structured as follows: in Section 2 we

illustrate the general architecture and the main fea-

tures of the knowledge-based system. In Section 3

we provide the results of a twofold experimentation

to assess the accuracy of the system in a categoriza-

tion task. Finally, we conclude by presenting the re-

lated work (Section 4) and outlining future work (Sec-

tion 5).

2 THE SYSTEM

Two cornerstones inspiring the current proposal are

the dual process theory and the heterogeneous ap-

proach to concepts in Cognitive Science. The theoret-

ical framework known as dual process theory postu-

lates the co-existence of two different types of cogni-

tive systems (Evans and Frankish, 2009; Kahneman,

2011). The systems of the ﬁrst type (type 1) are phy-

logenetically older, unconscious, automatic, associa-

tive, parallel and fast. The systems of the second type

(type 2) are more recent, conscious, sequential and

slow, and featured by explicit rule following.

We assume that each system type can be com-

posed by many sub-systems and processes; according

to the hypotheses in (Frixione and Lieto, 2012; Frix-

ione and Lieto, 2014), the conceptual representation

of our system includes two main sorts of components,

based on two sorts of processes. Type 1 processes are

used to perform fast and approximate categorization,

and beneﬁt from prototypical information associated

to concepts. Type 2 processes, used in classical in-

ference tasks, and not considering the contribution

of prototypical knowledge. The two sorts of system

processes are assumed to interact, since type 1 pro-

cesses are executed ﬁrst and their results are then re-

ﬁned by type 2 processes. We also based our work on

the heterogeneous approach to the concepts in Cogni-

tive Science (Machery, 2009). According to this per-

spective, concepts do not constitute a unitary element;

rather, artiﬁcial concepts can be composed by several

bodies of knowledge, each one conveying a speciﬁc

type of information.

The whole system embedding the proposed con-

ceptual architecture works as follows. The input to

the system is a simple description, like ‘The animal

that eats bananas’: and the expected output is the cat-

egory

monkey

. We devised an Information Extraction

(IE) step in order to map the linguistic input to an in-

ternal representation. In turn, the internal represen-

tation is fed into the proposed architecture, which is

concerned with the categorization task. Due to space

restrictions, we are presently concerned with the ar-

chitecture, deferring the details about the IE task to

future contributions.

2.1 Knowledge Base Architecture

We designed a hybrid conceptual architecture that

builds on a classical ontological component, and on

a typical one. Each component represents a speciﬁc

conceptual body of knowledge together with the re-

lated reasoning procedures as in the dual process per-

spective. A classical representation grounded on a DL

formalism is the base of the ontological component.

It permits listing necessary and/or sufﬁcient condi-

tions to deﬁne concepts. For example, if we consider

the concept

water

, the classical representation con-

tains the information that

water

is a natural substance,

whose chemical formula is H

O. On the other hand,

the prototypical traits include information about the

fact that water usually occurs in liquid state, and it is

mostly a tasteless, odorless and colorless ﬂuid.

According to the “dual process” approach, in

the implemented system the representational and

reasoning functions are assigned to the system 1

(hereafter S1), that executes processes of type 1,

and are associated to the Conceptual Spaces frame-

work (G¨ardenfors, 2000). On the other hand, the rea-

soning and representational functions are assigned to

the system 2 (hereafter S 2), to execute processes of

type 2, and are associated to a classical DL-based on-

tological representation.

2.1.1 The S 1 and S 2 Components

Conceptual spaces (CS) are a geometrical representa-

tional framework where knowledge is represented in

terms of as a set of quality dimensions (G¨ardenfors,

2000). A geometrical structure is associated to each

quality dimension. In this framework instances are

represented as points in a multidimensional space,

and their similarity can be computed as the interven-

ing distance, based on some suitable distance met-

rics. In this setting, concepts correspond to regions

and regions with different geometrical properties cor-

respond to different sorts of concepts.

Conceptual spaces are suitable to represent con-

cepts in “typical” terms, since the regions represent-

ing concepts can have soft boundaries. Prototypes

have a natural geometrical interpretation, in that they

correspond to the geometrical centre of the convex re-

gion; conversely, given a convex region we can pro-

vide each point with a certain centrality degree, that

can be interpreted as a measure of its typicality.

ADualProcessArchitectureforOntology-basedSystems

The conceptual space deﬁnes a metric space that

can be used to compute the proximity between any

two entities, and between entities and prototypes. To

compute the distance between two points p

, p

apply a distance metrics based on the combination of

the Euclidean distance and the angular distance inter-

vening between the points. Namely, we use Euclidean

metrics to compute within-domain distance, while for

dimensions from different domains we use the Man-

hattan distance metrics, as suggested in (G¨ardenfors,

2000; Adams and Raubal, 2009). Weights assigned to

domain dimensions are affected by the context, too,

so the resulting weighted Euclidean distance dist

computed as follows

dist

, p

,k) =

∑

i=1

1,i

− p

2,i

)

where i varies over the n domain dimensions, k is the

context that indicates the set of weights associated to

each domain, and w

are dimension weights.

We represent points as vectors (with as many

dimensions as required by the considered domain),

whose components correspond to the point coordi-

nates, so that a natural metrics to compute the simi-

larity between them is cosine similarity. In the metric

space being deﬁned, the distance between an individ-

ual and prototypes is computed with the Manhattan

distance, enriched with information about context k.

Also, the distance between any two concepts can be

computed as the distance between two regions in a

given domain Also, we can compute the distance be-

tween any two region prototypes, or the minimal dis-

tance between their individuals, or we can apply more

sophisticated algorithms. Further details about tech-

nical issues can be found in (Ghignone et al., 2013).

Inference in conceptual spaces is mostly per-

formed on incomplete and/or noisy information: that

is, it is frequent the case that only partial information

is available to categorize a given input individual, and

some points are not deﬁned for one or more dimen-

sions. Conceptual spaces are robust to this sort of lack

of information, which is conversely a decisive factor

in the context of formal ontologies. In these cases we

restrict to considering domains that contain points in

the input individual: if the description for a given in-

dividual does not contain points of some domains, the

distance for those domains is set to a default value.

The basic representational structure processed by

the system is named genericDescription; it encodes

the salient aspects of the entities being considered. A

genericDescription is a super-domain that hosts infor-

mation about physical and non physical features ar-

ranged into nine domains: size, shape, color, loca-

tion, feeding, locomotion, hasPart, partOf, manRela-

tionship. The size of entities is expressed through

the three Euclidean dimensions; the shape allows ex-

pressing that an object has circular, square, spherical,

cubic, etc., shape. The color space maps object’s fea-

tures onto the

⋆

color space.

⋆

(0 ≤ L ≤ 100)

is the correlate of lightness,

⋆

(−128 ≤ a ≤ 127) is

the chromaticity axis ranging from green to red, and

⋆

(−128 ≤ b ≤ 127) is the chromaticity axis ranging

from blue to yellow.

The location space indicates the place where the

object being modeled can be typically found. It actu-

ally results from the combination of ﬁve dimensions,

and namely: humidity, indicated as a percentage; tem-

perature, ranging in [−40

◦

,50

◦

]; altitude, ranging in

[−11000,8848]; vegetation, ranging in [0,100]; time.

In turn, time contains a partitioning of the hours of the

day into sunrise (4–6 AM), morning (6–12 AM), af-

ternoon (12–5 PM), evening (5–10 PM) and night (10

PM–4 AM).

The domain feeding is currently speciﬁc to ani-

mals, and it allows mapping an element over two di-

mensions, typeOfFood and amountOfFood. The type-

OfFood is associated to an integer indicating 1: her-

bivore, 2: lectivore, 3: detritivore, 4: necrophage,

5: carnivore. The underlying rationale is that close

elements (e.g., necrophage and carnivore, that are

one step apart in the proposed scale) are represented

as close in this space due to their proximity under

an ethological viewpoint, whilst different categories

(e.g., herbivore and carnivore) are featured by larger

distances in the considered scale (Getz, 2011).

Similar to the previous one, also the locomotion

domain combines two dimensions: the former dimen-

sion is used to account for the type of movement (1:

swim, 2: dig, 3: crawl, 4: walk, 5: run, 6: roll, 7:

jump, 8: ﬂy), and the latter one is used to account

for the speed, expressed in km/h (Bejan and Marden,

2006). Finally, the manRelationship space is used to

grasp entities as related to man by function (both a

train and a horse can be used as ‘transport’), product

(chicken produce ‘eggs’, and ‘chicken’ per se are a

food product), symbol (‘lion’ can be used as a symbol

for ‘strength’ and ‘royalty’). A simpliﬁed example of

the

lion

prototype information is reported below.

</feeding>

</genericPhysicalDescription>

<symbol id="08153437">royalty</symbol>

<symbol id="05036872">strength</symbol>

</manRelationship>

</object>

KEOD2014-InternationalConferenceonKnowledgeEngineeringandOntologyDevelopment

On the other hand, the representation of the clas-

sical component S2 is implemented through a formal

ontology. As already pointed out, the standard on-

tological formalisms leave unsolved the problem of

representing prototypical information. Furthermore,

it is not possible to execute non monotonic inference,

since classical ontology-based reasoning mechanisms

contemplate deductive processes.

In this setting we cannot represent even simple

prototypical information, such as ‘A typical rose is

red’. This is due to the fact that being red is neither a

necessary nor a sufﬁcient condition for being a rose,

and therefore it is not possible neither to represent and

to automatically identify a prototypical rose (let us as-

sume #roseP) nor to describe (and to learn from new

cases) the typical features of the class of prototypical

roses. Such aspect has, on the other hand, a natural

interpretation by using the conceptual spaces frame-

work.

2.2 Inference in the Hybrid System

Categorization (i.e., to classify a given data instance

into a predeﬁned set of categories) is one of the clas-

sical processes automatically performed both by sym-

bolic and sub-symbolic artiﬁcial systems. In our sys-

tem categorization is based on a two-step process in-

volving both the typical and the classical component

of the conceptual representation. These components

account for different types of categorization: approxi-

mate or non monotonic (performed on the conceptual

spaces), and classical or monotonic (performed on the

ontology). Different from classical ontological infer-

ence, in fact, categorization in conceptual spaces pro-

ceeds from prototypical values. In turn, prototypical

values need not be speciﬁed for all class individuals,

that vice versa can overwrite them: one typical exam-

ple is the case of birds that (by default) ﬂy, except for

special birds, like penguins, that do not ﬂy.

The whole categorization process can be summa-

rized as follows. The system takes in input a textual

description d and produces in output a pair of cate-

gories hc

,cci, the output of S 1 and S 2, respectively

(see Algorithm 1). If the S 2 system classiﬁes it as

consistent with the ontology, then the classiﬁcation

succeeded and the category provided by S 2 (cc) is re-

turned along with c

, the top scoring class returned by

S 1 (Algorithm 1: line 8). If cc –the class computed by

S 2– is a subclass of one of those identiﬁed by S 1 (c

both cc and c

are returned (Algorithm 1: line 11).

Thus, if S 2 provides more speciﬁc output, we follow

a speciﬁcity heuristics; otherwise, the output of S 2 is

returned, following the rationale that it is safer.

The output of S 2 cannot be wrong on a purely logical

Algorithm 1: Inference in the hybrid system.

input : textual description d

output : a class assignment, as computed by S 1 and

S 2

1: C ← S 1(d)

/* conceptual spaces output

2: for each c

∈ C do

3: cc ← S 2(hd,c

/* ontology based

output */

4: if cc == NULL then

5: continue

/* inconsistency detected

6: end if

7: if cc equals c

then

8: return hc

,cci

9: else

10: if cc is subclass of c

then

11: return hc

,cci

12: end if

13: end if

14: end for

15: cc ← S 2(hd, Thingi)

16: return hc

,cci

pair of results is always returned, including both the

output of S 1 and the output of S 2, thereby providing

typically valid answers (through S1) that are checked

against a logically valid reasoning conducted on the

ontological knowledge base (through S 2). In so do-

ing, we follow the rationale that despite the S1 output

can contain errors, it furnishes approximate answers

that cannot be obtained by resorting only to classical

ontological inference.

If all results in C are inconsistent with those com-

puted by S2, a pair of classes is returned including

and the output of S 2 having for actual parameters

d and Thing, the meta-class of all the classes in the

ontological formalism.

An important function provided by S 2 regards

the explanation of the detected inconsistencies. This

function is obtained by recurring to standard DL rea-

soners.

One main problem encountered in the expla-

nation of inconsistencies regards the fact that reason-

ers’ output is usually quite verbose, since it provides

the whole chain of all the possible reasons explain-

ing why a given model is not consistent w.r.t. the

perspective, in that it is the result of a deductive process.

The control strategy implements a tradeoff between onto-

logical inference and the output of S 1, which is more in-

formative but also less reliable from a formal point of view.

However, in next future we plan to explore different concil-

iation mechanisms to ground the overall control strategy.

To actually access the KBs we used the Jena frame-

work, https://jena.apache.org.

ADualProcessArchitectureforOntology-basedSystems

represented assertions. For example, let us suppose

that the ontological KB is provided with an assertion

about the fact that

whale isA ﬁsh

. Whales are not

ﬁshes: they are in the order of

cetacea

. If an incon-

sistency is detected, the initial results obtained by the

reasoner report the followingclauses: since

whale isA

mammal

, and since

mammal

is disjoint with other an-

imal classes, it must also be disjoint with the classes

ﬁsh

reptile

bird

and so on, as well as with all respec-

tive subclasses of

ﬁsh

reptile

bird

. Although fac-

tually correct and complete, this explanation is quite

long (in an ontological KB with good coverage each

class contains many subclasses) and it thus provides

information not very informative for the punctual ex-

planation of the raised inconsistency. The only rele-

vant information, in this case, regards the fact that the

tested class

whale

cannot be classiﬁed as a

ﬁsh

be-

cause

mammal

and

ﬁsh

are mutually disjoint classes.

The disjointness of the superclass

mammal

with the

other classes in the ontology is not contextually rele-

vant in this phase. Therefore we designed and imple-

mented a software layer that runs on top of Jena expla-

nation utilities to extract a laconic explanation from

the longer one: main focus of the laconic explanation

is to make apparent the cause of the inconsistency. In

so doing, we adopted a simple heuristic according to

which the only motivation reported is that focused on

the tuples of classes which are under investigation.

3 EXPERIMENTATION

We have designed a twofold experimentation on a cat-

egorization task. In the ﬁrst experiment we tested the

system over a restricted domain (the animal kingdom

domain) to compare it with the results obtained in a

previous experimentation with S 2 using a handcrafted

ontology. In the second experiment we tested the sys-

tem in a broader context, and we were interested in as-

sessing its robustness and the discriminative features

of the S 1 component in a multi-domain setting. Two

datasets, one composed of 39 “common-sense” lin-

guistic descriptions, and one composed of 40 linguis-

tic descriptions were used.

Each stimulus st = hd,Ti

is a pair of description and target, such as h‘The big

carnivore with yellow and black stripes’, ‘tiger’i. The

target T is the “prototypically correct” category, and

in the following it is referred to as the expected re-

sult. The set of stimuli was devised by a team of

neuropsychologists and philosophers in the frame of

a broader project aimed at investigating the role of vi-

The full list of the stimuli is available at the URL:

http://www.di.unito.it/∼radicion/datasets/keod 2014/

stimuli.txt.

sual load in concepts involved in inferential and ref-

erential tasks. The expected prototypical target cate-

gory represents a gold standard, since it corresponds

to the results provided within a psychological exper-

imentation. In this experimentation 30 subjects were

requested to provide the corresponding target concept

for each description. Such input was then used for

querying our system as in a typicality based question-

answering task. In Information Retrieval such queries

(which is by far the most common and complex to

interpret) are known to belong to the class of “infor-

mational queries”, i.e., queries where the user intends

to obtain information regarding a speciﬁc information

need (Jansen et al., 2008).

Experiment 1

The ﬁrst experiment was made to compare the re-

sults obtained in the previous study (Ghignone et al.,

2013) –where S 2 made use of a handcrafted ontolog-

ical knowledge base– to those obtained with S2 using

publicly available common sense domain ontologies.

With respect to the previous experiment a larger num-

ber of new descriptions were considered, all in the

animal domain. Two ontologies have been selected

for the comparison: the Animal in Context Ontology

(ACO), developed by the Veterinary Medical Infor-

matics Laboratory at the Virginia-Maryland Regional

College and the BBC WildLife Ontology.

They were

both retrieved by using a mixed search over Sindice

and Swoogle, and they were selected as guarantee-

ing a granularity of information similar to that of the

handcrafted ontology of our previous study.

The results concerning both experimental setups

(ACO and BBC ontologies within S 2 KB) are re-

ported in Table 1.

By considering the ACO ontology, in most cases

(37/39) S 1 and S 2 return the same category. Differ-

ently from the previous experiment, however, this is

given by the fact that such ontology, in certain cases,

is not precise enough to provide a different correct/-

better speciﬁed category w.r.t. S 1. In particular, in the

case of the description “The big ﬁsh eating plankton”,

the result

whale

obtained by S 1 is not detected as

inconsistent in such ontology, since the concept rep-

resentation along with its restrictions are not present

in the ontology itself. Therefore, checking the con-

sistency of the class (which was added to the ontol-

ogy without the conceptual constraints associated to

the concept) does not reveal inconsistencies with the

original model of the ontology. The only differences

w.r.t. the S 1 and S 2 output is obtained by consider-

Available at the URLs: http://vtsl.vetmed.vt.edu/aco/

Ontology/aco.zip and http://www.bbc.co.uk/ontologies/wo.

KEOD2014-InternationalConferenceonKnowledgeEngineeringandOntologyDevelopment

Table 1: The results of the ﬁrst experiment, with S 2 using external ontologies.

ACO BBC

Cases where S 1 and S 2 returned the same category 94.8% (37/39) 92.3% (36/39)

Cases where S 1 returned the expected category 94.8% (37/39) 94.8% (37/39)

Cases where S 2 returned the expected category 94.8% (37/39) 92.3% (36/39)

ing the two descriptions: “A honey producing insect.

(Bee)” and “A bird kept on a farm for its meat and

eggs. (Chicken)”. In both cases, more speciﬁc and

detailed results are provided by S 2 w.r.t. the already

correct categorization of S 1.

Analogous results, in terms of accuracy, havebeen

obtained by plugging the BBC Wildlife ontology into

S 2. The testing of such ontology has been possi-

ble thanks to a manual import of the missing class-

es/instances through the browsing of the BBC direc-

tory containing the full classiﬁcation considered by

the BBC archives.

Despite the two experiments pro-

duced similar ﬁgures, they also revealed some differ-

ences between the two ontologies. The answer

whale

for the above mentioned description is correctly de-

tected as inconsistent with the BBC ontology, and the

ontologically correct result

whale-shark

is returned.

Summing up: the results of this experiment show

that the conceptual categorization obtained by S 2 us-

ing external ontological knowledge bases are directly

comparable to those obtained in a previous study

where a handcrafted domain ontology was consid-

ered. A major problem encountered in using exter-

nal knowledge bases stems from axiomatization, that

is not always sufﬁcient to solve the task of detecting

and correcting errors produced by S 1. In particular,

in the case of ACO, a ﬁner-grained result is obtained

in two cases after a correct categorization already pro-

vided by S 1. This ﬁner-grained result is not obtained

by considering the BBC Wildlife ontology that, how-

ever, is able to detect the expected inconsistencies.

Experiment 2

In order to assess the accuracy of the system in a

more demanding experimental setting, we devised a

second experimental task, where S 2 used the knowl-

edge base OpenCyc.

OpenCyc is one of the largest

ontologies publicly available, in that it is an enor-

mous attempt at integrating many diverse semantic

resources (such as, e.g., WordNet, DBpedia, Wiki-

company, etc.). Its coverage and depth was therefore

its most attractive feature (it contains about 230, 000

concepts, 2,090,000 triples and 22,000 predicates).

http://www.bbc.co.uk/nature/wildlife.

http://www.cyc.com/platform/opencyc.

Table 2: The results of the second experiment.

Test cases categorized 40 100.0%

S 1-S 2 31 77.5%

Google 26 65.0%

Bing 23 57.5%

Additionally, we adopted OpenCyc to use a knowl-

edge base independent of our own representational

commitments in order to more effectively assess the

ﬂexibility of the proposed system when using general-

purpose, well known, existing resources.

A second dataset of 40 “common-sense” linguis-

tic descriptions was collected with the same rationale

considered for the ﬁrst experiment.

The obtained results are reported in Table 2.

Despite being less accurate with respect to the pre-

vious experiment, the hybrid knowledge based S1-S 2

system was able to categorize and retrieve most of

the new typicality-based stimuli provided as input

and still showed a better performance w.r.t. the gen-

eral purpose search engines Google and Bing used in

question-answering mode.

The major problems encountered in this experi-

ment derived from the difﬁculty of mapping the lin-

guistic structure of stimuli containing very abstract

meaning in the representational framework of S 1 as

they are actually encoded according to the conceptual

space. For example, it was impossible to map the

information contained in the description “the place

where kings, princes and princesses live in fairy tales”

onto the features used to characterize the prototypi-

cal representation of the concept

Castle

. Similarly,

the information extracted from the description “Giv-

ing something away for free to someone” could not

be mapped onto the features associated to the con-

cept

Gift

. On the other hand, the system shows good

performances when dealing with less abstract descrip-

tions based on perceptual features such as shape,

color, size, and with some typical information such

as function, symbol, product.

In this experiment, differently from the previous

one (e.g., that one in the case of

whale

for the BBC

ontology), S 2 did not reveal any inconsistency, in

that S 1 mostly provided an output coherent with the

An extended version of the present experiment is illus-

trated in (Lieto et al., 2014).

ADualProcessArchitectureforOntology-basedSystems

model in OpenCyc. However, by analyzing in detail

the different answers, we notice that at least one in-

consistency should have been detected by S 2. This is

the case of the description “An intelligent grey ﬁsh”

associated to the target concept

Dolphin

. In this case,

the S 1 system returned the expected target, but S2

did not raise the inconsistency since OpenCyc er-

roneously represents

Dolphin

as a subclass of

Fish

rather than a subclass of

Mammal

4 RELATED WORK

In the context of a different ﬁeld of application, a so-

lution similar to the one adopted here has been pro-

posed in (Chella et al., 1997). The main difference

with their proposal concerns the underlying assump-

tion on which the integration between symbolic and

sub-symbolic system is based. In our system the con-

ceptual spaces and the classical component are inte-

grated at the level of the representation of concepts,

and such components are assumed to convey differ-

ent –though complementary– conceptual information.

On the other hand, the previous proposal is mainly

used to interpret and ground raw data coming from

sensors in a high level symbolic system through the

mediation of conceptual spaces.

In other respects, our system is also akin to that

ones developed in the ﬁeld of the computational ap-

proach to the above mentioned dual process theo-

ries. A ﬁrst example of such “dual based systems”

is the mReasoner model (Khemlani and Johnson-

Laird, 2013), developed with the aim of providing a

computational architecture of reasoning based on the

mental models theory proposed by Philip Johnson-

Laird (Johnson-Laird, 1980). The mReasoner archi-

tecture is based on three components: a system 0, a

system 1 and a system 2. The last two systems corre-

spond to those hypothesized by the dual process ap-

proach. System 0 operates at the level of linguistic

pre-processing. System 1 uses this intensional rep-

resentation to build an extensional model, and uses

heuristics to provide rapid reasoning conclusions; ﬁ-

nally, system 2 carries out more demanding processes

to search for alternative models, if the initial conclu-

sion does not hold or if it is not satisfactory.

A second system that is close to our present work

has been proposed by (Pilato et al., 2012). The au-

thors do not explicitly mention the dual process ap-

proach; however, they build a system for conversa-

tional agents (chatbots) where agents’ background

knowledge is represented using both a symbolic and a

sub-symbolic approach. They also associate different

sorts of representation to different types of reasoning.

Namely, deterministic reasoning is associated to sym-

bolic (system 2) representations, and associative rea-

soning is accounted for by the sub-symbolic (system

1) component. Differently from our system, however,

the authors do not make any claim about the sequence

of activation and the conciliation strategy of the two

representational and reasoning processes. It is worth

noting that other examples of this type of systems can

be considered that are in some sense analogous to

the dual process proposal: for example, many hybrid,

symbolic-connectionist systems –including cognitive

architectures such as, for example, CLARION

–, in

which the connectionist component is used to model

fast, associative processes, while the symbolic com-

ponent is responsible for explicit, declarative compu-

tations. However, at the best of our knowledge, our

system is the only one that considers this hybridiza-

tion with a granularity at the level of individual con-

ceptual representations.

5 CONCLUSIONS

In this work we have presented a knowledge-based

system relying upon a cognitively inspired archi-

tecture for the representation of conceptual knowl-

edge. The system is grounded on a hybrid frame-

work coupling classical and prototypical representa-

tion and reasoning, and it aims at extending the rep-

resentational and reasoning capabilities of classical

ontological-based systems towards more realistic and

cognitively grounded scenarios, such as those envi-

sioned by the prototype theory. The results obtained

in a twofold experimentation show that in the re-

stricted domain of the animal kingdom the proposed

architecture is comparable with the results obtained

with a handcrafted domain ontology. The overall re-

sults corroborate the hypothesis that matching in con-

ceptual spaces can be fruitfully coupled to ontological

inference. However, the ontological inference would

not sufﬁce to categorize the presented common sense

stimuli presented in both experiments, with only one

exception.

In next future we plan to test the proposed ap-

proach in the area of biomedical domain to as-

sess disease diagnosis tasks by grounding S 2 on

SNOWMED,

and S 1 on conceptual spaces repre-

senting the typical symptoms of a given disease.

http://www.cogsci.rpi.edu/∼rsun/clarion.html

http://www.b2international.com/portal/snow-owl.

KEOD2014-InternationalConferenceonKnowledgeEngineeringandOntologyDevelopment

REFERENCES

Adams, B. and Raubal, M. (2009). A metric conceptual

space algebra. In Hornsby, K. et al., editors, COSIT,

volume 5756 of LNCS, pages 51–68. Springer.

Bejan, A. and Marden, J. H. (2006). Constructing animal

locomotion from new thermodynamics theory. Amer-

ican Scientist, 94(4):342.

Brachmann, R. J. and Schmolze, J. G. (1985). An overview

of the KL-ONE knowledge representation system.

Cognitive Science, 9(2):171–202.

Chella, A., Frixione, M., and Gaglio, S. (1997). A cognitive

architecture for artiﬁcial vision. Artiﬁcial Intelligence,

89(1–2):73 – 111.

Evans, J. S. B. and Frankish, K. E. (2009). In two minds:

Dual processes and beyond. Oxford University Press.

Frixione, M. and Lieto, A. (2010). The computational

representation of concepts in formal ontologies-some

general considerations. In KEOD.

Frixione, M. and Lieto, A. (2012). Representing concepts in

formal ontologies: Compositionality vs. typicality ef-

fects. Logic and Logical Philosophy, 21(4):391–414.

Frixione, M. and Lieto, A. (2014). Towards an Extended

Model of Conceptual Representations in Formal On-

tologies: A Typicality-Based Proposal. Journal of

Universal Computer Science, 20(3):257–276.

G¨ardenfors, P. (2000). Conceptual spaces: The geometry of

thought. MIT press.

Getz, W. M. (2011). Biomass transformation webs provide

a uniﬁed approach to consumer–resource modelling.

Ecology letters, 14(2):113–124.

Ghignone, L., Lieto, A., and Radicioni, D. P. (2013).

Typicality-Based Inference by Plugging Conceptual

Spaces Into Ontologies. In Lieto, A. and Cruciani, M.,

editors, Proceedings of the International Workshop on

Artiﬁcial Intelligence and Cognition. CEUR.

Giordano, L., Gliozzi, V., Olivetti, N., and Pozzato, G. L.

(2013). A non-monotonic description logic for reason-

ing about typicality. Artiﬁcial Intelligence, 195:165–

202.

Jansen, B. J., Booth, D. L., and Spink, A. (2008). Determin-

ing the informational, navigational, and transactional

intent of web queries. Information Processing & Man-

agement, 44(3):1251–1266.

Johnson-Laird, P. (1980). Mental models in cognitive sci-

ence. Cognitive Science, 4(1):71–115.

Kahneman, D. (2011). Thinking, fast and slow. Macmillan.

Khemlani, S. and Johnson-Laird, P. (2013). The processes

of inference. Argument & Computation, 4(1):4–20.

Lieto, A., Minieri, A., Piana, A., and Radicioni, D. P. (In

press, 2014). A knowledge-based system for proto-

typical reasoning. Connection Science.

Machery, E. (2009). Doing without concepts. OUP.

Minsky, M. (1975). A framework for representing knowl-

edge. In Winston, P., editor, The Psychology of

Computer Vision, pages 211–277. McGraw-Hill, New

York.

Nardi, D. and Brachman, R. J. (2003). An introduction

to description logics. In Description logic handbook,

pages 1–40.

Pilato, G., Augello, A., and Gaglio, S. (2012). A modular

system oriented to the design of versatile knowledge

bases for chatbots. ISRN Artiﬁcial Intelligence, 2012.

Rosch, E. (1975). Cognitive representations of semantic

categories. J. Exp. Psychol. Gen., 104(3):192–233.

Straccia, U. (2011). Reasoning within fuzzy description

logics. arXiv preprint arXiv:1106.0667.

ADualProcessArchitectureforOntology-basedSystems