Modeling and Simulation of Associative Reasoning

Jiří Jelínek

Institute of Applied Informatics, Faculty of Science, University of South Bohemia,

Branišovská 1760, České Budějovice, Czech Republic

Keywords: Associative Memory, Graph Structures, Human Reasoning, Knowledge Representation, Retrieval,

Processing.

Abstract: Modeling human behavior is a popular area of research. Special attention is then focused on activities related

to knowledge processing. It is the knowledge that has a fundamental influence on an individual's decision-

making and its dynamics. The subject of research is both the representation of knowledge and the procedures

of their processing. The processing also comprises associative reasoning. Associations significantly influence

the knowledge base used in processing stimuli and thus participate in creating a knowledge context that is

further used for knowledge derivation and decision making. This paper focuses on the area of associative

knowledge processing. There are already classical approaches associated with developing probabilistic neural

networks, which can also be used with modifications at a higher abstraction level. This paper aims to show

that associative processing of knowledge can be described with these approaches and simulated. The article

will present a possible implementation of the model of knowledge storage and associative processing on the

individual's knowledge base. The behavior of this model will be demonstrated in experiments.

1 INTRODUCTION

Modeling of human beings' behavior is a popular area

of research. The main goal here is to understand and

imitate human behavior to be further investigated and

eventually implemented in artificial systems to use its

positives. Special attention is then logically focused

on processes related to the individual's knowledge

and its use in processing external stimuli and realizing

decision-making processes. One possible way to learn

more about them is modeling brain activities focused

on creating and storing information and knowledge.

The subject of research is both the representation of

knowledge itself and the procedures of their

processing.

An essential part of these activities is associative

reasoning because it acts as a modifier in selecting

and processing knowledge. When processing the

external stimulus, this reasoning participates in

creating the knowledge context used for this

processing.

The internal representation of knowledge in the

brain is the active subject of intense research. The

theory about this topic can be found already in the 80s

(Warrington & McCarthy, 1983, 1987; Warrington &

https://orcid.org/0000-0002-1842-2055

Shallice, 1984).

The research of conceptual memory

and processing changes can also be found in

(Patterson et al., 2007), based on the study of neural

deceases. Other sources focus on examining sensory

data's internal representations on low-level (Smith et

al., 2012).

However, at a low level, it is clear that the brain

activities are based on massive parallelism in

a network of nodes (neurons), which have

a reasonably simple functionality. These nodes are

interconnected by many links of the same type with

unequal significance or permeability (influence of

synapses). The structure of these interconnections in

the brain is not flat, and we can find specialized areas

with a specific connection and purpose, but (to get

a notion) the above description is sufficient. The

number of neurons is in the hundreds of billions, and

the number of links is many times larger.

We do not know the concrete way of storing

knowledge in the above structure. However, if we use

knowledge about artificial neural network models, it

is probably a combination of a suitable

interconnection and setting the throughput (weights)

of connections. Thus, a very large-scale graph

structure is used.

716

Jelínek, J.

Modeling and Simulation of Associative Reasoning.

DOI: 10.5220/0010234707160723

In Proceedings of the 13th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2021) - Volume 2, pages 716-723

ISBN: 978-989-758-484-8

If we use a certain degree of abstraction, the same

form of knowledge representation can represent

concepts and their links. However, other approaches

could also be used, e.g., formal logic tools with

temporal and spatial extensions. Some principles,

known from database technologies, can also be used,

such as relational data representation. Also, a possible

way is to use the objects; this is understandable to

people due to its similarity to the environment in

which people live, and we can find it in (Caramazza

& Mahon, 2006). Others use the object-attribute-

relation approach to describe this representation

(Wang, 2007).

However, all these structures can be implemented

at a lower level using graphs with the appropriate

node and link types. The idea of graph knowledge

representation can be found, e.g., in (Hayes, 2003).

Therefore, in our paper, we will prefer a graph

representation of knowledge in the form of concepts

and their relationships.

The rest of the article is organized as follows. In

Section 2, we look at the basics of associative

knowledge processing. In Section 3, the proposed

simulation model will be presented. In Section 4,

some experiments performed with the model are

discussed. At the end of the article, we will

summarize the findings and discuss the results.

2 RELATED WORK

The human brain's associative abilities in working

with data can be viewed from different angles. They

can also be categorized as a part of the knowledge

retrieval area. If we focus on modeling human

behavior, then, based on the human brain structures,

it seems that a suitable universal basis for storing

knowledge is graph representation. That also allows,

among other things, a very variable application of

higher formal knowledge representations.

Modeling the process of associative reasoning has

a relatively long history in the field of artificial

intelligence. As early as 1982, Hopfield came up with

its single-layer artificial neural network (Hopfield,

1982), sometimes referred to as auto-associative

memory. It is a recurrent neural network, where the

information is stored in dynamically stable attractors

of the network state. The network is assumed to be

symmetric (weights from node i to j are the same as

from node j to node i, w

= 0). However, this is not

always the case with human reasoning. The network

learning mechanism then has two development

variants, binary and continuous. A continuous variant

is closer to our goals, whose dynamics can be

described according to formula (1).

𝑠



=𝑤



.𝑠





−𝜃



(1)

In formula (1), Ø

is the threshold of node j, w

is the

weight of the link between nodes i and j, and s

and s

are states (activations) of nodes j and i. The use of the

above formula for network state updates can be

synchronous (in a single moment for all network

nodes) or asynchronous (at a given time or step, one

selected node's status is updated).

If we want to use the Hopfield network as

associative memory, we will first teach it to represent

a stored pattern (information, knowledge). The stored

information could then be entirely recalled by setting

(activating) any node within the respective attractor

range. For the network learning, the generalized Hebb

learning rule (Hebb, 1949) described by formula (2)

can be used, based on the neurobiological observation

that the bond strength increases when neurons at both

ends are activated simultaneously. For this basic

learning algorithm, it was empirically found that it is

possible to store about 0.15 N patterns in the binary

Hopfield network, which can then be called by

association (the number of nodes in the network is

denoted as N) (Jain, 1996).

𝑤







=𝑤







.+α 𝑠



𝑠



(2)

In formula (2), w

is the weight of the link between

nodes i and j, s

and s

are both end neurons'

activations. The learning parameter α has a standard

meaning and allows us to adjust the learning speed

and strength. Formula (2) is one of the oldest

algorithms for learning neural networks.

The above-described approach shows that the

graphs or networks are not new structures when

examining associative behavior. The proposed model

also uses them but assumes asymmetric links and

their weights (w

≠

). Our model also uses

modifications of the formulas mentioned above. An

interesting application of associative processes is

mentioned in (Diehl, 2009), where they are used for

information management on PDA.

Another approach to using graphs for associative

reasoning is to move up on a higher level of

abstraction and examine the links between concepts

in the knowledge base. Here, a stochastic approach is

often applied, where the weights of nodes and edges

in the graph are interpreted as probabilities of

occurrence of these elements. So, we are talking

about so-called probabilistic networks, also referred

to as Bayesian networks.

Modeling and Simulation of Associative Reasoning

717

These networks also apply probabilistic

reasoning. The basis here is formula (3) for

conditional probability, where P(i) is the probability

of the concept represented by node i. The notation

P(j | i) is then the probability of concept j with the

condition of concept i and thus relates to the oriented

edge between the two nodes.

𝑃

(

𝑗

𝑖

)

=𝑃(

𝑗

,𝑖) 𝑃(𝑖)

⁄

(3)

For several currently valid independent conditions,

the relationship can be adjusted to formula (4).

𝑃

(

𝑗

𝑖,𝑘,𝑙

)

=𝑃(

𝑗

,𝑖,𝑘,𝑙) 𝑃(𝑖,𝑘,𝑙)

⁄

(4)

When having the number of occurrences c of

individual nodes, the given formula can be modified

to formula (5).

𝑃

(

𝑗

𝑖,𝑘,𝑙

)

𝑐(

𝑗

∧𝑖∧𝑘∧𝑙)

𝑐

(

𝑖∧𝑘∧𝑙

)

(5)

The precondition for using these formulas is the

independence of the concepts i, k, and l. This

condition is problematic and hard to meet in networks

with the complex interconnection of nodes. However,

we can learn from that. A similar approach is used in

the presented model, but the specific formulas for

calculating the weights of edges are modified.

Procedures from fuzzy logic are also applied.

The social contextual influence on perceptual

processes is discussed in (Otten et al., 2017). The

importance of context for knowledge processing is

also the base idea of our model.

2.1 Graphs and Terms

For a graph representation, the key is what the nodes

of the graph represent. At a higher level of

abstraction, these can be concepts with which the

individual works. With their help, the individual (or

agent simulating him) describes himself and his

surroundings. This assumption of concept nodes is

followed up by using links between them representing

different types of interrelationships.

The concepts can be described in various ways.

Their minimal description is the identification of the

concept (term). If we use the WordNet terminology

(Miller, 1995), this will be a so-called synset. If we

assume that a textual description of a term is uniquely

tied to a single synset, this synset can be identified by

this textual description. Otherwise, the text can be

extended with a unique part to meet the above-

assumed condition. The text is then the primary

identification of the term. Other data can further

extend the description of the concept. In different

representations, the added information may differ.

For example, in the object representation, it may be

properties or behavior associated with the object

identified by the term. The extension may also

involve the addition of metadata, i.e., spatial and

temporal data linked to the concept's observation.

The graphical representation used in the paper

allows supplementing these extensions; in the form of

other concepts and interconnections. Thus, this

representation is very flexible and will enable us to

work in a single knowledge base with several higher

knowledge representation paradigms.

3 PROPOSED MODEL

The presented model is based on a graph

representation of concepts and uses only a single type

of link necessary to simulate associative reasoning

(association relation). The model uses a general

approach and has only a very few requirements on

stored concepts; it just needs their text identification.

It aims to show the abilities of humans' associative

reasoning based on the simultaneous observation of

individual concepts. Concurrence can be in temporal

and spatial dimensions and, based on others, at first

glance, invisible connections. The agent does not

know how the observed concepts are ontologically

linked and therefore treats them in the same general

way.

3.1 Model Principles

The basis of the model's operation is the assumption

that the individual uses its sensors to receive stimuli

from the environment at a given point in time. In these

stimuli, he can recognize concrete perceptions,

objects, or entities (concepts) identified by their name

or at least by temporary identification (for new and as

yet unknown concepts). Information about concepts

is stored in the knowledge base, together with

numbers of concepts' occurrence and activation

levels. The counts of occurrences enable us to respect

the probabilistic relations. The activation is the

primary tool when modeling the dynamics of the

whole simulation. Its values for nodes and links are

updated in each step according to previous data and

current observation.

The model is based on the assumption that an

individual, when using his knowledge, does not

always use his entire knowledge base, but only a part

of it, which is currently activated by the observed

concepts. So, a specific knowledge context is creating

for subsequent work with knowledge. It is this

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

718

knowledge context creation process that the model

tries to implement.

The model's input is an observation represented

by a list of terms that describe concepts

simultaneously observed. However, some of these

terms (concepts) may not express what the individual

recorded but may also represent other metadata

(individual's position at the time of observation,

current time, source of observation, and others); these

metadata are then considered as separate concepts.

The count of occurrence is incremented for each

observed concept.

All simultaneously observed concepts are fully

interconnected by oriented relationships (links,

edges) whose certainty (probability) is continually

adjusted in the simulation, based on several factors

(see below). Based on observations, the individual

creates his internal knowledge base and a model of

the world around him.

3.2 Model Details

The activity of the model can be (similarly as for

artificial neural networks or other models from the

machine learning area) divided into two phases: the

reasoning phase (production) and the setting's

modification phase (learning). However, these two

phases cannot be separated from each other; even the

reasoning brings up changes in the activation of nodes

and connections, and the model thus modifies at the

same time.

Actions from both the above phases repeatedly

proceed within the simulation. Each simulation step

includes the activities shown in Fig. 1.

The activities in the picture can be divided into

four categories:

• Implementation of the forgetting process

• Management of the counts of occurrence

• Node and edge activation settings

• Derivation of the current context

The first action in each step is to model forgetting,

which considers the simulation time course. The

model assumes that the previously observed concepts

gradually lose their importance after a new

observation and are overlapped with the new ones.

Therefore, the activation of these nodes decreases.

A similar process takes place for links. However,

several aspects must be respected when adjusting link

activations. The forgetting is then realized by

calculation according to formulas (6).

𝑙





=𝑙





𝑓



𝑠





=𝑠





𝑓



(6)

Separate parameters (f

for concepts and f

for

relationships) were used for nodes and connections,

both with values in (0; 1). One means that the

forgetting process is excluded, and the zero value

represents a state where the individual forgets all

activation values from previous observations. Two

separate parameters are used due to their different

influence on the model's operation and thus the

possibility of their different settings in experiments.

Figure 1: Actions in one step of the simulation.

The next phase of each simulation step is to

update the counts of occurrence of nodes and edges.

This action only applies to newly observed objects

and their interconnections. Counts are then used for

setting the primary activations of the links.

The value of the new link activation must consider

more factors. Therefore, it is calculated in several

steps. Here, the lowest (basic) value p

of link activity

is calculated according to formula (7).

𝑝



𝑛



𝑛



=𝑃(

𝑗

|𝑖)

(7)

Modeling and Simulation of Associative Reasoning

719

In formula (7), n

denotes the occurrences of a given

link, and n

occurrences of the concept i. The p

values

are considered long-term and static, representing the

relationship's certainty or probability.

If nodes or edges with the same identification

already exist in the knowledge base, their occurrence

count is incremented. Otherwise, new objects are

added to the database with an occurrence of 1. Next,

the activations s

and l

of the newly observed

concepts and connections are set to 1.0 (very actual),

both for nodes and links.

The following phase of the simulation is the most

computationally demanding part of the model. Its

effectiveness depends very much on the appropriate

representation of the entire collection of nodes and

their connections. The activations of all nodes in the

knowledge base are being modified here using the

breadth-first search mechanism for selecting the

nodes with a start on currently observed ones. The

new values are calculated according to formula (8). If

the new activation value is lower than the original

one, the higher value is used. This procedure

guarantees that the new activation value will be the

highest possible, respecting all the links that point to

the node.

𝑠





=max



∈

𝑠





,𝑠



.𝑙





(8)

In formula (8), s

is the activation of the term j, s

the

activation of the term i, from which there is a direct

link to the term j, and l

the reliability of this link. The

P is a set of nodes preceding (in terms of existing

connections) the term j. The logic of using the highest

value of certainty of occurrence is based on the fact

that each manipulation with a given term causes its

reactivation. As a result of this step and disseminating

information about the new observation, the activation

of all terms associated with the observed ones by

correctly oriented links (also by multiple links across

other nodes) will be adjusted.

The next phase of the simulation step is the update

of the link activations. This step must respect both the

static probability of the link according to formula (7),

the process of forgetting according to formula (6) and

also it is necessary to respect the neurobiological

observation and Hebb's rule of learning expressed by

formula (2). Activation is then calculated according

to relation (9).

𝑙





=max𝑝



,𝑠



𝑠



,𝑙







(9)

In the formula, 𝑙





is the actual value of link activity

from the previous step after using the forgetting

mechanism, 𝑙





is the new link activation, p

is the

value of the conditional link probability. The s

and s

are the current activations of nodes i and j and product

. s

represents the Hebb rule.

All of the previous actions in the simulation step

result in the current setting of activations in the

knowledge base. Then only nodes and links with

activation higher than the selected threshold are

considered relevant and used in following cognitive

and decision-making processes (not presented in this

paper). These objects thus form the so-called

knowledge context. The thresholds may (according to

our needs) differ for nodes and links. Therefore, two

limit parameters a

and a

are used. Thus, only nodes

that satisfy the relationship are included in the current

context (10).

𝑠



>= 𝑎



(10)

For links, the fulfillment of the formula (11) is

required. The limit of link activation must be reached,

but both its ends must also be in the context.

𝑙



>= 𝑎



⋀

(

𝑠



>= 𝑎



)

⋀ 𝑠



>= 𝑎





(11)

The presented model works primarily with the

temporal concurrence of concepts. However, thanks

to the possibility of including other metadata (e.g., the

place of observation) as the additional concepts in

observation, the spatial and possible other

concurrences can also be respected.

3.3 Model Implementation

The described model was implemented in Java and

took full advantage of the running program threads in

parallel.

Two ways of working with the model were tested.

The first operation mode implements only a subset of

the activities of Fig. 1 and allows the model to be run

only in the input observation processing mode. In it,

only the counts of the occurrence of nodes and edges

are stored. Therefore, the created graph base is

focused on capturing long-term and statistically based

information, which, however, does not capture the

dynamics of the associative reasoning process. This

mode can thus be used to create a basis for the graph

knowledge base. The activation of objects is not set,

and the base is thus ready for further use without the

influence of its creation on the association process.

The mode can be therefore referred to as initial. This

mode may or may not be used, but its main advantage

is the high-speed processing of a large volume of

input observations.

There are many solutions for implementing graph

data storage from relational through various NoSQL

databases to their specific type, graph databases. An

internal representation storing the data in the working

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

720

memory was used to achieve the maximum speed of

data processing.

However, the model's primary operating mode is

the full mode implementing for each observation the

set of activities shown in Fig. 1. It ensures complete

capture of knowledge base development dynamics

and the activation of individual relationships and

concepts during the simulation. This mode can be

used right from the beginning of working with the

model, and during it, a graph base is created from

scratch. However, it can be used even after the initial

mode, and then it will work with the already prepared

basis of the knowledge base, and it will expand it

further. This operation mode of the model is thus

significantly oriented to the current state of

knowledge. It evokes a certain resemblance to short-

term memory, where nodes and links that do not have

a significant number of occurrences can be very

active during some time after we observe them.

4 EXPERIMENTS

The FreeBase dataset (Kuznetsova et al., 2018) was

used to test the model. It includes a database of

approximately 5.6 million annotated images. The

annotation is created both expertly and with the help

of automatic derivation. It is in the form of a list of

entities (concepts) identified in a particular image.

This description corresponds with the model's

requirements on input data. Every image was

supposed to be one observation of the world.

The dataset mentioned above was used in two

ways. The first one was testing of the model

performance, where the entire dataset was used.

During the initial mode was imported 19,508 nodes

and 79,399,030 links between them. A computer with

a minimum of 32 GB RAM was used to process this

volume of data. All data were imported within a few

tens of seconds.

The second way of using the dataset was to

examine the dynamics of association processes. The

following experiments were carried on the subset of

1,000 image annotations from the dataset with 2,304

nodes and 113,496 links.

The first experiment was focused on examining

the dynamics of knowledge base creation using the

model's main mode. The results can be seen in Fig. 2,

where both the total numbers of nodes and edges

(totals) and their numbers in the current context

(active objects) are displayed. The simulation was

performed in 1000 steps; in each step, an annotation

of one image from the above set (observation) was

used as input. The values a

= a

= 0.5 were used to

generate the context. Forgetting parameters were set

to f

= 0.95 and f

= 0.97.

Figure 2: Creation of graph base.

Figure 3: Comparison of contexts created with activation

limits of 0.1 and 0.5.

The previous experiment's context sizes were

compared with setting a

= a

= 0.1 to examine the

effect of the activation limit values on the context

size. The results are in Fig. 3, and the difference is

visible here; the lower values of limit cause more

nodes and edges in the context.

The next experiment was performed to verify the

influence of the initiation mode on the created

context. The initiation mode was followed by the

submission of all 1000 image annotations as in

previous experiments. The results are given in Fig. 4,

comparing this experiment with the first one (Fig. 2).

The differences in numbers of objects in contexts

are minimal both for nodes and edges. When using

the initiation mode, the counts of occurrence were

already set, and the following main mode further

increased them gradually. That caused the small

differences. The forgetting parameters' values did not

change; the activation limit values a

and a

were 0.5

both.

Modeling and Simulation of Associative Reasoning

721

Figure 4: Comparison of contexts created with initiation

mode (init) and without it (main).

Figure 5: The context based on a single term "Truck."

Figure 6: The context based on a sequence of terms ended

with the term "Truck."

Another experiment focused on examining the

internal composition of the context. The starting point

was the same knowledge base, as in the previous

cases created using the model's initial mode. The

model's main mode was subsequently used on the

knowledge base. Separate concepts were introduced

so that no new interconnections were added. In the

first case, the only term "Truck" was entered, and the

created context is shown in Figure 5.

The influence of the previous observations on the

context can be seen in Fig. 6, where before the term

"Truck," the words "Bus," "Motorcycle," "Man," and

"Plant" were entered. The context is much broader

and also contains terms not directly connected to the

word "Truck." The activation limit here was again

0.5.

Figure 7: The influence of forgetting factors f

and f

Both forgetting factors f

and f

also have a

significant influence on the context. This effect can

be demonstrated in Figure 7, where two different

settings are used: the original f

= 0.95 and f

= 0.97

used in previous experiments and the adjusted

= 0.995 and f

= 0.997. Less influence of forgetting

leads to compensation of fluctuations in context

creation dynamics and a more significant number of

included nodes and links.

5 CONCLUSIONS

The presented paper introduces a model of associative

reasoning used by humans, which simulates this

process, including its dynamics. The model works

with the individual's observations. These define the

concepts and their interconnections, which both have

to be stored. With the help of the knowledge context

created during associative reasoning, it is then

possible, in information and multiagent systems, to

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

722

consider the agent's personal history and thus better

capture the agent's individuality.

In its processes, the model respects long-term and

short-term links between concepts. Functionally the

model is partly based on Hopfield's auto-associative

memory and Hebb's rules of learning, whose it

modifies for given conditions and goals.

The model was tested on a dataset of annotated

images. The results demonstrate the model's ability to

perform associative reasoning and create a current

knowledge context usable for other agent processes.

That was the main goal of the work. The results also

show that the whole process is significantly affected

by global parameters, mainly forgetting coefficients

and limiting objects' activation when generating

context.

Future research will focus on further testing the

model, both in terms of its efficiency and

performance in processing large graph structures and

its compliance with the observed human behavior. In

a longer perspective, the goal is to use the model in

other projects focused on multiagent systems and

behavioral simulations.

Work on the model will also focus on the

possibilities of processing data of a different nature.

The annotation of the images probably best

demonstrates the model activity, but time series or,

more generally, data captured in relational structures

can also be processed. The model can then provide a

different view of this data and the possibilities of its

processing.

REFERENCES

Caramazza, A., Mahon, B.Z., 2006. The organisation of

conceptual knowledge in the brain: The future's past

and some future directions. Cognitive

Neuropsychology, 23(1), pp. 13-38.

Diehl, J., 2009. Associative personal information

management. CHI '09 Extended Abstracts on Human

Factors in Computing Systems (CHI EA '09). ACM,

New York, NY, USA, pp. 3101–3104.

Hayes, P. J., 2003. Knowledge representation.

Encyclopedia of Computer Science. John Wiley and

Sons Ltd., GBR, 947–951.

Hebb, D. O., 1949. The organization of behavior. John

Wiley & Sons, New York.

Hopfield, J. J., 1982. Neural networks and physical systems

with emergent collective computational abilities.

Proceedings of the national academy of sciences, 79(8),

pp. 2554-2558.

Jain, A. K., Mao, J., Mohiuddin, K. M., 1996. Artificial

neural networks: A tutorial. Computer, 29(3),

pp. 31- 44.

Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin,

I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M.,

Kolesnikov, A., Duerig, T., Ferrari, V. (2018). The

open images dataset v4: Unified image classification,

object detection, and visual relationship detection at

scale. arXiv preprint arXiv:1811.00982.

Miller, G. A., 1995. WordNet: A Lexical Database for

English. Communications of the ACM, Vol. 38, No. 11:

pp. 39-41.

Otten, M., Seth, A.K., Pinto, Y., 2017. A social Bayesian

brain: How social knowledge can shape visual

perception. Brain and cognition, 112, pp. 69-77.

Patterson, K., Nestor, P.J., Rogers, T.T., 2007. Where do

you know what you know? The representation of

semantic knowledge in the human brain. Nature

reviews neuroscience, 8(12), pp. 976-987.

Smith, M.L., Gosselin, F., Schyns, P.G., 2012. Measuring

internal representations from behavioral and brain

data. Current Biology, 22(3), pp. 191-196.

Wang, Y., 2007. The OAR model of neural informatics for

internal knowledge representation in the

brain. International Journal of Cognitive Informatics

and Natural Intelligence (IJCINI), 1(3), pp. 66-77.

Warrington, E.K., McCarthy, R., 1983. Category specific

access dysphasia. Brain, 106(4), pp. 859-878.

Warrington, E.K., McCarthy, R.A., 1987. Categories of

knowledge: Further fractionations and an attempted

integration. Brain, 110(5), pp. 1273-1296.

Warrington, E.K., Shallice, T., 1984. Category specific

semantic impairments. Brain, 107(3), pp. 829-853.

Modeling and Simulation of Associative Reasoning

723