INTERACTIONS BETWEEN HEMISPHERES

WHEN DISAMBIGUATING AMBIGUOUS HOMOGRAPH

WORDS DURING SILENT READING

Zohar Eviatar

, Hananel Hazan

, Larry Manevitz

, Orna Peleg

and Rom Timor

Institute of Information Processing and Decision Making, University of Haifa, Haifa, Israel

Department of Computer Science, University of Haifa, Haifa, Israel

Keywords: Simulation, Neural network, Corpus collosum.

Abstract: A model of certain aspects of the cortex related to reading is developed corresponding to ongoing

exploration of psychophysical and computational experiments on how the two hemispheres work in

humans. The connectivity arrangements between modelled areas of orthography, phonology and semantics

are according to the theories of Eviatar and Peleg, in particular with distinctions between the connectivity in

the right and left hemisphere. The two hemispheres are connected and interact both in training and testing in

a reasonably "natural" way. We found that the RH (right hemisphere) serves to maintain alternative

meanings under this arrangement longer than the LH for homophones. This corresponds to the usual

theories (about homographs) while, surprisingly, the LH maintains alternative meanings longer then the RH

for heterophones. This allows the two hemispheres, working together to resolve ambiguities regardless of

when the disambiguating information arrives. Human experiments carried out subsequent to these results

bear this surprising result out.

1 INTRODUCTION

Neuropsychological studies have shown that both

cerebral hemispheres process written words, but they

do it in somewhat different ways (e.g., Iacoboni &

Zaidel, 1996, Grindrod & Baum, 2003).

Previous simulation work has examined the

activation of meanings of ambiguous words with

polarized meanings (where one meaning is much

more frequent (dominant) in the language) and has

shown that transfer of information from a 'right

hemisphere' (RH) network to a 'left hemisphere'

(LH) network, when context biasing to the

nondominant meaning is presented after the initial

presentation of the word, is the most efficient

mechanism for "recovery" from erroneous activation

of the dominant meaning. That is, there are

systematic cases where the LH purported

architecture could not recover by itself; nor could

the RH perform at high levels of performance (Peleg

et al., 2007, 2010). Other simulation work (Weems

& Regggia, 2004) suggests that different

connections can produce different results.

This paper examines different possible

connections between networks representing the two

hemispheres and how these differences affect the

results of processing homophones. (Monaghan &

Pollmann, 2003) shows that when stimuli have to be

matched in a complex task (such as whether two

letters have the same name), performance is better

when stimuli are presented across the hemispheres

of the brain. Furthermore, they argue that for

simpler tasks (such as whether two letters have the

same shape), better performance is achieved when

stimuli are presented unilaterally. They show that

this bilateral distribution advantage effect emerged

spontaneously in a neural network model learning to

solve simple and complex tasks with separate input

layers and separate, but interconnected, resources in

a hidden layer. They also show that relating

computational models to behavioral and imaging

data helps to understand hemispheric processing and

generating testable hypotheses.

This paper presents the computational advantage

of having two networks that can exchange

information: LH fully connected (Orthography,

The authors are listed in alphabetical order. This work appears as

art of the M.Sc. thesis of Rom Timor. We thank the Caesare

Rothschild Institute of University of Haifa for its support.

271

Eviatar Z., Hazan H., Manevitz L., Peleg O. and Timor R..

INTERACTIONS BETWEEN HEMISPHERES WHEN DISAMBIGUATING AMBIGUOUS HOMOGRAPH WORDS DURING SILENT READING .

DOI: 10.5220/0003059802710278

In Proceedings of the International Conference on Fuzzy Computation and 2nd International Conference on Neural Computation (ICNC-2010), pages

271-278

ISBN: 978-989-8425-32-4

 2010 SCITEPRESS (Science and Technology Publications, Lda.)

Phonology and Semantics) and RH lack of

connection between Orthography and Phonology.

2 BACKGROUND

Behavioral studies have shown that the LH is more

influenced by the phonological aspect of written

words whereas lexical processing in the RH is more

sensitive to visual form. A large amount of

psycholinguistic literature indicates that readers

utilize both frequency and context to resolve lexical

ambiguity (e.g., Titone, 1998 , Peleg et al., 2004).

Although hemispheric specialization for LH in

language processing is assumed, it is also assumed

that the RH plays a significant role in language

function, especially when ambiguous words are

presented in context (e.g. Burgess & Simpson,

1988).

Behavioral studies examining the disambiguation

of homophones (e.g., “bank”) suggest that all

meanings of an ambiguous word are initially

activated in both hemispheres, but at different

speeds. While the LH quickly activates both

meanings and then selects one alternative (the

contextually compatible meaning when prior

contextual information is biased, or the salient, more

frequent meaning when embedded in non-

constraining contexts), the RH activates the

nondominant meaning more slowly, and maintains

both alternate meanings (including less salient,

subordinate and contextually inappropriate

meanings).

Previous studies also suggests that exchange of

information between the LH and the RH networks

will produce better performance and can help the LH

recover the subordinate meaning, when it is

appropriate to the context (This task the LH could

not perform in isolation.)

2.1 Research Goals

The main goal is to investigate how different types

of information (phonological, lexical and contextual)

are utilized during silent reading in the two

connected networks simulating the left and right

hemispheres. Specifically the results are crucial for

answers regarding inter-hemispheric relation during

the disambiguation process of homophones.

We achieved this goal by building a neural

network that can process word information

(phonological, lexical and contextual) and resolve

the meaning of ambiguous words in Hebrew. The

network is based on (Peleg et al., 2007 & 2010) LH

and RH networks architecture while adding

connections between regions (Orthography,

Phonology and Semantics) in various ways.

The connected networks after training,

demonstrate the effects of context and frequency on

the resolution of homophones. The computer model

consists of "weakly coupled" neural networks that

can deal with ambiguity of a written or a spoken

word in Hebrew. The main idea is to investigate

some questions regarding the "weakly coupled"

connection properties such as when, how and where

information is transferred and determines the degree

of transferred information while shedding light on

the division of labor between hemispheres. The

"weakly coupled" networks should support the same

properties when disconnected and additional or

improved properties when connected.

Furthermore, we measure the time it take for the

connected networks to resolve the meaning and the

paths the networks use to do so. Then we compare

the results to existing psycholinguistics theories of

how humans process the language. One of the

reasons to build computational models is the ability

to change parameters, aspects and connection

properties of the models in ways that are not

possible with human subjects. This provides us with

an insight into the mechanisms of reading and

understanding the meaning of words.

3 PREVIOUS WORK

3.1 Kawamoto’s network

Kawamoto (Kawamoto, 1993) designed his neural

network model in such a way that the entire word,

including its orthographic, phonological and

semantic features occurs as an “attractor” in the

recurrent network.

According to his model, the more frequent a

certain meaning of the word in a certain context is,

the stronger the attractor it will be, and the

completion of other features (semantic and

phonological) would usually fall into this attractor.

Another factor examined was the time lapse

between accessing the dominant meaning and the

time lapse of accessing the secondary (subordinate)

meaning (Kawamoto, 1993).

3.2 Hazan’s Network

Peleg, Eviatar, Hazan & Manevitz in (Peleg et al.,

2007 and (Peleg et al., 2010) designed a two-

hemisphere model based on Kawamoto’s model (see

ICFC 2010 - International Conference on Fuzzy Computation

272

Figure 1). The model includes two separate

networks. One network incorporates Kawamoto’s

version, and successfully simulates the time course

of lexical disambiguation in the LH. In the other

network based on the behavior of the disconnected

RH of split brain patients (Zaidel & Peters, 1981), a

change was made in Kawamoto's architecture,

removing the direct connections between

orthographic and phonological units. (Peleg et al,

2010)

Figure 1: Hazan's network architecture.

Figure 2: Illustration of network that include CC

connections between Corresponding regions of LH & RH.

3.3 Weems & Reggia Network

Weems & Reggia tested hemispheric specialization

and independence for word recognition while

comparing three computational models: Callosal

Relay (strong right to left, minimal left to right

connectivity, output from LH), Direct Access

(minimal connectivity between hemispheres,

separate outputs) and Cooperative (strong

connection, single output) and showed advantage for

the Cooperative model together with a slight

performance dropdown (Weems & Reggia, 2004).

4 COMPUTATIONAL

SIMULATION

The simulation is based on (Peleg et al., 2007) LH

and RH network which includes the implementation

as described by (Kawamoto, 1993) with some

changes in the encoding. The simulation includes the

“Corpus Callosum” (CC) that was implemented as a

connection from LH units to RH units in a various

ways including "One to One

", "One to Many

within regions and between regions

(See Figure 2).

3.4 The Learning Stage

The network was trained with a simple error

correction algorithm (Kawamoto, 1993) taking into

consideration a learning constant and the magnitude

of the error determining a bipolar activity of a single

unit. This activity is determined by the input from

the environment, the units connected to it (within the

hemisphere and from the CC) and a decay in its

current level of activity. The learning process was

achieved by altering the weights between the units

of the network to minimize the error between the

activation level and the network input.

∆



 

















, 











n – Learning constant.

– target activation levels of unit i and j.

– net value of unit i.

In a learning trial an entry was selected randomly

from the lexicon. Dominant and subordinate

meanings were selected with a ratio of 5 to 3

roughly based on linguistic considerations. The

learning phase was divided to the following steps:

A. Initialization of units with random values.

B. Random order of sets of words.

C. The network was trained with 48 words.

D. The network was tested if more training is

needed. If so another 48 words were chosen to

continue the training. The testing had to fulfill

these conditions:

• Presenting the orthographic part of word leads

the network to select the dominant meaning.

• Presenting the orthographic part of word with a

clue to the subordinate meaning leads the

network to select the subordinate meaning.

The learning was stopped when the conditions

were fulfilled for each group of words (homophones,

hetrophones and normal words) separately or when

the training set ended.

In a learning trial an entry was selected randomly

from the lexicon. Dominant and subordinate

meanings were selected with a ratio of 5 to 3

roughly based on linguistic considerations. We

performed different experiments that include

different learning stages. First, the learning stage

One to One: each neuron from LH/RH is connected to the

corresponding neuron in the other hemisphere.

One to Many: each neuron from LH/RH is connected to a group o

neurons in the corresponding area of the other hemisphere.

Regions: Orthography, Phonology and Semantics.

INTERACTIONS BETWEEN HEMISPHERES WHEN DISAMBIGUATING AMBIGUOUS HOMOGRAPH WORDS

DURING SILENT READING

273

was done while the LH and RH are disconnected.

We connected them only while testing the model.

Second, the learning stage was done when the LH

and RH are connected via the CC. This was

performed in two manners: free learning (no

restriction on the CC weights) and restricted

learning. In the restricted learning the weights on the

CC did change but were limited to 0.1 - 0.3.

3.5 Testing the Model

After the networks were trained they were tested by

presenting just the orthographic part of the entry as

the input (to simulate neutral context) or by

presenting part of the semantic (subordinate

meaning) sub-vector after presenting the

orthography (to simulate contextual bias). In each

simulation the input sets the initial activation of the

units.

Each unit was influenced from the following

sources:

A. External stimuli (orthophonic part of word or

clues).

B. Previous values from the last iteration

multipled by the decay rate.

C. Sum of the inner connected units output

multipled by the weights.

D. Sum of the inter-hemispheric connected units

output (Simulates the CC).

The activity of unit a at time t+1 is:

where:

δ – the decay variable.

The decay variable was set dynamically starting

from 0.6 , increasing while network is progressed

and ending at value of 1 when the run is completed.

In order to assess lexical access, the number of

iterations through the network requiered for all the

units in the spelling, pronunciation or meaning fields

to become saturated, was measured. A response was

considered an error if the pattern of activity did not

correspond with the input; non convergent if all the

units did not saturate after 50 iterations.

Testing was done after training the connected LH

and RH or after setting fixed weights on the CC. In

the latter case in different experiments the weights

were fixed uniformly at values that varied between

0.05 to 0.50 or one value was chosen for the weights

from LH to RH and a different one for RH to LH.

In order to test the maintenance of alternative

meanings, tests were run where no semantic clues

were given for various numbers of iterations (and

thus the networks started to converge towards the

dominant choice), and then clues for the subordinate

meaning were given. The differences in recovery of

the RH and LH in the different cases were measured.

3.6 Results and Analysis

In each simulation, 12 identical networks were used

to simulate 12 subjects in an experiment by varying

their training randomly. During the testing phase the

network received various inputs. First the

orthography of a word and then the inputs including

clues from the word meaning. The level was set to

+0.25 if the corresponding input feature was

positive, -0.25 if it was negative and 0 otherwise.

Result of each trial was recorded including the

number of iterations needed for coverage and the

total number of errors. The data was separated for:

• Group of words (homophones , hetrophones and

normal words).

• Type of clues (to subordinate or to dominate).

• Number of clues.

• CC weights or weight limitation.

• Mean and standard deviations were calculated.

In this work the focus was on the different type

of connections in the different ambiguity task

(hetrophonic vs. homophonic).

3.7 Results

Previous results of (Peleg et al., 2007 & 2010)

indicated that without transfer of data between LH

and RH the LH cannot recover to the subordinate

meaning after receiving semantic clues and thus

selects the dominant meaning. The RH was able to

perform this recovery and select the subordinate

meaning (See Figure 3). This phenomenon was

called the "Change of heart".

Figure 3: Network performance without CC. Only RH can

perform the "Change of heart" for homophones.

ICFC 2010 - International Conference on Fuzzy Computation

274

Our initial results indicated that when setting the

weights of CC from RH to LH to 0.25 in a "One to

One" connection the transfer of data from RH to LH

can help the LH perform the "Change of heart" and

select the subordinate meaning (See Figure 4).

Figure 4: Network performance with CC. RH & LH can

perform the "Change of heart" for homophones.

3.7.1 Homophones

Table 1 shows the results of average convergence

time

for LH and RH when presenting a homophonic

word without clues, the recovery status (in general)

when presenting a word with clues

to the

subordinate meaning and the sum of errors and non-

convergences.

Table 1: RH & LH convergence time (in iterations)

Homophones with no context (* Errors and Non-conv are

out of 96 in each hemisphere).

Network architecture LH RH

Errors*

LH/RH

Non-conv*

LH/RH

Without CC

40.32

(3.42)

41.54

(4.19)

29 0 37 0

With CC: Weights fixed at 0.25

(RH to LH)

39.18

(3.24)

41.06

(2.93)

11 0 14 0

With CC: Weights fixed at 0.25

(LH phonology to RH

phonology, RH semantics to LH

semantics).

39.77

(4.21)

40.69

(4.48)

23 12 19 7

With CC: Weights fixed at 0.25

RH to LH and 0.10 LH to RH.

39.84

(5.33)

40.14

(4.77)

21 9 17 11

With CC: Weights fixed at 0.25

(All, Both ways)

40.36

(6.03)

40.79

(5.64)

31 24 33 19

Connected learning yield the following results:

A. Free learning of CC weights caused the LH and

RH to lose their special properties. LH became

slower while selecting the dominant meaning and

the RH lost its ability to perform the "Change of

heart" when presented with clues to the

subordinate meaning (See Figure 5).

B. Restricted learning was able us to cause the LH

and RH to not lose their special properties. Both

RH and LH performed the "Change of heart" but

LH recovery is partial (See Figure 6).

Figure 5: Network performance with CC (Free learning).

RH & LH cannot perform the "Change of heart" for

homophones.

Figure 6: Network performance with CC (CC weights are

fixed at 0.25). RH & LH can perform the "Change of

heart". Note LH recovery is partial.

Table 2: Network performance in resolving hetrophone

(various architectures) [*Errors and Non-conv are out of

96 in each hemisphere]. Case 2 in this table above seems

to be the optimum for resolving hetrophone ambiguity.

Network architecture LH RH

Errors*

LH/RH

Non-conv*

LH/RH

Without CC

30.39

(4.88)

28.07

(5.14)

0 11 0 5

With CC: Weights fixed at

0.25 (LH to RH)

30.14

(5.11)

27.51

(5.37)

0 7 0 3

With CC: Weights fixed at

0.25 (RH phonology to LH

phonology, LH semantics to

RH semantics).

29.77

(6.52)

29.23

(5.93)

7 13 5 9

With CC: Weights fixed at

0.25 LH to RH and 0.10 RH

to LH.

29.63

(7.13)

28.36

(7.20)

9 16 4 2

With CC: Weights fixed at

0.25 (All, Both ways)

29.32

(9.31)

29.95

(8.67)

19 18 12 15

3.7.2 Hetrophones

Table 2 shows the results of average convergence

time

for LH and RH when presenting a hetrophonic

word without clues, the recovery status (in general)

when presenting the word with clues

to the

Standard deviation in parentheses.

Using different number of clues changed the results but in a unifor

way when comparing the results of different architectures. Th

result presented is for 4 clues out of 8.

Standard deviation in parentheses.

Using different number of clues changed the results but in a uniform

way when comparing the results of different architectures. The

result presented is for 4 clues out of 8.

INTERACTIONS BETWEEN HEMISPHERES WHEN DISAMBIGUATING AMBIGUOUS HOMOGRAPH WORDS

DURING SILENT READING

275

subordinate meaning and the sum of errors and non-

convergences.

Figure 7: Network performance when weights on CC are

fixed at 0.25 (LH to RH). Both LH and RH can perform

the "change of heart" for heterophones.

Figure 7 and 8 shows the time course of

convergence corresponding to case 2 in the table

above. Trails where performed for various CC

weights between 0.1 to 0.3 (with 0.05 intervals), one

way or both way, same regions or between regions.

Figure 8: The same diagram as the previous figure but

presented here with standard deviation.

4 DISCUSSION

4.1 Homophones vs. Hetrophones

Previous work (Peleg et al., 2007 & 2010) showed

that in the homophone case running the LH without

data transfer from RH has substantially worse

performance, both in number of iterations to

convergence and in the ability to perform the

"Change of heart" when presented with clues to the

subordinate meaning.

(Peleg et al., 2010) demonstrated the above by

transferring the data between the hemispheres

artificially. After some iterations the data from the

RH was copied to the LH and was clamped for

further iterations. Transfer of data from RH to LH in

homophones yieled better performance for the LH

even in cases when the RH has failed to perform the

recovery.

This work shows that:

1. Connecting the LH and RH in a more natural

way draws the same conclusions in

homophones (See Table 1 - Row 2 and Figure

3).

2. Data transfer in homophones is more beneficial

when done from RH to LH (See Table 1 - Row

2 compared to Table 1 Row 3-5).

3. Data transfer in hetrophones can be more

beneficial when done from LH to RH (See

Table 2 - Row 2 compared to Table 1 Row 3-

5). Note that results are less conclusive.

4. Connection between the hemispheres via the

CC is "weakly coupled" as compared to the

inner hemisphere connections (See Table 1 and

Table 2 Rows 2-5)

Word processing is different in LH and RH when

comparing different tasks such as homophone and

hetrophone disambiguate resolution. In homophones

the RH has less error and non-convergence cases

then LH but in the cost of convergence time.

Whereas in hetrophone the LH has less error and

non-convergence cases then RH but again in the cost

of convergence time

The convergence time

drawback in performance is an advantage when

trying to perform the "change of heart" from

dominate to subordinate meaning because then the

subordinate meaning is still available in the “slower”

hemisphere. This ability to perform the "change of

heart" more efficiently helps when transferring data

between hemispheres. The diffrence in convergence

time is due to the networks architectures.

4.2 Connected Learning vs. Separate

Learning

Results of connected learning also point out some

interesting facts. In general connected learning has

better performance in convergence time then with

separate learning.

Further, it is shown that free learning of the CC

weights causes the network to lose the "weakly

coupled" proportions and therefore the LH and RH

lose their special properties (convergence time and

Weights on the CC must be more than 0.05 in order to make

difference and less than 0.30 to prevent non convergence. Note that,

in contrast, inner hemispheric weights vary from -1 to 1, and forms

a relative strong intra-hemispheric connection between the

hemispheric regions orthography , phonology and semantics.

Note that in hetrophones the different time course of the LH is no

so significant than in homophones and therefore the results are no

as conclusive as in homophones.

ICFC 2010 - International Conference on Fuzzy Computation

276

"Change of heart"). Furthermore, learning with

bounded weights on the CC produces the desired

properties only if the CC bounded weights are less in

proportion to the interior hemispheric natural

boundary of weights (1 to -1), thus forming a

"weakly coupling" between the hemispheric

networks.

Results of LH and RH after connected learning

are slightly different then in separate learning. In

performance variables such as convergence time

there is a slight advantage to connected learning but

in errors measurements connected learning shows

worse results (in comparison to the results

demonstrated in separate learning).

As mentioned above the LH and RH has a

different time course and that each hemisphere has a

different time course in homophones and

heterophones. In separate learning it is shown that

the different between homophone and hetrophones

in the RH are not significant but are significant in

the LH. Further, separate learning shown than the

RH has a longer time course both in homophones

and in hetrophones. The different time course is

maintained in connected learning but it is noted that

the significant difference between homophones and

hetrophones is more prominent and that in the

connected learning the time course of RH is longer

only in homophones while in hetrophones the LH

has a longer time course.

In connected learning we can see that there is an

advantage to transfer data from RH to LH in

homophones and help the LH recover where in

hetrophone the transfer of data from LH to RH has

no significant effect. Note that in hetrophones

transfer of data from RH to LH has a negative effect

on the LH ability to recover.

4.3 Consequences for Human

Experiments

Recently, behavioral studies have been performed by

Peleg and Eviatar (Peleg & Eviatar, 2007 & 2010)

designed to test certain intra-hemispheric

connectivity assumptions that they put forward.

These studies combined divided visual field (DVF)

techniques with a semantic priming paradigm.

The behavioral studies were conducted in

Hebrew and combined a divided visual field (DVF)

technique with a semantic priming paradigm.

Subjects were asked to focus on the center of the

screen and to silently read sentences that were

presented centrally in two stages. First, the sentential

context was presented for 1500 ms and then the final

ambiguous prime was presented for 150 ms. After

the prime disappeared from the screen a target word

was presented to the left visual field (LVF) or the

right visual field (RVF) for the subject to make a

lexical decision. Targets were either related to the

dominant or the subordinate meaning or unrelated.

Magnitude of priming was calculated by subtracting

reaction time (RT) for related targets from RT to un-

related targets. The most interesting results were

observed in the subordinate-biasing context

condition (“The fisherman sat on the bank”): At 250

SOA both meanings (money and river) were still

activated in both hemispheres (Peleg & Eviatar,

2009). However, 750 ms later (1000 SOA), a

different pattern of results was seen in the two visual

fields. For homophones (e.g., “bank”), previous

results were replicated: the LH selected the

contextually appropriate meaning, whereas both

meanings were still activated in the RH These

studies, although limited to reaction time did

succeed in implying different patterns of activation

of both meanings in the two hemispheres. Our

simulations correspond to their intra-hemispheric

connectivity assumptions and produce results that fit

well with those human experiments and thereby

further support the theoretical underpinnings of

Peleg and Eviatar (Peleg & Eviatar, 2009). Here the

interpretation of the similarity of activation to

dominant and subordinate meanings at iterations is

taken as parallel to maintenance of the

corresponding meanings in the hemispheres.

Our work suggests a refinement of these

experiments to check as well the connectivity

strength between hemispheres. One possible method

to do this, would be to use Dynamic Causal

Modeling (Friston et al., 2003) to test the effective

connectivity between hemispheres during fMRI

studies. Such an experiment is currently being

prepared.

Our prediction as indicated above is that the RH

is functionally connected to the LH and vice versa

but in an asymmetric manner, with (1) the RH being

more strongly connected to LH than vice versa and

(2) the inter-hemispheric connections are relatively

weak compared to the intra-hemispheric

connections. In addition, our experiments indicate

that the major learning changes should be intra-

hemispheric.

5 SUMMARY

We implemented a model of both the RH and LH,

with architectural differences between the

hemispheres as proposed by the theories of Peleg

INTERACTIONS BETWEEN HEMISPHERES WHEN DISAMBIGUATING AMBIGUOUS HOMOGRAPH WORDS

DURING SILENT READING

277

and Eviatar (Peleg & Eviatar, 2009). The

hemispheres are linked together in a natural fashion,

both during learning and functioning. The results of

the simulations show that the connections between

the hemispheres allow additional functionality for

the LH as observed in humans ("change of heart");

and the hemispheres also perform at comparative

speeds that also qualitatively match human DVF

experiments.

Further, our work predicts connectivity strength

between the two hemispheres in architectural

regions; and thus suggests new human experiments.

REFERENCES

Iacoboni M, Zaidel E (1996). Hemispheric independence

in word recognition: evidence from unilateral and

bilateral presentations. Brain Lang 53:121–140 14.

Grindrod, C., & Baum, S. (2003). Sensitivity to local

sentence context information in lexical ambiguity

resolution: Evidence from left-right-hemisphere-

damaged individuals, Brain & Language, 85, 502-523.

Peleg, O., Eviatar, Z., Hazan, H, & Manevitz, L. (2007).

Differences and Interactions between Cerebral

Hemispheres When Processing Ambiguous

Homographs. Lecture Notes in Computer Science

(LNAI) publication no.4840, 'Attention in Cognitive

Systems', L. Paletta and E. Rome, Eds., pp. 367–380.

Weems, S. & Reggia, J. (2004). Hemispheric

specialization and independence for word recognition:

Acomparison of three computational models, Brain

and Language 89 554-568.

Monaghan, P. & Pollmann, S. (2003). Division of labour

between the hemispheres for complex but not simple

tasks: An implemented connectionist model. Journal

of Experimental Psychology: General, 132, 379-399.

Titone, D. A. 1998. Hemispheric differences in context

sensitivity during lexical ambiguity resolution. Brain

and Language, 65, 361-394.

Peleg, O., Giora, R., & Fein, O. (2004). Contextual

strength: The Whens and hows of context effects. In I.

Noveck & D. Sperber (Eds.), Experimental

Pragmatics (pp.172-186). Basingstoke: Pagrave.

Burgess, C., & Simpson, G. B. (1988). Cerebral

hemispheric mechanisms in the retrieval of ambiguous

word meanings. Brain and Language, 33 (pp. 86-103).

Kawamoto, A. H. (1993). Nonlinear dynamics in the

resolution of lexical ambiguity: A parallel distributed

processing account. Journal of Memory and

Language, 32, 474-516.

Zaidel, E., & Peters, A. M. (1981). phonological encoding

and ideographic reading by the disconnected right

hemisphere: Two case Studies. Brain & Language, 14,

205- 234.

Peleg O. and Eviatar Z (2009). Semantic asymmetries are

modulated by phonological asymmetries: Evidence

from the disambiguation of homophonic versus

heterophonic homographs. Brain and Cognition 70

(pp. 154–162).

Friston K. J., Harrison L. and Penny W. (2009). Dynamic

causal modelling. Neuroimage 2003 Aug;19(4) (pp.

1273-302).

Peleg, O., Manevitz, L., Hazan, H., Eviatar, Z., (2010).

Two Hemispheres - Two Networks A Computational

Model Explaining Hemispheric Asymmetries While

Reading Ambiguous Words. Annals of Mathematics

and Artificial Intelligence (AMAI). – to appear.

ICFC 2010 - International Conference on Fuzzy Computation

278