You are on page 1of 8

This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON CYBERNETICS 1

EEG-Based Classification of Implicit Intention


During Self-Relevant Sentence Reading
Suh-Yeon Dong, Bo-Kyeong Kim, and Soo-Young Lee

Abstract—From electroencephalography (EEG) data during We postulated that the implicit intentions of “agreement”
self-relevant sentence reading, we were able to discriminate two and “disagreement” elicit different brain signals, and a recent
implicit intentions: 1) “agreement” and 2) “disagreement” to the functional magnetic resonance imaging (fMRI) study showed
read sentence. To improve the classification accuracy, discrimi-
nant features were selected based on Fisher score among EEG different activation areas for the unexpressed intentions of
frequency bands and electrodes. Especially, the time-frequency agreement and disagreement [7]. In this paper, EEG signals are
representation with Morlet wavelet transforms showed clear dif- used to classify the agreement-versus-disagreement implicit
ferences in gamma, beta, and alpha band powers at frontocentral intention while reading a self-relevant sentence.
area, and theta band power at centroparietal area. The best clas- EEG is a useful technique because of its noninvasive nature
sification accuracy of 75.5% was obtained by a support vector
machine classifier with the gamma band features at frontocen- and high temporal resolution. In particular, EEG oscillatory
tral area. This result may enable a new intelligent user-interface responses are functionally related to cognitive processing [8].
which understands users’ implicit intention, i.e., unexpressed or Recently, changes in frequency band powers were reported
hidden intention. during language processing, specifically syntactic and seman-
Index Terms—Agreement/disagreement, electroencephalogra- tic processing at the sentence level. Increased power in
phy (EEG), implicit intention, self-relevance. the theta band was observed in response to grammatical
violations [9]–[11]. Most of the increase in the theta band
power was found at the temporal areas, whereas the power
I. I NTRODUCTION increase in the gamma band was also found at the frontal areas
VER the last 25 years, electroencephalography (EEG)- in sentence processing [12], [13]. Furthermore, the alpha band
O based brain–computer interfaces (BCIs) have been devel-
oped to provide assistance to physically impaired patients [1].
power decreased with semantic congruency [14], while the
beta band power decreased in response to semantic violations
Therefore, current BCI technologies had focused on rec- during reading Chinese sentences [15].
ognizing and responding to users’ explicitly expressed The EEG-based studies were further extended with self-
intentions [2]–[4]. On the other hand, we believe that intel- relevant sentences with stronger responses. The top-down
ligent user interface for common people should also be able attention to self-relevant information has been widely reported
to understand users’ unexpressed or hidden intention. We use in human perception. The most famous example of this phe-
the term implicit intention to denote this unexpressed or hidden nomenon is known as the “cocktail party effect” [16], whereby
intention. an individual’s auditory attention can be focused on a par-
However, surprisingly only a few research has been reported ticular stimulus (e.g., a conversation containing self-relevant
to recognize human implicit intention. Even all the previous items) while filtering out other stimuli (e.g., other conversa-
researches on the implicit intention focused on hidden inten- tions taking place nearby). More recently, based on both EEG
tion or lie detection, i.e., whether or not the user’s explicitly and fMRI data, many researchers have suggested that self-
expressed intention is the same as the actual intention [5], [6]. relevant information has preferential access to our perceptual
We focus on another type of the implicit intention, i.e., unex- systems. The exposure to one’s own name evokes stronger
pressed intention, especially whether or not a user agrees P300 and N250 responses than other stimuli [17]. Moreover,
with the others during conversation or sentence reading. it was reported that autobiographical words or phrases evoke
P300 responses [18], and that sentences written in the first per-
Manuscript received July 13, 2015; accepted September 9, 2015. This work son elicited early components (P1, N1, and P2) [19]. Also, one
was supported in part by the Brain Research Program through the National fMRI study compared brain responses between self-knowledge
Research Foundation of Korea funded by the Ministry of Science, ICT, and and semantic knowledge [20].
Future Planning under Grant 2013-035100 (2013.05.01-2014.04.30), and in
part by the U.S. Air Force Research Laboratory through the Asian Office However, the self-relevance may not be sufficient to excite
of Aerospace Research and Development. This paper was recommended by strong brain signals. If people do not agree with the state-
Associate Editor A. Cichocki. ment, the brain responses may not reflect the self-relevance.
The authors are with the Department of Electrical Engineering, Korea
Advanced Institute of Science and Technology, Daejeon 305-701, Even in cases with strong autobiographical words or phrases
Korea (e-mail: suhyeon.dong@gmail.com; kbghome@kaist.ac.kr; written in the first person, people may not consider them
sylee@kaist.ac.kr). self-relevant. Therefore, with respect to self-relevant stim-
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. uli, it is important to distinguish “agreed” responses from
Digital Object Identifier 10.1109/TCYB.2015.2479240 “disagreed” responses. However, there has been no reported
2168-2267 c 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON CYBERNETICS

C. Stimulating Statements
We aimed to discriminate EEG responses for agreement
and disagreement cases while reading self-relevant statements.
Self-relevant statements are those related to personal experi-
ences or opinions, and expected to generate stronger brain
signals than irrelevant sentences.
Seventy-four stimulating sentences were selected from the
Minnesota multiphasic personality inventory-II [26], one of
the most frequently used standardized psychometric tests.
All sentences were converted into Korean, a subject-object-
verb (SOV) language. Also, all sentences were ending with
Fig. 1. Experimental paradigm. One trial consisted of a fixation cross, main only one type of verb, i.e., existence verbs such as “do exist”
statement (contents), a positive or negative ending phrase (sentence ending), and “do not exist.” Therefore, an English sentence “I do (not)
and a subject’s response. For each subject, 74 trials were made in sequence
separated by a fixation cross. worry a great deal over money” was converted into an SOV
typology with a contents block “The experience of worrying
over money” and an ending block “Does (not) exist.” Here,
research to further discriminate “Yes” (or agreement) versus the negated sentence was formed by simply changing only the
“No” (or disagreement) responses on the self-relevant state- ending block from “Does exist” to “Does not exist.”
ments. Although a recent study investigated discriminating At the design stage of the experiment, we were not sure
brain responses between “Yes/No” thinking, which actually whether the positive and negative endings might result in dif-
came from a semantic violation [21]. ferent brain signals. Therefore, each sentence was decomposed
We had reported differences on fMRI [7], [22], [23] and into two blocks: 1) the contents block and 2) the ending block,
multichannel EEG signals [24], [25] between agreement and and EEG signals with the contents blocks and ending blocks
disagreement to self-relevant statements. This paper presents were investigated separately. We had two hypotheses to prove
a new time-frequency EEG analysis based on Morlet wavelet by this experiment. The first was that, as the main objective of
transform and discriminant electrode selection. this paper, the implicit intentions of agreement and disagree-
ment were elicited by reading only the contents block. The
II. M ETHOD AND M ATERIALS second was that the positive and negative ending blocks did
A. Subjects not make any difference.
Nine healthy right-handed Korean subjects (six males and
three females between 20 and 30 years old) were recruited D. EEG Recording and Preprocessing
for this paper. All were either undergraduate or graduate stu- The EEG data were recorded using the BrainAmp sys-
dents, and participated voluntarily. None of the subjects had tem (Brain Products GmbH, Germany) and an EEG cap
a history of mental disorder, significant physical illness, head with 32 electrodes (BrainCap) in an electromagnetically
injuries, neurological disorder, or alcohol or drug dependence. shielded room. Thirty electrodes were placed on each sub-
Written informed consent was obtained from all subjects, and ject’s scalp according to the international 10–20 system. One
this paper was approved by the institutional review board at electrode was positioned below subject’s left eye to record
the Korea Advanced Institute of Science and Technology. electrooculogram for eye movement. Another electrode was
placed above subject’s left collarbone to record electrocar-
B. Experimental Paradigm diogram. The impedance of each electrode was maintained
Each subject participated an experimental session of 74 tri- below 10 k.
als. As shown in Fig. 1, each trial started with a fixation cross Data were acquired at a sampling rate of 500 Hz with 60 Hz
and a beeping sound for the subject’s readiness, which was notch filter. The EEG data were measured with a reference at
followed by two images of a stimulating sentence to read. FCz electrode, but later converted into values with the aver-
Each statement consisted of a contents block and a sentence- age reference. High-pass filtering was implemented with 1 Hz
ending block. (The idea behind this division will be explained cutoff frequency and 0.2 Hz transition bandwidth. Independent
in Section II-C.) The contents block was shown for 4 s and the component analysis (ICA) based on extended Infomax algo-
sentence-ending block was shown for 2 s. Immediately after- rithm was used to identify artifacts. Then, artifacts from
wards, an asterisk image was presented for 2 s, and subjects eye blinking, eye movement, and heartbeats were manually
were instructed to answer whether or not they agreed with the selected and removed from the measured data [27]–[29].
statement. A blank image was presented for 2 s at the end of Since EEG data elicited by the contents block were of
each trial. our concern, the time reference was moved to the onset
During the experiment, subjects were asked to remain still of the contents block for each trial. Then, EEG data from
and to avoid eye blinking as much as possible, especially just −0.2 to 1 s were extracted, and the baseline correction was
after the onset of the contents block. Before, during, and after adopted by subtracting the mean amplitude within the range of
the experiment, the subjects were carefully monitored to make [−0.2 s, 0 s]. In Fig. 2, averages of the resulting EEG data, i.e.,
sure of their good performance without any difficulty. event-related potential (ERP), were shown for the agreement
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

DONG et al.: EEG-BASED CLASSIFICATION OF IMPLICIT INTENTION DURING SELF-RELEVANT SENTENCE READING 3

TABLE I
N UMBER OF S TATEMENTS W ITH U NWILLINGNESS
TO R ESPOND H ONESTLY

Fig. 2. Average ERPs for a representative electrode site (Cz) of subject 4 for
the agreement condition (black-dashed line) and the disagreement condition
(charcoal-gray solid line). ERPs were time-locked to the onset of the contents
block (vertical-dotted line). TABLE II
N UMBER OF T RIALS IN AGREEMENT AND
D ISAGREEMENT C LASSES

and disagreement trials. The disagreement ERP showed higher


amplitudes at two time intervals, i.e., [250 ms, 400 ms] and
[750 ms, 850 ms].

E. Morlet Wavelets for Time-Frequency Representation


To classify the implicit intentions into two classes from each
single-trial EEG, instead of the averaged ERP, frequency band
powers of event-related oscillatory responses were estimated
for each trial.
There are several methods for extracting time-frequency
joint representation from time-dependent ERP data. The spec-
trogram based on fast Fourier transform of time-framed
data is the popular choice, but the time and frequency resolu- probability of discrepancy between the implicit and explicit
tions become the same throughout the whole time-frequency intentions was estimated as 0.055(= (3 × 6 + 2 × 13 + 1 × 9)/
region. To obtain proper time and frequency resolutions for (13 × 74)). This probability may provide a limitation on the
each region, Morlet wavelet transform was adopted [30]. The final experimental accuracy, but may not be significant to
mother wavelet was designed with a seven-cycle width, and our binary classification accuracies. However, every efforts
scaled for center frequencies from 1 to 70 Hz. In this man- were made to assure anonymity of subjects throughout the
ner, a time-frequency representation (TFR) was obtained for experiments for reliable and honesty responses.
each trial and electrode. The power values in each TFR were The numbers of good trials for the agreement and disagree-
converted into percentage changes relative to the power in the ment classes for each subject are shown in Table II. Although
baseline interval between −0.2 and 0 s prior to the onset of 74 trials were given to all subjects, subject 5 missed one
the contents. trial and subject 6 misunderstood the task during the first few
trials. The numbers of trials become unbalanced between the
III. DATA A NALYSIS AND R ESULTS two classes. Therefore, the overall classification accuracy was
defined as an average classification rates of the two classes.
A. Labeling of Explicit Intention
In accordance with the user responses, the recorded EEG
for each trial was labeled as agreement or disagreement. B. Feature Selection: Time-Frequency Components
Because some sentences contain contents related to personal To understand the neural mechanism as well as to improve
privacy, this might cause a discrepancy between implicit and the classification accuracy with small number of available
explicit answers. The social-desirability bias might cause more data, feature selection was implemented in two steps. At the
answers of “to-be-viewed favorably by others” [31]. first step, several time-frequency components were selected
To test the validity of tagging implicit intentions with with high discrimination capacity between the agreement and
explicit answers, a survey was conducted for 13 subjects disagreement classes.
(seven males and six females between 20 and 30 years old), For this purpose, based on previous fMRI studies with sim-
who did not participate in the EEG experiment. For each of ilar experiments [7], five frontal electrodes close to superior
the 74 statements used in the EEG experiments, the subjects frontal gyrus and anterior cingulate cortex were considered
were asked to select either “be unwilling to respond honestly” as shown in Fig. 3(a). These areas were known to be related
or “do not care.” to decision-making [32], empathic judgments [33], and self-
Table I summarizes the numbers of subjects who were descriptive trait judgment [34].
unwilling to respond honestly for each statement. For six state- The TFRs based on Morlet wavelet transforms were aver-
ments, three subjects answered that they were unwilling to aged for all subjects, with agreement and disagreement classes
respond honestly. There were 13 and 9 statements, respec- separately, at the five electrodes. Then, the difference of
tively, to which two subjects and one subject showed their percentage changes between the two classes was plotted in
unwillingness to respond honestly. Using these results, the Fig. 3(b).
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON CYBERNETICS

Fig. 3. Average difference in TFRs between agreement and disagreement classes for five frontal electrodes. (a) Five considered frontal electrodes, i.e., F3, Fz,
F4, FC1, and FC2, in yellow color. (b) Percentage differences (agreement–disagreement) of Morlet wavelet-based TFRs. The x-axis and the y-axis represent
time and frequency, respectively. Boxes within the TFRs indicate the time-frequency components with significant differences. (c) Scalp topographic maps of
the five selected time-frequency components.

TABLE III
SNR OF E ACH T IME -F REQUENCY C OMPONENT

With a careful inspection for significant percentage changes


in Fig. 3(b), five time-frequency components were selected
as: 1) the gamma component (35–45 Hz) between 350 and
550 ms; 2) the beta2 component (20–26 Hz) between 300
and 450 ms; 3) the beta1 component (14–17 Hz) between
800 and 1000 ms; 4) the alpha component (9–12 Hz)
between 300 and 700 ms; and 5) the theta component (5–7 Hz) Fig. 4. Electrode (or channel) selection procedure for each time-frequency
between 400 and 1000 ms after the onset of the contents. component. Fisher scores were computed for all electrodes, and sorted in
For each time-frequency component, the signal-to-noise descending order. (a) Compute Fisher score for each channel. (b) Channel
selection based on Fisher score.
ratios (SNRs) were shown in Table III. The SNRs were calcu-
lated for each time-frequency component, but averaged over
subjects, trials, and electrodes. The noise power was estimated
frontocentral area. However, when subjects agreed, beta2 and
at the baseline time period [−0.2 s, 0 s], and the signal power
theta band activities increased at frontal area. Moreover, at
was estimated by subtracting the noise power from the total
low frequencies such as alpha and theta bands, some parietal
power. Due to the low SNR of entire time-frequency region, it
and occipital regions also showed large differences between
is advantageous to consider the specific time-frequency region.
agreement and disagreement.
Although the gamma component has negative SNR, it is still
discriminant in Fig. 3(b).
In Fig. 3(c), scalp topographic maps were shown for the five C. Feature Selection: Electrode Positions
selected time-frequency components using the average power For each of the five chosen time-frequency components, the
for each component [35]. When subjects disagreed, the aver- importance of each electrode positions were evaluated, and
age power increased in gamma, beta1, and alpha bands around only a subset was selected for efficient classification. As shown
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

DONG et al.: EEG-BASED CLASSIFICATION OF IMPLICIT INTENTION DURING SELF-RELEVANT SENTENCE READING 5

TABLE IV
F IVE B EST E LECTRODES FOR E ACH T IME -F REQUENCY C OMPONENT BASED ON F ISHER S CORE

in Fig. 3(c), there existed some high activation differences


outside of the frontal areas, and all the 30 electrode positions
were considered.
The Fisher score is one of the most widely used measures
of feature discriminant ability [36]. For each time-frequency
component, the Fisher score for the ith electrode (or channel)
is computed as
K  i 
i 2
k=1 nk μk − μ
Fi = K  i 2
k=1 nk σk

where nk is the number of trials of the kth class, and μik and
σki denote the mean and standard deviation for the kth class at
the ith electrode, respectively. Here, K = 2 is the number of
classes. Also, μi denotes the mean for the entire data at the ith
electrode. After computing the Fisher score for all electrodes,
top M ranked electrodes were selected and used as the inputs
to a classifier. The electrode selection procedure for each time-
frequency component is summarized in Fig. 4.
Table IV summarizes the five best electrodes for each time-
frequency component. For the gamma component, the selected
electrodes were located at the frontocentral regions (F3 and
FC5 on the left and FC2 on the right), the right tempo-
ral region (T8), and the left centroparietal region (CP5). For
the beta2 component, the selected electrodes were located
at the left central regions (C3 and CP5) and frontal regions
(FC1, Fp1, and Fp2). For the beta1 component, the selected
electrodes were located at the frontocentral regions (F3 and
FC1 on the left, and F4 on the right), the right tempo-
ral region (T8), and the left parietal region (P7). For the
alpha component, the selected electrodes were located at the
frontocentral regions (Fz, FC1, and F4) and left central regions
(C3 and Cp1). For the theta component, the selected electrodes
were located at the centroparietal regions (C3 and CP5 on
the left and CP2 on the right) and the left parietal region
(P3 and P7).
For the gamma to alpha components, the frontocentral elec- Fig. 5. Training (the top plot) and testing (the bottom plot) accuracies for
trodes were found to contain the discriminating features. By each time-frequency component with increasing number of selected electrodes
using (a) linear kernel and (b) RBF kernel for binary SVM classifiers. The x-
contrast, the theta component of the centroparietal electrodes axis represents the number of selected electrodes, and the y-axis represents the
was more discriminant. Moreover, the gamma component average classification accuracy (%). Accuracies were averaged with fivefold
had higher Fisher scores, and therefore higher classification cross-validations for each subject, and then averaged again for all nine sub-
jects. The line colors darken as the band frequency increases. The red-dotted
performance than the other components. line indicates the chance level, i.e., 50%.

D. Single-Trial Classification Results EEG data. Simple linear discriminant analysis (LDA) was
Our goal was to classify users’ implicit intention into first applied, but the results were not satisfactory. Therefore,
agreement and disagreement binary classes from single-trial support vector machine (SVM) was applied with two types
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON CYBERNETICS

TABLE V
M AXIMUM T ESTING ACCURACY (%) FOR E ACH C OMPONENT AND
C LASSIFIER . PARENTHESIS I NDICATES THE N UMBER OF
S ELECTED C HANNELS R EQUIRED TO M AKE THE
M AXIMUM T ESTING ACCURACY

obtained at 74.0% with four electrodes from gamma, beta1,


and beta2 components. However, the linear kernel required
20 electrodes to reach its highest accuracy at 64.7%. The rel-
atively poorer performance with all five components may come
Fig. 6. Training (the top plot) and testing (the bottom plot) accuracies
from overfitting.
using all five time-frequency components with increasing number of selected Table V summarizes the maximum single-trial testing accu-
electrodes using a linear and RBF kernel for SVM classifiers. The x-axis repre- racy and the number of electrodes for each time-frequency
sents the number of selected electrodes, and the y-axis represents the average
classification accuracy (%). Accuracies were averaged with fivefold cross-
component and with both kernel types for the SVM classifiers.
validations for each subject, and then averaged again for all nine subjects. The gamma component had higher Fisher scores than the other
The red-dotted line indicates the chance level, i.e., 50%. components, which resulted in a more accurate classification
performance.

of kernel functions. The LIBSVM toolbox was used in [37]. IV. D ISCUSSION
Considering the fact that LDA did not work well, the two In this paper, using single-trial multichannel EEG data, we
classes might not be linearly separable, and radial basis func- proposed a new method to classify human implicit intentions
tion (RBF) kernel was expected to perform better. The optimal into agreement and disagreement classes.
values for the SVM parameters were found through an exhaus-
tive grid search during the validation phase [38]. Due to the A. Concept of Implicit Intention
small data available, fivefold cross-validation was performed
The term “implicit intention” in this paper refers to the
using the test data for the validation.
unexpressed or hidden intention. Especially, we had focused
With the electrode selection in the previous section, the clas-
on the agreement versus disagreement classes among the
sification results were calculated as functions of the number
unexpressed intentions. On the other hand, a few others had
of selected electrodes, for each time-frequency component and
reported lie-detection, which is related to the hidden intention.
for all components.
The implicit agreement (or disagreement) to a given statement
Fig. 5 shows the training and testing accuracies versus the
may be regarded as the agreement (or disagreement) with
number of selected electrodes for SVM classifiers with linear
a counterpart in conversation. On the other hand, lie detec-
and RBF kernels. As the number of electrodes increased, the
tion is concerned with agreement (or disagreement) to ones
training accuracy improved for both SVMs. The testing accu-
expressed intention. While a vast body of research exists on
racies for most components also increased as more electrodes
lie detection, our concept of implicit intention on agreement
were used with the linear kernel, whereas the testing accu-
versus disagreement is novel.
racy for the gamma component decreased when more than
20 electrodes were used. By contrast, with the RBF kernel,
the testing accuracy started to decrease from smaller num- B. Experiment Paradigm
ber of electrodes. Actually, the gamma component with five In most experiments concerning lie detection, subjects are
most discriminant electrodes showed the best performance, asked to tell a lie in order to create a situation where neural
at 75.5% (for these five electrodes of the gamma component, activities related to lying may be assessed. However, there may
see Table IV). Over-fitting on the training data might be the be differences between this artificially generated and truly nat-
reason. ural cases. Our experiment was constructed to elicit subjects’
Furthermore, all the five time-frequency components were natural and reflexive intentions on self-relevant statements.
used for the classification. The 150 Fisher score values from The self-relevance had been reported to generate pronounced
5 components and 30 electrodes were sorted, and selected from intention of agreement or disagreement and stronger neural
the highest Fisher score. Fig. 6 shows the training and testing activities. Although Ruf et al. [21] had a similar experimental
classification rates (%) for both linear and RBF SVMs. Using paradigm, their yes thinking and no thinking were generated
the RBF kernel, the maximum test classification rate was by semantic violation. Our single-trial experiments yielded
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

DONG et al.: EEG-BASED CLASSIFICATION OF IMPLICIT INTENTION DURING SELF-RELEVANT SENTENCE READING 7

a superior classification performance in terms of classification human implicit intentions of agreement versus disagreement
accuracy, which probably came from the strong and profound were classified with 75.5% accuracy from single-trial EEG
neural signals to the self-relevant statements. signals. Also, it was shown that the gamma-band power from
Our experiment was designed specifically for Korean speak- 350 to 550 ms after the onset of stating sentence played an
ers, and took advantage of the SOV Korean grammar. The important role for the classification.
stating sentence was divided into two parts: 1) sentence con- These results suggest that different neuro-cognitive pro-
tents and 2) ending. The sentence ending part may be negated, cesses are involved in the two implicit intentions, i.e., agree-
ex., “does not.” It was shown that the neural activities related ment versus disagreement, when people read self-relevant
to the implicit intentions of agreement and disagreement were sentences. It may prove our hypothesis that agreement and
generated by the sentence contents. Also, the careful analy- disagreement consist of an axis on human implicit inten-
sis of EEG data at the sentence ending block showed that the tion space. Therefore, for a new intelligent human–machine
positive “does” and negative “does not” endings did not make interface, one may use EEG signals to generate labels for
any noticeable difference in our preliminary studies. It demon- the audio-visual training database for the classification of
strated that a user’s implicit intention is determined before the agreement versus disagreement during conversation. The other
sentence ending. For English speakers, the experimental design alternative may be self-report, which is impossible to be
may be modified to merge the contents and ending blocks with obtained in real-time and may not be reliable for sensitive per-
only positive verbs. sonal issues. Therefore, although the EEG-based labels may
include some error, it may be utilized at early developmental
C. Feature Selection in Time-Frequency Space Domain stage. Also, the accuracy may be further improved by online
incremental learning.
Unlike many other researches using ERP in time domain or
band powers in frequency domain, time-frequency joint repre-
sentation was adopted in this paper. Since the ERP and band R EFERENCES
powers may be obtained by integrating the time-frequency [1] J. R. Wolpaw, N. Birbaumer, D. J. McFarland, G. Pfurtscheller, and
spectrum in frequency and time domain, respectively, the TFR T. M. Vaughan, “Brain-computer interfaces for communication and
is more informative. Also, to obtain appropriate time and fre- control,” Clin. Neurophysiol., vol. 113, no. 6, pp. 767–791, Jun. 2002.
[2] B. Blankertz et al., “The Berlin brain-computer interface: EEG-based
quency resolutions throughout the whole time and frequency communication without subject training,” IEEE Trans. Neural Syst.
domain, Morlet wavelet transform was implemented. Rehabil. Eng., vol. 14, no. 2, pp. 147–152, Jun. 2006.
As a result, five time-frequency components were identified [3] D. Coyle, G. Prasad, and T. M. McGinnity, “Faster self-organizing fuzzy
to be discriminative features between the agreement and dis- neural network training and a hyperparameter analysis for a brain–
computer interface,” IEEE Trans. Syst., Man, Cybern. B, Cybern.,
agreement classes. The gamma and theta components within vol. 39, no. 6, pp. 1458–1471, Dec. 2009.
a similar latency range were also observed in the general- [4] C. Escolano, J. M. Antelis, and J. Minguez, “A telepresence mobile robot
knowledge violation condition for sentence comprehension controlled with a noninvasive brain–computer interface,” IEEE Trans.
Syst., Man, Cybern. B, Cybern., vol. 42, no. 3, pp. 793–804, Jun. 2012.
in [12] and [13]. In the context of agreement versus dis- [5] L. A. Farwell and E. Donchin, “The truth will out: Interrogative
agreement intention, the general-knowledge violation may be polygraphy (“lie detection”) with event−related brain potentials,”
regarded as a disagreement with the propositional statements. Psychophysiology, vol. 28, no. 5, pp. 531–547, Sep. 1991.
[6] C. Davatzikos et al., “Classifying spatial patterns of brain activity with
Also, the important electrodes were found in the same brain machine learning methods: Application to lie detection,” NeuroImage,
areas, i.e., the frontocentral areas for the gamma component. vol. 28, no. 3, pp. 663–668, Nov. 2005.
This similarity may come from the experimental design of [7] S.-Y. Dong, B.-K. Kim, and S.-Y. Lee, “Implicit agreeing/disagreeing
intention while reading self-relevant sentences: A human fMRI
agreement versus disagreement cases. However, the beta com- study,” Soc. Neurosci. [Online]. Available: http://dx.doi.org/10.1080/
ponent was not found in the knowledge violation experiments. 17470919.2015.1059362
This difference may result from the self-relevance versus [8] E. Başar, M. Schürmann, T. Demiralp, C. Başar-Eroglu, and
A. Ademoglu, “Event-related oscillations are ‘real brain responses’—
general knowledge of the stating sentences. Wavelet analysis and new strategies,” Int. J. Psychophysiol., vol. 39,
However, there is a drawback to our feature selection nos. 2–3, pp. 91–127, Jan. 2001.
method. The Fisher score was evaluated separately for each [9] M. C. M. Bastiaansen, J. J. A. van Berkum, and P. Hagoort, “Event-
related theta power increases in the human EEG during Online sentence
electrode, while some combination of time, frequency, and processing,” Neurosci. Lett., vol. 323, no. 1, pp. 13–16, Apr. 2002.
electrode position may contribute to the classification. One [10] D. Roehm, M. Schleschewsky, I. Bornkessel, S. Frisch, and H. Haider,
may try every possible feature combinations. Or, before feed- “Fractionating language comprehension via frequency characteristics of
the human EEG,” Neuroreport, vol. 15, no. 3, pp. 409–412, Mar. 2004.
ing to the SVM classifier, one may remove the dependency
[11] D. J. Davidson and P. Indefrey, “An inverse relation between event-
among features by using ICA. However, considering that related and time–frequency violation responses in sentence processing,”
an overfitting occurred with only five features and all pos- Brain Res., vol. 1158, pp. 81–92, Jul. 2007.
sible 150 features did not improve the testing accuracy in [12] P. Hagoort, L. Hald, M. Bastiaansen, and K. M. Petersson, “Integration
of word meaning and world knowledge in language comprehension,”
Fig. 6, no significant improvement is expected from the 75.5% Science, vol. 304, no. 5669, pp. 438–441, Apr. 2004.
classification rate with 5 gamma-band features. [13] L. A. Hald, M. C. Bastiaansen, and P. Hagoort, “EEG theta and
gamma responses to semantic violations in Online sentence processing,”
Brain Lang., vol. 96, no. 1, pp. 90–105, Jan. 2006.
V. C ONCLUSION [14] R. M. Willems, R. Oostenveld, and P. Hagoort, “Early decreases in
alpha and gamma band power distinguish linguistic from visual infor-
In this paper, by applying Morlet-wavelet time-frequency mation during spoken sentence comprehension,” Brain Res., vol. 1219,
analysis to multichannel EEG data, it was demonstrated that pp. 78–90, Jul. 2008.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON CYBERNETICS

[15] Y. Luo, Y. Zhang, X. Feng, and X. Zhou, “Electroencephalogram oscil- [37] C. C. Chang and C. J. Lin, “LIBSVM: A library for support vector
lations differentiate semantic and prosodic processes during sentence machines,” ACM Trans. Intell. Syst. Technol., vol. 2, no. 3, Apr. 2011.
reading,” Neuroscience, vol. 169, no. 2, pp. 654–664, Aug. 2010. [Online]. Available: http://www.csie.ntu.edu.tw/~cjlin/libsvm
[16] E. C. Cherry, “Some experiments on the recognition of speech, with one [38] C. W. Hsu, C. C. Chang, and C. J. Lin (Apr. 2010). A Practical
and with two ears,” J. Acoust. Soc. Am., vol. 25, no. 5, pp. 975–979, Guide to Support Vector Classification. [Online]. Available:
Sep. 1953. http://www.csie.ntu.edu.tw/~cjlin/papers/guide/guide.pdf
[17] K. Zhao, Q. Wu, H. D. Zimmer, and X. Fu, “Electrophysiological
correlates of visually processing subject’s own name,” Neurosci. Lett.,
vol. 491, no. 2, pp. 143–147, Mar. 2011.
[18] H. M. Gray, N. Ambady, W. T. Lowenthal, and P. Deldin, “P300 as an
index of attention to self-relevant stimuli,” J. Exp. Soc. Psychol., vol. 40,
Suh-Yeon Dong received the B.S. and M.S.
no. 2, pp. 216–224, Mar. 2004.
degrees from the Department of Electrical
[19] E. C. Fields and G. R. Kuperberg, “It’s all about you: An ERP study
Engineering, Korea Advanced Institute of Science
of emotion and self-relevance in discourse,” NeuroImage, vol. 62, no. 1,
and Technology, Daejeon, Korea, in 2010 and 2011,
pp. 562–574, Aug. 2012.
respectively, where she is currently pursuing the
[20] S. C. Johnson et al., “Neural correlates of self-reflection,” Brain,
Ph.D. degree in electrical engineering.
vol. 125, no. 8, pp. 1808–1814, Aug. 2002.
Her current research interests include brain
[21] C. A. Ruf et al., “Semantic classical conditioning and brain–computer
signal processing, implicit human intention, and the
interface control: Encoding of affirmative and negative thinking,” Front.
brain–computer interface.
Neurosci., vol. 7, no. 23, pp. 1–13, Mar. 2013.
[22] S.-Y. Dong and S.-Y. Lee, “Recognition of human implicit intention
based on fMRI and EEG,” presented at Neuroinformat., Stockholm,
Sweden, 2013. [Online]. Available: http://www.frontiersin.org/10.3389/
conf.fninf.2013.09.00044/event_abstract
[23] S.-Y. Dong and S.-Y. Lee, “Understanding human implicit intention
while reading self-relevant sentences: An fMRI study,” in presented at Bo-Kyeong Kim received the B.S. degree from
the 19th Annu. Meeting OHBM, Seattle, WA, USA, Jun. 2013. the Department of Bio and Brain Engineering,
[24] S.-Y. Dong and S.-Y. Lee, “Understanding human implicit intention Korea Advanced Institute of Science and
based on frontal electroencephalography (EEG),” in Proc. IJCNN, Technology, Daejeon, Korea, in 2011, where she
Brisbane, QLD, Australia, Jun. 2012, pp. 1–5. is currently pursuing the Ph.D. degree in electrical
[25] S.-Y. Dong, B.-K. Kim, and S.-Y. Lee, “Decoding and predicting engineering.
implicit agreeing/disagreeing intention based on electroencephalogra- Her current research interests include single
phy (EEG),” in Neural Information Processing (LNCS 8227). Berlin, EEG trial classification, feature extraction of brain
Germany: Springer, 2013, pp. 587–594. signals, and brain–computer interface.
[26] J. N. Butcher, W. G. Dahlstrom, J. R. Graham, A. Tellegen,
and B. Kaemmer, The Minnesota Multiphasic Personality
Inventory-2 (MMPI-2): Manual for Administration and Scoring.
Minneapolis, MN, USA: University of Minnesota Press, 1989.
[27] A. Delorme, J. Palmer, J. Onton, R. Oostenveld, and S. Makeig,
“Independent EEG sources are dipolar,” PLoS One, vol. 7, Feb. 2012,
Art. ID e30135. Soo-Young Lee received the B.S. degree from Seoul
[28] T. W. Lee, M. Girolami, A. J. Bell, and T. J. Sejnowski, “A unifying National University, Seoul, Korea, in 1975, and
information-theoretic framework for independent component analysis,” the Ph.D. degree from the Polytechnic Institute of
Comput. Math. Appl., vol. 39, pp. 1–21, Jun. 2000. New York University, Brooklyn, NY, USA, in 1984.
[29] A. Delorme and S. Makeig, “EEGLAB: An open source toolbox for He is currently a Full Professor with the
analysis of single-trial EEG dynamics including independent component Department of Electrical Engineering and Computer
analysis,” J. Neurosci. Meth., vol. 134, no. 1, pp. 9–21, Mar. 2004. Science, Korea Advanced Institute of Science and
[30] A. Grossmann and J. Morlet, “Decomposition of hardy functions into Technology, Daejeon, Korea. He has researched on
square integrable wavelets of constant shape,” SIAM J. Math. Anal., auditory models, information-theoretic processing,
vol. 15, no. 4, pp. 723–736, 1984. proactive knowledge development, and top-down
[31] R. J. Fisher, “Social desirability bias and the validity of indirect selective attention. His current research interests
questioning,” J. Consum. Res., vol. 20, no. 2, pp. 303–315, Sep. 1993. include artificial brains such as human-like intelligent systems based on
[32] C. S. Carter et al., “Anterior cingulate cortex, error detection, and biological information processing mechanisms, and combining cognitive neu-
the online monitoring of performance,” Science, vol. 280, no. 5364, roscience and information theory for artificial cognitive systems.
pp. 747–749, May 1998. Prof. Lee was a recipient of the International Neural Network
[33] T. F. Farrow et al., “Investigating the functional anatomy of empathy and Society (INNS) Leadership Award in 1994, the INNS Presidential Award in
forgiveness,” Neuroreport, vol. 12, no. 11, pp. 2433–2438, Aug. 2001. 2001, and the APPNA Service Award in 2004. He joined the INNS Governing
[34] I. I. Goldberg, M. Harel, and R. Malach, “When the brain loses its self: Board in 2012. He was the President of the Asia-Pacific Neural Network
Prefrontal inactivation during sensorimotor processing,” Neuron, vol. 50, Assembly. He is the Editor-in-Chief of Natural Intelligence and the INNS
no. 2, pp. 329–339, Apr. 2006. Magazine, and is on Editorial Boards of several other journals. In 1997, he
[35] Z. J. Koles, “The quantitative extraction and topographic mapping of established the Brain Science Research Center, and from 1998 to 2008 he
the abnormal components in the clinical EEG,” Electroen. Clin. Neuro., served as the Director and Principal Investigator of the Brain Neuroinformatics
vol. 79, no. 6, pp. 440–447, Dec. 1991. Research Program, the first interdisciplinary research program in Korea for
[36] R. Duda, P. Hart, and D. Stork, Pattern Classification. New York, NY, brain-inspired intelligent systems with perception, learning, inference, and
USA: Wiley, 2001. human-like behavior.

You might also like