A Large Scale Knowledge Base Representing the Base Form of Kaomoji
Noriyuki Okumura
Department of Electrical and Computer Engineering, National Institute of Technology, Akashi College,
679-3 Nishioka, Uozumi-cho, Akashi-City, Hyogo-Prefecture, Japan
Keywords:
Kaomoji, Emoticon, Original Form, N-gram, Kaomoji’s Dictionary, Annotation.
Abstract:
In this paper, we construct a large-scale knowledge base representing the base form of kaomoji (emoticon)
and other elements of kaomoji: eye, nose, mouth, and so on, to analyze features of kaomoji in detail. Previous
methods to analyze kaomoji mainly aim to extract kaomoji from sentences, paragraphs, or documents, or to
classify kaomoji into some emotion classes based on the emotion that kaomoji shows or potentially includes.
We define the base form of kaomoji for detailed kaomoji analytics. Application systems can estimate another
feature of derivative kaomoji based on its base form and other elements for sentiment analytics, emotion
extraction, or kaomoji classification. We annotated about 40,000 kinds of kaomoji for constructing a large-
scale knowledge base. The total number of extracted base forms is about 3,000. In experimental evaluations
based on cosine similarity using N-gram based features and simple Skip-gram based features, we show that
the model can estimate the base form of kaomoji with an accuracy of about 50%.
1 INTRODUCTION
How can we understand our real intentions, emo-
tions or feelings each other using only written words
in computer-mediated media such as Twitter
1
, Face-
book
2
or something like them? Can we send our ges-
tures using only characters? One of the solutions is
kaomoji. Kaomoji is a typical sequence that can send
writer’s expression, sign, emotion, and so on to the
reader with similar uses as an emoticon, a pictogram,
a smiley, and a stamp using only written words. Kao-
moji is widely used especially in Japanese culture. We
observed over 100,000 kinds of kaomoji in the web
pages written in the Japanese language.
Why kaomoji come into widespread use as com-
municating writers emotions, gestures and so on in
Japan? Because Japanese culture is the most high-
context culture all over the world. People living in
the high-context culture can communicate with each
other without expressing their opinions clearly on
a face-to-face basis because speakers and listeners
have a cultural context in common(Hall, 1976). On
the other hand, people living in the low-context cul-
ture (for example, English-speaking people, German-
speaking people, and so on) communicate with each
other using a clear explanation because they do not
1
https://twitter.com
2
http://www.facebook.com
have a cultural context in common. It is hard that
people do communication with each other at the
character-based communication on the internet when
using high-context language such as Japanese. People
can estimate the intension of speaker’s utterance with
observing speaker’s expression, gestures, and accent
at the face-to-face communication. Character-based
communication such as internet services has the po-
tential to make readers misunderstand their real in-
tention because of the limited information to express
their intention. Because of these situations, kaomoji
come into widespread use as communicating writer’s
implicit messages in the character-based communica-
tion.
kaomoji do exist over 100,000 kinds of their vari-
ations as previously noted. Nowadays, the number
of kaomoji increase day by day. The preferred sit-
uation is to keep renewing a large-scale knowledge
base kaomoji such as the proposed knowledge base
in this paper, but it is distant idea because construct-
ing large-scale knowledge base needs manual annota-
tions. Therefore, we need the method that estimates
the meaning that kaomoji expresses by extracting its
base form
3
such as lemmatization of verbs and the
other elements of it. For example, we can extract pos-
3
In this paper, we use this words (the base form) for kao-
moji despite the existence of “the original form” or “infini-
tive”
Okumura N.
A Large Scale Knowledge Base Representing the Base Form of Kaomoji.
DOI: 10.5220/0006517002460252
In Proceedings of the 9th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (KEOD 2017), pages 246-252
ISBN: 978-989-758-272-1
Copyright
c
2017 by SCITEPRESS Science and Technology Publications, Lda. All r ights reserved
Table 1: Rules and regulations to extract/make the base form of kaomoji.
kaomoji Original Form Rules and regulations
ˆ) (ˆˆ)
Remove all spaces
\
ˆ)/ ˆ)
Remove parts without borders, eyes, mouth and nose
ˆ)
Transform characters from two-byte characters to one-byte characters
(’‘) ()
If there are no corresponding one-byte characters, then the characters
remain in two-byte characters (in this case, the object is ω)
(T T)/˜˜˜ (T T)
Do not transform characters from uppercase/lowercase to lower-
case/uppercase (for instance, do not transform
(T T)
to
(t t)
)
(
>
-) (
> <
)
,
(- -)
If kaomoji is asymmetry, then make it symmetry
(;) (;;)
If the eyes or some parts are covered with arms or something, then com-
plement them using the other side of objective parts
ˆ; ˆ)
If kaomoji have odd border, then complement it using the other side of
border
( (;) ()
,
()
If kaomoji have multiple faces, then extract each original form
)ˆˆ) (ˆˆ)
If borders are asymmetry, then make border’s curve to outside
itive sentiment from kaomoji such as
(ˆ_ˆ)
because
we usually use this face as “smiling. If we attach the
new element (
*
) to this kaomoji, we transform
(ˆ_ˆ)
into
(*ˆ_ˆ*)
.
(*ˆ_ˆ*)
has more positive sentiment
than
(ˆ_ˆ)
. In fact, (
*
) has the role to emphasize
positive sentiment. By the same token, we can extract
negative sentiment from kaomoji such as
()
because
we usually use this face as “anger. If we attach the
new element (*) to this kaomoji, we transform
()
into
(**)
.
(**)
has more negative sentiment than
()
.
In fact, (
*
) has the role to emphasize positive/negative
sentiment. We can guess the sentiment of
(*ˆ_ˆ*)
or
(**)
using the sentiment information of
(ˆ_ˆ)
or
()
because we know the role of (
*
) whether we have or
do not have any information of
(*ˆ_ˆ*)
or
(**)
.
In this paper, we annotate about 40,000 kinds of
kaomoji collecting from websites
4
. We extract about
3,000 kinds of the base form of kaomoji using anno-
tated kaomoji. We also investigate whether N-gram
based model can estimate the base form of kaomoji.
In experimental evaluations based on cosine similar-
ity using N-gram based features and simple Skip-
gram based features, we show that the model can es-
timate the base form of kaomoji’ with an accuracy of
about 50%.
2 RELATED WORK
The research dealing with kaomoji is come across oc-
casionally. The research, however, that focuses on
the relation: the base form - the derivative form as
our study is in the minority. Ptaszynski et al. pro-
posed CAO: A Fully Automatic Emoticon Analysis
System to achieve kaomoji extraction with high ac-
curacy(Ptaszynski et al., 2010b; Ptaszynski et al.,
4
For example:http://www.kaomoji.sakura.ne.jp/
2010a). CAO system can manage over 10,000 kinds
of kaomoji, and the system can extract over a few mil-
lion of kaomoji. On the other hand, they do not focus
on kaomojis original form. CAO system only extracts
kaomoji from sentences using eye-mouth-eye triplet
in this scheme.
Bedrick et al. constructed a robust method to ex-
tract kaomoji from tweets on Twitter(Bedrick et al.,
2012). They proposed the extraction method of kao-
moji in consideration of symmetric property using
Probabilistic context-free grammar (PCFG). In our
study, we attach great importance to kaomojis sym-
metric property as their research.
Yamada et al. constructed the classification sys-
tem for about 700 kaomoji using N-gram based fea-
ture extracted each kaomoji(Yamada et al., 2007).
This system classifies kaomoji to 8 emotions (smile,
cry, anger, surprise, confuse, unsatisfied, anxiety, no
emotion) with an accuracy of 90%. They, however,
focus on only global classification without putting
each element of kaomoji to the proof.
Tanaka et al. reported the classification system us-
ing K-means and Support Vector Machines (SVMs)
that classify kaomoji to 6 emotions (pleasure, sadness,
anger, surprise, action, bitter smile) in the similar way
of Yamada’s research(Tanaka et al., 2005). These re-
searches mainly focus on emotions that kaomoji have,
but they do not use each element of kaomoji.
Kazama et al. construct a method to extract
kaomoij from Twitter
5
(Kazama et al., 2016). Their
method to extract kaomoji focuses on Unicode blocks
and Unicode character properties. The method can
extract over 50 million kinds of kaomoji from Twitter
(360 million tweets.) The method, however, cannot
5
This article is submitted to a Japanese conference and
it is written in Japanese, however, their extraction method
of kaomoji marked high accuracy. We dare to introduce this
article here.
deal with kaomoji that have comments such as
(**)
.
6
The other research about kaomoji are the recom-
mendation method of kaomoji(Urabe et al., 2013),
the method to extract kaomojis tense(Onishi and
Okumura, 2014) and so on. “Science of Emoti-
cons”(Ptaszynski et al., 2012) has a detailed knowl-
edge of general kaomoji extraction method.
3 CONSTRUCTION OF
LARGE-SCALE KNOWLEDGE
BASE REPRESENTING THE
BASE FORM OF KAOMOJI
In this study, we annotate 70,106 kinds of kaomoji
that are collected from websites. Annotators add tags
to kaomojis each element using the rules in the fol-
lowing subsections. The terms to annotate kaomoji is
from 1st October 2015 to 20th March 2016. Annota-
tors are six persons including five males and a female.
As a result, we annotated 43,373 kaomoji out of all of
the collected kaomoji in the experimental period.
3.1 Definition of the Base Form of
Kaomoji
It is hard that we analyze the meaning of each element
of kaomoji because kaomoji have various elements.
Therefore, we define the base form of kaomoji such as
stemming of verbs or lemmatization for easy handling
in natural language processing. The basic elements of
the base form of kaomoji are the triplet (eye-mouth-
eye: sometimes move away from mouth towards the
nose) that Ptaszynski et al. proposed. Besides, we fo-
cus on the borders of kaomoji because almost kaomoji
have borders. Table 1 shows the rules and regulations
to extract/make the base form of kaomoji.
We extracted 3,110 base forms of kaomoji in this
work. Table 2 shows the sample of the base form of
kaomoji in order of prevalence. The number of the
base forms that have only one derivative kaomoji is
1,173. Therefore, 1,973 kinds of the base form of kao-
moji have two or more derivative kaomoji. We show
the details of each kaomojis element in the following
subsection.
3.2 Annotated Elements of Kaomoji
We annotated the collected kaomoji to extract the base
form of kaomoji. We also annotated each element that
falls under the category of eye, nose, mouth, cheek,
6
means “Hello.
Table 2: Examples of kaomojis original form.
The base form Freq. The base form Freq.
(_)
1032
(ˆ-ˆ)
495
(-_-)
610
(00)
471
()
589
(w)
466
(__)
540
(o|o)
452
(>_<)
529
()
426
ear, forehead, border, arm, other body elements, cap-
tion, onomatopoeia, and the other expression. In this
section, we show the examples of these elements of
kaomoji.
3.2.1 Eyes
The smallest components of kaomoji are eyes. There-
fore, this research pushes aside emoticons such as
orz
or back shot that have no eyes because these emoti-
cons do not show their faces. kaomoji sometimes have
signs of inequality as their eyes such as
(>_<)
. These
signs indicate that eyes harmonized with arms. We
extract these signs as eyes in this study. Table 3 shows
the sample of eyes.
Table 3: Eyes: the elements of the base form of kaomoji.
kaomoji Eyes kaomoji Eyes
(--) - - ()
3.2.2 Mouth, Nose
The next important elements of kaomoji are mouth
and nose. The expressions dramatically increase with
mouth or nose as the elements of kaomoji. On the
other hand, a nose exists instead of mouth or a mouth
harmonized with a nose. Therefore, we extract a
mouth and a nose as core elements of kaomojis orig-
inal form by consensus of annotators. Table 4 shows
the samples of mouth and nose.
Table 4: Mouth and Nose: the second elements of the base
form of kaomoji.
kaomoji Mouth kaomoji Nose
((( ))) (*ˆˆ*) ˆˆ
( ’’) - () @@
3.2.3 Borders
Borders of kaomoji are made mostly of a parenthe-
sis. In other words, a parenthesis is not an essential
element of kaomoji regarding discriminating kaomoji
with a concept of IDF(Inverse Document Frequency
in information retrieval or natural language process-
ing). Borders, however,change the expression of kao-
moji or the combination of borders express the move-
ment of kaomoji. We extract borders as the element
parts of the base form of kaomoji because they have
possibilities to express a subtle sense. Table 5 shows
the samples of borders.
Table 5: Borders: the third elements of the base form of
kaomoji.
kaomoji Borders kaomoji Borders
() ( ) _
3.2.4 Cheek, Ears, Forehead
The face has the other parts such as cheek, ears and
forehead except for the elements as noted above. We
healthy people have these parts without any excep-
tions. On the other hand, kaomoji have these parts for
the rare occasion(frequency of these parts is low). We
extract the base form of kaomoji exclusive of these
parts because these parts have the role of decorating
kaomojis original form. Table 6 shows the samples
of cheek and ears.
Table 6: The samples of cheek and ears.
kaomoji Cheek kaomoji Ears
(*ˆˆ*) * * Ł(() )Ł Ł Ł
(;) ; ( )
3.2.5 Arms
The word kaomoji has two japanese words: “kao”
and “moji. The word “kao” means the “face” and
the word moji” means the “characters or symbols”
in Japanese language under normal circumstances.
Whether arms, legs or some parts exclusive of the
parts of face belong to kaomoji remains a matter of de-
bate. In fact, there are many something like kaomoji
that have arms, legs, and so on. We extract arms as
the parts that add value to kaomoji in this study. Arms
consist of symmetry parts or one side parts. There-
fore, we extract arms without considering symmetry
property. Table 7 shows the samples of arms.
Table 7: The samples of arms.
kaomoji Arms kaomoji Arms
|O| ( --)
3.2.6 Other Elements
Many various elements show something exclusive of
face or body as the elements of kaomoji. These el-
ements show the particular situation or emphasizing
emotions. Whether these parts belong to kaomoji re-
mains a matter of debate in a similar way to arms. In
this paper, the existential reason of kaomoji is to re-
solve the issue of context in Japanese culture as we
once remarked in Section 1. Therefore, this study de-
fines kaomoji as combined expression of faces and the
other particular situations. Table 8 shows the sample
of the elements that shows particular situation. In this
case,
˜
shows the situation: ”Do you have a cup of
Japanese tea?”.
Table 8: The other elements exclusive of face and arms.
kaomoji Elements
( -_-)˜ ˜
The manageability of these elements differs de-
pending on the position that kaomoji have these ele-
ments on the left side or right side. For example, it is
harder for CAO system or Bedrick’s PCFG to extract
sequences forward kaomoji than following kaomoji.
The reason is that kaomoji are used as the punctuation
of sentences. We can extract the sequences follow-
ing kaomoji using CAO system or Bedrick’s PCFG
because we usually use kaomoji at the end of sen-
tences. On the other hand, it is unobvious where we
should return to extract sequences forward kaomoji.
Our knowledge base of kaomoji discriminates the se-
quences that are at the left side or the right side.
3.2.7 Caption, Onomatopoeia
Kaomoji have every character and symbol as the ele-
ments. These elements construct “sentence” or “ono-
matopoeia in some instances. Sentences or ono-
matopoeia have detailed information such as emo-
tions or situations. We extract this information from
kaomoji in a proactive way. For example,
o(ˆ-ˆ)o
shows the wonderful situation with onomatopoeia
7
.
7
means wonderful.
3.2.8 Other Expression
There are the characteristic elements of kaomoji other
than described above. For example,
(((()))))))
shows shuddering situation with repeated use. Re-
peated use of a certain element shows this expression.
We extract these repeated use element as the other ex-
pression of kaomoji.
We Japanese people regularly use two-byte char-
acters in daily life. It is necessary to discriminate
between two-byte characters and one-byte characters.
For instance,
(ˆˆ)
and are similar looking faces be-
cause these kaomoji consist of same parts. These are,
however, different faces in a strict sense because of
the former consist of one-byte characters, the latter
consist of two-byte characters. The latter looks like
well-rounded than the former. In this paper, we trans-
form characters from two-byte characters to one-byte
characters when we extract/make kaomojis original
form because of choking off the manifold.
4 EXPERIMENTS FOR
ESTIMATING THE BASE FORM
OF KAOMOJI
We can refer the base form of kaomoji and the other
elements of kaomoji using the large-scale knowledge
base of kaomoji constructed in this paper. There is,
however, a high possibility of encountering unknown
kaomoji because of the number of kaomoji increase
day by day. We except UTF kaomoji despite the in-
crease in kaomoji because we find out the UTF kao-
moji that are outside our study. Figure 1 shows the
samples of UTF kaomoji.
Figure 1: kaomoji using UTF-8.
This paper investigates the method to estimate the
base form of kaomoji. If the method can estimate
the base form of kaomoji, then the method can also
estimate the information of kaomoji using the other
parts exclusive of the base form. In other words, the
method estimate emotions, situations, and gestures
using the base form of kaomoji and the other elements
of it. We describe the method of estimating original
form using N-gram based features, simple Skip-gram
features, and the combination of these features in the
following section.
4.1 Experiment to Estimate the Base
Form of Kaomoji using N-gram
Model
N-gram based features are widely used in natural lan-
guage processing. We can use this simple model for
analyzing kaomoji. We can calculate the similarity
between kaomoji using N-gram based features by sep-
arating kaomoji. In this paper, we calculated the ac-
curacy of estimating the base form of kaomoji using
N-gram based model (N=1 to 5)(Eq. 1). In the case
of Kaomoji that have multiple base forms, we count
it a correct answer if one of the each base form is ex-
tracted.
Accuracy =
num of kaomoji estimated correctly
num of all annotated kaomoji (43, 373)
(1)
Cosine similarity is used to calculate similarity be-
tween kaomoji(Eq. 2).
Cosine(A, B) =
A· B
|A||B|
(2)
Our method calculates similarity using Eq. 2 be-
tween input kaomoji and all of extracted/made the
base forms from large-scale knowledge base of kao-
moji (Section 3.1). The base form of kaomoji that
scored the highest similarity is the estimated base
form. Table 9 shows the result of estimation using
N-gram based features.
Table 9: Accuracy of estimating the base form of kaomoji
using N-gram based single feature.
N-gram Accuracy
1gram 0.244
2gram 0.386
3gram 0.440
4gram 0.336
5gram 0.019
The highest accuracy is 0.440 (using Trigram). On
the other hand, the lowest accuracy is 0.019 (using
5-gram). 5-gram model seems to be useless for esti-
mating original form. We investigate the estimation
of kaomojis original form exclusive of 5-gram in the
following section.
4.2 Experiment to Estimate the Base
Form of Kaomoji using Simple
Skip-gram Model
Skip-gram is used in Word2Vec model proposed by
Mikolov et al(Mikolov et al., 2013). Skip-gram is
the model that use co-occurrence words in config-
ured window size. If the window size equals 6,
then for example ”I go to school by bus. is sep-
arated to ”I - go”, ”I - to”, ”go - bus” , and so
on. In our study, we use more simple Skip-gram
model. We extend N-gram model to simple Skip-
gram model that does not use every combination of
co-occurrence words(characters) but only use skip-
ping N words (characters). If the window size equals
3(Skip 1 character model), then for example ”I go to
school by bus. is separated to ”I - to”, ”go - school”,
”to - by”, ”school - bus” by using our simple Skip-
gram model. We can change the expression of kao-
moji by inserting a certain element. For instance, we
transform
(ˆ_ˆ)
into
(*ˆ_ˆ*)
by inserting (
*
). Sim-
ple Skip-gram model is effective about this situation.
Table 10 shows the result using naive Skip-gram(Skip
N characters N = 1 to 3). The highest accuracy is
0.270(Skip 1 character model). However, the accu-
racy is lower than N-gram based model.
Table 10: Accuracy of estimating the base form of kaomoji
using simple Skip-gram based single feature.
naive Skip-gram Accuracy
Skip 1 character 0.270
Skip 2 characters 0.039
Skip 3 characters 0.018
4.3 Experiment to Estimate the Base
Form of Kaomoji using the
Combination of N-gram and Simple
Skip-gram based Features
In the above experiment, we use only individual fea-
tures such as only unigram, only trigram and so on.
We examine the multiple features to estimate the base
form of kaomoji. 5-gram model is removed in this
experiment as previously noted.
Table 11 shows the result of combination
model(N-gram based features). The highest accuracy
is 0.433(bigram + trigram model). However, this ac-
curacy is lower than using only trigram model despite
we form a hypothesis that the combination model im-
proves the accuracy.
Table 12 shows the result of combination model
(N-gram and naive Skip-gram model). The highest
accuracy is 0.489 (bigram + trigram + Skip 1 charac-
ter model) with contrary to expectations.
Figure 2 shows each the highest accuracy of es-
timation using N-gram and simple Skip-gram based
features.
Table 11: Accuracy of estimating the base form of kaomoji
using combination of N-gram based features.
Combinations Accuracy
1gram+2gram 0.381
1gram+3gram 0.406
1gram+4gram 0.339
2gram+3gram 0.433
2gram+4gram 0.410
3gram+4gram 0.425
1gram+2gram+3gram 0.422
1gram+2gram+4gram 0.401
1gram+3gram+4gram 0.414
2gram+3gram+4gram 0.427
1gram+2gram+3gram+4gram 0.431
Table 12: Accuracy of estimating the base form of kao-
moji using combination of N-gram based features and naive
Skip-gram based features.
Combinations Accuracy
2gram+Skip 1 character 0.470
3gram+Skip 1 character 0.404
2gram+3gram+Skip 1 character 0.489
5 CONCLUSION
In this paper, we annotated kaomoji and constructed
a large-scale knowledge base of kaomoji. The total
number of annotated kaomoji is 43,373 out of our col-
lected kaomoij. As a result, we extracted 3,110 kinds
of the base form of kaomoji. In experimental evalu-
ations, we achieved 0.489 accuracy of estimating the
base form kaomoji using bigram + trigram + Skipping
1 character model. Ih the case of random estimation,
the maximum probability is 1,032 / 43,373 (=0.024)
that Table 2 shows. We consider that our experiments
are efficient for estimating the base form of kaomoij.
44.0%
43.3%
48.9%
0%
10%
20%
30%
40%
50%
60%
3gram 2gram+3gram 2gram+3gram
+Skip 1 character
Accuracy
Figure 2: Result of experiments using N-gram and simple
Skip-gram based fearures.
As future works, we will finish annotating re-
maining kaomoji. We will attempt to construct clas-
sification method using deep learning tools such as
Chainer, TensorFlow, and so on to improve the ac-
curacy of estimation (c.f. emoji2vec(Eisner et al.,
2016)). In addition, we have to annotate kaomojis
emotions based on Plutchik model(Plutchik, 1980).
Because we do not extract emotions that kaomoji
shows in this paper.
ACKNOWLEDGEMENT
This work was supported by JSPS KAKENHI Grant
Number 15K21592.
REFERENCES
Bedrick, S., Beckley, R., Roark, B., and Sproat, R. (2012).
Robust kaomoji detection in twitter. In Proceedings
of the Second Workshop on Language in Social Me-
dia, LSM ’12, pages 56–64, Stroudsburg, PA, USA.
Association for Computational Linguistics.
Eisner, B., Rockt¨aschel, T., Augenstein, I., Bosnjak, M.,
and Riedel, S. (2016). emoji2vec: Learning emoji
representations from their description. Proceedings of
The Fourth International Workshop on Natural Lan-
guage Processing for Social Media, pages 48–54.
Hall, E. T. (1976). Beyond Culture. Anchor Books.
Kazama, K., Mizuki, S., and Sakaki, T. (2016). Study of
sentiment analysis using emoticons on twitter. The
30th Annual Conference of the Japanese Society for
Artificial Intelligence, 3H3-OS-17a-4.
Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S., and
Dean, J. (2013). Distributed representations of words
and phrases and their compositionality. In Burges, C.
J. C., Bottou, L., Welling, M., Ghahramani, Z., and
Weinberger, K. Q., editors, Advances in Neural In-
formation Processing Systems 26, pages 3111–3119.
Curran Associates, Inc.
Onishi, C. and Okumura, N. (2014). An inverstigation of the
usage of kaomoji for emotions judgment and kaomoji
recommendation. In The 13th IASTED International
Conference on Artificial Intelligence and Applications
AIA2014. #816-014.
Plutchik, R. (1980). Chapter 1 - a general psychoevolution-
ary theory of emotion. In Plutchik, R. and Kellerman,
H., editors, Theories of Emotion, pages 3 33. Aca-
demic Press.
Ptaszynski, M., Maciejewski, J., Dybala, P., Rzepka, R., and
Araki, K. (2010a). Cao: A fully automatic emoticon
analysis system. In Fox, M. and Poole, D., editors,
AAAI. AAAI Press.
Ptaszynski, M., Maciejewski, J., Dybala, P., Rzepka, R., and
Araki, K. (2010b). Cao: A fully automatic emoticon
analysis system based on theory of kinesics. IEEE
Transactions on Affective Computing, 1(1):46–59.
Ptaszynski, M., Maciejewski, J., Dybala, P., Rzepka, R.,
Araki, K., and Momouchi, Y. (2012). Science of
Emoticons. IGI Global.
Tanaka, Y., Takamura, H., and Okumura, M. (2005). Ex-
traction and classification of facemarks. In Proceed-
ings of the 10th International Conference on Intel-
ligent User Interfaces, IUI ’05, pages 28–34, New
York, NY, USA. ACM.
Urabe, Y., Rafal, R., and Araki, K. (2013). Emoticon rec-
ommendation for japanese computer-mediated com-
munication. In Semantic Computing (ICSC), 2013
IEEE Seventh International Conference on, pages 25
31.
Yamada, T., Tsuchiya, S., Kuroiwa, S., and Ren, F. (2007).
Classification of facemarks using n-gram. In 2007
International Conference on Natural Language Pro-
cessing and Knowledge Engineering, pages 322–327.