A Flick-based Japanese Tablet Keyboard using Direct Kanji Input

Yuya Nakamura

and Hiroshi Hosobe

Graduate School of Computer and Information Sciences, Hosei University, Tokyo, Japan

Faculty of Computer and Information Sciences, Hosei University, Tokyo, Japan

Keywords:

Text Entry Touch Screen, Touch Typing, Software Keyboard.

Abstract:

Tablets, as well as smartphones and personal computers, are popular as Internet clients. A split keyboard is a

software keyboard suitable for tablets with large screens. However, unlike other methods, the split keyboard

has space in the center of the screen, which makes the part of the screen for displaying suggestions small and

inconvenient. This paper proposes a Japanese input software keyboard that enables direct kanji input on a split

ﬂick keyboard. Once the user has mastered this keyboard, it allows the user to efﬁciently input Japanese text

while holding a tablet with both hands. The paper presents an implementation of the keyboard on Android

and reports the result of an experiment on its performance compared with existing methods. In addition, since

direct kanji input generally takes time for users to learn, one of the authors by himself has conducted a long-

term experiment to conﬁrm the possibility of its mastery. For 12 months, both the input speed and the error

rate have gradually improved.

1 INTRODUCTION

Japanese text is composed of Chinese-originated

“kanji” characters and Japanese “kana” characters

(Tamaoka, 2014). While a kanji character typi-

cally has a meaning, a kana character does not; in-

stead, a kana character is associated with a speech

sound. Computer users need to input both kanji and

kana characters for Japanese text. Many Japanese

keyboards use a conversion function from kana to

kanji. One of the most common kana-kanji conver-

sion methods used in Japan is the predictive conver-

sion that uses past inputs as well as dictionaries. By

contrast, direct kanji input lets users select kanji char-

acters by themselves. It is believed that once a user

mastered direct kanji input, the user can efﬁciently

and effortlessly input Japanese text because of the un-

necessity of judging conversion suggestions.

A split keyboard is a software keyboard for tablets

that improves display space efﬁciency. It allows its

user to put both hands at natural positions. The im-

provement of display space efﬁciency is good for mul-

titasking. Split keyboards are often used with the QW-

ERTY layout, and are adopted on Windows and iPad

tablets (Knox, 2012; Apple Inc., 2019). Current split

keyboards for Japanese input use common kana-kanji

conversion methods such as predictive conversion.

However, unlike other software keyboards, split key-

boards have space in the center of the screen, which

makes the part of the screen for displaying conversion

suggestions small and inconvenient.

In Japan, there are about as many users of “ﬂick”

keyboards as QWERTY keyboard users. Flick is a

gesture operation especially used for character input

on touch-screen devices such as tablets and smart-

phones. Flick input usually makes its user to touch

a key with a ﬁnger and then slide the ﬁnger upward,

downward, to the left, or to the right for input.

This paper proposes a Japanese input software

keyboard that enables direct kanji input on a split ﬂick

keyboard. We extend two-handed ﬂick input (Naka-

mura and Hosobe, 2020) to allow direct kanji input.

Once the user has mastered this keyboard, it allows

the user to efﬁciently input Japanese text while hold-

ing a tablet with both hands. We propose two layouts:

one uses elements of kanji characters called bushu;

the other uses elements called on’yomi. Kanji char-

acters are generally made up of dots and lines. The

bushu-based layout uses a common collection of such

dots and lines. The on’yomi-based layout arranges

kanji characters by using their readings. They cover

2136 kanji characters called joyo-kanji, a standard set

of kanji characters for social life in Japan.

For the objective evaluation of the proposed key-

board, we conducted comparative experiments. We

compared the bushu-based layout, the on’yomi-based

layout, two QWERTY software keyboards, and a ﬂick

keyboard. One of the QWERTY keyboards enabled

Nakamura, Y. and Hosobe, H.

A Flick-based Japanese Tablet Keyboard using Direct Kanji Input.

DOI: 10.5220/0010245800490059

In Proceedings of the 16th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2021) - Volume 2: HUCAPP, pages

49-59

ISBN: 978-989-758-488-6

learning in predictive conversion for individual par-

ticipants, and the other disabled it. We recruited 8

participants of ages ranging from 23 to 24. During the

experiments, they were seated and held a tablet with

both hands. The comparative experiments treated two

kinds of input: sentence input and kanji conversion-

required word input. The results showed that the

on’yomi-based layout was faster in the input speed

than the bushu-based layout. In addition, the results

of subjective evaluation using the User Experience

Questionnaire (UEQ) showed that the proposed key-

board was better in novelty than the existing methods

although it was inferior in perspicuity, efﬁciency, and

dependability.

Direct kanji input generally takes time for users

to learn. Therefore, one of the authors by himself has

conducted a long-term experiment to conﬁrm the pos-

sibility of the mastery of the proposed keyboard using

the bushu-based layout. The author has input ﬁve sen-

tences every day. For 12 months, the input speed has

increased, and the error rate has decreased.

2 RELATED WORK

In this paper, we extend Nakamura and Hosobe’s bi-

manual ﬂick software keyboard for tablets (Nakamura

and Hosobe, 2020). It improved screen space efﬁ-

ciency by splitting a ﬂick keyboard into the left and

the right. However, it was not suitable for kana-kanji

conversion because the conversion space generally

extends from the left to the right edge of the screen.

In this paper, we solve this problem by introducing

direct kanji input.

Many Japanese input method uses predictive con-

version, which presents conversion suggestions based

on previously used words. For example, Ichimura et

al. proposed a predictive kana-kanji conversion sys-

tem (Ichimura et al., 2000). It used the current main-

stream predictive conversion method that had been

proposed several years before, and reduced users’

keystrokes to 78 %.

Unlike predictive conversion that is used in many

current Japanese input methods, direct kanji input lets

the user select a kanji character. There are two types

of direct kanji input: associative and non-associative.

Associative direct kanji input has clear relationships

between keystrokes and kanji characters. This method

has the advantage of being more intuitive and easier

to use (T-Code Project, 2003). On the other hand,

non-associative direct kanji input does not have such

clear relationships between keystrokes and kanji char-

acters. From the user’s point of view, it is a random

key placement. For this reason, it is more difﬁcult

to use than the associative method. T-Code (T-Code

Project, 2003) is one of the most famous methods

of non-associative direct kanji input. T-Code uses a

combination of two keystrokes on the QWERTY key-

board to enter a kanji character. There is no regu-

larity in such key combinations, and the user needs

to ﬁrst learn them. The non-associative method takes

longer time to learn than the associative method, but

allows faster input (T-Code Project, 2003). This is be-

cause the associative method requires the user to as-

sociate kanji characters with keystrokes, but the non-

associative method does not. However, whichever

method is used, direct kanji input requires the user

to more practice than kana-kanji conversion-based in-

put. In this paper, direct kanji input keyboard was not

included in comparative experiment. The reason is

that we were not able to ﬁnd any available software

keyboards for tablets that used direct kanji input.

Research and development of input methods for

Chinese characters are not limited to the Japanese lan-

guage. Pinyin input is a widely used Chinese charac-

ter input method that uses Chinese readings of char-

acters (Li and Li, 2019). Cangjie is a direct Chinese

character input method used in Hong Kong. In this

method, users think of a Chinese character as a com-

bination of parts. A keystroke corresponds to such a

part, and a combination of keystrokes is used to input

a Chinese character. Liu and Lin (Liu and Lin, 2008)

proposed an extension of Cangjie to classify similar

Chinese characters. Niu et al. (Niu et al., 2010) pro-

posed Stroke++, a Chinese character input method for

mobile phones, in which an input is made by combin-

ing bushu elements.

Various research has been done on keyboards for

tablets. Sax et al. proposed an ergonomic QWERTY

tablet keyboard (Sax et al., 2011). Bi et al. proposed

a bimanual gesture keyboard to reduce display space

and to shorten ﬁnger movement (Bi et al., 2012).

Hasegawa et al. studied input of a software keyboard,

with a focus on aging effects and differences between

dominant and non-dominant hands (Hasegawa et al.,

2012). Odell studied feedbacks of software keyboards

(Odell, 2015). Takei and Hosobe proposed a Japanese

kana input keyboard that input one character with

two strokes by using 2 × 6 keys (Takei and Hosobe,

2018). Yajima and Hosobe proposed a Japanese soft-

ware keyboard for tablets that reduced user fatigue

(Yajima and Hosobe, 2018).

In Japan, much research on ﬂick keyboards has

been done. Sakurai and Masui proposed a QWERTY

ﬂick keyboard (Sakurai and Masui, 2013). This key-

board enabled input of Japanese kana characters and

English letters without mode changes. Fukatsu et al.

proposed an eyes-free Japanese kana input method

HUCAPP 2021 - 5th International Conference on Human Computer Interaction Theory and Applications

called no-look ﬂick (Fukatsu et al., 2013). This

method enabled ﬂick input for vowels and consonants

in two keystrokes. Hakoda et al. proposed a kana in-

put method using two ﬁngers for touch-panel devices

(Hakoda et al., 2013). This method was also an eyes-

free Japanese input method, but enabled gesture input

by two ﬁngers.

Our software keyboard requires users to have

more knowledge of kanji characters than exist-

ing methods, especially concerning their bushu and

on’yomi elements. Therefore, the continuous use of

our software keyboard could be regarded as a pro-

cess of learning kanji or Chinese characters. Much

research has also been done on learning Chinese char-

acters. Dragon Tale (Plecher et al., 2018) is an ad-

venture game in which players learn kanji characters.

Kanev et al. (Kanev et al., 2012) proposed a 3D

model to let students learn the elements of Chinese

characters. In Japan, standard kanji education of chil-

dren lasts for 9 years from the ﬁrst year of elemen-

tary schools to the third year of junior high schools,

which also has caused proposals of various methods

of teaching kanji characters outside the research com-

munity.

3 JAPANESE CHARACTERS AND

KEYBOARDS

3.1 Kana Characters

As previously described, Japanese text is composed

of Chinese-originated kanji characters and Japanese

kana characters. While a kanji character typically

has a meaning, a kana character is associated with a

speech sound. There are two kinds of kana characters

called hiragana and katakana. Although they are used

for different purposes, they correspond to each other;

for each hiragana character, there is a corresponding

katakana character, and vice versa. There are approxi-

mately 50 basic kana characters, which are further di-

vided into 10 groups that are ordered, each of which

typically consists of 5 characters. The ﬁrst group is

special because its 5 characters indicate 5 vowels that

are pronounced “a,” “i”, “u”, “e”, and “o”. The other 9

groups are associated with the basic consonants, “k”,

“s”, “t”, “n”, “h”, “m”, “y”, “r”, and “w”. A kana

character in these 9 groups forms the sound that com-

bines a consonant and a vowel. For example, the 5

characters of the “k” group are pronounced “ka”, “ki”,

“ku”, “ke”, and “ko”. This grouping of kana charac-

ters is basic knowledge of the Japanese language.

The “k,” “s,” “t,” and “h” groups have variants

called dakuon. Speciﬁcally, the dakuon variants of

“k”, “s”, “t”, and “h” are “g”, “z”, “d”, and “b” re-

spectively. In addition, the “h” group has another vari-

ant called handakuon, which is “p”. Certain charac-

ters have variants that are written in smaller shapes.

Sequences of kana characters can be expressed with

the Roman alphabet by using the standard Japanese

romanization system (ISO, 1989). This is widely used

for computer users to enter Japanese text with alpha-

bet keyboards such as QWERTY.

3.2 Elements of Kanji Characters

We explain bushu and on’yomi elements of kanji

characters that we use in our software keyboard.

3.2.1 Bushu

Bushu, also called a radical, indicates an element of

kanji characters. Kanji characters are generally made

up of dots and lines. A bushu element is a common

collection of such dots and lines. For example, Figure

1 shows kanji characters for a pine and cherry blos-

soms. The red boxes in the ﬁgure indicate their bushu

elements. This type of bushu is called “kihen” and

is typically used in kanji characters related to trees.

Other kanji characters that use kihen correspond to,

for example, a small forest and a bridge. The total

number of bushu elements used in Japan is 214.

Figure 1: Examples of bushu elements.

3.2.2 On’yomi

Kanji characters typically have two kinds of read-

ings, on’yomi and kun’yomi. The on’yomi of a kanji

character indicates its old Chinese reading, while the

kun’yomi indicates its Japanese reading. In general,

the on’yomi of a kanji character does not make sense

in Japanese while the kun’yomi does. For example,

the kun’yomi of the kanji character for cherry blos-

soms in the Figure 1 is “sakura”, which means cherry

blossoms by itself. By contrast, its on’yomi, which is

“ou”, has no meaning in Japanese.

A Flick-based Japanese Tablet Keyboard using Direct Kanji Input

3.3 Japanese Keyboards

Figure 2 shows a typical Japanese ﬂick keyboard. The

main key layout is composed of 4 × 3 keys. If a user

ﬂicks a key to the left, upward, to the right, or down-

ward with a thumb, the keyboard inputs a character

corresponding to the direction (Figure 3). The enter

key and the delete key are located on the right side of

the keyboard. When either the “123” or the “ABC”

key is pressed on the left side of the keyboard, the

Japanese keyboard is replaced with the English letter

keyboard or the number letter keyboard. A conver-

sion space is located at the top of the keyboard. When

a user touches a word, kana characters are converted

to kanji or other characters. If a user touches the up-

ward arrow, it will show other kanji candidates. In

addition, most of Japanese input keyboards provide a

function for learning conversions. This function al-

lows the user to quickly select recently used charac-

ters. Predictive conversion is a function for predicting

kana-kanji conversions from partial inputs. This func-

tion allows the user to more efﬁciently input text if the

prediction is successful. However, it is often difﬁcult

to perform the predictive conversion from only a few

kana characters.

Figure 2: Typical Japanese ﬂick keyboard (which is avail-

able on iOS).

Figure 3: A typical input ﬂow for a Japanese ﬂick keyboard

(which is available on iOS).

4 PROPOSED METHOD

4.1 Bimanual Flick

We propose a software keyboard for tablets that splits

a ﬂick keyboard into the left and right sides. It is

based on the bimanual split ﬂick keyboard proposed

by Nakamura and Hosobe (Nakamura and Hosobe,

2020) that uses normal kana-kanji conversion. Instead

of using such normal kana-kanji conversion, our new

software keyboard introduces direct kanji input. Since

the user can use the keyboard while holding the tablet

with both hands, the advantages of split keyboards are

not lost. If the user wants to input a kana character,

the user only needs to ﬂick a key on one side as with

other splits keyboards.

A primary reason for developing a new direct

kanji input method is that previous methods were de-

signed for desktop computers. Although these previ-

ous methods could be adapted to newer devices such

as tablets and smartphones, there has not been much

research, and their effectiveness is unclear. Also, the

proposed method might be applicable to the Chinese

language. This is because Cangjie is similar to our

bushu-based method and Pinyin input is similar to our

on’yomi-based method (although Japanese and Chi-

nese use different sets of standard Chinese characters,

which would require extra efforts.)

A typical input ﬂow for a kanji character is shown

in Figure 4. This example inputs the kanji character

meaning “cherry blossoms” previously shown in Fig-

ure 1. Figure 4-a shows a state in which no input is

made. This kanji character uses the bushu element

called “kihen”, which belongs to the “ki” group. The

“ki” group further belongs to the “k” group. There-

fore, the user ﬁrst touches the “ka” key that repre-

sents the “k” group (Figure 4-b). Then the user ﬂicks

the ﬁnger to the left to show the “ki” group (Figure

4-c). The user looks for the kanji character for cherry

blossoms, ﬁnding that it belongs to the top left key.

Therefore, the user touches the top left key (Figure 4-

d). Then the user ﬂicks the ﬁnger to the right (Figure

4-e), which completes the input of the kanji character.

Figure 4: A typical input ﬂow for a kanji character.

The number of kanji characters that can be in-

put with this method is 2136. These kanji characters,

called joyo-kanji, are indicated as the standard for us-

ing kanji characters in social life in Japan. Accord-

ing to a survey conducted by the Japanese Agency for

Cultural Affairs, more than 96 % of the total number

of kanji characters regularly used in Japanese society

are joyo-kanji characters (Takeda, 2019). It was pos-

HUCAPP 2021 - 5th International Conference on Human Computer Interaction Theory and Applications

sible to cover more kanji characters in the proposed

method. However, we decided that no more kanji

characters were needed. Other kanji and kana char-

acters can be input using the conversion function of

the previous bimanual ﬂick keyboard. It also should

be noted that our method covers more characters than

T-Code (T-Code Project, 2003), which covers 1600

characters.

4.2 Kanji Layouts

We propose two types of kanji layouts: the bushu-

based layout and the on’yomi-based layout. In

the comparative experiments, we compare these two

methods with existing ones and evaluate their perfor-

mance.

4.2.1 Bushu-based Layout

The bushu-based layout uses bushu elements of kanji

characters. The layout is shown in Figures 5 and

6. Each cross in Figure 5 indicates a key that can

be ﬂicked upward, downward, to the left, and to the

right, and the positions of the crosses correspond to

the shape of the keyboard. The bushu elements that

appear more than once in the ﬁgure are those that have

many kanji characters. On the contrary, the bushu ele-

ments shown in Figure 6 have a few kanji characters.

They are grouped together in the parts labeled with

the numbers in Figure 5. In the case of kanji charac-

ters with the same bushu element, they are arranged

by on’yomi from the top left. Also, since there is a

limit on the size of the keyboard, four bushu elements

with large numbers of characters are divided into two

keys. In this case, they are located symmetrically.

Figure 5: Placement of bushu elements.

Figure 6: Bushu elements with a few kanji characters.

4.2.2 On’yomi-based Layout

The on’yomi-based layout arranges kanji characters

by using their readings. Since dakuon is also present

in on’yomi, the “k,” “s,” “t,” and “h” groups can be re-

placed with the dakuon keyboard. Figure 7 shows the

process of inputting the kanji characters of the “ka”

group. If the user presses the top left key on the kanji

keyboard, it is replaced with the dakuon keyboard. If

the user presses the top left key on the dakuon key-

board, it is replaced with the original keyboard.

Figure 7: The on’yomi-based layout (for inputting a kanji

characters in the “ka” group).

4.3 Direct Kanji Input

Our keyboard is based on direct kanji input. Previous

direct kanji input methods were categorized into as-

sociative input and non-associative input. We regard

the bushu-based layout as being semi-associative be-

cause it is neither associative nor non-associative in

the original senses. It basically arranges kanji char-

acters by bushu elements, but there is no perfect cor-

respondence between kana characters and bushu ele-

ments. By contrast, the on’yomi layout is associative

since there is correspondence between kanji charac-

ters and kana characters.

A Flick-based Japanese Tablet Keyboard using Direct Kanji Input

4.4 Key Arrangement

Our software keyboard is of size of 240-px height

and 180-px width before conversion. This layout re-

duces the display space by 73 % in the portrait mode

and by 83 % in the landscape mode, compared with

the QWERTY keyboard with the maximum display

width. The size and the position of the bushu-based

layout are based on the heat map used in the design

of the Windows 8 touch keyboard (Knox, 2012). This

limited the bushu-based layout to 4 × 4. By contrast,

the on’yomi layout does not have this limit. There-

fore, in the case of the on’yomi-based layout, there

may be much more kanji characters for a kana char-

acter than in the case of the bushu-based layout. In the

case of the bushu-based layout, the maximum number

of kanji characters for a kana character is 80. Since

there are 50 kana characters in total, the total num-

ber of kanji characters that can be placed is 4000. We

adopted 2136 joyo-kanji characters. Kana keys with a

few kanji characters are composed of a 2× 2 kanji lay-

out. The keyboard in the bushu-based layout is of the

maximum size of 320-px height and 320-px width on

one side. The keyboard in the on’yomi-based layout

is of the maximum size of 480-px height and 400-px

width.

5 IMPLEMENTATION

We implemented the proposed software keyboard on

an ASUS ZenPad 10 tablet (Android OS 7.0, 1920 ×

1200-px screen) as shown in Figure 8. The keys are of

60 × 60 px, and the keyboard is placed symmetrically

at the lower ends of the screen. In the proposed key-

board, its position was adjustable with a bar at the bot-

tom of the screen. The red part in the ﬁgure indicates

the conversion space that is used to enter non-joyo-

kanji and katakana characters. The user can display

different characters by swiping the conversion space

to the left or to the right. The radio button in the center

of the ﬁgure allows the user to change the key layout.

6 COMPARATIVE

EXPERIMENTS

To evaluate the proposed software keyboard, we con-

ducted an experiment on its comparison with exist-

ing software keyboards. We compared the bushu-

based layout, the on’yomi-based layout, two QW-

ERTY keyboards, and a ﬂick keyboard. There are

two types of QWERTY keyboards, one with the learn-

ing of predictive conversion enabled and one with the

Figure 8: Implementation of the proposed software key-

board.

learning disabled. The reason why the direct input

method was not compared is that we were not able to

ﬁnd any available software keyboards for tablets that

used direct kanji input. The learning of the predictive

conversion is reset for each participant. The compar-

ative experiments treated the landscape mode of each

keyboard layout. In the proposed keyboard, its po-

sition was adjustable with a bar at the bottom of the

screen. The position of the keyboard was set by each

participant.

We recruited 8 participants who were Japanese

university students and workers. Their ages ranged

from 23 to 24, and all the participants were male.

They were seated on a chair and held a tablet in the

landscape mode with both hands. If participants were

not able to reach the center of the keyboard in us-

ing the QWERTY, they were allowed to release their

hands. The comparative experiments were composed

of two input experiments and subjective evaluation.

After the input experiments, we investigated subjec-

tive evaluation for each method. In addition to the

UEQ, free descriptions were also collected.

We measured the input speed and the error rate.

The input speed is measured by the number of charac-

ters per minute CPM, which is calculated as follows:

CPM =

T − E

× 60 (1)

where T is the length of the input string, S is the in-

put time, and E is the number of characters that were

wrong. This means the number of characters typed

correctly per minute. On the other hand, the error rate

ER is calculated as follows:

ER =

C + IF + INF

(2)

where C is the total number of correct words, IF

is the number of incorrect but ﬁxed (backspaced)

words, and INF is the number of incorrect (but not

ﬁxed) words. These equations are based on Bi et

al.’s research (Bi et al., 2012), and we adjust them to

Japanese character input.

HUCAPP 2021 - 5th International Conference on Human Computer Interaction Theory and Applications

The comparative experiments treated two kinds of

input: sentence input and kanji conversion-required

word input. In the following, we describe the details

and the results of the two experiments.

6.1 Sentence Input

In the sentence input experiment, ﬁve input sentences

were selected for each method from a book (Waragai,

2008) about learning joyo-kanji with example sen-

tences. One of the QWERTY keyboards and the ﬂick

keyboard enabled learning in predictive conversion,

but the sentences were speciﬁc to each method, and

therefore the learning was limited to the inside of a

method. Before the experiment, participants warmed

up with a few sentences for each method. They started

the experiment by pushing the start button on the up-

per left corner of the screen, and moved to the next

sentence by pushing the enter key. A target sentence

was displayed on the text ﬁeld at the top of the screen.

The sentences included in the list were of about 20-

character length and mixed kana and kanji characters.

Since the bushu-based layout generally takes long

time for users to learn, and also since the short warm-

up before the experiment was not sufﬁcient for the

participants’ learning, a support function was pro-

vided. It consisted of two hints: a hint for the bushu

element of a kanji character and a hint for the position

of a kanji character. The hint for bushu was displayed

by pressing the hint button in the upper right corner of

the screen (Figure 8). When the button was pressed,

the bushu elements of kanji characters in the target

sentence were displayed one by one. The hint for a

kanji position was shown in Figures 5 and 6. The par-

ticipants were able to look at this diagram if they did

not know the positions of kanji characters.

6.2 Result of the Sentence Input

Experiment

The results of the input speeds and the error rates are

shown in Figures 9 and 10 respectively. In the charts,

“QWERTY ON” and “Flick” indicate the QWERTY

keyboard and the ﬂick keyboard with learning in pre-

dictive conversion respectively. A higher input speed

is better, and a high error rate is worse. The results

show large differences between the proposed meth-

ods and the existing methods (shown as the three bars

on the right sides of the charts). The ANOVA on the

two proposed methods also showed a signiﬁcant dif-

ference in the input speeds (p < 0.05), but not in the

error rates. Among the existing methods, the ﬂick

keyboard was the fastest, and the two QWERTY key-

boards were of about the same speeds. Also, the ﬂick

keyboard was the highest in the error rates, and the

two QWERTY keyboards showed about the same er-

ror rates.

Figure 9: Input speeds. Figure 10: Error rates.

6.3 Kanji Conversion-required Word

Input

The experiment on the kanji conversion-required

word input treated the same layouts as the sentence

input experiment. Its procedure was almost the same

as that of the sentence input experiment, and the only

difference was what the participants input. In this

experiment, the participants input words written in

kana characters at the top of the screen and converted

them into kanji characters. The experiment was con-

ducted after the sentence input experiment and there

was no warm-up time. Five words were used for each

method, and when a participant pressed the enter key,

the next word was displayed. The target word was the

word that appeared in the sentence input experiment.

For this reason, the hints used in the sentence input

experiment were not used in the word input.

6.4 Result of the Kanji

Conversion-required Word Input

Experiment

The results of the input speeds and the error rates are

shown in Figures 11 and 12 respectively. The results

show large differences between the proposed meth-

ods and the existing methods. The ANOVA on the

two proposed methods showed a signiﬁcant differ-

ence in the input speeds (p < 0.05), but not in the er-

ror rates. Among the existing methods, the QWERTY

keyboard with learning in predictive conversion and

the ﬂick keyboard were the fastest and almost compa-

rable. The three existing methods showed similar low

error rates.

6.5 Subjective Evaluation

The subjective evaluation was performed by using the

UEQ (Schrepp et al., 2017) to investigate the two

proposed methods and the QWERTY keyboard with

learning. Figures 13 and 14 showed the results of

A Flick-based Japanese Tablet Keyboard using Direct Kanji Input

Figure 11: Input speeds. Figure 12: Error rates.

the subjective evaluation. In the bushu-based layout,

the UEQ showed an excellent rating for novelty al-

though the ratings of the other items were low. In

particular, ratings for perspicuity and efﬁciency were

very low. The on’yomi-based layout was rated higher

than the bushu-based layout except novelty. When

the on’yomi-based layout and the existing methods

were compared, the existing methods were better

in perspicuity, efﬁciency, and dependability, and the

on’yomi-based layout was better in novelty.

Figure 13: Result of the UEQ on the bushu-based layout.

Figure 14: Result of the UEQ on the on’yomi-based layout.

The comparison of the on’yomi-based layout and

the existing methods is shown in Figure 15. Both the

QWERTY and the ﬂick keyboard obtained higher rat-

ings in perspicuity, efﬁciency, and dependability. The

proposed method was rated higher in novelty.

7 LONG-TERM EXPERIMENT

7.1 Method

Direct kanji input is a method that takes time to learn.

Therefore, one of the authors by himself conducted an

experiment to conﬁrm its mastery by using the pro-

posed method for 12 months. The experiment con-

sisted of the input of sentences. Every day the author

entered ﬁve sentences chosen at random from 602

sentences in the literature (Waragai, 2008). The 602

sentences contain 68 % of the joyo-kanji characters.

The author is a 23-year-old male graduate student in

the ﬁeld of computer science. In the experiment, the

author was seated on a chair and held a tablet with

Figure 15: Result of the comparison of the on’yomi-based

layout, the QWERTY keyboard, and the ﬂick keyboard in

the UEQ.

both hands. The experiment was conducted with the

bushu-based layout, and a search function was used

when the location of a kanji character was not known.

7.2 Result

Figures 16 and 17 show the results of the input speeds

and the error rates respectively. In both charts, the

main line indicates the average value for a day. The

green bars represent the maximum and the minimum

values for the day. For comparison, the author en-

tered the same sentences by using the QWERTY key-

board on the ASUS ZenPad 10 (red line) and Naka-

mura and Hosobe’s bimanual ﬂick keyboard (Naka-

mura and Hosobe, 2020) (purple line). The QWERTY

keyboard on the ASUS Zenpad 10 also enabled learn-

ing in predictive conversion. According to the charts,

both the input speeds and the error rates gradually im-

proved day by day. The reason why the bimanual ﬂick

keyboard had a high error rate was because it used the

backspace key to make adjustments during the con-

version.

Figure 16: Input speeds (12 months).

Figure 17: Error rates (12 months).

HUCAPP 2021 - 5th International Conference on Human Computer Interaction Theory and Applications

8 DISCUSSION

8.1 Comparative Experiment

8.1.1 Sentence Input

The result of the sentence input experiment showed

that the proposed methods were inferior to existing

methods in terms of the input speeds. There was

also a signiﬁcant difference in the input speeds be-

tween the proposed methods. We think that this is

because the bushu-based layout is a semi-associative

direct kanji input method and the on’yomi-based lay-

out is associative. The semi-associative method takes

longer time for users to learn than the associative

method, but can be expected to enable faster input.

The experiment tried to compensate for the difference

by providing hints, but it was not successful.

The result also showed that the proposed meth-

ods were inferior to the existing methods in terms of

the error rates. However, there was no signiﬁcant

difference between the proposed methods in terms

of the error rates. We think that the error rates of

the on’yomi-based layout in the sentence input ex-

periment were higher than its error rates in the kanji

conversion-required word input experiment because

the given sentences were written in kanji and kana

characters. In this case, if a participant could not read

kanji, he would have trouble in input.

Therefore, to compare input speeds, users who

used the proposed method for a long time are needed.

One solution might be to distribute the software of the

proposed method. We could implement a function to

measure the input speed and evaluate it. However,

it would be difﬁcult to have many users of the pro-

posed method. Another solution might be to have a

small number of participants who would use the pro-

posed method for a long term. Either way, it would

be difﬁcult to measure the performance for one year.

Therefore, we think that the immediate solution is to

increase the efﬁciency of the learning of the proposed

method. Speciﬁcally, better key layouts and learning

support software are needed.

8.1.2 Kanji Conversion-required Word Input

The results of the experiments showed that the in-

put speeds of the proposed methods in the kanji

conversion-required word input were not very dif-

ferent from those in the sentence input. However,

there was a difference between the existing methods.

The QWERTY keyboard and the ﬂick keyboard with

learning in predictive conversion showed high input

speeds. As in the sentence input experiment, a sig-

niﬁcant difference in the input speeds was shown be-

tween the bushu-based layout and the on’yomi-based

layout.

Overall, the error rates in the kanji conversion-

required word input were lower than those in the sen-

tence input. We think that this is because of the small

number of characters entered. However, there was a

difference in the error rates between the bushu-based

layout and the on’yomi-based layout. We think that

the reason is that the participants using the on’yomi-

based layout were able to easily associate kanji char-

acters with kana characters, which was not applicable

to the bushu-based layout. However, there was still

no signiﬁcant difference between the bushu-based and

the on’yomi-based layout. We think that this is be-

cause the on’yomi-based layout is not always associa-

tive. Since the bushu-based layout is semi-associative

and the on’yomi-based layout is associative, the input

speeds and the error rates might be reversed depend-

ing on the degree of learning. For accurate assess-

ment, we need to ﬁnd out what is the ratio of the two

degrees of learnability.

There was no large difference in the error rates

between the on’yomi-based layout and any existing

method. According to this result, although it is dif-

ﬁcult to compare the input speeds without the long-

term use of the methods, we think that the error rates

of the methods could be compared even without the

long-term use. To reduce the error rate, another asso-

ciative layout should be proposed and compared with

the on’yomi-based layout.

8.1.3 Subjective Evaluation and Participants’

Comments

The subjective evaluation showed that the bushu-

based layout was of high novelty. However, all the

other ratings were low, and in particular, those of per-

spicuity and efﬁciency were very low. We think that

this is because of the complicated layout that is difﬁ-

cult to use without learning. We also think that this

is because the user evaluated the bushu-based and the

on’yomi-based layout by comparing them.

The on’yomi-based layout was higher than the

bushu-based layout in all the items except novelty.

Still, other than stimulation and novelty, it was rated

low. One major factor is the short amount of time

that the participants spent using the proposed meth-

ods. However, in order for users to use it for a long

time, the proposed methods should give a good im-

pression at the beginning. The goal is to improve

each item to the point that cause no large difference

between the proposed methods and the existing meth-

ods.

The participants’ comments mainly indicated the

A Flick-based Japanese Tablet Keyboard using Direct Kanji Input

difﬁculty of inputting with the proposed methods.

The comments included “It is like a mental exercise”,

“I’m tired”, and “It is not for me”. The on’yomi-based

layout had more positive comments than the bushu-

based layout.

8.2 Long-term Experiment

According to the result of the long-term experiment,

the input speed of the bushu-based layout after 12

months was about the same as that of the QWERTY

keyboard. The long-term experiment was conducted

by one of the authors alone, and therefore it is not ob-

jective. However, it shows that the input speed of the

bushu-based layout could improve with a long-term

use.

Based on the author’s experience, we think that

there are three stages of growth in the input speeds.

The ﬁrst is to learn the placement of bushu elements.

In other words, the user can remember the locations

of bushu elements without referring to Figure 5. The

next stage is to identify bushu elements of kanji char-

acters whose bushu is confusing. For example, a kanji

character is confusing if it has more than one bushu-

like element. The ﬁnal stage is to learn the location of

a kanji character. After this, the author knows where

frequently used kanji characters are located. We ex-

pect that the number of such kanji characters will in-

crease as the author further continues it.

The error rate of the bushu-based layout after 12

months is still higher than that of the QWERTY key-

board. There are several reasons. The ﬁrst is due to

a mistake in the bushu element of a kanji character.

Because of the speciﬁcation of the input, if the user

makes a mistake in the bushu, a kana character will be

input. The error rate also increases if the user makes a

mistake in the kanji character itself. Since some kanji

characters have similar shapes, users may make in-

putting errors. These problems might be solved after

a longer-term use.

We have not yet compared the bushu-based layout

with other methods with a long-term use. Especially,

it would be desirable to include the on’yomi-based

layout in the experiment.

9 CONCLUSIONS AND FUTURE

WORK

This paper proposed a ﬂick-based tablet software key-

board using direct kanji input. There are two types of

layouts: bushu-based and on’yomi-based. The results

of the comparative experiments showed that the pro-

posed methods were not as good as the existing meth-

ods, even if the hint function was provided. Since the

kanji direct input method generally takes long time

to learn, one of the authors by himself conducted a

long-term experiment. After 12 months, both the in-

put speed and the error rate improved. However, they

did not surpass those of the existing method.

Future directions include improving error rates,

improving ﬁrst impressions, evaluating learning efﬁ-

ciency, and exploring more thorough ways to evaluate

performance. We aim to improve both the error rate

and the ﬁrst impression by further examining the lay-

out. The evaluation of the input speed requires mul-

tiple users who use the proposed methods for a long

time. Therefore, by distributing an improved version

of the proposed methods, we hope to acquire users

who will use them for a long time.

REFERENCES

Apple Inc. (2019). Split and move the keyboard on your

iPad. https://support.apple.com/en-us/HT207521.

Bi, X., Chelba, C., Ouyang, T., Partridge, K., and Zhai, S.

(2012). Bimanual gesture keyboard. In Proc. UIST,

pages 137–146.

Fukatsu, Y., Shizuki, B., and Tanaka, J. (2013). No-look

ﬂick: Single-handed and eyes-free Japanese text input

system on touch screens of mobile devices. In Proc.

MobileHCI, pages 161–170.

Hakoda, H., Fukatsu, Y., Shizuki, B., and Tanaka, J. (2013).

An eyes-free kana input method using two ﬁngers for

touch-panel devices. IPSJ SIG Tech. Rep., 154(6):1–8.

In Japanese.

Hasegawa, A., Hasegawa, S., and Miyao, M. (2012). Char-

acteristics of the input on software keyboard of tablet

devices: Aging effects and differences between the

dominant and non-dominant hands for input. J. Mo-

bile Interact., 2(1):23–28.

Ichimura, Y., Saito, Y., Kimura, K., and Hirakawa, H.

(2000). Kana-kanji conversion system with input sup-

port based on prediction. In Proc. COLING, volume 1,

pages 341–347.

ISO (1989). Documentation—Romanization of Japanese

(kana script). ISO 3602.

Kanev, K., Oido, I., Yoshioka, R., and Mirenkov, N. (2012).

Employment of 3D printing for enhanced kanji learn-

ing. In Proc. HCCE, pages 165–170.

Knox, K. (2012). Designing the Windows 8 touch

keyboard. https://docs.microsoft.com/en-us/archive/

blogs/b8/designing-the-windows-8-touch-keyboard.

Li, G. and Li, Y. (2019). Chinese Pinyin input method in

smartphone era: A literature review study. In Proc.

HCI International, volume 11569 of LNCS, pages 34–

43.

Liu, C.-L. and Lin, J.-H. (2008). Using structural infor-

mation for identifying similar Chinese characters. In

Proc. ACL HLT, pages 93–96.

HUCAPP 2021 - 5th International Conference on Human Computer Interaction Theory and Applications

Nakamura, Y. and Hosobe, H. (2020). A Japanese bimanual

ﬂick keyboard for tablets that improves display space

efﬁciency. In Proc. VISIGRAPP, volume 2, pages

170–177.

Niu, J., Zhu, L., Yan, Q., Liu, Y., and Wang, K. (2010).

Stroke++: A hybrid Chinese input method for touch

screen mobile phones. In Proc. MobileHCI, pages

381–382.

Odell, D. (2015). On-screen keyboard: Does the presence

of feedback or tactile landmarks improve typing per-

formance? In Proc. MobileHCI, pages 131–136.

Plecher, D. A., Eichhorn, C., Kindl, J., Kreisig, S., Winterg-

erst, M., and Klinker, G. (2018). Dragon tale—a se-

rious game for learning Japanese kanji. In CHI PLAY

Extended Abstracts, pages 577–583.

Sakurai, Y. and Masui, T. (2013). A ﬂick-based Japanese

input system for a QWERTY software keyboard. IPSJ

SIG Tech. Rep., 154(5):1–4. In Japanese.

Sax, C., Lau, H., and Lawrence, E. (2011). LiquidKey-

board: An ergonomic, adaptive QWERTY keyboard

for touchscreens and surfaces. In Proc. ICDS, pages

117–122.

Schrepp, M., Hinderks, A., and Thomaschewski, J. (2017).

Construction of a benchmark for the user experience

questionnaire (UEQ). International Journal of Inter-

active Multimedia and Artiﬁcial Intelligence, 4(4):40–

44.

T-Code Project (2003). T-Code laboratory.

http://openlab.ring.gr.jp/tcode/index.html. In

Japanese.

Takeda, Y. (2019). A survey of the occurrence of kanji char-

acters. In 4th Meeting of the Working Group on the

Standard of Japanese Education. Japanese Agency for

Cultural Affairs. In Japanese.

Takei, K. and Hosobe, H. (2018). A 2-by-6 button Japanese

software keyboard for tablets. In Proc. VISIGRAPP,

volume 2, pages 147–154.

Tamaoka, K. (2014). The Japanese writing system and lex-

ical understanding. Japanese Language and Litera-

ture, 48(2):431–471.

Waragai, H. (2008). Learning Joyo-Kanji by Reading One

Sentence. Goto Shoin. In Japanese.

Yajima, T. and Hosobe, H. (2018). A Japanese software

keyboard for tablets that reduces user fatigue. In Proc.

COMPSAC, pages 339–346.

A Flick-based Japanese Tablet Keyboard using Direct Kanji Input