Kanji Storyteller

Kanji Storyteller: A Sketch-Based Multimedia Authoring and
Feedback Application to Reinforce Japanese Language Learning
Ross Peterson, Sung Kim, Ke Wang,
Gabriel Dzodom, Francisco Vides, Jimmy Ho, Hong-Hoe Kim
Department of Computer Science and Engineering, Texas A&M University – College Station
We present Kanji Storyteller, a kanji sketch recognition
system to educate and entertain students through
constructing a storyboard space. The current kanji set
includes symbols representing concrete objects that can
be included as actors, sprites or objects within the story
space. The story space can be annotated to allow the
author to create a story about the space they created.
When using the story space, the author selects a spot to
enter kanji and is presented with a kanji insertion popup
window. The user confirms their entry manually so the
system can attempt to recognize the kanji written and
present corresponding images or “stickers” that represent
the kanji’s actual meaning. We implement recognition of
sketched kanji and kana with 3 different template
matching methods; we use a new method for online kanji
recognition that we call, “Stroke by Stroke." We also use
the $p recognizer for kanji, as well as a modified $1
algorithm for kana recognition. Kanji Storyteller’s
storyboard space is a medium for educating users on the
meaning of traditional kanji as well as their transcription
and stroke-order.
sketch recognition, educational software, new media,
Japanese kanji
Novice students of the Japanese language must not only
learn and understand two separate syllabaries, but many
must also struggle to remember the pronunciation, stroke
order, radicals, and form of the Chinese characters used in
Japanese called, “Kanji”. Traditional studying methods
emphasize mnemonics, discipline, and repetition to
master kanji use. When these traditional methods fail,
many students stop studying altogether. As such, we have
focused on creating a fun and engaging system that allows
students to construct a kanji study space and create their
own stories and mnemonics, as well as immerse
themselves in an engaging set of visuals. By mapping
kanji sounds and visuals to the writing of the kanji, we
hope to improve the student’s ability to recall the kanji.
This application also enables students to simultaneously
study and express their own artistic visions within their
new understanding of the character set. The interface also
features a jukugo (compound kanji words) and radical
(commonly recurring parts of kanji) recognition system so
the student can explore those facets of kanji study. Kanji
Storyteller is also planned to feature a social panel in the
interface that will act as a common discussion space
which may provide opportunities for the students to
reinforce their learning by discussing mnemonics or
discussing the annotation and stories attached to a given
We have many educational applications that support
sketch-recognition in many fields.
Mechanix is an educational system for engineering
students to draw structure sketches on sketch pads
[Field11]. Students and instructors can draw their trusses
and free-body diagrams, along with other shapes. The
benefit of Mechanix is that the system provides feedback
to students when they want to check if their sketches are
correct. Additionally, Mechanix automatically scores the
students' assignments. The system's recognizer uses a
geometrical approach similar to LADDER [Hammond05].
Most of the shapes for engineering class have many
numbers of lines. For example, an arrow has a one
horizontal line and two small lines. This approach allows
Mechanix to recognize many kinds of composite or
complex shapes successfully.
iCanDraw is an application that teaches users how to
draw a human face [Dixon10]. The application shows
steps to draw faces and it provides instructions on which
shapes the users need to draw. The feedback in iCanDraw
is controlled by interactions conducted in a step-by-step
manner, and additional feedback is provided when the
system detects completion of a face sketch.
Application to teach how to draw characters
Taele introduced an application that teaches how to draw
the East Asian characters: Korean, Chinese, and Japanese
[Taele10]. The recognizer uses a low-level recognizer, i.e.
PaleoSketch, to analyze primitives in shapes [Paulson08]
and a high-level recognizer, which represents rules of
primitives to recognize the shapes. To teach how to draw
the characters, the application shows steps that users need
to draw. After the users finish their drawings, the
application summarizes the scores of their drawings.
Developers of educational software utilize a wide
assortment of multimedia to provide entertainment value
as a way of capturing the interest of students. Such genres
of educational software applications that blend
educational elements with entertainment are often labeled
as edutainment or learn to play[Rap06]. Positive
contributions of edutainment software include research
studies that have identified those core attributes that make
them ideal for educational purposes[Khi08]. Among those
positive contributions are the capabilities of entertaining
software to keep the attention of the learners, retain their
flow of concentration, and promote the development of
good learning by stimulating creativity (learn by doing).
While there are many kanji sketch recognition
applications, dictionaries, and Japanese language studying
tools [Ren09] [TNJ12], our prior works investigation has
not revealed any software tools that feature kanji to image
and audio transformations that our system provides, nor
have we found any applications that supports the
authoring of stories based on kanji studying.
Figure 0. Screenshot of “Writing Test” from Renshuu
Renshuu [Ren09], a simple and free, sketch based study
application drills its users into writing a target kanji by
giving them a fill-in-the-blank prompt and a sketch area.
The user sketches in the target kanji and then chooses
reveal the actual target kanji by pressing a reveal button.
The revealed kanji features the stroke count and stoke
order of the kanji along with onyomi (“on” pronunciation
used in compound kanji words) in katakana syllabary
script and kunyomi (“kun” pronunciation used in single
kanji words) in hiragana syllabary script along with a few
common jukugo that the kanji is found in. No actual
sketch recognition is performed by the application,
instead, the user marks whether they wrote the kanji
correctly or incorrectly. The system records the selfreported hit and miss rate at the bottom of the application.
Renshuu also lets the student search kanji or English
words by their JLPT level.
Tagai ni Jisho (“With Each Other Dictionary”) [TNJ12] is
a free, open-source Japanese dictionary and kanji lookup
tool that provides detailed information that includes
nanori (pronunciation of kanji used in names), dictionary
lookup codes such as Heisig’s, 4 corner, and SKIP, as
well as component radicals for kanji and words of the
Japanese language. Tagai ni Jisho is particularly
noteworthy for its completeness, speed, and a feature that
allows students to mark which kanji they have studied so
as to limit repetition.
In scholarly publications on Kanji study and sketch
recognition, Taele and Hammond [Taele09] note the
difficulty of remembering how to write and read kanji for
Japanese students due to shape collisions and the nonphonetic properties, respectively, of the logographic
script. Their kanji sketch recognition system, Hashigo,
aimed to improve student recall by acting as an interactive
instructor that provided feedback on the sketch provided
by the student. Our system instead takes a different
approach to resolving the difficulty of remembering how
to write kanji by instead mapping the written kanji,
however poorly written, to an image analogue that
represents the kanji and providing the option to see the
kanji as a part of the original image. This reduces the
effect of shape collisions and similarity between kanji by
disambiguation and association with the idea that the
kanji represents. Furthermore, by placing the image on a
permanent canvas, the user is exposed to the kanji for
longer periods of time when compared to kanji (flash
card) drills where the kanji is only seen for a few brief
Lin et al.’s research is perhaps the most relevant to our
system in they constructed a collaborative story based
kanji learning system for tabletop computers where
learners congregate around the tabletop, sketch kanji and
annotations as well as arrange computer recognized
radical cards that combine into larger kanji [Lin09]. Their
collaborative story construction table top system is what
has inspired us to create a networked desktop/laptop
analogue where students could share their stories as well
as search and construct kanji by the selection of radical
components. Given the current adoption rate of tabletop
systems, the fact that our system caters to owners of
laptops and desktops means that our system can reach a
much larger audience. Furthermore, our system does not
even necessarily require touch screens, cards, or camera
trackers which lowers the hardware cost of when
compared to Lin’s system. The Kanji Storyteller system
also features a dictionary that allows in depth studying of
the kanji that Lin’s system does not provide.
The graphical user interface for Kanji Storyteller is
divided into two views with 3 main panels:
Drawing Panel / Canvas (Right)
Dictionary Panel (Left) – Compose View Only
Story Panel (Bottom Center)
Social Panel (Right) – Explore View Only
Figure 1. Kanji Storyteller GUI
The drawing panel is the canvas where students can
construct their scene with the kanji and their associated
static images. Instead of only drawing pictures on the
canvas directly, clicking on any point on the drawing
panel creates a pop up with an area where they can draw a
kanji. If the kanji they draw is contained in our content
database and recognized by our system, an associated
sticker image can be chosen from the recognition popup
menu and placed on the screen. For example, if the user
draws the kanji of a dog (犬), the user can select from
several kinds of dogs and place that dog on the screen at
the point it was drawn. In this way, the users are be
rewarded when they correctly draw a kanji by having the
desired sticker in the canvas.
The dictionary panel allows the student to see an
overview of all kanji in the language and see the range of
possibilities that they can draw from to create their story
space. The dictionary allows the student to look up the
details of a particular kanji of interest so that they can
master it. The dictionary panel also allows students to sort
the fields and study the kanji along that field of interest.
For example, in preparation of the Japanese Language
Proficiency Test (JLPT), a student may want to sort the
JLPT level field and study all JLPT kanji starting from
level 4 up to 1. Finally, the dictionary panel also has a
search form that allows the student to find details on a
particular kanji directly rather than searching manually
through the table. This panel is visible in compose view.
The story panel is the section of the user interface that
enables the user to create annotations to supplement the
graphics on the canvas. We believe the companion
annotations furthers the user's understanding of context of
the kanji he has studied. On the user interface, the story
panel is directly below the drawing panel; it supports
basic formatting features like most standard editors.
The whole system is planned to be part of a social
network where users exchange feedback on stories and
handwritings shared by their peers. The social panel is
designed to support this interaction. Located at the right
side of the drawing panel, the social panel comprises two
major components: the display component that renders all
feedback connected to the story in context, and the entry
component that enables the user to write and post
The interface also features a control panel that acts as the
main menu for the interface. From left to right, the control
panel contains a file command button group, a sticker
depth position button group, a radical, jukugo, and kanato-word ‘grouped search’ button group, clear stickers and
clear background buttons, a help button, and a view
selection dropdown menu.
The file command button group consists of a new button,
a load button, and a save button. The sticker depth
position button group features a forward button, a
backward button, a to-front button, and a to-back button.
Since one of our major goals was to provide a complete
study suite for kanji, it was important that Kanji
Storyteller would have a database that includes at least the
most commonly used kanji that describes a lower limit for
fluency in Japanese. However, even this lower limit may
be too limiting for students who are passionate about
studying kanji. Therefore, we thought it would be ideal to
include close to all of the kanji in the Japanese language.
In line with this thinking, we chose to start with the
KANJIDIC database [KDIC] which is maintained by the
Electronic Dictionary Research and Development Group
at Monash University and contains 6355 kanji.
From this original database, we populated a Microsoft
Excel spreadsheet that is read into the Kanji Storyteller’s
dictionary panel application through the Java Excel API
[Jaxl]. We chose to use an excel document as our storage
structure with Java Excel API to access it since our kanji
database is mostly static (i.e. kanji are not frequently
added to the language) and the Java Excel API was well
documented and provided easy access to the excel
The database file itself contains only the information most
relevant to Japanese students: meanings, pronunciations,
as well as a few statistics and indicators on the difficulty
and frequency of use of the kanji. While the
KanjiDictionary spreadsheet is useful for storing text
based information, it is less than ideal for storing images
and audio that we plan to include for kanji that the
recognizer can transform. As such, we have constructed
several separate property text files that specify the relative
links to a local 1st level cache of the associated audio and
sticker/kanji images. Since only a relatively small number
of kanji are recognized by the recognizer, the costs of
maintaining and updating these separate associative
property files and image cache is relatively small. The
property files also have the advantage of pointing to a
small cache of easy to find and frequently used images. A
similar approach for was taken for specifying the audio
and images for the jukugo and radical recognition data
storage. The layout and operation of the property files is
discussed in more detail later.
A second level of caching has also been undertaken by
using the Google Web Search API search interface to
retrieve images for all 6355 kanji and all katakana and
hiragana. In total, this cache is about 15 gigabytes in size.
The second level sticker image cache did have a few kanji
that the Google based image caching failed to find images
for. So when the system attempts to retrieve an image for
a kanji at runtime, it first checks the 1st and 2nd level
caches for an image using the kanji itself as the search
key. If the lookup fails, the system uses the Google Web
Search API to retrieve an image dynamically at runtime.
Given our long-term goal to expand the recognizer to all
kanji has not yet been reach and our goal to supporting
learner creativity, any kanji is not yet sketch recognizable
can still be used to create sticker images by going to the
kanji’s dictionary entry and pressing a “get images from
internet” button. The button dynamically retrieves a set of
images using the Google Web Search API.
Since many results can be returned on any given radical,
jukugo, or kanji lookup, the significance of the caching
method above is only further stressed. With many results,
the recognition menu must populate and scale many
images. However, since the user is actively waiting for
the results from their sketch, the time to populate the
recognition menu had to be kept to a minimum.
Therefore, we implemented a ‘sliding window’ feature for
the recognition menu where only the first 5 images are
returned for all the kanji that are found. The remaining
kanji info is sent to the recognition menu where it can be
used to perform the image cache lookups and populate the
remaining kanji images if the user demands those results
and their associated images.
Figure 2. ‘Sliding Window’ Recognition Menu that
feeds the user more images on demand.
Drawing Panel
At the center of our interface we have our drawing panel.
This is the place where the main interaction takes place. It
is a blank area or canvas where the user authors a scene
by setting a background and placing stickers on it. In this
sense, it is no different from PowerPoint or similar slidepresentation authorship tools. However the way stickers
are put on screen is novel and unique in that it is designed
as a teaching tool for people who are learning kanji. The
input method for converting written Kanji to stickers is
explained below.
Figure 3. Compose view interface for Kanji Storyteller
with sketch input popup and recognition menu.
Sketch Window
Perhaps the most intuitive way of using the white canvas
would be to simply draw on it and have the system
automatically interpret drawings. But this poses several
complications such as finding out when the user has
stopped drawing to signal the end of a kanji as compared
to when the user has simply paused. In what spatial scope
should the recognizer work? When would the recognizer
determine when characters are a part of multiple kanji
jukugo? These are research problems in their own. In
order to keep the implementation feasible within the time
scope, we opted to give the user the power to explicitly
tell the system when they are done drawing. As soon as
the user clicks on any point on the drawing canvas the
sketch window pop up shows a small area where the user
can write the kanji using a stylus or mouse. Each of the
user’s input strokes are processed by the Stroke by Stroke
recognizer to give the user possible kanji as the user
writes. More details on the Stroke by Stroke recognizer
will be given in the Sketch Recognition section.
If the user does not see their desired result in the list given
to them by the Stroke by Stroke recognizer after they are
done sketching, they can either click the ‘recognize kanji’
or ‘recognize kana’ button to attempt to have their sketch
recognized all at once by the $p dollar or modified $1
recognizer, respectively. The result stickers are given on a
list next to the window showing the most likely symbols
recognized where the user can pick which one they
Dictionary Panel
The dictionary panel is populated from our
KanjiDictionary excel spreadsheet. Entries can be
selected and the details for that kanji can be viewed. The
table shows which kanji can be recognized through sketch
input and displayed by the recognizer. The table also
allows the student to look up the details of a particular
kanji of interest and sort the database along a parameter of
interest. If, for example, in preparation of the Japanese
Language Proficiency Test, a student wants to study all
JLPT level 1 kanji, the student can sort the JLPT level
field. The student could do this by clicking on the
associated header to sort in descending order. Finally, the
dictionary panel also has a search form that allows the
student to find details on a particular kanji directly rather
than searching manually through the table. Searching
works by finding a matching word within each cell in a
row. If there’s an input match in that row, the row is
selected and the details panel opened, replacing the
dictionary panel. If multiple kanji are found in the search,
then multiple dictionary entries are returned in order of
their frequency of use. The user can cycle through the
kanji with left and right navigation buttons. Searching
with both Japanese and English characters is supported.
Story Panel
The story panel includes the formatting controls and a text
area for writing the story. First, the formatting controls
located at the upper right of the panel allow users to
change the font, font color, and font size as well as to
bold, italicize, and underline text.
Second, the text area occupies most of the story panel and
is below the formatting controls. In this text area, the user
can enter their annotations or story.
Figure 1. The Kanji Dictionary Panel.
Details Panel
The details panel (Figure 3) shows several important
fields. It displays the kanji being studied in large and
easy-to-read KaiTi font that mimics good kanji
penmanship. Below the kanji is the English meaning for
that kanji. Below the kanji are boolean checkmark images
that indicate whether that kanji is a kokuji, whether the
kanji has local sticker content available, and whether the
sketch recognizer can recognize the kanji. To the right of
the kanji is a group of statistical information for the kanji
that includes the stroke count, JLPT level, grade level and
frequency of use ranking for the kanji. In the middle of
the details panel is the readings group that shows the
kunyomi, onyomi, and nanori pronunciations. Below the
readings panel is the jukugo section where the compound
words that include the kanji are provided for kanji case
studies. At the bottom right corner of the details panel is a
“Place on Canvas” button that brings up a popup window
that allows the user to choose an associated image to
place on the canvas. As previously mentioned, the details
panel also has a “Get Online Images” button to retrieve
images dynamically. The learner can also view the stroke
order (also retrieved dynamically) by pressing the “View
Stroke Order” button. Finally, there is a “Return to
Dictionary” button that hides the details panel and returns
the user to the dictionary panel.
Figure 4. Annotation Panel
Comments Panel
The current interface or the comments panel is a proof-of
concept/placeholder GUI that behaves like a feed reader
where the feed is the stream of comments related to the
story in context. The panel renders the stream as a vertical
list of comment cells sorted by date in descending order.
Each comment cell contains the commenter's username
and the comment's posted date at the top. The rest of the
comment cell displays the content of the comment. This
is shown in Figure 4.
Figure 5. Comments Panel showing comments related
to the current story.
Kanji and Kana Sketch Recognition
Sketch recognition generally can be divided into three
categories: gesture-based recognition [Rubine91][
Jacob07], vision-based recognition [Kara05], and
geometric-based recognition [Hammond05]. We chose a
vision-based recognition approach for our recognizers
which include $1, $p, and our proposed Stroke by Stroke
Figure 2. Details Panel showing important information
of selected kanji.
It was initially planned that a single recognizer would be
used as a general purpose recognizer for both kana and
kanji. As more symbols and training samples were added,
however, several problems became apparent:
1) Adding more symbols makes it less likely for the
recognizer to distinguish between those classes
2) Many symbols share the same parts or look
Ex: 刀 (blade), 力 (power), カ(‘KA’)
This observation makes sense given that the kana
were originally created by making them look
similar to kanji that have the same pronunciation.
Figure 6. Similarity of kana to parent kanji.
3) More training samples increases the cost of
Given the high degree of visual collisions between kana
and kanji, the increased time and accuracy cost of having
a single input symbol recognized between two distinct
classes of symbols, and the preexisting interface
affordance for all-at-once user initiated recognition, it was
decided that the recognition of kanji and kana should be
separate in the interface.
Much research has been dedicated to the recognition of
multi-stroke symbols that echoes our third observation.
For example, Wobbrock’s $N recognizer [Anthony10]
translates the multi-stroke symbols into the one-stroke
symbols. Unfortunately, $N recognizer has an exponential
time complexity. In fact, the recognizer needs O(n * S! *
2^S * T) time, where S is the number of strokes, T is the
number of templates, and n is the number of sampled
points. This time complexity is problematic for Japanese
characters which commonly consist of half a dozen to a
dozen strokes or, in the most extreme case, 84 strokes
(when properly written).
Given that the computational cost can be prohibitive for a
large number of template classes, up to 6355 classes for
kanji alone, we decided to use template based recognizers
that have low time complexity in execution. The most
reasonable approach would be to eliminate computation
where possible and reduce the set of templates that have
to be traversed in order to produce the kanji that the user
desires. This is the main motivation for our Stroke by
Stroke recognizer. It uses online recognition to
progressively reduce the set of kanji as each stroke is
given to it by the user.
Figure 7. CJK Strokes and Classification Groups
(Adapted From [CJKWiki13])
The user’s strokes are monitored for ‘Mouse Up’ events
that are used to determine the ending of a single stroke.
Each individual stroke is sent to the $p dollar recognizer
to be classified as one of the 37 unique CJK (China,
Japan, Korea) calligraphic strokes. Once the stroke is
classified, the CJK stroke’s classification group is
determined which is in turn used to determine which kanji
contain that kind of stroke for the current stroke number.
Any kanji that does not have that stroke appearing in the
correct order is eliminated from the candidate list and the
remaining kanji are returned as results for the user to
choose from. For example, if the user draws a horizontal
stroke as their first stroke, the Stroke by Stroke recognizer
returns kanji containing a horizontal stroke as their first
stroke, which would include kanji like: 土, 木, and 本.
Any further strokes made by the user would progressively
thin the result list. If the user enters more strokes than
exists in a given kanji, it is removed from the list.
Many of these CJK strokes were grouped together into the
same classification groups due to their visual similarity to
the untrained eye. For example, any CJK stroke that is a
repeat of a previous stroke with a ‘G’ (stylistic tail)
attached were grouped together. Other CJK groups were
formed from their visual similarity as determined by the
$1 and $p recognizers. Group 13 is one such group and
consists of CJK strokes N, P, D, and T. When written by
users within Kanji Storyteller’s small input sketch
window, the relative sizes between strokes N, P and D, T
were nonexistent which made distinguishing between
them difficult. A total of 11 distinct groups were made.
It is important to note that the grouping system we
implemented reduces our ability to distinguish between
similar kanji. As a result of our inclusive grouping of CJK
strokes, the Stroke by Stroke recognizer returns more
kanji than it would if each stroke were a different class.
At the interface level, this means that the user may have
to scroll through more results in the recognition window
to find their desired kanji. This problem is minimal
compared to the problems compared in other recognizers,
however, since any other recognizer would have to
classify between 6355 classes of kanji rather than just 27
unique strokes and would be more likely to return false
positives as the top results. When the stroke order,
number of strokes, and the classification groups are used
in concert, the Stroke by Stroke recognizer was able to
uniquely distinguish all possible of the given kanji
The $p recognizer was chosen as the recognizer for CJK
strokes through a grounded approach where the $p and $1
recognizers were evaluated for their “hit” accuracy on two
CJK stroke datasets. The same grounded approach was
taken for determining which recognizers would be used
for the kanji and kana data sets. The details of this
analysis is given the evaluation section.
As stated before, our goal was to minimize computation
time for symbol recognition while using a visual
approach. As such, $1 and $p were obvious choices for
their strong run time performances of O(n * T * R) and
O(n^2.5 * T), respectively. Note that the R term for $1 is
number of iterations used in Golden Section Search for
rotation invariance. Given that kanji and kana are only
‘valid’ in a single orientation, we were able to eliminate
this rotational term by implementing a modified $1 that
removes the rotation based scoring. $p did not need to be
modified since it was already rotation invariant. The
modified $1 recognizer finds the Euclidean distance
between the input and template after they have been
preprocessed, e.g., resampled to 64 points, scaled to a 500
by 500 square, and translated so that their to the origin.
Since kana are often multi stroke symbols, and since $1 is
meant to recognize single stroke symbols, the modified $1
algorithm simply concatenates the strokes to represent it
as a single stroke. This does mean that the user should
draw the kana in the right stroke order so that it will be
recognized. We felt that this was a reasonable restriction
given that one of Kanji Storyteller’s goals is to promote
good handwriting. The resulting time complexity of the
modified $1 algorithm is an impressive O(n*T), which is
linear with the number of input points rather than
quadratic like $p. The $p preprocessing is a slightly
different; it resamples to 32 points, scales the points of a
given sketch to the size of its bounding box (which
preserves the kanji’s shape), and translates the center to
origin. Details for the $1 and $p algorithms are provided
in [Jacob07] and [Vatavu12]
Figure 7. Examples of symbol templates.
Rejection Thresholds
For rejection, a set of empirically determined thresholds
were put in place for each recognizer. If the average
confidence of the highest scoring class is less than the
threshold, the Kanji Storyteller does not return any
results, buzzes the user, and sets the status label to inform
the user that rejection failed. The thresholds for each
recognizer were chosen by finding the unacceptable
misses that had the highest confidences. The rejection
thresholds values are as follows:
$p for kanji and CJK strokes: 1.5
$1 for kana: 0.6
It should be noted that $p scores are 0 or greater where 0
is a perfect match while $1 scores are from 0 to 1 where 1
is a perfect match.
No rejection is done for the results given by the Stroke X
Stroke recognizer. We left this step out to provide a more
inclusive set of results and to minimize the amount of
time the user spends waiting on the recognition menu.
However, this step could easily be implemented by
running $p on the examples for each kanji returned by
Stroke X Stroke and then rejecting any class (kanji) that
did not get a score higher than an empirically determined
Kanji Storyteller was formerly using the Tanimoto
coefficient and the modified Hausdoff distance to score
and classify kanji. However, after much testing we have
decided to use the $p recognizer for kanji recognition and
the modified $1 for kana recognition. The results and
justification for this switch is provided in the evaluation
and discussion section.
One benefit of visual based recognition of kanji, as
opposed to some gesture based recognition methods, is
that the recognition is not rotation invariant which would
otherwise cause confusion between some kanji like
“three” (三) and “river” (川). Another benefit is that most
visual recognizers are stroke-order invariant. Stroke-order
invariance allows the student to practice kanji and
seamlessly create scenes without being distracted by the
finer details of learning kanji. However, it may be
beneficial to enforce correct stroke-order to promote
penmanship. Unfortunately, visual recognition methods
tend to be sensitive to differences between training data
and sketch input data. In situations where there is a high
variability between acceptable instances of a given kanji,
recognition rates can falter unless the training data
adequately represents those differences in style. With
respect to kanji, there can be a high degree of variation.
Kanji can be written in different aspect ratios, with
differences in the ends of strokes with different brushes,
brush orientations, and writing pressure. Kanji can also be
written in “cursive” stylistic forms, or in more formal
blocky forms.
Kanji Database
This file contains 6355 kanji which were included the JISX-0208-1990 2 byte character set as specified by the
Japanese Industrial Standard [JISWiki][JISUni][JISN].
KANJIDIC’s file structure lists each kanji and its
information on a separate line with a mixture of English
respectively. For our purposes however, much of this
information was superfluous since it included fields for
over a dozen different dictionary indexes among multiple
other classification codes. Furthermore, the fields in
KANJIDIC were not included in any particular order per
key value for searching
Actual symbol
native Japanese reading; used
mostly for nouns and
Japanese interpretation of the
original Chinese pronunciation;
used mostly for jukugo (words
made of multiple kanji)
Pronunciation when used in
Number of strokes to write the
The Japanese Language
Proficiency Test level of the
Whether the kanji was
Whether there is image content
Whether the recognizer can
Name Key
A string identifier that links the
kanji to the sticker images,
audio, and kanjiImg content in
their respective properties
* on Grade Level (adapted from [1]):
drawn kanji into an image.
recognize and convert the
the "grade" of the kanji.*
available for the given kanji
than China.
The KanjiDictionary contains the following fields:
Field Name
originally made in Japan rather
Frequency of Use ranking for
the 2501 most used symbols
To reduce the information to a reasonable working set for
Kanji Storyteller as well as to have a file which was in a
easy to work with and format, we developed a parser in
java that output the desired fields, delineated by spaces,
into a separate file and which could then be imported into
an excel document called, “KanjiDictionary.xls”. The
parser also removed the indicators in front of each field
entry and consolidated the kunyomi, onyomi, nanori, and
English meanings into their own consolidated fields so
that they would be inserted into the appropriate column
when imported to the excel spreadsheet.
G1 to G6 indicates the grade level as specified by the Japanese
Ministry of Education for kanji that are to be taught in
elementary school (1006 Kanji). These are sometimes called the
"Kyouiku" (education) kanji and are part of the set of Jouyou
(daily use) kanji;
G8 indicates the remaining Jouyou kanji that are to be taught in
secondary school (additional 1130 Kanji);
G9 and G10 indicate Jinmeiyou ("for use in names") kanji
which in addition to the Jouyou kanji are approved for use in
family name registers and other official documents. G9 (774
kanji, of which 628 are in KANJIDIC) indicates the kanji is a
"regular" name kanji, and G10 (209 kanji of which 128 are in
KANJIDIC) indicates the kanji is a variant of a Jouyou kanji;
**On kokuji: This information was extracted from the KANJIDIC set of
English meanings fields.
Image, Audio, and Formula Property Files
The Kanji Storyteller application must regularly access
image and audio media files in order to populate
recognition popup menus where the user can select which
kanji and its associated images to place as stickers or set
as the background. The system knows which images and
audio to use by looking in property files that act as
hashmap where the kanji’s name key field acts as the
lookup key. The following are the list of property files
used and their specifications.
kanjiAudio – Specifies the file path to the .wav audio
file that is played when the kanji is pasted on the
drawing panel.
kanjiImages - Specifies the file paths to the
associated images for the kanji. The first file path
specified is to the image of the kanji itself in the
resources/content/kanjiImg directory. The following
file paths are the images with the kanji placed as a
visible tag in the bottom right hand corner of the
image. Finally, the file paths of the same images
without the kanji tag are included.
radical – Specifies the kanji (there could be many),
where this kanji is included as a radical. Each entry
consists of a formula where the kanji (as radicals) or
radicals on the left hand side are the other radicals
composing the target kanji and the right hand side of
the formula is the target kanji. The separation
between the left hand side and right hand side of a
formula is delineated by an underscore while
component radicals and kanji (as radicals) on the left
hand side are delineated by plus signs. An example
entry for kanji ‘ricefield’ would appear as follows:
jukugo – Specifies the kanji where this kanji is the
first kanji in the jukugo. Each entry consists of a
formula where the kanji on the left hand side, in the
order that they appear in the jukugo, are part of the
target jukugo word that the kanji represents and the
right hand side is the target word itself. The
separation between the left hand side and right hand
side of the formula is delineated by an underscore
while the kanji on the left hand side are delineated by
plus signs. An example entry for the word ‘queen’
would appear as follows: woman=king_queen
jukugoAudio – Specifies the file path to the .wav
audio file that is played when the jukugo is pasted on
the drawing panel.
jukugoImages – Specifies the file paths to the
associated images for the jukugo. The first file path
specified is to the image of the jukugo itself in the
following file paths are the images with the jukugo
placed as a visible tag in the bottom right hand corner
of the image. Finally, the file paths of the same
images without the jukugo tag are included.
twice. This approach reduced the time it takes to conduct
a radical search dramatically. Any formulas that remained
were used by taking the target kanji and using those kanji
as the keys to populate the popup selection menu where
the user could make their selection. As an example of
how this process worked, imagine that the user groups the
two kanji, each of which is the kanji for ‘tree’ ( 木 ),
executes a radical recognition/search. The first tree radical
is used as the lookup key which returns the following
formulas: table_desk, genius_lumber, tree_woods,
tree+tree_forest. The radical recognition/search algorithm
then uses the second tree that was selected to eliminate
the first two formulas and trim the remaining results
which narrows the results to _woods, +tree_forest. Since
there are no more selected kanji, the algorithm passes the
two target kanji, ‘woods’ and ‘forest’ from their formulas
to be used as keys to lookup their associated images and
While this approach file properties lookup approach was
fast and could identify the number of each kind of radical
in a given kanji, it had the disadvantage that the system
could only recognize kanji radicals from formulas that
were explicitly made in the file itself. Given the sheer
number of kanji (6355) and the number of radicals that
each could contain, this method was entirely too
restrictive. As such, Kanji Storyteller now takes
advantage of information retrieval methods to query the
online WWWJDIC dictionary’s backdoor API for
multiple radical searches. While this approach is slower, it
enables the user to lookup the radicals for any kanji. In
either method, because the radical search is reductive, the
more kanji that the user uses to search, the more narrow
the results are.
Radical Group Search
The Kanji Storyteller system allows students to group
selections of kanji that have been placed on the drawing
panel and use them as radicals (kanji parts) to search for
related kanji. As an example, the kanji for ‘fire’ is used as
a radical in kanji like ‘torch’, ‘farm’, ‘dry/parched’, and
‘flames/blaze’. In order to find kanji based on their
component radicals, the user middle clicks to group the
kanji to be included in the radical search.
In the previous version of Kanji Storyteller, the first
grouped radical would have been used as a key to lookup
formulas in the properties map file and the following
grouped kanji would have been used to reductively
narrow down the possible kanji that could be returned.
This was done by removing formulas if a selection kanji
is not used in the formula. However, if that kanji was
used, that kanji would have been removed from the
composing radicals in the formula so that it is not counted
Figure 8. Radical Group Search with 木+火
Jukugo Group Search
The Kanji Storyteller system also allows students to use
the same grouping mechanism, activated by the middle
mouse button, to select kanji to be used in a jukugo
search. With the jukugo, the order of the kanji matters.
For example, ‘sun’ (日) and ‘evening’ (夕) makes ‘night
fall’ ( 日 夕 ) which is pronounced ‘nisseki’ while
‘evening’ plus ‘day’ makes ‘setting sun’ (夕日) which is
pronounced ‘yuuhi’. The jukugo search mechanism
reflects this so that students must select kanji in the
correct order. To provide feedback to the user, the
grouping selection adds n-1 red boxes on the border of the
sticker to indicate that sticker’s order in the selection. The
jukugo recognition/search algorithm works by using the
first kanji selected as the lookup key for the jukugo
formulas. Then each formula is checked to see that the
following kanji were selected in the right order and are
the kanji that compose the left hand side of the formula. If
there’s a match, the target jukugo word is passed on to be
used as the key to lookup the associated audio and images
and populate the recognition selection popup menu.
Figure 9. Jukugo Group Search with 夕+日
Kana Group Search
Kana group searching works slightly differently than the
radical and jukugo group searches. In this case, the local
kanji dictionary is searched for entries that contain the
same kunyomi, onyomi, or nanori pronunciation as the
grouped kanji.
Figure 10: Kana Group Search with や+す+む
Save, Load, and New
In any authoring system, it is of utmost importance that
the author can save and share their works so that their
work can be viewed by others and so that self-expression
is not restricted. As such, we have used simpl
[Shahzad11] to serialize the sticker, canvas, and author
data into XML. Currently, files can be saved and loaded
from any directory. However, this may change in future
implementations should the system be exported to a
networked setting. In this case, the simpl serialization tool
should prove beneficial as it affords portability and
interoperability between languages and varying platforms.
Quantitative Kanji Sketch Recognizer Evaluation
A quantitative evaluation has been conducted to compare
the recognition rates between the Kanji Storyteller’s
Tanimoto & Modified Hausdorff (TaniHaus) recognizer,
$P recognizer, and the modified $1 recognizer. Given that
the interface returns the top three examples for the user to
choose from, we define an “acceptable accuracy” metric
as the ratio of input examples classified as the correct
kanji in the top three results. We also define a “hit” as an
input example that was correctly classified as the intended
kanji in top result. Consequently, an “acceptable miss” is
when the correct kanji is returned in as the 2 nd or 3rd most
confident result and an “unacceptable miss” as when the
correct kanji is not in the top 3 results. The correct kanji’s
position in the results list described as the “rank” where a
rank of 0 corresponds to a “hit”.
In the first part of this evaluation, the recognizers
distinguished between a mixed kana and kanji set that
consisted of 23 different kanji and 11 different hiragana
that end with the consonant “A” including “A” (あ) itself
along with “N” (ん). For this mixed kana and kanji set,
two users contributed 5 examples each for a total of 10
examples of each kanji which we will mark as S1. For the
other dataset, S2, a different user contributed 5 examples.
Both contributors had previous experience with writing
kanji; the contributor of the training data was a Chinese
language student while the other contributor was a
Japanese language student. The results are provided in
table R1.
In the second part of the evaluation, the recognizers
distinguished between an all kanji set that consisted of 51
different kanji, most of which were JLPT 1 level kanji.
The contributors for S1 were a Chinese language student
and a Japanese language student. The contributor for S2
was a Japanese language student. Each contributor
provided 5 examples for each kanji. The results are
provided in table R2.
In the third part of the evaluation, the recognizers
classified examples from all hiragana without dakuten and
handakuten (diacritic marks indicating that the sounds of
the kana should use the voiced voiced or ‘p’ plosive
sound, respectively). For S1, one user contributed 5
examples of each hiragana while another user drew 5
examples of each hiragana for S2. Both contributors had
previous experience with writing kanji, but the contributor
of S1 was a Chinese language student that had no
previous experiencing sketching hiragana. Conversely the
other contributor was a Japanese language student with
much experience in writing kana. The results are provided
in table R3
For all of the above evaluations, the contributors were
told to write all the kanji and kana in the correct stroke
In the fourth part of the evaluation, the recognizers
distinguished between the 11 CJK Stroke groups. For the
CJK stroke evaluation, we defined a “hit” as a correct
classification (True Positive) for the most confident result.
For the one CJK dataset, S1, we had one contributor, a
Japanese student, give 5 examples of each CJK stroke.
For the second CJK dataset, S2, a different Japanese
student also provided 5 examples of each CJK stroke. The
results are provided in Table R4.
Kana (11) and
Kanji (23)
Mod $1
0.84, 0.88
0.90, 0.85
0.78, 0.80
Hit Ratio
0.74, 0.75
0.86, 0.78
0.68, 0.66
Miss Ratio
0.10, 0.14
0.04, 0.07
0.10, 0.15
Miss Ratio
0.16, 0.12
0.10, 0.15
0.22, 0.20
Table R1: Mixed Kana and Kanji Results.
Kanji (51)
Mod $1
0.83, 0.88
0.83, 0.82
0.77, 0.82
Hit Ratio
0.68, 0.75
0.78, 0.78
0.63, 0.63
Miss Ratio
0.14, 0.13
0.05, 0.05
0.13, 019
Miss Ratio
0.17, 0.12
0.17, 0.18
0.23, 0.18
Table R2: Kanji Results
Hiragana (48)
No Dakuten or
Mod $1
0.60, 0.68
0.72, 0.79
0.50, 0.45
Hit Ratio
0.44, 0.49
0.57, 0.68
0.34, 0.28
Miss Ratio
0.16, 0.18
0.14, 0.11
0.16, 0.18
Miss Ratio
0.40, 0.32
0.28, 0.21
0.50, 0.55
Miss Ratio
0.04, 0.04
0.12, 0.12
Miss Ratio
0.04, 0.04
0.10, 0.08
Table R4: Mixed Kana and Kanji Results
Note: The results are given below are rounded to the second decimal
place. The variable ‘m’ is the number of resampling points used in
preprocessing. The results on the left in each cell is where S1 was the
training data. Results on the right side in each cell is where S2 was the
training data.
Discussion of Quantitative Results
It is interesting to note that the while the modified $1 does
produce a higher “hit” accuracy than the $p recognizer for
the kanji study, it did not perform as well with respect to
the overall “acceptable” accuracy. The results seem to
imply a tradeoff between hit accuracy and acceptable
accuracy. The reason for this perceived tradeoff is most
likely differences in stroke order between the input and
training samples. If the stroke order matches, then $1 will
likely give a good result, however, in cases where the
stroke order doesn’t match, the more robust $p can
compensate since it treats the points as “clouds” rather
than as a single consecutive set of points. Essentially the
direction invariance of $p makes it more robust.
Even though $1 had a higher hit ratio, we decided to use
the $p recognizer instead since the desired result was
likely to be included as a lower ranked result in the
recognition menu. In other words, if the result is not on
the top, then the user just has to scroll down to what they
were actually intending to draw. To make a more
grounded decision, it would be beneficial to perform a
user study evaluating users’ qualitative responses to
having their intended symbol be the top result as
compared to their result being further down the
recognition menu. It may be the case that having a
slightly more selective ‘all-or-nothing’ recognizer that
returns the top result more often would satisfy users more.
0.96, 0.96
Mod $1
0.90, 0.92
The $1 recognition had the highest hit rate for classifying
hiragana when comparing the 3 recognizers that we have
implemented. (Tani-Haus, $1, $p) We believe that the
simple nature of hiragana has allowed the $1 recognizer
to be very effective in this case. The stroke order for kana
is easy to remember and the shape and style of kana tends
to vary less form person to person which means that was
probably less error derived the sketch contributors. All of
the recognizers performed relatively badly with the
highest overall accuracy not even breaking 80% and with
the worst recognizer, the modified Tanimoto Hausdorff,
returning an unacceptable miss rate of 50%. We suspect
that the reason for the bad results was that one of the
contributors had never written hiragana before that point
resulting in irregular looking characters, even if they had
been drawn in the correct stroke order.
0.92, 0.91
0.77, 0.80
For the classification of CJK strokes, $p had the best
recognition rate of 91% hit rate. The $1 algorithm
Table R3: Mixed Kana and Kanji Results
CJK Stroke
Hit Ratio
provided an 83-85% hit rate in comparison. Since the
Stroke by Stroke algorithm only takes the top stroke result
to return a list of applicable kanji, the acceptable ratio was
not relevant and we only looked at the hit ratio. The $p
most likely outperformed our modified $1 due to the fact
that $p is stroke direction invariant since it processes
points as a cloud of points without any temporal
information. $1, on the other hand progresses through
strokes in order. Therefore, any strokes written in the
wrong direction would increase the probability of a stroke
miss. We also chose $p for recognition of CJK strokes
since it performs well on single stroke gestures and is
resistant to variations in sketches between different users.
We also conducted tests to see if the program was
providing us with all of the different possible kanji based
on the already entered strokes. Although we did not test in
an exhaustive manner, the program did appear to in
provide all the relevant kanji choices for a large set of test
set. Several iterations of accuracy evaluation were run for
the CJK Stroke classification to find the best groupings.
The results provided here are the results for the final
For all the quantitative evaluations, it was found that
modified Tanimoto Hausdorff recognizer was taking
longest, sometimes by a factor of 2 compared to
Combined with the fact that TaniHaus was returning
worst recognition results, it was easy to eliminate
recognizer from later evaluations.
Storyteller would be a beneficial tool for learning Kanji
but stressed that porting the software to a mobile platform
would be desirable. The professor also requested a way to
drill and test the students’ knowledge of the kanji that
they had studied from within the application.
In order to test the effectiveness of our program we
created a user interface evaluation for our project. Our test
sample population came from students taking Japanese
101, 102, and 201 at Texas A&M University. The
evaluation can be broken down into three different
sections: pre-questionnaire, user study, and postquestionnaire. The pre-questionnaire consisted of several
Likert scale questions that questioned the amount of time
the user spends studying Japanese as well as his/her
interest in learning Japanese. It also asked the user which
aspect of learning Japanese is most difficult for the user.
The format and questions used for the questionnaires can
be seen in the appendix.
The user study portion consisted of a set of tasks that the
user undertook while using the program. The tasks
included writing and searching for specific kanji and
The last portion of the study was the post-questionnaire.
This part of the study gauged how satisfied the users were
of our program itself. Half of the questions were open
ended as asked for comments for various aspects of the
program while the other half of the questions contained
Likert scale questions about the user’s satisfaction with
the software and their experience.
Interface Evaluation
Results and Discussion of Interface Evaluation
Before conducting an interface evaluation, we met with a
long-time Japanese professor at Texas A&M University.
In this meeting, we gave a demonstration of Kanji
Storyteller and subsequently interviewed the professor to
obtain a professional opinion on the quality of the
software as Japanese language teaching tool. The
professor claimed little experience in using any kind of
educational software for Japanese but did mention that
current students were using tablet and smart-phone
applications to provide on-the-fly lookups of kanji
information. When asked about what kind of study
methods were recommended to students for learning
kanji, we were told that repetition in writing kanji was
stressed. The professor gave positive feedback on the
completeness of the kanji dictionary and jukugo/radical
lookup features. When the professor was asked how we
would try to user Kanji Storyteller for his students, he in
return asked how we intended Kanji Storyteller to be used
as a study tool. We replied that we intended for the
software to be a supplemental study aid to reinforce
learning by providing visuals. We also explained to the
professor that the students could save and share their
canvases as .xml files. After some discussion, the
conclusion was made that it would be preferable to share
the xml files on a class server that the students could
access remotely. Overall, the professor agreed that Kanji
The professor’s feedback that repetition in was the
preferred study method for learning kanji was
substantiated by reports from the students that
participated in our user study and it would appear that
traditional rote repetition based study methods remain
prevalent. This partially validates Kanji Storyteller’s
claim to novelty as a kanji study tool. Kanji Storyteller’s
novelty, however, also presented unique challenges for
the users. Many users were unsure of how they would use
annotation feature to add to their learning experience.
Using the stickers to make scenes, while reported to
“cool”, entertaining, and memorable, was often seen as a
superfluous feature for studying.
The professor’s request to add a quizzing mechanism to
Kanji Storyteller was particularly significant and we have
come to the conclusion that such a mechanism would
structure the user’s learning and provide a set of both
short term goals that the user could aim for. We have also
decided that a quizzing feature would benefit from a user
progress tracking feature so that the learner could identify
kanji that they have encountered and know which parts of
their knowledge needs to be reinforced. This information
would further structure their studies and provide a set of
long term goals for learning Japanese kanji. From the
feedback we received form the professor, we added a
question to the post-questionnaire that asked whether a
quiz feature would be useful. The responses to this
question were overwhelmingly positive.
Our user study had 4 participants volunteer; 3 students
took the study in an observed laboratory setting while the
other participants took the study remotely and without
supervision. 3 of the students were Japanese language
students at Texas A&M University while another one was
a native Chinese speaker with interest in beginning to
study Japanese. Although our sample size was small, the
users all reported being satisfied with our program with
one exception; This exception was for a user who
performed the study remotely and most likely did not see
the last two questions which were on a different page. For
the pre-questionnaire, the average Japanese study time
was about 2-4 hours per week while the average kanji
study time was about 1-1.75 hours per week, which is
about 1/3 to 1/2 half of their study time. For the post
questionnaire, the average Likert scale score for kanji and
hiragana recognition satisfaction was 4 (usually
recognized correctly) while the ease of finding
information in dictionary had an average Likert scale
score of 4.5 (Easy to find info). The overall satisfaction
rating had an average of 5 (Very Satisfied). Based on the
Likert, results we can safely claim that our program was
successful in giving the users a novel and pleasurable
experience for studying kanji. Some interesting quotes
from the free responses included the following:
“Associating Kanji with images really helps
“Great way to discover new kanji”
“Good way to find kanji just by drawing”
“I like Kanji Storyteller because if I draw it
incorrectly, it doesn’t read it or reads a different
The above quotes help substantiate our claims that Kanji
Storyteller is a beneficial studying tool that helps by
reinforcing kanji through visuals, through unique
interactions, and through handwriting feedback.
On the negative side of the results, students were also
frustrated with being forced to use a mouse for sketching
the kanji and were suspicious that any misclassifications
by the recognizers were due to their inability to
adequately draw the kanji with a mouse. As such we
would like to add pen/stylus support for our application.
This commonly cited complaint also aligns well with the
professor’s recommendation to port Kanji Storyteller to a
tablet or smart phone since those devices natively support
touch based interaction.
It was also observed that the users wanted to drag-anddrop kanji directly from the dictionary panel onto the
drawing panel. Also, 2 of the users tried to use the
backspace key instead of the delete key to remove stickers
from the canvas. Another user reported that they wanted a
‘clear’ button (which had not been implemented at the
time). These demanded features were surprising from the
implementers’ point of view but were also simple to add
to the interface. This set of observations provided a
significant lesson on the value of performing an interface
Another major observation is that the grouping features
were largely ignored or were unknown to the users. 2 of
the users did middle click the stickers but did not press
any of the buttons to initiate recognition. Another user
requested that they would, “add a tutorial or some info
bubbles that pop up over the different buttons and
functions.” Based on these observations and feedback we
implemented to the help window, status panels, and ‘first
time popup’s to inform the user on how Kanji Storyteller
can be used without the need for intervention from a
trained observer. We also plan to add ‘popup buttons’
beneath the stickers when they have been grouped in
order to make a visual connection between the action of
grouping and the available jukugo, radica1s, and kana
lookup functions.
Proposed Long Term User Learning and Creativity
Our system’s intended purpose is two-fold. It serves as
both an educational tool and it also serves as a creative
authoring tool for visual mnemonics. One major idea
behind Kanji Storyteller is that the user can learn and
practice Japanese kanji while creating meaningful scenes
and annotations they can share with the community. Our
system also aims to enhance the writing skills of the
learner, but also to be a fun tool to use that can keep the
learner engaged at all times.
The proposed user learning evaluation we envision would
involve examining a set of Japanese students that are in
the stage of learning many new kanji characters in their
writing development. We would need to evaluate
attempting to answer the following research questions.
Do the writing and reading skills for Japanese
Kanji of the learner improve after use of our
system and to what degree with respect to time
of investment?
Does the system provide an intuitive and easy to
use authoring tool for visual stories?
Does the user find the system entertaining
enough to remain engaged into creating new
stories at the same time that they are learning?
Do students retain more with sounds?
Do students retain more with conversion to
images or visual representations?
Do students retain more by constructing a
canvas, scene, or story?
In order to accurately evaluate the system, it is important
that we define what it means to “learn” kanji. Considering
that not all kanji are used or encountered as frequently as
others, it follows then that less complete knowledge is
required from kanji that are used less. Consequently, our
definition of “learned kanji” varies with the JLPT level of
the kanji. Therefore, the student has learned a kanji when
the students knows:
JLPT 4 and Jouyou (common use) kanji:
• at least 90% of readings known
• all English meanings
• how to write
• at least 1 kunyomi (if applicable)
• at least 1 onyomi (if applicable)
• all English meanings
• how to write
• at least 1 reading (onyomi/kunyomi)
• at least 1 English meaning
• how to write
JLPT 1: (or above)
• at least 1 English meaning
• how to write
It is also important to define what it means to say that
students learn kanji “better” with Kanji Storyteller than
with more traditional methods. As such, we provide
several comparative aspects of kanji learning with each
aspect framed as a positive hypothesis for the
performance of our system. Each aspect is followed by a
possible means of measurement:
Difficulty: Students learn more difficult kanji on average
Record the JLPT level and grade level of the kanji
encountered, sketched, and posted by the users.
Speed: Students can write kanji faster
Record the average time to recall and write a kanji
Record the average time to complete a drill set of N kanji
Mastery: Students can recall more information for a given
set of kanji
Record the % recalled of: onyomi, kunyomi, English
meaning, jukugo
Breadth of Knowledge: Students know more kanji
Record the distribution of exploration vs difficulty of kanji
% Encountered versus mastered for a grade/JLPT level
Raw Frequency of kanji encountered and mastered
Readability: Students have better handwriting
Record and compare before and after samples of
Motivation: Students are more motivated to learn kanji
and related subjects
Questionnaire and surveys to get responses on:
Fun factor
What they got out of using it
Whether they want to study more
The actual evaluation we plan to implement would first
involve gathering a set of representative users for the
system consisting primarily of beginning Japanese
students. The first stage of the evaluation would involve
the observation and analysis of users’ qualitative
experiences with the free form use of the system. The
feedback from this first stage would allow us to refine the
look, feel, and operation of our program as well as give us
an understanding of how users approach and attempt to
use the system. The second stage of the evaluation would
entail a more rigorous quantitative analysis of user
knowledge of kanji and skill in reproduction after set
periods of Kanji Storyteller usage. A third stage of
evaluation should also be conducted to evaluate user
learning over a more representative use period over the
course of a semester. This could be accomplished by
measuring student knowledge of kanji based on the above
definitions of learning. The study participants should
ideally be members of the same section of a Japanese
course where one half of the students are placed in an
experimental group who study using Kanji Storyteller and
where the other half of the students are placed in a
control group that only uses traditional study methods.
There are a number of features and improvements that
could vastly change the experience with Kanji Storyteller.
In order to ease editing of the drawing panel, it would also
beneficial to be able to translate groups of stickers
simultaneously and move stickers relative with the
background when the size of the drawing panel is
It would also be beneficial to include freeform drawing
pens, brushes, and other direct manipulation editing tools
much like many image editing systems provide to allow
the authors to paint their own canvas background or
stickers and add annotations directly to the screen. Taking
this idea further, these stickers could also be dynamically
associated with a particular kanji, should the use wish it.
The feedback from the user studies and our goal to make
Kanji Storyteller an effective learning tool obligates us to
implement an quiz and testing feature to drill the user on
their knowledge. Extending this idea, we plan to
implement a learning tracking method that is integrated
with the quizzing and testing feature.
We also plan to implement a “hint” feature that gives tips
to the user when first opening Kanji Storyteller to inform
them on available features and make suggestions on
effective ways of studying and making mnemonics.
There was one particular goal of our interface that has not
been accomplished: Social network connection and
network deployment. This was a major piece of our
interface that would have eased the sharing of canvases
and promoted communication, expression, collaboration,
and constructive learning. Users are, however, currently
able to save and share their works by sharing the xml save
files. Making a port of Kanji Storyteller to mobile devices
would also meet reported usage patterns of kanji studying
Finally, the system could benefit from the inclusion of
more content in the form of more kanji for sketch
recognition, more audio and more flexibility in user
choice for the image content. Dynamic online retrieval of
audio files is difficult given the relative dearth of free
sound effects sites. More disconcerting, however, is the
problem of the disconnection between the metadata and
terms used to describe audio files and the terms that Kanji
Storyteller would be able to provide automatically. For
example, automatically retrieving sound clips for “fire”
would be more successful if the search term were instead
“crackling”, however Kanji Storyteller doesn’t have the
information to know to use the search term “crackling”
for “fire”. Sounds are usually tagged with a rich set of
adjectives while images are usually easy to find with
nouns. At any rate, more diverse audio and image content
would enrich the experience by offering both breadth of
content and depth of constructive possibilities for creative
expression to reinforce learning.
While we have not yet conducted the long term user
learning and creativity evaluation, our interface does look
promising so far from its limited exposure to beginning
students. Authoring interesting and amusing scenes is not
only possible, but almost unavoidable. Furthermore, the
audio and visual feedback on placement is both fun and
memorable, reinforcing the meaning of the kanji without
effort. The database is more complete than most online
resources and convergent searching as well as and
divergent exploration of the kanji database is strongly
afforded by the interface. As a study tool, the Kanji
Storyteller could benefit greatly from a quiz and
progression tracking feature. Finally, the sketch
recognizer is both accurate with the results it retrieves
with $p and the modified $1 and proactive in fetching
those results in a short amount of time with Stroke x
[Rap06] K. Rapeepisarn, K. W. Wong, C. C. Fung, A.
Depickere, (2006). Similarities and differences between
learn through play and edutainment. Proceedings of the
3rd Australasian conference on Interactive entertainment
(pp. 28-32). Murdoch University. Retrieved from
[Khi08] M. Khine,(2008). Core attributes of interactive
computer games and adaptive use for edutainment.
Transactions on edutainment I. Retrieved from
[Ren09] Renshuu, An application to practice drawing
Japanese kanji. Last updated on January 4, 2009,
Accessed and downloaded on January 17, 2012
[TNJ12] Tagaini Jisho, A free, open-source Japanese
dictionary and kanji lookup tool Accessed and
downloaded on January 18, 2012
[Taele09] P. Taele, T. Hammond, Hashigo: A NextGeneration Sketch Interactive System for Japanese
Kanji. Proceedings of the Innovative Applications of
Artificial Intelligence 2009, Texas A&M University,
Department of Computers Science, retrieved from
[Taele10] P. Taele, “Freehand Sketch Recognition for
Computer-Assisted Language Learning of Written East
Asian Languages,” Master’s thesis, Texas A&M
University, 2010.
[Paulson08] B. Paulson, T. Hammond, “Paleosketch:
Beautification,” Proc. 13th International Conference on
Intelligent User Interfaces, pp. 1-10, 2008.
[Lin09] N. Lin, S. Kajita, K. Mase, Collaborative StoryBased Kanji Learning Using an Augmented Tabletop
System, JALT (Japanese Association for Language
Teaching) Call Journal, 2009 (pp. 21-44) retrieved from
[Rubine91] D. Rubine, (1991) Specifying gestures by
example. In SIGGRAPH ’91: Proceeding of the 18th
annual conference on Computer graphics and interactive
techniques 329-337.
[Wobbrock07] J. Wobbrock, A. Wilson, Y. Li, (2007)
Gestures without libraries, toolkits, or training: a $1
recognizer for user interface prototypes. In UIST ’07:
Proceeding of the 20th annunology 159-168.
[Vatavu12] R. Vatavu, A. Li, J. Wobbrock, (2013)
Gestures as Point Clouds: A $P recognizer for User
Interface Prototypes. In Proceedings of the 14 th ACM
International Conference on Multimodal Interfaction
(ICMI ’12). ACM, New York, NY, USA, 273-280.
[Kara05] Kara, L., Thomas S (2005) An image-based,
trainable symbol recognizer for hand-drawn sketches.
Computers & Graphics 29(4):501-517.
[Hammond05] T. Hammond, R. Davis, (2005) LADDER,
a sketching language for user interace developers.
Computers & Graphics 29(4):518-532.
[Anthony10] L. Anthony, J. Wobbrock, (2010) A
lightweight multistroke recognizer for user interface
prototypes. Proceeding GI'10 Proceedings of Graphics
Interface 245-252
[Field11] M. Field, S. Valentine, J. Linsey, T. Hammond.
2011. Sketch recognition algorithms for comparing
complex and unpredictable shapes. In Proceedings of the
Twenty-Second international joint conference on
Artificial Intelligence – Volume Three(IJCAI'11), Toby
Walsh (Ed.), Vol. Volume Three. AAAI Press 2436-2441.
[Dixon10] D. Dixon, M. Prasad, T. Hammond, (2010)
iCanDraw: using sketch recognition and corrective
feedback to assist a user in drawing human faces.
In Proceedings of the SIGCHI Conference on Human
Factors in Computing Systems (CHI '10). ACM, New
[Shahzad11] N. Shahzad, S.IM.PL Serialization: Type
System Scopes Encapsulate Cross-Language, MultiFormat Information Binding, Texas A&M University
Masters Thesis, 2011.
[CJKWiki13] Stroke (CJKV character) Wikipedia Entry,
[KanaWiki13] Kana Wikipedia Entry, Accessed on: May
4, 2013; http://en.wikipedia.org/wiki/Kana
Have you ever studied Japanese?
a. Yes, I enjoy it
b. Yes, it’s ok
c. No, but I’m interested
d. No, I’m not into that
e. Other:
How regularly do you study Japanese?
(choose one)
 not at all
 less than an hour per week
 1 to 2 hrs per week
 2 to 3 hrs per week
 3 to 7 hrs per week
 more than 7 hours per week
 other:
How regularly do you study Kanji in particular?
(choose one)
 not at all
 less than an hour per week
 1 to 2 hrs per week
 2 to 3 hrs per week
 3 to 7 hrs per week
 more than 7 hours per week
 other:
Experimenter Guided User Study:
Play with the program and try out clicking the
canvas and the different buttons
Find a kanji you haven’t studied before and
study it in whatever way you would like to study
it, preferably your usual way.
Try sketching the symbol for “Tree” (木) with
the program
Try sketching the symbol for “Person” (人) with
the program
Try sketching the hiragana “KA” (か) with the
Make an interesting scene for yourself or to
show other people
Find as many kanji that have the “Fire” (火)
radical in it in 2 minutes
Find the kanji with the most strokes
Write down as much info as you can remember
about the kanji you studied in #2.
How do you study Kanji when you do study?
What difficulties do you usually have when trying to
use kanji?
(choose all that apply)
 Remembering how they look
 Remembering what they mean
 Remembering the right stroke order
 Remembering how to pronounce them (onyomi / kun-yomi)
 Remembering the jukugo
 Other:
How would you use Kanji Storyteller for your own study
and use? (if at all)
Would you prefer to use Kanji Storyteller or some other
tool for studying Kanji? If not KS, which tool and Why?
How easy is it to find the information you are looking for
in the dictionary panel?
Very Easy
Kind of easy
Kind of hard
Very Hard
Would you make scenes/canvases that would help you or
other students study? If so, how?
Overall, are you satisfied with Kanji Storyteller?
Is there anything that you would add to Kanji Storyteller?
If so, why?
Is there anything that you would change or take away
from Kanji Storyteller?
Would adding a drill or quiz feature help you study? Why
or why not?
Do you usually get the Kanji you want after drawing it as
often as you would like?
Usually not
More often than not
Do you usually get the Hiragana you want after drawing it
as often as you would like?
Usually not
More often than not
Not at all
Somewhat satisfied
Very satisfied
Extremely satisfied
How likely are you to recommend our software to others?
Not at all
Somewhat likely
Very Likely