KanjiCompass: An Etymology-Driven Adaptive Kanji Learning Tool

Keywords:

Abstract:

Learning Kanji is a complex and critical component of Japanese language acquisition, requiring learners to

understand its semantics, morphology, and phonology. Traditional rote memorization methods often overlook

Kanji’s etymological and structural nuances, limiting their effectiveness. This paper presents an etymology-

driven, adaptive Kanji learning tool designed to visualize Kanji relationships, reduce cognitive load, and enhance

learner engagement. The tool features interactive graph visualizations, personalized learning recommendations,

and integration with Anki ﬂashcards for explorative, self-regulated learning (SRL). The tool was evaluated for its

usability and adaptivity in a ﬁeld study with 19 participants. Overall, the tool’s usability was well-received, with

the detailed Kanji graph and Anki integration being commended for their clarity and ease of use. Personalized

learning recommendations were particularly valued for providing adaptive and targeted learning paths. However,

the macro-level perspective provided by the overall graph was found overwhelming by some users. Results also

indicate that learning goal motivation strongly inﬂuenced engagement, with motivated users beneﬁting more

from the tool’s adaptive features. Key contributions include methods for visualizing interconnected knowledge,

recommendations for personalized learning paths, and supporting tools for encoding and retrieval stages.

1 INTRODUCTION

In 2021, about 3.79 million people worldwide studied

Japanese formally, plus countless self-learners (The

Japan Foundation, 2023). A crucial and challenging

component of mastering Japanese is its writing sys-

tem, particularly the use of Kanji, which plays a fun-

damental role in both literacy and deeper language

comprehension.

Kanji, a morphographic script integral to the

Japanese writing system, poses signiﬁcant challenges

to learners (Rose, 2017). In contrast to a phonographic

script like the Latin alphabet, where every character

represents a sound, characters of a morphographic

script represent a meaning. There are 2,136 Kanji

for regular-use issued by the Japanese government,

so-called j

o Kanji, which are deemed essential for

functional literacy in Japanese, covering the vast ma-

jority of Kanji seen in everyday texts. Of these, the

500 most frequent characters account for 75 % of all

Kanji occurrences in written Japanese (Crowley, 1968).

The script’s complexity lies in its multilayered com-

position of semantics, morphology, and phonology,

making it distinct from other writing systems. As

a simple example the Kanji

土

has the meanings of

’soil, earth, ground, Turkey’

(semantics), looks simi-

lar to another Kanji

士

(morphology), and can be pro-

nounced “tsuchi”, “do”, and “to” (phonology). Even

though Kanji are rooted in Chinese characters, they

differ signiﬁcantly. One Kanji can have several pro-

nunciations, several meanings, and is normally com-

bined with other Kanji or Hiragana—another Japanese

script—to form words. This combination of factors

needs to be considered when learning Kanji, making it

impossible to learn Kanji as one would learn vocabu-

lary. Unlike vocabulary, where one usually only learns

one meaning and pronunciation, Kanji often have sev-

eral meanings and pronunciation options that depend

on the context.

While some materials integrate structural or etymo-

logical explanations, many Kanji learning resources

still emphasize rote memorization, which has been

criticized for limiting the learner’s ability to grasp the

structural nuances of Kanji (Mori, 2014). Existing

methods, such as visual imagery or narrative aids, at-

tempt to enhance memorization but often overlook the

potential of etymology as a pedagogical tool (Rose,

2017). By leveraging etymology, where the origin and

development of a character is explored, learners can

https://jisho.org/, last accessed: 2025-02-11

Etymology-Based Learning, Self-Regulated Learning, Graph-Based Interfaces, Kanji Learning.

Sigrid L. Klinger

and Sven Strickroth

LMU Munich, Munich, Germany

{sigrid.klinger, sven.strickroth}@iﬁ.lmu.de

https://orcid.org/0009-0006-9277-8355

https://orcid.org/0000-0002-9647-300X

472

Klinger, S. L. and Strickroth, S.

KanjiCompass: An Etymology-Driven Adaptive Kanji Learning Tool.

DOI: 10.5220/0013341000003932

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Computer Supported Education (CSEDU 2025) - Volume 1, pages 472-483

ISBN: 978-989-758-746-7; ISSN: 2184-5026

understand not only the meanings and pronunciations

of individual Kanji but also discern meaningful rela-

tionships among characters, fostering a more cohesive

learning experience.

Etymology not only enables learners to grasp the

meanings and pronunciations of individual Kanji, but

also facilitates the identiﬁcation of connections be-

tween known and unknown characters. These inher-

ent interrelations between Kanji, often disregarded in

conventional teaching materials, are nonetheless rec-

ognized by learners, who frequently form their own

assumptions about these links (Rose, 2017). Unfortu-

nately, these assumptions can lead to misconceptions

and frustration, diminishing motivation and the overall

learning experience.

To address these challenges, this paper introduces

an etymology-driven, adaptive learning tool, which

answers the following research questions:

•

How can the structure of a Kanji itself and its re-

lation to other Kanji be visualized in a learning

tool to minimize the cognitive load needed to learn

Kanji?

•

How can adaptive recommendations for learning

paths be given to guide learners in deciding which

Kanji to learn?

•

How do learners perceive and use the tool’s visual-

ization and recommendation features, and to what

extent do these features motivate and engage them

in exploring and deeply learning Kanji?

By emphasizing the interconnectedness of Kanji

through their etymological roots, the tool aims to pro-

vide both, self-learners and formal students, with a

deeper understanding of Kanji, reducing cognitive load

and preventing erroneous assumptions. This techno-

logical solution presents an opportunity to shift the

focus from isolated memorization to a more integrated,

explorative learning experience, better suited to the

complex nature of Kanji acquisition. The tool is not in-

tended to replace conventional and established learning

methods but to complement them. The contributions

are: (1) a purposeful layout with navigation elements

aimed at minimizing cognitive load and optimizing

the learning process, (2) an approach to visualizing

the interconnectedness of a learning subject, and (3) a

method for providing personalized learning paths by

giving unobtrusive learning recommendations.

The paper is structured as follows: Section 2 pro-

vides background on Kanji learning, Section 3 reviews

related work, Section 4 presents the prototype, Section

5 details evaluation and results, followed by discussion

and future research.

2 BACKGROUND

To provide an understanding of the current state of

Kanji learning this section will provide an overview

on Kanji and research concerning Kanji learning meth-

ods. The Japanese writing system is comprised of

four scripts: Hiragana, Katakana, R

omaji and Kanji.

Hiragana and Katakana each consist of 45 unique char-

acters, while R

omaji is used to transcribe Japanese

words with the Latin alphabet. Kanji, however, do

not have a set number of characters. Some online dic-

tionaries comprise more than 28,400 Kanji

, but for

learners, the 2,136 j

o Kanji are of primary impor-

tance. To be proﬁcient in Kanji, one needs to know the

meaning of a Kanji, its pronunciation, and the ability

to recognize and write it.

Kanji can be classiﬁed into six etymological cate-

gories based on their historical origins, with three ma-

jor categories alone accounting for 97.5 % (Tamaoka

et al., 2017) of all Kanji. This etymological perspective

is valuable because it not only aids memory by anchor-

ing characters in meaningful contexts but also helps

learners recognize patterns and connections among

Kanji. Examples of the three main categories include:

高

(“tall”), originally depicting a tower (pictorial),

森

(“forest”), formed by three “tree” (

木

) components

(metaphorical),

際

(“border,” pronounced sai), com-

bining 阝 (“hill”) and 祭 (also sai) (phonological).

Figure 1: A visualization of the two graphemes

阝

and

祭

forming the Kanji 際.

While some characters represent a single entity,

most are composed of multiple graphemes, including

one radical. A grapheme is “the smallest unit in a sys-

tem of writing a language that can express a difference

in sound or meaning”.

An example in the English

writing system are the words “shake” and “snake” only

differentiable by the graphemes “h” and “n”. Radi-

cals are foundational building blocks that help make

Kanji searchable and provide semantic clues, while

graphemes contribute directly to a character’s meaning

or phonetic reading. This leads to a natural parent-

https://kanji.jitenon.jp/, last accessed: 2025-02-11

https://dictionary.cambridge.org/dictionary/english/,

last accessed: 2025-02-11

KanjiCompass: An Etymology-Driven Adaptive Kanji Learning Tool

473

child relationship between graphemes and the Kanji

they form, with pictorial Kanji and radicals as roots.

Despite Kanji’s compositional structure, many

learning materials like JiShop

break them into their

smallest components without considering etymologi-

cal signiﬁcance. This poses a challenge for research

and pedagogy: distinguishing between graphemes,

which contribute to a Kanji’s meaning or reading, and

components, which may not. For instance, the Kanji

際

(pronounced “sai”) contains the phonetic grapheme

祭

(also “sai”) and the radical

阝

, but can also be

broken into smaller components (

ノ

二

小

阡

癶

示

)

, which offer no phonetic and no mnemonic value.

Distinguishing between etymologically meaningful

graphemes and irrelevant components is therefore key

to optimizing Kanji learning.

Some Kanji were simpliﬁed in the 1946 reform,

resulting in “old” versions being replaced with stan-

dardized forms. For instance,

国

(“country”) was his-

torically written as

國

, yet both versions belong to the

same etymological category of metaphors:

囗

(“bor-

der”) enclosing

或

(“estate”) or

玉

((king’s)“jewel”).

This demonstrates that an etymological analysis of

modern Kanji remains useful for learners, even with-

out referring explicitly to older forms. The structural

changes often preserve underlying semantic and pho-

netic relationships, making etymology a valuable tool

for Kanji learning.

For effective learning, it is important to differen-

tiate the three stages of memory: encoding, storage,

and retrieval. Understanding these stages is crucial,

as they highlight different approaches to Kanji learn-

ing—from initial character recognition to long-term

recall. Learning methods for the encoding stage in-

clude visual associations, components analysis, and

mnemonic sentences (Rose, 2017). The retrieval stage

involves methods such as rote memorization or re-

peated writing of characters. These methods are in-

corporated in some learning materials. “Kanji Look

and Learn” (Banno et al., 2009), for example, focuses

solely on visual aids, while Heisig (1994) relies en-

tirely on mnemonic phrases. These materials often

present their own learning method as “the one and

only easy method” to learn all Kanji, yet they typically

overlook the etymological creation of characters.

This overview highlights the complexity of Kanji

and the limitations of traditional methods, emphasizing

the need for an etymology-driven approach that can

make learning both more efﬁcient and meaningful.

http://www.jishop.com/, last accessed: 2025-02-11

https://jisho.org/, last accessed: 2025-02-11

3 RELATED WORK

This chapter explores Kanji learning tools, beginning

with graph-based representations to illustrate charac-

ter relationships, followed by ﬂashcard systems for

memorization, and concluding with adaptive learning

technologies for personalized support. Technologies

for language learning encompass a wide range of ap-

proaches, but most focus on vocabulary or grammar

(Cook, 2016) and are not readily applicable to Kanji

learning. Kanji learning requires tools that address its

unique characteristics, such as etymological structure

and compositional relationships.

As described, Kanji are either composed of gra-

phemes or are standalone, indivisible characters. The

structure of Kanji follows a hierarchy, with standalone

Kanji forming the top level, followed by combinations

of standalone and/or composite Kanji. This hierar-

chy can be represented by a directed graph, allowing

learners to better understand the relationships between

Kanji (Komarek et al., 2015). In this context, the ques-

tion arises of how to build and visualize a Kanji net-

work to support learning. Existing research on graph

models of Chinese (H

anz

ı) and Japanese characters

offers initial approaches, as Kanji and H

anz

ı share

structural similarities.

Li and Zhou (2007) modeled a H

anz

ı network with

6,652 characters and 1,624 components as nodes, ana-

lyzing its density, path length, clustering, degree distri-

bution, and assortativity. They found that the network

exhibits complex small-world properties. However, a

visualization of the entire network was found too com-

plex to be legible. Jeronimus et al. (2017) examined

a network of 1,945 Kanji, showing that some Kanji

had very high degrees due to extensive breakdown into

components and not graphemes. They also visualized

only a partial graph.

Yan et al. (2013) developed a H

anz

ı network to

create a learning sequence, ﬁnding that a few charac-

ters construct many others, while most characters form

only a few. This network utilizes color to indicate com-

positional features and node size to suggest learning

order based on stroke count (cf. Figure 2). However,

while this approach considers the graphemes of the

characters, it does not account for how learners can

internalize these relationships, nor does it adapt to the

learner’s existing knowledge base. The compositional

relationships are presented statically, and further ex-

planation is necessary for learners to fully understand

the characters’ structure and meaning. Although Yan

et al. (2013) addresses some aspects of character com-

position, their network lacks the adaptability necessary

to support personalized learning pathways. Kovacs

https://thekanjimap.com/, last accessed: 2025-02-11

CSEDU 2025 - 17th International Conference on Computer Supported Education

474

Figure 2: Network graph of the H

anz

音

from the “Meaning-

fully learning Chinese characters” tool by Yan et al. (2013)

created the web application The Kanji Map, using a

force-directed layout to visualize connections for a

given Kanji. However, this visualization does not ac-

count for the etymological structure of Kanji, limiting

its effectiveness for learning.

In language learning, ﬂashcard programs have been

widely recognized as effective tools for memorization

(Hanson and Brown, 2020). A widely used ﬂashcard

program, continuously updated since 2008, is Anki.

Anki (

暗記

, “memorization”) is an open-source ﬂash-

card software originally developed by a Japanese lan-

guage learner for studying Japanese and English but

became a generic tool for learning with ﬂashcards.

Anki still remains popular in the Japanese learning

community due to its many add-ons, tutorials, and

active user base. Users can create custom multime-

dia ﬂashcards with a front and back, customizable via

HTML and CSS. Anki’s algorithm schedules review

sessions using spaced repetition, based on the Super-

Memo 2 (Wozniak, 1990) and the extended version

Free Spaced Repetition Scheduler algorithms, though

users can adjust intervals manually.

When a card tran-

sitions from learning to review status, it is assigned an

ease factor, which reﬂects the learning difﬁculty level.

This factor starts at 250 % and can drop to a minimum

of 130 %, depending on past responses and repetition

frequency. Zimmerman and McMeekin (2020) suggest

that educators should recognize learners’ preference

for simple ﬂashcard programs when learning Kanji

and, thus, avoid investing signiﬁcant effort in creating

learning games.

https://apps.ankiweb.net/, last accessed: 2025-02-11

https://faqs.ankiweb.net/what-spaced-repetition-

algorithm.html, last accessed: 2025-02-11

One approach to implementing personalized learn-

ing experiences is through Intelligent Tutoring Sys-

tems (ITS). These systems simulate the role of a hu-

man tutor by providing tailored feedback and guidance

and have been successfully applied in computer-based

language learning (e.,g. Slavuj et al., 2015; Heilman

and Eskenazi, 2006). ITS leverage a well-established

framework that includes a student model, a domain

model, and a pedagogical model (Alhabbash et al.,

2016). These frameworks can also be adapted to

broader adaptive systems, incorporating instructional

models and adaptive engines to enhance personaliza-

tion (Martin et al., 2020).

Building on the principles of ITS, adaptive learn-

ing systems have shown particular promise for het-

erogeneous learner groups (Taylor et al., 2021), such

as Kanji learners. These systems adjust parameters

based on individual needs, personalizing the learn-

ing process through algorithmic changes and fostering

self-regulated learning (SRL) without requiring man-

ual customization. Given the complexities of Kanji

networks, such personalization could signiﬁcantly en-

hance learner engagement and accommodate diverse

needs. In language learning, adaptive systems have

been studied less frequently than in STEM education.

Kaur et al. (2023) reviewed 1,342 articles, ﬁnding

that most focus on English grammar and vocabulary,

with adaptations targeting problem-solving assistance,

learning path recommendations, content adjustments,

and domain model modiﬁcations. Adaptations can also

be based on factors such as knowledge level, perfor-

mance, cognitive abilities, learning style, and behavior.

There are speciﬁc adaptive learning tools for Japanese,

such as an app developed by Ng et al. (2015), designed

for Chinese native speakers to learn Japanese vocabu-

lary and writing. The app offered tasks like selecting

correct pronunciations from multiple-choice options

or practicing writing and adjusted learning paths based

on user performance. While a small study reported

positive outcomes, the criteria for determining when

to adjust the learning path were not clearly deﬁned.

Another example is an adaptive email-based program

by Li et al. (2009), which customized Kanji quizzes

based on user interests, performance, and preferred

times. The system dynamically adjusted test difﬁculty

and sent quizzes at individual times to encourage regu-

lar learning. Although the adaptive group in a small

study performed better, limited user engagement re-

duced the system’s overall effectiveness. These studies

demonstrate both the potential and the challenges of

applying adaptive methods in Japanese learning.

In summary, while existing methods address some

needs of Kanji learners, there remains a need for a tool

that effectively bridges the gap between network visu-

KanjiCompass: An Etymology-Driven Adaptive Kanji Learning Tool

475

alization, adaptive learning, and practical memoriza-

tion techniques. Building on the insight of Zimmerman

and McMeekin (2020), it is more practical to integrate

such a well-established tool like Anki into a uniﬁed

approach, combining its strengths with adaptivity and

etymological insights to provide comprehensive sup-

port for Kanji learners.

4 AN ETYMOLOGY-DRIVEN

ADAPTIVE KANJI LEARNING

TOOL

This section presents the design decisions for a proto-

type that addresses the research gap identiﬁed in the

previous section. The goal is to develop a tool for

o Kanji—characters used in everyday Japanese—

that supports mnemonic creation through etymological

insights and facilitate self-regulated learning (cf. Zim-

merman, 2000). Hence, the tool should be tailored to

accommodate diverse learning strategies and adapt to

users’ varying knowledge levels. It should encourage

exploratory learning by allowing users to search for

Kanji, explore related characters, and save them for

future review. An important consideration is how to

effectively present the interconnectedness of Kanji, en-

suring that learners can easily navigate relationships

between characters. Learners should be able to track

their progress by viewing previously studied Kanji

and receive personalized learning recommendations

based on their known characters. Additionally, the

tool should offer simple example sentences and high-

lights Kanji that have not yet been mastered. It is not

the goal, however, to build yet another rote memoriza-

tion or ﬂashcard learning tool (cf. Zimmerman and

McMeekin, 2020) but to allow explorative Kanji learn-

ing and to provide an interface to the established and

widely used ﬂashcard tool Anki.

The prototype offers three main functions: Look-

ing up and displaying all Kanji and their relationships

in a visual format, getting detailed information rele-

vant for learning on selected Kanji, exporting these

information to the ﬂashcard software Anki, and im-

porting previously learned Kanji via Anki. In addition,

the application reacts adaptively to the user’s level of

knowledge by making learning path recommendations

and providing simple example sentences generated

with GPT-4 Turbo. The prototype was developed as a

platform-independent single-page web application and

uses the MEN stack (MongoDB, Express.js, Node.js).

It offers various ways to look up Kanji. Users can enter

a Kanji using a digital Japanese keyboard or draw a

Kanji using their mouse/ﬁnger. This function uses Kan-

jicanvas.js (Klein, 2021), which works even with im-

precise strokes and displays a selection of recognized

Kanji from which the desired one can be selected.

A central component and the entry point of the

tool is an interactive Kanji graph. This graph includes

the J

o Kanji, their graphemes, and visualizes the

construction of the Kanji as a network by using its

graphemes as nodes and its etymological composition

as edges. The graph is fully interactive, supporting

click, drag, and zoom functionalities. The Kanji net-

work can be viewed as a complete graph with a con-

centric layout, placing Kanji most often used (i. e., the

graphemes) in the center. This gives learners the op-

portunity to understand that a few graphemes build

most of the Kanji and are therefore useful to learn to

gain a better understanding of many Kanji. This graph

addresses the gap that existing tools often fail to vi-

sualize the etymological relationships and structural

composition of Kanji in an accessible and interactive

manner, leaving learners without a clear understand-

ing of how graphemes and other Kanji connect to the

Kanji they form. The Kanji network is based on a cus-

tom database developed speciﬁcally for the prototype,

compiled from openly available resources, including

(Tamaoka et al., 2017) and CHISE.org. CHISE.org pro-

vides a grapheme-based decomposition of Kanji, iden-

tifying meaningful structural units rather than merely

the smallest components. However, it serves primarily

as a character processing database and does not struc-

ture this information for learning purposes. (Tamaoka

et al., 2017), in contrast, offers an etymological cate-

gorization of Kanji but does not include any decom-

position into components. Since neither source alone

fully supports a pedagogically guided decomposition,

our approach reﬁnes and integrates these resources to

ensure that Kanji are broken down into meaningful

and instructive graphemic units, preserving relevance

for learners.

The overall design of the tool is illustrated in Fig-

ure 3. At the top center (1), users can search for Kanji

using the search bar. The central area features the

Kanji graph (2), which displays either the entire set of

Kanji or a ﬁltered subset along with their relationships.

As detailed in Section 4.1, the graph adaptively high-

lights nodes corresponding to Kanji already known

to the user. On the lower left (3), a legend provides

a clear overview of the various node and edge types

based on the etymological categories. The legend is

visible by default but can be toggled as needed. On

the right (4), an information box, referred to as the

“Kanji card”, is displayed whenever a Kanji is selected,

offering additional details about the character.

Upon searching for a speciﬁc Kanji (see Figure 3),

the tool presents a tailored, detailed hierarchical graph

CSEDU 2025 - 17th International Conference on Computer Supported Education

476

Figure 3: The prototype’s user interface after searching for the Kanji 字; details are in the text

of its parent and child characters. It shows the Kanji

together with its direct and optionally also its indirect

neighbors. The user can switch between these views

to either reduce the number of shown Kanji or get a

better overview using the toggle button (5). The lay-

out is based on the Klay Layered algorithm (Schulze

et al., 2014), which arranges the nodes according to

the hierarchy of the Kanji’s graphemes. The selected

node is highlighted (6). A special feature of the de-

tail graph are the compound nodes, which represent

the relationship between radicals and their variants

by grouping radical variants under a common parent

node. This structure was chosen to make it easier to

understand the composition of Kanji and their variants.

The graphs are created using the Cytoscape.js library

(Franz et al., 2015), with various optimizations applied

to improve performance and reduce complexity.

The Kanji card (see Figure 4) presents detailed

Kanji information in an interactive, ﬂashcard-like for-

mat. Additional details are revealed through hover

interactions. At the top left (1), the usage frequency

of the Kanji is displayed (based on Tamaoka et al.

(2017)). To make the distribution of the Kanji easier to

understand, it is divided into quintiles and visualized

using one to ﬁve stars. Other important information

on the card includes the etymology category of the

Kanji, which is represented by icons (2). Two mean-

ings (3) of the Kanji in German are generated using

GPT-4, queried with a compact prompt and stored in

the database. GPT-4 was chosen over standard dic-

tionary resources due to its ﬂexibility in generating

concise, user-friendly meanings by providing only the

most important meaning of a Kanji rather than listing

all possible interpretations. This approach ensures a

consistent format and seamless integration into the

application while minimizing the manual effort re-

quired for curating deﬁnitions. Below this, the Kanji

is displayed in large letters (4), followed by a tabular

overview of the Japanese readings (5). In the center

(6) of the card is the structure display, which shows

either the ancient writing of the Kanji for clearer pic-

torial representation or its composition of graphemes.

Further down, the card contains example words and

sentences (7), also generated by GPT-4. These exam-

ple sentences are generated to provide clear context for

how the Kanji is used in practice, utilizing grammati-

cally correct and vocabulary-appropriate sentences at

a simple language level. The example sentences were

generated, when the user clicked on the correspond-

ing button. At the bottom is general data (8) such as

stroke count and the level on the Japanese Language

Proﬁciency Test (JLPT).

While the Kanji card allows learners to delve

deeply into the details of individual characters, this

represents only one facet of the prototype’s capabili-

ties. The tool is also designed to encourage broader

exploratory learning by adapting dynamically to the

user’s knowledge level and offering personalized rec-

ommendations. These adaptive functionalities, along-

side the seamless integration with external tools like

Anki, will be elaborated upon in the following section.

4.1 Adaptivity

The prototype adapts learning content to each user’s

knowledge level, following the adaptive learning

framework by Martin et al. (2020). In the prototype,

the Domain Model is represented by the database con-

taining the J

o Kanji, their meanings, and etymolo-

gies. The Student Model contains the set of Kanji

that the learner “knows”, and the Instructional Model

consists of algorithms for learning recommendations.

A central component of the prototype is its inte-

gration with the ﬂashcard software Anki, enabling the

KanjiCompass: An Etymology-Driven Adaptive Kanji Learning Tool

477

Figure 4: The information card for the Kanji 字

prototype to rely on already existing data about the

user’s learning progress. By utilizing Anki’s Ease Fac-

tor, a measure indicating how well a Kanji has been

learned, the prototype avoids the need to collect data

from scratch, thereby mitigating a potential cold start

problem. This integration allows the prototype to dy-

namically adjust its learning recommendations and

visualizations based on the user’s current knowledge.

The adaptive nature of the prototype stems from

its ability to import knowledge levels of Kanji the

user has already learned, as well as to export Kanji

for further practice in Anki. This dual functionality

ensures a seamless connection between the learning

tool and the user’s established study habits, enhancing

both ﬂexibility and personalization. Details regarding

the technical implementation of this integration are

provided in the following.

To assess the user’s knowledge level, the Anki-

Connect

plugin is used, which provides a HTTP API

allowing external applications on the same computer

such as webbrowsers to exchange data with Anki. The

prototype makes use of this API to allow users to ex-

port Kanji they want to learn as ﬂashcards to Anki or

to import the state of knowledge of Kanji they have

already learned. However, integrating this functional-

https://foosoft.net/projects/anki-connect/, last accessed:

2025-02-11

ity into a website requires the domain to be manually

added to an allow list in the plugin settings to conﬁgure

cross-origin resource sharing (CORS) headers

When exporting Kanji to Anki, a pop-up dialog al-

lows the user to select which information (e. g., mean-

ing, etymological details) should appear on the Anki

ﬂashcard. The ﬂashcards are automatically generated

using a predeﬁned HTML and CSS layout. For im-

porting Kanji—used to update the Student Model—the

prototype displays a list of Anki decks (folders) from

which the user can select. Kanji from these decks,

along with their corresponding knowledge levels, are

then imported into the prototype.

The prototype visualizes the user’s knowledge in

the interactive graph by highlighting learned Kanji.

Additionally, a special view displays all learned Kanji

in a grid layout, sorted by Anki’s Ease Factor. Kanji

considered to be poorly learned are marked in red.

The prototype suggests which of these Kanji should be

reviewed and encourages users to create personalized

mnemonics and adjust the ﬂashcards in Anki.

4.1.1 Learning Recommendations

The prototype generates learning recommendations

based on the user’s existing Kanji knowledge, promot-

ing an exploratory approach to learning. Once a Kanji

is exported to Anki, the application can suggest an-

other Kanji that aligns with the user’s learning progress

and encourages further discovery. These recommen-

dations leverage the structural relationships between

Kanji, prioritizing content that is both relevant and

conducive to deepening the learner’s understanding

through exploration.

There are four main categories of learning recom-

mendations in descending importance:

Recommendations based on radicals: Radicals and

their common variations are prioritized, as they appear

in many Kanji. The prototype’s database includes 208

radicals and 36 variations, with the 25 most frequent

radicals appearing in about 61 % of all Kanji. Hence,

mastery of these radicals is considered a foundation

for understanding a majority of Kanji.

Phonological graphemes: Kanji containing phono-

logical graphemes that inﬂuence the pronunciation of

other Kanji are recommended if they are deemed rel-

evant for daily use. Relevance is determined using

factors identiﬁed by Toyoda et al. (2013), who assign

scores to Kanji based on these criteria. A grapheme

is considered useful by the prototype if its combined

score exceeds the average across all factors.

Kanji in example words presented on the Kanji

card: If a user selects a sample word containing an

unknown Kanji during export, that Kanji is recom-

CSEDU 2025 - 17th International Conference on Computer Supported Education

478

mended for further study, reinforcing learning through

contextual connections.

Phonological usage: If the exported Kanji is used

as a phonological grapheme in other Kanji, one of

these Kanji is recommended for further learning to

strengthen pronunciation connections. Learning an-

other Kanji with the same grapheme and pronunciation

can enhance recall.

Through these adaptive learning recommendations

and the visualization of knowledge, the prototype pro-

vides a personalized learning experience tailored to

the user’s individual progress.

The prototype addresses key use cases, such as re-

peatedly searching for Kanji, exploring the component-

based structure, contextualizing known Kanji with new

characters, and using the provided resources for both

educational and personal interest.

5 EVALUATION

The goal of the evaluation was to understand how

learners perceive and use the tool’s visualization and

recommendation features in a realistic setting, where

they use the tool to study Kanji, and to determine to

what extent these features motivate and engage them

in exploring and deeply learning Kanji.

To ensure that the tool incorporated both pedagog-

ical insights and features that support self-regulated

learning, feedback was sought from a Kanji learning

expert for non-native speakers before the evaluation

took place. During an hour-long session, the expert

reviewed the prototype and suggested improvements,

including stroke order animations and a handwritten-

style font. She praised the integration of GPT, the

graphical representation of Kanji, and the Anki con-

nection as valuable features that promote a compre-

hensive and ﬂexible learning experience. The expert

emphasized the importance of self-regulated learning,

particularly for advanced learners, highlighting that

the tool effectively encourages users to actively think

and make decisions, which she deemed essential for

meaningful learning.

The developed prototype was investigated using a

user study in May 2024 with 19 participants. The eval-

uation methodology involved a ﬁeld study with two

versions of the Kanji learning tool: one with adaptive

features (the full version as described in Section 4) and

one without. The non-adaptive version was visually

indistinguishable, but does not provide the buttons to

export and import Kanji from and to Anki.

To minimize confounding variables like prior Kanji

knowledge or learning strategies, a within-subject de-

sign was used: All participants tested both versions in

randomized order over the course of two weeks with

one week for each version. At the end of each week

the participants ﬁlled out a questionnaire. The usage

of the prototype and which Kanji to look up were not

pre-speciﬁed, but instead the participants were asked

to integrate the tool into their already existing learning

routine. Usability and user experience were evaluated

using the System Usability Scale (SUS; Brooke, 1996)

and Likert scales, which measured satisfaction with

features such as the graph, sentence examples, and

open-ended questions. Participants were also asked

to state their learning goal for the week and rate their

motivation to achieve it on a Likert scale. Participants’

responses were standardized to a scale from 0 to 4,

where 0 represent the most negative rating and 4 the

most positive. The activity of the participants in the

prototype was logged by registering their clicks.

All 19 participants completed both week’s surveys.

Thirteen had formal Japanese education, while six

were self-taught. Recruitment was through personal

connections and snowball sampling, with all partici-

pants being German-speaking young adults. The me-

dian Kanji knowledge was 350, with ﬁve knowing over

1,000, covering a diverse learner range.

Non-parametric statistical tests, particularly the

Wilcoxon-Mann-Whitney and Wilcoxon Signed-Rank

tests, were applied to compare groups and conditions,

and Spearman’s rank correlation (

) was used to exam-

ine relationships. Typically, these tests are stricter than

their parametric counterparts Kaur and Kumar (2015),

but do not require e. g. normal distributed data. The

level of signiﬁcance is

α = .05

˜x

denotes the median.

5.1 Results

This section presents the results of the study, beginning

with overall ﬁndings on usability and user experience,

followed by an analysis of participant groups based on

their use of the tool’s adaptive features.

The participants used the prototype for a median of

57 minutes and rated it favorably, with a usability score

(SUS) of 72.1, indicating a good user experience (Ban-

gor et al., 2008). Ten participants rated the usability

as good, and four as excellent. SUS scores improved

over time, rising from 70.1 in the ﬁrst week to 74.1 in

the second week. Thirteen participants indicated they

would use the tool regularly (

˜x = 3.0

), ﬁnding it easy

to use and its functions well integrated ( ˜x = 3.0).

Participants rated the presentation, loading time,

and interactivity of the overall graph positively (

˜x =

3.0

). However, the majority found the graph over-

whelming, especially when they did not have a speciﬁc

learning objective (

˜x = 1.0

). In contrast, the detailed

graph received signiﬁcantly better ratings for presenta-

KanjiCompass: An Etymology-Driven Adaptive Kanji Learning Tool

479

tion, loading time, and interactivity (all

˜x ≥ 3

), and was

perceived as less overwhelming (

p < .001

;

r = .05

)

(see Figure 6). It was also seen as more motivating and

supportive for learning, particularly for participants

with clear learning goals and extended usage. The de-

tailed graph was praised for helping users understand

Kanji connections (

˜x = 4.0

) and the linkage of radicals

with their variants (˜x = 3.5).

Seven participants used the Anki export func-

tion. They rated it as very helpful (

˜x = 4.0

) and fast

(

˜x = 4.0

). The participants appreciated the ability to

create cards with customizable information (

˜x = 4.0

Two participants were motivated by the tool to install

Anki for the ﬁrst time and began using it regularly.

The other participants did not use the feature, citing

existing schemes for creating Anki cards (33.3 %) or a

lack of active engagement with Anki (33.3 %). Partici-

pants who used the export function received an aver-

age of 35.2 learning recommendations, of which they

followed 38.9 %. Moreover, participants rated these

recommendations as both highly helpful (

˜x = 4.0

) and

motivating (

˜x = 4.0

). They perceived the recommen-

dations as non-disruptive (

˜x = 3.5

) and easily com-

prehensible (

˜x = 3.5

). The participants rated the im-

port function as motivating (

˜x = 4.0

), as they can see

their progress, and helpful for understanding the re-

lationships between known and “new” Kanji. The

highlighting of poorly learned Kanji was perceived

as motivating (

˜x = 3.0

); however, only two partici-

pants actively clicked on and revised poorly learned

Kanji, as reﬂected by a low median agreement score

˜x = 2.0

for the statement: “I followed the request to

revise poorly learned Kanji.”

In the following, an in-depth analysis is conducted

and the participants are grouped based on their actual

use of the tool’s adaptive features: nine participants

used the adaptive functions, while ten did not (despite

being able to do so). These participants primarily used

the tool to look up Kanji as they were free to decide

how to integrate the prototype into their learning pro-

cess. For instance, 12 participants did not export Kanji

to Anki or 15 participants did not explore example

more than two sentences. The dimensions analyzed

included prior Kanji knowledge, learning goal motiva-

tion (whether a learning goal was set and how high the

motivation was to achieve it; rated 0–10), usage time,

SUS scores, and adaptivity usage.

The analysis revealed a strong positive correla-

tion between learning goal motivation and both the

time spent using the tool (

ρ = .66

p = .002

) and SUS

scores (

ρ = .59

p = .008

), as well as the adaptivity

usage (

ρ = .59

p = .007

) (see Figure 5). Additionally,

SUS scores were moderately correlated with usage

time (

ρ = .49

p = .034

) and adaptivity usage (

ρ = .44

p = .057

). Notably, participants who set a learning

goal rated the tool signiﬁcantly higher (SUS = 76.9)

than those without a goal (SUS = 62.5;

p = .025

They also rated the detailed graph lower across several

factors, as shown in Figure 6.

Moti

vation-Time

Moti

vation-SUS

Moti

vation-Adaptivity

SUS-T

ime

SUS-Adapti

vity

0.5

0.66

0.59 0.59

0.49

0.44

Correlation (ρ)

Figure 5: Correlation coefﬁcients (

) between key dimen-

sions: learning goal motivation (Motivation), system usabil-

ity scores (SUS), usage time (Time), and adaptivity usage

(Adaptivity).

Presen-

tation

Relation-

ships

Detail

Level

Over-

whelm-

ing

3.5

4 4 4

3 3 3

Median

Figure 6: Median ratings for the detailed graph by over-

all responses (Gray), with learning goal motivation (Cyan),

and without learning goal motivation (Teal). Categories:

Presentation reﬂects overall satisfaction with the graph’s pre-

sentation; Relationships assesses whether the graph helped

identify connections between Kanji; Detail Level refers to

the appropriateness of the graph’s level of detail; Overhelm-

ing measures whether the graph was overwhelming.

Finally, participants suggested improvements in

open-ended comments such as integrating stroke order

animations, using a font that looks like handwriting, or

providing an introductory video. Many appreciated the

tool for combining multiple resources. One participant

likened the application to Google Maps, describing

how it allowed her to recognize Kanji connections and

learn radicals from a “bird’s-eye view”.

CSEDU 2025 - 17th International Conference on Computer Supported Education

480

The study highlights the critical role of learning goal

motivation in driving engagement with the prototype.

Participants with a clear learning goal spent more time

exploring the features and using more adaptive func-

tions. This suggests that internal motivation, when

coupled with clear objectives, can amplify the effec-

tiveness of learning tools. Conversely, participants

without explicit goals demonstrated less commitment,

indicating the need for external incentives to sustain

their engagement. The goal of the study was to have

an authentic usage by the learners, hence, no clear

objectives for tool usage were provided. Having such

objectives may have increased tool usage—still, the

study provided interesting insights into the design of

such a learning tool. To support learners, the prototype

may be extended to provide help setting learning objec-

tives (cf. foresight phase in SRL; Zimmerman, 2000)

or to include gamiﬁcation elements such as progress

badges, streak rewards, or social competition, which

could provide motivate users to engage more actively.

While the SUS score was generally favorable, cer-

tain usability challenges emerged. The overall positive

rating reﬂects the intuitive interface and ease of use,

particularly for exploration and Anki integration. How-

ever, the adaptive features—designed to be a central

aspect of the tool—received less emphasis. This sug-

gests that, while users valued adaptability, the core

usability features were more immediately impactful.

Future research should introduce the adaptive features

through onboarding tutorials or speciﬁc prompts.

The overall Kanji network graph received mixed

feedback. While the majority of users appreciated

its interactivity, the sheer volume of information pre-

sented was overwhelming for some. Despite this, the

inclusion of the overall graph serves an important pur-

pose: it provides learners with a macro-level perspec-

tive of Kanji interconnections, enabling them to see

how a small subset of foundational Kanji forms the

basis for many others. This “big picture” view, though

initially daunting, can inspire curiosity and highlight

the systematic nature of Kanji learning. Hence, it was

initially included. Future reﬁnements could include a

personalized knowledge graph to reduce overwhelm

and individualize the user experience. This person-

alized graph could display only the Kanji known to

the user together with their parents and children. In

contrast, the detailed graph was rated signiﬁcantly

higher for its clarity and its more user-friendly, less

overwhelming design. The smaller number of nodes

and the use of highlighting for learned Kanji created

a more personalized and engaging experience. Partic-

ipants valued the motivational aspect of seeing their

progress through color-coded nodes and exploring ety-

mological relationships in smaller, manageable graphs.

The Anki export function was appreciated by those

who used it, as it simpliﬁed the process of creating

custom ﬂashcards. However, the relatively low adop-

tion rate suggests that the setup effort, particularly

in conﬁguring the CORS headers in the Anki HTTP

server, posed a barrier for some participants, despite

the availability of a tutorial. Simplifying the connec-

tion process could address these concerns. Requests

for improvement concerned technical aspects such as

destination selection and batch export. Although learn-

ing recommendations were not frequently used, their

quality was rated positively. While only 38.9 % of the

recommendations were followed, which may seem low

at ﬁrst glance, it is worth noting that learners typically

do not implement every suggestion in tools designed

for SRL, especially when presented with a high vol-

ume of recommendations. This implementation rate

demonstrates a meaningful level of engagement with

the tool and reﬂects the perceived relevance and feasi-

bility of the recommendations.

The import function was highly appreciated, as the

users found the display of previously learned Kanji

helpful. This feature helped learners recognize their

progress and adjust their priorities — both are impor-

tant aspects in SRL (cf. Zimmerman, 2000). However,

some participants suggested extending the function-

ality to provide methodical learning suggestions for

less experienced users. The knowledge level check

motivated some participants to adjust their learning

priorities, but was not consistently used. Adding

beginner-friendly learning strategies could support

users in building effective study habits.

7 THREATS TO VALIDITY

While the study provides valuable insights into the

usability and impact of the Kanji learning tool, several

threats to validity need to be acknowledged.

The sample consisted of only 19 participants. Al-

though efforts were made to include individuals with

varying levels of prior Kanji knowledge, all partic-

ipants shared a similar cultural and linguistic back-

ground as German-speaking learners of Japanese. This

homogeneity could inﬂuence the results, as the tool

may perform differently with users from other linguis-

tic or cultural contexts. Future studies should aim to

include a larger and more diverse sample to increase

the robustness and generalizability of the ﬁndings.

Participants tested the tool in their own homes

rather than in a controlled environment, allowing for

authentic insights into its usability in a real-world con-

6 DISCUSSION

KanjiCompass: An Etymology-Driven Adaptive Kanji Learning Tool

481

text. However, participants were free to use the tool

as they wished, without necessarily integrating it into

their study routines as suggested. This lack of over-

sight may have introduced inconsistencies that could

affect the validity of the ﬁndings.

To minimize bias, the order in which participants

used the adaptive and non-adaptive versions of the tool

was randomized. However, it is still possible that the

novelty or frustration experienced in the ﬁrst week

could have subtly inﬂuenced user behavior, percep-

tions, or exploration in the second week.

Although the study intentionally included partic-

ipants with varying levels of Kanji knowledge, their

existing familiarity with Kanji and learning tools likely

inﬂuenced their interaction with the prototype. For in-

stance, more experienced learners may have found

certain features redundant, while beginners may have

struggled to fully utilize advanced functionalities. For

example, a button was labeled with

例

(“example”) as

done in many books to indicate that clicking it would

provide an example. However, this label was not easily

understood by all beginners. These differences could

have affected their usability ratings and engagement

levels, highlighting the need for segmentation of user

feedback based on experience levels in future studies.

The study was conducted over a relatively short pe-

riod of two weeks. While this time frame allowed for

initial insights into user engagement and satisfaction,

it was not sufﬁcient to evaluate the tool’s impact on

sustained learning outcomes. A longer evaluation pe-

riod would be necessary to get a more comprehensive

understanding of the tool’s effectiveness and usability.

CONCLUSIONS AND OUTLOOK

In this paper, an interactive learning tool for Kanji

was proposed, combining graph-based visualizations

and adaptive features to support self-regulated and

exploratory learning. The tool enabled learners to

explore relationships between Kanji, track progress,

and connect new concepts to prior knowledge, while

adaptive recommendations aligned with their knowl-

edge levels. The ﬁndings underscore the importance of

balancing comprehensive features with a user-centric

design. Features like the overall graph, while ambi-

tious in scope, need to be complemented by mech-

anisms that simplify and personalize the user expe-

rience. Adaptive visualizations, streamlined integra-

tions, and guided tutorials are critical for accommodat-

ing diverse user needs and knowledge levels.

In terms of adaptivity, the tool highlighted the value

of context-sensitive recommendations of Kanji aligned

with learners’ knowledge levels, while preserving the

autonomy to engage with or bypass these suggestions.

Future developments could beneﬁt from implement-

ing adaptive features in a subtle, non-intrusive manner,

catering to learners who prefer ﬂexibility and mini-

mizing disruptions to established learning strategies.

The visualization approach and adaptive principles

are transferable to other domains, such as exploring

German word formations with preﬁxes and sufﬁxes,

visualizing chemical compounds and their functional

groups, or illustrating relationships in mathematical

formulas and transformations.

The prototype also supports encoding and retrieval;

this includes integration with an established tool fo-

cused on retrieval, demonstrating that combining spe-

cialized tools can enhance the overall learning pro-

cess. Future tools for supporting memorization may

beneﬁt from similar partnerships rather than attempt-

ing to create all-encompassing systems. Open Sci-

ence plays a crucial role in enabling such applica-

tions, preventing redundancy and fostering collab-

oration by allowing researchers to build on exist-

ing frameworks rather than reinventing tools and

methodologies. The project website of KanjiCompass

is https://www.tel.iﬁ.lmu.de/software/kanjicompass/,

where also the source code is available as open source.

Finally, motivation played a key role in tool usage.

Learners with clear goals found the tool beneﬁcial

for deepening their knowledge, while those lacking

intrinsic motivation were less likely to engage fully.

Future studies could explore how extrinsic motivation

techniques—such as gamiﬁcation or structured inte-

gration into formal learning settings—might increase

engagement, especially among learners without pre-

deﬁned goals. For these learners, supporting SRL,

particularly in the goal-setting and planning phases, is

essential. Future developments should focus on further

promoting learner motivation and strengthening SRL.

Finally, the study highlighted the potential of large

language models (LLMs) in generating example sen-

tences and translations. Future research should inves-

tigate their capabilities and quality further, exploring

how they can enhance learning tools across domains.

REFERENCES

Alhabbash, M. I., Mahdi, A. O., and Naser, S. S. A. (2016).

An intelligent tutoring system for teaching grammar

english tenses. European Academic Research, 4(9):1–

15.

Bangor, A., Kortum, P. T., and Miller, J. T. (2008). An

empirical evaluation of the system usability scale. Int.

J. Hum.-Comput. Interact., 24(6):574–594.

Banno, E., Ikeda, Y., Shinagawa, C., Tajima, K., and

Takashiki, K. (2009). Kanji Look and Learn: 512 Kanji

CSEDU 2025 - 17th International Conference on Computer Supported Education

482

with Illustrations and Mnemonic Hints -

イメージで

覚える［げんき］な漢字

512. The Japan Times, 5th

edition.

Brooke, J. (1996). Sus - a quick and dirty usability scale. In

Thomas, P. W. J. B., Weerdmeester, B., and McClel-

land, I. L., editors, Usability Evaluation In Industry,

pages 189–194. Taylor & Francis Ltd.

Cook, V. (2016). Second Language Learning and Language

Teaching: Fifth Edition. Routledge, 5 edition.

Crowley, D. P. (1968). The occurrence and function of

chinese characters in modern japanese orthography.

The Journal-Newsletter of the Association of Teachers

of Japanese, 5(3):1–9.

Franz, M., Lopes, C. T., Huck, G., Dong, Y., Sumer, O.,

and Bader, G. D. (2015). Cytoscape.js: a graph theory

library for visualisation and analysis. Bioinformatics,

32(2):309–311.

Hanson, A. E. S. and Brown, C. M. (2020). Enhancing l2

learning through a mobile assisted spaced-repetition

tool: an effective but bitter pill? Computer Assisted

Language Learning, 33(1-2):133–155.

Heilman, M. and Eskenazi, M. (2006). Language learn-

ing: Challenges for intelligent tutoring systems. In

Proc. Workshop of Intelligent Tutoring Systems for Ill-

Deﬁned Tutoring Systems., pages 20–28.

Heisig, J. W. (1994). Remembering the Kanji I: A complete

course on how not to forget the Meaning and Writing

of Japanese characters. Japan Publications Trading

Co., Ltd., 3rd edition. 12th printing.

Jeronimus, M., Westerveld, S., van Leeuwen, C., Bhulai, S.,

and van den Berg, D. (2017). Japanese kanji characters

are small-world connected through shared components.

In Bhulai, S. and Kardaras, D., editors, DATA ANA-

LYTICS 2017 : The Sixth International Conference on

Data Analytics, pages 53–58. IARIA.

Kaur, A. and Kumar, R. (2015). Comparative analysis of

parametric and non-parametric tests. Journal of com-

puter and mathematical sciences, 6(6):336–342.

Kaur, P., Kumar, H., and Kaushal, S. (2023). Technology-

assisted language learning adaptive systems: A com-

prehensive review. International Journal of Cognitive

Computing in Engineering, 4:301–313.

Klein, D. (2021). The stroke correspondence problem, revis-

ited. arXiv preprint arXiv:1909.11995v3.

Komarek, A., Pavlik, J., and Sobeslav, V. (2015). Network

visualization survey. In N

nez, M., Nguyen, N. T.,

Camacho, D., and Trawi

nski, B., editors, Computa-

tional Collective Intelligence, pages 275–284, Cham.

Springer International Publishing.

Li, J. and Zhou, J. (2007). Chinese character structure analy-

sis based on complex networks. Physica A: Statistical

Mechanics and its Applications, 380:629–638.

Li, M., Ogata, H., Hashimoto, S., and Yano, Y. (2009).

Adaptive kanji learning using mobile-based email. In

Proceedings of the 17th International Conference on

Computers in Education, ICCE 2009, pages 520–526.

Martin, F., Chen, Y., Moore, R. L., and Westine, C. D. (2020).

Systematic review of adaptive learning research de-

signs, context, strategies, and technologies from 2009

to 2018. Education Tech Research Dev, 68:1903–1929.

Mori, Y. (2014). Review of recent research on kanji process-

ing, learning, and instruction. Japanese Language and

Literature, 48(2):403–430.

Ng, S. C., Lui, A. K., and Wong, Y. K. (2015). An adap-

tive mobile learning application for beginners to learn

fundamental japanese language. In Li, K. C., Wong,

T.-L., Cheung, S. K. S., Lam, J., and Ng, K. K., editors,

Technology in Education. Transforming Educational

Practices with Technology, pages 20–32, Berlin, Hei-

delberg. Springer Berlin Heidelberg.

Rose, H. (2017). The Japanese Writing System: Challenges,

Strategies and Self-regulation for Learning Kanji, vol-

ume 116. Multilingual Matters.

Schulze, C. D., Sp

onemann, M., and von Hanxleden, R.

(2014). Drawing layered graphs with port constraints.

Journal of Visual Languages & Computing, 25(2):89–

106.

Slavuj, V., Kova

c, B., and Jugo, I. (2015). Intelligent tu-

toring systems for language learning. In 2015 38th

International Convention on Information and Commu-

nication Technology, Electronics and Microelectronics

(MIPRO), pages 814–819.

Tamaoka, K., Makioka, S., Sanders, S., and Verdonschot,

R. G. (2017). www.kanjidatabase.com: a new interac-

tive online database for psychological and linguistic

research on japanese kanji and their compound words.

Psychological research, 81:696–708.

Taylor, D. L., Yeung, M., and Bashet, A. (2021). Personal-

ized and adaptive learning. Innovative learning envi-

ronments in STEM higher education: Opportunities,

Challenges, and Looking Forward, pages 17–34.

The Japan Foundation (2023). Survey report on japanese-

language education abroad 2021. Technical report, The

Japan Foundation.

Toyoda, E., Firdaus, A. M., and Kano, C. (2013). Identifying

useful phonetic components of kanji for learners of

japanese. Japanese Language and Literature, pages

235–272.

Wozniak, P. A. (1990). Optimization of learning. Master’s

thesis, Poznan University of Technology, Poland.

Yan, X., Fan, Y., Di, Z., Havlin, S., and Wu, J. (2013).

Efﬁcient learning strategy of chinese characters based

on network approach. PloS one, 8(8):e69745.

Zimmerman, B. (2000). Attaining self-regulation: A so-

cial cognitive perspective. Self-regulation: Theory,

research, and applications/Academic.

Zimmerman, E. and McMeekin, A. (2020). A review of

japanese call for kanji, vocabulary, and reading: Find-

ings, best practices and future directions. JNCOLCTL,

29:34–68.

KanjiCompass: An Etymology-Driven Adaptive Kanji Learning Tool

483