Design an Intelligence System for Early Identification on

Developmental Dyslexia of Chinese Language

Man-Ching Yuen

, Ka-Fai Ng

, Ka-Ming Lau

, Chun-Wing Lam

and Ka-Yin Ng

iFREE GROUP Innovation and Research Centre, Department of Applied Data Science,

Hong Kong Shue Yan University, China

Department of Computer Science and Engineering, The Hong Kong University of Science and Technology, China

Keywords: Dyslexia, Traditional Chinese Character, App, Early Identification, Cloud.

Abstract: People with dyslexia have difficulties in fluently reading and writing characters which highly affect their

learning progress. It is very important to identify dyslexic students that need intervention and extra support

during their childhood. However, the waiting time for dyslexia assessment services is often long. To address

the above problem, we propose a cloud-based early identification system of dyslexia. We design and develop

a mobile app with AWS cloud platform as server. We have identified 27 representative traditional Chinese

characters for handwriting data collection. After the first round of the data collection, 66 children aged 5-7

were recruited in Hong Kong. We carry out K-means clustering algorithm to investigate the characteristics of

data points on a feature map for each character. We find out that some Chinese words contain more

distinguishable characteristics for identifying children with dyslexia. Since children aged 5-7 are still learning

how to write traditional Chinese characters properly, children with no risk of dyslexia still have certain

possibilities of writing characters with characteristics of handwriting by children with dyslexia. It increases

the difficulties of the early identification on developmental dyslexia of Chinese language. Finally, we present

our findings and future work.

1 INTRODUCTION

Dyslexia is a kind of life-span learning disability

disorder (Wang, J. and Perez, L., 2017). People with

dyslexia have difficulties in fluently reading and

writing characters which highly affect their learning

progress. Tsui et al. showed that Hong Kong primary

school students, aged 6-12, with dyslexia issue write

significantly slower and inaccurate (Tsui, C.M., Li-

Tsang, W.P.C. and Lung, P.Y., 2012). However,

Dyslexia is not an uncommon symptom. Based on an

earlier study (Chan, D.W., Ho, C.S.H., Tsang, S.M.,

Lee, S.H. and Chung, K.K., 2007), the prevalence rate

of dyslexia in Hong Kong was 9.7% (6.2% mild

severity, 2.2% moderate and 1.3% severe). In 2013,

Sprenger-Charolles et al. suggested that around 17%

of the world’s population experience dyslexia

(Sprenger-Charolles, L., Colé, P. and Serniclaes, W.,

2013).

It is very important to identify dyslexic students

that need intervention and extra support during their

https://orcid.org/0000-0003-2551-7746

childhood. However, the waiting time for dyslexia

assessment services is often long. Besides, traditional

psychological diagnosis consumes much time and

resources. For example, the assessment provided by

the British Dyslexia Association takes up to 3 hours

(Asvestopoulou, T., Manousaki, V., Psistakis, A.,

Smyrnakis, I., Andreadakis, V., Aslanides, I.M. and

Papadopouli, M., 2019). It is necessary to have an

early identification system of dyslexia for teachers

and parents, so that intervention can be provided as

soon as possible.

As machine learning has emerged into daily life,

researchers have applied different machine learning

models to analyse any specific patterns from

behaviors of dyslexic children especially for their

handwriting images. Various approaches have been

conducted to detect dyslexia using machine learning,

including eyeball movement tracking (Biswas, A. and

Islam, M.S., 2021), handwriting motion and pressure

(Isa, I.S., Rahimi, W.N.S., Ramlan, S.A. and

Sulaiman, S.N., 2019; Košak-Babuder, M., Kormos,

Yuen, M., Ng, K., Lau, K., Lam, C. and Ng, K.

Design an Intelligence System for Early Identiﬁcation on Developmental Dyslexia of Chinese Language.

DOI: 10.5220/0011281500003286

In Proceedings of the 19th International Conference on Wireless Networks and Mobile Systems (WINSYS 2022), pages 46-52

ISBN: 978-989-758-592-0; ISSN: 2184-948X

J., Ratajczak, M. and Pižorn, K., 2019), neuroimages

for biomarkers (Lam, S.S., Au, R.K., Leung, H.W.

and Li-Tsang, C.W., 2011), and brain images

(Usman, O.L. and Muniyandi, R.C., 2020). Although

the biomarkers are trackable, the assessment

equipment is expensive and the assessment process is

hard to scale up. These solutions are not suitable for

common usage at home or school.

There were studies conducted on convolution

analysis on handwritings of English, Spanish (Drotár,

P. and Dobeš, M., 2020) and Indian students

(Mahone, E.M. and Schneider, H.E., 2012). However,

there are only a few research works related to Chinese

characters, especially traditional Chinese characters

that are more difficult to analyse compared with

simplified Chinese characters. Tseng showed that

traditional Chinese characters contain sharp turns and

frequent pen lifts in which the symptoms should be

more critical (Tseng, M. H., 1998).

To address the above problems, we design and

develop a cloud-based system for early identification

of dyslexia, where machine learning methodology is

adopted to identify dyslexia involving traditional

Chinese characters. The main contributions are:

 To design and develop a mobile app with AWS

cloud platform as server. Our framework can

support real-time performance evaluation on

children handwriting wherever the number of

concurrent testers increases

 To identify 27 representative traditional

Chinese characters which are commonly taught

in kindergartens or training centers in Hong

Kong for data analysis experiments

 To collect the handwritings of 27 traditional

Chinese characters from 66 children, where 25

have dyslexia and 41 do not have dyslexia

 To carry out some preliminary experiments to

investigate the characteristics of the

handwritten character images

The organization of this paper is as follows.

Section 2 presents the related work. Section 3

describes our proposed dyslexia identification

system. Section 4 shows the preliminary experimental

result analysis. Section 5 draws out the conclusion

and the future work.

2 RELATED WORKS

Various approaches have been conducted to detect

dyslexia using machine learning. Thomais et al used

an eye-ball tracker to analyze the eyeball movement

(Biswas, A. and Islam, M.S., 2021) with the highest

accuracy of 89.39%. Several groups learnt features

from handwriting motion and pressure (Isa, I.S.,

Rahimi, W.N.S., Ramlan, S.A. and Sulaiman, S.N.,

2019; Košak-Babuder, M., Kormos, J., Ratajczak, M.

and Pižorn, K., 2019) showed promising results.

Some others applied CNN on neuroimages for

biomarkers and achieved accuracies of 73.2% (Lam,

S.S., Au, R.K., Leung, H.W. and Li-Tsang, C.W.,

2011). Another similar research analyzed brain

images while students were reading and resulted in an

accuracy of 72.73% (Usman, O.L. and Muniyandi,

R.C., 2020). A research in Malaysia tried to increase

the performance by applying the result OCR with a

73.77% accuracy. However, these solutions are not

suitable for common usage at home or school.

The above methods showed a very promising

result. However, these methods take too much time

and resources to sample a candidate. To solve this,

researchers studied detecting dyslexia via

handwriting. Xing et.al. showed it is reliable to

distinguish writers with handwritings using

convolutional neural networks (CNN), and the

proposed work, DeepWriter, achieved 99.01% on 301

writers and 97.03% on 657 writers (Xing, L. and

Qiao, Y., 2016). Later, several researchers studied

whether writers have dyslexia with a similar

approach. Spoon et.al. gathered students' English and

Spanish exercise books and applied CNN with Keras

to detect dyslexia. They achieved an accuracy of

77.6% with 1200 samples from K-6 students (Spoon,

K., Crandall, D. and Siek, K., 2019). Yogarajah et.al.

conducted a similar research for Hindi characters

achieving 86.14% (Yogarajah, P. and Bhushan, B.,

2020). Optical character recognition (OCR) with

Artificial neural network (ANN) focused on

analysing 8 characters and achieved a test accuracy of

57.5% (Wei, P., Li, H. and Hu, P., 2019).

3 DYSLEXIA IDENTIFICATION

SYSTEM

3.1 Overview of System Design

In this paper, we present an early identification

system for detecting dyslexia with traditional Chinese

characters, which is a cloud-based AI system. For

children doing the assessment test, parents have to

print out the worksheets and ask their children to

write the traditional Chinese characters on the

worksheets. After completion, parents have to take

pictures of the worksheets, crop the images, and

upload the cropped images to the cloud system for

Design an Intelligence System for Early Identiﬁcation on Developmental Dyslexia of Chinese Language

data analysis. Figure 1 shows the system architecture

of the early identification system of dyslexia. We

develop our machine learning by using AWS

SageMaker for data analysis.

Figure 1: System architecture of the early identification

system of dyslexia.

3.2 Mobile App Design

Figure 2 shows the web app interface in the web

browser. Since the user interface is in traditional

Chinese, we add some description in English. The

functions are providing user guide, downloading

worksheet, uploading worksheet, setting and

previewing previous testing results. Since the

duration of children handwriting is also a

performance indicator of dyslexia, the mobile app

allows parents to collect the time spent on

handwriting of their children by either using a system

timer or inputting the value manually (as shown in

Figure 3 and Figure 4). Before parents take pictures

and upload each of the 3 worksheets, parents have to

scan the QR code in the left corner of the worksheet

to make sure they are uploading the correct

worksheets (as shown in Figure 5). When the image

taken is shown on the mobile app, parents can crop

Figure 2: The web app interface in the web browser.

the image by adjusting the little yellow boxes at each

corner (as shown in Figure 6). The analysis carried

out by the cloud AI system indicates the risk of having

dyslexia for testers (as shown in Figure 7).

Figure 3: The mobile app can collect the time spent on

handwriting by using a system timer.

Figure 4: The mobile app can collect the time spent on

handwriting by inputting the value manually.

WINSYS 2022 - 19th International Conference on Wireless Networks and Mobile Systems

Figure 5: Scan the QR code in the left corner of the

worksheet to make sure parents are uploading the correct

worksheets.

Figure 6: Parents can crop the image by adjusting the little

yellow boxes at each corner.

Figure 7: The analysis result indicates the risk of having

dyslexia for the tester.

Figure 8: Project promotion website.

4 EXPERIMENTS

4.1 Participants and Tasks

To collect handwriting for model training, we plan to

invite about 30 children with dyslexia and 200

children without dyslexia from kindergartens.

Handwriting data is planned to be collected

periodically every 6-9 months. After the first round of

the data collection, a total of 66 children (aged 5-7)

were recruited from several kindergartens or training

centers of Yan Chai Hospital Social Services

Department (YCH) in Hong Kong. The children do

not have physical or mental disabilities which might

affect the handwriting performance and lead to bias

in the data collected for model training. Informed

consent of parents was obtained for all children. In

order to raise the awareness of dyslexia among

teachers and parents, we develop a website for project

promotion as shown in Figure 8.

We use the following 4 steps to carry out pre-

screening on the participating children in order to

distinguish whether they are dyslexia or not.

Design an Intelligence System for Early Identiﬁcation on Developmental Dyslexia of Chinese Language

1. Teachers will screen all participating children’s

performance on writing and reading traditional

Chinese words during class. Teachers will screen

out the students who are not suitable to

participate in the experiment.

2. Occupational Therapists (OTs) will carry out

Visual Perceptual (VP) assessment (only for

suspected cases of dyslexia).

3. Educational Psychologists (EPs) / Clinical

Psychologists (CPs) will carry out formal

assessments of dyslexia (only for suspected cases

of dyslexia in order to determine the level of

dyslexia).

4. The children are categorized into 3 groups:

(1) H – High risk of having dyslexia;

(2) L – Low risk of having dyslexia;

(3) N – No risk of having dyslexia / Normal.

Table 1 summarizes the age and gender for

children in the 3 groups, which are normal, low and

high risk of having dyslexia respectively.

Table 1: Age and gender of participating children.

High Low No

Number of

children

19 6 41

Mean age

(months)

67.2 69.0 69.9

Gender

(girls vs.

boys)

4 vs. 15 1 vs. 5 8 vs. 32

Figure 9 shows the 27 representative traditional

Chinese characters in the handwriting collection. The

27 characters are commonly taught in kindergartens

or training centers in Hong Kong. The words cover

six basic structures of Chinese characters (i.e.,

independent, left-right, above-bottom, above-middle-

bottom, left-middle-right and inside-outside) and

most basic stroke units to ensure that the selected

Chinese characters are sufficiently representative of

the characteristics of traditional Chinese. The idea on

how to select the words for data collection is also

inspired by Wu et al. presented in 2019 (Wu, Z., Lin,

T., and Li, M., 2019). The display sequences of the

words across 3 worksheets were randomized

regardless of the level of writing difficulties and the

structure of words.

We design a special set of worksheets for easier

and better quality sampling. A set of 3-page

worksheets contain the 27 representative traditional

Chinese characters as shown in Figure 10. On the

worksheet, each character is featured with a printed

character alongside and great margin. The printed

characters provide direct reference to the students to

maintain quality of the written text, especially for

normal students. Besides, the margin is designed to

capture all strokes and details in case students write

out of the box. The dimension of the box is 2.5cm x

2.5cm which is the same as the children learning

handwriting at the childhood at age 5-7 in Hong

Kong.

Figure 9: The 27 traditional Chinese characters used in the

experiment.

Figure 10: A set of 3-page worksheets contain the 27

representative traditional Chinese characters.

4.2 Observation from Screening

Based on our observation, for each traditional

Chinese character, some handwriting characters have

characteristics of handwriting by children with

dyslexia but these characters are written by children

without dyslexia. Moreover, children of high risk of

dyslexia have higher possibility of writing characters

with characteristics of handwriting by children with

dyslexia; while children of no risk of dyslexia still

have certain possibilities. The main reason behind

this is children aged 5-7 are still learning how to write

traditional Chinese characters properly.

For another finding, some handwriting characters

do not have characteristics of handwriting by children

with dyslexia but these characters are written by

children with dyslexia. It is because writing abilities

in children with dyslexia can be improved by training.

It demonstrates the difficulties and importance of this

project.

WINSYS 2022 - 19th International Conference on Wireless Networks and Mobile Systems

4.3 Preliminary Experimental Results

We carry out K-means clustering algorithm for each

traditional Chinese character by setting k as 3, plot all

data points on a 2D feature map. We summarize our

observations from the feature maps as follows.

First, we find that data points representing images

not having characteristics of handwriting by children

with dyslexia are usually in the same cluster. Use

character “丁” (Ding) as an example. The result of

character “ 丁” (Ding) are shown in Table 2 and

Figure 11. In Table 2, all data points in the cluster 0

do not have characteristics of handwriting by children

with dyslexia. Figure 11 shows the cluster

distribution of character “丁” (Ding) in the feature

map.

Second, some words contain more distinguishable

characteristics for identifying children with dyslexia,

such as “學” (Learn), “游” (Swim), “和” (And), while

some are not such as “樹” (Tree), “上” (Up). Figure

12 and Figure 13 show the cluster distribution of

character “和” (And) and “上” (Up) in the feature

maps respectively.

Table 2: Image of character “丁” (Ding) in 3 clusters.

Cluster Sample

Figure 11: Cluster distribution of character “丁” (Ding) in

the feature map.

Figure 12: Cluster distribution of character “和” (And) in

the feature map.

Figure 13: Cluster distribution of character “上” (Up) in the

feature map.

5 CONCLUSIONS AND FUTURE

WORK

In this paper, we have proposed a cloud-based early

identification system of dyslexia. We have designed

and developed a mobile app with AWS cloud

platform as server. For data collecting for model

training, we plan to invite about 30 children with

dyslexia and 200 children without dyslexia from

kindergartens. After the first round of the data

collection, 66 children aged 5-7 were recruited in

Hong Kong. We have identified 27 representative

traditional Chinese characters for handwriting data

collection. We have carried out K-means clustering

algorithm to investigate the characteristics of data

points on a feature map for each character.

In the future, we will design the machine learning

model of the identification system. Besides, we will

continue recruiting children to collect handwriting for

model training. Moreover, we consider oversampling

by applying data augmentation to enlarge the data

size and also look into different techniques to

generate more reliable samples for model training.

ACKNOWLEDGEMENTS

This research was in part supported by grants from

the Research Grants Council of the Hong Kong

Special Administrative Region, China (Project No.

UGC/FDS15/E02/20).

Besides, the authors would like to thank the

research team from Rehabilitation Services (Early

Education & Training) from Yan Chai Hospital

Design an Intelligence System for Early Identiﬁcation on Developmental Dyslexia of Chinese Language

Social Services Department in Hong Kong for their

professional suggestions and their time for carrying

out experiments. Besides, the authors would like to

thank the participating children and their parents who

volunteered their time and valuable feedback on this

study.

REFERENCES

Asvestopoulou, T., Manousaki, V., Psistakis, A.,

Smyrnakis, I., Andreadakis, V., Aslanides, I.M. and

Papadopouli, M. (2019). Dyslexml: Screening tool for

dyslexia using machine learning. arXiv preprint

arXiv:1903.06274.

Biswas, A. and Islam, M.S. (2021). An Efficient CNN

Model for Automated Digital Handwritten Digit

Classification. Journal of Information Systems

Engineering and Business Intelligence, 7(1), pp.42-55.

Chan, D.W., Ho, C.S.H., Tsang, S.M., Lee, S.H. and

Chung, K.K. (2007). Prevalence, gender ratio and

gender differences in reading‐related cognitive abilities

among Chinese children with dyslexia in Hong Kong.

Educational Studies, 33(2), pp.249-265.

Drotár, P. and Dobeš, M. (2020). Dysgraphia detection

through machine learning. Scientific reports, 10(1),

pp.1-11.

Isa, I.S., Rahimi, W.N.S., Ramlan, S.A. and Sulaiman, S.N.

(2019). Automated Detection of Dyslexia Symptom

Based on Handwriting Image for Primary School

Children. Procedia Computer Science, 163, pp.440-

449.

Košak-Babuder, M., Kormos, J., Ratajczak, M. and Pižorn,

K. (2019). The effect of read-aloud assistance on the

text comprehension of dyslexic and non-dyslexic

English language learners. Language Testing, 36(1),

pp.51-75.

Lam, S.S., Au, R.K., Leung, H.W. and Li-Tsang, C.W.

(2011). Chinese handwriting performance of primary

school children with dyslexia. Research in

developmental disabilities, 32(5), pp.1745-1756.

Mahone, E.M. and Schneider, H.E. (2012). Assessment of

attention in preschoolers. Neuropsychology review,

22(4), pp.361-383.

Spoon, K., Crandall, D. and Siek, K. (2019). Towards

Detecting Dyslexia in children’s handwriting using

neural networks. In Proceedings of the International

Conference on Machine Learning AI for Social Good

Workshop, Long Beach, CA, USA (pp. 1-5).

Sprenger-Charolles, L., Colé, P. and Serniclaes, W. (2013).

Reading acquisition and developmental dyslexia.

Psychology Press.

Tseng, M. H. (1998). Development of pencil grip position

in preschool children. Occupational Therapy Journal of

Research, 18, 207-224.

Tsui, C.M., Li-Tsang, W.P.C. and Lung, P.Y. (2012).

Dyslexia in Hong Kong: challenges and opportunities.

InTech.

Usman, O.L. and Muniyandi, R.C. (2020). CryptoDL:

Predicting Dyslexia Biomarkers from Encrypted

Neuroimaging Dataset Using Energy-Efficient Residue

Number System and Deep Convolutional Neural

Network. Symmetry, 12(5), p.836.

Wang, J. and Perez, L. (2017). The effectiveness of data

augmentation in image classification using deep

learning. Convolutional Neural Networks Vis.

Recognit, 11, pp.1-8.

Wei, P., Li, H. and Hu, P. (2019). Inverse discriminative

networks for handwritten signature verification. In

Proceedings of the IEEE/CVF Conference on

Computer Vision and Pattern Recognition.

Wu, Z., Lin, T., and Li, M. (2019). Automated Detection of

Children at Risk of Chinese Handwriting Difficulties

Using Handwriting Process Information: An

Exploratory Study. IEICE TRANSACTIONS on

Information and Systems, 102(1), 147-155.

Xing, L. and Qiao, Y. (2016). October. Deepwriter: A

multi-stream deep CNN for text-independent writer

identification. In 2016 15th International Conference on

Frontiers in Handwriting Recognition (ICFHR) (pp.

584-589). IEEE.

Yogarajah, P. and Bhushan, B. (2020). Deep Learning

Approach to Automated Detection of Dyslexia-

Dysgraphia. In The 25th IEEE International Conference

on Pattern Recognition.

WINSYS 2022 - 19th International Conference on Wireless Networks and Mobile Systems