https://doi.org/10.31449/inf.v43i2.2783 Informatica 43 (2019) 299–303 299 
Design of Intelligent English Writing Self-evaluation Auxiliary System 
Man Liu 
School of Foreign Languages, Changchun Institute of Technology, Changchun, Jilin 130012, China 
E-mail: manliu_lm@aliyun.com 
Keywords: English writing, self-evaluation, auxiliary system, writing teaching 
Received: May 23, 2019 
Since the reform and opening up, the exchanges between China and the world have become more and 
more frequent. English, as a widely used international language, plays an important role in international 
exchanges. English teaching includes five aspects, listening, speaking, reading, writing and translation. 
Writing teaching is very important but difficult. In order to improve students' autonomous writing ability, 
this paper briefly introduced the real-time multi-writing teaching mode and designed an automatic scoring 
algorithm of writing self-evaluation auxiliary system, random sampling based Bayesian classification and 
combinational algorithm. One thousand CET-4 and CET-6 compositions from Chinese Learner English 
Corpus (CLEC) were evaluated, and the scoring effect of Bayesian classification algorithm was also 
tested. The results showed that the accuracy rate, recall rate and F value of the proposed algorithm was 
better than that of Bayesian classification algorithm under 150 feature extraction dimensions, the two 
algorithms had improved scoring effect under the optimal feature extraction dimensions, and the 
improvement amplitude of the algorithm proposed in this study was larger. In summary, the random 
sampling based Bayesian classification and combinational algorithm is effective and feasible as an 
automatic scoring algorithm of writing self-evaluation auxiliary system. 
Povzetek: Za boljše učenje angleščine je za kitajske študente razvita vrsta pripomočkov v obliki 
informacijskih storitev. 
1 Introduction 
As the economic globalization deepens, the 
communication between China and other countries is 
more and more frequent, and the most frequently used 
language is English. English teaching includes five 
aspects, listening, speaking, reading, writing and 
translation, among which writing teaching is the most 
important and difficult part [1]. English writing ability can 
be improved through a large number of writing exercises 
and comments of teachers. However, the ratio of the 
number of teachers to the number of students is very small 
in China. Teachers can not provide guidance to all 
students. If students lack guidance about writing exercises, 
the improvement effect will be greatly reduced. Because 
of the rapid development of computer and natural 
language research, a computer intelligence based self-
correction system has been developed [2]. Despite the fact 
that the writing language is changeable and the effect is 
not ideal in practice, it is still of great help to lighten the 
burden on teachers and improve students' English writing 
level. Deng [3] put forward a design method of cloud 
service platform based intelligent English writing 
auxiliary teaching system, constructed the overall design 
model of English writing teaching system, improved the 
intelligence level of the English writing teaching system 
using Cloud-P2P fusion model, and found that the English 
writing teaching system had favourable adaptability, 
learning ability and reliability. Li [4] selected two parallel 
classes as the object, one was taught by the traditional 
business English writing teaching mode and the other was 
guided by computer assisted technology. The teaching 
quality was evaluated after one academic term, and it was 
found that computer assisted technology had positive 
effect on business English writing. Tsai [5] applied the 
blackboard course management system in English essay 
writing teaching and found that the teaching result of the 
experimental group was superior to that of the control 
group after two academic years. The questionnaire result 
suggested that most of the students had positive learning 
result, indicating the teaching mode could improve the 
effectiveness of English writing learning. This paper 
briefly introduced the real-time multi-writing teaching 
mode and designed the automatic scoring algorithm of the 
writing self-evaluation assistant system. One thousand 
CET-4 and CET-6 compositions from Chinese Learner 
English Corpus (CLEC) were scored. The scoring effect 
of the Bayesian classification algorithm was also tested 
and compared with the automatic scoring algorithm. 
2 English writing teaching mode 
The traditional teaching mode in China is mostly “duck-
stuffing”, which is similar to assembly line. Most colleges 
and universities regard English writing teaching as a part 
of English teaching, or as a subsidiary part, and moreover 
it is not paid much attentions to because of time and 
energy waste. The traditional teaching of English writing 
is usually conducted in the classroom, but teachers usually 
do not pay attention to whether students understand or not 
and only provide students with template sentences and 
simple explanation [6]. 

300 Informatica 43 (2019) 299–303 M. Liu  
 
Figure 1: The flow of real-time multi-writing teaching. 
Guided by the theory of constructivism [7], a new 
teaching mode, real-time multi-writing teaching mode [8] 
has been proposed, and its flow chart is shown in Figure 
1. The whole writing process is divided into 3 parts, 
writing preparation, writing modification, and writing 
summarization. Teachers and students participate in the 
whole process. Teachers take students as the center to 
teach in the classroom and on the network platform. 
Classroom teaching follows the principle of student-
oriented to implement the traditional teaching mode. On 
the network platform, student interaction, human-
computer interaction, teacher-student interaction and self-
interaction can be achieved because of the convenient 
Internet. Besides the Internet, the achievement of the 
above interaction also relies on the writing self-evaluation 
system. Through the objective evaluation of computer and 
based on the evaluation of students and teachers, 
comments and review comments are obtained. 
3 Writing self-evaluation system 
3.1 General structure of the system 
The general structure of the system is shown in Figure 2. 
The system in this paper is a Web system based on B/S 
mode [9]. The client used by user runs in the browser, 
while the business function of the system runs on the 
server. The overall structure of the system is divided into 
3 parts, user interface layer, business logic layer and data 
layer. The user interface is the web browser; the business 
logic layer contains all the functions of the system, and the 
data layer contains the data needed to run the system. 
Automatic scoring is the main function of the system. 
 
 
Figure 2: The overall structure of the system. 
3.2 Automatic scoring 
3.2.1 Feature extraction 
The features of compositions with different scores needed 
to be extracted through training corpus before use to 
facilitate the classification of compositions to be tested 
[10]. In this study, the information gain method was used 
to extract the features of compositions. The expression of 
the information gain method [11] is: 
  
= = =
+ + − =
k
i
i i
k
i
i i i
k
i
i
t a P t a P t P t a P t a P t P a P a P t IG
1 1 1
) ( log ) ( ) ( ) ( log ) ( ) ( )) ( log( ) ( ) (
(1) 
where t is the feature of adjacent binary phrase, 
i
a is 
the set of compositions with the i-th score, t stands for 
the condition in case of absence of feature t , ) (
i
a P stands 
for the possibility of score 
i
a in the training corpus, ) (t P 
stands for the possibility of composition containing 
feature t in the training corpus, ) ( t a P
i
 stands for the 
possibility of composition containing feature t and with 
score 
i
a , ) (t P stands for the possibility of composition 
not containing feature t , ) ( t a P
i
 stands for possibility of 
composition which is scored as 
i
a but not contains feature 
t , k stands for the number of score grade, 4 here (grade 1: 
1 ~ 5 points; grade 2: 6 ~ 9 points; grade 3: 10 ~ 13 points; 
grade 4: 13 ~ 14 points). 
3.2.2 Random sampling and Bayesian 
classification based composition scoring 
algorithms 
The flow of the scoring algorithm [12] is shown in Figure 
3. 

 
= =
m i
i
y I Y
1
) ( max arg 
 (2), where α refers to 
different score grades, 
i
y refers to the classification result 
of a random sampling, ) • ( I refers to indicator function, 1 
if the parameter is true and 0 if not. 
 
Figure 3: The flow of random sampling and Bayesian classification based composition scoring algorithms. 

Design of Intelligent English Writing Self-evaluation... Informatica 43 (2019) 299–303 301 
The calculation formula of Bayesian classification 
based algorithm [13] is: 

=
=
V
k k
N
i k
i
N
a P
a b P
k
1
!
) (
) (

 (3) 
V N
N
a P
i
a
ji
i j
+
+
=
1
) ( 
   (4) 
where b refers to the spatial vector of an English 
composition,  
m i
b       , , , ,
2 1
= , m refers to the 
number of features of adjacent binary phrase, 
i
θ refers to 
the weight of the i-th feature in composition b, V refers to 
the number of binary phase features, 
k
N refers to the 
number of times of the k-th feature appearing in b, 
) (
i
a b P refers to the possibility of a composition 
obtaining some score, ) (
i j
a θ P refers to the possibility of 
feature 
j
θ in a composition which is scored as 
i
a , 
ji
N 
refers to the number of times of the j-th feature appearing 
in a composition which is scored as the j-th feature, and 
i
a
N refers to the number of all the features of a 
composition which is scored as 
i
a . 
4 Automatic scoring test of the 
writing self-evaluation system 
4.1 Testing method 
Five hundred of CET-4 compositions which involved two 
themes and 500 CET-6 compositions with which involved 
two themes were selected from CLEC [14]. Before the 
test, the compositions were scored and classified into four 
score grades according to the scoring criteria of CET-4 
and CET-6 compositions. Based on the binary phrase 
features, the compositions which involved four themes 
were scored under 150 feature extraction dimensions, i.e., 
the 1000 compositions were classified into four score 
grades using the algorithm proposed in this study. Then 
the compositions were scored after the automatic 
calculation of optimal feature dimensions. The whole 
system program ran on a server in a lab. The server was 
configured with Windows 7 system, I7 processor and 16G 
memory. In order to increase the persuasiveness of the 
results, the test results of the Bayesian classification 
algorithm based scoring system was selected for 
comparison. The test method was the same as the system 
proposed in this study. 
4.2 Evaluation criteria 
The automatic scoring effect of the writing self-evaluation 
system was evaluated in the aspects of accuracy rate, recall 
rate and F value [15]. The accuracy rate could be 
calculated using the following formula: the accuracy rate 
= correctly recognized number/total recognized number, 
where correctly recognized number is the number of 
correctly classified compositions based on above scoring 
algorithm and the total recognized number is the total 
number of compositions identified by the scoring 
algorithm. 
The recall rate could be calculated using the following 
formula: recall rate = correctly recognized 
number/actually existed number, where actually existed 
number refers to the number of compositions which 
actually existed and ought to be recognized. 
F value (comprehensive evaluation index) could be 
calculated using the follow formula: F value = 2 × 
accuracy rate × recall rate /(accuracy rate + recall rate). 
4.3 Testing results 
As shown in Table 1, the accuracy rate, recall rate and F 
value of the scoring algorithm were about 14%, 28% and 
22% higher than those of Bayesian classification 
algorithm in scoring different themes of compositions. 
The accuracy rate, recall rate and F value of the Bayesian 
classification algorithm were 0.739, 0.661 and 0.690 
respectively; the accuracy rate, recall rate and F value of 
the algorithm proposed in this study were 0.845, 0.850 and 
0.844 respectively. It indicated that the algorithm 
proposed in this study was superior to the Bayesian 
classification algorithm in scoring compositions, and all 
Theme Bayesian classification algorithm Random sampling based Bayesian 
classification algorithm 
Accuracy 
rate (P) 
Recall rate 
(R) 
F value Accuracy 
rate (P) 
Recall rate 
(R) 
F value 
Theme 1 
(CET-4) 
0.719 0.626 0.660 0.849 0.881 0.862 
Theme 2 
(CET-4) 
0.750 0.679 0.708 0.838 0.833 0.832 
Theme 3 
(CET-6) 
0.759 0.710 0.730 0.836 0.831 0.831 
Theme 4 
(CET 6) 
0.726 0.630 0.661 0.857 0.853 0.851 
Average 
value 
0.739 0.661 0.690 0.845 0.850 0.844 
Table 1: The scoring test results of the two algorithms under 150 feature extraction dimensions. 

302 Informatica 43 (2019) 299–303 M. Liu  
the indicators were above 80%. Based on the binary phrase 
features extracted from the compositions, the algorithm 
could accurately classify the tested composition into the 
corresponding score grade and make accurate and 
reasonable evaluation on the content of compositions 
based on the score grade. 
As shown in Table 2, the optimal feature extraction 
dimension of theme 1, 2, 3 and 4 compositions was 390, 
350, 710 and 530 respectively. Considering the evaluation 
index values of the four themes, the accuracy rate, recall 
rate and F value of Bayesian classification algorithm were 
0.841, 0.821 and 0.799 respectively under the optimal 
feature extraction dimension, and the accuracy rate, recall 
rate and F value of the algorithm proposed in this study 
were 0.992, 0.991 and 0.991 respectively. 
As shown in Figure 4, the accuracy, recall rate and F 
value of the algorithm under the optimal feature extraction 
dimensions were better than those under 150 feature 
extraction dimensions. Moreover the three indicators of 
the algorithm proposed in this study were always the best, 
and the improvement of the algorithm was greater after 
changing feature extraction dimensions. That is to say, 
after the application of optimal feature extraction 
dimension, the algorithm proposed in this study could 
classify the compositions more accurately according to the 
binary phrase features, and make more accurate and 
reasonable evaluation according to the score grade. 
 
Figure 4: The comparison of testing results of the two 
algorithms under the fixed and optimal feature extraction 
dimensions. 
5 Conclusion 
This paper briefly introduced the real-time multi-writing 
teaching model and designed an automatic scoring 
algorithm for the writing self-evaluation system. Then 
1000 CET-4 and CET-6 compositions in CLEC were 
scored according to four grades. As a comparison, the 
scoring effect of the Bayesian classification algorithm was 
tested. The results showed that the accuracy rate, recall 
rate and F value of the Bayesian classification algorithm 
were 0.739, 0.661 and 0.690 respectively, and the 
corresponding data of the algorithm proposed in this study 
were 0.845, 0.850 and 0.844 respectively, indicating that 
the scoring effect of the algorithm proposed in this study 
was superior to that of the Bayesian classification 
algorithm. Under the optimal feature extraction 
dimensions, the accuracy rate, recall rate and F value of 
the Bayesian classification algorithm were 0.841, 0.821 
and 0.799 respectively and those of the algorithm 
proposed in this study were 0.992, 0.991 and 0.991 
respectively, which were improved compared to under the 
fixed feature extraction dimensions. Moreover the 
improvement amplitude of the algorithm proposed in this 
study was larger. 
Acknowledgement 
This study was supported by General Project of Higher 
Education Department of Education Department of Jilin 
Province in 2018, Research on English “Four-in-One” 
Teaching Reform in College English under the 
Background of Internet+, Key Project of Teaching Reform 
of Changchun University of Engineering, June 2018 to 
June 2020, under researching. 
References 
[1] Nattapong J, Rattanavich S (2015). The effects of 
computer-assisted instruction based on top-level 
structure method in English reading and writing 
abilities of Thai EFL students. English Language 
Theme Optimal 
feature 
dimension 
number 
Bayesian classification algorithm Random sampling based Bayesian 
classification algorithm 
Accuracy 
rate (P) 
Recall 
rate (R) 
F 
value 
Accuracy 
rate (P) 
Recall 
rate (R) 
F value 
Theme 1 
(CET-4) 
390 0.856 0.846 0.838 0.990 0.989 0.989 
Theme 2 
(CET-4) 
350 0.784 0.778 0.763 0.989 0.988 0.988 
Theme 3 
(CET 6) 
710 0.878 0.866 0.857 1 0.999 0.999 
Theme 4 
(CET 6) 
530 0.844 0.795 0.739 0.987 0.986 0.986 
Average 
value 
 0.841 0.821 0.799 0.992 0.991 0.991 
Table 2. The scoring test results of the two algorithms under optimal feature extraction dimensions. 

Design of Intelligent English Writing Self-evaluation... Informatica 43 (2019) 299–303 303 
Teaching, 8(11), pp. 231. 
https://doi.org/10.5539/elt.v8n11p231 
[2] Xia M, Zhang Y, Zhang C (2017). A TAM-based 
approach to explore the effect of online experience 
on destination image: A smartphone user's 
perspective. Journal of Destination Marketing & 
Management. 
https://doi.org/10.1016/j.jdmm.2017.05.002 
[3] Deng L (2018). Design of English writing assisted 
instruction teaching system based on intelligent 
cloud service platform. International Conference on 
Intelligent Transportation, Big Data & Smart City. 
IEEE Computer Society, pp. 287-290. 
https://doi.org/10.1109/ICITBS.2018.00080 
[4] Li X (2018) Influence of computer-aided instruction 
model on business English writing teaching effect. 
International Journal of Emerging Technologies in 
Learning, 13(3), pp. 197.  
[5] Tsai Y R (2015). Applying the technology 
acceptance model (TAM) to explore the effects of a 
course management system (CMS)-assisted EFL 
writing instruction. Calico Journal, 32(1). 
https://doi.org/10.1558/cj.v32i1.153-171 
[6] Tarhini A, Hassouna M, Abbasi M S, et al. (2015). 
Towards the acceptance of RSS to support learning: 
an empirical study to validate the technology 
acceptance model in Lebanon. Electronic Journal of 
e-Learning, 13(1), pp. 30-41.  
[7] Li L (2016). Design and implementation of higher 
vocational English writing and training testing 
system. International Conference on Materials 
Engineering, Manufacturing Technology and 
Control.   
[8] Chapela M E G (2016). A corpus-based analysis of 
post-auxiliary ellipsis in modern English: 
methodological and theoretical issues.  
[9] Ohta R (2017). The impact of an automated 
evaluation system on student-writing performance. 
Kate Bulletin, 22, pp. 23-33.  
[10] Ahsen M E, Ayvaci M, Raghunathan S (2017). 
When algorithmic predictions use human-generated 
data: a bias-aware classification algorithm for breast 
cancer diagnosis. Social Science Electronic 
Publishing. https://doi.org/10.2139/ssrn.3087467 
[11] Alcázar V, Fernández S, Borrajo D, et al. (2015). 
Using random sampling trees for automated 
planning. Ai Communications, 28(4), pp. 665-681.  
[12] Myttenaere A D, Golden B, Grand B L, et al. (2015). 
Study of a bias in the offline evaluation of a 
recommendation algorithm. Computer Science, 
79(3), pp. 263-72. https://doi.org/10.1016/0002-
9416(81)90074-9 
[13] Hao J, Niu K, Meng Z, et al. (2017). A collaborative 
filtering recommendation algorithm based on score 
classification. International Conference on Security, 
Privacy and Anonymity in Computation, 
Communication and Storage. Springer, Cham, pp. 
435-445. https://doi.org/10.1007/978-3-319-72395-
2_40 
[14] Santos D P D, André C P L F D C (2017). Automatic 
selection of learning bias for active sampling. 
Intelligent Systems. IEEE, Recife, Brazil, pp. 55-60. 
https://doi.org/10.1109/BRACIS.2016.021 
[15] Zoph B, Yuret D, May J, et al. (2016). Transfer 
learning for low-resource neural machine 
translation. Conference on Empirical Methods in 
Natural Language Processing, pp. 1568-1575. 
  

304 Informatica 43 (2019) 299–303 M. Liu