



























































































Yu Li et al. /Future Technology                                                                                        November 2025| Volume 04 | Issue 04 | Pages 
100-116 

100 
 

 

 

Article 

POA-MLSP: a multi-dimensional learning analytics 
framework for predicting CET4 writing 
performance based on a production-oriented 
approach and student engagement patterns 
Yu Li1, Nur Ainil BT. Sulaiman1*, Halizah BT. Omar2  
1Faculty of Education, Universiti Kebangsaan Malaysia, 43600 UKM Bangi, Selangor, Malaysia 
2Pusat Pengajian Citra Universiti, Universiti Kebangsaan Malaysia, 43600 UKM Bangi, Selangor, Malaysia 

A R T I C L E   I N F O 
 

Article history: 
Received 01 June 2025  
Received in revised form 
17 July 2025 
Accepted 06 August 2025 
 
Keywords:  
Production-oriented approach,  
Multi-dimensional learning analytics, 
CET-4 writing performance prediction, 
Student engagement patterns, 
Self-determination theory 
 
*Corresponding author 
Email address: 
nurainil@ukm.edu.my 
 
 
DOI: 10.55670/fpll.futech.4.4.9 

A B S T R A C T 
 

Contemporary College English Test Band 4 (CET-4) writing instruction faces 
significant challenges in accurately predicting student performance and 
providing timely pedagogical interventions. This study develops and validates 
the Production-Oriented Approach Multi-Dimensional Learning Analytics 
Framework for Student Performance (POA-MLSP) for predicting CET-4 writing 
performance across five dimensions through systematic integration of 
Production-Oriented Approach (POA) theory and Self-Determination Theory 
(SDT)-based engagement modeling. The framework implements a four-layer 
architecture incorporating Feature Adaptive Selection Mechanism and SDT-
Based Engagement Dynamic Modeling algorithms. Validation involves 124 
students during a 16-week semester, collecting multi-source data including 
Jacobs' five-dimensional assessments, Utrecht Work Engagement Scale-Student 
(UWES-S) engagement measurements, classroom observations, and digital 
platform interactions across experimental and control groups. POA-MLSP 
achieves R² = 0.75 overall prediction accuracy, outperforming linear regression 
(R² = 0.58), random forest (R² = 0.66), and support vector machines (R² = 0.63) 
by 17-29%. Content prediction reaches highest accuracy (R² = 0.78), while the 
framework identifies five distinct engagement profiles and achieves 78.4% ± 
2.1% early warning accuracy with 79.8% ± 2.9% teacher satisfaction. 
Educational theory-guided algorithms significantly enhance prediction 
performance while maintaining pedagogical interpretability, enabling 
proactive intervention through early warning systems with minimal 
implementation burden for authentic educational applications. 

1. Introduction 
The Production-Oriented Approach (POA) has emerged 

as a revolutionary pedagogical framework within Teaching 
English as a Foreign Language (EFL) contexts, representing a 
paradigmatic shift from input-based teaching methodologies 
toward a more integrative theory-driven pedagogical process 
designed to enhance comprehensive learning outcomes. 
Contemporary research demonstrates that POA 
implementation within tertiary educational contexts 
effectively bridges the gap between language learning and 
language utilization through its comprehensive three-stage 
instructional model encompassing motivating, enabling, and 
assessing phases [1]. The theoretical depth of POA comes 
through multi-theoretical fusion (systematic integration of 
complementary educational frameworks) unifying several of 

the most basic theoretical frameworks, such as Krashen's 
Input Hypothesis, Vygotsky´s Social Constructivism and 
cognitive processing theories of writing development, as well 
as sociocultural perspectives of language learning, in a multi-
faceted approach that attends to various dimensions of 
language learning paths at the same time [2]. Classroom 
practices and implementations suggest that POA-based 
instruction is highly effective at improving student writing 
performance on a range of competencies when partnered 
with reciprocal teaching models that leverage peer 
interaction and collaborative learning channels [3]. The 
capability of POA to encompass both general and business. 
English teaching, as well as specialized academic teaching, is 
a testimony to its theoretical robustness and practical fit for 
diverse learning goals and student groups [4].   

 

 

November 2025| Volume 04 | Issue 04 | Pages 100- 116 

ISSN 2832-0379 

Open Access Journal 
Journal homepage: https://fupubco.com/futech 

 

https://doi.org/10.55670/fpll.futech.4.4.9 

 

Future Technology 

mailto:nurainil@ukm.edu.my
https://fupubco.com/futech
https://doi.org/10.55670/fpll.futech.4.4.9


Yu Li et al. /Future Technology                                                                                        November 2025| Volume 04 | Issue 04 | Pages 100-116 

101 
 

 
Contrary to the studies described in the previous 

sections which investigated the effects of POA on language 
learning in comparison to that of TVA, recent POA 
experiments carried out in the context of university 
Classrooms in China show the success of POA in ameliorating 
learner motivation and engagement, factors that, up to now, 
have been depression paving the way for language learning 
progress [5]. The inclusion of components such as cultural 
features and technology-enhanced learning environments in 
POA frameworks reflects the comprehensiveness of the 
approach to change and its ability to evolve and respond to 
the needs of modern education [6, 7]. More advanced 
applications of POA in the flipped classroom, as well as 
teacher training programs, reveal the scalability and potential 
of the approach to revolutionize language education at 
various educational levels and professional development 
contexts [8, 9]. Theoretical underpinnings of student 
engagement in POA instruction. Based on Self-Determination 
Theory (SDT), which yields important understandings about 
the motivational factors that are associated with language 
learning effectiveness and the intricate relationship between 
psychological need fulfilment and academic achievement, 
student engagement in POA instruction has been extensively 
argued. Studies show that learning environments in which 
freedom of choice, competence, and socially relatedness are 
supported can be conducive to sustained involvement and 
enhanced learning outcomes [10].   

The difference between intrinsic and extrinsic 
motivation orientations is especially important in the context 
of a POA, as challenging activities on unlimited progression 
levels or limited progression levels must strike a balance 
between challenge and the capability level of the learner to 
ensure that motivation remains at a high level and no 
disengagement occurs [11]. DLAs provide opportunities for 
the facilitation of students’ engagement through well-
designed technological interventions, which are in alignment 
with SDT, especially in blended learning settings where the 
conventional classroom environment is complemented with 
online elements [12]. During this transition to the pandemic 
mode of online learning, a better understanding of what 
drives students to engage across different delivery and online 
modalities has also become clear, demonstrating the power of 
ensuring sustained psychological need satisfaction regardless 
of delivery form [13]. Collaborative learning approaches 

within English language programs demonstrate significant 
potential for enhancing engagement through peer support 
mechanisms, though the effectiveness of such interventions 
depends heavily on group dynamics and task design 
considerations [14]. The complex relationships between 
motivational factors, including the mediating effects of 
emotional states and the fulfillment of basic psychological 
needs, create dynamic engagement patterns that directly 
influence learning behaviors and academic outcomes [15]. 
Contemporary research into self-directed e-learning 
environments reveals the critical role of social support, self-
regulated learning strategies, and flow experiences in 
sustaining long-term engagement with language learning 
activities [16].  

The application of educational data mining and learning 
analytics to language instruction represents an emerging 
frontier that offers substantial potential for understanding 
and predicting student performance patterns, though the 
integration of these technologies with established 
pedagogical theories remains largely unexplored. Systematic 
reviews of predictive modeling applications in educational 
contexts demonstrate the effectiveness of data-driven 
approaches for early identification of at-risk students and 
personalized intervention strategies, particularly when 
applied to large-scale educational datasets [17]. Machine 
learning algorithms have shown remarkable success in 
predicting academic performance across diverse educational 
contexts, with particular effectiveness in identifying subtle 
patterns and relationships that traditional assessment 
methods fail to capture [18]. The design of complex 
forecasting models with the use of ensemble methods and 
highly developed feature selection techniques has resulted in 
substantial enhancements of forecasting accuracy in the case 
of the prediction of students’ performance, which creates 
significant opportunities in terms of educational applications 
[19]. By now, the incorporation of artificial intelligence in 
intelligent tutoring systems provides overwhelming evidence 
of the powerful effects of technology-based instruction, 
especially if developed with sustainability and adaptability in 
mind [20]. The recent introduction and rapid development of 
predictive learning analytics over the last ten years have laid 
the foundation for sound methodological techniques for 
analyzing educational data, but there is still a great deal of 
opportunity to leverage these technologies with theory-
informed instructional approaches [21].  

Recent advancements in the application of artificial 
intelligence technologies in education environments are 
indicative of the potential benefits offered as well as the 
challenges to the deployment of technology in traditional 
education [22]. State-of-the-art work in educational data 
mining considers complex student performance prediction 
algorithms with dynamic feature selection and ensemble 
evolution approaches to enable improved accuracy and 
interpretability in educational use cases [23, 24]. The 
assessment environment in the College English Test Band 4 
(CET-4) writing domain poses unique difficulties for 
performance prediction and instructional optimization. High-
stakes large-scale writing tests, which form an indispensable 
basis from which language assessment is conducted, manifest 
certain limitations in the scoring system framework with 
potential influence on the reliability and validity of scoring of 
performance [25].  

Comparative studies between automated evaluation and 
previous studies indicate that there is a large gap between the 
assessment results, and there are some data conflicts among 
them, so a more advanced mechanism should be developed 

Abbreviations 
AI  Artificial Intelligence  
CEFR  Common European Framework of Reference 
CET-4  College English Test Band 4 
DLAs  Digital Learning Activities 
EDA  Educational Data Analytics 
EFL  English as a Foreign Language 
MAE  Mean Absolute Error  
ML  Machine Learning 
MLSP  Multi-dimensional Learning Analytics  

Framework for Student Performance 
POA  Production-Oriented Approach 
R²  Coefficient of Determination  
RMSE  Root Mean Square Error  
SDT  Self-Determination Theory 
SVM  Support Vector Machine 
TVA  Traditional Approach 
UWES-S  Utrecht Work Engagement Scale-Student  
 



Yu Li et al. /Future Technology                                                                                        November 2025| Volume 04 | Issue 04 | Pages 100-116 

102 
 

for further study, to balance the relationship between high-
speed computing and the cost-effectiveness in the pedagogy 
for assessment [26]. Cross-sectional studies of CET-4 test 
components reveal the interdependency of language abilities 
and the significance of a comprehensive test approach 
focusing on various skill dimensions at the same time [27]. 
Current assessments of fair and non-discriminatory testing 
for the college English examinations raise concerns about 
underlying systemic biases that may disproportionately 
disadvantage certain groups of students, which require fairer 
and more compelling frameworks of assessment [28]. The 
introduction of peer assessment systems into college 
students’ English writing instruction provides potential 
options to replace “teacher-centered” assessment, but the 
reality of effectiveness in using peer assessment is also 
different among various types of educational systems and 
needs careful training and support [29]. Longitudinal studies 
of peer feedback in academic writing development have 
shown sustained improvements in student performance 
where the collaborative assessment was well-structured and 
supported [30].  

Collaborative writing approaches, which stress the role 
of peer feedback, interaction on dynamics, and learning from 
each other, have proven to result in increased learning 
outcomes owing to the social interaction and shared 
knowledge building by the group members [31]. Cross-
cultural investigations of valuations that students assign to 
peer feedback in contrasting LSEs have revealed that the 
emotional aspect of the emotional dimension of CLA, and the 
motivational factors that contribute to attitudes to peer 
feedback, affect the utility of collaborative peer assessment 
practices and have implications for the provision of culturally 
sensitive implementation resources [32].  

Despite extensive theoretical development of POA 
methodology and significant advances in educational data 
analytics, a substantial research gap persists between these 
two domains within CET-4 writing instruction contexts. 
Current predictive modeling approaches in language 
education predominantly rely on statistical correlations 
without incorporating established pedagogical theories, 
resulting in limited educational interpretability and reduced 
practical utility for classroom instruction. Furthermore, 
existing assessment frameworks fail to capture the dynamic 
nature of student engagement patterns as conceptualized 
through SDT principles, thereby limiting the effectiveness of 
personalized intervention strategies and compromising the 
potential for data-driven pedagogical decision-making in 
authentic educational environments. Although POA has been 
well developed in theory and educational data analytics 
(EDA) has become more advanced, there is still a big gap to 
bridge the two in CET-4 writing teaching.  

This investigation establishes three primary objectives: 
(1) to develop the POA-MLSP framework that systematically 
integrates POA theoretical constructs with advanced learning 
analytics for multi-dimensional writing performance 
prediction; (2) to establish empirical validation of SDT-based 
engagement modeling capabilities for early identification of 
at-risk students within CET-4 writing instruction contexts; 
and (3) to demonstrate the practical utility and educational 
interpretability of theory-informed machine learning 
approaches compared to traditional statistical prediction 
methods. The framework specifically targets the prediction of 
student writing performance across Jacobs' five assessment 
dimensions while maintaining pedagogical soundness and 
computational efficiency suitable for authentic classroom 
implementation.  

This novel approach obtains complete prediction models 
that facilitate decision-making of instructors in runtime and 
personal offers planning of intervention strategies, which can 
preserve the pedagogical soundness and make use of the 
computational power to better serve learners, and hence also 
provide a theoretical design for intelligent educational 
systems. 

2. Methods 
2.1 POA-MLSP framework educational theoretical 

foundation 
The theoretical foundation of POA-MLSP is set through 

the step-by-step embedding of Production-Oriented 
Approach principles and Self-Determination Theory 
mechanisms, which turns on to be a full-stack educational 
data modeling approach to handle the intricate dynamics of 
CET-4 writing instruction. The data modeling for POA theory 
leverages the abundant behavioral and cognitive data 
collected in the three-phase instructional cycle and translates 
qualitative pedagogical processes into machine learning-
friendly numerical data while maintaining the educationally 
valid aspects of the theoretical framework. In the Motivating 
phase, the behaviour aspects that the model itself is expected 
to indicate that the student’s learning-motivation has been 
activated are modelled in the form of preconditions that the 
variables are driven to some values by the data flows like the 
student response patterns to communicative scenarios, the 
engagement of working with authentic materials, and the 
frequency with which the student has entered into class 
discussion, as shown in Figure 1. 

Figure 1 shows the comprehensive POA-MLSP 
theoretical foundation and implementation framework, 
illustrating systematic data collection across POA's three 
phases (motivation activation, knowledge construction, and 
feedback interaction), SDT three-dimensional integration 
(autonomy, competence, relatedness), multi-dimensional 
data integration with temporal alignment and contextual 
factors, education-theory-guided feature engineering, and 
multi-theoretical integration processing, culminating in the 
POA-MLSP predictive framework with continuous model 
refinement capabilities. The Enabling phase quantification 
focuses on knowledge construction processes through 
participation measurement metrics that include collaborative 
task engagement duration, peer interaction frequency, 
scaffolding utilization patterns, and self-regulation behavior 
indicators captured through learning management system 
logs and classroom observation protocols. Assessment phase 
evaluation concentrates on feedback interaction quality 
through quantitative analysis of peer evaluation accuracy, 
self-assessment reliability coefficients, teacher feedback 
incorporation rates, and iterative improvement patterns 
demonstrated across multiple writing drafts.  

The multi-theoretical integration guidance within 
predictive modeling draws upon the systematic review 
findings that demonstrate enhanced effectiveness when POA 
incorporates complementary theoretical frameworks, 
including Input Hypothesis principles, Social Constructivism 
mechanisms, Cognitive Process Theory applications, and 
Sociocultural Theory perspectives. This integration approach 
creates feature engineering protocols that capture the 
synergistic effects of theoretical convergence, enabling the 
prediction model to account for the complex 
interdependencies between different learning mechanisms 
operating simultaneously within POA instruction.  

 
 



Yu Li et al. /Future Technology                                                                                        November 2025| Volume 04 | Issue 04 | Pages 100-116 

103 
 

 
 
 
 
The theoretical synthesis mechanism extends beyond 

simple additive combination to establish dynamic interaction 
models that recognize how Input Hypothesis comprehensible 
input requirements influence Social Constructivism 
collaborative learning effectiveness, while Cognitive Process 
Theory writing strategies interact with Sociocultural Theory 
contextual factors to create emergent learning behaviors that 
transcend individual theoretical predictions, as detailed in 
Table 1. The multi-theoretical integration framework 
establishes a systematic mapping between educational 
theories and system architecture components. Input 
Hypothesis principles directly inform comprehensible input 
processing within the Feature Adaptive Selection Mechanism. 
Social Constructivism guides collaborative pattern 
recognition algorithms in the processing layer. Cognitive 
Process Theory shapes metacognitive tracking mechanisms 
throughout the prediction framework. Meanwhile, 
Sociocultural Theory influences contextual adaptation 
parameters across all system layers. This theoretical 
convergence ensures that each computational component 
maintains educational validity while contributing to the 
overall predictive capability. POA-MLSP in combination with 
the Self Determination Theory brings up an 
operationalization of the three basic motivational needs 
based on elaborate measurements which map the abstract 
motivational constructs to concrete behavior indicators, 
adjusted to algorithm processability. Autonomy dimension 
measurement encompasses learning self-selection and 
control behavior indicators, including decision-making 
frequency within learning activities, self-directed learning 
time allocation, autonomous task initiation rates, and digital 
learning environment goal-setting behaviors captured 
through learning journal interactions and reflective practices. 

 
 
 
 
 Assessment of competence dimensions emphasizes 

measures for the efficacy of learning and the experience of 
achievement by performance confidence ratings, frequency of 
pursuit of challenges, indices of mastery goal orientation, and 
patterning of success attributions recorded in writing tasks 
and self-appraisal protocols. Meaning At the level of analysis 
of relatedness, peer interaction and teacher-student 
relationship assessment is being conducted via the method of 
social network analysis of classroom communication 
patterns, participation rates in collaboration learning, 
frequency of help-seeking behavior, and the sharing or 
reproducing of fostering conditions as interpreted through 
social support across group activities and peer feedback 
sessions.  

The 3D engagement model integrates these 
heterogeneous aspects through higher-level temporal 
103odelling techniques, in order to capture the evolutionary 
response of engagement patterns throughout the academic 
semester and, ultimately, to pinpoint those transition points 
where intervention mechanisms can be deployed more 
effectively without conflicting with students’ autonomous 
learning and intrinsic motivation. Sophisticated measure of 
psychological need satisfaction that is sensitive to situational 
factors that shape the dynamics of needs and motivation, e.g., 
how difficult the task is perceived, the effects of social 
comparison, receiving high quality feedback, and the 
availability of environmental support renders our model a 
powerful tool for predicting and understanding engagement 
taking into account the intricate relation between individual 
and instructional context variables. 

 

 

Motivating Phase
Learning Motivation Activatior

Enabling Phase
Knowledge Construction Process

Assessing Phase
Feedback Interaction Quality

Response patterns to
Communicative 

scenarios

Engagement levels 
with

authentic materials

Class discussion
participation 

frequency

Motivation activation
behavioral indicators

Collaborative Task
engagement duration

Peer interaction
frequency patterns

Scaffolding 
utilization

patterns (LMS logs)

Self-regulation
behavior indicators

Peer evaluation
accuracy metrics

Self assessment
reliability 

coefficients

Teacher feedback
incorporation rates

Iterative 
improvement

patterns across drafts

Multi-Dimensional Data Integration Layer
Temporal Alignment | Individual Baseline Normalization | Contextual Factors

Education-Theory-Guided 
Feature Engineering

Pattern Recognition 
&Educational 
Interpretability

Multi-Theoretical 
Integration Processing

Educational Domain
Knowledge Integration

POA-MLSP Predictive Framework
Five-Dimensional Writing Performance Prediction

SDT Three-Dimensional Integration
Autonomy| Competence | 

Relatedness

Continuous 
Model 

Refinemen

Figure 1. OA-MLSP theoretical foundation and implementation framework 



Yu Li et al. /Future Technology                                                                                        November 2025| Volume 04 | Issue 04 | Pages 100-116 

104 
 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

2.2 POA-MLSP intelligent prediction framework design 
The overarching design remains child-centric based on a 

four-layer model, which holds that pedagogical 
considerations would drive technological implementation 
and not vice versa.  The input layer allows multi-dimensional 
input data to be collected on the POA teaching activities, log 
data of real-time classroom interaction, engagement intensity 
on a digital learning platform, trajectory of the writing 
portfolio development, data on the network of peer 
collaboration, or on the teacher's observation records that 
follow the structuring principles of POA theories. 
Implementation of the processing layer. In our processing 
layer implementation, educational-theory-guided feature 
engineering and pattern recognition are emphasized, which 
can translate raw educational data into meaningful predictive 
features and maintain their interpretability in front of 
educational stakeholders, where the processed features can 
include domain knowledge from language learning research 
to warrant the feature relevance and pedagogical validity. 
Within the processing layer, specialized normalization 
techniques, based on data preprocessing protocols (e.g., to 
take into account of differences between individual learners 
on their initial level, on their learning trajectory or on their 
corresponding situation also called instructional context and 
generating factor), are implemented taking into account 
measurement validity between different groups of 
students/intonation context. Temporal alignment routines 
are used to observe data recorded from different phases of 
POA instruction remain temporally consistent while allowing 
for variations in individual and group rates of improvement 
and in the ebb and flow of interest and attention that is a 
hallmark of normal classroom settings.nDesign of prediction 
layer with the intention to implement teaching/application-
oriented five dimensions writing ability prediction based on 
which the granular performance prediction for the five 
dimensions (Content, Organization, Language Use, 
Vocabulary and Mechanics) which defined on Jacobs'  

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

assessment framework can be supported directly. We place 
emphasis on support for teacher decision-making, at the 
application layer, by interpreting recommendations and 
offering insights that are actionable for instructional changes, 
student groupings, and an individualization strategy (all in an 
interpretable and transparent manner about the level of 
confidence in the predictions and how uncertain 
predictions/conclusions are) as depicted in Figure 2. This 
Figure shows the comprehensive four-layer architecture 
demonstrating data flow from POA instructional activities 
through feature processing to prediction generation and 
educational application, emphasizing the bidirectional 
feedback mechanism that enables continuous model 
refinement based on educational outcomes. 

Core algorithm educational adaptation design replaces 
traditional attention mechanisms with education-oriented 
approaches that align with established pedagogical theories 
and classroom realities. The Feature Adaptive Selection 
Mechanism represents a novel alternative to conventional 
attention architectures, incorporating educational domain 
knowledge to guide feature importance assessment rather 
than relying solely on statistical correlations that may lack 
pedagogical meaning, as demonstrated in Algorithm 1. This 
mechanism implements a POA teaching principle-based 
feature importance adjustment that prioritizes educationally 
meaningful variables. The algorithm emphasizes student 
engagement indicators, collaborative learning participation 
rates, and progress trajectory patterns correlating with 
sustained learning improvement while maintaining 
pedagogical validity beyond statistical optimization. The 
multi-theoretical fusion (the systematic integration of 
complementary educational frameworks within algorithmic 
design) feature weight optimization algorithm integrates 
insights from Input Hypothesis, Social Constructivism, 
Cognitive Process Theory, and Sociocultural Theory to create 
balanced feature representations that reflect the complex 
interactions between different learning mechanisms 
operating within POA instruction. 

Table 1. Multi-theoretical integration feature mapping 

Theoretical Framework Feature Category Measurement Indicators POA Phase 
Integration 

Weight Optimization 

Input Hypothesis Comprehensible Input 
Processing 

• Input complexity levels 
• Comprehension accuracy 

rates 
• Input-output gap analysis 

Motivating → 
Enabling 

Dynamic based on 
proficiency 

Social Constructivism Collaborative Learning 
Patterns 

• Peer interaction frequency 
• Knowledge co-

construction events 
• Scaffolding effectiveness 

Enabling → 
Assessing 

Group dynamics 
weighted 

Cognitive Process Theory Writing Strategy 
Application 

• Planning behavior 
indicators 

• Revision pattern analysis 
• Metacognitive strategy use 

All phases Individual cognitive 
load 

Sociocultural Theory Contextual Mediation 
Factors 

• Cultural background 
influence 

• Social context adaptation 
• Tool-mediated learning 

Motivating + 
Assessing 

Context-sensitive 
adjustment 

Multi-Theory Synergy Cross-Framework 
Interactions 

• Input-collaboration 
correlation 

• Strategy-context 
alignment 

• Emergent learning 
behaviors 

Integrated across 
phases 

Synergistic 
amplification 

 



Yu Li et al. /Future Technology                                                                                        November 2025| Volume 04 | Issue 04 | Pages 100-116 

105 
 

 
Figure 2. POA-MLSP framework architecture 

Its real-time adaptation machinery allows for adapting 
prediction models to changing classroom dynamics and 
student response patterns, ensuring that the suggestions 
generated by the algorithm remain sensitive to the 
continuous evolution of educational scenarios as well as to 
the needs and rules of traditional pedagogical approaches. 
Cross-validation schemes such as the one used here and 
designed especially for educational settings are not only 
sensitive for the inherent temporal dependencies present in 
learning data, but they also allow us to have the evaluation of 
the performance of the model as close as possible to the 
expected performance of the model under a realistic use case, 
where prediction accuracy will need to be traded off with 
educational interpretability needs.  

Algorithm 2 employs three-dimensional temporal 
modeling through time-series analysis techniques to capture 
engagement pattern evolution across academic semesters. 
The computational framework utilizes automated change 
pattern recognition for each SDT dimension, incorporating 
individual baseline adjustments and environmental factor 
integration to detect declining motivation indicators 
preceding academic performance impacts.  

Automatic identification of engagement pattern changes 
enables proactive intervention through early warning models 
that predict motivational decline before academic 
performance impact occurs. Subject-specific alert thresholds, 
dynamic adaptation to individual differences, baseline, and 
learning trajectory sensitivity for intervention guidance, and 
absence of “false alarms”, so that the instructional staff are not 
overloaded and students with “typical” contextually 
performance levels do not become nervous. 

 
 
 

 
 
 

2.3 Five-dimensional writing performance prediction 
model 
The Jacobs framework-based multi-faceted prediction 

approach accounts for each dimension of writing competence 
through targeted sub-models that model the specificities of 
development and shaping of the proficiency related to 
different writing aspects. There’s No Dark Art to it. Content 
dimension prediction includes forecasting thought depth and 
logical coherence by applying natural language processing 
methods to the structural complexity of the argument, 
integration of evidence patterns, critical thinking signals, and 
conceptual development progression that can be traced using 
snippets across several writing samples. The organization 
dimension modeling targets predicting structural integrity 
and coherence at the discourse level and leverages discourse 
analysis algorithms that assess how well a paragraph 
transitions to another, how well a thesis is developed, how 
well a conclusion is synthesized, and the overall architectural 
soundness of the article.  

Language Use dimension forecasting emphasizes 
grammatical accuracy and syntactic complexity prediction 
through computational linguistics approaches that assess 
sentence structure variety, clause combination 
sophistication, error pattern identification, and grammatical 
development trajectories that indicate language proficiency 
advancement. Vocabulary dimension analysis concentrates 
on lexical richness and accuracy prediction through corpus-
based approaches that evaluate word choice appropriateness, 
semantic precision, vocabulary range expansion, and register 
consistency maintenance across different writing contexts 
and task requirements.  

 
 
 

Input Layer :POA leaching Activity Data Collection

Real-time Classroom
Interaction Logs

Digital Platform
Engagement Metrics

Writing Portfolio
Development 
Trajectories

Peer Collaboration
Network Data

Teacher Observation
Records

Processing Layer: Education-Theory-Guided Feature Engineering

Data Preprocessing& 
Normalization

Educational Feature
Engineering

Pattern Recognition& 
Temporal Alignment

Domain Knowledge
Integration

Prediction Layer:Five-Dimensional Writing Ability Forecasting

Content Prediction Organization 
Prediction

Language Use 
Prediction Vocabulary Prediction Mechanics Prediction

Application Layer: Teacher Decision-Support & Interpretable Recommendations

Co
nt

in
uo

us
 M

od
el

 R
ef

in
em

en
t

Data
Collection

Feature
Processing

Performance
Prediction

Educational
Application



Yu Li et al. /Future Technology                                                                                        November 2025| Volume 04 | Issue 04 | Pages 100-116 

106 
 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Mechanics dimension assessment targets spelling and 

punctuation norm compliance prediction through error 
detection algorithms that identify persistent mistake 
patterns, improvement trajectory analysis, and mechanical 
skill development indicators that correlate with overall 
writing proficiency advancement, as summarized in Table 2. 

Cross-dimensional interaction modeling reflects the 
reality that writing proficiencies manifest intricate 
interdependencies in which the enrichment of vocabulary 
profiles increases one's capacity to develop textual content 
and the fostering of organizational skills to help write more 
sophisticated grammar, with synergies generating additional 
returns that push overall writing performance beyond 
additive component contributions.  

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
These interaction patterns are revealed by advanced 

correlation analysis in order to guide systematic intervention 
strategies that exploit competency interrelationships for 
optimal learning effects. An ensemble learning prediction 
optimization strategy is implemented based on ensemble 
learning prediction optimization that integrates a variety of 
algorithm approaches so that the prediction accuracy can be 
improved without losing interpretability for educational 
applications. The ensemble learning approach combines 
gradient boosting for sequential learning pattern capture, 
random forest for high-dimensional educational variable 
handling, neural networks for non-linear cognitive 
relationship modeling, and support vector machines for 
robust classification boundary establishment. This 

Algorithm 1: Feature adaptive selection mechanism 
Input: Educational feature set F, POA phase indicators, SDT engagement data, Multi-theoretical weights, Learning improvement trajectories 
Output: Dynamically selected features prioritizing sustained learning improvement 
Initialize theoretical framework weights for Input Hypothesis, Social Constructivism, Cognitive Process, Sociocultural theories 
Define educationally meaningful variable categories: 
Student engagement indicators (autonomy, competence, relatedness) 

Collaborative learning participation rates 
Progress trajectory patterns 

for each POA phase (Motivating, Enabling, Assessing) do 
Compute phase-specific educational significance for each variable category 
Apply dynamic weight adjustment based on sustained learning improvement correlation 

end for 
Multi-theoretical fusion: 

Analyze complex interactions between learning mechanisms 
Balance feature representations across theoretical frameworks 
Prioritize sustained learning improvement over predictive accuracy 

for each feature do 
if feature belongs to educationally meaningful categories and 

correlates with sustained learning improvement then 
Select with POA-principle-based importance weight 

end if 
end for 
return Educational-priority features ensuring pedagogical meaning and learning improvement focus 

 

Algorithm 2: SDT-based engagement dynamic modeling 
Input: SDT three-dimensional engagement data, Individual baselines, Environmental factors 
Output: Dynamic engagement predictions, Early warning indicators, Personalized thresholds 
SDT Three-Dimensional Temporal Modeling: 
for each time period t do 

Collect autonomy, competence, relatedness measurements 
Account for individual differences in motivation development 
Account for environmental factors affecting psychological need satisfaction 

end for 
Engagement Pattern Evolution Capture: 
Apply time-series analysis to capture engagement pattern evolution across academic semester 
Identify critical transition points where intervention strategies can maximize effectiveness 
Automatic Change Pattern Recognition: 
for each SDT dimension do 

Detect declining motivation indicators before academic performance impact 
Enable proactive intervention through early warning systems 

end for 
Personalized Warning Threshold Adjustment: 
Account for individual baseline differences and learning trajectory variations 
Minimize false positive alerts while maintaining appropriate sensitivity levels 
return Dynamic engagement predictions with personalized intervention recommendations 

 



Yu Li et al. /Future Technology                                                                                        November 2025| Volume 04 | Issue 04 | Pages 100-116 

107 
 

combination leverages each algorithm's strengths: gradient 
boosting handles temporal dependencies in writing 
development, random forest manages missing data common 
in educational contexts, neural networks model complex 
motivation-performance relationships, while support vector 
machines provide stable decision boundaries across diverse 
student populations. Methods for predicting uncertainty 
estimation and visualization can be employed to generate a 
confidence interval for each prediction output. Stakeholders 
can then use this interval to inform evidence-based decisions 
regarding the reliability of the prediction and the uncertainty 
bounds, thereby guiding the timing and extent of 
intervention.  

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Validation models, as utilized by the approach, embed 
domain knowledge from the education sector in the form of 
expert teacher evaluations of the accuracy of the predictions 
and the relevance of the suggestions, thus ensuring that the 
algorithm outputs are consistent with experienced 
practitioner views on student needs or appropriate 
instructional responses. Longitudinal validation studies 
monitor the accuracy of the prediction over the long term to 
test whether the model remains stable and reliable in 
different educational contexts and with different types of 
students. Interindividual fit between educational applications 
and model interpretability optimization ensures that 
predictions of underperforming work are not merely output 
values but actionable outcomes that provide advice for 
improving the instruction process.  

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Table 2. Five-dimensional prediction model specifications 

Writing Dimension Technical Approach Key Assessment Indicators Prediction Focus Specialized Sub-
Model 

Content Natural Language 
Processing 

Argument structure complexity, 
Evidence integration patterns, Critical 

thinking indicators, Conceptual 
development progression 

Thought depth and 
logical coherence 

forecasting 

Multi-sample 
content analysis 

Organization Discourse Analysis 
Algorithms 

Paragraph transition effectiveness, 
Thesis development consistency, 

Conclusion synthesis quality, Overall 
architectural soundness 

Structural integrity 
and coherence 

prediction 

Compositional 
structure modeling 

Language Use Computational 
Linguistics 

Sentence structure variety, Clause 
combination sophistication, Error 

pattern identification, Grammatical 
development trajectories 

Grammatical 
accuracy and 

syntactic complexity 
prediction 

Language 
proficiency 

advancement 
tracking 

Vocabulary Corpus-Based 
Approaches 

Word choice appropriateness, Semantic 
precision, Vocabulary range expansion, 

Register consistency maintenance 

Lexical richness and 
accuracy prediction 

Cross-context 
vocabulary analysis 

Mechanics Error Detection 
Algorithms 

Persistent mistake patterns, 
Improvement trajectory analysis, 

Spelling compliance indicators, 
Punctuation norm adherence 

Spelling and 
punctuation norm 

compliance 
prediction 

Mechanical skill 
development 

modeling 

 

Table 3. Model interpretability components 

Interpretability 
Component 

Educational Purpose Output Format Teacher Decision 
Support 

Practical Utility 

Feature Importance 
Analysis 

Highlight behavioral 
indicators and learning 
patterns contributing to 

predictions 

Ranked importance scores 
with educational context 

Focus intervention 
efforts on high-impact 

areas 

Identify key factors 
affecting student 

performance 

Educational Rationale 
Generation 

Explain algorithmic 
recommendations using 
pedagogical principles 

Natural language 
explanations linked to POA 

theory 

Understand the 
reasoning behind 

intervention 
suggestions 

Bridge technical 
predictions with 
teaching practice 

Actionable Insight 
Extraction 

Transform statistical outputs 
into instructional 

improvement strategies 

Specific teaching 
recommendations with 
implementation steps 

Support instructional 
modifications and 
student grouping 

decisions 

Direct classroom 
application guidance 

Prediction Confidence 
Visualization 

Display uncertainty levels 
and reliability assessments 

Confidence intervals with 
educational interpretation 

Inform intervention 
timing and intensity 

decisions 

Enable evidence-
based teaching 

adjustments 
Learning Pattern 

Recognition 
Identify recurring behavioral 
and cognitive patterns across 

students 

Visual pattern summaries 
with trend analysis 

Recognize effective 
teaching strategies and 

problematic areas 

Systematic teaching 
approach 

optimization 
Intervention Strategy 

Mapping 
Connect predictions to 

personalized educational 
interventions 

Customized intervention 
recommendations with 

success probabilities 

Provide individualized 
student support 

strategies 

Practical intervention 
implementation 

framework 
 

 

 

 



Yu Li et al. /Future Technology                                                                                        November 2025| Volume 04 | Issue 04 | Pages 100-116 

108 
 

Feature importance analysis features Reflect on which 
individual behaviors and learning dynamics are most 
influential to predicting performance, where intervention 
activities should be prioritized, while understanding the 
pedagogical reasoning underpinning algorithmic suggestions 
(see Table 3). The POA-MLSP framework establishes a 
comprehensive four-layer architecture that integrates POA 
theoretical principles with advanced learning analytics 
through education-oriented algorithmic designs. By 
implementing the Feature Adaptive Selection Mechanism and 
SDT-based engagement modeling, this methodological 
foundation enables empirical validation within authentic 
CET-4 writing instruction contexts for enhanced educational 
outcomes. 

3. Results 
3.1 Data collection and preprocessing 

The POA-MLSP framework validation was conducted at S 
Normal University, involving 124 students distributed across 
POA experimental classes and traditional instruction control 
groups during a complete 16-week academic semester. The 
research design incorporated authentic educational 
environments while maintaining rigorous experimental 
controls necessary for robust statistical analysis, as outlined 
in Table 4. This Table shows comprehensive experimental 
specifications, including participant demographics, class 
distribution, temporal framework, and control variables that 
ensured ecological validity while enabling meaningful 
statistical comparisons between POA-enhanced and 
traditional instruction approaches. Multi-source educational 
data collection protocols captured behavioral, cognitive, and 
social dimensions of language learning within POA 
instructional contexts while maintaining manageable 
collection burdens for educational stakeholders, as 
systematized in Table 5. This Table displays a systematic data 
collection that includes writing performance assessments 
through Jacobs’ five-dimensional rubric, student engagement 
measurements by means of UWES-S scales, classroom 
observation, and POA instructional process documentation of 
teacher-student interactions, peer collaboration network, and 
resource utilization through digital learning platforms and 
structured observation protocols.  

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Preprocessing of the data included the use of advanced 

normalization techniques tailored to educational settings, 

taking care of integration of multi-scale measurement, time-
alignment effects, and individual-response baseline 
variations while maintaining pedagogical interpretability 
necessary for educational stakeholders to interpret, as 
described in Table 6. This Table outlines comprehensive 
quality control measures, including outlier detection, 
distinguishing educational phenomena from collection 
errors, normalization techniques preserving educationally 
meaningful variance, temporal alignment algorithms 
synchronizing multi-instrument data, and feature 
engineering processes that transformed raw educational data 
into analytically tractable variables while maintaining 
theoretical alignment with POA principles and SDT 
frameworks. 

3.2 Experimental design and model training 
The experimental framework implemented temporally-

aware methodological approaches, respecting chronological 
learning sequences while ensuring robust validation 
procedures. POA-MLSP model training procedures integrated 
hyperparameter optimization combining grid search 
exploration with Bayesian optimization techniques, as 
presented in Table 7. This Table demonstrates systematic 
model training, achieving 86.4% consistency across multiple 
random initializations, with an optimal hyperparameter 
configuration identified through 47 iterations of combined 
grid search and Bayesian optimization. Cross-validation 
yielded an average score of 0.742 ± 0.018 while maintaining 
prediction variance below 6.1% across all dimensions, 
ensuring acceptable model stability. Training convergence 
was achieved in 82.7% of runs with early stopping at epoch 
73, while 84.3% alignment with pedagogical expectations 
validates educational interpretability integration throughout 
the training process. 

3.3 Core framework component validation and 
educational application results 
The validation of the POA-MLSP model indicates that the 

system has achieved systematic effectiveness in CET-4 
writing teaching and learning by virtue of its integrated four-
layer architecture, which seamlessly fuses advanced learning 
analytics with established pedagogy.  

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Table 4. Experimental design and participant characteristics 

Experimental Component Specification Details 
Research Setting S Normal University CET-4 Writing Program Authentic classroom environments 
Study Duration 16-week academic semester Complete instructional cycle 

Total Participants 124 undergraduate students Representative CET-4 learner population 
Group Distribution POA Experimental: 62 students 

Traditional Control: 62 students 
Balanced group allocation 

Age Range 18-22 years (Mean: 19.8, SD: 1.2) Typical undergraduate demographic 
Gender Distribution Female: 68 (54.8%) 

Male: 56 (45.2%) 
Representative gender balance 

English Proficiency Level Intermediate (CEFR B1-B2 equivalent) Pre-CET-4 preparation level 
Academic Majors Engineering: 35% 

Liberal Arts: 28% 
Business: 22% 
Sciences: 15% 

Diverse academic backgrounds 

Prior CET-4 Experience First attempt: 89 students (71.8%) 
Repeat attempt: 35 students (28.2%) 

Mixed experience levels 

Class Schedule 4 hours/week writing instruction Consistent instructional time 
Instructor Qualifications Master's/PhD in Applied Linguistics 

5+ years CET-4 teaching experience 
Standardized expertise level 

Control Variables Same curriculum materials 
Identical assessment rubrics 

Matched class times 

Methodological rigor 

 



Yu Li et al. /Future Technology                                                                                        November 2025| Volume 04 | Issue 04 | Pages 100-116 

109 
 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Validation of Algorithm 1: The validation indicated a 

significant outperformance of traditional methods in finding 
educationally meaningful variables, such as student 
engagement indicators and collaborative learning patterns, 
which are correlated with learning improvement. The multi-
theoretical integration analysis found that the synthesis of the 
four theoretical perspectives of the Input Hypothesis, Social 
Constructivism, Cognitive Process Theory, and Sociocultural 
Theory provided a better explanatory power than the 
individual theoretical bases, which were able to support more 
authentic educational judgements.  

 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
The established framework had established good 

predictive power for each of the five dimensions of the writing 
assessment, but also maintained educational interpretability 
for practical classroom use, as shown in Table 8. This Table 
demonstrates POA-MLSP framework achieving overall R² = 
0.75 with 12.0% improvement over baseline approaches, 
while Content prediction reached highest accuracy (R² = 0.78, 
+12.3% gain) and Organization forecasting showed 
substantial enhancement (R² = 0.76, +11.7% gain), 
confirming framework effectiveness across all five writing 
competency dimensions with confidence intervals indicating 
robust statistical reliability. 

Table 5. Comprehensive educational data collection specifications 

Data Category Collection Method Frequency Measurement 
Instruments 

Data Types 

Writing Performance Jacobs Five-
Dimensional 
Assessment 

Pre/Mid/Post-test 
(Week 1, 8, 16) 

Jacobs Writing Rubric Content, Organization, 
Language Use, Vocabulary, 

Mechanics scores 
Writing Performance Process-Oriented 

Evaluation 
Weekly classroom 

exercises 
Structured 

assessment forms 
Improvement trajectories, 

error patterns 
Writing Performance Peer and Self-

Assessment 
Bi-weekly reflection 

sessions 
Standardized 

evaluation criteria 
Collaborative assessment 

data, metacognitive 
reflections 

Student Engagement UWES-S Scale 
Administration 

Bi-weekly surveys (8 
measurement points) 

Utrecht Work 
Engagement Scale-

Student 

Vigor, dedication, absorption 
scores 

Student Engagement Systematic Classroom 
Observation 

Daily during class 
sessions 

POA-specific 
observation protocol 

SDT three-dimensional 
behavioral indicators 

Student Engagement Learning Journal 
Analysis 

Weekly reflection 
entries 

Structured journal 
prompts 

Self-reported motivation, 
attitude changes 

POA Teaching Process Three-Phase Activity 
Documentation 

Continuous throughout 
semester 

Digital activity logging 
system 

Motivating, Enabling, 
Assessing phase records 

POA Teaching Process Interaction Behavior 
Logging 

Real-time during 
instruction 

Video recording and 
coding 

Teacher-student, peer 
collaboration patterns 

POA Teaching Process Resource Utilization 
Tracking 

Continuous digital 
monitoring 

Learning management 
system logs 

Platform engagement, task 
completion rates 

 

Table 6. Data processing and validation procedures 

Processing Stage Technique Purpose Implementation Quality Control 
Measures 

Data Cleaning Missing Value 
Imputation 

Handle incomplete data 
while preserving 

educational meaning 

Educational domain 
knowledge-guided 

interpolation 

Expert teacher validation 
of imputed patterns 

Data Cleaning Outlier Detection Distinguish genuine 
educational phenomena 

from collection errors 

Statistical threshold 
combined with pedagogical 

judgment 

Manual review of flagged 
cases by experienced 

instructors 
Normalization Multi-Scale 

Integration 
Enable cross-student 
comparisons across 

different instruments 

Z-score standardization with 
educational context 

adjustment 

Variance preservation 
validation for pedagogical 

meaningfulness 
Normalization Individual Baseline 

Adjustment 
Account for diverse 

students starting points 
and backgrounds 

Relative improvement 
calculation from personal 

baselines 

Baseline stability 
verification across 

measurement periods 
Temporal Alignment Cross-Instrument 

Synchronization 
Align data collected 

through different 
methods and timeframes 

Timestamp-based alignment 
with learning rhythm 

accommodation 

Temporal consistency 
validation across data 

sources 
Temporal Alignment Learning Pace 

Accommodation 
Respect natural 

variations in student 
progression patterns 

Adaptive time window 
adjustment for data 

aggregation 

Pedagogical validity check 
for temporal groupings 

Feature Engineering Educational 
Variable 

Construction 

Transform raw data into 
pedagogically meaningful 

predictors 

POA and SDT theory-guided 
feature derivation 

Theoretical alignment 
verification with domain 

experts 
Feature Engineering Interaction Feature 

Generation 
Capture emergent 

educational phenomena 
from data intersections 

Multi-dimensional 
correlation analysis with 

educational interpretation 

Expert validation of 
derived educational 

constructs 
Quality Validation Pedagogical 

Interpretability 
Check 

Ensure all processed 
variables maintain 

educational meaning 

Regular stakeholder review 
sessions with teachers 

Actionable insight 
generation capability 

assessment 
 

 

 



Yu Li et al. /Future Technology                                                                                        November 2025| Volume 04 | Issue 04 | Pages 100-116 

110 
 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Table 7. Model training configuration and performance metrics 

Training Component Configuration Specification Performance Result 
Hyperparameter 

Optimization 
Grid Search + Bayesian 

Optimization 
Learning rate: 0.001-0.01, Batch 
size: 16-64, Hidden layers: 2-5 

Optimal configuration identified in 
47 iterations 

Convergence Verification Multiple Random 
Initializations 

10 different random seeds for 
robustness testing 

86.4% consistency across 
initializations 

Training-Validation Split Temporal Sequence 
Preservation 

70% training, 20% validation, 10% 
testing 

Maintained chronological learning 
order 

Cross-Validation Strategy K-Fold for Educational 
Time-Series 

K=5 with temporal dependency 
preservation 

Average CV score: 0.742 ± 0.018 

Overfitting Prevention Early Stopping + 
Regularization 

Patience=15 epochs, L2 
regularization λ=0.01 

Training stopped at epoch 73, 
optimal validation loss 

Convergence Criteria Loss Stabilization 
Threshold 

Validation loss improvement < 0.001 
for 10 epochs 

Achieved stable Achieved stable 
convergence in 82.7% of runs 

Training Duration Educational Context 
Optimization 

Average: 4.6 hours per complete 
training cycle 

Suitable for educational 
implementation timelines 

Model Stability Performance Variance 
Analysis 

Standard deviation across training 
runs 

Prediction variance < 6.1% across all 
dimensions 

Educational Interpretability Feature Importance 
Validation 

Expert teacher evaluation of 
algorithmic outputs 

84.3% alignment with pedagogical 
expectations 

Training Performance Validation Loss 
Monitoring 

Final validation loss: 0.287, Training 
epochs: 73 

Optimal training convergence 
achieved 

 

Table 8. Five-dimensional writing performance prediction results 

Writing Dimension POA-MLSP 
Performance 

Baseline 
Performance 

Improvement 
Rate 

Confidence 
Interval 

Educational Interpretation 

Content R² = 0.78, RMSE = 
0.29 

R² = 0.65, RMSE = 
0.41 

+12.3% accuracy 
gain 

R² = 0.78 ± 
0.032 

Strong thought development 
and logical reasoning 
forecasting 

Organization R² = 0.76, RMSE = 
0.31 

R² = 0.64, RMSE = 
0.43 

+11.7% accuracy 
gain 

R² = 0.76 ± 
0.028 

Effective structural 
coherence and discourse 
pattern modeling 

Language Use R² = 0.73, RMSE = 
0.33 

R² = 0.63, RMSE = 
0.45 

+9.8% accuracy 
gain 

R² = 0.73 ± 
0.035 

Solid grammatical accuracy 
and syntactic complexity 
prediction 

Vocabulary R² = 0.74, RMSE = 
0.32 

R² = 0.64, RMSE = 
0.44 

+10.2% accuracy 
gain 

R² = 0.74 ± 
0.030 

Robust lexical richness and 
word choice appropriateness 
forecasting 

Mechanics R² = 0.69, RMSE = 
0.36 

R² = 0.61, RMSE = 
0.48 

+7.4% accuracy 
gain 

R² = 0.69 ± 
0.041 

Meaningful spelling and 
punctuation compliance 
prediction 

Overall Framework R² = 0.75, RMSE = 
0.31, MAE = 0.24 

R² = 0.63, RMSE = 
0.44, MAE = 0.35 

+12.0% overall 
improvement 

R² = 0.75 ± 
0.026 

Comprehensive multi-
dimensional writing ability 
forecasting 

 

Table 9. Ablation study performance analysis 

Component Configuration Overall 
R² 

Content 
R² 

Organization 
R² 

Language Use 
R² 

Vocabulary 
R² 

Mechanics 
R² 

Full POA-MLSP 0.75 0.78 0.76 0.73 0.74 0.69 

Without Feature Adaptive Selection 0.69 0.72 0.71 0.67 0.69 0.65 

Without SDT Modeling 0.72 0.75 0.73 0.70 0.71 0.67 

Without Multi-theoretical Integration 0.67 0.70 0.68 0.65 0.67 0.63 

 

 



Yu Li et al. /Future Technology                                                                                        November 2025| Volume 04 | Issue 04 | Pages 100-116 

111 
 

To validate the individual contribution of each 
framework component, comprehensive ablation studies were 
conducted by systematically removing key algorithmic 
elements and measuring resulting performance degradation 
across all writing dimensions, as presented in Table 9. Table 
9 reveals differential contributions of POA-MLSP components 
across writing dimensions. Feature Adaptive Selection 
Mechanism removal yields 6-point R² reduction (0.75 to 
0.69), with Language Use and Mechanics dimensions 
demonstrating greater sensitivity to educationally-informed 
variable selection. SDT-based engagement modeling 
elimination produces a 3-point performance decrease (0.75 
to 0.72), indicating a moderate but consistent impact across 
all assessment dimensions. Multi-theoretical integration 
removal generates the most substantial degradation (8-point 
reduction to 0.67), particularly affecting Content and 
Organization predictions, thereby confirming the critical role 
of theoretical convergence in complex writing assessment 
contexts. 

Algorithm 2 validation revealed solid capability for early 
identification of at-risk students through three-dimensional 
engagement pattern recognition. The SDT-based modeling 
demonstrated consistent effectiveness across all three 
psychological dimensions, with autonomy modeling 
successfully predicting self-directed learning behaviors, 
competence analysis effectively forecasting efficacy 
development, and relatedness assessment reliably identifying 
social engagement patterns, enabling comprehensive student 
profile construction that supports personalized intervention 
strategies. Building upon this three-dimensional engagement 
analysis, the framework's pattern recognition capabilities 
enabled systematic identification of distinct student 
motivational profiles. Student engagement pattern 
recognition identified five distinct motivational profiles 
among participants, providing insights into psychological 
need satisfaction diversity characterizing CET-4 writing 
learners, as illustrated in Figure 3. 

 

 
 

Figure 3 demonstrates engagement pattern diversity 
where High Sustained (28%) and Autonomy-Oriented (22%) 
profiles represent the largest student groups, while Figure 
3(a) shows balanced distribution across five motivational 
types and Figure 3(b) reveals distinct SDT dimensional 
characteristics with Autonomy-Oriented students achieving 
highest autonomy scores (9.1) and Relatedness-Dependent 
students displaying strongest social engagement patterns 
(9.3), confirming theoretical framework validity. To evaluate 
the practical effectiveness of POA-MLSP framework 
predictions for improving educational outcomes and validate 
the utility of data-informed pedagogical decision-making 
capabilities, a comprehensive intervention effectiveness 
analysis was conducted across multiple performance 
dimensions, including early warning system accuracy, 
teaching adjustment outcomes, and overall implementation 
feasibility, as detailed in Table 10. As shown in Table 10, the 
early warning system's performance accuracy (78.4% ± 
2.1%) of students at risk of engagement decline 3-4 weeks 
before score performance decline emerges through 
traditional assessment, which is a 9.5 percentage point 
improvement on the baseline method. Individualized 
threshold-adjustment algorithms achieved 24.3% lower false 
positive alerts vs static alarm systems, with target 
interventions averaged 82.1% ± 2.4%. Teaching intervention 
based on framework predictions achieved an average 
improvement of 8.7% ± 1.9% in performance over control 
groups; 79.8% ± 2.9% satisfaction of teachers with 
recommendations, indicating deployment of the results in the 
classroom and acceptance of the educational stakeholder. 
Effect size analysis demonstrates substantial educational 
impact beyond statistical significance, with Cohen's d values 
indicating large effects for writing performance improvement 
(d = 0.82), student engagement enhancement (d = 0.89), and 
teacher instructional effectiveness (d = 0.74).  

 
 
 

 
 

(a) Engagement Profile Distribution

High Sustained

Autonomy-Oriented

Relatedness-Dependent

Competence-Anxious Low-Risk

28%

22%

19%

18%

13%

Engagement Profile Types

0

5

10

15

20

25

30

35

Pe
rc

en
ta

ge
 o

f S
tu

de
nt

s 
(%

)

(b) SDT Three-Dimensional Characteristics

High Sustained

Autonomy-Oriented

Relatedness-Dependent

Competence-Anxious Low-Risk

8.2

8.5

8.1

9.1

7.9

5.8

5.4

7.2

9.3

6.3

4.1

6.8

3.8

3.5

4.2

Engagement Profile Types

0

1

2

3

4

5

6

7

8

9

10

11

SD
T 

D
im

en
si

on
 S

co
re

 (0
-1

0)

Autonomy

Competence

Relatedness

Figure 3. Student engagement profile distribution and characteristics 



Yu Li et al. /Future Technology                                                                                        November 2025| Volume 04 | Issue 04 | Pages 100-116 

112 
 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Sustained academic outcomes reveal framework-

supported students achieving 12.3% higher final CET-4 
writing scores while maintaining engagement levels 18.7% 
above baseline measurements throughout the academic year. 
Expanding on these intervention efficacy results, it also 
became critical to investigate the temporal changes over a 
complete academic semester of student learning trajectories 
to define how POA's three-phase structured approach is 
affecting developmental trends. This type of longitudinal 
evidence is essential for us to learn whether the theory-based 
assumptions of POA instruction are borne out in evidence of 
instructional-phase-specific acceleration of learning, as 
systematically collated in Table 11. This Table reveals 
systematic progression patterns where motivating phase 
interventions generated average engagement increases of 
23.8% within the first month, enabling phase activities 
produced sustained skill development with 31.2% 
improvement rates during mid-semester periods, and 
assessment integration phase created consolidation effects 
yielding 18.4% additional gains during final semester weeks, 
validating structured phase progression effectiveness 
compared to undifferentiated instructional approaches. 

3.4 Comparative analysis and advanced algorithm 
performance validation 
Following the comprehensive validation of POA-MLSP 

framework components and educational effectiveness 
demonstration, it became crucial to establish the framework's 
technical superiority through systematic comparison with 
established machine learning approaches commonly applied 
to educational prediction tasks. This comparative analysis 
validates that educational theory-guided design principles 
genuinely enhance prediction performance beyond purely 
statistical approaches, while demonstrating the practical 
advantages of the novel algorithmic components in multi-
dimensional writing performance prediction contexts. To 
comprehensively evaluate these technical advantages and 
validate the core research hypothesis that educational 
domain knowledge integration improves prediction accuracy, 
a systematic performance comparison was conducted against 
multiple baseline models, as systematically visualized in 
Figure 4. This Figure demonstrates the POA-MLSP 
framework's technical superiority through a comprehensive 
algorithmic comparison.  

 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Figure 4(a) shows overall performance where POA-

MLSP achieved R² = 0.75 ± 0.026, outperforming linear 
regression (R² = 0.58), random forest (R² = 0.66), and support 
vector machines (R² = 0.63) by 17-29%, validating 
educational theory integration benefits. Figure 4(b) reveals 
consistent advantages across all five writing dimensions, with 
Content (R² = 0.78) and Organization (R² = 0.76) showing the 
highest prediction accuracy. Figure 4(c) illustrates Feature 
Adaptive Selection Mechanism contributions, delivering 
notable improvements in Content prediction (+9.2%), 
Organization forecasting (+7.4%), and Language Use 
modeling (+6.1%), confirming pedagogically-informed 
algorithmic design effectiveness over traditional attention 
mechanisms. While technical performance validation 
demonstrates algorithmic superiority, the ultimate success of 
educational technology innovation depends on acceptance 
and practical utility among educational stakeholders who 
must integrate these tools into authentic teaching contexts. 
To assess whether the framework's technical capabilities 
translate into meaningful educational support that enhances 
instructional practice, a comprehensive stakeholder 
evaluation was conducted to validate practical 
implementation feasibility and educational value perception, 
as systematically documented in Figure 5. This Figure 
demonstrates comprehensive stakeholder validation through 
three key assessment dimensions. Figure 5(a) reveals strong 
teacher satisfaction with decision-making insights (79.8% ± 
2.9%), student grouping effectiveness (76.4% ± 3.1%), and 
continued utilization willingness (81.5% ± 2.4%), confirming 
educator acceptance. Figure 5(b) validates practical 
implementation feasibility, requiring minimal initial training 
(4.8 ± 1.2 hours) and weekly operation time (27.5 minutes), 
supporting widespread deployment viability. Figure 5(c) 
confirms framework utility across educational 
interpretability (77.6% ± 3.1%), practical utility (79.2% ± 
2.7%), and prediction reliability (82.4%), demonstrating 
successful translation of sophisticated algorithms into 
meaningful educational support tools. The comparative 
analysis demonstrates the POA-MLSP framework's technical 
superiority, achieving 17-29% performance improvements 
over traditional approaches while maintaining computational 
efficiency through educational domain knowledge 
integration.  

 
 

Table 10. Educational intervention effectiveness validation 

Intervention 
Component 

Performance Metric Framework 
Result 

Baseline/Control 
Result 

Improvement Statistical 
Significance 

Early Warning System At-risk Student 
Identification Accuracy 

78.4% ± 2.1% 68.9% ± 2.8% +9.5 percentage 
points 

p < 0.05 

Early Warning System Advance Warning Time 3-4 weeks 1-2 weeks +2 weeks average p < 0.05 
Alert Optimization False Positive Reduction 

Rate 
24.3% 

reduction 
Static threshold 

baseline 
-24.3% false alerts p < 0.05 

Teaching Adjustments Student Performance 
Improvement 

8.7% ± 1.9% 
gain 

Control group 
baseline 

+8.7% relative 
improvement 

p < 0.05 

Intervention Timing Proactive vs Reactive 
Success Rate 

71.2% ± 3.1% 58.6% ± 3.4% +12.6 percentage 
points 

p < 0.05 

Personalized Thresholds Intervention Targeting 
Precision 

82.1% ± 2.4% 74.3% ± 3.0% +7.8 percentage 
points 

p < 0.05 

Overall Framework Teacher Satisfaction with 
Recommendations 

79.8% ± 2.9% 61.2% ± 3.7% +18.6 percentage 
points 

p < 0.01 

Implementation 
Feasibility 

Successful Classroom 
Integration Rate 

83.7% ± 2.6% N/A High adoption 
success 

p < 0.01 

 



Yu Li et al. /Future Technology                                                                                        November 2025| Volume 04 | Issue 04 | Pages 100-116 

113 
 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Table 11. Temporal learning trajectory analysis across a 16-week period 

Time Period POA Phase Learning Indicator Framework 
Group 

Control 
Group 

Improvement 
Rate 

Statistical 
Significance 

Weeks 1-4 Motivating Student Engagement 
Level 

73.8% ± 2.4% 58.6% ± 
3.1% 

+23.8% increase p < 0.01 

Weeks 1-4 Motivating Learning Motivation 
Score 

7.2 ± 0.8 5.8 ± 0.9 +24.1% increase p < 0.05 

Weeks 5-8 Enabling 
(Early) 

Writing Skill 
Development 

6.8 ± 0.7 5.9 ± 0.8 +15.3% 
improvement 

p < 0.05 

Weeks 9-12 Enabling 
(Peak) 

Writing Performance 
Gains 

8.1 ± 0.6 6.2 ± 0.9 +31.2% 
improvement 

p < 0.01 

Weeks 9-12 Enabling 
(Peak) 

Collaborative 
Learning 

Participation 

84.7% ± 2.1% 64.5% ± 
3.4% 

+31.3% increase p < 0.01 

Weeks 13-16 Assessment 
Integration 

Learning 
Consolidation Effects 

7.9 ± 0.5 6.7 ± 0.7 +18.4% additional 
gains 

p < 0.05 

Weeks 13-16 Assessment 
Integration 

Self-Regulation 
Development 

78.2% ± 2.6% 66.0% ± 
3.2% 

+18.5% 
improvement 

p < 0.05 

Overall Semester Complete POA 
Cycle 

Cumulative Learning 
Progress 

82.4% ± 1.9% 65.1% ± 
2.8% 

+26.6% total 
improvement 

p < 0.001 

 

 
Figure 4. Algorithm performance comparison across writing dimensions 

 

Figure 5. Teacher satisfaction and framework usability assessment 

 

(a) Overall Algorithm Performance Comparison

Linear

Regression

Random

Forest

Support

Vector

Machine

POA-MLSP

Framework

0.580±0.031

0.660±0.025

0.630±0.029

0.750±0.026

Machine Learning Algorithms

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

R
² P

er
fo

rm
an

ce
 S

co
re

(b) Five-Dimensional Writing Performance Prediction

Con
ten

t

Orga
niz

ati
on

La
ng

ua
ge

Use
Voc

ab
ula

ry

Mec
ha

nic
s

0.62

0.78

0.64

0.76

0.63

0.73

0.64

0.74

0.61

0.69

Writing Dimensions (Jacobs Framework)

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

R
² P

re
di

ct
io

n 
A

cc
ur

ac
y

Baseline Average

POA-MLSP Framework

(c) Feature Adaptive Selection Mechanism Contribution

Con
ten

t

Orga
niz

ati
on La

ng
ua

ge

Use

Voc
ab

ula
ry

Mec
ha

nic
s

+9.2%

+7.4%

+6.1%

+5.8%

+4.9%

Writing Dimensions

0

2

4

6

8

10

12

Pe
rf

or
m

an
ce

 Im
pr

ov
em

en
t (

%
)

(a) Teacher Satisfaction Assessment

Decision-Making

Insights

Student

Grouping

Effectiveness

Actionable

Recommendations

Continued

Utilization

79.8%±2.9
76.4%±3.1

73.2%±2.7

81.5%±2.4

Satisfaction Assessment Dimensions

0

20

40

60

80

Te
ac

he
r S

at
is

fa
ct

io
n 

Sc
or

e 
(%

)

(b) Implementation Feasibility Analysis

Initial Training

Time

(hours)

Weekly

Operation

Time (minutes)

Technical

Support

Requests

(per month)

System Integration

Success Rate (%)

4.8±1.2

27.5±2.5

2.3±0.8

89.4%±2.1

Implementation Feasibility Factors

0

20

40

60

80

100

Im
pl

em
en

ta
tio

n 
M

et
ric

s

(c) Framework Practical Utility Assessment

Educational Interpretability Practical Utility Prediction Reliability Classroom Applicability

77.6%±3.1 79.2%±2.7
82.4%±2.3

78.9%±2.8

Practical Utility Assessment Dimensions

0

20

40

60

80

100

Fr
am

ew
or

k 
U

til
ity

 S
co

re
 (%

)



Yu Li et al. /Future Technology                                                                                        November 2025| Volume 04 | Issue 04 | Pages 100-116 

114 
 

Stakeholder evaluation reveals strong teacher 
acceptance with 79.8% ± 2.9% reporting valuable decision-
making insights and 81.5% ± 2.4% expressing continued 
utilization willingness. Implementation feasibility validation 
shows minimal training requirements (4.8 ± 1.2 hours) and 
practical utility ratings of 79.2% ± 2.7%, confirming 
successful translation of sophisticated algorithms into 
deployable educational tools ready for authentic CET-4 
writing instruction environments. 

4. Discussion 
The POA-MLSP model validity test indicates a significant 

move forward in educational predictions from the existing 
approaches, especially against the background of the recent 
trends and advancements in language learning analytics and 
AI. The framework’s ability to predict R² = 0.75 across five 
dimensions of writing is a strong update to and overall a more 
effective outcome than those productivity prediction trends 
reported in recent meta-analysis on POA implementation 
outcomes [33]. This performance improvement becomes 
particularly salient when compared to existing automated 
essay evaluation systems, which are predominantly based on 
large language models, of which recent comparative studies 
have shown there are often significant limitations in terms of 
capturing the rich developmental patterns associated with 
authentic writing progression [34]. The incorporation of 
domain knowledge in education with the help of Feature 
Adaptive Selection Mechanism bridges the gap between 
writing assessment research and deep learning, as traditional 
methods of deep learning often tend to focus on statistical 
precision over pedagogical interpretability, hampering their 
deployment for practical guidance [35]. The systematic 
identification of five distinct student engagement profiles 
through SDT-based modeling contributes meaningfully to 
existing theoretical understanding while providing practical 
frameworks for personalized intervention strategies that 
extend beyond traditional one-size-fits-all approaches. 
Contemporary systematic reviews of artificial intelligence 
applications in personalized learning highlight the persistent 
challenge of translating sophisticated algorithmic capabilities 
into educationally meaningful interventions that respect 
individual learner differences and maintain pedagogical 
authenticity [36]. The achievement of 78.4% ± 2.1% early 
warning accuracy coupled with 79.8% ± 2.9% teacher 
satisfaction by such an unassuming framework indicates that 
closure may be reached with little compromise between the 
high-tech technological sophistication and educational utility 
that is found in many modern educational technology 
implementations. When viewed in light of the recent studies 
of the effects of POA on student psychological factors, it shows 
that student autonomy and intrinsic motivation can indeed be 
increased, rather than decreased, by well-designed forms of 
technology exposure [37]. 

The temporal learning trajectory analysis identifying 
different patterns of effectiveness across POA’s three phases 
attests to the theoretical soundness of the approach, while 
providing empirical evidence in support of timely 
instructions strategies proposed in the literature that adds to 
existing systematic review evidence on the diversity of POA 
implementation and effectiveness across educational settings 
[38]. The framework’s unique contribution over transformer-
based multidimensional feedback systems is its theoretically 
informed stance on feature selection and interpretation, 
countering criticisms against recent AI-driven writing 
instruction tools on the possible hiatus between automated 
feedback generation and actual learning support [39]. Human 

expertise validation introduced at various stages of the 
algorithmic design makes certain that technological 
capabilities supplement, rather than supplant, professional 
pedagogical judgment and comes in light of the emerging 
evidence that the effective utilization of educational AI 
applications mandates a delicate balance between fully 
automated optimization and supervisory human control in 
order to sustain learning authenticity [40]. The framework's 
contribution extends beyond immediate performance 
improvements to establish methodological precedents for 
educational theory-informed machine learning that 
addresses fundamental challenges in learning analytics 
regarding the integration of sophisticated computational 
approaches with established pedagogical knowledge. The 
success of multi-theoretical integration within predictive 
modeling suggests promising directions for future 
educational technology development that prioritize 
theoretical coherence and practical utility over purely 
technical optimization metrics, thereby advancing the field 
toward more sustainable and educationally meaningful 
artificial intelligence applications in language instruction 
contexts. 

5. Conclusion 
The POA-MLSP model is an important step in the 

application of educational technology on language teaching, 
which successfully shows that the advanced machine learning 
techniques can be meaningfully combined with the traditional 
pedagogical theory for improving CET-4 writing performance 
prediction and assistance. Having produced an R² = 0.75 
across five dimensions of writing and while receiving teacher 
satisfaction of 79.8% ± 2.9%, it demonstrates that developing 
educational technology solutions that reflect pedagogical 
authenticity can be balanced with technical expressiveness, 
and can resolve a potent dilemma in learning analytics: the 
linkage of algorithms to the practice of a classroom. The 
purposeful and principled incorporation of POA theory with 
SDT-based engagement modeling sets into motion a novel 
methodological precedent among literatures that seek to 
predict in educational contexts by fusing domain-specific 
knowledge and multi-theoretical models directly into 
algorithmic design. The discovery of five student engagement 
profiles and the 78.4% ± 2.1% prediction accuracy achieved 
in early warning detection show that the framework has the 
potential to provide personalized intervention strategies by 
bridging student autonomy and intrinsic motivation 
development. These findings imply potential research 
directions for future ed-tech development that is theoretically 
consistent and practical rather than focusing only on 
technical optimization criteria with limited educational 
significance. Framework scalability considerations include 
modular deployment options for institutions with varying 
technical capabilities, automated data collection mechanisms 
reducing manual teacher workload to sustainable levels, and 
cross-cultural adaptation protocols for implementation 
beyond Chinese university contexts. The minimal training 
requirements (4.8 hours initial setup, 25-30 minutes weekly 
operation) and cloud-based deployment options facilitate 
broader institutional adoption while maintaining educational 
effectiveness. The framework's contribution transcends 
immediate performance improvements to establish 
sustainable pathways for intelligent educational system 
development that respects the fundamental nature of 
teaching and learning processes while harnessing 
technological capabilities for enhanced educational 
outcomes. The validation of temporal learning trajectory 



Yu Li et al. /Future Technology                                                                                        November 2025| Volume 04 | Issue 04 | Pages 100-116 

115 
 

patterns across POA's three phases provides empirical 
evidence for optimized instructional timing strategies that 
can inform broader curriculum design and implementation 
practices within English language education contexts. The 
successful deployment of the framework within authentic 
CET-4 writing instruction environments demonstrates the 
viability of theory-informed learning analytics for supporting 
data-driven pedagogical decision-making that enhances 
educational equity through personalized learning support 
while maintaining the essential human elements that 
characterize effective language instruction. 

Ethical issue 
The authors are aware of and comply with best practices in 
publication ethics, specifically with regard to authorship 
(avoidance of guest authorship), dual submission, 
manipulation of figures, competing interests, and compliance 
with policies on research ethics. The authors adhere to 
publication requirements that the submitted work is original 
and has not been published elsewhere. 

Data availability statement 
The manuscript contains all the data. However, more data will 
be available upon request from the authors. 

Conflict of interest 
The authors declare no potential conflict of interest. 

References 
[1] Qiu, L., Enabling in the production-oriented approach: 

Theoretical principles and classroom 
implementation. Chinese Journal of Applied 
Linguistics, 2020. 43(3): p. 284-304. 
http://dx.doi.org/10.1515/CJAL-2020-0019 

[2] Zhang, W., Effects of the production-oriented 
approach on EFL learners’ writing performance in 
China’s tertiary education. Chinese Journal of Applied 
Linguistics, 2020. 43(3): p. 323-341. 
http://dx.doi.org/10.1515/CJAL-2020-0021 

[3] Liu, G. and H. Cao, The application of poa-based 
reciprocal teaching model in Chinese senior high 
school English writing class. Theory and Practice in 
Language Studies, 2021. 11(8): p. 891-900. 
http://dx.doi.org/10.17507/tpls.1108.04 

[4] Xie, Q., Using production-oriented approach in 
business English courses: perceptions of China’s 
English-major and non-English-major 
undergraduates. Sage Open, 2021. 11(2): p. 
21582440211016553. 
http://dx.doi.org/10.1177/21582440211016553 

[5] Zhao, Y., N. Ainil Sulaiman, and W. Wahi, Exploring 
the Impact of the Production-Oriented Approach on 
Chinese University Students' Motivation for English 
Learning: A Mixed Methods Study. Arab World 
English Journal, 2024. 15(3): p. 10.24093.  
http://dx.doi.org/10.24093/awej/vol15no3.7 

[6] Cheng, J. Blended learning reform in English viewing, 
listening and speaking course based on the POA in the 
post-pandemic era. in Frontiers in Education. 2025. 
Frontiers Media SA. 
http://dx.doi.org/10.2991/icosihess-19.2019.9 

[7] He, T. and C. Li, An empirical study on the teaching 
mode of cultural translation in college English based 
on the Production Oriented Approach (POA). PloS 
one, 2025. 20(6): p. e0326127. 
http://dx.doi.org/10.1371/journal.pone.0326127 

[8] Zhang, L. and Y. Kong, Integrating Production-
Oriented Approach (POA) in Flipped Classrooms: An 
Action Research on Enhancing Spoken English 
Instruction for English Majors in China. Forum for 
Linguistic Studies, 2025. 7(2): p. 117–136.  
http://dx.doi.org/10.30564/fls.v7i2.8136 

[9] Yibin, H., Reconstructing “Learning-Application 
Integration” Practical Training Courses for English 
Teacher Trainees From the POA Perspective. US-
China Education Review A, 2025. 15. 
http://dx.doi.org/10.17265/2161-
623X/2025.04.005  

[10] Noels, K.A., et al., Self-determination and motivated 
engagement in language learning, in The Palgrave 
handbook of motivation for language learning. 2020, 
Springer. p. 95-115. http://dx.doi.org/10.1007/978-
3-030-28380-3_5 

[11] Ryan, R.M. and E.L. Deci, Intrinsic and extrinsic 
motivation from a self-determination theory 
perspective: Definitions, theory, practices, and future 
directions. Contemporary educational psychology, 
2020. 61: p. 101860. 
http://dx.doi.org/10.1037/mot0000234 

[12] Chiu, T.K., Digital support for student engagement in 
blended learning based on self-determination theory. 
Computers in Human Behavior, 2021. 124: p. 106909. 
http://dx.doi.org/10.1016/j.chb.2021.106909 

[13] Chiu, T.K., Applying the self-determination theory 
(SDT) to explain student engagement in online 
learning during the COVID-19 pandemic. Journal of 
research on Technology in Education, 2022. 
54(sup1): p. S14-S30. 
http://dx.doi.org/10.1080/15391523.2021.1891998 

[14] Li, H., Impact of collaborative learning on student 
engagement in college English programs: Mediating 
effect of peer support and moderating role of group 
size. Frontiers in Psychology, 2025. 16: p. 1525192. 
http://dx.doi.org/10.3389/fpsyg.2025.1525192 

[15] Liu, H., Y. Wang, and H. Wang, Exploring the 
mediating roles of motivation and boredom in basic 
psychological needs and behavioural engagement in 
English learning: a self-determination theory 
perspective. BMC psychology, 2025. 13(1): p. 179. 
http://dx.doi.org/10.1186/s40359-025-02524-3 

[16] Yang, Y., J. Chen, and X. Zhuang, Self-determination 
theory and the influence of social support, self-
regulated learning, and flow experience on student 
learning engagement in self-directed e-learning. 
Frontiers in Psychology, 2025. 16: p. 1545980. 
http://dx.doi.org/10.3389/fpsyg.2025.1545980 

[17] Namoun, A. and A. Alshanqiti, Predicting student 
performance using data mining and learning analytics 
techniques: A systematic literature review. Applied 
Sciences, 2020. 11(1): p. 237. 
http://dx.doi.org/10.3390/app11010237 

[18] Zhang, Y., et al., Educational data mining techniques 
for student performance prediction: method review 
and comparison analysis. Frontiers in psychology, 
2021. 12: p. 698490. 
http://dx.doi.org/10.3389/fpsyg.2021.698490 

[19] Yağcı, M., Educational data mining: prediction of 
students' academic performance using machine 
learning algorithms. Smart Learning Environments, 
2022. 9(1): p. 11. http://dx.doi.org/10.1186/s40561-
022-00192-z 



Yu Li et al. /Future Technology                                                                                        November 2025| Volume 04 | Issue 04 | Pages 100-116 

116 
 

[20] Lin, C.-C., A.Y. Huang, and O.H. Lu, Artificial 
intelligence in intelligent tutoring systems toward 
sustainable education: a systematic review. Smart 
Learning Environments, 2023. 10(1): p. 41. 
http://dx.doi.org/10.1186/s40561-023-00260-y 

[21] Sghir, N., A. Adadi, and M. Lahmer, Recent advances in 
Predictive Learning Analytics: A decade systematic 
review (2012–2022). Education and information 
technologies, 2023. 28(7): p. 8299-8333. 
http://dx.doi.org/10.1007/s10639-022-11536-0 

[22] Wang, S., et al., Artificial intelligence in education: A 
systematic literature review. Expert Systems with 
Applications, 2024. 252: p. 124167. 
http://dx.doi.org/10.1016/j.eswa.2024.124167 

[23] Malik, S., et al., Advancing educational data mining for 
enhanced student performance prediction: a fusion of 
feature selection algorithms and classification 
techniques with dynamic feature ensemble evolution. 
Scientific Reports, 2025. 15(1): p. 8738. 
http://dx.doi.org/10.1038/s41598-025-92324-x 

[24] Angeioplastis, A., et al., Predicting student 
performance and enhancing learning outcomes: a 
data-driven approach using educational data mining 
techniques. Computers, 2025. 14(3): p. 83. 
http://dx.doi.org/10.3390/computers14030083 

[25] Zhao, C. and J. Huang, The impact of the scoring 
system of a large-scale standardized EFL writing 
assessment on its score variability and reliability: 
Implications for assessment policy makers. Studies in 
Educational Evaluation, 2020. 67: p. 100911. 
http://dx.doi.org/10.1016/j.stueduc.2020.100911 

[26] Chen, H. and J. Pan, Computer or human: A 
comparative study of automated evaluation scoring 
and instructors’ feedback on Chinese college 
students’ English writing. Asian-Pacific Journal of 
Second and Foreign Language Education, 2022. 7(1): 
p. 34. http://dx.doi.org/10.1186/s40862-022-00171-
4 

[27] Huang, C., Assessing reading comprehension in CET-
4: A cross-sectional case study. Interactive Learning 
Environments, 2023. 31(10): p. 7149-7158. 
http://dx.doi.org/10.1080/10494820.2022.2061012 

[28] Li, K., et al. Test fairness of the in-house College 
English examination for Chinese non-English major 
undergraduates: a case study. in Frontiers in 
Education. 2025. Frontiers Media SA. 
http://dx.doi.org/10.3389/feduc.2025.1518315 

[29] Sun, Q., F. Chen, and S. Yin, The role and features of 
peer assessment feedback in college English writing. 
Frontiers in Psychology, 2023. 13: p. 1070618. 
http://dx.doi.org/10.1177/0265532220927487 

[30] Wu, C., Y.-W. Zhang, and A.W. Li, Peer feedback and 
Chinese medical students’ English academic writing 
development: a longitudinal intervention study. BMC 
Medical Education, 2023. 23(1): p. 578. 
http://dx.doi.org/10.1186/s12909-023-04574-w 

[31] de Caux, B.C. and L. Pretorius, Learning together 
through collaborative writing: The power of peer 
feedback and discussion in doctoral writing groups. 
Studies in Educational Evaluation, 2024. 83: p. 
101379. 
http://dx.doi.org/10.1016/j.stueduc.2024.101379 

[32] Peungcharoenkun, T. and B. Waluyo, Students’ 
affective engagements in peer feedback across offline 
and online English learning environments in Thai 
higher education. Asian-Pacific Journal of Second and 
Foreign Language Education, 2024. 9(1): p. 60. 
http://dx.doi.org/10.1186/s40862-024-00286-w 

[33] Wang, W. and C. Lyu, The Effectiveness of Production-
Oriented Approach on Students’ English Language 
Skills: A Meta-Analysis. The Asia-Pacific Education 
Researcher, 2025. 
http://dx.doi.org/https://doi.org/10.1007/s40299-
025-00990-2 

[34] Liew, P.Y. and I.K. Tan. On Automated Essay Grading 
using Large Language Models. in Proceedings of the 
2024 8th International Conference on Computer 
Science and Artificial Intelligence. 2024. 
http://dx.doi.org/10.1145/3709026.3709030 

[35] Misgna, H., et al., A survey on deep learning-based 
automated essay scoring and feedback generation. 
Artificial Intelligence Review, 2024. 58(2): p. 36. 
http://dx.doi.org/10.1007/s10462-024-11017-5 

[36] Merino-Campos, C., The impact of artificial 
intelligence on personalized learning in higher 
education: A systematic review. Trends in Higher 
Education, 2025. 4(2): p. 17. 
http://dx.doi.org/10.3390/higheredu4020017 

[37] Huang, T., et al., The Effect of Production-Oriented 
Approach on Chinese University Students’ Foreign 
Language Writing Anxiety and English Writing 
Performance: Evidence From a Longitudinal Study. 
SAGE Open, 2025. 15(2): p. 21582440251344016. 
http://dx.doi.org/10.1177/21582440251344016 

[38] Sun, L., H.H. Ismail, and A.A. Aziz, Current English 
Language Teaching Using Production-Oriented 
Approach: A Systematic Review. World Journal of 
English Language, 2024. 14(4). 
http://dx.doi.org/10.5430/wjel.v14n4p101 

[39] Zheng, X. and J. Zhang, The usage of a transformer 
based and artificial intelligence driven 
multidimensional feedback system in english writing 
instruction. Scientific Reports, 2025. 15(1): p. 19268. 
http://dx.doi.org/10.1038/s41598-025-05026-9 

[40] Shi, H., et al., Comparing the effects of ChatGPT and 
automated writing evaluation on students’ writing 
and ideal L2 writing self. Computer Assisted 
Language Learning, 2025: p. 1-28. 
http://dx.doi.org/10.1080/09588221.2025.2454541 

 
 

  

This article is an open-access article distributed under the 
terms and conditions of the Creative Commons Attribution 
(CC BY) license 
(https://creativecommons.org/licenses/by/4.0/). 

https://creativecommons.org/licenses/by/4.0/

	1. Introduction
	The Production-Oriented Approach (POA) has emerged as a revolutionary pedagogical framework within Teaching English as a Foreign Language (EFL) contexts, representing a paradigmatic shift from input-based teaching methodologies toward a more integrati...
	Contrary to the studies described in the previous sections which investigated the effects of POA on language learning in comparison to that of TVA, recent POA experiments carried out in the context of university Classrooms in China show the success of...
	The difference between intrinsic and extrinsic motivation orientations is especially important in the context of a POA, as challenging activities on unlimited progression levels or limited progression levels must strike a balance between challenge and...
	The application of educational data mining and learning analytics to language instruction represents an emerging frontier that offers substantial potential for understanding and predicting student performance patterns, though the integration of these ...
	Recent advancements in the application of artificial intelligence technologies in education environments are indicative of the potential benefits offered as well as the challenges to the deployment of technology in traditional education [22]. State-of...
	Comparative studies between automated evaluation and previous studies indicate that there is a large gap between the assessment results, and there are some data conflicts among them, so a more advanced mechanism should be developed for further study, ...
	Collaborative writing approaches, which stress the role of peer feedback, interaction on dynamics, and learning from each other, have proven to result in increased learning outcomes owing to the social interaction and shared knowledge building by the ...
	Despite extensive theoretical development of POA methodology and significant advances in educational data analytics, a substantial research gap persists between these two domains within CET-4 writing instruction contexts. Current predictive modeling a...
	This investigation establishes three primary objectives: (1) to develop the POA-MLSP framework that systematically integrates POA theoretical constructs with advanced learning analytics for multi-dimensional writing performance prediction; (2) to esta...
	This novel approach obtains complete prediction models that facilitate decision-making of instructors in runtime and personal offers planning of intervention strategies, which can preserve the pedagogical soundness and make use of the computational po...
	2. Methods
	2.1 POA-MLSP framework educational theoretical foundation
	The theoretical foundation of POA-MLSP is set through the step-by-step embedding of Production-Oriented Approach principles and Self-Determination Theory mechanisms, which turns on to be a full-stack educational data modeling approach to handle the in...
	Figure 1 shows the comprehensive POA-MLSP theoretical foundation and implementation framework, illustrating systematic data collection across POA's three phases (motivation activation, knowledge construction, and feedback interaction), SDT three-dimen...
	The multi-theoretical integration guidance within predictive modeling draws upon the systematic review findings that demonstrate enhanced effectiveness when POA incorporates complementary theoretical frameworks, including Input Hypothesis principles, ...
	The theoretical synthesis mechanism extends beyond simple additive combination to establish dynamic interaction models that recognize how Input Hypothesis comprehensible input requirements influence Social Constructivism collaborative learning effecti...
	Assessment of competence dimensions emphasizes measures for the efficacy of learning and the experience of achievement by performance confidence ratings, frequency of pursuit of challenges, indices of mastery goal orientation, and patterning of succe...
	The 3D engagement model integrates these heterogeneous aspects through higher-level temporal  odelling techniques, in order to capture the evolutionary response of engagement patterns throughout the academic semester and, ultimately, to pinpoint those...
	2.2 POA-MLSP intelligent prediction framework design
	The overarching design remains child-centric based on a four-layer model, which holds that pedagogical considerations would drive technological implementation and not vice versa.  The input layer allows multi-dimensional input data to be collected on ...
	assessment framework can be supported directly. We place emphasis on support for teacher decision-making, at the application layer, by interpreting recommendations and offering insights that are actionable for instructional changes, student groupings,...
	Core algorithm educational adaptation design replaces traditional attention mechanisms with education-oriented approaches that align with established pedagogical theories and classroom realities. The Feature Adaptive Selection Mechanism represents a n...
	Figure 2. POA-MLSP framework architecture
	Its real-time adaptation machinery allows for adapting prediction models to changing classroom dynamics and student response patterns, ensuring that the suggestions generated by the algorithm remain sensitive to the continuous evolution of educational...
	Algorithm 2 employs three-dimensional temporal modeling through time-series analysis techniques to capture engagement pattern evolution across academic semesters. The computational framework utilizes automated change pattern recognition for each SDT d...
	Automatic identification of engagement pattern changes enables proactive intervention through early warning models that predict motivational decline before academic performance impact occurs. Subject-specific alert thresholds, dynamic adaptation to in...
	2.3 Five-dimensional writing performance prediction model
	The Jacobs framework-based multi-faceted prediction approach accounts for each dimension of writing competence through targeted sub-models that model the specificities of development and shaping of the proficiency related to different writing aspects....
	Language Use dimension forecasting emphasizes grammatical accuracy and syntactic complexity prediction through computational linguistics approaches that assess sentence structure variety, clause combination sophistication, error pattern identification...
	Mechanics dimension assessment targets spelling and punctuation norm compliance prediction through error detection algorithms that identify persistent mistake patterns, improvement trajectory analysis, and mechanical skill development indicators that ...
	Cross-dimensional interaction modeling reflects the reality that writing proficiencies manifest intricate interdependencies in which the enrichment of vocabulary profiles increases one's capacity to develop textual content and the fostering of organiz...
	These interaction patterns are revealed by advanced correlation analysis in order to guide systematic intervention strategies that exploit competency interrelationships for optimal learning effects. An ensemble learning prediction optimization strateg...
	Validation models, as utilized by the approach, embed domain knowledge from the education sector in the form of expert teacher evaluations of the accuracy of the predictions and the relevance of the suggestions, thus ensuring that the algorithm output...
	Table 2. Five-dimensional prediction model specifications
	Table 3. Model interpretability components
	Feature importance analysis features Reflect on which individual behaviors and learning dynamics are most influential to predicting performance, where intervention activities should be prioritized, while understanding the pedagogical reasoning underpi...
	3. Results
	3.1 Data collection and preprocessing
	The POA-MLSP framework validation was conducted at S Normal University, involving 124 students distributed across POA experimental classes and traditional instruction control groups during a complete 16-week academic semester. The research design inco...
	Preprocessing of the data included the use of advanced normalization techniques tailored to educational settings, taking care of integration of multi-scale measurement, time-alignment effects, and individual-response baseline variations while maintain...
	3.2 Experimental design and model training
	The experimental framework implemented temporally-aware methodological approaches, respecting chronological learning sequences while ensuring robust validation procedures. POA-MLSP model training procedures integrated hyperparameter optimization combi...
	3.3 Core framework component validation and educational application results
	The validation of the POA-MLSP model indicates that the system has achieved systematic effectiveness in CET-4 writing teaching and learning by virtue of its integrated four-layer architecture, which seamlessly fuses advanced learning analytics with es...
	Validation of Algorithm 1: The validation indicated a significant outperformance of traditional methods in finding educationally meaningful variables, such as student engagement indicators and collaborative learning patterns, which are correlated with...
	The established framework had established good predictive power for each of the five dimensions of the writing assessment, but also maintained educational interpretability for practical classroom use, as shown in Table 8. This Table demonstrates POA-M...
	To validate the individual contribution of each framework component, comprehensive ablation studies were conducted by systematically removing key algorithmic elements and measuring resulting performance degradation across all writing dimensions, as pr...
	Algorithm 2 validation revealed solid capability for early identification of at-risk students through three-dimensional engagement pattern recognition. The SDT-based modeling demonstrated consistent effectiveness across all three psychological dimensi...
	Figure 3 demonstrates engagement pattern diversity where High Sustained (28%) and Autonomy-Oriented (22%) profiles represent the largest student groups, while Figure 3(a) shows balanced distribution across five motivational types and Figure 3(b) revea...
	Sustained academic outcomes reveal framework-supported students achieving 12.3% higher final CET-4 writing scores while maintaining engagement levels 18.7% above baseline measurements throughout the academic year. Expanding on these intervention effic...
	3.4 Comparative analysis and advanced algorithm performance validation
	Following the comprehensive validation of POA-MLSP framework components and educational effectiveness demonstration, it became crucial to establish the framework's technical superiority through systematic comparison with established machine learning a...
	Figure 4(a) shows overall performance where POA-MLSP achieved R² = 0.75 ± 0.026, outperforming linear regression (R² = 0.58), random forest (R² = 0.66), and support vector machines (R² = 0.63) by 17-29%, validating educational theory integration benef...
	Figure 4. Algorithm performance comparison across writing dimensions
	Figure 5. Teacher satisfaction and framework usability assessment
	Stakeholder evaluation reveals strong teacher acceptance with 79.8% ± 2.9% reporting valuable decision-making insights and 81.5% ± 2.4% expressing continued utilization willingness. Implementation feasibility validation shows minimal training requirem...
	4. Discussion
	The POA-MLSP model validity test indicates a significant move forward in educational predictions from the existing approaches, especially against the background of the recent trends and advancements in language learning analytics and AI. The framework...
	The temporal learning trajectory analysis identifying different patterns of effectiveness across POA’s three phases attests to the theoretical soundness of the approach, while providing empirical evidence in support of timely instructions strategies p...
	5. Conclusion
	The POA-MLSP model is an important step in the application of educational technology on language teaching, which successfully shows that the advanced machine learning techniques can be meaningfully combined with the traditional pedagogical theory for ...
	The manuscript contains all the data. However, more data will be available upon request from the authors.
	Conflict of interest
	The authors declare no potential conflict of interest.
	References
	[1] Qiu, L., Enabling in the production-oriented approach: Theoretical principles and classroom implementation. Chinese Journal of Applied Linguistics, 2020. 43(3): p. 284-304. http://dx.doi.org/10.1515/CJAL-2020-0019
	[2] Zhang, W., Effects of the production-oriented approach on EFL learners’ writing performance in China’s tertiary education. Chinese Journal of Applied Linguistics, 2020. 43(3): p. 323-341. http://dx.doi.org/10.1515/CJAL-2020-0021
	[3] Liu, G. and H. Cao, The application of poa-based reciprocal teaching model in Chinese senior high school English writing class. Theory and Practice in Language Studies, 2021. 11(8): p. 891-900. http://dx.doi.org/10.17507/tpls.1108.04
	[4] Xie, Q., Using production-oriented approach in business English courses: perceptions of China’s English-major and non-English-major undergraduates. Sage Open, 2021. 11(2): p. 21582440211016553. http://dx.doi.org/10.1177/21582440211016553
	[5] Zhao, Y., N. Ainil Sulaiman, and W. Wahi, Exploring the Impact of the Production-Oriented Approach on Chinese University Students' Motivation for English Learning: A Mixed Methods Study. Arab World English Journal, 2024. 15(3): p. 10.24093.  http:...
	[6] Cheng, J. Blended learning reform in English viewing, listening and speaking course based on the POA in the post-pandemic era. in Frontiers in Education. 2025. Frontiers Media SA. http://dx.doi.org/10.2991/icosihess-19.2019.9
	[7] He, T. and C. Li, An empirical study on the teaching mode of cultural translation in college English based on the Production Oriented Approach (POA). PloS one, 2025. 20(6): p. e0326127. http://dx.doi.org/10.1371/journal.pone.0326127
	[8] Zhang, L. and Y. Kong, Integrating Production-Oriented Approach (POA) in Flipped Classrooms: An Action Research on Enhancing Spoken English Instruction for English Majors in China. Forum for Linguistic Studies, 2025. 7(2): p. 117–136.  http://dx.d...
	[9] Yibin, H., Reconstructing “Learning-Application Integration” Practical Training Courses for English Teacher Trainees From the POA Perspective. US-China Education Review A, 2025. 15. http://dx.doi.org/10.17265/2161-623X/2025.04.005
	[10] Noels, K.A., et al., Self-determination and motivated engagement in language learning, in The Palgrave handbook of motivation for language learning. 2020, Springer. p. 95-115. http://dx.doi.org/10.1007/978-3-030-28380-3_5
	[11] Ryan, R.M. and E.L. Deci, Intrinsic and extrinsic motivation from a self-determination theory perspective: Definitions, theory, practices, and future directions. Contemporary educational psychology, 2020. 61: p. 101860. http://dx.doi.org/10.1037/...
	[12] Chiu, T.K., Digital support for student engagement in blended learning based on self-determination theory. Computers in Human Behavior, 2021. 124: p. 106909. http://dx.doi.org/10.1016/j.chb.2021.106909
	[13] Chiu, T.K., Applying the self-determination theory (SDT) to explain student engagement in online learning during the COVID-19 pandemic. Journal of research on Technology in Education, 2022. 54(sup1): p. S14-S30. http://dx.doi.org/10.1080/15391523...
	[14] Li, H., Impact of collaborative learning on student engagement in college English programs: Mediating effect of peer support and moderating role of group size. Frontiers in Psychology, 2025. 16: p. 1525192. http://dx.doi.org/10.3389/fpsyg.2025.15...
	[15] Liu, H., Y. Wang, and H. Wang, Exploring the mediating roles of motivation and boredom in basic psychological needs and behavioural engagement in English learning: a self-determination theory perspective. BMC psychology, 2025. 13(1): p. 179. http...
	[16] Yang, Y., J. Chen, and X. Zhuang, Self-determination theory and the influence of social support, self-regulated learning, and flow experience on student learning engagement in self-directed e-learning. Frontiers in Psychology, 2025. 16: p. 154598...
	[17] Namoun, A. and A. Alshanqiti, Predicting student performance using data mining and learning analytics techniques: A systematic literature review. Applied Sciences, 2020. 11(1): p. 237. http://dx.doi.org/10.3390/app11010237
	[18] Zhang, Y., et al., Educational data mining techniques for student performance prediction: method review and comparison analysis. Frontiers in psychology, 2021. 12: p. 698490. http://dx.doi.org/10.3389/fpsyg.2021.698490
	[19] Yağcı, M., Educational data mining: prediction of students' academic performance using machine learning algorithms. Smart Learning Environments, 2022. 9(1): p. 11. http://dx.doi.org/10.1186/s40561-022-00192-z
	[20] Lin, C.-C., A.Y. Huang, and O.H. Lu, Artificial intelligence in intelligent tutoring systems toward sustainable education: a systematic review. Smart Learning Environments, 2023. 10(1): p. 41. http://dx.doi.org/10.1186/s40561-023-00260-y
	[21] Sghir, N., A. Adadi, and M. Lahmer, Recent advances in Predictive Learning Analytics: A decade systematic review (2012–2022). Education and information technologies, 2023. 28(7): p. 8299-8333. http://dx.doi.org/10.1007/s10639-022-11536-0
	[22] Wang, S., et al., Artificial intelligence in education: A systematic literature review. Expert Systems with Applications, 2024. 252: p. 124167. http://dx.doi.org/10.1016/j.eswa.2024.124167
	[23] Malik, S., et al., Advancing educational data mining for enhanced student performance prediction: a fusion of feature selection algorithms and classification techniques with dynamic feature ensemble evolution. Scientific Reports, 2025. 15(1): p. ...
	[24] Angeioplastis, A., et al., Predicting student performance and enhancing learning outcomes: a data-driven approach using educational data mining techniques. Computers, 2025. 14(3): p. 83. http://dx.doi.org/10.3390/computers14030083
	[25] Zhao, C. and J. Huang, The impact of the scoring system of a large-scale standardized EFL writing assessment on its score variability and reliability: Implications for assessment policy makers. Studies in Educational Evaluation, 2020. 67: p. 1009...
	[26] Chen, H. and J. Pan, Computer or human: A comparative study of automated evaluation scoring and instructors’ feedback on Chinese college students’ English writing. Asian-Pacific Journal of Second and Foreign Language Education, 2022. 7(1): p. 34....
	[27] Huang, C., Assessing reading comprehension in CET-4: A cross-sectional case study. Interactive Learning Environments, 2023. 31(10): p. 7149-7158. http://dx.doi.org/10.1080/10494820.2022.2061012
	[28] Li, K., et al. Test fairness of the in-house College English examination for Chinese non-English major undergraduates: a case study. in Frontiers in Education. 2025. Frontiers Media SA. http://dx.doi.org/10.3389/feduc.2025.1518315
	[29] Sun, Q., F. Chen, and S. Yin, The role and features of peer assessment feedback in college English writing. Frontiers in Psychology, 2023. 13: p. 1070618. http://dx.doi.org/10.1177/0265532220927487
	[30] Wu, C., Y.-W. Zhang, and A.W. Li, Peer feedback and Chinese medical students’ English academic writing development: a longitudinal intervention study. BMC Medical Education, 2023. 23(1): p. 578. http://dx.doi.org/10.1186/s12909-023-04574-w
	[31] de Caux, B.C. and L. Pretorius, Learning together through collaborative writing: The power of peer feedback and discussion in doctoral writing groups. Studies in Educational Evaluation, 2024. 83: p. 101379. http://dx.doi.org/10.1016/j.stueduc.20...
	[32] Peungcharoenkun, T. and B. Waluyo, Students’ affective engagements in peer feedback across offline and online English learning environments in Thai higher education. Asian-Pacific Journal of Second and Foreign Language Education, 2024. 9(1): p. 6...
	[33] Wang, W. and C. Lyu, The Effectiveness of Production-Oriented Approach on Students’ English Language Skills: A Meta-Analysis. The Asia-Pacific Education Researcher, 2025. http://dx.doi.org/https://doi.org/10.1007/s40299-025-00990-2
	[34] Liew, P.Y. and I.K. Tan. On Automated Essay Grading using Large Language Models. in Proceedings of the 2024 8th International Conference on Computer Science and Artificial Intelligence. 2024. http://dx.doi.org/10.1145/3709026.3709030
	[35] Misgna, H., et al., A survey on deep learning-based automated essay scoring and feedback generation. Artificial Intelligence Review, 2024. 58(2): p. 36. http://dx.doi.org/10.1007/s10462-024-11017-5
	[36] Merino-Campos, C., The impact of artificial intelligence on personalized learning in higher education: A systematic review. Trends in Higher Education, 2025. 4(2): p. 17. http://dx.doi.org/10.3390/higheredu4020017
	[37] Huang, T., et al., The Effect of Production-Oriented Approach on Chinese University Students’ Foreign Language Writing Anxiety and English Writing Performance: Evidence From a Longitudinal Study. SAGE Open, 2025. 15(2): p. 21582440251344016. http...
	[38] Sun, L., H.H. Ismail, and A.A. Aziz, Current English Language Teaching Using Production-Oriented Approach: A Systematic Review. World Journal of English Language, 2024. 14(4). http://dx.doi.org/10.5430/wjel.v14n4p101
	[39] Zheng, X. and J. Zhang, The usage of a transformer based and artificial intelligence driven multidimensional feedback system in english writing instruction. Scientific Reports, 2025. 15(1): p. 19268. http://dx.doi.org/10.1038/s41598-025-05026-9
	[40] Shi, H., et al., Comparing the effects of ChatGPT and automated writing evaluation on students’ writing and ideal L2 writing self. Computer Assisted Language Learning, 2025: p. 1-28. http://dx.doi.org/10.1080/09588221.2025.2454541

