1. Introduction
Under the wave of global informatization, the education industry is undergoing unprecedented
changes, and intelligence and personalization have become the core trends of future
education development [1]. The research of intelligent educational content recommendation systems still needs
multiple challenges. On the one hand, it is difficult for the system to accurately
capture the deep semantics and knowledge structure of educational content, which leads
to the deviation between the recommended content and the actual needs of users. On
the other hand, the complexity and diversity of user behavior data also bring processing
difficulties to the system, which makes it difficult for personalized recommendation
algorithms to exert the best results.
As a branch of machine learning, deep learning can automatically learn high-level
abstract features in large-scale data sets by its multi-level nonlinear transformation
ability, especially in image recognition and speech analysis [2]. In recent years, with the improvement of computing power and the enrichment of data
resources, deep learning has begun to penetrate the field of education, showing great
potential to reshape the educational ecology. Deep learning also plays a vital role
in natural language processing (NLP). It achieves an in-depth understanding of text
data through advanced architectures such as recurrent neural networks (RNN), long-term
and short-term memory networks (LSTM), and Transformer. Moreover, generation provides
solid technical support for information retrieval, content analysis, and dialogue
interaction of intelligent education systems [3,
4].
The core goal of the intelligent educational content recommendation system is to dynamically
push the most appropriate learning materials and courses according to each student’s
unique learning needs and interests to stimulate learning motivation and improve learning
efficiency [5,
6]. The traditional recommendation system mainly screens the content according to the
user’s browsing history, scoring record, or behavior track. Although it can realize
personalized recommendations to a certain extent, it often needs to pay more attention
to the knowledge structure, difficulty level, and logical sequence of the learning
materials themselves, resulting in limited pertinence and practicality of the recommendation
results [7]. In contrast, the intelligent education recommendation system that integrates deep
learning and natural language processing technology can comprehensively consider multi-dimensional
information such as students’ ability level, knowledge mastery, and emotional state
from a more macro perspective and deeply analyze the internal connection and extension
meaning of learning content with the help of NLP technology. In turn, highly customized
learning path planning and resource suggestions are generated [?].
In the research of intelligent educational content recommendation system integrating
deep learning and natural language processing, there are innovations in technology
and methodology. At the technical level, this study abandons the traditional practice
of single or simple combination technology, deeply integrates deep learning and natural
language processing, uses deep learning to automatically extract complex latent features
in massive educational data, and relies on natural language processing to accurately
understand educational texts, so as to accurately model educational content and student
needs, such as predicting course interest tendencies and improving the accuracy of
recommendations by analyzing students’ online learning discussion texts combined with
deep learning models [10]. At the same time, it breaks through the limitations of focusing only on text data
in the past, and includes multi-modal data such as images and audio, such as comprehensive
understanding of the course by integrating image and audio information when processing
online video courses, so as to provide students with more suitable recommendations.
It also uses incremental learning technology to update the model in real time, reflect
the changes in students’ learning status and educational content in a timely manner,
and quickly adjust recommendations to improve the learning experience. In terms of
methodology, traditional recommendation systems are mostly based on group behavior
or simple user portraits, and this study uses deep learning and natural language processing
to construct students’ individual knowledge graphs, and tailors personalized learning
paths according to students’ learning goals and problem descriptions, such as recommending
personalized learning sequences for mathematical problems. The reinforcement learning
methodology is introduced, and the interaction between students and the recommendation
system is regarded as a dynamic decision-making process, and the recommendation strategy
is optimized according to the students’ feedback to achieve dynamic optimization.
In addition, the cross-domain knowledge transfer methodology is adopted to learn the
general knowledge representation and recommendation mode from vocational training
and other related fields and migrate to the intelligent education recommendation system
to broaden the recommendation ideas and methods.
This study focuses on the in-depth exploration of an intelligent educational content
recommendation system that integrates deep learning and natural language processing,
and innovatively constructs a multi-level technical architecture. At the level of
natural language processing, the Transformer-based Bidirectional Encoder Representation
Model (BERT) combined with dynamic word vector generation technology is used to conduct
in-depth semantic analysis of the text of educational resources, which can not only
accurately identify the logical association between knowledge points, but also mine
the potential knowledge context through the mask language model. In the field of deep
learning, a user portrait generation module is built through deep neural network (DNN),
the attention mechanism is used to weighted analysis of user historical learning behavior
data, personalized learning preferences are captured, and graph neural network (GNN)
is introduced to model the user-content interaction relationship, which effectively
solves the sparsity problem in collaborative filtering. At the same time, the multi-modal
data fusion technology is used to align and fuse the semantic information of the text,
the visual features of the image, and the dynamic information of the video, and realize
the dimensionality reduction and feature enhancement through the autoencoder. In terms
of recommendation strategy optimization, based on the reinforcement learning framework,
with the goal of maximizing user learning completion and knowledge mastery, a reward
function is constructed, combined with Deep Q Network (DQN) to realize dynamic policy
adjustment, and a cold-start recommendation module based on knowledge graph embedding
is designed for the cold start scenario of new users, and the entity relationship
completion technology is used to recommend relevant educational content, which effectively
alleviates the long-tail effect and significantly improves the accuracy, timeliness
and user learning experience of recommendation. It provides technical solutions with
both theoretical depth and practical value for personalized recommendation of intelligent
educational resources.
This study aims to build an intelligent education platform that can accurately match
students’ needs and maximize teaching effects by integrating deep learning and natural
language processing technology. Specifically, the system will rely on deep neural
networks’ powerful representation learning ability to deeply mine massive educational
data and extract the knowledge elements crucial to students’ academic growth and relationship
networks. At the same time, natural language processing technology will be used to
understand the main idea, difficulty, and applicable scenarios of text-based learning
materials and to analyze the semantic information in the interaction between teachers’
teaching content and students to establish a fine content index and labeling system
on this basis. Subsequently, the system continuously optimizes the recommendation
algorithm through deep reinforcement learning and other strategies so that it can
adjust the recommendation strategy in time according to students’ real-time feedback
and learning results, ensure that the recommended content always fits students’ learning
progress and personal interests, and promote the iterative upgrade of the personalized
learning experience. Intelligent educational content recommendation system involves
long-term projects such as data collection, model construction, and system testing.
Among them, high-quality and diversified educational data collection is particularly
critical. It must cover many learning subjects and forms to ensure the comprehensiveness
and timeliness of recommendation results. At the same time, privacy protection and
ethical issues cannot be ignored. In collecting and utilizing personal learning data,
the system must follow relevant laws and regulations and take measures such as encrypted
storage and anonymous processing to ensure users’ personal information security and
maintain an excellent educational and ecological environment.
2. Recommendation System Theory and Related Technologies
2.1. Deep Learning Technology
CNN is widely used in image feature extraction, classification, and scene recognition
because it can directly process original images [11]. The CNN structure includes input, convolution, pooling, fully connected, and output
layers, and the convolution and pooling layers are arranged alternately. CNN weight
sharing and local connection realize neuron connection and jointly reduce model complexity
and weight number [12].
The convolution kernel K of the two-dimensional convolution layer is usually regarded
as an interaction window to set the step size sliding. The convolutional layer output
is shown in Eq. (1), where $M$ is the size of the output image, $k$ is the size of the convolution kernel,
$s$ is the step size, and $I$ is the size of the input image.
A layer of pixels is filled around the input image matrix, and the commonly used element
0 is used as the padding. The output of the convolution layer containing the padding
is shown in Eq. (2), where $P$ is the number of pixels filled.
$N$ is the size of the output feature map, and $I$ is the size of the input image.
Pooling operations are divided into average pooling, maximum pooling, and random pooling.
Average pooling averages, retaining the background; Maximum pooling extracts peaks
and preserves textures; Random pooling samples according to probability, and probability
greatly affects the effect. RNN is good at processing sequence data and can flexibly
deal with sequences of different lengths and process extended sequence information
[13,
14]. The RNN can memorize information at each moment, and the current hidden layer depends
on the input and the hidden layer at the previous moment, as shown in formulas Eqs.
(3)-(4).
$O_t$ represents the output at the current time, and $V$ represents the weight matrix
from the hidden layer to the output layer. $S_t$ is the hidden state at the current
moment, and $g$ represents the feature extraction function. $U$ is the weight matrix
input to the hidden layer, $X_t$ is the input at the current time, $W$ is the weight
matrix from the hidden layer to the hidden layer, $S_{t-1}$ is the hidden state at
the previous time, and $f$ is the activation function. LSTM is a variant of RNN, which
can capture long-term dependencies and includes three control information flows: forgetting,
updating, and outputting. The first step is determining which information in the cell
state $C_{t-1}$ at the previous moment needs to be retained. The forgetting gate converts
the information in $h_{t-1}$ and $X_t$ into a value between 0 and 1 through the sigmoid
function. The size of this value determines which information is retained. The formula
is shown in Eq. (5), where $f_t$ represents the output of the forgetting gate, $W_f$ represents the weight
matrix of the forgetting gate, $b_f$ represents the deviation of the forgetting gate,
$h_{t-1}$ represents the hidden state at the previous time, $x_t$ represents the input
at the current time, and $\sigma$ represents the Sigmoid activation function.
The second step is to generate the information that needs to be updated. First, the
update gate decides which values must be updated through the sigmoid function. Then,
a new cell state $C_{2t}$ is generated using the tanh function, which may be added
to the cell state as the cell state generated at the current moment. The formulas
are shown in Eqs. (6)-(8).
$i_t$ represents the output of the input gate and determines the part to be updated.
$C_{2t}$ represents the new cell state candidate value generated by the tanh function,
and $C_t$ is the cell state at the current moment. $W_C$ represents the weight matrix,
$b_C$ represents the bias term used in the candidate cell state calculation process,
$b_i$ represents the bias term of the Input Gate, and $C_{t-1}$ and $C_{2-1}$ are
state values. The third step is to calculate the final output of the model. First,
we use the sigmoid function to get a preliminary output result. Next, using the tanh
function, the value of cell state $C_t$ is scaled, and its range is limited to $[-1,1]$.
This step helps to understand better and analyze the changes in cell state. Finally,
the scaled cell state value is multiplied by the initial output result to obtain a
more accurate and reliable final output result. The formulas are shown in Eqs. (9)-(10).
$W_o$ represents the weight matrix of the Output Gate. The Gated Recurrent Unit (GRU)
has two gates for resetting and updating, and its input and output structure is consistent
with that of the standard RNN—inputs and outputs of the GRU at time $t$. At time $t$,
the input of the GRU includes the hidden layer state $h_{t-1}$ and the input vector
$x_t$, while the output is only the hidden layer state $h_t$. The GRU workflow is
divided into two steps. The first step is to obtain a new hidden layer state, and
a new hidden layer state $h_{2t}$ is obtained according to $h_{t-1}$ at the previous
time and the current input $x_t$. The formula is shown in Eqs. (11)-(12).
$r_t$ stands for the output of the Reset Gate. $W_r$ is the weight matrix of the reset
gate, $b_r$ is the corresponding bias term, and $b_h$ represents the hidden state
of the candidate. The second step is to obtain the final hidden layer state, and the
update gate is used to obtain the final hidden layer state. $z_t$ represents the output
of the reset gate, and $b_z$ is the corresponding bias term. The formulas are shown
in Eqs. (13)-(14).
2.2. Natural Language Processing Technology
In the intelligent educational content recommendation system, ADGITN, LKS-Rec, DKT,
and MCMC algorithms each have their own key roles [15]. The ADGITN algorithm focuses on solving the problem of data sparsity and cold start,
and can dynamically adjust the graph structure according to the data in real time
by building an adaptive graph structure to capture the complex relationship between
students and educational content, making it flexible and scalable when processing
large-scale high-dimensional data. The LKS-Rec algorithm deeply integrates the knowledge
graph with the recommendation algorithm by mining the potential knowledge structure
of students, which can provide explanatory recommendation results and make the recommendation
association clearly visible.
The DKT algorithm, or deep knowledge tracking algorithm, can effectively model students’
learning process, predict students’ future knowledge mastery based on their past learning
performance, and provide a strong basis for personalized recommendations. The MCMC
algorithm is used to deal with complex probabilistic models in this system, which
can efficiently sample from high-dimensional distributions, estimate and optimize
model parameters, and show advantages in dealing with uncertainty problems and large-scale
data.
The integration point of deep learning with NLP components is that the powerful feature
extraction capabilities of deep learning are used to process educational text content,
and NLP technology is used for semantic understanding and analysis. In terms of combination
strategy, the combination of mathematical formulas such as collaborative filtering
formula and deep learning model loss function is used to realize the comprehensive
consideration of students’ interests and the characteristics of educational content.
In specific tasks, CNNs are suitable for processing image-based feature extraction
of educational resources, while RNN, LSTM, and GRU are selected for analyzing students’
learning behavior sequences and text content sequences due to their ability to process
sequence data, and each of them plays a unique advantage to help the efficient operation
of the intelligent educational content recommendation system.
Natural Language Processing (NLP) is a necessary discipline integrating computer science,
artificial intelligence, and linguistics. It is dedicated to enabling machines to
understand, interpret, and even generate natural language humans use [16]. The natural language computing architecture is shown in Fig. 1. The key to NLP lies in analyzing the complex structure of language, identifying
its semantics, and effectively extracting or generating information to realize barrier-free
information exchange between man and machine [17,
18].
Word segmentation is the most basic and crucial step in NLP, which involves segmenting
continuous text into independent lexical units. Different languages may have different
word segmentation rules. For example, English is usually automatically segmented according
to spaces, while Chinese needs to use unique algorithms to determine word boundaries
[19,
20]. Accurate word segmentation is the premise of subsequent grammatical and sentiment
analysis tasks. Part-of-speech tagging means that each word in a given sentence is
marked with its corresponding part of speech, such as noun, verb, adjective, etc.
This technology helps understand the grammatical structure of sentences and plays
a vital role in syntactic analysis and semantic parsing. Named entity recognition
focuses on recognizing proper nouns such as person names, place names, and organization
names in text and is the basis of advanced applications such as information extraction
and question-answering systems. NER technology can help us quickly locate essential
information and improve the efficiency and accuracy of information processing. The
syntactic analysis aims to parse the structure of sentences and construct a dendrogram
representing the relationship of sentence components, that is, a syntactic tree. This
analysis is crucial for understanding and generating complex language expressions,
especially in machine translation, intelligent question answering, and other scenarios.
Rather than focusing solely on parts of speech or syntactic structure, SRL looks at
understanding the logical relationships between individual components in a sentence,
i.e., who performs what action and impacts whom. This deep language understanding
is critical to high-order NLP tasks such as building knowledge graphs and event detection.
Sentiment analysis is a technique to evaluate the emotional tendency of text. It is
mainly used to judge whether comments, social media posts, etc., are positive, negative,
or neutral [21]. It has a broad application value in market research and public opinion monitoring.
With the mighty computing power of deep learning models, machine translation (MT)
has evolved from a simple statistical method to a neural network architecture, achieving
high-quality transformation across languages. MT technology has extensively promoted
the circulation and sharing of global information. The dialogue system, including
chat robot and voice assistant, is designed to simulate the human dialogue process
and provide information inquiry, transaction processing, and other functions [22]. They combine various NLP technologies, such as context understanding, intent recognition,
and reply generation, to provide users with an intelligent service experience.
Fig. 1. Natural language computing architecture.
2.3. Recommendation Algorithm
The content filtering algorithm realizes personalized recommendations based on an
in-depth analysis of users and projects, uses implicit feedback to collect information,
and constructs user and recommendation characteristics [23]. K-means grouping, iterative clustering optimization, and single-sided and double-sided
clustering in parallel can be used. The prediction performance is improved through
matrix factorization algorithms, such as principal component analysis, singular value
decomposition, non-negative matrix factorization, and the new algorithms, such as
maximum interval matrix factorization and probability matrix factorization. At the
same time, the advantages of the hybrid recommendation system to integrate content
and collaborative filtering basic algorithms are used to solve the problems of cold
start and data sparsity [24,
25]. The strategies include weighted fusion, scene switching, result mixing and reordering,
feature combination, algorithm cascade, and algorithm meta-level fusion. Mixed recommendations
are deepened by integrating expert systems, similarity, and probability frameworks.
3. Research on Educational Content Recommendation Based on the Integration of Deep
Learning and Natural Language Processing
3.1. Design and Implementation of Personalized Educational Content Recommendation
System
In the system, natural language processing uses pre-trained language models (such
as BERT) to perform deep semantic coding on the text of educational resources, and
combines named entity recognition and relationship extraction technology to construct
a knowledge graph in the field of education, which transforms unstructured content
into structured semantic vectors and solves the problem of semantic understanding
of educational content. At the same time, natural language processing technology is
used to analyze user query intent and tap potential needs. Deep learning uses deep
neural networks to construct user portraits, uses attention mechanism to weighted
and analyzes user historical learning behavior data, accurately captures personalized
preferences, and models user-content interaction with the help of graph neural networks,
effectively solving the problem of data sparsity. In addition, multi-modal data fusion
technology is used to integrate text, image, video and other features, and the recommendation
strategy is dynamically optimized through reinforcement learning, and the model parameters
are adjusted according to the real-time feedback of users. The synergy between the
two can achieve a deep understanding of educational content, accurate capture of user
needs, and dynamic optimization of recommendation strategies, significantly improving
the accuracy and personalization level of the recommendation system.
In the research field of intelligent educational content recommendation system, reinforcement
learning provides a powerful dynamic adaptive method for optimizing recommendation
strategies. At its core, it is about learning the optimal behavior strategy through
the interaction between the agent and the environment, continuous trial and error,
and feedback rewards.
Application of the Deep Q network model
In this intelligent educational content recommendation system, the Deep Q Network
(DQN) is used as the basic model for reinforcement learning. DQN cleverly combines
the powerful feature extraction capabilities of deep learning with the decision optimization
mechanism of Q learning algorithms. In the system, the neural network will take the
state of multi-dimensional information such as the user’s learning progress, knowledge
mastery, and learning habits as input, and then output the Q value corresponding to
each possible action (such as recommending different types of course videos, practice
questions, reading materials, and other educational content). This model structure
can effectively deal with the complex state space composed of rich user features and
diverse educational content in the intelligent education scenario, and accurately
respond to the recommendation task.
DQN uses online strategy learning in this system. During the agent’s real-time interaction
with the environment, the Q network parameters are continuously updated. For each
interaction, the agent selects one of many possible actions based on the current state,
and then observes the reward of the environmental feedback and the new state entered,
and uses the newly acquired information to update the Q network. This online strategy
enables agents to continuously adjust their recommendation strategies according to
real-time changes in the education scenario, and always maintain sensitivity and adaptability
to user needs.
By comparing the performance of intelligent educational content recommendation systems
with and without reinforcement learning, it can be found that the system integrated
with reinforcement learning performs better. In systems without reinforcement learning,
recommendation strategies are often based on fixed rules or historical data statistics,
which is difficult to flexibly adjust according to the user’s real-time learning situation.
With the introduction of reinforcement learning, the system can dynamically optimize
the recommended content based on the user’s immediate feedback during the learning
process. For example, when users have a slow grasp of a certain type of knowledge,
the system can automatically increase the recommendation of relevant exercises and
explanatory materials through reinforcement learning to improve the user’s learning
effect. Experimental data verify that the recommendation system using reinforcement
learning has a significant improvement in key indicators such as user retention rate
and learning efficiency improvement, which strongly proves the important value of
reinforcement learning in the intelligent education content recommendation system.
In the system, deep learning and natural language processing are tightly integrated,
and deep learning models such as multi-layer perceptrons, convolutional neural networks,
recurrent neural networks and their variants can automatically mine the learning style,
interest preferences, knowledge mastery and other characteristics from the students’
learning behavior data, provide key inputs for the recommendation algorithm, and improve
the accuracy and personalization of recommendations. Natural language processing technology
transforms educational content into structured representation through lexical, syntactic,
and semantic analysis, uses word vector model to calculate the semantic similarity
between students’ learning needs and educational content, realizes accurate matching
and recommendation, and can also classify and label content, and the two collaborate
with each other. The system faces many privacy risks in real-world data processing,
covering students’ personal information, learning records, behavioral data and educational
institution teaching resource data, etc., data collection, transmission, storage and
use of all aspects of privacy leakage, for which anonymization strategies can be adopted,
such as hashing of personal identity information, the use of generalization and suppression
technology to reduce data identification, and the system must follow the principles
of ethical compliance to ensure that the data collection is legal and proper, and
the purpose of use is consistent with the purpose of collection. Keep your data safe
and secure.
At present, educational platforms and resource libraries are surging, but unified
content is complex to meet individual needs. Currently, most platforms still need
to thoroughly consider students’ individual differences, resulting in low learning
efficiency [26]. Personalized services can establish accurate matching between users and resources,
optimize resource utilization and user experience, and enhance the platform’s attractiveness.
Online education personalization focuses on the interactive mapping between students
and resources, builds a linkage network, and improves the learning effect. The system
needs to face students’ individual differences, provide multiple services, and select
recommended resources according to students’ knowledge.
In the intelligent educational content recommendation system, data preprocessing model
selection and dataset segmentation criteria are crucial. In terms of data preprocessing,
the text cleaning model cleans the text from a wide range of sources and with noise
(such as HTML tags, special characters, garbled characters, etc.) in the field of
intelligent education, based on regular expressions and error correction with the
help of the pre-trained language model BERT. Because the educational texts involve
professional terms and subject vocabulary, the word segmentation model selects the
stammering word segmentation based on the combination of dictionary and statistics,
and enhances the word segmentation effect by expanding the custom dictionary. The
word vector model takes into account the specialized and domain characteristics of
educational texts, and uses the Word2vec or GloVe models to convert the text into
vectors. In terms of dataset segmentation standards, the user data of the intelligent
education system is separated by user type, and the student subset is subdivided by
grade and academic performance, and the teacher subset is divided by teaching experience
and taught disciplines, so as to accurately recommend different users; For example,
the data of the past year is selected as the training set, the next month is used
as the verification set, and the next half month is used as the test set to simulate
the real learning scenario to ensure that the model has adaptability and prediction
ability to the data that changes over time. According to the content type, the data
related to different types of intelligent education content, such as video courses,
e-books, and practice questions, are divided into different subsets, and the training
set, verification set, and test set are divided in proportion to each subset, and
the recommendation model is trained and optimized according to the characteristics
of different content types, so as to improve the accuracy and pertinence of recommendation.
This system adopts an MVC hierarchical structure, mainly realized by PHP, JavaScript,
and Ajax programming languages and technologies. The personalized system recommends
hidden question resources for registered students. After registering and logging in,
students choose courses and do preliminary tests to evaluate their knowledge level.
You can search for resources, collect, view history, and personalize learning on the
homepage. You can comment and collect during browsing, and the system records behaviors
for subsequent accurate recommendations. Users of this system are divided into two
categories: student users and administrators. Student accounts support registration
and login, including resource search, collection, discussion management, and personal
center. After logging in, browse and copy resources. The system recommends similar
test questions based on the algorithm and supports collection, help, and viewing history.
The personal center manages initial information. The background is divided into ordinary
and senior administrators. Ordinary administrators manage and optimize systems and
data and handle resource uploading, classification, and student problems. Advanced
administrators have additional permissions to modify and delete student accounts.
Fig. 2 illustrates the operation process of an intelligent educational content recommendation
system that integrates deep learning and natural language processing technology. The
first step is data collection, which covers all types of education-related data. Next,
the data preprocessing stage is used to clean and organize the data. Then, feature
engineering is carried out to prepare effective features for subsequent model training.
Uses deep learning and natural language processing modules to assist feature engineering
with technologies such as text classification, sentiment analysis, and semantic understanding.
After the feature preparation is completed, the model is trained, and the trained
model is evaluated. If the evaluation is up to standard, it will enter the online
recommendation link and push educational content for users; If the target is not met,
return to retrain the model. The feedback generated by users after receiving the recommended
content will be re-flowed into the data preprocessing stage, so as to continuously
optimize the recommendation system and improve the accuracy of the recommendation
and user experience.
Fig. 2. Flow chart of intelligent educational content recommendation system.
This system’s development tool and programming language is PHP, and the database is
MySQL. Students must register and log in to enjoy full functions. Otherwise, they
can only browse anonymously, without collection or comment permission, and no initial
test. After registering and logging in, operate the test questions resources, record
learning information systematically, recommend high-similarity test questions, improve
efficiency and knowledge mastery, and avoid the problem of massive resource screening.
To register, you must fill in your username, primary, email address, and password,
all required, and be format compliant. After completion, you will jump to the login
page. After logging in, students can browse the test questions, and their feedback
will be prompted by messages. Click on the details of the test questions, comment
in the database, and support the collection. Browse the extended information, search
for more related resources, and dig deeper into the topic. Login users click to get
recommendations, and the system will personalize the questions according to the answer
records. Students can search by knowledge points or browse similar test questions
to expand their learning. This interface manages the database, including test questions,
resources, information, and personal management. Administrators can upload, modify,
and delete test questions, manage student information (modify, delete, and freeze),
and handle messages and comments (reply and delete).
3.2. Key Technologies of the System Based on the Fusion of Deep Learning and Natural
Language Processing
The dataset used in this study is derived from multi-channel integration, including
real user data obtained in collaboration with online education platforms, and publicly
available educational resource libraries. The dataset size covers a sample of 1 million
students’ online learning records, involving more than 150 characteristics such as
basic user information, learning behavior logs (such as course browsing time, answer
records, and search keywords), and integrates 200,000 pieces of text information of
educational content, including course introductions, textbook chapters, exercise analysis
and other materials. In the pre-processing stage, the user behavior data is timestamped
and standardized, outliers are eliminated, and category features are processed by
One-Hot coding. For the content of educational texts, BERT is used for dynamic word
vector embedding, combined with stop word filtering and word reduction technology
to clean the text, named entity recognition and relationship extraction to construct
an educational knowledge graph, and finally feature normalization and principal component
analysis (PCA) to reduce the dimension to ensure data quality and model training efficiency,
so as to improve the reproducibility and reliability of the research.
In the intelligent educational content recommendation system, for CNN, the traditional
convolution operation for images treats the input as a vector sequence when processing
educational text data, and adjusts the convolution kernel dimension to capture local
features such as specific word combinations in the text sequence, and the pooling
layer also carries out maximum pooling according to the order of the text sequence,
reducing the feature dimension and retaining the key features. The basic RNN sets
the input as educational content or user behavior feature vectors to capture time
series relationships. LSTMs can better handle long-term dependencies through multiple
equations such as forgetting gates and input gates, such as analyzing students’ learning
trajectories and remembering students’ long-term learning preferences and knowledge
mastery. The GRU structure is relatively simple, the calculation efficiency is high,
and in the real-time recommendation scenario, the recommendation content can be quickly
adjusted according to the current and previous behavior of students. The agent training
using MCMC is based on the Bayesian inference framework, which treats the parameters
as random variables and updates the prior beliefs about the parameters based on the
observational data. MCMC constructs a Markov chain, and its stationary distribution
is a posteriori distribution that is to be sampled, and solves the high-dimensional
problem by randomly walking in the parameter space to avoid falling into the local
optimum. Taking the Metropolis-Hastings algorithm as an example, the parameter state
is initialized, the prior distribution, likelihood function and suggestion distribution
are defined, and then the candidate state is generated from the suggestion distribution,
the acceptance probability is calculated, and the acceptance is decided according
to the random number.
DeepFM and SHAN are two key technologies. DeepFM combines deep learning with a factorization
machine, in which the factorization machine handles the interaction of low-order features,
and deep learning is responsible for the interaction of higher-order features, which
can comprehensively capture the complex relationship between students’ learning behavior,
grades, interests, educational content and other characteristics, so as to accurately
predict students’ interest in educational content and learning effect, and achieve
personalized recommendations. Based on the self-attention mechanism, SHAN can automatically
learn the importance weights of each part of text data (such as course descriptions,
learning material texts, student learning feedback, etc.), process and learn data
from different granularities and levels through hierarchical aggregation, effectively
extract key information and semantic features, understand the context relationship,
provide more semantic understanding of feature representations for educational content
and student information, and help improve the accuracy and effectiveness of the recommendation
system.
Reinforcement learning-driven interactive recommendation (IR) simulates real-time
user interaction and dynamically updates strategies based on feedback. It is superior
to static content or collaborative filtering recommendations and fits real scenarios.
Based on Markov Decision Process (MDP), the user sequence behavior and preference
are dynamically modeled, the recommendation diversity is enhanced, the cumulative
benefit of the system is maximized, and the long-term feedback of users is paid attention
to [27,
28]. Using this model, we focus on user satisfaction and knowledge improvement, enhance
the accuracy of recommendations, and achieve the expected results.
In this paper, deep knowledge tracking (DKT) is used to model learners’ knowledge
status in a personalized test recommendation system, and combined with dynamic cognitive
diagnosis, students’ knowledge is tracked in real-time. Technically, semi-offline
and semi-real-time methods are adopted: offline regular training, long-term monitoring
of students’ problems throughout the learning process, and capturing changes in knowledge
status [29,
30]. Online real-time application immediately reflects students’ current knowledge status
and realize rapid responses to small data. The process of explicitly modeling students’
knowledge state is shown in Fig. 2.
Fig. 3 shows the deep knowledge tracking (DKT) model. First, a personalized knowledge model
should be built to save learner status and support knowledge level prediction. Second,
the knowledge state is obtained in real-time, and the environmental state of the reinforcement
learning agent is simulated for reward prediction.
Fig. 3. Knowledge tracking part models students’ knowledge status.
During online learning, the user interaction triggers the agent to select exercises
($q_1, q_2, \dots, q_n$) from the exercise set $A$. $N$ is often 1 or 2, as action
$a$. After the user answers the questions, the user feeds back the change of knowledge
state and rewards $r$ to the agent and recommends them in a cycle. Both knowledge
status and recommendation strategies influence learners’ mastery of topics. The action
set originates from the exercise resource library, and the agent selects the topic
$\{q_1, q_2, \dots, q_n\} \in Q$. Two optimization paths are adopted, and instant
rewards are used to maximize learning to find the best single step: Total reward maximization
MCMC algorithm, focusing on the whole system. In this paper, the MCMC algorithm is
used to train the agent. Due to the low efficiency and high cost of online direct
training agent recommendations, offline data is used for regular training and online
applications.
The process includes database extraction history records, file storage, formatting
input, and offline training after preprocessing. According to the online strategy
recommendation, facing the vast state and action space, calculate the similarity,
approximate processing, and adjust according to the user scale.
4. Experimental Results and Analysis
In terms of experimental setup, this study used a hierarchical sampling method to
divide the dataset into a training set and a test set in an 8 : 2 ratio, reduced the
randomness of model evaluation through 5-fold cross-validation, and used Adam optimizer
and Early Stopping strategy to prevent overfitting during the training process, set
the batch size to 128, and adopt a dynamic decay strategy for the learning rate. The
baseline comparison method selects collaborative filtering (CF), content-based recommendation
(CB), factorization machine (FM), deep neural collaborative filtering (NCF), and the
recent mainstream graph neural network recommendation model (LightGCN) to ensure that
the traditional and cutting-edge recommendation technologies are covered. The performance
evaluation focuses on two core metrics: AUC (area under the curve) and CTR (click-through
rate), which measures the model’s ability to distinguish between suitable and inappropriate
educational resources, and CTR, which directly reflects the appeal of recommended
content to users. The experimental results showed that compared with the baseline
method, the AUC value of the system increased by 23% and the CTR increased by 19%.
In order to verify the statistical significance of the results, the Paired Samples
t-test was used, and the results showed that the improvement of AUC and CTR passed
the significance test ($p < 0.01$) under the 95% confidence interval, which means
that the performance improvement was not caused by random factors, which further proved
the effectiveness and superiority of the recommendation system integrating deep learning
and natural language processing in educational scenarios.
In this study, experiments were carried out based on specific hardware and software
hyperparameter tuning. The hardware environment is a workstation equipped with an
NVIDIA RTX 3090 GPU (with 24GB of video memory, which can efficiently handle large-scale
matrix operations during deep learning model training), an Intel Xeon Platinum 8380
CPU (40 cores and 80 threads, providing stable and powerful computing support for
the overall operation of the system), and 128GB DDR4 high-speed memory (to ensure
fast data reading and storage, reducing computing bottlenecks).
In the research of intelligent educational content recommendation system, the commonly
used benchmarking methods include collaborative filtering, SLIM, DKN, NCF. According
to the user-item scoring matrix, collaborative filtering predicts the target user’s
score of unrated items by calculating the similarity between users or items (often
using cosine similarity), and in terms of hyperparameters, the number of neighbors
$k$ is generally 50, which affects the personalization and stability of recommendations,
and the cosine similarity is often selected for the similarity measure. SLIM transforms
the recommendation problem into a least-squares sparse optimization problem with regularization,
and controls the risk of model sparsity and overfitting by minimizing the weights
between items learned by the objective function, and the regularization parameters
range from 10-4 to 10-1. DKN combines knowledge graph and deep learning, uses CNN
to extract text features, fuses knowledge graph information, and outputs recommendation
scores through the fully connected layer, with 64 convolution kernels of size 3 and
a learning rate of 10-3 for hyperparameters. NCF models user-item interaction through
neural networks, maps user and project ID inputs to a low-dimensional vector space
through the embedding layer, and then transforms the neural network structure such
as MLP to output a predictive score, and the hyperparameters include the embedding
dimension $d = 64$ and the number of neurons and layers in the hidden layer.
The experiment uses Junyi academy and ASSISTments2009 public data sets derived from
Junyiacademy.org mathematics answer records and ASSISTments mathematics exercises,
respectively. In order to adapt to the requirements, the two data sets are preprocessed,
formatted, and filtered for sparse records. Table 1 is an overview of the details of the data set.
Table 1. Overview of dataset.
|
Dataset Overview
|
Junyi academy
|
ASSISTments2009
|
|
Number of students
|
71322
|
6245
|
|
Number of exercises
|
89688
|
26627
|
|
Number of knowledge points
|
1253
|
185
|
|
Number of exercise records
|
2193303
|
424658
|
|
Number of Learning Stages
|
97593
|
22580
|
Fig. 4 presents the change in MSE loss of the ADGITN model on the Junyi academy dataset
as training progresses. In the early stage, the part accounting for about two-thirds
of the model parameters cannot be updated due to the freezing of the word embedding
layer, resulting in limited learning and slow loss reduction. The word embedding layer
uses a pre-trained GloVe word vector. In order to avoid large gradients destroying
its information, it is frozen first and then thawed after the model parameters are
stable to ensure a better overall training effect. The loss shown in the figure is
the average of each training step.
Fig. 4. Variation diagram of main mean square error loss function value.
Fig. 5 depicts the change of KL loss on the item and user sides in the training of the ADGITN
model, and the trends of the two are similar. In the beginning, even if the word embedding
layer freezes, the loss drops rapidly, indicating that the model first learns to establish
attention distribution. The loss gradually increases with the deepening of training,
reaching a balance with other loss items. Although the KL loss increases slightly,
the overall loss decreases.
Fig. 5. Two training phases Variation plot of KL loss function value.
Fig. 6 compares the MSE losses of five models on the Junyi academy test set. Each model
adopts the strategy of first freezing and then thawing the word vector, resulting
in a gentle decrease in initial losses. Among them, the ADGITN model performs the
best, but the convergence rate is slightly lower.
Fig. 6. Variation plot of mean square error loss function value on test set.
The default experimental results in Table 2 show that on Junyi academy, the default model outperforms all comparison groups,
especially the noU2I structure, causing a sudden drop in performance, confirming the
critical role of attention-guided information transmission mechanism on RQ2. Auxiliary
task mining U2I comment value is better than the direct input method; SameSizeCNN
experiments prove that multi-scale CNN can capture more diverse features.
Table 2. Default experiments.
|
Model
|
ADGITN
|
noU2L
|
Add-w
|
Concat2Add
|
FM2FC
|
SameSizeCNN
|
ModRevTog
|
|
Mean square error value
|
1.5334
|
1.8601
|
1.5532
|
1.562
|
1.5807
|
1.562
|
1.5554
|
|
Percentage decrease (%)
|
0
|
19.3204
|
1.4025
|
2.0141
|
3.2912
|
2.0141
|
1.5554
|
Fig. 7 shows the comparison of similarity measures. High data sparsity leads to falsely
high similarity of user pair (item pair), and many zero values make the algorithm
misjudge the commonality and affect the clustering accuracy. The similarity measurement
method proposed in this paper can effectively deal with the sparsity problem and significantly
improve the identification of project similarity differences.
Fig. 7. Comparison of similarity measures.
Fig. 8 shows the model’s performance under the educational content data set. LKS-Rec represents
this paper’s reinforced recommendation system version, and LKS is the non-reinforced
version. The former is more efficient. The experiment found that strengthening recommendations
can significantly improve learning results and the state of knowledge combination.
Compared with the click-through rate of standard recommendation models and collaborative
filtering, it is found that the model in this paper has the best performance.
Fig. 8. Model performance under educational content data set.
Table 3 compares traditional recommendation methods with research recommendation systems
that fuse deep learning (DL) and natural language processing (NLP). In terms of resource
differentiation ability, the AUC of the experimental recommender system was increased
by 23% compared with the traditional method. In terms of click-through rate (CTR),
it is 19% higher than traditional collaborative filtering (CF) methods, and the recommended
content is more engaging. In addition, the advantages of the experimental recommendation
system in analyzing the semantics of educational content and understanding user behavior
are also reflected from the multi-dimensional comparison of personalized recommendation
accuracy, ability to process user behavior data, and ability to deal with cold start
problems. The effectiveness of the local migration algorithm is verified through experimental
comparison, and the experimental results are shown in Fig. 9. RBT and IWFQI are selected as references. The former is based on the similarity
migration model and reward function instance migration, while the latter considers
task differences and weights reuse trajectories. Experiments show that the proposed
algorithm converges faster, which proves that it can effectively extract information
from the source task and improve the learning efficiency of the target task.
Table 3. Comparison of recommended systems.
|
Comparison
|
Traditional referral methods
|
Experimental recommender system
|
|
AUC
|
-
|
23% improvement in resource discrimination
|
|
CTR
|
-
|
19% improvement over CF, more appealing recs
|
|
Ability to comprehend
|
Hard to capture edu content semantics
|
DL & NLP analyze text for learning material key info
|
|
Personalized recommendation accuracy
|
Limited pertinence and practicality
|
NLP - aided student - tailored content recs
|
|
Ability to process user behavior data
|
Difficult to deal with
|
Deep learning mines student data, NLP analyzes semantics to enhance user behavior
understanding.
|
|
Ability to deal with cold start issues
|
Poor testimonials
|
Pre - trained models + incremental learning for new users/small data recs
|
Fig. 9. Comparative experiment of local migration algorithm.
Fig. 10 shows the performance of non-personalized SLIM and NCF models in various metrics.
The model performance could be better based on users or projects, considering changes
in interest over time. Although DeepFM predicts the user’s click probability, it does
not fully use course attributes such as introduction and duration and has the worst
effect. The fusion method of deep learning and natural language processing proposed
in this paper is significantly superior to the first two in Precision and AUC, effectively
solving the “cold start” problem and proving that integrating course text information
can significantly improve recommendation accuracy. SHAN captures users’ dynamic interests
through a two-layer attention network, which is better than SLIM and NCF, especially
in indicators that reflect users’ long-term and short-term preferences. DKN uses CNN
to process text, optimally extract local features, and emphasize the value of content
information for recommendation.
Fig. 10. Performance of each model under indicators.
Variant testing was conducted in this study to evaluate the value of each module.
Fig. 11 shows that FIHA-T (no text information) and FIHA-A (single-layer attention) surpass
NCF, confirming that the proportion of short-term interest is high. FIHA-T is better
than DKN, and the project semantic information utility is more robust; FIHA-A is less
full-featured than FIHA, suggesting a better integration of long-term and short-term
interests. FIHA-T exceeds FIHA-A on P @ ten and AUC, indicating that curriculum semantic
features are crucial. FIHA integrates two components, and the recommendation effect
is the best.
Fig. 11. Ablation experiment.
5. Conclusion
Through the DL and NLP, this study builds an IECRS, aiming to accurately meet the
individual needs of learners and improve learning efficiency. After a series of rigorous
experimental design and data analysis, the system showed significant improvement effects:
In terms of recommendation accuracy, compared with the baseline model, the IECRS achieved
a 23% increase in AUC value on the test data set, indicating that the system can more
accurately distinguish educational resources that are suitable and unsuitable for
learners, significantly reducing the number of invalid recommendations, and improving
user satisfaction.
The appeal of recommended content has also been significantly enhanced. By optimizing
the recommendation strategy of the model, the CTR was successfully increased by about
19%, which means that more learners are willing to accept and actively interact with
the content recommended by the IECRS, which is crucial to stimulating learning interest
and promoting active learning.
This study effectively alleviates the cold start problem. Using the pre-trained deep
learning model and incremental learning strategy, the IECRS can still provide reasonable
and practical recommendation services when dealing with newly registered users or
learning fields with small data volumes. This system also embodies excellent scalability
and flexibility. The IECRS performs well when handling large-scale data streams or
adapting to diverse learning scenarios.
In terms of model efficiency, the training efficiency is limited by the demand for
a large amount of computing resources and time for deep learning models, and the long
training time when processing massive educational data will affect the system update
and adaptability, which can be improved with the help of distributed training technology
and efficient optimization algorithms. The inference efficiency is directly related
to the response speed of the recommender system, especially in the real-time recommendation
scenario, which can be improved through model compression and lightweight architecture
design. In terms of scalability, with the increase in the number of students and the
abundance of educational resources, the system needs to have the ability to cope with
the growth of data scale, and the use of distributed storage and computing frameworks
and the design of scalable data processing processes are effective solutions. The
feasibility of real-time deployment covers many aspects, and real-time data processing
requires efficient collection and analysis of newly generated data, which can be achieved
with the help of stream processing technology. System stability and reliability are
crucial, and it is necessary to establish and improve monitoring and fault tolerance
mechanisms, as well as redundant backup and fault recovery technologies to ensure
them. Resource management can be combined with the cloud computing platform through
containerization technology to dynamically adjust resources according to the load
and improve utilization. In short, despite facing many challenges, through reasonable
technology selection and architecture design, the intelligent education recommendation
system can break through the dilemma, better serve students’ personalized learning,
and promote the process of education intelligence.