Design and Effectiveness Evaluation of Data Mining-Based Artificial Intelligence-Driven Intelligent Teaching and Assisting System for Civics Classes

With the improvement of people’s quality of life and the level of consumption, the use of mobile smartphones has gained rapid popularity, and more and more people utilize fractional time for mobile learning [1]. Mobile learning refers to the use of wireless mobile communication network technology and wireless mobile devices to access learning information, learning resources and educational services of a new type of learning. In the face of a new generation of “digital natives” born under the Internet society, mobile learning is bound to become an inevitable choice for future learning [2].

On college campuses, while college students complete classroom learning, smart terminals have become their best companions. On the negative side, all kinds of smart terminals are coming out in an endless stream, especially some entertainment software is especially fascinating to young college students, and a lot of time and energy are consumed in it [3]. How to make college students better educated in the era of increasingly advanced smart terminals? In view of this, the concept of mobile teaching based on mobile learning has emerged. Of course, the development of mobile teaching is still relatively slow due to reasons such as hardware and network technology. However, with the arrival of the network era, all relevant parties, especially in the field of education, the development of technology in this area can be said to be rapidly changing, and now the price of mobile hardware equipment is gradually reduced, laying a good material foundation for mobile teaching [4].

The reform and improvement of the education and teaching of Civic and Political Science courses in colleges and universities is related to the overall development of Chinese higher education. From the overall perspective of the education and teaching of Civic and Political Science courses in colleges and universities, the classroom is still only a platform for the implementation of the curriculum, and only a way for teachers and students to communicate, and subject to the constraints of classroom time, space, number of students and other factors, the quality and quantity of this communication is very limited [5-6]. Therefore, in the era of information technology, the mobile teaching system of college civics class should be reconstructed and optimized in many aspects to form a new paradigm of education and teaching of college civics class in the era of “Internet+” [7]. Intelligent informatization technology will bring more new phenomena to the Civic and Political Science class in colleges and universities, and will be more conducive to the emergence of intelligent Civic and Political Science education with Chinese characteristics and in line with the world [8].

Sajja, R et al. conceptualized a personalized intelligent assistant for educational assistance in colleges and universities using artificial intelligence technology as the underlying architecture to provide personalized instructional support to students and effectively alleviate their learning burden [9]. Alam, A explored and dissected the practice of AI technology in the field of education involving smart campus, teacher assessment, and smart tutoring, pointing out that AI technology facilitates the improvement of teachers’ teaching quality and students’ learning outcomes [10]. Minn, S conducted a study on knowledge assessment and student dominant family model in the online teaching mode of AI using psychometric theory to analyze the reasons for the popularization and rapid development of AI in educational practice, which provides some new horizons and ideas for researchers in related fields [11]. Ahmad, S. F et al. elaborated that Artificial Intelligence technology empowers education and instructional management, significantly improves the effectiveness of education, and reduces the burden of instructional management for teachers and argued that related research needs to be validated with more quantitative studies [12]. Çela, E et al. empirically explored the attitudes and perceptions of Albanian students towards AI-assisted forms of learning and found that the frequency of use of AI learning tools showed a positive correlation with students’ grades, as well as with the efficiency of completing assignments, but that the excessive use of AI learning tools impeded the development of students’ critical thinking [13]. Li, X et al. built a framework for analyzing students’ innovation ability based on logistic regression method, and combined with the regression analysis model to explore how intelligent-assisted teaching affects students’ ability in terms of innovation, and the results of the study show that intelligent-assisted teaching significantly promotes the advancement of the level of students’ innovation ability [14]. The research of artificial intelligence technology in the field of education has attracted a lot of attention, and the direction of the research is mainly the path of artificial intelligence integration into education and optimization, and the assessment of the positive and negative impacts of artificial intelligence education on the main body of teaching and learning, etc., and the main methods used are empirical investigation, quantitative research and other social practice methods.

Civic education has an important impact on the formation of the three views of students in colleges and universities and their future comprehensive development, while the mainstream direction of research related to Civic education is to explore the development path of Civic education integrating innovative technologies and advanced educational theories, to find out the shortcomings of Civic education, and Civic teaching evaluation. Gao, C. et al. discussed the benefits of Internet technology for ideological education, which can strengthen the transmission of the spirit of ideology, enhance the precision and efficiency of ideological education, and also make the ideological education more timely and innovative, and finally gave important opinions for the development of ideological education combined with the big data analyzing technology and Internet technology [15]. Yang, H comprehensively analyzed the development path of intelligent teaching of civic and political education, and proposed an intelligent teaching system of civic and political education based on the background of innovation and entrepreneurship, and verified that the proposed innovation of civic and political education effectively enhanced the learning efficiency of students through the experimental class [16]. Wang, J emphasized the importance of innovation in Civic Education, as well as the relevance and effectiveness of Civic Education in the reform of Civic Education, and proposed to improve the political quality of Civic Teachers, to enrich the content and form of Civic Education, as well as to deepen the reform of the evaluation of Civic Education in the light of the actual situation [17]. Zhu, G et al. used qualitative research methods to elucidate the association between general history education and Civic and political education, and revealed the shortcomings of college education, such as utilitarianism, and put forward suggestions to improve college education, such as perfecting the evaluation mechanism of teachers and enriching the content of Civic and political education [18].

Facing the demand for intelligent teaching and tutoring in the field of Civics and Political Science education, this paper designs an intelligent teaching and tutoring system for Civics and Political Science courses with the main functions of deep knowledge tracking and exercise recommendation, which provides a direction for the construction of intelligent teaching and tutoring for Civics and Political Science courses in colleges and universities. Taking the DKVMN model as the framework, we use ResNet to model the three forgetting factors in depth, and use IRT to calculate the learning ability and the difficulty of exercises to construct a forgettable knowledge tracking model based on IRT. The Civics domain knowledge map is introduced to explore the implicit relationship between knowledge points, and the knowledge map embedding method is used to enhance the exercise representation, combine the knowledge tracking algorithm with the exercise recommendation algorithm, and set the reward category through the reward function to ensure that the reward value of learner’s participation is in a reasonable range, so as to complete the construction of the exercise recommendation model for Civics class. After the performance experiments of the in-depth knowledge tracking model proposed in this paper and the exercise recommendation model of the Civics class respectively, the first year class 1 of the Chinese language and literature major in the culture of H university is selected as the research object to carry out the application experiments of the intelligent teaching and learning aid system of the Civics class in this paper, and to explore the impact of this paper’s system on the cultivation of students’ Civic Culture Literacy and Civic Values Concepts.

2

Design of Intelligent Teaching and Support System for Civics Courses

2.1

System Functional Principle

2.1.1

IRT-based deep knowledge tracking model

This section describes a forgettable knowledge tracking model based on IRT, which is based on the DKVMN model and mainly consists of two parts: the deep forgetting modeling part and the deep IRT modeling part [19-20]. In the deep forgetting modeling process, the ResNet network is utilized to extract the features of forgetting factors and build the deep forgetting model. In the deep IRT modeling module, a feed-forward neural network is used to establish an item response theory network model to predict the accuracy of answers. [21]

1)

Model Structure

The model is based on the DKVMN model as a framework, and mainly consists of two parts: the deep forgetting modeling module and the deep IRT modeling module, where the deep IRT modeling module is divided into learning ability modeling and item difficulty modeling. The main workflow of the model is as follows: first, the answer input (q_t,r_t) is embedded to obtain the answer embedding vector ν_t, the forgetting vector F_t is obtained by combining the three forgetting factors, namely, the repetition time interval, the sequence time interval, and the number of answers in the past, and the feature extraction of F_t is performed to obtain the forgetting feature vector F by using the ResNet network. Then, the forgotten feature vector F and the answer embedding vector ν_t are combined to obtain the answer embedding vector $v_{t}^{F}$ after the forgetting process, from which the memory erasure vector and the memory enhancement vector are computed, and the attention weight w_t is computed according to the knowledge point storage matrix M^k, and the forgotten update of the knowledge mastery state matrix is realized by the memory erasure vector, the memory enhancement vector, the attention weight w_t and the LSTM module together. Finally, the probability of answering a question correctly was jointly predicted based on the learning ability coefficient and the item difficulty coefficient calculated in the IRT module.

The deep forgetting modeling module in the model is responsible for processing the answering information and forgetting information, and the input answering data and forgetting information data are trained by Res Net network to get the input data after forgetting processing. The deep IRT modeling module is divided into two parts, i.e., the learning ability modeling part and the item difficulty modeling part, which builds the neural network layer to calculate the learning ability and the item difficulty, respectively, and finally estimates the probability of the learner’s correct answer from these two parts.

2)

Deep item response theory modeling

In order to improve the interpretability and predictive performance of the knowledge tracking model, a new IRT model for deep knowledge tracking is proposed, which consists of two parts: learning ability network and item difficulty network. Among them, the item difficulty network is divided into two parts: the difficulty of the topic on which the learner takes the test and the difficulty of the knowledge skills related to the topic. Through the learning ability network and item difficulty network, the probability of the learner answering the question correctly can be calculated.

First, in the Item Difficulty Network, for the exercise numbered j, the Exercise Difficulty Coefficient $β_{i t e m}^{j}$ and the Difficulty Coefficient of the Knowledge Skills Required to Solve Exercise j, $β_{s k i l l}^{j}$ , are computed, where the Exercise Difficulty Coefficient denotes a difficulty parameter separate from the difficulty of the knowledge required for the topic. The final item difficulty is the sum of $β_{i t e m}^{j}$ and $β_{s k i l l}^{j}$ .

Before calculating the difficulty coefficient, the embedding vector k_j ∈ ▯^J obtained from Exercise q_j is uniquely encoded: (1) $q_{j m} = {\begin{array}{l} 1 & (j = m) \\ 0 & (o t h e r w i s e) \end{array}$

Here J denotes the total number of exercises and the item difficulty network contains n layers to compute the input items to the feedforward neural network: (2) $β_{1}^{j} = \tanh (W^{(q_{1})} k_{j} + τ^{(q_{1})})$ (3) $β_{l}^{j} = \tanh (W^{(β_{l})} β_{l - 1}^{j} + τ^{(β_{l})})$

where l = 2,.....,n, and then calculate the difficulty factor of the exercise $β_{i t e m}^{j}$ : (4) $β^{j} = [β_{1}^{j}, β_{2}^{j}, \dots, β_{n}^{j}]$ (5) $β_{i t e m}^{j} = \tanh (W^{(β_{i t e m})} β^{j} + τ^{(β_{i t e m})})$

The coefficient of difficulty of the knowledge points was then calculated in the same way, starting with solo heat coding for Knowledge Point S_j ∈ ▯^S: (6) $s_{j m} = {\begin{array}{l} 1 & (i t e m j r e q u i r e s s k i l l m) \\ 0 & (o t h e r w i s e) \end{array}$

Here S denotes the number of knowledge points, and the value of the input to the knowledge point skill difficulty network is calculated: (7) $γ_{1}^{j} = \tanh (W^{(γ_{1})} s_{j} + τ^{(γ_{1})})$ (8) $γ_{l}^{j} = \tanh (W^{(γ_{k})} γ_{l - 1}^{j} + τ^{(γ_{l})})$

where l = 2,.....,n. And then calculate the knowledge difficulty coefficient $β_{s k i l l}^{j}$ : (9) $γ^{j} = [γ_{1}^{j}, γ_{2}^{j}, \dots, γ_{l}^{j}]$ (10) $β_{s k i l l}^{j} = \tanh (W^{(β_{s k i l l)}} γ^{j} + τ^{(β_{s k i l l})})$

The input data for the learner’s learning ability network is then calculated: (11) $θ_{1}^{(t, j)} = \sum_{i = 1}^{N} M_{t}^{v} (i)$

Here $M_{t}^{v} (l)$ is the knowledge mastery state matrix in the DKVMN framework, followed by the sequential computation of all the values, and then the use of a feed-forward neural network to derive the learner learning ability θ(t,j): (12) $θ_{l}^{(t, j)} = \tanh (W^{(θ_{l})} θ_{l - 1}^{(t, j)} + τ^{(θ_{l})})$ (13) $θ^{(t, j)} = [θ_{1}^{(t, j)}, θ_{2}^{(t, j)}, \dots, θ_{l}^{(t, j)}]$ (14) $θ^{(t, j)} = \tanh (W^{(θ^{(t, j)})} θ^{(t, j)} + τ^{θ^{(t, j)}})$

Finally, the difference between a learner’s learning ability and item difficulty was used to predict the learner’s probability of answering a question correctly P_tj: (15) $P_{t j} = s i g m o i d (3.0 * θ^{(t, j)} - (β_{i t e m}^{j} + β_{s k i l l}^{j}))$

3)

Experimental results and analysis

Two datasets, ASSISTments 2009-2010 (ASSIST09) and ASSISTments 2012-2013 (ASSIST12), are selected as the experimental samples, and the deep knowledge tracking model proposed in this paper is subjected to comparison experiments and knowledge tracking experiments respectively.

(1)

Comparison Experiment

The comparison models selected for this comparison experiment can be categorized into three types, traditional models (DKT, DKVMN), KT models incorporating learned features (CKT, DKT-DT), and interpretable models (SAKT, Deep-IRT). The average experimental results of the five experiments are shown in Table 1. Analyzing the data in the table, it can be obtained that the IRT-based deep knowledge tracking model proposed in this paper achieves the highest prediction accuracy on all datasets. The AUC of this paper’s model is 82.77% and 80.3% on the ASSSIST09 dataset and ASSSIST12 dataset, respectively. The smaller the root mean square error value (i.e., RMSE), the closer the model’s prediction results are to the true value, and the RMSE values of this paper’s model on both the ASSSIST09 dataset and the ASSSIST12 dataset are the smallest 68.34% and 70.51%, which makes the model’s fitting even better.Higher values of the ACC and the R² similarly represent better model fitting, and the model of this paper on the ASSSIST09 dataset and ASSSIST12 dataset have ACC values of 76.36% and 76.31%, and R² values of 0.3098 and 0.2456, respectively, which are higher than the other comparative models, showing a better model fit.

Table 1.

Experimental results

Algorithm	RMSE		AUC		ACC		R²
Algorithm	ASSIST09	ASSIST12	ASSIST09	ASSIST12	ASSIST09	ASSIST12	ASSIST09	ASSIST12
DKT	0.9412	0.735	0.7634	0.6888	0.7246	0.7248	0.2091	0.1087
DKVMN	0.9138	0.7909	0.7384	0.6868	0.6926	0.6924	0.189	0.1437
CKT	0.9208	0.7434	0.7697	0.7067	0.7335	0.7346	0.2189	0.1236
DKT-DT	0.9018	0.7222	0.7683	0.6877	0.7286	0.7289	0.1974	0.1101
SAKT	0.9313	0.7544	0.7574	0.6993	0.7356	0.7356	0.2156	0.1853
Deep-IRT	0.9127	0.7127	0.7669	0.7021	0.7279	0.7282	0.211	0.2273
Model of this article	0.6834	0.7051	0.8277	0.803	0.7636	0.7631	0.3098	0.2456

(2)

Knowledge tracking experiment

The analysis of knowledge state can show how the model predicts the change of students’ knowledge over time and assess its interpretability. The visualization of the prediction results will improve the interpretability of the model, which will help students and teachers to effectively and intuitively understand the state of knowledge, and target to check and fill in the gaps. In this paper, we will evaluate the interpretability of the model based on the assessment point of “whether the model only updates the knowledge status of the relevant knowledge points answered at each time step, and whether the updating of the knowledge status is reasonable”.

In this paper, a student in the ASSIST09 dataset was randomly selected and changes in the student’s knowledge status were tracked over thirty time steps. The changes in this student’s mastery level of the concepts of the four Civics knowledge (knowledge points 1, 2, 3, and 4) over the 30 exercise interactions are shown specifically in Figure 1. The x-axis and y-axis in the figure show the time step and knowledge point respectively, and the cell color indicates the degree of change in proficiency level during the time step. The darker the color (the closer the number is to 1), the deeper the students’ mastery of the knowledge point. The first column shows the student’s initial mastery of the knowledge point, and as time passes, the student’s mastery of a certain knowledge will decrease, if a knowledge point is studied, when he answers a question correctly (or incorrectly), her mastery of the corresponding knowledge point increases (decreases) at the next moment. After completing the 30 exercises the student has a higher level of mastery of knowledge points 2, 3 and 4, while knowledge point 1 needs more practice for consolidation. Obviously, with the help of the in-depth knowledge tracking model in this paper, we can clearly understand the students’ mastery of the knowledge points of Civics, and accurately diagnose the weaknesses of their Civics learning.

2.1.2

Civics Exercise Recommendation Modeling

In this paper, we consider introducing external knowledge, i.e., the Knowledge Graph of the Civic and Political Science domain, and then preprocessing the knowledge points of the exercises using knowledge graph embedding techniques to enhance the representation of the states in the model and to mine the implicit relationships.

1)

Entity Definition and Relationship Extraction

Based on the logical structure of knowledge in the field of Civics and the functional requirements for domain knowledge mapping in the recommender system, three types of knowledge entities are defined: data structure entities, algorithm entities and problem entities, and the three types of knowledge concept sets are denoted as K_s, K_a, and K_p, respectively.

(1) Containment: ∀K_v, k_u and k_v ∈ K_s ∨ K_a ∨ K_p, ∃k_u ⊆ k_v are said to contain Knowledge Point k_v, which contains Knowledge Point k_v. For example, tree contains binary tree, sorting algorithm contains subsumption sort.

(2) Successor relationship: ∀k_u, k_v and k_v ∈ K_s ∨ K_a ∨ K_p, $\exists k_{u} \overset{n e x t}{\to} k_{v}$ that knowledge point k_u is the basis of knowledge point k_v, knowledge point k_v is said to be the successor knowledge point k_u, the two have a sequential relationship. For example: the circular chain table is the successor of the single chain table, the heap sort is the successor of the selection sort.

(3) Solve the relations: ∀k_u, k_v and k_u ∈ K_a, k_v ∈ K_p, $\exists k_{u} \overset{s o l v e}{\to} k_{v}$ , then the algorithm class entity k_u can solve the problem class entity k_v. e.g. DFS solves the shortest path problem.

(4) Realization relationship: ∀k_u, k_v and k_u ∈ K_s, k_v ∈ K_a, $\exists k_{u} \overset{i m p l e}{\to} k_{v}$ , then the data structure class entity k_u is said to be able to realize the algorithm class entity k_v. e.g.: queue realizes BFS.

2)

implicit relationship mining

Given a learner’s Civics learning sequence H_t = e₁,e₂,...,e_t and a minimum utility threshold E, traverse all learners with the following processing steps.

(1) Traverse the Civics learning interaction sequence and get the corresponding knowledge points according to the exercises to get the Civics knowledge point sequence K_t = k₁,k₂,...,k_t.

(2) Traverse the Civics knowledge point sequence, store two knowledge points as a group in set in sequential order and count them (the count is increased by 1 if the same group occurs once). If there are k_a and k_b in set in all the knowledge point sequences of utility value (the number of times they occur together) is greater than or equal to E, then it is stored as a high efficiency with 2-item set.

(3) Traverse the efficiently used 2-item set and if there exists a number of occurrences of [k_a,k_b] that is greater than or equal to the number of occurrences of [k_b,k_a], then create ternary (k_a,implication,k_b), and vice versa create ternary k_b,implication,k_a). This ternary shows that there is an implicit sequential relationship between k_b and k_a.

(4) All the generated ternary instances are added to the original knowledge graph and trained using TransE. The final knowledge embedding vector k containing the explicit relationships of the knowledge graph and the mined implicit relationships is obtained and used later.

3)

Model Architecture and Details

The DRSS model is mainly divided into policy generation module, state representation module and DRSS network module [22].

(1)

Policy Generation Module

Since the learner’s learning strategies may change to some extent as time changes during the process, in the strategy generation module, we use a sliding window of size T to generate the current learning strategy vector based on the learner’s knowledge vector of Civics within the time range n~n + T.

First, the strategy selection network receives the learner vector u_t and the average knowledge vector k_avg within the sliding window size of T as inputs, and this network makes a selection of learning strategies within the learner’s current time range T, i.e., it outputs the probability of each strategy. Then the one with the highest probability is selected as the current learning strategy.

(2)

State representation module

First, the sequence of exercise embedding vectors {e₁,e₂,...,e_t,...,n} and learner vectors u in the Civics learning sequence are used as inputs, and the sequence of exercise embedding vectors is obtained {g(e_i)|i = 1,...,n} through a weighted average pooling layer: (16) $g (e_{i}) = a v e (w_{i} e_{i}) | i = 1, \dots, n$

Then, the interaction vector is obtained by performing the elemental product operation of {g(e_i)|i = 1,...,n} with the learner vector. Finally, the learner vector, the interaction vector and the average pooling result are spliced. Since the current state representation of the learner is not only related to personal preferences but also influenced by the current learning strategy, the final form of the state representation is obtained by finally splicing with the learning strategy vector as follows: (17) $s_{t} = [u_{t}, u_{t} \otimes {g (e_{i}) | i = 1, \dots, n}, {g (e_{i}) | i = 1, \dots, n}, p_{t}]$

(3)

Actor Critic Network [23].

Actor network is mainly responsible for generating actions based on their states for a given learner. In this network, state s_t is used as input, and through two ReLU layers and one Tanh layer, state representation s_t is converted into action a = π_θ(s_t) as the output of Actor network. In order to avoid the algorithm from falling into a local optimum solution, as well as to increase the diversity of the recommendation results, we use Gaussian noise for exploration during the training process, i.e., a = π_θ(s_t) + N. Where action a ∈ ℝ^1×k is a continuous parameter vector that can be used to represent a ranking function that scores the exercises. Therefore, the action obtained after exploration is used to score exercise e_t to obtain score_t: (18) $s c o r e_{t} = e_{t} a^{•}$

Finally, the top n exercises with the highest rankings are recommended to the learner.

Critic networks are deep Q networks that utilize deep neural networks parameterized as Q_ω(s,a) to approximate the true state-action value function Q^π(s,a). In this network, the states s_t generated by the state representation module and the actions a_t generated by the Actor network are used as inputs, and the output is a scalar Q value, and then the parameters of the Actor network are updated based on the Q value towards improving the performance of the actions. Based on the deterministic policy gradient theorem, we can update the Actor by sampling the policy gradient shown in the following equation: (19) ${{\nabla_{θ} J (π_{θ}) \approx \frac{1}{N} \sum_{t} \nabla_{a} Q_{ω} (s, a) |}_{s = s_{t}, a = π_{θ} (s_{t})} \nabla_{θ} π_{θ} (s) |}_{s = s_{t}}$

where J(π_θ) is the expected value of all possible Q values that follow strategy π_θ. Here a small batch strategy is used and N denotes the batch size. In addition, we use temporal differencing to update the Critic network accordingly, minimizing the mean square error as follows: (20) $L = \frac{1}{N} \sum_{i} {(y_{i} - Q_{ω} (s_{i}, a_{i}))}^{2}$

Where, y_i = R_i + γQ_ω′(s_i+1,π_θ,(s_i+1)), ω′ and θ′ are the parameters of the target Critic network and Actor network.

(4)

Reward function

Based on the motivation of using a learner simulator to determine whether an exercise is challenging or not and update the reward value based on the learner’s performance in doing the problem, at time step t, we define the reward function by integrating three different reward functions: (21) $R (s_{t}, a_{t}) = γ_{1} R_{s e q} + γ_{2} R_{k g} + γ_{3} R_{k t}$

R_seq denotes the exercise sequence matching reward. We refer to the BLEU algorithm in machine translation, which employs an N-gram matching rule for the recommended sequences. Formally, given the actual interaction subsequence and the recommended subsequence, the exercise sequence matching reward function is defined as follows: (22) $R_{s e q} (e_{t : t + k}, {\hat{e}}_{t : t + k}) = \exp (\frac{1}{M} \sum_{m = 1}^{M} \log p r e c_{m})$

where prec_m is calculated as follows: (23) $p r e c_{m} = \frac{\sum_{p_{m} \in e_{t : t + k}} \min (# (p_{m}, e_{t : t + k}), # (p_{m}, {\hat{e}}_{t : t + k}))}{\sum_{p_{m} \in e_{t : t + k}} # (p_{m}, e_{t : t + k})}$

In the above equation, p_m denotes a subsequence of m levels of precision in e_t:t+k, and #(pm,e_t:t+k) denotes the number of times p_m appears in e_t:t+k.

R_seq denotes the knowledge matching reward. Given the actual and predicted subsequences, we use average aggregation to aggregate the knowledge embedding vectors of these two subsequences, denoted as k_t:t+k and ${\hat{k}}_{t : t + k}$ , respectively, and then use the cosine similarity measure to measure the difference between the two knowledge vectors as a reward function, which is calculated as follows: (24) $R_{k g} (k_{t : t + k}, {\hat{k}}_{t : t + k}) = \frac{k_{t : t + k} \cdot {\hat{k}}_{t : t + k}^{•}}{‖ k_{t, t + k} ‖ \cdot ‖ {\hat{k}}_{t, t + k} ‖}$

R_kt denotes the engagement reward. In order to simulate the real learning process of students, we elicited the knowledge tracking model as the student simulator and pre-trained it in advance. The recommended exercises are then used as inputs to the student simulator, and the probability of a student answering correctly is output. The reward value set for the output probability β obeys β ~ N(0.5,σ²), so that the final R_kt is expressed as follows: (25) $R_{k t} ({\hat{β}}_{t : t + k}) = \sum_{i = t}^{t + k} \frac{1}{\sqrt{2 π} σ} \exp (- \frac{{({\hat{β}}_{i} - 0.5)}^{2}}{2 σ^{2}})$

By integrating the above formulas, we can get the final reward function.

4)

Experimental results and analysis

In this section, the Civics exercise recommendation ability of the proposed Civics exercise recommendation model will be tested for different lengths of topic sequence recommendations. The relationship between the number of recommended Civics exercise topics and the reward change graph of rewards in training is specifically shown in Figure 2. The graphs (a) and (d) represent the effects of using the deep reinforcement learning recommendation model in different application environments. When there are few topics and they need to be recommended sequentially, the model converges quickly and stably. As the total number of topics increases and the number of topics to be recommended increases, convergence tends to be slow and the model does not converge accurately at the end of training. This phenomenon is due to the fact that the initial state of the student simulator is randomized each time, so the recommended topic scheme will not be the same. This also requires personalized recommendations for different students. From the analysis results, it can be concluded that the model in this paper is suitable for the scenario of the total number of exercises in the Civics course, which is in line with the teaching needs of the Civics course in colleges and universities.

2.2

System design

This study tracks the learners’ Civics knowledge mastery status with the help of the deep knowledge tracking model, understands the students’ Civics course learning status and Civics knowledge mastery, and uses the Civics exercise recommendation algorithm to realize the system’s personalized recommendation function for the students’ intelligent Civics course topics. On the basis of functional realization, this section will carry out the overall and architectural design of the Civics intelligent teaching and support system, and build an intelligent teaching and support system to provide certain support for the teaching of Civics in the future.

2.2.1

Overall design

System design is the foundation of developing an intelligent teaching and learning system, which plays a key role in the design of the system’s architecture, performance, and interface, and lays the foundation for the development and presentation of the intelligent teaching and learning system in the future. This section describes the working principle of the system in conjunction with the improved teaching process model, the keyword structure model, and the mastery competency model.

The functional components of a complete intelligent tutoring platform should mainly contain collecting learners’ process data during the teaching and learning process, analyzing and calculating the learner’s keyword mastery and generating student portraits and course portraits, recommending appropriate learning resources for the student to fit his/her learning path, and generating evaluative reports. The student portrait refers to the application of the results of each test, the results of questionnaires and the process data generated during self-study, including the collection of the length and number of the learner’s tests, the mastery assessment (i.e., the ability value of the mastery of various keywords) and the process evaluation (e.g., the learner’s learning efficiency and forgetting coefficients of the knowledge points of various subjects).

2.2.2

Architectural design

This paper utilizes a browser/server architecture. The teacher management interface (teacher-side) and student operation interface (i.e., student-side) are independent of each other, and the application data are provided through an API application programming interface.

According to the requirement analysis and architectural design described above, the system is set up as two major interfaces, the teacher side and the student side, as shown in Figure 3. The student side is a module for the learners to provide them with operational requirements, which includes two sub-modules: personal information settings and test and evaluation reports. The personal information module includes registration and login, personal information settings, notices, password reset, and course selection; the test and evaluation report module includes homework practice, test and quiz, report generation, and report viewing. Teacher’s side is a module for administrators and teachers to provide them with the functions they need to manage their operational needs, which includes five sub-modules: basic management, course management, question bank management, test management and system management. The basic management module includes interfaces for course list, knowledge keyword management, mastery management, etc. The course management module includes interfaces for course modeling and classroom management, the question bank management module includes interfaces for question bank management only, the exam management module includes interfaces for exam papers and quizzes, and the system management module includes interfaces for roles and user management.

3

Analysis of the Effectiveness of the Application of the Intelligent Teaching and Support System for Civic and Political Science Courses

This chapter will investigate the impact of the design of the intelligent teaching and support system for the Civics class constructed in this paper on the students’ Civics cultural literacy and Civics value concepts. The experimental object is all the students of the first year class 1 of the culture Chinese language and literature major in college H, i.e., the class is set as an experimental class. Based on the research design, data analysis of students’ pre- and post-test results was carried out with the help of SPSS22.0 software.

3.1

Analysis of Civic and Cultural Literacy Dimensions

The pre and post-test scores of the Civic and Cultural Literacy dimensions of the students in the experimental class and the control class are specifically shown in Table 2. From the table, it can be seen that the scores of the experimental class students in the rule of law awareness, public participation, and moral cultivation increased by 3.38, 2.54, and 1.53 respectively compared to the pre-test, and there is a significant difference between the scores of the pre and post-test (P<0.05). The difference between the scores of pre-test and post-test of cultural self-confidence is less than 1, which does not reach the significant level (P>0.05). The system described in this paper has a positive impact on the development of students’ civic and cultural literacy in areas such as rule of law awareness, public participation, and moral development.

Table 2.

Cultural literacy

Comparison	Variable	Test	Mean	SD	T	P
Comparison1	Consciousness of rule of law	Before experiment	10.12	2.361	-3.671	0.002
Comparison1	Consciousness of rule of law	After experiment	13.5	1.828	-3.671	0.002
Comparison2	Public participation	Before experiment	8.87	2.659	-6.22	0.006
Comparison2	Public participation	After experiment	11.41	1.808	-6.22	0.006
Comparison3	Cultural confidence	Before experiment	9.88	2.345	-1.741	0.156
Comparison3	Cultural confidence	After experiment	10.52	2.356	-1.741	0.156
Comparison4	Moral cultivation	Before experiment	13.65	3.517	-2.423	0.041
Comparison4	Moral cultivation	After experiment	15.18	3.114	-2.423	0.041

3.2

Analysis of the Dimension of Civic Value Concepts

The dimensions of the experimental class students’ civic values in the pre-test and post-test are specifically shown in Table 3. It can be seen that the scores of the pre-test and post-test on integrity, friendliness, family, and national sentiment did not reach a significant level of difference (P>0.05). The difference between the pre-test and post-test scores of the social responsibility dimension reaches 1.77, demonstrating a difference that reaches a significant level (P=0.001<0.05). The scores of the pre-test and post-test of ideal beliefs also reached the same significant level (P=0.022<0.05), with the difference between the scores of the pre-test and post-test of 1.45.The auxiliary application of this paper’s system in the Civic Science course has a good contribution to the cultivation of students’ Civic Science value concepts in the areas of ideal beliefs and social responsibility.

Table 3.

Ideological and political values

Comparison	Variable	Test	Mean	SD	T	P
Comparison1	Goodwill	Before experiment	15.39	2.388	-0.998	0.659
Comparison1	Goodwill	After experiment	15.87	2.527	-0.998	0.659
Comparison2	Social responsibility	Before experiment	14.52	2.472	-3.413	0.001
Comparison2	Social responsibility	After experiment	16.29	2.654	-3.413	0.001
Comparison3	National condition	Before experiment	11.22	2.322	-0.433	0.798
Comparison3	National condition	After experiment	11.4	1.652	-0.433	0.798
Comparison4	Ideal belief	Before experiment	11.48	1.814	-2.042	0.022
Comparison4	Ideal belief	After experiment	12.93	1.653	-2.042	0.022

On the whole, the intelligent teaching aid system designed in this paper enhances the teaching effect and quality of the Civics course, and plays a significant good effect on the cultivation of students’ Civics cultural literacy and Civics concept.

4

Conclusion

In this paper, we construct an intelligent teaching aid system for Civics courses based on the knowledge tracking and exercise recommendation functions of data mining to promote the construction of intelligent teaching aid for Civics courses in colleges and universities.

In the performance experiment of the in-depth knowledge tracking model constructed in this paper, the AUC of the model in this paper on the ASSSIST09 dataset and the ASSSIST12 dataset is 82.77% and 80.3%, the ACC value is 76.36% and 76.31%, and the R² value is 0.3098 and 0.2456, which is higher than that of the other comparative models, while the RMSE value is the smallest among all comparative models, 68.34%. The RMSE values are 68.34% and 70.51%, which are the smallest among all the comparison models, and overall, they demonstrate the highest prediction accuracy. In the knowledge tracking experiment, the model in this paper is also able to update the students’ knowledge status with the relevant knowledge points answered at each time step, realizing the precise diagnosis of students’ weaknesses in Civics learning.

The Civics exercise recommendation model constructed in the total number of topics and recommended topics is less, the model converges quickly and stably, and with the increase of the total number of topics and recommended topics, the convergence speed tends to be slow, and can’t be completely converged accurately, which is suitable for the scenario that the total number of topics of Civics is relatively more, and it can satisfy the needs of teaching Civics courses in colleges and universities.

Finally, we take the first year class 1 students of Chinese language and literature major in college H as the experimental object, apply the intelligent teaching aid system designed in this paper to the learning of Civics course, and analyze the evaluation changes of the dimensions of students’ Civics cultural literacy and Civics value concepts before and after the learning. In the dimension of Civic and Political Cultural Literacy, except for the dimension of Cultural Confidence, the pre- and post-test scores of Rule of Law Awareness, Public Participation, and Moral Cultivation are significantly different (P<0.05), and the post-test scores are increased by 3.38, 2.54, and 1.53 compared to the pre-test, respectively.The scores of the Civic and Political Values Concepts dimensions of Honesty and Friendliness and Family and Nation Sense in the pre-test and post-test do not reach a significant level (P>0.05), and the scores of Social Responsibility and Ideal Faith do not reach a significant level. ), while the difference between the pre-test and post-test scores of social responsibility and ideal belief reached 1.77 and 1.45, demonstrating a significant level difference (P<0.05).

To summarize, the intelligent teaching aid system for Civics designed in this paper has a positive role in promoting the learning of students’ Civics courses, which can promote the cultivation of students’ Civics cultural literacy and Civics values, and enhance the level of Civics course construction in colleges and universities.

Idioma:: Inglés

Calendario de la edición:: 1 veces al año
Temas de la revista:: Ciencias de la vida, Ciencias de la vida, otros, Matemáticas, Matemáticas aplicadas, Matemáticas generales, Física, Física, otros

RSS Feed de revista

Design and Effectiveness Evaluation of Data Mining-Based Artificial Intelligence-Driven Intelligent Teaching and Assisting System for Civics Classes

Yanghejia Wang

Publicado en línea: 19 mar 2025

Recibido: 30 oct 2024

Aceptado: 24 feb 2025

DOI: https://doi.org/10.2478/amns-2025-0544

Palabras claveDKVMN model, IRT modeling, Knowledge tracking, Knowledge mapping, Civics teaching and learning system

© 2025 Yanghejia Wang, published by Sciendo

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Palabras clave
DKVMN model, IRT modeling, Knowledge tracking, Knowledge mapping, Civics teaching and learning system