The Design of Intelligent Q&amp;A System for English Education Based on Artificial Intelligence Technology and the Cultivation of Students’ Autonomous Learning Ability

With the continuous development of artificial intelligence technology, AI education platform has gradually become an important part of modern education, and the AI English education intelligent question-answering system is a product of this era [1-3], which uses big data and machine learning algorithms to be able to provide personalized learning assistance according to the students’ learning situation and differentiated needs [4-6]. By analyzing students’ learning data, assessment tests, and problems, it provides insights into students’ learning habits and knowledge mastery, and then tailors learning plans and teaching resources to fit students’ learning styles and levels [7-10]. This kind of personalized learning assistance enables students to learn in a way that meets their own characteristics and needs, thus effectively improving their problem-solving ability, especially conducive to the development of students’ independent learning ability [11-14].

The AI English education intelligent question-answering system can provide personalized learning paths and resources for each student based on the student’s learning status and characteristics. It analyzes students’ learning data and behaviors, based on the targeted adjustment of teaching contents and methods according to students’ strengths and weaknesses [15-18]. And students can independently choose the learning content and learning mode according to their own learning progress and interest characteristics to improve the learning effect. In addition, the Q&A system can also provide targeted review and consolidation materials according to the students’ learning situation to help them learn independently [19-22].

In this paper, an intelligent Q&A system for English education is designed to improve students’ independent learning ability. Firstly, the design of the system functional modules is explained in detail, including the user registration and login module, the three types of user modules for teachers, students, and administrators, as well as the FAQ and Q&A list module. In addition to the database design required for the system. Then the TF-IDF is improved to work on all the words divided after the word division work as the feature items of the text after vectorization, and the concept of weight factor is introduced to ensure that the weights calculated traditionally still occupy a large proportion in the new weight calculation method, and at the same time, different weights are assigned to the specialized and non-specialized vocabularies. The system is applied to English teaching in a school to analyze the accuracy of question-answering and the effects of independent learning.

2

Design of an intelligent question-and-answer system for English language education

2.1

Module Functional Design

The English Teaching Intelligent Q&A System consists of three parts: the student user module, the teacher user module, and the administrator user module.For student users, they first register as users and then log in to the system. After entering the system, it will display three option modules: Q&A list, My questions, and I want to ask a question. When a student clicks on the My Questions module, all the questions and answers asked by the student will be displayed; the student can ask questions through the I Want to Ask module and wait for the teacher or the system to give an answer. For the teacher user, who also registers before logging in to the system, the teacher user includes two options, the Q&A list and the Waiting for Answers module. Teachers use the Waiting for Answers module to answer questions from students that cannot be answered by the system. The Q&A list displays questions and answers from all users who have accessed the system.The administrator user module is made up of two options modules: user management and user deletion.The system’s overall functional modules are depicted in Figure 1.

The system is designed using a modular design approach.

2.1.1

User registration and login module

1) User Registration Before users can use the system to answer questions, they first need to register as users on the Q&A website. The process of user registration is to log in to the website first, and then click on the user registration link to enter the registration page.

2) User Login When the user is successfully registered, the registered information can be used for user login.

2.1.2

Student user module

The student user module mainly consists of four sub-functional modules: Q&A list, FAQ, I want to ask a question, and My questions.

1) My QuestionsAfter logging into the system, students can view all the questions the student has asked by clicking the My Questions button, and the information is displayed on the page. The flowchart is shown in Figure 2:

2) I want to ask questions This module is the key technology of the intelligent question-answering system, students can ask questions through the I want to ask questions and get answers to questions.

2.1.3

Teacher user module

The Teacher User Module mainly consists of two sub-functional modules: Q&A List and Waiting for Answers. The main flow is as follows:

1) After logging in as a teacher, click the Waiting for Answer button on the top left corner of the page on the home page;

2) After clicking it, the backend will request to connect to the pending answer table in the database; the unanswered questions read will be fed back to the frontend and the page will be jumped;

3) The teacher selects the question to be answered and enters the answer page;

4) The teacher needs to enter the exact answer according to the pending question and subsequently submit the answer to realize the command of the system to update the answer;

5) After receiving the answer update submitted by the teacher, the answering system stores the updated question into the database together with the corresponding answer and jumps back to the Q&A list page.

The waiting answer flowchart is shown in Figure 3:

2.1.4

Administrator user module

The Administrator User module is mainly for managing accounts and is divided into two categories: user management and account deletion.User management enables you to modify user passwords and related attribute information. For accounts that are no longer in use, administrators can delete them by deleting users.

2.1.5

Q&A list module

When logging in to the system, users with teacher or student privileges will be able to display the Q&A list option button in the top left corner of the home page. Once clicked, all the questions asked by users who have accessed the system and the corresponding answers are displayed on the page.

The Q&A list flowchart is shown in Figure 4:

2.1.6

Database design

Database design is the creation of a database that matches the application system to which it corresponds, and is a top priority in the overall development of information.

Database design is mainly based on the following two main principles:

1) the logical design of the database is a mathematical modeling of data, a step in the establishment of the data model, this design and the database itself has no connection, but simply reflects the demand for the operation of the logical relationships, including all the actual use and interrelationships, belonging to the specific business and entities and relationships between the articulation of the correspondence.

2) The physical design of the database is based on the former part of the logical design decided in the physical design step of the two options, namely, relational database and object-oriented database, and further determine the database table structure in line with it. In the market, relational databases account for the vast majority.

2.2

Intelligent Q&A System Technical Preparation

2.2.1

Textual segmentation techniques

To achieve the natural semantic understanding of text, text segmentation is the first link to be solved, which is the prerequisite for the realization of upper-level processing. Chinese word separation techniques can be roughly divided into the following types:

1)

Statistical Text Segmentation Algorithm

The principle of its realization is based on the statistics of phrases. If two Chinese characters are presented together more frequently, it can show that they may be a word. Then, in the practical application, we can build a phrase knowledge base to analyze the keywords, and at the same time, we can use the automatic learning of new words, and then update them into the knowledge base. For example, “you, we, he said, this way, may”, etc., these words are frequently used in daily life, so they can be separated out during the semantic analysis.

2)

Text Segmentation Algorithm Based on String Matching

The technique of text segmentation is mainly to use and design a certain analysis strategy to analyze, and then based on these strategies to analyze the words and has been set to match the dictionary in the the thesaurus, in the process of matching, such as if there is a successful match or the matching degree in a certain value above the ratio, can be defined as a successful match, or else it is a match failure [23]. A few of the more commonly used mechanical segmentation methods are as follows:

(1) Forward Maximum Matching Algorithm

Forward Maximum Matching Method Its key design concept is to assume that maxlen is used to represent the maximum word length, and then in accordance with the order from left to right, in turn, from the text string to take out the substring whose size is maxlen, check the dictionary library, if there is such a word string in the dictionary library, the substring will be split, and then move backward by maxlen text after the next split.

(2) Reverse Maximum Matching Method

It is also called RMM method, the main design idea of RMM method is similar to MM method, the difference between them is that the direction of segmentation is opposite to the latter, and there is also a difference in the method of matching the thesaurus. The realization method is to be solved first from the end of the text to match the scanning, each time the end of the last 2n character is considered to be a matching field, such as if there is no successful match, then it is necessary to delete the first word of the field being matched, and then proceed to match.

(3) Minimum cut method: to minimize the number of words split out of each sentence.

(4) Bidirectional matching method: Forward maximum matching method and reverse maximum matching method are combined to split words. Firstly, the text should be initially roughly segmented according to the position of the punctuation marks, and the text is cut into several sub-sentences, and then the MM and RMM methods are used to scan and segment the sentences respectively.

The inverse maximum matching algorithm is used in this thesis.

2.2.2

Statement similarity algorithm

The most crucial technique in the study of natural language processing of text is the study of utterance similarity generally using some method to derive the value of similarity between the sentence and the clause, generally taking the value between zero and one [24]. If this value is larger, it means that the sentence and clause have a similar meaning in some way or another. They may be similar in word type or in semantic meaning. Based on the analysis of domestic and international research, there are several commonly used utterance similarity algorithms:

1)

TFIDF method based on vector space modeling

Assumption n is used to denote the total number of statements in the above statements, m is the number of statements contained in the question bank system, and M is the total number of statements contained in the question bank, which leads to M = {m₁,m₂,m₃,⋯m_n}, m_i ∈ M. Frequently used similarity algorithms are inner product method, dice coefficient, cosine coefficient, jaekard coefficient, etc., among them cosine coefficient is the more applied algorithm: (1) $s i m (a, m_{i}) = \cos (θ) = \frac{\sum_{j = 1}^{n} w_{n j}, w_{i j}}{\sqrt{\sum_{j = 1}^{n} w_{n j^{2}} \cdot \sum_{j = 1}^{n} w_{i j^{2}}}}$

Where, sim(a,m_i) is the similarity between a and m_i, e is the angle between vectors a and m_i, w_aj is the weight of word t_j in a, and W_ij is the weight of word t_j in m_i.

2)

Based on the semantic similarity calculation method

Assuming that the semantic distance between two words A and B is x, according to Prof. Wang Yang’s research, that is to say, the semantic similarity between A and B can be expressed as Eq: (2) $S (a, b) {\begin{matrix} H - X * \frac{(H - L)}{D} & P \neq \infty \\ 0 & P = \infty \end{matrix}$

There are two sentences E, F. If the words contained in these sentences are E₁,E₂,E₃⋯E_M and F₁,F₂,F₃⋯,F_N, then the following matrix formula can be obtained: (3) $S (E, F) = [\begin{array}{l} S (E_{1}, F_{1}), S (E_{1}, F_{2}), \dots \dots, & S (E_{1}, F_{N}) \\ S (E_{m}, F_{1}), S (E_{m}, F_{2}), \dots \dots, & S (E_{m}, F_{n}) \end{array}]$

Equation (4) can be obtained using matrix (3): (4) $S (E, F) = \frac{\sum_{i = 1}^{m} \max (S (A_{1}, B_{1}) S (A_{1}, B_{2}) \dots S (A_{1}, B_{N}))}{M}$

The similarity of two statements that need to be compared and matched can be found by using equations (2) to (4) respectively.

In order to make the algorithm as time efficient as possible, this paper uses the improved TF-IDF algorithm.

3

Intelligent Q&A algorithm design

In this section, we work on weight calculation, semantic similarity calculation, and answer extraction for the input questions.

3.1

Text vectorization

The first task to complete is the vectorization of the input text. Assuming input text T, it is determined that there are n different feature terms {t₁,t₂,……,t_n} within that text, which are represented in the text as: (5) $T {t_{1}, t_{2}, \dots \dots, t_{n}}$

In Eq. (5), each feature term t_i corresponds to its corresponding weight ω_i in the whole text T. Putting it into a n-dimensional coordinate system, feature term t_i corresponds to the coordinate axes of the coordinate system, while weight ω_i represents the coordinate values on each axis, so that the text T can be represented in a spatial coordinate system.

(6)

\vec{V_{T}} = (ω_{1}, ω_{2}, \dots \dots, ω_{n})

The $\vec{V_{T}}$ in Eq. represents the eigenvector of text T.

Combining Eq. (5) and Eq. (6), the vector matrix of text T can be derived as: (7) $T = [\begin{array}{l} \vec{V_{1}} \\ \vec{V_{2}} \\ ⋮ \\ \vec{V_{n}} \end{array}] = (t_{1}, t_{2}, \dots \dots, t_{n}) = (\begin{matrix} ω_{11} & \dots & ω_{1 n} \\ ⋮ & ⋱ & ⋮ \\ ω_{n 1} & \dots & ω_{n n} \end{matrix})$

After completing the text vectorization, what we need to proceed is to compute the weights of the feature terms.

3.2

Improvement of TF-IDF weight calculation method

3.2.1

Traditional TF-IDF weighting method

When we carry out the calculation in the weight of feature items, we aim at the characteristics that this system is a question-answering system, and the length of the input questions is not too long in general, so we propose to work on all the words that have been divided out by the word-splitting work after vectorization as the feature items of the text.

Among the selected feature items, we first calculate the weights of their feature items [25]. In calculating the feature term weights, the formula for calculating feature term weights in the TF-IDF weight calculation method is: (8) $T F - I D F (ω_{i}) = T F (ω_{i}) \times I D F (ω_{i}) = n_{i, k} \times \log \frac{N}{N_{i, k} + α}$

In the formula, although it has been more mature, there are still some defects: the algorithm only takes into account the two factors of word frequency and inverse text frequency, but in actual use, there may be a situation in which the word frequency of several nouns is the same, in which case the weight of specialized nouns is obviously higher than that of the commonly used nouns, but this cannot be reflected in the algorithm. Based on the fact that there may be different types of nouns with the same weight in the weight calculation method, this paper assigns different weights to different words to further refine the results calculated by the TF-IDF algorithm.

3.2.2

Improvement of TF-IDF weight calculation method

In this paper, the concept of weight factor is introduced here: the weight ω_i calculated by the traditional TF-IDF algorithm is assigned to weight α₁, and the lexical weight ω_p is assigned to weight α₂ The improved weight calculation formula is: (9) $ω_{n e w} = α_{1} \times ω_{i} + α_{2} \times ω_{p}$

Where, make α₁ + α₂ = 1, and assign α₁ = 0.6, α₂ = 0.4, so as to ensure that the weights calculated traditionally still occupy a large proportion in the new weight calculation method, and at the same time, the specialized and non-specialized words will be given different weights to increase the rationality of its weight calculation.

After the weight factor is introduced, we can further analyze the accuracy of weight calculation: when the weight ω₁ of the three words “stack”, “queue” and “difference” are all 0.3 after being calculated by the traditional TF-IDF algorithm, if the similarity calculation is carried out directly, the system will treat “stack”, “queue” and “difference” as equally important words when processing, but it is obvious that the importance of stack and queue is better than difference. After the improved TF-IDF algorithm, the weights of “stack” and “queue” are: (10) $ω_{n e w} = α_{1} \times ω_{i} + α_{2} \times ω_{p} = 0.6 \times 0.3 + 0.4 \times 0.6 = 0.42$

The word “distinction” is weighted: (11) $ω_{n e w} = α_{1} \times ω_{i} + α_{2} \times ω_{p} = 0.6 \times 0.3 + 0.4 \times 0.4 = 0.34$

Through this example we can see that the improved TF-IDF weight calculation method proposed in this paper improves the accuracy of the calculated weights to a certain extent, and next we will further analyze the reasonableness of the weight settings.

3.2.3

Analysis of the rationality of the improved TF-IDF weight calculation method

In this part, we will analyze the rationality of the proposed improved TF-IDF weight calculation method in terms of the weight assignment method and the rationality of the assignment size.

1)

Weight assignment method

At present, the commonly used weight assignment methods include two comprehensive weight calculation methods: additive synthesis method and multiplicative synthesis method. In this paper, the additive synthesis method is used to calculate the weights: the subjective weight vector ω_p and the objective weight vector ω_i, which are assigned on the basis of different weights, are directly added together under a certain weight distribution (α₁,α₂), which is widely used in the rating of the enterprise scale and the project evaluation indexes, but has not been introduced into the calculation of the similarity degree;

2)

Assignment size rationality

The objective weight vector ω_i calculated by the traditional algorithm is assigned a weight size of 0.6 to ensure that it is still dominant in the improved algorithm; the introduced lexical subjective weight vector ω_p is assigned a weight size of 0.4 to ensure that it has an impact in the improved algorithm but will not affect the results when the gap is too large;

The weights assigned to ω_p, 0.6 for specialized words and 0.4 for non-specialized words, are based on the current allocation of weights between the two factors commonly used in similarity calculations.

3.3

Problem similarity calculation

On the basis of calculating the weight of each feature term, it is necessary to calculate the similarity between it and the problem in the database. In this system, the cosine similarity calculation method is used: if the smaller the angle between the calculated feature vectors representing the two feature terms, the larger the cosine value between the two vectors, the larger the correlation between the two vectors, i.e., the higher the similarity between the two. On this basis, the similarity between text T₁, T₂ is calculated as: (12) $S I M (T_{1}, T_{2}) = \frac{ω_{1 k} \times ω_{2 k}}{‖ ω_{1 k} ‖ ‖ ω_{2 k} ‖} = \frac{\sum_{k = 1}^{n} ω_{1 k} \times ω_{2 k}}{\sqrt{(\sum_{k = 1}^{n} ω_{1 k}^{2}) (\sum_{k = 1}^{n} ω_{2 k}^{2})}}$

In Eq. (12), T₁, T₂ are the two texts to be compared, and ω_1k, ω_2k are the weights of the corresponding vectors t₁, t₂. In the calculated results, if the result is closer to 1, it means that the angle between the two vectors is closer to 0°, and the similarity between the two vectors is greater; if the result is closer to 0, it means that the angle between the two vectors is closer to 180°, and the similarity between the two vectors is less.

3.4

Answer Extraction

After completing the calculation of the similarity of the questions, the next work is to select some of the corresponding answers from the many similar questions calculated and feedback them to the students to complete the question-answering work of the students. In this work process, the main purpose is to select the best answer by comparing the similarity between each question. The process of extracting answers is roughly divided into two points:

1) If there are relatively few similar questions, then according to the similarity calculated by the system, all of them will be fed back to the students to complete the Q&A work for the students after being sorted according to the similarity level;

2) If there are more similar questions, a threshold is set (after testing through experiments, the value is set to 10), and when the similar questions exceed the threshold, only the top ten questions are selected to provide feedback for students.

After completing the answer extraction, the whole process of automatic Q&A is completed. The flowchart of automatic Q&A is shown in Figure 5.

4

Dynamic modeling of user questions

The model was applied to the teaching of English in a school and student questioning data was collected and analyzed.

4.1

Establishment of the questioning model

Hidden Markov models include general stochastic processes and Markov chains, with hidden states and observation sequences to describe the stochastic process. Transitions of states in a Markov chain are described in terms of transfer probabilities. Denote the Hidden Markov Model by the quintuple β = (M,Q,N,P,G).

1) M = {M₁,M₂,M₂⋯M₂} denotes that there are x hidden states in the model, M is the state, and x is the number;

2) Q = {Q₁,Q₂,Q₂⋯Q₂} denotes the possible observations of the state. y denotes the maximum number of possible observations;

3) N is a probability distribution, and N = {N} denotes the probability distribution of all initial states.

4) P is a probability matrix, P = {P_ij}, denoting the probability of a transfer from state M_i to state M_j at moment t;

5) G is a probability matrix, G = {G_ij} denoting the probability of observing an observation Q in state M_i at a given moment.

Hidden Markov models are used in intelligent question answering systems, by defining rules to make the constructed model conform to the Hidden Markov model conditions. This paper proposes a Hidden Markov Question Prediction Model, and the model construction steps are as follows:

1) Preliminarily construct the Markov model, and clarify the meaning of the state, observation values represent;

2) Collect historical data, statistic or define the initial state probability distribution and the probability distribution of the observation value in a state:

3) Calculate the transfer probability of a state based on historical data;

4) Debugging the parameters according to the algorithm, so that the model tends to stabilize;

5) Use the stabilized model for predictive analysis.

According to the construction steps of the questioning model, it is now stipulated that the hidden state is the classification of the question, that is, the four major classification types proposed in this paper: the observation value is expressed as a specific question asked by the student, and the knowledge point it belongs to can be identified according to the question labeling: the state transfer is described as the migration of the classification type of the question asked by the student, for example, the question asked in a certain moment from the “simple program derivation” to the “complex program derivation” conversion. In this paper, some concepts of the Hidden Markov Question Prediction Model are defined as follows:

Definition One Question Initial State Distribution Probability (N = {N_i}): Expressed as the probability that the question a student chooses to ask in the first instance is of type N_i.

It is calculated as: (13) $N_{i} = \frac{N u m b e r o f s t u d e n t s w h o s e f i r s t q u e s t i o n w a s o f t y p e M_{i}}{T o t a l q u e s t i o n e r s}$

Definition II Questioning State Transfer Probability (P = {P_i}): Expressed as the likelihood that a student’s questioning type will be transferred from M_i → M_j.

It is calculated as: (14) $P_{i j} = \frac{A s k t h e n u m b e r o f s t u d e n t s f r o m M_{i} \to M_{j}}{N u m b e r o f s t u d e n t s w h o w i l l a s k t h e n e x t q u e s t i o n}$

Definition III Probability Distribution of Observations (G = {G_ij}): Expressed as the probability that a student asks a question to Knowledge Point Q_j in Type M_i.

Calculated as: (15) $G_{i j} = \frac{\begin{matrix} T h e n u m b e r o f Q_{i} k n o w l e d g e p o int s \\ i n a l l q u e r i e s M_{i} i n a \sin g l e q u e s t i o n i n g s e s s i o n \end{matrix}}{N u m b e r o f q u e s t i o n s a s k e d i n a q u e s t i o n}$

In the model, the hidden state (M) and observation (Q) can be used as known conditions, by counting the questions asked by a number of students and the order in which they are asked, their question sequences are obtained, and these question sequences are counted and computed to obtain the probability values of N, P, G, and so on.

4.2

Generation of Questioning Model Parameters

In this paper, a sequence set of 100 students’ questions was collected, and the sequence set was preprocessed to remove irrelevant questions and uncollected questions from the question bank, and the authors who wrote the book were asked to classify the importance of each knowledge point under each of the four major classifications in this paper, and this weighted value is the observation value (G), and the probability distribution is shown in Table 1.

Table 1.

Observed probability matrix

Observation value probability matrix	1	2	3	4	5	6	7
A	0.06	0.21	0.26	0.17	0.14	0.16	0
B	0.11	0.12	0.23	0.22	0.32	0	0
C	0.11	0.16	0.13	0.22	0.17	0.21	0
D	0.03	0.04	0.09	0.11	0.15	0.21	0.37

The question classification to which the first question in the initial state belonged was counted for those 50 sets of questions. It is calculated that there are 24 students whose first question belongs to category A, 16 students whose first question belongs to category B, 8 students whose first question belongs to category C, and 2 students whose first question belongs to category D. Therefore, the initial state probability distribution is: N={0.52,0.30,0.14,0.04}. From Definition II, the state transfer probability can be obtained as shown in Table 2.

Table 2.

State transfer probability matrix

State transfer matrix	A	B	C	D
A	0.16	0.19	0.06	0.59
B	0.17	0.17	0.04	0.62
C	0.06	0.08	0.21	0.65
D	0.08	0.21	0.12	0.59

4.3

Prediction of student questioning

Based on the above parameters, the stabilized Markov question prediction model is obtained, as shown in Figure 6:

From the calculated state transfer probabilities and observation probabilities, the prediction process for the next question asked by the student is shown in Figure 7.

Based on the prediction process provided above, it is easy to calculate that after the fourth questioning, if the next questioning is conducted, it is most likely to ask the 7th knowledge point under the D classification.

Therefore, the 7th knowledge point under the D classification can be used as a recommended question to be fed back to the student. The stabilized Hidden Markov Model can not only be used to realize the prediction of students’ questions, but also to determine the possible knowledge points of the next question of the student according to the current question of the student, and select the top 5 questions under the knowledge point and return them to the student, which is recommended as a predicted question to the student. In addition, according to the user’s question sequence, it is easy to find out the transition between the hidden states, i.e., the degree of relevance of the knowledge point and the knowledge point articulation. The Viterbi algorithm is to explain the Hidden Markov Model prediction problem using dynamic programming, i.e., maximizing the path of probability using dynamic programming.

Viterbi algorithm:

Inputs: model γ = (P,G,N) and observation Q = (q₁,q₂,⋯,q_T);

Output: optimal path $I^{*} = (i_{1}^{*}, i_{2}^{*}, \dots, i_{T}^{*})$ .

1) Initialization (16) $δ_{1} (i) = π_{i} b_{i} (o_{1}), i = 1, 2, \dots, N$ (17) $ψ_{1} (i) = 0, i = 1, 2, \dots, N$

2) Recursion. For t = 2,3,⋯,T (18) $δ_{t} (i) = \max_{1 \leq j \leq N} [δ_{t - 1} (j) a_{j i}] b_{i} (O_{t}), i = 1, 2, \dots, N$ (19) $ψ_{t} (i) = \arg \max_{1 \leq j \leq N} [δ_{t - 1} (j) a_{j i}], i = 1, 2, \dots, N$

3) Termination (20) $P^{*} = \max_{1 \leq i \leq N} δ_{T} (i)$ (21) $i_{T}^{*} = \underset{1 \leq i \leq N}{\arg} [δ_{T} (i)]$

4) Optimal path backtracking for t = T–1,T–2,⋯,1 (22) $i_{t}^{*} = ψ_{t + 1} (i_{t + 1}^{*})$

Find the optimal path $I^{*} = (i_{1}^{*}, i_{2}^{*}, \dots, i_{T}^{*})$ .

For example, a student asks the following question: 1_7→3_6→4_2. Indicates that the student asked question 7 of the first knowledge point, question 6 of the third knowledge point, and question 2 of the fourth knowledge point. Now find what type of knowledge conversion is most likely to be involved in the student’s question, with the path shown in Figure 8.

As can be seen in Figure 8, three consecutive questions to the student’s

At t = 3, the optimal endpoint $i_{3}^{*} = \arg \max [δ_{3} (i)] = D$

From the optimal path endpoint $i_{3}^{*}$ , reverse to find $i_{2}^{*}$ , $i_{1}^{*}$ ;

At t = 2, $i_{2}^{*} = ψ_{3} (i_{3}^{*}) = D$

At t = 1, $i_{1}^{*} = ψ_{2} (i_{2}^{*}) = B$

Thus, the optimal path is found, and the optimal state sequence $I^{*} = (i_{1}^{*}, i_{2}^{*}, i_{3}^{*}) = (B, D, D)$ . i.e., the student most likely asked questions about the knowledge of category B, category D, and category D respectively.

According to the sequence of student errors, to find out the knowledge conversion process in the students difficult to understand the place, which has a guiding role in the division of the content of the teacher’s lecture focus. At the same time, according to the size of the conversion probability of the hidden state of the question, the side of the verification of which knowledge is more closely connected, which has a significant reference value for the teacher’s teaching content development.

5

Application and analysis of intelligent question-answering systems

5.1

System applications

In this paper, 10 members of a university’s Sino-British cooperative project team were asked to use and test the system, and the process and results of the experiment can be used as a measure of the reliability and effectiveness of this intelligent system.

5.1.1

System Q&A accuracy test

Different students logged in to test the system and the test results are shown in Table 3. As can be seen from the table, using the intelligent module of this system for answering questions about the knowledge of Fundamentals of Grammar, the average checking accuracy rate is more than 90%, and the checking completeness rate is 100%, and the module is able to satisfy the users well in answering questions about the conceptual knowledge and formulas and theorems.

Table 3.

Check rate test

Student	Question number	Number of correct answers	Accuracy ratio	Check rate	Question style
101	35	32	91.43%	100%	Natural language
102	24	23	95.83%	100%	Natural language
103	29	28	96.55%	100%	Natural language
104	28	26	92.86%	100%	Natural language
105	27	25	92.59%	100%	Natural language

5.1.2

Testing the effectiveness of systematic question prediction

According to the established Markov question prediction model, the statistical analysis of the question asked by the five students, this paper sets the system can predict the next question asked by students in a knowledge classification of a knowledge point in the question, the question and the next question asked by the students have a similarity of 80%, the question is an accurate prediction of the question asked by the students.

Statistics of the above five students’ questions and students’ next question and question prediction analysis and comparison results, as shown in Table 4. Different students logged in to test the system, and for each student question, the system gives 5 predictions for the next question.

Table 4.

Question prediction

Student	Question number	Accurately predict the number of questions	ACC
101	35	32	91.43%
102	37	34	91.89%
103	29	28	96.55%
104	32	32	100.00%
105	35	33	94.29%

It can be seen that using the intelligent module of this system for answering questions about the knowledge of Fundamentals of Grammar, the prediction of students’ questions, the average accuracy of question prediction is more than 90%, which can greatly improve the efficiency of the system’s question-answering.

5.2

Evaluation of English Independent Learning Benefits

When students use the Smart Q&A system for distance learning, the system’s back-end database processes data such as the length of online learning, the number of questions asked, and the content of the questions. We also recorded the students’ quiz scores at the end of the five sessions. Therefore, this section evaluates students’ learning gains in terms of their online learning engagement and quiz scores.

5.2.1

Analysis of Student Engagement in Online Learning

In this subsection, we will evaluate students’ online learning engagement by analyzing the data and content analysis of online learning hours, number of questions asked, and question content.First, we will provide an overall description of the online learning hours for the experimental and control classes using box plots.Box plots can better characterize the range variation of the whole data, and the box plots of online learning hours of the experimental and control classes are shown in Figure 9.

Online learning duration is an important indicator of students’ online learning engagement. In terms of the overall performance of the five courses, the length of study in the experimental class is much higher than that in the control class. Therefore, it can be shown that the intelligent question-answering system increases the breadth of students’ online learning engagement to a certain extent.

Second, we analyzed the number of questions asked by the students. The students in the experimental class used the intelligent question-answering system to ask questions, whereas the students in the control class used email to ask the teacher. By counting the number of students’ questions and teachers’ emails recorded in the system, a graph comparing the number of questions asked by the experimental class and the control class was produced, as shown in Figure 10.

Overall, the number of questions asked by the experimental class is much higher than that of the control class, which indicates that the smart Q&A system stimulates and encourages students to ask more questions. Therefore, it can be shown that the smart Q&A system increases the frequency of students’ online learning participation to a certain extent. However, some of the questions asked by the experimental class were questions unrelated to the course, such as “What is your name” and “Hello”. In contrast, although the control class asked fewer questions, most of the questions asked were highly relevant to the course. Therefore, a screening or reward/punishment mechanism for the Q&A system to help students focus more on the course itself is one of the directions that the Q&A system needs to consider.

5.2.2

Analysis of student test scores

In this experiment, a total of five course quizzes were conducted, and the full score of each quiz was 20 points, and we used the independent samples t-test to analyze the difference between the scores of the two classes, and the analysis of the scores of the course quizzes of the experimental class and the control class is shown in Table 5.

Table 5.

Analysis of course test results

Courses	Laboratory class		Cross-reference class		T	P
Courses	Mean	variance	Mean	variance	T	P
Course 1	14.35	3.08	13.15	4.48	2.068	.044*
Course 2	15.28	2.64	14.48	3.85	1.796	0.075
Course 3	14.37	3.65	12.69	4.49	2.094	.038*
Course 4	14.89	2.87	14.82	2.31	-0.038	0.92
Course 5	12.26	4.83	12.12	4.77	0.106	0.916

In Course 1 and Course 3, the observed values of t-statistics are sig(2-tailed)=0.044 and sig(2-tailed)=0.038, respectively, which are less than the level of significance of 0.05, and it can be assumed that there is a significant difference between the performance of the experimental class and the performance of the control class in Course 1 and Course 3. In contrast, in Course 2, Course 4 and Course 5, the t-statistic observations were greater than the significance level of 0.05 and there was no significant difference in performance.

Among the five courses, the experimental class has four courses with higher grade means than the control class, and there is a significant difference in the grades of two courses, so it can be concluded that there is a significant improvement in the academic performance of the experimental class and a significant increase in the independent learning ability of the students after the use of the intelligent question-answering system.

6

Conclusion

In this paper, we design an intelligent question-answering system for English education used to improve students’ independent learning ability, and design experiments to verify the accuracy of question-answering and the effect of students’ application, and draw the following conclusions:

1) The system in this paper can conveniently find out the transition between hidden states, i.e. the degree of correlation of knowledge points and the articulation of knowledge points, based on the sequence of questions asked by the user.

2) Using the intelligent module of this system to answer questions on the knowledge of “Grammar Fundamentals” in the subject of English, the average rate of checking accuracy is more than 90%, and the rate of checking completeness is 100%, and the module is able to satisfy the users’ answers to the conceptual knowledge and formulas and theorems very well.

3) The prediction of students’ questions in the Q&A of Fundamentals of Grammar, the average accuracy of question prediction is over 90%, which verifies that the system in this paper can greatly improve the efficiency of the system in answering questions.

4) In the 5 course quizzes, the experimental class has 4 courses with higher grade means than the control class, and there is a significant difference in the grades of 2 courses.

In conclusion, the intelligent Q&A system in this paper can improve students’ independent learning ability, and the system meets the design expectations.

Język:: Angielski

Częstotliwość wydawania:: 1 razy w roku
Dziedziny czasopisma:: Nauki biologiczne, Nauki biologiczne, inne, Matematyka, Matematyka stosowana, Matematyka ogólna, Fizyka, Fizyka, inne

Kanał RSS czasopisma

The Design of Intelligent Q&A System for English Education Based on Artificial Intelligence Technology and the Cultivation of Students’ Autonomous Learning Ability

Qian Xu

Yuxin Gan

Data publikacji: 21 mar 2025

Otrzymano: 28 paź 2024

Przyjęty: 08 lut 2025

DOI: https://doi.org/10.2478/amns-2025-0625

Słowa kluczoweDisambiguation technique, Utterance similarity, TF-IDF, Intelligent question-answering system

© 2025 Qian Xu et al., published by Sciendo

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

Słowa kluczowe
Disambiguation technique, Utterance similarity, TF-IDF, Intelligent question-answering system