Test paper generation method, device, storage medium and electronic equipment

Abstract

The present application discloses a test paper generation method, device, storage medium and electronic device, which relates to the field of data processing technology, including: processing question feature data and student feature data to obtain derived feature data for identifying student abilities and language feature data for analyzing question difficulty corresponding to target students; inputting question feature data, student feature data, derived feature data and language feature data into a deep learning main model and a linear regression auxiliary model respectively, predicting the matching degree between the target student's ability level and the question difficulty in the deep learning main model, determining a candidate question set corresponding to each target student based on the obtained first matching degree prediction result, and predicting the matching degree between the target student and the question in the linear regression auxiliary model to obtain a second matching degree prediction result; performing matching degree verification on the candidate question set according to the second matching degree prediction result to obtain a target question set. The present application can improve the adaptability of the test paper.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a test paper generation method, device, storage medium and electronic equipment. Background Art

[0002] The generation of examination papers is not only to evaluate students' learning outcomes, but also to improve teaching methods, education quality and service levels. Automated and intelligent examination paper generation systems can further optimize this process, making it more efficient, fair and accurate.

[0003] Currently, the question setter can select questions from the preset question bank and set marks, and then use the tool to input the marked questions in the preset question bank into the preset test paper template for filling in and formatting, and then generate student test papers.

[0004] However, when using this test paper generation method, the test questions are uniformly generated after being set by the test setter, which will result in uneven difficulty and quality of test papers determined by different test setters at different times, which will in turn lead to low adaptability between the generated test papers and students, and even affect students' enthusiasm for learning. Summary of the invention

[0005] In view of this, the present application provides a test paper generation method, device, storage medium and electronic device, the main purpose of which is to improve the current technical problem that the test questions of the existing test papers are uniformly generated after being set by the question setter, which will lead to uneven difficulty and quality of test papers determined by different question setters at different times, and then lead to low adaptability between the generated test papers and students, and even affect students' enthusiasm for learning.

[0006] In a first aspect, the present application provides a test paper generation method, comprising:

[0007] Obtaining the topic characteristic data and student characteristic data of the target students;

[0008] Processing the question feature data and the student feature data to obtain derived feature data for identifying student abilities and language feature data for analyzing question difficulty corresponding to the target student;

[0009] The topic feature data, student feature data, derived feature data and language feature data are respectively input into a deep learning main model and a linear regression auxiliary model, the matching degree between the target student's ability level and the topic difficulty is predicted in the deep learning main model, a set of candidate topics corresponding to each target student is determined based on the obtained first matching degree prediction result, and the matching degree between the target student and the topic is predicted in the linear regression auxiliary model to obtain a second matching degree prediction result; wherein the target student's ability level, the topic difficulty and the matching degree between the student's ability level and the topic difficulty are obtained based on training of different prediction models;

[0010] Performing a matching check on the candidate question set according to the second matching prediction result to obtain a target question set;

[0011] A target test paper is generated based on the test paper condition information and the target question set.

[0012] In a second aspect, the present application provides a test paper generating device, comprising:

[0013] An acquisition module, configured to acquire question feature data and student feature data of a target student;

[0014] A processing module is configured to process the question feature data and the student feature data to obtain the derived feature data for identifying the student's ability and the language feature data for analyzing the difficulty of the question corresponding to the target student;

[0015] The prediction module is configured to input the topic feature data, student feature data, derived feature data and language feature data into a deep learning main model and a linear regression auxiliary model respectively, predict the matching degree between the target student's ability level and the topic difficulty in the deep learning main model, determine a set of candidate topics corresponding to each target student based on the obtained first matching degree prediction result, and predict the matching degree between the target student and the topic in the linear regression auxiliary model to obtain a second matching degree prediction result; wherein the target student's ability level, the topic difficulty and the matching degree between the student's ability level and the topic difficulty are obtained based on training of different prediction models;

[0016] A verification module is configured to perform a matching check on the candidate question set according to the second matching prediction result to obtain a target question set;

[0017] The generating module is configured to generate a target test paper based on the test paper condition information and the target question set.

[0018] In a third aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the test paper generating method of the first aspect.

[0019] In a fourth aspect, the present application provides an electronic device, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor implements the test paper generation method of the first aspect when executing the computer program.

[0020] By means of the above technical scheme, the present application provides a test paper generation method, device, storage medium and electronic device. Compared with the current existing technology, the present application processes the student feature data and the question feature data to obtain the derived feature data and the language feature data, and can analyze the student and question data more deeply on the basis of the student feature data and the question feature data, and then can understand the student's learning and question-solving mode more deeply, and deepen the insight of educational data; by inputting the question feature data, the student feature data, the derived feature data and the language feature data into the deep learning main model and the linear regression auxiliary model respectively, a candidate question set and a second matching degree prediction result are obtained, and on the basis of obtaining the candidate question set, a linear regression model can be introduced to provide a second matching degree score for verification, and a target question set is obtained. It can more accurately match the student's ability and the difficulty of the question, so that the generated test paper is more in line with the needs of the target student, and then the learning efficiency and learning motivation of the target student can be improved; in addition, the test paper generation through multiple models can not only ensure the quality of the generated test paper, but also improve the efficiency of the generated test paper and shorten the test paper generation time when the data volume is large.

[0021] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0023] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0024] Figure 1 A schematic diagram of a process of generating a test paper provided in an embodiment of the present application is shown;

[0025] Figure 2A schematic diagram of a process of generating a test paper provided in an embodiment of the present application is shown;

[0026] Figure 3 A schematic diagram of an example process provided by an embodiment of the present application is shown;

[0027] Figure 4 A schematic diagram of a process of generating a test paper provided in an embodiment of the present application is shown;

[0028] Figure 5 A schematic diagram of a process of generating a test paper provided in an embodiment of the present application is shown;

[0029] Figure 6 A schematic diagram of a process of generating a test paper provided in an embodiment of the present application is shown;

[0030] Figure 7 A schematic diagram of an example process provided by an embodiment of the present application is shown;

[0031] Figure 8 A schematic diagram of the structure of a test paper generating device provided in an embodiment of the present application is shown;

[0032] Fig. 9 A schematic structural diagram of an electronic device provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0033] The embodiments of the present application will be described in more detail below with reference to the accompanying drawings. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.

[0034] In order to improve the current existing technical problem that the test questions of the test paper are uniformly generated after being set by the test setter, which will lead to uneven difficulty and quality of the test papers determined by different test setters at different times, and then lead to low adaptability between the generated test papers and students, and even affect the students' learning enthusiasm. This embodiment provides a test paper generation method, such as Figure 1 As shown, the method includes:

[0035] Step 101: Obtain the question feature data and student feature data of the target student.

[0036] In an embodiment of the present application, the target student may be one or more students who need to take an exam, one or more class students, one or more grade students, one or more school students, etc., and the scope of the target student is not specifically limited herein. Exemplarily, the target student may be a student who needs to take a mock exam in different grades and stages of study, or a student who needs to take a unified exam, or a student who needs to undergo review exam training, and so on. For example, if a unified exam is currently required, the target student may be a student in Class A, and it is necessary to obtain the topic feature data and student feature data corresponding to the entire Class A student body. Correspondingly, if an individual exam is currently required, the target student may be each student in Class A, and it is necessary to obtain the topic feature data and student feature data corresponding to each student in Class A separately.

[0037] For this embodiment, the question feature data may include the length of the question in the test paper, the knowledge points of the question, the question summary, the question type, the number of recorded commentaries on the question, the length of time, etc. Correspondingly, the student feature data may include the time it takes for the student to answer the question, the score per unit time, the correct rate of the student's answering, whether the student subjectively feedbacks the difficulty of the question, the relative score of the history paper, the learning progress, the knowledge points that are good and not good at, the habit of answering questions, etc. In some examples, the question feature data and student feature data of the target student can be obtained through the historical test data of the target student, and can also be obtained through other channels, which are not specifically limited here.

[0038] Step 102: Process the question feature data and the student feature data to obtain the derived feature data for identifying the student's ability and the language feature data for analyzing the difficulty of the question corresponding to the target student.

[0039] In an embodiment of the present application, the derived feature data and the language feature data can be obtained by further processing the student feature data and the topic feature data. Specifically, the derived feature data may include: topic difficulty index, student concentration index, learning progress speed, knowledge point mastery stability, etc. In an embodiment of the present application, creating derived features is a key step to improve the model's prediction ability and understand students' abilities. Correspondingly, the language feature data may include vocabulary richness, text sentiment analysis, keyword frequency-inverse document frequency, topic modeling, grammatical complexity, etc.

[0040] Step 103: Input the question feature data, student feature data, derived feature data and language feature data into the deep learning main model and the linear regression auxiliary model respectively, predict the matching degree between the target student's ability level and the question difficulty in the deep learning main model, determine the set of candidate questions corresponding to each target student based on the first matching degree prediction result, and predict the matching degree between the target student and the question in the linear regression auxiliary model to obtain a second matching degree prediction result.

[0041] Among them, the target student ability level, question difficulty, and the matching degree between student ability level and question difficulty are obtained based on different prediction model training.

[0042] In an embodiment of the present application, the structure of the deep learning main model includes but is not limited to: input layer -> embedding layer -> convolutional neural network (CNN), 32 convolution kernels of size 3 (convolution layer, pooling layer, Flatten layer) -> TransformerEncoder layer (6-layer attention mechanism, 8-head attention, residual connection and normalization) -> task-specific feature extraction layer -> output layer. In some examples, the ability level can be understood as predicting the student's mastery of each knowledge point and the overall ability level based on student characteristics and question interaction information. Correspondingly, the difficulty of the question can be understood as the predicted value of the difficulty of the question and the distribution of knowledge points. Optionally, the specific process of predicting the matching degree of the ability level and the difficulty of the question can be to calculate the adaptability score based on the overall ability level of the student and the difficulty of the question, which is the first matching degree prediction result.

[0043] As an optional method, the candidate question set is screened from the preset question bank according to the first matching degree prediction result. Specifically, the preset question bank is a question bank containing knowledge points of each grade. In order to enrich the content of the preset question bank, a large model can be used to generate questions of each academic stage and grade knowledge point. Exemplarily, a certain number of questions can be generated based on information such as grade, academic stage, subject, question classification, knowledge point, medium difficulty (with preliminary recall), etc. For this embodiment, the second matching degree prediction result is obtained as a continuous value, which indicates the degree of match between the student and a certain test paper or a certain question. The higher the value, the better the match.

[0044] Step 104: Perform a matching check on the candidate question set according to the second matching prediction result to obtain a target question set.

[0045] In the embodiment of the present application, the matching degree verification may be to re-match the candidate question set based on the second matching degree prediction result, and select the target question set from the candidate question set.

[0046] Step 105: Generate a target test paper based on the test paper condition information and the target question set.

[0047] In some examples, the test paper condition information may include, but is not limited to, the test paper length, test paper type, test paper specifications, and the like.

[0048] Compared with the current existing technology, this embodiment processes the student feature data and the question feature data to obtain the derived feature data and the language feature data, and can analyze the student and question data at a deeper level on the basis of the student feature data and the question feature data, and then can understand the student's learning and question-solving mode at a deeper level, and deepen the insight of educational data; by inputting the question feature data, the student feature data, the derived feature data and the language feature data into the deep learning main model and the linear regression auxiliary model respectively, a candidate question set and a second matching degree prediction result are obtained, and on the basis of obtaining the candidate question set, a linear regression model can be introduced to provide a second matching degree score for verification, and a target question set is obtained. It can more accurately match the student's ability and the difficulty of the question, so that the generated test paper is more in line with the needs of the target students, and then the learning efficiency and learning motivation of the target students can be improved; in addition, the test paper generation through multiple models can not only ensure the quality of the generated test paper, but also improve the efficiency of the generated test paper and shorten the test paper generation time when the data volume is large.

[0049] As a refinement and extension of the above embodiment, when acquiring the topic characteristic data and student characteristic data of the target student, the following contents may be included, but are not limited to:

[0050] Exemplarily, if the target student's question feature data and student feature data are obtained through the target student's historical test data, the data table included in the target student's historical test data can be obtained. Specifically, it can include: student answer details table, question table, knowledge point table, knowledge point question association table, question type table, subject table, review recording table, paper related question table, whether the student subjectively feedbacks the question difficulty, student paper score table, textbook catalog table, learning progress table, student ability score, etc. Specifically, the student answer details table can be shown in Table 1 below:

[0051] Table 1

[0052]

[0053] The topic table can be shown in Table 2 below:

[0054] Table 2

[0055]

[0056] Correspondingly, the knowledge point table can be shown in Table 3 below:

[0057] Table 3

[0058]

[0059] The knowledge point question association table can be shown in Table 4 below

[0060] Table 4

[0061]

[0062] Optionally, the question type table may be as shown in Table 5 below:

[0063] Table 5

[0064]

[0065] The chart of accounts can be shown in Table 6 below:

[0066] Table 6

[0067]

[0068] Accordingly, the commentary recording table can be shown in Table 7 below:

[0069] Table 7

[0070]

[0071] The paper related question table can be shown in Table 8 below

[0072] Table 8

[0073]

[0074] Optionally, whether students give subjective feedback on the difficulty of the questions can be shown in Table 9 below:

[0075] Table 9

[0076]

[0077] The student's test score table can be shown in Table 10 below:

[0078] Table 10

[0079]

[0080] Accordingly, the textbook catalogue can be shown in Table 11 below:

[0081] Table 11

[0082]

[0083] The learning progress table can be shown in Table 12 below:

[0084] Table 12

[0085]

[0086] The student ability scores can be shown in Table 13 below:

[0087] Table 13

[0088]

[0089] Furthermore, by further processing the data in Tables 1 to 13, student characteristic data and question characteristic data can be obtained, wherein obtaining the student characteristic data may include the following examples, but is not limited thereto:

[0090] (1) Answering time: The answering time for students 1, 2, and 3 can be: Student 1's answering time: 2024-06-12 10:15:10 - 2024-06-12 10:05:10 = 0.17h; Student 2's answering time: 2024-06-12 10:16:10 - 2024-06-12 10:02:10 = 0.23h; Student 3's answering time: 2024-06-12 10:20:10 -2024-06-12 10:03:10 = 0.28h; further, if Student 1 and Student 2 correspond to Class 1, and Student 3 corresponds to Class 2, then the answering time for students in Class 1 and Class 2 can be: answering time for Class 1: (0.17+0.23) / 2=0.2h; answering time for Class 2: 0.28 / 1=0.28h; further, if Student 1, Student 2, and Student 3 correspond to Grade 1 and School 1, then the answering time for students in Grade 1 and School 1 can be: answering time for Grade 1: (0.17+0.23+0.28) / 3=0.226; answering time for School 1: (0.17+0.23+0.28) / 3=0.226.

[0091] (2) Score per unit time: The score per unit time of Student 1, Class 1 and Grade 1 for Mathematics Multiple-Choice Question 1 can be: Student 1’s score per unit time for Question 1: 5 / 0.17=29.41 points / h; Class 1’s score per unit time for Question 1: (0 / 0.23+5 / 0.28) / 2=8.92 points / h; Grade 1’s score per unit time for Question 1: (5 / 0.17+0 / 0.23+5 / 0.28) / 3=(29.41+0+17.85) / 3=15.76.

[0092] (3) Students’ accuracy rate: The accuracy rate of students 1 and 2 on paper 1 can be: student 1’s accuracy rate on paper 1: 2 / 3 = 0.66; student 1’s accuracy rate on paper 1: 3 / 3 = 0.33; further, if class 1 includes students 1 and 2, the accuracy rate on question 1 in paper 1 can be: class 1’s accuracy rate on question 1 is the number of correct students in the class / the total number of correct students: 2 / 2 = 1; class 2’s accuracy rate on question 1 is 1 / 1 = 1; further, the accuracy rate of grade 1 and school 1 on question 1 in paper 1 can be: grade 1’s accuracy rate on question 1 is 3 / 3 = 1; school 1’s accuracy rate on question 1 is 3 / 3 = 1.

[0093] (4) Whether students give subjective feedback on the difficulty of the questions: Student 1 and Student 2. Whether students give subjective feedback on the difficulty of the questions can be: Student 1: Feedback that Question 2 in Test Paper 1 is difficult; Student 2: Feedback that Question 1 in Test Paper 1 is difficult.

[0094] (5) Relative score of the history paper: The relative score of the history paper can be determined by the formula 1 shown below. In formula 1, Average_Score_i is the average score; Max_Average is the maximum score.

[0095] Relative_Score_i=(Average_Score_i / Max_Average) * 100 (Formula 1)

[0096] Specifically, the relative scores of the history papers of Student 1, Student 2, and Student 3 can be: Student 1's average score in history paper: (90+96) / 2 = 93; Student 2's average score in history paper: (85+98) / 2 =91.5; Student 3's average score in history paper: (86+70) / 2 =78; further, the maximum scores of Student 1, Student 2, and Student 3: Max_Average = max(93, 91.5, 78)=93; Student 1's relative score in history paper: 93 / 93 * 100 = 100; Student 2's relative score in history paper: 91.5 / 93 * 100 = 98.38; Student 3's relative score in history paper: 78 / 93 * 100 = 83.87.

[0097] (6) Learning progress: The learning progress of Student 1, Class 1, and Grade 1 can be: Student 1’s learning progress: 2 / 3 = 0.66; Class 1’s learning progress: number of students who have learned / total number of students: 1 / 2 = 0.5; Grade 1’s learning progress: 2 / 3 = 0.66.

[0098] (7) Knowledge points that students are good at and not good at: Knowledge point 1 includes 1 question, knowledge point 2 includes 1 question, knowledge point 3 includes 2 and 3 questions, and knowledge point 4 includes 3 questions. Student 1 answered questions 1 and 3 correctly, but answered 2 incorrectly. Then, the student’s proficiency in knowledge point 1, knowledge point 2, knowledge point 3, and knowledge point 4 can be as follows: Student 1’s proficiency in knowledge point 1: the accuracy rate is 1, which means he is good at it; Student 1’s proficiency in knowledge point 2: the accuracy rate is 1, which means he is good at it; Student 1’s proficiency in knowledge point 3: the accuracy rate is 0.5, which means he is not good at it; Student 1’s proficiency in knowledge point 4: the accuracy rate is 1, which means he is good at it.

[0099] Accordingly, the acquisition of topic feature data may include the following examples, but is not limited thereto:

[0100] (1) Question length: The question length can be determined as follows: Question 1 length 27; Question 2 length 35; Question 3 length 40. (2) Question knowledge points: The knowledge points of Question 1, Question 2, and Question 3 can be: Question 1 knowledge points: absolute value, trigonometric function; Question 2 knowledge points: straight line; Question 3 knowledge points: straight line, vector. (3) Number and duration of recorded commentary: The number and duration of recorded commentary for Question 1, Question 2, and Question 3 can be: Number and duration of recorded commentary for Question 1: 2, total duration 270s; Number and duration of recorded commentary for Question 2: 1, total duration 30s; Number and duration of recorded commentary for Question 3: 1, total duration 60s.

[0101] In the embodiment of the present application, when predicting the matching degree between the target student's ability level and the difficulty of the question in the deep learning main model, and determining the set of candidate questions corresponding to each target student based on the obtained first matching degree prediction result, the following methods may be used but are not limited to: Figure 2 As shown, the method includes:

[0102] Step 201: Process the question feature data, student feature data, derived feature data and language feature data according to a predetermined data processing architecture to obtain feature interaction information between questions and students.

[0103] In an embodiment of the present application, the predetermined data processing architecture may be a Transformer Encoder architecture. Specifically, the core goal of the shared representation layer in the Transformer architecture is to create a unified feature space so that the language features of the question and the behavioral features of the students can effectively interact and compare within this space. The Transformer architecture, especially its Encoder part, is an ideal choice for achieving this goal because of its powerful sequence modeling capabilities and long-distance dependency learning capabilities. It can capture the contextual information of the question text and the complexity of the student's learning pattern, thereby promoting the model's in-depth understanding of individual learning needs.

[0104] (I) Input of the shared representation layer in the Transformer architecture: topic feature data, student feature data, derived feature data, and language feature data. Exemplarily, the input of the shared representation layer can be an input vector constructed based on the topic feature data, student feature data, derived feature data, and language feature data, which can specifically include the following examples, but are not limited to them. For example, the input vector constructed based on the topic feature data can include but is not limited to: (1) topic feature vector: ((Q_{emb})); (2) topic length: ((len_Q)); (3) knowledge point embedding: ((topic_{emb})); (4) topic summary embedding: ((summary_{emb})); (5) topic type embedding: ((type_{emb})); (6) review quantity and duration embedding: ((reviewCount_{emb}, reviewTime_{emb})). For example, the input vector constructed based on student feature data may include but is not limited to: (1) student feature vector: ((S_{emb})); (2) answering time: ((time_{ans})); (3) score per unit time: ((score_{unit})); (4) accuracy embedding: ((accuracy_{emb})); (5) question difficulty feedback: ((difficulty_{feedback})); (6) average score of historical grades: ((avgScore_{hist})); (7) learning progress embedding: ((progress_{emb})); (8) knowledge point mastery: ((knowledge_{proficiency})); (9) question-solving habit embedding: ((habit_{emb})). For example, the input vector constructed based on the derived feature data may include but is not limited to: (1) Question difficulty index: ((difficulty _{emb})); (2) Student concentration index: ((focus_{emb})); (3) Learning progress speed: (progressSpeed_{emb})); (4) Knowledge point mastery stability: ((stability_{emb})). For example, the input vector constructed based on the language feature data may include but is not limited to: (1) Language feature vector: ((language_{emb})).

[0105] In some examples, the input method of the shared representation layer may include direct concatenated input and separate input followed by fusion.

[0106] Among them, direct concatenation input is to directly concatenate the question feature vector (Q_{emb}) and the student feature vector (S_{emb}) into a long vector as the input of the Transformer Encoder, which can be specifically expressed by the following formula 2:

[0107] Input_{concat} = [Q_{emb}; S_{emb}] (Formula 2)

[0108] Correspondingly, the fusion after input can first process the question and student features independently through the Transformer Encoder, and then fuse the representations of the two through specific operations (such as weighted summation, attention mechanism, etc.) at the high level. It can be specifically expressed by the following formula 3:

[0109] Q_{encoded}=TransformerEncoder(Q_{emb})][S_{encoded}=TransformerEncoder(S_{emb})][SharedRep=Fusion(Q_{encoded},S_{encoded})] (Formula 3)

[0110] (II) The processing content of the shared representation layer in the Transformer architecture: Generate a shared representation vector based on the question feature data, student feature data, derived feature data and language feature data. As an optional method, the layer details in the TransformerEncoder architecture may include but are not limited to the following: (1) Add to the input vector during the position encoding (PE) process to preserve the order information of the elements in the sequence. (2) In the self-attention mechanism, each position in the input sequence can pay attention to the information of other positions. The body can be expressed by the following formula 4, where Q, K, and V are query, key, and value matrices respectively, and d_k is the dimension of the key vector.

[0111] Attention(Q,K,V) = softmax(\frac{QK^T}{\sqrt{d_k}})V (Formula 4)

[0112] (3) In Multi-Head Attention, the self-attention mechanism is extended to multiple parallel "heads" to capture different types of dependencies. (4) In the Feed Forward Network (FFN), after the self-attention layer, a combination of two linear transformations and activation functions is used to further process the information.

[0113] (III) Output of the shared representation layer in the Transformer architecture: a shared representation vector ((SharedRep)) that integrates the interactive information of the question and the student's features. Furthermore, the output of the shared representation layer can be a shared representation vector ((SharedRep)); specifically, the shared representation vector is an advanced feature representation vector that integrates the interactive information of the question and the student's features, and provides input for the subsequent task-specific feature extraction layer. It should be noted that this shared representation vector can better capture the degree of match between the difficulty of the question and the student's ability, and provide strong support for the selection and generation strategy of personalized questions.

[0114] Step 202: Input the feature interaction information into the question prediction model to predict the question difficulty and knowledge point distribution, and obtain the question difficulty prediction value and knowledge point distribution coverage.

[0115] In an embodiment of the present application, the question prediction model can specifically be a part of the task feature extraction layer in the deep learning main model. Specifically, the question prediction model is based on the output result of the shared representation layer in the Transformer architecture, and further extracts the features of characteristics such as question difficulty and knowledge point coverage through a series of fully connected (Dense) layers, and finally outputs the difficulty prediction value and knowledge point distribution. (I) Input of the question prediction model: shared representation vector ((Z)). Specifically, the shared representation vector ((Z)) is used as the input of the question prediction model. The vector combines the question features and student behavior features, and the size is (d_{model}), where (d_{model}) is the hidden layer size of the Transformer. (II) Processing content of the question prediction model: predict the question difficulty value and knowledge point distribution based on the shared representation vector. Specifically including: (1) In the question prediction model, one or more fully connected layers are used to further extract features related to question difficulty and knowledge point distribution in the shared representation vector ((Z)). For example, assume that the first fully connected layer has (d_{hidden}) hidden units with an activation function (such as ReLU), which is specifically expressed by Formula 5 as shown below. In Formula 5, (W_{d_1}) is the weight matrix and (b_{d_1}) is the bias term.

[0116] H_{difficulty} = ReLU(W_{d_1}Z + b_{d_1}) (Formula 5)

[0117] (2) In the question prediction model, the second fully connected layer is used to predict the difficulty value and knowledge point distribution. The difficulty prediction may be a single-value output, which can be calculated by the following formula 6. Correspondingly, the knowledge point distribution may be a softmax function applied to multiple output nodes to obtain a probability distribution, which can be calculated by the following formula 7. In formulas 6 and 7, (W_{d_2}), (b_{d_2}), (W_{kd}), (b_{kd}) are the weights and biases of the corresponding layers, respectively.

[0118] Difficulty_{pred} = W_{d_2}H_{difficulty} + b_{d_2} (Formula 6)

[0119] KnowledgeDist = Softmax(W_{kd}H_{difficulty} + b_{kd}) (Formula 7)

[0120] (III) Output of the question prediction model: difficulty prediction value ((Difficulty_{pred})) and knowledge point distribution ((KnowledgeDist)). (1) Difficulty prediction value ((Difficulty_{pred})): A scalar value indicating the predicted difficulty of the question. (2) Knowledge point distribution ((KnowledgeDist)): A vector, each element of which represents the probability of the corresponding knowledge point being covered.

[0121] Step 203: Input the feature interaction information into the student ability prediction model to predict the knowledge point mastery and student ability, and obtain the target student's mastery of each knowledge point and the student's comprehensive ability.

[0122] Specifically, the student ability prediction model is based on the output of the shared representation layer in the Transformer architecture. Through another fully connected layer, the student's mastery of each knowledge point and the overall ability level are predicted according to the student's characteristics and the interactive information of the question. (I) Input of the student ability prediction model: shared representation vector ((Z)). Specifically, the shared representation vector ((Z)) is used as the input of the question prediction model. The vector combines the question characteristics and the student's behavior characteristics, and the size is (d_{model}), where (d_{model}) is the hidden layer size of the Transformer. (II) Processing content of the student ability prediction model: Predict the student's mastery of each knowledge point and the overall ability level based on the shared representation vector. Specifically, it includes: (1) In the student ability prediction model, a group of fully connected layers focusing on student ability prediction are designed, and features are extracted through a fully connected layer with an activation function. Specifically, it can be expressed by the following formula 8, where (W_{a_1}) and (b_{a_1}) are the parameters of this layer.

[0123] H_{ability} = ReLU(W_{a_1}Z + b_{a_1}) (Formula 8)

[0124] (2) In the student ability prediction model, another fully connected layer is used to predict the student's mastery of each knowledge point and overall ability level. The mastery of the knowledge point can also be processed by the softmax function to obtain the probability distribution. The mastery of the knowledge point can be calculated by the following formula 9:

[0125] KnowledgeProficiency = Softmax(W_{kp}H_{ability} + b_{kp}) (Formula 9)

[0126] Accordingly, the overall capability level may be a single scalar output, which can be calculated using the following formula:

[0127] OverallAbility = W_{oa}H_{ability} + b_{oa} (Formula 10)

[0128] (III) Output of the student ability prediction model: knowledge proficiency and overall ability. (1) Knowledge proficiency: a vector representing the probability of a student mastering each knowledge point. (2) Overall ability: a scalar reflecting the student's comprehensive learning ability.

[0129] Step 204: Input the target student's mastery of each knowledge point, the student's comprehensive ability, the predicted value of question difficulty, and the knowledge point distribution coverage into the prediction sub-model, predict the matching degree between the target student's ability level and the question difficulty, and determine the set of candidate questions corresponding to each target student based on the first matching degree prediction result.

[0130] In an embodiment of the present application, in the process of generating a set of candidate questions, an attention mechanism is used to select the most suitable question for each student. Specifically, the prediction sub-model performs weighted scoring on the questions in the question bank according to the results of the student ability prediction, and selects the set of candidate questions that best suits the student. (I) Input of the prediction sub-model: the output results of the student ability prediction model and the output results of the question prediction model. (1) The results of the student ability prediction model ((OverallAbility), (KnowledgeProficiency)) include: the overall ability level of the student and the degree of mastery of each knowledge point. (2) The results of the question prediction model ((QuestionBank)) include: a set of features such as the difficulty of each question, knowledge point coverage, etc., each of which can be represented as a vector, such as ([Difficulty_i, TopicCoverage_i]), where (i) represents the index of the question.

[0131] Optionally, step 204 may adopt but is not limited to the following methods when executed, including: (ii) the processing content of the prediction sub-model: based on the prediction results of the topic prediction model and the prediction results of the student ability prediction model, a set of candidate topics is selected. Specifically, the content included in step 11 and step 12: Step 11, align the target student's mastery of each knowledge point with the knowledge point distribution coverage to obtain the degree of matching of the topic in terms of knowledge point coverage. Specifically, align the student's knowledge point mastery (KnowledgeProficiency) with the topic's knowledge point coverage (TopicCoverage), and calculate the similarity between the two. It can be measured using cosine similarity or dot product, which can be specifically calculated using formula 11 shown below. In formula 11, Similarity_{ik} represents the degree of matching between the student and the (i)th topic in terms of knowledge point coverage.

[0132] Similarity_{ik}=KnowledgeProficiency \cdot TopicCoverage_i (Formula 11)

[0133] Step 12: Calculate the matching degree of the student's comprehensive ability and the predicted value of the question difficulty to obtain a matching value.

[0134] Specifically, the adaptability score is calculated based on the student's overall ability level (OverallAbility), i.e., the student's comprehensive ability, and the question difficulty (Difficulty_i), i.e., the predicted value of the question difficulty. It can be calculated specifically by Formula 12 as shown below. In Formula 12, (w) is an adjustment factor that balances the proportion of knowledge point matching and difficulty adaptability. The higher the AdaptivityScore_i, the more suitable the question is for the student.

[0135] AdaptivityScore_i = w \times Similarity_{ik} + (1-w) \times (1 - |OverallAbility - Difficulty_i|) (Formula 12)

[0136] (III) Output of the prediction sub-model: a set of candidate questions selected based on the matching value. Specifically, the content included in step 13 is as follows: Step 13, select questions that meet the matching threshold in order of matching values ​​from high to low as the set of candidate questions corresponding to each target student. Specifically, calculate the adaptability score (AdaptivityScore_i) of each question in the question bank, and then sort them in descending order, select the questions with the highest scores to form a set as the set of candidate questions for the personalized test paper of the student. Exemplarily, a predetermined threshold can be set or a certain number of questions can be selected from the high-scoring questions to form the final test paper according to the length requirement of the generated test paper. In order to ensure the diversity and comprehensiveness of the test paper, other factors such as question type and question length can also be considered in the selection process for fine-tuning. Among them, the setting of the predetermined threshold can be dynamically adjusted, and the specific adjustment may include: (1) stratification setting: according to the student ability stratification, different thresholds are set for students of different ability segments. (2) Proportional method: according to the target total difficulty coefficient of the generated test paper (for example, the expected average difficulty of the test paper is equal to the predicted value of the student ability), the difficulty threshold of each question is inferred. (3) Gradual approximation: Use a looser threshold at the beginning, and gradually tighten the threshold as the topic selection process proceeds to ensure that the last topic added can accurately match the remaining personalized needs.

[0137] (IV) Loss function of the prediction sub-model: loss function after the integration of multi-task loss functions. (1) Prediction of student learning ability and question difficulty Prediction is a classification task divided by level, using cross-entropy loss (Cross-EntropyLoss); (2) The adaptive strategy model loss function uses supervised learning to evaluate the quality of the generation strategy (through students' feedback on the generated test papers). A customized loss function can be designed, such as negative log-likelihood (NLL) loss to measure the correctness or adaptability of the generated questions. (3) Multi-task loss function integration can be specifically expressed by Formula 13 as shown below. In Formula 13, \alpha_{*} is the weight of each task loss, which is used to balance the importance of different tasks. It should be noted that these weights are dynamically adjusted according to the training process, using an adaptive weight algorithm. Specifically, an additional optimization process is used to learn the task weights. For example, the task weights are regarded as part of the model parameters and updated together with other parameters through gradient descent.

[0138] L_{total}=\alpha_{ability}L_{ability}+\alpha_{difficulty}L_{difficulty}+\alpha_{adaptation} L_{adaptation} (Formula 13)

[0139] (V) Optimizer of the prediction sub-model: AdamW (1) Optimizer selection AdamW: AdamW is a widely used optimizer. It is a variant of Adam. It solves the incompatibility problem between weight decay and momentum by directly subtracting weight decay when updating weights. (2) Learning rate strategy: Using dynamic learning rate strategy, learning rate decay, and adaptive learning rate algorithm One-Cycle Policy can help the model converge quickly in the early stage of training, and make careful adjustments in the later stage to avoid falling into local minima. (3) Regularization: In order to prevent overfitting, L2 regularization can be combined. It should be noted that, if Figure 3 As shown, step 201 in the embodiment of the present application can be Figure 3 In the deep learning main model in, step 202 can be performed Figure 3 Step 203 can be performed in the question bank model in Figure 3 In the student competency model in , step 204 can be performed Figure 3 The sub-models in are fused.

[0140] Furthermore, when performing the prediction of the matching degree between the target student and the question in the linear regression auxiliary model to obtain the second matching degree prediction result, the following methods may be specifically adopted but are not limited to:

[0141] In the embodiment of the present application, when constructing a linear regression model, the question features and student features and their interaction terms are used as inputs, and the predicted matching score is used as output, which may specifically include but is not limited to: (i) Input of the linear regression model: question feature vector and student feature vector. (1) Question feature vector ((Q)) includes: question length, knowledge point coding (single hot), question type coding (single hot), number of question recording and commenting, duration and other features. (2) Student feature vector ((S)) includes: answering time, score per unit time, accuracy, subjective feedback (binary coding, such as 1 for difficult, 0 for not difficult), relative score of history test paper, learning progress (percentage), proficient knowledge point coding (single hot), not proficient knowledge point coding (single hot), question-solving habit coding, etc. (ii) Processing of the linear regression model: predict the matching degree between students and questions. Constructing a linear regression model can be specifically expressed by the following formula 14, where n_q and n_s are the number of question features and student features respectively, Q_i and S_j represent the (i)th and (j)th eigenvalues ​​in the question feature vector and the student feature vector respectively, Q_{intk} and S_{intk} represent the question feature part and the student feature part in the (k)th interaction feature, \beta_{qi}, \gamma_{sj} and \delta_{intk} are model parameters, corresponding to the question feature coefficient, student feature coefficient and interaction term coefficient respectively, \beta_0 is the intercept term, and \epsilon is the error term, representing the variation not explained by the model.

[0142] MatchScore=\beta_0+\sum_{i=1}^{n_q}\beta_{qi}Q_i+\sum_{j=1}^{n_s}\gamma_{sj}S_j + \sum_{k=1}^{n_{int}}\delta_{intk}(Q_{intk} \times S_{intk}) +\epsilon (Formula 14)

[0143] (III) Output of the linear regression model: the match score between the student and the question. Match score (MatchScore): a continuous value indicating the degree of match between the student and a test paper or a question. The higher the value, the better the match. For this embodiment, when performing a match check on the candidate question set according to the second match prediction result to obtain the target question set, the following methods may be used but are not limited to: Figure 4 As shown, the method includes:

[0144] Step 301: compare and verify the first matching degree prediction result of each question in the candidate question set according to the second matching degree prediction result.

[0145] Exemplarily, the matching result obtained by the linear regression model, namely the second matching result, may be verified with the adaptability scores of the questions in the candidate question set, namely the first matching result.

[0146] Step 302: Select topics whose comparison and verification results meet preset conditions from the candidate topic set to form a target topic set.

[0147] For example, the questions that meet the preset condition may be questions for which the similarity between the first matching result and the second matching result is greater than a predetermined threshold, or questions for which both the first matching result and the second matching result are greater than a predetermined threshold, etc. The specific content of the preset condition is not limited here. It should be noted that step 301 and step 302 may be Figure 3 The model fusion is performed in .

[0148] When processing the question feature data and the student feature data to obtain the derived feature data for identifying the student's ability and the language feature data for analyzing the difficulty of the question corresponding to the target student, the following methods may be used but are not limited to: Figure 5 As shown, the method includes:

[0149] Step 401: normalize the continuous features in the question feature data and the student feature data, and discretize the category features in the question feature data and the student feature data.

[0150] Step 21: Select continuous features and category features from the question feature data and student feature data for separate processing.

[0151] Specifically, if it is a continuous feature, execute step 22, and if it is a categorical feature, execute step 23; Step 22, normalize the continuous features in the question feature data and the student feature data, specifically map the original data to the [0, 1] interval through Min-Max Scaling, which can be calculated by the following formula 15, where X is the original data, (X_{min}) and (X_{max}) are the minimum and maximum values ​​in the data set, respectively.

[0152] X_{norm}=\frac{X-X_{min}}{X_{max}-X_{min}} (Formula 15)

[0153] Exemplarily, the normalization process may include the following examples but is not limited to them: (1) Answering time: Assuming that the answering time of student 1 is the minimum value (X_{min}=0.17h) and the answering time of student 3 is the maximum value (X_{max}=0.28h), then: the normalized answering time of student 1 is: (X_{norm1} = \frac{0.17- 0.17}{0.28 - 0.17} = 0); the normalized answering time of student 2 is: (X_{norm2} = \frac{0.23 -0.17}{0.28 - 0.17} = \frac{0.06}{0.11} \approx 0.545); the normalized answering time of student 3 is: (X_{norm3} = \frac{0.28 - 0.17}{0.28 - 0.17} = 1). (2) Unit time score: Assume that the minimum value of the unit time score is (X_{min}=0) and the maximum value is (X_{max}=30) (this is just an example value and should be determined based on the data). Specifically, if the average unit time score (15.76) is given, it is used as the original data for normalization, which can be: (X_{norm} = \frac{15.76 - 0}{30 - 0} = \frac{15.76}{30} \approx0.525).

[0154] Step 23: Discretize the continuous features in the question feature data and the student feature data, specifically by One-Hot Encoding, to facilitate model understanding. Exemplarily, the discretization process may include the following examples but is not limited to: (1) Question type: Assuming that the question types include single-choice questions, multiple-choice questions, fill-in-the-blank questions, short-answer questions, and cloze questions, for each question type, a dimension can be created, and the encoding includes single-choice questions: [1, 0, 0, 0, 0]; multiple-choice questions: [0, 1, 0, 0, 0]; fill-in-the-blank questions: [0, 0, 1, 0, 0]; short-answer questions: [0, 0, 0, 1, 0]; cloze questions: [0,0, 0, 0, 1].

[0155] Step 402: Fill missing values ​​and identify outliers on the normalized and / or discretized question feature data and student feature data to obtain derived feature data and language feature data.

[0156] Step 24: fill missing values ​​in the normalized and / or discretized question feature data and student feature data. Specifically, the missing data may be processed by filling in the mean, filling in the median, or filling in the model-based prediction.

[0157] Exemplarily, the missing value filling process may include the following examples but is not limited to them: (1) Answering time: Due to recording errors or system failures, the answering time data of some students may be missing, and the filling method can be median filling. Specifically, the median is more resistant to the influence of extreme values ​​than the mean, and is suitable for non-normally distributed data, such as the answering time may be affected by the speed of individual students. (2) Unit time score: If the score calculation is missing due to some reasons (for example, the student did not complete the question), the filling method can be model-based prediction filling. Specifically, a simple regression model is trained to use other complete features (such as answering time, accuracy, etc.) to predict the missing unit time score. (3) Student question accuracy: If the student did not participate in a test or part of the questions, resulting in incomplete accuracy data, the filling method can be mean filling or prediction based on the student's historical performance. Specifically, if the student's historical data is rich enough, the mean of its historical accuracy can be used as the filling value, and the regression model can also be used to predict the missing accuracy based on the student's performance in other tests. (4) Whether students provide subjective feedback on the difficulty of the questions: Students may not provide subjective feedback on the difficulty of all questions. The filling method can be mode filling. Specifically, if most students give consistent feedback on questions of similar difficulty, this feedback can be used to fill in the missing values. (5) Relative scores of history papers: New students or occasional absences may result in missing scores for a certain exam. The filling method can be median filling. Specifically, you can choose based on the distribution of existing data. From the perspective of student performance, the score distribution is relatively uniform, so the median is a better choice. (6) Learning progress: The learning progress records in the learning management system are incomplete. The filling method can be model-based predictive filling. Specifically, the learning progress may be related to factors such as students' historical learning behavior and course participation. A model can be built to predict the missing learning progress.

[0158] Step 25: Perform outlier detection and processing on the filled data. Specifically, the interquartile range (IQR) can be used to identify and process outliers. The identified outliers can be removed, replaced or estimated to ensure data quality.

[0159] In some examples, IQR is the difference between the upper quartile (Q3) and the lower quartile (Q1) in the data set, reflecting the distribution range of the middle 50% of the data, where: Q1 (first quartile): the upper limit of the smallest 25% of all values ​​in the data; Q3 (third quartile): the upper limit of the smallest 75% of all values ​​in the data. Outlier identification in the IQR method includes: lower boundary: Q1 - 1.5 * IQR; upper boundary: Q3 + 1.5 * IQR; specifically, values ​​falling outside these two boundaries are considered outliers.

[0160] Exemplarily, the outlier detection and processing process may include the following examples but is not limited to them: (1) Student answering time: Assuming that we calculated Q1 to be 0.15 hours and Q3 to be 0.30 hours, then the IQR is 0.15 hours. Accordingly, according to the IQR rule, any answering time less than 0.15 - 1.5 * 0.15 = -0.125 hours or greater than 0.30 + 1.5 * 0.15 = 0.525 hours will be considered an outlier.

[0161] Optionally, the acquisition of derived feature data may include the following examples, but is not limited to them: (1) Question difficulty index: The question difficulty index can comprehensively evaluate the difficulty of the question, taking into account the question length, accuracy and average answering time. It can be specifically calculated by Formula 16 as shown below. In Formula 16, D is the question difficulty index, L is the question length (such as the number of words, the number of characters or the estimated reading time), P is the average accuracy (between 0 and 1), T is the average answering time (unit time, such as minutes), w_1, w_2, and w_3 are the weights of length, accuracy and time, respectively, which are adjusted according to actual data and business understanding to ensure that the total is 1.

[0162] D=w_1\timesL+w_2\times(1-P)+w_3\timesT (Formula 16)

[0163] (2) Student concentration: Student concentration can be measured by combining the time it takes to answer questions and the score per unit time to measure the degree of concentration of students during the test. It can be calculated using Formula 17 as shown below. In Formula 17, F is the student concentration index, S is the score per unit time, T is the time it takes to answer questions, P_{avg} is the average accuracy of the student in the subject or knowledge point, and |P_{avg} - P| is the difference between the student's accuracy rate in this test and the average accuracy rate, reflecting volatility.

[0164] F=\frac{S}{T}\times(1-|P_{avg}-P|) (Formula 17)

[0165] (3) Learning progress rate: The learning progress rate can evaluate students’ learning efficiency and progress in a specific time period. It can be calculated using Formula 18 as shown below. In Formula 18, V is the learning progress rate, S_{t1} and S_{t2} are the total scores or average scores of students at two time points (such as two exams), and T_{t1} and T_{t2} correspond to the time point interval (which can be days, weeks, months, etc.).

[0166] V=\frac{(S_{t2}-S_{t1})}{(T_{t2}-T_{t1})} (Formula 18)

[0167] (4) Knowledge point mastery stability: Knowledge point mastery stability can reflect whether students’ mastery of a specific knowledge point is stable. It can be calculated using Formula 19 as shown below. In Formula 19, KSI is the knowledge point mastery stability index, N is the total number of questions attempted by students on this knowledge point, P_i is the accuracy of the ith attempt, P_{avg} is the average accuracy of this knowledge point, and P_{max} and P_{min} are the highest and lowest accuracy rates, which are used to normalize the fluctuation range.

[0168] KSI=\frac{1}{N}\sum_{i=1}^{N}\left(1-\frac{|P_i-P_{avg}|}{P_{max}-P_{min}}\right) (Formula 19)

[0169] Exemplarily, the acquisition of language feature data may include the following examples, but is not limited to them: (1) Lexical richness: Lexical richness can evaluate the diversity of vocabulary usage in a text and reflect the student's language expression ability. It can be specifically calculated using Formula 20 shown below, where LR represents the lexical richness score of each text sample.

[0170] LR=\frac{\text{UniqueWords}}{\text{TotalWords}}\times\frac{\text{TotalWords}}{\text{SentenceCount}} (Formula 20)

[0171] (2) Text sentiment analysis: Text sentiment analysis can use natural language processing technology to determine the sentiment tendency of the text and understand students' attitudes or emotional reactions to the learning content. It can be calculated specifically by using Formula 21 as shown below. In Formula 21, (SentimentScore) is the text sentiment score, w_i is the sentiment weight of word i, Polarity_i is the sentiment polarity of word i (positive, negative or neutral), N is the total number of words in the text, and SentimentScore is the sentiment score of each text, such as -1 to 1, where negative values ​​represent negative emotions and positive values ​​represent positive emotions. It should be noted that Formula 21 is a simplified example and may actually involve complex models, which are not limited here.

[0172] SentimentScore=\sum_{i=1}^{N}w_i\timesPolarity_i (Formula 21)

[0173] (3) Keyword frequency-inverse document frequency: Keyword frequency-inverse document frequency (TF-IDF) can identify which words in a text are the most representative and help distinguish different topics or knowledge points. It can be calculated using Formula 22 as shown below. In Formula 22, TF(w,d) is the word frequency of word w in document d, and IDF(w,D) is the inverse document frequency, which is equal to the total number of documents N divided by the number of documents containing word w, and then taking the logarithm. TFIDF(w,d,D) is the TF-IDF value of each word for each document, which is used to select the keyword that best represents the document.

[0174] TFIDF(w,d,D)=TF(w,d)\timesIDF(w,D) (Formula 22)

[0175] (4) Topic Modeling: Topic modeling can use methods such as Latent Dirichlet Allocation (LDA) to automatically discover latent topics from a large amount of text and obtain the relevance score of each text sample to multiple topics, reflecting the scope and focus of the topics covered by the text.

[0176] (5) Syntactic Complexity: Syntactic Complexity can be evaluated by analyzing sentence structure, such as average sentence length, number of clauses, etc. The grammatical complexity of a text can be calculated using Formula 23 as shown below. In Formula 23, SC represents the grammatical complexity index of the text, which reflects the student's level of use of language structure.

[0177] SC=\frac{\text{Total Clauses}}{\text{Sentence Count}} (Formula 23)

[0178] When generating a target test paper based on the test paper condition information and the target question set, the following methods may be used but are not limited to: Figure 6 As shown, the method includes:

[0179] Step 501: Classify the questions in the target question set based on the question type in the test paper condition information to obtain a target question set corresponding to each question type.

[0180] Step 502: Generate a target test paper based on the test paper length in the test paper condition information and the target question set corresponding to each question type.

[0181] When executing step 501 and step 502, the following methods may be used: Figure 7As shown, it specifically includes: determining the type of questions suitable for students, selecting a target test question set from a label question bank, i.e., a preset question bank, based on the question type, and generating a target test paper according to the test paper length. Among them, determining the type of questions suitable for students may include but is not limited to the following:

[0182] (1) By determining the Pearson correlation coefficient to analyze the correlation between students and question types, the numerator (the sum of the differences between the products of the two variables) and the denominator (the product of the square roots of their respective square differences) in the Pearson correlation coefficient formula can be determined. Specifically, it can be calculated using Formula 24 as shown below. In Formula 24, (x_i) and (y_i) are the two variable values ​​of the (i)th observation, (\bar{x}) and (\bar{y}) are the average values ​​of the two variables, and (n) is the total number of observations.

[0183] r= \frac{\sum_{i=1}^{n}(x_i - \bar{x})(y_i - \bar{y})}{\sqrt{\sum_{i=1}^{n}(x_i - \bar{x})^2}\sqrt{\sum_{i=1}^{n}(y_i - \bar{y})^2}} (Formula 24)

[0184] Exemplarily, the calculation of the Pearson correlation coefficient may include the following examples but is not limited thereto: The relationship between the time spent on answering questions and the accuracy rate: Explore the relationship between the time students spend on questions and their ability to answer questions correctly. In theory, if students spend more time reviewing questions and thinking, the accuracy rate may be higher, but it may also be reduced due to fatigue or distraction. The specific calculation process includes: recording the time it takes each student to answer each question (unit: minutes) and whether the question is correct or not (which can be converted to 0 error / 1 correct to calculate the accuracy rate); calculating the average answering time and average accuracy rate of all questions for each student.

[0185] (2) The correlation analysis between students and question types can be conducted through chi-square test. Specifically, the correlation between question types, knowledge points and students' correct rate can be tested to guide the selection and design of questions. For example, the chi-square test can include the following examples but is not limited to them: Analyze the relationship between single-choice questions, multiple-choice questions and fill-in-the-blank questions and students' correct rate: The existing data includes students' answers to each question type (correct or incorrect), where data preparation includes: dependent variable: question type (A=single-choice question, B=multiple-choice question, C=fill-in-the-blank question); outcome variable: correct (1) or incorrect (0); data table: construct a cross-tabulation to list the correct and incorrect frequencies for each question type, as shown in Table 14 below.

[0186] Table 14

[0187]

[0188] Furthermore, the chi-square test can be specifically calculated by Formula 25 as shown below. In Formula 25, (O) represents the observed frequency, that is, the actual frequency in the above Table 14, and (E) represents the expected frequency. It can be specifically calculated by Formula 26 as shown below. In Formula 26, \sum represents the sum of all cells.

[0189] \chi^2 = \sum \frac{(OE)^2}{E} (Formula 25)

[0190] E = (row total\times column total) / total number of samples (Formula 26)

[0191] Optionally, the output of the chi-square test may include: calculating the expected frequency: for each cell, calculating the expected frequency based on the marginal totals of the rows and columns; calculating the chi-square statistic: using formula twenty-five to calculate the chi-square value; looking up a table or calculating a p-value: based on the degrees of freedom (df = (number of rows - 1) * (number of columns - 1)) and the chi-square value, looking up a chi-square distribution table or using statistical software to calculate the corresponding p-value; making a decision: if the p-value is less than the set significance level (such as 0.05), reject the null hypothesis (i.e., assume that there is a correlation between the question type and the accuracy rate); otherwise, accept the null hypothesis (assume that there is no significant correlation between the two).

[0192] (3) Cluster analysis of students: Use K-means clustering algorithm as the clustering algorithm. Initialization: Select K initial centroids (center points), usually randomly selecting K samples from the data set. Assignment: Assign each sample to the nearest centroid to form K clusters, which can be calculated by the following formula 27. In formula 27, (x_i) is the i-th eigenvalue of the sample, and (c_{ki}) is the i-th eigenvalue of the current centroid of the k-th cluster.

[0193] \text{Distance} = \sqrt{\sum_{i=1}^{n}(x_i - c_{ki})^2} (Formula 27)

[0194] Update the centroid: recalculate the mean (or median) of each cluster and use it as the new centroid. Repeat the steps to assign and update the centroid: until the centroid no longer changes significantly or the preset maximum number of iterations is reached. The output of cluster analysis includes: Cluster division: each student is assigned to a cluster, and the output is a mapping table between the student ID and the cluster to which it belongs; Cluster center: the centroid of each cluster, representing the typical characteristics of the students in that cluster; Sum of squared distances within clusters (WCSS): the sum of squared distances from all points in each cluster to its centroid, which is an indicator for evaluating clustering effectiveness.

[0195] In the embodiment of the present application, the traditional test paper generation method mainly relies on the intuition and experience of education experts or teachers, and manually selects questions and arranges the test paper structure according to the curriculum standards, teaching materials and teaching progress. This process involves comprehensive coverage of knowledge points, reasonable distribution of difficulty levels, and diversified combinations of question types. However, this manual operation mode has obvious limitations and is highly dependent on manual work. It is not only time-consuming and laborious, but also requires teachers to have rich experience and in-depth subject knowledge, which limits the possibility of large-scale and efficient test paper generation. The personal preferences and judgments of teachers may affect the choice of questions, resulting in the fluctuation of the quality of the test paper by subjective factors, and it is difficult to achieve complete objectivity and fairness. Limited to the experience range of teachers, the question ideas and question types are often more traditional, and it is difficult to innovate frequently, which is not conducive to stimulating students' interest and challenging learning. It is impossible to customize the test paper according to the specific learning situation, ability level and progress of each student, resulting in non-personalized allocation of learning resources, affecting learning efficiency and effect.

[0196] Due to its manual nature, the existing test paper generation technology leads to a long time for manual question setting and test paper arrangement, which is difficult to adapt to the rapid response of teaching demand changes and the organization of large-scale examinations, especially in the context of the increasing scale of education and the increasing demand for personalized education. The quality of the test paper is subject to the personal ability and state of the teacher at the time, which may lead to uneven difficulty and quality of test papers in different classes and at different times, affecting educational equity. It is impossible to adjust the difficulty and type of questions according to the students' learning data in real time, which makes it difficult to accurately match the students' learning needs, and is not conducive to the planning and implementation of personalized learning paths. Traditional methods are difficult to frequently introduce new types of questions and technical means, which limits the modernization of educational content and the innovation of educational methods, and is not conducive to cultivating students' high-order thinking and innovation capabilities. The present invention aims to solve the problem that test paper generation in existing teaching scenarios lacks personalization and is difficult to accurately match students' ability levels. Traditional test paper distribution is usually based on a unified standard, ignoring the ability differences between students, resulting in inaccurate evaluation of learning effects and difficulty in effectively stimulating the potential and learning enthusiasm of each student.

[0197] Compared with the current existing technology, this embodiment processes the student feature data and the question feature data to obtain the derived feature data and the language feature data, and can analyze the student and question data at a deeper level on the basis of the student feature data and the question feature data, and then can understand the student's learning and question-solving mode at a deeper level, and deepen the insight of educational data; by inputting the question feature data, the student feature data, the derived feature data and the language feature data into the deep learning main model and the linear regression auxiliary model respectively, a candidate question set and a second matching degree prediction result are obtained, and on the basis of obtaining the candidate question set, a linear regression model can be introduced to provide a second matching degree score for verification, and a target question set is obtained. It can more accurately match the student's ability and the difficulty of the question, so that the generated test paper is more in line with the needs of the target students, and then the learning efficiency and learning motivation of the target students can be improved; in addition, the test paper generation through multiple models can not only ensure the quality of the generated test paper, but also improve the efficiency of the generated test paper and shorten the test paper generation time when the data volume is large.

[0198] It should be noted that deep learning technology can be used to build a main model to capture the complex characteristics of students' abilities, and combined with a linear regression model as an auxiliary tool to accurately predict students' mastery of different knowledge areas. By integrating the preset deep learning model with the preset linear regression model, a highly customized test paper generation engine is formed. The engine can automatically analyze the ability distribution of students and dynamically generate personalized test papers that match them. This not only significantly improves the pertinence and efficiency of teaching evaluation, but also promotes the development of students' personalized learning paths, further enhancing their learning motivation and sense of achievement.

[0199] Further, as Figure 1 and Figure 2 The specific implementation of the method shown in the embodiment provides a test paper generating device, such as Figure 8 As shown, the device includes: an acquisition module 61, a processing module 62, a prediction module 63, a verification module 64, and a generation module 65.

[0200] An acquisition module 61 is configured to acquire the topic characteristic data and the student characteristic data of the target student;

[0201] The processing module 62 is configured to process the question feature data and the student feature data to obtain the derived feature data for identifying the student's ability and the language feature data for analyzing the difficulty of the question corresponding to the target student;

[0202] The prediction module 63 is configured to input the topic feature data, student feature data, derived feature data and language feature data into the deep learning main model and the linear regression auxiliary model respectively, predict the matching degree between the target student's ability level and the topic difficulty in the deep learning main model, determine the candidate topic set corresponding to each target student based on the obtained first matching degree prediction result, and predict the matching degree between the target student and the topic in the linear regression auxiliary model to obtain a second matching degree prediction result; wherein the target student's ability level, the topic difficulty and the matching degree between the student's ability level and the topic difficulty are obtained based on training of different prediction models;

[0203] A verification module 64 is configured to perform a matching verification on the candidate question set according to the second matching prediction result to obtain a target question set;

[0204] The generating module 65 is configured to generate a target test paper based on the test paper condition information and the target question set.

[0205] In some examples of the present embodiment, the prediction module 63 is specifically configured to process the question feature data, student feature data, derived feature data and language feature data according to a predetermined data processing architecture to obtain feature interaction information between the question and the student; input the feature interaction information into the question prediction model to predict the question difficulty and the knowledge point distribution to obtain the predicted value of the question difficulty and the knowledge point distribution coverage; input the feature interaction information into the student ability prediction model to predict the knowledge point mastery and the student ability to obtain the target student's mastery of each knowledge point and the student's comprehensive ability; input the target student's mastery of each knowledge point, the student's comprehensive ability, the predicted value of the question difficulty and the knowledge point distribution coverage into the prediction sub-model to predict the matching degree between the target student's ability level and the question difficulty, and determine the set of candidate questions corresponding to each of the target students based on the first matching degree prediction result obtained.

[0206] In some examples of the present embodiment, the prediction module 63 is further specifically configured to align the target student's mastery of each knowledge point with the knowledge point distribution coverage to obtain the matching degree of the question in the knowledge point coverage; calculate the matching degree of the student's comprehensive ability and the predicted value of the question difficulty to obtain the matching value; and select questions that meet the matching threshold in order from high to low as the candidate question set corresponding to each of the target students.

[0207] In some examples of this embodiment, the verification module 64 is specifically configured to compare and verify the first matching degree prediction result of each question in the candidate question set according to the second matching degree prediction result; and select questions from the candidate question set whose comparison and verification results meet preset conditions to form the target question set.

[0208] In some examples of this embodiment, the processing module 62 is specifically configured to normalize the continuous features in the question feature data and the student feature data, and to discretize the category features in the question feature data and the student feature data; fill in missing values ​​and identify outliers on the normalized and / or discretized question feature data and student feature data to obtain the derived feature data and the language feature data.

[0209] In some examples of this embodiment, the generation module 65 is specifically configured to classify the questions in the target question set based on the question type in the test paper condition information to obtain a target question set corresponding to each question type; and generate a target test paper based on the test paper length in the test paper condition information and the target question set corresponding to each question type.

[0210] It should be noted that for other corresponding descriptions of the functional units involved in the test paper generating device provided in this embodiment, reference can be made to Figure 1 and Figure 2 The corresponding description in will not be repeated here.

[0211] Based on the above Figure 1 to Figure 2 The method shown in the embodiment also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the above-mentioned Figure 1 to Figure 2 The method shown.

[0212] Based on this understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, USB flash drive, mobile hard disk, etc.), including a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods of various implementation scenarios of the present application.

[0213] like Fig. 9 The figure shows a schematic diagram of the hardware structure of an electronic device of the present invention, comprising: at least one processor 701; and a memory 702 communicatively connected to at least one of the processors 701; wherein the memory 702 stores instructions executable by at least one of the processors, and the instructions are executed by at least one of the processors so that at least one of the processors can execute the test paper generation method as described above. Fig. 9 A processor 701 is used as an example. The electronic device may further include: an input device 703 and a display device 704. The processor 701, the memory 702, the input device 703 and the display device 704 may be connected via a bus or other means. Fig. 9 The memory 702 is a non-volatile computer-readable storage medium that can be used to store non-volatile software programs, non-volatile computer executable programs and modules, such as program instructions / modules corresponding to the test paper generation method in the embodiment of the present application, for example, Figure 1 , Figure 2 The method flow shown. The processor 701 executes various functional applications and data processing by running the non-volatile software programs, instructions and modules stored in the memory 702, that is, the test paper generation method in the above embodiment is realized. Optionally, the above-mentioned entity device may also include a user interface, a network interface, a camera, a radio frequency (RF) circuit, a sensor, an audio circuit, a WI-FI module, and the like. The user interface may include a display screen (Display), an input unit such as a keyboard (Keyboard), etc., and the optional user interface may also include a USB interface, a card reader interface, etc. The network interface may optionally include a standard wired interface, a wireless interface (such as a WI-FI interface), etc. Those skilled in the art will understand that the above-mentioned entity device structure provided in this embodiment does not constitute a limitation on the entity device, and may include more or fewer components, or a combination of certain components, or different component arrangements.

[0214] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the above-mentioned physical device, and supports the operation of the information processing program and other software and / or programs. The network communication module is used to realize the communication between the components inside the storage medium, and the communication with other hardware and software in the information processing physical device.

[0215] Through the description of the above implementation methods, the technicians in this field can clearly understand that the present application can be implemented by means of software plus the necessary general hardware platform, or by hardware. Compared with the current existing technology, this embodiment processes the student feature data and the question feature data to obtain the derived feature data and the language feature data, and can analyze the student and question data more deeply on the basis of the student feature data and the question feature data, and then can understand the student's learning and question-solving mode more deeply, and deepen the insight of educational data; by inputting the question feature data, the student feature data, the derived feature data and the language feature data into the deep learning main model and the linear regression auxiliary model respectively, the candidate question set and the second matching degree prediction result are obtained, and the linear regression model can be introduced on the basis of the candidate question set to provide the second matching degree score for verification, and the target question set is obtained. It can more accurately match the student ability and the question difficulty, so that the generated test paper is more in line with the needs of the target students, and then the learning efficiency of the target students can be improved, and the learning motivation of the target students can be improved; in addition, the test paper generation through multiple models can not only ensure the quality of the generated test paper, but also improve the efficiency of the generated test paper in the case of a large amount of data, and shorten the test paper generation time.

[0216] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0217] The above description is only a specific implementation of the present application, so that those skilled in the art can understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. The present application will not be limited to the embodiments described herein, but will conform to the widest scope consistent with the principles and novel features applied for herein.

Claims

1. A test paper generation method, characterized in that: include: Obtaining the topic characteristic data and student characteristic data of the target students; Processing the question feature data and the student feature data to obtain the derived feature data for identifying the student's ability and the language feature data for analyzing the difficulty of the question corresponding to the target student, including: normalizing the continuous features in the question feature data and the student feature data, and discretizing the category features in the question feature data and the student feature data; filling missing values ​​and identifying outliers on the normalized and / or discretized question feature data and student feature data to obtain the derived feature data and the language feature data; The question feature data, student feature data, derived feature data and language feature data are respectively input into a deep learning main model and a linear regression auxiliary model, the matching degree between the target student's ability level and the question difficulty is predicted in the deep learning main model, a set of candidate questions corresponding to each target student is determined based on the first matching degree prediction result, and the matching degree between the target student and the question is predicted in the linear regression auxiliary model to obtain a second matching degree prediction result; wherein, the target student's ability level, question difficulty and the matching degree between the student's ability level and question difficulty are obtained based on training of different prediction models, including: processing the question feature data, student feature data, derived feature data and language feature data according to a predetermined data processing architecture to obtain feature interaction information between the question and the student; inputting the feature interaction information into a question prediction model to predict question difficulty and knowledge point distribution to obtain a predicted value of question difficulty and knowledge point distribution coverage; inputting the feature interaction information into a student ability prediction model to perform knowledge point mastering The method comprises the following steps: predicting the mastery degree of each knowledge point of the target student and the student's ability, obtaining the mastery degree of each knowledge point and the student's comprehensive ability of the target student; inputting the mastery degree of each knowledge point of the target student, the student's comprehensive ability, the predicted value of the difficulty of the question, and the knowledge point distribution coverage into the prediction sub-model, predicting the matching degree of the target student's ability level and the difficulty of the question, and determining the candidate question set corresponding to each target student based on the first matching degree prediction result obtained, including: aligning the mastery degree of each knowledge point of the target student with the knowledge point distribution coverage, and obtaining the matching degree of the question in the knowledge point coverage, including: aligning the student's knowledge point mastery degree with the knowledge point coverage of the question, and calculating the similarity between the two; calculating the matching degree of the student's comprehensive ability and the predicted value of the difficulty of the question, and obtaining the matching value, including: calculating the adaptability score according to the student's overall ability level, i.e., the student's comprehensive ability, and the question difficulty, i.e., the predicted value of the difficulty of the question; selecting the questions that meet the matching degree threshold as the candidate question set corresponding to each target student in the order of matching degree values ​​from high to low; Performing a matching check on the candidate question set according to the second matching prediction result to obtain a target question set, including: comparing and checking the first matching prediction result of each question in the candidate question set according to the second matching prediction result; selecting questions from the candidate question set whose matching and checking results meet preset conditions to form the target question set; A target test paper is generated based on the test paper condition information and the target question set.

2. The method according to claim 1, characterized in that The generating of the target test paper based on the test paper condition information and the target question set includes: Classifying the questions in the target question set based on the question type in the test paper condition information to obtain a target question set corresponding to each question type; A target test paper is generated based on the test paper length in the test paper condition information and the target question set corresponding to each question type.

3. A test paper generating device, characterized in that: include: An acquisition module, configured to acquire question feature data and student feature data of a target student; A processing module is configured to process the question feature data and the student feature data to obtain the derived feature data for identifying the student's ability and the language feature data for analyzing the difficulty of the question corresponding to the target student, including: normalizing the continuous features in the question feature data and the student feature data, and discretizing the category features in the question feature data and the student feature data; filling missing values ​​and identifying outliers on the normalized and / or discretized question feature data and student feature data to obtain the derived feature data and the language feature data; The prediction module is configured to input the question feature data, student feature data, derived feature data and language feature data into a deep learning main model and a linear regression auxiliary model respectively, predict the matching degree between the target student's ability level and the question difficulty in the deep learning main model, determine the set of candidate questions corresponding to each target student based on the obtained first matching degree prediction result, and predict the matching degree between the target student and the question in the linear regression auxiliary model to obtain a second matching degree prediction result; wherein the target student's ability level, question difficulty and the matching degree between the student's ability level and question difficulty are obtained based on different prediction model training, including: processing the question feature data, student feature data, derived feature data and language feature data according to a predetermined data processing architecture to obtain feature interaction information between the question and the student; inputting the feature interaction information into the question prediction model to predict the question difficulty and knowledge point distribution to obtain the question difficulty prediction value and knowledge point distribution coverage; inputting the feature interaction information into the student ability prediction model to perform Prediction of the mastery of knowledge points and students' abilities, obtaining the mastery of each knowledge point and the comprehensive ability of the target students; inputting the mastery of each knowledge point, the comprehensive ability of the students, the predicted value of the difficulty of the questions, and the knowledge point distribution coverage of the target students into the prediction sub-model, predicting the matching degree of the target students' ability level and the difficulty of the questions, and determining the candidate question set corresponding to each of the target students based on the first matching degree prediction result obtained, including: aligning the mastery of each knowledge point of the target students with the knowledge point distribution coverage, and obtaining the matching degree of the questions in the knowledge point coverage, including: aligning the mastery of the students' knowledge points with the knowledge point coverage of the questions, and calculating the similarity between the two; calculating the matching degree of the students' comprehensive ability and the predicted value of the difficulty of the questions, and obtaining the matching degree value, including: calculating the adaptability score according to the overall ability level of the students, i.e., the comprehensive ability of the students, and the difficulty of the questions, i.e., the predicted value of the difficulty of the questions; selecting the questions that meet the matching degree threshold in the order of matching degree values ​​from high to low as the candidate question set corresponding to each of the target students; The verification module is configured to perform a matching verification on the candidate question set according to the second matching prediction result to obtain a target question set, including: comparing and verifying the first matching prediction result of each question in the candidate question set according to the second matching prediction result; selecting questions from the candidate question set whose matching verification results meet preset conditions to form the target question set; The generating module is configured to generate a target test paper based on the test paper condition information and the target question set.

4. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 2 is implemented.

5. An electronic device comprising a storage medium, a processor, and a computer program stored in the storage medium and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 2 is implemented.

6. A computer program product, comprising a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 2 is implemented.

Images