Method for obtaining test paper test questions, readable storage medium and computer equipment

Abstract

The embodiment of the invention discloses a method for obtaining test paper test questions, a readable storage medium and computer equipment, which are used for realizing accurate adaptation and intelligent optimization of personalized teaching resources through deep fusion of multi-source learning behavior data and a dynamic knowledge graph. According to the embodiment of the invention, through dual mechanisms of data driving and knowledge reasoning, chained feedback from behavior analysis to resource optimization is realized, and the intelligence degree of generation of the test question text of the differentiated test paper is improved. In addition, the test question text of the differentiated test paper can be subjected to structured semantic mining, and the semantic knowledge graph is stored, so that the test question text of the differentiated test paper can be efficiently updated subsequently.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of text analysis, and in particular to a method for obtaining test paper questions, a readable storage medium, and a computer device. Background Art

[0002] Traditional online education platforms usually push standardized question banks based on single-dimensional data such as students' correct answer rate or learning progress, lacking in-depth mining of the multi-dimensional characteristics of learning behavior. In the existing technology, the question generation method mostly relies on manually preset difficulty labels and knowledge point classifications, which is difficult to dynamically adapt to students' cognitive changes; at the same time, the question update mechanism often adopts a fixed-cycle replacement strategy, which is unable to reorganize the content according to the relevance of the knowledge system. In addition, most systems have not established an explicit association between the question content and the knowledge structure, which makes it difficult for the question bank to evolve automatically as the syllabus expands. The above problems make it difficult for existing technologies to meet the needs of precise and adaptive resource supply in complex teaching scenarios. Therefore, how to achieve differentiated and intelligent test paper question text generation based on student portraits is a technical problem that needs to be solved at present. Summary of the invention

[0003] The embodiments of the present invention provide a method for obtaining test papers and test questions, a readable storage medium and a computer device, which are used to achieve accurate adaptation and intelligent optimization of personalized teaching resources through deep integration of multi-source learning behavior data and dynamic knowledge graphs.

[0004] In a first aspect, an embodiment of the present invention provides a method for obtaining test papers and test questions, which is applied to a computer device, and the method comprises: obtaining a set of learning operation behavior data generated by a target student on an online teaching platform, wherein the set of learning operation behavior data comprises a plurality of interactive operation sequences, and each interactive operation sequence comprises at least one teaching resource access record, a teaching video viewing record, and a real-time answer feedback record; performing feature extraction processing on the set of learning operation behavior data to generate a multidimensional learning behavior portrait of the target student, wherein the multidimensional learning behavior portrait comprises a knowledge mastery degree feature, a learning concentration feature, and a wrong question association feature; based on the predicted A differentiated test question generation model is set up to perform dynamic strategy analysis on the multi-dimensional learning behavior portrait to generate a differentiated test paper question text that matches the target student; structured semantic mining is performed on the differentiated test paper question text to generate a semantic knowledge graph corresponding to the differentiated test paper question text, and the semantic knowledge graph is stored in a test question knowledge base; the differentiated test paper question text is dynamically updated according to the semantic knowledge graph stored in the test question knowledge base to generate an optimized differentiated test paper question text, and the optimized differentiated test paper question text is pushed to the online teaching terminal corresponding to the target student.

[0005] In a second aspect, an embodiment of the present invention provides a computer device, including: processor; a storage device having a computer program stored thereon, When the computer program is executed by the processor, the processor implements any of the methods for obtaining test paper questions.

[0006] An embodiment of the present invention provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the method for obtaining test paper questions are implemented.

[0007] It can be seen that the embodiments of the present invention have the following beneficial effects: the embodiments of the present invention realize the precise adaptation and intelligent optimization of personalized teaching resources by deeply integrating multi-source learning behavior data and dynamic knowledge graphs. First, based on the dynamic construction of multi-dimensional learning behavior portraits, it can capture the individual characteristics of students in knowledge mastery, attention distribution and wrong question association patterns in real time, breaking through the limitations of traditional static evaluation models. Secondly, through the strategic analysis of learning behavior characteristics by the differentiated test question generation model, a test question combination that is highly adapted to the student's cognitive state is generated, which significantly improves the matching accuracy between the test question and the learner's ability curve. Furthermore, the structured semantic mining technology converts the test question content into an extensible semantic knowledge graph, realizes the explicit expression and dynamic storage of the logical association between knowledge points, and provides a traceable knowledge evolution path for test question updates. Finally, combined with the associative reasoning ability of the semantic knowledge graph, it can actively identify knowledge loopholes and trigger the dynamic reorganization of the test question content, ensuring that teaching resources always fit the students' real-time learning needs. In summary, through the dual mechanisms of data-driven and knowledge reasoning, chain feedback from behavior analysis to resource optimization is realized, and the intelligent degree of differentiated test paper question text generation is improved. In addition, structured semantic mining can be performed on the differentiated test paper question text and the semantic knowledge graph can be stored to facilitate the subsequent efficient updating of the differentiated test paper question text. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 The present invention provides a flowchart of a method for obtaining test questions on a test paper.

[0009] Figure 2 A schematic diagram of the basic structure of a computer device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0010] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0011] See also Figure 1As shown in FIG. 1 , this figure is a flow chart of a method for obtaining test questions provided by an embodiment of the present invention, which can be applied to a computer device. Figure 1 As shown, the method may include steps 110 to 150.

[0012] Step 110: Obtain a learning operation behavior data set generated by the target students on the online teaching platform, wherein the learning operation behavior data set includes multiple interactive operation sequences, each interactive operation sequence consisting of at least one teaching resource access record, teaching video viewing record, and real-time answer feedback record.

[0013] In the embodiment of the present invention, the learning operation behavior data set refers to a set of digital operation records generated by students when they conduct learning activities on the online teaching platform, and its core element is a sequence of interactive operations arranged in chronological order. In this application scenario, the teaching resource access record is specifically manifested as the behavior track of students accessing courseware, downloading learning materials, or browsing auxiliary materials on the platform. Each record contains an access timestamp, a resource identifier, and a length of stay.

[0014] For example, when the target student logs into the online teaching platform, the computer device automatically records the start time, course number and total browsing time of the courseware of "Basics of Computer Networks", forming a record of teaching resource access. The teaching video viewing record reflects the student's learning of the video teaching content on the platform. Each record includes the video identifier, playback progress, number of pauses and the time interval for repeated viewing of the segment.

[0015] For example, when students are learning the teaching video "Detailed Explanation of TCP / IP Protocol Stack", the computer device records the first playback time, total viewing time, and the operation details of jumping to the 15th minute and repeating the video three times. Real-time answer feedback records refer to the answer data generated by students when they complete classroom tests or after-class exercises. Each record contains the question number, answer options, correctness mark, and answer time.

[0016] For another example, in the test of the "Network Layer Routing Algorithm" chapter, the student selected option B for question number Q2023. The computer device determined that the answer was wrong and recorded that it took 45 seconds. Each interactive operation sequence is composed of the above three types of records in chronological order, which fully reflects the student's behavior trajectory in a single learning session. For example, a certain interactive operation sequence includes: accessing the "Data Encapsulation and Decapsulation" courseware at 08:30 (staying for 12 minutes), starting to watch the corresponding teaching video at 08:42 (full playback without jumps), and completing 10 in-class test questions at 09:05 (correct rate 70%). By obtaining these fine-grained operation data, computer equipment can build a multi-dimensional foundation for student behavior analysis.

[0017] Step 120: Perform feature extraction processing on the learning operation behavior data set to generate a multidimensional learning behavior portrait of the target student, wherein the multidimensional learning behavior portrait includes knowledge mastery characteristics, learning concentration characteristics, and wrong question association characteristics.

[0018] In the embodiment of the present invention, feature extraction processing refers to the calculation process of extracting key indicators that characterize the student's learning status from the original operation data. The knowledge mastery level feature is obtained by quantifying the student's understanding level of each knowledge point. The specific implementation method includes: counting the correct answer rate of students under different knowledge units, calculating the distribution density of knowledge points involved in their wrong questions, and analyzing the topic relevance of their repeated visits to teaching resources.

[0019] For example, for the knowledge point of "Subnet Division", the computer equipment statistics show that the student's correct rate for related questions is 62%, and the wrong questions are mainly concentrated in questions related to CIDR notation, and he repeatedly watched the third chapter of the video of this knowledge point. The learning concentration characteristics are evaluated through a time series analysis model, including the duration of a single learning session, the frequency of operation switching, and the proportion of effective learning time. For example, the student's average session duration during the evening study period is 90 minutes, of which 80% of the time is spent on continuously watching videos and answering questions, and there are only 3 page switching operations, indicating that he has a high degree of concentration during this period. The wrong question association feature uses a graph neural network to mine the potential connections between wrong questions, and construct a correlation map with wrong question knowledge points as nodes and co-occurrence frequencies as edges.

[0020] For another example, the student's wrong answers on the two knowledge points of "OSI Model" and "Physical Layer Transmission Medium" are strongly correlated, and the computer equipment recognizes that the two form a cross-error pattern in test questions Q3056 and Q4102. The multi-dimensional learning behavior portrait forms a structured data representation by integrating the above features, such as generating a comprehensive portrait including a knowledge mastery coefficient matrix (such as ["IPv4 Addressing": 0.73, "Routing Protocol": 0.58]), a concentration time series curve (such as the concentration value peaks at 14:00-16:00 every day) and a wrong question correlation map. This portrait not only reflects the student's current learning status, but also reveals their cognitive weaknesses and potential knowledge structure defects.

[0021] Step 130: Based on a preset differentiated test question generation model, a dynamic strategy analysis process is performed on the multi-dimensional learning behavior portrait to generate a differentiated test paper test question text that matches the target student.

[0022] In this embodiment, the differentiated test question generation model refers to a personalized test question generation system built based on a machine learning algorithm, and its core mechanism includes a strategy parsing engine and a test question synthesis module. The dynamic strategy parsing process first converts the multidimensional learning behavior portrait into a strategy parameter vector, such as marking knowledge points with a knowledge mastery coefficient lower than 0.6 as priority reinforcement areas, and setting the high-frequency wrong question association paths as cross-examination focuses. The differentiated test question generation model then dynamically adjusts the cognitive complexity and examination angle of the test questions according to the preset difficulty gradient rules (such as the cognitive levels of Bloom's taxonomy) and knowledge point coverage requirements.

[0023] For example, given that the student's mastery of the knowledge point "VLAN Division" has a characteristic value of 0.55, the model gives priority to practical questions at the application level (such as requiring VLAN configuration given a topology diagram) rather than conceptual multiple-choice questions at the memory level. The test question synthesis module generates test question texts that meet policy requirements based on the semantic template library and knowledge graph. For example, based on the strong correlation between "STP Protocol" and "Network Loop" in the student's wrong question association characteristics, the model generates a comprehensive application question that combines the two knowledge points: "A certain enterprise network has a broadcast storm due to an STP configuration error. Please explain the cause of the failure and provide a solution based on the working principle of the protocol." The generated differentiated test paper question text contains attributes such as question type distribution, knowledge point weight, and difficulty gradient to ensure that the test paper not only covers the requirements of the syllabus, but also makes adaptive adjustments based on individual learning characteristics.

[0024] Step 140: Perform structured semantic mining on the differentiated test paper question text, generate a semantic knowledge graph corresponding to the differentiated test paper question text, and store the semantic knowledge graph in a test question knowledge base.

[0025] In the embodiment of the present invention, structured semantic mining refers to the use of natural language processing technology to analyze the deep semantic relationship of the test text. The specific implementation process includes: extracting the main components of the test questions through dependency syntax analysis, annotating knowledge point entities using named entity recognition, and establishing logical associations between knowledge points based on the relationship extraction model.

[0026] For example, after parsing the test question "Briefly describe the routing update mechanism of the RIP protocol and its maximum hop limit", the computer equipment identifies the core entities "RIP protocol", "routing update mechanism", and "maximum hop count", and establishes semantic relationships such as "belongs to", "has attributes", and "is limited by". The generated semantic knowledge graph is stored in the form of RDF triples. The nodes include elements such as knowledge points, test ability dimensions, and test question difficulty levels, and the edges represent the logical relationship between the elements. For example, the nodes "OSPF protocol" and "link state routing algorithm" are connected through the "belongs to instance" edge, and "configure ACL rules" and "application layer security" are associated through the "involved" edge. The semantic knowledge graph stored in the test question knowledge base realizes the standardized representation of test question resources, supports test question retrieval and association analysis based on semantic similarity, and lays the data foundation for subsequent dynamic updates.

[0027] Step 150: Dynamically update the differentiated test paper question text according to the semantic knowledge graph stored in the test question knowledge base, generate optimized differentiated test paper question text, and push the optimized differentiated test paper question text to the online teaching terminal corresponding to the target student.

[0028] In this embodiment, dynamic update processing refers to an iterative process of continuously optimizing the test question set based on the semantic reasoning ability of the knowledge graph. The computer device first compares the semantic relevance of the current test paper question text with the latest teaching resources in the knowledge base, for example, detecting whether there are differences in the expression of knowledge points caused by the update of the course outline. Then, the graph path analysis is used to identify alternative test questions. If it is found that the knowledge point tested by a certain test question has been replaced by the new version of the protocol, a similar difficulty test question containing updated content is automatically recommended. The optimization process also takes into account the student's latest learning progress. For example, when it is monitored that the student's mastery of the knowledge point "IPv6 Address Configuration" in the recent test has increased to 0.85, the computer device reduces the proportion of questions for this knowledge point, and at the same time increases the examination intensity of its associated knowledge point "Dual Stack Technology Deployment". The generated optimized differentiated test paper question text maintains individual adaptability characteristics, such as replacing two basic-level "DNS Resolution" multiple-choice questions in the original test paper with an application question combined with "DNSSEC Security Extension". Finally, the optimized test papers will be pushed to the students in the form of encrypted data packets through the API interface of the online teaching terminal, ensuring that they can obtain updated personalized learning materials in their personal learning space in a timely manner.

[0029] In an optional embodiment, the step 120 of performing feature extraction processing on the learning operation behavior data set to generate a multi-dimensional learning behavior portrait of the target student includes: Step 121: Perform time dimension analysis on the teaching resource access records in the interactive operation sequence to obtain the resource access duration distribution characteristics of the target students in different knowledge modules.

[0030] Among them, time dimension analysis and processing refers to mining the behavioral patterns of students accessing teaching resources based on time series data, which is specifically achieved by modular classification and duration statistics of the access timestamps of teaching resource access records. For example, the teaching resource access records generated by the target students in the "Computer Network Basics" course contain resource identifiers CS101-2 (data link layer), CS101-3 (network layer), CS101-5 (application layer), etc. The computer device (which can be understood as a differentiated test paper question generation system, hereinafter referred to as the system) maps the resource identifiers to the corresponding knowledge modules according to the preset knowledge module classification rules (such as course chapter division). Taking the "Network Layer" module as an example, the access time period set contains three independent access records: the first access starts at 2023-05-10 09:00:00 and ends at 09:30:00 (lasting 30 minutes); the second access starts at 14:20:00 and ends at 14:50:00 (lasting 30 minutes); the third access starts at 19:15:00 and ends at 19:45:00 (lasting 30 minutes). By calculating and accumulating the time difference of each access, the total access time of the module is 90 minutes. The system compares the actual total time of each module with the preset standard learning time threshold (such as the standard time of the "Network Layer" module is 120 minutes) to generate a time deviation value (90-120=-30 minutes). The duration deviation values ​​of all knowledge modules are normalized to form a resource access duration distribution feature vector. For example, the deviation value of the "network layer" module is normalized to 0.25, and the deviation value of the "data link layer" module is normalized to 0.45 (because the actual duration exceeds the standard value). This distribution feature vector reflects the differences in students' learning input in different knowledge modules. The modules with larger deviation values ​​have higher weights in subsequent feature fusion.

[0031] Step 122: The teaching video viewing record is segmented and parsed to extract the repeated viewing times and pause interval duration characteristics of each teaching video segment. Among them, the segmented analysis process is to divide the continuous video viewing record into fixed-length segment units and analyze the learning behavior indicators of each segment. For example, when the target student watches the "TCP / IP Protocol Stack" video (total length 60 minutes), the system divides it into 12 segment units (V1-V12) at 5-minute intervals. For each segment unit, the system monitors its playback times and pause intervals: for example, segment V3 (10:00-15:00) is played 3 times (the standard threshold is 1 time), and the pause intervals are 8 seconds, 12 seconds, and 15 seconds (the average is 11.67 seconds). The repeated viewing times feature is generated by counting the number of times exceeding the threshold (3-1=2 times, normalized to 0.67), and the pause interval duration feature is generated by calculating the offset between the mean and the standard response time (5 seconds) (11.67-5=6.67 seconds, normalized to 0.78). These two features together reflect the difficulty of students' understanding of specific knowledge points and the fluctuation of their concentration.

[0032] Step 123: performing wrong question clustering processing on the real-time answer feedback record to generate the target student's wrong question knowledge point distribution characteristics and wrong question correction timeliness characteristics. Among them, the wrong question clustering process uses natural language processing technology to identify the relevance of knowledge points in wrong questions. For example, the wrong questions of the target students in the "Network Layer" chapter involve the knowledge points of "IP address classification" and "subnet division". The system classifies them into the same cluster through text similarity calculation. The timeliness feature of wrong question correction is generated by analyzing the time interval between the first answer to the wrong question and the final correct answer. For example, the average correction time of a cluster is 48 hours (standard deviation ±6 hours), indicating that students have a low correction efficiency for this knowledge point cluster.

[0033] Step 124: The resource access duration distribution characteristics, the repeated viewing times characteristics, the pause interval characteristics, the wrong question knowledge point distribution characteristics and the wrong question correction timeliness characteristics are subjected to multi-dimensional fusion processing to generate the target student's knowledge mastery characteristics, learning concentration characteristics and wrong question association characteristics. Among them, multi-dimensional fusion processing adopts feature weighted splicing and principal component analysis (PCA) dimensionality reduction method. For example, the resource access time distribution feature (weight 0.3), repeated viewing number feature (weight 0.2), and wrong question knowledge point distribution feature (weight 0.5) are linearly weighted to generate the knowledge mastery feature vector [0.75, 0.62, 0.33], which represents the student's mastery level of each knowledge module; the learning concentration feature is generated by the fusion of the pause interval duration feature and the resource access deviation value; the wrong question association feature uses the graph convolutional network (GCN) to mine the topological relationship between wrong question knowledge points and generate an association strength matrix (such as the association strength of "IP address classification" and "subnet division" is 0.79). The final generated multi-dimensional learning behavior portrait provides data support for the subsequent generation of differentiated test questions.

[0034] In a preferred embodiment, the step 121 of performing time dimension analysis on the teaching resource access records in the interactive operation sequence to obtain the resource access duration distribution characteristics of the target students in different knowledge modules includes: Step 1211: Divide the teaching resource access records into access time period sets corresponding to multiple knowledge modules according to a preset knowledge module classification rule, each access time period set including an access start time point and an access end time point. Exemplarily, the knowledge module classification rule is a hierarchical classification standard defined according to the course knowledge system, for example, the "Network Engineering" course is divided into modules such as "Network Protocol", "Routing Algorithm", and "Network Security". The system completes the classification through a mapping table of resource identifiers and knowledge modules. For example, resource identifiers CS201-1 to CS201-3 belong to the "Network Protocol" module. The generation process of the access time period set is: traverse all teaching resource access records, extract the starting time point (such as 2023-05-1214:30:00) and end time point (such as 2023-05-1215:00:00) of each record, and classify them by module. For example, the access time period set of the "Network Protocol" module contains three time periods: [14:30-15:00], [16:20-17:00], and [20:15-21:00], each of which represents a continuous resource access behavior.

[0035] Step 1212: Calculate the time difference between the access start time point and the access end time point in the access time period set corresponding to each knowledge module to generate the total resource access duration corresponding to each knowledge module. Exemplarily, the time difference calculation adopts a discretized accumulation algorithm to sum the difference between the end time point and the start time point of all access time periods in the same module. For example, the three time periods of the "Network Protocol" module last for 30 minutes, 40 minutes and 45 minutes respectively, with a total duration of 115 minutes. The calculation process is accurate to the second level. For example, if the start time of a certain access is 14:30:15 and the end time is 14:55:30, the duration is 25 minutes and 15 seconds (converted to 25.25 minutes). The total duration calculation result is stored as a module-duration key-value pair, for example {"Network Protocol": 115, "Routing Algorithm": 80}.

[0036] Step 1213: Compare the total resource access duration corresponding to each knowledge module with a preset standard learning duration threshold to generate a resource access duration deviation value corresponding to each knowledge module. For example, the standard learning time threshold is derived from the syllabus recommendations or historical data statistics, for example, the standard learning time for the "Network Protocol" module is 120 minutes. The deviation value is calculated as the difference between the actual total time and the standard time, for example, 115-120=-5 minutes (negative value indicates insufficient learning). If the actual time of a module is 150 minutes (such as the standard time of the "Network Security" module is 100 minutes), the deviation value is +50 minutes (positive value indicates over-investment). The deviation value is used to quantify the balance of students' learning of each module.

[0037] Step 1214: Perform normalized distribution processing according to the resource access time deviation values ​​corresponding to all knowledge modules to generate the resource access time distribution characteristics of the target student under different knowledge modules, wherein the knowledge module with a larger time deviation value has a higher weight in the distribution characteristics. Exemplarily, the normalization process uses the Min-Max standardization method to map all deviation values ​​to the [0, 1] interval. For example, the maximum deviation value is +50 minutes ("Network Security" module), and the minimum deviation value is -30 minutes ("Data Link Layer" module), then the normalization value of the "Network Security" module is 1.0, and the "Data Link Layer" module is 0.0. The final generated distribution feature vector is arranged in module order, for example [0.1, 0.8, 0.5, 1.0], where 1.0 corresponds to the module with the largest deviation. This feature vector is used to adjust the weight ratio of each module in the subsequent fusion process.

[0038] In a preferred embodiment, the step 122 of performing segmented parsing on the teaching video viewing record to extract the repeated viewing times and pause interval duration features of each teaching video segment includes: Step 1221: Divide each teaching video viewing record into multiple video segment units according to preset time intervals. For example, the preset time interval is set according to the knowledge point distribution of the video content, for example, each segment unit corresponds to an independent knowledge point explanation paragraph. The system divides the segments according to fixed duration (such as 5 minutes) or dynamic semantic segmentation (such as switching according to subtitle keywords). For example, the "OSI Model" video is divided into 7 segment units, among which segment V4 (20:00-25:00) corresponds to the "Transport Layer Function" knowledge point.

[0039] Step 1222: Real-time monitoring and processing of the playback progress data of each video segment unit is performed to extract the number of repeated playbacks corresponding to the video segment unit and the duration of the pause operation interval within the segment. For example, the playback progress data is obtained through the API interface of the video player, recording the start time, end time and jump event of each playback operation. For example, segment V3 is jumped and replayed 3 times, and two pause operations are generated during each playback (with intervals of 8 seconds and 12 seconds respectively). The system counts the number of repeated playbacks as 3 times, and the pause interval is 10 seconds (average).

[0040] Step 1223: Compare the number of repeated playbacks with a preset attention threshold to generate a knowledge comprehension difficulty score for the target student on the video clip unit. For example, the attention threshold is set based on the group average data, for example, the threshold for repeated playback is 2 times. If the actual number of times is 3 times, the difficulty score = (3-2) / (5-2)×10=3.3 points (out of 10 points), indicating that students have moderate difficulty in understanding this segment.

[0041] Step 1224: Perform a difference calculation based on the pause operation interval duration and a preset standard response duration to generate an offset of the target student's learning concentration on the video clip unit. For example, the standard response time reflects the average pause interval in the ideal state of concentration (e.g., 5 seconds). If the actual mean is 12 seconds, the offset is 7 seconds, and after normalization (assuming the maximum allowable offset is 20 seconds), the characteristic value is 0.35, indicating that the concentration is lower than expected.

[0042] Step 1225: normalize the knowledge comprehension difficulty score and the learning concentration offset to generate the repeated viewing count feature and the pause interval duration feature. For example, normalization uses the Z-score method. For example, the original difficulty score value 3.3 (μ=2.5, σ=1.2) is converted to (3.3-2.5) / 1.2=0.67, and then linearly mapped to the interval [0, 1] to obtain 0.67; the offset 7 seconds (μ=5, σ=3) is converted to (7-5) / 3=0.67, and finally the feature vector [0.67, 0.67] is generated.

[0043] In a preferred embodiment, the step 123 of performing wrong question clustering processing on the real-time answer feedback record to generate the distribution characteristics of wrong question knowledge points and the timeliness characteristics of wrong question correction of the target student includes: Step 1231: extracting a wrong question knowledge point label set from the real-time answer feedback record, wherein the wrong question knowledge point label set includes a teaching chapter identifier and a knowledge point weight value corresponding to each wrong question.

[0044] In this embodiment, the knowledge point label is generated by matching the question stem keywords with the knowledge graph. For example, the wrong question "Calculate the number of hosts in the 192.168.1.0 / 24 subnet" is associated with the knowledge point "Subnet Division" (section identifier CS203-5), and the knowledge point weight value is 0.7 (set according to the assessment score ratio).

[0045] Step 1232: Perform similarity matching processing on the wrong question knowledge point label set to generate a cluster of wrong question knowledge points with the same teaching chapter identifier. In this embodiment, the similarity matching uses the cosine similarity algorithm to calculate the angle cosine value of the knowledge point label vector. For example, the similarity between wrong question Q3056 (label vector [0, 0.7, 0]) and Q4102 (label vector [0, 0.6, 0.3]) is 0.92, which exceeds the threshold of 0.85 and is classified into the same cluster C1.

[0046] Step 1233: Perform time series analysis on the wrong question correction timestamps in each wrong question knowledge point cluster, and generate the average correction response time of the target student in each knowledge point cluster.

[0047] In this embodiment, the time series analysis uses the sliding window method to count the correction time interval. For example, cluster C1 contains 5 wrong questions, the first wrong answer timestamp is 2023-05-10 14:00, the final correct answer timestamp is 2023-05-12 09:30, and the average correction response time is 46 hours.

[0048] Step 1234: Perform weighted calculation processing based on the knowledge point weight value and the average correction response time to generate the wrong question knowledge point distribution characteristics and the wrong question correction timeliness characteristics. In this embodiment, the time decay function is introduced into the weighted calculation. For example, the product of the knowledge point weight of 0.7 and the correction time of 46 hours (normalized to 0.6) is 0.42, which reflects the comprehensive index of the mastery degree and correction efficiency of the knowledge point cluster.

[0049] As an optional implementation, the preset differentiated test question generation model in step 130 performs dynamic strategy analysis on the multi-dimensional learning behavior portrait to generate differentiated test paper texts matching the target students, including: Step 131: Input the knowledge mastery level characteristics into the knowledge point coverage analysis module of the differentiated test question generation model to generate a priority sequence of knowledge points to be strengthened for the target students.

[0050] In this implementation, the knowledge point coverage analysis module dynamically determines the priority ranking system of knowledge points that need to be strengthened by analyzing the correlation between the characteristics of students' knowledge mastery and the coverage of test questions. For example, the characteristic vector of the target student's knowledge mastery is [Subnet Division: 0.58, Routing Protocol: 0.62, Network Security: 0.43]. Based on historical teaching data, the module finds that the coverage rate of the knowledge point "Subnet Division" in the test paper is 35% (for example, 3.5 out of 10 questions involve this knowledge point), and the low mastery of students in this knowledge point will significantly affect the overall score rate. By calculating the impact factor of the test question coverage of each knowledge point (such as the impact factor of "Subnet Division" = 0.58×35%=0.203), the module generates a priority sequence of "Subnet Division" > "Routing Protocol" > "Network Security", indicating that the student's mastery of "Subnet Division" needs to be strengthened first.

[0051] Step 132: Input the learning concentration feature into the test question difficulty adaptation module of the differentiated test question generation model to generate a test question difficulty gradient parameter that matches the current learning status of the target student. In this implementation, the test difficulty adaptation module dynamically adjusts the test difficulty distribution according to the learning concentration characteristics. For example, the learning concentration characteristics of the target student show that his concentration score in the evening is 0.88 (higher than the daily average of 0.65). The module combines the concentration-difficulty mapping rule (when the concentration is > 0.8, the proportion of adaptive challenging questions increases by 20%) to generate difficulty gradient parameters: basic questions 30%, application questions 50%, and comprehensive questions 20%. If the student's concentration drops to 0.5, the gradient parameters are adjusted to: basic questions 50%, application questions 40%, and comprehensive questions 10%, to ensure that the difficulty of the test questions matches his attention level.

[0052] Step 133: Input the wrong question association features into the wrong question derivation module of the differentiated question generation model to generate a set of derived questions that have the same knowledge points as the wrong questions of the target student's history. In this implementation, the wrong question derivation module automatically generates variant questions by analyzing the topological relationship of knowledge points in the wrong question association features. For example, the target student's wrong question association strength on the knowledge points of "OSI Model" and "Physical Layer Transmission Medium" is 0.79. The module generates derived questions based on semantic templates: "Combining the functions of each layer of the OSI model, analyze the performance differences and applicable scenarios of optical fiber and twisted pair in physical layer transmission." This type of question not only retains the core knowledge points of the original wrong question, but also strengthens the relevance of knowledge points through cross-examination.

[0053] Step 134: Optimizing the question combination according to the knowledge point priority sequence, the question difficulty gradient parameter and the derived question set to generate the differentiated test paper question text. In this implementation, the combination of test questions is optimized using a multi-objective constraint algorithm to ensure that the test paper meets the requirements of knowledge point coverage, difficulty gradient and derivative examination. For example, the priority sequence requires that "Subnet Division" accounts for 35%, the difficulty gradient parameter limits the comprehensive questions to 20%, and the derivative test questions must include at least 2 cross-knowledge point questions. The system selects test questions that meet the conditions from the question bank and optimizes the arrangement order, and finally generates a test paper containing 15 questions: 4 basic questions (such as "CIDR notation basic calculation"), 8 application questions (such as "given topology design subnet division plan"), and 3 comprehensive questions (such as "combining VLAN and subnet division to optimize enterprise network"), of which derivative questions account for 13.3%.

[0054] The method for constructing the knowledge point coverage analysis module includes: Step 1311: Obtain a data set of correlation between the knowledge mastery characteristics of different students in the history teaching data and the corresponding test score rates. In specific implementation, the history teaching data set is constructed by integrating the knowledge mastery feature vectors of previous students and their test score records. For example, the data set contains 1,000 records, each of which stores the mastery of a knowledge point (such as "Subnet Division": 0.6) and the corresponding test score rate (such as 65%). Through statistics, it is found that when the mastery of "Subnet Division" is increased to 0.8, the average score rate rises to 85%, revealing a positive correlation between the two.

[0055] Step 1312: Perform a linear regression analysis on the knowledge point mastery characteristics and test question score rates in the association relationship data set to generate a test question coverage influencing factor corresponding to each knowledge point. In the specific implementation, linear regression analysis uses the degree of mastery of knowledge points as the independent variable and the score rate as the dependent variable, and calculates the regression coefficient as the influencing factor. For example, the regression equation of "Subnet Division" is score rate = 0.25 × degree of mastery + 0.5 (R² = 0.82), and the influencing factor is 0.25, indicating that for every 0.1 increase in the degree of mastery of this knowledge point, the score rate is expected to increase by 2.5%.

[0056] Step 1313: Sort the knowledge points according to the question coverage impact factor, and generate a knowledge point priority sorting rule, wherein a knowledge point with a larger question coverage impact factor has a higher priority in the sorting rule. In specific implementation, the sorting rule is to arrange the knowledge points in descending order according to the impact factor. For example, if the impact factor list is "Subnet Division": 0.25, "Routing Protocol": 0.18, "Network Security": 0.12, then the priority sequence is "Subnet Division" > "Routing Protocol" > "Network Security", giving priority to strengthening the knowledge points that have a greater impact on the total score.

[0057] Step 1314: Match the knowledge point priority sorting rules with the target student's knowledge mastery level characteristics to generate the knowledge point coverage analysis module. In specific implementation, the module dynamically combines individual characteristics of students with global sorting rules. For example, the mastery level of the target student's "Subnet Division" is 0.58 (lower than the group mean of 0.65). The module calculates the priority weight = 0.58 × 0.25 = 0.145 based on its influence factor 0.25, which is higher than 0.62 × 0.18 = 0.112 of "Routing Protocol". Finally, the decision instruction to strengthen "Subnet Division" is generated in the parsing module.

[0058] In one implementation, the step 140 of performing structured semantic mining on the differentiated test paper question text to generate a semantic knowledge graph corresponding to the differentiated test paper question text includes: Step 141: Perform semantic word segmentation processing on the test question stem and option content in the differentiated test paper test question text to generate multiple test question semantic units. Based on this implementation, semantic word segmentation uses natural language processing technology to decompose the test text. For example, the test question "Briefly describe the routing update mechanism and hop limit of the RIP protocol" is segmented into semantic units ["RIP protocol", "routing update mechanism", "hop limit"], and the option "A. Broadcast the routing table every 30 seconds" is parsed into ["30 seconds", "broadcast", "routing table"].

[0059] Step 142: Perform context dependency analysis on the test question semantic units to generate the logical association between each test question semantic unit. Based on this implementation, dependency parsing analyzes the modification relationship between semantic units through the syntax tree. For example, "RIP protocol" and "routing update mechanism" form a subject-predicate relationship (correlation 0.9), "hop limit" is an attribute of "mechanism" (correlation 0.7), and "30 seconds" modifies the frequency of "broadcast" (correlation 0.6).

[0060] Step 143: construct a hierarchical relationship network between test question knowledge points based on the logical association, and map the hierarchical relationship network into the semantic knowledge graph; wherein the nodes in the semantic knowledge graph represent test question knowledge points, and the edges represent the logical associations and derivative relationships between knowledge points. Based on this implementation, the hierarchical relationship network constructs a directed graph with knowledge points as nodes and relevance as edge weights. For example, the node "RIP Protocol" is connected to "Distance Vector Protocol" (weight 0.9) through the "belongs to" edge, and to "maximum hops 15" (weight 0.8) through the "has attributes" edge. The semantic knowledge graph is stored as RDF triples, supporting query and reasoning based on the graph structure.

[0061] In one implementation, the step 150 of dynamically updating the differentiated test paper question text according to the semantic knowledge graph stored in the test question knowledge base to generate an optimized differentiated test paper question text includes: Step 151: monitor the newly added teaching resource data in the test question knowledge base, and perform semantic parsing on the newly added teaching resource data to generate newly added knowledge point nodes and associated edges. In this implementation, after the new teaching resources such as the "IPv6 address configuration" courseware are parsed, the knowledge point node "IPv6 address format" (weight 0.7) and its associated edge "Extended type" (weight 0.6) with the "IP address classification" node are generated. At the same time, the dependency relationship between this knowledge point and "dual stack technology" is parsed, and a new edge "Dependency" (weight 0.8) is added.

[0062] Step 152: Perform similarity matching processing on the newly added knowledge point node and the original node in the semantic knowledge graph to determine the insertion position of the newly added knowledge point node in the semantic knowledge graph. In this implementation, similarity matching uses a graph embedding algorithm to calculate the vector distance between the newly added node and the original node. For example, the cosine similarity between "IPv6 address format" and "IP address classification" is 0.85 (threshold 0.7), so it is inserted as a child node of "IP address classification" and inherits its hierarchical relationship.

[0063] Step 153: Perform semantic expansion processing on the test stem in the differentiated test paper question text according to the insertion position to generate an optimized differentiated test paper question text containing the newly added knowledge points; wherein the semantic expansion processing includes reconstruction of the test stem, supplementation of options and adjustment of difficulty parameters. In this implementation, semantic expansion integrates new knowledge points into the original test questions. For example, the original test question "Briefly describe the classification of IPv4 addresses" was reconstructed into "Compare the differences between IPv4 and IPv6 address formats, and explain the application scenarios of dual-stack technology", and the option "IPv6 uses 128-bit hexadecimal representation" was added. The difficulty parameter was adjusted from basic questions to application questions to match the increased complexity of knowledge points. The optimized test paper ensures the timeliness of the content and the integrity of the knowledge system.

[0064] In one implementation, the step 151 of performing semantic parsing on the newly added teaching resource data to generate newly added knowledge point nodes and associated edges includes: Step 1511: extract the core knowledge point description text and the related case description text in the newly added teaching resource data. Specifically, the core knowledge point description text refers to the structured text paragraphs that define the core concepts of the knowledge domain in the newly added teaching resources, such as the explanatory text in the newly added "IPv6 Address Configuration" chapter that "IPv6 addresses are represented in 128-bit hexadecimal, and include three parts: global routing prefix, subnet identifier, and interface identifier"; the associated case description text is the example text that demonstrates the application of knowledge points, such as the operation case of "a certain enterprise network needs to convert IPv4 address 192.168.1.0 / 24 to IPv6 address 2001:db8:: / 32, and configure dual-stack devices to achieve smooth transition." Two types of text are separated from the teaching resources through text positioning algorithms (such as based on title hierarchy and keyword matching), such as extracting chapter summaries from PDF documents as core knowledge point description texts, and extracting configuration cases from experimental instruction manuals as associated case description texts.

[0065] Step 1512: Perform semantic entity recognition and dependency analysis on the core knowledge point description text to generate a complete identifier and attribute description information for the newly added knowledge point node. In detail, semantic entity recognition uses the named entity recognition (NER) model to annotate core knowledge points and their attributes. For example, it identifies the entities "IPv6 address" (knowledge point entity), "global routing prefix" (attribute entity), and "64 bits" (attribute value) from "IPv6 address contains global routing prefix (64 bits), subnet identifier (16 bits), and interface identifier (64 bits)". Dependency resolution constructs modification relationships between entities through syntactic analysis, such as "global routing prefix" and "64 bits" constitute a "bit attribute" relationship. The complete identifier is generated according to the "knowledge point entity-attribute type: attribute value" rule, such as "IPv6 address-component: global routing prefix-bit: 64", and the attribute description information includes the attribute type (such as "bit"), value range (such as "64-bit fixed length") and association strength (such as the dependency strength with "subnet identifier" is 0.8).

[0066] Step 1513: Match the test question stem conversion rules on the associated case description text to generate a set of candidate test question stems associated with the newly added knowledge point node. In detail, the test stem conversion rules define the conversion logic from case text to test questions, for example, converting the configuration case "Convert IPv4 subnet 192.168.1.0 / 24 to IPv6 address 2001:db8:: / 32" into a multiple-choice question stem: "An enterprise needs to upgrade IPv4 subnet 192.168.1.0 / 24 to IPv6. Which of the following configurations complies with the global routing prefix allocation principle?" The candidate test stem set is generated by template filling, for example, using the fill-in-the-blank question template "In IPv6 address 2001:db8:: / 32, the subnet identifier occupies ____ bits", and automatically filling in the answer based on the value in the case (such as "16 bits").

[0067] Step 1514: verify the logical relevance of the candidate test question stem set with the original knowledge point nodes in the semantic knowledge graph, and screen out the candidate test question stems whose logical relevance meets the preset threshold. In detail, the logical relevance verification is achieved by calculating the semantic similarity between the knowledge points in the candidate questions and the existing nodes in the semantic knowledge graph. For example, the question "The role of IPv6 global routing prefix" has a relevance of 0.92 (exceeding the threshold of 0.85) with the "IPv6 address structure" node in the graph, while the relevance with the "IPv4 NAT technology" node is 0.15 (below the threshold). After screening, the questions with high relevance are retained, such as the relevance of 0.88 between the "Generation method of interface identifiers in IPv6 addresses" and the "MAC address conversion" node, which meets the requirements.

[0068] Step 1515: Generate associated edges based on the screened candidate test questions and the attribute description information of the newly added knowledge point nodes, and insert the newly added knowledge point nodes and associated edges into the semantic knowledge graph. In detail, the associated edge generation is based on the dependency relationship between the knowledge points and attributes in the question stem. For example, the question stem "IPv6 subnet identifier length" corresponds to the attribute "subnet identifier: 16 bits", and the edge "examination attribute" is generated and the weight is set to 0.9. The newly added knowledge point node "IPv6 address-subnet identifier" is connected to the original node "IPv6 address structure" through the edge "contained in", with a weight of 0.95. The insertion operation follows the graph consistency verification rules, such as checking whether the "IPv6 address-subnet identifier" node already exists. If not, add the node and associated edge to the graph.

[0069] In one implementation, the step 1512 of performing semantic entity recognition and dependency analysis on the core knowledge point description text to generate a complete identifier and attribute description information of the newly added knowledge point node includes: Step 15121: Perform multi-level word segmentation processing on the core knowledge point description text to generate a set of vocabulary units consisting of knowledge point keywords, attribute modifiers and associated relationship words, wherein the knowledge point keywords are words that describe the core concepts of the knowledge field, the attribute modifiers are words that describe the characteristics or scope of the knowledge points, and the associated relationship words are logical relationship words that connect different knowledge points or attributes. Among them, the multi-level word segmentation process adopts a strategy combining domain dictionaries and rule engines. For example, the sentence "IPv6 address consists of global routing prefix (64 bits), subnet identifier (16 bits) and interface identifier (64 bits)" is segmented into knowledge point keywords "IPv6 address", attribute modifiers "global routing prefix", "64 bits", "subnet identifier", "16 bits", and association relationship words "consist of..." and "and". The vocabulary unit collection is stored according to grammatical roles, such as "IPv6 address" as the core concept, "64 bits" as the quantitative attribute, and "consist of..." to indicate the composition relationship.

[0070] Step 15122: Perform contextual semantic weight analysis on the knowledge point keywords in the vocabulary unit set, extract the vocabulary that appears repeatedly in the core knowledge point description text and matches the standard knowledge point names in the preset knowledge base as candidate knowledge point entities, and generate extended attribute description fragments and association relationship description fragments of the candidate knowledge point entity based on the attribute modifiers and association relationship words. Among them, semantic weight analysis calculates the importance of keywords through the TF-IDF algorithm. For example, "IPv6 address" appears 5 times in the text and completely matches the standard word in the knowledge base, and the weight value is 0.9 (highest). The extended attribute description fragment is generated by combining attribute modifiers and association relationship words. For example, the "global routing prefix (64 bits)" fragment describes the "prefix length attribute", and the association relationship fragment "together with the subnet identifier to form an address structure" describes the composition relationship.

[0071] Step 15123: Perform dependency syntactic structure analysis on the candidate knowledge point entity and its extended attribute description fragments, construct a dependency tree with the candidate knowledge point entity as the root node, attribute modifiers as child nodes, and associated relationship words as edge labels, and extract the attribute modifiers directly connected to the candidate knowledge point entity in the dependency tree as the core attribute entity set. Among them, dependency parsing generates a tree structure, for example, the root node "IPv6 address" is connected to the child node "global routing prefix" through the edge "attribute: component", and the child node is further connected to the leaf node "64 bits" through the edge "attribute: number of bits". The core attribute entity set extracts directly connected child nodes, such as "global routing prefix", "subnet identifier", and "interface identifier", and ignores the secondary nodes "64 bits" and "16 bits".

[0072] Step 15124: According to the hierarchical structure of the dependency tree, the candidate knowledge point entity is combined and concatenated with the attribute modifiers in the core attribute entity set to generate a complete identifier for the newly added knowledge point node, wherein the complete identifier is composed of the standard name of the candidate knowledge point entity, the type label of the core attribute entity, and the attribute value concatenated in a preset order. The concatenation rule is "knowledge point entity-attribute type: attribute value". For example, "IPv6 address-component: global routing prefix-bits: 64" means that the global routing prefix of the IPv6 address described by this node is 64 bits. The complete identifier must be unique. For example, "IPv6 address-component: subnet identifier-bits: 16" and "IPv6 address-component: interface identifier-bits: 64" are two independent nodes.

[0073] Step 15125: Perform logical connector parsing on the association relationship words in the association relationship description fragment, extract the dependency path between the candidate knowledge point entity and the core attribute entity set, and map the attribute modifiers in the extended attribute description fragment to the attribute description information of the newly added knowledge point node according to the dependency path, wherein the attribute description information includes the attribute type, attribute value range and attribute association strength. Among them, logical connectors are parsed to identify the relationship type, such as "consisting of..." is mapped to "composition relationship", and "value range is" is mapped to "value constraint". The dependency path "IPv6 address → composition → global routing prefix → number of bits → 64 bits" is converted into attribute description information: attribute type "number of bits", value range "64-bit fixed length", and association strength 0.9 (based on co-occurrence frequency).

[0074] Step 15126: Structurally encapsulate the complete identifier and the attribute description information to generate a complete data object of the newly added knowledge point node including a unique identifier field, an attribute type field, an attribute value range field and an association strength field, and perform a uniqueness verification on the complete data object and the original knowledge point node in the preset knowledge base and then store it in the test question knowledge base. The structured package uses the JSON format, for example: json { "Unique Identifier": "IPv6 Address - Component: Global Routing Prefix - Number of bits: 64", "Attribute Type": "Digit Attribute", "Value range": "64 bits", "Association Strength": {"Subnet ID": 0.7, "Interface ID": 0.8} } Among them, the uniqueness check is achieved by comparing the identifier hash value. If there is no duplication in the knowledge base, the data object is inserted into the graph node list and an associated edge with the "IPv6 address structure" node is established.

[0075] Based on the above technical solution, in an optional embodiment, the method further includes: Monitoring the real-time answer data stream of the target student to the optimized differentiated test paper question text, and extracting the test question correctness distribution characteristics and knowledge point gap marks in the real-time answer data stream; Dynamically attenuate the node weights in the semantic knowledge graph according to the test question accuracy distribution characteristics, thereby reducing the association strength of knowledge point nodes with accuracy rates higher than a preset threshold in the semantic knowledge graph; Based on the knowledge point vulnerability mark, topological reconstruction is performed on the associated edges in the semantic knowledge graph to generate an updated semantic knowledge graph containing a newly added vulnerability knowledge point path; The updated semantic knowledge graph is input into the differentiated test question generation model, triggering a secondary correction process of the knowledge mastery level feature, generating iteratively optimized differentiated test paper question texts and pushing them to the online teaching terminal.

[0076] In this embodiment, the real-time answer data stream refers to the time-series answer record generated by students in the process of completing the optimized differentiated test paper, including the answer result of each test question, the time spent on answering the questions and the knowledge point label. For example, the target student's correct rate in the test questions related to "IPv6 Address Configuration" is 95% (exceeding the preset threshold of 90%), and the system extracts the correct rate distribution characteristics of the knowledge point and marks it as a "high mastery knowledge point". The dynamic decay processing uses an exponential decay algorithm to reduce the association strength of such nodes from the initial value of 0.9 to 0.6, reducing its examination weight in subsequent test papers. The knowledge point vulnerability marker is generated by identifying knowledge points with continuous errors or too long correction time. For example, if a student answers incorrectly three times in a row in the "Dual Stack Technology Deployment" test question, the system marks the knowledge point as a vulnerability node and adds a new associated edge "Dependency Repair Path" in the semantic knowledge graph to connect to the "IPv6 Address Configuration" node (the original associated edge strength 0.7 is increased to 0.85). The updated semantic knowledge graph triggers the differentiated test question generation model to recalculate the knowledge mastery level characteristics, and increases the proportion of derivative questions of "Dual Stack Technology Deployment" to 25% when generating iterative test papers, while reducing the number of questions on "IPv6 Address Configuration".

[0077] Based on the above technical solution, in an optional embodiment, the method further includes: Performing time series analysis on the push feedback records of the optimized differentiated test paper test question texts to extract the test question response delay characteristics and continuous answer interval fluctuation characteristics of the target students; Performing time sensitivity labeling processing on the knowledge point nodes in the semantic knowledge graph according to the test question response delay characteristics, and generating a dynamic knowledge point priority sequence labeled with time constraints; Based on the fluctuation characteristics of the intervals between consecutive answers, the gradient parameters in the test question difficulty adaptation module are nonlinearly scaled to generate a smooth difficulty transition curve that adapts to the real-time attention of the target student; The dynamic knowledge point priority sequence and the smooth difficulty transition curve are subjected to multi-dimensional cross-matching processing to generate a reinforcement test paper text with time-adaptive characteristics and push it to the online teaching terminal.

[0078] In this embodiment, the time-series analysis of the push feedback record is achieved by analyzing the timestamp sequence of students opening the test paper and submitting answers. For example, the average delay in the target student's test response in the evening period of 20:00-21:00 is 5 seconds (3 seconds during the day). The system extracts this period as a "high delay interval" and marks the time sensitivity of the relevant knowledge point nodes (such as "Network Topology Design") as "night priority". The fluctuation characteristics of the interval between consecutive answers are characterized by the stability of attention by calculating the standard deviation of the interval between answers to adjacent questions (such as 15±8 seconds). If the fluctuation value exceeds the threshold (such as 10 seconds), the difficulty gradient parameter is nonlinearly scaled: the original gradient parameters of "basic questions 40%, application questions 50%, comprehensive questions 10%" are adjusted to "basic questions 30%, application questions 45%, comprehensive questions 25%" to match the need to increase cognitive challenges when attention is distracted. The multi-dimensional cross-matching of the dynamic knowledge point priority sequence and the smooth difficulty curve is achieved through a matrix alignment algorithm. For example, the time-sensitive "Routing Protocol" knowledge point (priority 0.8) is bound to the "comprehensive questions" interval in the difficulty curve, and the proportion of comprehensive questions for this knowledge point in the generated intensive test paper is increased to 30%.

[0079] Based on the above technical solution, in an optional embodiment, the method further includes: Capturing terminal interaction trajectory data of the target student when receiving the optimized differentiated test paper question text, and parsing the screen focus movement path and touch operation density distribution in the interaction trajectory data; Performing spatiotemporal calibration processing on the learning concentration feature in the multi-dimensional learning behavior portrait according to the screen focus movement path, and generating an updated learning concentration feature including a time-division attention decay curve; Using the touch operation density distribution, probability weighted processing is performed on the node connection strength in the semantic knowledge graph to generate a reinforced knowledge point association network that reflects the operation habit preference; The updated learning concentration feature and the reinforced knowledge point association network are input into the wrong question derivation module to generate a dynamic wrong question derivation question set based on attention-driven and push it to the online teaching terminal.

[0080] In this embodiment, the terminal interaction trajectory data collects the screen touch coordinates and the focus stay time through the buried point technology. For example, the focus movement path of the target student on the "VLAN Configuration" test page shows multiple switches between the "Topology Map Area" and the "Command Line Example Area" (switching frequency 3 times / second), and the touch operation density reaches a peak at the "Command Line Example" button (10 clicks). The spatiotemporal calibration process maps such behaviors into time-divided attention decay curves: the concentration is maintained at 0.9 in the first 5 minutes, and then drops to 0.6 due to frequent switching. The touch operation density distribution generates hot spots through kernel density estimation, and the connection strength of the associated knowledge point nodes (such as "VLAN Command Line Configuration") is weighted to 0.9 (originally 0.7). Strengthen the knowledge point association network to drive the wrong question derivation module to generate dynamic test questions. For example, combined with the command line operation preference with high touch density, the test question "Write VLAN batch configuration script according to the topology map" is derived and the interactive command line simulator option is set. The proportion of such questions in the updated test paper has increased to 40%.

[0081] Based on the above technical solution, in an optional embodiment, the method further includes: Collecting in real time the rendering and presentation data of the optimized differentiated test paper question text on the online teaching terminal, and extracting the question stem layout parameters and option visual significance indicators from the rendering and presentation data; Performing spatial matching analysis on the wrong question association features in the multi-dimensional learning behavior portrait according to the question stem layout parameters, and generating a wrong question knowledge point reorganization sequence adapted to different screen resolutions; Based on the option visual saliency index, the gradient parameters in the test question difficulty adaptation module are subjected to visual interference factor compensation processing to generate an objective difficulty evaluation matrix that eliminates interface preference; The reorganized sequence of wrong question knowledge points is superimposed and fused with the objective difficulty assessment matrix to generate a visual test paper question text with an adaptively optimized interface and push it to the online teaching terminal.

[0082] In this embodiment, the rendering presentation data includes the position of the question stem on the screen (such as center or sidebar), font size (such as 14pt or 12pt) and option color contrast (such as #FF0000 and #FFFFFF). For example, on the mobile screen (resolution 720×1280), the question stem of "Subnet Division" is difficult to read due to too many line breaks. The system extracts the layout parameter "line spacing 8pt" and analyzes its spatial matching degree with the characteristics associated with the wrong question (matching value 0.4), and then generates a wrong question knowledge point reorganization sequence adapted to the tablet terminal (resolution 1200×1600), and adjusts the relevant questions of "Subnet Division" to a single column layout. The visual saliency index is calculated by an eye tracking simulation algorithm. For example, the visual saliency score of the red option (#FF0000) is 0.9, which may cause students to make a wrong choice. The system compensates for the interference factor in the difficulty assessment matrix (such as correcting the original difficulty of 0.6 to 0.5). The interface adaptively optimizes the test paper by overlaying the layout reorganization sequence and the revised difficulty matrix to generate a visual test text. For example, a two-column layout is used on the PC to display the comprehensive question "IPv6 Address", and the color saturation of high-significance options is reduced to reduce interference.

[0083] The embodiment of the present invention realizes the precise adaptation and intelligent optimization of personalized teaching resources by deeply integrating multi-source learning behavior data and dynamic knowledge graph. First, based on the dynamic construction of multi-dimensional learning behavior portraits, it can capture the individual characteristics of students in knowledge mastery, attention distribution and wrong question association patterns in real time, breaking through the limitations of traditional static evaluation models. Secondly, through the strategic analysis of learning behavior characteristics by the differentiated test question generation model, a test question combination that is highly adapted to the student's cognitive state is generated, which significantly improves the matching accuracy between the test question and the learner's ability curve. Furthermore, the structured semantic mining technology converts the test question content into an extensible semantic knowledge graph, realizes the explicit expression and dynamic storage of the logical association between knowledge points, and provides a traceable knowledge evolution path for test question updates. Finally, combined with the associative reasoning ability of the semantic knowledge graph, it can actively identify knowledge loopholes and trigger the dynamic reorganization of the test question content, ensuring that teaching resources always fit the students' real-time learning needs. In summary, through the dual mechanisms of data-driven and knowledge reasoning, chain feedback from behavior analysis to resource optimization is realized, and the intelligent level of differentiated test paper question text generation is improved. In addition, structured semantic mining can be performed on the differentiated test paper question text and the semantic knowledge graph can be stored to facilitate the subsequent efficient updating of the differentiated test paper question text.

[0084] See also Figure 2 As shown, this figure is a schematic diagram of the basic structure of a computer device 200 provided by an embodiment of the present invention, and the computer device 200 includes: Processor 201; a storage device 202 on which a computer program 2020 is stored; When the computer program 2020 is executed by the processor 201, the processor 201 implements any of the methods for obtaining test paper questions.

[0085] Based on the above, a readable storage medium is provided, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the above method are implemented.

[0086] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other. For the system or device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

Claims

1. A method for obtaining test questions, characterized in that: The method comprises: Acquire a learning operation behavior data set generated by a target student on an online teaching platform, wherein the learning operation behavior data set includes a plurality of interactive operation sequences, each interactive operation sequence consisting of at least one teaching resource access record, a teaching video viewing record, and a real-time answer feedback record; Performing feature extraction processing on the learning operation behavior data set to generate a multi-dimensional learning behavior portrait of the target student, wherein the multi-dimensional learning behavior portrait includes a knowledge mastery degree feature, a learning concentration feature, and a wrong question association feature; Based on a preset differentiated test question generation model, a dynamic strategy analysis process is performed on the multi-dimensional learning behavior portrait to generate differentiated test paper test question texts matching the target students; Performing structured semantic mining processing on the differentiated test paper question text, generating a semantic knowledge graph corresponding to the differentiated test paper question text, and storing the semantic knowledge graph in a test question knowledge base; The differentiated test paper question text is dynamically updated according to the semantic knowledge graph stored in the test question knowledge base to generate optimized differentiated test paper question text, and the optimized differentiated test paper question text is pushed to the online teaching terminal corresponding to the target student.

2. The method according to claim 1, characterized in that The step of performing feature extraction processing on the learning operation behavior data set to generate a multi-dimensional learning behavior portrait of the target student includes: Performing time dimension analysis on the teaching resource access records in the interactive operation sequence to obtain the resource access time distribution characteristics of the target students in different knowledge modules; The teaching video viewing record is analyzed in segments to extract the repeated viewing times and pause interval length features of each teaching video segment; Performing wrong question clustering processing on the real-time answer feedback record to generate the distribution characteristics of wrong question knowledge points and the timeliness characteristics of wrong question correction of the target student; The resource access time distribution characteristics, the repeated viewing times characteristics, the pause interval characteristics, the wrong question knowledge point distribution characteristics and the wrong question correction timeliness characteristics are subjected to multi-dimensional fusion processing to generate the target student's knowledge mastery characteristics, learning concentration characteristics and wrong question association characteristics.

3. The method according to claim 2, characterized in that The time dimension analysis and processing of the teaching resource access records in the interactive operation sequence to obtain the resource access time distribution characteristics of the target students in different knowledge modules includes: Dividing the teaching resource access records into access time period sets corresponding to a plurality of knowledge modules according to a preset knowledge module classification rule, each access time period set including an access start time point and an access end time point; Calculate the time difference between the access start time point and the access end time point in the access time period set corresponding to each knowledge module to generate the total resource access duration corresponding to each knowledge module; Compare the total resource access time corresponding to each knowledge module with the preset standard learning time threshold to generate a resource access time deviation value corresponding to each knowledge module; Normalized distribution processing is performed according to the resource access time deviation values ​​corresponding to all knowledge modules to generate the resource access time distribution characteristics of the target students under different knowledge modules, wherein the knowledge module with a larger time deviation value has a higher weight in the distribution characteristics.

4. The method according to claim 2, characterized in that: The segmented parsing process of the teaching video viewing record to extract the repeated viewing times and pause interval duration features of each teaching video segment includes: Divide each teaching video viewing record into a plurality of video segment units according to a preset time interval; Performing real-time monitoring and processing on the playback progress data of each video segment unit, and extracting the number of repeated playbacks corresponding to the video segment unit and the interval length of the pause operation within the segment; Compare the number of repeated playbacks with a preset attention threshold to generate a knowledge comprehension difficulty score for the target student for the video clip unit; Performing a difference calculation process based on the pause operation interval duration and a preset standard response duration to generate a learning concentration offset of the target student on the video clip unit; The knowledge comprehension difficulty score and the learning concentration offset are normalized to generate the repeated viewing number feature and the pause interval duration feature.

5. The method according to claim 2, characterized in that: The performing wrong question clustering processing on the real-time answer feedback record to generate the distribution characteristics of wrong question knowledge points and the timeliness characteristics of wrong question correction of the target student includes: Extracting a wrong question knowledge point label set from the real-time answer feedback record, wherein the wrong question knowledge point label set includes a teaching chapter identifier and a knowledge point weight value corresponding to each wrong question; Performing similarity matching processing on the wrong question knowledge point label set to generate wrong question knowledge point clusters with the same teaching chapter identifier; Performing time series analysis on the wrong question correction timestamps in each wrong question knowledge point cluster to generate the average correction response time of the target student in each knowledge point cluster; A weighted calculation process is performed based on the knowledge point weight value and the average correction response time to generate the wrong question knowledge point distribution characteristics and the wrong question correction timeliness characteristics.

6. The method according to claim 1, characterized in that The method of performing dynamic strategy analysis on the multi-dimensional learning behavior portrait based on the preset differentiated test question generation model to generate differentiated test paper texts matching the target students includes: Inputting the knowledge mastery level characteristics into the knowledge point coverage parsing module of the differentiated test question generation model to generate a priority sequence of knowledge points to be strengthened for the target students; Inputting the learning concentration feature into the test question difficulty adaptation module of the differentiated test question generation model to generate a test question difficulty gradient parameter that matches the current learning state of the target student; Inputting the wrong question association features into the wrong question derivation module of the differentiated question generation model to generate a set of derived questions with the same knowledge points as the wrong questions of the target student's history questions; Performing test question combination optimization processing according to the knowledge point priority sequence, the test question difficulty gradient parameter and the derived test question set to generate the differentiated test paper test question text; The method for constructing the knowledge point coverage analysis module includes: Obtain a data set of correlation between the knowledge mastery characteristics of different students and the corresponding test score rates in history teaching data; Performing linear regression analysis on the knowledge point mastery characteristics and test question score rates in the association relationship data set to generate a test question coverage influencing factor corresponding to each knowledge point; Sort the knowledge points according to the question coverage impact factor to generate a knowledge point priority sorting rule, where a knowledge point with a greater question coverage impact factor has a higher priority in the sorting rule; The knowledge point priority sorting rules are matched with the knowledge mastery level characteristics of the target students to generate the knowledge point coverage analysis module.

7. The method according to claim 1, characterized in that The performing structured semantic mining processing on the differentiated test paper question text to generate a semantic knowledge graph corresponding to the differentiated test paper question text includes: Performing semantic word segmentation processing on the test question stem and option contents in the differentiated test paper test question text to generate multiple test question semantic units; Performing context dependency analysis on the test question semantic units to generate a logical correlation between each test question semantic unit; Constructing a hierarchical relationship network between test question knowledge points according to the logical association, and mapping the hierarchical relationship network into the semantic knowledge graph; wherein the nodes in the semantic knowledge graph represent test question knowledge points, and the edges represent the logical associations and derivative relationships between the knowledge points; The dynamically updating the differentiated test paper question text according to the semantic knowledge graph stored in the test question knowledge base to generate an optimized differentiated test paper question text includes: Monitoring newly added teaching resource data in the test question knowledge base, and performing semantic parsing processing on the newly added teaching resource data to generate newly added knowledge point nodes and associated edges; Perform similarity matching processing on the newly added knowledge point node and the original node in the semantic knowledge graph to determine the insertion position of the newly added knowledge point node in the semantic knowledge graph; The test question stem in the differentiated test paper question text is semantically expanded according to the insertion position to generate an optimized differentiated test paper question text containing the newly added knowledge points; wherein the semantic expansion processing includes reconstruction of the test question stem, supplementation of options and adjustment of difficulty parameters.

8. The method according to claim 7, characterized in that The semantic parsing of the newly added teaching resource data to generate newly added knowledge point nodes and associated edges includes: Extracting the core knowledge point description text and the related case description text from the newly added teaching resource data; Performing semantic entity recognition and dependency analysis on the core knowledge point description text to generate a complete identifier and attribute description information of the newly added knowledge point node; Match the test question stem conversion rule on the associated case description text to generate a candidate test question stem set associated with the newly added knowledge point node; Verify the logical relevance of the candidate test question stem set with the original knowledge point nodes in the semantic knowledge graph, and screen out the candidate test question stems whose logical relevance meets the preset threshold; Generate associated edges based on the selected candidate test questions and the attribute description information of the newly added knowledge point nodes, and insert the newly added knowledge point nodes and associated edges into the semantic knowledge graph; The performing of semantic entity recognition and dependency analysis on the core knowledge point description text to generate a complete identifier and attribute description information of the newly added knowledge point node includes: Performing multi-level word segmentation processing on the core knowledge point description text to generate a vocabulary unit set consisting of knowledge point keywords, attribute modifiers and association relationship words, wherein the knowledge point keywords are words that describe the core concepts of the knowledge field, the attribute modifiers are words that describe the characteristics or scope of the knowledge points, and the association relationship words are logical relationship words that connect different knowledge points or attributes; Performing contextual semantic weight analysis on the knowledge point keywords in the vocabulary unit set, extracting the words that appear repeatedly in the core knowledge point description text and match the standard knowledge point names in the preset knowledge base as candidate knowledge point entities, and generating extended attribute description segments and association relationship description segments of the candidate knowledge point entities based on the attribute modifiers and association relationship words; Perform dependency syntactic structure analysis on the candidate knowledge point entity and its extended attribute description fragments, construct a dependency tree with the candidate knowledge point entity as the root node, attribute modifiers as child nodes, and association words as edge labels, and extract attribute modifiers directly connected to the candidate knowledge point entity in the dependency tree as a core attribute entity set; According to the hierarchical structure of the dependency tree, the candidate knowledge point entity is combined and spliced ​​with the attribute modifiers in the core attribute entity set to generate a complete identifier of the newly added knowledge point node, wherein the complete identifier is composed of the standard name of the candidate knowledge point entity, the type label of the core attribute entity and the attribute value spliced ​​in a preset order; Performing logical connector parsing on the association relationship words in the association relationship description segment, extracting the dependency path between the candidate knowledge point entity and the core attribute entity set, and mapping the attribute modifiers in the extended attribute description segment to the attribute description information of the newly added knowledge point node according to the dependency path, wherein the attribute description information includes the attribute type, the attribute value range and the attribute association strength; The complete identifier and the attribute description information are structured and encapsulated to generate a complete data object of a newly added knowledge point node including a unique identifier field, an attribute type field, an attribute value range field and an association strength field, and the complete data object is verified for uniqueness with the original knowledge point nodes in the preset knowledge base and then stored in the test question knowledge base.

9. A readable storage medium, characterized in that: The readable storage medium stores programs or instructions, and when the programs or instructions are executed by the processor, the method for obtaining test paper questions as described in any one of claims 1-8 is implemented.

10. A computer device, characterized in that: include: processor; A storage device having a computer program stored thereon, wherein when the computer program is executed by the processor, the processor implements the method for obtaining test paper questions as described in any one of claims 1-8.

Images