A method for generating intelligent practice content by dynamically adjusting the weights of question extraction

By constructing a database of question proficiency and dynamically adjusting question weights, the problem of the single evaluation dimension in the existing system has been solved. This enables accurate assessment of users' cognitive levels and refined adjustment of personalized learning paths, thereby improving the utilization efficiency and adaptability of learning resources.

CN123068620APending Publication Date: 2026-11-10GUANGZHOU MINGZHI EDUCATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611247145.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-17
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing intelligent learning systems have a single evaluation dimension and insufficient accuracy, making it difficult to distinguish different cognitive levels of users, resulting in inaccurate personalized learning path recommendations.

Method used

By acquiring users' historical answer records, a database of question proficiency is constructed. Combining answer results and time taken, the question extraction weight is dynamically adjusted. A differentiated extraction weight allocation model is used to generate practice content. Answer time is introduced as a process-oriented quantitative indicator of cognitive proficiency, thereby achieving refined adjustment of proficiency levels.

Benefits of technology

It enables accurate assessment of users' cognitive levels, avoids repeatedly pushing content that users have already mastered, ensures that practice content targets users' knowledge gaps, and improves the efficiency of learning resource utilization and the level of personalized adaptation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123068620A_ABST
    Figure CN123068620A_ABST
Patent Text Reader

Abstract

This invention relates to an intelligent practice content generation method that dynamically adjusts the weighting of extracted questions. The technical solution includes: extracting question identifiers and answer results from users' historical answer records; constructing and dynamically maintaining a question proficiency database; receiving answer results and time taken in real time; adjusting the proficiency level based on answer accuracy and time taken; and recording learning progress; combining the proficiency level distribution and question priority, using a differentiated extraction weighting allocation model to perform weighted random sampling of the question bank, and supplementing the sampled questions with a priority replenishment mechanism when necessary to meet the required number of questions. This method achieves personalized and accurate question recommendation and closed-loop proficiency evaluation, improving the adaptability of the intelligent practice system and the user's learning effectiveness.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent education assessment and personalized learning content generation technology, and in particular to an intelligent practice content generation method that dynamically adjusts the weight of question extraction. Background Technology

[0002] Currently, the field of intelligent learning and education technology is developing rapidly, especially in areas such as intelligent generation of practice content, accurate assessment of users' knowledge mastery, and personalized recommendations. Various practice content generation platforms and products based on answer data analysis have emerged. The common practice in current mainstream industry solutions is to simply label each question as "correct / incorrect," adding incorrect questions to a collection of incorrect questions for priority recommendation, while correctly answered questions are marked as mastered and pushed with reduced priority, or implementing a traversal learning approach based on a fixed order.

[0003] Most existing systems determine mastery solely based on whether a user's answers are "correct" or "incorrect." Incorrect questions are categorized into a "wrong question" set and prioritized for recommendation, while correct questions are marked as "mastered" and have their recommendation weight reduced. This "rework incorrect questions" model is currently the most prevalent solution—some platforms claim their core selling point is "recommending practice questions based on your weaknesses." Some platforms have made limited extensions, such as labeling each question with different reasons for error or reconstructing knowledge maps using these reasons; however, overall, they still rely on a binary result as the primary decision-making basis. An edtech company points out that most educational software on the market "only provides answers but lacks problem-solving guidance," leading students into a "sea of ​​questions" approach with little to show for it.

[0004] Some systems attempt to introduce the concept of knowledge mastery, but the evaluation logic is rather crude. For example, some patented solutions only "reflect the user's level of knowledge mastery based on the error rate; the higher the error rate, the weaker the mastery of that knowledge point." Other solutions simply divide mastery into two states: "mastered" and "not mastered." This coarse-grained evaluation method makes it difficult to distinguish between different cognitive levels such as "proficient," "mastered," and "guessed correctly."

[0005] The existing technology has the following prominent problems:

[0006] The assessment dimensions are too limited and the accuracy is insufficient. Most current systems only judge mastery based on whether the user answers correctly or incorrectly, lacking consideration of process indicators such as answering time and thinking process. This results in the inability to distinguish between different cognitive levels such as "proficiency, mastery, and guessing correctly," making the proficiency assessment too crude and hindering the delivery of personalized and efficient learning paths. Summary of the Invention

[0007] This application provides an intelligent practice content generation method that dynamically adjusts the weights of question extraction, aiming to solve the problems of single evaluation dimensions and insufficient accuracy mentioned in the background technology.

[0008] This application provides a method for generating intelligent practice content that dynamically adjusts the weights of question extraction, specifically including:

[0009] S1: Obtain the user's historical answer record data, and construct a question proficiency database including a question proficiency level field based on the question identifier and answer result in the answer record data;

[0010] S2: Receive the real-time answer result signal and the corresponding answer duration value of the current question, and judge the correctness of the answer based on the real-time answer result signal. If the answer is determined to be incorrect, perform downgrade processing on the corresponding question proficiency level in the question proficiency database and generate a downgraded proficiency level.

[0011] S3: If the answer is determined to be correct, the answer duration value is compared and matched with a preset time threshold interval set. The proficiency upgrade range is determined according to the matched time threshold interval. Based on the upgrade range, the corresponding question proficiency level in the question proficiency database is upgraded to generate an upgraded proficiency level.

[0012] S4: Write the downgraded proficiency level or the upgraded proficiency level into the question proficiency database to overwrite the original data, and synchronously record the learning progress log data to form a complete dynamic evaluation closed-loop data source.

[0013] S5: The dynamic evaluation closed-loop data source is combined with the priority classification tags of the questions in the question bank to construct a differentiated extraction weight allocation model;

[0014] S6: Use the differentiated extraction weight allocation model to perform weighted random sampling on the candidate question set in the question bank to generate an initial question sequence, and count whether the actual number of questions in the initial question sequence meets the preset question number parameter requirements.

[0015] S7: If the actual number of questions does not meet the requirement of the number of questions to be drawn, the number of questions will be supplemented and incorporated into the initial question sequence to obtain the final intelligent practice content.

[0016] This application provides a method for generating intelligent practice content that dynamically adjusts the weights of question extraction, which has the following beneficial effects:

[0017] First, regarding assessment accuracy, this application breaks through the limitations of the traditional binary "right / wrong" judgment by introducing answering time as a process-oriented quantitative indicator of cognitive proficiency. By matching answering time with a preset time threshold range, it can effectively distinguish different cognitive levels such as "quick and proficient mastery," "general mastery," and "guessing correctly," and performs downgrading processing on incorrect answers, achieving refined and dynamic adjustment of proficiency levels. This solves the technical problem of traditional systems having a single assessment dimension and insufficient accuracy. Second, regarding personalized learning paths, the differentiated extraction weight allocation model constructed in this application can achieve dynamic mapping of question extraction weights based on a user's individual proficiency database. This mechanism can accurately focus on weak links, effectively avoid repeatedly pushing content that has already been mastered, and ensure that each exercise targets the user's current knowledge gaps, significantly improving the utilization efficiency and personalized adaptation level of learning resources. Finally, in terms of intelligent content generation, this application ensures that the generated practice content meets the preset requirements in terms of the number of questions and the distribution of difficulty through weighted random sampling and automatic supplementation mechanism. This not only guarantees the randomness and diversity of the questions but also achieves targeted reinforcement of weak knowledge points, truly achieving the intelligent practice goal of "identifying and filling gaps and consolidating knowledge efficiently," and providing solid technical support for personalized adaptive learning. Attached Figure Description

[0018] Figure 1 This is the main flowchart of an intelligent practice content generation method that dynamically adjusts the weights extracted from questions.

[0019] Figure 2 This is a sub-flowchart of an intelligent practice content generation method that dynamically adjusts the weights extracted from questions.

[0020] Figure 3 This is another sub-flowchart of an intelligent practice content generation method that dynamically adjusts the weights extracted from questions.

[0021] Figure 4 This is a block diagram of an intelligent practice content generation system that dynamically adjusts the weights extracted from questions. (Diagrams of the system modules have been provided; however, the specific application scenario diagram is currently unclear. For example, a software interface diagram cannot be provided. The attached diagrams are sufficient to clearly illustrate the patent content.) Detailed Implementation

[0022] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0023] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0024] like Figure 1 As shown, this application provides an intelligent practice content generation method that dynamically adjusts the weights of question extraction, specifically including:

[0025] S1: Obtain the user's historical answer record data, and construct a question proficiency database including a question proficiency level field based on the question identifier and answer result in the answer record data.

[0026] S2: Receive the real-time answer result signal and the corresponding answer duration value of the current question, determine the correctness of the answer based on the real-time answer result signal, and if the answer is determined to be incorrect, perform downgrade processing on the corresponding question proficiency level in the question proficiency database and generate a downgraded proficiency level.

[0027] S3: If the answer is determined to be correct, the answer duration value is compared and matched with a preset time threshold interval set. The proficiency upgrade range is determined according to the matched time threshold interval. Based on the upgrade range, the corresponding question proficiency level in the question proficiency database is upgraded to generate an upgraded proficiency level.

[0028] S4: Write the downgraded proficiency level or the upgraded proficiency level into the question proficiency database to overwrite the original data, and synchronously record the learning progress log data to form a complete dynamic evaluation closed-loop data source.

[0029] S5: The dynamic evaluation closed-loop data source is combined with the priority classification labels of questions in the question bank to construct a differentiated extraction weight allocation model.

[0030] S6: Use the differentiated extraction weight allocation model to perform weighted random sampling on the candidate question set in the question bank to generate an initial question sequence, and count whether the actual number of questions in the initial question sequence meets the preset question number parameter requirements.

[0031] S7: If the actual number of questions does not meet the requirement of the number of questions to be drawn, the number of questions will be supplemented and incorporated into the initial question sequence to obtain the final intelligent practice content.

[0032] Step S1: Obtain user's historical answer record data, and based on the question identifiers and answer results in the answer record data, construct a question proficiency database including a question proficiency level field. Specifically, this includes:

[0033] S1.1: Obtain user's historical answer record data, parse the original log entries in the data, extract the question identifier field and the answer result status field, and generate a set of structured answer feature vectors to establish the basic data object for subsequent proficiency assessment.

[0034] Read the user's historical answer log file from the storage medium. The original log file contains unstructured text stream data or semi-structured JSON format data strings as the initial input source for data processing.

[0035] The log parsing engine is invoked to perform line-by-line scanning and syntax verification on the original log file, identifying and removing incomplete log entries caused by network transmission interruptions or system anomalies, ensuring the integrity and legality of the data to be processed.

[0036] For valid log entries that pass the verification, use regular expression matching or JSON path query language to locate and extract key fields, including the question identifier field that uniquely identifies the question, the binary answer result status field that represents the correctness of the answer, and the timestamp difference field that records the answer time.

[0037] The extracted discrete field data undergoes type conversion processing: the string-format question identifier is mapped to an integer index ID, the text-format answer result "correct" or "incorrect" is mapped to a boolean value 1 or 0, and the time string is converted to a millisecond-level value.

[0038] Construct a structured question-answering feature vector, where each vector element corresponds to an independent question-answering event, including the question identifier, the question-answering result flag, and the question-answering duration value, forming an ordered sequence of feature vectors.

[0039] Data cleaning is performed on the feature vector sequence to remove abnormally fast answering records with answering time less than a preset minimum threshold (e.g., 1 second) and abnormally delayed answering records with answering time greater than a preset maximum threshold (e.g., 300 seconds), thereby eliminating the interference of noisy data on the proficiency assessment model.

[0040] Through the above-described parsing, extraction, transformation, and cleaning processes, the unstructured raw logs from the previous step are transformed into a standardized set of structured answer feature vectors, realizing the transformation of multi-source heterogeneous data into a unified evaluation data object, and providing high-precision and highly consistent basic data support for the subsequent construction of a question proficiency database.

[0041] S1.2: Based on the structured answer feature vector set, the duplicate question identifier fields are deduplicated and merged using the uniqueness mapping method to generate a deduplicated independent question index list, so as to ensure that the evaluation object of each knowledge point in the question proficiency database is unique.

[0042] Receive the set of structured answer feature vectors generated in the previous steps. This set contains multiple historical answer records, where the same question identifier may correspond to multiple answer behaviors.

[0043] Traverse the structured answer feature vector set, extract the question identifier field from each record, and construct a temporary mapping table to count the frequency of occurrence of each question identifier and the associated answer state sequence.

[0044] For duplicate question identifiers that appear more than once, perform a data aggregation operation to arrange multiple answers to the same question in timestamp order, forming a complete answer history for that question.

[0045] Based on a pre-defined uniqueness mapping method, feature fusion processing is performed on the answer history trajectory corresponding to duplicate question identifiers, and the most recent answer status or comprehensive historical accuracy rate is extracted as the representative feature index of the question.

[0046] The fused representative feature indicators are re-bound to the original question identifiers to generate a deduplicated list of independent question indexes, ensuring that each question identifier in the list retains only one cleaned and aggregated standard record.

[0047] By using unique mapping and data aggregation processing, the redundant structured answer feature vector set from the previous step is transformed into an independent question index list with no duplicates, thus establishing the uniqueness of the question proficiency assessment object and providing a standardized input data foundation for subsequent database instantiation.

[0048] S1.3: Based on the independent question index list, instantiate a question proficiency database storage structure, define a question proficiency level field in the storage structure, set the initial value of the field to an undefined status flag, and construct an empty question proficiency database framework with dynamic update capability.

[0049] The system receives the deduplicated list of independent question indexes generated in the previous steps as the structured input object for constructing the question proficiency database. Based on a relational database architecture or key-value pair storage model, it instantiates the physical storage structure of the question proficiency database, defining the primary key field as the question identifier to ensure the uniqueness and efficiency of data retrieval. It then extends the storage structure with a question proficiency level field, which characterizes the user's mastery of a specific question. The data type is set to integer or enumeration to support subsequent numerical comparisons and state transition operations. Each question identifier in the independent question index list is traversed, and a corresponding record entry is created in the question proficiency database. The initial value of the question proficiency level field is uniformly set to an undefined state flag, such as Null or a specific negative value, to distinguish between evaluated and unevaluated questions. A strong association constraint is established between the question proficiency level field and the question identifier, ensuring that each question instance has an independent state storage space, forming a dynamically updated empty question proficiency database framework. By using the above database instantiation and field initialization methods, discrete question indexes are transformed into structured state storage objects, realizing the empty modeling of users' knowledge mastery status, and providing a standardized data container for subsequent cold start strategy execution and real-time dynamic proficiency assessment.

[0050] S1.4: Traverse the undefined status identifier entries in the empty question proficiency database framework, and replace the undefined status identifiers with default low-level status values ​​in batches according to the preset cold start strategy, so as to generate a pre-initialized question proficiency database containing the initial proficiency benchmark value.

[0051] The system iterates through undefined status identifier entries in the empty question proficiency database framework, extracting a set of records to be initialized containing question identifiers. User historical behavior data mining is performed on this set of records to retrieve historical answer logs for specific question identifiers. If historical answer logs are detected, the accuracy rate and average answer time parameters in the logs are parsed. Based on the parsed accuracy rate and average answer time, the four-level proficiency mapping method is called to calculate the initial proficiency level value. For newly added question records without historical answer logs, a system-preset default low-level status value is directly assigned, corresponding to Level-0 (unproficient). The calculated initial proficiency level value or default low-level status value is written into the proficiency level field of the corresponding question. The written data undergoes type validation to ensure that all proficiency level fields are integer values ​​within the valid range of 0 to 3. Through batch replacement processing, the empty framework from the previous step is transformed into a pre-initialized question proficiency database containing initial proficiency benchmark values, enabling rapid modeling of user knowledge states in a cold start state.

[0052] S1.5: Perform an integrity check operation on the pre-initialized question proficiency database to verify the consistency of the association between the question identifier field and the question proficiency level field, and output the final usable question proficiency database as the sole data source for subsequent real-time answer result signal processing and proficiency upgrade / downgrade calculation.

[0053] Step S2: Receive the real-time answer result signal and corresponding answer duration value for the current question; determine the correctness of the answer based on the real-time answer result signal; if the answer is determined to be incorrect, perform a downgrade processing on the corresponding question proficiency level in the question proficiency database, generating a downgraded proficiency level. Specifically, this includes:

[0054] S2.1: Obtain the real-time answer result signal and the corresponding answer duration value of the current question, perform Boolean logic verification processing on the real-time answer result signal to generate a correct answer judgment flag bit, wherein the correct answer judgment flag bit is used as the trigger condition for subsequent branch logic execution.

[0055] The system receives real-time answer result signals and corresponding millisecond-level answer duration values ​​reported by the client, serving as the raw input data source for this step. The real-time answer result signals undergo standardized parsing to extract the user identifier, question identifier, and original answer status code, constructing an answer event object to be verified. Based on a preset standard answer database index, a set of correct options matching the current question identifier is retrieved, establishing a truth comparison benchmark. The user-submitted original answer status code is XORed bit-by-bit with the set of correct options in the standard answer database, generating an initial intermediate variable for truth / false determination. This intermediate variable undergoes non-empty verification and type consistency checks to eliminate interference from abnormal signals caused by network packet loss or data format errors. If an empty or invalid answer status code is detected, an exception handling mechanism is triggered, marking the answer result as invalid and terminating subsequent processes; if the data is valid, the intermediate variable for truth / false determination is mapped to a binary answer correctness flag. A flag value of 1 indicates a correct answer, and a flag value of 0 indicates an incorrect answer. This flag serves as a crucial trigger for subsequent branch logic execution, directly determining whether the system enters the proficiency downgrade or upgrade processing path. Through the aforementioned Boolean logic verification and anomaly filtering methods, the unstructured real-time answer signal is transformed into a deterministic binary answer correctness judgment flag. This achieves precise quantification and noise isolation of the user's answer results, ensuring the accuracy and robustness of the input data for the subsequent dynamic proficiency evaluation model.

[0056] S2.2: Based on the state that the correctness judgment flag is incorrect, read the current proficiency level of the question to be updated from the question proficiency database, and perform index matching and level decrement operation on the current question proficiency level using a preset downgrade mapping rule table to generate a target downgrade proficiency level, wherein the target downgrade proficiency level reflects the user's ability regression state after making an incorrect answer.

[0057] Based on the state where the correctness judgment flag is incorrect, the current proficiency level value of the question to be updated is retrieved from the question proficiency database to establish the baseline input parameters for level adjustment.

[0058] The pre-stored downgrade mapping rule table is invoked. This rule table defines the state transition logic in the four-level proficiency level system, where Level-3 represents mastery, Level-2 represents proficiency, Level-1 represents relative proficiency, and Level-0 represents lack of proficiency.

[0059] Using the current proficiency level of the question as the index key, perform an exact match query in the downgrade mapping rule table to obtain the corresponding target downgrade proficiency level identifier.

[0060] If the current level is Level-3, it will be updated to Level-2 according to the mapping rules, reflecting a slight regression in the user's mastery of the knowledge point.

[0061] If the current level is Level-2, it will be updated to Level-1 according to the mapping rules, indicating that the user's proficiency has dropped to the relatively proficient range.

[0062] If the current level is Level-1, it will be updated to Level-0 according to the mapping rules, indicating that the user's mastery of the question has deteriorated to unfamiliarity.

[0063] If the current level is already Level-0, according to the boundary preservation principle, Level-0 will be maintained to prevent the proficiency level from becoming negative or invalid.

[0064] By using table lookup matching and state transition processing, the error signal from the previous step is transformed into specific target downgrade proficiency level data, thereby realizing the regression correction of the user's knowledge mastery status and ensuring that the proficiency model can accurately respond to the user's negative feedback.

[0065] S2.3: Perform boundary constraint detection processing on the target downgraded proficiency level. If the target downgraded proficiency level is detected to be lower than the minimum proficiency threshold defined by the system, the target downgraded proficiency level is forcibly reset to the minimum proficiency level to generate the final downgraded proficiency level, ensuring that the problem proficiency level is always within the effective four-level system.

[0066] The system receives the target downgraded proficiency level value generated from the previous steps and logically compares it with the lowest threshold constant in the system's predefined four-level proficiency level system. This lowest threshold constant corresponds to the Level-0 state, representing the user's basic cognitive level of being completely unfamiliar with or unaware of the question.

[0067] A boundary constraint detection logic module is constructed to read the target downgraded proficiency level value and perform a less-than comparison operation to determine whether the value has fallen below the minimum threshold. If the target downgraded proficiency level value is detected to be strictly less than the integer identifier value corresponding to Level-0, the current level state is determined to be out of bounds, triggering a forced reset mechanism.

[0068] The level truncation function is invoked to directly assign the out-of-bounds target proficiency level value to the standard identifier value of Level-0. This operation eliminates the risk of negative level overflow caused by consecutive incorrect answers, ensuring that the proficiency model always operates within the non-negative integer domain.

[0069] If the detected target downgrade proficiency level value is greater than or equal to the integer identifier value corresponding to Level-0, then this value is maintained and directly marked as the final downgrade proficiency level. This branch processing ensures the data integrity of the normal downgrade path and avoids unnecessary computational overhead.

[0070] Output the final downgraded proficiency level after boundary validation. This level data has a definite physical meaning, that is, it represents the lowest effective mastery level that the user is in after making the current incorrect answer.

[0071] By using boundary constraint detection and forced reset processing, illegal negative values ​​or out-of-bounds levels that may have been generated in the previous step are transformed into valid level data that conforms to the four-level system specifications, thereby enhancing the robustness of the proficiency assessment model and ensuring the closure of the state space.

[0072] For example, four proficiency levels are defined as Level-3, Level-2, Level-1, and Level-0, corresponding to integer values ​​3, 2, 1, and 0, respectively. When a user's current proficiency level is Level-0, a wrong answer triggers a downgrade rule, and the target downgrade proficiency level is calculated as -1. The system performs boundary detection, recognizing that -1 is less than the minimum threshold of 0, and thus triggers a reset mechanism, forcibly correcting the final downgrade proficiency level to 0. If the user's current proficiency level is Level-1, and the target downgrade level is 0 after a wrong answer, the system checks that 0 is not less than 0, so the final downgrade proficiency level remains 0. This mechanism ensures the numerical stability of fields in the database and prevents abnormal data from contaminating subsequent weight calculation modules.

[0073] S2.4: Write the final downgraded proficiency level into the question proficiency database to overwrite the original current question proficiency level, and simultaneously generate a status update instruction containing the error event identifier of this answer and the final downgraded proficiency level to complete the dynamic correction operation of the question proficiency model.

[0074] It receives the final downgraded proficiency level value and the corresponding unique question identifier generated by the preceding boundary constraint detection process, and constructs a data payload object containing state update instructions. This data payload object encapsulates the target level value to be written, the timestamp index, and the operation type identifier, ensuring the atomicity and consistency of the database transaction.

[0075] Establish a connection session with the question proficiency database storage engine and locate the specific data record row pointed to by the question identifier. Use the primary key index mechanism to quickly retrieve the target record, lock the memory address or disk sector where the current question proficiency level field is located, and prepare to perform an overwrite operation.

[0076] The logic for overwriting and updating field values ​​is executed, replacing the original current proficiency level value with the final downgraded proficiency level value. During this process, the database's transaction log recording mechanism is triggered, generating Before-Image (original state image) and After-Image (updated state image) snapshots to support potential data rollback and audit trail requirements.

[0077] A status update instruction package containing the identifier of the current answer error event is constructed synchronously. This instruction package integrates the question identifier, the final downgraded proficiency level, the boolean value of the answer result, and the original data of the answer duration to form a structured log entry.

[0078] The status update command package is pushed to the learning progress log storage module and persistently saved via asynchronous I / O channel. This operation ensures the integrity of the dynamic evaluation closed-loop data source, enabling the subsequent question selection strategy module to calculate weights based on the latest user ability status.

[0079] By combining database overwrite updates with asynchronous log persistence, the degradation results calculated in the previous step are transformed into real-time status data and traceable historical behavior logs in the question proficiency database, thereby realizing real-time dynamic correction and data closed-loop maintenance of the user's knowledge mastery status model.

[0080] like Figure 2As shown, step S3: If the answer is determined to be correct, the answer duration value is compared and matched with a preset time threshold interval set. The proficiency upgrade level is determined based on the matched time threshold interval. Based on the upgrade level, the corresponding question proficiency level in the question proficiency database is upgraded to generate an upgraded proficiency level. Specifically, this includes:

[0081] S3.1: Obtain the real-time answer duration value and the preset time threshold interval set. Based on the lower limit threshold and upper limit threshold parameters contained in the preset time threshold interval set, perform boundary matching processing on the real-time answer duration value to generate time matching result data containing the target interval identifier.

[0082] S3.2: Receive the target interval identifier in the time matching result data, and perform a table lookup retrieval process on the target interval identifier based on the pre-stored interval-to-amplitude mapping rule table to generate a quantitative upgrade amplitude parameter representing the proficiency improvement level.

[0083] The system receives the time matching result data output from step S3.1, which includes a target interval identifier identifying the interval to which the current answering time belongs. The target interval identifier is parsed and input as an index key into a pre-stored interval-to-amplitude mapping rule table. This mapping rule table, stored in system memory, defines a static correspondence between different time threshold intervals and the proficiency upgrade level. A table lookup operation is performed, traversing the index columns of the mapping rule table to locate records that perfectly match the target interval identifier. The corresponding quantitative upgrade level parameter value is extracted from the matching record. This parameter value is an integer representing the number of levels the user's proficiency level should increase under the current time performance. The extracted quantitative upgrade level parameter is validated to ensure its value is non-negative and does not exceed the maximum single jump step size allowed by the system. If the validation passes, the quantitative upgrade level parameter is encapsulated into a standard data object. Through table lookup and validity validation, the time interval matching result from the previous step is transformed into a quantitative upgrade level parameter representing the proficiency upgrade level, achieving the expected technical effect of refined ability assessment based on answering efficiency.

[0084] S3.3: Extract the original proficiency level field value of the current question from the question proficiency database, and combine it with the quantitative upgrade range parameter. Based on the capping constraint logic of the four-level proficiency level system, perform level transition calculation processing on the original proficiency level field value to generate candidate upgrade proficiency level values.

[0085] The original proficiency level field value of the current question is read from the question proficiency database and parsed into an integer status identifier, which is used as the baseline input for level jump calculation.

[0086] Obtain the quantitative upgrade magnitude parameter generated in the previous steps. This parameter represents the step size of the proficiency improvement level determined based on the matching results of the answer duration.

[0087] A level transition calculation logic is constructed, which performs arithmetic addition on the original proficiency level value and the quantified upgrade magnitude parameter to generate preliminary candidate upgrade proficiency level values.

[0088] The following formula is used to perform level transition calculations:

[0089]

[0090] in, The original proficiency level. To quantify the upgrade magnitude parameter, Candidates for upgrading proficiency level.

[0091] S3.4: Obtain the candidate upgrade proficiency level value and the system-defined highest proficiency level threshold. Based on the maximum value truncation method, perform out-of-bounds verification processing on the candidate upgrade proficiency level value to generate a target upgrade proficiency level that conforms to the level system specification.

[0092] Obtain the candidate upgrade proficiency level values ​​and the system-defined highest proficiency level threshold as input parameters for out-of-bounds verification.

[0093] Read the upper limit of the four-level proficiency system preset in the system configuration file. This value is fixed at Level-3, which represents the state of complete mastery.

[0094] Construct a maximum value truncation logic judgment unit to compare the candidate upgrade proficiency level values ​​with the highest proficiency level threshold.

[0095] If the candidate proficiency level value is less than or equal to the highest proficiency level threshold, it is determined that no boundary has been crossed, and the candidate value is directly confirmed as the target proficiency level.

[0096] If the candidate's proficiency level value is greater than the highest proficiency level threshold, it is determined that a positive out-of-bounds event has occurred, triggering the capping protection mechanism.

[0097] Perform a forced assignment operation to reset the target's proficiency level to the highest proficiency level threshold, i.e., Level-3.

[0098] By using the maximum value truncation method, the potential overflow level data generated in the previous step is transformed into the target upgrade proficiency level that conforms to the four-level system specification, thereby realizing the boundary constraints and standardized output of the proficiency level status and ensuring the data legality of subsequent database write operations.

[0099] S3.5: Read the target proficiency level and the corresponding question identifier information, and based on the database writing protocol, perform an overwrite update process on the status field of the corresponding record in the question proficiency database to generate an updated question proficiency database record containing the latest mastery status information.

[0100] like Figure 3 As shown, step S4 involves writing the downgraded proficiency level or the upgraded proficiency level into the question proficiency database to overwrite the original data, and simultaneously recording learning progress log data to form a complete dynamic evaluation closed-loop data source. Specifically, this includes:

[0101] S4.1: Obtain the downgraded or upgraded proficiency level and the corresponding question identifier generated by the previous steps, and perform an overwrite update process on the target record in the question proficiency database based on the database writing protocol to generate an updated question proficiency dataset with the latest status marker.

[0102] Receive the target proficiency level value and corresponding unique question identifier from the previous steps, and construct a data payload object containing status update instructions. This data payload object serves as the input parameter for database transaction operations, ensuring the atomicity and consistency of subsequent write operations.

[0103] Based on the question identifier, a primary key index retrieval is performed in the question proficiency database to locate the corresponding specific record row in the storage structure. The version control field of the current record and the original proficiency level value are read to establish the baseline state before data overwriting.

[0104] The update interface in the database write protocol is called to map the target proficiency level value to the proficiency level field in the record, while updating the last modified timestamp field to the current system time, and generating a transaction change set to be committed.

[0105] Perform a transaction commit operation to persistently write the transaction change set to the physical storage medium. If the write is successful, a confirmation signal is returned. If a conflict occurs, a retry mechanism or exception handling process is triggered to ensure data consistency and integrity.

[0106] Read the completed question proficiency database records, verify whether the proficiency level field value is consistent with the target proficiency level value, and generate an updated question proficiency dataset with the latest status marker, which serves as the authoritative data source for subsequent log recording and weight calculation.

[0107] Through the aforementioned database transaction processing and status verification mechanism, the dynamic evaluation results calculated in the previous step are transformed into persistent user knowledge status data, enabling real-time and accurate updates to the question proficiency model and providing a reliable data foundation for building a complete dynamic evaluation closed loop.

[0108] S4.2: Read the real-time answer result signal, answer duration value, and updated question proficiency dataset output in step S4.1 during the current question-answering process. Use the log serialization method to encapsulate the above multi-source heterogeneous data to generate a structured learning progress original log entry containing a complete answer context.

[0109] Read the updated question proficiency dataset output in step S4.1, and extract the question identifier, the latest proficiency level value, and the status update timestamp contained therein, as the basic metadata object for log encapsulation.

[0110] Obtain the real-time answer result signal in the current answer session, convert it into a boolean type answer correctness flag, and associate it with the corresponding user ID and session ID (a unique identifier generated by the server to identify and track the session state between the user and the website, usually stored in the browser) to build an answer behavior context index.

[0111] Accurate timing data was collected during the question-answering process. The original millisecond-level timestamp difference was standardized into a question-answering duration value in seconds, retaining two decimal places of precision to eliminate measurement noise interference.

[0112] Call the log serialization method interface to initialize the JSON structure container, and map the question identifier, the correctness flag, the answer duration, and the latest proficiency level to predefined data key-value pairs.

[0113] A multi-dimensional data validation mechanism is introduced to check the non-nullability and type consistency of each field in the encapsulated data. If an abnormal missing field is detected, a default invalid value is filled in to ensure the integrity of the log structure.

[0114] The validated structured data is encoded using UTF-8 to generate a binary byte stream containing the complete response context, and then serialized into a storable structured learning progress log entry.

[0115] By using log serialization, the multi-source heterogeneous data from the previous step is transformed into standardized, structured raw log entries of the learning progress, thus enabling persistent storage of answer behavior data and laying the foundation for subsequent time-series analysis.

[0116] S4.3: Receive the original structured learning progress log entries, perform time-series alignment and integrity verification on the entries based on the timestamp index mechanism, and generate valid learning progress log data units with unique index identifiers.

[0117] The system receives raw, structured learning progress log entries generated by a log serialization method. Each entry includes a question identifier, a boolean value for the answer result, a millisecond-level answer duration, and an updated proficiency level. A globally unique timestamp field is extracted from the log entry and used as the primary key index for time-series alignment. This timestamp is then differentially calculated with the timestamp of the last record in the user's historical answer trajectory database. If the current timestamp is detected to be earlier than or equal to the timestamp of the last record, it is considered a time-series anomaly or duplicate submission. A data cleaning mechanism is triggered to discard the anomalous entry and generate an error flag. For log entries that pass time-series verification, a field integrity scan is performed to verify that the question identifier is not empty, the answer duration is within a reasonable physical range, and the proficiency level is within the preset four-level system. For semi-structured data lacking key fields, missing values ​​are filled in according to the default filling strategy, or the data is marked as awaiting manual review to ensure the atomicity and consistency of data units. A unique index identifier based on a hash algorithm is generated for valid log data units that pass integrity verification. This identifier is generated by encrypting the user ID, question ID, and timestamp to prevent redundant data storage. Valid learning progress log data units with unique index identifiers are written to a temporary buffer, ready for subsequent incremental aggregation processing. Through the above-described time-series alignment and integrity verification methods, the original heterogeneous logs generated in the previous step are transformed into valid learning progress log data units with spatiotemporal consistency and data integrity, achieving the expected technical effect of high reliability and traceability of the closed-loop data source for dynamic evaluation.

[0118] S4.4: Extract the final proficiency level distribution features from the effective learning progress log data units, combine them with the user's historical answer trajectory data, and perform dynamic fusion and update processing on the user's overall learning progress data to generate a panoramic view of the user's learning status that reflects the current overall knowledge mastery.

[0119] Extract the proficiency level identifiers from the valid learning progress log data units and construct a local proficiency state vector within the current time window. Perform dimension mapping on this local proficiency state vector to convert it into a frequency statistical distribution corresponding to the four proficiency levels.

[0120] The system retrieves the global proficiency distribution baseline value stored in the user's historical answer trajectory database. This baseline value represents the steady-state characteristics of the user's knowledge mastery over a long period. Based on the exponentially weighted moving average method, a dynamic fusion model is established between the global baseline value and the local frequency statistical distribution.

[0121] The updated overall user learning progress metric is calculated using the following formula:

[0122]

[0123] in, The updated overall user learning progress metrics. This represents the local proficiency state vector within the current time window. This serves as the baseline value for the historical global proficiency distribution. This is the forgetting factor coefficient, with a value ranging from 0 to 1.

[0124] For the calculated Normalization is performed to eliminate the influence of dimensional differences caused by variations in the number of questions for different knowledge points, generating a standardized panoramic view data matrix of user learning status.

[0125] By using an exponentially weighted moving average, the local answer results from the previous step are transformed into a panoramic view of the user's learning status that reflects long-term trends. This enables smooth processing and precise tracking of fluctuations in the user's knowledge mastery, significantly improving the stability of the evaluation results.

[0126] S4.5: Integrate the updated question proficiency dataset with the panoramic view of user learning status, and push the above data stream to the dynamic evaluation closed-loop data source storage area based on the data bus broadcast mechanism to generate a complete dynamic evaluation closed-loop data source that can be directly called by the subsequent question extraction strategy module.

[0127] Step S5: The dynamic evaluation closed-loop data source, combined with the priority classification tags of questions in the question bank, constructs a differentiated extraction weight allocation model. Specifically, this includes:

[0128] S5.1: Aggregate the learning progress log data in the dynamic evaluation closed-loop data source, extract the final proficiency level identifier for each question, and generate a proficiency level distribution dataset containing the mapping relationship between question identifiers and corresponding proficiency levels, so as to establish the quantitative basis for the current user's knowledge status.

[0129] Receive structured learning progress log data from the dynamic evaluation closed-loop data source. This data includes user ID, question identifier, answer timestamp, and latest proficiency level status.

[0130] The learning progress log data is subjected to a partitioning and aggregation operation based on user ID. The hash index mechanism is used to quickly locate the most recent valid answer record of a specific user, and redundant historical logs are removed to ensure data timeliness.

[0131] Extract the question identifier field and the corresponding final proficiency level field from the aggregated records, and construct a raw mapping set in the form of key-value pairs, where the key is a unique question identifier and the value is an enumerated value of the four proficiency levels.

[0132] Traverse the entire metadata set of the question bank, identify missing question identifiers in the original mapping set as questions that have not been asked, and assign them specific undefined level identifiers to complete the full coverage of the user's knowledge graph.

[0133] For questions with undefined level labels, a state transition is performed, mapping them uniformly to the "not done" category, while retaining the answered question categories from Level-0 to Level-3, forming a five-class discrete proficiency distribution dataset.

[0134] Through the data cleaning and state mapping processes described above, the multi-source heterogeneous logs from the previous step are transformed into a well-structured and categorized proficiency level distribution dataset, thereby establishing a quantitative foundation for the current user's knowledge state and providing accurate input for subsequent weight allocation.

[0135] S5.2: Based on the proficiency level distribution dataset, the priority classification label of each question is obtained by calling the question bank metadata index. The proficiency level identifier is associated and matched with the priority classification label to generate a multi-dimensional attribute question set that integrates the user's mastery level and the importance of the question, so as to construct the input object for weight calculation.

[0136] Receive the proficiency level distribution dataset generated in step S5.1. This dataset contains the mapping relationship between question identifiers and corresponding final proficiency levels, and serves as the initial input object for weight calculation.

[0137] Call the question bank metadata index interface to retrieve question attribute information in batches based on question identifiers in the dataset, and extract the priority classification label for each question. The priority classification label represents the importance of the question in the knowledge system or the frequency of examination.

[0138] The extracted priority category tags are matched with the proficiency level identifiers at the field level to construct a multi-dimensional attribute question set containing question identifiers, proficiency levels, and priority tags, thereby achieving data fusion between the user's mastery status and the inherent attributes of the questions.

[0139] Perform integrity checks on the multi-dimensional attribute question set, remove dirty data records with missing priority labels or abnormal proficiency levels, and ensure that subsequent weight allocation is based on high-quality structured data.

[0140] Through the above data processing flow, single proficiency level data is transformed into a multi-dimensional set of questions that integrates the user's dynamic mastery level and the static importance of the questions. This provides a precise input object for building a differentiated extraction weight allocation model, significantly improving the scientific nature and pertinence of the question extraction strategy.

[0141] S5.3: The questions in the multi-dimensional attribute question set are grouped and clustered according to the status of never done, level 0 proficiency, level 1 proficiency, level 2 proficiency and level 3 proficiency. The preset weight configuration rules are used to assign a specific initial probability weight value to each group, and the original weight allocation scheme divided by proficiency level is generated to achieve differentiated treatment of questions with different levels of mastery.

[0142] The system receives a multi-dimensional set of questions with attributes that integrate user proficiency and question importance, serving as input data for weight allocation. For each question in the multi-dimensional set, a status identifier is parsed to extract its corresponding proficiency level field value, dividing the questions into five mutually exclusive logical groups: never done before, level 0 proficiency, level 1 proficiency, level 2 proficiency, and level 3 proficiency. A pre-defined weight configuration rule table is established, defining the basic importance coefficients of different proficiency groups in the question selection strategy. Questions never done before and level 0 proficiency questions are assigned higher initial weights to strengthen new knowledge learning and consolidate weak areas, while high-proficiency questions are assigned lower weights to maintain memory retention and avoid ineffective repetition. Each logical group is traversed, and the number of questions in each group is counted. If the number of questions in a group is zero, its initial probability weight value is temporarily stored as zero to prevent division by zero errors or null pointer exceptions in subsequent calculations. For non-empty groups, the corresponding baseline weight value is read from the weight configuration rule table. This baseline weight value is multiplied by the number of questions in the group to generate the original weighted total value for that group, thus reflecting the impact of the number of questions on the overall sampling probability. The original weighted total values ​​of all groups are summed to obtain the global weight sum, which is used for subsequent normalization processing. Through the above group clustering and baseline weight mapping processing method, the multi-dimensional attribute question set of the previous step is transformed into an original weight allocation scheme divided by proficiency level, realizing differentiated treatment of questions with different levels of mastery, and laying the data foundation for building an accurate personalized question sampling model.

[0143] For example, suppose the current multi-dimensional attribute question set contains 100 candidate questions. After analysis, there are 20 questions that have never been asked, 30 questions at level 0 proficiency, 25 questions at level 1 proficiency, 15 questions at level 2 proficiency, and 10 questions at level 3 proficiency. According to the preset weight configuration rules, the baseline weight for questions that have never been asked is set to 3.0, level 0 to 3.0, level 1 to 2.0, level 2 to 1.0, and level 3 to 1.0. The original weighted total value for each group is calculated as follows: the group that has never been asked is 20 multiplied by 3.0, which equals 60.0; the group at level 0 is 30 multiplied by 3.0, which equals 90.0; the group at level 1 is 25 multiplied by 2.0, which equals 50.0; the group at level 2 is 15 multiplied by 1.0, which equals 15.0; and the group at level 3 is 10 multiplied by 1.0, which equals 10.0. The total global weight is 60.0 + 90.0 + 50.0 + 15.0 + 10.0, which equals 225.0. In the generated original weight allocation scheme, the group of questions that have not been asked before occupies the highest weight ratio, which significantly increases the exposure rate of new questions, effectively solves the problem of imbalance between the ratio of new and old questions in traditional random question selection, and greatly improves learning efficiency.

[0144] S5.4: Based on the initial probability weight values ​​in the original weight allocation scheme, perform normalization verification and dynamic balancing methods to adjust the weight ratio of each group to ensure that the sum is one and meets the constraints on the number of items to be sampled, and generate an optimized standardized differentiated sampling weight allocation model to output the final decision parameters that can be directly used for weighted random sampling.

[0145] Receive the original weight allocation scheme generated by S5.3, extract the initial probability weight values ​​and corresponding number of questions for each group of untested and proficiency levels from level 0 to level 3, and construct the weight vector and sample capacity matrix to be verified.

[0146] Perform a sum consistency check on the original weight vector, calculate the sum of the initial weights of all groups, and determine whether it is strictly equal to the unit probability value of 1. If there is a floating-point operation deviation, record the normalization correction coefficient.

[0147] A dynamic balancing method is introduced to perform weight truncation on proficiency groups with scarce question quantity. When the number of available questions in a certain group is less than the theoretical expected extraction value calculated based on the current weight, the upper limit of the weight of that group is constrained to the actual question proportion.

[0148] An iterative proportional fitting method is adopted to redistribute the remaining weight quotas while satisfying the hard constraint of the number of questions in each group. This ensures that the weights of high-priority groups with insufficient questions do not overflow, while proportionally compensating for the extraction probability of other groups.

[0149] Perform a global normalization operation, divide the adjusted weights of each group by the new sum of weights, and generate a standardized differential extraction weight allocation model to ensure that the sum of the extraction probabilities of each level of questions in the final output strictly converges to 1.

[0150] Through the above normalization verification and dynamic balancing process, the original weight configuration of the previous step is transformed into standardized decision parameters that conform to the actual distribution constraints of the question bank and are probability-conserving, thereby achieving a precise match between theoretical preferences and actual resources for the question selection strategy.

[0151] For example, suppose the initial weights of unsolved questions and Level-0 to Level-3 questions are 0.3, 0.3, 0.2, 0.1, and 0.1, respectively, corresponding to 5, 10, 50, 100, and 200 available questions in the question bank. If the preset number of questions to be drawn is 20, the theoretical expectation for Level-0 is 6 questions, and 10 questions are actually available, satisfying the constraint. If only 2 questions remain for a certain less popular knowledge point at Level-0, a truncation mechanism is triggered, adjusting its upper limit of weight to 2 / 20 = 0.1. The 0.2 difference in the original weight of 0.3 is distributed to other groups in a ratio of 0.3:0.2:0.1:0.1. After iterative calculation and normalization, the final weight vector is output, ensuring that the probability distribution of question drawing remains valid and optimal even in scenarios with limited question resources.

[0152] Step S6: Using the differentiated extraction weight allocation model, perform weighted random sampling on the candidate question set in the question bank to generate an initial question sequence, and count whether the actual number of questions in the initial question sequence meets the preset question quantity parameter requirement. Specifically, this includes:

[0153] S6.1: Obtain the probability weight values ​​of unsolved questions, zero-level proficiency questions, first-level proficiency questions, second-level proficiency questions, and third-level proficiency questions output by the differentiated extraction weight allocation model. Combine these with the question priority classification tags in the question bank metadata to perform hierarchical mapping processing on the candidate question set, so as to generate a hierarchical candidate question index set containing weight attributes.

[0154] Step S6.1 aims to transform the abstract weight allocation model into a concrete set of question indexes, providing a structured data foundation for subsequent weighted sampling. It receives the differentiated extraction weight allocation model output from the preceding step S5, which includes probability weight values ​​for questions not yet answered and questions at proficiency levels zero through three. It calls the question bank metadata index interface to obtain the unique identifier, current proficiency level status, and priority classification label for each question in the candidate question set. It iterates through the candidate question set, mapping each question to its corresponding proficiency group category based on its proficiency level status field. For questions not yet answered, it marks their group as "New Question Group" and associates it with the probability weight value of questions not yet answered. For questions at proficiency level-0, it marks their group as "Weak Group" and associates it with the probability weight value of level-0 proficiency questions. For questions at proficiency level-1, it marks their group as "Advanced Group" and associates it with the probability weight value of level-1 proficiency questions. For questions at proficiency level 2, their group is labeled "Consolidation Group," and associated with the probability weight value of level 2 proficiency questions. For questions at proficiency level 3, their group is labeled "Mastery Group," and associated with the probability weight value of level 3 proficiency questions. During the mapping process, the priority classification label of each question is read synchronously, and this label is bound as an additional attribute to the corresponding question index object. A hierarchical candidate question index set is constructed, with proficiency group as the key and a list of objects containing question IDs, weight values, and priority labels as the values. Through the above hierarchical mapping process, the weight parameters from the previous step are logically bound to specific question entities, generating a hierarchical candidate question index set with clear weight attributes. This achieves precise implementation from macro-level weighting strategies to micro-level question objects, ensuring that subsequent sampling operations strictly follow the preset probability distribution.

[0155] For example, the total number of questions selected is set to 10, and the weighting model is configured as follows: 0.3 for questions not yet answered, 0.3 for Level-0, 0.2 for Level-1, 0.1 for Level-2, and 0.1 for Level-3. The system iterates through a candidate question pool containing 500 questions, identifying 50 questions not yet answered, 20 Level-0 questions, 30 Level-1 questions, 100 Level-2 questions, and 300 Level-3 questions. The system assigns the 50 questions not yet answered to a new question group with a weight of 0.3; the 20 Level-0 questions to a weak group with a weight of 0.3; the 30 Level-1 questions to an advanced group with a weight of 0.2; the 100 Level-2 questions to a consolidation group with a weight of 0.1; and the 300 Level-3 questions to a mastery group with a weight of 0.1. Simultaneously, the priority label (such as core test point, secondary test point) of each question is read and stored in an index object. The final generated hierarchical candidate question index set clearly divides the questions into five levels of subsets. Each subset of questions carries a corresponding global weight value, providing standardized input data for constructing the cumulative probability distribution table in step S6.2, which significantly improves the execution efficiency and accuracy of the question extraction strategy.

[0156] S6.2: Based on the weight attributes in the hierarchical candidate question index set, call the pseudo-random number generator to perform normalized cumulative probability calculation on the question subsets corresponding to each proficiency level, so as to construct a weighted probability distribution interval mapping table covering the entire range of candidate questions.

[0157] Receive the stratified candidate question index set, extract the number of questions and corresponding standardized weight values ​​for each proficiency level subset. For the five groups—no questions done, proficiency level zero to three—calculate the proportion of each group's weight in the total weight, and generate normalized probability distribution parameters.

[0158] Based on the normalized probability distribution parameters, a cumulative probability interval is constructed. The initial cumulative value is set to 0, and the normalized weights of each group are sequentially added to form a monotonically increasing cumulative probability sequence. This sequence divides the [0,1] interval into five non-overlapping sub-intervals, with the length of each sub-interval corresponding to the extraction probability of the corresponding proficiency level.

[0159] Call the pseudo-random number generator to generate a sequence of random numbers that follow a uniform distribution. For each question position to be extracted, generate a random floating-point number in the range [0,1) as a sampling pointer.

[0160] The sampling pointer is compared with the cumulative probability interval mapping table. Using a binary search or linear traversal, the specific sub-interval in which the sampling pointer falls is determined. The proficiency level corresponding to this sub-interval is the target level category for the current sampling round.

[0161] Within the subset of questions corresponding to the defined target level category, uniform random selection is performed. If the subset is empty, the sampling is marked as invalid and a new random number is generated or the sampling is skipped to ensure that the selected questions actually exist and meet the weight distribution requirements.

[0162] Through the above-mentioned weighted probability distribution interval mapping and random sampling processing, the weight allocation scheme of the previous step is transformed into a specific question extraction decision instruction, thereby achieving the expected technical effect of intelligent question extraction based on user proficiency differences.

[0163] For example, let the weight of a question not yet answered be 0.3, Level-0 be 0.3, Level-1 be 0.2, Level-2 be 0.1, and Level-3 be 0.1. Construct cumulative probability intervals: [0, 0.3) for questions not yet answered, [0.3, 0.6) for Level-0, [0.6, 0.8) for Level-1, [0.8, 0.9) for Level-2, and [0.9, 1.0) for Level-3. Generate a random number 0.45, falling within the interval [0.3, 0.6), and determine the target level as Level-0. Randomly select a question from the Level-0 question pool. If a random number 0.95 is generated, falling within the interval [0.9, 1.0), the target level is determined as Level-3, and a question is randomly selected from the Level-3 question pool. This process ensures that high-weight questions are extracted more frequently, significantly improving the relevance of the practice content.

[0164] S6.3: Using a weighted probability distribution interval mapping table as the sampling basis, perform multiple rounds of weighted random sampling on the hierarchical candidate question index set to generate an initial question sequence that conforms to the preset probability distribution characteristics.

[0165] The system receives a weighted probability distribution interval mapping table and a tiered candidate question index set, and initializes a blank initial question sequence container. It constructs a cumulative probability pointer and iterates through the weight intervals of the question subsets corresponding to each proficiency level. It generates a pseudo-random floating-point number conforming to a uniform distribution as a sampling anchor. It compares and retrieves the sampling anchor with the cumulative boundary values ​​in the weighted probability distribution interval mapping table. It locates the specific proficiency level weight interval identifier into which the sampling anchor falls. Within the hit weight interval, it maps the relative position to the specific question identifier under that level. It extracts the metadata information and unique identifier ID of the hit question. It executes deduplication verification logic to determine if the currently hit question already exists in the initial question sequence. If a duplicate exists, it discards the current sampling result and regenerates a pseudo-random number for the next round of sampling. If no duplicate exists, it appends the current question identifier to the end of the initial question sequence. It repeats the above sampling, location, verification, and appending operations until the length of the initial question sequence reaches a preset question quantity threshold or the available questions of the corresponding weight in the question bank are exhausted. By using weighted random sampling, the probability distribution model from the previous step is transformed into a sequence of questions that matches the user's proficiency level, thus achieving the expected technical effect of personalized and accurate question recommendation based on cognitive state.

[0166] For example, the preset number of questions is set to 10. Questions not previously answered have a weight of 30%, Level-0 has a weight of 30%, Level-1 has a weight of 20%, Level-2 has a weight of 10%, and Level-3 has a weight of 10%. A cumulative probability interval is constructed: [0, 0.3) for questions not previously answered, [0.3, 0.6) for Level-0, [0.6, 0.8) for Level-1, [0.8, 0.9) for Level-2, and [0.9, 1.0] for Level-3. The first random number 0.45 is generated, falling within the interval [0.3, 0.6), and is determined to be Level-0. Question ID_101 is randomly selected from the Level-0 candidate question pool, verified to be unique, and added to the sequence. The second random number 0.85 is generated, falling within the interval [0.8, 0.9), and is determined to be Level-2. Question ID_205 is randomly selected from the Level-2 candidate question pool, verified to be unique, and added to the sequence. The process continues until the sequence contains 10 questions. If only one question remains in the Level-3 pool and has already been selected, subsequent random numbers falling into that range will trigger a retry or skip mechanism, ensuring that the output questions are not repeated and strictly follow the weight distribution, significantly improving the scientific rigor and relevance of question selection.

[0167] S6.4: Receive the initial question sequence and the preset question quantity parameter requirements, perform sequence length traversal statistical processing to generate the actual question quantity statistics reflecting whether the current sampling results meet the standard.

[0168] Receives an initial sequence of questions object generated by a preceding weighted random sampling step. This sequence contains a list of unique identifiers for the extracted questions and their corresponding metadata attributes.

[0169] Call the sequence length retrieval interface to perform a traversal operation on the initial question sequence, count the total number of question elements actually contained in the sequence, and generate the original count value.

[0170] The system reads the preset question quantity parameter requirement, which defines the total number of standard questions required for a single practice task, serving as a benchmark threshold for quantity compliance.

[0171] The generated original count value and the preset number of questions to be drawn are loaded into the comparison logic unit, and the numerical comparison operation is performed.

[0172] If the original count value is less than the preset threshold, calculate the difference between the two and mark it as insufficient quantity; if the original count value is greater than or equal to the preset threshold, mark it as sufficient quantity.

[0173] Based on the comparison results, a status signal containing the actual number of questions is encapsulated. This signal clearly indicates whether the current sampling results meet the business constraints.

[0174] By using sequence traversal statistics and threshold comparison processing, the sampling results of the previous step are transformed into quantified actual question quantity statistics and compliance status indicators, realizing automated verification of the integrity of the intelligent question quantity, and providing accurate judgment basis for triggering the supplement mechanism or directly outputting the results.

[0175] For example, the preset question selection parameter is set to 20 questions. The system iterates through the initial question sequence and counts 18 questions. Comparing 18 with 20, if the number is insufficient, an actual question quantity of 18 is generated, along with an insufficient status marker. If the actual number of questions is 20, the number is satisfied, and an actual question quantity of 20 is generated, along with a qualified status marker. If the actual number of questions is 22, exceeding the maximum selection quantity, the excess questions are automatically removed. This process ensures precise control over the number of questions selected, significantly improving the standardization of exercise content generation.

[0176] S6.5: Based on the statistical value of the actual number of questions, a logical comparison is performed with the preset question quantity parameter requirement. If it is determined that the quantity is insufficient, a trigger signal is marked. If it is determined that the quantity is sufficient, the final initial question sequence is output to complete the sampling quantity compliance verification process.

[0177] The system receives the actual number of questions in the initial question sequence and the preset question selection parameters as input data for compliance verification.

[0178] Perform a numerical reading operation on the actual number of questions, extract the preset question quantity parameter requirements from the system configuration module, and establish a comparison benchmark threshold.

[0179] Using subtraction logic, the difference between the preset number of questions required and the actual number of questions is calculated to generate a quantity difference index.

[0180] Based on the aforementioned quantity difference index, a Boolean logic judgment condition is constructed. If the difference is greater than zero, it is determined that the current sampling result does not meet the quantity constraint.

[0181] When the quantity is determined to be insufficient, a trigger signal object is instantiated, the signal is marked as high, and associated with the control interface of the question supplementation mechanism.

[0182] If the difference is less than or equal to zero, the current sampling result is determined to meet the quantity constraint, the branch jump logic is executed, and the supplementation mechanism is skipped.

[0183] When the quantity is determined to be sufficient, the data pointer of the initial question sequence is locked to prevent subsequent operations from tampering with the generated sequence.

[0184] The locked initial question sequence is encapsulated into a standard output data packet, and an integrity check code is added to ensure the security of data transmission.

[0185] Through the above logical comparison and branching process, the statistical results of the previous step are transformed into clear execution instruction signals or the final question sequence, thereby realizing automated verification of the compliance of the number of questions drawn and process diversion.

[0186] Step S7: If the actual number of questions does not meet the required number of questions to be drawn, then supplement the number of questions and incorporate them into the initial question sequence to obtain the final intelligent practice content. Specifically, this includes:

[0187] S7.1: Compare and calculate the actual number of questions in the initial question sequence with the preset question quantity parameter requirement, and generate a quantity difference identification signal to determine the execution conditions for triggering the question supplementation mechanism.

[0188] Receives an initial question sequence object generated by a preceding weighted random sampling step. This sequence contains a set of unique identifiers for the extracted questions and the current sequence length attribute. It then calls the system's preset question quantity parameter configuration item to read the threshold for the total number of questions required for the target practice content, using this as a benchmark value for determining quantity compliance.

[0189] An integer subtraction operation is performed, subtracting the current actual length of the initial question sequence from the threshold of the total number of questions required for the target practice content, to calculate the remaining question gap. This calculation process aims to quantify the quantitative difference between the current sampling results and user needs, providing accurate data support for subsequent supplementation mechanisms.

[0190] Based on the calculated remaining question gap, a logical condition judgment operation is performed. If the remaining question gap is greater than zero, it is determined that the current number of questions drawn is insufficient, and a Boolean value of true is generated as a quantity difference indicator signal, which serves as the control command to trigger the question replenishment mechanism. If the remaining question gap is less than or equal to zero, it is determined that the number of questions drawn meets the requirements, a Boolean value of false is generated, and the replenishment process is terminated.

[0191] The generated quantity difference indicator signal is encapsulated into a standardized control message packet, which includes the specific size of the shortage value and the trigger status flag. This control message packet is sent to the question supplementation module via the internal bus to ensure that the downstream processing unit can accurately identify and respond to the insufficient quantity scenario.

[0192] Through the above comparison calculation and logical judgment processing, the sampling results of the previous step are transformed into a clear quantity difference identification signal, realizing the automatic verification of the compliance of the number of questions and the precise establishment of the triggering conditions of the supplement mechanism, thus ensuring the strict consistency of the final output number of questions.

[0193] For example, the preset number of questions to be drawn for a single practice session is 10. After weighted random sampling, the initial question sequence contains 8 questions. The system performs a subtraction operation: 10 minus 8 equals 2, resulting in a remaining question gap of 2. Since 2 is greater than 0, the system generates a Boolean value of true to indicate the quantity difference and writes the gap value of 2 into the signal payload. This signal triggers the subsequent selection of 2 high-priority questions from the remaining candidate question pool to supplement the gap, ensuring that the practice content received by the end user terminal strictly meets the requirement of 10 questions, avoiding the problem of insufficient practice due to uneven distribution of the question bank.

[0194] S7.2: Based on the quantity difference identification signal, remove the question identifiers that are already included in the initial question sequence from the entire question bank set, and construct the remaining candidate question pool as the data source object for subsequent supplementary operations.

[0195] Receive the quantity difference identifier signal generated in step S7.1. This signal carries the difference information between the current initial question sequence and the preset question quantity parameter, and serves as a trigger instruction to construct the remaining candidate question pool.

[0196] Call the full set of question bank data index to obtain a metadata set containing all available questions. The metadata set contains at least a unique question identifier field, a question content field, and an associated knowledge graph tag field, thus establishing the full range of data to be processed.

[0197] Extract the initial question sequence generated in step S6.3, parse the unique question identifiers of the selected questions in the sequence, and construct an exclusion set containing all extracted question identifiers to define the boundaries of questions that cannot be reused in this supplementary operation.

[0198] Based on the exclusion set, a set difference operation is performed on the index of the entire question bank data. A hash table is used to quickly match and remove question identifiers that exist in the exclusion set in the entire question bank, ensuring that the remaining dataset does not contain any questions that have appeared in the initial sequence.

[0199] The metadata of the remaining questions after the elimination process is re-encapsulated to generate a structured pool of remaining candidate questions. This object retains the original question priority classification tags and proficiency level attributes, serving as the data input source for priority sorting in the subsequent step S7.3.

[0200] By using the set difference operation and hash matching method described above, the result of the quantity difference judgment in the previous step is transformed into a clean and non-repeating pool of remaining candidate questions. This achieves deduplication and isolation of question resources, providing a high-quality data foundation for subsequent accurate supplementation of questions based on priority, and effectively avoiding redundancy and duplication of practice content.

[0201] For example, the preset number of questions is 20, and the initial question sequence contains 15 questions, identified as Q001 to Q015. The system receives a quantity difference indicator signal and determines that 5 questions need to be added. The entire question bank is retrieved, assuming there are 1000 questions in the question bank, identified as Q001 to Q1000. An exclusion set {Q001, Q002, ..., Q015} is constructed. The entire question bank is traversed or hashed, and the question records corresponding to these 15 identifiers are removed. The remaining candidate question pool contains 985 questions, identified as Q016 to Q1000 (assuming the first 15 were removed, it actually includes all questions except Q001-Q015). This remaining candidate question pool is passed as input to the next step, ensuring that the supplemented questions do not overlap with the initial questions, significantly improving the diversity and coverage of the practice content.

[0202] S7.3: Sort each candidate question in the remaining candidate question pool in descending order according to its priority classification label to generate an ordered sequence of candidate questions arranged in priority order, so as to establish the selection order of supplementary questions.

[0203] The system receives the remaining candidate question pool data. This dataset contains question objects that were not selected in the initial question selection sequence and their associated priority classification labels. The priority classification label represents the coreness of the question in the knowledge graph or the business recommendation weight, and is usually defined as an integer value, with a larger value indicating a higher priority.

[0204] Iterate through each question object in the remaining candidate question pool and extract its corresponding priority category label value. Establish a temporary sorting index structure, mapping the unique identifier of the question to the extracted priority value as key-value pairs, and construct a set of data tuples to be sorted.

[0205] The system invokes either quicksort or mergesort, using priority values ​​as the key comparison field, to sort the data tuple set in descending order. During the sorting process, if objects with the same priority value are encountered, a secondary sorting factor is introduced, such as the last update timestamp of the question or the question difficulty coefficient, to ensure the determinism and stability of the sorting result.

[0206] Generate an ordered sequence of candidate questions arranged from highest to lowest priority. In this sequence, high-priority questions are located at the beginning of the sequence, and low-priority questions are located at the end of the sequence, forming a linear data structure with a clear gradient characteristic.

[0207] By using a descending sorting method, the unordered remaining candidate question pool obtained in the previous step is transformed into an ordered sequence of candidate questions with a clear selection order. This achieves logical standardization of the selection of supplementary questions, ensuring that subsequent steps can efficiently extract the highest-value supplementary questions based on priority, thereby maximizing the educational value of the practice content and user benefits.

[0208] S7.4: Based on the gap value indicated by the quantity difference identification signal, extract the corresponding number of supplementary questions from the head of the ordered sequence of candidate questions arranged in priority order to generate a set of supplementary questions to be merged.

[0209] For example, suppose the preset number of questions is 20, and the initial question sequence actually contains 15 questions. Then, the gap value indicated by the quantity difference signal is 5. The ordered sequence of candidate questions contains 100 questions, arranged from high to low priority P. Among them, there are 20 questions with P=5 and 30 questions with P=4. The system extracts the first 5 questions from the beginning of the sequence. These 5 questions all have the highest priority P=5. These 5 questions are added to the question set to be merged and supplemented. Finally, this set contains 5 high-priority questions, ensuring that users can still obtain the most valuable practice content when there are insufficient questions, significantly improving practice efficiency.

[0210] S7.5: The set of questions to be merged and supplemented is incorporated into the initial question sequence, the sequence fusion operation is performed, the final intelligent practice content generation result is generated and output to the user terminal, and the closed-loop generation of practice content is completed.

[0211] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0212] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0213] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for generating intelligent practice content by dynamically adjusting the weights of question extraction, characterized in that, Specifically, it includes: S1: Obtain the user's historical answer record data, and construct a question proficiency database including a question proficiency level field based on the question identifier and answer result in the answer record data; S2: Receive the real-time answer result signal and the corresponding answer duration value of the current question, and judge the correctness of the answer based on the real-time answer result signal. If the answer is determined to be incorrect, perform downgrade processing on the corresponding question proficiency level in the question proficiency database and generate a downgraded proficiency level. S3: If the answer is determined to be correct, the answer duration value is compared and matched with a preset time threshold interval set. The proficiency upgrade range is determined according to the matched time threshold interval. Based on the upgrade range, the corresponding question proficiency level in the question proficiency database is upgraded to generate an upgraded proficiency level. S4: Write the downgraded proficiency level or the upgraded proficiency level into the question proficiency database to overwrite the original data, and synchronously record the learning progress log data to form a complete dynamic evaluation closed-loop data source. S5: The dynamic evaluation closed-loop data source is combined with the priority classification tags of the questions in the question bank to construct a differentiated extraction weight allocation model; S6: Use the differentiated extraction weight allocation model to perform weighted random sampling on the candidate question set in the question bank to generate an initial question sequence, and count whether the actual number of questions in the initial question sequence meets the preset question number parameter requirements. S7: If the actual number of questions does not meet the requirement of the number of questions to be drawn, the number of questions is supplemented and incorporated into the initial question sequence to obtain the final intelligent practice content.

2. The intelligent practice content generation method for dynamically adjusting question extraction weights according to claim 1, characterized in that, The question proficiency level field is initialized to an undefined state or a default low level state.

3. The intelligent practice content generation method for dynamically adjusting question extraction weights according to claim 1, characterized in that, The learning progress log data includes the answer result, answer time, and final proficiency level.

4. The intelligent practice content generation method for dynamically adjusting question extraction weights according to claim 1, characterized in that, The differentiated extraction weight allocation model assigns specific probability weight values ​​to questions of different proficiency levels.

5. The intelligent practice content generation method for dynamically adjusting question extraction weights according to claim 1, characterized in that, S1 specifically includes: Obtain user's historical answer record data, parse the original log entries in the data, extract the question identifier field and the answer result status field, and generate a set of structured answer feature vectors; Based on the structured answer feature vector set, duplicate question identifier fields are deduplicated and merged to generate a deduplicated independent question index list. Based on the independent question index list, instantiate a question proficiency database storage structure, define a question proficiency level field in the storage structure, set the initial value of the field to an undefined status flag, and construct an empty question proficiency database framework. Traverse the undefined status identifier entries in the empty question proficiency database framework, batch replace the undefined status identifiers with default low-level status values, and generate a pre-initialized question proficiency database containing the initial proficiency benchmark value. Perform an integrity check on the pre-initialized question proficiency database to verify the consistency of the association between the question identifier field and the question proficiency level field, and output the final usable question proficiency database.

6. The intelligent practice content generation method for dynamically adjusting question extraction weights according to claim 5, characterized in that, The deduplication and merging process for the duplicate question identifier field is performed using a uniqueness mapping method.

7. The intelligent practice content generation method for dynamically adjusting question extraction weights according to claim 1, characterized in that, S5 specifically includes: The learning progress log data in the dynamic evaluation closed-loop data source is aggregated and processed to extract the final proficiency level identifier for each question, generating a proficiency level distribution dataset. Based on the proficiency level distribution dataset and the priority classification label of each question, the proficiency level identifier is associated and matched with the priority classification label to generate a multi-dimensional attribute question set that integrates the user's mastery level and the importance of the question. The questions in the multi-dimensional attribute question set are grouped and clustered. Each group is assigned a specific initial probability weight value using a preset weight configuration rule to generate an original weight allocation scheme divided by proficiency level. Based on the initial probability weight values ​​in the original weight allocation scheme, the weight ratios of each group are adjusted to ensure that the sum is one and meets the constraint of the number of questions to be drawn, thereby generating the differentiated extraction weight allocation model.

8. The intelligent practice content generation method for dynamically adjusting question extraction weights according to claim 7, characterized in that, The proficiency level distribution dataset includes a mapping relationship between question identifiers and corresponding proficiency levels.

9. The intelligent practice content generation method for dynamically adjusting question extraction weights according to claim 7, characterized in that, The specific steps for grouping and clustering questions in the multi-dimensional attribute question set are as follows: The questions in the multi-dimensional attribute question set are grouped and clustered according to the status of never having been asked, level 0 proficiency, level 1 proficiency, level 2 proficiency, and level 3 proficiency.

10. An intelligent practice content generation system that dynamically adjusts the weights of question extraction, characterized in that, include: The database construction module is used to obtain users' historical answer records data, and construct a question proficiency database including a question proficiency level field based on the question identifiers and answer results in the answer records data; The downgrade processing module is used to receive the real-time answer result signal of the current question and the corresponding answer duration value, and judge the correctness of the answer based on the real-time answer result signal. If the answer is determined to be incorrect, the module performs downgrade processing on the corresponding question proficiency level in the question proficiency database and generates a downgraded proficiency level. The upgrade processing module is used to compare and match the answer duration value with a preset time threshold interval set if the answer is determined to be correct, determine the proficiency upgrade range based on the matched time threshold interval, and perform upgrade processing on the corresponding question proficiency level in the question proficiency database based on the upgrade range to generate an upgraded proficiency level. The dynamic evaluation module is used to write the downgraded proficiency level or the upgraded proficiency level into the question proficiency database to overwrite the original data, and synchronously record the learning progress log data to form a complete dynamic evaluation closed-loop data source. The model building module is used to construct a differentiated extraction weight allocation model by combining the dynamic evaluation closed-loop data source with the priority classification tags of questions in the question bank. The sampling and quantity statistics module is used to perform weighted random sampling on the candidate question set in the question bank using the differentiated extraction weight allocation model, generate an initial question sequence, and count whether the actual number of questions in the initial question sequence meets the preset question quantity parameter requirements. The question completion module is used to supplement the number of questions and incorporate them into the initial question sequence if the actual number of questions does not meet the requirement of the number of questions to be drawn, so as to obtain the final intelligent practice content.