An AI online education intelligentized question and answer information processing method

Abstract

The application discloses an AI online education intelligent question and answer information processing method, which aligns and maps a cold start interaction set time to a knowledge point index, combines a teaching structured constraint set to complete prerequisite dependency binding to form a cold start sequence; based on an evidence group identifier, retrieves an evidence list and performs effective evidence screening, calculates a sample support degree index and compares it with a support degree threshold to output a to-be-verified label set; obtains a question discrimination degree level and a source reliability level and performs supervised learning regression inference to output an upper bound value of a wrong label risk, compares the upper bound value with a wrong label threshold to determine a stable label and adds the stable label to a recommendation set; generates an alternative label for a forbidden solidification result, retrieves a differentiated interaction task set, collects a verification evidence sequence, calculates a consistency index and compares the consistency index with a consistency threshold to output a pseudo-stable label mark, outputs a label decision result and updates the recommendation set, and reduces a recommendation path locking risk.

Description

Technical Field

[0001] This invention relates to the field of information processing technology, specifically to an AI-powered intelligent question-and-answer information processing method for online education. Background Technology

[0002] In the cold start phase of new users, existing AI-powered online education's intelligent question-answering and personalized recommendations often generate and solidify learning tags directly from a small number of questions and initial learning behaviors. When the cold start interaction sample size is insufficient, there is a lack of quantitative constraints on the scale of evidence, the scope of hierarchical coverage, and the consistency of source caliber. This leads to occasional behaviors being written into long-term profiles, causing learning tags to deviate from the structured constraints of the course syllabus. At the same time, the determination of tag stability mostly relies on fixed-time smoothing or single consistency conditions, lacking a mechanism for identifying and labeling pseudo-stable tags when sample support is insufficient. Tags trigger recommendations based on a single tag before a retrievable mislabeling risk assessment result is formed, lacking constraints on the risk of locking in the recommendation path, which in turn leads to the solidification of learning paths, deviation of recommended content from the syllabus, and risks to educational equity.

[0003] Therefore, proposing an AI-powered intelligent question-and-answer information processing method for online education to address the difficulties of existing technologies is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] The purpose of this invention is to provide an AI-powered intelligent question-and-answer information processing method for online education, in order to address the shortcomings of the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an AI-powered intelligent question-and-answer information processing method for online education, comprising:

[0006] Obtain the cold start interaction set and teaching structured constraint set of new users, process them, and output the cold start sequence;

[0007] Based on the cold start sequence, evidence aggregation is performed to generate a candidate learning label set; each candidate learning label item includes at least an evidence group identifier field. Based on the evidence group identifier field, the evidence list is retrieved from the cold start sequence and effective evidence screening is performed. Based on the screening results, the sample support index is calculated and compared with the support threshold, and a set of labels to be verified is output.

[0008] For the set of tags to be verified, the question discrimination level and source reliability level are obtained and supervised learning regression inference is performed to output the upper bound of mislabeling risk. The upper bound of mislabeling risk is compared with the mislabeling threshold to obtain the results that are allowed to be solidified and the results that are prohibited from being solidified. The tags to be verified corresponding to the results that are allowed to be solidified are determined as stable tags and added to the recommendation set.

[0009] For the unverified labels corresponding to the prohibited solidification results, alternative labels are generated and the interactive task set is retrieved and distinguished. The verification evidence sequence is collected, the consistency index is calculated and compared with the consistency threshold to output pseudo-stable label markings. The label decision results are output by comprehensively comparing the upper limit of mislabeling risk, consistency index and pseudo-stable label markings, and the recommendation set is updated accordingly.

[0010] Furthermore, the cold start interaction set includes question and answer records, practice answer records, course browsing records and dwell time records, and the teaching structure constraint set includes knowledge point index, parent knowledge point index, prerequisite knowledge point index set, required path sequence number and difficulty gradient identifier, as well as a corresponding list of question item identifiers and knowledge point indexes, a corresponding list of content item identifiers and knowledge point indexes, and a corresponding list of question and answer item identifiers and knowledge point indexes.

[0011] Time alignment is performed based on the cold start interaction set to form a time alignment sequence. Then, based on the corresponding list, the question item identifier, content item identifier, and question and answer item identifier are mapped to knowledge point indexes and the knowledge point index field is filled to obtain the time alignment mapping sequence.

[0012] Based on the knowledge point index field in the time-aligned mapping sequence, retrieve the teaching structured constraint set, extract the prerequisite knowledge point index set, parent knowledge point index, required path number and difficulty gradient identifier, fill in the prerequisite dependency field, level pointing field, path position field and difficulty gradient field, and output the cold start sequence.

[0013] Furthermore, methods for generating candidate learning label sets based on evidence aggregation performed on cold-start sequences include:

[0014] Receive the cold start sequence and arrange the cold start sequence from earliest to latest according to the recorded timestamp within the same new user ID range;

[0015] Based on the permutation results, the cold start sequence is merged according to the knowledge point index field to obtain evidence groups with the knowledge point index field as the primary key, and an evidence group identifier is generated for each evidence group.

[0016] Based on the evidence group, the record type field, source type field, and correct / incorrect field are extracted. The number of correct records, the number of incorrect records, the number of questions, the number of answers, the number of entries, the number of exits, the cumulative stay time, and the maximum stay time are counted respectively, and a source type sequence is formed.

[0017] Candidate learning tag entries are generated based on statistical results. Candidate learning tag entries include tag identifier field, tag type field, knowledge point index field, prerequisite dependency field, and evidence group identifier field. The tag type field is determined by comparing the number of erroneous entries with the preset error number threshold, the number of correct entries with the preset correct number threshold, the number of entries entered with the preset browsing entry threshold, the cumulative dwell time in seconds with the preset cumulative dwell time threshold, and the maximum dwell time in seconds with the preset maximum dwell time threshold. The candidate learning tag entries are then summarized to obtain a candidate learning tag set.

[0018] Furthermore, methods for screening valid evidence include:

[0019] Receive the set of candidate learning labels, and read the evidence group identifier field of each candidate learning label item one by one. Based on the evidence group identifier field, retrieve the corresponding evidence from the cold start sequence to form an evidence list.

[0020] Based on the evidence list, the effective evidence is filtered item by item. The effective evidence filtering includes source type filtering, field completeness filtering, record type matching filtering, and short-term duplicate filtering. The output is a set of effective evidence and a set of ineffective evidence.

[0021] The statistical fields of candidate learning label entries are updated based on the valid evidence set. The statistical fields include the number segments of valid evidence entries, correct entries, incorrect entries, question entries, answer entries, entry entries, exit entries, cumulative stay seconds, and maximum stay seconds. The source type field in the valid evidence set is deduplicated to obtain the valid source type set and the source type sequence field is updated.

[0022] Furthermore, methods for outputting the set of tags to be verified include:

[0023] Receive the set of candidate learning labels after the effective evidence screening is completed, and generate a sample support index for each candidate learning label item in the set;

[0024] Based on the candidate learning label entries, the effective evidence number segment is extracted as the effective evidence number. Based on the knowledge point index field, the hierarchical number is retrieved in the knowledge point hierarchy relationship of the teaching structured constraint set to form a hierarchical number set. The hierarchical number set is deduplicated to obtain the hierarchical coverage number and written into the hierarchical coverage number segment. The effective evidence number and the hierarchical coverage number constitute the sample support index.

[0025] Support threshold comparison is performed based on the sample support index. The support threshold consists of a preset support number threshold and a preset support level threshold. When the number of valid evidence items meets or exceeds the preset support number threshold, and the level coverage meets or exceeds the preset support level threshold, the candidate learning label item is not included in the label set to be verified. When the number of valid evidence items does not meet or exceeds the preset support number threshold, or the level coverage does not meet or exceeds the preset support level threshold, the candidate learning label item is included in the label set to be verified, and the following fields are retained: label identifier field, label type field, knowledge point index field, prerequisite dependency field, evidence group identifier field, valid evidence item number field, level coverage number field, correct item number field, incorrect item number field, entry item number field, cumulative stay seconds number field, question item number field, source type sequence field, record timestamp start field, and record timestamp end field.

[0026] Furthermore, methods for obtaining the question discrimination level and source reliability level for each item in the set of tags to be verified include:

[0027] Receive the set of tags to be verified, read the evidence group identifier field of each tag to be verified, and retrieve the corresponding list of valid evidence based on the evidence group identifier field;

[0028] Based on the list of valid evidence, evidence entries whose record type field is practice answer records and include the question entry identifier field are selected to form a question entry identifier list. Based on the question entry identifier list, the question bank statistical entries are read one by one, and the question discrimination level is found in the discrimination level table according to the difference range between the high score segment accuracy field and the low score segment accuracy field in the question bank statistical entries. The question discrimination level is written into the question discrimination level sequence field.

[0029] Based on the list of valid evidence, the source type sequence field is deduplicated to obtain the source type list. The reliability level field in the source reliability level table is read one by one, and the source reliability level summary record is formed by combining the number of occurrences of the source type in the list of valid evidence. The source reliability level summary record is written into the source reliability level sequence field.

[0030] Furthermore, methods for adding recommendations to a set include:

[0031] Receive a set of tags to be verified, and construct a regression inference input record for each tag in a fixed field order. The regression inference input record includes a tag type field, a knowledge point index field, a valid evidence bar number field, a correct bar number field, an incorrect bar number field, a level coverage bar number field, a question discrimination level sequence field, a source reliability level sequence field, a record time span field, and a source type diversity field. The regression inference execution process outputs the upper bound value of the mislabeling risk.

[0032] Read the upper bound field of mislabeling risk, the tag type field, the knowledge point index field, and the level coverage number field of the tag to be verified in the tag set one by one, and select the applicable mislabeling threshold field from the mislabeling threshold list based on the tag type field and the level coverage number field.

[0033] Based on the applicable mislabeling threshold field, the upper limit of mislabeling risk is compared with the mislabeling threshold. When the upper limit of mislabeling risk is not greater than the mislabeling threshold, the result of allowing solidification is output. When the upper limit of mislabeling risk is greater than the mislabeling threshold, the result of prohibiting solidification is output.

[0034] Based on the results of allowing solidification, the tag status field of the tag to be verified is updated to a stable tag, and the tag identifier field, tag type field, knowledge point index field, prerequisite dependency field, source type sequence field, record timestamp start field, and record timestamp end field of the stable tag are written into the recommendation set generation list.

[0035] Furthermore, the method for generating alternative labels for the unverified labels in the list of prohibited solidified results and retrieving and distinguishing the set of interactive tasks includes:

[0036] Receive the list of prohibited solidification results and its corresponding set of tags to be verified, read the tag identifier field, tag type field, knowledge point index field and evidence group identifier field of the tags to be verified one by one, and retrieve the hierarchical number field of the teaching structured constraint set based on the knowledge point index field.

[0037] Based on the tag type field and the level number field, the original tag type field and the applicable knowledge point level field are matched in the list of alternative tag rules. The alternative tag type field is read and an alternative tag is generated. The alternative tag includes an alternative tag identifier field, an alternative tag type field, a knowledge point index field and an evidence group identifier field. The alternative tag identifier field and the tag identifier field are combined to form a tag pair identifier field.

[0038] Based on the tag pair identifier field, knowledge point index field, tag type field, and alternative tag type field, the task template library is used to retrieve and differentiate the interactive task set. The template identifier field, task content item identifier field, and task source type field are read to form the task list field. Based on the task list field and the tag pair identifier field, a task delivery record is formed. The task delivery record includes the new user identifier field and the delivery timestamp field.

[0039] Furthermore, methods for obtaining pseudo-stable label markers include:

[0040] Based on the task delivery record, interactive tasks corresponding to the task content item identifier fields are presented, and data is collected based on the task output field list to form a verification evidence sequence. The verification evidence sequence includes new user identifier field, tag pair identifier field, template identifier field, task content item identifier field, record timestamp field, record type field, knowledge point index field, source type field, answer result field, answer duration field, skip mark field, and correct / incorrect field.

[0041] Based on the verification evidence sequence, the task result judgment rule field is retrieved by the template identifier field, and the task support category field is generated according to the correct / incorrect field, the response time field, and the skip flag field. The task support category field has the values ​​of supporting the label to be verified, supporting the alternative label, and not distinguishable. The number of items supporting the label to be verified, the number of items supporting the alternative label, and the number of items not distinguishable are counted to form a consistent summary record.

[0042] Based on the consistency summary records, the consistency index field is determined in the consistency level table, and the applicable consistency threshold level field is determined based on the preset consistency threshold list. The consistency index field and the applicable consistency threshold level field are compared in order of rank, and the pseudo-stable label field is output. The pseudo-stable label field can be either pseudo-stable or non-pseudo-stable.

[0043] Furthermore, methods for updating the recommendation set include:

[0044] Read the upper bound field of mislabeling risk, consistency index field, applicable consistency threshold level field, pseudo-stable label mark field, label identifier field, knowledge point index field and prerequisite dependency field of the label to be verified corresponding to the prohibited solidification result one by one, and retrieve the mislabeling threshold field from the mislabeling threshold list based on the label type field and knowledge point index field;

[0045] The label decision result field is determined by comparing the upper limit value field of mislabeling risk with the mislabeling threshold field, comparing the level order based on the consistency index field and the applicable consistency threshold level field, and combining the pseudo-stable label marking field. The label decision result field takes the value of either a label to be verified or a label that is prohibited from being recommended.

[0046] The recommendation set is updated based on the tag decision result field. When the tag decision result field is a prohibited recommendation tag, the entries in the recommendation entry sequence whose associated tag identifier field is equal to the tag identifier field and whose knowledge point index field is equal to the knowledge point index field are removed from the recommendation entry sequence. Learning content entries are then retrieved from the course content index list based on the prerequisite dependency field and written into the recommendation entry sequence. When the tag decision result field is a tag to be verified, the learning content entries in the recommendation entry sequence whose associated tag identifier field is equal to the tag identifier field are retained. Interactive task entries are then added to the recommendation set based on the task delivery record.

[0047] The technical effects and advantages of the AI-powered intelligent question-and-answer information processing method for online education provided by this invention are as follows:

[0048] By acquiring the cold-start interaction set of new users and the structured constraint set of the teaching curriculum, time alignment and knowledge point index mapping are performed on the cold-start interaction set. Prerequisite dependency binding is then completed using the structured constraint set, forming a cold-start sequence with knowledge point index fields and prerequisite dependency fields. Since subsequent evidence aggregation, tag generation, and recommendation adjustments all use the knowledge point index field as the unified index key and continuously reference the prerequisite dependency field to limit knowledge point relationships, a small number of question-and-answer sessions or initial learning behaviors can fall within the structured scope of the course syllabus. This reduces the path of occasional behaviors being amplified into long-term stable profiles, thus methodologically lowering the probability of learning tags deviating from the course syllabus.

[0049] In the candidate learning label stage, a valid evidence screening standard is introduced. Source type, field completeness, record type matching, and short-term repetition are treated consistently. The valid evidence segment, correct evidence segment, incorrect evidence segment, and source type sequence field are recalculated using the valid evidence set to ensure the comparability and readability of statistical fields. Based on this, a sample support index is generated, consisting of the number of valid evidence entries and the number of hierarchical coverages. This index is compared with a support threshold to output the set of labels to be verified. This diversion addresses sample uncertainty by comparing countable fields with a threshold, including entries with insufficient evidence or insufficient hierarchical coverage in the set of labels to be verified. This avoids directly solidifying learning labels based on a small number of questions or initial learning behaviors, thus specifically addressing the source of misjudgments caused by insufficient samples in the cold start phase.

[0050] This method obtains the question discrimination level and source reliability level for the set of tags to be verified, and constructs regression inference input records with a fixed field order, outputting an upper bound value for mislabeling risk. The upper bound value is then compared with the mislabeling threshold to obtain the results for permissible and prohibited permissible ...

[0051] For the unverified labels corresponding to the prohibited solidification results, alternative labels are generated, and the interactive task set is retrieved and distinguished. Verification evidence sequences are collected, and a consistency index is calculated. Each piece of verification evidence is categorized into three types based on the task support category field: supporting the unverified label, supporting the alternative label, or indistinguishable. The number of each type of evidence is summarized and mapped to a consistency level table to obtain a consistency index level. This level is then compared with the applicable consistency threshold level field determined by the consistency threshold list to output a pseudo-stable label. This process explicitly identifies pseudo-stable labels caused by insufficient samples by comparing discrete levels with thresholds, compensating for the identification gaps caused by existing schemes that rely on time window smoothing or single consistency conditions.

[0052] On the recommendation side, the upper bound of mislabeling risk, mislabeling threshold, consistency index, consistency threshold, and pseudo-stable label marking are used as constraints to output label decision results. The label decision results are limited to labels to be verified or labels to be prohibited from recommendation, and the recommendation set is updated accordingly. Prohibiting a recommended label triggers the removal of the entries corresponding to the label identifier field and knowledge point index field in the recommendation entry sequence, and inserts prerequisite content entries based on the prerequisite dependency field; labels to be verified trigger the retention of related learning content entries and the addition of interactive task entries to differentiate them. This strategy avoids locking the recommendation path based on a single label in the stage where the label is not fully verified, increases the contribution of subsequent interaction data to label confirmation, and reduces the risks of learning path solidification, teaching content deviation from the syllabus, and educational equity. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0054] Figure 1This is a schematic diagram of an AI-powered intelligent question-and-answer information processing method for online education according to the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] like Figure 1 As shown in this embodiment, an AI-powered intelligent question-and-answer information processing method for online education includes:

[0057] Step 1: Obtain the cold start interaction set and the teaching structured constraint set for new users; perform time alignment and knowledge point index mapping on the cold start interaction set, complete the prerequisite dependency binding using the teaching structured constraint set, and output the cold start sequence.

[0058] Obtain the cold start interaction set and teaching structured constraint set of new users. Specifically, the starting time when a new user first enters the course page is recorded as the cold start start time. During the preset cold start period after the cold start start time, extract four types of original records from the data storage according to the new user identifier: question and answer records, practice answer records, course browsing sequence, and dwell time sequence. Then, organize the four types of original records into an interaction record set with the same field caliber.

[0059] Question and answer records, practice answer records, course browsing records, and dwell time records were selected as cold start interaction data because they provide complementary and retrievable evidence: question and answer records expose user-initiated questions, practice answer records provide quantitative evidence with correct and incorrect results, course browsing records reflect content selection paths, and dwell time records quantify attention intensity. When combined, these four data can cover the main observable learning behaviors in the cold start phase and form countable, aggregateable, and threshold-comparable statistical fields, providing stable input for subsequent sample support, mislabeling risk, and consistency judgments.

[0060] Each Q&A record must include at least a new user identifier, Q&A entry identifier, question text, answer text, record timestamp, and source identifier; each exercise answer record must include at least a new user identifier, question entry identifier, answer options, answer result field, submission timestamp, and source identifier; each course browsing record must include at least a new user identifier, content entry identifier, entry timestamp, exit timestamp, and source identifier; and each dwell time record must include at least a new user identifier, content entry identifier, dwell start timestamp, dwell end timestamp, dwell time in seconds field, and source identifier.

[0061] After the above fields are complete, a record type field is added to each record. The record type field can take one of four values: question and answer record, exercise answer record, course browsing record, or dwell time record, to ensure that different records can be distinguished on the same timeline later. In the same batch processing, the teaching structured constraint set is extracted. The teaching structured constraint set is stored one by one according to knowledge point items. Each knowledge point item includes at least a knowledge point index, knowledge point name, parent knowledge point index, prerequisite knowledge point index set, required path number, and difficulty gradient identifier. The knowledge point index is an integer number, the parent knowledge point index is an integer number, the prerequisite knowledge point index set is a combination of multiple integer numbers, the required path number is an integer number, and the difficulty gradient identifier is an integer number. At the same time, the corresponding lists of question item identifiers and knowledge point indexes, the corresponding lists of content item identifiers and knowledge point indexes, and the corresponding lists of question and answer item identifiers and knowledge point indexes are extracted and fixed as part of the teaching structured constraint set, so that the subsequent knowledge point index mapping is completed under the same numbering system.

[0062] The reason why the knowledge point index, parent knowledge point index, prerequisite knowledge point index set, required path number, and difficulty gradient identifier are used as the core fields of the teaching structure constraint set is that this scheme uses the knowledge point index as the unified anchor point for evidence merging and tag placement, the parent knowledge point index to determine the hierarchical direction and support the number of hierarchical coverage, the prerequisite knowledge point index set to bind prerequisite dependencies and constrain the order of tags and recommendations, the required path number to align recommendations with the course structure to avoid skipping, and the difficulty gradient identifier to limit the difficulty range that distinguishes interactive tasks and content delivery to ensure comparability, thereby incorporating cold start evidence into a constrainable, hierarchical, and traceable syllabus framework.

[0063] Based on the set of interaction records, the cold start interaction set is time-aligned and mapped to knowledge point indexes. Specifically, the record timestamps are used as a unified time base, as follows:

[0064] The submission timestamp of practice answer records is written to the record timestamp field, the question timestamp of Q&A records is written to the record timestamp field, the entry timestamp of course browsing records is written to the record timestamp field, and the stay start timestamp of stay duration records is written to the record timestamp field, so that all four types of records have a sortable record timestamp field; then the set of interaction records is sorted from earliest to latest according to the record timestamp.

[0065] When records have the same timestamp, they are arranged in a fixed order according to the record type field: practice answer records first, question and answer records second, course browsing records third, and dwell time records last. The sorted result is used as a time-aligned sequence. After obtaining the time-aligned sequence, knowledge point index mapping is performed record by record: for practice answer records, the target knowledge point index is retrieved from the corresponding list of question entry identifiers and knowledge point indexes based on the question entry identifier, and written to the knowledge point index field of that record; for course browsing records, the target knowledge point index is retrieved from the corresponding list of content entry identifiers and knowledge point indexes based on the content entry identifier, and written to the knowledge point index field of that record; for dwell time records, ... Based on the content entry identifier, the target knowledge point index is retrieved from the corresponding list of content entry identifiers and knowledge point indexes, and written into the knowledge point index field of the record. Simultaneously, a field representing the dwell time in seconds is retained and stored alongside the knowledge point index field to characterize the dwell intensity corresponding to that knowledge point index. For question-and-answer records, the target knowledge point index is retrieved first from the corresponding list of question-and-answer entry identifiers and knowledge point indexes based on the question-and-answer entry identifier and written into the knowledge point index field. If no corresponding entry is found for the question-and-answer entry identifier, the target knowledge point index is retrieved from the corresponding list of content entry identifiers and knowledge point indexes based on the content entry identifier in the question-and-answer record and written into the knowledge point index field, thus ensuring that the question-and-answer record also falls within the knowledge point index numbering system.

[0066] The reason why the correspondence lists of question item identifiers and knowledge point indexes, the correspondence lists of content item identifiers and knowledge point indexes, and the correspondence lists of question and answer item identifiers and knowledge point indexes are necessary is that the carriers of cold start interaction records are questions, content, and question and answer items, respectively. The evidence aggregation, question discrimination level acquisition, source reliability level summary, and recommendation set generation in the solution all use the knowledge point index as a unified index key. Through these correspondence lists, interaction records from different carriers can be mapped to the same knowledge point index field and cross-source merging and retrieval can be achieved.

[0067] After the above writing is completed, a time-aligned mapping sequence is formed, including new user identifier, record timestamp, record type, content entry identifier or question entry identifier or question and answer entry identifier, source identifier, knowledge point index, and answer result field of practice answer record or dwell time record dwell time second number segment.

[0068] The teaching structure constraint set is used to complete the prerequisite dependency binding on the time-aligned mapping sequence and output the cold start sequence, as follows:

[0069] The process involves reading the knowledge point index fields in the time-aligned mapping sequence one by one, locating knowledge point entries with the same knowledge point index in the teaching structured constraint set, extracting the prerequisite knowledge point index set for that knowledge point entry, and writing the prerequisite knowledge point index set into the prerequisite dependency field of that record, so that each record forms a definite correspondence between the current knowledge point index and the prerequisite dependency. The process also involves extracting the parent knowledge point index from the same knowledge point entry and writing it into the hierarchy pointer field, extracting the required path number from the same knowledge point entry and writing it into the path position field, and extracting the difficulty gradient identifier from the same knowledge point entry and writing it into the difficulty gradient field, so that each record, in addition to the prerequisite dependency field, also carries the hierarchy and path semantics consistent with the teaching structured constraint set. Finally, the source identifier field is converted into a source type field according to the source identifier mapping list, with the source type field taking values ​​such as in-class exercises, chapter quizzes, and homework. The system identifies one of six categories: free questioning, directory browsing, and video playback, enabling subsequent processing to differentiate evidence sources. The answer result fields of practice response records are then converted into "correct" or "incorrect" fields based on their value ranges. These fields accept one of three values: correct, incorrect, or no result. "No result" is used for non-practice response records, ensuring consistent fields across different record types within the same evidence structure. Finally, the cold start sequence is output sequentially according to the time-aligned mapping sequence. Each cold start evidence entry includes at least a new user identifier, record timestamp, record type, knowledge point index, prerequisite dependencies, source type, and correct / incorrect fields, as well as hierarchical pointing fields, path location fields, and difficulty gradient fields. This ensures that the cold start sequence simultaneously reflects the time alignment results of the cold start interaction set, the knowledge point index mapping results, and the prerequisite dependency binding results of the teaching structure constraint set at the single record level.

[0070] Step 2: Generate a candidate learning label set by performing evidence aggregation based on the cold start sequence; each candidate learning label entry in the candidate learning label set includes at least an evidence group identifier field. Retrieve the evidence list from the cold start sequence based on the evidence group identifier field and perform effective evidence screening. Calculate the sample support index based on the effective evidence screening results, compare it with the support threshold, and output the label set to be verified.

[0071] A candidate set of learned labels is generated based on evidence aggregation performed according to the cold start sequence, as follows:

[0072] To receive the output cold start sequence, the cold start sequence is stored as a single piece of evidence. A single piece of evidence includes a new user identifier field, a record timestamp field, a record type field, a knowledge point index field, a prerequisite dependency field, a source type field, a correct / incorrect field, a hierarchy pointer field, a path location field, and a difficulty gradient field. The record type field has four categories: question and answer records, practice answer records, course browsing records, and dwell time records. The source type field has six categories: in-class exercises, chapter quizzes, homework, open questions, directory browsing, and video playback. The correct / incorrect field has three categories: correct, incorrect, and no result. Within the same new user identifier range, all evidence is first arranged from earliest to latest by record timestamp, then merged by the knowledge point index field. Evidence with the same knowledge point index field is grouped into the same evidence group. Each evidence group corresponds to one knowledge point index field, and this knowledge point index field is retained as evidence. Based on the primary key of the evidence group, an evidence group identifier is generated for each evidence group. The evidence group identifier is formed by concatenating the new user identifier field, the knowledge point index field, and the earliest record timestamp field within the evidence group. The evidence group identifier is then filled into the evidence group identifier field of each piece of evidence within the evidence group to ensure that evidence can be retrieved through the evidence group identifier later. Within each evidence group, the records are arranged again from earliest to latest according to the record timestamp. The correct and incorrect fields of the practice answer records are extracted one by one, and the number of correct and incorrect records are counted respectively. The question and answer records are extracted one by one, and the number of questions and answers is counted. The course browsing records are extracted one by one, and the number of entries and exits is counted. The dwell time records are extracted one by one, and the dwell time in seconds is accumulated and the maximum dwell time in seconds is recorded. At the same time, the source type field is collected one by one to form a source type sequence, and the level pointing field, path location field, and difficulty gradient field are collected one by one to form a set of values ​​for level pointing value, path location value, and difficulty gradient value.

[0073] After completing the above summary, a candidate learning tag entry is generated for this evidence group. The candidate learning tag entry includes the following fields: tag identifier, tag type, knowledge point index, prerequisite dependency, evidence group identifier, start and end timestamp fields, correct answer field, incorrect answer field, question field, answer field, entry field, exit field, cumulative stay in seconds, maximum stay in seconds, source type sequence field, level pointer value field, path location value field, difficulty gradient value field, and valid evidence point field. Specifically, the start and end timestamp fields are the earliest recorded timestamps within the evidence group, and the prerequisite dependency field is the corresponding prerequisite dependency field for the knowledge point index in the cold start sequence. The tag type field is assigned values ​​according to a fixed tag type mapping rule, which is as follows:

[0074] When the number of erroneous entries is no less than the preset error threshold (e.g., 3), and the number of erroneous entries is greater than the number of correct entries, the tag type field is assigned the "Weakness" tag. When the number of correct entries is no less than the preset correct entry threshold, and the number of correct entries is greater than the number of erroneous entries, the tag type field is assigned the "Knowledge Point Mastery" tag. When the number of entries entered is no less than the preset browsing entry threshold, and the cumulative stop... If the number of seconds spent on the platform meets the preset cumulative dwell time threshold and the maximum number of seconds spent on the platform meets the preset maximum dwell time threshold, the tag type field is assigned a learning preference tag. If none of the above three threshold comparison conditions are met, the tag type field is assigned a pending tag, and a tag identifier field is assigned a value. The tag identifier field is formed by concatenating the new user identifier field, the knowledge point index field, and the tag type field, so that the same new user can form different candidate learning tag entries under different knowledge point indices.

[0075] Candidate learning labels are generated for each evidence group in the manner described above, and the candidate learning labels are aggregated to obtain a set of candidate learning labels. The set of candidate learning labels maintains a one-to-one correspondence with the evidence group identifier, and the traceability relationship between the set of candidate learning labels and the cold start sequence is maintained.

[0076] The following steps are taken to screen for valid evidence:

[0077] The system receives a set of candidate learning tags, reads the evidence group identifier field of each candidate learning tag entry, and retrieves all evidence corresponding to that evidence group identifier from the cold start sequence to form an evidence list. It then performs valid evidence filtering on each piece of evidence in the evidence list, using a fixed filtering criterion to produce two types of output: a valid evidence set and an invalid evidence set. The first filtering is source type filtering, where the source type field is read one by one. If the source type field value is one of in-class exercises, chapter quizzes, homework, open questions, table of contents browsing, or video playback, the evidence is retained; if the source type field value is empty or any other value, the evidence is discarded and added to the invalid evidence set. The second filtering is field completeness filtering, where the new user identifier field, record timestamp field, knowledge point index field, and prerequisite dependency field are read one by one. If all four fields exist, the evidence is retained; if any field is empty, the evidence is discarded and added to the invalid evidence set. The third filter is a record type matching filter. When the record type field is a practice answer record, the correct / incorrect field is also read. If the correct / incorrect field value is correct or incorrect, it is retained. If the correct / incorrect field value is empty, it is removed and added to the invalid evidence set. When the record type field is a stay duration record, the stay duration in seconds is also read. If the stay duration in seconds is an integer, it is retained. If the stay duration in seconds is empty, it is removed and added to the invalid evidence set. When the record type field is a question and answer record or a course browsing record, it is retained, but the correct / incorrect field value is kept as empty. The fourth filter is a short-term repetition filter (e.g., 30 seconds or 5 minutes). Within the same evidence list, if there are multiple pieces of evidence with the same record type field, the same knowledge point index field, the same source type field, and the record timestamps appearing repeatedly within 120 seconds, only the record with the earlier timestamp is retained as valid evidence, and the remaining evidence is added to the invalid evidence set.

[0078] After completing the above four screening steps, a valid evidence set is obtained. Then, returning to the candidate learning tag entries, the statistical fields written to the candidate learning tag entries are recalculated and overwritten using the valid evidence set as the statistical caliber: the valid evidence count field is the number of valid evidence entries; the correct count field is the number of practice answer records in the valid evidence set where the correct / incorrect field is correct; the incorrect count field is the number of practice answer records in the valid evidence set where the correct / incorrect field is incorrect; the question count field is the number of question-and-answer records in the valid evidence set; and the response count field is the number of question-and-answer records in the valid evidence set where a response text exists. The "Entry" field represents the number of course browsing records in the valid evidence set; the "Exit" field represents the number of course browsing records in the valid evidence set with an exit timestamp; the "Cumulative Stay Seconds" field represents the sum of the stay seconds fields in the stay duration records in the valid evidence set; and the "Maximum Stay Seconds" field represents the maximum stay seconds field in the stay duration records in the valid evidence set. Simultaneously, the "Source Type" field in the valid evidence set is deduplicated to form a valid source type set, and the "Source Type Sequence" field is updated. This ensures that candidate learning tag entries only reference field values ​​under the valid evidence caliber during subsequent support threshold comparisons.

[0079] Calculate the sample support index, compare it with the support threshold, and output the set of labels to be validated, as follows:

[0080] After receiving the candidate learning label set after the effective evidence screening is completed, a sample support index is generated for each candidate learning label item in the candidate learning label set, and the set of labels to be verified is output based on the comparison result of the sample support index and the support threshold.

[0081] The sample support index consists of two components: the number of valid evidence entries and the level coverage. The number of valid evidence entries is directly obtained from the number of valid evidence entries in the candidate learning label entries. The level coverage is obtained by retrieving the corresponding level number from the knowledge point hierarchy relationship in the teaching structured constraint set based on the knowledge point index field in the candidate learning label entries, forming a level number set of all level numbers associated with the candidate learning label entries, deduplicating the level number set to obtain the level coverage, and filling the level coverage number into the level coverage number field of the candidate learning label entries. Subsequently, a support threshold comparison is performed on the sample support index of each candidate learning label entry. The support threshold consists of a preset support number threshold and a preset support level threshold. The comparison rule is limited to: when the number of valid evidence entries meets the preset support number threshold and the level coverage meets the preset support level threshold, the candidate learning label entry is determined to meet the support threshold and is not included in the label set to be verified; when there are If the number of valid evidence entries does not meet the preset support threshold, or the number of hierarchical coverage entries does not meet the preset support level threshold, the candidate learning label entry is deemed not to meet the support threshold and is included in the label set to be verified. For candidate learning label entries included in the label set to be verified, the following fields are retained: label identifier field, label type field, knowledge point index field, prerequisite dependency field, evidence group identifier field, valid evidence entry number field, hierarchical coverage number field, correct entry number field, incorrect entry number field, entry entry number field, cumulative stay seconds number field, question entry number field, source type sequence field, record timestamp start field, and record timestamp end field. These retained fields are used as record fields in the label set to be verified. For candidate learning label entries not included in the label set to be verified, they are retained in the candidate learning label set and the statistical fields under the valid evidence caliber are maintained. The label set to be verified is output by generating sample support indicators, comparing support thresholds, and performing set splitting for each candidate learning label entry.

[0082] For example, taking a candidate learning tag entry identified as "User A" as the object, after filtering for valid evidence, the valid evidence entry for this candidate learning tag entry has the number field "7", the knowledge point index field "K17", and the prerequisite dependency fields "K05" and "K09". Based on this, the sample support index is calculated as follows:

[0083] First, the number of valid evidence entries is extracted. The number segment "7" of valid evidence entries is taken as the first component of the sample support index, denoted as Valid Evidence Entries = 7. Second, the level coverage is calculated. Based on the knowledge point index K17, the level number is retrieved from the knowledge point hierarchy relationship in the teaching structured constraint set to obtain "3". Based on the prerequisite dependency K05, the level number is retrieved to obtain "2". Based on the prerequisite dependency K09, the level number is retrieved to obtain "2", forming a level number set {3, 2, 2}. The level number set is deduplicated to obtain {2, 3}, and the number of elements after deduplication is "2". This number of elements is taken as the second component of the sample support index, denoted as Level Coverage = 2. Thus, the sample support index is obtained as "(Valid Evidence Entries = 7, Level Coverage = 2)". The sample support index is then compared with the support threshold: when the preset support number threshold is "6" and the preset support level threshold is "2", the number of valid evidence items in the sample support index = 7 satisfies the requirement of not less than 6, and the level coverage = 2 satisfies the requirement of not less than 2. Therefore, the sample support index meets the support threshold, and the candidate learned label item is not included in the label set to be verified; when the preset support number threshold is "8" and the preset support level threshold is "2", the number of valid evidence items in the sample support index = 7 does not satisfy the requirement of not less than 8. Therefore, the sample support index does not meet the support threshold, and the candidate learned label item is included in the label set to be verified.

[0084] It should be noted that the preset support count threshold and preset support level threshold are provided by the support threshold list, which is stored according to the tag type field and the knowledge point level field. The support count threshold is determined by enumerating candidate values ​​of the valid evidence number segments in the historical user samples, comparing the proportion of mislabeled fields in the sample subset corresponding to each candidate value with the preset mislabeling ratio upper limit, and selecting the smallest candidate value that meets the comparison conditions. The support level threshold is determined by enumerating candidate values ​​of the level-covered number segments in the historical user samples, comparing the proportion of mislabeled fields in the sample subset corresponding to each candidate value with the preset mislabeling ratio upper limit, and selecting the smallest candidate value that meets the comparison conditions. When the historical user sample size does not meet the preset sample size threshold, the threshold value corresponding to the previous effective time is used.

[0085] Excluding a candidate learning label from the set of labels to be verified indicates that the sample support index of the candidate learning label item has met the support threshold. That is, the number of valid evidence items and the number of hierarchical coverages have reached the preset thresholds, and there is a reproducible statistical basis. Therefore, it will not enter the subsequent processing queue, which is centered on obtaining the question discrimination level, obtaining the source reliability level, supervised learning regression inference, and the placement of differentiation interaction tasks. This avoids repeatedly triggering high-cost data collection and calculation for low-risk items. At the same time, this processing is not equivalent to directly solidifying it into a stable label. Instead, resources are prioritized for items with insufficient support or insufficient coverage, so that the insufficient sample screening and mislabeling risk assessment are carried out in a tiered manner.

[0086] Step 3: For the set of labels to be verified, obtain the question discrimination level and source reliability level, and perform supervised learning regression inference to output the upper bound of mislabeling risk; compare the upper bound of mislabeling risk with the mislabeling threshold to obtain the results that are allowed to be fixed and the results that are prohibited from being fixed. The labels to be verified corresponding to the results that are allowed to be fixed are determined as stable labels and added to the recommendation set.

[0087] For each tag in the set to be verified, the question discrimination level and source reliability level are obtained, as follows:

[0088] The system receives a set of tags to be verified and reads each tag's evidence group identifier field, tag type field, knowledge point index field, source type sequence field, valid evidence segment number field, correct evidence segment number field, incorrect evidence segment number field, record timestamp start field, and record timestamp end field. Based on the evidence group identifier field, it retrieves the list of valid evidence corresponding to that tag. From the list of valid evidence, it filters out evidence entries whose record type field is practice answer records and includes a question entry identifier field, forming a question entry identifier list. Then, it reads each question bank statistical entry from the question entry identifier list. These statistical entries are stored indexed by the question entry identifier and include a statistical period field and a valid answer sample size. The system consists of four fields: high score accuracy, low score accuracy, and statistical period. The statistical period field takes a fixed start and end date pair, the effective sample size field takes an integer value, and the high score accuracy and low score accuracy fields take a ratio. For each question item identifier, based on the difference range between the high score accuracy and low score accuracy fields in the question bank statistics, the corresponding question discrimination level is retrieved from the discrimination level table. The discrimination level table uses the difference range as the key and outputs a discrete level identifier, which takes one of five categories: level 1, level 2, level 3, and level 4. The question discrimination level of the question item identifier is then recorded in the question discrimination level sequence field of the label to be verified.

[0089] It should be noted that the discrimination level table is built using question entry identifiers as indexes, and the data source is limited to question bank statistical entries. The statistical period field of the question bank statistical entries uses fixed start and end date pairs, and levels are generated only for questions whose valid answer sample size field meets the preset sample size threshold. Answer samples within the same statistical period are divided into high-score and low-score sets according to preset grouping rules, resulting in high-score accuracy fields and low-score accuracy fields respectively. The difference between these two is taken as the discrimination input. The discrimination level table pre-fixes a one-to-one correspondence between the difference range key and the discrete level identifier, which is limited to levels one to five. When the difference falls within the difference range key, the corresponding level is output and written to the statistical period field, effective time field, and version entry field. If the sample size threshold is not met, the previous version entry is used; if no previous version entry exists, a preset default level identifier is written.

[0090] Subsequently, the source reliability level is obtained. Specifically, the source type sequence field of the label to be verified is deduplicated to obtain a source type list. Each source type in the source type list is read one by one, and the corresponding entry in the source reliability level table is read. The source reliability level table is stored with the source type as the index and includes four fields: statistical period field, statistical sample size field, historical label prediction accuracy field, and reliability level field. The reliability level field takes one of five categories: level one, level two, level three, and level four. For the same label to be verified, the reliability level field of each source type in the source type list and the number of occurrences of that source type in the list of valid evidence are combined to form a source reliability level summary record. The source reliability level summary record includes three fields: source type, reliability level, and number of occurrences. This source reliability level summary record is written into the source reliability level sequence field of the label to be verified. Thus, both the question discrimination level and the source reliability level are obtained by the method of retrievable list fields - level table entries - discrete level identifier. The acquisition process continuously references the evidence group identifier field and the source type sequence field in the set of labels to be verified to ensure the consistency of the fields in the preceding and following steps.

[0091] The supervised learning regression inference is performed, and the upper bound of the mislabeling risk is output as follows:

[0092] For each tag in the set of tags to be verified, a regression inference input record is constructed. This record is written in a fixed field order, including: tag type field, knowledge point index field, valid evidence number field, correct evidence number field, incorrect evidence number field, level coverage number field, question discrimination level sequence field, source reliability level sequence field, record time span field, and source type diversity field. The record time span field is formed by the record timestamp end field and the record timestamp start field, and the source type diversity field is formed by the number of entries in the source type list. After the regression inference input record is formed, the regression inference parameter set matching the tag type field is read from the regression parameter storage area. The regression inference parameter set is indexed by the tag type. The system stores and includes four fields: parameter group version identifier, parameter group effective time, parameter group field sequence list, and parameter group value. The parameter group field sequence list is used to match the order of the regression inference input record fields, and the parameter group value field is used to carry the mapping relationship used in supervised learning regression inference. Subsequently, the regression inference input record and the regression inference parameter group are submitted to the regression inference execution process. The regression inference execution process outputs the upper bound value of mislabeling risk and writes the upper bound value of mislabeling risk to the upper bound value field of the label to be verified. At the same time, the parameter group version identifier and parameter group effective time are written to the inference version field of the label to be verified to ensure that the upper bound value of mislabeling risk for the same label to be verified has a traceable parameter source.

[0093] The regression inference parameter set is derived from the fixed output of the offline training phase. During offline training, historical user samples are used to construct training sample records, and the input fields of these records are consistent with the regression inference input record fields. The output field of the training sample records is the mislabeling field, whose value is limited to either mislabeled or not mislabeled. The mislabeling field is generated based on the consistency between early label results and subsequent stable results: early label results are taken from label records formed during the cold start phase, and these records include at least a knowledge point index field and a label type field; subsequent stable results are taken from the final label records of the same new user after additional interaction data reaches the preset additional sample conditions. The records include at least a knowledge point index field and a label type field. Early label results and subsequent stable results are aligned according to the knowledge point index field and then compared for consistency. When the label type field under the same knowledge point index field is inconsistent, the mislabeling marker field is assigned a mislabel value; when the label type field under the same knowledge point index field is consistent, the mislabeling marker field is assigned a non-mislabel value. Simultaneously, a sample time window field and a sampling interval field are synchronously written into the training sample records as the data retrieval criteria. The sample time window field takes a fixed start and end time period pair, and the sampling interval field takes a fixed number of seconds, thus ensuring that the input and output fields of supervised learning regression inference are consistent with the data retrieval criteria and are reproducible.

[0094] It should be noted that the number of erroneous and correct entries in the regression inference input records jointly define the direction of evidence. When the difference between the number of erroneous and correct entries leads to an increase in the degree of conflict, the upper bound of mislabeling risk shifts towards the risk side. The number of valid pieces of evidence in the regression inference input records defines the scale of evidence. When the number of valid pieces of evidence falls into the insufficient support range, the upper bound of mislabeling risk shifts towards the risk side. The hierarchical coverage in the regression inference input records defines the hierarchical span of evidence coverage. When the hierarchical coverage only covers a single level or the coverage is insufficient, the upper bound of mislabeling risk shifts towards the risk side. The question discrimination level in the regression inference input records defines the source of the discriminative power of the practice answer evidence. When the gradation level falls into the low-gradation-level range, the upper limit of mislabeling risk shifts towards the risk side; the source reliability level in the regression inference input records limits the historical stability of the evidence source, and when the source reliability level falls into the low-reliability-level range, the upper limit of mislabeling risk shifts towards the risk side; the recording time span in the regression inference input records limits the time distribution range of the evidence, and when the recording time span is too short, resulting in evidence concentrated at a single point in time, the upper limit of mislabeling risk shifts towards the risk side; the source type diversity in the regression inference input records limits the dispersion of the evidence source, and when insufficient source type diversity results in evidence concentrated at a single source type, the upper limit of mislabeling risk shifts towards the risk side. The determination of the above "shift towards the risk side" is not subjectively set in the online stage.

[0095] By comparing the upper bound of the mislabeling risk with the mislabeling threshold, we obtain the allowed and prohibited solidification results. The labels corresponding to the allowed solidification results are determined as stable labels and added to the recommendation set, as detailed below:

[0096] To read the upper bound of the mislabeling risk field, the label type field, the knowledge point index field, and the level coverage number field for each label in the set of labels to be verified, and to select the applicable mislabeling threshold from the mislabeling threshold list, which is indexed by label type and level number and includes four fields: mislabeling threshold field, effective time field, applicable level range field, and applicable label type field, with the applicable level range field represented by the start and end level numbers; for each label to be verified, first match the applicable label type field based on the label type field, and then match the level coverage number field. The corresponding level number falls into the applicable level range field, determining the value of the applicable mislabeling threshold field. The upper limit of the mislabeling risk is compared with the applicable mislabeling threshold. If the upper limit of the mislabeling risk is not greater than the mislabeling threshold, an "allow solidification" result is output; if the upper limit of the mislabeling risk is greater than the mislabeling threshold, a "prohibit solidification" result is output. For the unverified tags that output an "allow solidification" result, their tag status field is updated to "stable tag," and the tag identifier field, tag type field, knowledge point index field, prerequisite dependency field, source type sequence field, and record timestamp start field of the stable tag are updated. The timestamp end field is copied to the recommendation set generation list, which is indexed by the new user identifier and includes a stable tag entry set field. For tags to be verified that are prohibited from being solidified, their tag status field is kept as "tag to be verified," and the tag identifier field, mislabeling risk upper limit field, and applicable mislabeling threshold field are written to the prohibited solidification record list, which is indexed by the new user identifier and includes a prohibited solidification entry set field. During the recommendation set generation stage, the stable tag entry set field in the recommendation set generation list is read, and the corresponding content entries are extracted from the course content index list using the knowledge point index field as the content index key. The course content index list is stored according to the knowledge point index field and includes four fields: content entry identifier field, content type field, difficulty gradient field, and required path number field. Then, the content entries are arranged in the prerequisite order according to the prerequisite dependency field of the stable tag to form a recommendation set entry sequence. The recommendation set entry sequence is then associated with the tag identifier field of the stable tag, thereby determining the tag corresponding to the allowed solidification result as a stable tag and adding it to the recommendation set.

[0097] The mislabeling threshold field is determined by enumerating candidate thresholds for the upper bound of mislabeling risk within the corresponding groups of historical user samples, comparing the proportion of mislabeling markers in the sample subset corresponding to the candidate threshold with the preset upper limit of mislabeling proportion, and selecting candidate thresholds that meet the comparison conditions. In the online stage, the hierarchical number field is retrieved based on the tag type field and knowledge point index field of the tag to be verified, and the applicable entries are matched in the mislabeling threshold list and the mislabeling threshold field is read as the applicable mislabeling threshold field.

[0098] Step 4: Generate alternative labels for the labels to be verified corresponding to the prohibited solidification results and retrieve the set of distinguishing interaction tasks. Calculate the consistency index based on the verification evidence sequence collected from the set of distinguishing interaction tasks and compare it with the consistency threshold to obtain pseudo-stable label markings. Perform a threshold comparison between the upper bound of the mislabeling risk, the consistency index, and the pseudo-stable label markings for the labels to be verified corresponding to the prohibited solidification results, and output the label decision results. The label decision results include labels to be verified or prohibited recommended labels. Update the recommendation set based on the label decision results.

[0099] For the unverified tags corresponding to the prohibited solidification results, alternative tags are generated and the set of interactive tasks is retrieved and differentiated, as follows:

[0100] Based on the list of prohibited solidification results and its corresponding set of tags to be verified, for each tag to be verified, the tag identifier field, tag type field, knowledge point index field, evidence group identifier field, source type sequence field, valid evidence segment number field, correct segment number field, incorrect segment number field, mislabeling risk upper limit value field, record timestamp start field, and record timestamp end field are read sequentially, and the tag identifier field is used as the identifier of the processing object in this step; within the same tag to be verified range, alternative tags are generated according to the alternative tag rule list, where the alternative tag rule list is stored in five fields: original tag type field, alternative tag type field, applicable knowledge point level field, applicable record type field, and applicable source type field. The original tag type field is limited to one of three categories: knowledge point mastery tag, weak link tag, and learning preference tag, and the alternative tag type field is limited to one category that is mutually exclusive with the original tag type field.

[0101] For example, if the original tag type field is a knowledge point mastery tag, the substitute tag type field is the weak link tag; if the original tag type field is a weak link tag, the substitute tag type field is the knowledge point mastery tag; if the original tag type field is a learning preference tag, the substitute tag type field is the non-preference tag; the applicable knowledge point level field takes the value of the level number range in the teaching structured constraint set; the applicable record type field is limited to one or a combination of four categories: practice answer record, course browsing record, dwell time record, and question and answer record; the applicable source type field is limited to one or a combination of six categories: in-class exercises, chapter quizzes, homework, free questions, table of contents browsing, and video playback; for the tag to be verified, first retrieve the level number field in the teaching structured constraint set based on the knowledge point index field, and then match the original tag type field with the tag type field. A unique alternative label rule entry is selected based on the condition that the level number field falls within the applicable knowledge point level field. The alternative label type field is read from the selected entry, and an alternative label is generated. The alternative label includes at least an alternative label identifier field, an alternative label type field, a knowledge point index field, and an evidence group identifier field. The alternative label identifier field is formed by concatenating the label identifier field of the label to be verified with the alternative label type field. The alternative label identifier field and the label identifier field to be verified are written together into the label pair identifier field. The label pair identifier field is used to bind the differentiation interaction task and the verification evidence sequence to the same opposing label relationship in the future. After the alternative label is generated, the differentiation interaction task set is retrieved based on the knowledge point index field, the label type field of the label to be verified, and the alternative label type field of the alternative label. The differentiation interaction task set is obtained from the differentiation task template library.

[0102] The task template library is stored item by item with the template identifier field as the primary key. Each template item includes at least the template identifier field, the applicable knowledge point index field, the applicable original tag type field, the applicable alternative tag type field, the task content item identifier field, the task presentation method field, the task input field list, the task output field list, the task result judgment rule field, the task difficulty gradient field, and the task source type field. The task presentation method field is limited to one of five categories: single question answering, multiple question answering, short question answering, answering after short reading, and answering after video clips. The task input field list clearly records the field names collected during interaction, the task output field list clearly records the field names written in the output evidence items, and the task result judgment rule field is written in the form of discrete rule items.

[0103] For example, if the output is correct and the response time falls within the rule range, it is recorded as supporting the original tag; if the output is incorrect and the error type falls within the rule range, it is recorded as supporting the alternative tag; if the output is skipped or timed out, it is recorded as indistinguishable. The task source type field is limited to in-class exercises or chapter quizzes to ensure consistent source criteria. During retrieval, a candidate template list is obtained by matching the applicable knowledge point index field with the knowledge point index field, the applicable original tag type field with the tag type field, and the applicable alternative tag type field with the alternative tag type field. The candidate template list is then gradient-bound according to the task difficulty gradient field and the difficulty gradient field in the teaching structure constraint set. The gradient binding process involves reading the difficulty gradient field value corresponding to the knowledge point index field, prioritizing templates in the candidate template list whose task difficulty gradient field value is the same as that difficulty gradient field value, arranging templates with adjacent values ​​in order, and writing the sorted template identifier field into the task list field. Finally, the tag pair identifier field, template identifier field, task content item identifier field, and task source type field are written together into the task delivery record. The task delivery record is stored with the new user identifier field as the primary key and includes a delivery timestamp field.

[0104] Collect and verify evidence sequences, calculate consistency indices, and compare them with consistency thresholds to obtain pseudo-stable labels, as detailed below:

[0105] Based on the task delivery records, interactive tasks corresponding to the task content item identifier fields are presented one by one. When a user completes an interaction, data is collected according to the task output field list to form a verification evidence sequence. Each verification evidence sequence consists of one verification evidence item per task interaction. A single verification evidence item includes at least the following fields: New User Identifier, Tag Pair Identifier, Template Identifier, Task Content Item Identifier, Record Timestamp, Record Type, Knowledge Point Index, Source Type, Response Result, Response Duration, Skip Flag, and Correct / Incorrect. The Record Type field is limited to either practice response records or question-and-answer records to match the template presentation. The Source Type field is written by the Task Source Type field. The Response Result field contains options or text according to the task input / output definition. The Response Duration field records the number of seconds from presentation to submission. The Skip Flag field is limited to skipped or not skipped. The Correct / Incorrect field is limited to correct, incorrect, or no result. After the verification evidence sequence is formed, a consistency index is calculated for the same tag pair identifier fields. The consistency index is not expressed by a formula but is generated using a task support classification-counting-level mapping method.

[0106] First, retrieve the task result judgment rule field from the task template library according to the template identifier field. Apply the task result judgment rule field to each verification evidence item based on its correct / incorrect field, response time field, and skip flag field to obtain the task support category field. The value of the task support category field is limited to three categories: support for the label to be verified, support for the alternative label, and cannot be distinguished. Write the task support category field back to the verification evidence item.

[0107] Second, for the same label pair identifier field, count the number of verification evidence entries that support the label to be verified, the number of entries that support the alternative label, and the number of entries that cannot be distinguished, and write the three types of entries into the consistency summary record. The consistency summary record includes at least the label pair identifier field, the number field of the number of entries that support the label to be verified, the number field of the number ...

[0108] Third, the consistency summary records are mapped to consistency indicator fields. These consistency indicator fields use discrete level identifiers provided by a consistency level table. The consistency level table uses the following input keys: the range of differences between the number of records to be verified and the number of records to be substituted; the range of indistinguishable proportions of records; and the range of valid task records. The output is a consistency level identifier, which is limited to one of five levels: Level 1, Level 2, Level 3, Level 4, and Level 5. This consistency level identifier is written into the consistency indicator field of the label to be verified. After the consistency indicator field is written, a consistency threshold comparison is performed. The consistency threshold is provided by a consistency threshold list, which stores records by label type field and knowledge point level field, and includes a consistency threshold level field and an effective time field. The consistency threshold level field takes one of the aforementioned consistency level identifiers. For the label to be verified, its label type field is matched against the label type field in the consistency threshold list. Then, based on the knowledge point index field, the level number field is retrieved from the teaching structured constraint set and matched against the knowledge point level field in the consistency threshold list to obtain the applicable consistency threshold level field. Finally, the consistency indicator field and the consistency threshold level field are compared in order of rank.

[0109] When the consistency index field does not meet the requirement of not less than the consistency threshold level field, a pseudo-stable label field is generated and assigned the value of pseudo-stable. When the consistency index field meets the requirement of not less than the consistency threshold level field, the pseudo-stable label field is assigned the value of non-pseudo-stable. The consistency index field, the applicable consistency threshold level field, the number segments to be verified, the number segments to be replaced, and the number segments that cannot be distinguished are all written into the comparison record field.

[0110] It should be noted that the consistency threshold list is grouped by tag type field and knowledge point level field. Each group provides a consistency threshold level field as the applicable threshold. The consistency threshold level field uses the mislabeling mark field of historical user samples as the constraint target. Within the same group, candidate consistency threshold levels are enumerated, and the proportion of mislabeling mark fields in the sample subset that meets the candidate level is compared with the preset mislabeling proportion upper limit. Candidate levels that meet the comparison conditions are selected and written into the consistency threshold list. The consistency threshold list also limits the sample records participating in the setting to meet the condition that the number of valid tasks reaches the preset task quantity condition and the proportion of indistinguishable tasks falls within the preset grading range. The consistency threshold list entries include an effective time field. When updating, the consistency threshold level field is replaced only when the statistical sample size field is not less than the preset sample size threshold; otherwise, the previous effective time entry is used.

[0111] The updated recommendation set is as follows:

[0112] For each prohibited solidification result, the following fields are read for the label to be verified: upper bound of mislabeling risk, applicable key of mislabeling threshold, consistency index, applicable consistency threshold level, pseudo-stable label mark, label identifier, knowledge point index, prerequisite dependency, and source type sequence. The applicable mislabeling threshold is retrieved from the mislabeling threshold list, which is stored item by item according to label type and knowledge point level, and includes the mislabeling threshold and effective time fields. The selection process for the applicable mislabeling threshold involves matching the label type field in the mislabeling threshold list with the label type field, and then retrieving the level number field from the teaching structured constraint set based on the knowledge point index field and matching it with the knowledge point level field in the mislabeling threshold list to obtain the mislabeling threshold value. Subsequently, a joint threshold comparison is performed, and the label decision result is output. The joint threshold comparison includes two comparisons and one constraint.

[0113] The first comparison is between the upper limit of the mislabeling risk and the mislabeling threshold field. When the upper limit of the mislabeling risk does not meet the requirement of not being greater than the mislabeling threshold field, it is recorded as a failure to pass the mislabeling risk test; when the upper limit of the mislabeling risk meets the requirement of not being greater than the mislabeling threshold field, it is recorded as a success to pass the mislabeling risk test.

[0114] The second comparison is between the consistency indicator field and the applicable consistency threshold level field. If the consistency indicator field does not meet the requirement of being no less than the applicable consistency threshold level field, it is recorded as consistency failure; if the consistency indicator field meets the requirement of being no less than the applicable consistency threshold level field, it is recorded as consistency success.

[0115] The constraint term is a pseudo-stability label field. When the pseudo-stability label field is pseudo-stable, it is directly recorded as a pseudo-stability constraint being triggered; when the pseudo-stability label field is non-pseudo-stability, it is recorded as a pseudo-stability constraint not being triggered.

[0116] After the three judgment results are formed, the label decision result field is output. The value of the label decision result field is limited to one of two categories: label to be verified or label to be prohibited from recommendation. When the mislabeling risk fails, the consistency fails, or the pseudo-stability constraint is triggered, the label decision result field is assigned the value of label to be prohibited from recommendation. When the mislabeling risk passes, the consistency passes, and the pseudo-stability constraint is not triggered, the label decision result field is assigned the value of label to be verified. The above three judgment results, along with the mislabeling risk upper limit field, mislabeling threshold field, consistency index field, and applicable consistency threshold level field used for comparison, are written into the label decision record field. After the label decision result field is written, the recommendation set is updated. The recommendation set is stored with the new user identifier field as the primary key and includes the recommendation item sequence field. The recommendation item sequence is stored in units of items. Each item includes at least the content item identifier field, knowledge point index field, associated label identifier field, item type field, sorting field, and insertion source field. The item type field is limited to one of two categories: learning content or distinguishing interactive tasks.

[0117] When the tag decision result field is a prohibited tag, iterate through the recommended item sequence and remove items whose associated tag identifier field is equal to the tag identifier field of the tag to be verified and whose knowledge point index field is equal to the knowledge point index field of the tag to be verified from the recommended item sequence. At the same time, based on the prerequisite dependency field of the tag to be verified, retrieve the learning content items corresponding to its prerequisite knowledge point index in the course content index list. Sort the retrieved learning content items by the required path number field in ascending order and insert them into the recommended item sequence, writing "prohibited tag replacement" in the insertion source field. When the tag decision result field is a tag to be verified, retain the learning content items in the recommended item sequence whose associated tag identifier field is equal to the tag identifier field of the tag to be verified, and add the tag pair identifier field of the task delivery record in the previous sub-step that is equal to the tag pair corresponding to the tag to be verified. The identification field distinguishes interactive task entries and adds them to the recommended entry sequence. When adding, the entry type field is assigned the value of the distinguishing interactive task, and the insertion source field is written with the tag to be verified to supplement the distinguishing task. The addition order follows the order of the delivery timestamp field in the task delivery record from earliest to latest. Finally, the updated recommended entry sequence is written back to the recommended set, and the content entry identification field, associated tag identification field, and triggered tag decision record field of the removed and inserted entries in the recommended entry sequence are written to the recommended change record list. This makes the entire process of generating alternative tags and retrieving the distinguishing interactive task set, collecting verification evidence sequences and calculating consistency indicators and comparing them with the consistency threshold to obtain pseudo-stable tag markings, comparing the comprehensive mislabeling risk upper bound with the consistency indicator execution threshold, outputting tag decision results, and updating the recommended set a readable logic at the field level.

[0118] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent question-and-answer information processing in AI-powered online education, characterized in that, include: Obtain the cold start interaction set and teaching structured constraint set of new users, process them, and output the cold start sequence; Generate a candidate learning label set based on evidence aggregation performed on the cold start sequence; Candidate learning label entries must include at least an evidence group identifier field. Based on the evidence group identifier field, the evidence list is retrieved from the cold start sequence and effective evidence screening is performed. Based on the screening results, the sample support index is calculated and compared with the support threshold, and the set of labels to be verified is output. Methods for outputting the set of tags to be verified include: Receive the set of candidate learning labels after the effective evidence screening is completed, and generate a sample support index for each candidate learning label item in the set; Based on the candidate learning label entries, the effective evidence number segment is extracted as the effective evidence number. Based on the knowledge point index field, the hierarchical number is retrieved in the knowledge point hierarchy relationship of the teaching structured constraint set to form a hierarchical number set. The hierarchical number set is deduplicated to obtain the hierarchical coverage number and written into the hierarchical coverage number segment. The effective evidence number and the hierarchical coverage number constitute the sample support index. Support threshold comparison is performed based on the sample support index. The support threshold consists of a preset support number threshold and a preset support level threshold. When the number of valid evidence items meets or exceeds the preset support number threshold, and the level coverage meets or exceeds the preset support level threshold, the candidate learning label item is not included in the label set to be verified. When the number of valid evidence items does not meet or exceeds the preset support number threshold, or the level coverage does not meet or exceeds the preset support level threshold, the candidate learning label item is included in the label set to be verified, and the following fields are retained: label identifier field, label type field, knowledge point index field, prerequisite dependency field, evidence group identifier field, valid evidence item number field, level coverage number field, correct item number field, incorrect item number field, entry item number field, cumulative stay seconds number field, question item number field, source type sequence field, record timestamp start field, and record timestamp end field. For the set of tags to be verified, the question discrimination level and source reliability level are obtained and supervised learning regression inference is performed to output the upper bound of mislabeling risk. The upper bound of mislabeling risk is compared with the mislabeling threshold to obtain the results that are allowed to be solidified and the results that are prohibited from being solidified. The tags to be verified corresponding to the results that are allowed to be solidified are determined as stable tags and added to the recommendation set. For the unverified labels corresponding to the prohibited solidification results, alternative labels are generated and the interactive task set is retrieved and distinguished. The verification evidence sequence is collected, the consistency index is calculated and compared with the consistency threshold to output pseudo-stable label markings. The label decision results are output by comprehensively comparing the upper limit of mislabeling risk, consistency index and pseudo-stable label markings, and the recommendation set is updated accordingly.

2. The AI-powered intelligent question-and-answer information processing method for online education according to claim 1, characterized in that, The cold start interaction set includes question and answer records, practice answer records, course browsing records and dwell time records. The teaching structure constraint set includes knowledge point index, parent knowledge point index, prerequisite knowledge point index set, required path sequence number and difficulty gradient identifier. It also includes a corresponding list of question item identifiers and knowledge point indexes, a corresponding list of content item identifiers and knowledge point indexes, and a corresponding list of question and answer item identifiers and knowledge point indexes. Time alignment is performed based on the cold start interaction set to form a time alignment sequence. Then, based on the corresponding list, the question item identifier, content item identifier, and question and answer item identifier are mapped to knowledge point indexes and the knowledge point index field is filled to obtain the time alignment mapping sequence. Based on the knowledge point index field in the time-aligned mapping sequence, retrieve the teaching structured constraint set, extract the prerequisite knowledge point index set, parent knowledge point index, required path number and difficulty gradient identifier, fill in the prerequisite dependency field, level pointing field, path position field and difficulty gradient field, and output the cold start sequence.

3. The AI-powered intelligent question-and-answer information processing method for online education according to claim 1, characterized in that, Methods for generating candidate learning label sets based on evidence aggregation of cold-start sequences include: Receive the cold start sequence and arrange the cold start sequence from earliest to latest according to the recorded timestamp within the same new user ID range; Based on the permutation results, the cold start sequence is merged according to the knowledge point index field to obtain evidence groups with the knowledge point index field as the primary key, and an evidence group identifier is generated for each evidence group. Based on the evidence group, the record type field, source type field, and correct / incorrect field are extracted. The number of correct records, the number of incorrect records, the number of questions, the number of answers, the number of entries, the number of exits, the cumulative stay time, and the maximum stay time are counted respectively, and a source type sequence is formed. Candidate learning tag entries are generated based on statistical results. Candidate learning tag entries include tag identifier field, tag type field, knowledge point index field, prerequisite dependency field, and evidence group identifier field. The tag type field is determined by comparing the number of erroneous entries with the preset error number threshold, the number of correct entries with the preset correct number threshold, the number of entries entered with the preset browsing entry threshold, the cumulative dwell time in seconds with the preset cumulative dwell time threshold, and the maximum dwell time in seconds with the preset maximum dwell time threshold. The candidate learning tag entries are then summarized to obtain a candidate learning tag set.

4. The AI-powered intelligent question-and-answer information processing method for online education according to claim 3, characterized in that, Methods for screening valid evidence include: Receive the set of candidate learning labels, and read the evidence group identifier field of each candidate learning label item one by one. Based on the evidence group identifier field, retrieve the corresponding evidence from the cold start sequence to form an evidence list. Based on the evidence list, the effective evidence is filtered item by item. The effective evidence filtering includes source type filtering, field completeness filtering, record type matching filtering, and short-term duplicate filtering. The output is a set of effective evidence and a set of ineffective evidence. The statistical fields of candidate learning label entries are updated based on the valid evidence set. The statistical fields include the number segments of valid evidence entries, correct entries, incorrect entries, question entries, answer entries, entry entries, exit entries, cumulative stay seconds, and maximum stay seconds. The source type field in the valid evidence set is deduplicated to obtain the valid source type set and the source type sequence field is updated.

5. The AI-powered intelligent question-and-answer information processing method for online education according to claim 4, characterized in that, Methods for obtaining the discrimination level and source reliability level of each item in the set of tags to be verified include: Receive the set of tags to be verified, read the evidence group identifier field of each tag to be verified, and retrieve the corresponding list of valid evidence based on the evidence group identifier field; Based on the list of valid evidence, evidence entries whose record type field is practice answer records and include the question entry identifier field are selected to form a question entry identifier list. Based on the question entry identifier list, the question bank statistical entries are read one by one, and the question discrimination level is found in the discrimination level table according to the difference range between the high score segment accuracy field and the low score segment accuracy field in the question bank statistical entries. The question discrimination level is written into the question discrimination level sequence field. Based on the list of valid evidence, the source type sequence field is deduplicated to obtain the source type list. The reliability level field in the source reliability level table is read one by one, and the source reliability level summary record is formed by combining the number of occurrences of the source type in the list of valid evidence. The source reliability level summary record is written into the source reliability level sequence field.

6. The AI-powered intelligent question-and-answer information processing method for online education according to claim 5, characterized in that, Methods for adding items to a recommendation set include: Receive a set of tags to be verified, and construct a regression inference input record for each tag in a fixed field order. The regression inference input record includes a tag type field, a knowledge point index field, a valid evidence bar number field, a correct bar number field, an incorrect bar number field, a level coverage bar number field, a question discrimination level sequence field, a source reliability level sequence field, a record time span field, and a source type diversity field. The regression inference execution process outputs the upper bound value of the mislabeling risk. Read the upper bound field of mislabeling risk, the tag type field, the knowledge point index field, and the level coverage number field of the tag to be verified in the tag set one by one, and select the applicable mislabeling threshold field from the mislabeling threshold list based on the tag type field and the level coverage number field. Based on the applicable mislabeling threshold field, the upper limit of mislabeling risk is compared with the mislabeling threshold. When the upper limit of mislabeling risk is not greater than the mislabeling threshold, the result of allowing solidification is output. When the upper limit of mislabeling risk is greater than the mislabeling threshold, the result of prohibiting solidification is output. Based on the results of allowing solidification, the tag status field of the tag to be verified is updated to a stable tag, and the tag identifier field, tag type field, knowledge point index field, prerequisite dependency field, source type sequence field, record timestamp start field, and record timestamp end field of the stable tag are written into the recommendation set generation list.

7. The AI-powered intelligent question-and-answer information processing method for online education according to claim 6, characterized in that, Methods for generating alternative labels for the unverified labels corresponding to the prohibited solidification results and for retrieving and distinguishing the set of interactive tasks include: Receive the list of prohibited solidification results and its corresponding set of tags to be verified, read the tag identifier field, tag type field, knowledge point index field and evidence group identifier field of the tags to be verified one by one, and retrieve the hierarchical number field of the teaching structured constraint set based on the knowledge point index field. Based on the tag type field and the level number field, the original tag type field and the applicable knowledge point level field are matched in the list of alternative tag rules. The alternative tag type field is read and an alternative tag is generated. The alternative tag includes an alternative tag identifier field, an alternative tag type field, a knowledge point index field and an evidence group identifier field. The alternative tag identifier field and the tag identifier field are combined to form a tag pair identifier field. Based on the tag pair identifier field, knowledge point index field, tag type field, and alternative tag type field, the task template library is used to retrieve and differentiate the interactive task set. The template identifier field, task content item identifier field, and task source type field are read to form the task list field. Based on the task list field and the tag pair identifier field, a task delivery record is formed. The task delivery record includes the new user identifier field and the delivery timestamp field.

8. The AI-powered intelligent question-and-answer information processing method for online education according to claim 7, characterized in that, Methods for obtaining pseudo-stable labels include: Based on the task delivery record, interactive tasks corresponding to the task content item identifier fields are presented, and data is collected based on the task output field list to form a verification evidence sequence. The verification evidence sequence includes new user identifier field, tag pair identifier field, template identifier field, task content item identifier field, record timestamp field, record type field, knowledge point index field, source type field, answer result field, answer duration field, skip mark field, and correct / incorrect field. Based on the verification evidence sequence, the task result judgment rule field is retrieved by the template identifier field, and the task support category field is generated according to the correct / incorrect field, the response time field, and the skip flag field. The task support category field has the values ​​of supporting the label to be verified, supporting the alternative label, and not distinguishable. The number of items supporting the label to be verified, the number of items supporting the alternative label, and the number of items not distinguishable are counted to form a consistent summary record. Based on the consistency summary records, the consistency index field is determined in the consistency level table, and the applicable consistency threshold level field is determined based on the preset consistency threshold list. The consistency index field and the applicable consistency threshold level field are compared in order of rank, and the pseudo-stable label field is output. The pseudo-stable label field can be either pseudo-stable or non-pseudo-stable.

9. The AI-powered intelligent question-and-answer information processing method for online education according to claim 8, characterized in that, Methods for updating the recommendation set include: Read the upper bound field of mislabeling risk, consistency index field, applicable consistency threshold level field, pseudo-stable label mark field, label identifier field, knowledge point index field and prerequisite dependency field of the label to be verified corresponding to the prohibited solidification result one by one, and retrieve the mislabeling threshold field from the mislabeling threshold list based on the label type field and knowledge point index field; The label decision result field is determined by comparing the upper limit value field of mislabeling risk with the mislabeling threshold field, comparing the level order based on the consistency index field and the applicable consistency threshold level field, and combining the pseudo-stable label marking field. The label decision result field takes the value of either a label to be verified or a label that is prohibited from being recommended. The recommendation set is updated based on the tag decision result field. When the tag decision result field is a prohibited recommendation tag, the entries in the recommendation entry sequence whose associated tag identifier field is equal to the tag identifier field and whose knowledge point index field is equal to the knowledge point index field are removed from the recommendation entry sequence. Learning content entries are then retrieved from the course content index list based on the prerequisite dependency field and written into the recommendation entry sequence. When the tag decision result field is a tag to be verified, the learning content entries in the recommendation entry sequence whose associated tag identifier field is equal to the tag identifier field are retained. Interactive task entries are then added to the recommendation set based on the task delivery record.

Images