A dynamic multi-path retrieval method and system for semantic profiling of primary and secondary school test questions

By generating semantic profiles with source and confidence levels and educational hard constraint bitmaps, and dynamically controlling retrieval channel quotas, the problem of unified representation of multimodal queries and verification of educational attributes in primary and secondary school test question retrieval systems is solved, thereby improving retrieval efficiency and accuracy.

CN123068604APending Publication Date: 2026-11-10SHENZHEN ERYI EDUCATION CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611249287.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-18
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing test question retrieval systems for primary and secondary schools cannot effectively unify multimodal query representations, making it difficult to verify the search intent and educational attributes. Fixed multi-path retrieval results in a high proportion of unqualified candidates, duplicate aggregation, and large computational overhead, and it is impossible to adjust resource allocation according to query intent and real-time status.

Method used

By generating a semantic profile of test questions with source and confidence level through a unified query object, and combining the candidate bitmap of education hard constraints and multi-way execution plan, the quota of retrieval channels is dynamically controlled to ensure that the questions are recalled in batches and finely ranked within the scope of education constraints.

Benefits of technology

It achieves efficient and accurate retrieval within educational constraints, reduces unqualified candidates across disciplines and grades, improves the quality and efficiency of retrieval results, and controls computational overhead.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123068604A_ABST
    Figure CN123068604A_ABST
Patent Text Reader

Abstract

This invention discloses a dynamic multi-path retrieval method and system for semantic profiling of primary and secondary school test questions, belonging to the field of educational informatization and information retrieval. To address the problems of inconsistent processing of multimodal test question queries, misinterpretation of educational constraints, and false recall, duplicate aggregation, tail delay, and excessive computational cost caused by fixed multi-path retrieval, this invention generates verifiable semantic profiles, compiles educational hard constraints into candidate bitmaps, dynamically selects compatible retrieval channels, and allocates retrieval quotas based on batch marginal utility and completion probability. Within a common deadline, it completes evidence fusion, question family aggregation, fine ranking, and constraint verification, achieving multimodal test question retrieval controlled by educational hard constraints.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of educational informatization and information retrieval, and in particular to a dynamic multi-path retrieval method and system for semantic profiling of primary and secondary school test questions. Background Technology

[0002] As the scale of digital teaching resources in primary and secondary schools continues to grow, test question banks typically contain heterogeneous content such as structured question fields, natural language question stems, scanned images, page layouts, and mathematical formulas. Existing retrieval systems mostly perform recall using a combination of keywords, vector similarity, or fixed indexes. Different input modalities often employ independent preprocessing and retrieval paths, making it difficult to form a unified query representation that can be reused across different retrieval channels.

[0003] In scenarios where original question retrieval, similar question retrieval, and teaching demand retrieval coexist, query intent and educational attributes such as subject, grade level, grade, textbook version, and knowledge point may come from different input fragments and have different levels of confidence. If the model output is directly used as a filtering condition, it is easy to cause missed recalls or false recalls across subjects or grades due to attribute conflicts or insufficient confidence. If educational constraint filtering is not performed, semantic computation and fine ranking need to be performed on a large number of unqualified candidates.

[0004] Furthermore, fixed-channel or parallel retrieval across all channels cannot adjust resources based on query intent, input modality, profile completeness, and real-time execution status. Duplicate questions returned by different channels are prone to clustering, and slow channels can amplify tail latency; simple timeout degradation, if it relaxes the candidate range, may compromise the consistency of educational hard constraints.

[0005] Therefore, there is a need for a dynamic multi-path retrieval method and system for semantic profiling of primary and secondary school test questions that can overcome the shortcomings of the existing technologies. Summary of the Invention

[0006] One objective of this invention is to propose a dynamic multi-path retrieval method and system for semantic profiling of primary and secondary school test questions. Addressing the problems in existing technologies such as the difficulty in uniformly representing multimodal test question queries, the difficulty in verifying retrieval intent and educational attributes, the high proportion of unqualified candidates due to fixed multi-path retrieval, duplicate aggregation, and excessive tail latency and computational overhead, this invention proposes a technical solution consisting of a unified query object, a semantic profiling of test question queries with source and confidence levels, an educational hard constraint candidate bitmap, a compatible channel execution plan, and dynamic batch quota control. This invention enables different indexes to be recalled in batches within the same educational constraint range, and adjusts quotas based on the batch's qualified output and estimated time consumption, completing question family aggregation, fine-tuning, and constraint verification within the cutoff boundary.

[0007] This invention provides a dynamic multi-path retrieval method for semantic profiling of primary and secondary school test questions, including:

[0008] S1. Obtain multimodal test question queries, perform normalization and structural parsing on them to obtain a unified query object;

[0009] S2. Input the unified query object into the multi-task semantic model to generate retrieval intent and educational attribute candidates with source fragments, confidence levels and constraint tags, thus forming a semantic profile of test question query;

[0010] S3. Verify the attribute relationships and confidence levels in the semantic profile based on the educational constraint relationship data. Compile the educational attribute candidates that pass the verification and are marked as hard constraints into a candidate bitmap of a unified question identifier space, so that the attributes that fail the verification do not participate in the compilation, and obtain the verified semantic profile.

[0011] S4. Based on the intent probability, input modality, and completeness of the verified semantic profile, select a compatible channel from the retrieval channels corresponding to the heterogeneous test question index, and generate a multi-path execution plan with initial quota and common deadline.

[0012] S5. According to the multi-path execution plan, each compatible channel returns candidate questions with channel scores, hit evidence and question family identifiers in batches within the candidate bitmap, forming a batch candidate set;

[0013] S6. Review hard constraints batch by batch, and determine the marginal utility and completion probability of each compatible channel based on qualified output and estimated time consumption. Accordingly, add, reclaim, or migrate quotas to compatible channels that maintain the same or more stringent candidate bitmaps until the common deadline. Based on the evidence fusion according to the intent probability, perform topic family aggregation, fine ranking, and constraint review on the retained candidates, and output the search results.

[0014] Optionally, S1 includes: character encoding unification, symbol normalization, paragraph segmentation, and noise character filtering for natural language queries and question stem fragments;

[0015] The test question image is subjected to layout area detection, and a region sequence is generated according to the question stem area, option area, illustration area and formula area. Optical character recognition is performed on the question stem area and option area.

[0016] The identification results of the formula object or the formula region are parsed for operators, operands and hierarchical relationships to generate a formula structure tree.

[0017] The standardized text, region sequence, recognized text, formula structure tree, and the position of each content in the original input are associated and written into the unified query object.

[0018] Optionally, S2 includes: the multi-task semantic model outputting the probabilities of original question retrieval intent, similar question retrieval intent, and teaching demand intent, as well as candidate educational attributes such as subject, grade level, grade, textbook version, knowledge point, question type, and difficulty.

[0019] The text range, page area, or formula structure node in the unified query object that generates each candidate is used as the source fragment;

[0020] Confidence scores are generated based on the normalized scores output by the corresponding task.

[0021] When educational attribute candidates are obtained by user-specified fields, unique mappings from textbook catalogs, or unique mappings from grade level, a hard constraint flag is set to enable the corresponding attribute to participate in the intersection operation of the candidate bitmap. Otherwise, a non-hard constraint flag is set so that the corresponding attribute is only used as a fine-ranking feature in step S6.

[0022] Optionally, S3 includes: verifying the subordinate relationship between subjects and knowledge points in the education ontology, verifying the inclusion relationship between school stages and grades in the school stage and grade mapping, and verifying the association between textbook version, volume and knowledge points in the textbook catalog version data;

[0023] The educational attribute candidates with hard constraint tags and that have passed the verification are converted into their respective corresponding question identifier bitmaps, and the intersection operation is performed on the question identifier bitmaps to obtain the candidate bitmaps;

[0024] Educational attribute candidates whose confidence level does not reach the corresponding attribute confidence threshold, educational attribute candidates that fail the association verification, and their source fragments are written into the conflict table, and their hard constraint labels are converted into non-hard constraint labels, so that the corresponding attributes are only used as fine ranking features in step S6.

[0025] The attribute confidence threshold is determined by the lower bound of the confidence level when the false acceptance rate of the corresponding educational attribute on the validation set does not exceed the preset upper limit of the false acceptance rate.

[0026] Optionally, S4 includes: the heterogeneous test question index includes a full-text index, a question fingerprint index, a learning-type sparse index, a dense vector index, an educational tag index, an optical character recognition layout index, and a formula structure index;

[0027] The necessary profile attribute set is pre-associated for each retrieval channel. The retrieval channels associated with retrieval intents whose intent probability is not less than the corresponding intent selection threshold are taken as the first channel set. The retrieval channels that support the input modality are taken as the second channel set. The retrieval channels that have been verified to have attributes that include the corresponding necessary profile attribute set are taken as the third channel set. The compatible retrieval channel is obtained from the intersection of the first channel set, the second channel set, and the third channel set.

[0028] When the intersection is empty, the baseline retrieval channel that supports the input modality and preserves the candidate bitmap is determined as the compatible retrieval channel;

[0029] The intent selection threshold is the upper bound of the probability threshold allowed when the recall rate of the corresponding retrieval intent on the validation set is not lower than the preset recall rate lower limit.

[0030] The initial quota and common deadline are determined based on the historical qualified question family output rate and historical time quantile of each compatible retrieval channel, and a compatible downgrade channel identifier that maintains the same or more stringent candidate bitmap is written for each compatible retrieval channel.

[0031] Optionally, S5 includes: sending the candidate bitmap as a filtering condition to the corresponding compatible retrieval channel along with the channel query;

[0032] Each compatible retrieval channel divides the recall batches according to the initial quota, and returns the candidate question identifier, normalized channel score, corresponding text hit interval, image layout area or formula structure node and question family identifier for each recall batch.

[0033] Before writing the batch candidate set, perform a bitwise AND check between the returned candidate topic identifier and the candidate bitmap, and discard candidate topic identifiers with a check result of zero.

[0034] Optionally, S6 includes: counting the number of question families, knowledge point coverage increment, and cross-channel duplication number newly added between adjacent recall batches that have passed hard constraint review; and obtaining the batch utility increment by subtracting the penalty term for the cross-channel duplication number from the weighted sum of the number of newly added qualified question families and knowledge point coverage increment.

[0035] Based on the historical time quantile of the corresponding compatible retrieval channel, the real-time queue length, the current batch time, and the remaining time of the common deadline, the probability of the next recall batch being completed before the common deadline is estimated, and the ratio of the batch utility increment to the estimated time of the next recall batch is used as the marginal utility.

[0036] For compatible search channels with marginal utility not less than the utility threshold and completion probability not less than the probability threshold, add a recall batch quota. For compatible search channels that have not added qualified question families in two consecutive recall batches or whose completion probability is less than the probability threshold, reclaim unexecuted quotas. According to the compatible downgrade channel identifier, migrate the reclaimed unexecuted quotas to compatible search channels with marginal utility not less than the utility threshold and using the same or more stringent candidate bitmaps.

[0037] The utility threshold and probability threshold are determined by the lower bound of marginal utility and the lower bound of completion probability when the coverage rate of the qualified question family per unit time on the validation set reaches the preset lower limit of coverage, respectively.

[0038] Furthermore, the channel scores and hit evidence of each compatible retrieval channel are weighted and fused according to the intent probabilities corresponding to the original question retrieval intent, similar question retrieval intent, and teaching demand intent, respectively.

[0039] Candidate questions within the same question family are aggregated according to the question family identifier, and candidate questions whose fusion scores meet the fine ranking quota are retained from each question family as a limited candidate set;

[0040] The finite candidate set is sorted and hard-constrained. Candidate topics that pass the hard-constraint review are marked as education-consistent. The degradation flag in the multi-way execution plan is associated with the corresponding candidate topic and then output.

[0041] Furthermore, when the common deadline is reached, a cancellation instruction is sent to the recall batches that have not yet been completed, while the candidate questions returned by the completed recall batches are retained, and the unexecuted quota is recorded.

[0042] When an anomaly occurs in the retrieval channel before the common deadline, the unexecuted quota will be migrated to the compatible downgrade channel only if the expected batch time of the compatible downgrade channel is not greater than the remaining time of the common deadline and the candidate bitmap after migration is a subset of the candidate bitmap before migration or is the same as the candidate bitmap before migration; otherwise, the corresponding batch will be canceled.

[0043] An execution record is generated, which includes the original retrieval channel identifier, the target retrieval channel identifier, the migration quota, the cancellation batch identifier, the candidate bitmap summary before migration, the candidate bitmap summary after migration, the constraint retention verification result, and the downgrade status.

[0044] On the other hand, the present invention also provides a dynamic multi-path retrieval system for semantic profiling of primary and secondary school test questions, comprising:

[0045] The unified query object generation module is used to execute step S1; the semantic profile generation module is used to execute step S2; the education constraint verification module is used to execute step S3; the multi-way execution plan generation module is used to execute step S4; the bitmap-limited recall module is used to execute step S5; and the dynamic quota control and result output module is used to execute step S6.

[0046] The beneficial effects of this invention are:

[0047] 1. By standardizing and structurally parsing text, images, and formula inputs, and retaining source fragments, confidence levels, and constraint tags in semantic profiling, the retrieval intent and educational attributes have a traceable and verifiable processing foundation.

[0048] 2. By compiling the verified educational hard constraints into a candidate bitmap in a unified question identifier space, and ensuring that each compatible retrieval channel returns candidates only within the scope of this bitmap, it is possible to reduce the number of unqualified candidates from different disciplines or grades from entering the subsequent fusion and ranking.

[0049] 3. By dynamically adding, withdrawing, or migrating quotas based on batch qualified output, estimated time, marginal utility, and completion probability, and requiring downgrade channels to maintain the same or more stringent candidate bitmap, it is possible to improve the coverage of qualified non-repeating question families while controlling common deadlines. Attached Figure Description

[0050] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0051] Figure 1 This is a flowchart of a dynamic multi-path retrieval method for semantic profiling of primary and secondary school test questions according to the present invention.

[0052] Figure 2 This is a flowchart of the dynamic quota control and result output of the present invention, S6. Detailed Implementation

[0053] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0054] refer to Figures 1-2 A dynamic multi-path retrieval method for semantic profiling of primary and secondary school test questions, comprising:

[0055] S1. Obtain multimodal test question queries, perform normalization and structural parsing on them to obtain a unified query object;

[0056] S2. Input the unified query object into the multi-task semantic model to generate retrieval intent and educational attribute candidates with source fragments, confidence levels and constraint tags, thus forming a semantic profile of test question query;

[0057] S3. Verify the attribute relationships and confidence levels in the semantic profile based on the educational constraint relationship data. Compile the educational attribute candidates that pass the verification and are marked as hard constraints into a candidate bitmap of a unified question identifier space, so that the attributes that fail the verification do not participate in the compilation, and obtain the verified semantic profile.

[0058] S4. Based on the intent probability, input modality, and completeness of the verified semantic profile, select a compatible channel from the retrieval channels corresponding to the heterogeneous test question index, and generate a multi-path execution plan with initial quota and common deadline.

[0059] S5. According to the multi-path execution plan, each compatible channel returns candidate questions with channel scores, hit evidence and question family identifiers in batches within the candidate bitmap, forming a batch candidate set;

[0060] S6. Review hard constraints batch by batch, and determine the marginal utility and completion probability of each compatible channel based on qualified output and estimated time consumption. Accordingly, add, reclaim, or migrate quotas to compatible channels that maintain the same or more stringent candidate bitmaps until the common deadline. Based on the evidence fusion according to the intent probability, perform topic family aggregation, fine ranking, and constraint review on the retained candidates, and output the search results.

[0061] In this specific embodiment, S1 includes:

[0062] In this specific implementation, the unified query object generation module of the test question retrieval server performs multimodal test question query access. The multimodal test question query comes from the teacher's search box, question stem fragments in the question bank editor, uploaded test question images, or editable formula objects. The module generates a globally unique query identifier and a millisecond-level receiving timestamp for each query, and writes the input modality, original byte summary, and user-specified subject, grade level, grade, and textbook version fields into the query header record. When the same query contains text, images, and formulas, a position association key is established according to the character offset, image page number, and area coordinates in the original input, so that the subsequently formed text intervals, page areas, and formula structure nodes can all be traced back to the same original input.

[0063] For natural language queries and question fragments, the module first converts the input to UTF-8 encoding, merges equivalent characters using Unicode normalized combination, maps full-width letters, numbers, parentheses, and operators to corresponding half-width characters, and preserves the category differences between the mathematical minus sign and the hyphen. Then, it performs paragraph segmentation based on newline characters, question number patterns, and consecutive whitespace. The filtering rules only delete zero-width characters, control characters, and decorative symbols that appear more than three times consecutively and do not belong to mathematical expressions. The original character start and end offsets, the normalized character start and end offsets, and the mapping type are written into the character mapping table. Missing text is written as an empty text marker instead of fabricating content.

[0064] For the test question images, the module scales the images proportionally by 1280 pixels along the longest side and records the scaling ratio. It then uses a convolutional feature pyramid detector trained with four types of page labels to output the bounding boxes and detection confidence scores for the question stem region, option region, illustration region, and formula region. The regions are then sorted according to page number, upper boundary, and left boundary to form a region sequence. The detector training samples come from test question pages within the platform that have been reviewed by teaching and research personnel. The labels include the region category and rectangle coordinates. The model version and confidence threshold are stored in the page model configuration record. In this specific implementation, the detection confidence threshold is set to 0.60. Regions below this threshold are retained as regions to be parsed but are not directly assigned the four types of labels.

[0065] The module employs an optical character recognition path that connects a convolutional feature extractor and a bidirectional sequence encoder to a CTC decoder for the question stem and option regions. The input is a grayscale line image with a height normalized to 48 pixels, and the output includes the recognized text, character probability, character box, and region identifier. Adjacent lines are merged according to the vertical overlap rate and left boundary difference. Characters with a character probability lower than 0.50 are placed as unknown characters and their character boxes are retained. The recognized text is written into the region record after undergoing the same character normalization rules as natural language queries. The illustration region only saves the image feature extraction entry and coordinates, and does not mistakenly write the image content as text.

[0066] For regions where the detection label is a formula region and the detection confidence is not lower than 0.60, the unified query object generation module calls the formula visual recognizer to scale the region grayscale image by a height of 64 pixels and retain the affine mapping from the original image to the scaled image. The formula visual recognizer consists of a convolutional visual encoder and an autoregressive structure decoder. It outputs numbers, variables, operators, parentheses, fraction lines, square roots, and subscript / superscript relationship sequences in the order from left to right and from baseline to subscript / superscript. It also outputs the confidence and scaled image coordinates for each symbol and structural relationship. The module writes the region identifier, recognition sequence, node-by-node confidence, original image coordinates, formula visual recognizer version, and anomaly marker into a versioned formula recognition record. After restoring the original image coordinates through inverse affine transformation, this record is used as the input for the subsequent lexical scanner. When the confidence of a symbol or structural relationship is lower than 0.55, it is retained as an unknown node and marked as having an incomplete formula structure. This node is not allowed to participate in the hard formula equivalence judgment.

[0067] For editable formula objects or formula region recognition results, the module uses a lexical scanner to recognize numbers, variables, operators, parentheses, fraction lines, square roots, and subscript / superscript relationships. Then, it constructs an ordered formula structure tree according to the rules of parentheses priority, exponentiation priority, and multiplication and division priority over addition and subtraction. Each tree node at least stores the node identifier, node type, normalized symbol, child node number, parent node identifier, original character range or formula region coordinates, and recognition confidence. When parentheses are not closed or operands are missing, the corresponding subtree is marked as incomplete and the original recognition sequence is retained to avoid incomplete structures being treated as hard equivalence conditions.

[0068] Region sequences are sorted by executable sort key Stable arrangement, in which Indicates the page number of the image. Represents the normalized coordinates of the upper boundary of the region. Represents the normalized coordinates of the left boundary of the region. The system represents the region categories that are sequentially coded for the question stem, options, illustrations, and formulas. When the difference between the vertical coordinates of two regions does not exceed 0.5% of the image height, they are sorted by their left boundaries. When the area of ​​overlap between two regions is divided by the minimum area of ​​the two regions and is not less than 0.80 and the regions are of the same category, the region with the highest detection confidence value is retained. When the regions are of different categories, they are all retained and their overlap relationship is recorded in the position association array.

[0069] The unified query object is carried by versioned JSON records. The records contain query headers, normalized text arrays, region sequences, recognized text arrays, formula tree structure arrays, user-specified fields, location association arrays, and anomaly flag arrays. The location association array consists of a source identifier, source type, original location, normalized location, and target object identifier. The text representation, region representation, and formula tree representation of the same content share the source identifier. Duplicate uploaded images are deduplicated based on the original byte digest but retain the input location each time.

[0070] The character mapping table performs segment-by-segment integrity checks: each normalized character must correspond to a character range or a clearly merged character range in the original input, and adjacent normalized ranges must not be in reverse order; the region record performs boundary checks, requiring that both the horizontal and vertical coordinates fall within the image range; the formula tree structure performs acyclic checks, requiring that each node except the root node has only one parent node, and if any check fails, the corresponding object is marked as not being able to participate in the generation of hard constraints in the model, but the original load reference and the reason for failure are still retained;

[0071] Before writing, the unified query object generates an idempotent key based on the query identifier, object mode version, and original byte digest. If the idempotent key already exists and the content digest is consistent, the existing object is reused. If the idempotent key is the same but the digest is inconsistent, a new revision number is generated. The module also counts the number of available text characters, the number of valid regions, and the number of complete formula subtrees and writes them into the completeness original field. These counts only reflect the actual parsing results and are used by step S2 to generate modal missing markers. They are not used to delete user-explicit fields.

[0072] When an image cannot be decoded, the layout detection has no valid area, or the formula parsing is incomplete, the module sets a flag indicating that the image is unavailable, the area is pending parsing, or the formula structure is incomplete, respectively. It continues to retain the text and user-defined fields that can be successfully obtained. The query is terminated and an invalid input status is returned only when all input payloads are empty. After processing, the module writes the unified query object into the query object cache. The cache key is a globally unique query identifier, and the value stores both the object mode version and the model version. The cache key is then passed to the multi-task semantic model call interface in step S2.

[0073] In this specific embodiment, S2 includes:

[0074] The semantic profiling generation module takes the unified query object written in step S1 as input, concatenates the normalized text, question stem, and option recognition text into a text sequence according to their positional association, encodes the operator path of the formula structure tree into a structure tag sequence, and uses region category, normalized bounding box coordinates, and illustration visual vector as additional modal features; the text sequence is truncated or padded to 512 sub-word tags, the formula structure tags retain a maximum of 128 nodes, the illustration visual vector is generated by a visual encoder trained on the question image into a 256-dimensional vector, and the missing modality uses independent missing tags instead of using zero vectors to impersonate valid content;

[0075] The multi-task semantic model employs a shared Transformer encoder connecting a retrieval intent classification head and seven sets of educational attribute classification heads. The retrieval intent classification head outputs three probabilities: original question retrieval intent, similar question retrieval intent, and teaching demand intent. The educational attribute classification heads output candidate scores for subject, grade level, grade, textbook version, knowledge point, question type, and difficulty, respectively. The model training data consists of historical retrieval logs from the question bank and annotations reviewed by teaching and research personnel. Original question pairs, heterogeneous question pairs with the same knowledge point, and teaching demand queries are used to construct three types of intent samples. Educational attribute tags are taken from the metadata of the self-reviewed question bank. The training objective is the sum of the cross-entropy of each task, and the model version is selected based on the macro average recall rate of the validation set.

[0076] The shared encoder simultaneously outputs query feature carriers for use by the retrieval adapter: the pooling results of the last layer are normalized to form a dense query vector by unit length; the terms with weights not less than the threshold released with the model version in the output of the sparse projection layer are retained in descending order of weight to form a sparse term weight array, and the vector dimension, vocabulary identifier, shared encoder version and sparse projection layer version are recorded respectively; the module also generates text fingerprints based on the normalized text character triples in step S1, and generates formula structure fingerprints based on the operator path and subtree summary of the formula structure tree. The fingerprint record saves the fingerprint algorithm version, unified query object mode version and input summary; the above vectors, sparse weights and two types of fingerprints are written into the query feature side table with the query identifier as the key, and their immutable references and content summaries are written into the test question query semantic profile. Step S5 can only read the reference. When the summary or version is inconsistent, it is deterministically recalculated according to the same shared encoder version, vocabulary version, fingerprint algorithm version and unified query object mode version and the recalculated summary is written into the channel query.

[0077] The module performs temperature scaling on the raw scores of each task header to obtain the calibration confidence level, calculated as follows: , Indicates candidate The calibration confidence level is a dimensionless number between 0 and 1. This represents the category index in the current task header candidate category set. Its value range is a stable integer index from the task header label dictionary, which is derived from the label dictionary released with the model version. Because it is only used for discrete localization, it has no engineering units. This indicates that the multi-task semantic model supports candidate... The dimensionless raw score is output. This represents the dimensionless temperature parameter that is greater than zero, obtained by fitting the corresponding task header onto the validation set. This indicates the dimensionless bias obtained by fitting the corresponding task head on the validation set. The temperature parameters and biases are stored in the model calibration record with the task head identifier as the key and are refitted when the validation set distribution drifts or the model version is updated.

[0078] The three calibration confidence scores of the intent classification head are retrieved and then normalized to the intent probability by summing the three scores. The sum of the three scores is 1. If the sum of the three scores is less than the preset value protection value due to an abnormal value, the query is marked as a model abnormality and handed over to step S4 to enable only the baseline channel that supports the current input modality and maintains the candidate bitmap. The education attribute classification head retains no more than three candidates in descending order of calibration confidence score. If the confidence scores are the same, they are sorted in ascending order according to the stability identifier in the education attribute dictionary. The candidate record saves the attribute type, attribute value identifier, calibration confidence score and task head version.

[0079] For each intent or educational attribute candidate, the module uses the text tag interval with the highest and continuous attention contribution, the page area with the largest overlap with the visual feature response, or the formula structure node that participates in the formula structure attention aggregation as the source fragment; the source fragment record includes the source type, the target object identifier in the unified query object, the start and end character offset or region coordinates, the formula node identifier and the contribution score. When multiple sources jointly generate a candidate, up to five source fragments are saved in descending order of contribution score, so that the conflict record in step S3 and the hit evidence fusion in step S6 can locate the original input;

[0080] The semantic profiling generation module generates source candidates in addition to model candidates: It performs dictionary validation on each explicit user field in the unified query object; when validation passes, candidates are generated based on field type and value, with the source primary key being the record identifier from the query header field, and a quality score set to 1.00. It queries the current textbook catalog version using the textbook version and catalog node combination key; only when the query result is exactly one record is a corresponding knowledge point candidate generated. It uses the same uniqueness judgment for the range from grade level to grade level and the mapping from grade level to grade level, with the source primary key being the catalog record or mapping record identifier, and the quality score being the mapping accuracy rate obtained from the teaching and research sampling in the corresponding published record. If no accuracy rate is configured, it is set to pending validation without hard constraints. The source candidates and model candidates are deduplicated based on attribute type and value. For identical values, the source is merged and the maximum value between the quality score and the model calibration confidence score is taken as the candidate confidence score. Different values ​​of the same attribute do not overlap. The conflict priority is recorded in the order of user explicit, unique directory mapping, unique grade level mapping, and model inference, and all are marked as pending relationship verification in step S3. When the dictionary verification of the user explicit value fails or the unique mapping query returns zero or more records, no hard constraint candidate is generated, but the original value and query key are written into the conflict record. The above candidate generation record saves the generation rule version, data version, input key, output value, source primary key, quality score, candidate confidence score and conflict priority, and is passed to step S3 along with the semantic profile, so that the source candidate and the model candidate use the same confidence score field and threshold verification entry.

[0081] The module sets constraint tags based on the candidate source: hard constraint tags are set when the user-explicit field has a non-null value and passes the field dictionary validation; hard constraint tags are set for knowledge points in the textbook catalog that are uniquely mapped from the textbook version and the catalog node; hard constraint tags are set when the grade range is uniquely determined by the grade in the grade-level mapping or the grade is uniquely determined by the grade; and non-hard constraint tags are set for other candidates obtained by model inference, non-unique catalog mapping, or multi-value mapping. Hard constraint records also store the source primary key of the explicit field, catalog record, or mapping record, while non-hard constraint records store the feature name used only for fine-tuning.

[0082] After the model input is constructed, consistency gating is performed, requiring that the text mark position can be mapped to the character range, the visual region identifier can be mapped to the region sequence, and the formula node identifier can be mapped to the formula structure tree. When any source reference becomes invalid, only the source feature is removed and the source is marked as unavailable. Candidates can still be generated from other valid sources, but educational attribute candidates without any valid sources are not written into the test question query semantic profile, thereby ensuring that each retained candidate has at least one traceable source fragment.

[0083] The model version record includes the shared encoder version, each task head version, vocabulary summary, training sample time range, and calibration version. Online monitoring will trigger the training of a new candidate version when the average macro recall rate of any task drops by more than 3 percentage points for seven consecutive days. A new version will only be released when the recall rate of each retrieval intent on the fixed validation set is not lower than the current version and the attribute error acceptance rate does not exceed the corresponding upper limit. If it fails, the current version will continue to be used and the failure index will be written into the model release record.

[0084] The semantic profile for test question queries uses the query identifier as the primary key and includes the input modality set, completeness field, probabilities of three search intents, candidate array of educational attributes, array of source fragments, confidence score, constraint markers, model version, and calibration version. The completeness field generates bit markers based on the actual usability of four types of payloads: text, layout, formula, and user-defined fields. Missing modalities are not counted as failures, and when there are multiple value conflicts of the same type among attribute candidates, they are all retained and set to a pending verification status.

[0085] After the profile is constructed, the module writes the semantic profile of the question query into the semantic profile storage area, and writes the hard-constrained candidate index and the non-hard-constrained fine-ranking feature index into the profile side table respectively. Candidates with a calibration confidence of less than 0.20 and no explicit user source are not included in the subsequent attribute relationship verification, but are retained in the model diagnostic record. Other candidates, along with the source fragment, confidence and constraint label, are passed to step S3.

[0086] In this specific embodiment, S3 includes:

[0087] The education constraint verification module reads the semantic profile of the test question query formed in step S2, as well as the education ontology, grade level mapping, and textbook catalog version data of the same release version. All three types of education constraint relationship data use stable identifiers as connection keys. The education ontology edge records contain subject identifiers, knowledge point identifiers, subordinate relationship types, and effective versions. The grade level mapping records contain grade level identifiers, grade identifiers, inclusion status, and applicable regions. The textbook catalog records contain textbook version identifiers, volume identifiers, catalog node identifiers, knowledge point identifiers, and validity periods. During release, foreign key integrity checks are performed and a version summary of the relationship data is generated.

[0088] The module first reads candidates from the profile according to attribute type, performs educational ontology edge existence check on subject and knowledge point, performs inclusion relationship check on grade and grade level, and performs directory association chain check on textbook version, volume and knowledge point. When the same candidate involves multiple relationships, the relationship check is only considered passed if all required relationships exist and are in an active state. If the candidate key is not matched, the relationship version is inconsistent, or any required edge is disabled, it is considered to fail, and the missing connection key and relationship version are written into the check reason field.

[0089] The attribute confidence threshold is calibrated by the independent validation set according to the attribute type. The calibration program scans the calibration confidence of the attribute from high to low, divides the number of incorrectly accepted candidates by the total number of accepted candidates to obtain the false acceptance rate, and selects the lowest confidence that can be obtained when the false acceptance rate does not exceed 5% as the confidence threshold for the attribute. The threshold record includes the attribute type, lower bound of confidence, upper bound of false acceptance rate, number of samples, model version, calibration version, and effective time. The threshold is recalculated when the validation set is increased by no less than 20% of the existing sample size or when the model version is changed. For attributes with insufficient samples, a more stringent global threshold than that in the task header is used.

[0090] For candidates marked with hard constraints, the module also requires that the relationship verification pass and the calibration confidence level is not lower than the corresponding attribute confidence level threshold. When the conditions are met, the module searches for the question identifier bitmap directory according to the attribute type and attribute value identifier. The bitmap directory records and saves the attribute key, the unified question identifier space version, the compressed bitmap file identifier, the number of question identifiers and the content summary. Candidates that do not hit the directory or have different bitmap space versions are not included in the compilation and enter the conflict handling process to avoid performing intersection on incomparable bitmaps.

[0091] Candidate bitmap passed generate, This represents the candidate bitmap after all valid hard constraints have been applied. It originates from the output of this relation validation and the bitmap directory of the same version. The number of bits is equal to the space capacity of the unified question identifier. Because it is represented by a binary set, it has no engineering units. This represents the set of candidate indexes for the education attribute that have passed relation validation, reached the confidence threshold, and are marked with hard constraints. Members are derived from the records validated in this instance. Because it is a finite set of indexes, it does not have engineering units. This represents the candidate index in the set of valid hard constraint candidates, with a value range of 1. Because it is used for discrete positioning and therefore has no engineering units, Indicates candidate The question identifier bitmap found in the bitmap catalog that has the same question identifier space version as the verified semantic profile is represented by a binary set and therefore has no engineering units. The module performs the intersection according to the bitmap cardinality from smallest to largest to reduce intermediate results. If the set of valid hard constraints is empty, the full set of valid question identifier bitmaps corresponding to the current question bank version will be used as candidate bitmaps without introducing unverified attribute restrictions.

[0092] After each bitmap intersection, the participating attributes, cardinality before and after the intersection, and bitmap summary are recorded. When the intersection result is empty, the hard constraints that have passed the verification are not revoked. Instead, the empty bitmap state is written into the verified semantic profile, so that step S4 can only select the baseline channel that can maintain the empty candidate bitmap and quickly return the empty result. The unified question identifier space adopts continuous integer question identifiers. When a question is added, removed, or its attributes are changed, the question bank publishes a task to atomically update each attribute bitmap and generate a new space version. The query only reads the same version snapshot.

[0093] The task of publishing relational data uses the audited question bank metadata and textbook catalog import records as its source. First, it checks that the subject identifier, grade level identifier, textbook version identifier, volume identifier, and knowledge point identifier all exist. Then, it deletes invalid edges and adds audited relational edges. Each publication uses an incrementing version number and effective time. The three types of relational data are switched together only when the reference integrity rate is 100% and there are no mutually exclusive directory node conflicts for the same textbook version and volume. If the verification fails, the previous version is retained and a publication failure record is generated.

[0094] The attribute bitmap directory is built along with the question bank release task. For each attribute key, scan for question identifiers that are valid and have that attribute, set the corresponding bit to 1, and keep the other bits 0. Then, write it into the compressed bitmap using the attribute key and the unified question identifier space version as the joint key. After construction, randomly select no less than 1,000 question identifiers and check them item by item with the question bank attribute records. If any inconsistency is found, the bitmap version is prohibited from taking effect and the previous valid version continues to be used.

[0095] Educational attribute candidates whose confidence level does not reach the corresponding attribute confidence threshold, fail relationship verification, or cannot obtain the same version bitmap, along with their source fragments, are written into the conflict table. The conflict table fields include query identifier, candidate attribute type, candidate value, source fragment, original constraint tag, confidence level, threshold, failed relationship key, relationship data version, and processing status. The module converts the hard constraint tags of these candidates into non-hard constraint tags, retains their confidence level and source fragment, and registers them in the profile side table as fine-ranking features readable in step S6. Attributes that fail verification are not included in the effective hard constraint set.

[0096] The module writes the verified attribute relationships, candidate bitmaps and their unified title identifier space versions, conflict table references, non-hard constraint fine-sorting features and relation data versions into the verified semantic profile. Before writing, it verifies that the candidate bitmap summary is consistent with the bitmap directory summary. If it fails, it stops the publication of the version profile and returns the version conflict status. If it succeeds, it passes the verified semantic profile identifier to the multi-way execution plan generation module in step S4.

[0097] In this specific embodiment, S4 includes:

[0098] The multi-path execution plan generation module reads the verified semantic profile output from step S3 and retrieves the retrieval channel records corresponding to the full-text index, question fingerprint index, learned sparse index, dense vector index, educational tag index, optical character recognition layout index, and formula structure index from the channel registry. Each record contains a channel identifier, index version, associated intent set, supported input modality set, required profile attribute set, baseline channel tag, batch size, maximum concurrency, compatible downgraded channel identifier, and the question identifier space version used. Only the index version is available, and the question identifier space version matches the candidate position. Figure 1 The desired channel enters the candidate channel pool;

[0099] The module reads the intent probability and intent selection threshold item by item according to the search intent. Channels with intent probabilities not less than the corresponding threshold and whose channel-related intent sets contain the intent are added to the first channel set. The intent selection threshold is scanned on the validation set from high to low probability. The recall rate is calculated by dividing the number of recalled questions related to the intent by the total number of related questions. The maximum probability threshold allowed when the recall rate is not less than 95% is selected. The threshold record is recorded with intent type and model version as keys, including the lower limit of recall rate, number of samples and effective version. It is recalibrated when the model or question bank distribution is updated.

[0100] The second channel set consists of channels that support the input modality set and include the currently available modalities. Pure text queries do not require image or formula modalities, while queries with formula structure trees are allowed to use the formula structure index as the formula channel. Optical character recognition page indexes are allowed to be used as page channels only when there is a valid page area. The third channel set consists of channels where the required portrait attribute set is empty, or where the set is completely included by the already possessed attribute set of the verified semantic portrait. Attributes in non-hard constraint candidates and conflict tables are not included in the already possessed attribute set.

[0101] The construction rule for the first channel set is to iterate through the intents whose probabilities reach the corresponding threshold one by one, add the channels associated with the intent and whose status is available in the channel registry to the set and remove duplicates; when no intent reaches the threshold, the intent with the highest intent probability value is used as the only active intent, and when the probabilities are tied, the intents are selected in a fixed order of original question search intent, similar question search intent, and teaching need intent, so that the first channel set has a definite output.

[0102] The construction rule for the second channel set is to check whether the set of input modalities supported by each channel covers all the required modalities of this query. When normalized text exists, the text modal is required; when a complete formula structure tree exists, the formula modal is required; when a valid layout area exists and text is unavailable, the image layout modal is required. Incomplete formula structure trees and regions to be parsed are only registered as additional modalities and are not excluded from the list of channels due to the absence of the corresponding channel.

[0103] The construction rule for the third channel set is to perform an inclusion judgment on each attribute key in the channel's required profile attribute set. If all attribute keys exist in the verified semantic profile's existing attribute set, the channel is retained. If the required profile attribute set is empty, it is directly retained. If any required attribute only appears in the conflict table or non-hard constraint features, the channel is excluded and the missing attribute key is recorded in the planned audit field.

[0104] Compatible search channels by Sure, This represents a set of compatible search channels composed of channel identifiers, sourced from the results of this three-category screening. Because it is a discrete set of identifiers, it lacks engineering units. This represents the first set of channel identifiers selected by the channel registry based on intent probability and intent selection threshold. Because it is a discrete set of identifiers, it has no engineering units. This represents the second set of channel identifiers filtered by the channel registry according to the actual available input modalities. Because it is a discrete set of identifiers, it has no engineering units. This represents the third channel identifier set obtained by filtering according to the condition that the verified semantic profile has the attribute coverage of the required profile attributes of the channel. Since it is a discrete identifier set, it has no engineering unit. All four sets use the current channel registry version. The channels in the intersection are arranged in ascending order of channel identifier to form a stable execution sequence. If the intersection is empty, the module selects the historical 95th percentile time value from the baseline channels that support the current input modality, use the same question identifier space version, and can receive candidate bitmap filtering conditions. If the time is the same, the channel with the smallest channel identifier value is selected, and a baseline degradation flag is set in the multi-way execution plan. If the candidate bitmap is empty, an empty result plan without dispatching recall tasks is generated.

[0105] The statistics table uses channel identifier, intent type, input modality combination, and candidate bitmap cardinality as joint keys to store the following data for the past 30 days: qualified question family output rate, knowledge point coverage increment rate, cross-channel duplication rate, 50th and 95th percentiles of batch time, standard deviation of time residual, anomaly rate, number of statistical samples, and version. The qualified question family output rate is equal to the number of newly added question families that have passed hard constraint review divided by the number of valid candidates. The knowledge point coverage increment rate is equal to the number of newly added knowledge points relative to the previous qualified set divided by the number of valid candidates. The cross-channel duplication rate is equal to the number of question identifiers that are duplicated with candidates already retained in other channels divided by the number of valid candidates. All three are dimensionless statistics ranging from 0 to 1. The time field is in milliseconds. When the number of samples is less than 200, a coarser-grained record that does not distinguish bitmap cardinality in the same channel is read back. The statistics task is updated daily and the previous valid version is retained for anomaly rollback.

[0106] The module aims for a limited candidate size of 120 in this specific implementation, with the initial batch number of channels calculated as follows: Sure, Indicates channel The initial planned number of batches, in units of batches, ranging from 1 to... Positive integers between Indicates channel The upper limit of the initial planned batch number in a single query, in batches, is derived from the channel registry and is set to 2 batches in this specific implementation. This indicates a finite number of candidate targets, expressed in units of items. Indicates channel The historical qualified question family output rate for each batch in the current version of the runtime statistics table, based on the same intent type, input modality combination, and candidate bitmap cardinality buckets, is expressed in units of each family per batch and is not less than zero. Indicates channel The batch size in the channel registry is a positive value, expressed as a single entry per batch. Represents a set of compatible search channels The unassigned channel index in the database, because it is used for discrete positioning, has no engineering units. This refers to the aforementioned set of compatible retrieval channel identifiers, which, being a discrete set of identifiers, lacks engineering units. This indicates that the set is traversed when summing. The channel index, because it is used for discrete positioning, has no engineering units. Indicates channel The historical output rate of qualified question families per batch within the same running statistical version and the same bucket, in units of families per batch and not less than zero; the initial quota is equal to the initial batch number multiplied by the batch size;

[0107] When the sum of the historical qualified question families output rates of all channels is zero, the module allocates the number of batches corresponding to the target size according to the average of the channels. The remaining amount is then allocated sequentially to the channels with the smaller historical 95th percentile time consumption and the lower anomaly rate. After the initial allocation, a quota constraint check is performed. If the total quota is lower than the target size, it is replenished batch by batch in the same order, but the number of batches after replenishment for any channel cannot exceed the number of batches for that channel. All channels have reached their respective destinations. If the total post-quota remains below the target size, replenishment ceases, and the initial plan with limited output capacity is recorded, along with the difference between the target size and the allocable size. If the post-quota exceeds the target size by more than one maximum batch, batches are gradually recovered from the channel with the lowest expected output, but no compatible channel may have fewer than one initial batch. After replenishment or recovery is completed, the final planned batch number for each channel is written back. Subsequent quotas, deadlines, and plan status will only be read from this write-back value;

[0108] Available execution time long press Sure, This represents the available execution time, measured in milliseconds, starting from the time the plan was generated. This refers to the aforementioned set of compatible retrieval channels consisting of channel identifiers and without engineering units. Represents a set The channel index in the code lacks engineering units because it is used for discrete positioning. This indicates the channel that has been written back after the aforementioned supplementation or recycling verification. The final planned number of batches, in units of batches, is taken from 1 to... Positive integers between Indicates channel The historical single-batch execution time, expressed in milliseconds, is the 95th percentile value in the current version and current bucket of the runtime statistics table. This indicates the shortest execution time in this specific implementation method, in milliseconds, and is derived from the server-side shortest scheduling window configuration. The user interface response budget in this specific implementation is expressed in milliseconds and is sourced from the service level configuration of the calling end. The common deadline is equal to the planned generation time plus the available execution time. For channels lacking effective time consumption statistics, the conservative 95th percentile value of the same index type is used. The planned record simultaneously writes the statistical version, response budget, and monotonic clock start point for step S6 to calculate the remaining time.

[0109] The module organizes the input modality set, profile completeness bit markers, intent probability vectors, candidate bitmap cardinality, compatible channel set, quotas for each channel, and remaining execution time into a plan status record. It defines adding a batch, reclaiming undistributed batches, migrating unexecuted quotas, and maintaining the current plan as subsequent allowed actions. During the plan generation phase, only the initial quota is issued. In step S6, each time a batch completion event is received, an action is selected based on marginal utility and completion probability. All actions are subject to the common deadline, maximum concurrency of channels, and no relaxation of the candidate bitmap constraints.

[0110] The initial dispatch rule of the execution plan is to first dispatch one batch to each compatible channel in the stable execution sequence, and then register the batches to be dispatched for channels that still have initial quotas according to the stable execution sequence. The maximum concurrency in the channel registry indicates the capacity of the number of batches that the channel can process simultaneously across different queries. The batches of the same query are always serialized because they share the paging cursor. The maximum concurrency is only used to determine the global capacity before the adapter dispatches and does not indicate that two batches will be dispatched simultaneously in this query. If the index health status of a channel becomes unavailable before dispatch, the batches of that channel will not be dispatched. Instead, the unexecuted quota will be handed over to step S6 for processing according to the verified compatible downgraded channel identifier.

[0111] The plan persistence uses a query identifier and plan version as a composite key. If the composite key does not exist, a new plan is written. If the plan already exists and the plan summary is the same, the existing plan is returned. If the plan already exists but the summary is different, the plan version is incremented and the old version is set to undistributable. Step S5 only consumes the plan that is published and has the largest version number to prevent the same query from generating two sets of initial quotas due to duplicate requests.

[0112] Once the common deadline is released, it will not be extended due to the early completion of a single channel. All channels will be compared using the same deadline. If all initial batches are completed before the deadline and no channel meets the subsequent additional conditions, the plan status will be set to recall completed and the evidence fusion in step S6 will be initiated immediately. If there are still channels that meet the additional conditions, the additional batches will be dispatched in step S6 within the remaining time constraints.

[0113] For each compatible channel, the module verifies the target channel's compatibility degradation channel identifier in the channel registry. The target channel must support the current input modality, use the same question identifier space version, and its candidate bitmap must be a subset of the original candidate bitmap or the same as the original candidate bitmap. Only identifiers that pass the verification are written into the plan. If no qualified degradation channel is found, the target identifier is set to empty and migration is prohibited. If an exception occurs, step S6 cancels the corresponding batch without transferring it to the unverified channel.

[0114] The final multi-path execution plan includes query identifier, verified semantic profile identifier, candidate bitmap reference, compatible channel sequence, initial quota and batch size of each channel, common deadline, compatible downgraded channel identifier, statistical version and downgrade identifier; the module performs four checks on the plan execution quota, deadline later than the generation time, channel version consistency and downgraded bitmap not relaxed. After all checks are passed, it is written to the execution plan storage area and passed to step S5.

[0115] In this specific embodiment, S5 includes:

[0116] The bitmap-limited recall module reads the multi-path execution plan generated in step S4 and constructs a channel query for each compatible retrieval channel. The channel query includes query identifier, channel identifier, index version, candidate bitmap file identifier and summary, unified title identifier spatial version, retrieval intent probability, corresponding modal payload reference, query feature side table reference and summary from step S2, shared encoder version, sparse projection layer version, vocabulary version, text fingerprint and formula structure fingerprint algorithm version, unified query object pattern version, batch size, initial quota, and common deadline. The candidate bitmap is sent with the query as an indispensable filtering condition. Before executing the index recall, the channel adapter simultaneously verifies the bitmap summary, spatial version, and query feature summary. Channel queries that fail the verification are not executed and the channel abnormality status is returned to step S6.

[0117] Full-text index and learning-based sparse index retrieve the inverted index using normalized text terms and sparse term weight arrays in the query feature side table of step S2, respectively. Title fingerprint index finds near duplicate buckets using text fingerprints and formula structure fingerprints combined keys in the same side table. Dense vector index performs nearest neighbor retrieval using unit-length dense query vectors in the same side table. Education tag index retrieves by valid education attribute keys. Optical character recognition layout index retrieves by combining region category and recognized text. Formula structure index retrieves by operator path and subtree summary. When the adapter finds missing query features, it can only recalculate according to the fixed step S2 model, vocabulary, fingerprint algorithm, and unified query object mode version of the channel query record, and executes after the recalculated summary is consistent with the planned summary. Each adapter first obtains the index candidate identifier, and then performs filtering with the candidate bitmap. It is prohibited to return candidates outside the bitmap first and then have the upper layer relax the processing.

[0118] Each compatible retrieval channel divides the recall batches according to the initial quota, with the number of batches calculated as follows: Sure, Indicates channel The number of recalled batches, expressed as a positive integer. This refers to the aforementioned set of compatible search channels. The channel index in the code, because it is used for discrete positioning, does not have engineering units. Indicates channel The initial quota is in units of strips. Indicates channel The batch size is specified and the unit is each batch. The number of requests in the last batch is equal to the initial quota minus the number of requests in previous batches. Batch tasks are written to a concurrent queue, and the fields include channel identifier, batch number, cursor, number of requests, dispatch time, deadline time, and status.

[0119] The queue scheduling follows these rules: Within a single channel, the next batch is dispatched only after the previous batch returns to the cursor or is explicitly completed. Different channels execute in parallel within the maximum concurrency range of the registry. If the current time plus the historical 50th percentile of the time taken for that channel exceeds the common deadline, no new batch is dispatched. Undispatched batches that have reached the common deadline are directly converted to unexecuted quotas. Under the same dispatch conditions, the queue is dequeued in ascending order according to the planned stable channel sequence and batch number.

[0120] The raw scores returned by each channel are normalized using the lower and higher quantiles from their respective statistical versions, calculated as follows: , Indicates channel Returned candidates The normalized channel score is a dimensionless number between 0 and 1. This refers to the aforementioned set of compatible search channels. The channel index in the data takes a value range of a set. Furthermore, because it is used for discrete positioning, it lacks engineering units. Indicates channel The returned candidate array contains candidate question identifier indices mapped to the current unified question identifier space. The value range is the intersection of the candidate array and the current unified question identifier space, and because it is used for discrete positioning, it does not have engineering units. This represents the dimensionless raw score output for the corresponding index. Indicates channel The 5th percentile of the dimensionless score distribution for relevant questions in the validation set, sourced from the current version of channel statistics, which is dimensionless. Indicates channel The 95th percentile of the dimensionless score distribution for related questions on the same validation set, sourced from the same statistical version from the same channel and being dimensionless. Indicates channel The strictly positive dimensionless numerical protection quantity is taken as one millionth of the absolute value of the channel score span. The maximum value in, This specific implementation method is intended to ensure... The dimensionless lower limit for strict positive values ​​is derived from the channel score normalization configuration; when The aforementioned division will not be calculated and The value was set to 0.5 to avoid uncertain results caused by a zero span;

[0121] Each recall batch returns candidate topic identifiers, normalized channel scores, hit evidence, and topic family identifiers. Hit evidence simultaneously saves mappings of candidate-side hit segments and query-side contribution segments: the full-text or sparse channel saves candidate text hit intervals, hit terms, and query text tag intervals obtained by aligning common term positions; the layout channel saves candidate page numbers and rectangle coordinates, as well as the matched query region identifiers after normalizing their respective image width and height to 0-1 coordinates; the formula channel saves candidate formula structure nodes and matching subtree summaries, as well as query formula node identifier pairs returned by tree editing matching; the dense vector channel saves vector similarity and the identifier of the query source segment with the largest contribution; and the topic fingerprint channel saves candidate... The text fingerprint hit position, formula fingerprint similar substructure and its corresponding query text character range or query formula node identifier; the education tag channel saves the hit attribute key, candidate attribute value and the source fragment identifier of the same attribute candidate in the verified semantic profile; the mapping record includes the mapping type, query side source identifier, candidate side position or attribute value, overlap weight and generation adapter version. If it is indeed impossible to generate a query side mapping, a mapping missing flag is set and the corresponding evidence consistency degree is set to zero in step S6; the question family identifier comes from the family record generated by the question bank when it is published based on the mother question, adaptation relationship and normalized fingerprint. Candidates with missing family identifiers generate a single-member question family with their own question identifier and set a pending review flag;

[0122] Before writing the batch candidate set, the module maps each returned candidate question identifier to its bit order in a unified question identifier space, and reads the corresponding bits of the candidate bitmap to perform a bitwise AND check, retaining the predicate as... If and only if the question identifier Existence, consistent spatial version, and corresponding bit in the candidate bitmap is 1. Indicates candidate topics The dimensionless binary value retains the label and takes the value 0 or 1. This represents the candidate question identifier index that comes from the candidate array returned by the channel and belongs to the current unified question identifier space. Since it is used for discrete positioning, it has no engineering units. Candidates marked with 0 are retained and written to the audit count along with the reason for rejection. They are not included in the batch candidate set.

[0123] When the same question identifier is returned repeatedly in the same batch of the same channel, the normalized channel score is compared first and the one with the largest value is retained. If the scores are the same, the number of non-empty hit evidence fields is compared and the one with the largest number of fields is retained. If they are still the same, the one with the smallest internal index number is retained. When different channels hit the same question identifier, the channel records are retained for step S6 to fuse according to the intent probability. Evidence is not deleted across channels in step S5.

[0124] The batch candidate set uses the query identifier and the batch global sequence number as the primary keys, and records the channel identifier, channel batch sequence number, dispatch time, completion time, candidate array, number of bitmap verification passed, number of discarded, exception code, and next cursor. A batch is only marked as completed after the candidate array and statistics fields are atomically written. When the channel times out or there is an exception, the completed batches are retained and the writing of incomplete batches is stopped. The unexecuted quota and the exception status are submitted to the dynamic quota controller in step S6.

[0125] Upon completion of each recall batch, the module immediately publishes a batch completion event, which carries the batch candidate set reference, actual time consumed, remaining cursor, executed quota, and unexecuted quota. Step S6 then performs hard constraint review and dynamic quota decision-making for each batch accordingly. After the common deadline is reached, the bitmap-limited recall module only accepts cancellation instructions and batches that have completed atomic writing before the deadline, and no longer appends late results to the batch candidate set.

[0126] In this specific embodiment, S6 includes:

[0127] The dynamic quota control and result output module subscribes to the batch completion event in step S5, reads the batch candidate set, verified semantic profile, candidate bitmap, multi-way execution plan, and conflict table, and re-examines the candidate bitmap bit value and all valid hard constraint attributes for each new candidate. Both batch review and final review read the same question bank release version, the same relational data version, and the same unified question identifier space snapshot bound when the candidate bitmap was generated in step S3. The module verifies the association between textbook version, volume, and knowledge point in the textbook catalog version data. Only when the bitmap bit value is 1 and the attribute record in the frozen snapshot satisfies all valid hard constraints in subject, grade level, grade, textbook version, knowledge point, question type, and difficulty is it counted as a qualified candidate. When the frozen version is unavailable or any version summary is inconsistent, the module outputs the version conflict status and stops subsequent distribution. The latest real-time attribute records are not mixed. Candidates that fail the review are written into the batch elimination record and do not participate in question family production and knowledge point coverage statistics.

[0128] The controller uses two adjacent completed recall batches as statistical windows to calculate the number of newly added and qualified question families, the number of knowledge points not covered in the previous qualified set but added in this batch, and the number of question identifiers that are duplicated with candidates retained by other channels in this batch, and divides them by the number of valid candidates returned in this batch. Indicates channel The trigger batch has been completed. The newly qualified question family rate is a dimensionless number between 0 and 1, derived from the number of newly qualified question families divided by the number of valid candidates in this batch. Indicates channel In batch The knowledge point coverage increment rate is a dimensionless number ranging from 0 to 1, derived from the number of newly added knowledge points divided by the number of valid candidates in this batch. Indicates channel In batch The cross-channel repetition rate is a dimensionless number ranging from 0 to 1, derived from the number of cross-channel repetitive question markers divided by the number of valid candidates in this batch. This indicates compatibility with the retrieval channel index and the absence of engineering units. This indicates that the channel has completed and triggered the recall batch index for this decision and there are no engineering units; when the number of valid candidates returned in this batch is zero, all three items are recorded as zero and the count of consecutive batches without new additions is incremented;

[0129] Batch utility increment by calculate, Indicates channel batch The resulting dimensionless utility increment The positive weight representing the rate of newly added qualified question families. The positive weight representing the incremental rate of knowledge point coverage. The penalty weights for cross-channel repetition rate are represented by three dimensionless numbers ranging from 0 to 1, which sum to 1. In this specific implementation, they are set to 0.50, 0.30, and 0.20, respectively. A higher utility indicates more qualified non-repetition coverage added per unit batch, while a negative value indicates that the repetition penalty exceeds the new benefit.

[0130] The controller estimates the estimated time for the next batch by summing the 50th percentile of historical batch time, the real-time queue length multiplied by that percentile, the current batch time, and the channel anomaly correction amount. It also uses the standard deviation of the residuals of the most recent 200 completed batches from the same channel to describe the estimated fluctuation. The completion probability is calculated as follows: calculate, Indicates channel The trigger batch has been completed. The probability that the next batch to be dispatched will be completed before the common deadline is a dimensionless number between 0 and 1. Indicates channel In batch The estimated time for the next batch is then given, in milliseconds. Indicates the remaining duration of the common deadline, in milliseconds. Indicates channel The time consumption residual standard deviation, in milliseconds, is limited to a minimum of 1 millisecond. This indicates compatibility with the retrieval channel index and the absence of engineering units. This indicates that the batch indexing for this prediction has been completed and triggered, and there are no engineering units.

[0131] Marginal utility according to calculate, Indicates channel Completed batches The marginal utility after that, and the unit is per millisecond. This represents the channel obtained from the aforementioned batch utility increment formula. batch Dimensionless utility increment The channel used in the aforementioned completion probability formula represents In batch The estimated time for the next batch is in milliseconds. This indicates compatibility with the retrieval channel index and the absence of engineering units. This indicates a batch index that has been completed and triggered a decision, but has no project units. Indicates and The lower limit of the numerical protection is of the same dimension, with the unit being milliseconds. In this specific implementation, it is set to 1 millisecond and is derived from the dynamic quota control configuration. The utility threshold is determined by the lower bound of the marginal utility when the coverage rate of the qualified question family per unit time on the validation set reaches the lower limit of 80% coverage and is in milliseconds. The probability threshold is determined by the lower bound of the completion probability when the same validation set reaches the lower limit of coverage and is in the range of 0 to 1. The threshold record saves the validation set version, the number of statistical samples, the lower limit of coverage, the threshold, and the effective time.

[0132] Quota actions are executed according to deterministic predicates: when the marginal utility is not less than the utility threshold and the completion probability is not less than the probability threshold, the controller adds a quota for a recall batch to the corresponding channel and writes it into the next batch of tasks; when a channel has no new qualified question families in two consecutive recall batches, or the completion probability is less than the probability threshold, the controller reclaims the quota that has not yet been assigned to the channel. Assigned but uncompleted batches are only canceled when there is a common deadline or when an abnormal rule is triggered; when multiple channels meet the addition conditions at the same time, the order is determined by descending marginal utility value, descending completion probability value, and ascending channel identifier, and the estimated time after addition must not exceed the remaining time.

[0133] Each batch completion event triggers a quota decision only once. The controller uses the query identifier, channel identifier, and batch number to form an idempotent key. When the same event arrives repeatedly, the recorded append, reclaim, migration, or hold actions are reused without changing the quota repeatedly. When the event batch number is less than the channel processed number, only candidate evidence is merged without retroactively modifying the dispatched task.

[0134] In the migration decision, the controller first excludes target channels that may have their candidate range broadened based on bitmap subset relationships, then excludes target channels that are not expected to complete a batch based on the remaining time, and finally selects the first item from the remaining targets in descending order of marginal utility value, descending order of completion probability value, and ascending order of target channel identifier; if there are no remaining targets, the unexecuted quota is registered as a cancelled quota and the migration is stopped.

[0135] Migration quotas are rounded down to the nearest whole number based on the target channel batch size. Quotas less than one target batch are recorded as unexecuted quotas and not added to the target batch. At the decision point, the available global concurrent slots and capacity snapshot version are first read from the channel scheduler after deducting slots already occupied by other queries from the target channel's maximum concurrency. However, the single channel for this query is still serially dispatched according to the shared cursor rules of step S4. The remaining dispatchable rounds are calculated by dividing the remaining duration of the common deadline by the target channel's historical 95th percentile single-batch time, rounded down. If the historical 95th percentile single-batch time is less than 1... If the result is zero, the target capacity is zero. Otherwise, the maximum number of complete batches that can be migrated in this query is the minimum value allowed by the following three conditions: the remaining number of dispatch rounds, the maximum number of batches per query in the channel registry, and at least one available global concurrent slot. Then, the number of batches that have been executed, are being executed, and have not been dispatched in this query is deducted. For the part that exceeds the capacity, the next target will be tried in the migration candidate order. If the total capacity of all targets is insufficient, the remaining part will be canceled, and the available slots, remaining duration, conservative time consumption, rounded result, and capacity snapshot version will be written to the migration record.

[0136] The migration target metrics are initialized in a fixed order: First, the marginal utility and completion probability of the target channel in the most recent completed batch in this query are read; if there are no completed batches for the target channel in this query, the historical qualified question family output rate, historical knowledge point coverage increment rate, historical cross-channel repetition rate, batch time 50th percentile, 95th percentile, and residual standard deviation are read from the runtime statistics table used in step S4, along with the same data for the real-time batches. , and Calculate the prior utility, and then calculate the prior index according to the aforementioned marginal utility and completion probability relationship; when the number of samples in this bin is less than 200, use the same channel without distinguishing the bitmap cardinality and the conservative record with all three utility statistics complete. If there is still no valid record or any utility statistics are missing, migration to this channel is prohibited and the recovery quota is registered as cancelled.

[0137] The reclaimed quota is only migrated to the compatible downgraded channel recorded in the multi-way execution plan. Before migration, it is verified that the candidate bitmap of the target channel is the same as the candidate bitmap of the original channel, or that the target bitmap is equal to the original bitmap after bitwise AND, thus proving that the target bitmap is a subset of the original bitmap. The target channel must also meet the following requirements: marginal utility is not less than the utility threshold, the expected batch time is not greater than the remaining time, and the index version is available. When there are multiple targets, the first one is selected in descending order of marginal utility value. If these conditions cannot be met, the reclaimed quota is recorded as canceled without relaxing the hard constraints.

[0138] After each action is recorded in the database, the controller recalculates the executed quota, executed quota, unassigned quota, and canceled quota for each channel, and requires that the sum of the four equals the sum of the initial quota, the additional quota, and the inbound quota for that channel minus the outbound quota. If the equation is not true, the subsequent allocation of the query is stopped, the plan status is set to quota ledger abnormal, and the completed batches are retained to avoid duplicate additions or loss of recovered quotas.

[0139] When the common deadline is reached, the controller sends a cancellation instruction with query identifier and batch identifier to the recall batches that have not yet been completed. It retains the batch candidates that have completed atomic writing before the deadline and records the unexecuted quota. If the retrieval channel is abnormal before the deadline, the unexecuted quota is migrated only if the expected batch time of the compatible downgrade channel is not greater than the remaining time and the bitmap subset verification passes. Otherwise, the corresponding batch is cancelled. Each migration or cancellation generates an execution record with fields including the original retrieval channel identifier, the target retrieval channel identifier, the migration quota, the cancelled batch identifier, the candidate bitmap summary before and after migration, the constraint retention verification result, and the downgrade status.

[0140] The deadline determination is uniformly read from the monotonic clock start point and available execution time in the multi-way execution plan. If the current monotonic clock difference is not less than the available execution time, the deadline action is triggered. After the cancellation instruction is issued, the corresponding batch is set to the cancellation status. After the channel confirmation or after more than 50 milliseconds of the cancellation confirmation window, it is set to canceled. Candidates that return in the confirmation window but whose completion time is later than the common deadline time are still not included in the reserve set.

[0141] The batch qualified set is deduplicated using question identifiers, the question family increment is deduplicated using question family identifiers, and the knowledge point coverage increment is calculated using the difference between the knowledge point identifier sets that have passed hard constraint verification. When the same candidate is returned from multiple channels, only one new qualified question family and one knowledge point coverage are counted. However, each additional channel hit is counted as the cross-channel duplicate quantity to ensure that the positive and penalty terms in the batch utility will not use the same counting caliber repeatedly.

[0142] The termination conditions of the dynamic control loop are: the current time reaches the common deadline, all channels have no quota to assign and no batches in execution, or the candidate bitmap is empty; when any of the conditions are met, the addition and migration stop, the completed batch candidate set is frozen and evidence fusion begins, and the batches in execution are only sent a cancellation instruction when the deadline condition is met, and the other termination conditions are frozen after the batches in execution are completed.

[0143] During the evidence fusion phase, the module normalizes both the channel score and the consistency of the hit evidence to a range of 0 to 1, and then... Calculate candidate The fusion score, Indicates candidate The dimensionless fusion score is better the larger it is. This represents the finite set of indexes corresponding to the original question retrieval intent, similar question retrieval intent, and teaching need intent output in step S2. Its members are limited to these three types of intents, and because it is a discrete set of indexes, it has no engineering units. Represents a set The intent index in the data has a value range of 100%. Furthermore, because it is used for discrete positioning, it lacks engineering units. This represents the index of candidate topics in the frozen batch candidate set. Because it is used for discrete location, it does not have engineering units. This indicates the intention obtained after normalizing the three calibration confidence levels in step S2. The probability is a dimensionless number between 0 and 1, and the sum of the probabilities of all intentions is 1. This indicates the intent of validation set calibration. The lower channel score has a weight between 0 and 1. Indicates candidate Intent The maximum normalized channel score in the associated channels, Indicates candidate Intent The degree of consistency between the query source fragment and the channel hit evidence is a dimensionless number from 0 to 1, calculated by the unified query side mapping and intersection-union ratio rule of the immediately following segment;

[0144] Evidence consistency is determined by a deterministic matching process within a unified query-side comparison space: for textual evidence, the character intersection-union ratio (IU) is calculated using the query text marker interval obtained from step S5 and the source text marker interval from step S2; for layout evidence, the rectangle intersection-union ratio is calculated using the mapped region normalized to the 0-1 coordinates of the query image and the source region; for formula evidence, the node intersection-union ratio is calculated using the query formula node set returned by tree editing matching and the source formula node set. When multiple sources exist simultaneously, a weighted average is calculated based on the normalized source contribution score from S2. If the matching evidence lacks query-side mapping, has inconsistent mapping versions, or does not contain similar sources, the corresponding... Set to zero and record the reason; the fusion weight record uses intent type, model version and index version as keys, is fitted by normalized depreciation cumulative gain on the validation set and recalibrated when any version is updated; when the same candidate is hit by multiple consensus graph channels, the maximum channel score is retained and non-duplicative evidence is merged.

[0145] The module aggregates candidates within the same question family according to the question family identifier. First, it sorts them in descending order of fusion score, then in descending order of the number of matching education attributes, and finally in ascending order of question identifier. For each question family, it retains the top three candidates with the highest fusion scores. Then, it takes one candidate from each question family in a round-robin fashion until it reaches the 200 fine-ranking quota to form a limited candidate set. This rule prevents a single question family from filling the fine-ranking quota. Candidates with the same fusion score are processed in a defined stable order. Non-hard constraint attributes in the conflict table are only used as input for subsequent fine-ranking and do not change the candidate bitmap.

[0146] When constructing a finite candidate set, the question families are first sorted according to the fusion score of the first candidate of each question family, and then the family-by-family round-robin is performed. If there are fewer than 200 candidates, all candidates are output and retained. After reaching 200 candidates, the addition stops, but the sorting position of the unselected candidates is retained for auditing. Single-member question families generated due to missing question family identifiers use the same rules as ordinary question families, and their pending review mark enters the fine sorting machine.

[0147] The fine ranking algorithm uses a cross-encoding Transformer, which encodes the normalized query text, candidate question stem text, formula structure similarity features, non-hard constraint educational attribute matching features, fusion score, and downgrade label as input, and outputs a relevance score from 0 to 1. The model is trained using relevant questions, irrelevant questions, and difficult negative samples of the same question family confirmed by teaching and research personnel, and optimized with pairwise ranking loss. The model version is selected by the first 20 normalized loss cumulative gain of the independent validation set. After inference, it is sorted in descending order of relevance score, descending order of fusion score, and ascending order of question label.

[0148] Before output, the module reads the attribute records and candidate bitmaps from the same question bank release version, the same relational data version, and the same unified question identifier space snapshot that are bound to the candidate bitmap in step S3 and used for batch review. Hard constraint review is performed on the candidates that pass the review. Candidates that pass the review are marked as consistent with education. Those that fail are deleted from the result list and written to the final elimination record. If any version summary changes, the output version will conflict but the real-time version will not be changed. The failure of hard constraint will not be compensated by the non-hard constraint score. When the number of results exceeds the upper limit of the caller's request, they are truncated in descending order of the fine-ranked score value, descending order of the fused score value, and ascending order of the question identifier. When the upper limit of the request is empty, this specific implementation uses 20 results.

[0149] The final search results include the question identifier, question family identifier, ranking score, fusion score, hit evidence, education consistency marker, channel identifier used, and degradation marker in the multi-way execution plan. These, along with the execution record identifier and the actual completion time, are written to the search results storage area. The completion status is returned to the calling end only after successful writing. If writing fails, it is retried once with the query identifier. If it still fails, a storage exception is returned and the search results are not falsified.

[0150] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0151] This invention connects multimodal preprocessing and multi-task semantic models with a unified query object, connects educational attribute verification and retrieval channel selection with verified semantic profiles, and constrains the recall range of each retrieval channel with a candidate bitmap in a unified question identifier space, so that retrieval intent, educational attributes, hit evidence, and question family identifiers can participate in fusion and verification in the same execution chain.

[0152] The improvement of the multi-way retrieval structure in this invention lies in the introduction of batch quota control under a common deadline. Based on the qualified output of the batch and the expected time consumption, the marginal utility and completion probability are formed, and the quota of the saturated or expected timeout channel is dynamically reclaimed. The channel is migrated only to compatible channels that maintain the same or more stringent candidate bitmap, thereby avoiding the relaxation of educational hard constraints during the downgrade process.

Claims

1. A dynamic multi-path retrieval method for semantic profiling of primary and secondary school test questions, characterized in that, include: S1. Obtain multimodal test question queries, perform normalization and structural parsing on them to obtain a unified query object; S2. Input the unified query object into the multi-task semantic model to generate retrieval intent and educational attribute candidates with source fragments, confidence scores and constraint tags, thus forming a semantic profile of the test question query; S3. Verify the attribute relationships and confidence levels in the semantic profile based on the educational constraint relationship data. Compile the educational attribute candidates that pass the verification and are marked as hard constraints into a candidate bitmap of the unified question identifier space, so that the attributes that fail the verification do not participate in the compilation, and obtain the verified semantic profile. S4. Based on the intent probability, input modality, and completeness of the verified semantic profile, select a compatible channel from the retrieval channels corresponding to the heterogeneous test question index, and generate a multi-path execution plan with initial quota and common deadline. S5. According to the multi-path execution plan, each compatible channel returns candidate questions with channel scores, hit evidence and question family identifiers in batches within the candidate bitmap, forming a batch candidate set; S6. Review hard constraints in batches, and determine the marginal utility and completion probability of each compatible channel based on qualified output and estimated time consumption. Based on this, add, reclaim, or migrate quotas to compatible channels that maintain the same or more stringent candidate bitmaps until the common deadline. Fuse evidence according to intent probability, aggregate, rank, and review constraints for retained candidates, and output search results.

2. The dynamic multi-path retrieval method for semantic profiling of primary and secondary school test questions according to claim 1, characterized in that, S1 includes: performing character encoding unification, symbol normalization, paragraph segmentation, and noise character filtering on natural language queries and question stem fragments; performing layout area detection on the question image and generating area sequences according to the question stem area, option area, illustration area, and formula area; performing optical character recognition on the question stem area and option area; parsing the operators, operands, and hierarchical relationships of the recognition results of the formula object or the formula area to generate a formula structure tree; and associating the normalized text, area sequence, recognized text, formula structure tree, and the position of each content in the original input into the unified query object.

3. The dynamic multi-path retrieval method for semantic profiling of primary and secondary school test questions according to claim 1, characterized in that, S2 includes: outputting the probabilities of original question retrieval intent, similar question retrieval intent, and teaching demand intent from the multi-task semantic model, as well as educational attribute candidates for subject, grade level, grade, textbook version, knowledge point, question type, and difficulty; using text intervals, page areas, or formula structure nodes in the unified query object that generate each candidate as source fragments; generating confidence scores based on the normalized scores output by the corresponding task; setting hard constraint flags when educational attribute candidates are obtained by user-explicit fields, unique mappings from textbook catalogs, or unique mappings from grade level and grade level, so that the corresponding attributes participate in the intersection operation of the candidate bitmap; otherwise, setting non-hard constraint flags so that the corresponding attributes are only used as fine-ranking features in step S6.

4. The dynamic multi-path retrieval method for semantic profiling of primary and secondary school test questions according to claim 1, characterized in that, S3 includes: verifying the subordinate relationship between subjects and knowledge points in the education ontology, verifying the inclusion relationship between school stages and grades in the school stage and grade mapping, and verifying the association relationship between textbook version, volume, and knowledge points in the textbook catalog version data; converting the educational attribute candidates with hard constraint tags that have passed the verification into their respective corresponding question identifier bitmaps, and performing an intersection operation on the question identifier bitmaps to obtain the candidate bitmaps; writing the educational attribute candidates whose confidence level does not reach the corresponding attribute confidence threshold, the educational attribute candidates that have not passed the association relationship verification, and their source fragments into a conflict table, and converting their hard constraint tags into non-hard constraint tags, so that the corresponding attributes are only used as fine ranking features in step S6; the attribute confidence threshold is determined by the lower bound of the confidence level corresponding to the educational attribute on the validation set when the false acceptance rate does not exceed the preset false acceptance rate upper limit.

5. The dynamic multi-path retrieval method for semantic profiling of primary and secondary school test questions according to claim 1, characterized in that, In step S4, the heterogeneous test question index includes a full-text index, a question fingerprint index, a learning-based sparse index, a dense vector index, an educational tag index, an optical character recognition layout index, and a formula structure index; a set of necessary profile attributes is pre-associated for each retrieval channel; retrieval channels associated with retrieval intentions whose intent probability is not less than the corresponding intent selection threshold are designated as the first channel set; retrieval channels supporting the input modality are designated as the second channel set; and retrieval channels whose verified semantic profiles already contain the corresponding set of necessary profile attributes are designated as the third channel set. The first channel... The intersection of the first set of channels, the second set of channels, and the third set of channels yields the compatible retrieval channel. When the intersection is empty, the baseline retrieval channel that supports the input modality and maintains the candidate bitmap is determined as the compatible retrieval channel. The intent selection threshold is the upper bound of the probability threshold allowed when the recall rate of the corresponding retrieval intent on the validation set is not lower than the preset recall rate lower limit. The initial quota and common deadline are determined based on the historical qualified question family output rate and historical time quantile of each compatible retrieval channel, and a compatible downgrade channel identifier that maintains the same or more stringent candidate bitmap is written for each compatible retrieval channel.

6. The dynamic multi-path retrieval method for semantic profiling of primary and secondary school test questions according to claim 1, characterized in that, S5 includes: sending the candidate bitmap as a filtering condition to the corresponding compatible retrieval channel along with the channel query; each compatible retrieval channel divides the recall batch according to the initial quota, and returns the candidate question identifier, normalized channel score, corresponding text hit interval, image layout area or formula structure node and question family identifier for each recall batch; before writing the batch candidate set, performing a bitwise AND check between the returned candidate question identifier and the candidate bitmap, and discarding candidate question identifiers with a check result of zero.

7. The dynamic multi-path retrieval method for semantic profiling of primary and secondary school test questions according to claim 5, characterized in that, In step S6, the number of question families newly added and verified by hard constraints, the knowledge point coverage increment, and the number of cross-channel repetitions between adjacent recall batches are counted. The batch utility increment is obtained by subtracting the penalty term for the number of cross-channel repetitions from the weighted sum of the number of newly added qualified question families and the knowledge point coverage increment. Based on the historical time quantiles, real-time queue length, current batch time, and remaining time before the common deadline for the corresponding compatible retrieval channel, the probability of the next recall batch being completed before the common deadline is estimated. The ratio of the batch utility increment to the estimated time of the next recall batch is used as the marginal utility. For compatible retrieval channels with marginal utility not less than the utility threshold and completion probability not less than the probability threshold, a recall batch quota is added. For compatible retrieval channels that have not added any qualified question families in two consecutive recall batches or whose completion probability is less than the probability threshold, unexecuted quotas are revoked. According to the compatible downgrade channel identifier, the revoked unexecuted quotas are migrated to compatible retrieval channels with marginal utility not less than the utility threshold and using the same or more stringent candidate bitmaps. The utility threshold and probability threshold are determined by the lower bound of marginal utility and the lower bound of completion probability when the qualified question family coverage per unit time on the validation set reaches a preset coverage lower limit, respectively.

8. The dynamic multi-path retrieval method for semantic profiling of primary and secondary school test questions according to claim 7, characterized in that, In step S6, the channel scores and hit evidence of each compatible retrieval channel are weighted and fused according to the intent probabilities corresponding to the original question retrieval intent, similar question retrieval intent, and teaching demand intent, respectively; candidate questions within the same question family are aggregated according to the question family identifier, and candidate questions whose fusion scores meet the fine-ranking quota are retained from each question family as a limited candidate set; the limited candidate set is fine-ranked and hard-constrained, and the candidate questions that pass the hard-constraint review are marked as educationally consistent, and the downgrade identifier in the multi-path execution plan is associated with the corresponding candidate question before being output.

9. The dynamic multi-path retrieval method for semantic profiling of primary and secondary school test questions according to claim 7, characterized in that, When the common deadline is reached, a cancellation instruction is sent to the recall batches that have not yet been completed, the candidate questions returned by the completed recall batches are retained, and the unexecuted quota is recorded; when the retrieval channel experiences an anomaly before the common deadline, the unexecuted quota is migrated to the compatible downgrade channel only if the expected batch time of the compatible downgrade channel is not greater than the remaining time of the common deadline and the candidate bitmap after migration is a subset of the candidate bitmap before migration or is the same as the candidate bitmap before migration; otherwise, the corresponding batch is cancelled. An execution record is generated, which includes the original retrieval channel identifier, the target retrieval channel identifier, the migration quota, the cancellation batch identifier, the candidate bitmap summary before migration, the candidate bitmap summary after migration, the constraint retention verification result, and the downgrade status.

10. A dynamic multi-path retrieval system for semantic profiling of primary and secondary school test questions, used to execute the dynamic multi-path retrieval method for semantic profiling of primary and secondary school test questions as described in any one of claims 1 to 9, characterized in that, include: The unified query object generation module is used to execute step S1; The semantic profile generation module is used to execute step S2; the education constraint verification module is used to execute step S3; the multi-path execution plan generation module is used to execute step S4; the bitmap-limited recall module is used to execute step S5; and the dynamic quota control and result output module is used to execute step S6.