An intelligent teaching resource integration and personalized pushing system and method

Abstract

The application provides a kind of intelligent teaching resource integration and individualized pushing system and method, method includes: obtaining the teaching resource to be integrated, executes format uniformity, text transcription, paragraph segmentation and field extraction, generates resource knowledge unit set;According to resource knowledge unit set, execute grouping according to knowledge point mark, semantic alignment and consistency verification, generate semantic invariant set;Resource knowledge unit set and semantic invariant set are carried out field alignment, identify condition missing, meaning deviation, boundary weakening and exception missing, generate semantic drift record set and resource disposal task set;According to resource disposal task set, semantic drift record set and imitation learning resource disposal model, corresponding resource knowledge unit is executed isolation, correction or weight reduction processing, and is combined learner current learning stage, historical answer record and mastered knowledge record, generates candidate resource set and push constraint set;According to candidate resource set and push constraint set generates individualized push queue.

Description

Technical Field

[0001] This invention relates to the field of teaching resource data processing technology, and in particular to an intelligent teaching resource integration and personalized push system and method. Background Technology

[0002] With the development of educational informatization and intelligent teaching platforms, multi-source teaching resources such as textbooks, courseware, and micro-lecture videos are being widely integrated into unified teaching platforms. Existing technologies typically perform preliminary integration and delivery of teaching resources based on knowledge point tags, course chapters, and learner profiles to alleviate the problem of fragmented teaching resources.

[0003] However, most existing intelligent teaching resource integration and delivery technologies only focus on superficial features such as unified resource format and tag matching. They lack effective underlying constraints regarding whether the core propositions, boundary limitations, preconditions, and exceptions of the same knowledge point remain consistent across different resources. Especially after generative technologies are involved in rewriting and expanding teaching content, the same knowledge point is often generated into multiple versions. Although these versions share the same theme, they are prone to omissions of preconditions, alterations of applicable scope, and weakening of core propositions. While a single resource may appear to have superficial completeness, when the platform continuously integrates and alternately delivers multiple versions of resources, learners are highly susceptible to receiving inconsistent expressions of knowledge, leading to confusion in their understanding.

[0004] Therefore, this invention proposes an intelligent teaching resource integration and personalized recommendation system and method. The information disclosed in the background section is only for enhancing understanding of the background of this disclosure, and therefore may contain prior art information that is not common knowledge to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing an intelligent teaching resource integration and personalized push system and method, thereby resolving the technical problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The first part of this invention provides a method for intelligent teaching resource integration and personalized delivery, comprising the following steps: S1. Obtain the teaching resources to be integrated, perform format unification, text transcription, paragraph segmentation and field extraction on the teaching resources to be integrated, extract knowledge point identifiers, explanation content, scope of application, preconditions and exceptions, and generate a set of resource knowledge units. S2. Based on the resource knowledge unit set, perform grouping, semantic alignment and consistency verification according to knowledge point identifiers, extract core propositions, boundary constraints, preconditions and exceptions, and generate a set of semantic invariants; S3. Align the resource knowledge unit set with the semantic invariant set by field, identify missing conditions, meaning shifts, weakened boundaries and omissions of exceptions, and generate a semantic drift record set and a resource disposal task set. S4. Based on the resource disposal task set, semantic drift record set, and the imitation learning resource disposal model trained on the historical resource correction sample set, the corresponding resource knowledge units are isolated, corrected, or downweighted. Combined with the learner's current learning stage, historical answer records, and mastered knowledge records, a candidate resource set and a push constraint set are generated. S5. Generate a personalized push queue based on the candidate resource set and the push constraint set, push the target teaching resources to the learner's terminal, collect the resource effect feedback set, and update the semantic invariant set and the candidate resource set based on the resource effect feedback set.

[0007] S1 specifically includes: acquiring teaching resources to be integrated from textbook courseware, question bank analysis, micro-lecture video text, classroom handouts, and online teaching documents; performing format unification, version identification completion, and text transcription processing on each teaching resource to be integrated to generate a standardized resource record set; performing paragraph segmentation, title alignment, and cross-modal content alignment based on the standardized resource record set to generate a set of candidate knowledge unit fragments that retain resource identifiers, fragment positions, original expressions, and adjacent contexts; extracting knowledge point identifiers, explanation content, scope of application, preconditions, and exceptions from the set of candidate knowledge unit fragments, deleting invalid candidate fragments that lack knowledge point identifiers or explanation content, and generating a set of resource knowledge units.

[0008] S2 specifically includes: reading the resource knowledge unit set, directly grouping resource knowledge units with the same standard knowledge point identifier, and performing grouping similarity calculation on resource knowledge units with temporary knowledge point identifiers to generate a set of similar knowledge unit groups; performing semantic alignment on each similar knowledge unit group in the similar knowledge unit group according to the core proposition field, boundary constraint field, precondition field, and exception description field, extracting repeated and semantically stable fragments to generate a semantic skeleton candidate set; performing consistency verification based on the semantic skeleton candidate set, retaining candidate fragments that remain stable in multiple source resources and are not negated by conflicting expressions, deleting candidate fragments that cannot be supported by other similar resources, and generating a set of semantic invariants.

[0009] S3 specifically includes: reading the resource knowledge unit set and the semantic invariant set; aligning each resource knowledge unit with the semantic invariant records under the same knowledge point identifier; calculating the completeness of the explanation content, boundary retention, consistency of preconditions, and retention of exception descriptions; and generating a semantic difference feature set by combining the strength of field missing markers and conflict markers; identifying condition missing, meaning shift, boundary weakening, and exception omission based on the semantic difference feature set; determining the main drift type and additional drift type; and generating a semantic drift record set; and generating corresponding isolation tasks, correction tasks, or de-weighting tasks based on the semantic drift record set according to the drift intensity, main drift type, and resource exposure, thus forming a resource disposal task set.

[0010] S4 specifically includes: reading the resource disposal task set, semantic drift record set, and historical resource correction sample set; training an imitation learning resource disposal model based on historical manual disposal trajectories; and outputting a resource disposal action sequence based on the imitation learning resource disposal model; performing isolation, correction, or deweighting processing on the corresponding resource knowledge units according to the resource disposal action sequence, wherein the correction processing includes at least supplementing missing preconditions, restoring boundary constraints, and supplementing exception descriptions; and performing consistency verification on the processed resource knowledge units to generate a candidate resource set; reading the learner's current learning stage, historical answer records, and mastered knowledge records, and combining them with the candidate resource set to generate a set of push constraints used to restrict the eligibility, order, and frequency of candidate resources entering the personalized push queue.

[0011] S5 specifically includes: reading the candidate resource set and the push constraint set; performing filtering, sorting, and queue arrangement on the candidate resources to generate a personalized push queue; pushing the target teaching resources to the learner's terminal according to the personalized push queue; collecting the learner's reading completion status, answer results, dwell time, number of repeated openings, and error correction feedback on the target teaching resources to generate a resource effect feedback set; calculating the feedback reliability based on the resource effect feedback set, determining the valid feedback records, and performing field updates, resource weight adjustments, and resource disposal task write-back on the semantic invariant set and the candidate resource set based on the valid feedback records.

[0012] The second part of this invention provides an intelligent teaching resource integration and personalized push system, comprising: The resource knowledge unit generation module is used to acquire teaching resources to be integrated, perform format unification, text transcription, paragraph segmentation and field extraction on the teaching resources to be integrated, extract knowledge point identifiers, explanation content, scope of application, preconditions and exceptions, and generate a set of resource knowledge units. The semantic invariant generation module is used to perform grouping, semantic alignment and consistency verification based on the resource knowledge unit set according to the knowledge point identifier, extract core propositions, boundary constraints, preconditions and exceptions, and generate a set of semantic invariants. The semantic drift identification module is used to align the resource knowledge unit set with the semantic invariant set, identify missing conditions, meaning shifts, weakened boundaries and omissions of exceptions, and generate a semantic drift record set and a resource disposal task set. The resource disposal and constraint generation module is used to perform isolation, correction or weight reduction processing on the corresponding resource knowledge units based on the resource disposal task set, semantic drift record set and imitation learning resource disposal model trained based on historical resource correction sample set, and combine the learner's current learning stage, historical answer record and mastered knowledge record to generate candidate resource set and push constraint set; The personalized push and feedback update module is used to generate a personalized push queue based on the candidate resource set and the push constraint set, push the target teaching resources to the learner's terminal, collect the resource effect feedback set, and update the semantic invariant set and the candidate resource set based on the resource effect feedback set.

[0013] The beneficial effects of this invention are as follows: This invention constructs a resource knowledge unit set, a semantic invariant set, and a semantic drift record set, which enables consistency verification of the core propositions, boundary constraints, preconditions, and exceptions of teaching resources from different sources for the same knowledge point, thereby reducing the problem of inconsistent interpretations among multiple versions of teaching resources.

[0014] This invention identifies missing conditions, semantic shifts, weakened boundaries, and omitted exceptions, and generates corresponding resource disposal task sets. It can isolate, correct, or downgrade teaching resources with semantic drift, thereby improving the reliability of candidate resource sets.

[0015] This invention introduces an imitation learning resource disposal model to generate a sequence of resource disposal actions based on historical manual disposal trajectories. This improves the automation and consistency of the resource correction process, thereby reducing the cost of manual screening and correction.

[0016] This invention generates a set of push constraints by combining the learner's current learning stage, historical answer records, and mastered knowledge records, and constructs a personalized push queue accordingly. This improves the matching degree between candidate resources and the learner's current learning status, thereby improving the accuracy of personalized pushes.

[0017] This invention improves the system's ability to continuously control the semantic drift of teaching resources by collecting resource effect feedback sets and updating the semantic invariant set, candidate resource set, and resource disposal tasks. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of an intelligent teaching resource integration and personalized push method according to the present invention; Figure 2 This is a schematic diagram of the framework of an intelligent teaching resource integration and personalized push system according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1: As Figure 1 As shown in the figure, this embodiment provides a method for intelligent teaching resource integration and personalized push, including the following steps: S1. Obtain the teaching resources to be integrated, perform format unification, text transcription, paragraph segmentation and field extraction on the teaching resources to be integrated, extract knowledge point identifiers, explanation content, scope of application, preconditions and exceptions, and generate a set of resource knowledge units. S2. Based on the resource knowledge unit set, perform grouping, semantic alignment and consistency verification according to knowledge point identifiers, extract core propositions, boundary constraints, preconditions and exceptions, and generate a set of semantic invariants; S3. Align the resource knowledge unit set with the semantic invariant set by field, identify missing conditions, meaning shifts, weakened boundaries and omissions of exceptions, and generate a semantic drift record set and a resource disposal task set. S4. Based on the resource disposal task set, semantic drift record set, and the imitation learning resource disposal model trained on the historical resource correction sample set, the corresponding resource knowledge units are isolated, corrected, or downweighted. Combined with the learner's current learning stage, historical answer records, and mastered knowledge records, a candidate resource set and a push constraint set are generated. S5. Generate a personalized push queue based on the candidate resource set and the push constraint set, push the target teaching resources to the learner's terminal, collect the resource effect feedback set, and update the semantic invariant set and the candidate resource set based on the resource effect feedback set.

[0021] S1 specifically includes the following sub-steps: S110. Obtain teaching resources to be integrated from textbook courseware, question bank analysis, micro-lecture video text, classroom handouts and online teaching documents. Perform format unification, version identification completion and text transcription processing on each teaching resource to be integrated to generate a standardized resource record set.

[0022] Specifically, the system extracts titles, body text, page numbers, paragraph numbers, and file generation times from textbooks, courseware, lecture notes, and online teaching documents; it extracts question identifiers, question stems, answer texts, explanation texts, and question type tags from question bank analyses; and it extracts video identifiers, original subtitles, speech-to-text transcripts, and sentence / segment start and end times from micro-lecture videos. Subsequently, the content from different sources is uniformly converted into structured text records, ensuring that each record contains at least a resource identifier, source type, version number, text content, paragraph position, and timestamp. The paragraph position identifies the page number, paragraph number, or time segment position of the text content in the original resource, while the timestamp identifies the record generation time or transcription time.

[0023] For the version identifier completion process, the explicit version string in the resource file name, cover date, header and footer is read first. When the explicit version string is missing, a machine-completed version number is generated according to the course batch, upload time and source type, and the version source mark is written at the same time to distinguish between manually marked version numbers and machine-completed version numbers.

[0024] For the text transcription of micro-lesson videos, semantic pauses and fixed duration windows are used to constrain the generation of transcription segments. Segmentation is prioritized at points where the sentence meaning is complete. When the sentence meaning is continuous and the pause time is less than the preset threshold, the segment is merged into the previous segment. For example, a 34-second lecture is divided into a first transcription segment with a start time of 12 seconds and an end time of 19 seconds, and a second transcription segment with a start time of 19 seconds and an end time of 34 seconds, and written to the corresponding paragraph positions respectively.

[0025] For abnormal records, delete records whose text content length is below a preset lower limit, whose title field is empty and whose body field is empty, whose timeline is reversed, and whose text content is duplicated with existing record resource identifiers. After the above processing, a standardized resource record set is output, which is available for S120 to use.

[0026] S120. Based on the standardized resource record set, perform paragraph segmentation, title alignment, and cross-modal content alignment on each resource record to obtain a set of candidate knowledge unit fragments that correspond one-to-one with the knowledge representation position. Specifically, first calculate the segmentation score for each standardized resource record:

[0027] in, To score by segmentation; This is the value indicating a change in heading level. This is a value indicating a change in the serial number; This is the pagination change indicator value; This is the normalized value for the pause time; , , , These are the corresponding weights.

[0028] In practice, the above indicator values ​​are discretized and assigned as follows: when a change in the heading level of the text is detected, H is 1, otherwise it is 0; when an increment of bullet point or number sequence is detected, N is 1, otherwise it is 0; when crossing physical page boundaries, P is 1, otherwise it is 0.

[0029] For the pause time normalized value T, the maximum and minimum value normalization process is used, that is:

[0030] Where t is the currently detected pause time. and These are the maximum and minimum pause times within a preset duration window, respectively. During implementation, the values ​​can be set to... , , , When the segmentation score is not lower than the preset segmentation threshold, the current position is determined as a candidate segmentation point, thereby forming multiple knowledge representation segments.

[0031] Subsequently, a standard glossary of titles was established, grouping "basic concepts," "concept definitions," and "definition descriptions" into the same title category, and grouping "applicable conditions" and "scope of application" into the same title category. Based on this, title alignment was performed, so that knowledge expression fragments with different titles but similar semantics entered into the same title alignment set.

[0032] For cross-modal content alignment, the alignment score is calculated using micro-lecture video transcripts as the first object and textbook or lecture slides as the second object:

[0033] in, To score for the position; Title similarity; For contextual similarity; For time proximity; , , These are the corresponding weights.

[0034] To address the issue of second objects such as textbooks, courseware, or online documents lacking native timestamps, time proximity is calculated. At that time, based on the total number of pages or paragraphs of the second object, it is linearly mapped to the total duration of the micro-lesson video, generating virtual timestamps for each page or paragraph.

[0035] Then, the reciprocal normalized value of the time difference between the start time of the video transcription segment and the virtual timestamp is calculated as... If the second object is completely unable to generate temporal features, then the weight configuration is dynamically adjusted to make... And enlarged proportionally and The weight.

[0036] When implementing, it is possible to take , , When the alignment score is not lower than the preset alignment threshold, a cross-modal alignment relationship is established; for example, if the title of the video transcription segment is "Definition of Newton's Second Law" and the title of the courseware page is "Newton's Second Law", and both of them contain the words "force", "mass" and "acceleration" in their context, then an alignment relationship is established.

[0037] Each generated knowledge unit candidate fragment retains its resource identifier, fragment position, original description, and adjacent context. The adjacent context retains at least one preceding and one following paragraph of text for subsequent extraction of preconditions and exception descriptions. After the above processing, a knowledge unit candidate fragment set is output, which is available for use by S130.

[0038] S130. Extract knowledge point identifiers, explanation content, scope of application, preconditions and exceptions from the knowledge unit candidate fragment set, delete invalid candidate fragments with missing knowledge point identifiers or missing explanation content, and output the resource knowledge unit set.

[0039] Specifically, each candidate knowledge unit fragment is first matched with the standard knowledge point name in the course knowledge graph. When the standard knowledge point name or its standard alias appears in the candidate fragment, the corresponding knowledge point identifier is written. When the standard knowledge point name is not matched but the title category, body keywords and adjacent context all point to the same topic, a temporary knowledge point identifier is generated and the temporary knowledge point identifier is bound to the resource identifier and fragment position.

[0040] Subsequently, the original statement is used as the main field of the explanation content, and the scope of application, preconditions and exceptions are identified in the adjacent context. Among them, statements containing "applies to", "under the circumstances of", "for the purpose of", are written as the scope of application, statements containing "provided that", "should be completed first", "need to be mastered first" are written as the preconditions, and statements containing "except", "however", "special cases" are written as the exceptions.

[0041] To avoid inaccurate comparisons caused by a single segment corresponding to multiple core topics, when the same candidate segment involves two or more knowledge point identifiers, it is split into multiple resource knowledge units according to the knowledge point identifiers, inheriting the same resource identifier but assigned different sub-segment numbers. For deletion rules, in addition to candidate segments lacking knowledge point identifiers or explanation content, candidate segments that only contain titles without body text, only contain examples without definitions, or have conflicting contexts that cannot be resolved based on title alignment results are also deleted.

[0042] The final output is a set of resource knowledge units. Each resource knowledge unit in the set is bound to a unique resource identifier and a unique fragment location, and includes at least a knowledge point identifier, explanation content, scope of application, preconditions, and exception descriptions. The set of resource knowledge units is available for use by S210 and S310.

[0043] S2 specifically includes the following sub-steps: S210: Read the resource knowledge unit set output by S130, group multiple resource knowledge units according to the knowledge point identifier, and generate a knowledge unit group set with the same point.

[0044] In practice, the knowledge point identifiers in the resource knowledge unit set are first divided into standard knowledge point identifiers and temporary knowledge point identifiers. The standard knowledge point identifier is an identifier that corresponds one-to-one with the standard nodes in the course knowledge graph, while the temporary knowledge point identifier is an identifier to be confirmed generated by S130 when a standard node is not hit.

[0045] For resource knowledge units with the same standard knowledge point identifier, they are directly written into the same knowledge unit group. For resource knowledge units with temporary knowledge point identifiers, they are not directly grouped independently. Instead, the group similarity of the resource knowledge unit and the group title center and group context center records of the corresponding knowledge unit groups with standard knowledge point identifiers is calculated. The group similarity calculation formula is as follows:

[0046] in, Indicates group similarity; This indicates the matching degree of knowledge point identifiers, which is used to characterize the degree of name overlap or alias mapping between temporary knowledge point identifiers and standard knowledge point identifiers. Title similarity is used to characterize the degree of consistency between the title of a resource knowledge unit and the central theme of the title of a target group. Context similarity is used to characterize the degree of consistency between the adjacent contexts of a resource knowledge unit and the context center of the target group; , , These represent the corresponding weights, and the sum of the three is 1.

[0047] When implementing, it is possible to take , , To ensure that the matching degree of knowledge point identifiers is dominant. If the grouping similarity between a resource knowledge unit corresponding to a temporary knowledge point identifier and the group corresponding to a standard knowledge point identifier is not lower than the preset grouping threshold of 0.78, it is merged into the same knowledge unit group corresponding to the standard knowledge point identifier; if both are lower than the preset grouping threshold, it is retained as a group of same knowledge units to be confirmed, and a confirmation mark is written in the group record.

[0048] To prevent the same resource knowledge unit from being repeatedly assigned to multiple groups, each resource knowledge unit is only allowed to be assigned to one group of similar knowledge units; when it meets the grouping threshold of multiple groups at the same time, only the grouping result with the highest grouping similarity is retained.

[0049] After the above processing, a set of knowledge units with the same point is output. Each knowledge unit in the set of knowledge units with the same point contains at least a group identifier, a knowledge point identifier, a list of resource knowledge units in the group, a group title center, and a group context center. The set of knowledge units with the same point is available for S220 to call.

[0050] S220. Perform semantic alignment on each knowledge unit group in the same knowledge unit group set, extract the core proposition fragments, boundary limiting fragments, precondition fragments and exception explanation fragments that appear repeatedly in multiple resource knowledge units and are semantically stable, and generate a semantic skeleton candidate set.

[0051] In practice, the resource knowledge units within each knowledge unit group are first split into core proposition fields, boundary constraint fields, precondition fields, and exception description fields. Then, pre-clustering is performed within each field according to the title category, so that field records with different sources but the same title category enter the same aligned subset.

[0052] Subsequently, semantic similarity is calculated for the field text in each alignment subset, and field texts with semantic similarity not lower than a preset alignment threshold of 0.75 are grouped into the same candidate fragment cluster. The retention of candidate fragment clusters is not based on the completeness of a single source, but rather on the number of supporting sources and the stability of the representation within the group; a candidate fragment cluster is only allowed to enter the semantic skeleton candidate set if the number of supporting sources is at least 2. If the title field and body field corresponding to the same candidate fragment cluster conflict, the body field takes precedence, and the title field is retained only as auxiliary positioning information; if the boundary constraint field and the core proposition field have inconsistent constraint strengths, the boundary constraint field with stronger constraints is retained first, and a conflict marker is written into the candidate fragment cluster record for consistency verification in S230.

[0053] For example, if in a set of resource knowledge units, three sources all state "the denominator should be removed before solving a fractional equation", and another source only states "fractional equations can be directly rearranged", then the former statement enters the candidate fragment cluster of preconditions, and the latter statement forms a conflict fragment cluster separately because it is inconsistent with the majority of sources.

[0054] In this embodiment, performing semantic similarity calculation specifically includes: This method calls a pre-trained language representation model (such as BERT or RoBERTa) to extract the hidden layer output corresponding to the [CLS] tags of the text from both comparison fields as high-dimensional sentence vector features. Then, the cosine similarity between these two high-dimensional sentence vector features is calculated. Compared to traditional literal overlap calculation, this mechanism can accurately identify candidate segments that are different in expression but semantically equivalent.

[0055] After the above processing, a semantic skeleton candidate set is output. Each candidate record in the semantic skeleton candidate set contains at least a knowledge point identifier, a fragment cluster identifier, a fragment type, a candidate text, the number of supporting sources, a conflict marker, and a source list. The semantic skeleton candidate set is available for S230 to call.

[0056] S230. Perform consistency verification on each candidate fragment based on the semantic skeleton candidate set, retain candidate fragments that remain stable in multiple source resources and are not negated by conflicting representations, delete candidate fragments that only appear in isolated resources and cannot be supported by other resources of the same point, and output the semantic invariant set.

[0057] In practice, an invariant retention score is calculated for each candidate fragment cluster. The formula for calculating the invariant retention score is as follows:

[0058] in, Indicates that the score is retained for invariants; This indicates the proportion of supporting sources, representing the percentage of sources supporting the candidate fragment cluster relative to the total number of sources in the same knowledge unit group. Indicates the consistency of expression within a group, used to characterize the degree of consistency among the text expressions within a candidate fragment cluster; Indicates the conflict ratio, used to characterize the proportion of field records that conflict with the candidate fragment cluster within the group; , , These represent the corresponding weights.

[0059] When implementing, it is possible to take , , When the invariant retention score is not lower than the preset retention threshold of 0.72, the corresponding candidate fragment is retained as a semantic invariant; when the invariant retention score is lower than the preset retention threshold and the number of supported sources is only 1, the corresponding candidate fragment is deleted; when the invariant retention score is lower than the preset retention threshold, but the source type belongs to the preset high-confidence source and there is no reverse conflict fragment, the candidate fragment is temporarily stored as a low-confidence semantic invariant and the confidence level field is written into the record.

[0060] If there are two mutually exclusive candidate segments under the same segment type, and the difference between their scores does not exceed the preset difference threshold of 0.05, then neither of them will be directly written into the semantic invariant set. Instead, they will be retained in the current knowledge unit group with the same point as conflict markers and candidate records, which will be used as reference inputs when S310 performs semantic drift recognition.

[0061] The final output is a set of semantic invariants. Each semantic invariant record in the set contains a knowledge point identifier, core proposition, boundary constraints, preconditions, exception descriptions, confidence level, and source tracing information. The set of semantic invariants is available for use by S310, S420, and S530.

[0062] S3 specifically includes the following sub-steps: S310: Read the resource knowledge unit set output by S130 and the semantic invariant set output by S230, align the fields of each resource knowledge unit with the semantic invariant records under the same knowledge point identifier, calculate the completeness of the explanation content, the boundary retention, the consistency of the preconditions, and the retention of the exception description, and generate a semantic difference feature set.

[0063] In practice, a one-to-one comparison relationship is first established using the knowledge point identifier as the primary key. When there are multiple semantic invariant records under the same knowledge point identifier, the semantic invariant record with the highest confidence level and without conflict markers is selected as the current comparison benchmark. When there are only semantic invariant records with conflict markers, the conflict markers are retained and field alignment is continued so that subsequent drift recognition can identify abnormal differences caused by conflict semantic skeletons.

[0064] When aligning fields, the same-name fields are compared according to the core propositions of the explanation content, the boundary limits of the applicable scope, the preconditions of the preconditions, and the exceptions of the exceptions. The same field comparison also uses a cosine similarity calculation mechanism based on a pre-trained language model. By quantifying the cosine value of the angle between the vector spaces, the completeness of the explanation content, the degree of boundary preservation, the consistency of the preconditions, and the degree of exception preservation are obtained respectively.

[0065] When a resource knowledge unit is missing a field but the corresponding field exists in the semantic invariant record, the missing field is recorded and the corresponding consistency index is set to 0.

[0066] For semantic invariant fields with low confidence levels, they are not deleted; instead, their contribution to the comprehensive calculation is reduced to avoid amplifying the drift judgment result from a single low-confidence field. The completeness of the explanation content, boundary retention, consistency of preconditions, and retention of exceptions are all represented by normalized values ​​from 0 to 1, where 1 indicates that the field corresponding to the resource knowledge unit is consistent with the field corresponding to the semantic invariant record, and 0 indicates complete absence or obvious reversal.

[0067] Then, the normalization results of each field and the collision marker strength are combined for calculation to obtain the drift strength. The drift strength calculation formula is:

[0068] in, Indicates drift intensity; Indicates the completeness of the explanation content; Indicates the degree of boundary retention; Indicates the consistency of preconditions; Indicate the degree of retention for exceptions; Indicates the intensity of the conflict marker; , , , , These represent the corresponding weights.

[0069] When implementing, it is possible to take , , , , For example, if a resource knowledge unit has a completeness of 0.92, a boundary retention of 0.40, a precondition consistency of 0.85, and an exception retention of 0, then its drift intensity will be significantly higher than that of similar records with complete boundaries and complete exceptions.

[0070] After the above processing, a semantic difference feature set is output. Each feature record in the semantic difference feature set contains at least the resource identifier, knowledge point identifier, fragment position, completeness of explanation content, boundary retention, consistency of preconditions, retention of exception description, conflict marker strength, field missing marker, and drift strength. The semantic difference feature set is available for S320 to use.

[0071] S320. Identify missing conditions, semantic shifts, weakened boundaries, and omissions of exceptions in each resource knowledge unit based on the semantic difference feature set, and form a semantic drift record set.

[0072] In practice, a two-layer identification method combining field priority and drift intensity verification is adopted. First, field priority identification is performed: when a precondition field of a resource knowledge unit is missing, but the corresponding semantic invariant record contains a precondition field, it is preferentially determined as a missing condition; when an exception description field of a resource knowledge unit is missing, but the corresponding semantic invariant record contains an exception description field, it is preferentially determined as an omitted exception; when a resource knowledge unit retains an applicable scope field, but its applicable object, applicable conditions, or applicable boundary is narrowed, broadened, or replaced compared to the semantic invariant record, it is determined as a weakened boundary; when the content of the resource knowledge unit's explanation changes from the core proposition in the semantic invariant record in terms of subject-predicate relationship, causal relationship, or object of action, it is determined as a semantic shift.

[0073] Then perform drift intensity verification: when the drift intensity corresponding to the drift type identified by the field is lower than the preset type establishment threshold, the drift type is not directly deleted, but written to the additional drift type field; when the drift intensity is not lower than the preset type establishment threshold, the drift type is written to the main drift type field.

[0074] In implementation, 0.35 can be set as the threshold for type validity. For resource knowledge units that simultaneously meet two or more drift conditions, they are not forcibly compressed into a single label. Instead, multi-label results are retained, and the class with the highest drift intensity contribution value is taken as the primary drift type. For example, if a resource knowledge unit simultaneously lacks an exception description and changes the causal direction in the core proposition, then the meaning shift is written as the primary drift type, and the exception omission is written as the secondary drift type.

[0075] The final output is a semantic drift record set. Each semantic drift record in the semantic drift record set contains at least a resource identifier, a knowledge point identifier, a fragment position, a main drift type, an additional drift type, a drift intensity, and a corresponding reference invariant identifier. The semantic drift record set is available for use by S330 and S410.

[0076] S330. Generate a resource disposal task set for each resource knowledge unit based on the semantic drift record set. Among them, generate isolation tasks for resource knowledge units whose drift intensity reaches the isolation threshold, generate correction tasks for resource knowledge units whose drift intensity is in the correction range, and generate deweighting tasks for resource knowledge units whose drift intensity is in the deweighting range.

[0077] In practice, the task range is first divided according to the drift intensity: when the drift intensity is not lower than 0.75, an isolation task is generated; when the drift intensity is not lower than 0.45 but lower than 0.75, a correction task is generated; when the drift intensity is not lower than 0.20 but lower than 0.45, a deweighting task is generated; when the drift intensity is lower than 0.20, no resource disposal task is generated, and only the monitoring mark is retained.

[0078] Subsequently, a type linkage rule was introduced: when the main drift type is meaning offset, even if the drift intensity falls into the correction range, the generation of isolation tasks is promoted; when the main drift type is exception omission or condition missing, and the drift intensity is in the correction range, the generation of correction tasks is maintained; when the main drift type is only boundary weakening, and the drift intensity is in the deweighting range, the generation of deweighting tasks is maintained.

[0079] To determine the execution order within the same task type, a resource disposal priority is calculated for each task record. The formula for calculating the resource disposal priority is:

[0080] in, Indicates the priority of resource disposal; Indicates drift intensity; This indicates the resource exposure level, used to characterize the range within which the corresponding resource knowledge unit has entered the candidate resource set or has been invoked by the terminal. This represents the source risk coefficient, used to characterize the risk level corresponding to the type of resource source. , , These represent the corresponding weights.

[0081] When implementing, it is possible to take , , The higher the resource exposure, the greater the drift intensity, and the higher the source risk coefficient, the higher the priority of resource disposal.

[0082] The final output is a resource disposal task set. Each task record in the resource disposal task set includes at least the resource identifier, knowledge point identifier, fragment position, main drift type, additional drift type, drift intensity, task type, resource disposal priority, and task generation time. The resource disposal task set is available for S410 to call.

[0083] S4 specifically includes the following sub-steps: S410 reads the resource disposal task set and semantic drift record set output by S330, and reads the historical resource correction sample set. Based on the historical manual disposal trajectory, it trains the imitation learning resource disposal model and outputs the resource disposal action sequence.

[0084] In practice, each sample in the historical resource correction sample set is uniformly organized into three types of associated records: status record, expert action record, and verification result record. The status record includes at least the drift type, drift intensity, drift position, field missing marker, resource source type, and reference semantic invariant identifier. The expert action record includes at least the action category, action execution order, and action effect field. The verification result record includes at least the corrected consistency score and whether it has entered the candidate resource set.

[0085] Subsequently, each task record in the resource disposal task set is concatenated with its corresponding semantic drift record and quantized into the current state input. The specific vectorization process includes: The main drift type, additional drift types, and field missing markers are one-hot encoded and then concatenated with continuous variables (such as drift intensity and resource exposure) to generate a one-dimensional state feature vector. This state feature vector is then input into the imitation learning resource disposal model.

[0086] In this embodiment, the imitation learning resource disposal model adopts a deep feedforward neural network (FNN) structure, including an input layer, multiple fully connected hidden layers (using the ReLU activation function), and a softmax output layer to output discrete probabilities of each candidate action. .

[0087] The imitation learning resource disposal model is trained offline based on the behavior cloning paradigm: expert action records in historical manual disposal trajectories are used as hard labels, the cross-entropy loss between the model's predicted probability distribution and the expert action labels is calculated, and the Adam optimizer is used to continuously update the network parameters through backpropagation until the model's action prediction accuracy on the validation set meets the convergence condition.

[0088] The candidate action set of the imitation learning resource disposal model includes at least isolating resource knowledge units, filling in missing preconditions, restoring boundary constraints, filling in exception descriptions, adjusting resource weights, and retaining observation tags.

[0089] For each candidate action, calculate the overall action score. The formula for calculating the overall action score is:

[0090] Where k is the index number of the candidate action. Indicates the first The overall score of each candidate action; The output of the imitation learning resource disposal model represents the first... The probability of each candidate action; Indicates the first The degree of matching between each candidate action and the current drift type and field missing marker; Indicates the first The degree of preservation of semantic invariant constraints after each candidate action is executed; , , These represent the corresponding weights.

[0091] When implementing, it is possible to take , , When the overall action score is not lower than the preset action threshold of 0.65, the corresponding candidate action is written into the resource disposal action sequence. When multiple candidate actions simultaneously meet the threshold, they are sorted by overall action score from highest to lowest and written sequentially. If the highest overall action score is lower than the preset action threshold, or if the current state input contains a drift type combination not present in the training set, the model output is not used directly. Instead, a preset deterministic rule is invoked to generate a resource disposal action sequence to ensure the executability of the resource disposal action sequence.

[0092] The final output resource disposal action sequence includes at least the fields of resource identifier, knowledge point identifier, reference semantic invariant identifier, fragment position, action category, action order, and action effect. The resource disposal action sequence is available for S420 to call.

[0093] S420. Based on the resource disposal action sequence, perform isolation, correction or deweighting processing on the corresponding resource knowledge unit, and output the candidate resource set.

[0094] In practice, resource knowledge units are used as the basic processing objects. For records with the action category of "isolate resource knowledge unit", the corresponding resource knowledge unit is marked as isolated and prohibited from being written to the candidate resource set, while the isolation reason is written. For records with the action categories of "fill in missing preconditions", "restore boundary constraints", and "fill in exception descriptions", field-level write-back processing is performed according to the action order recorded in the resource processing action sequence. Among them, filling in missing preconditions involves writing the precondition fields in the semantic invariant set into the precondition fields of the corresponding resource knowledge unit, restoring boundary constraints involves writing the boundary constraint content in the semantic invariant set into the applicable scope field of the corresponding resource knowledge unit, and filling in exception descriptions involves writing the exception description content in the semantic invariant set into the exception description field of the corresponding resource knowledge unit. When the same resource knowledge unit corresponds to multiple correction actions at the same time, the filling in missing preconditions is executed first, followed by the restoration of boundary constraints, and finally the filling in exception descriptions.

[0095] For records whose action category is "adjust resource weight," only the resource weight of the corresponding resource knowledge unit is updated, without rewriting its text fields. A corresponding candidate resource record is formed when this resource knowledge unit is written into the candidate resource set. After completing the field-level write-back, the post-processed content, post-processed version number, and processing source marker are regenerated, and the post-processed content is re-performed in a consistency comparison with the corresponding semantic invariant record. Only when the corrected consistency score is not lower than the preset correction pass threshold of 0.70 is the resource knowledge unit written into the candidate resource set; otherwise, it remains isolated.

[0096] Each candidate resource record in the final output candidate resource set contains at least the resource identifier, knowledge point identifier, post-processing content, resource weight, post-processing version number, processing source marker, and the most recent consistency verification time. The candidate resource set is available for use by S430 and S510.

[0097] S430: Read the learner's current learning stage, historical answer records, and mastered knowledge records, and combine them with the candidate resource set to generate a push constraint set.

[0098] In practice, the learner's current learning stage is first encoded as a stage encoding field, the correct answer rate, number of consecutive errors and the most recent answer time in the historical answer records are extracted as answer status fields, and the knowledge point identifier and mastery level in the mastered knowledge records are extracted as mastery status fields.

[0099] Subsequently, using the knowledge point identifiers, resource weights, and disposal source markers from the candidate resource set as resource-side inputs, and the stage coding field, response status field, and mastery status field as learner-side inputs, the priority of candidate resources entering the personalized push queue is calculated. The push priority calculation formula is as follows:

[0100] in, Indicates the priority of pushing candidate resources; This indicates the degree of matching between the candidate resources and the learner's current learning stage and existing knowledge records; Represents the resource weight of the candidate resource; This indicates the degree to which candidate resources are repeatedly pushed within a preset time window; Indicates the semantic risk level of the candidate resource; , , , These represent the corresponding weights.

[0101] When implementing, it is possible to take , , When a candidate resource meets the preset queuing conditions and its push priority is not lower than the preset queuing threshold of 0.55, it is allowed to enter the subsequent queue orchestration set.

[0102] Based on this, a push constraint set is generated. This set includes at least the set of knowledge points allowed to enter the personalized push queue, the lower limit of the resource weight corresponding to each knowledge point, the minimum repeated push interval for the same resource, and a prohibition flag for high-risk resources. The push constraint set is not used to express teaching rules, but rather for direct use by the S510 to restrict the eligibility, order, and frequency of candidate resources entering the personalized push queue. The final push constraint set is then output for use by the S510.

[0103] S5 specifically includes the following sub-steps: S510 reads the candidate resource set output by S420 and the push constraint set output by S430, performs filtering, sorting and queue arrangement on the candidate resources, and generates a personalized push queue.

[0104] In practice, each candidate resource record in the candidate resource set is first used as the basic enqueue object, and each record is verified to meet the knowledge point adaptation constraint, stage matching constraint, resource weight lower limit constraint, and repeated push interval constraint in the push constraint set. For candidate resource records that do not meet any of the constraints, a prohibition enqueue mark is directly written and they are not included in the subsequent sorting process. For candidate resource records that meet the above constraints at the same time, they are written into the unsorted set.

[0105] Subsequently, an enqueue score is calculated for each candidate resource record in the set to be sorted. The enqueue score is calculated as follows:

[0106] in, Indicates joining the team and scoring points; This indicates the matching degree between the candidate resources and the learner's current learning status. The matching degree is calculated based on whether the knowledge point identifier is in the set of knowledge points that are allowed to enter the personalized push queue, whether the stage coding field matches, and whether the mastery level meets the reinforcement conditions. Indicates resource weight; This indicates the recent validity of a resource, and is used to characterize the validity of the candidate resource based on the most recent consistency check time and the most recent consistency check result. The repetition level is used to characterize the degree of content overlap between the candidate resource and the candidate resource records that have already entered the knowledge point sub-queue. , , , These represent the corresponding weights.

[0107] When implementing, it is possible to take , , , When the entry score is not lower than the preset entry threshold of 0.60, the corresponding candidate resource record is allowed to enter the knowledge point sub-queue; when there are multiple candidate resource records with the same source type and overlapping content after processing under the same knowledge point, only the candidate resource record with the highest entry score is retained to avoid duplicate push of homogeneous resources.

[0108] After the knowledge points are sorted internally, multiple sub-queues of knowledge points are alternately arranged according to their learning priority, so that candidate resources corresponding to preceding knowledge points are enqueued before candidate resources corresponding to subsequent knowledge points. For example, if the current learning stage corresponds to "basic learning", and "decomposition of forces" and "Newton's second law" both meet the enqueueing conditions, if the former is a preceding knowledge point of the latter, then the candidate resource corresponding to "decomposition of forces" will be written to the front of the personalized push queue first.

[0109] After the above processing, a personalized push queue is output, which is then available for S520 to use.

[0110] S520. Push the target teaching resources to the learner's terminal according to the personalized push queue, and collect the learner's subsequent answer results, dwell behavior and error correction feedback on the target teaching resources to generate a resource effect feedback set.

[0111] In practice, the target teaching resources are sent to the learner's terminal one by one according to the queue order in the personalized push queue, and a push timestamp is written when the target teaching resources enter the terminal display interface.

[0112] For each pushed target teaching resource, we continuously collect its corresponding reading completion status, answer results, dwell time, number of times it is opened repeatedly, error correction feedback mark, and feedback time. Among them, the reading completion status indicates whether the learner has completed browsing the main content of the target teaching resource; the answer results indicate the learner's correctness in the related exercises of the target teaching resource; the dwell time indicates the continuous time the learner stays on the target teaching resource interface; the number of times it is opened repeatedly indicates the number of times the learner accesses the same target teaching resource repeatedly within the preset observation window; and the error correction feedback mark indicates whether the learner or teacher has submitted content corrections for the target teaching resource.

[0113] For records where the dwell time is less than the preset minimum time and no response is triggered, an invalid feedback flag is written, and the record will not be considered as valid feedback in subsequent updates. For multiple opening behaviors of the same resource identifier within the same observation window, they are first merged into a single feedback record and then written into the resource effect feedback set to avoid repeatedly amplifying the feedback weight of a single resource.

[0114] The final output is a resource effect feedback set. Each feedback record in the resource effect feedback set includes at least the resource identifier, knowledge point identifier, push time, reading completion status, answer result, dwell time, number of times it is opened repeatedly, error correction feedback mark, feedback time and invalid feedback mark. The resource effect feedback set is available for S530 to call.

[0115] S530. Update the semantic invariant set and candidate resource set based on the resource effect feedback set. For candidate resources that still have semantic offsets after error correction feedback, regenerate resource disposal tasks. Increase the resource weight of candidate resources that have been verified by continuous and stable feedback. Then return the updated semantic invariant set and the updated candidate resource set to the processing links corresponding to S310 and S510.

[0116] In practice, the reliability of each feedback record in the resource effect feedback set is first calculated. The formula for calculating the reliability of feedback is:

[0117] in, Indicates the reliability of the feedback; It indicates the validity of the answer, and is used to characterize the degree to which the answer reflects the mastery of the knowledge points; This indicates the effectiveness of dwell time behavior, used to characterize how well dwell time and the number of times the page is repeatedly opened reflect the completeness of resource reading; Indicates the effectiveness of error correction feedback, used to characterize whether the error correction feedback points to a specific field error or semantic deviation; , , These represent the corresponding weights.

[0118] When implementing, it is possible to take , , When the feedback reliability is not lower than the preset feedback threshold of 0.55 and the corresponding feedback record has not been marked as invalid, the feedback record is determined as a valid feedback record; feedback records below the preset feedback threshold are only marked as observation records and are not included in the update calculation.

[0119] Subsequently, valid feedback records are merged according to knowledge point identifiers and resource identifiers: when multiple valid feedback records under the same knowledge point continuously point to misunderstanding of the same field, or continuously generate error correction feedback for the same field, only the corresponding field in the semantic invariant set is updated, and the entire semantic invariant record is not rewritten. When the same candidate resource record receives consistently stable and valid feedback records without any error correction feedback flags, its resource weight is increased; when the same candidate resource record receives consistently high-reliability error correction feedback, or when the response results are consistently abnormal and the corresponding fields are inconsistent with the semantic invariant records, its resource weight is decreased. Specifically, the resource weight adjustment is performed using an exponential smoothing update logic: Updated resource weights:

[0120] Where t represents the current feedback update round, As the current resource weight, To ensure the reliability of the feedback, This is a smoothing coefficient (e.g., a value of 0.8). To adjust the direction factor; when the boosting weight rule is triggered. When the weight reduction rule is triggered This allows for smooth, dynamic, and adaptive adjustment of weights based on feedback from the learner group. Simultaneously, for candidate resources that still exhibit semantic shifts as confirmed by error correction feedback, a new resource disposal task is generated, written into the resource disposal task set, and returned to S410.

[0121] After the update is completed, the updated semantic invariant set is written to the semantic invariant set storage area for subsequent calls by S310 and S420; the updated candidate resource set is written to the candidate resource set storage area for subsequent calls by S430 and S510. In this way, resource consistency verification, resource disposal, resource push, and feedback update form a continuous closed loop.

[0122] Example 2: Figure 2 As shown, this embodiment provides an intelligent teaching resource integration and personalized push system, including: The resource knowledge unit generation module is used to acquire teaching resources to be integrated, perform format unification, text transcription, paragraph segmentation and field extraction on the teaching resources to be integrated, extract knowledge point identifiers, explanation content, scope of application, preconditions and exceptions, and generate a set of resource knowledge units. The semantic invariant generation module is used to perform grouping, semantic alignment and consistency verification based on the resource knowledge unit set according to the knowledge point identifier, extract core propositions, boundary constraints, preconditions and exceptions, and generate a set of semantic invariants. The semantic drift identification module is used to align the resource knowledge unit set with the semantic invariant set, identify missing conditions, meaning shifts, weakened boundaries and omissions of exceptions, and generate a semantic drift record set and a resource disposal task set. The resource disposal and constraint generation module is used to perform isolation, correction or weight reduction processing on the corresponding resource knowledge units based on the resource disposal task set, semantic drift record set and imitation learning resource disposal model trained based on historical resource correction sample set, and combine the learner's current learning stage, historical answer record and mastered knowledge record to generate candidate resource set and push constraint set; The personalized push and feedback update module is used to generate a personalized push queue based on the candidate resource set and the push constraint set, push the target teaching resources to the learner's terminal, collect the resource effect feedback set, and update the semantic invariant set and the candidate resource set based on the resource effect feedback set.

[0123] All the above formulas are performed using dimensionless numerical calculations; the relevant formulas are based on empirical models that approximate the real situation, obtained through extensive data collection and software simulation fitting. The preset parameters and thresholds involved in the formulas can be conventionally set and adjusted by those skilled in the art according to the physical constraints of the actual application scenario.

[0124] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0125] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0126] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent integration and personalized delivery of teaching resources, characterized in that, Includes the following steps: S1. Obtain the teaching resources to be integrated, perform format unification, text transcription, paragraph segmentation and field extraction on the teaching resources to be integrated, extract knowledge point identifiers, explanation content, scope of application, preconditions and exceptions, and generate a set of resource knowledge units. S2. Based on the resource knowledge unit set, perform grouping, semantic alignment and consistency verification according to knowledge point identifiers, extract core propositions, boundary constraints, preconditions and exceptions, and generate a set of semantic invariants; S3. Align the resource knowledge unit set with the semantic invariant set by field, identify missing conditions, meaning shifts, weakened boundaries and omissions of exceptions, and generate a semantic drift record set and a resource disposal task set. S4. Based on the resource disposal task set, semantic drift record set, and the imitation learning resource disposal model trained on the historical resource correction sample set, the corresponding resource knowledge units are isolated, corrected, or downweighted. Combined with the learner's current learning stage, historical answer records, and mastered knowledge records, a candidate resource set and a push constraint set are generated.

2. The method for intelligent teaching resource integration and personalized delivery according to claim 1, characterized in that, Also includes: S5. Generate a personalized push queue based on the candidate resource set and the push constraint set, push the target teaching resources to the learner's terminal, collect the resource effect feedback set, and update the semantic invariant set and the candidate resource set based on the resource effect feedback set.

3. The method for intelligent teaching resource integration and personalized delivery according to claim 1, characterized in that, S1 specifically includes: Acquire teaching resources to be integrated from textbooks, courseware, question bank analysis, micro-lecture video texts, classroom handouts, and online teaching documents; perform format unification, version identification completion, and text transcription processing on each teaching resource to be integrated to generate a standardized resource record set; Based on a standardized resource record set, paragraph segmentation, heading alignment, and cross-modal content alignment are performed to generate a set of candidate knowledge units that retain resource identifiers, fragment positions, original descriptions, and adjacent contexts. Based on the candidate fragment set of knowledge units, extract knowledge point identifiers, explanation content, scope of application, preconditions and exceptions, delete invalid candidate fragments that are missing knowledge point identifiers or explanation content, and generate a resource knowledge unit set.

4. The method for intelligent teaching resource integration and personalized delivery according to claim 1, characterized in that, S2 specifically includes: Read the resource knowledge unit set, directly group resource knowledge units with the same standard knowledge point identifier, and perform grouping similarity calculation on resource knowledge units with temporary knowledge point identifier to generate a set of knowledge units with the same point. Semantic alignment is performed on each knowledge unit group in the same knowledge unit group set according to the core proposition field, boundary constraint field, precondition field and exception description field. Repeated and semantically stable fragments are extracted to generate a semantic skeleton candidate set.

5. The method for intelligent teaching resource integration and personalized delivery according to claim 4, characterized in that, Also includes: Consistency checks are performed on the semantic skeleton candidate set. Candidate segments that remain stable across multiple source resources and are not negated by conflicting representations are retained, while candidate segments that cannot be supported by other resources of the same point are deleted, and a set of semantic invariants is generated.

6. The method for intelligent teaching resource integration and personalized delivery according to claim 1, characterized in that, S3 specifically includes: Read the resource knowledge unit set and semantic invariant set, align each resource knowledge unit with the semantic invariant record under the same knowledge point identifier, calculate the completeness of the explanation content, the boundary retention, the consistency of the preconditions, and the retention of the exception description, and generate a semantic difference feature set by combining the strength of the field missing marker and the conflict marker. Based on the semantic difference feature set, conditions are missing, meaning shifts, boundary weakening and exception omissions are identified, the main drift type and the additional drift type are determined, and a semantic drift record set is generated; Based on the semantic drift record set, corresponding isolation tasks, correction tasks, or de-weighting tasks are generated according to the drift intensity, main drift type, and resource exposure, forming a resource disposal task set.

7. The method for intelligent teaching resource integration and personalized delivery according to claim 1, characterized in that, S4 specifically includes: Read the resource disposal task set, semantic drift record set, and historical resource correction sample set; train the imitation learning resource disposal model based on the historical manual disposal trajectory; and output the resource disposal action sequence based on the imitation learning resource disposal model. Based on the sequence of resource disposal actions, the corresponding resource knowledge units are isolated, corrected, or downgraded. The correction process includes at least adding missing preconditions, restoring boundary constraints, and adding exception descriptions. Consistency checks are then performed on the processed resource knowledge units to generate a candidate resource set.

8. The method for intelligent teaching resource integration and personalized push according to claim 7, characterized in that, Also includes: Read the learner's current learning stage, historical answer records, and mastered knowledge records, and combine them with the candidate resource set to generate a set of push constraints to restrict the eligibility, order, and frequency of candidate resources entering the personalized push queue.

9. The method for intelligent teaching resource integration and personalized delivery according to claim 1, characterized in that, S5 specifically includes: Read the candidate resource set and push constraint set, perform filtering, sorting and queue orchestration on the candidate resources, and generate a personalized push queue; The target teaching resources are pushed to learners' terminals according to the personalized push queue, and the learners' reading completion status, answer results, dwell time, number of times they open the target teaching resources and error correction feedback are collected to generate a resource effect feedback set. The reliability of feedback is calculated based on the resource effect feedback set, the valid feedback records are determined, and the field update, resource weight adjustment and resource disposal tasks are written back to the semantic invariant set and candidate resource set based on the valid feedback records.

10. An intelligent teaching resource integration and personalized push system, employing the intelligent teaching resource integration and personalized push method according to any one of claims 1 to 9, characterized in that, include: The resource knowledge unit generation module is used to acquire teaching resources to be integrated, and to perform format unification, text transcription, paragraph segmentation and field extraction on the teaching resources to be integrated, and to extract knowledge point identifiers, explanation content, scope of application, preconditions and exceptions. The semantic invariant generation module is used to perform grouping, semantic alignment, and consistency verification based on the resource knowledge unit set according to the knowledge point identifier; The semantic drift recognition module is used to align the resource knowledge unit set with the semantic invariant set, and to identify missing conditions, meaning shifts, weakened boundaries, and omissions of exceptions. The resource disposal and constraint generation module is used to learn the resource disposal model based on the resource disposal task set, the semantic drift record set, and the imitation learning resource disposal model trained based on the historical resource correction sample set. The personalized push and feedback update module is used to generate a personalized push queue based on the candidate resource set and the push constraint set, and push the target teaching resources to the learner's terminal.

Images