A deep learning-based course data intelligent multi-dimensional retrieval system

Abstract

The application discloses a kind of course data intelligent multidimensional retrieval system based on deep learning, it is related to deep learning information processing and educational data retrieval technical field, including: retrieval pedestal module, for forming course retrieval pedestal;Evidence demarcation module, for forming course evidence section set;Evidence sequence module, for obtaining course evidence token sequence set;Query sequence module, for receiving user course retrieval sentence, and focus on retrieval is arranged as query token sequence;ColBERT retrieval module, for inputting query token sequence and course evidence token sequence set into improved ColBERT model, form course evidence matching score set;Target evidence module, for forming target course evidence set;Result generation module, for generating course data intelligent multidimensional retrieval result.The application improves course evidence positioning accuracy, and supports multidimensional course retrieval for learning state.

Description

Technical Field

[0001] This invention relates to the field of deep learning information processing and educational data retrieval technology, and in particular to a deep learning-based intelligent multidimensional retrieval system for course data. Background Technology

[0002] With the development of online course platforms, smart teaching systems, and corporate training resource databases, course resources are typically stored in various forms, such as course catalogs, video-transcribed text, courseware text, exercise solutions, knowledge point configurations, and resource location indexes. Existing course retrieval methods largely rely on keyword inverted indexes, title matching, or ordinary semantic vector retrieval, which can achieve basic resource recall. However, when faced with search statements containing question descriptions, knowledge point associations, or specific learning tasks, problems arise such as hit content only containing keywords, evidence segments spanning teaching units, and inaccurate resource location back-indexing. While ordinary neural retrieval models can improve semantic matching capabilities, they typically treat course segments as ordinary text, lacking the coordinated utilization of course levels, teaching unit boundaries, and resource locations. Late Interaction retrieval structures, represented by the ColBERT model, can preserve token-level matching responses, but the original single-path token projection and single-peak MaxSim aggregation methods primarily focus on the strongest matching response, easily weakening the scattered derivation conditions, question conditions, and conclusions within the same course's evidence segment.

[0003] In practical teaching applications, user search objects also include question analysis, learning bottlenecks, learning evidence with similar cognitive levels, teaching materials that match interests and preferences, and course content corresponding to specific learning progress. Existing course retrieval typically only performs keyword matching and semantic recall on the course text or question text. Cognitive level, interests and preferences, learning bottlenecks, and learning progress are mostly stored as display fields in the business database, without being encoded into the evidence sequence and query sequence. This makes it difficult for the system to screen course evidence according to learning status conditions, and also makes it difficult to uniformly return course location and learning dimension information in the same search results.

[0004] Therefore, how to provide an intelligent multidimensional retrieval system for course data based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent multidimensional retrieval system for course data based on deep learning. This invention organizes course content, course structure, and course resource locations into searchable course evidence segments through a course retrieval foundation. These course evidence segments retain course level positions, content boundary positions, and course resource location indexes, enabling retrieval results to be routed back to specific courses, chapters, teaching units, and resource locations, thus improving the accuracy of course retrieval results. Simultaneously, a hierarchical residual projection layer is set at the encoder output of the ColBERT model, mapping the hidden state of course evidence tokens, course level positions, and content boundary positions together into hierarchical residual retrieval vectors. This allows course structure information and content boundary information to participate in the vector representation process of the deep retrieval model. Furthermore, a main peak-residual peak interaction aggregation structure replaces the single-peak MaxSim aggregation path in the Late Interaction layer, incorporating the highest matching response and the second-highest matching response into the course evidence matching score, while retaining the hit token position. This improves the recall effect of scattered teaching content, the accuracy of hit position labeling, and the traceability of course retrieval results.

[0006] According to an embodiment of the present invention, a deep learning-based intelligent multidimensional retrieval system for course data includes: The retrieval base module is used to obtain course retrieval related data, organize course content, course structure, course resource location, and learning retrieval dimensions related to course content, and form a course retrieval base. The evidence delineation module is used to delineate course evidence segments with pedagogical meaning boundaries in the course retrieval base, record the course level position and content boundary position of the course evidence segment, and establish a connection between the course evidence segment and the corresponding learning retrieval dimension to form a set of course evidence segments. The evidence sequence module is used to perform retrieval granular compression on the course evidence segments in the course evidence segment set, and organize the continuously unfolded teaching content within the course evidence segment into a set of course evidence token sequences carrying course level bits, content boundary bits and learning retrieval dimension markers; The query sequence module receives the user's course search statement and search filter conditions, identifies the search focus from the user's course search statement, and marks and organizes the query dimensions corresponding to the search focus and search filter conditions into a query token sequence. The ColBERT retrieval module is used to input the query token sequence and the course evidence token sequence set into the improved ColBERT model. It replaces the single token projection path at the encoder output with a hierarchical residual projection layer and replaces the single-peak MaxSim aggregation path with a main peak-residual peak interactive aggregation structure to form a course evidence matching score set. The target evidence module is used to screen candidate course evidence segments according to the course evidence matching score set and query dimension label, and to back-row the candidate course evidence segments to the corresponding course level position and content boundary position to form the target course evidence set. The results generation module is used to organize the retrieval dimensions, learning retrieval dimensions, hit locations, and course resource locations corresponding to the evidence set of the target course, and generate intelligent multidimensional retrieval results for course data.

[0007] Optionally, the course retrieval associated data includes course content data, course structure data, course resource location data, and learning status associated data; Course content data corresponds to the course text, question content, question analysis, step-by-step guidance content or teaching material description; course structure data corresponds to the hierarchical relationship between courses, chapters, teaching units and knowledge points and the order of teaching development; course resource location data corresponds to the source position, page position, time position, paragraph position or question position of the teaching expression content in the course resources. The learning status-related data includes cognitive level tags, interest preference tags, learning checkpoint tags, and learning progress tags. The learning retrieval condition data consists of at least one of the retrieval filtering conditions or learning status-related data.

[0008] Optionally, the formation of the course retrieval platform specifically includes: The retrieval base module reads the course catalog, chapter hierarchy, and teaching unit sequence from the course structure data, and establishes a course hierarchy framework composed of course positions, chapter positions, and teaching unit positions. Identify title boundaries, paragraph boundaries, or question number boundaries in the course content data, organize continuous content within the same boundary interval into content segments, and place the content segments into the corresponding teaching unit positions; Write the course resource location data into the content segment to form a course retrieval base that includes a course hierarchy framework, content segments, content boundary positions, and resource location indexes.

[0009] Optionally, the evidence delineation module specifically comprises: The evidence delineation module uses the teaching unit corresponding to the course structure as the delineation scope in the course retrieval base, and scans continuous content sequentially along the course content within the teaching unit. Course content that is within the same teaching unit, has continuous content boundaries, and has continuous resource locations is classified into the same candidate evidence segment, and the candidate evidence segment is truncated at the point of teaching unit switch, content boundary interruption, or resource location jump. The candidate evidence segments are identified as course evidence segments, the teaching unit positions corresponding to the course evidence segments are written as course level positions, and the start and end boundaries of the course evidence segments are written as content boundary positions, thus forming a set of course evidence segments.

[0010] Optionally, the evidence sequence module specifically comprises: The evidence sequence module processes course evidence segments as units, reads the continuously unfolding teaching content within a course evidence segment, and retains the course level position and content boundary position corresponding to the course evidence segment. The teaching content is segmented into search terms, and whitespace, format control characters and duplicate delimiters are removed. The search terms are arranged according to the order of the teaching content in the course evidence segment to form a course evidence term sequence. The WordPiece word segmentation mapping method, which shares a vocabulary with the document encoder of the improved ColBERT model, is adopted. The hierarchical position markers, boundary position markers, and learning retrieval dimension markers associated with the course evidence segments are written into the course evidence term sequence and converted into a course evidence token sequence to form a set of course evidence token sequences. The learning retrieval dimension markers include cognitive level markers, interest preference markers, learning checkpoint markers, or learning progress markers. Course evidence segments not associated with the corresponding learning retrieval dimension will not be written with the corresponding markers.

[0011] Optionally, the query sequence module specifically includes: Receive user course search statements and search filtering conditions, clean up whitespace, format control characters and duplicate delimiters in user course search statements, and retain search terms related to course content, knowledge points, resource locations and problem descriptions; The search focus is determined by the continuity of terms in the user's course search statement, and consecutively appearing course terms, knowledge point terms, or question description terms are grouped into the same search focus segment. Convert the cognitive level conditions, interest preference conditions, learning bottleneck conditions, or learning progress conditions in the search filtering conditions into query dimension tags, and arrange the query dimension tags before the corresponding search focus segment; We adopt the WordPiece sub-word segmentation mapping method, which shares a vocabulary with the query encoder of the improved ColBERT model, to convert the retrieval focus segment and query dimension tags into a query token sequence, and retain the start and end positions of the retrieval focus segment in the query token sequence.

[0012] Optionally, the improved ColBERT model includes a query encoder, a document encoder, a hierarchical residual projection layer, a main peak-residual peak interaction aggregation structure, and a scoring output layer; The hierarchical residual projection layer includes a semantic projection branch and a hierarchical residual projection branch; The query encoder receives a sequence of query tokens and outputs the hidden state of the query tokens; the document encoder receives a set of course evidence token sequences and outputs the hidden state of the course evidence tokens. The hierarchical residual projection layer performs semantic projection on the query token hidden state and the course evidence token hidden state to form the query semantic retrieval vector and the course evidence semantic retrieval vector, and maps the course evidence token hidden state, course level bit and content boundary bit together into the hierarchical residual retrieval vector. The main peak-residual peak interaction aggregation structure constructs a semantic matching response using query semantic retrieval vectors and course evidence semantic retrieval vectors, and constructs a hierarchical residual matching response using query semantic retrieval vectors and hierarchical residual retrieval vectors; the main peak matching branch extracts the highest matching response corresponding to the query token from the semantic matching response, and the residual peak matching branch compresses the secondary matching response corresponding to the query token from the hierarchical residual matching response. The scoring output layer synthesizes the highest matching response and the second-highest matching response to obtain the course evidence matching score corresponding to each course evidence token sequence. It retains the course evidence token positions corresponding to the highest matching response and the second-highest matching response and summarizes them to form a course evidence matching score set.

[0013] Optionally, the hierarchical residual projection layer performs semantic projection on the query token hidden state and the course evidence token hidden state to form a query semantic retrieval vector and a course evidence semantic retrieval vector, and maps the course evidence token hidden state, course level bits, and content boundary bits together into a hierarchical residual retrieval vector, specifically: The semantic projection branch uses a trainable semantic projection matrix to perform linear projection and normalization processing on the query token hidden state and the course evidence token hidden state, respectively, to form the query semantic retrieval vector and the course evidence semantic retrieval vector. The hierarchical residual projection branch reads the course level bit and content boundary bit corresponding to the hidden state of the course evidence token. Within the same content boundary bit limit, it performs local difference deduction on the hidden states of adjacent course evidence tokens to form the local residual state of the course evidence. The hierarchical residual projection branch uses a trainable position embedding table to map course hierarchical positions and content boundary positions into position embeddings, and concatenates the local residual states of course evidence with the position embeddings in dimensional order to form a residual position combination vector. The hierarchical residual projection branch uses a trainable residual projection matrix to perform linear projection and normalization on the residual position combination vector to form a hierarchical residual retrieval vector. The trainable semantic projection matrix, trainable position embedding table, and trainable residual projection matrix are all trainable parameters in the improved ColBERT model.

[0014] Optionally, the target evidence module specifically comprises: The target evidence module writes back the course evidence matching score to the corresponding course evidence segment according to the correspondence between the course evidence matching score and the course evidence token sequence in the course evidence matching score set; When a query token sequence carries one or more query dimension tags, check the query dimension tags against the learning retrieval dimension tags in the course evidence token sequence item by item according to the tag type. If each query dimension tag has a learning retrieval dimension tag with the same tag type and tag number, retain the corresponding course evidence segment. If the query token sequence does not carry query dimension tags, retain all course evidence segments that have been written to the course evidence matching score. The retained course evidence segments are sorted by descending score, and the course evidence segments with the highest scores are selected to form candidate course evidence segments. Perform duplicate boundary checks on candidate course evidence segments, and retain the candidate course evidence segment with the highest course evidence matching score that has the same course level position and overlapping content boundary positions. The candidate course evidence segments that have undergone repeated boundary checks are back rowed to the corresponding course level position and content boundary position in the course retrieval base to form the target course evidence set.

[0015] Optionally, the result generation module specifically includes: Read the course level position, content boundary position, course evidence matching score, course resource location and learning retrieval dimension markers corresponding to each target course evidence segment in the target course evidence set; Organize the courses, chapters or teaching units corresponding to the course level into course search dimension labels, and organize the learning search dimension labels into cognitive level labels, interest preference labels, learning checkpoint labels or learning progress labels. In the evidence segment of the target course, mark the start and end positions of the hit that match the query token sequence to form a hit position label; The course retrieval dimension label, learning retrieval dimension label, hit location label, course resource location, and course evidence matching score are written into the same retrieval result item, and the retrieval result items are arranged according to the course evidence matching score to generate intelligent multidimensional retrieval results for course data.

[0016] The beneficial effects of this invention are: 1. This invention places course content into course positions, chapter positions, and teaching unit positions through a course retrieval base, and simultaneously records content boundary positions and course resource location indexes. This ensures that course evidence segments maintain clear course affiliation, teaching boundaries, and resource locations during the delineation, retrieval, back-sorting, and result output processes, thereby improving the positioning accuracy and traceability of course retrieval results.

[0017] 2. This invention converts cognitive level, interest preference, learning bottleneck and learning progress into learning retrieval dimension tags and writes them into the course evidence token sequence. It also converts user search filtering conditions into query dimension tags. By verifying the tag type and tag number, the course evidence segment is selected. This allows course retrieval to combine the semantics of teaching content and the user's learning status at the same time, reducing the problem of returning unsuitable teaching content by relying solely on keywords or ordinary semantic similarity.

[0018] 3. This invention sets up a hierarchical residual projection layer at the encoder output of the ColBERT model and replaces the single-peak MaxSim aggregation path with a main peak-residual peak interactive aggregation structure, so that the course level, content boundary, highest semantic hit and secondary residual hit within the boundary jointly participate in course evidence matching, thereby improving the recall stability and hit position annotation accuracy of the dispersed teaching evidence. Attached Figure Description

[0019] 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: Figure 1 This is a structural diagram of a deep learning-based intelligent multidimensional retrieval system for course data proposed in this invention. Figure 2 This invention presents an improved ColBERT model structure and matching scoring flowchart for a deep learning-based intelligent multidimensional retrieval system for course data. Detailed Implementation

[0020] 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.

[0021] refer to Figures 1-2 A deep learning-based intelligent multidimensional retrieval system for course data includes: The retrieval base module is used to obtain course retrieval related data, organize course content, course structure, course resource location, and learning retrieval dimensions related to course content, and form a course retrieval base. The evidence delineation module is used to delineate course evidence segments with pedagogical meaning boundaries in the course retrieval base, record the course level position and content boundary position of the course evidence segment, and establish a connection between the course evidence segment and the corresponding learning retrieval dimension to form a set of course evidence segments. The evidence sequence module is used to perform retrieval granular compression on the course evidence segments in the course evidence segment set, and organize the continuously unfolded teaching content within the course evidence segment into a set of course evidence token sequences carrying course level bits, content boundary bits and learning retrieval dimension markers; The query sequence module receives the user's course search statement and search filter conditions, identifies the search focus from the user's course search statement, and marks and organizes the query dimensions corresponding to the search focus and search filter conditions into a query token sequence. The ColBERT retrieval module is used to input the query token sequence and the course evidence token sequence set into the improved ColBERT model. It replaces the single token projection path at the encoder output with a hierarchical residual projection layer and replaces the single-peak MaxSim aggregation path with a main peak-residual peak interactive aggregation structure to form a course evidence matching score set. The target evidence module is used to screen candidate course evidence segments according to the course evidence matching score set and query dimension label, and to back-row the candidate course evidence segments to the corresponding course level position and content boundary position to form the target course evidence set. The results generation module is used to organize the retrieval dimensions, learning retrieval dimensions, hit locations, and course resource locations corresponding to the evidence set of the target course, and generate intelligent multidimensional retrieval results for course data.

[0022] In this embodiment, the course retrieval associated data includes course content data, course structure data, course resource location data, and learning status associated data; Course content data corresponds to the course text, question content, question analysis, step-by-step guidance content or teaching material description; course structure data corresponds to the hierarchical relationship between courses, chapters, teaching units and knowledge points and the order of teaching development; course resource location data corresponds to the source position, page position, time position, paragraph position or question position of the teaching expression content in the course resources. The learning status-related data includes cognitive level tags, interest preference tags, learning checkpoint tags, and learning progress tags. The learning retrieval condition data consists of at least one of the retrieval filtering conditions or learning status-related data.

[0023] In the specific implementation process, course retrieval data is collected separately according to course resource data and learning status data. The course content data, including the course text, question content, question analysis, step-by-step guidance content, and teaching material descriptions, retains the corresponding course number, chapter number, teaching unit number, knowledge point number, and resource number, respectively. The course structure data records the hierarchical relationships between courses, chapters, teaching units, and knowledge points, as well as the order of teaching. The course resource location data records the source position, page position, time position, paragraph position, or question position of each teaching content in the courseware, video, question bank, or teaching material library.

[0024] Learning status data is obtained from user exam parameters, answer records, interest tag records, problem-solving bottleneck records, and learning training records. Cognitive level tags include deep cognitive tags and shallow cognitive tags. Deep cognitive tags include the ability to ask questions, distinguish, unify, and create; shallow cognitive tags include the ability to memorize, connect, unify, master, and teach. When a user has answer records in both assisted and unassisted modes, the difference in assisted accuracy rate ΔA and the assisted completion quality ratio ΔQ are retained. The assisted completion quality ratio is the ratio of the completion quality in assisted mode to the completion quality in unassisted mode. When ΔA is greater than 25 percentage points and ΔQ is greater than 1.4, a pseudo-shallow tag is written. Interest preference tags are obtained from user input or user interest tag records. Learning bottleneck tags are obtained from bottleneck records corresponding to answering timeouts, no answers, consecutive errors, or interruptions in problem-solving steps. Learning progress tags are obtained from convergence progress, the set of mastered question types, and the set of question types to be mastered.

[0025] When course content data has been configured with cognitive levels, interest themes, learning checkpoint types, or applicable learning progress, the corresponding tags, along with the course number, teaching unit number, knowledge point number, or question number, are written into the course retrieval base. When course content data has not been configured with a certain type of tag, the corresponding tag field remains empty. The cognitive level tags, interest preference tags, learning checkpoint tags, and learning progress tags in the learning status-related data serve as user-side learning retrieval conditions, while the corresponding tags in the course content data serve as course evidence-side learning retrieval dimensions. Both sides use the same tag type and tag number for recording.

[0026] When a user selects cognitive level, interest preference, learning bottleneck, or learning progress as filter conditions in the search interface, the selected tags are directly used to form the learning search conditions. If the user does not select any search filter conditions, the learning status tags already recorded for the current user under the corresponding knowledge point or question type are read. If no corresponding learning status tag exists, this type of learning search condition is not generated. The user's course search statement is entered into the query sequence module as query text, and the learning search condition data is converted into query dimension tags separately.

[0027] In this embodiment, a course retrieval base is formed, specifically as follows: The retrieval base module reads the course catalog, chapter hierarchy, and teaching unit sequence from the course structure data, and establishes a course hierarchy framework composed of course positions, chapter positions, and teaching unit positions. Identify title boundaries, paragraph boundaries, or question number boundaries in the course content data, organize continuous content within the same boundary interval into content segments, and place the content segments into the corresponding teaching unit positions; Write the course resource location data into the content segment to form a course retrieval base that includes a course hierarchy framework, content segments, content boundary positions, and resource location indexes.

[0028] In the specific implementation process, after receiving the course retrieval association data, the retrieval base module configures location numbers for courses, chapters, and teaching units respectively. Course positions are represented by a course identifier, chapter positions by a combination of the course identifier and chapter number, and teaching unit positions by a combination of the chapter number and teaching unit number. When the course catalog has multiple levels of chapters, the lowest-level chapter serves as the content placement level, and the preceding chapters are written with the prefix information of their chapter positions. The order of teaching units prioritizes the arrangement already configured in the course structure data; if the course structure data does not configure the order of teaching units, the course content within the same chapter position is arranged according to the order in which the course resources appear.

[0029] Before course content data enters the course hierarchy framework, boundary identification is performed according to resource type. For page-type resources, priority is given to identifying title fields, title numbers, and page number changes; for text-type resources, priority is given to identifying title fields, paragraph numbers, line break marks, and paragraph ending punctuation; for audio / video-type resources, priority is given to identifying subtitle timestamps and title timestamps in the transcribed text; for question-type resources, priority is given to identifying question number markers, the starting position of the question stem, and the starting position of the explanation. Continuous content between the same question number and the next question number is organized into question intervals; continuous content between the same title and the next title is organized into title intervals; continuous text without question numbers or titles is organized into content segments according to paragraph boundaries.

[0030] When a content segment is incorporated into the course hierarchy framework, the retrieval base module verifies the correspondence between the content segment and the teaching unit position. Content segments whose course structure data already specifies their teaching unit or knowledge point affiliation are placed in the corresponding teaching unit position. Content segments whose course structure data does not specify affiliation are searched for backwards within the same chapter position, following the order of course resource appearance, and placed in the nearest teaching unit position. When a content segment covers two teaching unit positions, the content segment is split into two segments using the title boundary, time position, page position, or question number boundary corresponding to the teaching unit switching position as the dividing point, and the start and end boundaries are recorded for each segment.

[0031] When writing resource locations to the content segment, page-type resources are written with the source and page positions, audio / video resources with the source and time positions, text resources with the source and paragraph positions, and question-type resources with the source and question positions. After writing is complete, the content segment saves the course position, chapter position, teaching unit position, start boundary, end boundary, and resource location index. The course position, chapter position, and teaching unit position constitute the course hierarchy positions, the start and end boundaries constitute the content boundary positions, and the resource location index participates in the course evidence segment back-sorting and retrieval result location.

[0032] In this embodiment, the evidence delineation module specifically comprises: The evidence delineation module uses the teaching unit corresponding to the course structure as the delineation scope in the course retrieval base, and scans continuous content sequentially along the course content within the teaching unit. Course content that is within the same teaching unit, has continuous content boundaries, and has continuous resource locations is classified into the same candidate evidence segment, and the candidate evidence segment is truncated at the point of teaching unit switch, content boundary interruption, or resource location jump. The candidate evidence segments are identified as course evidence segments, the teaching unit positions corresponding to the course evidence segments are written as course level positions, and the start and end boundaries of the course evidence segments are written as content boundary positions, thus forming a set of course evidence segments.

[0033] In practice, the evidence delineation module uses the teaching unit position in the course retrieval base as the boundary for defining evidence segments. Content segments within the same teaching unit position are arranged in the order of their starting content boundary positions. Each content segment carries a course level position, a starting boundary, an ending boundary, and a resource location index. The scanning starts from the content segment with the highest position within the same teaching unit position, taking the first content segment as the starting content segment of the current candidate evidence segment, and then sequentially checking the course level position, content boundary position, and resource location index between adjacent content segments.

[0034] If adjacent content segments have the same course level, the termination boundary of the preceding content segment is adjacent to the start boundary of the following content segment, and the resource location index remains continuous, the following content segment will be merged into the current candidate evidence segment. In page-type resources, if the page positions are the same or the page position of the following content segment is the next page of the preceding content segment, the resource location index is considered continuous. In audio / video resources, if the interval between the start time of the following content segment and the end time of the preceding content segment does not exceed 5 seconds, the resource location index is considered continuous. In text-type resources, if the paragraph positions are adjacent, the resource location index is considered continuous. In question-type resources, if the question positions are the same or the question numbers are consecutive, the resource location index is considered continuous.

[0035] When any of the following situations occur between adjacent content segments: a switch in teaching unit position, a non-connected content boundary position, or a discontinuous resource location index, the evidence delineation module truncates the current candidate evidence segment at the termination boundary of the previous content segment and starts delineating a new candidate evidence segment from the next content segment. After the candidate evidence segment is formed, the evidence delineation module checks whether it contains a complete title interval, paragraph interval, or question interval. If the candidate evidence segment does not cover a complete interval, the evidence delineation module fills in adjacent content segments along the content expansion sequence within the same teaching unit position until the candidate evidence segment covers a complete interval. The filling process is still subject to the continuity of teaching unit positions and resource locations.

[0036] Candidate evidence segments that have undergone truncation and complete interval checks are identified as course evidence segments. The course hierarchy position of a course evidence segment consists of a course position, a chapter position, and a teaching unit position. The content boundary position of a course evidence segment consists of the start boundary of the first content segment and the end boundary of the last content segment within the course evidence segment. The evidence delineation module writes the course evidence segments, course hierarchy positions, content boundary positions, and resource location indexes into the same evidence segment record, forming a set of course evidence segments.

[0037] In this embodiment, the evidence sequence module specifically includes: The evidence sequence module processes course evidence segments as units, reads the continuously unfolding teaching content within a course evidence segment, and retains the course level position and content boundary position corresponding to the course evidence segment. The teaching content is segmented into search terms, and whitespace, format control characters and duplicate delimiters are removed. The search terms are arranged according to the order of the teaching content in the course evidence segment to form a course evidence term sequence. The WordPiece word segmentation mapping method, which shares a vocabulary with the document encoder of the improved ColBERT model, is adopted. The hierarchical position markers, boundary position markers, and learning retrieval dimension markers associated with the course evidence segments are written into the course evidence term sequence and converted into a course evidence token sequence to form a set of course evidence token sequences. The learning retrieval dimension markers include cognitive level markers, interest preference markers, learning checkpoint markers, or learning progress markers. Course evidence segments not associated with the corresponding learning retrieval dimension will not be written with the corresponding markers.

[0038] In the specific implementation process, the evidence sequence module reads the course evidence-side learning retrieval dimensions from the course evidence segment records. The cognitive level tags corresponding to the course content are converted into cognitive level markers, the interest topics corresponding to the teaching materials are converted into interest preference markers, the checkpoint types corresponding to the question analysis or step-by-step guidance content are converted into learning checkpoint markers, and the applicable learning progress and question type status corresponding to the course content are converted into learning progress markers; duplicate markers of the same type are deleted, and different markers are retained separately.

[0039] The pseudo-shallow labels on the course evidence side are obtained from the applicable cognitive labels already configured in the course content; the pseudo-shallow labels on the user side are determined by the difference in assist accuracy ΔA between the assisted mode and the unassisted mode and the assist completion quality ratio ΔQ. Pseudo-shallow query labels are generated when ΔA is greater than 25 percentage points and ΔQ is greater than 1.4.

[0040] In this embodiment, the query sequence module specifically includes: Receive user course search statements and search filtering conditions, clean up whitespace, format control characters and duplicate delimiters in user course search statements, and retain search terms related to course content, knowledge points, resource locations and problem descriptions; The search focus is determined by the continuity of terms in the user's course search statement, and consecutively appearing course terms, knowledge point terms, or question description terms are grouped into the same search focus segment. Convert the cognitive level conditions, interest preference conditions, learning bottleneck conditions, or learning progress conditions in the search filtering conditions into query dimension tags, and arrange the query dimension tags before the corresponding search focus segment; We adopt the WordPiece sub-word segmentation mapping method, which shares a vocabulary with the query encoder of the improved ColBERT model, to convert the retrieval focus segment and query dimension tags into a query token sequence, and retain the start and end positions of the retrieval focus segment in the query token sequence.

[0041] In the specific implementation process, the query sequence module prioritizes reading the search filtering conditions selected by the user in the search interface; if the user has not selected search filtering conditions, it reads the learning status tags from the current user's cognitive assessment record, interest tag record, problem-solving checkpoint record, and learning training record. Deep cognition, shallow cognition, and pseudo-shallow conditions are converted into cognitive query tags, user interest tags are converted into interest query tags, checkpoint types are converted into checkpoint query tags, and convergence progress and the status of mastered and unmastered question types are converted into progress query tags.

[0042] Query dimension tags and course evidence-based learning retrieval dimension tags use the same tag type and tag number; duplicate conditions of the same type are deleted, while different conditions are retained. Each query dimension tag is arranged as a complete token in the shared thesaurus before the corresponding search focus segment, without performing sub-word splitting; if the corresponding search condition is not obtained, it is not written to the corresponding query dimension tag. Query dimension tags participate in query encoding and evidence matching, but are not included in the restoration of the search focus start position, end position, and hit content position.

[0043] In this embodiment, the improved ColBERT model includes a query encoder, a document encoder, a hierarchical residual projection layer, a main peak-residual peak interaction aggregation structure, and a scoring output layer. The hierarchical residual projection layer includes a semantic projection branch and a hierarchical residual projection branch; The query encoder receives a sequence of query tokens and outputs the hidden state of the query tokens; the document encoder receives a set of course evidence token sequences and outputs the hidden state of the course evidence tokens. The hierarchical residual projection layer performs semantic projection on the query token hidden state and the course evidence token hidden state to form the query semantic retrieval vector and the course evidence semantic retrieval vector, and maps the course evidence token hidden state, course level bit and content boundary bit together into the hierarchical residual retrieval vector. The main peak-residual peak interaction aggregation structure constructs a semantic matching response using query semantic retrieval vectors and course evidence semantic retrieval vectors, and constructs a hierarchical residual matching response using query semantic retrieval vectors and hierarchical residual retrieval vectors; the main peak matching branch extracts the highest matching response corresponding to the query token from the semantic matching response, and the residual peak matching branch compresses the secondary matching response corresponding to the query token from the hierarchical residual matching response. The scoring output layer synthesizes the highest matching response and the second-highest matching response to obtain the course evidence matching score corresponding to each course evidence token sequence. It retains the course evidence token positions corresponding to the highest matching response and the second-highest matching response and summarizes them to form a course evidence matching score set.

[0044] In this embodiment, the hierarchical residual projection layer performs semantic projection on the query token hidden state and the course evidence token hidden state to form a query semantic retrieval vector and a course evidence semantic retrieval vector. The course evidence token hidden state, course level bits, and content boundary bits are then mapped together into a hierarchical residual retrieval vector. Specifically: The semantic projection branch uses a trainable semantic projection matrix to perform linear projection and normalization processing on the query token hidden state and the course evidence token hidden state, respectively, to form the query semantic retrieval vector and the course evidence semantic retrieval vector. The hierarchical residual projection branch reads the course level bit and content boundary bit corresponding to the hidden state of the course evidence token. Within the same content boundary bit limit, it performs local difference deduction on the hidden states of adjacent course evidence tokens to form the local residual state of the course evidence. The hierarchical residual projection branch uses a trainable position embedding table to map course hierarchical positions and content boundary positions into position embeddings, and concatenates the local residual states of course evidence with the position embeddings in dimensional order to form a residual position combination vector. The hierarchical residual projection branch uses a trainable residual projection matrix to perform linear projection and normalization on the residual position combination vector to form a hierarchical residual retrieval vector. The trainable semantic projection matrix, trainable position embedding table, and trainable residual projection matrix are all trainable parameters in the improved ColBERT model.

[0045] In the specific implementation, the improved ColBERT model retains the query encoder, document encoder, and Late Interaction matching concepts of the original ColBERT model. The original ColBERT model uses a single-path token projection at the encoder output, mapping the latent states of query tokens and document tokens to single retrieval vectors respectively, and extracting the highest MaxSim response for each query token in the Late Interaction layer as the primary matching criterion. In course data retrieval, the same course evidence segment typically contains concept descriptions, derivation conditions, question conditions, and conclusions. A single-peak MaxSim response tends to retain only the strongest term hit, weakening the contribution of adjacent teaching content, hierarchical positions, and boundary positions within the same course evidence segment to the retrieval score. This implementation sets a hierarchical residual projection layer at the encoder output and replaces the single-peak MaxSim aggregation path in the Late Interaction layer with a main peak-residual peak interactive aggregation structure, allowing the highest semantic hit and boundary residual hits to both contribute to the course evidence matching score.

[0046] The improved ColBERT model includes a query encoder, a document encoder, a hierarchical residual projection layer, a peak-residual peak interaction aggregation structure, and a scoring output layer. The query encoder receives a sequence of query tokens and outputs the query token latent state. The document encoder receives a set of course evidence token sequences and outputs the course evidence token latent state. Both the query encoder and document encoder use a Transformer encoder with a shared vocabulary. The token latent state dimension is set to 768, and the retrieval vector dimension output by the hierarchical residual projection layer is set to 128. The start and end positions of the retrieval focus in the query token sequence are entered into the query encoder along with the query token. The course hierarchy bits and content boundary bits in the course evidence token sequence are entered into the document encoder along with the course evidence token. Each course evidence token within the same course evidence segment inherits the corresponding course hierarchy bit. The content boundary bits are written into the start and end token bits of the course evidence segment. Intermediate tokens located between the start and end token bits share the same boundary interval number.

[0047] The hierarchical residual projection layer comprises a semantic projection branch and a hierarchical residual projection branch. The semantic projection branch uses a trainable semantic projection matrix to process the query token latent state and the course evidence token latent state, respectively. For each query token latent state or course evidence token latent state, the semantic projection branch inputs the 768-dimensional token latent state into the trainable semantic projection matrix in dimensional order, and performs a weighted summation of the 768 input dimensions on each output dimension of the trainable semantic projection matrix to obtain a 128-dimensional semantic projection vector. Subsequently, the semantic projection branch performs L2 normalization on the 128-dimensional semantic projection vector. During L2 normalization, if the L2 norm of the 128-dimensional semantic projection vector is zero, the normalization result remains a zero vector; if the L2 norm of the 128-dimensional semantic projection vector is not zero, each component in the 128-dimensional semantic projection vector is divided by the L2 norm, so that the semantic projection vector corresponding to each token has a uniform vector scale, forming the query semantic retrieval vector and the course evidence semantic retrieval vector, respectively. The trainable semantic projection matrix has an input dimension of 768 and an output dimension of 128. The matrix parameters are used as trainable parameters in the improved ColBERT model during training. The course evidence semantic retrieval vector is generated offline from a set of course evidence token sequences and written into the retrieval index during the course resource storage phase. The query semantic retrieval vector is generated online from a sequence of query tokens after receiving the user's course retrieval statement.

[0048] The hierarchical residual projection branch receives the latent state of the course evidence token, the course hierarchy bit, and the content boundary bit. The hierarchical residual projection branch limits the local difference reduction range according to the content boundary bit. The hierarchy position marker, boundary position marker, and learning retrieval dimension marker do not participate in local difference reduction; only the latent state difference is calculated between adjacent teaching content tokens within the same boundary interval number. For course evidence tokens located within the boundary interval, the hierarchical residual projection branch calculates the difference between the current course evidence token's latent state and the mean of the previous and next adjacent course evidence token's latent states, forming the local residual state of the course evidence. For course evidence tokens located at the start or end token positions, only the latent states of adjacent course evidence tokens within the same boundary interval number are used for difference reduction. This processing ensures that the local residual state of the course evidence retains the fine-grained changes of adjacent teaching content within the course evidence segment, avoiding the introduction of irrelevant residual information across teaching boundaries.

[0049] The hierarchical residual projection branch uses a trainable position embedding table to process course hierarchical bits and content boundary bits. The trainable position embedding table includes a hierarchical position embedding area and a boundary position embedding area. The hierarchical position embedding area receives the course bit number, chapter bit number, and teaching unit bit number, while the boundary position embedding area receives the start token bit number, end token bit number, and boundary interval number. For each course evidence token in the same course evidence token sequence, the hierarchical residual projection branch reads the inherited course bit number, chapter bit number, and teaching unit bit number, and retrieves the course position embedding, chapter position embedding, and teaching unit position embedding respectively in the hierarchical position embedding area. The hierarchical residual projection branch performs a positional addition on the course position embedding, chapter position embedding, and teaching unit position embedding to obtain the hierarchical position embedding corresponding to the course evidence token. In this embodiment, the dimensions of the course position embedding, chapter position embedding, and teaching unit position embedding are all set to 64 dimensions, and the hierarchical position embedding after positional addition is still 64 dimensions.

[0050] The hierarchical residual projection branch reads the start token bit number, end token bit number, and boundary interval number corresponding to the course evidence token, and retrieves the start boundary embedding, end boundary embedding, and interval boundary embedding respectively in the boundary position embedding area. The start token bit number corresponds to the first token position of the course evidence segment in the course evidence token sequence, the end token bit number corresponds to the last token position of the course evidence segment in the course evidence token sequence, and the boundary interval number is used to mark the content boundary interval to which the course evidence token belongs. The hierarchical residual projection branch performs a 64-dimensional sum on the start boundary embedding, end boundary embedding, and interval boundary embedding to obtain the boundary position embedding corresponding to the course evidence token. The dimensions of the start boundary embedding, end boundary embedding, and interval boundary embedding are all set to 64-dimensional, and the boundary position embedding after 64-dimensional summation is still 64-dimensional.

[0051] The hierarchical residual projection branch concatenates the hierarchical position embedding and the boundary position embedding in dimensional order to obtain a 128-dimensional position embedding. Subsequently, it concatenates the local residual state of the course evidence token with the 128-dimensional position embedding in dimensional order. The dimension of the local residual state of the course evidence is the same as the dimension of the latent state of the course evidence token, set to 768 dimensions. After concatenation, an 896-dimensional residual position combination vector is obtained. This residual position combination vector simultaneously contains the local differences between the course evidence token and its adjacent tokens, the course hierarchical position, and the content boundary position.

[0052] After the residual position combination vector is fed into the trainable residual projection matrix, the hierarchical residual projection branch performs a weighted summation of the 896-dimensional residual position combination vector on each output dimension of the trainable residual projection matrix, resulting in a 128-dimensional residual projection vector. Subsequently, L2 normalization is performed on the 128-dimensional residual projection vector. If the L2 norm of the 128-dimensional residual projection vector is zero, the hierarchical residual retrieval vector is set to zero; if the L2 norm is not zero, each component in the 128-dimensional residual projection vector is divided by its L2 norm to form the hierarchical residual retrieval vector. The trainable residual projection matrix has an input dimension of 896 and an output dimension of 128, and its parameters are used as trainable parameters in the improved ColBERT model during training.

[0053] The main peak-residual peak interaction aggregation structure receives query semantic retrieval vectors, course evidence semantic retrieval vectors, and hierarchical residual retrieval vectors. The semantic matching response is obtained by performing a token-by-token dot product between the query semantic retrieval vector and the course evidence semantic retrieval vector, and the hierarchical residual matching response is also obtained by performing a token-by-token dot product between the query semantic retrieval vector and the hierarchical residual retrieval vector. For each query token, the main peak matching branch extracts the maximum response value from the semantic matching response as the highest matching response and records the position of the course evidence token corresponding to the highest matching response. The residual peak matching branch masks the position of the course evidence token corresponding to the highest matching response in the hierarchical residual matching response, then selects the two highest-ranked response values ​​from the remaining responses and performs mean compression on the two response values ​​to form a secondary matching response; if there is only one remaining response, this remaining response is used as the secondary matching response; if there are zero remaining responses, the secondary matching response is set to zero.

[0054] The scoring output layer synthesizes the highest-matching response and the second-highest-matching response in a ratio of 0.8:0.2 to obtain the combined matching response for each query token. The scoring output layer sums the combined matching responses for each query token within the same query token sequence to obtain the course evidence matching score for each course evidence token sequence. The scoring output layer also retains the course evidence token positions corresponding to the highest-matching response and the second-highest-matching response, and writes the course evidence matching score, course evidence token sequence number, and course evidence token position together into the course evidence matching score set.

[0055] During the model training phase, training samples are composed of query statements, positive course evidence segments, and negative course evidence segments. When the training query carries conditions such as cognitive level, interest preference, learning bottleneck, or learning progress, each retrieval condition is converted into query dimension tags. Positive course evidence segments, in addition to answering the query statement, also carry learning retrieval dimension tags that are consistent with the query dimension tag type and tag number item by item. Negative course evidence segments retain course evidence segments with semantically similar teaching content but at least one inconsistent learning retrieval dimension tag. When the training query does not carry retrieval conditions, positive and negative course evidence segments are still selected based on the answer relationship between the teaching content and the query statement.

[0056] The token embeddings corresponding to the query dimension tags and the learning retrieval dimension tags, together with the query encoder parameters and the document encoder parameters, participate in the training, so that course evidence segments with different cognitive levels, interest topics, bottleneck types or learning progress under the same course topic can generate distinguishable token latent states.

[0057] Before training, the trainable semantic projection matrix, trainable position embedding table, and trainable residual projection matrix are all initialized as trainable parameters in the improved ColBERT model, and participate in training together with the query encoder parameters and document encoder parameters.

[0058] When constructing the training samples, each query statement is configured with one positive course evidence segment and seven negative course evidence segments. Positive course evidence segments are selected based on the source of the query statement: when the query statement originates from manually retrieved and labeled data, the course evidence segment selected by the labeler and labeled as "completely answered" is used as the positive course evidence segment; a "completely answered" label means that the teaching content in the course evidence segment covers the core course terms, knowledge point terms, or question descriptions in the query statement, and includes explanations, derivations, operational content, or question analysis content related to the query statement; when the query statement originates from a question-and-answer pair, the course evidence segment containing the question-and-answer pair's answer text is used as the positive course evidence segment; when the query statement originates from question analysis data, the course evidence segment corresponding to the question position of the question analysis text is used as the positive course evidence segment; if the query statement does not have the above correspondence, it is not included in the training samples.

[0059] Negative course evidence segments are selected from the set of course evidence segments. Before selection, positive course evidence segments, as well as those with the same course level and overlapping content boundaries as positive course evidence segments, are deleted. Then, seven negative course evidence segments are selected from the remaining course evidence segments. Three of these negative course evidence segments are selected from course evidence segments in the same course as the positive course evidence segment but at a different teaching unit level; two are selected from course evidence segments in the same course as the positive course evidence segment but at a different chapter level; and two are selected from course evidence segments from different courses. If the number of course evidence segments within a certain source range is insufficient, they are supplemented from the remaining course evidence segments in order of course evidence segment number. Negative course evidence segments must not contain the text of the answer to the query statement, the text of the question explanation, or content manually marked as "complete answer". When determining whether a negative course evidence segment contains the answer text or question analysis text corresponding to the query statement, first check the question position, course question-answer pair number, and manual annotation number corresponding to the negative course evidence segment; if the numbers are inconsistent, then check the overlap of consecutive terms between the negative course evidence segment text and the answer text or question analysis text. If the number of consecutive overlapping terms reaches 5 or more, the course evidence segment will not be regarded as a negative course evidence segment.

[0060] After the positive and negative course evidence segments are screened, the query statement is converted into a query token sequence through the query sequence module. Positive and negative course evidence segments are then converted into course evidence token sequences through the evidence sequence module. Each course evidence token sequence retains the course level position, content boundary position, and the correspondence between the course evidence token sequence and the course evidence segment. After the training samples are constructed, they are divided into a training set, a validation set, and a test set in an 8:1:1 ratio. The training set is used for model parameter updates, the validation set is used to calculate MRR@10 and trigger the training completion condition, and the test set is used for retrieval performance statistics in the implementation example.

[0061] During training, each training sample forms a sorting group, which includes one query token sequence, one positive course evidence token sequence, and seven negative course evidence token sequences. The query token sequence, positive course evidence token sequence, and negative course evidence token sequence are fed into the query encoder, document encoder, hierarchical residual projection layer, main peak-residual peak interaction aggregation structure, and scoring output layer, respectively. Following the aforementioned course evidence matching score calculation process, one positive sample matching score and seven negative sample matching scores are obtained.

[0062] The positive sample matching scores and the seven negative sample matching scores form the same ranking group. The scoring output layer performs exponential operations on each of the eight matching scores, and then divides the exponential result of each matching score by the sum of the exponential results of the eight matching scores to obtain eight normalized probabilities. The normalized probability corresponding to the positive course evidence segment is used as the positive sample probability, and the cross-entropy ranking loss is taken as the negative logarithm of the positive sample probability.

[0063] During parameter updates, the AdamW optimization method is used to process the gradient corresponding to the cross-entropy ranking loss. Each training batch contains 32 ranking groups. The cross-entropy ranking loss is first averaged over the 32 ranking groups to obtain the batch loss. The gradient of the batch loss is then calculated layer by layer along the scoring output layer, the main peak-residual peak interaction convergence structure, the hierarchical residual projection layer, the query encoder, and the document encoder. The trainable semantic projection matrix, trainable position embedding table, and trainable residual projection matrix are updated using a first learning rate of 1×10^-4. The query encoder parameters and document encoder parameters are updated using a second learning rate of 2×10^-5. The AdamW optimization method maintains a first-order momentum cache and a second-order momentum cache for each trainable parameter. At the end of each training batch, the current gradient is written to the first-order momentum cache, the square of the current gradient is written to the second-order momentum cache, and the parameter update amount is calculated according to the first-order momentum cache, the second-order momentum cache, and the corresponding learning rate. Subsequently, the parameter update amount is subtracted from the current parameter value, and the parameter magnitude is reduced by a weight decay coefficient of 0.01. When the gradient norm exceeds 1.0, the gradients of the current training batch are pruned before parameter updates. After a parameter update, the updated trainable semantic projection matrix continues to participate in the linear projection of the semantic projection branch, the updated trainable position embedding table continues to participate in the embedding lookup of course level positions and content boundary positions, and the updated trainable residual projection matrix continues to participate in the generation of level residual retrieval vectors. The training completion condition is set to no longer improve the MRR@10 on the validation set for 3 consecutive rounds, or the number of training rounds reaches 20 rounds.

[0064] After training, the updated trainable semantic projection matrix participates in the generation of query semantic retrieval vectors and course evidence semantic retrieval vectors, while the updated trainable position embedding table and trainable residual projection matrix participate in the generation of hierarchical residual retrieval vectors. During the course resource entry phase, the course evidence token sequence set is input offline into the document encoder, semantic projection branch, and hierarchical residual projection branch to generate course evidence semantic retrieval vectors and hierarchical residual retrieval vectors. The course evidence token sequence number, course evidence segment number, course evidence token bit, course level bit, content boundary bit, course resource location, course evidence semantic retrieval vector, and hierarchical residual retrieval vector are written into the retrieval index. Upon receiving a user's course retrieval statement, the query token sequence is input online into the query encoder and semantic projection branch to generate query semantic retrieval vectors. The main peak-residual peak interactive aggregation structure performs the aforementioned interactive aggregation and scoring output processing on the query semantic retrieval vectors and course evidence semantic retrieval vectors and hierarchical residual retrieval vectors in the retrieval index, forming a course evidence matching scoring set.

[0065] During the hit location reconstruction, the scoring output layer reads the course evidence token positions corresponding to the highest and second-highest matching responses. The hit start position is taken from the smallest token position among the course evidence token positions participating in the highest and second-highest matching responses, and the hit end position is taken from the largest token position among the course evidence token positions participating in the highest and second-highest matching responses. A mapping relationship between tokens and original content positions is maintained between the course evidence token sequence and the course evidence segment. The hit start and hit end positions are used to reconstruct the content hit position within the course evidence segment through this mapping relationship and participate in the hit location labeling in the result generation module.

[0066] In this embodiment, the target evidence module specifically refers to: The target evidence module writes back the course evidence matching score to the corresponding course evidence segment according to the correspondence between the course evidence matching score and the course evidence token sequence in the course evidence matching score set; When a query token sequence carries one or more query dimension tags, check the query dimension tags against the learning retrieval dimension tags in the course evidence token sequence item by item according to the tag type. If each query dimension tag has a learning retrieval dimension tag with the same tag type and tag number, retain the corresponding course evidence segment. If the query token sequence does not carry query dimension tags, retain all course evidence segments that have been written to the course evidence matching score. The retained course evidence segments are sorted by descending score, and the course evidence segments with the highest scores are selected to form candidate course evidence segments. Perform duplicate boundary checks on candidate course evidence segments, and retain the candidate course evidence segment with the highest course evidence matching score that has the same course level position and overlapping content boundary positions. The candidate course evidence segments that have undergone repeated boundary checks are back rowed to the corresponding course level position and content boundary position in the course retrieval base to form the target course evidence set.

[0067] In practice, query dimension tags and learning retrieval dimension tags are checked against each other based on tag type and label number. Cognitive query tags are checked only against cognitive level tags, interest query tags are checked only against interest preference tags, checkpoint query tags are checked only against learning checkpoint tags, and progress query tags are checked only against learning progress tags. If both tag type and label number are the same, the two tags are considered to be identical.

[0068] When a query token sequence carries one or more query dimension tags, check the query dimension tags against the learning retrieval dimension tags in the course evidence token sequence according to the tag type. If each query dimension tag has a learning retrieval dimension tag with the same tag type and tag number, retain the corresponding course evidence segment. If the query token sequence does not carry query dimension tags, retain all course evidence segments written with the course evidence matching score.

[0069] The learning retrieval dimension labels only participate in the selection of candidate course evidence segments and do not rewrite the course evidence matching score output by the improved ColBERT model. Course evidence segments that pass the dimension check are further sorted according to their course evidence matching scores and duplicate boundary checks are performed, while course evidence segments that fail the dimension check are removed from the candidate range.

[0070] In this embodiment, the result generation module is specifically as follows: Read the course level position, content boundary position, course evidence matching score, course resource location and learning retrieval dimension markers corresponding to each target course evidence segment in the target course evidence set; Organize the courses, chapters or teaching units corresponding to the course level into course search dimension labels, and organize the learning search dimension labels into cognitive level labels, interest preference labels, learning checkpoint labels or learning progress labels. In the evidence segment of the target course, mark the start and end positions of the hit that match the query token sequence to form a hit position label; The course retrieval dimension label, learning retrieval dimension label, hit location label, course resource location, and course evidence matching score are written into the same retrieval result item, and the retrieval result items are arranged according to the course evidence matching score to generate intelligent multidimensional retrieval results for course data.

[0071] In the specific implementation process, the results generation module queries the tag dictionary according to the type and tag number of the learning retrieval dimension markers. It restores cognitive level markers to labels such as "able to ask," "able to distinguish," "able to unify," "able to create," "able to remember," "able to connect," "able to summarize," "able to master," "able to teach," or pseudo-shallow labels. Interest preference markers are restored to corresponding interest topics. Learning bottleneck markers are restored to labels indicating missing main concepts, unfamiliarity with shallow techniques, insufficient proficiency, or pseudo-shallow dependencies. Learning progress markers are restored to convergence progress and labels indicating mastered and yet-to-be-mastered question types. Duplicate markers of the same type are deleted, while different markers are retained. If the evidence section of the target course does not carry a certain type of marker, no corresponding marker is generated.

[0072] When restoring the hit location, hierarchical location markers, boundary location markers, and learning retrieval dimension markers are excluded. Only the original text character position corresponding to the teaching content token is used to determine the start and end hit positions. When the course resource location is a courseware page, video time segment, text paragraph, or question number, the corresponding resource location information is written respectively. When the target course evidence segment belongs to teaching materials or step-by-step guidance content, the material address or guidance step position is also written.

[0073] Example 1: To verify the feasibility of this invention in practice, it was applied to an online course resource platform of a university. The course retrieval platform provides course content, question analysis, step-by-step guidance, and teaching material retrieval services for public and professional foundation courses such as Computer Fundamentals, Database Principles, Machine Learning, University Physics, and Advanced Mathematics. The course retrieval platform sets up resource access services, course evidence processing services, model retrieval services, and result display services. The resource access service receives course catalogs, courseware, video-transcribed text, exercises and solutions, step-by-step guidance text, and descriptions of teaching materials. The course evidence processing service completes course structure organization, evidence segment delineation, learning retrieval dimension attachment, and course evidence token sequence generation. The model retrieval service performs vector index construction, online matching scoring, and target evidence filtering. The result display service returns the course location, hit content, resource entry point, and learning retrieval dimension.

[0074] This embodiment selects 24 courses from the course platform as the search objects, including 1268 chapters, 4315 teaching units, approximately 864,000 searchable text segments, 612 hours of video transcription content, 3487 courseware pages, and 21,000 question explanations. The course catalog, chapter hierarchy, teaching units, and knowledge point affiliations are organized into course structure data; the course text, video transcription text, courseware text, question content, question explanations, step-by-step guidance content, and teaching material descriptions are organized into course content data; and page numbers, video duration, paragraph numbers, question numbers, and material addresses are organized into course resource location data.

[0075] The course platform synchronously collects user cognitive assessment records, interest tag records, answer process records, and learning training records. Deep cognition, shallow cognition, and pseudo-shallow cognition states in the cognitive assessment records are converted into cognitive level tags; topic preferences in the interest tag records are converted into interest preference tags; the types of obstacles corresponding to answering timeouts, no answers, consecutive errors, and interruptions in problem-solving steps are converted into learning obstacle tags; convergence progress, mastered question types, and question types yet to be mastered are converted into learning progress tags. The course text, question explanations, step-by-step guidance content, and teaching material descriptions are configured with corresponding cognitive levels, interest topics, obstacle types, and applicable learning progress according to knowledge point numbers, question numbers, or material numbers, forming the course's evidence-based learning retrieval dimensions.

[0076] The retrieval base module establishes course positions, chapter positions, and teaching unit positions according to the course catalog. It identifies title boundaries, paragraph boundaries, subtitle time boundaries, and question number boundaries in the course content, organizes continuous content into content segments, and writes them into source positions, page positions, time positions, paragraph positions, or question positions. Learning retrieval dimensions are linked to the corresponding content segments according to knowledge point numbers, question numbers, or material numbers. The evidence delineation module uses teaching units as the delineation scope, merges content segments with continuous content boundaries and continuous resource locations into course evidence segments, and summarizes the cognitive level tags, interest preference tags, learning checkpoint tags, and learning progress tags associated with the course evidence segments, ultimately forming 74,236 course evidence segments.

[0077] The evidence sequence module performs term segmentation on the teaching content in the course evidence segments, converting course level positions, content boundary positions, and learning retrieval dimensions into corresponding tags. It then generates a course evidence token sequence using a WordPiece sub-word segmentation mapping method that shares a vocabulary with the improved ColBERT model document encoder. The course evidence token sequence, course evidence segment number, course level position, content boundary position, course resource location, and learning retrieval dimension tags are jointly written into the course evidence index. The course evidence token sequence is input offline into the document encoder and the hierarchical residual projection layer to generate course evidence semantic retrieval vectors and hierarchical residual retrieval vectors, which are then stored in the vector index.

[0078] Users enter their course search query on the course search interface and select filtering conditions such as cognitive level, interest preference, learning bottleneck, or learning progress. If the user does not actively select filtering conditions, the query sequence module reads the learning status tags already recorded for the current user under the corresponding knowledge point or question type. The query sequence module identifies the search focus in the course search query, converts the filtering conditions into query dimension tags, and generates a query token sequence using a shared WordPiece vocabulary. The query token sequence is then input into the online query encoder to generate a semantic retrieval vector.

[0079] The improved ColBERT model retains the query encoder and document encoder, adds a hierarchical residual projection layer at the encoder output, and replaces the single-peak MaxSim aggregation path with a main-peak-residual-peak interactive aggregation structure. The model retrieval service performs token-by-token matching between the query semantic retrieval vector and the course evidence semantic retrieval vector and hierarchical residual retrieval vector in the course evidence index, extracting the highest semantic matching response and the secondary residual matching response within the boundary to form the course evidence matching score.

[0080] The target evidence module writes the course evidence matching score back to the course evidence segment. When querying a token sequence carrying query dimension tags, the target evidence module verifies that the query dimension tags match the learning retrieval dimension tags in the course evidence segment, retaining course evidence segments with consistent tag types and label numbers. Pseudo-shallow search conditions correspond to course evidence segments carrying pseudo-shallow tags; interest preference search conditions correspond to teaching materials carrying tags for the same interest topics; learning checkpoint search conditions correspond to question explanations or step-by-step guidance content carrying tags for the same checkpoint type; and learning progress search conditions correspond to course evidence segments carrying the same progress status or question type mastery status.

[0081] The course evidence segments verified through the learning dimension are sorted in descending order of course evidence matching score, and the top 20 course evidence segments are selected as candidate course evidence segments. Candidate course evidence segments with the same course level and overlapping content boundaries are grouped into the same duplicate boundary group, and the course evidence segment with the highest course evidence matching score is retained in each group. The target evidence module then reorders the retained course evidence segments back into the original course content according to the course level and content boundary, and reads the corresponding page number, video time, paragraph number, question number or material address.

[0082] The results display service outputs the course name, chapter name, teaching unit name, matched content, course resource location, and course evidence matching score, and simultaneously outputs cognitive level, interest preference, learning checkpoints, or learning progress markers. When restoring the matched location, the hierarchical location marker, boundary location marker, and learning retrieval dimension marker are excluded; only the original character position corresponding to the teaching content token is used to determine the start and end positions of the match.

[0083] For comparative testing, 3200 user search queries were extracted from the course platform's historical search logs. Three individuals with teaching assistant experience labeled the queries with correct course evidence, course level location, resource location, and applicable learning search dimensions. Search queries with differing opinions among the labelers were re-verified, ultimately retaining 2850 valid test queries. These included 1046 concept explanation queries, 732 question analysis queries, 418 experiment operation queries, 369 formula derivation queries, and 285 resource location queries. When test queries included cognitive level, interest preferences, learning bottlenecks, or learning progress conditions, corresponding tag types and tag numbers were also labeled.

[0084] The comparative tests used the same course data, the same test queries, and the same number of returned results, running keyword inverted index retrieval, original ColBERT retrieval, and the intelligent multidimensional retrieval of course data from this invention, respectively. Test metrics included Top-5 hit rate, MRR@10, evidence segment location accuracy, and average retrieval time. The test results are shown in Table 1.

[0085] Table 1. Comparison of Course Evidence Retrieval Results

[0086] As shown in Table 1, the system of this invention outperforms the keyword inverted index retrieval method and the original ColBERT retrieval method in all three core metrics: Top-5 hit rate, MRR@10, and evidence segment location accuracy. The keyword inverted index retrieval method relies on term matching, and when faced with course queries containing question descriptions, it tends to return content with similar keywords but incomplete teaching meaning. Its Top-5 hit rate is 72.84%, and its evidence segment location accuracy is 68.39%. The original ColBERT retrieval method improves semantic matching capabilities through token-level late interaction, increasing the Top-5 hit rate to 89.16%. However, the original ColBERT model still primarily relies on a single-peaked MaxSim response, failing to adequately retain the scattered conditions, derivations, and conclusions within the same course's evidence segment. Its evidence segment location accuracy is 84.77%. The system of this invention organizes course resources into evidence through course evidence segments, course level positions, and content boundary positions. It also models the highest semantic hit and boundary residual hit through a hierarchical residual projection layer and a main peak-residual peak interaction aggregation structure, achieving a Top-5 hit rate of 94.91%, an MRR@10 of 0.887, and an evidence segment positioning accuracy of 92.35%.

[0087] Among them, Top-5 hit rate represents the proportion of queries where the correct course evidence segment appears in the first 5 search results, reflecting the system's ability to recall valid course evidence from a limited number of returned results; MRR@10 represents the average of the last rank of the first correct course evidence segment in the first 10 results, measuring the ranking position of the correct evidence segment; Evidence segment location accuracy represents the proportion of course, chapter, teaching unit, and resource locations in the system's returned results that are consistent with the manually annotated locations, measuring the course source traceability capability of the search results; Average retrieval time represents the average time from receiving the returned search results for a single query. The average retrieval time of the system in this invention is 151ms, slightly higher than the original ColBERT retrieval method, but still meets the interactive retrieval requirements of online course platforms, and the location accuracy is significantly improved.

[0088] This embodiment demonstrates that the system of the present invention does not simply input course resources into a neural retrieval model. Instead, it first organizes the course content into course evidence segments carrying course hierarchy positions and content boundary positions, and then performs fine-grained matching through a hierarchical residual projection layer and a peak-residual peak interaction aggregation structure in the improved ColBERT model. The hierarchical residual projection layer allows course hierarchy positions and content boundary positions to participate in the construction of course evidence token representation, while the peak-residual peak interaction aggregation structure allows the highest semantic hit and the secondary residual hits within the same evidence segment to participate in the scoring. Through this processing, the system can reduce the omission of scattered teaching evidence by single-peak matching paths, improve the stability of course evidence segment recall, ranking, and position back-ranking, and is suitable for application in university online course platforms, corporate training course libraries, and smart teaching resource retrieval scenarios.

[0089] 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.

Claims

1. A deep learning-based intelligent multidimensional retrieval system for course data, characterized in that, include: The retrieval base module is used to obtain course retrieval related data, organize course content, course structure, course resource location, and learning retrieval dimensions related to course content, and form a course retrieval base. The evidence delineation module is used to delineate course evidence segments with pedagogical meaning boundaries in the course retrieval base, record the course level position and content boundary position of the course evidence segment, and establish a connection between the course evidence segment and the corresponding learning retrieval dimension to form a set of course evidence segments. The evidence sequence module is used to perform retrieval granular compression on the course evidence segments in the course evidence segment set, and organize the continuously unfolded teaching content within the course evidence segment into a set of course evidence token sequences carrying course level bits, content boundary bits and learning retrieval dimension markers; The query sequence module receives the user's course search statement and search filter conditions, identifies the search focus from the user's course search statement, and marks and organizes the query dimensions corresponding to the search focus and search filter conditions into a query token sequence. The ColBERT retrieval module is used to input the query token sequence and the course evidence token sequence set into the improved ColBERT model. It replaces the single token projection path at the encoder output with a hierarchical residual projection layer and replaces the single-peak MaxSim aggregation path with a main peak-residual peak interactive aggregation structure to form a course evidence matching score set. The target evidence module is used to screen candidate course evidence segments according to the course evidence matching score set and query dimension label, and to back-row the candidate course evidence segments to the corresponding course level position and content boundary position to form the target course evidence set. The results generation module is used to organize the retrieval dimensions, learning retrieval dimensions, hit locations, and course resource locations corresponding to the evidence set of the target course, and generate intelligent multidimensional retrieval results for course data.

2. The intelligent multidimensional retrieval system for course data based on deep learning according to claim 1, characterized in that, The course retrieval associated data includes course content data, course structure data, course resource location data, and learning status associated data; Course content data corresponds to the course text, question content, question analysis, step-by-step guidance content or teaching material description; course structure data corresponds to the hierarchical relationship between courses, chapters, teaching units and knowledge points and the order of teaching development; course resource location data corresponds to the source position, page position, time position, paragraph position or question position of the teaching expression content in the course resources. The learning status-related data includes cognitive level tags, interest preference tags, learning checkpoint tags, and learning progress tags. The learning retrieval condition data consists of at least one of the retrieval filtering conditions or learning status-related data.

3. The intelligent multidimensional retrieval system for course data based on deep learning according to claim 1, characterized in that, The formation of the course retrieval platform specifically includes: The retrieval base module reads the course catalog, chapter hierarchy, and teaching unit sequence from the course structure data, and establishes a course hierarchy framework composed of course positions, chapter positions, and teaching unit positions. Identify title boundaries, paragraph boundaries, or question number boundaries in the course content data, organize continuous content within the same boundary interval into content segments, and place the content segments into the corresponding teaching unit positions; Write the course resource location data into the content segment to form a course retrieval base that includes a course hierarchy framework, content segments, content boundary positions, and resource location indexes.

4. The intelligent multidimensional retrieval system for course data based on deep learning according to claim 1, characterized in that, The evidence delineation module specifically comprises: The evidence delineation module uses the teaching unit corresponding to the course structure as the delineation scope in the course retrieval base, and scans continuous content sequentially along the course content within the teaching unit. Course content that is within the same teaching unit, has continuous content boundaries, and has continuous resource locations is classified into the same candidate evidence segment, and the candidate evidence segment is truncated at the point of teaching unit switch, content boundary interruption, or resource location jump. The candidate evidence segments are identified as course evidence segments, the teaching unit positions corresponding to the course evidence segments are written as course level positions, and the start and end boundaries of the course evidence segments are written as content boundary positions, thus forming a set of course evidence segments.

5. The intelligent multidimensional retrieval system for course data based on deep learning according to claim 1, characterized in that, The evidence sequence module is specifically as follows: The evidence sequence module processes course evidence segments as units, reads the continuously unfolding teaching content within a course evidence segment, and retains the course level position and content boundary position corresponding to the course evidence segment. The teaching content is segmented into search terms, and whitespace, format control characters and duplicate delimiters are removed. The search terms are arranged according to the order of the teaching content in the course evidence segment to form a course evidence term sequence. The WordPiece word segmentation mapping method, which shares a vocabulary with the document encoder of the improved ColBERT model, is adopted. The hierarchical position markers, boundary position markers, and learning retrieval dimension markers associated with the course evidence segments are written into the course evidence term sequence and converted into a course evidence token sequence to form a set of course evidence token sequences. The learning retrieval dimension markers include cognitive level markers, interest preference markers, learning checkpoint markers, or learning progress markers. Course evidence segments not associated with the corresponding learning retrieval dimension will not be written with the corresponding markers.

6. The intelligent multidimensional retrieval system for course data based on deep learning according to claim 1, characterized in that, The query sequence module is specifically as follows: Receive user course search statements and search filtering conditions, clean up whitespace, format control characters and duplicate delimiters in user course search statements, and retain search terms related to course content, knowledge points, resource locations and problem descriptions; The search focus is determined by the continuity of terms in the user's course search statement, and consecutively appearing course terms, knowledge point terms, or question description terms are grouped into the same search focus segment. Convert the cognitive level conditions, interest preference conditions, learning bottleneck conditions, or learning progress conditions in the search filtering conditions into query dimension tags, and arrange the query dimension tags before the corresponding search focus segment; We adopt the WordPiece sub-word segmentation mapping method, which shares a vocabulary with the query encoder of the improved ColBERT model, to convert the retrieval focus segment and query dimension tags into a query token sequence, and retain the start and end positions of the retrieval focus segment in the query token sequence.

7. The intelligent multidimensional retrieval system for course data based on deep learning according to claim 1, characterized in that, The improved ColBERT model includes a query encoder, a document encoder, a hierarchical residual projection layer, a main peak-residual peak interaction aggregation structure, and a scoring output layer. The hierarchical residual projection layer includes a semantic projection branch and a hierarchical residual projection branch; The query encoder receives a sequence of query tokens and outputs the hidden state of the query tokens; the document encoder receives a set of course evidence token sequences and outputs the hidden state of the course evidence tokens. The hierarchical residual projection layer performs semantic projection on the query token hidden state and the course evidence token hidden state to form the query semantic retrieval vector and the course evidence semantic retrieval vector, and maps the course evidence token hidden state, course level bit and content boundary bit together into the hierarchical residual retrieval vector. The main peak-residual peak interaction aggregation structure constructs a semantic matching response using query semantic retrieval vectors and course evidence semantic retrieval vectors, and constructs a hierarchical residual matching response using query semantic retrieval vectors and hierarchical residual retrieval vectors; the main peak matching branch extracts the highest matching response corresponding to the query token from the semantic matching response, and the residual peak matching branch compresses the secondary matching response corresponding to the query token from the hierarchical residual matching response. The scoring output layer synthesizes the highest matching response and the second-highest matching response to obtain the course evidence matching score corresponding to each course evidence token sequence. It retains the course evidence token positions corresponding to the highest matching response and the second-highest matching response and summarizes them to form a course evidence matching score set.

8. The intelligent multidimensional retrieval system for course data based on deep learning according to claim 7, characterized in that, The hierarchical residual projection layer performs semantic projection on the query token hidden state and the course evidence token hidden state to form a query semantic retrieval vector and a course evidence semantic retrieval vector. It also maps the course evidence token hidden state, course level bits, and content boundary bits together into a hierarchical residual retrieval vector. Specifically: The semantic projection branch uses a trainable semantic projection matrix to perform linear projection and normalization processing on the query token hidden state and the course evidence token hidden state, respectively, to form the query semantic retrieval vector and the course evidence semantic retrieval vector. The hierarchical residual projection branch reads the course level bit and content boundary bit corresponding to the hidden state of the course evidence token. Within the same content boundary bit limit, it performs local difference deduction on the hidden states of adjacent course evidence tokens to form the local residual state of the course evidence. The hierarchical residual projection branch uses a trainable position embedding table to map course hierarchical positions and content boundary positions into position embeddings, and concatenates the local residual states of course evidence with the position embeddings in dimensional order to form a residual position combination vector. The hierarchical residual projection branch uses a trainable residual projection matrix to perform linear projection and normalization on the residual position combination vector to form a hierarchical residual retrieval vector. The trainable semantic projection matrix, trainable position embedding table, and trainable residual projection matrix are all trainable parameters in the improved ColBERT model.

9. The intelligent multidimensional retrieval system for course data based on deep learning according to claim 1, characterized in that, The target evidence module is specifically as follows: The target evidence module writes back the course evidence matching score to the corresponding course evidence segment according to the correspondence between the course evidence matching score and the course evidence token sequence in the course evidence matching score set; When a query token sequence carries one or more query dimension tags, check the query dimension tags against the learning retrieval dimension tags in the course evidence token sequence item by item according to the tag type. If each query dimension tag has a learning retrieval dimension tag with the same tag type and tag number, retain the corresponding course evidence segment. If the query token sequence does not carry query dimension tags, retain all course evidence segments that have been written to the course evidence matching score. The retained course evidence segments are sorted by descending score, and the course evidence segments with the highest scores are selected to form candidate course evidence segments. Perform duplicate boundary checks on candidate course evidence segments, and retain the candidate course evidence segment with the highest course evidence matching score that has the same course level position and overlapping content boundary positions. The candidate course evidence segments that have undergone repeated boundary checks are back rowed to the corresponding course level position and content boundary position in the course retrieval base to form the target course evidence set.

10. The intelligent multidimensional retrieval system for course data based on deep learning according to claim 1, characterized in that, The result generation module specifically comprises: Read the course level position, content boundary position, course evidence matching score, course resource location and learning retrieval dimension markers corresponding to each target course evidence segment in the target course evidence set; Organize the courses, chapters or teaching units corresponding to the course level into course search dimension labels, and organize the learning search dimension labels into cognitive level labels, interest preference labels, learning checkpoint labels or learning progress labels. In the evidence segment of the target course, mark the start and end positions of the hit that match the query token sequence to form a hit position label; The course retrieval dimension label, learning retrieval dimension label, hit location label, course resource location, and course evidence matching score are written into the same retrieval result item, and the retrieval result items are arranged according to the course evidence matching score to generate intelligent multidimensional retrieval results for course data.

Images