A teacher report adaptive generation method and system based on a data base

Abstract

The application provides a kind of based on data base's teacher report adaptive generation method and system, the method includes: collecting teacher business data and carrying out data cleaning and entity alignment, constructs the data base including physical layer and logic layer;The report to be filled is input into visual semantic analysis engine, and the layout and table header semantics of report are jointly analyzed, to obtain the table header semantic tree;According to the table header semantic tree, different report fields are respectively matched by semantic type mixed matching strategy and knowledge graph path reasoning to query and deduce the data base, to obtain the filling result set of multiple fields;According to the filling result set, prefill report is generated, and user check and correction interaction processing is received, to generate archive report.The application improves data consistency and integrity.

Description

Technical Field

[0001] This invention relates to the field of educational informatization technology, and in particular to a method and system for adaptive generation of teacher reports based on a data foundation. Background Technology

[0002] With the deepening of educational informatization, modern universities have generally established multiple vertical business systems, such as academic affairs, personnel, scientific research, and finance. In scenarios such as annual performance appraisals of teachers, professional title evaluation applications, and discipline assessments, it is necessary to aggregate and fill in multi-dimensional teacher data scattered across these systems to generate various statistical reports. These reports cover multiple dimensions, including teaching workload, research achievements, awards, and professional development experience. The data sources are dispersed, and the statistical definitions are complex, requiring a high degree of accuracy and consistency in data processing.

[0003] In existing technologies, the generation and management of teacher report data mainly rely on traditional ETL tools or automatic filling systems based on preset templates. These systems extract and fill structured data by using fixed mapping rules between predefined database fields and report cells. However, these solutions heavily rely on hard-coded "field-coordinate" mapping rules and lack the ability to understand the visual features and deep semantics of the report layout. Once the report layout is adjusted or involves implicit complex logical reasoning, the original mapping rules become invalid. At the same time, existing systems only support unidirectional data flow from the database to the report front end. User corrections on the front end cannot be written back to the underlying data base, resulting in long-term uncorrectable errors in the source data. Furthermore, the system's intelligence level cannot continuously improve with increased usage time. Summary of the Invention

[0004] This invention provides a data-based adaptive generation method and system for teacher reports to address the shortcomings of existing technologies.

[0005] This invention provides a data-based adaptive method for generating teacher reports, comprising: S1. Collect teachers' business data and perform data cleaning and entity alignment to build a data base that includes physical and logical layers; S2. Input the form to be filled into the visual semantic parsing engine, and perform joint parsing of the form layout and header semantics to obtain the header semantic tree; S3. Based on the semantic tree of the table header, query and infer the data base for different report fields using a semantic type hybrid matching strategy and knowledge graph path reasoning respectively, to obtain a set of populated results for multiple fields; S4. Generate a pre-filled report based on the filled result set, and receive user verification and correction interaction processing to generate an archived report.

[0006] According to the adaptive generation method for teacher reports based on a data base provided by the present invention, step S1 further includes: S11. Extract teacher business data from the teaching affairs system, personnel system, financial system, scientific research system, external academic databases and unstructured documents using ETL tools, and perform regular expression cleaning on the teacher business data to obtain a standardized dataset. S12. For teacher nodes in multiple source systems in the standardized dataset, the corresponding name semantic vector, title one-hot encoding vector and department embedding are concatenated to construct the initial feature vector of multiple nodes. S13. Input the initial feature vector into the graph attention network, and aggregate the neighbor node information by weighting the attention coefficients to obtain the updated feature vectors of multiple nodes; S14. Based on the updated feature vector, calculate the alignment similarity score of cross-system node pairs, and filter out cross-system node pairs of the same entity by comparing the alignment similarity score with a preset threshold. S15. Assign globally unique identifiers to cross-system node pairs, store the aligned data in the physical layer of the data base, and construct a semantic relationship graph between entities in the logical layer.

[0007] According to the adaptive generation method of teacher reports based on data base provided by the present invention, in step S13, the attention coefficient between two adjacent nodes is obtained by performing nonlinear transformation processing on the concatenation result of the learnable attention vector and the linearly transformed node features through a linear unit activation function with leakage correction; in step S14, the alignment similarity score is obtained by weighted summation of the cosine similarity of the updated feature vector of the cross-system node pair and the Jaccard similarity of the attribute set.

[0008] According to the adaptive generation method for teacher reports based on a data base provided by the present invention, step S2 further includes: S21. Input the form to be filled into the document layout analysis model, and perform multimodal fusion of the text content, page coordinates and visual image features in the form to obtain the page feature representation of multiple cells; S22. Based on the layout feature representation, the hierarchical relationship between multi-level headers and merged cells is parsed and processed. The complete semantic paths of multiple fill slots from the root node to the leaf node are merged to obtain an ambiguous header semantic tree.

[0009] According to the adaptive generation method for teacher reports based on a data base provided by the present invention, step S3 further includes: S31. For the direct fields in the header semantic tree, the semantic feature vectors of multiple fields are compared with the semantic feature vectors of candidate data fields in the data base. A comprehensive score is calculated in combination with data type constraints. The optimal mapping result of the direct fields is obtained by selecting the candidate field with the highest comprehensive score. S32. For the complex logical fields in the semantic tree of the table header, starting from the entity in the logical layer of the data base, the entity relationship graph is traversed through multi-hop path search, and the subgraphs that meet the preset attribute condition constraints are aggregated and calculated to obtain the statistical values ​​of the complex logical fields. S33. Combine the optimal mapping result with the statistical value to obtain a filled result set.

[0010] According to the adaptive generation method for teacher reports based on a data base provided by the present invention, step S4 further includes: S41. Based on the filling result set, fill multiple field values ​​into the report template, and highlight the filling items whose confidence scores are lower than the preset threshold to obtain a pre-filled report. S42. Present the pre-filled report on the interactive front end, receive the user's verification and correction operations on the highlighted items and other fields, and package and archive the final report data to obtain the archived report.

[0011] The adaptive generation method for teacher reports based on a data base provided by the present invention further includes: S5. Capture user correction behavior and generate correction logs. Based on the correction logs, determine the type of correction behavior and trigger online incremental update processing of model parameters and reverse data governance processing respectively to continuously optimize the visual semantic parsing engine and the data base.

[0012] According to the adaptive generation method for teacher reports based on a data base provided by the present invention, step S5 further includes: S51. Collect the user-inputted correction value, original fill value, field identifier, and timestamp, encapsulate and process them to obtain the correction log; S52. Based on the metadata attributes of the modified fields in the correction log, determine the type of the correction behavior and obtain the determination result; the determination result includes semantic understanding deviation and factual data error; S53. Based on the determination result, the visual semantic parsing engine and the data base are continuously optimized through online incremental update of model parameters and reverse data governance.

[0013] According to the adaptive generation method for teacher reports based on a data base provided by the present invention, step S53 specifically includes: When the judgment result is a semantic understanding deviation, the correction log is converted into a supervised training sample, and the parameters of the document layout analysis model in the visual semantic parsing engine are updated by gradient using an online incremental learning algorithm with a feedback loss function to obtain updated model parameters. When the determination result indicates that the factual data is incorrect, the supporting materials uploaded by the user are used as the identification text. The consistency between the identification text and the corrected value, the user's historical credit score, the corrected value, and the degree of conflict recorded in the data base are weighted and calculated to obtain a verification credibility score. When the verification credibility score is greater than or equal to a preset credibility threshold, a reverse governance instruction is generated to synchronously update the original records in the physical and logical layers of the data base. When the verification credibility score is less than the preset credibility threshold, the corrected value is retained in the archived report, and an abnormal work order is generated for verification processing.

[0014] This invention also provides a data-based adaptive teacher report generation system for executing a data-based adaptive teacher report generation method as described in any of the preceding claims, comprising: Building module: Used to collect teachers' business data and perform data cleaning and entity alignment, building a data base that includes physical and logical layers; The parsing module is configured as a visual semantic parsing engine, which is used to receive the report to be filled, and jointly parse the layout and header semantics of the report to obtain the header semantic tree. Query module: Based on the semantic tree of the table header, it performs query inference on the data base for different report fields using a semantic type hybrid matching strategy and knowledge graph path reasoning, and obtains a set of populated results for multiple fields; Correction module: Used to generate a pre-filled report based on the filled result set, and to receive user verification and correction interaction processing to generate archived reports.

[0015] This invention constructs a one-person-one-file data foundation comprising physical and logical layers, and introduces a graph attention network to align heterogeneous data from multiple sources across systems. This effectively resolves the long-standing data silo problem in systems such as academic affairs, personnel, finance, and research. Data records of the same teacher in different business systems can be accurately linked using a globally unique identifier, fundamentally eliminating ambiguity in cross-system entities such as those with the same name but different identities, ensuring consistency in report statistics and traceability of data sources. Secondly, in the report parsing stage, a visual semantic parsing engine jointly parses the report layout and header semantics and constructs a header semantic tree. Combined with a semantic type hybrid matching strategy and a knowledge graph multi-hop path reasoning mechanism, the system can handle non-standard layouts, The system's ability to handle multi-level nested headers and complex statistical logic is significantly enhanced. Even when faced with complex filtering conditions, it can automatically complete calculations through dynamic configuration of the rule base, greatly reducing the frequency of manual intervention and the probability of errors in data entry. In addition, during user interaction, the confidence score-driven highlighting mechanism allows users to focus their verification efforts on high-risk fields, effectively shortening the report verification cycle. Furthermore, this invention uses a correction log-driven dual-loop mechanism to simultaneously transform each user's correction action into a supervisory signal for online incremental updates of model parameters and a trigger for reverse governance of the data base. This allows the semantic parsing accuracy and underlying data quality of the system to continuously improve with increased actual use, achieving a fundamental shift from passive data entry to proactive governance. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 A schematic diagram of a teacher report adaptive generation method based on a data base provided by the present invention; Figure 2 This invention provides a schematic diagram of a teacher report adaptive generation system based on a data base. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0019] To better understand this invention, the technical terms appearing in the embodiments of this invention will be explained in detail below.

[0020] One Person, One File Data Base: refers to a logical and physical two-layer storage architecture that takes a natural person as the core entity and associates their entire lifecycle business data (teaching, scientific research, personnel, etc.) with a globally unique identifier (GUID).

[0021] Graph Attention Networks (GAT): A type of graph neural network based on attention mechanisms, used to handle node classification or link prediction tasks in graph-structured data.

[0022] Visual semantic analysis: refers to the technology that combines computer vision and natural language processing techniques, and uses both textual and layout information of a document to understand it.

[0023] ETL: refers to the data extraction, transformation, and loading processes in data warehouse technology.

[0024] BERT stands for Bidirectional Encoder Representations from Transformers, which refers to a pre-trained model of bidirectional encoder representation based on the Transformer architecture.

[0025] LayoutLMv3: refers to a general-purpose pre-trained model for multimodal document understanding.

[0026] OCR stands for Optical Character Recognition, which refers to optical character recognition technology.

[0027] The embodiments of the present invention are described below with reference to the figures.

[0028] like Figure 1As shown, this invention provides a data-based adaptive teacher report generation method, including: S1. Collect teachers' business data, perform data cleaning and entity alignment, and construct a data base that includes physical and logical layers.

[0029] Step S1 further includes: S11. Extract teacher business data from the teaching affairs system, personnel system, financial system, scientific research system, external academic databases and unstructured documents using ETL tools, and perform regular expression cleaning on the teacher business data to obtain a standardized dataset.

[0030] Furthermore, in step S11, the present invention first uses an ETL tool to periodically extract full or incremental teacher business data from the university's internal academic affairs system, personnel system, financial system, research system, external academic databases (such as Web of Science), and unstructured documents (such as PDF papers and scanned certificates). After extraction, to address the common problem of dirty data in the original data, the present invention uses regular expressions to clean each field. For example, for the mobile phone number field, by matching and removing all non-numeric characters (hyphens, spaces, and other interfering symbols), it is uniformly converted into a pure numeric format, ultimately obtaining a standardized dataset.

[0031] S12. For teacher nodes in multiple source systems in the standardized dataset, the corresponding name semantic vector, title one-hot encoding vector and department embedding are concatenated to construct the initial feature vectors of multiple nodes.

[0032] In step S12, for teacher nodes from different source systems in the standardized dataset, this invention constructs an initial feature vector for each node. Specifically, this invention encodes the teacher's name text using a BERT pre-trained model to obtain a name semantic vector; it performs one-hot encoding on the teacher's title to obtain a title one-hot encoded vector; and it maps the department identifier to a low-dimensional dense representation using a learnable lookup table to obtain a department embedding vector. Finally, the above three types of vectors are concatenated to form the initial feature vector for each teacher node.

[0033] Specifically, for any teacher node in the graph Constructed initial feature vector This vector is composed of semantic features and structural features, and is represented as: ; in, This represents a vector concatenation operation. This represents the semantic vector of the name text extracted through the BERT pre-trained model. This represents one-hot encoding of teacher professional titles, transforming discrete title categories into orthogonal sparse vectors. This represents entity embeddings that map discrete department identifiers to low-dimensional dense vectors using a learnable lookup table.

[0034] S13. Input the initial feature vector into the graph attention network, and aggregate the neighbor node information by weighting the attention coefficients to obtain the updated feature vectors of multiple nodes; wherein, the attention coefficients between two adjacent nodes are obtained by performing a nonlinear transformation on the concatenation result of the learnable attention vector and the linearly transformed node features by the linear unit activation function with leakage correction.

[0035] Furthermore, after the initial feature vector is constructed in step S12, the present invention inputs it into a graph attention network (GAT). GAT is a graph neural network based on an attention mechanism. Its core idea is to assign different attention weights to the neighboring nodes of each node in the graph structure, thereby differentially aggregating neighbor information.

[0036] Specifically, for node i and its neighbor node j, this invention concatenates their linearly transformed feature vectors and feeds them into a LeakyReLU activation function for nonlinear transformation to obtain the original attention coefficients. Then, the attention coefficients are normalized using the Softmax function to obtain the attention weights of each neighbor. Subsequently, the feature vectors of node i are weighted and summed using the attention weights of each neighbor to obtain the updated feature vector of node i. This updated feature vector comprehensively reflects the node's own attributes and its social relationship structure information.

[0037] Furthermore, nodes with his neighbors Attention coefficient between The calculation formula is: ; in, Used to measure nodes For nodes The importance of For the modified linear unit activation function with leakage, This represents a learnable attention vector. This represents the learnable linear transformation weight matrix. , Representing nodes respectively and nodes eigenvectors, This represents a vector concatenation operation. This indicates the transpose operation.

[0038] S14. Calculate the alignment similarity score of cross-system node pairs based on the updated feature vectors, and filter cross-system node pairs of the same entity by comparing the alignment similarity score with a preset threshold; wherein, the alignment similarity score is obtained by weighted summation of the cosine similarity of the updated feature vectors of the cross-system node pairs and the Jaccard similarity of the attribute set; S15. Assign globally unique identifiers to cross-system node pairs, store the aligned data in the physical layer of the data base, and construct a semantic relationship graph between entities in the logical layer.

[0039] In steps S14-S15, this invention performs cross-system entity alignment on node pairs from different source systems. The specific alignment similarity score is obtained by weighted summation of two parts: first, the cosine similarity of the updated feature vectors of the two nodes, used to measure the directional consistency between them in the feature space; second, the Jaccard similarity of the attribute sets of the two nodes, used to measure the degree of overlap of their attribute values. When the alignment similarity score of a node pair exceeds a preset threshold (e.g., 0.85), this invention determines that the node pair refers to the same physical entity, assigns it a globally unique identifier (GUID), and persistently stores the aligned structured data and unstructured files in the physical layer of the data base. Simultaneously, in the logical layer, a full-domain knowledge graph is constructed with entities as nodes and business relationships as edges, forming a semantic relationship network of teachers covering multiple dimensions such as teaching, scientific research, and human resources.

[0040] Furthermore, for nodes from the personnel system graph and nodes from the scientific research system map The calculated alignment similarity score is expressed as: ; in, Represents a node and nodes The final alignment similarity score between them , These are the weight coefficients for the structural similarity term and the attribute similarity term, respectively. This represents the cosine similarity calculation function, used to measure the directional consistency of vectors in the feature space. , Representing nodes respectively and nodes Feature vector updated after GAT layer This represents the Jaccard similarity calculation function, used to measure the degree of overlap between two sets. , Representing nodes respectively and nodes The set of attributes.

[0041] S2. Input the form to be filled into the visual semantic parsing engine, and perform joint parsing of the form layout and header semantics to obtain the header semantic tree.

[0042] Step S2 further includes: S21. Input the form to be filled into the document layout analysis model, and perform multimodal fusion of the text content, page coordinates and visual image features in the form to obtain the page feature representation of multiple cells.

[0043] In step S21, the present invention inputs the form to be filled (Excel or Word format) into the document layout analysis model. The document layout analysis model adopts a multimodal pre-trained model such as LayoutLMv3. This type of model can simultaneously process three heterogeneous inputs: text, coordinate position, and visual image.

[0044] Specifically, the model extracts the text content of each cell from the report, the bounding box coordinates of the cell on the page, and the visual image features of the corresponding area. Subsequently, the three types of features are integrated through a multimodal fusion mechanism to obtain the layout feature representation of each cell. This representation encodes the semantic content of the cell and its spatial position relationship in the report layout.

[0045] S22. Based on the layout feature representation, the hierarchical relationship between multi-level headers and merged cells is parsed and processed. The complete semantic paths of multiple fill slots from the root node to the leaf node are merged to obtain an ambiguous header semantic tree.

[0046] In step S22, this invention parses the hierarchical relationship between multi-level headers and merged cells based on the layout features of each cell. For common cross-row and cross-column merged cells in reports, this invention determines the parent header node of each cell based on spatial proximity and visual hierarchy, establishing the hierarchical relationship between headers layer by layer. For each blank data slot to be filled, this invention traces upwards along its column and row, fusing the semantic feature vectors of all nodes along the path from the root node (top-level header) to the leaf node (directly corresponding to the field name of the slot), obtaining the path semantic fusion vector for that slot. Finally, the path semantic fusion vectors of all slots and their hierarchical structure together constitute the header semantic tree, where each leaf node corresponds to a field to be filled and its unambiguous and complete semantic description.

[0047] S3. Based on the semantic tree of the table header, query and infer the data base for different report fields using a semantic type hybrid matching strategy and knowledge graph path reasoning to obtain a set of populated results for multiple fields.

[0048] Step S3 further includes: S31. For the direct fields in the semantic tree of the table header, the semantic feature vectors of multiple fields are compared with the semantic feature vectors of candidate data fields in the data base. The comprehensive score is calculated in combination with the data type constraints. The optimal mapping result of the direct fields is obtained by selecting the candidate field with the highest comprehensive score.

[0049] In step S3, the present invention classifies report fields into two categories, direct fields and complex logical fields, based on the semantic features of the fields corresponding to each leaf node in the header semantic tree, and processes them using different data query and deduction paths.

[0050] In step S31, for direct fields, this invention calculates the semantic similarity between the path semantic fusion vector of the leaf node of the field and the semantic feature vectors of all candidate data fields in the physical layer of the data base, one by one. Simultaneously, it performs constraint matching on the data type of the field (e.g., a date type field only matches a date format candidate field; a score of 1 is given if the types match, otherwise 0). The semantic similarity and type matching scores are weighted and summed to obtain a comprehensive score. After traversing all candidate fields, the candidate field with the highest comprehensive score is determined as the optimal mapping result for that direct field, and the corresponding attribute value is directly extracted from the data base as the fill value.

[0051] S32. For the complex logical fields in the semantic tree of the table header, starting from the entities in the logical layer of the data base, the entity relationship graph is traversed through multi-hop path search, and the subgraphs that meet the preset attribute condition constraints are aggregated and calculated to obtain the statistical values ​​of the complex logical fields.

[0052] In step S32, for complex logical fields (e.g., "the number of papers published in Journal A as the first author in the past three years"), this invention starts with the teacher entity node in the knowledge graph of the data foundation logical layer, and performs a multi-hop path search along relational edges such as "published", traversing the paper entity nodes associated with the teacher node. During the traversal, this invention filters the attributes of each paper node item by item according to preset attribute conditions (time range limited to the past three years, author role limited to first author, and inclusion type limited to Journal A), retaining the subgraphs that simultaneously meet all conditions, and then counting or numerical aggregation calculations are performed on the nodes in the subgraphs to obtain the statistical value of the complex logical field.

[0053] S33. Combine the optimal mapping result with the statistical value to obtain a filled result set.

[0054] In step S33, the present invention aligns and merges the optimal mapping results of all direct fields with the statistical values ​​of all complex logical fields according to the slot index of each field in the semantic tree of the table header. At the same time, it adds the confidence score corresponding to the comprehensive score in step S31 to each filling result to obtain a complete filling result set, wherein each record contains three items: field identifier, filling value and confidence score.

[0055] S4. Generate a pre-filled report based on the filled result set, and receive user verification and correction interaction processing to generate an archived report.

[0056] Step S4 further includes: S41. Based on the filling result set, fill multiple field values ​​into the report template, and highlight the filling items whose confidence scores are lower than the preset threshold to obtain a pre-filled report.

[0057] In step S41, based on the fill result set output in step S3, the present invention writes the field identifiers and fill values ​​of each record into the report template one by one according to the corresponding slot index, completing the batch filling of field values. Subsequently, the present invention compares the confidence score attached to each record in the fill result set with a preset threshold, and highlights the fill items with confidence scores lower than the preset threshold in the report template to obtain a pre-filled report. In addition, the judgment basis for highlighting is the comprehensive score obtained by weighted summation of semantic similarity and type matching degree in step S31. A low comprehensive score means that the automatic mapping result of the field has a high degree of uncertainty, and the user needs to be prompted to check it carefully.

[0058] S42. Present the pre-filled report on the interactive front end, receive the user's verification and correction operations on the highlighted items and other fields, and package and archive the final report data to obtain the archived report.

[0059] In step S42, after the pre-filled report is generated, the present invention presents it on the interactive front end, which includes three functional modules: an online preview unit, an intelligent verification unit, and a user correction unit. The online preview unit fully displays the entire content of the pre-filled report; the intelligent verification unit highlights the marked items in a visually differentiated manner to guide the user to prioritize verification; the user correction unit provides an editing entry, allowing the user to modify the filled values ​​of any field or supplement missing data. After the user completes all verification and correction operations and confirms that there are no errors, the present invention packages and archives the final report data to obtain an archived report.

[0060] This also includes: S5. Capture user correction behavior and generate correction logs. Based on the correction logs, determine the type of correction behavior and trigger online incremental update processing of model parameters and reverse data governance processing respectively to continuously optimize the visual semantic parsing engine and the data base.

[0061] Step S5 further includes: S51. Collect the user-inputted correction value, original fill value, field identifier, and timestamp, encapsulate and process them to obtain the correction log.

[0062] In step S51, during the process of the user correcting the pre-filled report through the interactive front end, the front-end monitor of this invention collects the data involved in each correction action in real time, specifically including: the user's unique identifier for performing the correction operation, the field identifier of the field being corrected, the original filled value before correction, the correction value entered by the user, and the timestamp of the correction operation. After capture, this invention encapsulates the above five data items to form a correction log record in a five-tuple format and sends it to the back-end processing center.

[0063] S52. Based on the metadata attributes of the modified fields in the correction log, determine the type of the correction behavior and obtain the determination result; the determination result includes semantic understanding deviation and factual data error.

[0064] In step S52, after the correction log arrives at the backend processing center, this invention determines the type of the correction action based on the metadata attributes of the corrected field in the correction log. The metadata attributes describe the business meaning and data nature of the field, such as whether the field is a mapping field (corresponding to semantic comprehension deviation) or a fact-recording field (corresponding to factual data error). The final determination result is divided into two categories: semantic comprehension deviation and factual data error. Subsequently, different processing paths are entered for each category of determination result.

[0065] S53. Based on the determination result, the visual semantic parsing engine and the data base are continuously optimized through online incremental update of model parameters and reverse data governance.

[0066] Specifically, step S53 includes: When the judgment result is a semantic understanding deviation, the correction log is converted into supervised training samples, and the parameters of the document layout analysis model in the visual semantic parsing engine are updated using an online incremental learning algorithm with a feedback loss function to obtain updated model parameters. When the judgment result is a factual data error, the supporting materials uploaded by the user are used as the identification text, and the consistency between the identification text and the correction value, the user's historical credit score, the correction value, and the conflict degree recorded in the data base are weighted and calculated to obtain a verification credibility score. When the verification credibility score is greater than or equal to a preset credibility threshold, a reverse governance instruction is generated to synchronously update the original records in the physical and logical layers of the data base. When the verification credibility score is less than the preset credibility threshold, the correction value is retained in the archived report, and an abnormal work order is generated for verification processing.

[0067] In step S53, the present invention aims to continuously optimize the visual semantic parsing engine and the data base.

[0068] Specifically, when the judgment result is semantic understanding bias, it indicates that the visual semantic parsing engine in step S2 has a bias in its semantic understanding of the report fields, resulting in incorrect field mapping. This invention transforms the correction log into a supervised training sample, where the original filled value before correction corresponds to the model's incorrect prediction, and the user-inputted corrected value corresponds to the correct label. Subsequently, this invention uses an online incremental learning algorithm to input this supervised training sample into the document layout analysis model, constructing a feedback loss function to measure the difference between the current model's predicted distribution and the user's correction intention. This feedback loss function consists of two parts: a cross-entropy loss term and an L2 regularization term. The L2 regularization term constrains the offset between the new parameters and the original parameters, preventing the model from forgetting its original semantic understanding ability when learning new samples. Based on this feedback loss function, this invention performs gradient descent updates on the parameters of the document layout analysis model to obtain the updated model parameters, and writes them back to the visual-semantic parsing engine, enabling the engine to automatically adapt to the user's filling preferences or correct this type of semantic understanding bias in subsequent parsing of similar reports.

[0069] Specifically, the invention constructs a specific feedback loss function. The formula used to measure the difference between the current model's predicted distribution and the user's intended correction is as follows: ; in, This represents the total loss function value of the feedback mechanism. This represents the total number of semantic categories for elements in the document layout. This indicates the first [number] determined based on the user's correction log. Correct semantic tags for class This indicates the document layout analysis model under the current parameters. Predicting input Belongs to semantic category The conditional probability, This represents the weight coefficient of the regularization term. The L2 norm regularization term, representing parameter offsets, is used to prevent catastrophic forgetting in the process of adapting the model to new data.

[0070] Based on the above feedback loss function, the specific parameter update formula is as follows: ; in, Indicates the first The model parameters are updated in the next iteration. Indicates the first The current model parameters at the next iteration The learning rate is used to control the step size of parameter updates. Indicates about parameters gradient calculation operator, The model represents the current parameters. For input samples The predicted output, Indicates input sample The corresponding real tags.

[0071] When the judgment result indicates that the factual data is incorrect, it means that the original data stored in the data base itself is incorrect and needs to be verified before deciding whether to write it back for correction. This invention first requires users to upload supporting materials related to the correction value (such as certificate images, thesis screenshots, etc.). Then, optical character recognition (OCR) technology is used to extract text from the supporting material images to obtain the recognized text. After obtaining the recognized text, this invention performs a weighted calculation on the following three data items: first, the text similarity between the recognized text and the user's correction value, used to measure the consistency between the content of the supporting materials and the correction value; second, the user's historical credit score or permission level, reflecting the credibility of the user's past correction behavior; and third, the degree of conflict between the correction value and existing records in the data base. The higher the degree of conflict, the greater the contradiction between the correction value and existing facts, and the corresponding weight is negative in the calculation. Subsequently, the above three items are weighted and summed to obtain the verification credibility score.

[0072] Specifically, a verification score is calculated by combining the OCR recognition results and the user's credit weight. This scoring formula differs from conventional text matching; its calculation formula is as follows: ; in, This indicates the verification credibility score of the corrected data. , , These represent the weight coefficients for the text consistency item, the user authority item, and the data conflict item, respectively. This represents a text similarity calculation function used to measure the consistency between the corrected values ​​submitted by the user and the identified content in the supporting materials. This represents the target correction value input by the user. This refers to the supporting images uploaded by the user. The text result obtained after optical character recognition. Indicates the user who performed the correction operation. Historical credit score or privilege level, This represents the conflict degree calculation function, used to evaluate the correction value. With database base The degree of contradiction between facts or logical rules.

[0073] Subsequently, the invention compares the verification credibility score with a preset credibility threshold. When the verification credibility score is not lower than the preset credibility threshold, the invention determines that the correction has passed verification, generates a reverse governance instruction, constructs an SQL update statement and a graph database update instruction, persistently writes the user's correction value to the corresponding structured storage record in the physical layer of the data base, and simultaneously updates the attribute values ​​of the corresponding entity nodes in the knowledge graph of the logical layer, completing the synchronous update process between the physical layer and the logical layer. When the verification credibility score is lower than the preset credibility threshold, or the user refuses to upload supporting materials, the invention only retains the user's correction value in the report view of the current archived report, does not perform any write operations on the data base, and simultaneously generates a highlighted exception work order in the background, records the correction log content and the reason for verification failure, and pushes it to the management end for manual intervention and verification.

[0074] like Figure 2 As shown, the present invention also provides a data-based adaptive teacher report generation system for executing a data-based adaptive teacher report generation method as described in any of the above claims, comprising: Module 100: Used to collect teachers' business data and perform data cleaning and entity alignment to build a data base that includes physical and logical layers; The parsing module 200 is configured as a visual semantic parsing engine, used to receive the report to be filled, and jointly parse the layout and header semantics of the report to obtain the header semantic tree. Query module 300: Based on the semantic tree of the table header, it performs query inference on the data base for different report fields using a semantic type hybrid matching strategy and knowledge graph path reasoning, and obtains a set of populated results for multiple fields; Correction module 400: Used to generate a pre-filled report based on the filled result set, and receive user verification and correction interaction processing to generate an archived report.

[0075] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0076] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the adaptive generation method of teacher reports based on a data base described in various embodiments or some parts of embodiments.

[0077] The beneficial effects of this invention are that it effectively solves the pain points of traditional teacher report generation, such as scattered data, tedious manual filling, easy errors, and low efficiency. It realizes the automation, intelligence and accuracy of report generation. Furthermore, by building a unified data base, it integrates teacher business data from multiple systems and completes entity alignment, breaks down data silos, ensures data consistency and integrity, and provides reliable data support for report filling. Furthermore, this invention leverages a visual semantic parsing engine to accurately analyze report layout and header semantics, combining semantic matching and knowledge graph reasoning to achieve automatic field mapping and calculation, significantly reducing manual intervention and improving report generation efficiency. This invention also continuously optimizes models and data driven by user verification and correction logs, constantly improving report population accuracy and reducing manual verification costs. Meanwhile, the data reverse governance mechanism of this invention further improves the data foundation, forming a closed loop of "generation-verification-optimization", which is compatible with different types and formats of teacher reports. It is highly versatile and can be widely used in teaching management scenarios, reducing the workload of management personnel and improving the digitalization and refinement of education management.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A data-based adaptive method for generating teacher reports, characterized in that, include: S1. Collect teachers' business data and perform data cleaning and entity alignment to build a data base that includes physical and logical layers; S2. Input the form to be filled into the visual semantic parsing engine, and perform joint parsing of the form layout and header semantics to obtain the header semantic tree; S3. Based on the semantic tree of the table header, query and infer the data base for different report fields using a semantic type hybrid matching strategy and knowledge graph path reasoning respectively, to obtain a set of populated results for multiple fields; S4. Generate a pre-filled report based on the filled result set, and receive user verification and correction interaction processing to generate an archived report.

2. The adaptive generation method for teacher reports based on a data base according to claim 1, characterized in that, Step S1 further includes: S11. Extract teacher business data from the teaching affairs system, personnel system, financial system, scientific research system, external academic databases and unstructured documents using ETL tools, and perform regular expression cleaning on the teacher business data to obtain a standardized dataset. S12. For teacher nodes in multiple source systems in the standardized dataset, the corresponding name semantic vector, title one-hot encoding vector and department embedding are concatenated to construct the initial feature vector of multiple nodes. S13. Input the initial feature vector into the graph attention network, and aggregate the neighbor node information by weighting the attention coefficients to obtain the updated feature vectors of multiple nodes; S14. Based on the updated feature vector, calculate the alignment similarity score of cross-system node pairs, and filter out cross-system node pairs of the same entity by comparing the alignment similarity score with a preset threshold. S15. Assign globally unique identifiers to cross-system node pairs, store the aligned data in the physical layer of the data base, and construct a semantic relationship graph between entities in the logical layer.

3. The adaptive generation method for teacher reports based on a data base according to claim 2, characterized in that, In step S13, the attention coefficients between two adjacent nodes are obtained by performing a nonlinear transformation on the concatenation result of the learnable attention vector and the linearly transformed node features using a linear unit activation function with leakage correction; in step S14, the alignment similarity score is obtained by weighted summation of the cosine similarity of the updated feature vectors of cross-system node pairs and the Jaccard similarity of the attribute set.

4. The adaptive generation method for teacher reports based on a data base according to claim 1, characterized in that, Step S2 further includes: S21. Input the form to be filled into the document layout analysis model, and perform multimodal fusion of the text content, page coordinates and visual image features in the form to obtain the page feature representation of multiple cells; S22. Based on the layout feature representation, the hierarchical relationship between multi-level headers and merged cells is parsed and processed. The complete semantic paths of multiple fill slots from the root node to the leaf node are merged to obtain an ambiguous header semantic tree.

5. The adaptive generation method for teacher reports based on a data base according to claim 1, characterized in that, Step S3 further includes: S31. For the direct fields in the header semantic tree, the semantic feature vectors of multiple fields are compared with the semantic feature vectors of candidate data fields in the data base. A comprehensive score is calculated in combination with data type constraints. The optimal mapping result of the direct fields is obtained by selecting the candidate field with the highest comprehensive score. S32. For the complex logical fields in the semantic tree of the table header, starting from the entity in the logical layer of the data base, the entity relationship graph is traversed through multi-hop path search, and the subgraphs that meet the preset attribute condition constraints are aggregated and calculated to obtain the statistical values ​​of the complex logical fields. S33. Combine the optimal mapping result with the statistical value to obtain a filled result set.

6. The adaptive generation method for teacher reports based on a data base according to claim 1, characterized in that, Step S4 further includes: S41. Based on the filling result set, fill multiple field values ​​into the report template, and highlight the filling items whose confidence scores are lower than the preset threshold to obtain a pre-filled report. S42. Present the pre-filled report on the interactive front end, receive the user's verification and correction operations on the highlighted items and other fields, and package and archive the final report data to obtain the archived report.

7. The adaptive generation method for teacher reports based on a data base according to claim 1, characterized in that, Also includes: S5. Capture user correction behavior and generate correction logs. Based on the correction logs, determine the type of correction behavior and trigger online incremental update processing of model parameters and reverse data governance processing respectively to continuously optimize the visual semantic parsing engine and the data base.

8. The adaptive generation method for teacher reports based on a data base according to claim 7, characterized in that, Step S5 further includes: S51. Collect the user-inputted correction value, original fill value, field identifier, and timestamp, encapsulate and process them to obtain the correction log; S52. Based on the metadata attributes of the modified fields in the correction log, determine the type of the correction behavior and obtain the determination result; the determination result includes semantic understanding deviation and factual data error; S53. Based on the determination result, the visual semantic parsing engine and the data base are continuously optimized through online incremental update of model parameters and reverse data governance.

9. The adaptive generation method for teacher reports based on a data base according to claim 8, characterized in that, Step S53 specifically includes: When the judgment result is a semantic understanding deviation, the correction log is converted into a supervised training sample, and the parameters of the document layout analysis model in the visual semantic parsing engine are updated by gradient using an online incremental learning algorithm with a feedback loss function to obtain updated model parameters. When the determination result indicates that the factual data is incorrect, the supporting materials uploaded by the user are used as the identification text. The consistency between the identification text and the corrected value, the user's historical credit score, the corrected value, and the degree of conflict recorded in the data base are weighted and calculated to obtain a verification credibility score. When the verification credibility score is greater than or equal to a preset credibility threshold, a reverse governance instruction is generated to synchronously update the original records in the physical and logical layers of the data base. When the verification credibility score is less than the preset credibility threshold, the corrected value is retained in the archived report, and an abnormal work order is generated for verification processing.

10. A data-based adaptive teacher report generation system, used to execute the data-based adaptive teacher report generation method as described in any one of claims 1 to 9, characterized in that, include: Building module: Used to collect teachers' business data and perform data cleaning and entity alignment, building a data base that includes physical and logical layers; The parsing module is configured as a visual semantic parsing engine, which is used to receive the report to be filled, and jointly parse the layout and header semantics of the report to obtain the header semantic tree. Query module: Based on the semantic tree of the table header, it performs query inference on the data base for different report fields using a semantic type hybrid matching strategy and knowledge graph path reasoning, and obtains a set of populated results for multiple fields; Correction module: Used to generate a pre-filled report based on the filled result set, and to receive user verification and correction interaction processing to generate archived reports.

Images