A hierarchical semantic embedding-based enterprise system change training task triggering method and enterprise training management system

Abstract

This invention relates to a method for triggering training tasks related to corporate policy changes and a corporate training management system based on hierarchical semantic embedding. The method involves acquiring the old and new versions of the corporate policy text and a pre-defined association graph. After clause segmentation and comparison, the changed clauses are determined. The method also involves acquiring the text before and after the change, and identifying candidate knowledge points through multi-level semantic matching based on the expression of the change intent and the embedding of knowledge points. A versioned, heterogeneous, traceable impact graph is constructed. Candidate knowledge points are then subjected to structural consistency judgment within the impact graph. Affected knowledge points that simultaneously satisfy semantic matching and graph structure consistency constraints, along with their corresponding traceable paths, are selected. The affected training tasks are then identified and associated with their corresponding traceable paths, generating training task update trigger instructions. The system is implemented based on this method. This invention balances recall efficiency and matching accuracy, significantly reducing the risk of false triggering and improving the explainability and audit reproducibility of the trigger source.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing, and in particular to a method for triggering enterprise system change training tasks and an enterprise training management system based on hierarchical semantic embedding. Background Technology

[0002] Corporate employee training is typically based on management documents such as regulations and work instructions. These documents break down the requirements into job competency elements and assessable knowledge points, further developing into courses, question banks, or training task lists. As compliance requirements, processes, and risk control points are continuously updated, these regulations are frequently iterated, and clauses may be modified, added, or deleted.

[0003] To ensure that training content remains consistent with the latest regulations, companies need to promptly identify the affected knowledge points and corresponding training tasks after a change in regulations, and provide verifiable evidence (such as a traceability link from "changed clauses - knowledge points - training tasks") so that the training management system can trigger update, review, and retraining processes, thus avoiding management problems such as outdated training content, incomplete evidence chains, and difficulty in tracing responsibility.

[0004] In existing technical solutions, a common approach is to first compare the old and new versions of the policy by clauses or keywords to identify changed clauses. Then, based on a rule base (such as clause number mapping, keyword dictionary, manually maintained traceability matrix) or a text similarity model (such as TF-IDF, topic model, BERT general semantic vector, etc.), the "new version clause text" is matched with knowledge points, course or training task descriptions to identify potentially affected objects. Subsequently, the system generates a notification or work order, which is reviewed by training specialists to update courses and tasks. Some solutions also introduce knowledge graphs or traceability tables to establish static associations between clauses and knowledge points or courses. When a clause change is detected, the update process of the corresponding training task is directly triggered according to the association relationship.

[0005] The main shortcomings of the above solutions are as follows: First, they rely primarily on semantic similarity or keyword matching in the full text of the new clauses, making it difficult to depict the direction and extent of changes from old to new clauses. When the clauses are lengthy and contain extensive background descriptions, the matching scope may be expanded to knowledge points unrelated to the changes, or knowledge points corresponding to newly added or tightened requirements may be omitted. Second, even if knowledge points or training tasks are matched, there is often a lack of structured verification under the constraints of the company's existing traceability relationships, making it difficult to output verifiable traceability path evidence, resulting in insufficient interpretability and auditability of the updates. Summary of the Invention

[0006] This invention solves the problems existing in the prior art and provides a method for triggering enterprise system change training tasks and an enterprise training management system based on hierarchical semantic embedding.

[0007] The technical solution adopted in this invention is a method for triggering enterprise system change training tasks based on hierarchical semantic embedding, comprising the following steps:

[0008] S1 retrieves the old and new versions of the corporate policy text, as well as the preset relationship graph;

[0009] S2 divides and compares the old and new versions of the text into clauses, identifies the changed clauses, and obtains the text before and after the change for each changed clause.

[0010] S3 generates a change intent representation based on the text before and after the change to characterize the direction and magnitude of the change. Based on the change intent representation and the embedding of knowledge points, candidate knowledge points that meet the matching conditions are determined through multi-level semantic matching.

[0011] S4 constructs a versioned heterogeneous traceable impact graph based on change clauses, knowledge points, training tasks and association graphs. Candidate knowledge points are judged for structural consistency in the impact graph, and the affected knowledge points and their corresponding traceable paths that simultaneously meet the semantic matching and graph structure consistency constraints are selected.

[0012] S5 determines the affected training tasks based on the affected knowledge points and the coverage relationship between training tasks and knowledge points, and associates the corresponding traceable path with each affected training task.

[0013] S6 generates a training task update trigger instruction containing the identifier of the affected training task and its associated change clauses, along with a traceable path.

[0014] Preferably, in S1, the association graph includes knowledge points, the coverage relationship between training tasks and knowledge points, and the initial tracing relationship between clauses and knowledge points.

[0015] Preferably, S3 includes the following steps:

[0016] S3.1 Generate a first embedding and a second embedding for the text before and after the change, respectively;

[0017] S3.2 The first embedding, the second embedding, and the differential and interactive information derived from them are fused and transformed by the first neural network to output a representation of the change intention;

[0018] S3.3 Based on the change intent representation and knowledge point embedding, in the same semantic space, first use the low-dimensional representation to perform coarse-grained matching to filter out candidate knowledge points, and then use the high-dimensional representation to update the fine-grained semantic matching values ​​of the candidate knowledge points.

[0019] S3.4 outputs candidate knowledge points that meet the matching threshold.

[0020] Preferably, the first neural network concatenates the first embedding, the second embedding, the differential information, and the interaction information, and then performs a nonlinear transformation through a multi-layer fully connected network to extract semantic features that characterize the essence of the clause change, as a representation of the change intention.

[0021] Preferably, both the knowledge point embedding and the change intent representation are generated using a hierarchical semantic embedding model with truncation consistency, so that the low-dimensional representation is a prefix sub-vector of the high-dimensional representation;

[0022] When updating the semantic matching judgment value based on the high-dimensional representation, the fusion weights of the low-dimensional representation and the high-dimensional representation in the updated judgment value are adaptively adjusted according to the change magnitude indicated by the difference information.

[0023] Preferably, S4 includes the following steps:

[0024] S4.1 Using change clauses, knowledge points, and training tasks as nodes, and initial traceability relationships and coverage relationships as edges, construct a versioned heterogeneous traceable impact graph;

[0025] S4.2 In the influence graph, retrieve the traceable path from the node of the changed clause to the node of the candidate knowledge point, and remove the candidate knowledge points that do not have a path;

[0026] S4.3 For candidate knowledge points with traceable paths, the graph structure representation of nodes on the traceable path and the change intention representation are integrated to generate path consistency features, which are then input into the second neural network to calculate the structural consistency judgment value used to characterize the degree of relevance between the path and the clause change.

[0027] S4.4 Filter candidate knowledge points that meet the preset consistency threshold as affected knowledge points and output their corresponding traceable paths.

[0028] Preferably, when integrating the graph structure representation and change intention representation of nodes on a traceable path, the graph structure representation of nodes on the path is weighted and aggregated based on the semantic similarity of the truncated representations in the graph structure representation and change intention representation of each node to generate a path representation.

[0029] Preferably, in S5, each affected training task is associated with its corresponding traceability path, multiple traceability paths associated with the same affected training task are merged and deduplicated, and organized according to the identifier of the change clause of the triggering source to generate a structured traceability basis.

[0030] Preferably, in S6, the generated training task update trigger instruction includes structured information; the structured information includes a task identifier field, a trigger clause field, and a traceability path field.

[0031] An enterprise training management system includes a processor and a memory. The memory stores a computer program, and when the processor executes the computer program, it implements the enterprise policy change training task triggering method based on hierarchical semantic embedding.

[0032] This invention relates to a method for triggering training tasks for enterprise policy changes based on hierarchical semantic embedding and an enterprise training management system. The method involves acquiring the old and new versions of the enterprise policy text, as well as a pre-defined association graph. After clause segmentation and comparison, the changed clauses are determined. For each changed clause, the pre-change and post-change texts are obtained, generating a change intent representation to characterize the direction and magnitude of the change. Based on the change intent representation and the embedding of knowledge points, multi-level semantic matching is used to determine candidate knowledge points that meet the matching conditions. Based on the changed clauses, knowledge points, training tasks, and association graph, a versioned heterogeneous traceable impact graph is constructed. Candidate knowledge points are subjected to structural consistency judgment in the impact graph, and affected knowledge points and their corresponding traceable paths that simultaneously satisfy semantic matching and graph structure consistency constraints are selected. The affected training tasks are determined and associated with their corresponding traceable paths, generating a training task update trigger instruction. The system is implemented based on this method.

[0033] The beneficial effects of this invention are as follows:

[0034] (1) By generating a change intent representation by combining the text of the clause before and after the change, the matching judgment focuses on the changed part of the clause rather than the overall description, effectively solving the problem of scattered hits and omissions caused by the background information of long clauses. By utilizing the truncation consistency of Matryoshka Representation Learning, in the same semantic space, we first recall with coarse-grained truncation embedding and then recalculate with fine-grained non-truncation embedding to achieve homologous multi-granularity matching. Without introducing an additional semantic model, we can balance recall efficiency and matching accuracy, and significantly reduce the risk of false triggering.

[0035] (2) Construct a versioned heterogeneous traceable impact graph, introduce a graph adaptation layer to determine the structural consistency of candidate knowledge points, not only check whether there is an existing traceability path from the change clause to the candidate knowledge point, but also integrate the node representation in the path with the change intention representation to calculate the consistency value, put the semantic candidate back under the existing traceability edge constraint of the enterprise for secondary delimitation, ensure that the affected knowledge points simultaneously meet the semantic change constraint and traceability structure constraint, output verifiable traceability path as evidence, and improve the interpretability and audit reproducibility of the trigger source from the mechanism.

[0036] (3) A complete closed loop is formed from clause change analysis to training task triggering. The final triggering instruction carries complete "clause-knowledge point-task" path evidence, which is convenient for the system side to directly locate, verify and connect with automated updates. It is more in line with the core needs of reliably determining the scope of impact and providing verifiable evidence in the existing traceability system under the scenario of system change. Attached Figure Description

[0037] Figure 1 This is a flowchart of the method of the present invention;

[0038] Figure 2 This is a comparison diagram showing different matching methods.

[0039] Figure 3 This is a schematic diagram illustrating an embodiment of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] This invention relates to a method for triggering training tasks on enterprise policy changes based on hierarchical semantic embedding, comprising the following steps:

[0042] S1 retrieves the old and new versions of the corporate policy text, as well as the preset relationship graph;

[0043] S2 divides and compares the old and new versions of the text into clauses, identifies the changed clauses, and obtains the text before and after the change for each changed clause.

[0044] S3 generates a change intent representation based on the text before and after the change to characterize the direction and magnitude of the change. Based on the change intent representation and the embedding of knowledge points, candidate knowledge points that meet the matching conditions are determined through multi-level semantic matching.

[0045] S4 constructs a versioned heterogeneous traceable impact graph based on change clauses, knowledge points, training tasks and association graphs. Candidate knowledge points are judged for structural consistency in the impact graph, and the affected knowledge points and their corresponding traceable paths that simultaneously meet the semantic matching and graph structure consistency constraints are selected.

[0046] S5 determines the affected training tasks based on the affected knowledge points and the coverage relationship between training tasks and knowledge points, and associates the corresponding traceable path with each affected training task.

[0047] S6 generates a training task update trigger instruction containing the identifier of the affected training task and its associated change clauses, along with a traceable path.

[0048] The steps of the method are described in detail below with reference to the embodiments.

[0049] S1 retrieves the old and new versions of the corporate policy text, as well as the preset relationship graph;

[0050] In S1, the association graph includes knowledge points, the coverage relationship between training tasks and knowledge points, and the initial tracing relationship between clauses and knowledge points.

[0051] Specifically, the old version of the system or work instruction text is recorded as The new version text is denoted as Let K be the set of texts for each knowledge point unit, and let K be the coverage relationship between training task units and knowledge point units. The initial tracing relationship from clause unit to knowledge point unit is denoted as ;in, This is used for subsequent reverse lookup of affected training task units from affected knowledge point units. This is used to subsequently retrieve traceable paths within the versioned heterogeneous traceable influence graph and determine structural consistency.

[0052] In this embodiment, the association graph is composed of the coverage relationship between knowledge point unit K, training task unit, and knowledge point. The initial tracing relationship from clause unit to knowledge point The structured knowledge representation formed by these elements provides foundational data support for subsequent steps.

[0053] In this invention, the knowledge points originate from the text set of knowledge points in the company's existing training system. Generally, this set is pre-stored as preset basic data in the association graph. Similarly, the training tasks originate from the company's existing training management data. The knowledge points, training tasks, the coverage relationship between training tasks and knowledge points, and the initial traceability relationship between clauses and knowledge points in the association graph all belong to the company's preset structured basic data. Existing structured data can be directly read and imported from the training management system or system management materials. Missing parts can be manually supplemented according to the company's current traceability relationship. Therefore, the acquisition method can be system import, manual maintenance, or a combination of both. This invention does not limit this, only requiring that the association graph can be provided during execution. Based on this, this method constructs a versioned heterogeneous traceable influence graph and performs structural consistency judgment on candidate knowledge points, thereby filtering out the affected knowledge points and their corresponding traceable paths. Therefore, it does not rely on manual specification in this step.

[0054] S2 divides and compares the old and new versions of the text into clauses, identifies the changed clauses, and obtains the text before and after the change for each changed clause.

[0055] Specifically, S2 includes the following steps:

[0056] S2.1 to and The clause structure is analyzed separately. Clause structure analysis uses parsing rules based on paragraph header numbering and heading lines to extract clause hierarchy, clause number, clause title, and clause body. Clause numbering covers both consecutive numbering and chapter / section numbering patterns. Clause hierarchy is determined by the number of paragraphs separating the hierarchy within the clause number. A clause unit identifier (ID) is generated for each clause. The document identifier used to construct the ID is obtained from the system or work instruction identifier, and the version identifier is obtained from the file metadata or the version field in the filename. The clause title and clause body are concatenated sequentially to form the clause unit text, creating separate sets of old version clause units. With the new version of the terms and conditions unit set ;

[0057] S2.2 Use the clause unit identifier ID as the primary key pair and Align them one by one;

[0058] For clause units with inconsistent identifiers, the hierarchical position difference is first calculated based on their hierarchical location. This difference is obtained by comparing the sequence number of the old version clause unit within its hierarchical level with the sequence number of the new version clause unit within its hierarchical level. Then, the larger of the total number of clauses in the old and new versions at that level is used as the normalization benchmark to normalize the hierarchical position difference into a hierarchical position difference value. Next, the clause titles are segmented and word frequency vectors are constructed. Cosine similarity is used to calculate the text similarity of the clause titles. The text similarity of the clause titles and the hierarchical position difference value are weighted and synthesized according to preset weight parameters to obtain an alignment score. Old version clause units and new version clause units whose alignment scores meet the alignment threshold constitute alignment candidate pairs.

[0059] For example, in the old version, a clause unit titled "Requirements for Approval of Confined Space Operations" had an order number of 5 at that level. In the new version, the clause unit titled "Requirements for Permitting Confined Space Operations" had an order number of 6 at that level. If the old version had 12 clauses at that level and the new version had 13, the difference in level position would be the difference in their order numbers. Taking 13 as the normalization benchmark, the difference in level position would be 1 / 13. Furthermore, regarding the two clauses mentioned above... The clause titles are segmented into words and word frequency vectors are constructed. Cosine similarity is used to calculate the text similarity of the clause titles, assuming a result of 0.87. In the preset weight parameters, if the weight of the text similarity of the clause titles is 0.8 and the weight of the difference in hierarchical position is 0.2, then the alignment score of the two clause units can be obtained by weighting. If the alignment score meets the preset alignment threshold, then the old version clause unit and the new version clause unit are identified as alignment candidate pairs; if the preset alignment threshold is not met, then they are not considered as alignment candidate pairs for further processing.

[0060] To further demonstrate that the closer the positions and the more similar the titles, the higher the alignment score, the hierarchical position difference value can also be evaluated. The smaller the value, the closer the positions of the two clause units are in the original document structure. Combined with the weighted weight of the similarity of the clause title text, the resulting alignment score can be used to more stably identify clause units with changes in clause numbering and slight adjustments in position, but whose substantive content still corresponds.

[0061] The alignment map M is generated by selecting alignment candidate pairs in descending order of alignment score. During the generation process, each old version clause unit is constrained to match at most one new version clause unit and each new version clause unit is constrained to match at most one old version clause unit. Unmatched old version clause units are marked as candidates to be deleted and unmatched new version clause units are marked as candidates to be added.

[0062] S2.3 Based on the alignment mapping M, a text difference comparison is performed on each pair of clause units. The text difference comparison uses character-level edit distance to calculate the minimum number of editing operations between the old version clause unit text and the new version clause unit text. Editing operations include insertion, deletion and replacement, and the larger of the two text lengths is used as the normalization benchmark to obtain the difference degree. The difference degree is compared with the preset difference judgment threshold. Clause units with a difference degree lower than the difference judgment threshold are judged as unchanged clause units, and clause units with a difference degree not lower than the difference judgment threshold are judged as modified clause units.

[0063] Newly added candidates are classified as newly added clause units, and candidates to be deleted are classified as deleted clause units, thus forming a classification result of unchanged clause units, modified clause units, newly added clause units, and deleted clause units;

[0064] S2.4 grouped the modified clause units, added clause units, and deleted clause units into a set of change clause units. For each changed clause unit, generate both the original and revised texts. The original text for the revised clause unit is taken from... The corresponding clause unit text and the modified text are taken from [the relevant source]. The corresponding clause unit text in the middle, the text before the change of the newly added clause unit is set to empty text and the text after the change is taken from the text before the change. The modified text of the deleted clause unit is set to empty text, and the original text is taken from [the previous text]. Meanwhile, based on the unit identifier of the change clause in The process retrieves and binds the traceability edges associated with the change clause unit identifier, outputs a change clause unit record containing the change clause unit identifier, the text before the change, the text after the change, and the corresponding traceability edge. This serves as the input for S3 to generate the pre-change embedding and post-change embedding and form a change intent representation, and as the input for S4 to construct a versioned heterogeneous traceable influence graph and perform a traceable path existence determination.

[0065] S3 generates a change intent representation based on the text before and after the change to characterize the direction and magnitude of the change. Based on the change intent representation and the embedding of knowledge points, candidate knowledge points that meet the matching conditions are determined through multi-level semantic matching.

[0066] Specifically, it includes the following steps:

[0067] S3.1 Generate a first embedding and a second embedding for the text before and after the change, respectively;

[0068] Set of change clause units output by S2 Execute one by one; for any change clause unit, denote the text before the change as P and the text after the change as Q. When P is empty text or Q is empty text, replace the preset empty text marker sequence with empty text as input text, so that the subsequent embedding model outputs deterministic embedding results, ensuring that the newly added clause unit and the deleted clause unit enter a unified subsequent semantic matching link.

[0069] The first embedding is generated using a hierarchical semantic embedding model with truncation consistency, with P as input. (e.g., 1024-dimensional), using Q as input, the same model is used to generate a second embedding. (e.g., 1024 dimensions); This hierarchical semantic embedding model can adopt the Matryoshka Representation Learning architecture, which is characterized by the same text encoder outputting high-dimensional vectors and satisfying the truncation consistency constraint, so that the first low-dimensional sub-vectors of any output vector are in the same semantic space as the full-dimensional vector and can directly participate in similarity calculation.

[0070] S3.2 The first embedding, the second embedding, and the differential and interactive information derived from them are fused and transformed by the first neural network to output a representation of the change intention;

[0071] The first neural network concatenates the first embedding, the second embedding, the differential information, and the interaction information, and then performs a nonlinear transformation through a multi-layer fully connected network to extract semantic features that characterize the essence of the clause change, as a representation of the change intention.

[0072] Specifically, based on and Construct four features and concatenate them sequentially to form the change intent feature F. The four features are as follows: , , and Element-wise difference vector (difference information) and The element-wise interaction vector (interaction information) is obtained by subtracting the element-wise difference vector in each dimension and by multiplying the element-wise interaction vector in each dimension, so that F can simultaneously express the semantics before the change, the semantics after the change, the direction of change and the coupling relationship of change; F is input into the first neural network (i.e. the change intention representation generation network) to obtain the change intention representation Z.

[0073] In this embodiment, the first neural network is a multi-layer fully connected network, which sequentially includes an input layer, several hidden layers, and an output layer. For example, if its dimension is set to 4096, after passing through a 2048-dimensional first hidden layer, a 1024-dimensional second hidden layer, and a 512-dimensional third hidden layer, it finally outputs a 512-dimensional representation of the change intention Z. Each hidden layer uses a non-linear activation function to output the layer-by-layer transformation result, and the output Z serves as the unique semantic anchor point for subsequent semantic matching determination.

[0074] S3.3 Based on the change intent representation and knowledge point embedding, in the same semantic space, first use the low-dimensional representation to perform coarse-grained matching to filter out candidate knowledge points, and then use the high-dimensional representation to update the fine-grained semantic matching values ​​of the candidate knowledge points.

[0075] Both the knowledge point embedding and the change intent representation are generated using a hierarchical semantic embedding model with truncation consistency, so that the low-dimensional representation is a prefix sub-vector of the high-dimensional representation;

[0076] When updating the semantic matching judgment value based on the high-dimensional representation, the fusion weights of the low-dimensional representation and the high-dimensional representation in the updated judgment value are adaptively adjusted according to the change magnitude indicated by the difference information.

[0077] Specifically, S3.3 utilizes the truncation consistency property of the hierarchical semantic embedding model to generate a low-dimensional representation of Z, such as a 256-dimensional truncated representation. , and high-dimensional representation For example, a 1024-dimensional non-truncated representation One approach is to use a fully connected layer to map Z into a high-dimensional vector. The weight matrix of the fully connected mapping has a dimension of 512×1024, and the bias vector has a dimension of 1024. Then take... The first 256 dimensions as ,make and Satisfies the same-origin truncation constraint;

[0078] For each knowledge point unit text, Matryoshka Representation Learning (using the same hierarchical semantic embedding model) is used to generate knowledge point embeddings, denoted as the 1024-dimensional high-dimensional representation. 256-dimensional truncated embedding as Where k is the knowledge point unit identifier, Taken from The first 256 dimensions.

[0079] by With all Calculate coarse-grained semantic matching decision value for input. . Cosine similarity can be used for calculation. First, for and Perform L2 normalization on each component, then calculate the dot product of the normalized vectors as the score, and classify all knowledge point units according to... Sort the knowledge points from highest to lowest, and select the top N knowledge point units corresponding to the preset candidate quantity threshold N to form a candidate knowledge point unit set. .

[0080] exist Calculate the updated semantic matching judgment value for candidate knowledge point units within the scope. To ensure that fine-grained matching inherits the coverage constraints of coarse-grained matching and adaptively converges towards higher-dimensional semantic anchors as the clause changes, the fusion weights of low-dimensional and high-dimensional representations in the updated decision value can be adaptively adjusted based on the change magnitude indicated by the difference information. Specifically, the change magnitude m can be taken as the mean of the absolute values ​​of each dimension of the difference vector. The weight γ is obtained by normalizing m with a preset change magnitude threshold τ and truncating the interval. Then, the following formula can be used to calculate... And update the ranking results of the candidate knowledge point units.

[0081]

[0082] in, Candidate knowledge point units The updated semantic matching judgment value; Used as a unit identifier for knowledge points; For the magnitude of change With the threshold of change Determined weights; A 1024-dimensional non-truncated representation of the intended change; A 256-dimensional truncated representation of the intended change; Knowledge point unit 1024-dimensional non-truncated embedding; Knowledge point unit 256-dimensional truncated embedding; symbol Represents the vector dot product; symbol This represents the 2-norm of a vector.

[0083] In practical implementation, weight To determine dynamically, the difference information is first obtained from the first and second embeddings, and then the change magnitude is calculated based on the mean of the absolute values ​​of each dimension of the difference vector. and will Compared with the preset change threshold Normalization and interval truncation are performed to obtain the weights. ,like ,make It will automatically adjust according to the degree of change in the current terms; when the change is minor, Smaller, more semantic matching judgment values ​​are inherited after the update. and The coarse-grained matching results; when the change is large, The value is relatively large, and the updated semantic matching judgment value relies more on... and The fine-grained matching results; among which, the change magnitude threshold These can be given as preset parameters, preferably set by combining the difference distribution of historical system version samples, or determined through optimization during the training process; however, in actual implementation, the weights It is automatically calculated from the differential information of the current change clause unit.

[0084] S3.4 outputs candidate knowledge points that meet the matching threshold.

[0085] Update the semantic matching judgment value Matching threshold Compare and filter to find the satisfactory Not less than The candidate knowledge point units are output as the input candidate knowledge point unit set of S3. And carry the corresponding knowledge point unit for each output. For use in subsequent structural consistency determination.

[0086] like Figure 2 As shown, (a) uses the "overall semantics of the new clause" for matching, and the hit points in the semantic space are scattered over a wide area. It is easily biased by background semantics and redundant expressions. The candidate knowledge points are noisy, and although the recall is high, the precision is low. (b) first generates the "intention to change Δ" from the clauses before / after the change. The matching only focuses on the changed part, and the hits converge significantly in the semantic space. This method shifts the search focus from the "entire clause" to the "changed components", reducing false positives and manual screening costs, and improving the interpretability and stability of the candidate set.

[0087] S4 constructs a versioned heterogeneous traceable impact graph based on change clauses, knowledge points, training tasks and association graphs. Candidate knowledge points are judged for structural consistency in the impact graph, and the affected knowledge points and their corresponding traceable paths that simultaneously meet the semantic matching and graph structure consistency constraints are selected.

[0088] Specifically, it includes the following steps:

[0089] S4.1 Using change clauses, knowledge points, and training tasks as nodes, and initial traceability relationships and coverage relationships as edges, construct a versioned heterogeneous traceable impact graph;

[0090] S4.1.1 Construct a versioned, heterogeneous, traceable impact graph G;

[0091] The old and new version clause units are constructed as clause unit nodes, each carrying a clause unit identifier and a version identifier. Knowledge point units are constructed as knowledge point unit nodes, and training task units are constructed as training task unit nodes. The initial traceability relationship is established. Each traceability relationship is constructed as a traceability edge, with the starting point of the traceability edge being the clause unit node and the ending point being the knowledge point unit node. Version identifiers are added to the clause unit node and the traceability edge to distinguish between old and new versions. The coverage relationship between training task units and knowledge points is constructed as a traceability edge, with the starting point of the traceability edge being the training task unit node and the ending point being the knowledge point unit node. This traceability edge type is marked as a coverage traceability edge.

[0092] S4.1.2 configures the initial node characteristics for G;

[0093] For clause unit nodes and knowledge point unit nodes, the initial node features are generated using high-dimensional embeddings (e.g., 1024-dimensional) generated by S3. For training task unit nodes, the set of knowledge point unit nodes covered by the training task unit node is determined based on the coverage relationship. The high-dimensional embeddings of each knowledge point unit node in the set are aggregated (e.g., mean pooling) to obtain the initial node features. To distinguish the three types of nodes, a node type vector (e.g., 32-dimensional) can be generated for each type of node, and then concatenated with the initial node features to obtain the final node features.

[0094] S4.2 In the influence graph, retrieve the traceable path from the node of the changed clause to the node of the candidate knowledge point, and remove the candidate knowledge points that do not have a path;

[0095] For candidate knowledge point unit set Perform traceable path existence determination one by one; for each candidate knowledge point unit k, locate the corresponding knowledge point unit node in G, and take the clause unit node corresponding to the current changed clause unit as the starting node; retrieve the set of traceable paths from the clause unit node to the knowledge point unit node based on G. The retrieval can be performed by breadth-first traversal along the tracing edge direction, and a maximum path length L is set to limit the traversal depth. When Path(k) is empty, the candidate knowledge point unit k is removed.

[0096] S4.3 For candidate knowledge points with traceable paths, the graph structure representation of nodes on the traceable path and the change intention representation are integrated to generate path consistency features, which are then input into the second neural network to calculate the structural consistency judgment value used to characterize the degree of relevance between the path and the clause change.

[0097] When fusing the graph structure representation and change intention representation of nodes on a traceable path, the graph structure representation of nodes on the path is weighted and aggregated based on the semantic similarity of the truncated representations in the graph structure representation and change intention representation of each node, to generate the path representation.

[0098] Specifically, a graph neural network is used to update the node representation of G, resulting in a graph structure representation of the nodes. This graph structure representation is not manually set, but is obtained by updating the versioned heterogeneous traceable influence graph through a graph adaptation layer. In this embodiment, a relational graph convolutional network can be used as the graph adaptation layer, containing two relational graph convolutional layers. Each layer uses the traceability edge type as a relation parameter, aggregates the features of its incoming neighbor nodes according to the traceability edge type for each node, performs a relation-specific linear transformation, and then fuses it with the node's own features and outputs an updated node representation after nonlinear activation (used for path consistency determination). After processing, the in-graph representation of the clause unit node is obtained. Knowledge point unit node diagram representation Training task unit node diagram representation ;

[0099] For candidate knowledge points with traceable paths, perform path consistency determination; for any traceable path in Path(k), extract the in-graph representation of the knowledge point unit nodes appearing in the path in the path order, denoted as... Based on the semantic similarity between the graph structure representation of each node and the truncated representation in the change intention representation, the graph structure representation of the nodes on the path is weighted and aggregated to generate the path representation. Specifically, with As a semantic anchor point for path aggregation, for each Calculate the cosine similarity and perform exponential normalization to obtain the weight coefficients, then perform weighted aggregation to obtain the path representation. The following formula is used for calculation:

[0100]

[0101] in, It is a 256-dimensional path representation; The number of knowledge point unit nodes contained in the traceable path; The 256-dimensional truncated representation of the intended change; For the first Each knowledge point unit node is represented within the diagram; For the first Each knowledge point unit node is represented within the diagram; An index for traversing knowledge point unit nodes within a traceable path; An index for normalizing and summing knowledge point unit nodes within a traceable path; symbol Dot product of vectors; symbol It is the vector norm 2; This is for exponential operations.

[0102] Represent the path The path consistency feature V is obtained by concatenating the change intention representation Z with the path consistency feature V. V is then input into a second neural network (i.e., a path consistency determination network) to calculate the structural consistency determination value s. The second neural network can be a multi-layer fully connected network, consisting of an input layer, several hidden layers, and an output layer. The output layer can be a single neuron, outputting the value s, which, after sigmoid activation, ranges from 0 to 1, directly representing the degree of consistency between the traceable path and the current clause change. When multiple traceable paths exist for the same candidate knowledge point unit, s is calculated for each traceable path, and the traceable path with the largest s is selected as the corresponding traceable path for that candidate knowledge point unit.

[0103] S4.4 Filter candidate knowledge points that meet the preset consistency threshold as affected knowledge points and output their corresponding traceable paths.

[0104] The structural consistency judgment value s is compared with the consistency threshold η. Candidate knowledge points that satisfy s not less than η and have a corresponding traceability path are selected and output as the affected knowledge point unit. At the same time, the traceability path corresponding to the affected knowledge point unit is also output as the input of S5.

[0105] S5 determines the affected training tasks based on the affected knowledge points and the coverage relationship between training tasks and knowledge points, and associates the corresponding traceable path with each affected training task.

[0106] In S5, each affected training task is associated with its corresponding traceability path. Multiple traceability paths associated with the same affected training task are merged and deduplicated, and then organized according to the identifier of the change clause of the triggering source to generate a structured traceability basis.

[0107] Specifically, the affected knowledge point units and their corresponding traceability paths output by S4, as well as the coverage relationship between training task units and knowledge points, are used as inputs to complete the traceable aggregation mapping from affected knowledge point units to affected training task units.

[0108] Let the set of affected knowledge point units be denoted as For each affected knowledge point unit k, a corresponding tracing path is given. Record the coverage relationship between training task units and knowledge points as follows: Each relation record contains at least the training task unit identifier t and the covered knowledge point unit identifier k; Structured parsing is performed to obtain the set of knowledge point unit identifiers covered by each training task unit, and an inverted index is constructed. This refers to the set of training task unit identifiers covering knowledge point unit k. For each affected knowledge point unit k, based on... Retrieve training task units that cover the affected knowledge point unit and generate a mapping result from the affected knowledge point unit to the training task unit.

[0109] The mapping results are then merged according to training task units to obtain the set of affected training task units. The system aggregates the traceability paths corresponding to the affected knowledge point units it covers. For each affected training task unit t, it aggregates the traceability paths corresponding to the affected knowledge point units it covers, forming the association result from the training task unit to the traceability path. When the same training task unit t corresponds to When duplicate tracing paths exist, one path is retained and duplicates are removed to ensure that the output association results remain unique at the path granularity.

[0110] S6 generates a training task update trigger instruction containing the identifier of the affected training task and its associated change clauses, along with a traceable path.

[0111] In S6, the generated training task update trigger instruction includes structured information; the structured information includes a task identifier field, a trigger clause field, and a traceability path field.

[0112] Here, the set of affected training task units output by S5 and the result of the tracing path association are used as input to generate a training task update trigger instruction.

[0113] For each affected training task unit t and its associated tracing path For each tracing path, the change clause unit identifier u corresponding to the starting clause unit node of the tracing path is obtained through parsing. Specifically, the first node identifier is read from the node identifier sequence of the tracing path, and its recorded clause unit identifier is parsed as u; tracing paths associated with the same affected training task unit are merged according to the change clause unit identifier to obtain the set of tracing paths corresponding to training task unit t under the triggering of change clause unit identifier u. All occurrences of change clause unit identifiers are aggregated into a set of triggered change clause unit identifiers. .

[0114] Generate a training task update trigger instruction for each affected training task unit. . It contains structured information, specifically including:

[0115] Task Identifier Field: Enter the identifier t of the affected training task unit;

[0116] Trigger Clause Field: Contains the set of identifiers for the triggered change clause. ;

[0117] Traceability Path Field: Contains the set of traceability paths corresponding to each change clause unit identifier u. It can be stored using a structured record method with u as the key and the set of tracing paths as the value.

[0118] The generated training task update trigger instruction is sent to the training management system to trigger the training task update process corresponding to the affected training task unit. Specifically, the instruction message can be encapsulated according to the preset instruction field order, a transmission session is established using the interface address and authentication parameters agreed upon with the training management system, and the instruction message is sent to the trigger receiving end of the training management system. The training management system performs field integrity verification on the instruction message, locates the corresponding training task unit through the affected training task unit identifier, locates the trigger source through the triggered change clause unit identifier and the corresponding traceability path, calls the training task update process entry, and uses the corresponding traceability path as the traceability basis input for the update process to start the update process.

[0119] In this invention, it should be noted that a joint training method is adopted, in which the graph adaptation layer, the first neural network, and the second neural network participate in the training together. That is, the first neural network is not trained independently, but is trained together with the multi-level semantic matching process based on the change intent representation and knowledge point embedding, and the structural consistency judgment process of the second neural network. The data flow is as follows: the first embedding and the second embedding are generated from the text before and after the change. The change intent representation Z is obtained by the first neural network. Then, the semantic matching judgment value is calculated by Z and the knowledge point embedding. The structural consistency judgment value is calculated by combining the traceable path in the versioned heterogeneous traceable influence graph. The two together constrain the judgment result of the affected knowledge points.

[0120] Training data is preferably derived from historical enterprise policy version data, including at least old and new version texts, initial traceability relationships between clauses and knowledge points, coverage relationships between training tasks and knowledge points, and confirmed affected knowledge points or corresponding traceable paths in the historical update process. Supervision signals originate from these confirmed relationships. Truly related clauses and knowledge points, and truly traceable paths, are considered positive samples, while unrelated combinations are considered negative samples. The overall training objective, semantic consistency L, can be expressed as:

[0121]

[0122] in, and Both can employ binary cross-entropy loss to constrain semantic matching and structural consistency judgment values ​​respectively; thus, the effectiveness of the change intention representation Z is not set in isolation, but is directly guaranteed by the joint training process of subsequent knowledge point judgment results.

[0123] For the second neural network, the confirmed real traceable paths can be used as positive samples, and the candidate paths that do not correspond to the current clause change can be used as negative samples. The parameters are updated by the loss between the structural consistency judgment value and the real label. The node graph structure representation, change intention representation Z output by the graph adaptation layer and the judgment result of the second neural network converge together in the same training process. Its effectiveness is jointly guaranteed by the judgment results of the final affected knowledge points and traceable paths.

[0124] The present invention also relates to an enterprise training management system, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the enterprise system change training task triggering method based on hierarchical semantic embedding.

[0125] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0126] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0127] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0128] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0129] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0130] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for triggering enterprise system change training tasks based on hierarchical semantic embedding, characterized in that: Includes the following steps: S1 retrieves the old and new versions of the corporate policy text, as well as the preset relationship graph; S2 divides and compares the old and new versions of the text into clauses, identifies the changed clauses, and obtains the text before and after the change for each changed clause. S3 generates a change intent representation based on the text before and after the change to characterize the direction and magnitude of the change. Based on the change intent representation and the embedding of knowledge points, candidate knowledge points that meet the matching conditions are determined through multi-level semantic matching. S4 constructs a versioned heterogeneous traceable impact graph based on change clauses, knowledge points, training tasks and association graphs. Candidate knowledge points are judged for structural consistency in the impact graph, and the affected knowledge points and their corresponding traceable paths that simultaneously meet the semantic matching and graph structure consistency constraints are selected. S5 determines the affected training tasks based on the affected knowledge points and the coverage relationship between training tasks and knowledge points, and associates the corresponding traceable path with each affected training task. S6 generates a training task update trigger instruction containing the identifier of the affected training task and its associated change clauses, along with a traceable path.

2. The method for triggering enterprise system change training tasks based on hierarchical semantic embedding according to claim 1, characterized in that: In S1, the association graph includes knowledge points, the coverage relationship between training tasks and knowledge points, and the initial tracing relationship between clauses and knowledge points.

3. The method for triggering enterprise system change training tasks based on hierarchical semantic embedding according to claim 1, characterized in that: S3 includes the following steps: S3.1 Generate a first embedding and a second embedding for the text before and after the change, respectively; S3.2 The first embedding, the second embedding, and the differential and interactive information derived from them are fused and transformed by the first neural network to output a representation of the change intention; S3.3 Based on the change intent representation and knowledge point embedding, in the same semantic space, first use the low-dimensional representation to perform coarse-grained matching to filter out candidate knowledge points, and then use the high-dimensional representation to update the fine-grained semantic matching values ​​of the candidate knowledge points. S3.4 outputs candidate knowledge points that meet the matching threshold.

4. The method for triggering enterprise system change training tasks based on hierarchical semantic embedding according to claim 3, characterized in that: The first neural network concatenates the first embedding, the second embedding, the differential information, and the interaction information, and then performs a nonlinear transformation through a multi-layer fully connected network to extract semantic features that characterize the essence of the clause change, as a representation of the change intention.

5. The method for triggering enterprise system change training tasks based on hierarchical semantic embedding according to claim 3, characterized in that: Both the knowledge point embedding and the change intent representation are generated using a hierarchical semantic embedding model with truncation consistency, so that the low-dimensional representation is a prefix sub-vector of the high-dimensional representation; When updating the semantic matching judgment value based on the high-dimensional representation, the fusion weights of the low-dimensional representation and the high-dimensional representation in the updated judgment value are adaptively adjusted according to the change magnitude indicated by the difference information.

6. The method for triggering enterprise system change training tasks based on hierarchical semantic embedding according to claim 2, characterized in that: S4 includes the following steps: S4.1 Using change clauses, knowledge points, and training tasks as nodes, and initial traceability relationships and coverage relationships as edges, construct a versioned heterogeneous traceable impact graph; S4.2 In the influence graph, retrieve the traceable path from the node of the changed clause to the node of the candidate knowledge point, and remove the candidate knowledge points that do not have a path; S4.3 For candidate knowledge points with traceable paths, the graph structure representation of nodes on the traceable path and the change intention representation are integrated to generate path consistency features, which are then input into the second neural network to calculate the structural consistency judgment value used to characterize the degree of relevance between the path and the clause change. S4.4 Filter candidate knowledge points that meet the preset consistency threshold as affected knowledge points and output their corresponding traceable paths.

7. The method for triggering enterprise system change training tasks based on hierarchical semantic embedding according to claim 6, characterized in that: When fusing the graph structure representation and change intention representation of nodes on a traceable path, the graph structure representation of nodes on the path is weighted and aggregated based on the semantic similarity of the truncated representations in the graph structure representation and change intention representation of each node, to generate the path representation.

8. The method for triggering enterprise system change training tasks based on hierarchical semantic embedding according to claim 1, characterized in that: In S5, each affected training task is associated with its corresponding traceability path. Multiple traceability paths associated with the same affected training task are merged and deduplicated, and then organized according to the identifier of the change clause of the triggering source to generate a structured traceability basis.

9. The method for triggering enterprise system change training tasks based on hierarchical semantic embedding according to claim 1, characterized in that: In S6, the generated training task update trigger instruction includes structured information; the structured information includes a task identifier field, a trigger clause field, and a traceability path field.

10. A corporate training management system, characterized in that: It includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the enterprise system change training task triggering method based on hierarchical semantic embedding as described in any one of claims 1 to 9.

Images