Intelligence research-oriented natural language case-related element analysis method and system

Abstract

The application relates to the technical field of data processing and natural language understanding, and provides a natural language case-related element analysis method and system for intelligence research and judgment, which comprises the following steps: independently analyzing field annotations of target data in parallel through at least two analysis layers, outputting corresponding candidate results and layer confidence; writing the candidate results into a knowledge graph as candidate nodes to form a topological structure reflecting the context association of the data; performing message propagation on the knowledge graph through a graph neural network, and calculating a cross-layer consistency score based on the path of the candidate nodes in the topological structure; and performing fusion decision based on the layer confidence and the cross-layer consistency score, and outputting a data element analysis result of the field annotations. The application realizes context constraint and conflict resolution of multi-strategy analysis results, avoids error accumulation of a linear pipeline, and improves analysis accuracy and robustness in a complex scene.

Description

Technical Field

[0001] This application relates to the fields of data processing and natural language understanding, and in particular to a method and system for analyzing elements involved in a case using natural language for intelligence analysis. Background Technology

[0002] In intelligence analysis and data governance scenarios, business databases store massive amounts of structured data, and their field annotations typically contain rich natural language descriptions, which are key to understanding the semantics of the data. Existing technologies mainly employ methods such as dictionary matching, rule template splitting, or pre-trained model semantic classification to parse field annotations, attempting to transform unstructured annotation text into machine-understandable element tags to support subsequent data analysis and knowledge construction.

[0003] However, in actual deployment and operation, due to inconsistent annotation standards across different business systems, the same field often has multiple variations or near-synonyms, causing single matching strategies to frequently fail. Simultaneously, existing parsing architectures mostly employ a linear pipeline model of splitting, identifying, and extracting, allowing errors in earlier stages to propagate unidirectionally and amplify at each level. Furthermore, fields are parsed in isolation, making it impossible to utilize contextual information such as foreign keys between tables or relationships between fields within the same table for cross-validation and conflict resolution. This mismatch between the linear processing mechanism and the demands of complex semantic understanding leads to irreversible accumulation of parsing errors and the inability to detect cross-field semantic conflicts, severely restricting the accuracy and reliability of intelligence analysis. Therefore, a data element parsing solution that can integrate multi-source contextual information and break the linear error propagation chain is urgently needed. Summary of the Invention

[0004] To address the issues of error accumulation and lack of context fusion mechanism in existing linear parsing architectures, this application proposes a natural language analysis method and system for analyzing elements involved in cases for intelligence assessment, which improves parsing accuracy and enhances robustness.

[0005] To achieve the above objectives, this application adopts the following technical solution: The target data's field annotations are independently parsed in parallel using at least two parsing layers, outputting corresponding candidate results and layer-inherent confidence scores. These candidate results are then written into a knowledge graph containing data table nodes, field nodes, and inter-field edges to form a topological structure reflecting the data's contextual relationships. A graph neural network is used for message propagation on the knowledge graph, and cross-layer consistency scores are calculated based on the paths of candidate nodes within the topological structure. Finally, a fusion decision is made based on the layer-inherent confidence scores and cross-layer consistency scores, outputting the data element parsing results for the field annotations.

[0006] The above scheme maps multi-layer parsing results to knowledge graph nodes and uses the message propagation of graph neural networks on the topology to quantify cross-layer consistency, thereby realizing contextual constraints and conflict resolution of multi-strategy parsing results and avoiding error accumulation in linear pipelines.

[0007] In conjunction with the first aspect mentioned above, in one possible implementation, the parsing layer includes at least two of the following: a rule parsing layer, a semantic vector matching layer, and a large model inference parsing layer: the selected at least two parsing layers are executed in parallel, and each outputs candidate results independently.

[0008] In conjunction with the first aspect mentioned above, in one possible implementation, the internal processing flow of the rule parsing layer, semantic vector matching layer, and large model inference parsing layer includes: the rule parsing layer performs adaptive delimiter recognition by querying delimiter pattern nodes in the knowledge graph, splits field annotations, matches the split natural language parts with a pre-built set of synonyms, calculates the edit distance between annotations that do not match precisely and standard words, and calculates the layer's built-in confidence based on the matching results and edit distance; the semantic vector matching layer uses a semantic encoding model to encode field annotations into annotation semantic vectors, calculates the cosine similarity between the annotation semantic vectors and each vector in the feature type vector library, selects the feature type with the highest similarity as the candidate result, and uses the cosine similarity as the layer's built-in confidence; the large model inference parsing layer constructs prompt words containing contextual information from the knowledge graph, calls the large language model for generative parsing, obtains structured output, and calculates the layer's built-in confidence based on the large language model's self-assessment confidence and rule validation pass rate.

[0009] In conjunction with the first aspect mentioned above, in one possible implementation, the process of calculating the cross-layer consistency score based on the path of the candidate node in the topology includes: the graph neural network employs a hybrid strategy of attention-weighted summation and mean aggregation to aggregate messages of the neighboring nodes of the candidate node to obtain updated node features; based on the updated node features, the semantic graph distance between candidate node pairs corresponding to the same field annotation is calculated; and based on the semantic graph distance and the layer-inherent confidence of each candidate node pair, the cross-layer consistency score is calculated.

[0010] In conjunction with the first aspect mentioned above, in one possible implementation, based on the updated node features, the semantic graph distance between candidate node pairs corresponding to the same field annotation is calculated, including: defining the semantic graph distance as the sum of the edge weights of the shortest path between the two candidate nodes in the candidate node pair, calculated using the following formula: ; in, Candidate nodes and candidate nodes Semantic graph distance between them Candidate nodes With candidate nodes The set of all paths between them for One of the paths, For the edges in the path, for Preset weights; Based on semantic graph distance and candidate nodes' built-in confidence in their respective layers, cross-layer consistency scores are calculated, including: The semantic graph distance is mapped to a consistency score using an exponential decay function, and a weighted average is calculated using the lower layer-built-in confidence among candidate node pairs. The formula is as follows:

[0011] in, Cross-level consistency score for field annotations. For consistent distance scale parameters, and Candidate nodes and The built-in confidence level of the layer.

[0012] In conjunction with the first aspect mentioned above, in one possible implementation, the knowledge graph also includes standard reference nodes, which are constructed based on industry data standards. The method also includes: calculating the graph distance from the candidate node to the nearest standard reference node, calculating the anchor matching score based on the graph distance using an exponential decay function, and incorporating the anchor matching score into the fusion decision-making considerations.

[0013] In conjunction with the first aspect mentioned above, one possible implementation incorporates anchor point matching scores into the fusion decision-making considerations. This includes: weighting and fusing layer-embedded confidence, cross-layer consistency scores, and anchor point matching scores to obtain a comprehensive confidence score; selecting the candidate node with the highest comprehensive confidence score as the data element parsing result output; and triggering Bayesian optimization to adaptively update the weights of layer-embedded confidence, cross-layer consistency scores, and anchor point matching scores in response to the accumulated manually reviewed data reaching a preset threshold.

[0014] In conjunction with the first aspect mentioned above, one possible implementation also includes a reverse feedback mechanism: For any two different parsing layers among at least two parsing layers, when the confidence level of the first candidate node output by the first parsing layer, obtained by fusing the layer-embedded confidence level and the cross-layer consistency score, is higher than the corresponding confidence level of the second candidate node output by the second parsing layer, it is determined whether the semantic graph distance between the first candidate node and the second candidate node in the knowledge graph is greater than a preset conflict threshold; if it is greater than the preset conflict threshold, the signal of the first parsing layer is transmitted to the second parsing layer through the reverse edge in the knowledge graph; the layer-embedded confidence level of the second candidate node is corrected based on the transmitted signal of the first parsing layer, and the corrected layer-embedded confidence level is re-substituted into the fusion decision; wherein, the semantic graph distance is calculated based on the path of the candidate node in the topology.

[0015] In conjunction with the first aspect mentioned above, one possible implementation also includes: when the cross-layer consistency score is lower than a preset confidence threshold, locating the conflict path that causes the cross-layer consistency score to decrease through a graph neural network and triggering a manual review process; receiving the manual review correction results for the conflict path and injecting the manual review correction results into the knowledge graph in the form of new triples to achieve incremental learning of the knowledge graph.

[0016] Furthermore, this application also provides a natural language analysis system for crime-related elements oriented to intelligence assessment, comprising: a multi-layer parallel parsing module, used to independently parse field annotations of target data in parallel through at least two parsing layers, and output corresponding candidate results and layer-embedded confidence scores; a knowledge graph construction module, used to write candidate results as candidate nodes into a knowledge graph, the knowledge graph containing a topological structure reflecting the contextual relationships of the data; a cross-layer consistency evaluation module, used to perform message propagation on the knowledge graph through a graph neural network, and calculate cross-layer consistency scores based on the paths of candidate nodes in the topological structure; and a fusion decision module, used to make fusion decisions based on layer-embedded confidence scores and cross-layer consistency scores, and output the data element parsing results of field annotations.

[0017] This application provides a method and system for parsing natural language elements involved in intelligence analysis. It transforms the task of parsing discrete field annotations into a node reasoning problem on a graph topology by constructing a knowledge graph containing data tables, fields, and related edges. This enriches the previously isolated parsing objects with rich contextual semantic support. Furthermore, it employs at least two parallel and independent parsing layers, outputting candidate results with built-in layer confidence scores. This overcomes the inherent flaw of unidirectional error propagation and accumulation in traditional linear pipelines, ensuring that local errors in any parsing layer do not directly contaminate subsequent stages. Simultaneously, it utilizes graph neural networks for message propagation on the knowledge graph, calculating cross-layer consistency scores based on the topological paths between candidate nodes. These scores quantify the degree of mutual corroboration between different parsing strategies in a unified semantic space. When multi-layer results converge in the graph structure, the cross-layer consistency score increases the credibility of the parsing results; conversely, it suppresses conflicting results. Ultimately, by making decisions using the built-in reliability of the fusion layer and the cross-layer consistency score, we can achieve dual verification of single-layer technical reliability and multi-layer semantic consistency, thereby improving parsing accuracy and system robustness in complex and non-standard annotation scenarios while ensuring parsing efficiency. Attached Figure Description

[0018] Figure 1 A flowchart illustrating the natural language analysis method for intelligence assessment provided in this application embodiment; Figure 2 This application provides a diagram illustrating the internal processing logic of a multi-layered parallel parsing layer in a natural language analysis method for intelligence assessment. Figure 3 Detailed flowchart of cross-layer consistency assessment in the natural language crime element parsing method for intelligence analysis provided in the embodiments of this application; Figure 4 This application provides a logical diagram for calculating semantic graph distance and cross-layer consistency score in a natural language element parsing method for intelligence analysis. Figure 5 A flowchart illustrating the introduction of standard reference nodes and the three-dimensional fusion decision-making process in the natural language-based crime element analysis method for intelligence assessment provided in this application embodiment; Figure 6 A flowchart illustrating the reverse feedback mechanism and incremental learning process in the natural language crime element analysis method for intelligence assessment provided in this application embodiment; Figure 7 An architecture diagram of a natural language crime element parsing system for intelligence analysis provided in this application embodiment; Figure 8 This is an example diagram illustrating an application scenario for a natural language-based crime element analysis system for intelligence assessment provided in this application embodiment. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0021] Example 1: like Figure 1 As shown, this embodiment provides a natural language analysis method for crime-related elements in intelligence assessment. This method achieves accurate semantic understanding of complex field annotations by constructing a multi-layered parallel parsing and graph topology fusion mechanism. Specifically, the method includes the following steps: Step S100: The field annotations of the target data are independently parsed in parallel by at least two parsing layers, and the corresponding candidate results and layer built-in confidence are output.

[0022] Parallelism of at least two parsing layers means that multiple parsing strategies process the same field annotation at the same time dimension. There is no data dependency, calling relationship or order between the parsing layers. That is, it is a non-serial pipeline mode and a non-cascaded filtering mode.

[0023] Layer built-in reliability is a quantitative evaluation index that characterizes the credibility of the output result of a single parsing layer based on its own algorithmic logic, reflecting the inherent reliability of this specific technical means under the current input conditions.

[0024] In some implementations, after receiving the target data field annotations from the business database, the system simultaneously distributes them to multiple configured heterogeneous parsing layers. At this time, each parsing layer will independently execute the parsing task according to its own preset parsing logic (such as rule-based matching, semantic vector-based retrieval, or large model inference), generate one or more candidate results without interference, and calculate the corresponding layer's built-in confidence based on its own matching score, similarity value, or generation probability and other internal indicators.

[0025] For example, one parsing layer might output "personnel identifier" as a candidate result and give a layer built-in confidence score of 0.85; another parsing layer might output "identity number" as a candidate result and give a layer built-in confidence score of 0.72.

[0026] It should be noted that this parallel independent parsing design breaks the chain of unidirectional error propagation in traditional linear pipeline architectures. Through parallel mechanisms, it ensures that a local failure or misjudgment in any parsing layer will not directly affect the normal operation of other parsing layers, preserving multiple independent evidence pathways. This provides the necessary preconditions for subsequent cross-validation and error correction using multi-source information, thereby improving the system's fault tolerance and robustness when facing non-standard annotations.

[0027] Step S200: Write the candidate results as candidate nodes into the knowledge graph. The knowledge graph includes data table nodes, field nodes, and inter-field association edges to form a topological structure that reflects the contextual relationships of the data.

[0028] In this context, a candidate node refers to an entity node in a knowledge graph that carries and represents the candidate results and their metadata (such as source level, layer built-in confidence, etc.) output by each parsing layer. It serves as a semantic bridge connecting unstructured text parsing results with structured graph knowledge.

[0029] The topology that reflects the contextual relationships of data refers to a network diagram structure composed of data table nodes, field nodes, and their hierarchical relationships, foreign key references, semantic relationships, etc. This structure explicitly encodes the inherent connections between data elements at the business logic level.

[0030] In some implementations, after obtaining the candidate results output by each parsing layer, the system dynamically creates corresponding candidate nodes in the knowledge graph and associates them with the field nodes corresponding to the field annotations through specific edges. Simultaneously, the knowledge graph maintains rich contextual edges in advance or in real-time. For example, field nodes are connected to their respective data table nodes through attribution edges, foreign key relationships are established between field nodes of different tables through reference edges, and co-occurrence constraints are established between fields of the same table through semantic edges.

[0031] It should be understood that although this embodiment describes specific node and edge types, other graph model structures that can express data context can also be used in other implementations, as long as they can provide sufficient topological information for subsequent graph neural network inference.

[0032] It's important to note that this step maps discrete candidate results to nodes in a graph, essentially transforming the originally isolated text classification problem into a node-based reasoning problem within a structured space. This ensures that the parsing results are no longer suspended in a vacuum but are anchored within a specific business context. When the annotation of a field is ambiguous, its connected table structures, related fields, business domain information, and other topological neighbors in the graph can serve as strong constraints for disambiguation, thereby resolving the issues of synonym confusion and cross-field semantic conflicts caused by the lack of context awareness in existing technologies.

[0033] Step S300: Message propagation is performed on the knowledge graph through a graph neural network, and cross-layer consistency scores are calculated based on the paths of candidate nodes in the topology.

[0034] Among them, the cross-layer consistency score is an indicator used to quantify the degree to which the candidate results output by different parsing layers for the same field annotation mutually corroborate each other in the semantic space of the knowledge graph. Its value reflects the level of consensus reached by multi-strategy parsing results under the constraints of structured context.

[0035] Message propagation in graph neural networks is a computational process of iterative information aggregation and feature updating on a graph structure. By passing and aggregating information from neighboring nodes along the edges, the feature vector of each node can encode its local and even global topological context information.

[0036] In some implementations, the system inputs a knowledge graph containing candidate nodes and their layer-specific built-in confidence scores into a pre-trained graph neural network model. The model performs multiple rounds of message propagation. In each round, candidate nodes aggregate feature information from their neighboring nodes (including field nodes, table nodes, other candidate nodes, etc.). After several rounds of propagation, the feature vectors of candidate nodes not only contain their own parsing information but also incorporate their contextual semantics along the graph topological path.

[0037] Subsequently, based on the updated candidate node features, the system calculates the semantic relevance or distance between different candidate node pairs under the same field, and combines their respective layer-inherent confidence scores to transform them into cross-layer consistency scores through a preset mapping function (such as an exponential decay function).

[0038] For example, if the ID card number output by the rule layer and the citizen identity number output by the large model layer point to the same standard data element node in the graph through the belonging relationship edge, and there is a strong correlation edge between the field nodes where the two are located, then the semantic graph distance between them is short, and the calculated cross-layer consistency score is high; conversely, if the two have no correlation in the graph or even point to conflicting concepts, then the consistency score is low.

[0039] It should be noted that this step utilizes graph neural networks to mine the structured prior knowledge implicit in the graph topology, placing multiple independent parsing results in a unified semantic space for cross-validation. When the results of multiple heterogeneous strategies tend to converge in the graph structure, it indicates that the result is supported by multi-dimensional evidence, and its credibility is structurally enhanced; conversely, conflicting results are identified and suppressed by the topological structure. This joint evaluation mechanism based on structured context effectively overcomes the blind spots of single strategies, improves the accuracy and stability of parsing results, and achieves a leap from single-layer self-verification to multi-layer mutual verification.

[0040] Step S400: Based on the layer-inherent confidence and cross-layer consistency score, perform a fusion decision and output the data element parsing results of the field annotations.

[0041] Among them, fusion decision refers to the process of selecting the optimal parsing result by comprehensively considering the technical reliability of the candidate result itself (layer-in-layer reliability) and the consistency support of the external structured environment (cross-layer consistency score) through weighted calculation or other decision logic.

[0042] In some implementations, the system performs a weighted sum of the layer-specific confidence score and the corresponding cross-layer consistency score for each candidate result under the same field, according to preset weights, to obtain a comprehensive confidence score. Then, the candidate result with the highest comprehensive confidence score is selected as the final data element parsing result output for that field annotation.

[0043] If the highest overall confidence level is still lower than the preset safety threshold, it can be marked as a low confidence result and trigger manual review or further processing.

[0044] It should be understood that the specific form of fusion decision-making is not limited to weighted summation. It can also be based on ranking learning, classification models, or other multi-criteria decision-making methods, as long as it utilizes information from both the layer-inherent reliability and cross-layer consistency scores.

[0045] It should be noted that the dual-track fusion decision-making mechanism introduces global semantic constraints provided by the graph topology as a correction signal (reflected through cross-layer consistency scores). When a layer gives a low layer-inherent confidence due to extremely non-standard annotations, but its results are highly consistent with the results of other layers in the graph structure and conform to business logic, the cross-layer consistency score can compensate for and improve it, preventing correct results from being mistakenly rejected. Conversely, when a layer gives a high layer-inherent confidence, but its results seriously conflict with the context in the graph, the cross-layer consistency score will penalize it, preventing high-confidence erroneous results from misleading downstream applications. This dynamic and adaptive decision-making mechanism ensures that the final output parsing results are not only statistically reliable but also semantically reasonable, thereby significantly improving the overall quality and usability of data element parsing in intelligence analysis scenarios.

[0046] Based on the above technical solution, this embodiment maps candidate results from multi-layer parallel parsing to knowledge graph nodes and utilizes message propagation in the topology of graph neural networks to quantify cross-layer consistency, constructing a complete technical closed loop from parallel parsing to graph carrying, then to topological inference, and finally to dual-track fusion. This breaks through the persistent error accumulation problem of linear pipelines at the architectural level. It preserves multiple sources of evidence through parallel mechanisms, injects contextual constraints through the graph structure, achieves multi-strategy mutual verification through GNN inference, and finally outputs parsing results with both technical reliability and semantic rationality through fusion decision-making. This achieves synergy, enabling the system to maintain high accuracy and strong robustness even when facing complex challenges commonly found in real intelligence analysis environments, such as non-standard annotations, diverse expressions, and cross-field conflicts.

[0047] Example 2: like Figure 2 As shown, this embodiment is a further detailed explanation of step S100 in embodiment 1. Based on embodiment 1, this application provides a preferred combination of parsing layers, namely, the parsing layers include at least two of the following: rule parsing layer, semantic vector matching layer, and large model inference parsing layer. The selected at least two parsing layers are executed in parallel, and each outputs candidate results independently. This allows for the construction of a multi-dimensional evidence capture network by utilizing the positive complementarity of different technical means in the feature space.

[0048] Specifically, the rule parsing layer excels at capturing literal patterns and structured tags, the semantic vector matching layer excels at understanding abstract semantics and synonymous expressions, and the large model inference parsing layer possesses the cognitive ability to handle complex contexts and long-tail ambiguities. These three layers are not simply functional overlaps, but rather cover the complete parsing spectrum from precise matching to fuzzy semantics to deep inference.

[0049] More importantly, each parsing layer does not block or depend on the others during execution. The output of any layer is written into the knowledge graph as an independent candidate node. This ensures that even if a layer fails due to a specific type of noise, other layers can still retain the correct parsing path. This provides sufficient and diverse decision-making basis for subsequent cross-layer consistency assessment, avoiding the risk of full-link collapse caused by a single point of failure in the traditional serial pipeline.

[0050] The specific implementation logic and built-in confidence calculation mechanism of the three parsing layers mentioned above are described in detail in this embodiment as follows: For the rule parsing layer, its internal processing flow includes: By querying delimiter pattern nodes in the knowledge graph, delimiter adaptive recognition is performed, field annotations are split, the split natural language parts are matched with a pre-built set of synonyms, the edit distance between the annotations that do not match precisely and the standard words is calculated, and the built-in confidence of the layer is calculated based on the matching results and edit distance.

[0051] Among them, adaptive delimiter recognition means that the system does not rely on hard-coded fixed symbols (such as only recognizing "|" or ";"), but dynamically accesses the pre-statistical delimiter configuration nodes in the knowledge graph to obtain the delimiter patterns and priorities that frequently appear in the current data source or similar tables, and then performs tentative splitting of field annotations accordingly.

[0052] Edit distance is a quantitative metric that measures the degree of difference between two strings. It is usually expressed as the minimum number of single-character edit operations required to convert one string into another.

[0053] In some implementations, the rule parsing layer first attempts to segment the original annotation text sequentially based on the delimiter priority sequence fed back from the graph, separating the natural language description part from the structured metadata part. Then, the natural language part is precisely matched against a pre-defined standard element thesaurus; if a match is successful, a high built-in layer confidence score (e.g., 0.95) is directly assigned. If no precise match is found, the edit distance between the annotation text and each term in the standard thesaurus is calculated, and the term with the smallest distance and below a preset threshold (e.g., 30% of the maximum length) is selected as the fuzzy matching result. The confidence score is then calculated based on the normalized edit distance, using the following formula: ; In the formula, The edit distance after normalization. To edit distance, and These represent the lengths of the annotation text and the standard terms, respectively.

[0054] It should be noted that this layer effectively solves the splitting error problem caused by inconsistent delimiters in existing technologies by introducing a graph-driven delimiter adaptive mechanism, enabling the rule engine to dynamically adapt to the evolution of the data source. Meanwhile, the confidence calculation based on edit distance provides a fine-grained literal similarity measure, allowing the system to still provide discriminative confidence scores when faced with typos, mixed simplified and traditional Chinese characters, or slight variations in expression, rather than a simple binary right or wrong judgment.

[0055] For the semantic vector matching layer, its internal processing flow includes: The semantic encoding model is used to encode field annotations into annotation semantic vectors, and the cosine similarity between the annotation semantic vectors and each vector in the feature type vector library is calculated. The feature type with the highest similarity is selected as the candidate result, and the cosine similarity is used as the layer built-in confidence.

[0056] Among them, semantic coding model specifically refers to a pre-trained language model that has undergone sentence-level semantic fine-tuning. It can map variable-length natural language text into dense vectors of fixed dimensions, making semantically similar texts geometrically closer in vector space.

[0057] Cosine similarity is an indicator that measures the consistency of the directions of two vectors. Its value ranges from [-1, 1] and it directly reflects the degree of semantic overlap in the normalized vector space.

[0058] In some implementations, the system pre-uses the same semantic encoding model to encode all predefined feature type names and their standard descriptions into vectors, thus constructing a feature type vector library.

[0059] During the parsing phase, the annotations of the fields to be processed are input into the same model to obtain the annotation semantic vectors. Then, through approximate nearest neighbor search or brute-force retrieval, the cosine similarity between the vector vector and all vectors in the vector library is calculated, using the formula:

[0060] In the formula, For cosine similarity, For annotation semantic vectors, This is a vector of feature types.

[0061] The system selects the element type with the highest similarity that exceeds a preset threshold (e.g., 0.75) as the candidate result, and uses this similarity value directly as the layer's built-in confidence.

[0062] It should be noted that when field annotations use expressions that are completely different from the literal meaning of standard terms but semantically equivalent (such as unique personnel identifiers, identity IDs, and national ID numbers), the rule layer may fail, but the vector layer, due to its generalization ability in continuous semantic space, can still accurately capture the associations. Using cosine similarity directly as the confidence score ensures that the output of this layer includes not only classification results but also a quantitative representation of semantic certainty. This confidence assessment based on high-dimensional geometric distance and the rule layer's assessment based on discrete string edit distance form a perfect positive interaction, enhancing the system's robustness to diverse expressions.

[0063] For the large model inference parsing layer, its internal processing flow includes: Construct prompt words containing contextual information from the knowledge graph, call the large language model for generative parsing to obtain structured output, and calculate the pass rate based on the self-assessment confidence and rule verification of the large language model, with built-in confidence in the layer.

[0064] Among them, the prompt words containing contextual information from the knowledge graph refer to the fact that when constructing the input of a large model, not only are the annotations of the fields to be parsed included, but also relevant environmental information retrieved from the knowledge graph is dynamically injected, such as the parsed results of other fields in the same table, the foreign key references of the field, and the business domain tags of the table to which it belongs, so that the large model can perform reasoning with a local and global perspective.

[0065] Structured output refers to forcing large models to return parsing results in a predefined JSON Schema format, including fields such as feature type, association relationship, and self-assessment confidence level, to ensure that the output can be automatically processed by downstream programs.

[0066] In some implementations, the system first extracts a contextual summary based on the neighbor relationships of the current field node in the graph and assembles it into the example-guided prompt. Then, it calls a privately deployed large language model to generate the parsing results.

[0067] After the model outputs, the system performs a double check: first, it extracts the confidence score of the model's self-assessment. Second, run the rule validator to check whether the output element type is within the valid set, whether the associated target exists, and whether the format is compliant, and calculate the validation pass rate. (Number of valid items / Total number of valid items), final layer built-in confidence The calculation formula is: .

[0068] It's worth noting that by injecting graph context into the example-guided prompts, the problem of general-purpose large models easily generating illusions in vertical domains is effectively mitigated, making their reasoning process constrained by the business structure. Furthermore, a confidence calculation strategy combining self-assessment and external assessment is employed. This leverages the large model's own probability estimation capabilities while objectively calibrating it through external rule verification, preventing the propagation of errors caused by the model's blind confidence. This design transforms the large model layer from an uncontrollable black box into a quantifiable and verifiable intelligent parsing component, whose output confidence truly reflects the reliability of the generated results within the current business context.

[0069] Based on the above technical solution, this embodiment constructs a three-layer heterogeneous parallel parsing system consisting of rules, vectors, and a large model. Each layer is designed with a differentiated built-in confidence calculation mechanism to match its technical characteristics, achieving full-coverage parsing capabilities from literal matching and semantic understanding to cognitive reasoning. The three layers are independent and complementary, with each layer capable of independently producing candidate nodes with confidence labels. This not only breaks the error accumulation chain of linear pipelines but also provides rich, diverse, and measurable input signals for subsequent cross-layer consistency evaluation of graph neural networks. This refined parallel parsing design ensures that the system can always find at least one effective parsing path when facing field annotations in real-world intelligence analysis scenarios. Furthermore, the confidence mechanism explicitly expresses uncertainty, laying a solid foundation for the final high-precision fusion decision.

[0070] Example 3: like Figure 3-4 As shown, this embodiment is a further detailed explanation of step S300 in Embodiment 1. Based on Embodiment 1, this application provides a specific implementation of cross-layer consistency evaluation based on graph neural network topology propagation.

[0071] Specifically, the process of calculating cross-layer consistency scores based on the paths of candidate nodes in the topology includes: the graph neural network employs a hybrid strategy of attention-weighted summation and mean aggregation to aggregate messages of the candidate node's neighboring nodes, obtaining updated node features; based on the updated node features, the semantic graph distance between candidate node pairs corresponding to the same field annotation is calculated; and based on the semantic graph distance and the layer-specific built-in confidence of each candidate node pair, the cross-layer consistency score is calculated. This transforms the abstract graph topology into a quantifiable mathematical indicator, providing precise numerical evidence for the fusion decision-making of multi-strategy analysis results.

[0072] Step S301: The graph neural network uses a hybrid strategy of attention-weighted summation and mean aggregation to aggregate messages of the neighboring nodes of the candidate node to obtain the updated node features.

[0073] The hybrid strategy of attention-weighted summation and mean aggregation refers to using two complementary aggregation mechanisms to process neighbor node information during the message passing stage of the graph neural network: on the one hand, the attention mechanism dynamically learns and assigns higher weights to important neighbor nodes to capture key semantic association signals; on the other hand, mean aggregation preserves the statistical distribution characteristics of all neighbor nodes to prevent the loss of global context due to excessive attention focus.

[0074] The updated node features refer to the node vector representation that integrates local neighborhood information and global topology after multiple layers of message propagation and aggregation operations. This vector not only contains the node's own attributes, but also encodes its structural role in the graph.

[0075] In some implementations, when the system performs graph neural network inference, for each target candidate node, it first collects the features of all its directly connected neighbor nodes, and then performs two-way aggregation calculations in parallel: the first path calculates attention weights by concatenating the features of the target node and its neighbor nodes, performing a nonlinear transformation to obtain attention coefficients, and then performing Softmax normalization, thereby weighting and summing the neighbor messages to highlight the contributions of nodes that are more semantically relevant to the current parsing task (such as foreign key fields in a table, standard data element anchors, etc.); the second path directly calculates the arithmetic mean of all neighbor messages as a supplementary smoothing background information.

[0076] Finally, the two aggregation results are fused according to a preset ratio (e.g., 0.7:0.3), and the feature vector of the target node is updated through a gated recurrent unit or residual connection. This process is usually iterated for 3 layers to ensure that the information can cover the neighbor range within 3 hops, avoiding information silos caused by too shallow layers and preventing oversmoothing problems caused by too deep layers.

[0077] It should be noted that pure attention mechanisms are prone to failure when faced with sparse connections or noisy edges, while pure mean aggregation struggles to distinguish between core and peripheral associations. This hybrid aggregation strategy effectively addresses the limitations of single aggregation mechanisms in intelligence analysis scenarios. Through this hybrid strategy, the system can accurately capture the strong reference relationship between a suspect's ID number and the personnel table (.SFZH), while also understanding the overall business domain atmosphere of the table containing that field. This results in updated node features possessing richer and more robust semantic representation capabilities, laying a high-quality feature foundation for subsequent distance calculations.

[0078] Step S302: Based on the updated node features, calculate the semantic graph distance between candidate node pairs corresponding to the same field annotation.

[0079] Semantic graph distance is a quantitative indicator used to measure the degree of semantic connectivity between two candidate nodes in the knowledge graph topology. Its value reflects the cost or resistance required to establish a semantic association between the two. The smaller the distance, the closer the two are in terms of business logic.

[0080] In this embodiment, the semantic graph distance is defined as the sum of the edge weights of the shortest path between two candidate nodes in a candidate node pair, and the calculation formula is as follows:

[0081] In the formula, This represents the semantic graph distance between candidate node Ri and candidate node Rj. This represents the set of all paths between candidate node Ri and candidate node Rj. Indicates one of the paths. Represents the edges in path P. Representing an edge The preset weights.

[0082] In some implementations, after obtaining the updated node features, the system performs a restricted shortest path search algorithm (such as Dijkstra's algorithm or breadth-first search) for each pair of candidate nodes under the same field annotation.

[0083] During the search process, the weight of the edges Instead of a uniform weight of 1, the weight is differentiated based on the semantic strength represented by the edge type. For example, an edge indicating that a candidate node directly labels a certain feature type has the strongest semantic association and is weighted at 1. The edge indicating that the feature type is mapped to the standard data element represents an indirect association, and its weight is set to 0.5. An edge indicates that the two elements have appeared consistently in historical data, and its weight is set to 0.3. Edges representing foreign key references between fields, which are related but cross entity boundaries, have a weight of 1.5; edges representing relationships between tables to which fields belong are mainly used for structural location rather than semantic matching, and have a weight of 2.

[0084] The search depth is limited to 3 hops to avoid introducing excessive irrelevant noise. If two candidate nodes are unreachable within the limited depth, a very large penalty value or the default maximum distance is applied.

[0085] It's important to note that this distance calculation method based on weighted shortest paths essentially transforms the discrete graph topology into a continuous semantic metric space. Compared to simple hop count statistics, introducing edge weights can more precisely represent the semantic differences between different association types. This allows two candidates directly pointing to the same standard data element to receive a closer distance score than two candidates belonging only to the same table. This ensures that cross-layer consistency assessment accurately reflects the true association strength at the business logic level, rather than merely the proximity at the physical storage level, thereby improving the accuracy of conflict resolution.

[0086] Step S303: Calculate the cross-layer consistency score based on the semantic graph distance and the candidate node’s built-in confidence in its respective layer.

[0087] Among them, the cross-layer consistency score is a composite evaluation index that combines the proximity of candidate node pairs in the graph topology with the credibility of their respective analysis results. It is used to screen out those combinations of analysis results that are both mutually corroborative and individually reliable.

[0088] In this embodiment, the semantic graph distance is mapped to a consistency score using an exponential decay function, and a weighted average is calculated using the lower layer-inherent confidence among candidate node pairs as the weight. The calculation formula is as follows:

[0089] In the formula, Indicates field comments Cross-level consistency score; This is a consistent distance scale parameter used to control the distance decay rate; and Representing candidate nodes respectively and candidate nodes The built-in confidence level of the layer.

[0090] In some implementations, the system first utilizes an exponential decay function. The semantic graph distance calculated in step S302 is converted into a similarity score within the interval [0,1]. The parameter σ is usually calibrated based on the average path length of the graph (e.g., set to 2.0) so that the similarity under the typical association distance is in the sensitive interval.

[0091] Subsequently, for each pair of candidate nodes, the smaller of their built-in layer confidence scores is taken. This serves as the consistency weight for the pair. Then, the sum of "similarity × weight" for all pairs is divided by the sum of "weights" to obtain the normalized cross-layer consistency score.

[0092] It's important to note that this formula not only uses a min function instead of a product or average to combine layer-embedded confidence, but also reflects that if one of the two candidate nodes has extremely low layer-embedded confidence (e.g., 0.3 given by a failed match in the rule layer), its contribution to the overall consistency score will be strictly suppressed, even if they are close in the graph. This effectively prevents false consistency from misleading the fusion decision caused by two low-quality results accidentally being close in the graph. Furthermore, the use of exponential decay rather than nonlinear decay to map distance aligns with the cognitive pattern that semantic association weakens sharply with path growth. When two candidate nodes are slightly far apart, their consistency scores drop rapidly, thus strengthening the discriminative power of closely related signals. Therefore, utilizing this dual constraint mechanism, the cross-layer consistency score becomes a highly robust quality filter, only giving a high score when the multi-strategy results are both highly confident and topologically tightly converged, thus providing an extremely reliable safety guarantee for the final fusion decision.

[0093] Based on the above technical solutions, this embodiment constructs a rigorous and interpretable cross-layer evaluation mathematical model by combining the hybrid aggregation of graph neural networks, weighted path distance calculation, and consensus fusion formula under Min constraints. This model not only transforms abstract graph topology into precise numerical signals, but also, through the dual mechanisms of the Min function and exponential decay, eliminates interference from low-quality noise and long-distance spurious correlations. This enables the system to accurately identify truly credible multi-strategy consensus results in complex and ever-changing intelligence analysis data environments.

[0094] Example 4: like Figure 5 As shown, this embodiment is a further extension based on the aforementioned embodiments, aiming to introduce external industry standards as constraints and references for the analysis results.

[0095] Specifically, the knowledge graph also includes standard reference nodes, which are constructed based on industry data standards. The method also includes: calculating the graph distance from the candidate node to the nearest standard reference node, calculating the anchor matching score based on the graph distance using an exponential decay function, and incorporating the anchor matching score into the fusion decision-making considerations.

[0096] Among them, standard reference nodes refer to entity nodes in the knowledge graph that are specifically used to carry and represent data elements, qualifiers or business codes defined in authoritative data standards in a specific field (such as the GA / T series standards in the public security industry, the JR / T series standards in the financial industry, etc.), and constitute the semantic benchmark and compliance anchor of the entire parsing system.

[0097] Anchor point matching score is an indicator used to quantify the degree of alignment between candidate parsing results and industry authoritative standards, reflecting the legitimacy and standardization level of the results at the business specification level.

[0098] In some implementations, during the initialization phase of building the knowledge graph, the system automatically loads industry standard documents for the target domain, parses the data element identifiers, names, data types, and hierarchical relationships defined therein, creates a corresponding standard reference node for each standard entry, and establishes hierarchical edges (such as "belongs to" and "limited") between standard nodes as well as mapping edges (MAPS_TO) between standard nodes and feature type nodes.

[0099] During the parsing phase, for each generated candidate node, the system searches the graph for the shortest path to all standard reference nodes and selects the minimum graph distance. This serves as a measure of the deviation between the candidate system and the standard system. Then, it is further assessed using an exponential decay function. Convert it to an anchor matching score in the [0,1] interval. ,in A scale parameter to control the decay rate (e.g., set to 1.5); This is a standard reference node layer for all standard data element nodes (such as "DE00023 ID number") pre-built based on industry data standards (such as Public Security GA / T, Financial JR / T, etc.) in the knowledge graph; Candidate nodes The nearest graph distance (minimum graph distance) to the standard reference node layer (AnchorLayer) in the knowledge graph.

[0100] It should be noted that by introducing external truth constraints, a dual-track evaluation system, distinct from the one in Example 3 which relies solely on internal topology, is formed. In traditional technical solutions, the parsing results are often merely classification labels at the algorithm level, creating a semantic gap with the actual business standards implemented in the industry, leading to significant manual standardization conversion required in downstream systems. This embodiment, however, explicitly models the standard system as anchor layers in a graph and uses graph distance to quantify alignment, giving the parsing results business compliance attributes. When a candidate result appears semantically reasonable (high built-in layer confidence) and consistent with other fields in the local topology (high cross-layer consistency), but cannot be mapped to any valid standard data element, its anchor matching score will approach zero, effectively suppressing it in the fusion decision. This ensures seamless integration of the parsing results with the industry data standard system, reducing data governance costs.

[0101] Furthermore, the anchor matching score is incorporated into the fusion decision-making considerations, including: weighting and fusing the layer-inherent confidence score, cross-layer consistency score, and anchor matching score to obtain a comprehensive confidence score; selecting the candidate node with the highest comprehensive confidence score as the data element parsing result output; and triggering Bayesian optimization to adaptively update the weights of the layer-inherent confidence score, cross-layer consistency score, and anchor matching score in response to the accumulated manually reviewed data reaching a preset threshold.

[0102] Among them, the overall confidence level is the final decision index calculated by weighting and fusing the three dimensions of technical reliability, structural consistency and standard compliance (with weighting coefficients of α, β and γ).

[0103] In some implementations, the system sets initial weights to empirical values ​​(e.g., α=0.35, β=0.40, γ=0.25) and continuously collects correction records from human reviewers for low-confidence or conflicting results during operation. When the accumulated number of valid review samples reaches a preset threshold (e.g., 50), the system automatically triggers a round of weight optimization: using manually confirmed correct results from historical review samples as labels, and maximizing the overall confidence ranking of correct results as the objective function, a Bayesian optimizer is used to sample and evaluate in the weight space, ultimately updating the values ​​of α, β, and γ.

[0104] It should be understood that, in addition to Bayesian optimization, other parameter optimization methods such as grid search, genetic algorithms, or gradient-based online learning can also be used, as long as they can dynamically adjust the fusion weights using human feedback data.

[0105] It should be noted that in actual intelligence analysis environments, the emphasis on technical accuracy, structural consistency, and standard compliance of data from different business units and historical periods often changes dynamically. For example, in the early stages of data governance, standard alignment (higher γ) may be more important; while in the in-depth analysis stage, the consistency of cross-table relationships (higher β) may be more important. By introducing closed-loop feedback driven by manual review and Bayesian optimization, the system can automatically perceive and adapt to this shift in business focus, ensuring that the fusion weights always approach the optimal solution for the current scenario. This not only avoids the high operational costs of repeated manual parameter tuning but also allows the analytical model to become more accurate with use over time.

[0106] Based on the above technical solution, this embodiment constructs a standard reference node anchor layer and designs a three-dimensional fusion decision-making mechanism. This internalizes industry data standards into the core constraints of the parsing algorithm, solving the problem of the disconnect between existing technical parsing results and business standards. Simultaneously, by leveraging Bayesian optimization to achieve adaptive weight updates, a closed-loop self-evolutionary path is constructed from parsing to review and then to optimization. This enables the system to maintain optimal long-term performance and standard compliance in complex and ever-changing real-world business environments, enhancing the practical value and lifecycle of intelligence analysis data element parsing.

[0107] Example 5: like Figure 6 As shown, this embodiment, based on the previous embodiments, further introduces a dynamic correction and self-evolution mechanism, aiming to solve the problem that static analytical models are difficult to self-correct and continuously optimize when faced with complex and ever-changing real data.

[0108] Specifically, this application provides an implementation method including a reverse feedback mechanism: for any two different parsing layers among at least two parsing layers, when the confidence of the first candidate node output by the first parsing layer, obtained by fusing the layer-inherent confidence and cross-layer consistency score, is higher than the corresponding confidence of the second candidate node output by the second parsing layer, it is determined whether the semantic graph distance between the first candidate node and the second candidate node in the knowledge graph is greater than a preset conflict threshold; if it is greater than the preset conflict threshold, the signal of the first parsing layer is transmitted to the second parsing layer through the reverse edge in the knowledge graph; the layer-inherent confidence of the second candidate node is corrected based on the transmitted signal of the first parsing layer, and the corrected layer-inherent confidence is re-substituted into the fusion decision.

[0109] Among them, the reverse feedback mechanism refers to a dynamic adjustment strategy in a multi-layer parallel parsing architecture that allows high-confidence parsing results to influence and correct low-confidence parsing results in reverse. Its essence is to use the verified strong semantic associations in the graph structure to repair the cognitive biases of local parsing layers.

[0110] The preset conflict threshold is a critical value used to determine whether there is a substantial semantic contradiction between two candidate nodes. Only when the distance between the two nodes in the graph is far enough (i.e. the semantic difference is significant) and the confidence difference is large enough, is it considered a valid conflict rather than a normal fluctuation.

[0111] After the system completes the initial fusion decision, it will perform pairwise comparisons on all candidate node pairs under the same field: Suppose that the overall confidence of candidate node A output by the first parsing layer (such as the large model layer) is 0.85, while the overall confidence of candidate node B output by the second parsing layer (such as the rule layer) is only 0.45, and the semantic graph distance between the two in the knowledge graph is... The calculated result is 3.2, which exceeds the preset conflict threshold of 3.0.

[0112] At this point, the system determines that a significant conflict exists and that the higher-level results are more credible, thus activating the reverse feedback process: The system utilizes a pre-defined cross-layer reverse feedback path within the knowledge graph to transmit the high-confidence analytical signal, verified by the fusion decision, back to the analytical layer that generated the low-confidence candidate results, thus passing the high-confidence signal from node A to node B. The correction process uses a weighted interpolation formula: , in, The layer-inherent confidence is corrected for the second candidate node B. The original layer-built-in confidence level for the second candidate node B. The overall confidence level of the first candidate node A. The feedback learning rate (e.g., set to 0.3) is used to control the correction magnitude to prevent overcorrection. The corrected layer built-in confidence will replace the original value and re-participate in the fusion decision calculation in step S400.

[0113] It should be noted that in existing technologies, once the underlying rule parsing fails, the upper-level model often has no choice but to continue reasoning based on the erroneous input, leading to error accumulation. This embodiment, however, introduces a bidirectional signal path, enabling high-confidence upper-level results to backtrack and correct lower-level misjudgments, achieving dynamic balance and mutual verification between layers. Simultaneously, the design of dual triggering conditions (confidence difference + semantic conflict) effectively avoids erroneous corrections caused by accidental noise, ensuring the accuracy and security of feedback signals and improving the system's robustness in extremely non-standard annotation scenarios.

[0114] As another optional implementation, in order to deal with difficult cases with extremely low cross-layer consistency that cannot be automatically resolved, this application further includes: when the cross-layer consistency score is lower than a preset confidence threshold, locating the conflict path that causes the cross-layer consistency score to decrease through a graph neural network and triggering a manual review process; receiving the manual review correction results for the conflict path, and injecting the manual review correction results into the knowledge graph in the form of new triples to achieve incremental learning of the knowledge graph.

[0115] Among them, conflict paths refer to subgraph structures in a knowledge graph that connect multiple contradictory candidate nodes and make a major negative contribution to cross-layer consistency scores. They are used to accurately locate the root cause of the parsing divergence (such as missing mapping relationships or incorrect association edges).

[0116] Incremental learning refers to the ability to update the system's knowledge reserves in real time by adding new entities or relationships to the knowledge graph without retraining the entire graph neural network model.

[0117] In some implementations, when a field's annotation has a cross-level consistency score, When the score falls below a preset confidence threshold (e.g., 0.4), the system determines that the field is seriously ambiguous and the automatic fusion mechanism fails. At this point, the graph neural network will analyze the attention weight distribution or gradient contribution during message propagation, trace back and locate the key edges or node sequences that lead to the score reduction, and highlight them as conflict paths for human reviewers.

[0118] At this point, after reviewing the conflicting paths and related context on the interface, the reviewer provides the correct element type or association. Upon receiving the manual correction, the system converts it into standard triples, which are then written to the knowledge graph database in real time. These newly injected triples immediately alter the graph's topology, enabling the graph neural network to calculate shorter semantic graph distances and higher cross-layer consistency scores through the newly added paths when parsing identical or similar fields, thus automatically outputting the correct results.

[0119] It's important to note that by constructing a closed-loop evolutionary pathway from machine analysis to human oversight and knowledge accumulation, this approach distinguishes itself from traditional technologies where human review is merely used for post-production quality control. Instead, it directly embeds human review into the production process, instantly transforming human judgment into machine-understandable graph knowledge. This not only solves the problem of insufficient capacity to handle long-tail, obscure samples in the early stages of model deployment but also enables the system to self-evolve, becoming more accurate with use, significantly reducing the reliance on computing power and labeled data in long-term operation and maintenance.

[0120] Based on the above technical solution, this embodiment constructs a dual dynamic correction mechanism of reverse feedback and incremental learning, endowing the data element analysis system with the ability to correct and iterate. The reverse feedback mechanism utilizes the reverse edges in the graph topology to achieve bidirectional information flow between levels, effectively blocking the error accumulation chain of the linear pipeline, enabling high-confidence results to correct low-confidence biases in real time. The incremental learning mechanism transforms manual review from an offline process to an online knowledge injection interface, continuously optimizing the graph topology structure through the real-time injection of new triples, enabling the system to continuously adapt to the evolution of business standards and changes in data forms. The synergistic effect of these two mechanisms can resolve the contradiction between the accuracy bottleneck and high maintenance cost of static analysis models in complex intelligence analysis environments, ensuring high accuracy and high availability of data element analysis results during long-term operation.

[0121] Example 6: like Figure 7 As shown, this embodiment provides a natural language-based crime element analysis system for intelligence assessment. At the hardware level, the system includes a processor and a memory. The memory stores a computer program, and the processor executes this computer program to implement the steps in any of the aforementioned method embodiments. From a functional architecture perspective, the system mainly includes a multi-layer parallel parsing module, a knowledge graph construction module, a cross-layer consistency assessment module, and a fusion decision module.

[0122] The multi-layer parallel parsing module is used to independently parse the field annotations of the target data in parallel through at least two parsing layers, and output the corresponding candidate results and the layer's built-in confidence.

[0123] Among them, the multi-layer parallel parsing module is the data access and preliminary semantic extraction unit of the system. It integrates a variety of heterogeneous parsing engines, each of which is independent of each other and supports concurrent execution.

[0124] In some implementations, this module receives target data field annotations from the business database via message queues or shared memory mechanisms and simultaneously distributes them to multiple configured parsing service instances. For example, the rule parsing service, semantic vector matching service, and large model inference service can be deployed in different processes of the same server or on different servers. They listen for the same data input events, and after independently completing their parsing tasks, they encapsulate the generated candidate results and their layer-specific built-in confidence scores into standardized data structures (such as JSON objects) and return them to downstream modules.

[0125] It should be understood that although this embodiment lists three specific parsing service types, in other implementations, the number of parsing layers can be increased or decreased or the specific parsing algorithm can be replaced according to business needs, as long as parallel and independent parsing is satisfied.

[0126] It should be noted that this module adopts a parallel architecture design, which enables the system to make full use of multi-core CPUs or distributed computing resources when processing massive field annotations, thereby improving data throughput. At the same time, since the parsing layers do not block each other, the delay or failure of a single parsing service will not cause the entire parsing chain to stop, thus ensuring the stability and response speed of the system under high load conditions.

[0127] The knowledge graph construction module is used to write candidate results as candidate nodes into the knowledge graph, which contains a topological structure that reflects the contextual relationships of the data.

[0128] The knowledge graph construction module is the semantic storage and context management unit of the system. It is responsible for maintaining a dynamically updated graph structure database. This database not only stores temporary candidate nodes generated by parsing, but also persistently saves entities such as data tables, fields, and standard data elements, as well as the relationships between them, such as ownership, references, and mappings.

[0129] In some implementations, this module provides graph data writing and graph query interfaces. When it receives candidate results from the multi-layer parallel parsing module, it automatically creates corresponding candidate nodes and associates them with the corresponding field nodes to form relationship edges. At the same time, it is also responsible for synchronously updating the table structure and field relationships during system initialization or data source changes to ensure that the topology remains consistent with the underlying business database.

[0130] In terms of specific technology selection, the underlying layer of this module can use native graph databases such as Neo4j and JanusGraph, or it can be implemented through virtual layers based on relational databases, as long as it can efficiently support multi-hop path queries and neighbor node traversal.

[0131] It should be noted that this module anchors discrete parsing results within a structured context, providing the necessary topological foundation for subsequent graph neural network inference. By unifying the physical relationships (such as foreign keys) and semantic relationships (such as standard mappings) of business data as edges in a graph, the system can go beyond simple text matching and utilize global structural information to assist in local semantic disambiguation.

[0132] The cross-layer consistency evaluation module is used to propagate messages on the knowledge graph through graph neural networks and calculate cross-layer consistency scores based on the paths of candidate nodes in the topology.

[0133] Among them, the cross-layer consistency evaluation module is the core inference engine of the system. It loads a pre-trained graph neural network model and has the ability to dynamically extract subgraphs from the knowledge graph, execute message passing algorithms, and perform numerical calculations.

[0134] In some implementations, the module obtains candidate nodes to be evaluated and their neighborhood subgraphs from the knowledge graph construction module in batches, and converts them into tensor format input graph neural network models.

[0135] After the model performs message aggregation and feature updates at a preset number of layers, it outputs an enhanced feature representation for each candidate node. Subsequently, the distance calculation unit inside the module calculates the semantic graph distance between candidate node pairs under the same field based on these features and the edge weights in the graph. Through mathematical operations such as exponential decay mapping and Min weighted average, it finally generates a cross-layer consistency score.

[0136] It should be understood that although the specific calculation formulas are described in detail in the above method embodiments, at the system implementation level, these logics can be encapsulated as independent operator libraries or custom layers in deep learning frameworks, and the modules only need to focus on the data stream transformation of input and output.

[0137] It should be noted that this module transforms abstract topological structures into quantifiable quality indicators, achieving a leap from data to knowledge and then to decision-making basis. By engineering and modularizing the graph neural network inference process, complex semantic consistency verification can be embedded into the real-time parsing process with millisecond-level latency, ensuring both the depth of evaluation and meeting the performance requirements of the production environment.

[0138] The fusion decision module is used to make fusion decisions based on the layer's built-in confidence and cross-layer consistency scores, and outputs the data element parsing results of the field annotations.

[0139] Among them, the fusion decision module is the final decision and output control unit of the system. It is responsible for comprehensively evaluating multi-dimensional signals to determine the optimal parsing result and executing the diversion strategy according to the confidence level.

[0140] In some implementations, the module receives the layer-in-layer confidence from the multi-layer parallel parsing module and the cross-layer consistency score from the cross-layer consistency evaluation module (and also the anchor matching score if standard anchor evaluation is enabled), and performs weighted fusion according to preset or dynamically optimized weight parameters to obtain the comprehensive confidence.

[0141] For candidate nodes with a comprehensive confidence level higher than the safety threshold, they are directly output as the final data element parsing result to downstream applications; for cases with a confidence level lower than the threshold, an audit work order is automatically generated and pushed to the manual review interface, while the manual feedback result is written back to the knowledge graph to trigger incremental learning.

[0142] In addition, in some implementations, this module can also integrate reverse feedback logic, which calls the correction interface to update the lower-level confidence level when a significant conflict is detected.

[0143] It should be noted that this module acts as a safety valve and adapter for the system, ensuring the quality of the output results is controllable and business compliant. Through flexible threshold configuration and weight management mechanisms, the system can automatically balance accuracy and recall in different business scenarios, avoiding the contamination of downstream intelligence analysis by low-quality results, and achieving continuous evolution through a human-machine collaborative closed loop.

[0144] Based on the above technical solution, this embodiment constructs a complete data element parsing system architecture by defining functional modules and their interaction interfaces. Each module is physically implemented using a processor and memory, and logically forms a tight closed loop from parallel parsing to graph carrying, then to topological reasoning, and finally to fusion decision-making. This modular design allows each component of the system to be developed, tested, and upgraded independently, reducing engineering complexity and maintenance costs.

[0145] Example 7: like Figure 8 As shown, this embodiment demonstrates a specific application scenario based on the aforementioned method embodiments. It aims to verify the effectiveness of the multi-layer parallel parsing and graph topology fusion mechanism in solving complex field annotation parsing problems through real-world data processing cases in intelligence analysis. It should be understood that this embodiment is merely an illustrative example and not a limitation on the scope of protection of this application. Other scenarios involving the parsing of multi-source heterogeneous data elements (such as financial risk control, medical record structuring, etc.) can be implemented with reference to the logic of this embodiment.

[0146] In this application scenario, assuming the target data is a case database in a public security intelligence analysis system, there is a case information table. The original annotation for a certain field in this table is "Suspect ID|ref=Personnel Table.SFZH". This annotation contains both the non-standard natural language expression "Suspect ID" (the standard term should be "Citizen Identity Number" or "ID Card Number") and the structured metadata "ref=Personnel Table.SFZH" indicating cross-table relationships. Under a traditional linear pipeline parsing architecture, this field is highly susceptible to misparsing. Specifically, traditional methods typically perform delimiter splitting first. If the preset delimiter library does not contain "|" or the splitting logic fails to correctly handle mixed patterns, the entire annotation will be treated as plain text. Even if the splitting is successful, subsequent keyword matching will often classify it as "plain text" or "unknown type" because "Suspect ID" does not match the standard element dictionary. Due to the sequential dependency of each stage in the pipeline, misjudgments in the preceding stages will irreversibly propagate to the downstream stages, causing this field to lose its crucial personnel identity semantics, thus rendering intelligence correlation analysis based on this field ineffective. This parsing failure is caused by non-standard expression and disconnect from context.

[0147] When processing the annotation of this field using the parsing method provided in this application, the system parses it independently in parallel through at least two parsing layers. Specifically, the rule parsing layer first queries the delimiter pattern node in the knowledge graph, adaptively identifies "|" as a valid delimiter, and splits the annotation into the natural language part "Suspect ID" and the metadata part "ref=Personnel Table.SFZH". When performing synonym matching on "Suspect ID", since the standard word is not accurately matched, the rule layer calculates that its edit distance from the standard word is large, and can only output the candidate result "text" based on the fallback rule, and gives a low layer built-in confidence (e.g., 0.3). At the same time, the semantic vector matching layer is not affected by literal differences. It maps "Suspect ID" into a vector through a semantic encoding model, and retrieves the "Personnel Identifier" type that is semantically closest to it in the feature type vector library, outputs the candidate result "Personnel Element", and uses a high cosine similarity (e.g., 0.88) as the layer built-in confidence.

[0148] The large model reasoning parsing layer constructs a hint word containing the table name to which the field belongs, information of adjacent fields, and metadata content. It calls the large language model to perform deep reasoning. The model understands that "ref=personnel table.SFZH" clearly points to the ID number field of the personnel table, thus confirming that the current field is the personnel identification. It outputs the candidate result "personnel element" and gives a high layer built-in confidence (e.g., 0.92) by combining the self-assessed confidence and rule validation pass rate.

[0149] It should be noted that in this stage, the three parsing layers completed their respective judgments without interfering with each other. Although the rule layer made a misjudgment, the vector layer and the large model layer independently captured the correct semantic signals. This parallel mechanism preserved multiple evidence paths, avoided the collapse of the entire link caused by a single point of failure, and laid the data foundation for subsequent fusion and correction.

[0150] Subsequently, the system writes the three candidate results as candidate nodes into the knowledge graph and performs message propagation and cross-layer consistency evaluation on the topology using a graph neural network. In the knowledge graph subgraph of this case, the "Case Table.Suspect ID" field node is directly connected to the "Personnel Table.SFZH" field node through an edge representing the foreign key reference relationship between fields. The "Personnel Table.SFZH" node is in turn directly connected to the standard reference node "DE00023 (ID number)" through an edge representing the relationship between feature type and standard data element.

[0151] When the graph neural network performs message propagation, the candidate nodes for "personnel elements" output by the vector layer and the candidate nodes for "personnel elements" output by the large model layer are not only semantically similar, but also share a short path to the same standard reference node in the graph topology. At this time, the system calculates the semantic graph distance based on the updated node features and finds that the sum of the path weights between these two candidate nodes is extremely small (e.g., they only pass through a relation edge indicating that they have appeared consistently in historical data or shared neighbors). After exponential decay mapping, an extremely high cross-layer consistency score (e.g., 0.95) is obtained.

[0152] Conversely, the "text" candidate nodes output by the rule layer lack effective connections with "personnel table.SFZH" or standard data elements in the graph, and their semantic graph distance from the other two candidate nodes is extremely large, resulting in the corresponding cross-layer consistency score approaching zero.

[0153] It should be noted that this step fully utilizes the structured prior knowledge inherent in the database itself. The edges connecting foreign key references between fields in the graph are not only mappings of physical foreign keys, but also serve as strong constraint signals for semantic verification. Thus, the parsing results from the vector layer and the large model layer closely match the topological structure of the graph, structurally amplifying its credibility; while isolated misjudgments at the rule layer are automatically suppressed because they lack topological support. This context-based joint evaluation mechanism solves the problem that traditional methods, lacking a global perspective, cannot resolve local ambiguities.

[0154] Finally, the system makes a fusion decision based on the layer's built-in confidence and cross-layer consistency score, outputting the data element parsing result for the field annotation. In this case, although the layer's built-in confidence is low, the high built-in confidence of the vector layer and the large model layer, along with their extremely high cross-layer consistency score, results in a comprehensive confidence score for "Personnel Element" calculated by the fusion formula that far exceeds the safety threshold (e.g., 0.89), while the comprehensive confidence score for the "Text" candidate is extremely low (e.g., 0.15). The system ultimately selects "Personnel Element" as the parsing result output and can further accurately map it to the standard data element "DE00023" based on the anchor point matching score.

[0155] If the reverse feedback mechanism of Example 5 is adopted, the system can also use the high-confidence "personnel element" result to correct the built-in confidence of the rule layer through the reverse edge, or inject the correspondence between the "suspect ID" and "personnel identifier" as a new triple into the graph, so that the rule layer can directly hit the same expression in the future.

[0156] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of variations or substitutions within the technical scope disclosed in this application, such as adding or removing parsing layer types, adjusting graph neural network aggregation strategies, equivalent transformations of confidence fusion formulas, adaptive modifications to the knowledge graph topology, and recombination of features of various embodiments without departing from the core technical concept of this application. All of these should be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A natural language analysis method for crime-related elements in intelligence assessment, characterized in that, include: The field annotations of the target data are independently parsed in parallel by at least two parsing layers, and the corresponding candidate results and the built-in confidence of the layers are output. The candidate results are written into the knowledge graph as candidate nodes. The knowledge graph includes data table nodes, field nodes, and inter-field association edges to form a topological structure that reflects the contextual relationships of the data. Message propagation is performed on the knowledge graph using a graph neural network, and cross-layer consistency scores are calculated based on the paths of the candidate nodes in the topology. Based on the layer's built-in confidence and the cross-layer consistency score, a fusion decision is made, and the data element parsing results of the field annotations are output.

2. The natural language analysis method for crime-related elements oriented to intelligence assessment according to claim 1, characterized in that, The parsing layer includes at least two of the following: a rule parsing layer, a semantic vector matching layer, and a large model inference parsing layer: The selected at least two parsing layers are executed in parallel, and each outputs the candidate results independently.

3. The natural language analysis method for crime-related elements oriented towards intelligence assessment according to claim 2, characterized in that, The internal processing flow of the rule parsing layer, the semantic vector matching layer, and the large model inference parsing layer includes: The rule parsing layer performs adaptive delimiter recognition by querying delimiter pattern nodes in the knowledge graph, splits the field annotations, matches the split natural language parts with a pre-built set of synonyms, calculates the edit distance between the annotations that do not match precisely and the standard words, and calculates the built-in confidence of the layer based on the matching results and the edit distance. The semantic vector matching layer uses a semantic encoding model to encode the field annotations into annotation semantic vectors, calculates the cosine similarity between the annotation semantic vectors and each vector in the feature type vector library, selects the feature type with the highest similarity as the candidate result, and uses the cosine similarity as the built-in confidence of the layer. The large model reasoning parsing layer constructs prompt words containing contextual information from the knowledge graph, calls the large language model for generative parsing, obtains structured output, and calculates the built-in confidence of the layer based on the self-assessment confidence of the large language model and the rule validation pass rate.

4. The natural language analysis method for crime-related elements oriented to intelligence assessment according to claim 1, characterized in that, The process of calculating the cross-layer consistency score based on the path of the candidate node in the topology includes: The graph neural network employs a hybrid strategy of attention-weighted summation and mean aggregation to aggregate messages from the neighboring nodes of the candidate node, thereby obtaining updated node features. Based on the updated node features, calculate the semantic graph distance between candidate node pairs corresponding to the same field annotation; The cross-layer consistency score is calculated based on the semantic graph distance and the candidate node’s confidence in its respective layer.

5. The natural language analysis method for case-related elements oriented to intelligence assessment according to claim 4, characterized in that, The process of calculating the semantic graph distance between candidate node pairs corresponding to the same field annotation based on the updated node features includes: The semantic graph distance is defined as the sum of the edge weights of the shortest path between the two candidate nodes in the candidate node pair, and the calculation formula is as follows: ； in, Candidate nodes and candidate nodes Semantic graph distance between them Candidate nodes With candidate nodes The set of all paths between them for One of the paths, For the edges in the path, for Preset weights; The calculation of the cross-layer consistency score based on the semantic graph distance and the candidate node's built-in confidence for each layer includes: The semantic graph distance is mapped to a consistency score using an exponential decay function, and a weighted average is calculated using the lower layer-built-in confidence among the candidate node pairs. The calculation formula is as follows: in, Cross-level consistency score for field annotations. For consistent distance scale parameters, and Candidate nodes and The built-in confidence level of the layer.

6. The natural language analysis method for crime-related elements oriented to intelligence analysis according to claim 1, characterized in that, The knowledge graph also includes standard reference nodes, which are constructed based on industry data standards. The method further includes: Calculate the graph distance from the candidate node to the nearest standard reference node, and calculate the anchor matching score based on the graph distance using an exponential decay function; The anchor point matching score is incorporated into the fusion decision-making considerations.

7. The natural language analysis method for case-related elements oriented to intelligence assessment according to claim 6, characterized in that, The process of incorporating the anchor point matching score into the fusion decision considerations includes: The layer-in-the-layer confidence score, the cross-layer consistency score, and the anchor point matching score are weighted and fused to obtain the comprehensive confidence score. The candidate node with the highest overall confidence level is selected as the data element parsing result output. In response to the accumulation of manually reviewed data reaching a preset threshold, Bayesian optimization is triggered to adaptively update the weights of the layer's built-in confidence, the cross-layer consistency score, and the anchor point matching score.

8. The natural language analysis method for crime-related elements oriented to intelligence assessment according to claim 1, characterized in that, It also includes a reverse feedback mechanism: For any two different parsing layers among the at least two parsing layers, when the confidence score of the first candidate node output by the first parsing layer, obtained by fusing the layer's built-in confidence score and the cross-layer consistency score, is higher than the corresponding confidence score of the second candidate node output by the second parsing layer, it is determined whether the semantic graph distance between the first candidate node and the second candidate node in the knowledge graph is greater than a preset conflict threshold; if it is greater than the preset conflict threshold, the signal of the first parsing layer is transmitted to the second parsing layer through the reverse edge in the knowledge graph. The layer-inherent confidence of the second candidate node is corrected based on the transmitted signal from the first parsing layer, and the corrected layer-inherent confidence is then substituted back into the fusion decision. The semantic graph distance is calculated based on the path of the candidate node in the topology.

9. The natural language analysis method for crime-related elements oriented to intelligence assessment according to claim 1, characterized in that, Also includes: When the cross-layer consistency score is lower than a preset confidence threshold, the graph neural network is used to locate the conflict path that caused the cross-layer consistency score to decrease, and a manual review process is triggered. The system receives the manual review and correction results for the conflicting paths, and injects the manual review and correction results into the knowledge graph in the form of new triples to enable incremental learning of the knowledge graph.

10. A natural language-based crime element analysis system for intelligence assessment, characterized in that: The natural language analysis method for crime-related elements, applied to any one of claims 1-9, specifically includes: The multi-layer parallel parsing module is used to independently parse the field annotations of the target data in parallel through at least two parsing layers, and output the corresponding candidate results and the layer's built-in confidence. A knowledge graph construction module is used to write the candidate results as candidate nodes into a knowledge graph, wherein the knowledge graph contains a topological structure that reflects the contextual relationships of the data. The cross-layer consistency evaluation module is used to perform message propagation on the knowledge graph through a graph neural network and calculate the cross-layer consistency score based on the path of the candidate node in the topology. The fusion decision module is used to make fusion decisions based on the layer's built-in confidence and the cross-layer consistency score, and output the data element parsing results of the field annotations.

Images