Intelligent review system for system conflict points based on knowledge graph and large language model

Abstract

The application relates to the technical field of electric digital data processing, and discloses a system for intelligently examining system conflict points based on a knowledge graph and a large language model, which comprises the following steps: constructing a dual-mode storage space containing an unstructured index and a structured logical graph, analyzing a target text to extract features and triggering generation logic based on the graph; converting the topological structure of an associated subgraph into a natural language instruction sequence to construct a mandatory logical constraint template, filling the template with the text and inputting the pre-trained language model to generate a verification result; and writing the result back to the graph. According to the application, the discrete graph topological mapping is converted into linear logical constraints, the random divergence of the generative model is inhibited in the calculation principle, and the precise decoupling and dynamic evolution of unstructured semantics and structured logic are realized.

Description

Technical Field

[0001] This invention relates to an intelligent review system for institutional conflict points based on knowledge graphs and large language models, belonging to the field of electronic digital data processing technology. Background Technology

[0002] Currently, in the field of digital data processing, deep learning models based on the Transformer architecture, with their attention mechanism and massive parameter scale, are the core tools for processing large-scale unstructured text data. By encoding and decoding high-dimensional semantic features of natural language, they achieve semantic understanding and content generation, and are applied in general data processing scenarios such as information extraction, automatic question answering, and decision support.

[0003] However, when statistical probability generative models are applied to highly constrained tasks with stringent requirements for logical consistency and rule boundaries, such as consistency verification or conflict analysis of normative documents with complex reference relationships, the underlying prediction mechanism exhibits systematic defects due to its randomness. During the autoregressive generation process, the model tends to fit high-frequency co-occurring language patterns rather than following logical deduction paths, leading to uncontrollable divergences or factual illusions in the output results. Although vector retrieval enhancement techniques are introduced to supplement external knowledge, capturing textual semantic relevance based on vector space distance similarity matching cannot maintain the inherent hierarchical structure and reasoning topology of complex rule systems. This results in noise introduced due to retrieval granularity imbalance or loss of logical structure when processing long-chain logical reasoning, making it difficult for the computing system to guarantee semantic accuracy. While maintaining the breadth of recall, the accuracy of logical deduction is also important. For example, Chinese invention patent CN120833124A discloses a method and system for intelligent review of power grid drawings based on knowledge graphs. This solution focuses on extracting discrete knowledge triples from the normative text and determining compliance by comparing the drawing elements with the attribute values ​​of the knowledge base. Although this technical approach achieves preliminary structured storage of knowledge, the core operating mechanism relies on static entity attribute matching and shallow numerical retrieval. It fails to transform the implicit nested logic between rules, priority constraints, and complex deduction paths into constraints that generative models follow when reasoning. That is, when such systems face deep review tasks with conflicting higher and lower-level laws or multiple exception clauses, they lack the means to control the reasoning process and cannot avoid the inherent random divergence risk of generative models from the perspective of computational principles.

[0004] Therefore, the technical problem to be solved by this invention is how to construct a method that suppresses the random divergence of generative models from the perspective of computational principles, preserves the general semantic understanding ability of models while imposing deterministic logical topological constraints, and achieves efficient decoupling and dynamic interaction between unstructured semantic data and structured logical rules. Summary of the Invention

[0005] To address the problems raised in the background art, the technical solution of this invention is as follows: An intelligent review system for institutional conflict points based on knowledge graphs and large language models, the system comprising:

[0006] The dual-mode storage architecture maintenance module is used to build and maintain a heterogeneous storage space containing unstructured data indexes and structured logical graphs. The structured logical graph defines the logical deduction sequence and constraint relationships between nodes of the review rules through directed edges.

[0007] The data stream feature parsing module is used to receive the target text data stream to be reviewed and perform feature extraction operations. Based on the extracted feature vectors, it identifies the processing intent of the target text data stream and triggers graph constraint-based generation logic when the processing intent indicates depth verification.

[0008] The logical topology translation module is used to traverse the structured logical graph based on feature vectors and extract subgraph data that are semantically related to the target text data stream. The logical translation program is executed to transform the directed edge topology in the subgraph data into a linearized natural language instruction sequence, and the review rule nodes are mapped to variable slots in the instruction sequence to carry data, thereby constructing a mandatory logical constraint template.

[0009] The constraint generation and verification module is used to fill the target text data stream into the variable slots of the forced logic constraint template to synthesize composite prompt data, and input the composite prompt data into the pre-trained language model to obtain the logic consistency verification result generated by the pre-trained language model within the logical boundary defined by the forced logic constraint template.

[0010] The dynamic knowledge evolution module is used to parse the logical consistency verification results to extract incremental knowledge triples, and write the incremental knowledge triples back to the structured logic graph to update node attributes and connection relationships.

[0011] Preferably, the dual-mode storage architecture maintenance module includes: a vectorization processing unit, used to segment and clean the input unstructured policy documents, convert text fragments into high-dimensional dense vectors using an embedding model, and establish a vector index based on approximate nearest neighbor search; and a graph ontology construction unit, used to define a graph ontology structure containing review rule nodes, audit element entities, and logical relationship edges, and to structure expert knowledge into triple data stored in a structured logical graph, thereby achieving decoupled storage of sparse logical rules and dense text vectors.

[0012] Preferably, the logical topology translation module is specifically used to perform the following operations: identify the start node and end node in the subgraph data, and calculate the shortest logical reasoning path between them; extract all logical predicate edges on the logical reasoning path, and convert each logical predicate edge into a condition constraint statement described in natural language; linearly concatenate all condition constraint statements into a natural language instruction sequence according to the temporal direction of the logical reasoning path; identify the review rule nodes on the logical reasoning path as data gaps to be filled, and define the data type constraints of the data gaps in the forced logical constraint template.

[0013] Preferably, the system also includes a confidence assessment module, used to calculate the logical confidence of the incremental knowledge triples before the dynamic knowledge evolution module performs the write-back operation. : ,in, This represents the vector similarity between the incremental knowledge triples and existing nodes in the structured logic graph. This refers to the shortest closed-loop path length formed after the incremental knowledge triples are integrated into the structured logic graph. and The preset weighting coefficients are used; the confidence assessment module only evaluates the calculated logical confidence level. When the value exceeds the preset write-back threshold, the dynamic knowledge evolution module is allowed to perform a write-back operation.

[0014] Preferably, the constraint generation and verification module is also used to: perform semantic segmentation and type verification on the target text data stream based on the attribute definition of the variable slot before filling the target text data stream into the variable slot; select only text fragments that meet the data type requirements of the variable slot to fill the corresponding variable slot, and filter out noise data that is irrelevant to the current logical reasoning step; and monitor the logical coherence of the generated content in real time during the process of generating verification results by the pre-trained language model. If the generated content is detected to deviate from the trajectory defined by the instruction sequence, the generation is truncated and a retry mechanism is triggered.

[0015] Preferably, when handling complex review tasks involving multiple levels of higher-level legal basis, the logical topology translation module is specifically used to: extract multiple parallel logical branch paths from the structured logical graph, each logical branch path corresponding to the deduction logic of a higher-level legal basis; construct sub-constraint templates for each logical branch path and define priority arbitration rules between each sub-constraint template; combine multiple sub-constraint templates and priority arbitration rules through logical connectors to construct a composite mandatory logical constraint template containing a hierarchical structure.

[0016] Preferably, the system also includes a conflict detection module, used to: analyze the logical consistency verification results and identify the institutional conflicts indicated therein; based on the position of the institutional conflict in the structured logical graph, backtrack to query the superior law review rule node that caused the institutional conflict; construct a conflict evidence chain containing the institutional conflict, the target review text, and the superior law review rule node, and format the conflict evidence chain into a structured review report.

[0017] Preferably, the write-back operation performed by the dynamic knowledge evolution module specifically includes: converting the unstructured description in the logical consistency verification result into standardized entity and relation data; detecting whether the entity and relation data already exist in the structured logical graph; if not, creating a new review rule node or logical relation edge, and assigning a unique identifier and timestamp attribute to the new element; if it already exists, updating the attribute value of the corresponding node or edge, and increasing the reference count to strengthen the weight of the logical path.

[0018] Preferably, the data stream feature parsing module is also used to: classify the target text data stream by intent, distinguishing it as a fact query intent or a logical review intent; only when it is identified as a logical review intent, call the logical topology translation module to start the template construction process; if it is identified as a fact query intent, directly retrieve the relevant text fragments from the dual-mode storage architecture maintenance module through vector retrieval and return them.

[0019] Preferably, the mandatory logical constraint template includes a role definition domain, a background knowledge domain, a logical instruction domain, and an output format domain; the logical topology translation module maps the node attributes in the subgraph data to the background knowledge domain, maps the directed edge topology to the logical instruction domain, and maps the preset review expert role characteristics to the role definition domain, so as to construct a fully constrained generation context environment.

[0020] Compared with the prior art, the beneficial effects of the present invention are:

[0021] 1. In knowledge graphs and large language models, a constraint generation mechanism based on graph topology is adopted to map the discrete entity relationship network of the knowledge graph into a linear logic instruction template. A deterministic reasoning path boundary is constructed at the input end of the generative large model. By converting graph structure data into serialized context constraint features, the model can perform probability prediction and content filling within a preset logical skeleton. This suppresses the inherent divergence of statistical language models when handling long-chain, strong logic tasks from the perspective of computational principles, ensuring the convergence and consistency of the logical structure of the output data.

[0022] 2. Construct a semantic and logical decoupled dual-modal storage architecture. Utilize a vector database to store high-dimensional dense vectors of unstructured text, and a knowledge graph to store sparse topological relationships of entities based on abstract rules. This heterogeneous data separation and processing changes the traditional single-dimensional recall mode. During the inference stage, prioritize locking the logical subgraph based on graph traversal, and then accurately fill the semantics of the vector database within the limited subgraph scope. The logical skeleton guides the semantic filling calculation process, avoiding noise interference and computational redundancy in full-volume text retrieval, reducing the dimensionality of the retrieval space, and ensuring the completeness of complex logical deduction and the accuracy of retrieval results.

[0023] 3. Establish a dynamic write-back mechanism based on confidence verification to achieve a closed-loop flow from unstructured generated results to structured graph ontology data. The system will parse the structured data generated through reasoning into triples that conform to the ontology definition and inject them incrementally into the original graph structure as new logical nodes. The feedback update mechanism enables the static knowledge storage structure to have the ability to dynamically adapt as the task is executed. By continuously converting the calculation results into prior logical constraints that can be called again, the information entropy of the system when processing similar complex tasks is reduced, and the adaptive optimization of the data processing system and automatic completion of the logic library are realized. Attached Figure Description

[0024] Figure 1 This is a closed-loop flowchart of the core business logic of the system based on dual-mode storage and graph constraints according to the present invention;

[0025] Figure 2 This is an experimental comparison chart of key review performance indicators of the present invention sample and the traditional RAG scheme under complex working conditions.

[0026] Figure 3 This is a diagram showing the physical deployment architecture and component interaction topology of the system across the security network boundary according to the present invention. Detailed Implementation

[0027] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0028] This invention discloses an intelligent review system for institutional conflict points based on knowledge graphs and large language models. It comprises a dual-mode storage architecture maintenance module, a data flow feature parsing module, a logical topology translation module, a constraint generation and verification module, a dynamic knowledge evolution module, and a confidence assessment module. The dual-mode storage architecture maintenance module is responsible for constructing and maintaining a heterogeneous storage space containing unstructured data indexes and structured logical graphs. This module performs preprocessing procedures on collected higher-level policy documents. The system calls the BGM-3 embedding model to segment and clean the input policy text, mapping variable-length text fragments into dense vectors of fixed dimensions, and establishing a vector index based on approximate nearest neighbor search. This index is used to quickly locate semantically adjacent points in high-dimensional space during subsequent retrieval. Simultaneously, this module utilizes a graph database to store the structured logical graph. The system transforms audit element entities and review rules verified by domain experts into triplet data containing subjects, predicates, and objects. The directed edges in this graph define the logical deduction sequence and constraint relationships between nodes of different review rules, thereby achieving separate storage and linked indexing of sparse logical rules and dense text vectors.

[0029] The data stream feature parsing module receives the target text data stream to be reviewed and parses its processing intent. This module calculates the cosine similarity between the feature vector of the input text and the center of the preset intent vector cluster. When the calculation result indicates that the target text belongs to the logical review intent, the system triggers the subsequent graph constraint generation logic; when it indicates the fact query intent, the system directly executes the vector retrieval process. During the retrieval, the system adopts a hybrid retrieval strategy, simultaneously performing keyword retrieval based on inverted index and semantic retrieval based on vector similarity. The system weights and sorts the two retrieval results, setting the weight of full-text semantic retrieval to 0.7 and the weight of vector retrieval to 0.3. The system filters out noisy data with similarity scores below 0.5 and selects the top 3 text fragments in the sorted results as the context for subsequent processing. The logical topology translation module connects the structured graph and the generated... The system employs a model to construct a mandatory logical constraint template. Based on the previously extracted feature vectors, it traverses the structured logical graph and extracts sub-graph data that are semantically related to the target text. The logical translation program identifies the start and end nodes in the sub-graph and uses a breadth-first search algorithm to calculate the shortest logical reasoning path between them. The system extracts all logical predicate edges on the path and converts the topological structure of each directed edge into a conditional constraint statement described in natural language. For example, the system translates the graph structure from node A to node B into a linear instruction that B can only be executed after A meets the conditions. The system maps the review rule nodes on the path to variable slots in the mandatory logical constraint template. This template includes a role definition domain, a background knowledge domain, a logical instruction domain, and an output format domain, thereby transforming the discrete graph topology into a linearized natural language instruction sequence.

[0030] The constraint generation and verification module performs review analysis and conflict detection. This module fills the variable slots of the mandatory logical constraint template with the target text data stream to be reviewed, synthesizes composite prompt data containing complete context and logical constraints, and inputs it into the pre-trained language model. Since the input data contains embedded logical path constraints translated from the graph, the pre-trained language model is restricted to the logical boundaries defined by the template during the generation process. The model outputs the logical consistency verification result, indicating whether there are any institutional conflicts between the target text and the superior legal basis. The dynamic knowledge evolution module and the confidence assessment module work together to realize the adaptive update of the system. The dynamic knowledge evolution module parses the logical consistency verification result and extracts potential incremental knowledge triples. Before performing the write-back update, the confidence assessment module calculates the logical confidence of the incremental knowledge triple. The calculation is performed based on the following formula: ,in, and In this embodiment, the preset weighting coefficients are used. The value is 0.6. The value is 0.4; The vector similarity between the incremental knowledge triple and the existing nodes in the structured logic graph is used to characterize the semantic fit between the new knowledge and the existing knowledge system. This is the shortest closed-loop path length formed after the triple is incorporated into the structured logic graph, used to characterize the compactness of the logical deduction, and is applied to the weighting coefficients in the confidence assessment formula. and It is determined that a pre-set standard dataset containing confirmatory review conclusions is used, with a step size of 0.05, to traverse within the closed interval from 0 to 1. Values ​​are determined based on Constraint Synchronization Adjustment For each set of parameters, the harmonic mean of precision and recall of the standard dataset is calculated, and the parameter combination corresponding to the peak value of the indicator is selected as the benchmark for the current application scenario. By periodically introducing new review case data for iteration, the weight coefficients are dynamically adapted to the logical preference characteristics of different legal fields.

[0031] Path length in the formula Defined as the shortest topological distance from the incremental triple access graph backtracking along the directed logical edge to the preset authoritative anchor node, the node to be verified and the existing nodes in the graph form a logical verification closed loop. The closed loop is characterized by the number of nodes it contains, quantifying the compactness of the logical deduction. If the backtracking path does not contain a closed loop, Dijkstra's algorithm is used to calculate the number of hops from a node to its nearest parent node or constitutional anchor node. Numerical values ​​are used to prevent computational divergence. A maximum search depth threshold is set, for example, 10 hops. If the path length exceeds this threshold, the search will be terminated immediately. Setting terms to zero filters out interference from long-chain weakly correlated logic on confidence scores; vector similarity. A high-dimensional cosine similarity algorithm is used for calculation. Normalization is performed to eliminate the influence of vector magnitude differences. The input unstructured text and graph node text are mapped to fixed-dimensional dense vectors through the same embedding model. The cosine value of the angle between the two vectors is calculated as the original similarity. It is ensured that the calculation result falls within the required interval of 0 to 1 in the formula. Linear pruning and mapping are performed on the original cosine value. If it is lower than a preset semantic noise threshold, such as 0.4, the result is reset to zero. Only high-confidence semantic association features are retained for subsequent weighted calculations to ensure the final logical confidence. Physical meaning and comparability, when the calculated logical confidence level is... When the value exceeds the preset write-back threshold of 0.75, the system allows the write-back operation to be performed. The write-back operation includes creating new review rule nodes or logical relationship edges in the graph and assigning them unique identifiers and timestamps, or updating the attribute values ​​of existing nodes and increasing the reference count, thereby realizing the dynamic evolution of the knowledge base.

[0032] Example 1: In compliance review scenarios involving multi-level financial regulatory policies, immediate review of newly released regulations such as the "Regulations on the Management of Cross-border Capital Flows" is required. This necessitates comparing thousands of scattered clauses in both the higher-level "Cybersecurity Law" and the lower-level "Guidelines for Data Compliance of Commercial Banks." In this context, traditional solutions relying solely on vector retrieval often suffer from logical divergence in long-chain reasoning due to a lack of structured constraints. Solutions relying solely on rule engines, on the other hand, fail to understand the subtle semantic differences in natural language, leading to high false negative rates. When the system receives a target text data stream containing the aforementioned regulations, the data stream... The feature parsing module performs feature extraction. This module calculates that the cosine similarity between the feature vector of the target text and the preset logical review intent vector cluster is higher than that of the fact query intent, thus determining that the current task is a deep logical verification task. This activates the generation logic based on graph constraints. To solve the noise interference problem in semantic recall, the system executes a hybrid retrieval strategy at this stage. The full-text keyword retrieval results and the vector semantic retrieval results are re-ranked with weights of 0.7 and 0.3, respectively. The top 3 text fragments in the ranking are extracted as the basic context. This process ensures the semantic relevance of the input data and provides a corpus basis for subsequent processing.

[0033] Subsequently, the logical topology translation module intervenes in the core processing flow, performing decoupling and reorganization of structured logic and unstructured semantics. Based on feature vectors, this module accurately locates and extracts sub-graph data containing key nodes such as data export, security assessment, and storage period from the structured logical graph maintained by the dual-mode storage architecture maintenance module. The system uses a breadth-first search algorithm to calculate the shortest logical reasoning path from the starting point of compliance obligations to the specific implementation details, and transforms each directed edge representing a logical predicate on this path into a linear condition constraint statement described in natural language. For example, the topology structure of security assessment -> [preceded] -> data export in the graph is translated into an instruction to complete the security assessment before performing the data export operation, and filled into the logical instruction domain of the logical constraint template. Together with the role definition domain and background knowledge domain, it constitutes a fully qualified generative context environment. This step maps the discrete graph topology to the linear input constraints of the pre-trained language model, suppressing the probabilistic divergence of the generative model in the autoregressive process from a computational perspective, so that the model performs semantic filling within the given logical skeleton.

[0034] In the constraint generation and verification module, the pre-trained language model outputs a logical consistency verification result based on the aforementioned composite prompt data. It identifies a conflict between the clause in the "Regulations on the Management of Cross-border Capital Flows" regarding the retention period of transaction records and higher-level laws. The dynamic knowledge evolution module analyzes this verification result and extracts incremental knowledge triples regarding new cross-border transaction data retention rules. The confidence assessment module then calculates the logical confidence of these triples. In the specific calculation of this embodiment, the vector similarity between the triplet and existing nodes in the graph is measured. The value is 0.85, representing the shortest closed-loop path length formed after its integration into the graph. The value is 4, according to the formula. Perform calculations to obtain the logical confidence level. The value is 0.61. Since this value is less than the preset write-back threshold of 0.75, the system determines that the logical stability of the incremental knowledge has not yet met the storage standard, and temporarily suspends the write-back operation and triggers the manual review process. This mechanism effectively prevents low-confidence noise data from polluting the core logic graph and ensures the rigor and purity of the knowledge system in the storage space.

[0035] Example 2: To verify the actual performance and engineering effectiveness of the intelligent review system for institutional conflicts proposed in this invention, a verification platform simulating a real financial regulatory review scenario was built in a controlled laboratory environment. It was deployed on a high-performance computing cluster equipped with eight NVIDIA A100 GPUs, running on Linux Ubuntu 20.04 LTS, based on the PyTorch 1.13 deep learning framework. The verification process selected a test dataset containing 2000 historical policy documents and 500 known conflicts annotated by experts as the benchmark test set. The core objective of the experiment was to quantitatively evaluate the accuracy and recall of this system when handling highly complex logical verification tasks, and to verify its superiority over traditional retrieval-enhanced RAG generation schemes. Therefore, the experiment included a comparative test between the present invention's sample group and a traditional RAG control group. The present invention's sample group adopted a complete dual-mode storage and logical topology translation architecture; while the traditional RAG control group removed the logical topology translation module, retaining only the vector similarity-based retrieval generation mechanism, serving as a benchmark representing the current technological level.

[0036] To simulate semantic noise interference in a real-world environment, 20% of irrelevant policy text fragments and paraphrased texts were randomly injected into the test dataset to construct a noise interference scenario. During the experiment, the system was required to automatically review 50 pre-set complex compliance issues, each involving logical reasoning across multiple levels of regulations. Two key performance indicators were recorded: logic verification accuracy and reactivity recall. Accuracy was defined as the proportion of the number of reactivity points correctly identified by the system to the total number of reactivity points output by the system, and recall was defined as the proportion of the number of reactivity points correctly identified by the system to the total number of known expert-annotated reactivity points in the test set. Table 1 shows the comparison data of the review performance of different schemes under the noise interference scenario.

[0037] Table 1: Review Performance Comparison Data Table

[0038]

[0039] Referring to Table 1, the data shows that under complex working conditions with introduced noise interference, the logic verification accuracy of the sample group of this invention reached 89.2%, which is higher than the 62.5% of the traditional RAG control group. This result indicates that by transforming the discrete graph structure into linear forced logic constraints through the logic topology translation module, the illusion phenomenon of generative models when facing noisy data is effectively suppressed, ensuring that the inference process follows the preset compliance logic. At the same time, the recall rate of the conflict point increased from 58.4% to 91.5%, proving that the dual-mode storage architecture does not sacrifice the coverage of long-tail knowledge while maintaining high accuracy. Although the average inference time increased slightly (+0.3s), this is mainly due to the additional overhead of logic path calculation. To further verify the write-back threshold parameter... and To verify the rationality of the write-back threshold setting, a gradient verification experiment was designed. In the experiment, the write-back threshold was fixed at 0.75, and the weighting coefficients were adjusted. (Vector similarity weights) and The ratio of (path length weight) is used to observe the effective incremental rate of the system knowledge base, that is, the proportion of knowledge items that are rewritten and verified by humans to be correct. Table 2 shows the impact of the weight coefficient ratio on the effective incremental rate.

[0040] Table 2: Data Table on the Influence of Weighting Coefficient Ratio

[0041]

[0042] See Table 2, when : When the ratio is set to 0.6:0.4, the effective incremental rate reaches a peak of 94.3%, while the knowledge base noise rate remains at a minimum level of 2.1%. This data reveals that in the process of knowledge evolution, it is necessary to consider not only the semantic similarity between new and old knowledge ( ), and is more based on the compactness of the logical closed loop ( Sufficient weights are used to prevent the injection of semantically relevant but logically loosely logical erroneous knowledge, which deviates from the optimal match, such as 0.8:0.2. Although this improves the sensitivity of semantic matching, it causes the noise rate to rise to 15.4% because it ignores the constraints of logical structure.

[0043] Example 3: This example combines Figures 1 to 3 An explanation of the intelligent review system for institutional conflict points based on knowledge graphs and large language models, such as... Figure 1 As shown, the data flow feature parsing module receives input and extracts feature vectors, identifies processing intentions to trigger graph-based generation logic, and performs feature comparison with the dual-mode storage architecture maintenance module. The dual-mode storage architecture maintenance module maintains a heterogeneous storage space containing unstructured data indexes and structured logic graphs, and provides the extracted associated sub-graphs to the logic topology translation module. The logic topology translation module traverses the associated sub-graphs, translates the topology into instruction sequences, constructs a mandatory logical constraint template, and transmits it to the constraint generation and verification module. In this module, text is filled into variable slots, a pre-trained model is used to generate verification results and obtain logical consistency judgments. The verification results are input to the dynamic knowledge evolution module, which parses the verification results, extracts incremental knowledge triples, and performs a write-back update operation to the dual-mode storage architecture maintenance module. Finally, the system outputs logical consistency verification results including institutional conflict point reports.

[0044] like Figure 2As shown in the chart, the bar chart compares the traditional RAG control group and the sample group of this invention in three key performance indicators. The vertical axis represents percentage (%) and time (s). In terms of logic verification accuracy, the bar height of the sample group of this invention is higher than that of the traditional RAG control group, with a value close to 90%. In terms of repulsion recall, the sample group of this invention also shows a significant advantage. In terms of average inference time (s), both have low values ​​and small differences, with the sample group of this invention slightly higher than the control group. Figure 3 As shown, the intelligent review core computing power cluster is deployed inside the secure network boundary; the review interaction terminal is located outside, submitting data to the cluster. The core of the cluster is a deep learning inference environment for large model generation and verification services. Its left and right sides are respectively connected to the feature parsing service responsible for semantic vectorization and the topology translation service responsible for logical rule mapping. The core service connects downward to the bottom storage layer, which includes a vector database cluster storing unstructured data indexes and a graph database cluster storing structured logical graphs. The data flow shows that the feature parsing service performs retrieval on the vector database, the topology translation service performs traversal on the graph database, and the large model service performs dynamic knowledge write-back, forming a closed-loop data interaction network.

[0045] Example 4: Regarding the write-back threshold disclosed in Example 2 : The optimal ratio of 0.6:0.4 is further disclosed in this embodiment, which provides a set of standardized engineering calibration procedures for determining this key parameter ratio. This provides a deterministic parameter optimization method for system deployment in different business scenarios, ensuring that the system achieves the best balance between semantic recall and logical constraints. The execution environment of the procedure is set as an isolated sandbox test area, which is equipped with hardware resources consistent with the production environment and a gold standard dataset containing 1,000 standard test cases. Each test case includes the input policy text, the expected compliance conclusion, and the correct logical deduction path marked by manual annotation.

[0046] The first stage of the calibration procedure is the extreme value benchmark test, which will Set to 1.0. Setting it to 0 means relying entirely on vector similarity for write-back decisions. Under this configuration, the test set is run, and the system records two key metrics: recall and noise rate. Set to 0, The initial value is set to 1.0, meaning the decision is entirely based on the logical path length. The aforementioned metrics are also recorded. The purpose of this stage is to establish the physical boundaries of the parameter values ​​and their extreme impact on system performance, defining an effective range for subsequent refined optimization. The second stage is gradient step optimization, within the range of 0.1 to 0.9, using a step size of 0.1. The constraints generate 9 sets of parameter ratios (e.g., 0.1:0.9, 0.2:0.8...0.9:0.1). For each set of ratios, the system automatically executes all test cases and calculates the overall performance score. The score is calculated using the following formula: in, The recall rate of effective incremental knowledge. The noise rate of erroneous knowledge. and To determine the penalty coefficient based on business preferences, in this embodiment, it is set based on the extremely low tolerance for errors in compliance reviews. , The system plots the nine calculated scores as curves to find the parameter ratio range corresponding to the highest score. The third stage is fine-tuning and locking. Within the optimal range determined in the second stage, such as 0.5:0.5 to 0.7:0.3, the step size is further reduced to 0.05 for fine-tuning tests. Finally, the comprehensive performance score is selected. The parameter ratio at which the global maximum value is reached is used as a fixed parameter for this specific business scenario. In the financial regulatory scenario of this embodiment, this procedure determines and ultimately locks the parameter ratio. , .

[0047] Example 5: In the scenario of in-depth compliance review of large-scale government procurement projects, the system needs to solve the technical problem of parsing and executing complex logical rules containing multiple nested conditions and exception clauses. In such scenarios, there is a logical dependency between the superior law, the Government Procurement Law, and the specific implementation rules. For example, when the procurement amount exceeds a threshold and the supplier is a foreign-funded enterprise, a security review is performed. Unless the project belongs to the exemption list, linear text retrieval is difficult to maintain the three-dimensional logical structure of condition A and condition B, and non-condition C, which can easily lead to omissions in the review. The logical topology translation module executes a set of structured mapping procedures based on depth-first traversal. The system locates the root node representing the procurement rule in the structured logical graph and traverses down along the directed edges to all leaf nodes. During this process, the system calls a pre-set logical semantic mapping table to transform the topological structure in the graph. For logical connectors in natural language, if the logical attribute of two parallel incoming edges pointing to the same child node in the graph is marked as a necessary condition, the system translates it into a logical AND relation connector, i.e., embeds AND in the generated constraint instructions; if it is marked as a selection condition, it is translated into a prohibition or unless negative instruction for edges marked as exclusive constraints. For example, for the aforementioned procurement rule, the system translates the two solid edges of the amount node and foreign investment node pointing to the security audit node in the graph, combined with the dashed blocking edge of the exemption node pointing to the security audit node, into a linear instruction sequence that follows the review logic: while confirming that the procurement amount meets the standard and the supplier attribute is foreign investment, verify whether there is an exemption situation, and fill it into the logical instruction domain of the logical constraint template. When generating the review conclusion, the pre-trained language model reproduces the logical topology structure defined in the graph.

[0048] After completing the review and generating a text report, the dynamic knowledge evolution module initiates a pattern-constrained reverse extraction procedure to address the structural black box problem when unstructured text flows back to a structured graph. The system loads a regularized extraction template aligned with the graph ontology definition, limiting the output data to include a standard triplet structure of subject-predicate-object, and constraining the predicates to be selected from a predefined set of relations in the graph, such as contain, precede, and conflict with. The system performs syntactic dependency analysis on the logical consistency verification results, identifies the core sentence segments describing institutional conflicts, and maps them to the extraction template. For example, from the text where the requirement of regulation A regarding the time limit conflicts with the superior law B, candidate triplets of (regulation A, conflict with, superior law B) are parsed out. The system calls the optimal parameter ratio determined in Example 3 ( The confidence level of the candidate triple is calculated. Only when the confidence level meets the standard is it added to the graph as a formal node. This procedure constructs a logical closed loop by applying topological mapping constraints at the input end and pattern extraction constraints at the output end, ensuring the transmission and evolution of unstructured semantics and structured logic in the interaction process.

[0049] Example 6: To eliminate the potential risk of bias in the confidence assessment model due to insufficient historical write-back data during the system cold start phase, this example constructs an offline calibration and data filling procedure for the confidence assessment module. Before the initial system deployment, an offline data preparation phase is performed. This phase selects no fewer than 10,000 text pairs with clear causal reasoning structures from publicly available legal case libraries, compliance Q&A sets, and declassified past review records to build a seed knowledge base. For each seed data point, the system automatically extracts its triple structure and assigns an initial logical confidence label using pre-set scoring scripts and those based on human expert rules. This batch of seed data with high confidence labels is then used to calibrate the confidence assessment module. and The parameters are trained through supervised learning, enabling them to have a basic ability to judge the logical rationality of new knowledge before being connected to real-time business flows, thus avoiding the randomness of decision-making during the cold start period.

[0050] In response to the significant differences in regulatory rules across different industries, this embodiment defines a pre-deployment calibration procedure. When the system is delivered to a specific industry customer environment, this procedure is executed. The 50 most representative current core regulations of that industry are selected as the calibration sample set. The system performs a full-process review of the sample set and records the intermediate constraint instruction sequence generated by the logical topology translation module. Industry experts score the semantic fidelity of 100 randomly selected instructions. If the average score is lower than the preset acceptance threshold, the system will automatically trigger a fine-tuning loop for that industry: adjust the translation rule weights in the logical semantic mapping table and rerun the calibration process until the score meets the standard.

[0051] Example 7: This example discloses a standardized system deployment pre-deployment calibration and model building procedure. For vector index construction in a dual-mode storage architecture, a dimension-precision adaptive calibration process is defined. When the process starts, the hardware computing power (FLOPS) and memory bandwidth of the target deployment environment are benchmarked. Using a pre-set standard semantic dataset containing 100,000 general texts, the vector dimensions are adjusted in a stepwise manner, such as from 128 dimensions to 1024 dimensions, with a step size of 64 dimensions. The retrieval precision (Recall@K) and query latency under different dimensions are recorded. The system constructs a three-dimensional response surface of precision-latency-dimension and automatically searches for the optimal vector dimension parameters that can provide the highest retrieval precision on this surface based on a preset maximum tolerable latency threshold, such as 50ms.

[0052] Secondly, for the path search algorithm in the logical topology translation module, an adaptive procedure for the truncation threshold of depth-first traversal (DFS) is established. Based on the differences in the length of legal logical chains in different domains, the system pre-configures a logical depth detector. During the initialization phase, the detector randomly selects 1000 logical nodes from the target domain's legal library and calculates their average out-degree and maximum connected subgraph depth. Based on statistical distribution, the system sets the maximum depth threshold for DFS search to 1.2 times the depth of this maximum connected subgraph. This ensures the integrity of long-chain reasoning while effectively preventing performance issues caused by search space explosion. It can crash; finally, to ensure the robustness of the confidence assessment model, boundary reinforcement training based on adversarial examples is implemented. After initial training using the seed knowledge base, the system automatically generates a batch of logical adversarial examples. These examples are generated by making small semantic perturbations to real compliant triples, such as replacing synonyms but reversing logical predicates. These adversarial examples are mixed into the training set for secondary fine-tuning, forcing the model to learn more granular logical discriminative features. The end of training is marked by the model's F1 score on the adversarial validation set no longer improving for 5 consecutive epochs and remaining above 0.85.

[0053] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A smart review system for institutional conflict points based on knowledge graphs and large language models, characterized in that the system... include: The dual-mode storage architecture maintenance module is used to build and maintain a heterogeneous storage space containing unstructured data indexes and structured logical graphs. The structured logical graph defines the logical deduction sequence and constraint relationships between nodes of the review rules through directed edges. The data stream feature parsing module is used to receive the target text data stream to be reviewed and perform feature extraction operations. Based on the extracted feature vectors, it identifies the processing intent of the target text data stream and triggers graph constraint-based generation logic when the processing intent indicates depth verification. The logical topology translation module is used to traverse the structured logical graph based on feature vectors and extract subgraph data that are semantically related to the target text data stream. The logical translation program is executed to transform the directed edge topology in the subgraph data into a linearized natural language instruction sequence, and the review rule nodes are mapped to variable slots in the instruction sequence to carry data, thereby constructing a mandatory logical constraint template. The constraint generation and verification module is used to fill the target text data stream into the variable slots of the forced logic constraint template to synthesize composite prompt data, and input the composite prompt data into the pre-trained language model to obtain the logic consistency verification result generated by the pre-trained language model within the logical boundary defined by the forced logic constraint template. The dynamic knowledge evolution module is used to parse the logical consistency verification results to extract incremental knowledge triples, and write the incremental knowledge triples back to the structured logic graph to update node attributes and connection relationships. The logical topology translation module is specifically used to perform the following operations: identify the start and end nodes in the subgraph data and calculate the shortest logical reasoning path between them; extract all logical predicate edges on the logical reasoning path and convert each logical predicate edge into a condition constraint statement described in natural language; linearly concatenate all condition constraint statements into a natural language instruction sequence according to the temporal direction of the logical reasoning path; identify the review rule nodes on the logical reasoning path as data gaps to be filled and define the data type constraints of the data gaps in the forced logical constraint template; The system also includes a confidence assessment module, which calculates the logical confidence of incremental knowledge triples before the dynamic knowledge evolution module performs the write-back operation. : ,in, This represents the vector similarity between the incremental knowledge triples and existing nodes in the structured logic graph. This refers to the shortest closed-loop path length formed after the incremental knowledge triples are integrated into the structured logic graph. and The preset weighting coefficients are used; the confidence assessment module only evaluates the calculated logical confidence level. When the value exceeds the preset write-back threshold, the dynamic knowledge evolution module is allowed to perform a write-back operation; When handling complex review tasks involving multiple levels of higher-level legal basis, the logical topology translation module is specifically used to: extract multiple parallel logical branch paths from the structured logic graph, with each logical branch path corresponding to the deduction logic of a higher-level legal basis; construct sub-constraint templates for each logical branch path and define priority arbitration rules between each sub-constraint template; and combine multiple sub-constraint templates and priority arbitration rules through logical connectors to construct a composite mandatory logical constraint template containing a hierarchical structure.

2. The intelligent review system for institutional conflict points based on knowledge graphs and large language models according to claim 1, characterized in that, The dual-mode storage architecture maintenance module includes: a vectorization processing unit, which is used to segment and clean the input unstructured policy documents, convert text fragments into high-dimensional dense vectors using an embedding model, and establish a vector index based on approximate nearest neighbor search; and a graph ontology construction unit, which is used to define the graph ontology structure containing review rule nodes, audit element entities, and logical relationship edges, and to structure expert knowledge into triple data stored in the structured logical graph.

3. The intelligent review system for institutional conflict points based on knowledge graphs and large language models according to claim 1, characterized in that, The constraint generation and verification module is also used to: perform semantic segmentation and type verification on the target text data stream based on the attribute definition of the variable slots before filling the target text data stream into the variable slots; select only text fragments that meet the data type requirements of the variable slots to fill the corresponding variable slots, and filter out noise data that is irrelevant to the current logical reasoning step; and monitor the logical coherence of the generated content in real time during the process of generating verification results by the pre-trained language model. If the generated content is detected to deviate from the trajectory defined by the instruction sequence, the generation is truncated and a retry mechanism is triggered.

4. The intelligent review system for institutional conflict points based on knowledge graphs and large language models according to claim 1, characterized in that, The system also includes a conflict detection module, which is used to: analyze the logical consistency verification results and identify the institutional conflicts indicated therein; based on the position of the institutional conflicts in the structured logical graph, backtrack to query the superior law review rule node that caused the institutional conflict; construct a conflict evidence chain containing the institutional conflicts, the target review text, and the superior law review rule node, and format the conflict evidence chain into a structured review report.

5. The intelligent review system for institutional conflict points based on knowledge graphs and large language models according to claim 1, characterized in that, The write-back operation performed by the dynamic knowledge evolution module specifically includes: converting the unstructured description in the logical consistency verification result into standardized entity and relation data; detecting whether the entity and relation data already exist in the structured logical graph; if not, creating new review rule nodes or logical relation edges, and assigning unique identifiers and timestamp attributes to the new elements; if they already exist, updating the attribute values ​​of the corresponding nodes or edges, and increasing the reference count to strengthen the weight of the logical path.

6. The intelligent review system for institutional conflict points based on knowledge graphs and large language models according to claim 1, characterized in that, The data stream feature parsing module is also used to: classify the intent of the target text data stream, distinguishing it as a fact query intent or a logical review intent; The template construction process is initiated by calling the logical topology translation module only when the intent is identified as a logical review; if the intent is identified as a fact query, the relevant text fragments are directly retrieved from the dual-mode storage architecture maintenance module via vector retrieval and returned.

7. The intelligent review system for institutional conflict points based on knowledge graphs and large language models according to claim 1, characterized in that, The mandatory logic constraint template includes a role definition domain, a background knowledge domain, a logic instruction domain, and an output format domain; The logical topology translation module maps node attributes in the subgraph data to the background knowledge domain, maps directed edge topology to the logical instruction domain, and maps preset review expert role characteristics to the role definition domain, in order to construct a fully qualified generation context environment.

Images