Audit path decision system and method based on dynamic switching of rule tree and semantics

Abstract

This invention relates to the field of artificial intelligence data processing technology, specifically disclosing a review path decision-making system and method based on rule trees and dynamic semantic switching. The system includes an unstructured data processing module, a rule tree configuration module, a semantic vector analysis module, a dynamic switching strategy controller, and a closed-loop feedback learning module. The unstructured data processing module receives and splits upstream form data, sending the resulting feature stream to the rule tree configuration module. The rule tree configuration module, through the dynamic switching strategy controller, calls the semantic vector analysis module to convert semantic vectors into virtual rule factors and uses the closed-loop feedback learning module to inject them back into the rule tree engine, achieving bidirectional collaboration. This invention, through its innovative architecture of rule trees and dynamic semantic switching, combined with a large-scale model illusion-based full-process prevention mechanism, achieves multi-dimensional technological upgrades compared to existing technologies. Strictly adhering to the MECE principle, it can be applied to multiple scenarios, covering all categories and meeting high compliance requirements.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence data processing technology, specifically to an intelligent decision-making technology for complex unstructured review scenarios with high compliance, strong risk control, and particularly to a review path decision-making system and method based on rule trees and dynamic semantic switching. Background Technology

[0002] Intelligent compliance auditing technology is a core supporting technology for digital transformation in various fields, including compliance management of official document circulation in large state-owned enterprises, group-level penetrating risk management, credit risk management in the financial industry, full-process auditing of underwriting and claims in the insurance industry, and compliance management of content security and information security across all scenarios. Its core value lies in replacing repetitive manual auditing tasks with standardized and automated decision-making mechanisms, thereby improving business processing efficiency, reducing human resource maintenance costs, ensuring consistency of audit results, and achieving closed-loop management of compliance risks throughout the entire process. Currently, mainstream automated auditing systems in the industry still rely on traditional rule-based decision-making systems as their core architecture. Although some solutions have attempted to introduce AI models to empower the semantic parsing process with the iterative evolution of artificial intelligence technologies such as large language models and multimodal algorithms, they have not built a compliant and strongly constrained collaborative decision-making mechanism. The overall technical solutions have significant inherent defects. A detailed analysis of the mainstream technical paths is as follows: ① Decision-making based on hard-coded or traditional rule engines: Linear rule engines were the core technology for early automated review. This solution pre-builds an "IF-THEN" paradigm linear rule base, breaking down the review logic into single-dimensional condition judgments and action execution instructions. The rule base adopts a flat storage structure with no hierarchical relationships, no branch nesting, and no logical linkage. When the system processes tasks to be reviewed, it matches rules one by one according to preset priorities or a fixed traversal order. After a task matches one or more rules, it directly executes the bound action (automatic approval, rejection, transfer to manual review, etc.). This solution has the advantages of simple architecture, low deployment threshold, and short response latency, and is suitable for basic review scenarios that are simple, standardized, and without branches; however, its core defects are prominent: the logical coupling between rules is extremely low, making it unable to express complex business logic such as branch nesting, process convergence, and multi-condition linkage. Its scenario adaptability has inherent shortcomings, and it completely fails when facing non-standardized and ambiguous tasks.

[0003] ② Static Decision Tree Solution: The static decision tree is an iterative optimization of the linear rule engine. This solution abstracts discretized review rules into a hierarchical tree data structure, with the root node as the task entry point. Internal nodes correspond to multi-dimensional condition judgments (review amount threshold, user qualification level, business type code, document confidentiality level, risk level, etc.), branch links represent the condition judgment results, and leaf nodes map the final review conclusion or sub-process entry point. Compared to the linear rule engine, the static decision tree can intuitively support moderately complex branched review logic, with higher visibility of the decision path and improved intuitiveness of rule maintenance. However, its core drawback is that the decision logic and flow path are pre-coded and fixed, lacking dynamic adjustment and adaptive iteration capabilities. It cannot cope with dynamic changes in business scenarios, policy and rule iteration updates, and non-standardized edge tasks. Furthermore, rule modifications require reconstruction of the tree structure, resulting in extremely high full lifecycle maintenance costs.

[0004] ③ Fully Manual Pre-review and Manual Assignment Solution: For complex tasks containing a large amount of unstructured data (such as user-filled remarks, appeal reasons, chat logs, and text in uploaded evidence), since the rule engine cannot parse it, a "manual pre-review / initial review" position is usually set up. After reviewing all materials, the human pre-reviewer manually judges them based on experience and assigns them to subsequent professional review positions (such as anti-fraud specialists, senior legal specialists, etc.).

[0005] ④ Current Status of Irrational Application of Large AI Models: To improve the intelligence level of review systems, the industry generally attempts to integrate AI technologies such as large language models and specialized algorithm models into the review process. However, these applications lack a strong constraint and coordination mechanism with the traditional rule system, exhibiting typical characteristics of "illusionary irrational application": First, unconstrained model decision-making relies on large model outputs directly as the core basis for review, without setting up verification loops to address the inherent illusionary defects of the models. This leads to frequent problems such as models fabricating business parameters, misjudging risk levels, generating violation handling actions, and misinterpreting compliance clauses; Second, undifferentiated models... The scheduling logic blindly calls large models to process simple structured tasks, complex unstructured tasks, and edge-prone tasks, resulting in redundant and wasted computing resources and further amplifying the risk of illusion. Third, the output acceptance mechanism lacks conflict resolution; when the model output contradicts traditional rule-based judgments, there is no priority determination or confidence verification logic, directly accepting the model results, leading to disordered decision-making logic and significantly reduced audit reliability. Fourth, there is no domain-adaptive optimization; the general capabilities of the large model cannot match the stringent rule requirements of vertical fields such as central and state-owned asset compliance, financial risk control, insurance underwriting, and content security, resulting in insufficient decision-making accuracy. In large-scale document review, "illusion" refers not only to factual errors but also to false logical chains generated by the model based on probability without rule support, which appear reasonable but violate the rigid requirements of state-owned asset compliance.

[0006] Based on the implementation mechanisms of existing mainstream technologies, the limitations of comparative patents, and the stringent application pain points in vertical fields such as document circulation in large state-owned enterprises, group-level penetrating risk control, financial credit risk control, insurance underwriting and claims settlement, and full-scenario content security management, through academic deduction, existing technologies have the following inherent defects that cannot be solved through local optimization: (1) The decision-making mechanism is static and rigid, lacking adaptability and self-adaptability to edge scenarios: The decision paths of existing rule engines and comparative patents are mostly preset linear logic or static decision trees, lacking dynamic perception of context changes. When faced with edge scenarios such as ambiguous expressions not covered by rules, hidden penetration risks, or sudden emergencies (such as medical green channels), the system cannot intelligently switch paths and can only adopt a "one-size-fits-all" manual fallback or direct false negatives, resulting in a high workload for manual review and damage to business flexibility and customer experience.

[0007] (2) The dual "explosion" of rules and logic leads to an exponential increase in system operation and maintenance costs: In order to achieve full-scenario coverage, operation and maintenance personnel need to continuously hard-code new rules, resulting in an infinite expansion of the rule base. The hidden logical conflicts and cross-over issues among the massive number of rules make the underlying logic of the system extremely obscure. A single business adjustment can easily trigger global flow anomalies, and the system lacks self-learning and optimization capabilities, resulting in extremely high manual and R&D costs for troubleshooting, debugging, and expansion.

[0008] (3) The analysis dimension is too narrow, and there are shortcomings in the deep semantic perception of unstructured data: Traditional solutions rely too much on structured fields such as amount, code, and security level, and cannot effectively extract deep semantic information from unstructured data such as the intent of official documents, implicit expressions of risks and hidden dangers, and the substantive connotation of underwriting materials. The one-sidedness of the decision-making basis leads to insufficient semantic recognition accuracy when the system processes complex business documents, resulting in a high risk of misjudgment and omission, and making it difficult to support compliance audits with high accuracy requirements.

[0009] (4) Lack of a large model illusion prevention mechanism, resulting in uncontrollable compliance and reliability of decision-making: Existing solutions that introduce large models generally lack a sound constraint mechanism. The inherent factual errors, logical fabrications, and rule deviations of large models, which are considered "illusions," are directly transmitted to the decision-making process, leading to distorted audit results. At the same time, due to the lack of verification and dynamic scheduling, indiscriminate use of large models not only wastes computing resources but also makes it impossible to substantially guarantee the reliability of business operations in high-compliance scenarios.

[0010] (5) Insufficient adaptability to vertical domains and lack of a unified constraint framework across scenarios: Existing technologies mostly focus on a single vertical domain (such as simple document or paper review), lacking the ability to customize and adapt to the needs of central and state-owned assets, finance, insurance and other full-scenario applications. The system architecture lacks a unified rule tree constraint framework and dynamic scheduling mechanism, which makes it impossible to achieve cross-domain logic reuse and group-level penetrating control, and makes it difficult to cope with complex and ever-changing diversified business needs.

[0011] In summary, existing technologies are merely simple concatenations of "rules + AI," weighted voting, or one-way calls, lacking a two-way collaborative mechanism and failing to meet the adaptability requirements of AI data processing solutions across various vertical fields. Summary of the Invention

[0012] The purpose of this invention is to provide a review path decision system and method based on rule tree and dynamic semantic switching, so as to solve the problems encountered in the above-mentioned background art.

[0013] To achieve the above objectives, the technical solution of the present invention is as follows: A review path decision-making system based on rule trees and dynamic semantic switching includes an unstructured data processing module, a rule tree configuration module, a semantic vector analysis module, a dynamic switching strategy controller, and a closed-loop feedback learning module. The unstructured data processing module receives and splits upstream form data, sends the split feature stream to the rule tree configuration module, and the rule tree configuration module calls the semantic vector analysis module through the dynamic switching strategy controller to generate semantic vectors. The semantic vectors are then converted into virtual rule factors and injected back into the rule tree engine using the closed-loop feedback learning module to achieve bidirectional collaboration.

[0014] In the above scheme, the system adopts a dual-engine coupled architecture; wherein: The unstructured data processing module receives upstream form data and breaks it down into feature streams with strong, medium, and weak fields. The rule tree configuration module stores strong constraint rules based on business logic through a rule tree layer, where each node represents a decision condition. It then uses a traditional rule engine to quickly execute deterministic logic operations at the rule tree layer. The semantic vector analysis module incorporates a large language model or deep NLP model to segment and clean the unstructured text, mapping it to a high-dimensional semantic vector space for identifying the specific intent, sentiment polarity, and key entities of the content. The dynamic switching strategy controller monitors the execution status of the rule tree in real time, calculates the current matching "confidence," and dynamically decides whether to interrupt / suspend the current rule tree flow based on a set threshold, thereby triggering the semantic analysis layer to intervene. The closed-loop feedback learning module collects the differences between the final human review results and the machine decision results, optimizes the vector weights of the semantic analysis layer through gradient updates, and recommends new rule nodes to the rule tree layer.

[0015] A method for review path decision-making based on rule trees and dynamic semantic switching involves extracting feature vectors for the task through a data preprocessing module; when the rule matching confidence is below a threshold, a semantic analysis engine is invoked to transform the feature vectors into high-dimensional semantic vectors, and further, intent vectors are generated based on business context clustering; the review path decision-maker generates discrete virtual rule factors based on the intent vectors to achieve dynamic switching of the review path; specifically, it includes the following steps: Step 1: Task Access and Bidirectional Feature Streaming: The system receives a task package containing document content, attachments, and metadata, and uses the data feature deconstruction module to vectorize the features. Step 2: Static depth-first traversal of the rule tree and quantitative determination of confidence: The system extracts the structured feature vector of the task to be reviewed through the data feature deconstruction module. The rule engine traverses from the L1 root node and introduces a confidence scoring function. ,in For the current node; if an ambiguous field is encountered or an L4 anchor node is entered, calculate... ,in, As weight, As the matching factor, For index variables; if , Preset safety threshold; Step 3: Semantic Vector Analysis and Intent Extraction. The semantic layer extracts the intent vector from the unstructured text in the task. ; Step 4: Virtual rule factor generation and injection. The semantic vector is transformed into virtual rule factors and injected back into the rule tree engine to achieve bidirectional collaboration. Step 5: Joint Decision Making, Path Redirection, and Closed-Loop Learning. Based on the injected complete factors, the rule tree generates the final decision path. The system records the decision log and adjusts the mapping function through human feedback. Make minor adjustments.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. Addressing the problem of "static and rigid decision-making mechanisms with a lack of adaptability". Solution: This system introduces a dynamic switching strategy controller as the decision-making center, breaking the constraints of linear decision-making. The system monitors the confidence level of the feature stream and environmental parameters in real time. Once the rule tree touches the "logical boundary" (such as encountering special cases like ambiguous intent, unstructured attachments, or medical green channels), the controller immediately triggers a dynamic jump to the semantic engine, transforming the deep semantic parsing results into "virtual rule factors" and feeding them back to the decision-making layer in real time.

[0017] Beneficial effects: It endows the system with human-like "context awareness" capabilities. In document circulation or credit review, the system no longer mechanically executes "pass / reject" but can spontaneously adjust the review path based on the business context, which greatly reduces the frequency of manual intervention and improves the automation adaptation rate of edge scenarios by more than 60%.

[0018] 2. Addressing the issue of "a double explosion of rules and logic, leading to an exponential increase in operational costs". Solution: By employing a closed-loop feedback learning module combined with a pruning-optimized rule tree architecture, the system no longer relies on endlessly stacking hard-coded rules. Instead, it utilizes a semantic engine to automatically cluster and extract logic from long-tail samples. Through a "semantic-to-rule conversion mechanism," the system can identify and streamline repetitive and redundant rules, automatically detect and warn of logical conflicts between rules, achieving a "slimming down" of the rule base and logical self-healing.

[0019] Beneficial effects: It effectively suppressed the "rule explosion" phenomenon, transforming the labor cost of system maintenance from "growing in tandem with business volume" to "growing steadily at a low rate." When dealing with complex cross-dimensional business logic, the system's scalability was significantly enhanced, and the deployment cycle for a single business logic adjustment was shortened by more than 70%.

[0020] 3. Addressing the issue of "single analytical dimension and shortcomings in understanding unstructured data". Solution: This system constructs a multi-dimensional feature fusion preprocessing mechanism, utilizing a built-in deep semantic vector analysis module to extract full features from unstructured data such as the intent of official documents, the substantive meaning of contract terms, and the evidentiary value of underwriting materials. The system can map this deep semantic information to a high-dimensional vector space and perform cross-dimensional correlation verification with the structured constraint fields in the rule tree.

[0021] Beneficial effects: This system significantly improves the depth and accuracy of the review. In the scenarios of reviewing official documents of central and state-owned assets and identifying financial risks, the system can detect substantive compliance risks hidden in text descriptions. The false judgment rate and the missed judgment rate are reduced by more than 45% compared with traditional solutions, providing underlying support for high-precision compliance.

[0022] 4. Regarding the issue of "lack of prevention and control measures for large-scale hallucinations and uncontrollable compliance and reliability". Solution: This system establishes a rule-semantic interlocking illusion prevention mechanism, strictly prohibiting large models from directly generating decision instructions, and instead using them as semantic extractors. All semantic outputs must undergo legality verification through a "rigid guardrail" formed by the rule tree. Simultaneously, the system introduces "fact consistency verification stubs" to backtrack and verify the intermediate logic generated by the large model, ensuring it does not deviate from the underlying business logic and regulatory red lines.

[0023] Beneficial effects: It completely solves the problem of "not daring to use" generative AI in demanding business scenarios. By constraining the randomness of AI through the certainty of rules, it ensures 100% compliance and traceability of decision results, and provides an AI application template with industrial-grade reliability for zero-tolerance scenarios such as financial lending and insurance claims.

[0024] 5. Addressing the issue of "insufficient adaptation to vertical domains and lack of a unified constraint framework". Solution: This invention proposes a cross-scenario universal rule tree constraint framework. This framework, through abstract logical interfaces, enables unified modeling of different vertical domains (such as document approval workflows, penetrating regulatory logic, and content security classification). A dynamic scheduling mechanism can flexibly load weight operators for specific business domains based on domain tags, achieving seamless migration of a single architecture across group-based and diversified businesses.

[0025] Beneficial effects: It has achieved "logic reuse" and "unified supervision" across fields and levels, which not only meets the penetrating requirements of central state-owned asset groups for collaborative risk control between parent and subsidiary companies, but also provides large financial institutions with unified compliance standards across product lines, greatly improving the efficiency of overall management and control and the rate of technology reuse in group scenarios.

[0026] This solution, based on vector dimensionality reduction theory, data standardization theory, and rule-grammar mapping theory, pioneers a bidirectional collaborative technology mechanism that uses semantic feedback to support rules. For the high-dimensional semantic feature vectors output by the semantic engine, it performs three academically sound and reproducible technical operations: First, vector dimensionality reduction, using Principal Component Analysis (PCA) to remove redundant features, compressing the high-dimensional semantic vectors into a low-dimensional computable space, eliminating the curse of feature dimensionality; Second, quantization and standardization mapping, using the Min-Max normalization algorithm to transform abstract semantic features into numerical standard variables in the [0,1] interval, achieving quantitative representation of semantic features; Third, rule-grammar mapping, based on the formal grammar specifications of the rule engine, transforming standardized numerical variables into decision factors identifiable by rule nodes, generating computable, compatible, and executable virtual rule factors. These virtual rule factors are then injected back into the rule tree decision nodes in real time through a dedicated data interface, serving as supplementary decision conditions for rigid rules, constructing a bidirectional collaborative closed loop of "semantic parsing - factor generation - back injection - joint decision-making," completely breaking down the technical barriers between rules and semantics. This invention is not a simple result superposition, but rather achieves collaboration through physical-level data exchange. First, the semantic features of unstructured text are normalized and discretized, transforming them from probabilistic vector representations into deterministic symbolic representations. Second, using memory mapping technology, the generated virtual rule factors (VRFs) are injected into the execution stack of the rule engine to complete the rule tree's missing judgments in the unstructured dimension.

[0027] In terms of substantial contributions, the technology breaks through inherent technical biases in the field, pioneering an academic path for transforming unstructured semantics into rule-based decision factors, and realizing the academic transformation and technical implementation of "probabilistic semantic features" into "deterministic rule variables." This completely solves the industry's academic challenge of integrating unstructured data into rigid rule-based decision-making. It achieves a substantial transformation from "semantic parsing results" to "rule-based decision variables." Experimental data shows that this closed-loop injection mechanism increases the system's accuracy in complex scenarios from 85.6% to 96.8%, while reducing the rate of human intervention by approximately 77%.

[0028] This invention achieves multi-dimensional technological upgrades compared to existing technologies through an innovative architecture of rule tree and dynamic semantic switching, combined with a large-scale model illusion prevention and control mechanism. It strictly follows the MECE principle (mutually exclusive and collectively exhaustive) to decompose eight major vertical application scenarios, with no business overlap between the scenarios and covering the high compliance requirements of all categories. Attached Figure Description

[0029] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein: Figure 1 This is an architecture diagram of an audit path decision system based on rule tree and dynamic semantic switching according to the present invention; Figure 2 This is a flowchart of a review path decision-making method based on rule tree and dynamic semantic switching according to the present invention; Detailed Implementation

[0030] To make the technical means, creative features, achieved objectives, and effects of this invention readily understandable, the invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the relevant components of the invention.

[0031] According to the technical solution of the present invention, without changing the essential spirit of the present invention, those skilled in the art can propose various interchangeable structural methods and implementations. Therefore, the following detailed embodiments and accompanying drawings are merely exemplary descriptions of the technical solution of the present invention, and should not be regarded as the entirety of the present invention or as a limitation or restriction of the technical solution of the present invention.

[0032] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0033] Example 1, such as Figure 1As shown, an audit path decision system based on rule tree and semantic dynamic switching includes an unstructured data processing module, a rule tree configuration module, a semantic vector analysis module, a dynamic switching strategy controller, and a closed-loop feedback learning module. The unstructured data processing module receives and splits upstream form data, and sends the split feature stream to the rule tree configuration module. The rule tree configuration module calls the semantic vector analysis module through the dynamic switching strategy controller to generate semantic vectors, and converts the semantic vectors into virtual rule factors and uses the closed-loop feedback learning module to back-inject them into the rule tree engine to achieve bidirectional collaboration.

[0034] In the above scheme, the system adopts a dual-engine coupled architecture. Wherein: The unstructured data processing module is responsible for receiving upstream form data and splitting it into feature streams with strong, medium, and weak fields.

[0035] As a preferred approach, the unstructured data processing module utilizes its internal data feature deconstruction module to decompose the input task into three types of features: strongly structured features, moderately structured features, and weakly unstructured features, based on ontology library identification. These three types of features are then sent as a feature stream to the rule tree configuration module.

[0036] Strongly structured features can be IDs, numerical values, or enumerations; moderately structured features can be tables or lists; and weakly structured features can be free text or OCR results.

[0037] The rule tree configuration module stores strong constraint rules based on business logic through a rule tree layer, with each node representing a decision condition, and uses a traditional rule engine to quickly perform deterministic logic operations in the rule tree layer.

[0038] As a preferred solution, the rule tree configuration module sets up a four-level hierarchical rule tree, which are as follows: L1 Global Policy Layer - Root Node: Determines the basic task classification and is used for global policy; L2 Compliance Element Layer - Branch Node: Corresponds to specific compliance items, used for compliance elements; L3 Decision Action Layer - Leaf Nodes: Map initial judgments such as pass and rejection, used for decision actions; L4 Dynamic Intervention Point - Anchor Node: A preset low-confidence monitoring point, specifically used to trigger semantic switching logic and for dynamic intervention point and low-confidence monitoring.

[0039] The semantic vector analysis module has a built-in large language model or deep NLP model to segment and clean unstructured text, and map it into a high-dimensional semantic vector space to generate semantic vectors, which are used to identify the specific intent, sentiment polarity and key entities of the content.

[0040] As a preferred approach, the semantic vector analysis module incorporates a model routing algorithm that schedules models from lightweight to ultra-large-scale based on task complexity.

[0041] The dynamic switching strategy controller is used to monitor the execution status of the rule tree in real time and calculate the "confidence" of the current match. Based on the set threshold, it dynamically decides whether to interrupt / suspend the current rule tree flow, thereby triggering the intervention of the semantic analysis layer.

[0042] The so-called "confidence level" is the information entropy. According to the set threshold, when the entropy value is too high, it will trigger the semantic analysis layer to intervene and activate the semantic branch.

[0043] Therefore, as a preferred solution, the dynamic switching strategy controller calculates the information entropy of the rule path in real time, and activates the semantic branch when the entropy value is too high.

[0044] The closed-loop feedback learning module collects the differences between the final human review results and the machine decision results, optimizes the vector weights of the semantic analysis layer through gradient updates, and recommends new rule nodes to the rule tree layer.

[0045] As a preferred solution, the closed-loop feedback learning module incorporates a comprehensive prevention and control system for large-scale model illusions, including pre-contextual cue constraints, mid-context logic verification, and post-context conflict adjudication. Pre-contextual cue constraints are used to inject the context of the current path in the rule tree; mid-context logic verification is used to perform triple verification of the variables output by the model ("type-range-logic"); and post-context conflict adjudication is used to make deterministic decisions based on the priority weight matrix.

[0046] It should also be noted that in this scheme, the closed-loop feedback learning module introduces a dual verification operator in the semantic processing stage: Consistency check operator: For the same input source, perform multiple inferences using different random seeds or temperature coefficients. The system calculates the semantic entropy value of the output result; if the entropy value exceeds a preset threshold... If it is not a hallucination, the weight factor of the output will be automatically reduced. ; External knowledge graph comparison operator: This operator compares the entity relationships extracted from semantic parsing with the business domain knowledge graph. If the parsing result violates common sense or logical reasoning, the system will immediately trigger an alarm and suspend the decision.

[0047] It should also be noted that in this scheme, the closed-loop feedback learning module has the ability to continuously optimize and evolve itself. Contrastive learning-based boundary fine-tuning: Using the final human audit result as the "gold standard," a contrastive learning loss function is used to fine-tune the classification boundary of the semantic space, improving subsequent... Accuracy of extraction; In the rule tree configuration module, an automatic rule accumulation mechanism is set up: the system analyzes the combination of frequently occurring virtual rule factors through clustering; when a certain semantic feature pattern reaches the statistical significance level, the system automatically generates the corresponding static rule suggestions, which are then accumulated as L2 and L3 level hard-coded rules after being confirmed by the administrator, realizing the dynamic migration from "flexible semantics" to "rigid rules".

[0048] Example 2: To verify the effectiveness and engineering application value of the "Review Path Decision System Based on Rule Tree and Semantic Dynamic Switching" proposed in this invention, the experiment selected the document compliance review and related financial business risk control scenario of a large central state-owned enterprise group as the test benchmark.

[0049] Data Source and Timeliness: The test dataset is entirely derived from the real business flow logs of a large group from Q4 2023 to Q2 2024. After de-identification and sensitive field obfuscation (anonymization), the raw data is constructed into a benchmark library containing 150,000 samples to ensure that the experimental environment highly replicates the real production scenario.

[0050] Strict data isolation (to prevent overfitting): The experiment strictly adheres to the principle of "physical isolation between the training and test sets." Test set samples are completely invisible during model training. Furthermore, to eliminate data randomness bias, the experiment employs a 5-fold cross-validation mechanism.

[0051] Sample Distribution and Expert Annotation Standards: This dataset strictly covers three types of samples: simple compliance (explicit rules), moderate complexity (semantic ambiguity), and high complexity (logical conflicts), with a distribution ratio of 5:3:2. Among them, the medium and high complexity scenario samples (accounting for 50% in total) were strictly manually cross-blinded and annotated by senior business audit experts according to three high-risk judgment criteria: "rule coverage blind spots, core semantic ambiguity, and multi-dimensional conditional logical conflicts," in order to establish the "gold standard" (Ground Truth) of the experiment.

[0052] Simulation Expansion and A / B Testing Verification: Given the low trigger frequency of extreme long-tail edge scenarios (such as sudden policy adjustments and highly concealed chain-penetration violations) in real logs, this experiment introduced a Monte Carlo simulation-based anomaly data injection mechanism to comprehensively verify the system's robustness. Simultaneously, in the later stages of the experiment, the system was integrated with the group's core business line's Shadow Mode for online A / B testing. By comparing the actual bypass traffic intercepted by traditional hard-coded rule engines, the quantitative credibility of this system in reducing manual intervention rates and controlling false negative rates was further solidified.

[0053] Comparison Plan (Baseline): Comparison with Solution A: Traditional hard-coded rule engine (pure rule-based decision-making).

[0054] Comparison with Option B: Single deep learning model audit (pure semantic analysis).

[0055] Comparison with Scheme C: Serial combination scheme (rule filtering directly enters the semantic model without dynamic feedback).

[0056] The present invention provides a review path decision system based on rule trees and dynamic semantic switching (including virtual factor injection and dual-weighted scoring).

[0057] The experiment quantitatively compared the accuracy of the review, system response latency, and the rate of secondary human intervention. The results of the comparison of the core performance indicators are shown in the table below: Overall audit accuracy: This accuracy is a stable result obtained through 5-fold cross-validation on a completely unseen test set, excluding the model's memory effect on specific data.

[0058] Recall rate in complex scenarios: The recall rate of 94.5% specifically refers to the performance in "high-difficulty conflict scenarios" annotated by experts, which proves the superiority of the semantic engine and the dynamic switching mechanism of the rule tree in handling logical contradictions.

[0059] This demonstrates a significant improvement in review accuracy and recall capabilities. Experimental results show that the solution of this invention achieves an overall accuracy of 96.8%, which is far higher than the comparative solution.

[0060] The experimental results will be analyzed below: Analysis of the reasons: Compared to Solution A, which is unable to handle unstructured semantics, Solution B is almost ineffective in complex scenarios; while Solution B can understand semantics, it is prone to misjudgment at the rigid compliance boundaries of "strong risk control" requirements. This invention, through a dynamic switching mechanism of rule tree initial screening and semantic engine intervention, retains the rigor of the rules while leveraging the understanding power of semantics. In particular, the virtual rule factor reverse injection technology enables the semantic analysis results to be transformed into rule judgments, eliminating the decision-making gap between the two systems.

[0061] Experimental results show that the present invention significantly improves the balance between decision-making efficiency and computing power consumption.

[0062] In terms of response latency, the proposed solution (185ms) is significantly better than the pure semantic solution (850ms) and the serial solution (910ms).

[0063] Analysis of the reasons: Pure semantic solutions involve expensive deep learning inference for all data. This invention employs a dual-stream data splitting approach, where approximately 60% of simple, compliant data completes decisions at the rule tree stage, eliminating the need for the semantic engine; the semantic engine is only triggered for complex data with low confidence. This on-demand computing power allocation strategy significantly improves processing capabilities in high-concurrency scenarios while maintaining accuracy.

[0064] Experimental results show that the system's self-evolution capability is verified more efficiently.

[0065] Tests on the "closed-loop feedback iterative optimization" step show that the convergence speed of the present invention is extremely fast.

[0066] Analysis of the reasons: The traditional model (Solution B) requires a large amount of labeled data for retraining (about 15 days). In contrast, this invention, through decision log feedback and dynamic weight adjustment, can adjust the dual-weighted scoring model in real time based on the auditor's corrective actions, and complete the incremental update of rule tree nodes in a short period of time, giving the system stronger business adaptability.

[0067] In summary, the system and method described in this invention, while ensuring extremely high review accuracy, effectively solve the contradiction of "rules that are too rigid and semantics that are too vague" in complex unstructured review scenarios, significantly reducing the workload of manual review, and have significant technical advancement and practical application value.

[0068] Example 3: A review path decision-making method based on rule trees and dynamic semantic switching. The method extracts feature vectors for the task through a data preprocessing module. When the rule matching confidence is below a threshold, a semantic analysis engine is invoked to transform the feature vectors into high-dimensional semantic vectors, and further, intent vectors are generated based on business context clustering. The review path decision-maker generates discrete virtual rule factors based on the intent vectors to achieve dynamic switching of the review path. Specifically, the method includes the following steps: Step 1: Task Access and Two-Way Feature Streaming: The system receives a task package containing document content, attachments, and metadata, and uses the data feature deconstruction module to vectorize the features.

[0069] Step 2: Static depth-first traversal of the rule tree and quantitative determination of confidence: The system extracts the structured feature vector of the task to be reviewed through the data feature deconstruction module. The rule engine traverses from the L1 root node and introduces a confidence scoring function. ,in For the current node; if an ambiguous field is encountered or an L4 anchor node is entered, calculate... ,in, As weight, Matching factor; This is an index variable representing the current rule node. The included first One audit factor; if , This is a preset safety threshold.

[0070] If no fuzzy field is encountered or no L4 anchor node is entered, continue the static rule tree traversal and proceed directly to step five. If a fuzzy field is encountered or an L4 anchor node is entered, If a preset safety threshold is set, the switch will be triggered, and the process will proceed to step three.

[0071] The rule tree confidence scoring algorithm calculates the initial confidence level using a multi-factor weighted scoring model during the initial rule tree screening stage. Its calculation formula is expressed as: in, The first one representing the data to be reviewed Each structured feature value is a strong structured feature extracted from the official document or financial document to be reviewed through the "data feature deconstruction module"; This represents the corresponding preset rule logic; k is the total number of structured feature items participating in the confidence evaluation; m is the total number of preset rigid compliance constraints. This is a feature matching function used to calculate the degree of overlap between feature values ​​and rules; For the preset weighting coefficients, satisfy ; It is a rigid constraint operator, and its value is a binary logic value. .

[0072] Step 3: Semantic Vector Analysis and Intent Extraction: Extracting the intent vector from unstructured text in the semantic layer task. After the switch is triggered, the semantic layer takes over the semantic branch and extracts the unstructured text intent vector.

[0073] Step 4: Virtual Rule Factors Generation and Injection: Semantic vectors are transformed into virtual rule factors and injected back into the rule tree engine to achieve bidirectional collaboration.

[0074] Its main steps involve transforming semantic vectors through a mapping function to generate discrete virtual rule factors. and virtual rule factors The missing dimension of the reverse injection rule tree.

[0075] As a preferred approach, the generation and injection of virtual rule factors consists of three steps: the identifier identification stage, the mapping invocation stage, and the injection execution stage. Identification Phase: When loading the rule tree model, the system automatically identifies virtual rule factor nodes in the configuration file using a pre-defined specific prefix identifier. For example, it automatically identifies virtual rule factor nodes in XML or JSON configuration files using a pre-defined specific prefix identifier (such as V_Factor_).

[0076] Mapping and Invocation Phase: Based on the name of the virtual rule factor, the system matches the corresponding logical function or SQL query template in the operator library, mapping the abstract factor name to an executable underlying operator.

[0077] During the mapping invocation phase: For unstructured text, the semantic engine transforms it into an intent vector. Then through the mapping function Converted into virtual factors that the rule engine can recognize: in, The learned projection matrix is ​​used to reduce the dimensionality of high-dimensional semantic features to the business logic dimension. This is an activation function used to filter out noisy features; This is a bias term vector used to adjust the judgment threshold benchmark under different business scenarios; The quantization function transforms a continuous vector into a discrete state enumeration code through Top-K sampling or threshold truncation.

[0078] Injection execution phase: When the rule tree traverses to the factor node, the execution unit dynamically pulls the specified dimension data from the intent vector to be reviewed as input parameters; injects the input parameters into the mapped operator to perform Boolean value calculation or interval matching; and feeds the calculation results back to the decision path branch to determine the next jump or termination of the rule tree.

[0079] Step 5: Joint Decision Making, Path Redirection, and Closed-Loop Learning: Based on the injected complete factors, the rule tree generates the final decision path. The system records the decision log, and the mapping function is adjusted through human feedback. Fine-tuning is then performed. The final decision includes options such as approval or rejection.

[0080] After injection, the decision dimensions of the rule tree expand from simple "numerical / enumeration" to "numerical / enumeration + semantic intent". At this point, the rule tree possesses all the dimensional information required to process the task, hence it is called a "complete factor".

[0081] Additionally, it should be noted that this system employs a conflict resolution evaluation model based on preference relationships: in, This represents the value determined by the rule path; Represents the semantic path probability distribution; The credibility factor for hallucination prevention is dynamically adjusted based on the model's historical correction rate and logical verification score. This represents the stiffness / flexibility coefficient, which is determined by the relevant business area.

[0082] Example 4, based on Examples 1-3, provides an audit path decision system based on rule tree and dynamic semantic switching, comprising an unstructured data processing module, a rule tree configuration module, a semantic vector analysis module, and a scheduling decision module. The unstructured data processing module converts the original document to be audited into a structured document to be audited vector using OCR or text extraction operators. The rule tree configuration module constructs a hierarchical decision tree containing multiple logical nodes, embedding virtual rule factors for specific compliance items within the logical nodes. The semantic vector analysis module uses a preset language model to perform semantic compliance detection on the document to be audited and outputs a semantic confidence score. The scheduling decision module integrates a logic gate, which switches the execution path between the semantic vector analysis module and the rule tree configuration module based on the comparison result of the semantic confidence score and a preset threshold.

[0083] Among them, the unstructured data processing module, rule tree configuration module, semantic vector analysis module, and scheduling decision module establish data communication links between each module through remote procedure calls or distributed message queues to realize the issuance of instructions and the return of processing results; the system uses the scheduling decision module to perform rule compensation for branches with low semantic confidence to realize automated decision-making in compliant scenarios.

[0084] In implementation, the rule tree configuration module and semantic vector analysis module, in processing the static depth traversal of the rule tree and the quantitative determination of confidence, the system extracts the structured feature vector of the task to be reviewed through the data feature deconstruction module. ; using preset weighting coefficients Calculate feature values ​​and compliance rules using matching functions The degree of overlap is determined, and a rigid constraint operator is introduced. Binary determination is performed on the status of documents or mandatory compliance items; if any core compliance item does not match, it will result in... If this occurs, the system will forcibly intercept the current decision path and set an alarm.

[0085] In implementation, the semantic vector analysis module extracts deep intent vectors from unstructured text using a pre-trained language model. The system uses a projection matrix. Reduce its dimensionality to the business logic dimension and utilize quantization functions. Continuous semantic features are transformed into discretized state enumeration codes that can be recognized by the rule engine, generating virtual rule factors. Then, the virtual symbol table interface was used to... Write the data in real time to the runtime context storage area of ​​the rule engine, so that it can be used as the native logical input variable for subsequent nodes in the rule tree.

[0086] In implementation, the rule tree configuration module and the semantic vector analysis module adopt a layered decoupling and bidirectional linkage coupling interaction mechanism, and realize the split and collaborative processing of structured data and unstructured data through a standardized communication protocol.

[0087] The scheduling decision module constructs a three-dimensional confidence evaluation model based on rule matching degree, scenario complexity, and compliance level through a dynamic switching strategy controller, so as to realize dynamic path switching control of the entire process, including rule tree execution suspension, semantic engine wake-up, and process recovery.

[0088] The internal logic of the scheduling decision module involves outputting a confidence score S from the semantic vector analysis module and judging whether S < T in the confidence comparator, where T is a preset threshold and the default semantic path. Then, a control signal is output to the path switching switch. After passing through the logical compensation path of the rule tree configuration module, compensation is finally triggered if S < T, and the result is sent to the weight fusion unit. The weight fusion unit outputs the final decision result and feeds it back to the semantic vector analysis module.

[0089] Among them, the confidence comparator is used to receive semantic scores and perform real-time difference calculations; the path switching switch determines whether to activate the rule compensation path based on the comparator results; and the weight fusion unit is used to logically weight multiple signals to ensure compliance red lines in high-risk scenarios.

[0090] The generation mechanism of the virtual rule factors includes: principal component analysis dimensionality reduction, Min-Max quantization standardization mapping, rule syntax adaptation and transformation, and the factors are back-injected into the rule tree decision node for execution through a dedicated interface.

[0091] In implementation, the semantic vector analysis module uses a dynamic weight adaptive adjustment algorithm through a dual-weighted joint scoring model, and realizes the fusion calculation of rule scores and semantic scores through a weighted summation formula, and has a built-in compliance red line veto mechanism for high-risk scenarios.

[0092] Example 5, based on Example 4, provides an audit path decision system based on dynamic switching between rule trees and semantics. The system includes an unstructured data processing module, a rule tree configuration module, a semantic vector analysis module, a scheduling decision module, and a closed-loop feedback learning module. The closed-loop feedback learning module, based on closed-loop feedback control theory, implements a full-process iterative evolution mechanism encompassing the full collection of system decision logs, semantic model gradient optimization, and incremental self-updating of rule tree nodes.

[0093] Example 6: A method for review path decision-making based on rule tree and dynamic semantic switching, based on the system described in Example 4, includes the following steps: Step 1: Convert the original document to be reviewed into a structured document to be reviewed vector using OCR or text extraction operators; Step 2: Construct a hierarchical decision tree containing multiple logical nodes, and embed virtual rule factors for specific compliance items into the logical nodes; Step 3: Use a pre-set language model to perform semantic compliance detection on the vector to be reviewed, and output a semantic confidence score; Step 4: Set up a logic gate to switch the execution path between the semantic vector analysis module and the rule tree configuration module based on the comparison result of the semantic confidence score and the preset threshold.

[0094] Example 7, from an academic perspective, analyzes existing intelligent review systems and finds that they generally suffer from underlying architectural defects such as "lack of decoupling and insufficient collaboration between rigid decision-making modules and flexible intelligent modules": Static rule engines, based on discrete logical reasoning, can only achieve deterministic matching of structured data and lack the ability to semantically perceive and extract features from unstructured data; Artificial intelligence semantic models, based on probabilistic representation and distributed representation, while possessing the ability to parse unstructured data, lack compliance constraints and deterministic decision boundaries, making them prone to logical illusions and decision biases; Existing solutions only adopt a shallow fusion approach with unidirectional serial connection and have not built a standardized layered coupling mechanism, resulting in the inability of the two types of modules to complement each other's technical advantages, low efficiency in the collaborative processing of structured and unstructured data, and difficulty in meeting the stringent requirements of high compliance fields in terms of decision reliability and scenario adaptability.

[0095] This solution, based on layered architecture design theory and modular collaboration, proposes a dual-engine coupled architecture with two-layer decoupling, bidirectional controllability, and standardized interaction: The first layer is a rigid decision engine based on a rule tree, which uses a formal rule expression with a tree topology structure and incorporates predicate logic and condition judgment mechanisms. It specifically processes strongly structured field data such as amount, code, security level, and qualification level, solidifying mandatory compliance clauses and risk control thresholds to ensure the certainty and rigidity of decision-making. The second layer is a flexible semantic analysis engine, built on a pre-trained language model and a vector space model (VSM). It is equipped with word segmentation and noise reduction, semantic embedding, and feature extraction algorithms to specifically parse unstructured data such as official documents, document notes, and risk descriptions, achieving quantitative representation of deep semantic features. The two engines build a standardized communication interface through RPC remote call protocol and message queue middleware, defining a unified data interaction format and instruction scheduling specification to achieve bidirectional controllable linkage of instruction issuance, data feedback, and result fusion. This differs from the simple unidirectional serial mode of existing technologies, forming a rigid-flexible underlying decision framework.

[0096] In terms of substantial contributions, this technology breaks through the technical limitations of a single decision-making architecture and constructs a complementary decision-making underlying framework of "deterministic rules + probabilistic semantics". It provides architectural support and module foundation for subsequent core technologies such as dynamic scheduling, virtual factor injection, and joint scoring, realizes the split and collaborative processing of structured and unstructured data, solves the industry academic problem of the inability to balance "rigid compliance" and "intelligent parsing" in high compliance scenarios, and lays the technical theoretical foundation for the entire invention.

[0097] Example 8, from the perspective of decision scheduling theory, traditional rule engines adopt a static traversal linear scheduling mode, which lacks the ability to quantitatively perceive and adaptively control the decision scenario: it cannot quantitatively evaluate the confidence of the rule matching effect, and cannot achieve intelligent diversion for edge-ambiguous scenarios and non-standardized tasks. It either directly triggers manual intervention, resulting in low scheduling efficiency, or blindly executes rule decisions, leading to misjudgments. At the same time, the undifferentiated AI model scheduling causes redundant waste of computing resources, which violates the principles of optimality and economy of intelligent scheduling. The flexibility and adaptability of the decision link have inherent defects.

[0098] This solution, based on multi-dimensional comprehensive evaluation theory and dynamic scheduling algorithm, constructs a three-dimensional confidence quantification evaluation model. It integrates three core indicators: rule matching fit, business scenario complexity, and compliance level strickenness. The weight of each indicator is determined by the entropy weight method, realizing the 0-1 interval confidence quantification calculation of the initial judgment result of the rule tree. Based on threshold control theory, a dynamic scheduling threshold is set, and the threshold is adaptively fine-tuned in combination with the business characteristics of high compliance fields. When the confidence is higher than the threshold, the closed-loop judgment is executed according to the preset decision path of the rule tree. When the confidence is lower than the threshold, the execution of the rule tree is suspended by an interruption command, and a parsing scheduling command is sent to the semantic engine at the same time. After the semantic engine completes the extraction of unstructured data features, the rule tree process is restarted by a recovery command. This realizes the interruptible, recoverable, and switchable dynamic scheduling of the decision path, forming a complete scheduling closed loop of "perception-judgment-scheduling-recovery".

[0099] In terms of substantial contributions, this technology upgrades the static decision-making chain into an adaptive dynamic decision-making system, realizing precise diversion and intelligent scheduling of decision-making tasks, significantly improving the adaptability to edge scenarios and the efficiency of computing power utilization, providing a prerequisite for the implementation of the core original technology, improving the dynamic decision-making logic of the entire invention, and filling the academic gap in the field of intelligent scheduling and quantitative control of existing technologies.

[0100] Example 9, from the perspective of data fusion and decision collaboration theory, reveals inherent technical biases and academic bottlenecks in this field: existing "rule + AI" fusion solutions all adopt a shallow fusion mode of one-way weighting and result voting, which can only realize the auxiliary reference of AI semantic results for rule decisions, and cannot transform unstructured semantic features into quantitative judgment factors that can be executed by the rule engine. There is a technical barrier between semantic information and the rule decision chain; those skilled in the art generally believe that rule engines can only process structured hard rules, and unstructured semantic results cannot be integrated into rigid decision logic. This technical bias prevents the full release of the deep value of unstructured data, and the problem of the single dimension of rule decision-making remains unresolved. The solution addresses the technical bottleneck in traditional solutions where unstructured semantic information (such as implicit violation expressions) cannot directly drive rigid rule engines to make logical jumps.

[0101] This solution, based on vector dimensionality reduction theory, data standardization theory, and rule-grammar mapping theory, pioneers a bidirectional collaborative technology mechanism that uses semantic feedback to support rules. For the high-dimensional semantic feature vectors output by the semantic engine, it performs three academically sound and reproducible technical operations: First, vector dimensionality reduction, using Principal Component Analysis (PCA) to remove redundant features, compressing the high-dimensional semantic vectors into a low-dimensional computable space, eliminating the curse of feature dimensionality; Second, quantization and standardization mapping, using the Min-Max normalization algorithm to transform abstract semantic features into numerical standard variables in the [0,1] interval, achieving quantitative representation of semantic features; Third, rule-grammar mapping, based on the formal grammar specifications of the rule engine, transforming standardized numerical variables into decision factors identifiable by rule nodes, generating computable, compatible, and executable virtual rule factors. These virtual rule factors are then injected back into the rule tree decision nodes in real time through a dedicated data interface, serving as supplementary decision conditions for rigid rules, constructing a bidirectional collaborative closed loop of "semantic parsing - factor generation - back injection - joint decision-making," completely breaking down the technical barriers between rules and semantics. This invention is not a simple result superposition, but rather achieves collaboration through physical-level data exchange. First, the semantic features of unstructured text are normalized and discretized, transforming them from probabilistic vector representations into deterministic symbolic representations. Second, using memory mapping technology, the generated virtual rule factors (VRFs) are injected into the execution stack of the rule engine to complete the rule tree's missing judgments in the unstructured dimension.

[0102] In terms of substantial contributions, the technology breaks through inherent technical biases in the field, pioneering an academic path for transforming unstructured semantics into rule-based decision factors, and realizing the academic transformation and technical implementation of "probabilistic semantic features" into "deterministic rule variables." This completely solves the industry's academic challenge of integrating unstructured data into rigid rule-based decision-making. It achieves a substantial transformation from "semantic parsing results" to "rule-based decision variables." Experimental data shows that this closed-loop injection mechanism increases the system's accuracy in complex scenarios from 85.6% to 96.8%, while reducing the rate of human intervention by approximately 77%.

[0103] Example 10, from the perspective of multi-source decision fusion theory, shows that the decision results of the rule engine and the semantic engine have dimensional differences and logical conflicts. Existing technologies lack a scientific conflict resolution mechanism: either using a single-dimensional decision leads to one-sided results, or using fixed weights results in poor scenario adaptability. The decision results lack interpretability and repeatability, and cannot meet the dual requirements of decision accuracy and reliability in high compliance scenarios. The fusion and resolution of multi-source decision results has become a key academic pain point restricting the performance of intelligent auditing systems.

[0104] This solution, based on multi-attribute decision theory and dynamic weighting algorithms, constructs a two-dimensional adaptive weighted scoring model: It extracts the rule matching score from the rule engine and the parsing confidence score from the semantic engine as core decision attributes. A dynamic weight adjustment function is built based on scenario type, compliance level, and data characteristics to achieve adaptive allocation of the two score weights. The comprehensive decision score is calculated using a weighted summation formula, expressed as follows: (in For the overall score, For dynamic weighting coefficients, To score points according to the rules, (Semantic score); the optimal decision path is selected based on the comprehensive score ranking, while a compliance hard red line constraint mechanism is built in, setting a veto threshold for high-risk scenarios to prevent illegal decision deviations and achieve scientific integration and conflict resolution of multi-source decision results.

[0105] In terms of substantive contributions, this technology establishes an interpretable, computable, and repeatable multi-source decision conflict adjudication system, realizes the academic integration of rule-based decision-making and semantic decision-making, takes into account both decision accuracy and compliance rigidity, improves the decision output logic of the entire invention, enhances the reliability and scenario adaptability of system decisions, and provides an academic solution for intelligent decision integration in highly compliant fields.

[0106] Example 11: From the perspective of system evolution and control theory, the existing review system lacks a closed-loop feedback and self-evolution mechanism, and belongs to an open-loop decision-making system: the rule base adopts a static hard-coded mode, which is prone to rule expansion and logical conflicts in long-term iteration; the semantic model parameters are fixed and cannot adapt to the dynamic changes of business scenarios and compliance policies, the system's full life cycle operation and maintenance costs increase exponentially, and it is impossible to achieve continuous iterative optimization of technical performance, which violates the academic concept of adaptive evolution of intelligent systems.

[0107] This solution, based on closed-loop feedback control theory and incremental learning algorithms, constructs a full-lifecycle self-optimizing system: First, it establishes a full-process decision log collection module, collecting comprehensive data from a time-series database, including manual review results, decision deviation data, illusion correction records, and scheduling logs, to build a standardized training dataset. Second, it employs a gradient descent optimization algorithm to iteratively update the vector weights, confidence thresholds, and dynamic weighting coefficients of the semantic model, improving model parsing and scheduling accuracy. Third, it uses an association rule mining algorithm to extract features from historical decision data, enabling incremental optimization, addition, and redundancy removal of rule tree nodes, thus curbing rule expansion. Fourth, it synchronizes the optimized model parameters and rule base to the dual-engine module, forming a closed-loop feedback chain of "data collection—model optimization—rule update—decision implementation," achieving adaptive evolution of the system.

[0108] In terms of substantial contributions, this technology upgrades the open-loop decision-making system into a closed-loop adaptive evolutionary system, solving the common academic problems of rule expansion and model solidification in the industry. It enables continuous iterative optimization of system performance, improves the execution accuracy of the preceding core technologies, extends the technology life cycle, and has significant academic value and industrial application prospects.

[0109] In summary, this invention, through its innovative architecture of rule trees and dynamic semantic switching, combined with a large-scale model illusion prevention mechanism, achieves multi-dimensional technological upgrades compared to existing technologies. It strictly adheres to the MECE principle (mutually exclusive and collectively exhaustive) to decompose eight major vertical application scenarios, each with no business overlap and covering all product categories' high compliance requirements. Specific implementation details are shown in the table below: The above scenarios fully cover the needs of high compliance, strong risk control, and complex unstructured audits. This invention can specifically improve the automation rate of complex edge tasks, break the rigidity of static decision-making, eradicate the problem of rule expansion, accurately prevent and control the illusion of large models, and optimize the allocation of computing resources, and has the value of large-scale industrial application.

[0110] Notes, Glossary of Terms: Rule tree-semantic dual-engine coupling: a two-layer decision architecture with deep collaboration and division of labor between the rigid rule tree decision engine and the flexible semantic analysis engine.

[0111] Confidence threshold: A quantitative indicator of the credibility of the decision-making rule. If the value is lower than this, the semantic engine will be triggered to intervene dynamically.

[0112] Virtual rule factors: Standardized variables generated by dimensionality reduction of semantic analysis results, used for back-injection into the rule tree to achieve semantic feedback decision-making.

[0113] Dual-weighted joint scoring model: A dynamic decision-making algorithm that integrates rule matching score and semantic relevance score, achieving a balance between rigid and flexible decision-making through weight coefficients.

[0114] Group-level penetrating risk management: a management model that conducts full-chain risk screening and auditing of subsidiaries and business lines at all levels of a large group.

[0115] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A review path decision system based on rule tree and dynamic semantic switching, characterized in that: It includes an unstructured data processing module, a rule tree configuration module, a semantic vector analysis module, a dynamic switching strategy controller, and a closed-loop feedback learning module. The unstructured data processing module receives and splits upstream form data, and sends the split feature stream to the rule tree configuration module. The rule tree configuration module calls the semantic vector analysis module through the dynamic switching strategy controller to generate semantic vectors, and converts the semantic vectors into virtual rule factors. The closed-loop feedback learning module then injects these virtual rules back into the rule tree configuration module to achieve bidirectional collaboration.

2. The review path decision system based on rule tree and dynamic semantic switching according to claim 1, characterized in that: The system adopts a dual-engine coupled architecture; wherein: The unstructured data processing module is responsible for receiving upstream form data and splitting it into feature streams with strong, medium, and weak fields. The rule tree configuration module stores strong constraint rules based on business logic through a rule tree layer, with each node representing a decision condition, and uses a traditional rule engine to quickly perform deterministic logic operations in the rule tree layer. The semantic vector analysis module has a built-in large language model or deep NLP model to segment and clean unstructured text and map it into a high-dimensional semantic vector space for identifying the specific intent, sentiment polarity and key entities of the content. The dynamic switching strategy controller is used to monitor the execution status of the rule tree in real time, calculate the "confidence" of the current match, and dynamically decide whether to interrupt / suspend the current rule tree flow based on the set threshold, thereby triggering the intervention of the semantic analysis layer; The closed-loop feedback learning module collects the differences between the final human review results and the machine decision results, optimizes the vector weights of the semantic analysis layer through gradient updates, and recommends new rule nodes to the rule tree layer.

3. The review path decision system based on rule tree and dynamic semantic switching according to claim 2, characterized in that: The unstructured data processing module uses its internal data feature deconstruction module to identify the input task through the ontology library, decomposing it into: strongly structured features, moderately structured features, and weakly unstructured features, and then sending these three types of features as a feature stream to the rule tree configuration module. The rule tree configuration module sets up a four-level hierarchical rule tree: L1 Global Strategy Layer - Root Node: determines the basic category of tasks; L2 Compliance Element Layer - Branch Node: corresponds to specific compliance items; L3 Decision Action Layer - Leaf Node: maps to initial judgments such as pass and rejection; L4 Dynamic Intervention Point - Anchor Point Node: preset low confidence monitoring points, specifically used to trigger semantic switching logic. The semantic vector analysis module has a built-in model routing algorithm that schedules models from lightweight to ultra-large-scale based on task complexity. The dynamic switching strategy controller calculates the information entropy of the rule path in real time, and activates the semantic branch when the entropy value is too high. The closed-loop feedback learning module incorporates a comprehensive prevention and control system for large-scale model illusions, including pre-contextualized prompt constraints, mid-contextualized logic checks, and post-contextualized conflict resolution. Pre-contextualized prompt constraints are used to inject the context of the current path in the rule tree; mid-contextualized logic checks are used to perform triple checks on the variables output by the model, including "type-range-logic"; and post-contextualized conflict resolution is used to make deterministic decisions based on the priority weight matrix.

4. The review path decision system based on rule tree and dynamic semantic switching according to claim 3, characterized in that: The closed-loop feedback learning module introduces a dual verification operator in the semantic processing stage: Consistency check operator: For the same input source, perform multiple inferences using different random seeds or temperature coefficients. The system calculates the semantic entropy value of the output result; if the entropy value exceeds a preset threshold... If it is not a hallucination, the weight factor of the output will be automatically reduced. ; External knowledge graph comparison operator: This operator compares the entity relationships extracted from semantic parsing with the business domain knowledge graph. If the parsing result violates common sense or logical reasoning, the system will immediately trigger an alarm and suspend the decision.

5. The review path decision system based on rule tree and dynamic semantic switching according to claim 3, characterized in that: The closed-loop feedback learning module has the ability to continuously optimize and evolve itself. Contrastive learning-based boundary fine-tuning: Using the final human audit result as the "gold standard," a contrastive learning loss function is used to fine-tune the classification boundary of the semantic space, improving subsequent... Accuracy of extraction; Automatic rule accumulation mechanism: The system analyzes frequently occurring virtual rule factor combinations through clustering; when a certain semantic feature pattern reaches the statistical significance level, the system automatically generates corresponding static rule suggestions, which are then accumulated as L2 / L3 level hard-coded rules after being confirmed by the administrator, realizing the dynamic migration from "flexible semantics" to "rigid rules".

6. A method for review path decision-making based on rule tree and dynamic semantic switching, based on the review path decision-making system based on rule tree and dynamic semantic switching as described in any one of claims 1-5, characterized in that: The feature vector of the task is extracted through the data preprocessing module; when the rule matching confidence is lower than the threshold, the semantic analysis engine is called to convert the feature vector into a high-dimensional semantic vector, and further generate an intent vector based on business context clustering. The review path decision-maker generates discrete virtual rule factors based on the intent vector to achieve dynamic switching of the review path; specifically, it includes the following steps: Step 1: Task Access and Bidirectional Feature Streaming: The system receives a task package containing document content, attachments, and metadata, and uses the data feature deconstruction module to vectorize the features. Step 2: Static depth-first traversal of the rule tree and quantitative determination of confidence: The system extracts the structured feature vector of the task to be reviewed through the data feature deconstruction module. The rule engine traverses from the L1 root node and introduces a confidence scoring function. ,in For the current node; if an ambiguous field is encountered or an L4 anchor node is entered, calculate... ,in, As weight, As the matching factor, For index variables; if , Preset safety threshold; Step 3: Semantic Vector Analysis and Intent Extraction. The semantic layer extracts the intent vector from the unstructured text in the task. ; Step 4: Virtual rule factor generation and injection. The semantic vector is transformed into virtual rule factors and injected back into the rule tree engine to achieve bidirectional collaboration. Step 5: Joint Decision Making, Path Redirection, and Closed-Loop Learning. Based on the injected complete factors, the rule tree generates the final decision path. The system records the decision log and adjusts the mapping function through human feedback. Make minor adjustments.

7. The method for review path decision-making based on rule tree and dynamic semantic switching according to claim 6, characterized in that: In step four, the virtual rule factor generation and injection are divided into three steps: the identifier identification stage, the mapping invocation stage, and the injection execution stage, wherein: Identification phase: When loading the rule tree model, the system automatically identifies the virtual rule factor nodes in the configuration file using a preset specific prefix identifier; Mapping and Invocation Phase: Based on the name of the virtual rule factor, the system matches the corresponding logical function or SQL query template in the operator library, mapping the abstract factor name to an executable underlying operator; Injection execution phase: When the rule tree traverses to the factor node, the execution unit dynamically pulls the specified dimension data from the intent vector to be reviewed as input parameters; injects the input parameters into the mapped operator to perform Boolean value calculation or interval matching; and feeds the calculation results back to the decision path branch to determine the next jump or termination of the rule tree.

8. The method for review path decision-making based on rule tree and dynamic semantic switching according to claim 6, characterized in that: During the mapping invocation phase, for unstructured text, the semantic engine transforms it into an intent vector. Then through the mapping function Converted into virtual factors that can be recognized by the rule engine: in, The learned projection matrix is ​​used to reduce the dimensionality of high-dimensional semantic features to the business logic dimension. This is an activation function used to filter out noisy features; This is a bias term vector used to adjust the judgment threshold benchmark under different business scenarios; The quantization function transforms a continuous vector into a discrete state enumeration code through Top-K sampling or threshold truncation.

9. The method for review path decision-making based on rule tree and dynamic semantic switching according to claim 6, characterized in that: In step two, the rule tree confidence scoring algorithm calculates the initial confidence level using a multi-factor weighted scoring model during the initial rule tree screening stage. ; Its calculation formula is expressed as follows: in, The first one representing the data to be reviewed Each structured feature value is a strong structured feature extracted from the official document or financial document to be reviewed through the "data feature deconstruction module"; This represents the corresponding preset rule logic; k is the total number of structured feature items participating in the confidence evaluation; m represents the total number of preset rigid compliance constraints; This is a feature matching function used to calculate the degree of overlap between feature values ​​and rules; For the preset weighting coefficients, satisfy ; It is a rigid constraint operator, and its value is a binary logic value. .

10. The method for review path decision-making based on rule tree and dynamic semantic switching according to claim 6, characterized in that: This system adopts a conflict resolution evaluation model based on preference relationships: in, This represents the value determined by the rule path; Represents the semantic path probability distribution; The credibility factor for hallucination prevention is dynamically adjusted based on the model's historical correction rate and logical verification score. This represents the stiffness / flexibility coefficient, which is determined by the relevant business area.

Images