An illusion suppression intelligent auxiliary decision-making method and device fusing a dynamic knowledge graph and a time and space snapshot mechanism, equipment and a storage medium

Abstract

This invention provides a method, apparatus, device, and storage medium for intelligent auxiliary decision-making to suppress hallucinations, integrating a dynamic knowledge graph and a spatiotemporal snapshot mechanism. By constructing a dynamic knowledge graph with timestamps and security level labels, a read-only snapshot of the graph is frozen at the time of the request upon receiving a decision request. A secure, isolated subgraph is generated based on user permissions, ensuring the spatiotemporal consistency and security of the data from the source. Decision suggestions generated based on the subgraph are decomposed into atomic propositions, which are quantitatively verified using topological similarity, numerical matching, and multimodal consistency to obtain a comprehensive confidence level. Low-confidence propositions drive iterative correction of the large language model, achieving hallucination recognition and suppression. Simultaneously, user-corrected information is used to update the graph structure and weights, and a hash chain records the version and verification results, ensuring the reliability, traceability, and continuous optimization of the decision-making process.

Description

Technical Field

[0001] This invention relates to the intersection of artificial intelligence and command and control, and in particular to a method, apparatus, device and storage medium for intelligent decision-making assistance in hallucination suppression that integrates dynamic knowledge graphs and spatiotemporal snapshot mechanisms. Background Technology

[0002] As large language models are increasingly applied in high-value scenarios such as command and control, emergency dispatch, and defense intelligence analysis, intelligent auxiliary decision-making systems based on retrieval-enhanced generation have become an important technological approach in this field. However, these scenarios place stringent demands on the factual accuracy, timeliness, and security of decision recommendations, which existing technologies struggle to fully meet.

[0003] Existing intelligent decision-making assistance solutions based on knowledge base retrieval typically employ a sequential architecture of "real-time retrieval + large language model generation." A typical problem with this architecture is that the underlying knowledge base is not static during a single decision-making process; new information is injected at any time, and old data is overwritten and updated. Furthermore, the reasoning process often involves multiple rounds of retrieval, multiple steps of generation, and multiple verifications, with an execution cycle longer than the knowledge base's update cycle. This leads to different stages of the same decision-making process potentially referencing different versions of data from the knowledge base. For example, the generation stage might use older terrain data, while the verification stage might compare it with newer electromagnetic environment data, causing a "temporal discrepancy" between the reasoning context and the final facts. This results in contradictory or even completely invalid decision recommendations.

[0004] Furthermore, even with the introduction of data snapshots to alleviate the aforementioned temporal discrepancies, three deep-seated obstacles remain: First, specialized domain data typically contains multiple security levels, and a simple full snapshot would expose high-security data to the context of prompts in large language models, posing a risk of leakage through side channels of inference traces; second, traditional hallucination verification relies solely on Boolean matching of text similarity, which cannot handle ambiguous situations such as sensor data noise and estimation errors, and lacks the ability to identify hallucinations at the topological, numerical, and multimodal levels of the graph; third, the implicit knowledge contained in the correction operations of experts to the system output during actual decision-making cannot be absorbed by the system in real time and transformed into structural updates to the knowledge base, preventing the system from continuously evolving.

[0005] In view of the above, this application is hereby submitted. Summary of the Invention

[0006] This invention discloses an intelligent decision-making aid method, device, equipment, and storage medium for hallucination suppression that integrates dynamic knowledge graphs and spatiotemporal snapshot mechanisms. It aims to address the difficulty in considering data visibility during decision-making and reasoning in dynamic, multi-level knowledge graph environments. Figure 1 Consistency and security isolation lead to the problem that the illusions output by large language models cannot be effectively checked and continuously corrected.

[0007] The first embodiment of the present invention provides an intelligent auxiliary decision-making method for hallucination suppression that integrates dynamic knowledge graphs and spatiotemporal snapshot mechanisms, including: A multimodal dynamic knowledge graph is constructed, wherein each node and edge in the dynamic knowledge graph is marked with a timestamp and a security level label; In response to the user's decision request, the dynamic knowledge graph is frozen into a read-only graph snapshot at the time of the request, and a unique version number is assigned to the read-only graph snapshot; according to the user's security permission level, nodes in the read-only graph snapshot with a security level label higher than the user's security permission level and their associated edges are masked in memory to generate a security isolation subgraph; The security isolation subgraph is serialized into prompt words, and combined with anti-illusion constraint rules, a large language model is invoked to generate preliminary decision suggestions associated with the version number; The preliminary decision recommendations are broken down into atomic propositions. A three-dimensional quantitative check is performed on each atomic proposition on the secure isolation subgraph to obtain three sub-scores: topological similarity, numerical matching degree, and multimodal consistency. The scores are then weighted and summed to obtain the comprehensive confidence score. When the comprehensive confidence score is lower than a preset threshold, the conflict dimension information is fed back to the large language model for iterative correction. Output the final decision suggestion carrying the version number, obtain the user's correction operation on the final decision suggestion, extract the difference feature vector, and update the topology or edge weight of the dynamic knowledge graph based on the difference feature vector. The version number, the scores of each round of verification, and the correction records are stored together using a hash chain.

[0008] The second embodiment of the present invention provides an intelligent auxiliary decision-making device for hallucination suppression that integrates dynamic knowledge graphs and spatiotemporal snapshot mechanisms, comprising: The graph construction module is used to construct a multimodal dynamic knowledge graph, in which each node and edge is marked with a timestamp and a security level label; The security isolation module is used to respond to the user's decision request, freeze the dynamic knowledge graph into a read-only graph snapshot based on the request time, and assign a unique version number to the read-only graph snapshot; according to the user's security permission level, the nodes in the read-only graph snapshot with a security level label higher than the user's security permission level and their associated edges are blocked in memory to generate a security isolation subgraph; A constraint generation engine is used to serialize the security isolation subgraph into prompt words, and combine it with anti-illusion constraint rules to call a large language model to generate preliminary decision suggestions associated with the version number; The iterative correction module is used to decompose the preliminary decision suggestion into atomic propositions, perform three-dimensional quantitative verification on each atomic proposition on the secure isolation subgraph, and obtain three sub-scores: topological similarity, numerical matching degree, and multimodal consistency. The scores are then weighted and summed to obtain the comprehensive confidence score. When the comprehensive confidence score is lower than a preset threshold, the conflict dimension information is fed back to the large language model for iterative correction. The self-evolution module is used to output a final decision suggestion carrying the version number, obtain the user's correction operation on the final decision suggestion, extract the difference feature vector, and update the topology or edge weight of the dynamic knowledge graph based on the difference feature vector. The trusted audit module is used to concatenate and store the version number, the scores of each round of verification, and the correction records through a hash chain.

[0009] The third embodiment of the present invention provides an intelligent auxiliary decision-making device for hallucination suppression that integrates dynamic knowledge graph and spatiotemporal snapshot mechanism, including a memory and a processor. The memory stores a computer program, which can be executed by the processor to implement the intelligent auxiliary decision-making method for hallucination suppression that integrates dynamic knowledge graph and spatiotemporal snapshot mechanism as described in any of the above embodiments.

[0010] The fourth embodiment of the present invention provides a computer-readable storage medium storing a computer program, which can be executed by the processor of the device where the computer-readable storage medium is located, to implement the hallucination suppression intelligent auxiliary decision-making method that integrates dynamic knowledge graph and spatiotemporal snapshot mechanism as described in any of the above embodiments.

[0011] This invention provides an intelligent auxiliary decision-making method, device, equipment, and storage medium for hallucination suppression that integrates dynamic knowledge graphs and spatiotemporal snapshot mechanisms. By constructing a dynamic knowledge graph marked with timestamps and security level labels, upon receiving a decision request, the graph is frozen into a read-only snapshot with a unique version number based on the request time. High-security nodes are shielded according to the user's security permission level to generate a secure isolation subgraph, thus ensuring both spatiotemporal consistency and security isolation of the data view at the source. The initial decision suggestions generated based on the secure isolation subgraph are decomposed into atomic propositions. A comprehensive confidence level is obtained through three-dimensional quantitative verification of topological similarity, numerical matching degree, and multimodal consistency. Propositions below a threshold drive iterative correction of the large language model, achieving multi-dimensional quantitative identification and suppression of hallucinations. By extracting the difference feature vectors from the user's correction operations, the graph topology and edge weights are updated in reverse. The version number and the scores of each round of verification are concatenated and stored via a hash chain, ultimately achieving credible, traceable, and evolvable decision suggestions. Attached Figure Description

[0012] Figure 1This is a flowchart illustrating an intelligent auxiliary decision-making method for hallucination suppression that integrates dynamic knowledge graphs and spatiotemporal snapshot mechanisms, provided in the first embodiment of the present invention. Figure 2 This is a schematic diagram of a module of an intelligent auxiliary decision-making device for hallucination suppression that integrates dynamic knowledge graph and spatiotemporal snapshot mechanism, provided in the second embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] To better understand the technical solution of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0015] This invention discloses an intelligent decision-making aid method, device, equipment, and storage medium for hallucination suppression that integrates dynamic knowledge graphs and spatiotemporal snapshot mechanisms. It aims to address the difficulty in considering data visibility during decision-making and reasoning in dynamic, multi-level knowledge graph environments. Figure 1 Consistency and security isolation lead to the problem that the illusions output by large language models cannot be effectively checked and continuously corrected.

[0016] Please see Figure 1 The first embodiment of the present invention provides an intelligent auxiliary decision-making method for hallucination suppression that integrates dynamic knowledge graphs and spatiotemporal snapshot mechanisms. This method can be executed by a decision-making device (or system), specifically by one or more processors within the decision-making device, to at least implement the following steps: S101, Construct a multimodal dynamic knowledge graph, wherein each node and edge in the dynamic knowledge graph is marked with a timestamp and a security level label; In this embodiment, the system constructs and updates a dynamic knowledge graph in real time through a multimodal intelligence access and dynamic graph construction module. Specifically, based on the Spring Cloud Stream framework, it subscribes to a Kafka topic for multi-source intelligence, continuously receiving intelligence input from various heterogeneous data sources such as text messages, structured data interfaces, and image / video streams. For text data, it calls a named entity recognition and relation extraction model to parse and obtain entity triples. For structured data, it extracts entities and their attribute values ​​according to preset mapping rules. For image / video stream data, it calls a multimodal feature extraction interface to identify target entities and output corresponding feature embedding vectors. These feature embedding vectors are retained as multimodal attributes of nodes for subsequent multimodal consistency verification. After extraction, the module uses Neo4j JavaDriver to write the entities, relations, and attributes into the graph database in the form of triples. During the writing process, two types of metadata tags are attached to each node and edge: one is a timestamp, recording the creation time and the most recent update time of the node or edge, which serves as the basis for version freezing based on the request time; the other is a security level tag, automatically assigned according to the sensitivity level of the intelligence source, which serves as the basis for comparing with the user's security permission level to perform node blocking. The dynamic knowledge graph continuously receives the writing of new intelligence and the overwriting and updating of old data during system operation, thus forming a dynamic knowledge base with multimodal characteristics, traceable versions, and fine-grained access control capabilities.

[0017] S102, in response to the user's decision request, the dynamic knowledge graph is frozen into a read-only graph snapshot based on the request time, and a unique version number is assigned to the read-only graph snapshot; according to the user's security permission level, nodes in the read-only graph snapshot with a security level label higher than the user's security permission level and their associated edges are masked in memory to generate a security isolation subgraph; In this embodiment, the system sequentially generates a spatiotemporally consistent snapshot and constructs a secure isolation subgraph in response to user decision requests. When a decision request is received, the system first records the current system timestamp as the request time and obtains the security permission level of the user initiating the request through the Spring Security security context. Then, based on the request time, a spatiotemporal locking process is initiated: the underlying Neo4j graph database's Multi-Version Concurrency Control (MVCC) interface is called to apply a version lock to the dynamic knowledge graph's storage engine. A set of read-only pointers pointing to all data pages at that time is created based on the request time, and a unique context version number (VID) is assigned to this set. This set of read-only pointers serves as the read-only graph snapshot used in this decision-making process. During the existence of the read-only graph snapshot, the underlying storage of the dynamic knowledge graph still allows external intelligence sources to continue writing new data, but all newly written data is appended to new data pages with a higher version number. The content of the data pages pointed to by the original set of read-only pointers remains unchanged, thereby avoiding inference context jumps caused by real-time graph updates. The pointers corresponding to the VIDs are released by the resource recycling thread after the decision-making process ends. After generating the read-only graph snapshot, a virtual view object is further constructed in memory. This virtual view object encapsulates the read-only graph snapshot using a proxy pattern and provides a graph traversal interface consistent with the original graph. When the large language model initiates a traversal request through this interface in subsequent steps, the virtual view object compares the security level label of the target node with the user's security permission level before forwarding the request. If the security level label is higher than the user's security permission level, it directly returns a null value to the caller without performing any physical deletion or modification operations on the underlying data. This ensures that the large language model cannot perceive the existence of high-security-level nodes and their associated edges during the construction and verification of prompt words. The virtual view object after completing the above interception configuration is output as a secure isolation subgraph, thus obtaining a subgraph view that simultaneously possesses spatiotemporal consistency and security isolation, serving as a unified data benchmark for subsequent inference and verification.

[0018] S103, the security isolation subgraph is serialized into prompt words, and combined with the anti-illusion constraint rules, the large language model is called to generate preliminary decision suggestions associated with the version number; In this embodiment, the system generates preliminary decision suggestions based on a secure isolation subgraph using a constraint generation engine. Specifically, the constraint generation engine first calls the graph traversal interface to extract subgraph data related to the current decision request from the secure isolation subgraph, including relevant entity nodes, relation edges, and their attribute values. It then converts this subgraph data into structured text fragments using a preset serialization template. These text fragments are organized in the form of triple lists or attribute key-value pairs, facilitating parsing by the large language model. Subsequently, the constraint generation engine assembles a prompt word template based on LangChain4j. This prompt word template consists of three parts: the first part is the structured text fragment and its corresponding version number, used to clearly inform the large language model of the data source and version upon which this reasoning is based; the second part is preset anti-illusion constraint rules, which at least include constraints such as "reasoning is only based on the subgraph data identified by the version number," "creating entities, relations, or values ​​outside the subgraph is strictly prohibited," and "missing information in the subgraph must be explicitly declared and cannot be filled in by the user"; the third part is the user's original decision problem description. After assembly, the constraint generation engine submits the complete prompt words to the large language model via an API interface, and sets the temperature parameter in the model's inference parameters to a low value of 0.2 to reduce the randomness of the generated results and improve the stability of the inference. Based on the prompt words, the large language model outputs preliminary decision suggestions in text form. The preliminary decision suggestions are associated with and stored in the metadata as a version number, serving as the basis for subsequent multi-dimensional quantitative verification and hash chain evidence storage.

[0019] S104, the preliminary decision suggestion is decomposed into atomic propositions, and a three-dimensional quantitative check is performed on each atomic proposition on the security isolation subgraph to obtain three sub-scores: topological similarity, numerical matching degree and multimodal consistency. The scores are weighted and summed to obtain the comprehensive confidence score. When the comprehensive confidence score is lower than a preset threshold, the conflict dimension information is fed back to the large language model for iterative correction. In this embodiment, a lightweight language model or the Drools rule engine is first invoked to decompose the preliminary decision suggestion into a list of atomic propositions. Each atomic proposition represents an independently verifiable factual assertion, including a subject entity, relational or attribute predicate, and the corresponding object value or numerical value. Subsequently, a three-dimensional quantization check is performed on each atomic proposition sequentially on the output secure isolation subgraph. In the topological dimension check, based on the JGraphT graph computing library, the first-order neighbor set of the entity in the atomic proposition in the secure isolation subgraph and the entity set associated with the entity in the generated content context are extracted respectively, and the Jaccard similarity between the two is calculated as the topological similarity. In the numerical dimension verification, the module pre-sets a basic threshold. (e.g., 0.05), and dynamically determine the adjustment factor α based on the confidence level of the intelligence source (e.g., α=1.2 for high-confidence satellite intelligence, and α=0.8 for manually entered intelligence), and call the Apache Commons Math library to calculate the numerical values ​​in the atomic proposition. Corresponding values ​​in the security isolation subgraph The deviation, when Seasonal numerical matching degree Otherwise This achieves fuzzy tolerance to sensor noise and estimation errors. In multimodal dimension verification, the module calls a pre-trained embedding model to calculate the embedding vector of the atomic proposition text, and performs cosine similarity calculation between this vector and the pre-stored multimodal feature embedding vector in the secure isolated subgraph node to obtain multimodal consistency. Subsequently, the module determines the weights based on preset weights (e.g., ...). , , The overall confidence level is obtained by weighted summation of the scores for the three sub-items. After the three-dimensional quantization verification is completed, the module further performs logical connectivity verification on atomic propositions involving action routes or resource scheduling. That is, on the secure isolation subgraph, it calls the A* or Dijkstra graph search algorithm based on the JGraphT library to verify whether the path of the atomic proposition is topologically connected, and verifies the attributes of each edge on the path (such as traffic capacity, load limit, electromagnetic environment, etc.) to see if they meet the proposition constraints. When the path is not connected or the attribute does not meet the constraints, regardless of whether the comprehensive confidence score reaches the preset threshold (e.g., 0.75), the module directly judges the atomic proposition as a logical illusion and adds the specific blocking point information of the disconnection and the unmet attribute conditions to the conflict dimension information. When the overall confidence score is below the threshold or logical illusion exists, the module encapsulates the conflict dimension information (including the scores of specific conflict items, conflict entities, conflict values, and blocking points) into correction prompts and feeds them back to the large language model to drive its regeneration. The iterative correction process is executed for a maximum of N rounds (e.g., 3 rounds) until the overall confidence score reaches the target or the number of iterations is exhausted, thereby completing the multi-dimensional quantitative illusion identification and suppression of the preliminary decision suggestion.

[0020] S105, output the final decision suggestion carrying the version number, obtain the user's correction operation on the final decision suggestion, extract the difference feature vector, and update the topology or edge weight of the dynamic knowledge graph based on the difference feature vector. In this embodiment, after the output decision suggestion reaches a preset threshold of comprehensive confidence, the human-machine collaborative self-evolution module encapsulates the decision suggestion along with its referenced graph node ID list and version number into a structured result, and pushes it to the front-end display interface via a WebSocket channel. When the user views the decision suggestion on the front-end interface, they can click the evidence chain entry to highlight the subgraph region supporting the decision and the associated multimodal data. The front-end interface also provides an editing entry, allowing users to modify, supplement, or delete specific content in the decision suggestion. The human-machine collaborative self-evolution module captures user modification events submitted by the front-end through an event listening mechanism, and sends the original decision content generated by the large language model and the content finally confirmed by the user into the pre-trained embedding model, calculating the text embedding vectors Embedding(Text_AI) and Embedding(Text_user) respectively. The difference between the two is used as the difference feature vector Δ. After the user's decision suggestion is adopted and executed, the human-machine collaborative self-evolution module continuously monitors the feedback status of the task execution. When the feedback status is "task successful," the module performs semantic direction analysis on the difference feature vector Δ. If the semantic direction of Δ points to a missing relation type in the dynamic knowledge graph, a new triple (containing a subject entity, a new relation type, and an object entity) is automatically constructed and written into the graph through the Neo4j Java Driver. The source attribute of the new triple is marked as "user-corrected source," and its confidence weight is initialized to a value higher than the default weight of automatically extracted data (for example, if the default weight of automatically extracted data is set to 0.7, then the initial weight of the user-corrected source is set to 1.0). If Δ only corrects the attribute values ​​of existing nodes, the module directly updates the attributes of the corresponding nodes and synchronously refreshes their source tags. The updated dynamic knowledge graph serves as the latest data basis for generating a read-only graph snapshot in step S102 of the subsequent decision request, thereby enabling the system to continuously absorb implicit knowledge from the user's actual operations and achieve closed-loop adaptive evolution of the knowledge graph.

[0021] S106, The version number, the scores of each round of verification, and the correction records are stored in a chain using a hash chain.

[0022] In this embodiment, at the end of each decision-making process, the key data generated in this process is automatically aggregated and assembled into an audit record. The audit records This includes at least the version number generated in step S102, the preliminary decision suggestion output in step S103, the sub-scores and overall confidence levels of each round of three-dimensional quantitative verification in step S104, conflict dimension information generated during the correction process, and the final decision suggestion and difference feature vector Δ output in step S105. The above fields are serialized into a byte stream according to a preset format, and the SHA-256 algorithm in the Java standard encryption library MessageDigest is used to perform a hash operation on the byte stream. During the operation, the hash value of the previous audit record is used. Compared with the current audit record The byte streams are concatenated and then calculated, that is... This cryptographically binds each newly generated audit record to the historical record, ensuring that any alteration to the historical record will result in a mismatch of all subsequent hash values, thus achieving tamper-proof chain-like evidence storage. After the calculation is complete, the trusted audit module will store the hash value... and its corresponding audit records The data is written into the audit database, and a retrieval relationship is established using the version number as the primary key. In post-event review or compliance audit scenarios, the audit database provides a reverse traceability query interface based on the version number. The caller can use this interface to retrieve the audit record corresponding to any decision-making process and verify the integrity of the record step by step according to the hash chain, thereby restoring the complete evidence chain of this decision, from the generation of the subgraph snapshot, the generation of preliminary suggestions, multi-dimensional quantitative verification, iterative correction to the final suggestion output and user feedback.

[0023] Please see Figure 2 The second embodiment of the present invention provides an intelligent auxiliary decision-making device for hallucination suppression that integrates dynamic knowledge graphs and spatiotemporal snapshot mechanisms, comprising: The graph construction module 201 is used to construct a multimodal dynamic knowledge graph, wherein each node and edge in the dynamic knowledge graph is marked with a timestamp and a security level label; Security isolation module 202 is used to respond to the user's decision request, freeze the dynamic knowledge graph as a read-only graph snapshot based on the request time, and assign a unique version number to the read-only graph snapshot; according to the user's security permission level, block the nodes in the read-only graph snapshot with a security level label higher than the user's security permission level and their associated edges in memory, and generate a security isolation subgraph; The constraint generation engine 203 is used to serialize the security isolation subgraph into prompt words, and combine the anti-illusion constraint rules to call the large language model to generate preliminary decision suggestions associated with the version number; The iterative correction module 204 is used to decompose the preliminary decision suggestion into atomic propositions, perform three-dimensional quantitative verification on each atomic proposition on the secure isolation subgraph, and obtain three sub-scores: topological similarity, numerical matching degree and multimodal consistency. The scores are then weighted and summed to obtain the comprehensive confidence score. When the comprehensive confidence score is lower than a preset threshold, the conflict dimension information is fed back to the large language model for iterative correction. The self-evolution module 205 is used to output a final decision suggestion carrying the version number, obtain the user's correction operation on the final decision suggestion, extract the difference feature vector, and update the topology or edge weight of the dynamic knowledge graph based on the difference feature vector. The trusted audit module 206 is used to concatenate and store the version number, the scores of each round of verification, and the correction records through a hash chain.

[0024] The third embodiment of the present invention provides an intelligent auxiliary decision-making device for hallucination suppression that integrates dynamic knowledge graph and spatiotemporal snapshot mechanism, including a memory and a processor. The memory stores a computer program, which can be executed by the processor to implement the intelligent auxiliary decision-making method for hallucination suppression that integrates dynamic knowledge graph and spatiotemporal snapshot mechanism as described in any of the above embodiments.

[0025] The fourth embodiment of the present invention provides a computer-readable storage medium storing a computer program, which can be executed by the processor of the device where the computer-readable storage medium is located, to implement the hallucination suppression intelligent auxiliary decision-making method that integrates dynamic knowledge graph and spatiotemporal snapshot mechanism as described in any of the above embodiments.

[0026] This invention provides an intelligent auxiliary decision-making method, device, equipment, and storage medium for hallucination suppression that integrates dynamic knowledge graphs and spatiotemporal snapshot mechanisms. By constructing a dynamic knowledge graph marked with timestamps and security level labels, upon receiving a decision request, the graph is frozen into a read-only snapshot with a unique version number based on the request time. High-security nodes are shielded according to the user's security permission level to generate a secure isolation subgraph, thus ensuring both spatiotemporal consistency and security isolation of the data view at the source. The initial decision suggestions generated based on the secure isolation subgraph are decomposed into atomic propositions. A comprehensive confidence level is obtained through three-dimensional quantitative verification of topological similarity, numerical matching degree, and multimodal consistency. Propositions below a threshold drive iterative correction of the large language model, achieving multi-dimensional quantitative identification and suppression of hallucinations. By extracting the difference feature vectors from the user's correction operations, the graph topology and edge weights are updated in reverse. The version number and the scores of each round of verification are concatenated and stored via a hash chain, ultimately achieving credible, traceable, and evolvable decision suggestions.

[0027] Exemplary examples show that the computer program described in the third and fourth embodiments of the present invention can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in implementing an intelligent auxiliary decision-making device for hallucination suppression that integrates dynamic knowledge graphs and spatiotemporal snapshot mechanisms. For example, the apparatus described in the second embodiment of the present invention.

[0028] The processor referred to can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. This processor is the control center of the aforementioned hallucination suppression intelligent auxiliary decision-making method that integrates dynamic knowledge graphs and spatiotemporal snapshot mechanisms. It connects various parts of the entire hallucination suppression intelligent auxiliary decision-making method using various interfaces and lines.

[0029] The memory can be used to store the computer programs and / or modules. The processor, by running or executing the computer programs and / or modules stored in the memory, and by calling the data stored in the memory, realizes various functions of an intelligent auxiliary decision-making method for hallucination suppression that integrates dynamic knowledge graphs and spatiotemporal snapshot mechanisms. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, text conversion function, etc.), etc.; the data storage area may store data created based on the use of the mobile phone (such as audio data, text message data, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0030] If the implemented module is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0031] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0032] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A hallucination suppression intelligent auxiliary decision-making method integrating dynamic knowledge graphs and spatiotemporal snapshot mechanisms, characterized in that, include: A multimodal dynamic knowledge graph is constructed, wherein each node and edge in the dynamic knowledge graph is marked with a timestamp and a security level label; In response to a user's decision request, the dynamic knowledge graph is frozen into a read-only graph snapshot based on the request time, and a unique version number is assigned to the read-only graph snapshot; Based on the user's security permission level, nodes with security level labels higher than the user's security permission level and their associated edges in the read-only knowledge graph snapshot are masked in memory to generate a security-isolated subgraph. Specifically, freezing the dynamic knowledge graph into a read-only knowledge graph snapshot based on the request time involves: at the request time, applying a version lock based on Multi-Version Concurrency Control (MVCC) to the underlying storage engine of the dynamic knowledge graph, and creating a set of read-only pointers pointing to the data pages at that time as the read-only knowledge graph snapshot; during the existence of the read-only knowledge graph snapshot, the underlying storage of the dynamic knowledge graph still allows receiving new intelligence write requests, and the newly written data is only appended to a new data page with a higher version number. The content of the data pages pointed to by the set of read-only pointers remains static until the pointers corresponding to the version number are released after the current decision-making process ends. The security isolation subgraph is serialized into prompt words, and combined with anti-illusion constraint rules, a large language model is invoked to generate preliminary decision suggestions associated with the version number; The preliminary decision recommendations are broken down into atomic propositions. A three-dimensional quantitative check is performed on each atomic proposition on the secure isolation subgraph to obtain three sub-scores: topological similarity, numerical matching degree, and multimodal consistency. These scores are then weighted and summed to obtain the overall confidence score. When the overall confidence score is lower than a preset threshold, conflict dimension information is fed back to the large language model for iterative correction. The numerical matching degree is calculated as follows: a basic threshold is set. And dynamic adjustment factors determined based on intelligence source confidence. When the numerical value in an atomic proposition The corresponding values ​​in the security isolation subgraph satisfy When, let the numerical matching degree Otherwise, let the numerical matching degree The numerical matching degree It is used as one of the input components of the overall confidence score in the weighted summation; The system outputs a final decision suggestion carrying the version number, obtains the user's correction operation on the final decision suggestion, extracts the difference feature vector, and updates the topology or edge weights of the dynamic knowledge graph based on the difference feature vector. Specifically, it calculates the text embedding vectors of the content generated by the large language model and the content confirmed by the user through a pre-trained embedding model, and the difference between the two is the difference feature vector. When the task execution feedback is successful and the semantic direction of the difference feature vector points to the missing relation type in the dynamic knowledge graph, a new triple is automatically instantiated and written into the dynamic knowledge graph. The source attribute of the new triple is marked as the user's correction source, and its confidence weight is initialized to a value higher than the default weight of the automatically extracted data. The updated dynamic knowledge graph serves as the basis for generating the read-only graph snapshot in subsequent decision requests. The version number, the scores of each round of verification, and the correction records are stored together using a hash chain.

2. The hallucination suppression intelligent auxiliary decision-making method integrating dynamic knowledge graph and spatiotemporal snapshot mechanism according to claim 1, characterized in that, Based on the user's security permission level, nodes with security level labels higher than the user's security permission level in the read-only graph snapshot and their associated edges are masked in memory to generate a secure isolation subgraph, specifically: A virtual view object is constructed in memory, which encapsulates the read-only graph snapshot and intercepts all graph traversal requests for the read-only graph snapshot; When the graph traversal request accesses a node with a security level label higher than the user's security permission level, the virtual view object returns a null value and does not perform a physical deletion operation on the underlying data; The virtual view object after the interception configuration is completed is output as the security isolation subgraph, so that the large language model cannot detect the existence of high-density nodes during the construction and verification of prompt words.

3. The hallucination suppression intelligent auxiliary decision-making method integrating dynamic knowledge graph and spatiotemporal snapshot mechanism according to claim 1, characterized in that, After performing a three-dimensional quantization check on each atomic proposition on the secure isolation subgraph, logical connectivity verification is also included: When the atomic proposition involves action routes or resource scheduling, a graph search algorithm is run on the safe isolation subgraph to verify whether the path described by the atomic proposition is topologically connected and whether the attributes of each edge on the path satisfy the proposition constraints. When the path is not connected or the attribute does not meet the constraints, regardless of whether the comprehensive confidence score reaches the preset threshold, the atomic proposition is determined to be a logical illusion, and the specific blocking point information is added to the conflict dimension information and fed back to the large language model for iterative correction.

4. The hallucination suppression intelligent auxiliary decision-making method integrating dynamic knowledge graph and spatiotemporal snapshot mechanism according to claim 1, characterized in that, The hash chain concatenation evidence storage specifically refers to: An audit record is generated for each decision-making process, and the audit record includes the version number, the preliminary decision recommendation, the scores of each round of verification, the final decision recommendation, and the difference feature vector. The hash value is calculated sequentially using the SHA-256 algorithm. Its expression is: ,in, The hash value of the previous audit record is used to... and the corresponding audit records Store in the audit database; The audit database provides a reverse traceability query interface based on the version number, which is used to restore the complete evidence chain of any decision-making process.

5. A hallucination suppression intelligent auxiliary decision-making device integrating dynamic knowledge graph and spatiotemporal snapshot mechanism, characterized in that, include: The graph construction module is used to construct a multimodal dynamic knowledge graph, in which each node and edge is marked with a timestamp and a security level label; The security isolation module is used to freeze the dynamic knowledge graph into a read-only graph snapshot based on the time of the request in response to the user's decision request, and assign a unique version number to the read-only graph snapshot; Based on the user's security permission level, nodes with security level labels higher than the user's security permission level and their associated edges in the read-only knowledge graph snapshot are masked in memory to generate a security-isolated subgraph. Specifically, freezing the dynamic knowledge graph into a read-only knowledge graph snapshot based on the request time involves: at the request time, applying a version lock based on Multi-Version Concurrency Control (MVCC) to the underlying storage engine of the dynamic knowledge graph, and creating a set of read-only pointers pointing to the data pages at that time as the read-only knowledge graph snapshot; during the existence of the read-only knowledge graph snapshot, the underlying storage of the dynamic knowledge graph still allows receiving new intelligence write requests, and the newly written data is only appended to a new data page with a higher version number. The content of the data pages pointed to by the set of read-only pointers remains static until the pointers corresponding to the version number are released after the current decision-making process ends. A constraint generation engine is used to serialize the security isolation subgraph into prompt words, and combine anti-illusion constraint rules to call a large language model to generate preliminary decision suggestions associated with the version number; The iterative correction module decomposes the preliminary decision recommendations into atomic propositions, performs a three-dimensional quantitative check on each atomic proposition on the secure isolation subgraph, and obtains three sub-scores: topological similarity, numerical matching degree, and multimodal consistency. These scores are then weighted and summed to obtain a comprehensive confidence score. When the comprehensive confidence score is lower than a preset threshold, conflict dimension information is fed back to the large language model for iterative correction. The numerical matching degree is calculated as follows: a basic threshold is set. And dynamic adjustment factors determined based on intelligence source confidence. When the numerical value in an atomic proposition The corresponding values ​​in the security isolation subgraph satisfy When, let the numerical matching degree Otherwise, let the numerical matching degree The numerical matching degree It is used as one of the input components of the overall confidence score in the weighted summation; The self-evolution module is used to output a final decision suggestion carrying the version number, obtain the user's correction operation on the final decision suggestion, extract the difference feature vector, and update the topology or edge weight of the dynamic knowledge graph based on the difference feature vector. Specifically, it calculates the text embedding vectors of the content generated by the large language model and the content confirmed by the user through a pre-trained embedding model, and the difference between the two is the difference feature vector. When the task execution feedback is successful and the semantic direction of the difference feature vector points to the missing relation type in the dynamic knowledge graph, a new triple is automatically instantiated and written into the dynamic knowledge graph. The source attribute of the new triple is marked as the user correction source, and its confidence weight is initialized to a value higher than the default weight of the automatically extracted data. The updated dynamic knowledge graph serves as the basis for generating the read-only graph snapshot in subsequent decision requests. The trusted audit module is used to concatenate and store the version number, the scores of each round of verification, and the correction records through a hash chain.

6. A hallucination suppression intelligent auxiliary decision-making device integrating dynamic knowledge graph and spatiotemporal snapshot mechanism, characterized in that, The system includes a memory and a processor. The memory stores a computer program that can be executed by the processor to implement an intelligent auxiliary decision-making method for hallucination suppression that integrates dynamic knowledge graphs and spatiotemporal snapshot mechanisms as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The device contains a computer program that can be executed by a processor of the device in which the computer-readable storage medium is located, to implement the hallucination suppression intelligent auxiliary decision-making method that integrates dynamic knowledge graph and spatiotemporal snapshot mechanism as described in any one of claims 1 to 4.

Images