Method, apparatus and electronic device for updating world model of service network

Abstract

This invention provides a method, apparatus, and electronic device for updating the world model of a business network. The method includes: acquiring newly added business knowledge vectors and the current model; calculating the knowledge reconstruction error using a variational autoencoder; verifying the newly added knowledge vectors when the error is less than a threshold; and finally fusing the newly added knowledge vectors with the current model based on the verification results to obtain an updated world model. The method provided by this invention uses a variational autoencoder to perform distribution sensing on the newly added business knowledge vectors to calculate the reconstruction error, and, under the premise of controllable error, combines a logical verification mechanism to obtain the verification result. Finally, based on this result, a conflict-free fusion with the current world model is achieved. This not only adaptively quantifies the novelty and uncertainty of input data in open environments but also effectively solves the illusion problem of generative knowledge in large language models, ensuring that the business network can achieve efficient, accurate, and logically coherent continuous evolution when dealing with multi-source event streams.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, and electronic device for updating a world model of a business network. Background Technology

[0002] In open-world environments, complex business networks, such as intelligent urban transportation and global supply chain monitoring, continuously receive massive amounts of multi-source, heterogeneous event data streams. To maintain real-time evolution and provide accurate intelligent decision support for complex events, it is objectively necessary to dynamically and continuously build and update the underlying world model or knowledge graph to ensure that the model can accurately and timely reflect the latest operational rules and business status of the real world. Existing technical solutions typically adopt a static, one-time full-scale construction mode, or, when handling incremental updates, directly use generative artificial intelligence technologies such as large language models to extract entity states and business rules from the newly input event stream. Then, by setting basic overwrite rules or simple logical judgment conditions, the extracted incremental data is directly written into the existing world model or knowledge graph to achieve information iteration of the underlying model of the business network.

[0003] However, existing technology update mechanisms face the problem of noise pollution caused by the inability to quantify the uncertainty of data distribution, and the defects of deep logic extraction that are prone to producing illusions and false causality, resulting in logical fragmentation and decision distortion of the entire world model. Summary of the Invention

[0004] This invention provides a method, apparatus, and electronic device for updating the world model of a business network, in order to solve the problems of noise pollution caused by the inability to quantify the uncertainty of data distribution in the existing update mechanism, and the defects of illusion and false causality that are easily generated by deep logic extraction.

[0005] This invention provides a method for updating the world model of a business network, comprising: Obtain new business knowledge vectors and the current world model of the business network; The newly added business knowledge vector is input into the variational autoencoder to obtain the knowledge reconstruction error calculated by the variational autoencoder; the variational autoencoder is obtained based on training the initial encoder and decoder; If the knowledge reconstruction error is less than a preset error threshold, the newly added business knowledge vector is verified to obtain the verification result. Based on the verification results, the newly added business knowledge vector and the current world model are fused to obtain the updated world model of the business network.

[0006] According to the present invention, a method for updating a world model of a business network, wherein verifying the newly added business knowledge vector to obtain a verification result includes: Based on the newly added business knowledge vector, generate causal rules to be verified; The semantic consistency ratio of the causal rule to be verified is obtained by sampling the rule multiple times. If the semantic consistency ratio is greater than or equal to a preset consistency threshold, the causal rule to be verified is subjected to a counterfactual simulation test in a preset sandbox environment to obtain the causal strength. The verification result is determined based on the causal strength.

[0007] According to a method for updating a world model of a business network provided by the present invention, the step of performing counterfactual simulation tests on the causal rule to be verified in a preset sandbox environment to obtain the causal strength includes: Construct a first factual scenario that meets the conditions of the causal rule to be verified, and a second counterfactual scenario that modifies the conditions of the causal rule to be verified; In the sandbox environment, the probability of the first outcome in the first factual scenario and the probability of the second outcome in the second counterfactual scenario are respectively derived. The probability difference between the probability of the first outcome and the probability of the second outcome is calculated to obtain the causal strength.

[0008] According to a method for updating a world model of a business network provided by the present invention, the step of fusing the newly added business knowledge vector and the current world model based on the verification result to obtain an updated world model of the business network includes: If the verification result indicates that the verification is successful, information propagation between nodes is performed between the newly added subgraph corresponding to the newly added business knowledge vector and the current world model to obtain a node representation; Based on the node representation, the fusion score between the newly added subgraph and the current world model is calculated; Based on the fusion score, a target fusion strategy is determined from a preset set of fusion strategies. The target fusion strategy is then used to update the new business knowledge vector into the current world model, resulting in the updated world model.

[0009] According to a method for updating a world model of a business network provided by the present invention, the step of calculating the fusion score between the newly added subgraph and the current world model based on the node representation includes: Based on the node representation, the semantic similarity and structural matching degree between the newly added subgraph and the current world model are calculated respectively; Evaluate the penalty value for logical conflicts between the newly added subgraph and the current world model; The fusion score is calculated by weighting and summing the semantic similarity, structural matching degree, and logical conflict penalty value.

[0010] According to the present invention, a method for updating a world model of a business network is provided, wherein the current world model includes an ontology layer, a state layer, and a causal layer; The newly added business knowledge vectors include newly added entity vectors, newly added state vectors, and newly added rule vectors; The step of using the target fusion strategy to update the new business knowledge vector into the current world model to obtain the updated world model of the business network includes: Using the target fusion strategy, the newly added entity vector is updated to the ontology layer, the newly added state vector is updated to the state layer, and the newly added rule vector is updated to the causal layer to obtain the updated world model.

[0011] The present invention also provides an apparatus for updating a world model of a business network, comprising: The acquisition unit acquires new business knowledge vectors and the current world model from the business network. The gating unit inputs the newly added business knowledge vector into the variational autoencoder to obtain the knowledge reconstruction error calculated by the variational autoencoder; the variational autoencoder is obtained based on training the initial encoder-decoder; The verification unit verifies the newly added business knowledge vector when the knowledge reconstruction error is less than a preset error threshold, and obtains the verification result. The fusion unit, based on the verification results, fuses the newly added business knowledge vector and the current world model to obtain the updated world model of the business network.

[0012] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a method for updating a world model of a business network as described above.

[0013] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for updating the world model of a business network as described in any of the above.

[0014] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements a method for updating the world model of a business network as described above.

[0015] The present invention provides a method, apparatus, and electronic device for updating the world model of a business network. It uses a variational autoencoder to perform distribution perception on newly added business knowledge vectors to calculate reconstruction errors. Under the premise of controllable errors, it combines a logical verification mechanism to obtain verification results. Finally, based on the results, it achieves conflict-free fusion with the current world model. This not only adaptively quantifies the novelty and uncertainty of input data in open environments, but also effectively solves the illusion problem of generative knowledge in large language models. It ensures that the business network can achieve efficient, accurate, and logically coherent continuous evolution when dealing with multi-source event streams. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is one of the flowcharts illustrating the method for updating the world model of a business network provided by the present invention; Figure 2 This is a flowchart illustrating the dual verification method for rules provided by the present invention; Figure 3 This is a flowchart illustrating the knowledge fusion method provided by the present invention; Figure 4 This is a schematic diagram of the architecture of the world model provided by this invention; Figure 5 This is the second flowchart illustrating the method for updating the world model of a business network provided by the present invention; Figure 6 This is a flowchart illustrating the gating mechanism for newly added business knowledge provided by the present invention. Figure 7 This is a schematic diagram of the structure of the world model update device for the business network provided by the present invention; Figure 8 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0019] It should be noted that all actions involving the acquisition of signals, information, or data in this application are carried out in compliance with the relevant data protection laws and policies of the country where the application is located, and with the authorization granted by the owner of the relevant device.

[0020] To address the aforementioned issues, this invention provides a method for updating the world model of a business network, enabling adaptive filtering, verification, and conflict-free fusion of continuously incoming multi-source data in an open-world environment, thereby improving the accuracy and robustness of the world model update. Figure 1 This is one of the flowcharts illustrating the method for updating the world model of a business network provided by the present invention, such as... Figure 1 As shown, the method includes: Step 110: Obtain the new business knowledge vectors and the current world model of the business network.

[0021] Here, "business network" refers to a complex network system that maps real-world business entities, states, and their logical relationships, such as an intelligent urban traffic command network, a global electronics manufacturing supply chain monitoring network, or a digital twin system. Here, "new business knowledge vector" refers to a high-dimensional vector representation obtained by extracting and mapping structured data from multi-source heterogeneous event data streams captured from an open-world environment using a unified semantic mapping method. New business knowledge refers to incremental information continuously input and captured in the open-world environment, which can encompass structured, semi-structured, or unstructured data extracted from multi-source event data streams, including but not limited to entities, relationships, timestamped state attributes, and potential causal rules. For example, in the case of an intelligent urban traffic command network, the new business knowledge vector here could be "ice formation on elevated roads in the city, causing multiple vehicles to skid."

[0022] In addition, the current world model here refers to the dynamic knowledge representation architecture that has been built and maintained by the business network before the current update cycle. It usually includes separate hierarchical structures such as ontology layer, state layer and causal layer.

[0023] Specifically, when acquiring new business knowledge vectors, raw information is first captured from multimodal event data streams. For example, in the extreme weather traffic assessment scenario in the public safety field, raw information includes text data from social media, such as icy roads and multiple vehicles skidding on urban overpasses, and numerical data from IoT sensors, such as temperatures of -3 degrees Celsius and a 90% probability of precipitation. In the supply chain risk monitoring scenario, raw information includes breaking news reports, such as a fire at a chip supplier's factory, and interface data from the enterprise resource planning system, such as current inventory being sufficient and delivery plans being normal.

[0024] Next, a large language model can be used to identify entity sets, relation sets, timestamped state attribute sets, and potential causal rule sets from the aforementioned text or structured data. Subsequently, using pre-trained embedding models, such as BERT (Bidirectional Encoder Representations from Transformers) or TransE (Translating Embeddings), the above four symbolic elements are uniformly mapped into numerical representations in a high-dimensional vector space, thereby obtaining new business knowledge vectors.

[0025] Step 120: Input the newly added business knowledge vector into the variational autoencoder to obtain the knowledge reconstruction error calculated by the variational autoencoder.

[0026] The variational autoencoder is obtained by training an initial encoder-decoder.

[0027] Here, variational autoencoder refers to a deep generative model used to learn the potential data distribution of an existing knowledge base. Knowledge reconstruction error refers to the distance metric between a newly added business knowledge vector and its output vector obtained after reconstruction by variational autoencoder; this metric reflects the novelty or uncertainty of the input data compared to existing distribution patterns.

[0028] Specifically, the acquired new business knowledge vector is input into a variational autoencoder. In the variational autoencoder, the encoder maps the input vector to the distribution parameters of the latent space, which are then sampled and reconstructed by the decoder to obtain the reconstructed vector. Then, the Euclidean distance between the original input new business knowledge vector and the reconstructed vector is calculated as the knowledge reconstruction error. For example, if the input is common low-temperature or low-speed traffic data, the reconstruction error is usually small; if the input is a long-tail, low-frequency event in the supply chain, such as a fire, the reconstruction error will be in a medium to high range. Here, the knowledge reconstruction error can be calculated based on the following formula, as shown below: ; In the formula, Indicates the error in knowledge reconstruction; This represents a newly added business knowledge vector; This represents the reconstructed vector output by the decoder of the variational autoencoder for the newly added business knowledge vector.

[0029] Here, the variational autoencoder can be an initial encoder-decoder whose parameters are to be adjusted, which is then pre-trained on the existing global knowledge graph of the initial autoencoder to obtain a variational autoencoder that has learned the data distribution of the current world model.

[0030] It should be noted that newly added business knowledge vectors often contain a large amount of noise or out-of-distribution data. Existing methods lack an effective pre-gating mechanism and cannot quantify the novelty and uncertainty of the input at the data distribution level before logical fusion, resulting in noise contamination of the graph. The method provided in this embodiment of the invention constructs an effective data pre-gating mechanism by introducing a variational autoencoder to calculate the reconstruction error. This mechanism can quantify the uncertainty of the input data at the distribution level, thereby identifying out-of-distribution abnormal noise or novel knowledge.

[0031] Step 130: If the knowledge reconstruction error is less than a preset error threshold, the newly added business knowledge vector is verified to obtain the verification result.

[0032] Here, the preset error threshold refers to a pre-set critical value used to distinguish between known distributions and unknown noise. Here, the verification result refers to the validity conclusion drawn after verifying the authenticity of the logical connections carried by the newly added business knowledge vectors, especially the causal rules generated by the large language model.

[0033] Specifically, firstly, the knowledge reconstruction error can be compared with a preset error threshold. When the comparison determines that the knowledge reconstruction error is less than the threshold, it means that the knowledge vector belongs to a known pattern or belongs to the long-tail category that needs to be fused, at which point the verification process is triggered. The verification process here aims to eliminate the illusion problem of generative knowledge. For example, causal rules generated by a large language model can be sampled multiple times independently to verify their semantic consistency, or counterfactual scenarios can be simulated in a preset sandbox environment. In the traffic assessment scenario, the probability of road icing can be significantly reduced when the temperature is five degrees Celsius, thereby calculating the strength of the causal effect.

[0034] Finally, based on whether the consistency ratio or causal strength meets the standard, a verification result is generated indicating whether the knowledge has real validity.

[0035] It should be noted that by performing logical verification under specific reconstruction error conditions, it is possible to ensure that the rules written into the world model have real causal effect, effectively suppressing the knowledge illusion pollution caused by the probability of generative models.

[0036] Step 140: Based on the verification results, the newly added business knowledge vector and the current world model are fused to obtain the updated world model of the business network.

[0037] Here, updating the world model refers to merging the newly added business knowledge, after verification, into the existing architecture through a structured strategy to form a world model version with the latest logic and state.

[0038] Specifically, based on the knowledge validity indicated by the verification results, a compatibility score between new and old knowledge is calculated at the graph topology level. For example, a graph neural network is used to extract global contextual information from the current world model, and the semantic similarity, structural matching, and logical conflicts between the newly added subgraph and the existing graph are evaluated. Based on the score, an adaptive decision is made between a strong fusion strategy that directly updates the state and a weak fusion strategy that stores the subgraph in a candidate pool for observation.

[0039] Finally, based on the fusion decision, the ontology layer, state layer, or causal layer are atomically updated to generate corresponding incremental logs and obtain the updated world model.

[0040] It should be noted that, through structured fusion based on verification results, the world model can be dynamically evolved in an automated and interpretable manner while maintaining the logical consistency of the knowledge system, thus avoiding structural conflicts that may arise during the fusion of new and old knowledge.

[0041] The method provided in this invention uses a variational autoencoder to perform distribution perception on newly added business knowledge vectors to calculate reconstruction errors. Under the premise of controllable errors, it combines a logical verification mechanism to obtain verification results. Finally, based on the results, it achieves conflict-free fusion with the current world model. This method can not only adaptively quantify the novelty and uncertainty of input data in open environments, but also effectively solve the illusion problem of generative knowledge in large language models. It ensures that the business network can achieve efficient, accurate and logically coherent continuous evolution when dealing with multi-source event streams.

[0042] It should be noted that rules extracted or generated using large language models, such as "A causes B," often contain probabilistic illusions or correlation fallacies. Existing technologies lack effective verification methods and struggle to distinguish between strong causality and weak correlation. To address this issue, based on any of the above embodiments, in step 130, the newly added business knowledge vector is verified to obtain verification results, including: Based on the newly added business knowledge vector, generate causal rules to be verified; The semantic consistency ratio of the causal rule to be verified is obtained by sampling the rule multiple times. If the semantic consistency ratio is greater than or equal to a preset consistency threshold, the causal rule to be verified is subjected to a counterfactual simulation test in a preset sandbox environment to obtain the causal strength. The verification result is determined based on the causal strength.

[0043] Here, the causal rules to be verified refer to the candidate rules that represent the logical driving relationship between things, which are identified and reconstructed from the newly added business knowledge vector. They are usually expressed in the conditional logic form of "if A then B".

[0044] Specifically, firstly, a large language model can be used to perform reverse analysis or correlation mining on the newly added business knowledge vectors to generate causal rules to be verified. In the urban traffic assessment scenario, if the newly added business knowledge vectors contain features such as temperatures below zero degrees Celsius, high probability of precipitation, and decreased vehicle speed, then a causal rule to be verified, such as "If the temperature is below zero degrees Celsius and there is precipitation, then road icing will lead to a decrease in traffic capacity," can be generated based on the large language model. In the supply chain scenario, if the vector features include factory fires and production capacity forecasts, the generated causal rule to be verified might be "Factory fires lead to supply disruptions."

[0045] It should be noted that, based on the newly added business knowledge vector, causal rules to be verified are generated, making the implicit correlations carried by the high-dimensional vector explicit into causal rules to be verified with logical structure, providing an analysis object for subsequent deterministic logic verification.

[0046] Next, the causal rule to be verified is sampled multiple times to obtain the semantic consistency ratio of the causal rule to be verified. Here, the semantic consistency ratio refers to the frequency with which the semantics of the rules generated multiple times by the large language model remain the same or similar under the same input conditions, and is used to measure the stability of the output of the generative model.

[0047] In detail, the same rule generation task is sampled multiple times independently. For example, the number of samplings is set to five, and the generation temperature parameter is adjusted to maintain a certain degree of randomness. If, in four out of the five samplings, the generated rules have the same semantic meaning, such as all pointing to freezing due to temperature and precipitation, then the semantic consistency ratio of the causal rule to be verified is calculated to be 80%.

[0048] It should be noted that by calculating the semantic consistency ratio, the self-consistency voting mechanism of the large language model is used to initially filter out accidental errors caused by probabilistic output, effectively suppressing the rudimentary knowledge illusion generated by the generative model.

[0049] Furthermore, when the semantic consistency ratio is greater than or equal to a preset consistency threshold, counterfactual simulation tests are performed on the causal rules to be verified in a preset sandbox environment to obtain the causal strength. Here, the preset consistency threshold is a critical indicator used to determine whether a rule possesses at least basic stability. The preset sandbox environment refers to a virtual simulation space used for secure simulation logic deduction, detached from the business network production environment. Additionally, the counterfactual simulation test here refers to a test method that determines whether a true causal relationship exists between A and B by changing a certain condition in known facts, such as constructing hypothetical scenarios of "if A does not occur" or "if A becomes A'", and observing the degree of change in result B. Furthermore, the causal strength here is a numerical indicator that quantifies the magnitude of this causal driving force.

[0050] In detail, when the semantic consistency ratio meets the standard, deep logic verification is initiated. Parallel scenarios are constructed in a sandbox environment. For example, the factual scenario is "the temperature is -3 degrees Celsius, and the probability of road icing is 95%"; the counterfactual scenario forcibly intervenes in the temperature variable, simulating "if the temperature rises to 5 degrees Celsius, with other conditions remaining unchanged," and deduction shows that the probability of road icing drops to 5%. The probability difference between the two scenarios is calculated, and the resulting difference is the causal strength.

[0051] It should be noted that by performing counterfactual simulation tests in a sandbox environment, it is possible to distinguish between strong causality and weak correlation at the level of logical mechanism, ensuring that the rules to be merged are not only statistically stable, but also have rigorous causal driving power in objective logic.

[0052] Finally, the verification result is determined based on the causal strength. Specifically, the calculated causal strength is compared with a preset causal threshold. If the causal strength exceeds the threshold, the rule is determined to be a strong causal rule, the verification result is passed, and the newly added business knowledge vector is allowed to enter the subsequent fusion stage. If the causal strength is below the threshold, the rule is determined to be only a weak correlation or a pseudo-causal relationship, the verification result is failed, and the newly added business knowledge vector will be downgraded or discarded directly.

[0053] The method provided in this invention generates causal rules to be verified and performs preliminary screening by combining the semantic consistency ratio of multiple samples. Then, it uses counterfactual simulation tests in a sandbox environment to quantify the causal strength to determine the final verification result. This forms a dual verification mechanism that goes from the surface to the core, from statistical stability to logical determinism. It effectively solves the illusion problem when large language models generate business rules and ensures that every rule written into the business network world model has real business guidance value.

[0054] Based on any of the above embodiments, counterfactual simulation tests are performed on the causal rule to be verified in a preset sandbox environment to obtain the causal strength, including: Construct a first factual scenario that meets the conditions of the causal rule to be verified, and a second counterfactual scenario that modifies the conditions of the causal rule to be verified; In the sandbox environment, the probability of the first outcome in the first factual scenario and the probability of the second outcome in the second counterfactual scenario are respectively derived. The probability difference between the probability of the first outcome and the probability of the second outcome is calculated to obtain the causal strength.

[0055] Specifically, firstly, a first-fact scenario conforming to the conditions of the causal rule to be verified is constructed, and a second counterfactual scenario with modified conditions is also constructed. The first-fact scenario refers to a benchmark testing environment constructed based on currently observed input data, where the condition variables are in their true states. The second counterfactual scenario, on the other hand, refers to a parallel comparison environment constructed by forcibly intervening in the core condition variables of the first-fact scenario, such as inverting, modifying, or setting them to mutually exclusive states, while keeping other background variables unchanged.

[0056] Specifically, for causal rules to be verified, such as "if condition X, then result Y", two parallel simulation scenarios are constructed in a sandbox environment. In the extreme weather traffic assessment scenario in the field of public safety, for the rule "if the temperature is below zero degrees Celsius and there is precipitation, then the road will be icy", the first factual scenario is constructed based on the actual input meteorological conditions, such as a temperature of -3 degrees Celsius and a 90% probability of precipitation. Then, counterfactual intervention is applied to the numerical conditional variable of temperature, modifying it to a value significantly different from the factual value, such as setting the temperature to five degrees Celsius, while keeping other conditions such as precipitation unchanged, thereby constructing the second counterfactual scenario.

[0057] For multiple enumeration or Boolean variables, such as the rule "If a fire occurs in the factory, the supply will be interrupted" in a supply chain scenario, counterfactual intervention is to modify it into a mutually exclusive enumeration state, that is, to construct a second counterfactual scenario of "no fire".

[0058] Then, in a sandbox environment, the probability of the first outcome in the first factual scenario and the probability of the second outcome in the second counterfactual scenario are derived. The probability of the first outcome refers to the mathematical probability of the target event or outcome state occurring in the system simulation under real observation conditions. The probability of the second outcome refers to the corresponding probability of the same target event or outcome state occurring after human intervention and modification of the conditions.

[0059] In detail, the model was used to perform logical deduction calculations on the two parallel scenarios mentioned above in an isolated sandbox environment. Taking the traffic assessment scenario as an example, in the first factual scenario, i.e. the temperature is -3 degrees Celsius, the probability of the first outcome, "road icing," is 95%; while in the second counterfactual scenario, i.e. the temperature is forcibly intervened at 5 degrees Celsius, the probability of the second outcome, "road icing," drops sharply to 5%.

[0060] For example, in the context of supply chain monitoring, in the first factual scenario of "factory fire", the probability of "supply interruption" is deduced to be extremely high; while in the second counterfactual scenario of "no fire", the probability of "supply interruption" is deduced to be extremely low, which corresponds to an extremely high probability of normal supply.

[0061] It should be noted that by extrapolating the probability of the outcome in two parallel scenarios, the abstract causal logical relationship is transformed into calculable and comparable quantitative numerical characteristics, providing reliable data support for the subsequent accurate measurement of the strength of causal effects.

[0062] Furthermore, the probability difference between the probability of the first outcome and the probability of the second outcome is calculated to obtain the causal strength. Specifically, the probability difference between the calculated probability of the first outcome and the probability of the second outcome represents the causal strength. For example, in the scenario of icy roads mentioned above, subtracting the 5% probability of the second outcome from the 95% probability of the first outcome yields a probability difference of 0.9, which is the causal strength. It can be understood that this causal strength objectively reflects the net impact of a change in a specific conditional variable on the final result; a larger value indicates a stronger causal relationship between the condition and the outcome.

[0063] It should be noted that by calculating the probability difference to obtain the causal strength, the causal power of the rule can be accurately quantified, effectively filtering out pseudo-causal rules that only have data co-occurrence but no actual logical driving relationship, thus ensuring objective and rigorous knowledge evaluation.

[0064] The method provided in this invention constructs parallel scenarios of fact and counterfactual in a sandbox environment and calculates the difference in the probability of occurrence of the results before and after intervention in the core conditions to obtain a quantified causal strength. It introduces a causal inference mechanism of counterfactual simulation, transforms the generation probability of the large language model into a true causal power indicator that has been rigorously deduced, and completely removes false correlations in complex data, thereby significantly improving the absolute credibility and decision-making guidance value of the causal rule base in the dynamic world model.

[0065] In one embodiment, Figure 2 This is a flowchart illustrating the dual verification method for rules provided by this invention, as shown below. Figure 2 As shown, this method demonstrates a dual mechanism for rigorously verifying the causal rules generated by large language models to eliminate relevance fallacies and ensure that the rules have genuine causal validity.

[0066] First, the causal rule R generated by the LLM (Large Language Model) is input, which is the causal rule to be verified. Then, the first verification stage, the self-consistency voting stage, is entered. In this stage, the same rule generation task is sampled N times independently, and a relatively high temperature parameter (e.g., T>0.7) is usually set to introduce a certain degree of randomness to observe the stability of the model output. Next, the semantic consistency ratio of these N sampling results is calculated, and it is determined whether the consistency ratio is greater than a preset consistency threshold T_cons. If the consistency ratio fails to reach the threshold, it is judged as "no," indicating that the generated rule has poor stability and may contain serious illusions, and it will be directly discarded. If the consistency ratio is greater than the threshold, it is judged as "yes," and the causal rule R passes the first verification and enters the second verification stage, namely, counterfactual simulation in a sandbox.

[0067] In the second verification phase, for the causal rule to be verified, "If X then Y", two parallel scenarios are constructed simultaneously in the sandbox environment. On one hand, a factual scenario is constructed, setting Set X = True, and the probability of outcome Y occurring under this factual scenario is deduced, i.e., the probability of outcome Y is deduced as P(Y|do(X=True)). On the other hand, the conditional variable is forcibly intervened, a counterfactual scenario is constructed, Intervene do(X = False) is executed, and the corresponding probability of outcome Y occurring under this counterfactual scenario is deduced, i.e., the probability of outcome Y is deduced as P(Y|do(X=False)).

[0068] After obtaining the probabilities of the two scenarios, the causal effect strength ΔP is calculated, which is equal to the probability P(fact) in the factual scenario minus the probability P(counter) in the counterfactual scenario. Then, the calculated causal effect strength is... With the preset causal threshold Perform a comparison. If ΔP is less than the causality threshold... If the judgment is negative, then the causal rule is deemed invalid and downgraded to a weak association; if... Greater than or equal to the causal threshold If the condition is met, the rule is considered to have strong causality, and the validated valid rule R is output. This dual-logic verification module transforms the generation probability of the large model into rigorously verified causal strength, significantly suppressing knowledge illusions and ensuring that incremental rules possess genuine causal efficacy.

[0069] Based on any of the above embodiments, step 140 includes: If the verification result indicates that the verification is successful, information propagation between nodes is performed between the newly added subgraph corresponding to the newly added business knowledge vector and the current world model to obtain a node representation; Based on the node representation, the fusion score between the newly added subgraph and the current world model is calculated; Based on the fusion score, a target fusion strategy is determined from a preset set of fusion strategies. The target fusion strategy is then used to update the new business knowledge vector into the current world model, resulting in the updated world model.

[0070] Here, the newly added subgraph refers to the local network topology formed by the structured organization of newly added business knowledge vectors, which includes newly added entity nodes, attribute nodes, and their associated edges. Here, node representation refers to a deep, high-dimensional feature vector that integrates local features and global network context information.

[0071] In detail, when the verification results indicate that the newly added knowledge possesses genuine logical or causal validity, it is allowed to enter the structured fusion stage. First, the newly added subgraph corresponding to the new business knowledge vector is temporarily integrated into the global knowledge graph of the current world model. Then, a K-layer graph neural network is used to propagate and aggregate information between the nodes of the newly added subgraph and the existing nodes of the current world model. This ensures that each node not only contains its own original attribute features but also absorbs the features of its neighbors and contextual information in the global topology, thus outputting a node representation containing global contextual information. In this way, isolated new knowledge is mapped to the global network topology and interacts with information, enabling a global understanding of the structural position of new knowledge within the existing knowledge system.

[0072] Next, based on the node representation, the fusion score between the newly added subgraph and the current world model is calculated. Here, the fusion score is a comprehensive quantitative indicator used to fully evaluate the compatibility and fit between the new and old knowledge at the semantic, structural, and logical rule levels.

[0073] In detail, based on the node representations output by the graph neural network and combined with a preset scoring function, the matching degree between the node representations of the newly added subgraph and the existing node representations associated with it in the current world model can be calculated to obtain a fusion score that reflects the cost and compatibility of integrating new and old knowledge. For example, in a traffic assessment scenario, if the detection of reduced vehicle speed and road icing is highly consistent in spatial topology and logic and does not conflict with other states, a higher fusion score can be calculated.

[0074] Furthermore, based on the fusion score, a target fusion strategy is determined from a pre-defined set of fusion strategies. Using this target fusion strategy, the newly added business knowledge vectors are updated into the current world model, resulting in an updated world model. The pre-defined set of fusion strategies typically includes different response mechanisms such as strong fusion, weak fusion, and rejection. Here, the target fusion strategy is the final execution plan selected based on the specific range the fusion score falls into.

[0075] In detail, the target fusion strategy is determined based on the interval in which the calculated fusion score falls. For example, if the fusion score is greater than or equal to the strong fusion threshold, strong fusion is determined as the target fusion strategy, allowing direct overwriting or updating of the state in the current world model. If the fusion score is in the weak fusion interval, such as greater than the weak fusion threshold but less than the strong fusion threshold, which usually occurs when there is some conflict between information sources but they are still of reference value, such as the conflict between news reports and enterprise interface data in a supply chain scenario, a weak fusion strategy is adopted, storing the knowledge in a pending candidate pool or attaching a shadow state, without overwriting existing high-confidence knowledge. If the score is lower than the weak fusion threshold, a rejection strategy is adopted.

[0076] Finally, the update operation is performed according to the determined target fusion strategy to obtain the updated world model.

[0077] It should be noted that by dynamically selecting the target fusion strategy based on the fusion score, the current world model has great flexibility and robustness when facing incremental data with different compatibility. It can not only agilely absorb high-value knowledge, but also effectively deal with conflicting information through a weak fusion mechanism, thus achieving a balance between risk warning and business continuity.

[0078] The method provided in this invention, after verification, utilizes a graph neural network to propagate node information and calculates a fusion score to adaptively select a target fusion strategy for model updating. It quantifies the compatibility and conflict cost of new and old knowledge at the graph topology level, realizes automated and highly interpretable fusion decision-making, and significantly enhances the robustness and global consistency of the world model update method in complex conflict environments.

[0079] Based on any of the above embodiments, and based on the node representation, the fusion score between the newly added subgraph and the current world model is calculated, including: Based on the node representation, the semantic similarity and structural matching degree between the newly added subgraph and the current world model are calculated respectively; Evaluate the penalty value for logical conflicts between the newly added subgraph and the current world model; The fusion score is calculated by weighting and summing the semantic similarity, structural matching degree, and logical conflict penalty value.

[0080] Here, semantic similarity measures the degree of similarity between new and old knowledge in terms of content meaning and feature representation. Structural matching assesses the degree of fit between new and old knowledge in terms of network topology, neighbor relationships, and connection patterns.

[0081] Specifically, firstly, semantic similarity can be calculated using the original vectors of the newly added subgraph and the original vectors of the corresponding nodes in the current world model, for example, by employing a cosine similarity algorithm. Simultaneously, based on the node representations containing contextual information output by the graph neural network, the structural matching degree between the newly added subgraph and the current world model at the graph structure level is calculated to determine the naturalness and appropriateness of the new node's embedding into the current network topology.

[0082] Understandably, by calculating the matching metrics for these two dimensions separately, it is ensured that the integration of new knowledge is not only relevant in terms of literal semantics, but also perfectly matched in terms of system architecture and relational topology.

[0083] Next, the logical conflict penalty value between the newly added subgraph and the current world model is evaluated. The logical conflict penalty value is a scalar value that quantifies and penalizes objective contradictions between new and old knowledge based on predefined physical or business logic mutual exclusion constraints.

[0084] In detail, conflict detection is performed based on predefined domain business rules, and the penalty value for logical conflicts between the newly added subgraph and the current world model is evaluated. For example, predefined logical constraints include that entity A cannot be located in both B and C simultaneously, or that a road segment cannot be both free-flowing and extremely congested within the same time period. When a logical conflict or temporal misalignment is detected between the state introduced by the newly added subgraph and an existing high-confidence state in the current world model, the corresponding scalar penalty value is calculated; if there is no conflict, the value is zero.

[0085] It should be noted that when new knowledge contradicts existing knowledge systems, such as mutually exclusive states or temporal misalignments, existing methods lack a unified fusion scoring standard based on graph topology, and can only perform simple overwriting or discarding, leading to logical fragmentation of the knowledge system. The method provided in this invention, however, introduces a logical conflict penalty value, overcoming the shortcomings of traditional graph overwriting, enabling it to keenly capture and quantify irreconcilable business logic contradictions, thus preventing logical conflicts.

[0086] Then, the semantic similarity, structural matching degree, and logical conflict penalty value are weighted and summed to calculate the fusion score. Specifically, corresponding weight coefficients can be configured for semantic similarity, structural matching degree, and logical conflict penalty value. The first two are multiplied by their corresponding weights and then added together. Finally, the product of the logical conflict penalty value and its weight is subtracted, and the final result is the comprehensive fusion score.

[0087] The method provided in this invention comprehensively considers the semantic similarity and structural matching degree between nodes, and introduces a logical conflict penalty value to calculate the fusion score through weighted summation. This not only comprehensively evaluates the relevance of knowledge, but also severely punishes underlying logical contradictions, making the fusion scoring mechanism more rigorous in multiple dimensions and greatly improving the logic and accuracy of the world model fusion decision.

[0088] In one embodiment, Figure 3 This is a flowchart illustrating the knowledge fusion method provided by the present invention, as shown below. Figure 3 As shown, the method includes: First, obtain the new subgraph to be merged. and the existing maps in the current world model. This refers to the current world model. Next, the newly added subgraph is temporarily integrated into the existing graph, and a K-layer GNN (Graph Neural Network) is used for message passing and context aggregation between nodes, enabling nodes to interact with each other within the topology, thereby obtaining node representations containing global context information.

[0089] Based on this, a fusion score can be calculated according to a preset scoring function. Specifically, the calculation of this fusion score comprehensively considers three dimensions of quantitative features: semantic similarity calculated based on semantic features, structural matching degree calculated based on graph neural network aggregation features, and logical conflict penalty value calculated based on predefined logical mutual exclusion constraints. The weighted values ​​of semantic similarity and structural matching degree are added together, and then the weighted value of the logical conflict penalty value is subtracted to obtain the final fusion score. The fusion score can be calculated using the following formula, as shown below: ; In the formula, Indicates the fusion score; These represent the weight coefficients corresponding to semantic similarity, structural matching degree, and logical conflict penalty value, respectively. The semantic vector representing the newly added subgraph semantic vectors of the current world model Semantic similarity between them; Represents the structural feature vector of the newly added subgraph Structural feature vectors of the current world model The degree of structural matching between them; This represents the penalty value for logical conflicts.

[0090] Finally, a score range determination is performed, and differentiated target fusion strategies are automatically executed based on the numerical range in which the fusion score falls. When the fusion score is greater than or equal to the set strong fusion threshold T_strong, a strong fusion strategy is determined and executed, allowing new knowledge to directly overwrite or update the state in the existing graph. When the fusion score is greater than or equal to the set weak fusion threshold T_weak and less than the strong fusion threshold T_strong, a weak fusion strategy is determined and executed, storing the new knowledge in the candidate pool as pending data without overwriting high-confidence knowledge in the existing graph. When the fusion score is less than the weak fusion threshold T_weak, a rejection strategy is determined and executed, directly discarding the new knowledge and recording a conflict log.

[0091] Understandably, through the aforementioned closed-loop process, the compatibility and conflict costs of new and old knowledge can be accurately quantified at the graph topology level, achieving adaptive conflict-free fusion with high interpretability and global consistency.

[0092] Based on any of the above embodiments, the current world model includes an ontology layer, a state layer, and a causal layer; The newly added business knowledge vectors include newly added entity vectors, newly added state vectors, and newly added rule vectors.

[0093] Here, the ontology layer is the stable underlying architecture that defines the concepts, entity categories, and their hierarchical patterns in the business network; the state layer is a collection of instance snapshots that record the specific attribute values ​​of entities at a specific time point, belonging to a high-frequency update layer; and the causal layer is a rule system that stores the evolutionary patterns and decision-making logic between entity states. Correspondingly, the newly added entity vector, the newly added state vector, and the newly added rule vector represent the high-dimensional feature representations of incremental data in these three different abstract dimensions.

[0094] Correspondingly, the step of adopting the target fusion strategy to update the new business knowledge vector into the current world model to obtain the updated world model of the business network includes: Using the target fusion strategy, the newly added entity vector is updated to the ontology layer, the newly added state vector is updated to the state layer, and the newly added rule vector is updated to the causal layer to obtain the updated world model.

[0095] Specifically, during update operations, the three-layer structure is updated hierarchically and atomically according to the issued target fusion strategy instructions. In the ontology layer update, if new knowledge triggers ontology expansion and the fusion score allows, new entity vectors are written to expand the system's concept boundaries. In the state layer update, new state vectors are written, automatically appending the current system timestamp and confidence score, and the replaced old states are archived as historical snapshots for agile response. In the causal layer update, new rule vectors are written; if a new rule partially conflicts with an old rule but is determined to be strongly fused, the scope of application of the old rule is adjusted or its weight is reduced accordingly. Simultaneously, the entire update process employs a transaction mechanism to ensure that all three layers of updates either succeed or are rolled back, preventing data inconsistency, and finally outputting the update.

[0096] The method provided in this invention is based on a three-layer architecture consisting of an ontology layer, a state layer, and a causal layer. It decouples and updates newly added entity vectors, state vectors, and rule vectors using a target fusion strategy. This ensures both high-frequency and agile response to sudden events and dynamic states, and maintains the long-term steady state of the underlying ontology knowledge system, thereby achieving sustainable and collaborative evolution of the dynamic world model.

[0097] In one embodiment, Figure 4 This is a schematic diagram of the architecture of the world model provided by the present invention, such as... Figure 4 As shown, taking the extreme weather traffic assessment scenario in the field of public safety as an example, the closed-loop working process from multi-source data stream input to the decoupling and updating of the three-layer structure of the world model is fully presented.

[0098] First, at the input end of the current world model, it receives a continuous influx of multi-source input streams, namely the newly added knowledge. This includes not only unstructured social media text (such as breaking news such as "the road surface of the urban overpass is icy and many vehicles are skidding"), but also structured IoT sensor data (such as real-time numerical information such as the temperature of -3 degrees Celsius, the probability of precipitation of 90%, and the average vehicle speed of 15 kilometers per hour).

[0099] These complex, heterogeneous data from multiple sources are uniformly fed into the preceding processing pipeline, which includes processing steps such as parsing and vectorization, distribution uncertainty gating, dual logic verification, and structured fusion of graph neural networks. After information extraction, filtering, and rigorous verification, the data is transformed into structured knowledge with high confidence and is adaptively injected into the corresponding management levels of the world model through rule generation and state update operation paths.

[0100] Next, within the core architecture of the dynamic world model, a refined design employs a three-layer decoupled and collaboratively evolving ontology-state-causality framework. The ontology pattern layer (layer L1 in the diagram), located at the bottom of the architecture, forms the stable foundation of the entire graph knowledge system. It defines the basic concepts of the business domain and their macroscopic relational patterns. For example, the diagram clearly establishes the topological logic framework that the concept of "road facilities" is "affected by meteorological conditions," thus providing standardized definitional support for the construction of specific instances and logical rules at the upper layers. The causal rule layer (layer L3 in the diagram), located at the top of the architecture, is specifically responsible for storing and managing business logic rules derived from incremental data and rigorously verified through counterfactual simulation. As shown in the diagram, a new rule numbered "Rule_Ice_01" can be automatically generated and written. This rule clearly defines that when the triggering condition is "temperature below zero degrees Celsius and precipitation probability greater than 80%", the specific result will be "road surface state is icy," and further infers that this will lead to the subsequent impact of "reduced traffic capacity or reduced vehicle speed." The rule is marked with an extremely high verification metric (the probability difference in the causal strength test reaches 0.9), proving that it belongs to a strong causal relationship that has withstood logical verification, and is thus used to perform reliable causal inferences at the lower level.

[0101] Finally, the state instance layer (L2 layer in the diagram), located in the middle layer of the architecture, serves as a dynamic data carrier for agile response to emergencies, recording and maintaining the current snapshot information of specific business entities in real time. Under the structural constraints of the underlying ontology pattern and the deductive guidance of the top-level causal rules, for the specific entity object "urban elevated highway main road," the latest acquired state parameters are precisely written and automatically appended with a current system timestamp accurate to the minute. Simultaneously, driven by strong causal logic, the overall monitoring status of this specific road segment is highlighted and updated to a danger or congestion level. Based on the organic collaboration of the above three-layer structure and the real-time update of atomic states, a warning output of "Icing leads to accident risk" was successfully triggered and sent out without human intervention. This fully verifies that the dynamic world model architecture proposed in this invention possesses both agile real-time tracking capabilities and accurate and reliable automatic decision-making and early warning capabilities when dealing with multi-source heterogeneous data streams.

[0102] Based on any of the above embodiments Figure 5 This is the second flowchart illustrating the method for updating the world model of a business network provided by this invention. Figure 5 As shown, the method includes: First, the system begins running and receiving multimodal event data streams, which may contain unstructured or semi-structured information such as text news, system logs, and sensor time-series data from different data sources. Then, it proceeds to step S1, the multi-granularity parsing and vectorized embedding stage. In this stage, LLM is used to extract the input stream (new business knowledge) based on the current ontology model, identifying elements such as entities, relationships, state attributes, and potential causal rules. Pre-trained embedding models such as BERT (Bidirectional Encoder Representations from Transformers) or TransE (Translating Embeddings) are then used to uniformly transform these symbolic elements into high-dimensional vectorized embedding representations, i.e., the new business knowledge vectors. This transforms the complex multi-source input into a unified, structured representation that can be processed by a computer.

[0103] Next, the process enters the VAE (Variational Auto-Encoder) distribution uncertainty gating stage in step S2. The newly added vectorized embedding is input into the VAE model pre-trained on the existing knowledge base, and the reconstruction error of the knowledge is calculated through the encoding and decoding process. Then, based on this reconstruction error, the data distribution is determined: if the reconstruction error is extremely large, it is determined to be new OOD (Out-of-Distribution) pattern data, and the ontology expansion process will be bypassed for manual review or the establishment of a new pattern to avoid abnormal noise directly impacting the existing map; if the reconstruction error is small or moderate, i.e., it is determined to be a known pattern or a long-tail pattern, the incremental knowledge is allowed to continue into the main processing pipeline.

[0104] For data that successfully passes the gating, the process proceeds to step S3, the rule generation and dual logical verification stage. In this stage, the focus is on rigorous logical verification of the causal rules extracted or generated by the large language model. A dual verification mechanism is employed: self-consistency voting (i.e., multiple independent sampling to determine the consistency ratio of generated semantics) and counterfactual simulation (i.e., constructing parallel scenarios of facts and counterfactuals in a sandbox environment for deduction to quantify causal strength). This eliminates probabilistic illusions and relevance fallacies, ensuring that the rules in the incremental knowledge base possess genuine causal validity.

[0105] After successful verification, the process proceeds to step S4, the GNN (Graph Neural Network) structured fusion and conflict detection stage. The newly added subgraph corresponding to the new knowledge is temporarily integrated into the existing global knowledge graph. The graph neural network is used for information propagation between nodes to obtain a global context representation. A comprehensive fusion score is calculated by considering semantic similarity, structural matching, and logical conflict penalties. Based on this score, an automatic determination is made whether to adopt a strong fusion strategy (e.g., direct writing or state overwriting) or a weak fusion strategy (e.g., storing in a candidate pool without directly covering high-confidence knowledge), thereby achieving a fusion decision with interpretability and global consistency.

[0106] Finally, the system proceeds to step S5, the three-layer dynamic world model hierarchical update stage. Based on the fusion operation instructions generated by the preceding modules, atomic operations are performed on the underlying architecture of the world model. Specifically, this includes expanding the pattern definitions at the ontology layer, updating instance snapshots at the state layer, and overwriting the rule logic at the causal layer. Ensuring data consistency across the three layers, the update process concludes, and the system successfully generates a dynamic world model containing the latest multi-source knowledge and resolving conflicts. This achieves highly robust and logically rigorous continuous evolution for open-world environments.

[0107] In one specific embodiment, data source A (news): breaking news report "A fire broke out at the factory of a core chip supplier, X, with an estimated 30-day production disruption." Data source B (ERP system): API data from supplier X shows "Current inventory is sufficient, and the delivery plan is normal." Initial state: Supplier X's state in the world model is "healthy," which is a key node. Therefore, firstly, S1 parsing and vectorization are performed: extracting the news event (entity: factory X, event: fire, consequence: production disruption). Extracting the ERP state (entity: factory X, attribute: inventory = high, state: normal).

[0108] Then, S2 VAE uncertainty gating is performed: "Fire" is a long-tailed, low-frequency event for the supply chain entity, and the reconstruction error l_rec calculated by VAE is in the medium range ( <l_rec≤ This knowledge should be marked as requiring careful handling.

[0109] Next, S3 rule generation and double-fact verification are performed. Specifically, a causal rule is generated using LLM: "Factory fire -> Supply interruption". Then, counterfactual verification is performed: if "no fire", then "supply is normal". The causal logic holds. The rule itself is valid.

[0110] Next, perform S4 GNN structured fusion and conflict detection. Specifically, map the two-source data into a graph. GNN detects a logical conflict, that is, there is a time misalignment between the "production capacity interruption (future state)" derived from the news and the "normal shipment (current state)" reported by the ERP report, but there is a strong mutual exclusion signal in the attribute of "supplier health". In addition, calculate the conflict penalty. That is, since the weights of the two information sources (news vs official API) are equivalent and the content is mutually exclusive, the fusion score Score is pulled down and falls into the interval of T_weak≤Score<T_strong. Finally, judge and execute the weak fusion strategy based on the calculated fusion score. Finally, perform S5 hierarchical update. That is, at the state layer: do not overwrite the current "normal shipment" state (to prevent incorrect production line shutdown), but attach a shadow state of "high-risk pending verification" next to this node and subscribe to the news stream in the next 24 hours. At the causal layer: temporarily do not activate the decision rule of "switching to alternative suppliers", but trigger the temporary strategy of "increasing the inventory monitoring frequency".

[0111] The method provided by the embodiment of the present invention has the following advantages compared with the prior art: highly robust incremental update; that is, introducing VAE as a pre-distribution gate to effectively isolate the impact of out-of-distribution noise data on the existing knowledge system and achieve adaptive filtering of open-world event streams. Strongly credible causal reasoning; that is, introducing a dual verification mechanism of counterfactual simulation and self-consistent voting to transform the generation probability of the large model into a rigorously verified causal strength, significantly suppressing knowledge hallucinations. Structured conflict resolution; that is, using GNN to calculate the fusion score at the graph topology level, considering not only semantic similarity but also structural matching degree and logical conflict penalty, making the fusion decision interpretable and globally consistent. A sustainable evolving architecture; that is, through an update strategy that separates the ontology, state, and causality into three layers, it not only ensures an agile response to emergencies (state layer) but also maintains the long-term stability of the knowledge system (ontology layer), supporting the long-term online operation of the system.

[0112] In one embodiment Figure 6 is a schematic flowchart of the gating mechanism for new business knowledge provided by the present invention, as Figure 6 shown. First, receive the new business knowledge vector (input vector Z_new). This vector contains the structured representations extracted from multi-source event data streams and uniformly mapped. , where respectively represent entity embedding vectors, relationship embedding vectors, timestamped state attribute embeddings, and rule embeddings.

[0113] Next, the input vector Z_new (the newly added business knowledge vector) is fed into a pre-trained variational autoencoder (VAE) model. This VAE has been trained on an existing knowledge base and includes an encoder, latent variables, and a decoder structure. The input vector is mapped to the latent variables by the encoder, and then reconstructed by the decoder.

[0114] Subsequently, the reconstruction error is calculated. Specifically, the square of the Euclidean distance between the original input vector Z_new and the reconstructed output vector Z' is calculated and used as the reconstruction error L_rec (i.e., L_rec = ||Z - Z'||^2), thereby quantifying the uncertainty and novelty of the input knowledge at the data distribution level.

[0115] Then, based on the preset dual thresholds ( and ,in < The hierarchical path decision is performed. First, it is determined whether the reconstruction error L_rec is less than or equal to the first threshold. (L_rec ≤ If the judgment result is yes, then the newly added business knowledge is determined to belong to the known pattern (ACCEPT), indicating that its confidence level is high, and it is allowed to directly enter the subsequent dual logic verification module, that is, directly enter S3. If the reconstruction error L_rec is greater than the first threshold Then it is further determined whether it is greater than the second threshold. (L_rec> If the judgment result is negative, that is, the reconstruction error L_rec is between the first threshold and the second threshold (?). <L_rec ≤ If the condition is met, the knowledge is determined to belong to the long-tail pattern or fuzzy pattern (WEAK), indicating that its confidence level is moderate. At this point, the knowledge is allowed to enter the subsequent dual logic verification module (i.e., enter S3), but it will be marked as weakly fused (WEAK_FUSION), and should be handled with care during subsequent fusion.

[0116] If the judgment result is yes, that is, the reconstruction error L_rec is greater than the second threshold. If the newly added business knowledge is determined to be a completely unknown off-distribution ontology (EXPAND), it is marked as EXPAND_ONTOLOGY, thereby initially filtering out off-distribution noise and triggering the ontology expansion process to update the underlying architecture of the world model.

[0117] Based on any of the above embodiments Figure 7This is a schematic diagram of the structure of the world model update device for the business network provided by the present invention, as shown below. Figure 7 As shown, the device includes: Unit 710 acquires the newly added business knowledge vectors and the current world model of the business network; The gating unit 720 inputs the newly added business knowledge vector into the variational autoencoder to obtain the knowledge reconstruction error calculated by the variational autoencoder; the variational autoencoder is obtained based on training the initial encoder and decoder; The verification unit 730 verifies the newly added business knowledge vector when the knowledge reconstruction error is less than a preset error threshold, and obtains the verification result. The fusion unit 740, based on the verification results, fuses the newly added business knowledge vector and the current world model to obtain the updated world model of the business network.

[0118] The apparatus provided in this invention uses a variational autoencoder to perform distribution perception on newly added business knowledge vectors to calculate reconstruction errors. Under the premise of controllable errors, it combines a logical verification mechanism to obtain verification results. Finally, based on the results, it achieves conflict-free fusion with the current world model. This not only adaptively quantifies the novelty and uncertainty of input data in open environments, but also effectively solves the illusion problem of generative knowledge in large language models, ensuring that the business network can achieve efficient, accurate, and logically coherent continuous evolution when dealing with multi-source event streams.

[0119] Based on any of the above embodiments, the verification unit is specifically used for: Based on the newly added business knowledge vector, generate causal rules to be verified; The semantic consistency ratio of the causal rule to be verified is obtained by sampling the rule multiple times. If the semantic consistency ratio is greater than or equal to a preset consistency threshold, the causal rule to be verified is subjected to a counterfactual simulation test in a preset sandbox environment to obtain the causal strength. The verification result is determined based on the causal strength.

[0120] Based on any of the above embodiments, the verification unit is further specifically used for: Construct a first factual scenario that meets the conditions of the causal rule to be verified, and a second counterfactual scenario that modifies the conditions of the causal rule to be verified; In the sandbox environment, the probability of the first outcome in the first factual scenario and the probability of the second outcome in the second counterfactual scenario are respectively derived. The probability difference between the probability of the first outcome and the probability of the second outcome is calculated to obtain the causal strength.

[0121] Based on any of the above embodiments, the fusion unit is specifically used for: If the verification result indicates that the verification is successful, information propagation between nodes is performed between the newly added subgraph corresponding to the newly added business knowledge vector and the current world model to obtain a node representation; Based on the node representation, the fusion score between the newly added subgraph and the current world model is calculated; Based on the fusion score, a target fusion strategy is determined from a preset set of fusion strategies. The target fusion strategy is then used to update the new business knowledge vector into the current world model, resulting in the updated world model.

[0122] Based on any of the above embodiments, the fusion unit is further specifically used for: Based on the node representation, the semantic similarity and structural matching degree between the newly added subgraph and the current world model are calculated respectively; Evaluate the penalty value for logical conflicts between the newly added subgraph and the current world model; The fusion score is calculated by weighting and summing the semantic similarity, structural matching degree, and logical conflict penalty value.

[0123] Based on any of the above embodiments, the current world model includes an ontology layer, a state layer, and a causal layer; The newly added business knowledge vectors include newly added entity vectors, newly added state vectors, and newly added rule vectors; The fusion unit is also specifically used for: Using the target fusion strategy, the newly added entity vector is updated to the ontology layer, the newly added state vector is updated to the state layer, and the newly added rule vector is updated to the causal layer to obtain the updated world model.

[0124] Figure 8 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 8As shown, the electronic device may include a processor 810, a communication interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute a method for updating the world model of the business network. This method includes: acquiring a newly added business knowledge vector and the current world model of the business network; inputting the newly added business knowledge vector into a variational autoencoder to obtain a knowledge reconstruction error calculated by the variational autoencoder; the variational autoencoder being trained based on an initial encoder-decoder; verifying the newly added business knowledge vector if the knowledge reconstruction error is less than a preset error threshold to obtain a verification result; and fusing the newly added business knowledge vector and the current world model based on the verification result to obtain an updated world model of the business network.

[0125] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0126] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the world model update method for the business network provided by the above methods. The method includes: acquiring a newly added business knowledge vector and a current world model of the business network; inputting the newly added business knowledge vector into a variational autoencoder to obtain a knowledge reconstruction error calculated by the variational autoencoder; the variational autoencoder being trained based on an initial encoder-decoder; verifying the newly added business knowledge vector when the knowledge reconstruction error is less than a preset error threshold to obtain a verification result; and fusing the newly added business knowledge vector and the current world model based on the verification result to obtain an updated world model of the business network.

[0127] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements a method for updating the world model of a business network provided by the methods described above. The method includes: acquiring a newly added business knowledge vector and a current world model of the business network; inputting the newly added business knowledge vector into a variational autoencoder to obtain a knowledge reconstruction error calculated by the variational autoencoder; the variational autoencoder being trained based on an initial encoder-decoder; verifying the newly added business knowledge vector when the knowledge reconstruction error is less than a preset error threshold to obtain a verification result; and fusing the newly added business knowledge vector and the current world model based on the verification result to obtain an updated world model of the business network.

[0128] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. 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. Those skilled in the art can understand and implement this without any creative effort.

[0129] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for updating a world model of a business network, characterized in that, include: Obtain new business knowledge vectors and the current world model of the business network; The newly added business knowledge vector is input into the variational autoencoder to obtain the knowledge reconstruction error calculated by the variational autoencoder; the variational autoencoder is obtained based on training the initial encoder and decoder; If the knowledge reconstruction error is less than a preset error threshold, the newly added business knowledge vector is verified to obtain the verification result. Based on the verification results, the newly added business knowledge vector and the current world model are fused to obtain the updated world model of the business network.

2. The method for updating the world model of a business network according to claim 1, characterized in that, The verification of the newly added business knowledge vector, to obtain the verification result, includes: Based on the newly added business knowledge vector, generate causal rules to be verified; The semantic consistency ratio of the causal rule to be verified is obtained by sampling the rule multiple times. If the semantic consistency ratio is greater than or equal to a preset consistency threshold, the causal rule to be verified is subjected to a counterfactual simulation test in a preset sandbox environment to obtain the causal strength. The verification result is determined based on the causal strength.

3. The method for updating the world model of a business network according to claim 2, characterized in that, The step of performing counterfactual simulation tests on the causal rule to be verified in a preset sandbox environment to obtain the causal strength includes: Construct a first factual scenario that meets the conditions of the causal rule to be verified, and a second counterfactual scenario that modifies the conditions of the causal rule to be verified; In the sandbox environment, the probability of the first outcome in the first factual scenario and the probability of the second outcome in the second counterfactual scenario are respectively derived. The probability difference between the probability of the first outcome and the probability of the second outcome is calculated to obtain the causal strength.

4. The method for updating the world model of a business network according to any one of claims 1 to 3, characterized in that, Based on the verification results, the newly added business knowledge vector and the current world model are fused to obtain the updated world model of the business network, including: If the verification result indicates that the verification is successful, information propagation between nodes is performed between the newly added subgraph corresponding to the newly added business knowledge vector and the current world model to obtain a node representation; Based on the node representation, the fusion score between the newly added subgraph and the current world model is calculated; Based on the fusion score, a target fusion strategy is determined from a preset set of fusion strategies. The target fusion strategy is then used to update the new business knowledge vector into the current world model, resulting in the updated world model.

5. The method for updating the world model of a business network according to claim 4, characterized in that, The calculation of the fusion score between the newly added subgraph and the current world model based on the node representation includes: Based on the node representation, the semantic similarity and structural matching degree between the newly added subgraph and the current world model are calculated respectively; Evaluate the penalty value for logical conflicts between the newly added subgraph and the current world model; The fusion score is calculated by weighting and summing the semantic similarity, structural matching degree, and logical conflict penalty value.

6. The method for updating the world model of a business network according to claim 4, characterized in that, The current world model includes an ontology layer, a state layer, and a causal layer; The newly added business knowledge vectors include newly added entity vectors, newly added state vectors, and newly added rule vectors; The step of using the target fusion strategy to update the new business knowledge vector into the current world model to obtain the updated world model of the business network includes: Using the target fusion strategy, the newly added entity vector is updated to the ontology layer, the newly added state vector is updated to the state layer, and the newly added rule vector is updated to the causal layer to obtain the updated world model.

7. An apparatus for updating a world model of a business network, characterized in that, include: The acquisition unit acquires new business knowledge vectors and the current world model from the business network. The gating unit inputs the newly added business knowledge vector into the variational autoencoder to obtain the knowledge reconstruction error calculated by the variational autoencoder; the variational autoencoder is obtained based on training the initial encoder-decoder; The verification unit verifies the newly added business knowledge vector when the knowledge reconstruction error is less than a preset error threshold, and obtains the verification result. The fusion unit, based on the verification results, fuses the newly added business knowledge vector and the current world model to obtain the updated world model of the business network.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a method for updating the world model of the business network as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements a method for updating the world model of the business network as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements a method for updating the world model of the business network as described in any one of claims 1 to 6.

Images