Knowledge graph-based large model inference enhancement system

Abstract

The application relates to the technical field of large model reasoning and discloses a large model reasoning enhancement system based on a knowledge graph. The system comprises a knowledge graph construction module, a knowledge retrieval module and a reasoning enhancement module. The knowledge graph construction module comprises an entity relationship analysis unit and a knowledge diffusion degree evaluation unit, the former analyzes entities and associated relationships in input data, and the latter determines reasoning state parameters of the knowledge graph based on the output of the former. The knowledge retrieval module determines whether to trigger knowledge retrieval according to the reasoning state parameters, and retrieves supplementary knowledge related to the current reasoning task from a multi-source knowledge base when triggering. The reasoning enhancement module generates an optimal reasoning path of the large model based on the supplementary knowledge, and adjusts the reasoning decision mechanism of the large model accordingly. Through deep mining, accurate retrieval and reasoning mechanism optimization of knowledge, the system improves the adaptability and effectiveness of large model reasoning, and better copes with complex reasoning scenarios.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of large model reasoning, in particular to a large model reasoning enhancement system based on a knowledge graph. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, large models have been widely applied in natural language processing, intelligent question answering, decision support and many other fields. However, the existing large models still face many challenges in the reasoning process: the reasoning ability of large models is highly dependent on their training data, and when encountering knowledge or complex logical relationships that are not fully covered in the training data, reasoning bias or errors are likely to occur. For example, when dealing with problems in professional fields, due to the scarcity and particularity of professional knowledge, large models often have difficulty in accurately understanding and using relevant knowledge for effective reasoning; the reasoning process of large models lacks effective mining and utilization of knowledge association. Traditional reasoning methods are mostly based on statistical rules or pattern matching, which are difficult to deeply understand the potential relationships between entities, resulting in single reasoning path and insufficient reliability of reasoning results. At the same time, when faced with massive multi-source data, large models cannot quickly and accurately filter out the knowledge related to the current reasoning task, and the interference of a large amount of irrelevant information further reduces the reasoning efficiency and accuracy. The existing reasoning systems mostly lack the ability to dynamically adjust the reasoning decision mechanism, and cannot optimize the reasoning path according to the real-time information and knowledge changes in the reasoning process, making the reasoning process have certain limitations and rigidity, and being difficult to adapt to complex and variable actual application scenarios. SUMMARY

[0003] The purpose of the present application is to provide a large model reasoning enhancement system based on a knowledge graph to solve the problems raised in the background.

[0004] To achieve the above purpose, the present application provides a large model reasoning enhancement system based on a knowledge graph, which comprises:

[0005] A knowledge graph construction module, comprising an entity relationship analysis unit and a knowledge diffusion degree evaluation unit, the entity relationship analysis unit is used to analyze entities and their associated relationships in input data, and the knowledge diffusion degree evaluation unit is used to determine reasoning state parameters of the knowledge graph based on the output of the entity relationship analysis unit; a knowledge retrieval module, used to determine whether to trigger a knowledge retrieval operation based on the reasoning state parameters generated by the knowledge diffusion degree evaluation unit, and when it is determined to trigger the knowledge retrieval operation, supplementary knowledge related to the current reasoning task is retrieved from a multi-source knowledge base; a reasoning enhancement module, used to generate an optimal reasoning path of a large model based on the supplementary knowledge obtained by the knowledge retrieval module, and adjust the reasoning decision mechanism of the large model based on the optimal reasoning path.

[0006] Preferably, the entity relationship analysis unit comprises an entity relationship density calculation subunit configured to extract an entity network topology of the input data, the entity network topology comprising preset semantic correlation marked nodes.

[0007] Preferably, the knowledge diffusion degree evaluation unit determines the inference state parameter of the knowledge graph based on the output of the entity relationship analysis unit, including: for each semantic correlation marked node in the entity network topology, performing modal decomposition on the connection strength signal of the semantic correlation marked node in the continuous inference period to generate a core modal component and a noise residual component; generating an optimized connection strength signal of the continuous inference period based on the core modal component and the noise residual component through a noise reduction model; and calculating the inference state parameter of the knowledge graph based on the optimized connection strength signal corresponding to each semantic correlation marked node.

[0008] Preferably, the knowledge diffusion degree evaluation unit calculates the inference state parameter of the knowledge graph based on the optimized connection strength signal corresponding to each semantic correlation marked node, including: for each semantic correlation marked node, performing time domain sampling on the optimized connection strength signal to obtain connection strength values at multiple sampling time points; for each sampling time point, calculating a single-point inference state value based on the connection strength values of all semantic correlation marked nodes at the sampling time point; and aggregating the single-point inference state values at all sampling time points to generate the inference state parameter of the knowledge graph.

[0009] Preferably, the knowledge retrieval module determines whether to trigger a knowledge retrieval operation based on the inference state parameter generated by the knowledge diffusion degree evaluation unit, including: calculating the difference value of the inference state parameter of any two semantic correlation marked nodes; generating a global inference state difference degree based on all the difference values of the inference state parameter; and triggering the knowledge retrieval operation when the global inference state difference degree is less than a preset difference threshold and there is at least one semantic correlation marked node whose inference state parameter is greater than a preset state threshold.

[0010] Preferably, the entity network topology further comprises a plurality of knowledge anchor points; the knowledge retrieval module retrieves supplementary knowledge related to the current inference task from a multi-source knowledge base, including: performing cross-library retrieval operations under a plurality of knowledge retrieval dimensions; and the inference enhancement module generates an optimal inference path of the large model based on the supplementary knowledge obtained by the knowledge retrieval module, including: for each knowledge retrieval dimension, determining the mapping relationship between the knowledge anchor points and the entity association paths based on the supplementary knowledge retrieved in the dimension; fusing the mapping relationship between the knowledge anchor points and the entity association paths corresponding to all the knowledge retrieval dimensions to generate a global knowledge fusion mapping relationship; and determining the optimal inference path of the large model based on the global knowledge fusion mapping relationship.

[0011] Preferably, the inference enhancement module determines the optimal inference path of the large model based on the global knowledge fusion mapping relationship, comprising: generating a plurality of candidate inference paths based on the global knowledge fusion mapping relationship; obtaining decision confidence information of the large model in a continuous inference period; obtaining context complexity information corresponding to the period; screening the plurality of candidate inference paths based on the decision confidence information and the context complexity information to determine the optimal inference path.

[0012] Preferably, the system further comprises an inference optimization module for dynamically updating the inference decision mechanism of the large model based on the decision confidence information, the context complexity information and the optimal inference path.

[0013] Preferably, the inference enhancement module adjusts the inference decision mechanism of the large model based on the optimal inference path, comprising: generating a preset inference strategy based on historical inference paths, the preset inference strategy containing default decision logic of different task scenarios; when the deviation degree of the optimal inference path and the default decision logic of the current task scenario is greater than a preset deviation threshold, correcting the real-time decision logic of the large model and the preset inference strategy according to the optimal inference path.

[0014] Preferably, the inference optimization module dynamically updates the inference decision mechanism of the large model, comprising: generating a decision logic incremental parameter according to the optimal inference path; based on integral separation control algorithm, integrating the decision logic incremental parameter into the real-time decision logic of the large model; when detecting a context complexity mutation, resetting the fusion weight of the decision logic incremental parameter.

[0015] Compared with the prior art, the beneficial effects of the present application are that through the entity relationship analysis unit in the knowledge graph construction module, the entities and their associated relationships in the input data can be deeply analyzed, providing a more detailed and comprehensive knowledge base for the subsequent reasoning process. The knowledge diffusion degree evaluation unit determines the reasoning state parameters based on the results of entity relationship analysis, enabling the system to have a clear understanding of the current reasoning state, providing a reasonable trigger basis for subsequent knowledge retrieval operations. The knowledge retrieval module determines whether to trigger knowledge retrieval based on the reasoning state parameters, accurately obtaining relevant supplementary knowledge from multiple knowledge bases when needed, effectively avoiding irrelevant information interference, enabling the large model to obtain the required additional knowledge in the reasoning process, making up for possible deficiencies in the large model's training data. This targeted knowledge supplement can enrich the available knowledge resources in the reasoning process, broadening the perspective and dimension of reasoning. The reasoning enhancement module generates an optimal reasoning path based on the supplementary knowledge and adjusts the reasoning decision mechanism of the large model accordingly, making the reasoning process more flexible and intelligent. The generation of the optimal reasoning path can guide the large model to reason in a more reasonable and effective direction, reducing unnecessary reasoning steps and improving the coherence and logic of reasoning. The dynamic adjustment of the reasoning decision mechanism enables the large model to adaptively change the reasoning strategy according to different reasoning tasks and knowledge supplement conditions, better coping with complex and variable reasoning scenarios and improving the rationality and reliability of the reasoning results. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 a timing diagram of the large model reasoning enhancement system based on a knowledge graph according to the present application;

[0017] Figure 2 a flowchart for determining the reasoning state parameters for the knowledge diffusion degree evaluation unit;

[0018] Figure 3 a flowchart for calculating the reasoning state parameters;

[0019] Figure 4 a flowchart for multi-dimensional knowledge retrieval and reasoning path generation;

[0020] Figure 5 a flowchart for adjusting the reasoning decision mechanism. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0022] Referring to Figure 1 The present application provides a knowledge graph-based large model reasoning enhancement system, which comprises a knowledge graph construction module, a knowledge retrieval module and a reasoning enhancement module.

[0023] The knowledge graph construction module includes an entity relationship analysis unit and a knowledge diffusion degree evaluation unit. The entity relationship analysis unit is used to analyze the entities and their associated relationships in the input data, and to extract the semantic connections and structural features between entities. The knowledge diffusion degree evaluation unit calculates the reasoning state parameters of the knowledge graph based on the output of the entity relationship analysis unit, which quantifies the state stability and integrity of the knowledge graph during the reasoning process. The knowledge retrieval module determines whether to trigger the knowledge retrieval operation based on the reasoning state parameters, and retrieves supplementary knowledge related to the current reasoning task from multiple knowledge bases when triggered. The reasoning enhancement module generates the optimal reasoning path of the large model using the retrieved supplementary knowledge, and adjusts the reasoning decision mechanism of the large model based on the path to improve the reasoning accuracy and efficiency.

[0024] Embodiment 1: Referring to Figure 2 , involves the collaborative operation of the entity relationship analysis unit and the knowledge diffusion degree evaluation unit. The entity relationship analysis unit includes an entity relationship density calculation subunit, which is responsible for processing input data and extracting the entity network topology structure therein. The input data usually comes from the text, image or multi-modal information currently processed by the large model, and is converted into a structured entity-relation set after preprocessing. The entity relationship density calculation subunit identifies the explicit and implicit associations between entities by analyzing the entity co-occurrence frequency, semantic similarity and predefined domain rules. For example, when processing a scientific literature about climate change, the subunit will identify entities such as "carbon dioxide", "global warming", "greenhouse effect", etc., and build an association network between these entities according to the logical discussion in the literature and the domain knowledge base. This network not only contains entity nodes and relationship edges, but also embeds preset semantic association marker nodes. These marker nodes are key hub nodes set according to prior knowledge, used to identify specific types of semantic relationships or domain core concepts. In the climate change example, semantic association marker nodes may be set as "greenhouse gases" or "temperature rise", to mark key semantic nodes in the network that have a dominant role in the reasoning process.

[0025] The construction process of the entity network topology structure adopts a graph structure representation, in which nodes represent entities, edges represent relationships, and semantic association marker nodes are assigned special type attributes. The entity relationship density calculation subunit calculates the connection density of each region, i.e. the number of entities and relationships in a unit region, to identify the core and edge regions in the network. This process helps the subsequent units focus more on in-depth analysis of high-density regions.

[0026] The knowledge diffusion evaluation unit further determines the reasoning state parameters of the knowledge graph based on the generated entity network topology. The processing of this unit is centered on the semantic association marked nodes. For each semantic association marked node, the unit extracts its connection strength signal within the continuous reasoning period. This signal is a time series data that reflects the changes in the connection strength between the node and other nodes over time. The calculation of the connection strength takes into account the weight of the relationship, the interaction frequency, and the semantic consistency. Within the continuous reasoning period, the large model may conduct multiple rounds of reasoning or gradually in-depth analysis on the same problem, and the connection strength signal captures the dynamic changes in the association activity of the marked node during this process.

[0027] The knowledge diffusion evaluation unit performs modal decomposition on the connection strength signal of each semantic association marked node. The purpose of modal decomposition is to decompose the original signal into different modal components, so as to separate the stable trend and random fluctuations in the signal. The empirical mode decomposition method adopted can adaptively decompose the signal into a group of core modal components and a noise residual component. The core modal component represents the stable, periodic or trend part of the signal, reflecting the continuous and reliable association pattern of the marked node during reasoning. The noise residual component contains high-frequency noise, abnormal fluctuations or temporary interference in the signal, which may be caused by noise in the input data or uncertainty in the model reasoning process.

[0028] After obtaining the core modal component and the noise residual component, the original connection strength signal is optimized through a denoising model. The denoising model uses an adaptive filtering algorithm based on signal processing technology. This algorithm dynamically adjusts the filtering parameters according to the statistical characteristics of the noise residual component, such as variance and spectral distribution. The purpose is to maximize the suppression of noise components while retaining the useful information carried by the core modal component. The optimized connection strength signal is smoother and more reliable, and can more truly reflect the actual connection state of the semantic association marked node during reasoning. This step is crucial for accurately calculating the reasoning state parameters later, as it eliminates random interference in the signal and highlights the potential regularity pattern.

[0029] Based on the optimized connection strength signals of each semantic association marker node, the knowledge diffusion degree evaluation unit begins to calculate the reasoning state parameter of the knowledge graph. First, each optimized signal is standardized to eliminate the dimensional influence caused by the difference in basic connection strength between different nodes. Standardization usually uses the min-max scaling or Z-score standardization method to adjust the numerical range of all signals to a unified interval. Subsequently, based on the standardized signals of all semantic association marker nodes, the final reasoning state parameter is generated by weighted average method. The weight distribution is not fixed, but is dynamically adjusted according to the importance of each semantic association marker node in the network. The evaluation of node importance includes its connectivity (how many edges are connected to it), feature vector centrality (whether it is connected to other important nodes), and pre-defined domain importance label. For example, in a medical diagnosis knowledge graph, the "clinical symptoms" marker node may be assigned a higher weight than the "patient demographics" marker node, because the former has a more direct and significant impact on the reasoning process.

[0030] The final generated reasoning state parameter is a comprehensive scalar value that quantifies the overall state of the current knowledge graph in the reasoning process. The parameter value reflects the activity, stability and potential support for reasoning of the knowledge graph. A higher value usually indicates that the knowledge graph is in an active and stable state, with rich and reliable semantic associations; a lower value may indicate that the knowledge graph has problems such as sparse information, weak association or large noise interference. This parameter provides a decision basis for the subsequent knowledge retrieval module, helping the system to judge whether to obtain supplementary knowledge from external knowledge base to enhance reasoning.

[0031] The entire implementation process embodies the analysis path from micro to macro: starting from the signal processing of individual semantic association marker nodes, gradually aggregating to the evaluation of global state. This method not only captures the dynamic characteristics of the knowledge graph in detail, but also ensures the robustness and representativeness of the evaluation results through noise reduction and weighting mechanisms.

[0032] Example 2: see Figure 3 , focusing on the refinement of the reasoning state parameter and the decision mechanism of knowledge retrieval triggering. This implementation takes semantic association marker nodes as the processing core, which play a key role in the entity network topology structure. For example, in the financial risk analysis scenario, the input data may contain entities such as "transaction account", "money flow", "merchant type", and the pre-set semantic association marker nodes include "abnormal transaction pattern" and "credit risk level", which connect multiple risk analysis dimension entities in the network.

[0033] The knowledge diffusion degree evaluation unit first processes the optimized connection strength signal corresponding to each semantic association labeled node. This signal is the time series data after the noise reduction processing of Embodiment 1, representing the stable trend of the node connection state. Taking the "abnormal transaction pattern" node as an example, its connection strength signal may reflect the change in the association strength of this node with entities such as "high-frequency transfer" and "cross-border payment" in continuous reasoning periods. Time-domain sampling is performed on this signal, using a fixed time interval sampling strategy. The sampling frequency is dynamically set according to the real-time requirements of the large model reasoning task, for example, millisecond-level sampling in high-frequency transaction monitoring, and minute-level sampling in credit evaluation. The connection strength value obtained at each sampling point represents the connection activity of the node at a specific time. In the financial case, the strength value of the "abnormal transaction pattern" node may suddenly increase at a certain sampling time, corresponding to the large model analyzing a suspicious transaction cluster.

[0034] For each sampling time, the system calculates a single-point reasoning state value, which is based on the connection strength values of all semantic association labeled nodes at the current time. First, the arithmetic mean of these strength values is calculated as the basic state indicator. However, the simple average value cannot reflect the synergistic effect between nodes, so a semantic association degree correction mechanism is introduced. The semantic association degree is measured by the co-occurrence frequency and semantic similarity between nodes. In the financial network, the "abnormal transaction pattern" node has a strong semantic association with the "credit risk level" node, as abnormal transactions often accompany credit deterioration. The correction process enhances the contribution weight of such strongly associated node combinations and weakens the influence of unrelated node combinations. The corrected single-point reasoning state value more comprehensively represents the overall reasoning state characteristics of the knowledge graph at that time.

[0035] The single-point reasoning state values of multiple sampling times are aggregated to generate the final reasoning state parameter, and the aggregation process uses time series analysis methods to perform data smoothing on the single-point value sequence. The moving average algorithm is used to eliminate short-term fluctuations, and the window size is adaptively adjusted according to the duration of the reasoning task. For example, in a 10-minute credit approval reasoning, the system may use a 3-minute sliding window to balance real-time and stability. The smoothed sequence is identified by trend extraction technology to identify the main change direction, and the final parameter takes the weighted average value of the entire sequence, with the time weight increasing as the sampling time approaches the current time. In the financial case, this parameter may be displayed as 0.78 (range 0-1), reflecting that the current risk analysis knowledge graph is in a highly active and stable state.

[0036] The knowledge retrieval module determines whether to trigger a retrieval operation based on the above reasoning state parameters. The process first calculates the parameter difference value of any two semantically associated marked nodes. In the financial network, the difference between the "abnormal transaction pattern" node parameter 0.85 and the "credit risk level" node parameter 0.72 can be compared. The difference is calculated in the form of squared difference to amplify significant differences, such as (0.85-0.72)²=0.0169. The difference values of all node pairs form a difference matrix, and after normalization to eliminate the influence of the number of nodes, the weighted sum is calculated to generate the global reasoning state difference. The weight allocation considers the structural importance of the node in the network, and the central nodes such as "abnormal transaction pattern" are usually given higher weights.

[0037] The trigger decision adopts a dual condition mechanism, triggering retrieval when the global difference is less than the preset threshold (indicating that the knowledge graph state is stable), and at least one node parameter is greater than the state threshold (indicating local high activity). In the financial case, the preset threshold is set through historical data analysis: the difference threshold 0.05 is determined based on the parameter distribution of normal transactions and fraudulent transactions; the state threshold 0.8 is calibrated according to the node activity when high-risk events occur. If the "abnormal transaction pattern" node parameter 0.85>0.8 and the global difference 0.03<0.05 are detected, knowledge retrieval is triggered. At this time, the system retrieves supplementary information from multiple knowledge bases such as anti-money laundering databases and credit investigation systems, such as the latest fraud pattern characteristics or the credit changes of associated accounts.

[0038] Embodiment 3: see Figure 4 , focusing on knowledge anchor-driven cross-dimensional retrieval and reasoning path optimization. This implementation is based on an entity network topology that includes semantically associated marked nodes and knowledge anchors. Knowledge anchors are special nodes in the network with high connectivity and semantic centrality, usually representing core concepts or key reasoning nodes in the field. Taking the medical diagnosis scenario as an example, the input data includes "patient complaints", "laboratory indicators", "imaging features", etc. The preset knowledge anchors include "initial diagnosis hypothesis" and "treatment options", which connect the key path of symptoms, test results, and treatment decisions.

[0039] The knowledge retrieval module obtains supplementary knowledge from multiple knowledge sources after triggering the retrieval operation. The retrieval process performs cross-knowledge base queries in parallel in three preset dimensions: the semantic dimension retrieves based on the semantic similarity between entities, such as matching the association knowledge of "chest pain" and "myocardial infarction" through a medical ontology library; the time sequence dimension focuses on the time effectiveness of knowledge, retrieving the myocardial infarction diagnosis and treatment standards updated in the past three years from the latest clinical guideline library; and the structure dimension retrieves case knowledge with similar symptom combinations from the electronic medical record library according to the topological characteristics of entity networks. Cross-knowledge base retrieval uses a distributed query engine to simultaneously access heterogeneous sources such as clinical knowledge graphs, scientific literature libraries, and drug databases, and solves the terminology differences of different knowledge bases through entity alignment technology, such as mapping "CK-MB" and "creatine kinase isoenzyme" uniformly.

[0040] The reasoning enhancement module generates the optimal reasoning path based on the retrieval results. For each retrieval dimension, the module establishes a mapping relationship between knowledge anchors and entity association paths. In the semantic dimension, the retrieved "myocardial infarction diagnosis standard" is converted into an entity path: chest pain → ECG ST segment elevation → troponin elevation → confirmed myocardial infarction, and is mapped to the "initial diagnosis hypothesis" anchor. The mapping is based on the semantic consistency of the path endpoint and the anchor reaching a threshold. In the time sequence dimension, the "direct PCI treatment path" recommended by the latest guidelines is mapped to the "treatment plan option" anchor, which includes the sequence of entities such as onset time window evaluation → catheter room activation → antiplatelet drug administration. In the structure dimension, the diagnosis and treatment path of similar cases "chest pain → coronary angiography → stent implantation" is extracted and mapped to two knowledge anchors at the same time because it involves both diagnosis and treatment.

[0041] The global knowledge fusion mapping relationship is generated by fusing the mapping relationships of multiple dimensions. The fusion process uses a conflict resolution mechanism: when there is a contradiction between the mapping relationships of different dimensions (such as recommending direct PCI in the time sequence dimension and conservative drug treatment in the structure dimension), the system assigns a credibility weight based on the authority and timeliness of the mapping source. The timeliness weight of the clinical guidelines is set to 0.8, and the universality weight of the case library is set to 0.6, and the conflict is eliminated by weighted fusion. The final global mapping relationship is represented as:

[0042]

[0043] Where: represents the global knowledge fusion mapping relationship, is the total number of knowledge retrieval dimensions (D=3 in this example), represents the fusion weight of the th dimension (dynamically adjusted according to the importance of the dimension), represents the set of knowledge anchor and entity association path mapping relationships of the th dimension. This formula ensures that the contributions of each dimension are integrated into a unified mapping framework according to the weights.

[0044] Based on the global mapping relationship, multiple candidate reasoning paths are generated, which may produce three paths in a medical case: path A (guideline priority path) follows the sequence of "symptom evaluation → rapid detection → PCI treatment"; path B (case similarity path) adopts "differential diagnosis → invasive examination → individualized treatment"; and path C (comprehensive path) combines the timeliness of A and the individualization of B. The system obtains the decision confidence information of the large model in the continuous reasoning period, which is calculated through the probability distribution variance of the model output. For example, the decision probability of "PCI treatment" in path A reaches 92%, so the confidence is high, and the probability of "individualized treatment" in path B is scattered between 65%-78%, so the confidence is medium. At the same time, the context complexity information is obtained, which is calculated through the entropy value of the current input data: if the patient has multiple complications such as diabetes and renal dysfunction, the complexity entropy value is significantly higher than that of a single disease.

[0045] When screening the optimal reasoning path, the system first excludes paths with decision confidence lower than the threshold (such as <70%). Then the matching degree of the remaining paths and the context complexity is calculated: path A is suitable for low-to-medium complexity cases (entropy value <2.0), and path B is suitable for high complexity cases (entropy value ≥2.0). If the current case entropy value is 2.3 (an elderly patient with multiple underlying diseases), path B gets a higher matching score. Finally, the optimal path is determined by weighted comprehensive score, with decision confidence weight 40% and complexity matching degree weight 60%, ensuring that in complex medical scenarios, the reasoning path with stronger adaptability is preferentially selected.

[0046] Example 4: see Figure 5 , focusing on the dynamic optimization and strategy update of the reasoning decision mechanism, the implementation is based on the reasoning optimization module, which continuously monitors the reasoning state of the large model, and dynamically adjusts the decision mechanism according to the decision confidence information, context complexity information and optimal reasoning path. Taking an intelligent investment decision support system as an example, the system processes input such as financial market data, company financial reports and macroeconomic indicators, and the entities in its knowledge graph include "stock code", "industry board", "financial indicators", and the pre-set reasoning strategies include default decision logic in different market environments, such as "momentum strategy in bull market trend" and "mean reversion strategy in volatile market".

[0047] The reasoning enhancement module first generates preset reasoning strategies based on historical reasoning paths. The historical paths are extracted from the past decision records of the large model, and the strategy templates are formed through pattern recognition and sequence analysis. In the investment case, the system analyzes tens of thousands of buy and sell decision records in the past five years and finds that when the "P / E ratio is lower than the industry average" and "cash flow growth is more than 20% year-on-year", the model has a 78% probability of making a "buy" decision. This pattern is summarized as the "value investment" default logic. Another common pattern is when "market volatility surges" and "risk-averse asset prices rise", the model has a 65% probability of starting "risk-avoidance" logic to reduce stock positions. These patterns are formalized as a preset strategy library containing trigger conditions, decision rules, and execution priorities after clustering and verification.

[0048] The system compares the deviation degree of the optimal reasoning path and the default decision logic of the current task scene in real time. The deviation degree calculation uses a comprehensive measure of structural difference and semantic distance. In a certain trading period, the market suddenly has a policy benefit, and the optimal reasoning path (generated by embodiment 3) suggests immediately overallocating to the beneficiary industry, while the current default "cautious" logic suggests maintaining the position based on historical volatility data. Path analysis shows that the structural difference is reflected in the operation sequence. The optimal path contains a rapid response chain of "identifying policy benefits → calculating industry elasticity → adjusting the investment portfolio", while the default logic follows a conservative sequence of "waiting for market confirmation → gradually building positions". The semantic distance is reflected in the decision goal. The optimal path emphasizes opportunity capture, and the default logic focuses on risk control. After calculation, the deviation degree of the two reaches 0.47 (range 0-1), far exceeding the system's preset deviation threshold of 0.3.

[0049] When the deviation degree exceeds the threshold, the system corrects the real-time decision logic of the large model according to the optimal reasoning path. The correction process has two steps: first, convert the optimal path into executable decision rules, such as decomposing "immediately overallocate to the beneficiary industry" into specific instructions such as "increase the relevant industry position to 20% within 1 hour after the policy is released"; then inject these rules into the real-time decision logic to cover or supplement existing rules. At the same time, the preset reasoning strategy is also updated synchronously. The system adds a strategy template under the "policy market response strategy" category, including "quickly identify policy impact industries", "elasticity coefficient calculation model", "position adjustment threshold", and other elements for future similar scenarios.

[0050] The inference optimization module dynamically updates the inference decision mechanism of the large model, and this process generates decision logic incremental parameters, which represent the decision rules and their weights that need to be added, deleted, or adjusted. In the investment case, the incremental parameters include: adding rule parameters (such as setting the "policy benefit response rule" weight to 0.8), deleting rule parameters (such as reducing the "wait and see rule" weight to 0.2), and adjusting parameters (such as adding a policy sensitivity factor in the "industry elasticity coefficient" calculation). The generation of incremental parameters is achieved by comparing the differences between the optimal path and the current logic: the system identifies that the "elasticity coefficient calculation" step is included in the optimal path but missing in the current logic, and parameterizes it as a new rule; at the same time, it finds that the "minimum holding limit" in the current logic is ignored in the optimal path, and its weight is lowered.

[0051] The integral separation control algorithm is used to integrate the incremental parameters into the real-time decision logic, which decomposes the incremental parameters into instantaneous components and cumulative components. The instantaneous component is used to respond to emergency changes, such as immediately enabling the "policy response rule" when a market policy suddenly occurs; the cumulative component is used for gradual adjustment, such as gradually increasing the weight of the "industry rotation strategy" to avoid excessive volatility. Algorithm parameters are dynamically adjusted according to decision confidence and context complexity: when the policy benefit is clear (confidence 0.9) and the market environment is simple (low complexity), the instantaneous component weight is set to 0.9 to achieve fast response; when the signal is ambiguous (confidence 0.6) and the market is complex (multiple factors intertwined), the cumulative component weight increases to 0.7, adopting a cautious adjustment mode.

[0052] The system continuously monitors the context complexity and resets the fusion weight when a mutation is detected. Complexity mutations are identified by calculating the entropy of context information in real time: if the entropy value changes at a rate exceeding the threshold (such as 30% / minute) within a short period of time, it is determined to be a mutation event. For example, a sudden geopolitical conflict breaks out during trading hours, causing the market complexity entropy value to soar from 1.2 to 2.8, and the system immediately resets the fusion weight of the incremental parameters. The reset process includes: clearing the historical cumulative value of the current cumulative component, temporarily increasing the application proportion of the instantaneous component to 95%, and reinitializing the rate parameter of the integral channel. After resetting, the system reevaluates the effectiveness of the incremental parameters based on the latest context, such as suspending the "policy response rule" in a conflict environment, and generating a new "risk avoidance rule" incremental parameter, as shown in Table 1.

[0053] Table 1: Decision logic incremental parameter table

[0054] Parameter type Target rule Weight change Trigger condition Effective range New parameter Policy response rule +0.80 Policy published and industry elasticity > 0.6 Equity assets Delete parameter Wait and see rule -0.60 Market volatility <0.5 All assets Adjust parameter Industry rotation rate +0.35 Macro economy in expansion Industry fund New parameter Conflict hedge rule +0.75 Geopolitical risk index > 0.7 Global assets Adjust parameter Position adjustment threshold -0.20 Market liquidity shortage High-risk assets

[0055] Embodiment 5: Fine-grained dynamic update focusing on reasoning decision mechanism. This implementation is achieved through a reasoning optimization module that continuously processes optimal reasoning paths, decision confidence information, and context complexity information to adaptively adjust the real-time decision logic of large models. Taking an autonomous driving decision support system as an example, the system processes inputs such as vehicle sensor data, traffic environment information, and navigation instructions. The entities in its knowledge graph include "traffic signals", "obstacle types", and "road topologies". The pre-set decision logic includes behavior rules for different driving scenarios, such as "following distance control" and "intersection passing strategy".

[0056] The reasoning optimization module first generates decision logic incremental parameters based on the optimal reasoning path. These parameters represent the decision rules that need to be added, modified, or deleted and their impact weights. In the autonomous driving scenario, when the optimal path suggests "starting braking 50 meters in advance on wet and slippery road surface" and the current decision logic only contains "standard braking distance of 30 meters on dry road surface", the system analyzes the conditional judgment sequence and operation instructions in the path and converts them into structured parameters. Incremental parameters include three categories: rule addition parameters (such as adding a "wet and slippery road surface braking coefficient" rule), rule deletion parameters (such as abandoning the "constant following time distance" rule), and rule weight adjustment parameters (such as increasing the weight of the "rain visibility compensation" rule from 0.6 to 0.8). The parameter generation process is achieved through a difference comparison engine that analyzes the semantic tree structure differences between the optimal path and the current decision logic layer by layer and quantifies the differences into a parameter set. For example, if it is identified that the optimal path contains a new conditional branch "activate special braking when road surface humidity > 0.7", a corresponding rule addition parameter is generated.

[0057] Based on the integral separation control algorithm, the system integrates the decision logic incremental parameters into the real-time decision logic of the large model. This algorithm uses a dual-channel processing architecture: the high-frequency channel processes changes that require immediate response, and the low-frequency channel processes changes that require gradual adjustment. In the scenario where the vehicle is about to enter a wet and slippery curve, the high-frequency channel directly applies the instantaneous component, immediately loading the "wet and slippery road surface braking coefficient" rule and setting the initial weight to 0.9, so that the vehicle can apply the new rule in the next decision cycle. The low-frequency channel gradually applies the cumulative component through time integration. For example, the weight of the "dynamic adjustment of following time distance" rule is gradually increased from 0.5 to 0.7, with an increase of 0.1 weight every 2 seconds, avoiding sudden changes in driving strategy that may cause passenger discomfort. Algorithm parameters are dynamically configured according to real-time environment: when sensors detect an emergency situation (decision confidence 0.95) and the road environment is simple (low complexity), the high-frequency channel proportion increases to 80%; when the road conditions are complex (multiple target conflicts) and the confidence is moderate (0.65), the low-frequency channel proportion increases to 70%, achieving smooth transition.

[0058] The system continuously monitors the context complexity information and calculates the environmental entropy in real time through dedicated sensors and data analysis modules. When a sudden change in context complexity is detected, the fusion weight reset mechanism is triggered immediately. During highway driving, a sudden rainstorm is encountered. The meteorological sensor and visual system simultaneously feedback a sharp decrease in visibility and a sharp decrease in road friction coefficient. The environmental entropy jumps from 1.2 to 3.5 within 3 seconds, exceeding the preset mutation threshold of 2.0. The system immediately performs weight reset: clear all parameter adjustment history accumulated by the low-frequency channel; temporarily increase the instantaneous application ratio of the high-frequency channel to 95%; reinitialize the integral rate parameter to set a more conservative adjustment step (e.g., from 0.1 / s to 0.05 / s) based on the latest environmental settings. After reset, the system re-evaluates the effectiveness of the incremental parameters, suspends the "high-speed cruise optimization" parameters generated before the rainstorm, and generates new "low-visibility following" incremental parameters based on the current visibility data, including reducing the upper limit of vehicle speed and increasing the following distance.

[0059] The entire updating process maintains stability through a closed-loop control mechanism. After each parameter adjustment, the system collects driving behavior feedback data (such as brake smoothness and trajectory deviation) and evaluates the adjustment effect in combination with decision confidence information. If it is detected that the "wet road braking coefficient" rule is applied after a sudden braking (confidence drops to 0.6), the weight value is automatically adjusted and the corrected parameters are generated. At the same time, the preset strategy library is updated collaboratively: when a new rule is successfully verified in 10 consecutive wet road scenarios, it is solidified as a sub-rule of the standard strategy "rainy day driving mode"; when a rule repeatedly causes decision conflicts (confidence consistently <0.5) in complex intersection scenarios, it is marked as a rule to be discarded and its weight is gradually reduced until it is zero.

[0060] This dynamic updating mechanism enables the automatic driving system to adapt to the ever-changing road environment, ensuring safety while continuously evolving the decision logic. Through fine-tuned incremental parameters and rapid reset of mutation responses, the system maintains decision coherence and reliability in extreme scenarios such as sudden rainstorms, avoiding logic confusion or inappropriate behavior caused by sudden environmental changes.

[0061] It should be noted that, in this text, relational terms such as first and second are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device.

[0062] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A knowledge graph-based large model inference enhancement system, characterized in that, Comprise: A knowledge graph construction module, comprising an entity relationship analysis unit and a knowledge diffusion degree evaluation unit, the entity relationship analysis unit is used to analyze the entity and its associated relationship in the input data, and the knowledge diffusion degree evaluation unit is used to determine the reasoning state parameter of the knowledge graph based on the output of the entity relationship analysis unit; A knowledge retrieval module for determining whether to trigger a knowledge retrieval operation based on the reasoning state parameter generated by the knowledge diffusion degree evaluation unit, when determining to trigger the knowledge retrieval operation, retrieving supplementary knowledge related to the current reasoning task from a multi-source knowledge base; A reasoning enhancement module for generating an optimal reasoning path of a large model based on the supplementary knowledge obtained by the knowledge retrieval module, and adjusting the reasoning decision mechanism of the large model based on the optimal reasoning path; The entity relationship analysis unit comprises an entity relationship density calculation subunit, which is used to extract the entity network topology structure of the input data, and the entity network topology structure contains a preset semantic association marked node; The knowledge diffusion degree evaluation unit determines the reasoning state parameter of the knowledge graph based on the output of the entity relationship analysis unit, comprising: for each semantic association marked node in the entity network topology structure, the connection strength signal of the semantic association marked node in the continuous reasoning period is modally decomposed to generate core modal components and noise residual components; through a noise reduction model, the core modal components and the noise residual components are used to generate an optimized connection strength signal of the continuous reasoning period; based on the optimized connection strength signal corresponding to each semantic association marked node, the reasoning state parameter of the knowledge graph is calculated; The knowledge diffusion degree evaluation unit calculates the reasoning state parameter of the knowledge graph based on the optimized connection strength signal corresponding to each semantic association marked node, comprising: for each semantic association marked node, the optimized connection strength signal is time-domain sampled to obtain connection strength values at multiple sampling times; for each sampling time, based on the connection strength values of all semantic association marked nodes at the sampling time, a single-point reasoning state value is calculated; all single-point reasoning state values at all sampling times are aggregated to generate the reasoning state parameter of the knowledge graph.

2. The knowledge graph-based large model inference enhancement system according to claim 1, wherein, The knowledge retrieval module determines whether to trigger a knowledge retrieval operation based on the reasoning state parameter generated by the knowledge diffusion degree evaluation unit, comprising: calculating the reasoning state parameter difference value of any two semantic association marked nodes; based on all the reasoning state parameter difference values, a global reasoning state difference degree is generated; when the global reasoning state difference degree is less than a preset difference threshold, and there is at least one semantic association marked node whose reasoning state parameter is greater than a preset state threshold, the knowledge retrieval operation is triggered.

3. The knowledge graph-based large model inference enhancement system according to claim 1, wherein, The entity network topology structure further comprises a plurality of knowledge anchors; The knowledge retrieval module retrieves supplementary knowledge related to the current reasoning task from a multi-source knowledge base, comprising: performing cross-library retrieval operation under multiple knowledge retrieval dimensions; The reasoning enhancement module generates an optimal reasoning path of the large model based on the supplementary knowledge obtained by the knowledge retrieval module, including: for each knowledge retrieval dimension, based on the supplementary knowledge retrieved in this dimension, determining the mapping relationship between the knowledge anchor point and the entity association path; fusing the mapping relationship between the knowledge anchor point and the entity association path corresponding to all knowledge retrieval dimensions to generate a global knowledge fusion mapping relationship; determining the optimal reasoning path of the large model based on the global knowledge fusion mapping relationship.

4. The knowledge graph-based large model inference enhancement system according to claim 3, wherein, The reasoning enhancement module determines the optimal reasoning path of the large model based on the global knowledge fusion mapping relationship, including: generating a plurality of candidate reasoning paths based on the global knowledge fusion mapping relationship; obtaining decision confidence information of the large model in a continuous reasoning period; obtaining context complexity information corresponding to the period; screening the plurality of candidate reasoning paths based on the decision confidence information and context complexity information to determine the optimal reasoning path.

5. The knowledge graph-based large model inference enhancement system according to claim 4, wherein, Further comprising a reasoning optimization module for dynamically updating the reasoning decision mechanism of the large model based on the decision confidence information, context complexity information and the optimal reasoning path.

6. The knowledge graph-based large model inference enhancement system according to claim 5, wherein, The reasoning enhancement module adjusts the reasoning decision mechanism of the large model based on the optimal reasoning path, including: generating a preset reasoning strategy based on historical reasoning paths, the preset reasoning strategy containing default decision logic of different task scenarios; when the deviation degree of the optimal reasoning path and the default decision logic of the current task scenario is greater than a preset deviation threshold, correcting the real-time decision logic of the large model and the preset reasoning strategy according to the optimal reasoning path.

7. The knowledge graph-based large model inference enhancement system according to claim 6, wherein, The reasoning optimization module dynamically updates the reasoning decision mechanism of the large model, including: generating a decision logic incremental parameter according to the optimal reasoning path; based on integral separation control algorithm, the decision logic incremental parameter is integrated into the real-time decision logic of the large model; when detecting context complexity mutation, resetting the fusion weight of the decision logic incremental parameter.

Images