Knowledge graph combined large model logic reasoning optimization method and system

Abstract

This invention provides a method and system for optimizing large-scale model logical reasoning by combining knowledge graphs, relating to the field of artificial intelligence technology. First, it acquires the logical reasoning task to be processed by the large model and its corresponding structured knowledge graph, which includes knowledge units, relationships, and relationship strength. The logical reasoning task includes reasoning objectives, premises, and constraints. Next, based on adaptability analysis, it obtains a set of reasoning guidance rules for the large model and injects them into the reasoning process. It captures and adjusts intermediate reasoning nodes in stages, correcting reasoning branches that deviate from the relationships. Then, it performs conflict resolution and propagation prediction on the reasoning process after stage guidance, identifies inconsistent conclusions, verifies the rationality of conflicts, predicts propagation stages, and generates multiple rounds of correction schemes. Finally, iteratively optimizes the reasoning path based on the reasoning process after conflict resolution, evaluates fit and efficiency, adjusts jump order and reference priority, integrates information to obtain the optimization result, and effectively improves the quality of large-scale model logical reasoning.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method and system for optimizing large-scale model logical reasoning by combining knowledge graphs. Background Technology

[0002] In an era of widespread application of large-scale models, logical reasoning ability is one of their core functions, playing a crucial role in many fields such as natural language processing, intelligent decision-making, and knowledge-based question answering. However, existing large-scale model logical reasoning processes have some problems that urgently need to be addressed.

[0003] On the one hand, large models often lack a precise grasp of the internal connections within a knowledge system when performing logical reasoning. Structured knowledge graphs contain rich knowledge units and the relationships and strengths between them, but traditional methods have failed to effectively utilize this information to guide the reasoning process, which may lead to deviations from the logic of the knowledge during the reasoning process, resulting in inaccurate and unreliable reasoning results.

[0004] On the other hand, during the reasoning process, when inference conclusions that are inconsistent with the knowledge graph occur, existing technologies lack effective conflict resolution mechanisms. They cannot promptly identify conflicts and reasonably verify the rationality of conflicting reasoning conclusions, nor can they predict the possible propagation of conflicts in the reasoning stages. Consequently, they cannot effectively correct the reasoning process, affecting the quality and efficiency of logical reasoning in large-scale models. Summary of the Invention

[0005] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, the present invention provides a method for optimizing large-scale model logical reasoning by incorporating knowledge graphs, the method comprising:

[0006] Obtain the logical reasoning task to be processed in the large model and the corresponding structured knowledge graph. The structured knowledge graph includes knowledge units, the relationships between knowledge units and the strength of the relationships. The logical reasoning task includes reasoning objectives, reasoning preconditions and reasoning constraints.

[0007] Based on the adaptability analysis of the structured knowledge graph and logical reasoning tasks, a set of large-scale model reasoning guidance rules is obtained.

[0008] The large model reasoning guidance rule set is injected into the large model reasoning process, intermediate reasoning nodes in the large model reasoning process are captured in stages, the jump direction of intermediate reasoning nodes is adjusted according to the guidance rules of different reasoning stages, and reasoning branches that deviate from the knowledge unit association relationship are corrected to obtain the large model reasoning process after stage guidance.

[0009] The large model reasoning process after the stage guidance is subjected to conflict resolution and propagation prediction. In the reasoning process, inconsistencies with the structured knowledge graph are identified. The relationship between knowledge units is called to verify the rationality of the conflict reasoning conclusions. The reasoning links where the conflict may propagate are predicted. Multiple conflict correction schemes are generated and applied to the reasoning process in turn to obtain the large model reasoning process after conflict resolution.

[0010] Based on the large model reasoning process after conflict resolution, the reasoning path is iteratively optimized, the fit between the reasoning path and the knowledge unit association and the reasoning efficiency are evaluated, the jump order of reasoning nodes and the reference priority of knowledge units are adjusted, and the optimized reasoning path, knowledge graph application records and conflict correction information are integrated to obtain the large model logical reasoning optimization results.

[0011] In another aspect, the present invention also provides a large model logic reasoning optimization system that combines knowledge graphs, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions or code, and the processor is used to execute the programs, instructions or code in the machine-readable storage medium to implement the above-mentioned method.

[0012] Based on the above, by acquiring the logical reasoning task to be processed in the large model and the corresponding structured knowledge graph, and then generating a set of reasoning guidance rules based on the adaptability analysis of the structured knowledge graph and the logical reasoning task, this set is injected into the reasoning process. Intermediate reasoning nodes are captured in stages, and the jump direction is adjusted. Reasoning branches that deviate from the relationships between knowledge units are corrected, ensuring the consistency between the reasoning direction and the knowledge logic from the source of reasoning, effectively improving the accuracy of reasoning. Conflict resolution and propagation prediction are performed on the large model reasoning process after stage guidance, accurately identifying reasoning conclusions inconsistent with the knowledge graph. By calling the relationships between knowledge units, the rationality of conflicts is verified and the propagation stage is predicted. Multi-round conflict correction schemes are generated and applied round by round, timely resolving conflicts in the reasoning process and ensuring the reliability of the reasoning results. Finally, based on the reasoning process after conflict resolution, the reasoning path is iteratively optimized, the fit and reasoning efficiency are evaluated, the jump order of reasoning nodes and the priority of knowledge unit references are adjusted, and relevant information is integrated to obtain optimized results, comprehensively improving the quality and efficiency of the large model's logical reasoning. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the execution flow of the large-scale model logical reasoning optimization method combining knowledge graphs provided in the embodiments of the present invention.

[0014] Figure 2 This is a schematic diagram of exemplary hardware and software components of the large-scale model logical reasoning optimization system that combines knowledge graphs, provided in an embodiment of the present invention. Detailed Implementation

[0015] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a large-scale logical reasoning optimization method combining knowledge graphs, provided in one embodiment of the present invention. The following is a detailed description of this large-scale logical reasoning optimization method combining knowledge graphs.

[0016] Step S110: Obtain the logical reasoning task to be processed in the large model and the corresponding structured knowledge graph. The structured knowledge graph includes knowledge units, the relationships between knowledge units and the strength of the relationships. The logical reasoning task includes reasoning objectives, reasoning preconditions and reasoning constraints.

[0017] In the subject-specific problem-solving scenarios within the field of smart education, the first step is to define the logical reasoning task that the large model needs to handle. This logical reasoning task represents a specific subject-specific problem encountered by students during their learning process, such as comprehensive mechanics problems in physics or geometric proof problems in mathematics. The reasoning objective refers to the final conclusion or result that the subject-specific problem needs to achieve, such as "determining the trajectory of an object under specific force conditions" or "proving the properties of a certain geometric figure." The reasoning premises are the known conditions upon which the reasoning is based; these conditions form the foundation for solving the problem and may include specific data given in the problem, existing theorems and laws, and relevant conceptual definitions. The reasoning constraints are the limitations or norms imposed on the reasoning process, such as "ignoring the influence of secondary factors," "using specified theorems for proof," and "the reasoning process must conform to the normative expression of the subject."

[0018] The corresponding structured knowledge graph is a structured representation of the knowledge system related to the discipline. Knowledge units are the basic elements constituting the knowledge graph, covering concepts, theorems, formulas, experimental phenomena, problem-solving methods, etc., within the discipline. The relationships between knowledge units reflect the inherent connections between different knowledge units, such as the inclusion relationship between concepts, the derivation relationship between theorems, the relationship between formulas, and the application relationship between concepts and theorems. The strength of the association measures the closeness or importance of the relationship between knowledge units. Its magnitude is determined by factors such as the logical importance and frequency of application of the association within the discipline's knowledge system. For example, the association strength between a theorem and the axioms directly derived from it is higher than the association strength between a theorem and indirectly related examples.

[0019] Step S120: Based on the adaptability analysis of the structured knowledge graph and the logical reasoning task, a set of large-scale model reasoning guidance rules is obtained.

[0020] After obtaining the logical reasoning task and the structured knowledge graph, a fit analysis is needed. The purpose of this analysis is to determine which knowledge units in the structured knowledge graph, and the relationships between them, are most relevant to the current logical reasoning task, and to what extent. Through this analysis, it becomes clear which knowledge units should be prioritized and which relationships should be followed during the reasoning process, thus providing effective guidance for the reasoning process of the large-scale model. Based on the results of the fit analysis, the above-mentioned guidelines and norms for guiding the large-scale model's reasoning are compiled into a set of large-scale model reasoning guidance rules.

[0021] Step S121: Perform reasoning task adaptability annotation on the knowledge units in the structured knowledge graph, analyze the degree of correlation between each knowledge unit and the reasoning target and reasoning premise in the logical reasoning task, mark the core adaptable knowledge units, auxiliary adaptable knowledge units and edge adaptable knowledge units, and obtain the knowledge unit adaptability annotation results.

[0022] To more accurately leverage structured knowledge graphs to assist in solving logical reasoning tasks, it is necessary to perform adaptability annotation on the knowledge units within the knowledge graph. This step involves analyzing the degree of correlation between each knowledge unit and the reasoning goal and premises in the logical reasoning task, categorizing knowledge units into different adaptability types. Core adaptability knowledge units are those directly related to the reasoning goal and play a decisive role in reaching the reasoning conclusion; auxiliary adaptability knowledge units supplement the core adaptability knowledge units, provide logical connections, or support intermediate derivation steps; peripheral adaptability knowledge units contribute to the refinement of the reasoning process and the supplementation of background knowledge, but are not essential to the reasoning process. Through this annotation, the role and importance of different knowledge units in the knowledge graph within the current reasoning task can be identified.

[0023] Step S1211: Extract the reasoning target keywords and reasoning premise keywords from the logical reasoning task, and construct a reasoning task keyword set. The reasoning target keywords contain the core attributes that need to be satisfied to achieve the reasoning conclusion, and the reasoning premise keywords contain the basic conditions that the reasoning process needs to be based on.

[0024] From the description of the logical reasoning task, accurately extract reasoning target keywords that reflect the core attributes of the reasoning objective and reasoning premise keywords that embody the basic conditions of the reasoning premises. Reasoning target keywords should directly point to the key characteristics of the final result to be achieved through reasoning. For example, in mathematical proof problems, reasoning target keywords might include the core concepts and relationships in the propositional conclusion to be proven. Reasoning premise keywords are the key identifiers of the known information upon which the reasoning process depends, such as the known data given in the problem and the core concepts in the known conditions. Organize the extracted reasoning target keywords and reasoning premise keywords, removing duplicates and irrelevant words, to construct a reasoning task keyword set. The keywords in this reasoning task keyword set will serve as the benchmark for subsequent knowledge unit matching, used to determine the degree of relevance between the knowledge unit and the reasoning task.

[0025] Step S1212: Perform keyword matching on each knowledge unit in the structured knowledge graph, count the number and matching depth of reasoning target keywords and reasoning premise keywords contained in the content of the knowledge unit, and obtain the keyword matching statistics of the knowledge unit.

[0026] Each knowledge unit in the structured knowledge graph is traversed, and the content of each knowledge unit is matched with keywords in the keyword set for the reasoning task. During the matching process, not only is the number of reasoning target keywords and reasoning premise keywords contained in the content of each knowledge unit counted, but the matching depth is also considered. Matching depth can be understood as the importance or relevance of keywords in the content of the knowledge unit; for example, keywords appearing in the core definition of a knowledge unit have a higher matching depth than those appearing in the extended application examples. For each knowledge unit, the number of reasoning target keywords, the number of reasoning premise keywords, and the matching depth of each keyword need to be recorded in detail. After summarizing and organizing the above information, the keyword matching statistics for the knowledge unit are obtained.

[0027] Step S1213: Calculate the fit between the knowledge unit and the reasoning task based on the keyword matching statistics of the knowledge unit. The fit calculation combines the number of keyword matches and the matching depth. The number of keyword matches is positively correlated with the fit. Set a fit classification standard. Mark the knowledge unit that directly supports the realization of the reasoning goal as the core fit knowledge unit, mark the knowledge unit that supplements the function of the core fit knowledge unit as the auxiliary fit knowledge unit, and mark the knowledge unit that improves the details of the reasoning process as the edge fit knowledge unit.

[0028] Step S1213-1: Set keyword matching weight parameters. The matching weight parameter of the inference target keyword is higher than the matching weight parameter of the inference premise keyword. The weight parameter of exact match is higher than the weight parameter of semantic similarity match. The weight parameter of exact match of the inference target keyword is set according to the first association level. The weight parameter of semantic similarity match of the inference target keyword is set according to the second association level. The weight parameter of exact match of the inference premise keyword is set according to the third association level. The weight parameter of semantic similarity match of the inference premise keyword is set according to the fourth association level.

[0029] To accurately calculate the fit between knowledge units and reasoning tasks, reasonable keyword matching weight parameters need to be set. Since the target keywords are more crucial for achieving the reasoning conclusion, their matching weight parameters should be higher than those of the premise keywords. Furthermore, different keyword matching methods have different levels of importance; exact matching more directly reflects the association between knowledge units and keywords, therefore its weight parameter is higher than that of semantic similarity matching. Specifically, exact matching of target keywords is set as the first association level and assigned a corresponding weight parameter; semantic similarity matching of target keywords is set as the second association level, with a lower weight parameter than the first association level; exact matching of premise keywords is set as the third association level, with a lower weight parameter than the second association level; and semantic similarity matching of premise keywords is set as the fourth association level, with the lowest weight parameter. These weight parameters will be used in subsequent fit calculations to reflect the different impacts of different types of keyword matching on the fit of knowledge units.

[0030] Step S1213-2: Perform a weighted calculation on the keyword matching statistics of each knowledge unit, and count the number and average matching depth of each matching type in the knowledge unit. The matching types include exact matching of the reasoning target, semantic similarity matching of the reasoning target, exact matching of the reasoning premise, and semantic similarity matching of the reasoning premise. Multiply the number of each matching type by the average matching depth, then multiply by the corresponding weight parameter and sum them to obtain the initial fit of the knowledge unit.

[0031] For each knowledge unit, based on its keyword matching statistics, different matching types are distinguished: exact match of the reasoning target, semantically similar match of the reasoning target, exact match of the reasoning premise, and semantically similar match of the reasoning premise. For each matching type, the number of keywords matching that type in the knowledge unit and the average matching depth of these keywords are counted. Then, the number of each matching type is multiplied by the average matching depth to obtain the weighted number of that matching type. This weighted number is then multiplied by the corresponding weight parameter for that matching type to obtain the contribution value of that matching type to the fitness score. Finally, the contribution values ​​of all matching types are summed to obtain the initial fitness score of the knowledge unit. The magnitude of the initial fitness score reflects the initial degree of association between the knowledge unit and the reasoning task.

[0032] Step S1213-3: Normalize the initial fit degree, map the fit degree value to a preset value range, eliminate the influence of the difference in content length of different knowledge units on the fit degree, and obtain the normalized fit degree of the knowledge unit.

[0033] Because the content length of different knowledge units may vary, longer knowledge units may contain more keywords, leading to a higher initial fit score. This can affect the objectivity and accuracy of the fit score. Therefore, it is necessary to normalize the initial fit score. The purpose of normalization is to map the initial fit score value to a preset numerical range, such as [0,1]. Through normalization, the influence of differences in the content length of knowledge units on the fit score can be eliminated, making the fit scores of different knowledge units comparable. Common normalization methods such as max-min normalization can be used. The appropriate method should be selected based on the actual situation. After processing, the normalized fit score of the knowledge unit is obtained.

[0034] Step S1213-4: Set the adaptation classification criteria. Taking into account the complexity of the reasoning task and the scale of the knowledge graph, determine the basis for dividing core adapted knowledge units, auxiliary adapted knowledge units, and marginal adapted knowledge units. Knowledge units that directly support the realization of the reasoning goal are classified as core adapted knowledge units. Knowledge units that supplement the functions of core adapted knowledge units are classified as auxiliary adapted knowledge units. Knowledge units that improve the details of the reasoning process are classified as marginal adapted knowledge units. The remaining knowledge units are classified as non-adapted knowledge units. Non-adapted knowledge units are excluded from the scope of reference in the reasoning process.

[0035] Considering the complexity of current logical reasoning tasks and the overall scale of structured knowledge graphs, a classification standard for fit is established. The formulation of this standard requires clarifying the criteria for dividing core fit knowledge units, auxiliary fit knowledge units, marginal fit knowledge units, and non-fit knowledge units. Core fit knowledge units refer to those with the highest normalized fit, directly supporting the achievement of the reasoning goal; auxiliary fit knowledge units have a slightly lower normalized fit, primarily supplementing the functions of core fit knowledge units and facilitating the smooth progress of the reasoning process; marginal fit knowledge units have a lower normalized fit, mainly used to refine the details of the reasoning process, but are not essential for reasoning; and non-fit knowledge units have the lowest normalized fit, with extremely low or no relevance to the reasoning task, and are excluded from the scope of reference in the reasoning process. The specific thresholds for this classification need to be set and adjusted according to the actual needs of the reasoning task and the characteristics of the knowledge graph to ensure the accuracy and rationality of the classification.

[0036] Step S1213-5: Compare the normalized fit of each knowledge unit with the fit classification standard to determine the fit type of each knowledge unit. Record the normalized fit, the comparison results of the classification standard, and the fit type label of each knowledge unit to form the preliminary results of knowledge unit fit labeling.

[0037] The normalized fit of each knowledge unit is compared with the established fit classification criteria, and the fit type of the knowledge unit is determined based on the range of the normalized fit. For example, if the normalized fit is greater than or equal to the core fit threshold, it is marked as a core fit knowledge unit; if the normalized fit is less than the core fit threshold but greater than or equal to the auxiliary fit threshold, it is marked as an auxiliary fit knowledge unit, and so on. Simultaneously, the specific value of the normalized fit for each knowledge unit, the detailed results of the comparison with the classification criteria, and the final fit type label are recorded. This information is then organized according to knowledge units to form preliminary results of knowledge unit fit labeling.

[0038] Step S1214: Perform secondary verification on the marked knowledge units, analyze whether the core adapted knowledge units directly support the realization of the reasoning goal, whether the auxiliary adapted knowledge units supplement or logically connect the core adapted knowledge units, and whether the edge adapted knowledge units help to improve the details of the reasoning process. Remove knowledge units that do not conform to the adaptation type definition and re-label them.

[0039] To ensure the accuracy of the initial results of knowledge unit adaptability labeling, secondary verification of the labeled knowledge units is necessary. For those labeled as core adaptable knowledge units, verification is needed to confirm whether they truly and directly support the achievement of the reasoning objective and are indispensable key knowledge units in the reasoning process. For auxiliary adaptable knowledge units, verification is needed to confirm whether they can supplement the core adaptable knowledge units, provide logical connections, or offer necessary intermediate steps. For peripheral adaptable knowledge units, verification is needed to confirm whether they provide practical help in refining the details of the reasoning process and supplementing background information. During the verification process, if the adaptability type label of a knowledge unit is found to be inconsistent with the actual situation, i.e., it does not conform to the definition of that adaptability type, it needs to be removed, and the adaptability type should be re-judged and labeled according to its actual situation. Through secondary verification, the reliability of the knowledge unit adaptability labeling results can be further improved.

[0040] Step S1215: Record the adaptation type, adaptation value, matched keywords and secondary verification results of each knowledge unit, classify and organize them according to the adaptation type, and obtain the knowledge unit adaptation labeling results.

[0041] After secondary verification and re-annotation, detailed information for each knowledge unit is recorded. This includes the knowledge unit's unique identifier, adaptation type (core adaptation, auxiliary adaptation, edge adaptation), normalized adaptation score, matched reasoning task keywords (including keyword type and matching depth), and secondary verification results (e.g., verification passed, verification failed, and re-annotation status). Then, the knowledge units are categorized according to their adaptation type, grouping units of the same type together to form the final knowledge unit adaptation annotation result. This result demonstrates the adaptation status of each knowledge unit in the knowledge graph to the current logical reasoning task.

[0042] Step S122: Based on the knowledge unit adaptability annotation results, analyze the relationship between knowledge units, identify the supporting relationship between core adaptability knowledge units and auxiliary adaptability knowledge units, the supplementary relationship between auxiliary adaptability knowledge units and edge adaptability knowledge units, and the indirect relationship between core adaptability knowledge units and edge adaptability knowledge units, and construct a knowledge unit relationship network.

[0043] Based on the knowledge unit adaptability annotation results, the relationships between knowledge units are analyzed in depth. The focus is on identifying specific relationships between knowledge units of different adaptability types: supporting relationships between core adaptability knowledge units and auxiliary adaptability knowledge units, where auxiliary adaptability knowledge units provide support for the application, understanding, or derivation of core adaptability knowledge units; supplementary relationships between auxiliary adaptability knowledge units and peripheral adaptability knowledge units, where peripheral adaptability knowledge units supplement or refine the content of auxiliary adaptability knowledge units; and indirect relationships between core adaptability knowledge units and peripheral adaptability knowledge units, which are indirect relationships established between them through intermediate knowledge units such as auxiliary adaptability knowledge units. By identifying these relationships, the various knowledge units are connected according to their relationships, constructing a structured knowledge unit relationship network. This knowledge unit relationship network can intuitively demonstrate the interrelationships between knowledge units and the position and role of different adaptability types of knowledge units within the network.

[0044] Step S123: Determine the priority direction of the reasoning path based on the knowledge unit association network, take the core adaptive knowledge unit as the starting node of reasoning, prioritize the auxiliary adaptive knowledge unit corresponding to the supporting association as the next hop reasoning node, and select the edge adaptive knowledge unit based on the supplementary association to improve the reasoning chain, forming a phased reasoning path priority order table.

[0045] Based on the constructed knowledge unit relationship network, the priority direction of the reasoning path is determined. The starting node of the reasoning path is set as the core adapted knowledge unit, because the core adapted knowledge unit is directly related to the reasoning goal and is the foundation and key to the reasoning. Starting from the core adapted knowledge unit, when selecting the next hop reasoning node, priority is given to auxiliary adapted knowledge units that have a supporting relationship with the current core adapted knowledge unit, because these auxiliary adapted knowledge units can provide direct support for the application of the core adapted knowledge unit. During the extension of the reasoning chain, based on the supplementary relationship between auxiliary adapted knowledge units and edge adapted knowledge units, edge adapted knowledge units are selected in a timely manner to improve the details of the reasoning chain, making the reasoning process more complete and rigorous. According to the characteristics and needs of different reasoning stages, the priority selection order of the above reasoning paths is organized to form a phased reasoning path priority order table. This phased reasoning path priority order table clarifies which knowledge units and relationships should be prioritized for reasoning jumps at different stages of the reasoning process, so as to improve the efficiency and accuracy of reasoning.

[0046] Step S124: Extract the correlation strength between knowledge units in the structured knowledge graph, and dynamically set the jump threshold of the reasoning node based on the adaptation labeling results of the knowledge unit and the reasoning task. The jump threshold between core adapted knowledge units is set according to the first priority, the jump threshold between auxiliary adapted knowledge units and core adapted knowledge units is set according to the second priority, and the jump threshold between edge adapted knowledge units and other knowledge units is set according to the third priority. Record the dynamic threshold corresponding to each correlation to obtain the staged reasoning node jump threshold table.

[0047] The association strength information between knowledge units is extracted from the structured knowledge graph. Association strength reflects the closeness or importance of the relationship between two knowledge units. Based on the adaptation annotation results between knowledge units and the reasoning task, the jump threshold of reasoning nodes is dynamically set. The jump threshold is the critical value for determining whether to allow a jump from one reasoning node to another. Different jump thresholds are set according to the adaptation type of the knowledge unit and the priority of the association: the association between core adapted knowledge units is the most important, so a high jump threshold is set according to the first priority; the supporting association between auxiliary adapted knowledge units and core adapted knowledge units is the next most important, so a medium jump threshold is set according to the second priority; the association between edge adapted knowledge units and other knowledge units is less important, so a low jump threshold is set according to the third priority. For each type of association, a specific jump threshold needs to be dynamically determined based on its association strength and corresponding priority, and recorded to form a phased reasoning node jump threshold table. This phased reasoning node jump threshold table is used to determine whether the currently planned reasoning node jump is reasonable and meets the association strength requirements during the reasoning process.

[0048] Step S125: The reasoning constraints are layered according to the reasoning stage. The initial reasoning stage constraints focus on the compliance of the reasoning premises, the intermediate reasoning stage constraints focus on the rationality of the knowledge unit references, and the final reasoning stage constraints focus on the fit between the reasoning conclusion and the goal. Each layer of constraints corresponds to the adaptation requirements of the reasoning steps, including the stage prohibited association list, the range of the number of stage steps, and the stage conclusion attribute standards.

[0049] In logical reasoning tasks, reasoning constraints are layered according to the reasoning stages. Different reasoning stages have different characteristics and objectives, therefore the focus of the constraints also differs. The initial reasoning stage marks the beginning of reasoning; constraints should focus on the compliance of the reasoning premises, i.e., whether the reasoning process strictly adheres to the given premises and meets the constraints of the premises. The intermediate reasoning stage extends and develops the reasoning chain; constraints should focus on the rationality of knowledge unit citations, i.e., whether the cited knowledge units are relevant to the current reasoning step, whether they conform to the adaptation annotation results, and whether the citation method is correct. The final reasoning stage is where the reasoning conclusion is reached; constraints should focus on the fit between the reasoning conclusion and the reasoning objective, i.e., whether the reasoning conclusion meets the core attribute requirements of the reasoning objective and whether the reasoning objective is accurately and completely achieved. For each layer of constraints, corresponding inference step adaptation requirements are formulated, including a prohibited association list for each stage (clearly defining the associations of knowledge units that are not allowed to be referenced in this stage), a range of steps for each stage (specifying the approximate range of the number of inference steps in this stage to avoid too many or too few steps), and attribute standards for the conclusions of each stage (clarifying the attribute characteristics that the intermediate conclusions of the inference in this stage should possess), so as to standardize the inference process and ensure that the inference proceeds in the correct direction.

[0050] Step S126: Integrate the phased inference path priority order table, the phased inference node jump threshold table, and the hierarchical inference step adaptation requirements, divide the rule entries according to the inference stage, and the rule entries of each inference stage include path priority, dynamic jump threshold and hierarchical adaptation requirements to obtain the large model inference guidance rule set.

[0051] The previously obtained phased inference path priority list, phased inference node jump threshold list, and hierarchical inference step adaptation requirements are integrated. According to the division of inference stages, the above information is organized into rule entries. Each rule entry for an inference stage includes path priority information for that stage (from the phased inference path priority list), dynamic jump thresholds (from the phased inference node jump threshold list), and hierarchical adaptation requirements (from the hierarchical inference step adaptation requirements). Through this integration, a complete, phased large-scale model inference guidance rule set is formed. This large-scale model inference guidance rule set provides clear inference guidance for the large model at different inference stages, guiding the large model on how to select inference paths, judge the rationality of node jumps, and follow which inference step requirements, thereby effectively guiding the large model's inference process and improving the accuracy and reliability of inference.

[0052] Step S130: Inject the large model reasoning guidance rule set into the large model reasoning process, capture intermediate reasoning nodes in the large model reasoning process in stages, adjust the jump direction of intermediate reasoning nodes according to the guidance rules of different reasoning stages, correct reasoning branches that deviate from the knowledge unit association relationship, and obtain the large model reasoning process after stage guidance.

[0053] A pre-constructed set of large-scale model reasoning guidance rules is injected into the large-scale model's reasoning process, enabling the model to be guided and constrained by these rules. During the large-scale model's reasoning process, corresponding mechanisms are set up according to the reasoning stage to capture intermediate reasoning nodes. These intermediate reasoning nodes are key steps and intermediate results in the reasoning process; capturing these nodes allows for real-time monitoring of the reasoning progress and direction. For each captured intermediate reasoning node, the jump direction is checked against the guidance rules for the corresponding reasoning stage in the large-scale model's reasoning guidance rule set to ensure it conforms to the rule requirements. If the jump direction deviates from the knowledge unit association relationship or does not comply with rules such as path priority or dynamic jump threshold, adjustments and corrections are made promptly, changing the direction of the reasoning branch to bring it back to the correct reasoning track. Through phased capture, comparison, and correction, the final large-scale model reasoning process, guided by these stages, is obtained. This process better follows the knowledge structure and reasoning rules of the knowledge graph, improving the effectiveness of the reasoning.

[0054] Step S131: During the large model inference process, set the inference node capture window according to the inference stage. In the initial inference stage, capture the intermediate inference node once every fixed inference step. In the intermediate inference stage, capture the intermediate inference node once every dynamic inference step. In the final inference stage, capture the intermediate inference node in real time and record the content, generation time, unique identifier of the preceding inference node and the inference stage to which each intermediate inference node belongs, to obtain the set of intermediate inference nodes in each stage.

[0055] During the reasoning process of the large model, corresponding reasoning node capture windows are set according to different reasoning stages. In the initial reasoning stage, since the reasoning process is relatively stable and the reasoning steps and rhythm are relatively fixed, intermediate reasoning nodes can be captured once every preset fixed reasoning step interval. In the intermediate reasoning stage, the reasoning process may have multiple possibilities, and the number and rhythm of reasoning steps may change. Therefore, a preset dynamic reasoning step interval is used to capture intermediate reasoning nodes. The dynamic interval can be adjusted according to factors such as the complexity of the reasoning and the importance of the current reasoning node. In the final reasoning stage, to ensure the accuracy and reliability of the reasoning conclusion, intermediate reasoning nodes need to be captured in real time to promptly identify and address potential problems. For each captured intermediate reasoning node, its specific content, generation timestamp, unique identifier of the preceding reasoning node (used to trace the reasoning path), and the reasoning stage to which the node belongs are recorded in detail. This information is categorized and organized according to the reasoning stage to form a set of intermediate reasoning nodes for each stage.

[0056] Step S132: Match the intermediate inference nodes of each inference stage in the set of intermediate inference nodes in the phased process with the inference path priority order table of the corresponding stage in the set of large model inference guidance rules, identify the next-hop priority knowledge unit type corresponding to the current intermediate inference node, the priority knowledge unit type includes core adaptation, auxiliary adaptation and edge adaptation, determine the reasonable inference jump direction for the inference stage, and obtain the phased node jump direction suggestion.

[0057] For each intermediate inference node in the set of intermediate inference nodes in each inference stage, a matching process is performed with the inference path priority order table for the corresponding stage in the large model inference guidance rule set. Through matching, the type of the next priority knowledge unit that the current intermediate inference node should select in that inference stage, based on path priority, is identified. Priority knowledge unit types include core adaptive knowledge units, auxiliary adaptive knowledge units, and edge adaptive knowledge units; their specific priority order is already specified in the inference path priority order table. Based on the identified priority knowledge unit type, a reasonable inference jump direction for that inference stage is determined, i.e., which type of adaptive knowledge unit should be prioritized for the next jump.

[0058] Step S133: Compare the next-hop inference node currently planned by the large model with the priority knowledge unit type in the proposed jump direction of the stage node. If they are consistent, retain the current inference path. If they are inconsistent, extract the corresponding dynamic threshold from the jump threshold table of the inference node in the corresponding inference stage, and determine whether the association strength from the current inference node to the planned next-hop node meets the threshold requirements of the inference stage.

[0059] The knowledge unit type of the next-hop inference node currently planned by the large model is compared with the priority knowledge unit type in the phased node jump direction suggestions. If the two types match, it means that the current inference path planned by the large model conforms to the priority direction of the guidance rules, so the current inference path is retained. If the two types do not match, it is necessary to further determine whether the inconsistency is reasonable. At this time, the dynamic threshold corresponding to the association relationship between the current inference node and the planned next-hop node is extracted from the inference node jump threshold table of the corresponding inference stage, and then the actual association strength between the current inference node and the planned next-hop node is compared with the size of the dynamic threshold. If the actual association strength meets or exceeds the dynamic threshold, it means that although the planned next-hop node type is not a priority type, it may still have a certain degree of rationality due to its sufficiently high association strength, and further analysis is needed; if the actual association strength does not meet the dynamic threshold, it means that the planned next-hop node does not meet the requirements and needs to be corrected.

[0060] Step S134: If the correlation strength between the current inference node and the planned next hop node meets the threshold requirement of this inference stage, then generate a stage path adjustment suggestion, explaining the functional characteristics of the priority knowledge unit type in the current inference stage, its correlation with the inference goal, and its supporting role in subsequent inference stages, guiding the large model to adjust the jump direction to the priority knowledge unit.

[0061] When it is determined that the correlation strength between the current inference node and the planned next-hop node meets the dynamic threshold requirement of this inference stage, even if the planned next-hop node type is not a priority knowledge unit type, a stage path adjustment suggestion needs to be generated to better follow the inference guidance rules and improve the efficiency and accuracy of inference. This stage path adjustment suggestion details the functional characteristics of the priority knowledge unit type in the current inference stage, such as the key supporting role of core adaptive knowledge units and the supplementary connecting role of auxiliary adaptive knowledge units; it elucidates the close correlation between the priority knowledge unit type and the inference goal, and the supporting role and positive impact of selecting the priority knowledge unit type as the next-hop node on subsequent inference stages. Through these explanations, the large model is guided to recognize the advantages and importance of the priority knowledge unit type, thereby proactively adjusting the jump direction of the inference node to align with the priority knowledge unit type.

[0062] Step S1341: Extract the type of priority knowledge unit corresponding to the current reasoning stage in the large model reasoning guidance rule set, analyze the functional characteristics of this type of knowledge unit in the current reasoning stage, summarize the functional differences of this type of knowledge unit compared with the type of the next hop node, and obtain the stage priority type advantage analysis results. The priority knowledge unit type includes core adaptation knowledge unit type, auxiliary adaptation knowledge unit type and edge adaptation knowledge unit type. The core adaptation knowledge unit type has logical rigor, the auxiliary adaptation knowledge unit type has smooth connection, and the edge adaptation knowledge unit type has detailed completeness.

[0063] The priority knowledge unit type corresponding to the current inference stage is extracted from the large model's inference guidance rule set. For each priority knowledge unit type, its functional characteristics in the current inference stage are analyzed in depth. For example, the core adaptation knowledge unit type has logical rigor; the auxiliary adaptation knowledge unit type has smooth transitions, enabling natural transitions between different stages of the inference process; and the edge adaptation knowledge unit type has detailed completeness, supplementing detailed information in the inference process. Simultaneously, this priority knowledge unit type is compared with the knowledge unit type of the next-hop node currently planned by the large model, summarizing the functional differences between the two and highlighting the advantages of the priority knowledge unit type. Through the above analysis, the results of the stage priority type advantage analysis are formed.

[0064] Step S1342: Combining the reasoning objectives in the logical reasoning task, the role of priority knowledge unit types in promoting the achievement of reasoning objectives is explained in stages, and the results of the stage objective correlation analysis are obtained. The initial reasoning stage explains the fit between priority knowledge unit types and reasoning premises, the intermediate reasoning stage explains the support of priority knowledge unit types for the extension of the reasoning chain, and the final reasoning stage explains the role of priority knowledge unit types in ensuring the accuracy of reasoning conclusions.

[0065] Closely integrating with the reasoning objectives of logical reasoning tasks, this paper elucidates the specific role of preferred knowledge unit types in achieving these objectives across different reasoning stages. In the initial reasoning stage, the focus is on the fit between preferred knowledge units and the reasoning premises—that is, how preferred knowledge units accurately develop reasoning based on the premises. In the intermediate reasoning stage, the emphasis is on the support provided by preferred knowledge units for extending the reasoning chain—that is, how the selection and application of preferred knowledge units ensures the continuous and effective extension of the reasoning chain, preventing interruptions or deviations from the intended direction. In the final reasoning stage, the paper emphasizes the role of preferred knowledge units in guaranteeing the accuracy of the reasoning conclusion—that is, how preferred knowledge units provide crucial support for arriving at accurate and reliable reasoning conclusions. Through this phased explanation, the paper presents a correlation analysis of the stage objectives, further demonstrating the necessity of selecting preferred knowledge unit types.

[0066] Step S1343: Based on the results of the stage priority type advantage analysis and the stage goal association analysis, generate a stage path adjustment suggestion document. The stage path adjustment suggestion document includes current inference stage information, current inference node information, next hop planning node information, priority knowledge unit type information, stage advantage comparison, stage goal association description and expected effect of the adjusted inference path, determine the suggested adjustment jump direction and form a target record.

[0067] Based on the combined results of the priority type advantage analysis and the correlation analysis of stage goals, a formal stage path adjustment suggestion document is generated. This document contains rich information, including the current inference stage, detailed content and identifiers of the current inference node, detailed content and identifiers of the planned next-hop node, specific information on the priority knowledge unit type, a comparison of stage advantages (i.e., a comparison of the functional advantages of the priority knowledge unit type and the type of the planned next-hop node), an explanation of the correlation with stage goals (i.e., how the priority knowledge unit type drives the achievement of inference goals), and the expected effects of the adjusted inference path (such as improved inference efficiency and accuracy). The document clearly defines the jump direction of the suggested adjustment and records the above information as goals so that the large model can clearly understand the content and basis of the adjustment suggestions.

[0068] Step S1344: Input the stage path adjustment suggestion document into the large model inference interaction interface to trigger the stage adaptability evaluation process of the large model for the path adjustment suggestion. Call the large model to re-evaluate the rationality of the jump direction by combining the task focus of the current inference stage, its own inference logic and suggestion content.

[0069] The generated stage path adjustment suggestion document is input into the large model through its inference interaction interface. This input operation triggers the stage adaptability evaluation process of the path adjustment suggestion within the large model. In this process, the large model will re-evaluate and judge the suggested jump direction based on the task focus of the current inference stage (e.g., the initial stage focuses on premise fit, the intermediate stage focuses on chain extension, etc.), its existing inference logic, and the content of the stage path adjustment suggestion document. The large model will analyze the rationality, feasibility, and compatibility of the suggestion with its own inference approach to decide whether to accept the adjustment suggestion.

[0070] Step S1345: If the large model is detected to accept the stage path adjustment suggestion, the next hop inference node is adjusted to a specific knowledge unit in the priority knowledge unit type, and the adjustment basis, suggestion reference information, and stage adaptability evaluation results are marked in the inference process record; if the large model is detected to temporarily not accept the suggestion, a suggestion feedback description is generated, the reasons for non-acceptance mentioned in the suggestion feedback description are analyzed, additional functional evidence and logical connection schemes for the priority knowledge unit type of the inference stage are supplemented, and the stage path adjustment suggestion is submitted again until the large model completes the jump direction adjustment. The reasons for non-acceptance include insufficient stage adaptability or logical connection problems.

[0071] The feedback mechanism of the large model is used to detect whether the large model accepts the stage path adjustment suggestion. If the large model accepts the suggestion, the next hop inference node is adjusted to a specific knowledge unit in the priority knowledge unit type according to the suggestion. At the same time, the basis for this adjustment (i.e., the key content in the stage path adjustment suggestion document), the suggested information cited, and the large model's evaluation results of stage adaptability are marked in detail in the large model's inference process record for subsequent traceability and analysis. If the large model does not accept the suggestion, it will generate a suggestion feedback explanation, which includes the specific reasons for not accepting the suggestion, such as insufficient stage adaptability or problems in logical connection. At this time, it is necessary to conduct an in-depth analysis of the reasons for not accepting the suggestion feedback explanation, and supplement the above reasons with additional functional evidence of the priority knowledge unit type of the inference stage, such as more application cases, more in-depth logical analysis, etc., and provide a more complete logical connection solution to solve the logical connection problems raised by the large model. Then, the supplemented and improved stage path adjustment suggestion is submitted to the large model again, and the above detection and adjustment process is repeated until the large model finally accepts the suggestion and completes the adjustment of the jump direction.

[0072] Step S135: If the association strength from the current inference node to the planned next-hop node does not meet the threshold requirement of this inference stage, then directly correct the inference branch, replace the planned next-hop node with a specific knowledge unit in the priority knowledge unit type, and update the node association record and stage inference log in the inference process.

[0073] When it is determined that the association strength between the current inference node and the planned next-hop node does not meet the dynamic threshold requirements of this inference stage, it indicates that the planned next-hop node does not meet the basic requirements of the inference guidance rules, and the inference branch needs to be directly corrected. The correction method involves replacing the planned next-hop node with a specific knowledge unit from the priority knowledge unit type. These specific knowledge units are determined based on the phased inference path priority order table and the knowledge unit association network, and are closely related to the current inference node and meet the priority type requirements. After the replacement, the node association records in the large model inference process are updated promptly, recording the new node associations and jump information. Simultaneously, the phase inference log is updated, detailing the operation, reason, and corrected inference node status to ensure the traceability of the inference process.

[0074] Step S136: After completing the capture, matching, comparison and correction of intermediate inference nodes in the current inference stage, proceed to the next inference stage. For the next inference node, re-execute the entire process from setting the inference node capture window to generating node jump direction suggestions, performing jump comparison, and adjusting or directly correcting the path based on the comparison results. Repeat this process until all inference stages have been processed. Record the path adjustment record, node correction information and stage inference effect for each inference stage to obtain the large model inference process after stage guidance.

[0075] After capturing all intermediate inference nodes in the current inference stage, matching them with the guidance rules, comparing jump directions, and performing necessary corrections, the large model's inference process enters the next inference stage. For the starting inference node of the next inference stage (i.e., the final node after correction in the previous stage), a series of operations starting from setting the inference node capture window are re-executed: setting the capture window parameters for this stage, capturing intermediate inference nodes, matching intermediate inference nodes with the guidance rules for this stage to generate node jump direction suggestions, comparing the next jump node planned by the large model with the suggested direction, and adjusting the path based on the comparison results (including sub-steps such as generating suggestions and guidance adjustments) or directly correcting the inference branches. This process is repeated in sequence, processing each stage according to the order of the inference stages, until all inference stages are successfully completed. Throughout the process, detailed records are kept of the path adjustment records for each inference stage (including the number of adjustments, reasons, and content), node correction information (corrected nodes, comparison before and after correction, and basis for correction), and stage inference effects (inference progress, accuracy, and proximity to the target). Finally, this information is integrated to obtain the large model's inference process after stage guidance.

[0076] Step S140: Perform conflict resolution and propagation prediction on the large model reasoning process after the stage guidance, identify reasoning conclusions that are inconsistent with the structured knowledge graph during the reasoning process, call the correlation between knowledge units to verify the rationality of conflict reasoning conclusions, predict the reasoning links where conflicts may propagate, generate multiple rounds of conflict correction schemes and apply them to the reasoning process round by round to obtain the large model reasoning process after conflict resolution.

[0077] Further optimization of the large-scale model reasoning process after phased guidance is performed, mainly in two aspects: conflict resolution and propagation prediction. First, inconsistencies between the inference conclusions and the structured knowledge graph generated during the large-scale model reasoning process after phased guidance are identified; these inconsistent conclusions are conflicting inference conclusions. Then, the relationships between knowledge units in the structured knowledge graph are invoked to verify the rationality of the conflicting inference conclusions, analyzing the causes and essence of the conflicts. Simultaneously, based on the knowledge unit relationship network, the potential propagation of the conflicting inference conclusions to subsequent reasoning stages is predicted, marking the conflict propagation inference nodes requiring key attention. Based on the causes of the conflicts and the propagation prediction results, multiple rounds of conflict correction schemes are generated, correcting the conflicts from different levels and angles. These multiple rounds of conflict correction schemes are applied sequentially to the large-scale model reasoning process. After each round of application, the reasoning process is re-examined and verified to ensure effective conflict resolution and the absence of new conflicts. Through these multiple rounds of correction, the final large-scale model reasoning process after conflict resolution is obtained. The reasoning conclusions in this conflict-resolved large-scale model reasoning process are consistent with the knowledge system of the knowledge graph, making the reasoning process more reliable.

[0078] Step S141: Extract all reasoning conclusions in the large model reasoning process after the stage guidance, classify and organize them into initial stage reasoning conclusions, intermediate stage reasoning conclusions and final stage reasoning conclusions according to the reasoning stage, record the generation basis, associated knowledge unit and to which the reasoning link belongs for each reasoning conclusion, and obtain a set of staged reasoning conclusions.

[0079] All reasoning conclusions are extracted from the large model reasoning process after stage-guided reasoning, including intermediate and final conclusions generated at each reasoning stage. These conclusions are categorized into initial stage reasoning conclusions, intermediate stage reasoning conclusions, and final stage reasoning conclusions according to the reasoning stages. For each conclusion, the basis for its generation is recorded in detail, including the preconditions, knowledge units used, and reasoning rules applied; the associated knowledge units are recorded, clarifying which knowledge units in the knowledge graph the conclusion is related to; and the specific reasoning stage to which the conclusion belongs, i.e., at which step or node of the reasoning process it was generated. The categorized and organized reasoning conclusions and their related information are then summarized to form a set of staged reasoning conclusions.

[0080] Step S142: Compare each reasoning conclusion in the set of phased reasoning conclusions with the knowledge units and relationships in the structured knowledge graph, identify reasoning conclusions that contradict the content of the knowledge units or conflict with the relationships, mark them as conflicting reasoning conclusions, and organize them into a set of phased conflicting reasoning conclusions according to the reasoning stage in which the conflict occurred.

[0081] Each reasoning conclusion in the set of staged reasoning conclusions is compared one by one with the structured knowledge graph. The comparison includes two aspects: first, whether there are contradictions between the reasoning conclusion and the content of the knowledge unit, such as whether the reasoning conclusion contradicts the definitions, theorems, or facts in the knowledge unit; second, whether there are conflicts between the reasoning conclusion and the relationships between knowledge units, such as whether the relationships between knowledge units implied by the reasoning conclusion are inconsistent with the explicit relationships in the knowledge graph. Through comparison, all contradictory or conflicting reasoning conclusions are identified and marked as conflicting reasoning conclusions. Then, they are categorized and organized according to the reasoning stage in which the conflicting reasoning conclusions arose, forming a set of staged conflicting reasoning conclusions. This set of staged conflicting reasoning conclusions demonstrates the conflict issues that arise at different reasoning stages, facilitating subsequent targeted conflict tracing and resolution.

[0082] Step S143: Trace the source of each conflicting reasoning conclusion in the set of phased conflicting reasoning conclusions, find the reasoning steps that generated the conflicting reasoning conclusion, the intermediate reasoning nodes that depend on it, the knowledge units that are referenced and the related reasoning rules, determine the specific reasoning link that caused the conflict and the corresponding error cause, and obtain the phased conflict tracing result. The error cause includes reasoning logic error or knowledge reference deviation.

[0083] A thorough source analysis is conducted on each conflicting reasoning conclusion in the set of staged conflicting reasoning conclusions. The generation process of each conflicting conclusion is traced to identify the specific reasoning step in which it arose, the intermediate reasoning nodes upon which that step relied, and the knowledge units and related reasoning rules referenced during the reasoning process. Through the analysis of these factors, the specific reasoning step in which the conflict arose is determined, i.e., at which step of the reasoning process the problem occurred. Simultaneously, the type of error causing the conflict is identified, primarily including logical errors in reasoning (such as invalid logical relationships between reasoning steps, violations of logical reasoning rules, etc.) and knowledge reference biases (such as referencing incorrect knowledge units, incorrect understanding and application of knowledge units, incorrect reference to relationships between knowledge units, etc.). The source information and error causes of each conflicting reasoning conclusion are recorded to form the staged conflict source tracing results.

[0084] Step S144: Call the knowledge unit association relationships related to the conflict reasoning conclusion in the structured knowledge graph, verify the contradictions between the conflict reasoning conclusion and the association relationships, analyze the essential reasons for the contradictions, predict the subsequent reasoning links that the conflict may propagate in combination with the knowledge unit association relationship network, mark the key reasoning nodes for conflict propagation, and obtain the conflict propagation prediction results.

[0085] By invoking the relationships between knowledge units in a structured knowledge graph that are associated with the conflicting reasoning conclusion, the contradictions between the conclusion and these relationships are verified in detail. The root causes of the contradictions are analyzed in depth: are they due to misunderstandings or misapplications of the relationships, or confusion caused by the complexity of the relationships themselves? Based on clarifying the root causes of the contradictions, and considering the structure and characteristics of the knowledge unit relationship network, the potential propagation of the conflicting reasoning conclusion to subsequent reasoning stages if it is not corrected is predicted. Based on the likelihood and extent of propagation, key reasoning nodes in the conflict propagation path are marked; these nodes are critical and require close monitoring.

[0086] Step S145: Based on the phased conflict tracing results and conflict propagation prediction results, generate multiple rounds of conflict correction schemes. The first round of correction schemes focuses on the correction of the conflict reasoning conclusion itself. The second round of correction schemes focuses on the pre-correction of the reasoning nodes of concern in conflict propagation. The third round of correction schemes focuses on the adaptation adjustment of the corrected reasoning conclusion with the surrounding knowledge units. Each round of correction scheme includes the node adjustment direction, the requirements for the correction of the correlation relationship, and the suggestions for the optimization of the reasoning steps.

[0087] Based on the phased conflict tracing results and conflict propagation prediction results, a multi-round conflict correction plan was formulated. The first round of correction plan mainly focuses on the direct correction of the conflict reasoning conclusion itself, proposing how to modify the conflict reasoning conclusion to be consistent with the knowledge units and relationships of the knowledge graph, targeting the specific reasoning steps and error causes that led to the conflict. The second round of correction plan focuses on the reasoning nodes that the conflict may propagate to, and pre-corrects these nodes to prevent the conflict from further spreading and affecting other reasoning conclusions in subsequent reasoning processes. The third round of correction plan focuses on the adaptability adjustment of the corrected reasoning conclusion with surrounding knowledge units, ensuring that the corrected conclusion can coexist harmoniously with related knowledge units and has good compatibility and consistency in the knowledge system. Each round of correction plan includes detailed information on node adjustment directions (such as replacing nodes, adding nodes, deleting nodes, etc.), relationship correction requirements (such as correcting relationship strength, adjusting relationship type, adding or deleting relationships, etc.), and reasoning step optimization suggestions (such as adjusting reasoning order, supplementing reasoning steps, simplifying reasoning steps, etc.).

[0088] Step S1451: Regarding the first round of conflict correction scheme, based on the reasons for errors in the conflict reasoning conclusions in the phased conflict tracing results, if the error is a logical reasoning error, then reconstruct the logical reasoning chain based on the relationship between knowledge units, correct the incorrect order of reasoning steps, supplement missing reasoning links, and determine the adjustment direction of reasoning nodes. Specific operations include replacing incorrect nodes or adding connecting nodes. Determine the correction requirements for relationships, specifically strengthening relationships whose strength does not meet the requirements of the current reasoning link or correcting incorrect relationships. Determine optimization suggestions for reasoning steps, specifically splitting complex steps or merging redundant steps. If the error is a knowledge reference deviation, then rematch knowledge units that are compatible with the reasoning target, determine the replacement direction of knowledge units, specifically replacing peripheral compatible knowledge units with core compatible knowledge units. Determine the requirements for re-establishing relationships, specifically establishing new supporting relationships or supplementary relationships. Determine the adaptation adjustment suggestions for reasoning steps, specifically adjusting the reference order of knowledge units.

[0089] For the first round of conflict correction, targeted corrections were made based on the identified causes of errors in the conflict reasoning conclusions from the phased conflict tracing results. If the error was due to a logical inference flaw, the reasoning logic chain needed to be reconstructed. The new reasoning logic chain must be firmly based on the relationships between knowledge units to ensure logical rigor. Specific operations included correcting the incorrect order of reasoning steps, adjusting reversed or unreasonable steps to the correct sequence; supplementing necessary missing links in the reasoning process to make the reasoning chain complete and coherent. Simultaneously, the direction of adjustment for reasoning nodes was determined, such as replacing erroneous nodes or adding necessary connecting nodes to link broken reasoning steps. For relationships, correction requirements were proposed, such as strengthening relationships whose strength does not meet the requirements of the current reasoning stage to provide sufficient support for reasoning; correcting erroneous relationships by removing incorrect relationships or correcting the type and direction of relationships. Regarding reasoning steps, optimization suggestions were given, such as breaking down overly complex steps into multiple simple sub-steps to improve the clarity of reasoning; merging redundant or unnecessary steps to improve reasoning efficiency.

[0090] If the error is due to knowledge citation bias, then it's necessary to rematch knowledge units that are compatible with the reasoning objective. Determine the direction for replacing knowledge units; for example, if bias is caused by citing marginally compatible knowledge units, replace them with core compatible knowledge units to ensure the accuracy and relevance of the knowledge citations. For relationships, propose rebuilding them; based on the correct connections between knowledge units, establish new supporting or supplementary connections to replace the original erroneous connections. Regarding the reasoning steps, provide adaptation adjustment suggestions, such as adjusting the order of knowledge unit citations to make the application of knowledge more consistent with the reasoning logic and the internal structure of the knowledge system.

[0091] Step S1452: For the second round of conflict correction scheme, based on the focus inference nodes in the conflict propagation prediction results, analyze the correlation between the focus inference nodes and the conflict inference conclusions, and formulate pre-correction measures for each focus inference node. If the focus inference node is a direct follow-up node to the conflict inference conclusion, adjust the inference basis of the focus inference node in advance. Specific operations include changing the referenced knowledge unit, determining the pre-adjustment direction of the node, determining the pre-correction requirements of the correlation, specifically adjusting the correlation strength that may be affected by the conflict to a preset safe range, and determining the pre-optimization suggestions for the inference steps, specifically inserting verification steps in advance. If the focus inference node is an indirect related node to the conflict inference conclusion, adjust the correlation strength between the focus inference node and other nodes to reduce the probability of conflict propagation, determine the adjustment direction of the node correlation strength, determine the protection requirements of the correlation, and determine the buffer suggestions for the inference steps, specifically adding intermediate verification steps.

[0092] For the second round of conflict correction schemes, pre-corrections are made based on the key inference nodes in the conflict propagation prediction results. First, the correlation between each key inference node and the conflict inference conclusion is analyzed; the higher the correlation, the greater the likelihood of being affected by conflict propagation. For each key inference node, corresponding pre-correction measures are formulated.

[0093] If the node in question is a direct successor to a conflicting inference conclusion—meaning its reasoning directly depends on that conclusion—then the node's reasoning basis needs to be adjusted beforehand. Specific actions include replacing the knowledge units referenced by the node with correct units consistent with the knowledge graph, thus altering its reasoning foundation. The pre-adjustment direction for the node should be determined, clarifying how its content should be modified to avoid the impact of conflict. For relationships, pre-correction requirements should be proposed, adjusting relationships whose strength might be affected by conflict to a preset safe range to enhance reasoning stability. Regarding reasoning steps, pre-optimization suggestions should be provided, such as inserting verification steps early in the node's reasoning process to verify the reasoning basis and intermediate results, ensuring their correctness.

[0094] If the node of interest is an indirectly related node to the conflicting inference conclusion—meaning it is indirectly connected to the conclusion through other nodes—then the strength of the association between that node and the other nodes needs to be adjusted. Reducing the association strength decreases the probability of conflict propagating to that node through these associations. The direction of the association strength adjustment should be determined, clarifying which associations should be strengthened or weakened. Protective requirements for these associations should be proposed, such as imposing stricter restrictions on their application conditions. Buffering suggestions should be given for the inference steps, such as adding intermediate verification steps between the node and nodes that may propagate conflict, filtering and verifying information to prevent conflict propagation.

[0095] Step S1453: For the third round of conflict correction scheme, based on the reasoning conclusions after the first two rounds of correction, analyze the adaptability of the corrected reasoning conclusions with the surrounding knowledge units. If the correlation strength between the corrected conclusions and the surrounding knowledge units does not meet the requirements, then strengthen the correlation between the two. Specific operations include supplementing intermediate knowledge units to establish new correlations, determining the direction of strengthening the correlation, determining the adaptability adjustment requirements of nodes, specifically adjusting the content of the nodes to improve adaptability, and determining suggestions for improving the reasoning steps, specifically adding correlation verification steps. If there are redundant correlations between the corrected conclusions and the surrounding knowledge units, then simplify the correlation. Specific operations include deleting invalid correlations or merging duplicate correlations, determining the simplification direction of the correlation, determining the simplification requirements of the nodes, and determining suggestions for optimizing the reasoning steps, specifically merging correlation verification steps.

[0096] For the third round of conflict correction, based on the first two rounds of correction, the adaptability of the corrected reasoning conclusions with surrounding knowledge units is analyzed. Surrounding knowledge units refer to other knowledge units that are directly or indirectly related to the corrected reasoning conclusions. If it is found that the strength of the association between the corrected conclusions and surrounding knowledge units does not meet the reasoning requirements—that is, the association is insufficient to support the reliability of the conclusions or the smooth progress of the reasoning—then the association needs to be strengthened. Specific operations may include supplementing intermediate knowledge units, establishing new association paths between the corrected conclusions and surrounding knowledge units through these new intermediate knowledge units, thereby enhancing the overall association strength. The direction of strengthening the association is determined, clarifying which aspects of the association should be strengthened. Adaptability adjustment requirements are proposed for relevant nodes, such as fine-tuning the content description of adjustment points to make their association with surrounding knowledge units closer and improve adaptability. Regarding the reasoning steps, improvement suggestions are given, such as adding an association verification step to verify the strengthened associations and ensure their effectiveness.

[0097] If redundant connections are found between the revised conclusion and surrounding knowledge units—that is, connections that do not substantially help the reasoning process or are repetitive—then these connections need to be simplified. Specific operations include deleting invalid connections (those that do not affect the reasoning conclusion and have no practical use) or merging duplicate connections (multiple connections expressing the same meaning or function). The direction of simplification should be determined, clarifying which connections should be streamlined. Simplification requirements should be proposed for relevant nodes, removing unnecessary content related to redundant connections. Optimization suggestions should be given for the reasoning steps, such as merging connection verification steps, reducing unnecessary verification steps, and improving reasoning efficiency.

[0098] Step S1454: Integrate the specific content of each round of conflict correction schemes, mark the execution order, applicable reasoning stage, expected correction effect and verification criteria of each round of conflict correction schemes, and form a multi-round conflict correction scheme document.

[0099] The specific details of the first, second, and third rounds of conflict correction plans were integrated. During integration, the execution order of each round of correction plans was clearly marked to ensure that the correction work was carried out sequentially. Simultaneously, the applicable inference stage for each round of correction plans was noted, enabling the correction work to be applied specifically to the corresponding inference stages. For each round of correction plans, the expected correction effect was set, such as the proportion of conflict elimination and the improvement in inference accuracy, and corresponding verification criteria were established, such as how to determine if a conflict has been eliminated and whether the inference accuracy has met expectations. This information was systematically organized to form a complete multi-round conflict correction plan document.

[0100] Step S146: Apply the multi-round conflict correction scheme sequentially: Apply the first-round correction scheme to correct the large model reasoning process after the stage guidance. Starting from the reasoning link where the conflict occurred, re-execute all subsequent reasoning steps, and re-execute the extraction, classification, and comparison with the knowledge graph of the reasoning conclusion to verify the correction effect and identify new conflicts; verify whether the correction effect meets the preset standard. If it does not meet the preset standard, adjust the details of the second-round correction scheme based on the verification results and apply it again, and repeat this verification process; after completing the application and verification of the second-round correction scheme, continue to apply and verify the third-round correction scheme; repeat this cycle until all rounds of correction schemes have been applied, and the conflict is completely resolved and no new conflict is generated, thus obtaining the large model reasoning process after conflict resolution.

[0101] Following the execution order outlined in the multi-round conflict correction plan document, the correction plan is applied round by round to the large model inference process after phase guidance. First, the first-round correction plan is applied, starting with the inference step that generated the conflict identified in the phased conflict tracing results. This step and all subsequent inference steps are re-executed, adjusting the inference nodes, relationships, and steps according to the requirements of the first-round correction plan. After correction, the extraction and classification of inference conclusions are performed again, and the new inference conclusions are compared with the structured knowledge graph to verify the effectiveness of the first-round correction plan and check for any new conflicts.

[0102] The effectiveness of the first round of corrections is evaluated based on preset verification criteria (such as the number of conflicts eliminated, the types of remaining conflicts, etc.). If the preset criteria are not met, the unresolved conflict issues in the verification results are analyzed, and the details of the second round of conflict correction plan are adjusted based on these issues. For example, the pre-correction measures of certain nodes are strengthened or the pre-correction requirements of the correlation are adjusted. Then the adjusted second round of correction plan is applied, and the above verification process is repeated (re-execute the reasoning steps, extract the comparison conclusions, and evaluate the effect).

[0103] After the second round of revision schemes has been applied and verified, and the expected results have been achieved, the third round of conflict revision schemes is applied and verified in the same way. This process is repeated, applying each round of revision schemes in turn and performing rigorous verification, until all rounds of revision schemes have been applied. Finally, when all conflicts have been completely resolved and no new conflicts have arisen during the verification process, the revision process is stopped, and the large model inference process after conflict resolution is obtained.

[0104] Step S150: Based on the large model reasoning process after conflict resolution, iteratively optimize the reasoning path, evaluate the fit between the reasoning path and the knowledge unit association and the reasoning efficiency, adjust the jump order of reasoning nodes and the reference priority of knowledge units, integrate the optimized reasoning path, knowledge graph application records and conflict correction information, and obtain the large model logical reasoning optimization result.

[0105] After obtaining the large-scale model reasoning process after conflict resolution, its reasoning path is iteratively optimized. The optimization of the reasoning path aims to further improve the quality and efficiency of reasoning, making it more consistent with the knowledge structure of the knowledge graph and the requirements of the reasoning task. By evaluating the fit between the reasoning path and the relationships between knowledge units, as well as the reasoning efficiency, shortcomings and areas for improvement in the reasoning path are identified. Then, the jump order of reasoning nodes and the reference priority of knowledge units are adjusted to optimize the reasoning path. Finally, the optimized reasoning path, the application records of the knowledge graph throughout the reasoning process, and the conflict correction information are integrated to form the final optimized result of the large-scale model logical reasoning.

[0106] Step S151: Extract the complete reasoning path in the large model reasoning process after conflict resolution, and divide it into initial stage path, intermediate stage path and final stage path according to the reasoning stage. Each path contains the reasoning node sequence, the relationship type between nodes and the knowledge unit reference record, to obtain the complete reasoning path in stages.

[0107] The complete reasoning path is extracted from the large model's reasoning process after conflict resolution. This reasoning path is a sequence of all reasoning nodes and their interrelationships from the start of reasoning to the final conclusion. Based on the division of reasoning stages, the complete reasoning path is broken down into initial stage path, intermediate stage path, and final stage path. Each stage path includes a detailed sequence of reasoning nodes within that stage (the order and specific content of the nodes), the type of interrelationships between nodes (e.g., supporting relationships, supplementary relationships, deductive relationships, etc.), and records of knowledge units referenced during the reasoning process (identifier of the referenced knowledge unit, its location, purpose, etc.). Through this breakdown and detailed recording, a complete staged reasoning path is obtained.

[0108] Step S152: Construct a knowledge base for evaluating reasoning paths, quantitatively evaluate the complete reasoning paths in stages, and obtain the evaluation results of the stages. The knowledge base for evaluating reasoning paths includes the fit index of the relationship between paths and knowledge units and the reasoning efficiency index. The fit index includes the core association ratio and the correct association rate. The reasoning efficiency index includes the number of reasoning steps and the time spent jumping between nodes.

[0109] For example, in step S1521: Construct a knowledge base for fit indexes. The core association ratio index calculates the proportion of the number of associations between core adapted knowledge units in the reasoning path to the total number of associations. The correct association rate index calculates the proportion of the number of associations in the reasoning path that conform to the knowledge unit association network to the total number of associations.

[0110] A knowledge base of metrics is constructed to evaluate the fit between inference paths and knowledge unit relationships. The core relationship ratio metric is defined as the proportion of relationships between core-fitting knowledge units within a inference path to the total number of relationships in that path. Core-fitting knowledge units are key knowledge units directly related to the inference objective; a high proportion of their relationships indicates that the inference path focuses on core knowledge and has a high degree of fit. The correct relationship rate metric is defined as the proportion of relationships within a inference path that conform to the established relationships in the knowledge unit relationship network to the total number of relationships in the path. This correct relationship rate reflects the correctness and compliance of relationships within the inference path; a higher proportion indicates a better fit between the inference path and the knowledge graph.

[0111] Step S1522: Construct a knowledge base for reasoning efficiency indicators. The reasoning step quantity indicator counts the total number of reasoning steps in each reasoning stage. The total number of reasoning steps does not include redundant steps. The node jump time indicator counts the average time spent jumping between reasoning nodes within each reasoning stage. The average time spent is the average time difference from the generation of the current node to the generation of the next jump node.

[0112] Construct a knowledge base of metrics for evaluating inference efficiency. The number of inference steps metric is used to count the total number of inference steps in each inference stage. Redundant steps (such as repeated verification steps or corrected error steps) must be excluded during the calculation. A lower total number of steps indicates a more concise and efficient inference process. The node jump time metric is used to count the average time spent jumping between inference nodes within each inference stage. This is calculated by averaging the time difference between the generation of the current node and the generation of the next hop node for all adjacent inference nodes within that stage. A shorter average time indicates higher jump efficiency between nodes and better smoothness of the inference process.

[0113] Step S1523: Set the weight parameters for each evaluation indicator. Based on the key requirements of the reasoning task, when the accuracy of the reasoning conclusion is given priority, the weight parameter of the fit indicator is higher than that of the reasoning efficiency indicator; when the reasoning speed is given priority, the weight parameter of the reasoning efficiency indicator is higher than that of the fit indicator; when the accuracy of the reasoning conclusion is given priority, the weight parameter of the core association ratio is set according to the first importance level, the weight parameter of the correct association rate is set according to the second importance level, the weight parameter of the number of reasoning steps is set according to the third importance level, and the weight parameter of the node jump time is set according to the fourth importance level.

[0114] Based on the key requirements of the logical reasoning task, corresponding weight parameters are set for each constructed evaluation indicator. When the reasoning task has high requirements for the accuracy of the reasoning conclusion, the weight parameters of the fit index (the proportion of core associations and the correct association rate) should be higher than those of the reasoning efficiency index (the number of reasoning steps and the time spent on node jumps) to highlight the importance of the fit between the reasoning path and the knowledge graph. When the reasoning task focuses more on reasoning speed and efficiency, the weight parameters of the reasoning efficiency index are higher than those of the fit index.

[0115] Prioritizing the accuracy of the reasoning conclusion, the weights within the fit index are further divided: the core association percentage is assigned the highest weight, set as the first level of importance; the correct association rate is assigned the second highest weight, set as the second most important; the number of reasoning steps is assigned the third highest weight; and the node jump time is assigned the lowest weight, set as the fourth highest weight. By setting these weight parameters, the relative importance of different indicators can be reflected in the comprehensive evaluation.

[0116] Step S1524: Calculate the indicators for the complete inference path in stages. Statistically calculate the core association ratio, correct association rate, number of inference steps, and node jump time for the initial stage, intermediate stage, and final stage respectively. Normalize each indicator value and then calculate the comprehensive evaluation score for each inference stage by combining the indicator weight parameters.

[0117] Based on the initial, intermediate, and final stages, various evaluation indicators are calculated for each stage of the complete reasoning path. For the core association ratio, the ratio of the number of associations between core adapted knowledge units in each stage to the total number of associations is calculated. For the correct association rate, the ratio of the number of associations conforming to the knowledge unit association network in each stage to the total number of associations is calculated. For the number of reasoning steps, the total number of steps in each stage without redundant steps is directly calculated. For node jump time, the average jump time difference between nodes within each stage is calculated.

[0118] Each calculated indicator value is normalized to map values ​​of different dimensions and orders of magnitude to a unified numerical range (e.g., [0,1]) to eliminate the impact of dimensional differences on the overall evaluation. After normalization, each indicator value is multiplied by its corresponding weight parameter, and then the weighted values ​​of all indicators are summed to obtain the overall evaluation score for each inference stage. The overall evaluation score reflects the overall quality of the inference path at that stage.

[0119] Step S1525: Record the values ​​of each indicator, the comprehensive evaluation score and the score level for each reasoning stage to form the phased path evaluation results.

[0120] The specific values, normalized values, comprehensive evaluation scores, and score levels (e.g., excellent, good, average, poor) of each inference stage, including the core association percentage, correct association rate, number of inference steps, and node jump time, are recorded in detail for each inference stage. These records are then organized according to the inference stage to form a phased path evaluation result, which comprehensively reflects the performance of each stage's inference path in terms of fit and efficiency.

[0121] Step S153: Based on the phased path evaluation results, if the fit index does not meet the preset standard, the jump order of the reasoning nodes is adjusted, and the jump direction of the node corresponding to the core association of the knowledge unit is selected first, so that the core association ratio meets the preset standard; if the reasoning efficiency index does not meet the preset standard, the reference priority of the knowledge unit is adjusted, and the knowledge units whose retrieval response time meets the preset reasoning stage standard and whose association strength meets the requirements of the current reasoning link in the knowledge unit association relationship network are referenced first, the number of reasoning steps is controlled within the preset reasoning stage steps range, and the node jump time is controlled within the preset reasoning stage time range, so as to obtain the adjusted phased reasoning path.

[0122] Based on the phased path evaluation results, determine whether each indicator meets the preset standards. If the fit index (core association ratio or correct association rate) does not meet the preset standards, it indicates that the reasoning path is insufficient in terms of its fit with the knowledge unit relationships. In this case, adjust the jump order of the reasoning nodes. During the reasoning process, when faced with multiple possible next-hop node choices, prioritize the jump direction of nodes that can form a core association with the current node (i.e., the association between core-fitting knowledge units). Through the above adjustment, increase the number of core associations in the reasoning path, thereby increasing the core association ratio and making it meet the preset standards.

[0123] If the reasoning efficiency metrics (number of reasoning steps or node jump time) do not meet the preset standards, it indicates that the reasoning process is not efficient enough. In this case, the priority of knowledge unit citations is adjusted. When citing knowledge units, priority is given to those whose retrieval response time meets the preset standards for this reasoning stage (e.g., faster response time is required in intermediate reasoning stages) and whose association strength meets the requirements of the current reasoning stage in the knowledge unit association network (e.g., the current stage requires support from knowledge units with high-strength associations). By optimizing the citation order and selection of knowledge units, the number of reasoning steps is controlled within the preset range for this reasoning stage, while the node jump time is controlled within the preset time range. After the above adjustments, the adjusted phased reasoning path is obtained.

[0124] Step S154: Iteratively verify the adjusted phased reasoning path, re-execute the complete reasoning process, and perform quantitative evaluation again based on the reasoning path obtained after re-execution to obtain a new path evaluation result; determine whether the new path evaluation result meets the preset standard; if it does not meet the standard, perform the operation of adjusting the jump order of reasoning nodes and the priority of knowledge unit reference again, and then perform iterative verification again; repeat this process until the path evaluation result meets the requirements and the iteratively optimized phased reasoning path is obtained.

[0125] The adjusted phased inference path is iteratively verified. Following the adjusted path, the entire large-scale model inference process is re-executed, including all inference steps from the initial stage to the final stage. Based on the new inference path obtained after re-execution, the quantitative evaluation of various indicators is performed again according to the method in step S152, calculating the core association ratio, correct association rate, number of inference steps, node jump time, and comprehensive evaluation score to obtain a new path evaluation result.

[0126] The new path evaluation results are compared with preset standards to determine whether all indicators have met the preset standards. If any indicators still fail to meet the standards, the steps to adjust the jump order of inference nodes and the priority of knowledge unit citations are performed again based on the problems reflected in the new evaluation results. For example, the jump order may be further optimized to increase the proportion of core associations, or the citation priority may be adjusted to further reduce the number of steps. Then, the adjusted path is iteratively verified again (inference is re-executed and the evaluation results are re-evaluated). This process is repeated, continuously adjusting and verifying, until all indicators in the phased path evaluation results meet the preset requirements, resulting in the iteratively optimized phased inference path.

[0127] Step S155: Collect knowledge graph application records throughout the reasoning process, including the types of knowledge units referenced at each stage, the types of relationships, the number of applications, and the application effects, and organize them into a knowledge graph application list according to the reasoning stage; summarize conflict correction information, including the content of each round of conflict correction schemes, execution effects, verification results, and adjustment records, and organize them into a conflict correction report according to the correction round.

[0128] Throughout the large-scale model inference optimization process, application records of the knowledge graph are collected. These records include the types of knowledge units referenced at each stage of inference (core adaptation, auxiliary adaptation, edge adaptation), the types of relationships between referenced knowledge units (supporting relationships, supplementary relationships, indirect relationships, etc.), the number of times each knowledge unit and relationship is applied, and their application effects during inference (e.g., whether they effectively support the inference steps, whether they cause conflicts, etc.). These records are then categorized and organized according to the inference stage to form a knowledge graph application list, clearly demonstrating the application of the knowledge graph during the inference process.

[0129] Simultaneously, conflict correction information is compiled, including the specific content of each round of conflict correction schemes (direction of node adjustment, requirements for correction of relationships, suggestions for optimization of reasoning steps, etc.), the execution effect of each round of correction schemes (number of conflicts eliminated, remaining conflicts), verification results (whether the preset standards were met, reasons for non-compliance), and records of adjustments made to the schemes during the correction process. This information is then organized by correction round to form a conflict correction report, detailing the entire conflict correction process and results.

[0130] Step S156: Integrate the iteratively optimized phased reasoning paths, knowledge graph application list, and conflict correction report; supplement the reasoning goal achievement analysis results; and form a large model logic reasoning optimization result containing all key information. The reasoning goal achievement analysis results include the fit between the reasoning conclusion and the goal, the efficiency of the reasoning process in utilizing the knowledge graph, and a supplementary summary of the reasoning optimization effect, which includes the improvement in fit and the improvement in reasoning efficiency.

[0131] The iteratively optimized phased reasoning paths, knowledge graph application lists, and conflict correction reports are integrated. Based on this, the results of the reasoning goal achievement analysis are added. This analysis includes the degree of alignment between the reasoning conclusions and the reasoning goals (e.g., whether the reasoning conclusions fully meet the core attribute requirements of the reasoning goals and whether they are accurate) and the efficiency of the reasoning process in utilizing the knowledge graph (e.g., the accuracy of knowledge unit citations and the effective utilization rate of relationships).

[0132] Further supplement the summary of reasoning optimization effects, quantifying the changes before and after optimization, such as the improvement in fit (the ratio of the difference between the optimized fit score and the unoptimized score to the unoptimized score), and the improvement in reasoning efficiency (the ratio of the difference between the number of reasoning steps before and after optimization to the number of steps before optimization, or the ratio of the difference between the time spent on node jumps before and after optimization to the time spent before optimization). Organize all this information organically to form a comprehensive optimization result for the large-scale model's logical reasoning, containing all key information. This optimization result fully reflects the final state and effect of the large-scale model's reasoning process after optimization.

[0133] Figure 2 The illustration shows exemplary hardware and software components of a large model logic reasoning optimization system 100 incorporating knowledge graphs, which can implement the ideas of this application, according to some embodiments of this application. For example, a processor 120 can be used in the large model logic reasoning optimization system 100 incorporating knowledge graphs and to perform the functions in this application.

[0134] For example, the large-scale model logic reasoning optimization system 100 incorporating knowledge graphs may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the large-scale model logic reasoning optimization system 100 incorporating knowledge graphs may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The methods of this application can be implemented according to these program instructions. The large-scale model logic reasoning optimization system 100 incorporating knowledge graphs also includes an I / O interface 150 between the computer and other input / output devices.

[0135] Furthermore, this embodiment of the invention also provides a readable storage medium, wherein computer-executable instructions are preset in the readable storage medium, and when the processor executes the computer-executable instructions, the above-mentioned large model logical reasoning optimization method combining knowledge graph is implemented.

[0136] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for optimizing logical reasoning of a large model combined with a knowledge graph, characterized in that, The method comprises: Obtaining a logical reasoning task to be processed by a large model and a corresponding structured knowledge graph, wherein the structured knowledge graph comprises knowledge units, association relationships and association strengths between the knowledge units, the logical reasoning task comprises a reasoning target, reasoning premises and reasoning constraint conditions, the logical reasoning task is a specific subject problem encountered by a student in a learning process, the reasoning target refers to a final conclusion or result to be achieved by the subject problem, the reasoning premises are known conditions for performing reasoning, including specific data given in the problem, existing theorems and laws, related concept definitions, the reasoning constraint conditions are restrictions or specifications for the reasoning process, the corresponding structured knowledge graph is a structured representation of a knowledge system related to the subject, the knowledge units are basic elements constituting the structured knowledge graph, covering concepts, theorems, formulas, experimental phenomena and problem solving methods in the subject, the association relationships between the knowledge units reflect the internal relationships between different knowledge units, including a containing relationship between concepts, a derivation relationship between theorems, an association relationship between formulas, and an application relationship between concepts and theorems, and the association strength is used to measure the closeness or importance of the association relationship between the knowledge units; Based on the adaptability analysis of the structured knowledge graph and the logical reasoning task, a large model reasoning guide rule set is obtained; The large model reasoning guide rule set is injected into a large model reasoning process, intermediate reasoning nodes in the large model reasoning process are captured in stages, the jump direction of the intermediate reasoning nodes is adjusted by referring to the guide rules of different reasoning stages, the reasoning branches deviating from the association relationships of the knowledge units are corrected, and a large model reasoning process after stage guidance is obtained; Conflict resolution and propagation prediction are performed on the large model reasoning process after stage guidance, an inconsistent reasoning conclusion in the reasoning process with the structured knowledge graph is identified, the association relationships between the knowledge units are called to verify the rationality of the conflict reasoning conclusion, reasoning links where the conflict may propagate are predicted, a multi-round conflict correction scheme is generated and applied to the reasoning process round by round, and a large model reasoning process after conflict resolution is obtained; Based on the large model reasoning process after conflict resolution, reasoning path iteration optimization is performed, the fit degree of the reasoning path and the association relationships of the knowledge units and the reasoning efficiency are evaluated, the jump order of the reasoning nodes and the reference priority of the knowledge units are adjusted, and the optimized reasoning path, knowledge graph application record and conflict correction information are integrated, and a large model logical reasoning optimization result is obtained.

2. The method of claim 1, wherein the method further comprises: The adaptability analysis of the structured knowledge graph and the logical reasoning task to obtain the large model reasoning guide rule set comprises: The knowledge units in the structured knowledge graph are annotated for reasoning task adaptability, the association degree of each knowledge unit and the reasoning target and the reasoning premises in the logical reasoning task is analyzed, the core adaptive knowledge units, the auxiliary adaptive knowledge units and the edge adaptive knowledge units are marked, and a knowledge unit adaptability annotation result is obtained. The association relationship between the knowledge units is parsed based on the knowledge unit adaptability annotation result, support association between the core adaptive knowledge unit and the auxiliary adaptive knowledge unit, supplementary association between the auxiliary adaptive knowledge unit and the marginal adaptive knowledge unit, and indirect association between the core adaptive knowledge unit and the marginal adaptive knowledge unit are identified, and a knowledge unit association relationship network is constructed; A priority direction of a reasoning path is determined according to the knowledge unit association relationship network, the core adaptive knowledge unit is taken as a reasoning starting node, the auxiliary adaptive knowledge unit corresponding to the support association is preferentially selected as a next-hop reasoning node, and the marginal adaptive knowledge unit is selected based on the supplementary association to perfect the reasoning chain, and a staged reasoning path priority order table is formed; The association strength between the knowledge units in the structured knowledge graph is extracted, the jump threshold of the reasoning node is dynamically set in combination with the adaptability annotation result of the knowledge unit and the reasoning task, the jump threshold is set in a first priority between the core adaptive knowledge units, the jump threshold is set in a second priority between the auxiliary adaptive knowledge units and the core adaptive knowledge units, the jump threshold is set in a third priority between the marginal adaptive knowledge units and other knowledge units, the dynamic threshold corresponding to each association relationship is recorded, and a staged reasoning node jump threshold table is obtained; The reasoning constraint condition is layered according to the reasoning stage, the constraint condition of the initial reasoning stage focuses on the compliance of the reasoning premise, the constraint condition of the intermediate reasoning stage focuses on the rationality of the knowledge unit reference, and the constraint condition of the final reasoning stage focuses on the fitting degree of the reasoning conclusion and the target, each layer of constraint condition corresponds to form the adaptation requirement of the reasoning step, including the stage prohibited association list, the stage step number range and the stage conclusion attribute standard; The staged reasoning path priority order table, the staged reasoning node jump threshold table and the layered reasoning step adaptation requirement are integrated, the reasoning stage division rule item is followed, each reasoning stage rule item contains path priority, dynamic jump threshold and layered adaptation requirement, and a large model reasoning guidance rule set is obtained.

3. The method of claim 1, wherein the method further comprises: The large model reasoning guidance rule set is injected into the large model reasoning process, the intermediate reasoning node in the large model reasoning process is captured in stages, the jump direction of the intermediate reasoning node is adjusted by comparing the guidance rules of different reasoning stages, the reasoning branch deviating from the knowledge unit association relationship is corrected, and a stage-guided large model reasoning process is obtained, including: In the large model reasoning process, a reasoning node capture window is set according to the reasoning stage, the intermediate reasoning node is captured once every interval of a preset fixed reasoning step in the initial reasoning stage, the intermediate reasoning node is captured once every interval of a preset dynamic reasoning step in the intermediate reasoning stage, and the intermediate reasoning node is captured in real time in the final reasoning stage, the content, generation time, previous reasoning node unique mark and belonging reasoning stage of each intermediate reasoning node are recorded, and a staged intermediate reasoning node set is obtained; matching each intermediate reasoning node of each reasoning stage in the set of staged intermediate reasoning nodes with the reasoning path priority order table of the corresponding stage in the set of large model reasoning guidance rules, identifying the next hop priority knowledge unit type corresponding to the current intermediate reasoning node, the priority knowledge unit type including core adaptation, auxiliary adaptation, and edge adaptation, determining a reasonable reasoning jump direction for the reasoning stage, and obtaining a staged node jump direction suggestion; comparing the next hop reasoning node currently planned by the large model with the priority knowledge unit type in the staged node jump direction suggestion, if they are consistent, retaining the current reasoning path, if they are inconsistent, extracting the corresponding dynamic threshold value in the reasoning node jump threshold value table of the corresponding reasoning stage, and judging whether the correlation strength from the current reasoning node to the planned next hop node meets the threshold requirement of the reasoning stage; if the correlation strength from the current reasoning node to the planned next hop node meets the threshold requirement of the reasoning stage, generating a stage path adjustment suggestion, explaining the functional characteristics of the priority knowledge unit type in the current reasoning stage, the relevance to the reasoning target, and the supporting role for the subsequent reasoning link, guiding the large model to adjust the jump direction to the priority knowledge unit; if the correlation strength from the current reasoning node to the planned next hop node does not meet the threshold requirement of the reasoning stage, directly modifying the reasoning branch, replacing the planned next hop node with a specific knowledge unit in the priority knowledge unit type, and updating the node correlation record and stage reasoning log in the reasoning process; after completing the intermediate reasoning node capture, matching, comparison, and modification of the current reasoning stage, entering the next reasoning stage, and for the next reasoning node, re-executing the entire process from setting the reasoning node capture window to generating the node jump direction suggestion, performing jump comparison, and making path adjustment or direct modification according to the comparison result, until all reasoning stages are processed, recording the path adjustment record, node modification information, and stage reasoning effect of each reasoning stage, and obtaining the large model reasoning process after stage guidance.

4. The method of claim 1, wherein the method further comprises: The stage-guided large model reasoning process is subjected to conflict resolution and propagation prediction, the reasoning conclusions inconsistent with the structured knowledge graph in the reasoning process are identified, the association relationship between knowledge units is called to verify the rationality of the conflict reasoning conclusion, the reasoning link where the conflict may propagate is predicted, and multiple rounds of conflict modification schemes are generated and applied to the reasoning process round by round, obtaining the large model reasoning process after conflict resolution, including: extracting all reasoning conclusions in the stage-guided large model reasoning process, classifying them into initial stage reasoning conclusions, intermediate stage reasoning conclusions, and final stage reasoning conclusions according to the reasoning stage, recording the generation basis, associated knowledge units, and belonging reasoning link of each reasoning conclusion, and obtaining a set of staged reasoning conclusions; comparing each reasoning conclusion in the set of staged reasoning conclusions with the knowledge units and association relationships in the structured knowledge graph, identifying the reasoning conclusions that are contradictory to the content of the knowledge units and conflict with the association relationships, marking them as conflict reasoning conclusions, and classifying them into a set of staged conflict reasoning conclusions according to the reasoning stage where the conflict occurs. trace each conflict reasoning conclusion in the set of the staged conflict reasoning conclusions, find the reasoning step of generating the conflict reasoning conclusion, the dependent intermediate reasoning node, the cited knowledge unit and the associated reasoning rule, determine the specific reasoning link of the conflict and the corresponding error cause, obtain the staged conflict trace result, and the error cause includes a reasoning logic error or a knowledge reference deviation; call the knowledge unit association relationship related to the conflict reasoning conclusion in the structured knowledge graph, verify the contradiction points of the conflict reasoning conclusion and the association relationship, analyze the essential cause of the contradiction, predict the subsequent reasoning link that the conflict may propagate in combination with the knowledge unit association relationship network, mark the attention reasoning node of the conflict propagation, and obtain the conflict propagation prediction result; based on the staged conflict trace result and the conflict propagation prediction result, generate a multi-round conflict correction scheme, a first round correction scheme focuses on the correction of the conflict reasoning conclusion itself, a second round correction scheme focuses on the pre-correction of the attention reasoning node of the conflict propagation, and a third round correction scheme focuses on the adaptability adjustment of the corrected reasoning conclusion and the surrounding knowledge unit, each round correction scheme includes a node adjustment direction, an association relationship correction requirement and a reasoning step optimization suggestion; apply the multi-round conflict correction scheme in sequence: apply the first round correction scheme to the stage guided large model reasoning process, start from the reasoning link where the conflict occurs, re-execute all subsequent reasoning steps, and execute the extraction, classification and comparison operation of the reasoning conclusion with the knowledge graph again to verify the correction effect and identify new conflicts; verify whether the correction effect meets the preset standard, if the preset standard is not met, adjust the details of the second round correction scheme based on the verification result and then apply it, and repeat the verification process; after completing the application and verification of the second round correction scheme, continue to apply and verify the third round correction scheme; until all rounds of correction schemes are applied, and the conflict is completely eliminated and no new conflict is generated, obtain the conflict-eliminated large model reasoning process.

5. The method of claim 2, wherein the method further comprises: The knowledge unit in the structured knowledge graph is annotated for reasoning task adaptability, the correlation degree of each knowledge unit with a reasoning target and a reasoning premise in a logical reasoning task is analyzed, core adaptive knowledge units, auxiliary adaptive knowledge units and edge adaptive knowledge units are marked, and a knowledge unit adaptability annotation result is obtained, including: extract the reasoning target keywords and reasoning premise keywords in the logical reasoning task, construct a reasoning task keyword set, the reasoning target keywords contain core attributes that need to be met to achieve a reasoning conclusion, and the reasoning premise keywords contain basic conditions that need to be based on in a reasoning process; perform keyword matching on each knowledge unit in the structured knowledge graph, and count the number and matching depth of the reasoning target keywords and the reasoning premise keywords contained in the knowledge unit content to obtain a knowledge unit keyword matching statistical result; Calculate the adaptation degree of the knowledge unit to the reasoning task based on the keyword matching statistical results of the knowledge unit, the adaptation degree calculation combines the keyword matching quantity and the matching depth, the keyword matching quantity is positively correlated with the adaptation degree, and the adaptation degree classification standard is set, the knowledge unit directly supporting the realization of the reasoning target is marked as the core adaptation knowledge unit, the knowledge unit supplementing the function of the core adaptation knowledge unit is marked as the auxiliary adaptation knowledge unit, and the knowledge unit perfecting the details of the reasoning process is marked as the edge adaptation knowledge unit; Secondly verify the marked knowledge unit, analyze whether the core adaptation knowledge unit directly supports the realization of the reasoning target, whether the auxiliary adaptation knowledge unit supplements the core adaptation knowledge unit or plays a logical connection role, and whether the edge adaptation knowledge unit is helpful to the details of the reasoning process, and eliminate the knowledge unit that does not meet the definition of the adaptation type and re-label it; Record the adaptation type, adaptation degree value, matched keyword and secondary verification result of each knowledge unit, sort by adaptation type to get the knowledge unit adaptation labeling result.

6. The method of claim 3, wherein the method further comprises: If the association strength of the current reasoning node to the planned next hop node meets the threshold requirement of the reasoning stage, a stage path adjustment suggestion is generated, which describes the functional characteristics of the priority knowledge unit type in the current reasoning stage, the relevance to the reasoning target and the support to the subsequent reasoning link, and guides the large model to adjust the jump direction to the priority knowledge unit, including: Extract the type of the priority knowledge unit corresponding to the current reasoning stage in the large model reasoning guide rule set, analyze the functional characteristics of the type of knowledge unit in the current reasoning stage, summarize the functional differences of the type of knowledge unit compared with the type of the planned next hop node, and get the stage priority type advantage analysis result, the priority knowledge unit type includes the core adaptation knowledge unit type, the auxiliary adaptation knowledge unit type and the edge adaptation knowledge unit type, the core adaptation knowledge unit type has logical rigor, the auxiliary adaptation knowledge unit type has smooth connection, and the edge adaptation knowledge unit type has detailed integrity; Combined with the reasoning target in the logical reasoning task, the driving effect of the priority knowledge unit type on achieving the reasoning target is described stage by stage to get the stage target association analysis result, the initial reasoning stage describes the fit degree of the priority type knowledge unit and the reasoning premise, the intermediate reasoning stage describes the support of the priority type knowledge unit to the extension of the reasoning chain, and the final reasoning stage describes the guarantee effect of the priority type knowledge unit to the accuracy of the reasoning conclusion; Based on the stage priority type advantage analysis result and the stage target association analysis result, a stage path adjustment suggestion document is generated, which contains current reasoning stage information, current reasoning node information, planned next hop node information, priority knowledge unit type information, stage advantage comparison, stage target association description and expected effect of adjusted reasoning path, determines the suggested adjustment jump direction and forms the target record; The stage path adjustment suggestion document is input into a large model inference interaction interface, triggering a large model stage adaptability evaluation process for path adjustment suggestions, calling the large model to re-judge the rationality of the jump direction in combination with the task focus of the current inference stage, the inference logic of the large model, and the suggestion content; If it is detected that the large model accepts the stage path adjustment suggestion, the next jump inference node is adjusted to a specific knowledge unit in the preferred knowledge unit type, and the adjustment basis, suggestion reference information, and stage adaptability evaluation results are marked in the inference process record; if it is detected that the large model does not accept the suggestion, a suggestion feedback statement is generated, the non-acceptance reasons mentioned in the suggestion feedback statement are analyzed, additional function evidence and logic connection schemes for the preferred knowledge unit type in the inference stage are supplemented, the stage path adjustment suggestion is submitted again, and the jump direction adjustment of the large model is completed, the non-acceptance reasons including insufficient stage adaptability or logic connection problems.

7. The method of claim 4, wherein the method further comprises: Based on the stage-by-stage conflict tracing results and conflict propagation prediction results, a multi-round conflict correction scheme is generated, the first round correction scheme focuses on the correction of the conflict inference conclusion itself, the second round correction scheme focuses on the pre-correction of the attention inference node of the conflict propagation, and the third round correction scheme focuses on the adaptability adjustment of the corrected inference conclusion and the surrounding knowledge units, each round of correction scheme includes node adjustment direction, association relationship correction requirement, and inference step optimization suggestion, including: For the first round of conflict correction scheme, based on the error reasons of the conflict inference conclusion in the stage-by-stage conflict tracing results, if the inference logic is wrong, the inference logic chain based on the knowledge unit association relationship is reconstructed, the error inference step order is corrected, the missing inference link is supplemented, and the adjustment direction of the inference node is determined, the specific operations including replacing the error node or adding a connection node, determining the correction requirement of the association relationship, the specific operations including strengthening the association relationship whose association strength does not meet the current inference link requirement or correcting the wrong association, determining the optimization suggestion of the inference step, the specific operations including splitting the complex step or merging the redundant step; if the knowledge reference deviation, the knowledge unit adapted to the inference target is re-matched, the replacement direction of the knowledge unit is determined, the specific operations including replacing the edge adaptation knowledge unit with the core adaptation knowledge unit, the re-establishment requirement of the association relationship is determined, the specific operations including establishing a new support association or supplementing the association, the adaptability adjustment suggestion of the inference step is determined, the specific operations including adjusting the knowledge unit reference order; For the second round of conflict correction scheme, based on the attention reasoning node in the conflict propagation prediction result, the correlation degree of the attention reasoning node and the conflict reasoning conclusion is analyzed, and the pre-correction measures are formulated for each attention reasoning node. If the attention reasoning node is the direct subsequent node of the conflict reasoning conclusion, the reasoning basis of the attention reasoning node is adjusted in advance, including replacing the cited knowledge unit, determining the pre-adjustment direction of the node, determining the pre-correction requirement of the correlation, including adjusting the correlation strength of the correlation that may be affected by the conflict to the preset safety range, determining the pre-optimization suggestion of the reasoning step, including inserting the verification step in advance; if the attention reasoning node is an indirectly associated node of the conflict reasoning conclusion, the correlation strength between the attention reasoning node and other nodes is adjusted to reduce the conflict propagation probability, the adjustment direction of the node correlation strength is determined, the protection requirement of the correlation is determined, and the buffering suggestion of the reasoning step is determined, including adding an intermediate verification link; For the third round of conflict correction scheme, based on the reasoning conclusion after the first two rounds of correction, the adaptability of the corrected reasoning conclusion and the surrounding knowledge unit is analyzed. If the corrected conclusion and the surrounding knowledge unit have a correlation strength that does not meet the requirements, the correlation between the two is strengthened, including supplementing intermediate knowledge units to establish new correlations, determining the strengthening direction of the correlation, determining the adaptability adjustment requirement of the node, including fine-tuning the node content to improve the adaptability, and determining the perfect suggestion of the reasoning step, including adding an association verification step; if the corrected conclusion and the surrounding knowledge unit have redundant associations, the correlation is simplified, including deleting invalid associations or merging duplicate associations, determining the simplification direction of the correlation, determining the simplification requirement of the node, and determining the optimization suggestion of the reasoning step, including merging the association verification step; Integrate the specific content of each round of conflict correction scheme, mark the execution order of each round of conflict correction scheme, the applicable reasoning stage, the expected correction effect and the verification standard, and form a multi-round conflict correction scheme document.

8. The method of claim 1, wherein the method further comprises: The reasoning path iteration optimization is performed based on the large model reasoning process after the conflict resolution, the fit degree of the reasoning path and the knowledge unit correlation and the reasoning efficiency are evaluated, the jump order of the reasoning node and the knowledge unit reference priority are adjusted, the optimized reasoning path, the knowledge graph application record and the conflict correction information are integrated, and the large model logical reasoning optimization result is obtained, including: Extract the complete reasoning path in the large model reasoning process after the conflict resolution, split it into initial stage path, intermediate stage path and final stage path according to the reasoning stage, each path contains reasoning node sequence, correlation type between nodes and knowledge unit reference record, and get the complete reasoning path in stages; A reasoning path evaluation index knowledge base is constructed to quantitatively evaluate the complete reasoning path in stages to obtain the path evaluation result in stages. The reasoning path evaluation index knowledge base includes the fit degree index of the path and the knowledge unit correlation and the reasoning efficiency index. The fit degree index includes the core correlation proportion and the correct correlation rate, and the reasoning efficiency index includes the number of reasoning steps and the node jump time consumption. Based on the evaluation results of the staged path, if the fit index does not reach the preset standard, the jump order of the reasoning node is adjusted, the node jump direction corresponding to the core association of the knowledge unit is preferentially selected, and the core association proportion is made to meet the preset standard; if the reasoning efficiency index does not reach the preset standard, the knowledge unit reference priority is adjusted, the knowledge unit with a search response time meeting the preset reasoning stage standard and an association strength meeting the requirement of the current reasoning link in the knowledge unit association relationship network is preferentially referenced, the number of reasoning steps is controlled within the preset reasoning stage step range, and the node jump time consumption is controlled within the preset reasoning stage time consumption range, to obtain an adjusted staged reasoning path; The adjusted staged reasoning path is iteratively verified, a complete reasoning process is re-executed, and based on the reasoning path obtained after re-execution, quantitative evaluation is performed again to obtain new path evaluation results; it is judged whether the new path evaluation results meet the preset standard; if not, the operation of adjusting the jump order of the reasoning node and the reference priority of the knowledge unit is performed again, and then iterative verification is performed again; until the path evaluation results meet the requirement, an iteratively optimized staged reasoning path is obtained; Knowledge graph application records in the entire reasoning process are collected, including the types of the knowledge units referenced in each stage, the types of the association relationships, the number of applications, and the application effects, and the knowledge graph application list is arranged according to the reasoning stages; conflict correction information is summarized, including the content, execution effect, verification result and adjustment record of each round of conflict correction scheme, and the conflict correction report is arranged according to the correction rounds; The iteratively optimized staged reasoning path, the knowledge graph application list and the conflict correction report are integrated, the reasoning target achievement situation analysis result is supplemented, the big model logical reasoning optimization result containing all the key information is formed, the reasoning target achievement situation analysis result includes the fit degree of the reasoning conclusion and the target and the utilization efficiency of the knowledge graph in the reasoning process, the reasoning optimization effect summary is supplemented, and the reasoning optimization effect summary includes the fit degree improvement amplitude and the reasoning efficiency improvement proportion.

9. The method of claim 5, wherein the method further comprises: The adaptation degree of the knowledge unit and the reasoning task is calculated based on the keyword matching statistical result of the knowledge unit, the adaptation degree calculation combines the keyword matching quantity and the matching depth, the keyword matching quantity is positively correlated with the adaptation degree, and the adaptation degree classification standard is set, the knowledge unit directly supporting the realization of the reasoning target is marked as a core adaptation knowledge unit, the knowledge unit supplementing the function of the core adaptation knowledge unit is marked as an auxiliary adaptation knowledge unit, and the knowledge unit perfecting the details of the reasoning process is marked as an edge adaptation knowledge unit, including: The keyword matching weight parameter is set, the matching weight parameter of the reasoning target keyword is higher than that of the reasoning premise keyword, the weight parameter of the accurate matching is higher than that of the semantic similar matching, the accurate matching of the reasoning target keyword is set with the weight parameter of the first association level, the semantic similar matching of the reasoning target keyword is set with the weight parameter of the second association level, the accurate matching of the reasoning premise keyword is set with the weight parameter of the third association level, and the semantic similar matching of the reasoning premise keyword is set with the weight parameter of the fourth association level; The keyword matching statistical results of each knowledge unit are weighted calculated, the number of each matching type in the knowledge unit and the average matching depth are counted, the matching types include reasoning target accurate matching, reasoning target semantic similar matching, reasoning premise accurate matching, reasoning premise semantic similar matching, the number of each matching type is multiplied by the average matching depth, then multiplied by the corresponding weight parameter and summed to obtain the initial adaptation degree of the knowledge unit; The initial adaptation degree is normalized to map the adaptation degree value to a preset numerical interval, eliminating the influence of the content length difference of different knowledge units on the adaptation degree, and obtaining the normalized adaptation degree of the knowledge unit; The classification standard of the adaptation degree is set, the complexity of the reasoning task and the scale of the knowledge graph are referred to, the division basis of the core adaptation knowledge unit, the auxiliary adaptation knowledge unit and the edge adaptation knowledge unit is determined, the knowledge unit directly supporting the realization of the reasoning target is divided into the core adaptation knowledge unit, the knowledge unit supplementing the function of the core adaptation knowledge unit is divided into the auxiliary adaptation knowledge unit, the knowledge unit perfecting the details of the reasoning process is divided into the edge adaptation knowledge unit, and the rest of the knowledge unit is divided into the non-adaptation knowledge unit, which is excluded from the reasoning process reference range; The normalized adaptation degree of each knowledge unit is compared with the adaptation degree classification standard to determine the adaptation type of each knowledge unit, the comparison results of the normalized adaptation degree, the classification standard and the adaptation type mark of each knowledge unit are recorded to form the preliminary adaptation annotation results of the knowledge unit.

10. A large model logic reasoning optimization system combined with a knowledge graph, characterized in that, The large model logical reasoning optimization system combined with the knowledge graph includes a processor and a memory, the memory and the processor are connected, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to realize the large model logical reasoning optimization method combined with the knowledge graph in any one of claims 1-9.

Images