An intelligent decision-making method and system based on a knowledge graph

Abstract

This invention relates to the field of intelligent decision-making technology, and discloses an intelligent decision-making method and system based on knowledge graphs. The method includes acquiring input data of a decision-making scenario, matching it with similar cases in a business knowledge base, extracting factual information, industry experience, and historical rules to form a set of knowledge nodes; analyzing the logical dependencies and temporal consistency of nodes to construct a complete logical chain; evaluating the contribution of nodes to filter key node sequences, converting them into sentence fragments and adjusting their coherence, detecting and correcting logical contradictions; integrating stable expression results with decision conclusions, and generating a complete explanatory text. This method can achieve interpretability and logical consistency in intelligent decision-making, meeting the multiple requirements of intelligent decision-making for transparency, credibility, and compliance.

Description

Technical Field

[0001] This invention relates to the field of intelligent decision-making technology, and in particular to an intelligent decision-making method and system based on knowledge graphs. Background Technology

[0002] Currently, in the field of intelligent decision-making technology, with the continuous increase in the complexity of business scenarios and the continuous strengthening of decision compliance requirements, knowledge graph-based decision support, as a core means to improve the credibility of judgments, has its interpretability and logic directly related to the acceptance of decision results and the efficiency of business implementation.

[0003] Existing intelligent decision-making methods in the industry mainly rely on single data-driven approaches or simple rule matching. For example, they may directly output decision results through machine learning models, generate explanatory text using fixed templates, or ignore the deep logical dependencies between nodes in a knowledge graph. However, this approach is clearly inadequate in complex operating environments. Single data-driven approaches lack business knowledge support, resulting in unconvincing explanations; fixed templates have poor adaptability and struggle to handle diverse decision-making scenarios; and they fail to fully utilize the interconnected advantages of knowledge graphs. Especially in complex decision-making scenarios with multiple coupled factors, it is difficult to clearly trace the basis for decisions, leading to contradictory explanations or logical breaks, and ultimately failing to meet the requirements for compliant and trustworthy decision-making.

[0004] In summary, existing technologies struggle to achieve interpretability and logical consistency in intelligent decision-making, failing to meet the multiple demands for transparency, credibility, and compliance in the field of intelligent decision-making technology. Summary of the Invention

[0005] This invention provides a knowledge graph-based intelligent decision-making method and system to achieve interpretability and logical consistency in intelligent decision-making, and to meet the multiple requirements of the field of intelligent decision-making technology for decision transparency, credibility and compliance.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides an intelligent decision-making method based on knowledge graphs, comprising: Obtain the input data for the current decision-making scenario; The input data is matched with similar cases in a preset business knowledge base, and the corresponding factual information, industry experience and historical rules are extracted and structured and integrated to obtain a set of matched knowledge nodes. Analyze the logical dependencies and temporal consistency among the knowledge nodes in the knowledge node set, determine the connection order of the nodes, and construct a complete logical chain structure according to the connection order; Based on the topological depth, node betweenness, and propagation probability of each node in the complete logical chain structure, the contribution of each knowledge node is evaluated. If the contribution exceeds a preset contribution judgment threshold, the corresponding knowledge node is retained to obtain a key knowledge node sequence. Based on the sequence of key knowledge nodes, the logical relationships between each node are transformed into sentence fragments, and semantically complete fragments are collected to form an initial set of expressions; Based on the initial set of expressions, the semantic correlation between each of the statement fragments is calculated. The order of the statements is determined by combining the priority of business logic and the temporal sequence, and the preset logical connectors are matched to obtain a coherent and natural language draft. Based on the language draft, detect logical contradictions in the preceding and following statements. If contradictions exist, correct the corresponding segments to obtain a stable statement result. Based on the complete logical chain structure, the decision conclusion is derived, and the semantic features and logical paths of the stable expression result and the decision conclusion are integrated to generate a complete explanatory text.

[0007] Secondly, the present invention provides an intelligent decision-making system based on knowledge graphs, comprising: The data acquisition module is used to acquire input data for the current decision-making scenario; The knowledge matching module is used to match the input data with similar cases in a preset business knowledge base, extract corresponding factual information, industry experience and historical rules and integrate them in a structured manner to obtain a set of matched knowledge nodes; The chain construction module is used to analyze the logical dependencies and temporal consistency between knowledge nodes in the knowledge node set, determine the connection order of the nodes, and construct a complete logical chain structure according to the connection order. The node filtering module is used to evaluate the contribution of each knowledge node based on the topological depth, node betweenness and propagation probability of each node in the complete logical chain structure. If the contribution exceeds a preset contribution judgment threshold, the corresponding knowledge node is retained to obtain a key knowledge node sequence. The expression generation module is used to convert the logical relationships between the nodes into sentence fragments based on the sequence of key knowledge nodes, and to collect semantically complete fragments to form an initial expression set. The coherence adjustment module is used to calculate the semantic relevance between each of the statement fragments based on the initial set of statements, determine the order of statements by combining business logic priority and temporal rules, and match preset logical connectors to obtain a coherent and natural language draft. The contradiction correction module is used to detect logical contradictions in the preceding and following statements based on the language draft. If contradictions exist, the corresponding segments are corrected to obtain a stable statement result. The text generation module is used to deduce decision conclusions based on the complete logical chain structure, integrate the semantic features and logical paths of the stable expression results and the decision conclusions, and generate complete explanatory text.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention obtains input data from decision-making scenarios, matches it with similar cases in the business knowledge base, extracts factual information, industry experience and historical rules and integrates them in a structured manner, analyzes the logical dependencies and time consistency between knowledge nodes, constructs a complete logical chain, breaks through the limitations of traditional single data-driven approaches that lack business knowledge support, mines the correlation features of multi-source knowledge, eliminates the interference of scattered heterogeneous information, provides high-precision logical foundation support for intelligent decision-making, effectively improves the credibility of decision-making basis, and solves the problem of explanations lacking business persuasiveness.

[0009] (2) This invention evaluates the contribution of knowledge nodes based on logical chains, selects key node sequences, transforms the logical relationship of nodes into sentence fragments, calculates semantic relevance and matches logical connectors with business priority and temporal patterns, breaks through the limitations of poor adaptability of traditional fixed templates, accurately captures the adaptation features of knowledge logic and natural language, provides multi-dimensional basis for the generation of explanatory text, significantly improves the coherence and comprehensibility of decision explanation, and makes up for the defects of rigid explanation and logical breakage in existing technologies.

[0010] (3) This invention detects and corrects logical contradictions in language drafts, derives decision conclusions based on complete logical chains, integrates the semantic features and logical paths of stable expression results and decision conclusions, generates complete explanatory texts, breaks through the limitations of traditional lack of logical consistency verification, provides a traceable and explanatory complete basis for decision-making, solves the problems of explanatory contradictions and insufficient compliance, takes into account the transparency and credibility of decision-making, and meets the dual requirements of compliance and credibility in the field of intelligent decision-making. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of a knowledge graph-based intelligent decision-making method provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of a knowledge graph-based intelligent decision-making system structure provided in the second embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Reference Figure 1 The first embodiment of the present invention provides an intelligent decision-making method based on knowledge graphs, comprising the following steps: S101, Obtain the input data for the current decision-making scenario; S102, match the input data with similar cases in the preset business knowledge base, extract the corresponding factual information, industry experience and historical rules and integrate them in a structured manner to obtain a set of matched knowledge nodes; S103, Analyze the logical dependencies and temporal consistency between knowledge nodes in the knowledge node set, determine the connection order of the nodes, and construct a complete logical chain structure according to the connection order; S104. Based on the topological depth, node betweenness and propagation probability of each node in the complete logical chain structure, evaluate the contribution of each knowledge node. If the contribution exceeds a preset contribution judgment threshold, retain the corresponding knowledge node to obtain a key knowledge node sequence. S105, Based on the sequence of key knowledge nodes, the logical relationships between each node are transformed into sentence fragments, and semantically complete fragments are collected to form an initial expression set; S106. Based on the initial set of expressions, calculate the semantic correlation between each of the statement fragments, determine the order of statements by combining business logic priority and temporal rules, and match preset logical connectors to obtain a coherent and natural language draft. S107. Based on the language draft, detect logical contradictions in the preceding and following statements. If contradictions exist, correct the corresponding segments to obtain a stable statement result. S108, based on the complete logical chain structure, derive the decision conclusion, integrate the semantic features and logical path of the stable expression result and the decision conclusion, and generate a complete explanatory text.

[0014] In step S101, obtaining the input data for the current decision-making scenario includes: Collect the raw input information of the current decision-making scenario, including business parameters, scenario attributes and constraints; The original input information is denoised to remove invalid data and duplicate records, resulting in cleaned input data. The cleaned input data is standardized in format and dimensionally regularized to construct structured input data; Extract the core features of the structured input data and determine the core features as the input data for the current decision-making scenario.

[0015] It should be noted that, firstly, when collecting the raw input information for the current decision-making scenario, business parameters are obtained through business system interfaces and data acquisition sensors (such as financial data acquisition sensors and environmental status sensors). Scenario attributes are collected through the system's built-in scenario classification and recognition module and a manual-assisted input terminal. Constraints are extracted from pre-set business rule documents and compliance requirement manuals. Business parameters cover quantitative indicators (such as amount, proportion, and quantity) and qualitative indicators (such as industry type and business status). Scenario attributes include the scenario's domain, time dimension, and spatial scope. Constraints clearly define the boundary limitations of the decision-making process (such as time limit and amount limit).

[0016] For example, in corporate credit decision-making scenarios, the business parameters collected include debt-to-asset ratio, current ratio, and annual revenue. The scenario attributes are manufacturing, established for 5 years, and located in East China. The constraints are that the loan term does not exceed 3 years and the single loan amount does not exceed 50 million yuan.

[0017] Next, the original input information undergoes noise reduction processing to remove invalid and duplicate data. Invalid data removal employs a dual filtering mechanism based on statistical rules and business logic. The statistical rules define reasonable value ranges for each field, using the field mean ± 3 standard deviations as the basic reasonable range based on statistical results from similar business data over the past three years. For special industries (such as emerging technology industries), the range can be extended to mean ± 4 standard deviations. Those skilled in the art can flexibly adjust this according to the business type. Business logic filtering verifies the logical consistency between fields, such as the reasonable correlation between revenue growth rate and profit growth rate. Duplicate record deletion uses a field hash comparison method, calculating the global hash value of each record, retaining the first occurrence of the record, and deleting subsequent entries with duplicate hash values ​​to ensure data uniqueness.

[0018] For example, if a loan application has a debt-to-asset ratio of 180%, which exceeds the manufacturing average by ±3 standard deviations (the reasonable range is 0-120%), it will be deemed invalid and removed. If two application records have completely identical core fields and the same hash value, the first record will be kept and the second record will be deleted.

[0019] Next, the cleaned input data underwent format standardization and dimensional normalization. When constructing structured input data, format standardization unified the representation of various types of fields. Date fields were standardized to YYYY-MM-DD format, numeric fields were uniformly retained to two decimal places, text fields were uniformly converted to lowercase and leading and trailing spaces were removed, and coded fields were uniformly mapped according to a preset coding dictionary (e.g., the industry type "Manufacturing" was mapped to "01"). Dimensional normalization used mean imputation to handle a small number of missing values. The mean of each field was calculated based on historical data of the same scenario and scale, and missing items were imputed. For high-dimensional data, principal component analysis (PCA) was used for dimensionality reduction, selecting the top k principal components whose cumulative contribution rate first exceeded 85%, ensuring data dimensionality uniformity and preserving core information. For example, the establishment date of a certain data is 2020 / 8 / 5, which is standardized to 2020-08-05; the annual revenue field is missing, so it is filled with the average annual revenue of 80 million yuan of companies in the same industry with the same establishment year; the 30-dimensional financial indicators are reduced to 15 dimensions through PCA, with a cumulative contribution rate of 88%, thus constructing structured input data.

[0020] First, features are extracted from the structured input data to form the input data representing the current decision-making scenario. The feature extraction uses the Gradient Boosting Tree (XGBoost) algorithm, which effectively captures the non-linear correlations and importance differences between features. The training set contains over 80,000 structured data points from historical decision-making scenarios, covering multiple business areas such as credit, supply chain, and risk control. These are divided into training and validation sets in a 7:3 ratio. The algorithm is set with a learning rate of 0.05, a maximum tree depth of 8, a minimum number of splits of 6 samples, a minimum number of leaf nodes of 3 samples, regularization parameters lambda of 0.1 and alpha of 0.05, and a logarithmic loss function. During training, the AUC value of the validation set is monitored in real time. Iteration stops when the AUC value fluctuates less than 0.002 for 12 consecutive rounds. Finally, the top 25 features by importance score are extracted to form a fixed-dimensional feature vector, which serves as the input data representing the current decision-making scenario.

[0021] In step S102, the input data is matched with similar cases in a preset business knowledge base, and corresponding factual information, industry experience, and historical rules are extracted and structurally integrated to obtain a set of matched knowledge nodes, including: The input data is converted into a feature vector, and the similarity between the feature vector and the feature vector of historical cases in the preset business knowledge base is calculated. If the similarity exceeds the preset similarity judgment threshold, the corresponding historical case is marked as a candidate similar case, and objective factual information, industry experience text and historical rule entries are extracted from the candidate similar cases. The factual information, the industry experience text, and the historical rule entries are semantically aligned and categorized. The categorized information is then structurally integrated to obtain a set of matching knowledge nodes.

[0022] It should be noted that, firstly, when converting the input data into feature vectors, the features of each dimension of the structured input data are first subjected to min-max normalization, mapped to the [0,1] interval, to eliminate interference from features of different magnitudes. The Word2Vec model is used to encode textual features (such as industry type and business description), using the CBOW architecture, with a word vector dimension of 64, a context window size of 4, and a negative sampling number of 3. The training set contains over 50,000 business text data entries (covering various decision-making scenarios). The initial learning rate is 0.02, decreasing by 0.1 every 25 rounds until it reaches 0.001. Iteration stops when the loss fluctuation is less than 0.0005 for 10 consecutive rounds. Numerical features directly retain their normalized values ​​and are concatenated with the text encoding vector to form a feature vector with a fixed dimension of 128. Similarity calculation uses the cosine similarity algorithm, quantifying the similarity between the input data feature vector and the historical case feature vectors in the business knowledge base by calculating the cosine of the angle. The value ranges from 0 to 1, with values ​​closer to 1 indicating higher similarity.

[0023] If the similarity exceeds a preset similarity threshold, the corresponding historical case is marked as a candidate similar case. The similarity threshold is set based on the statistical effect of historical case matching. Analyzing decision-making data from the past two years, it was found that when the similarity exceeds 0.7, the case's decision-making logic and the input scenario have an adaptation rate of over 80%. Therefore, the basic threshold is set at 0.7. For complex decision-making scenarios (such as supply chain optimization with multiple coupled factors), the threshold can be increased to 0.75 to ensure the accuracy of matched cases; for simple scenarios (such as routine approvals), it can be decreased to 0.65 to expand the candidate range. When extracting information from candidate similar cases, objective factual information is extracted through keyword matching (such as "number of overdue payments" and "revenue scale"); industry experience text uses semantic parsing technology (based on the LSTM model) to extract core viewpoints; and historical rule entries are extracted through regular expression matching of rule sentences such as "if...then..." and "when...when...". For example, if the similarity of a candidate similar case is 0.85, which exceeds the basic threshold of 0.7, the extracted objective factual information is "two overdue payments in the past six months", the industry experience text is "strict control of credit limits is required during the downturn of the manufacturing industry", and the historical rule entry is "credit rating will be downgraded if there are ≥ two overdue payments".

[0024] Next, when semantically aligning factual information, industry experience text, and historical rule entries, the BERT model was used to achieve semantic unification across cases. The base version of the BERT model was selected, with 768 hidden layer dimensions and 12 attention heads. The training set contained over 30,000 labeled business knowledge texts, divided into training and validation sets in an 8:2 ratio. The optimizer used was AdamW, with a learning rate of 2e-5, 30 training epochs, GELU activation function, and cross-entropy loss function. After semantic alignment, the knowledge was categorized and labeled according to knowledge type, into three categories: factual, experience, and rule. Structured integration used key-value pairs, with each knowledge entry containing a tag, content, and source case identifier, generating a set of matching knowledge nodes that supports fast retrieval by type.

[0025] For example, after semantic alignment and classification labeling, a certain knowledge node set includes fact-type nodes such as "two overdue payments in the past six months", experience-type nodes such as "strict control of credit limits is required during the downturn of the manufacturing industry", and rule-type nodes such as "credit rating will be downgraded if there are ≥ two overdue payments", which fully covers the various types of knowledge required for decision-making.

[0026] In step S103, analyzing the logical dependencies and temporal consistency between knowledge nodes in the knowledge node set, determining the connection order of the nodes, and constructing a complete logical chain structure according to the connection order includes: Calculate the interaction strength between each knowledge node in the knowledge node set, and construct the dependency matrix of each knowledge node based on the interaction strength; An initial topology is generated based on the dependency matrix, and the potential logical connection paths of each knowledge node are traced based on the initial topology to obtain a set of candidate tracing paths. Extract the timestamp information of each candidate tracing path in the candidate tracing path set, verify the time sequence of nodes in the path, and if a candidate tracing path has a time sequence conflict, remove the path and retain the time sequence compliant paths to form a compliant path set. By integrating the common nodes and logical relationships in the set of compliant paths, a reasoning subgraph is constructed. Traverse the reasoning subgraph, sort out the node connection order according to logical causal relationship and time sequence, and form a complete logical chain structure.

[0027] It should be noted that, firstly, when calculating the interaction strength between knowledge nodes based on the knowledge node set, the interaction strength is obtained by weighted summation of co-occurrence probability and semantic relevance. Co-occurrence probability is calculated based on the frequency of co-occurrence of knowledge nodes in historical decision-making cases, using data from over 100,000 cases in the business knowledge base over the past three years; higher frequency of occurrence indicates a higher co-occurrence probability. Semantic relevance is calculated using the BERT model, specifically the base version with 768 hidden layer dimensions and 12 attention heads. The training set contains over 50,000 annotated business knowledge-related texts, divided into training and validation sets in an 8:2 ratio. The optimizer is AdamW, the learning rate is 2e-5, training is conducted for 30 epochs, the activation function is GELU, and the loss function is cross-entropy loss. The output similarity value is the semantic relevance. The weights are assigned as 0.5 for co-occurrence probability and 0.5 for semantic relevance. Both are normalized to the [0,1] interval and then weighted summed to obtain the interaction strength (value 0-1). The dependency matrix is ​​an n×n matrix (n is the number of knowledge nodes), and the matrix elements are the interaction strength between two corresponding nodes. The diagonal elements are set to 0 (nodes have no dependencies on themselves).

[0028] For example, in a credit scenario, the co-occurrence probability of "two overdue payments in the past six months" and "downgrade of credit rating" is 0.8, the semantic relevance is 0.9, the interaction strength is 0.8×0.5 + 0.9×0.5=0.85, and the corresponding element value in the dependency matrix is ​​0.85.

[0029] Subsequently, when generating the initial topology based on the dependency matrix, an adjacency matrix to graph structure method was used to establish directed edges (edge ​​weights equal to interaction strengths) between nodes with interaction strengths exceeding 0.3, forming the initial topology. A depth-first search (DFS) algorithm was employed to trace potential logical connection paths, with a maximum path length of 5 (to avoid logical breaks due to excessively long paths). Starting from each node, all reachable paths were traversed to obtain a set of candidate source paths. The algorithm's training set contained over 20,000 historical logical path data points. By adjusting the search depth and pruning strategies, the path recall rate was stabilized at over 88%.

[0030] For example, the initial topology includes nodes A (overdue twice), B (downward trend in manufacturing), C (reduced credit line), and D (strict credit limit control). The DFS algorithm traces the candidate source paths to: A→C, A→B→C, and B→D→C.

[0031] Subsequently, timestamp information for each candidate tracing path is extracted. When verifying the chronological order of nodes within a path, timestamps are extracted from the source cases of the knowledge nodes, recording the event occurrence time (accurate to the day) corresponding to each node. The verification rule is that the timestamps of nodes within a path must increase in ascending order according to logical causal relationships; that is, the timestamp of the cause node must be earlier than the timestamp of the result node. If a node's timestamp is later than a subsequent result node, it is determined to be a chronological conflict, and the path is directly eliminated. All chronologically compliant paths are retained to form a set of compliant paths, ensuring that the paths conform to objective time patterns.

[0032] For example, in a candidate path A (2024-03-10) → B (2024-02-15) → C (2024-03-15), B's time is earlier than A's, which violates the causal time order and is judged as a conflicting path and removed; the path A (2024-03-10) → C (2024-03-15) is compliant with the time order and is retained in the compliant path set.

[0033] Next, when constructing the inference subgraph by integrating common nodes and logical relationships from the compliant path set, common nodes are matched and merged according to their node identifiers, and identical logical relationships (such as the causal relationship A→C) are merged by averaging the edge weights. The GraphSAGE model is used to optimize the graph structure of the merged paths. This model has 3 aggregation layers, a hidden layer dimension of 64, and uses ReLU as the activation function. The training set contains over 30,000 business inference subgraph data points. The initial learning rate is 0.01, decaying by 0.1 every 15 rounds until it reaches 0.001. Training stops when the loss fluctuation is less than 0.001 for 10 consecutive rounds. The optimized inference subgraph retains core logical relationships and common nodes, while removing redundant edges and isolated nodes.

[0034] For example, in the set of compliant paths A→C, A→B→C, B→D→C, after merging the common nodes A, B, C, D and their logical relationships, the constructed reasoning subgraph contains directed edges A→B, A→C, B→C, B→D, and D→C, clearly presenting the multiple logical connections between nodes.

[0035] Finally, the reasoning subgraph is traversed, and the node connection order is sorted according to logical causal relationships and chronological order to form a complete logical chain structure. First, the edges are sorted from high to low according to their weights (interaction strength), prioritizing the retention of strongly related logical paths. Then, the nodes are sorted from small to large according to their timestamps to ensure that the chain conforms to the temporal sequence. During the sorting process, if there are multiple parallel logical paths, they are integrated according to the hierarchical relationship of "cause node → intermediate node → result node" to form a tree-like logical structure, and then expanded into a linear chain according to the hierarchical order.

[0036] For example, after sorting out the reasoning subgraph, the complete logical chain formed is: Downward trend in manufacturing (2024-02-15) → Two overdue payments in the past six months (2024-03-10) → Strict control of credit limit (2024-03-12) → Downgrade of credit rating (2024-03-15), which not only reflects the causal relationship, but also follows the chronological order.

[0037] In step S104, the contribution level of each knowledge node is evaluated based on the topological depth, node betweenness, and propagation probability of each node in the complete logical chain structure. If the contribution level exceeds a preset contribution judgment threshold, the corresponding knowledge node is retained to obtain a key knowledge node sequence, including: Traverse the complete logical chain structure and calculate the topological depth, node betweenness, and propagation probability of each knowledge node; The contribution score of each knowledge node is obtained by weighting and summing the topology depth, the node betweenness, and the propagation probability using preset weighting coefficients. The contribution score is compared with a preset contribution judgment threshold. If the contribution score exceeds the contribution judgment threshold, the corresponding knowledge node is marked as a core node. Based on the temporal relationship and logical order in the complete logical chain structure, the core nodes are reorganized and sorted to obtain a sequence of key knowledge nodes.

[0038] It should be noted that, firstly, when traversing the complete logical chain structure to calculate the topological depth, betweenness, and propagation probability of each knowledge node, the topological depth is based on the starting node of the logical chain, with the starting node's topological depth set to 1. Subsequent nodes increase sequentially according to the shortest path length; that is, nodes directly connected to the starting node have a depth of 2, and indirectly connected nodes have depths accumulated according to path length, intuitively reflecting the node's hierarchical position in the logical chain. The betweenness centrality algorithm of the NetworkX library is used to calculate the betweenness centrality, counting the proportion of all shortest paths passing through the node to the total number of shortest paths in the chain. A higher proportion indicates a more significant role for the node as a "logical hub." Before calculation, the logical chain structure is converted into an undirected graph to ensure the comprehensiveness of path statistics. The propagation probability is based on the interaction strength between nodes and is calculated using the PageRank algorithm. The damping coefficient of this algorithm is set to 0.85, the number of iterations is set to 100, and the training set contains propagation data of over 50,000 historical logical chains. By adjusting the parameters, the propagation probability accurately reflects the node's ability to transmit influence to subsequent nodes. All three indicators are mapped to the [0,1] interval through min-max normalization to eliminate differences in magnitude.

[0039] For example, in the logical chain of credit decision-making, "two overdue payments in the past six months" is a node directly connected to the starting node with a topological depth of 2; this node is a necessary node in the three shortest paths with a node betweenness of 0.6; calculated by the PageRank algorithm, the propagation probability is 0.75, and all three indicators remain at their original values ​​after normalization.

[0040] Next, pre-defined weighting coefficients are assigned to topology depth, node betweenness, and propagation probability. When performing a weighted summation to obtain the contribution score of each knowledge node, the weighting is based on the degree of influence of each indicator on the decision conclusion. The node betweenness weight is set to 0.4 because it directly reflects the core hub role of the node in logical transmission and has the strongest correlation with the decision basis; the topology depth weight is set to 0.3, as nodes at higher levels have a more fundamental influence on subsequent logic; the propagation probability weight is set to 0.3, reflecting the efficiency of the node's influence on the decision conclusion. This weight combination has been verified by more than 10,000 historical decision cases and can accurately distinguish the differences in node contributions, making the contribution score of core nodes significantly higher than that of redundant nodes. During the weighted summation, the three normalized indicators are multiplied by their corresponding weights, and the summation yields the contribution score, which ranges from 0 to 1. The larger the value, the higher the degree of contribution of the node to the decision. For example, a node with a topological depth of 0.6, a node betweenness of 0.8, and a propagation probability of 0.7 has a weighted sum of contribution scores of 0.6×0.3 + 0.8×0.4 + 0.7×0.3 = 0.18 + 0.32 + 0.21 = 0.71, which reflects that the node has a high decision contribution.

[0041] Subsequently, the contribution score is compared with a preset contribution judgment threshold. If the contribution score exceeds the threshold, the corresponding knowledge node is marked as an important node. The contribution judgment threshold is set based on the contribution scores of core nodes in historical decisions. Analysis of decision data from the past two years revealed that the contribution scores of important nodes generally exceed 0.6, while nodes below this value are mostly redundant information or secondary related nodes. Therefore, the basic threshold is set at 0.6. In complex decision-making scenarios (such as supply chain optimization with multiple coupled factors), the threshold can be increased to 0.65 to improve screening accuracy; in simple decision-making scenarios (such as routine approvals), it can be decreased to 0.55. Those skilled in the art can adjust it within the range of 0.55-0.65 according to the complexity of the scenario.

[0042] For example, a node with a contribution score of 0.71, which exceeds the basic threshold of 0.6, is marked as an important node; another node with a contribution score of 0.58, which is below the threshold, is judged as a minor node and is not retained.

[0043] In this implementation case, important nodes are reorganized and sorted according to the temporal relationship and logical order in the complete logical chain structure. When obtaining the sequence of key knowledge nodes, they are first sorted by their timestamps from smallest to largest to ensure that the sequence conforms to the chronological order of events. Then, they are adjusted according to logical causal relationships, so that cause nodes come before result nodes. If parallel logical paths exist, they are sorted by contribution scores from highest to lowest, prioritizing the logical order of nodes with high contributions. The sorted sequence follows both objective temporal order and logical connections, ensuring the systematic nature of the decision-making basis.

[0044] For example, the important nodes are sorted by timestamp as "Manufacturing downturn" (2024-02-15), "Two overdue payments in the past six months" (2024-03-10), and "Credit rating downgraded" (2024-03-15). After adjusting for logical causal relationships, the final sequence of key knowledge nodes is "Manufacturing downturn - Two overdue payments in the past six months - Credit rating downgraded".

[0045] In step S105, the step of converting the logical relationships between the nodes into sentence fragments based on the key knowledge node sequence, and collecting semantically complete fragments to form an initial expression set, includes: Traverse the sequence of key knowledge nodes and extract the logical relationship types between adjacent nodes. The logical relationship types include causal relationships, conditional relationships, and progressive relationships. Based on the logical relation type, match the corresponding syntactic template to construct the statement skeleton; The entity information of the key knowledge nodes is filled into the statement skeleton to generate a statement fragment; The semantic integrity of the statement fragment is checked, and an integrity score is calculated. If the integrity score exceeds a preset integrity threshold, the statement fragment is retained, and all qualified fragments are collected to form an initial expression set.

[0046] It should be noted that, firstly, when traversing the key knowledge node sequence to extract the logical relationship types between adjacent nodes, a BiLSTM model is used to complete the relationship identification. This model can effectively capture sequence dependency features to adapt to logical relationship determination. The training set contains over 60,000 business texts labeled with logical relationships, covering causal, conditional, and progressive relationships, divided into training and validation sets in a 7:3 ratio. The model has two hidden layers, each with 256 nodes, using tanh as the activation function, Adam as the optimizer, an initial learning rate of 0.001, decaying by 0.1 every 20 rounds until reaching 0.0001, and cross-entropy loss as the loss function. Training stops when the accuracy fluctuation on the validation set is less than 0.002 for 15 consecutive rounds. By analyzing the semantic vector similarity, co-occurrence patterns, and implicit connection features of adjacent nodes, combined with the judgment rules obtained from training, the three types of logical relationships are accurately distinguished, ensuring the consistency and accuracy of relationship identification. For example, the key knowledge node sequence in the credit scenario, "Manufacturing downturn - 2 overdue payments in the past six months - credit rating downgrade", is analyzed by the model. The first two are causal, and the latter two are conditional, clearly defining the logical relationship between the nodes.

[0047] It's worth noting that when constructing the statement skeleton by matching the corresponding syntactic templates based on the logical relationship type, the syntactic templates are extracted from over 50,000 business domain specification texts and stored categorized by logical relationship type. Causal relationships are adapted to high-frequency templates such as "Because of X, Y results" and "X is the main cause of Y"; conditional relationships are adapted to templates such as "If X is true, then Y takes effect" and "When X is satisfied, Y is executed"; and progressive relationships are adapted to templates such as "Not only X, but also Y" and "X further triggers Y". Template selection prioritizes matching frequently used and semantically natural sentence structures to ensure that the generated statements conform to business expression habits and reduce the cost of understanding.

[0048] For example, the causal relationship is matched with the template "because of X, Y is caused", and the statement skeleton "because of X, Y is caused" is constructed; the conditional relationship is matched with the template "if X is true, then Y is effective", and the statement skeleton "if X is true, then Y is effective" is constructed.

[0049] In this embodiment, when filling entity information of key knowledge nodes into the sentence skeleton to generate sentence fragments, placeholders are assigned according to the role of the nodes in the logical relationship: cause nodes and condition nodes correspond to X in the template, and result nodes and progressive nodes correspond to Y in the template. During the filling process, word order and auxiliary words are automatically adjusted to ensure that the entity text matches the template's grammatical structure, avoiding semantic ambiguity or awkward sentence structure. For nodes containing multiple entities, they are filled sequentially according to semantic order to ensure smooth and coherent sentence expression.

[0050] For example, filling "Manufacturing downturn" in X and "2 overdue payments in the past six months" in Y will generate the statement fragment "Due to the manufacturing downturn, there have been 2 overdue payments in the past six months"; filling "2 overdue payments in the past six months" in X and "downgrade credit rating" in Y will generate the statement fragment "If the 2 overdue payments in the past six months are true, the credit rating will be downgraded".

[0051] Finally, semantic integrity is checked for sentence fragments to calculate integrity scores. If a sentence fragment exceeds a preset integrity threshold, it is retained. The semantic integrity detection uses the BERT-base model. The training set contains over 30,000 sentence fragments labeled with integrity tags, including 20,000 complete fragments and 10,000 incomplete fragments, divided into training and validation sets in an 8:2 ratio. The model output layer uses the sigmoid activation function to output an integrity score between 0 and 1. A higher score indicates a more complete semantic meaning. The integrity threshold is set based on the ROC curve analysis of the training set, taking the score of 0.8, where the Youden index is at its maximum, as the base threshold. For formal document scenarios (such as contracts and approval reports), this can be increased to 0.85, while for conversational business scenarios, it can be decreased to 0.75. Those skilled in the art can flexibly adjust it according to the needs of the scenario. All sentence fragments with scores exceeding the threshold are collected to form an initial set of expressions.

[0052] For example, the completeness score of "due to the downturn in manufacturing, there have been two overdue payments in the past six months" is 0.88, which exceeds the basic threshold of 0.8 and is retained; the score of "if there are two overdue payments in the past six months, the credit rating will be downgraded" is 0.83, which also meets the requirements. Together, they constitute the initial statement set.

[0053] In step S106, the step of calculating the semantic relevance between each statement fragment based on the initial set of expressions, determining the order of statements by combining business logic priority and temporal patterns, and matching preset logical connectors to obtain a coherent and natural language draft includes: Calculate the semantic association degree between any two sentence fragments in the initial set of expressions, and construct a semantic association matrix; Based on the semantic association matrix, core sentence fragments are identified, and the order of the core sentence fragments is determined by combining business logic priority and temporal patterns, generating a sentence order list. Based on the logical relationship of the sentence fragments in the word order list, preset logical connectors are matched, including causal, progressive, and adversative logical connectors. The logical connectors are embedded between corresponding sentence fragments and then pieced together in sequence to form a coherent and natural language draft.

[0054] It should be noted that, firstly, when calculating the semantic relevance between any two sentence fragments in the initial statement set, the Sentence-BERT model is used for precise quantification. This model is specifically optimized for sentence-level semantic similarity calculation and can effectively capture deep semantic relationships between sentences. The training set contains over 40,000 business scenario sentence pairs, covering areas such as credit, supply chain, and risk control, and is divided into training and validation sets in a 7:3 ratio. The model uses distilbert-base-nli-stsb-mean-tokens pre-trained weights. During fine-tuning, the batch size is set to 32, the initial learning rate is 0.0001, and the training epochs are 20. The optimizer used is AdamW, and the loss function is contrastive loss. During training, the Pearson correlation coefficient of the validation set is monitored in real time, and iteration stops when the fluctuation is less than 0.003 for 8 consecutive epochs. The semantic relevance is obtained by calculating the cosine similarity of the sentence embedding vectors, with a value ranging from 0 to 1. The closer the value is to 1, the stronger the semantic relationship. The semantic association matrix is ​​an n×n matrix (n is the number of sentence fragments), and the matrix elements are the semantic association degree between corresponding two fragments. The diagonal elements are set to 1 (the similarity between fragments is 1).

[0055] For example, the two fragments in the initial statement set, "Due to the downturn in manufacturing, there have been two overdue payments in the past six months" and "If there are two overdue payments in the past six months, the credit rating will be downgraded", have a semantic relevance of 0.85 calculated by the model, and the corresponding element value in the matrix is ​​0.85.

[0056] Subsequently, important sentence fragments are identified based on the semantic association matrix. The order is determined by combining business logic priority and temporal patterns. When generating the sentence order list, important sentence fragments are filtered by calculating the sum of their semantic associations with all other fragments. A higher sum indicates a more significant semantic hub role for that fragment in the set. A threshold for the sum of associations is set. Based on historical sentence sorting statistics, the basic threshold is set at 0.7. For semantically dense decision-making scenarios (such as multi-condition approval), this threshold can be increased to 0.75, and for simple scenarios, it can be decreased to 0.65. Those skilled in the art can adjust this flexibly. Business logic priority is sorted according to the contribution score of the knowledge node corresponding to the sentence fragment; the higher the contribution score, the higher the priority. Temporal patterns are sorted according to the event timestamps corresponding to the sentence fragments, from smallest to largest. Combining both, a preliminary sort is first made according to business logic priority, and then fine-tuned according to temporal patterns to ensure that the order conforms to both business importance and chronological order, thus generating the sentence order list.

[0057] For example, the total relevance of the three sentence fragments is 0.88, 0.75, and 0.62, respectively, and their contribution scores are 0.71, 0.65, and 0.60, respectively. The timestamps correspond to the event order as fragment 1 → fragment 2 → fragment 3, and the final generated sentence order list is fragment 1 → fragment 2 → fragment 3.

[0058] Next, based on the logical relationships between the sentence fragments in the word order list, when matching preset logical connectors, these connectors are stored according to their functions: causal connectors include "therefore," "thus," and "hence," progressive connectors include "further," "in addition," and "more importantly," and adversative connectors include "but," "however," and "however." The matching rules are based on the logical relationship type of adjacent sentences: causal relationships are preferentially matched with "therefore" and "thus," progressive relationships with "further" and "in addition," and adversative relationships with "but" and "however." If the logical relationship between sentences is unclear, high-frequency collocation connectors are selected based on semantic relevance: causal connectors are preferred when the relevance is higher than 0.8, progressive connectors are selected when it is between 0.6 and 0.8, and adversative connectors are selected when it is lower than 0.6. For example, in the word order list, fragment 1 and fragment 2 have a causal relationship, so "therefore" is matched; fragment 2 and fragment 3 have a progressive relationship, so "further" is matched, clearly indicating the logical connection between the sentences.

[0059] It should be noted that when embedding logical connectors between corresponding sentence segments and piecing them together in sequence to form a coherent and natural language draft, the embedding position is at the connection point of adjacent sentence segments. At the same time, the word order and auxiliary words should be adjusted to avoid grammatical errors or awkward expressions. During the splicing process, repetitive expressions between segments should be removed, and necessary transition words should be added to ensure that the overall language is fluent, logically coherent, and conforms to the expression habits of business scenarios.

[0060] For example, the following phrases in the sentence order list, "Due to the downturn in manufacturing, there have been two overdue payments in the past six months," "If the two overdue payments in the past six months are true, the credit rating will be downgraded," and "Reflects the high credit risk of the enterprise," can be combined with conjunctions to form "Due to the downturn in manufacturing, there have been two overdue payments in the past six months. Therefore, if the two overdue payments in the past six months are true, the credit rating will be downgraded, further reflecting the high credit risk of the enterprise," thus forming a coherent and natural language draft.

[0061] In step S107, the step of detecting logical contradictions in the preceding and following statements based on the language draft, and correcting the corresponding segments if contradictions exist to obtain a stable statement, includes: Semantic parsing is performed on the language draft to extract entity, attribute, and relation information, and a text logic graph is constructed. Traverse the text logic graph to detect entity attribute conflicts, relationship contradictions, and temporal logic errors; If a logical contradiction is detected, the text segment containing the contradiction is located, the relevant knowledge nodes and logical relationships are extracted, and a contradiction correction plan is generated based on the preset business knowledge base and logical rule base. The contradictory text fragments are adjusted according to the aforementioned contradiction correction scheme. A global logical consistency check is performed on the corrected language draft. If the consistency index exceeds the preset consistency threshold, a stable expression result is obtained.

[0062] It should be noted that, firstly, when semantically parsing the language draft to extract entity, attribute, and relation information and constructing the text logic graph, a combination of a BERT-based Semantic Role Labeling (SRL) model and dependency parsing was adopted. The SRL model used BERT-based Chinese pre-trained weights, and the training set contained over 30,000 business texts labeled with entities, attributes, and relations, divided into training and validation sets in a 7:3 ratio. The model settings included a hidden layer dimension of 768, 12 attention heads, an AdamW optimizer, a learning rate of 2e-5, 25 training epochs, GELU activation function, and cross-entropy loss. During training, the F1 score of the validation set was monitored in real-time, and iteration stopped when the fluctuation was less than 0.003 for 10 consecutive epochs. Dependency parsing used the Stanford Parser tool to extract subject-verb, verb-object, and causal relationships between entities in the sentences. The text logic graph uses entities as nodes, attributes as node attributes, and relations as directed edges, clearly presenting the deep logical structure of the language draft. For example, in a credit scenario, the language draft "Due to the downturn in the manufacturing industry, there have been two overdue payments in the past six months. Therefore, if the two overdue payments in the past six months are valid, the credit rating will be downgraded." After parsing, the entities "manufacturing industry," "number of overdue payments in the past six months," and "credit rating" are extracted, along with the attributes "downturn," "two times," and "downgrade," and the relationships "cause" and "condition trigger," to construct a text logic graph that includes the relationship between the three.

[0063] Next, the text logic graph is traversed. When detecting entity attribute conflicts, relationship contradictions, and time sequence logic errors, the entity attribute conflict detection verifies whether different attributes of the same entity are compatible. Based on a preset attribute compatibility rule base (e.g., "profit" and "loss" cannot coexist), the specific construction method of the attribute compatibility rule base is as follows: historical compliance documents, standard operating procedures, and logical error corpora that have been manually verified and corrected within a specific business domain are collected as the original data source; frequent itemset mining algorithms (e.g., FP-Growth algorithm) are used to perform pattern mining on the data source to extract attribute word pairs whose co-occurrence probability of the same entity under a single time slice or the same business logic is lower than a preset minimum threshold; at the same time, a contrastive learning framework is used to fine-tune the pre-trained language model to maximize the distance of mutually exclusive attribute pairs (e.g., "increase" and "decrease") in the feature vector space, and extract the attribute set whose vector distance is greater than the preset mutual exclusion threshold; the mined attribute word pairs are multimodal fused with the attribute set, and combined with the prior knowledge graph of domain experts for logical verification and deduplication, finally generating structured mutually exclusive rule entries and storing them in the attribute compatibility rule base.

[0064] During the detection process, the system retrieves the rule base in real time and calculates the semantic difference of attributes using cosine similarity. A difference greater than 0.7 is considered a conflict. The threshold of 0.7 is determined based on statistical analysis of a massive historical test dataset. Specifically, 50,000 manually labeled conflicting and non-conflicting attribute pairs within a specific business domain are collected as verification samples. The cosine similarity difference of each sample pair in the feature vector space is calculated, and receiver operating characteristic (ROC) curves are plotted to analyze the recognition effect. Statistical results show that when the difference threshold is 0.7, the Youden Index reaches its maximum value. The system can filter out misjudgments caused by conventional semantic fluctuations at the optimal balance between accuracy and recall, achieving accurate identification of genuine conflicts. Therefore, if the calculated difference is greater than 0.7, it is considered a conflict. Relationship contradiction detection compares whether the relationship type matches the entity semantics. For example, if the semantic correlation between the cause entity and the result entity in a causal relationship is less than 0.5, it is considered a contradiction. Temporal logic error detection verifies the timestamp order of event nodes. If there is a "result node time earlier than cause node time," it is considered an error. During the detection process, a combination of rule matching and BiLSTM classification model was used. The model training set contained more than 20,000 texts labeled with logical errors to ensure detection accuracy.

[0065] For example, in a certain language draft, "The company made a profit of 5 million in the past year and lost 3 million in the past year," the semantic difference between the attributes "profit of 5 million" and "loss of 3 million" of the same entity "company" is 0.85. Since it is greater than the judgment threshold of 0.7, it is judged as an entity attribute conflict. "Due to the increase in product sales, the production capacity was insufficient in the early stage." The cause node time (sales growth) is later than the result node time (insufficient production capacity), so it is judged as a temporal logic error.

[0066] If a logical contradiction is detected, the text segment containing the contradiction is located, relevant knowledge nodes and logical relationships are extracted, and a contradiction correction plan is generated based on a pre-set business knowledge base and logical rule base. The process involves tracing back the corresponding text segments through the conflict nodes in the text logic graph to clarify the specific statements where the contradiction occurred. The business knowledge base contains over 50,000 historical decision-making cases, industry standards, and a data dictionary. The logical rule base covers general logic such as causal transmission, conditional constraints, and temporal sequence, as well as domain-specific rules (such as the "correlation rule between overdue payment frequency and credit rating" in credit risk control). The correction plan generation uses a combination of case matching and rule reasoning, prioritizing matching with similar historical contradictions in the amendment examples. If no match is found, intermediate nodes are added, relationship types are adjusted, or attribute values ​​are corrected based on the logical rule base to ensure a logical closed loop after correction.

[0067] For example, if the entity attribute conflict of "the company's revenue has increased by 20% in the past year, but its operating cash flow continues to dry up" is detected, the conflicting nodes "revenue" and "operating cash flow" and the relationship "turning point" are extracted. The business knowledge base is consulted and it is found that the correction method for similar cases is to add the intermediate node "aggressive credit sales lead to delayed payment collection". Combining the rule in the logic rule base that "there must be an intermediate transmission factor between revenue growth and cash flow depletion", a correction solution is generated: add the statement fragment "due to the adoption of aggressive credit sales strategy, delayed payment collection" to connect the two sides of the conflict.

[0068] Subsequently, the contradictory text fragments were adjusted according to the contradiction correction scheme. A full-domain logical consistency check was performed on the corrected language draft. If the consistency index exceeded the preset consistency threshold, a stable expression result was obtained. Adjustments included replacing contradictory attributes, supplementing intermediate logical nodes, and correcting relationship types to ensure natural and logically coherent language expression. The full-domain logical consistency check used the logical conflict rate as the core indicator, which equals the number of detected logical conflicts divided by the total number of logical relationships. The consistency threshold was set based on historical verification data, with a base threshold of 0.9 (i.e., conflict rate ≤ 0.1). For complex decision-making scenarios (such as supply chain optimization with multiple coupled factors), this threshold could be increased to 0.95, and for simple scenarios, it could be decreased to 0.85. Those skilled in the art can flexibly adjust this threshold according to the complexity of the scenario. This threshold has been verified through extensive historical text verification; over 92% of texts exceeding the threshold did not present logical disputes in subsequent decision interpretations.

[0069] For example, the revised statement is "The company's revenue has increased by 20% in the past year, but due to the adoption of an aggressive credit sales strategy, the collection of payments has been delayed, and the operating cash flow has continued to dry up." After full-domain logic consistency verification, the logic conflict rate is 0.05, which is lower than the basic threshold of 0.9, so it is judged to be consistent and a stable expression result is obtained.

[0070] In step S108, the decision conclusion is derived based on the complete logical chain structure. The semantic features and logical paths of the stable representation result and the decision conclusion are integrated to generate a complete explanatory text, including: Based on the core logic and constraints of the complete logical chain structure, the decision conclusion is derived. The stable representation result and the decision conclusion are semantically encoded to obtain the representation feature vector and the conclusion feature vector, respectively. Calculate the fusion weights of the expression feature vector and the conclusion feature vector, and construct a fusion matrix; Based on the fusion matrix, the stable representation result and the decision conclusion are semantically fused to form a preliminary explanatory text; The preliminary explanation text is then syntactically optimized and its fluency adjusted to generate a complete explanation text.

[0071] It should be noted that, firstly, when deriving decision conclusions based on the core logic and constraints of the complete logical chain structure, the core logic is extracted from the causal relationships and conditional connections within the chain, focusing on the transmission path between key nodes; the constraints include business compliance requirements and data threshold limits (such as the threshold for the number of overdue payments and the upper limit of the amount in credit scenarios), all derived from a pre-set business rule base. The derivation is implemented using the Rete rule reasoning engine, which can efficiently match complex rules. The training set contains logical chains and corresponding conclusions from over 60,000 historical decision cases, divided into training and validation sets in a 7:3 ratio. The engine's rule base covers industry standards, compliance clauses, and historical decision rules. By matching the complete logical chain with the rule base, corresponding decision conclusions are triggered, ensuring a high degree of consistency between the conclusion and the logical chain, with no logical breaks.

[0072] For example, the complete logical chain of a credit scenario is "downturn in manufacturing → two overdue payments in the past six months → downgrade of credit rating", with the constraint that "no high-amount credit will be granted if there are two or more overdue payments". The decision conclusion derived through rule matching is: reject the company's credit application of 5 million yuan or more.

[0073] It is worth noting that the Sentence-BERT model was used to semantically encode the stable statements and decision conclusions to obtain the statement feature vector and conclusion feature vector, respectively. This model can accurately capture the deep semantic features of the text. The training set contains more than 50,000 stable statement texts and decision conclusion texts from business scenarios, covering areas such as credit, supply chain, and risk control. These were divided into training and validation sets in a 7:3 ratio. The model used distilbert-base-nli-stsb-mean-tokens pre-trained weights. During fine-tuning, the batch size was set to 32, the initial learning rate was 0.0001, and the training epochs were 20. The optimizer used was AdamW, the loss function was contrastive loss, and the activation function was GELU. The semantic similarity recall rate of the validation set was monitored in real time during training, and iteration was stopped when the fluctuation was less than 0.003 for 10 consecutive epochs. After encoding, 512-dimensional statement feature vectors and conclusion feature vectors were obtained, both of which were mapped to the [0,1] interval through min-max normalization to eliminate differences in magnitude.

[0074] For example, the stable statement "due to the downturn in manufacturing, there have been two overdue payments in the past six months, and the aggressive credit sales strategy has led to delayed payments" is encoded as a 512-dimensional vector, and the decision conclusion "reject credit applications of 5 million yuan or more" is also encoded as a vector of the same dimension, providing a basis for subsequent integration.

[0075] Subsequently, the fusion weights of the representation feature vector and the conclusion feature vector are calculated. When constructing the fusion matrix, the fusion weights are obtained by weighted summation of semantic relevance and logical path overlap. Semantic relevance is obtained by calculating the cosine similarity of the two feature vectors, with a value ranging from 0 to 1. Logical path overlap is calculated by the proportion of common nodes between the logical chain corresponding to the stable representation and the decision conclusion derivation path; the higher the proportion, the higher the overlap. The weight allocation is 0.6 for semantic relevance and 0.4 for logical path overlap. This combination is verified based on historical fusion results. Semantic relevance directly reflects the semantic fit between the two and is given a higher weight; logical path overlap ensures that the fusion does not deviate from the core logic. The fusion matrix is ​​a 2×2 matrix, with elements representing the fusion weights and interaction coefficients of the corresponding vectors, clearly defining the fusion rules.

[0076] For example, in a certain scenario, the semantic relevance is 0.85, the logical path overlap is 0.9, the fusion weight is calculated as 0.85×0.6 + 0.9×0.4=0.87, and the corresponding element value in the fusion matrix is ​​0.87, indicating that the fusion compatibility between the two is high.

[0077] Next, based on the fusion matrix, the stable statements and decision conclusions are semantically fused to form the preliminary explanatory text. A weighted summation fusion method is used, where the statement feature vector and the conclusion feature vector are weighted according to the fusion weights and then transformed into natural language fragments through linear mapping. During the fusion process, the logical process of the stable statements and the core viewpoints of the decision conclusions are retained first, and necessary logical connectors (such as "in summary" and "therefore") are added to ensure that the preliminary text contains both the basis for the decision and a clear understanding of the decision result, maintaining logical coherence.

[0078] For example, the stable statement "due to the downturn in the manufacturing industry, there have been two overdue payments in the past six months, and the aggressive credit sales strategy has led to delayed payments" is combined with the decision conclusion "reject credit applications of 5 million yuan or more". The preliminary interpretation text is "due to the downturn in the manufacturing industry, there have been two overdue payments in the past six months, and the aggressive credit sales strategy has led to delayed payments. Therefore, the credit application of 5 million yuan or more of this company is rejected."

[0079] Finally, the initial explanatory text underwent syntactic optimization and fluency adjustments. When generating the complete explanatory text, syntactic optimization employed a fine-tuned model based on the T5 model, specifically designed for text polishing. The training set contained over 30,000 fluent explanatory texts from various business domains, divided into training and validation sets in a 7:3 ratio. The model was configured with a hidden layer dimension of 512, 8 attention heads, an initial learning rate of 0.0001, 15 training epochs, and the AdamW optimizer. The cross-entropy loss function was used, and the fluency score of the validation set was monitored in real-time during training. The fluency score was based on grammatical correctness, word order, and naturalness of expression, with a fluency threshold of 0.85. This threshold was based on statistical analysis of human evaluations of historical optimized texts; texts exceeding the threshold achieved a human approval rate of over 90%. For complex business scenarios, the threshold could be increased to 0.9, while for simple scenarios, it could be decreased to 0.8. Adjustments included correcting grammatical errors, optimizing word order, removing repetitive expressions, and adding transitional words to ensure the final text was fluent, natural, and logically rigorous.

[0080] For example, after optimization, the complete explanation text is: "Affected by the downturn in the manufacturing industry, the company has defaulted twice in the past six months, and the collection of payments has been delayed due to the adoption of an aggressive credit sales strategy. Considering its credit status and operational risks, the company's credit application of RMB 5 million or more is rejected."

[0081] In summary, this invention discloses an intelligent decision-making method based on knowledge graphs, including: acquiring input data for a decision-making scenario; matching it with similar cases in a business knowledge base; extracting factual information, industry experience, and historical rules to form a set of knowledge nodes; analyzing the logical dependencies and temporal consistency of nodes to construct a complete logical chain; evaluating the contribution of nodes to filter key node sequences, converting them into sentence fragments and adjusting their coherence, and detecting and correcting logical contradictions; and integrating stable expression results with decision conclusions to generate a complete explanatory text. This achieves decision interpretability and logical consistency, meeting the requirements for decision transparency, credibility, and compliance.

[0082] Reference Figure 2 The second embodiment of the present invention provides an intelligent decision-making system based on knowledge graphs, comprising: The data acquisition module is used to acquire input data for the current decision-making scenario; The knowledge matching module is used to match the input data with similar cases in a preset business knowledge base, extract the corresponding factual information, industry experience and historical rules and integrate them in a structured manner to obtain a set of matched knowledge nodes; The chain construction module is used to analyze the logical dependencies and temporal consistency between knowledge nodes in the knowledge node set, determine the connection order of the nodes, and construct a complete logical chain structure according to the connection order. The node filtering module is used to evaluate the contribution of each knowledge node based on the topological depth, node betweenness and propagation probability of each node in the complete logical chain structure. If the contribution exceeds a preset contribution judgment threshold, the corresponding knowledge node is retained to obtain a key knowledge node sequence. The expression generation module is used to convert the logical relationships between the nodes into sentence fragments based on the sequence of key knowledge nodes, and to collect semantically complete fragments to form an initial expression set. The coherence adjustment module is used to calculate the semantic relevance between each of the statement fragments based on the initial set of statements, determine the order of statements by combining business logic priority and temporal rules, and match preset logical connectors to obtain a coherent and natural language draft. The contradiction correction module is used to detect logical contradictions in the preceding and following statements based on the language draft. If contradictions exist, the corresponding segments are corrected to obtain a stable statement result. The text generation module is used to deduce decision conclusions based on the complete logical chain structure, integrate the semantic features and logical paths of the stable expression results and the decision conclusions, and generate complete explanatory text.

[0083] It should be noted that the knowledge graph-based intelligent decision-making system provided in this embodiment of the invention is used to execute all the process steps of the knowledge graph-based intelligent decision-making method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0084] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0085] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A knowledge graph-based intelligent decision-making method, characterized in that, include: Obtain the input data for the current decision-making scenario; The input data is matched with similar cases in a preset business knowledge base, and the corresponding factual information, industry experience and historical rules are extracted and structured and integrated to obtain a set of matched knowledge nodes. Analyze the logical dependencies and temporal consistency among the knowledge nodes in the knowledge node set, determine the connection order of the nodes, and construct a complete logical chain structure according to the connection order; Based on the topological depth, node betweenness and propagation probability of each node in the complete logical chain structure, the contribution of each knowledge node is evaluated. If the contribution exceeds a preset contribution judgment threshold, the corresponding knowledge node is retained to obtain a key knowledge node sequence. Based on the sequence of key knowledge nodes, the logical relationships between each node are transformed into sentence fragments, and semantically complete fragments are collected to form an initial set of expressions; Based on the initial set of expressions, the semantic correlation between each of the statement fragments is calculated. The order of the statements is determined by combining the priority of business logic and the temporal sequence, and preset logical connectors are matched to obtain a coherent and natural language draft. Based on the language draft, detect logical contradictions in the preceding and following statements. If contradictions exist, correct the corresponding segments to obtain a stable statement result. Based on the complete logical chain structure, the decision conclusion is derived, and the semantic features and logical paths of the stable expression result and the decision conclusion are integrated to generate a complete explanatory text.

2. The knowledge graph-based intelligent decision-making method according to claim 1, characterized in that, The process of obtaining input data for the current decision-making scenario includes: Collect the raw input information of the current decision-making scenario, including business parameters, scenario attributes and constraints; The original input information is denoised to remove invalid data and duplicate records, resulting in cleaned input data. The cleaned input data is standardized in format and dimensionally regularized to construct structured input data; Extract the core features of the structured input data and determine the core features as the input data for the current decision-making scenario.

3. The knowledge graph-based intelligent decision-making method according to claim 1, characterized in that, The process involves matching the input data with similar cases in a pre-defined business knowledge base, extracting corresponding factual information, industry experience, and historical rules, and then structurally integrating these to obtain a set of matched knowledge nodes, including: The input data is converted into a feature vector, and the similarity between the feature vector and the feature vector of historical cases in the preset business knowledge base is calculated. If the similarity exceeds the preset similarity judgment threshold, the corresponding historical case is marked as a candidate similar case, and objective factual information, industry experience text and historical rule entries are extracted from the candidate similar cases. The factual information, the industry experience text, and the historical rule entries are semantically aligned and categorized. The categorized information is then structurally integrated to obtain a set of matching knowledge nodes.

4. The knowledge graph-based intelligent decision-making method according to claim 1, characterized in that, The process of analyzing the logical dependencies and temporal consistency among knowledge nodes in the knowledge node set, determining the connection order of the nodes, and constructing a complete logical chain structure according to the connection order includes: Calculate the interaction strength between each knowledge node in the knowledge node set, and construct the dependency matrix of each knowledge node based on the interaction strength; An initial topology is generated based on the dependency matrix, and the potential logical connection paths of each knowledge node are traced based on the initial topology to obtain a set of candidate tracing paths. Extract the timestamp information of each candidate tracing path in the candidate tracing path set, verify the time sequence of nodes in the path, and if a candidate tracing path has a time sequence conflict, remove the path and retain the time sequence compliant paths to form a compliant path set. By integrating the common nodes and logical relationships in the set of compliant paths, a reasoning subgraph is constructed. Traverse the reasoning subgraph, sort out the node connection order according to logical causal relationship and time sequence, and form a complete logical chain structure.

5. The knowledge graph-based intelligent decision-making method according to claim 1, characterized in that, The contribution of each knowledge node is evaluated based on its topological depth, betweenness number, and propagation probability in the complete logical chain structure. If the contribution exceeds a preset contribution threshold, the corresponding knowledge node is retained, resulting in a key knowledge node sequence, including: Traverse the complete logical chain structure and calculate the topological depth, node betweenness, and propagation probability of each knowledge node; The contribution score of each knowledge node is obtained by weighting and summing the topology depth, the node betweenness, and the propagation probability using preset weighting coefficients. The contribution score is compared with a preset contribution judgment threshold. If the contribution score exceeds the contribution judgment threshold, the corresponding knowledge node is marked as a core node. Based on the temporal relationship and logical order in the complete logical chain structure, the core nodes are reorganized and sorted to obtain a sequence of key knowledge nodes.

6. The knowledge graph-based intelligent decision-making method according to claim 1, characterized in that, The step of transforming the logical relationships between key knowledge nodes into sentence fragments based on the sequence of key knowledge nodes, and collecting semantically complete fragments to form an initial set of expressions, includes: Traverse the sequence of key knowledge nodes and extract the logical relationship types between adjacent nodes. The logical relationship types include causal relationships, conditional relationships, and progressive relationships. Based on the logical relation type, match the corresponding syntactic template to construct the statement skeleton; The entity information of the key knowledge nodes is filled into the statement skeleton to generate a statement fragment; The semantic integrity of the statement fragment is checked, and an integrity score is calculated. If the integrity score exceeds a preset integrity threshold, the statement fragment is retained, and all qualified fragments are collected to form an initial expression set.

7. The knowledge graph-based intelligent decision-making method according to claim 1, characterized in that, The process involves calculating the semantic relevance between each statement fragment based on the initial set of expressions, determining the order of statements by combining business logic priority and temporal patterns, and matching preset logical connectors to obtain a coherent and natural language draft, including: Calculate the semantic association degree between any two sentence fragments in the initial set of expressions, and construct a semantic association matrix; Based on the semantic association matrix, core sentence fragments are identified. Combining business logic priority and temporal sequence, the arrangement order of the core sentence fragments is determined, and a sentence order list is generated. Based on the logical relationship of the sentence fragments in the word order list, preset logical connectors are matched, including causal, progressive, and adversative logical connectors. The logical connectors are embedded between corresponding sentence fragments and then pieced together in sequence to form a coherent and natural language draft.

8. The knowledge graph-based intelligent decision-making method according to claim 1, characterized in that, The process of detecting logical inconsistencies in the preceding and following statements based on the language draft, and correcting the corresponding segments to obtain a stable expression result, includes: Semantic parsing is performed on the language draft to extract entity, attribute, and relation information, and a text logic graph is constructed. Traverse the text logic graph to detect entity attribute conflicts, relationship contradictions, and temporal logic errors; If a logical contradiction is detected, the text segment containing the contradiction is located, the relevant knowledge nodes and logical relationships are extracted, and a contradiction correction plan is generated based on the preset business knowledge base and logical rule base. The contradictory text fragments are adjusted according to the aforementioned contradiction correction scheme. A global logical consistency check is performed on the corrected language draft. If the consistency index exceeds the preset consistency threshold, a stable expression result is obtained.

9. The knowledge graph-based intelligent decision-making method according to claim 1, characterized in that, The decision conclusion derived based on the complete logical chain structure, integrating the semantic features and logical paths of the stable representation results and the decision conclusion, generates a complete explanatory text, including: Based on the core logic and constraints of the complete logical chain structure, the decision conclusion is derived. The stable representation result and the decision conclusion are semantically encoded to obtain the representation feature vector and the conclusion feature vector, respectively. Calculate the fusion weights of the expression feature vector and the conclusion feature vector, and construct a fusion matrix; Based on the fusion matrix, the stable representation result and the decision conclusion are semantically fused to form a preliminary explanatory text; The preliminary explanation text is then syntactically optimized and its fluency adjusted to generate a complete explanation text.

10. A knowledge graph-based intelligent decision-making system, characterized in that, include: The data acquisition module is used to acquire input data for the current decision-making scenario; The knowledge matching module is used to match the input data with similar cases in a preset business knowledge base, extract the corresponding factual information, industry experience and historical rules and integrate them in a structured manner to obtain a set of matched knowledge nodes; The chain construction module is used to analyze the logical dependencies and temporal consistency between knowledge nodes in the knowledge node set, determine the connection order of the nodes, and construct a complete logical chain structure according to the connection order. The node filtering module is used to evaluate the contribution of each knowledge node based on the topological depth, node betweenness and propagation probability of each node in the complete logical chain structure. If the contribution exceeds a preset contribution judgment threshold, the corresponding knowledge node is retained to obtain a key knowledge node sequence. The expression generation module is used to convert the logical relationships between the nodes into sentence fragments based on the sequence of key knowledge nodes, and to collect semantically complete fragments to form an initial expression set. The coherence adjustment module is used to calculate the semantic relevance between each of the statement fragments based on the initial set of statements, determine the order of statements by combining business logic priority and temporal rules, and match preset logical connectors to obtain a coherent and natural language draft. The contradiction correction module is used to detect logical contradictions in the preceding and following statements based on the language draft. If contradictions exist, the corresponding segments are corrected to obtain a stable statement result. The text generation module is used to deduce decision conclusions based on the complete logical chain structure, integrate the semantic features and logical paths of the stable expression results and the decision conclusions, and generate complete explanatory text.

Images