AI-driven enterprise big data intelligent analysis service methods and systems

By constructing an enterprise domain knowledge graph and utilizing graph attention propagation and user feedback to optimize the model, the problem of existing systems being unable to dynamically adapt to user needs was solved, enabling efficient and in-depth data analysis and deep insights.

CN122673309BActive Publication Date: 2026-11-17BEIJING MINGQI INT TECH GRP CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202610791909.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-06-03
Publication Date
2026-11-17
Estimated Expiration
2046-06-03

AI Technical Summary

Technical Problem

Existing enterprise big data intelligent analysis service systems cannot dynamically adapt to users' personalized analysis needs, lack interactive feedback mechanisms and in-depth mining of multi-hop related information, resulting in analysis results that cannot be adjusted in real time and provide in-depth business insights.

Method used

Construct an enterprise domain knowledge graph, perform multi-hop reasoning through graph attention propagation, generate candidate analysis paths, and use user feedback to train a reward model to optimize attention weights, generating a comprehensive analysis report.

Benefits of technology

It enables dynamic adjustment of the analysis path based on user preferences, improving the accuracy and depth of search results and lowering the barrier for users to obtain in-depth insights.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122673309B_ABST
    Figure CN122673309B_ABST
Patent Text Reader

Abstract

This invention relates to the field of enterprise big data intelligent analysis technology, and in particular to an AI-driven enterprise big data intelligent analysis service method and system. The method constructs an enterprise domain knowledge graph and receives user natural language queries, maps them to the knowledge graph, performs graph attention propagation to obtain an activated subgraph, extracts extended nodes from it to generate candidate analysis paths, executes the main query and converts the candidate paths into structured questions, trains a reward model to update attention parameters based on user selection feedback, executes related queries based on the received paths and merges them with the direct query results to generate a comprehensive analysis report, thereby improving the accuracy and personalization of enterprise big data analysis.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of enterprise big data intelligent analysis technology, and in particular to AI-driven enterprise big data intelligent analysis service methods and systems. Background Technology

[0002] Current enterprise big data intelligent analytics services generally employ predefined analysis templates or fixed rule engines, allowing users to obtain reports by selecting dimensions and indicator combinations. Common practices include building multi-dimensional data warehouses, utilizing OLAP multidimensional analysis techniques for operations such as slicing and drill-down, or using keyword-matching search engines to return structured query results. Some systems introduce domain ontology or simple entity relationship graphs to assist in basic semantic retrieval. However, these solutions rely on manually preset logical paths and cannot dynamically adapt to the personalized analysis needs of different users.

[0003] One drawback of conventional approaches is the lack of interactive feedback mechanisms in the analytical reasoning process. After a user submits a query, the system typically returns a fixed result all at once, failing to adjust subsequent reasoning directions in real time based on the user's preference for intermediate results. For example, when a user shows interest in a particular association analysis path, the existing system cannot proactively capture this signal and optimize the reasoning weights, causing subsequent queries to still use the initial parameters, making it difficult to converge to the user's true intent in the long run.

[0004] Another drawback is the lack of in-depth mining of multi-hop relational information. Existing technologies for reasoning in knowledge graphs are often limited to directly connected nodes or preset rules, failing to automatically discover semantic relationships across levels and types. When users conduct open-ended explorations, the system only returns directly matched entities, ignoring potential indirect causal relationships or trend associations. This keeps the analysis reports at a superficial statistical level, making it difficult to provide derivative conclusions with business insight value. Furthermore, the lack of a reinforcement learning mechanism for user feedback prevents the reasoning model from gradually improving its accuracy with accumulated usage, limiting the system's adaptability to complex queries. Summary of the Invention

[0005] This invention provides an AI-driven enterprise big data intelligent analysis service method and system, which can solve the problems in the prior art.

[0006] A first aspect of the present invention provides an AI-driven enterprise big data intelligent analysis service method, comprising:

[0007] Construct an enterprise domain knowledge graph, receive users' natural language query input, extract core query entities and query relationships, and obtain an initial query representation;

[0008] The initial query representation is mapped to the corresponding node of the enterprise domain knowledge graph. Multi-hop reasoning based on graph attention propagation is performed starting from the corresponding node. Attention weights are propagated layer by layer by calculating the semantic association strength between nodes and activating relevant subgraph regions to obtain activated subgraphs.

[0009] Extended nodes are extracted from the activated subgraph, and multiple candidate analysis paths are generated based on the semantic relationships of the extended nodes. The main query task is executed based on the initial query representation to obtain direct query results. At the same time, the multiple candidate analysis paths are converted into structured questions and presented to the user.

[0010] Receive user feedback on the structured question, record the candidate analysis paths of user acceptance and rejection and label them as positive and negative samples, construct a user preference sample set, train a reward model using the user preference sample set, and use the output of the reward model as a reinforcement learning signal to update the attention weight calculation parameters of the graph attention propagation.

[0011] Based on the candidate analysis path accepted by the user, an associated query task is executed to obtain extended query results. The extended query results are then fused with the direct query results according to semantic relationships to generate a comprehensive analysis report.

[0012] The initial query representation is mapped to the corresponding node in the enterprise domain knowledge graph. Multi-hop inference based on graph attention propagation is then performed starting from the corresponding node. Attention weights are propagated layer by layer by calculating the semantic association strength between nodes, and relevant subgraph regions are activated. The resulting activated subgraphs include:

[0013] Discriminative semantic representations are generated by comparing the initial query representation with the counterexample query using an adversarial dual encoder. In the enterprise domain knowledge graph, the node is jointly ranked by subgraph isomorphic matching and node role similarity. The node with the highest ranking is constructed as an attention source with an initial potential field as the corresponding node.

[0014] Obtain the set of first-order neighbor nodes of the corresponding node in the enterprise domain knowledge graph connected by semantic association edges, and parse the relation type and direction attribute of each semantic association edge according to the semantic dependency tree decomposition of the edge. Based on the semantic role of the relation type and the second-order matching of the query slot in the initial query representation, generate initial attention weights for each node in the set of first-order neighbor nodes.

[0015] Nodes whose initial attention weights exceed the propagation threshold in the first-order neighbor node set are selected as activation nodes. The second-order neighbor nodes of the activation nodes are obtained. The path semantic consistency is aggregated with the initial attention weights of the activation nodes through semantic interaction modeling in the tensor product space, and a hop count decay coefficient is applied to obtain the propagation attention weights of the second-order neighbor nodes.

[0016] Adaptive step-size diffusion is performed along the information entropy gradient direction to perform filtering and weight calculation until the attention weight of all nodes is lower than the propagation threshold or reaches the preset inference hop count. All nodes in each layer whose attention weight exceeds the activation threshold and their connecting edges are then combined to form an activation subgraph.

[0017] By modeling semantic interactions in tensor product space, the path semantic consistency is aggregated with the initial attention weights of the activated nodes, and a hop count decay coefficient is applied to obtain the propagation attention weights of the second-order neighbor nodes, including:

[0018] For the reasoning path from the corresponding node through the activated node to the second-order neighbor node, a relational algebraic expression of the path is constructed. By decomposing the relational algebraic expression into the minimum non-redundant relation atom set, the satisfiability of the logical implication relationship between each relation atom and the query predicate in the initial query representation is solved. The semantic consistency of the path is quantified based on the proportion of relation atoms that satisfy the constraints in the solution results.

[0019] The path semantic consistency is encoded as a diagonalized semantic operator in the tensor product space, and the initial attention weights of the activated nodes are embedded as eigenstate vectors in the tensor product space. The spectral decomposition and reconstruction of the eigenstate vectors by the diagonalized semantic operator are used to extract the dominant mode contribution of path semantics to attention distribution.

[0020] Based on the hop count of the inference path, construct an information decay Markov chain, and extract the transition probability related to the hop count from the stationary distribution of the Markov chain as the hop count decay coefficient.

[0021] The dominant mode contribution and the hop count decay coefficient are fused in Bayesian mode in the posterior distribution space using a variational inference framework to obtain the propagation attention weights of the second-order neighbor nodes.

[0022] Extended nodes are extracted from the activated subgraph, and multiple candidate analysis paths are generated based on the semantic relationships of the extended nodes. The main query task is executed based on the initial query representation to obtain direct query results. Simultaneously, the multiple candidate analysis paths are converted into structured questions, including:

[0023] Nodes whose propagation attention weight gradient change rate exceeds the inflection point detection threshold are selected from the activation subgraph as expansion nodes. Semantic role annotation is performed on the expansion nodes. A role constraint graph is constructed based on the semantic role annotation. On the role constraint graph, the node sequence that satisfies role compatibility is enumerated through the constraint satisfaction propagation algorithm. The node sequence is used as multiple candidate analysis paths.

[0024] For each of the multiple candidate analysis paths, the topological features and semantic relation type combinations of the candidate analysis path are extracted. The topological features are mapped to the syntactic framework of the problem, and the semantic relation type combinations are mapped to the predicate slots in the syntactic framework. A structured problem is generated by assembling the syntactic framework and the predicate slots.

[0025] The extended nodes are semantically labeled, and a role constraint graph is constructed based on the semantic role labels. On the role constraint graph, a constraint satisfaction propagation algorithm is used to enumerate the sequence of nodes that satisfy role compatibility, including:

[0026] Obtain the relational dependency pattern of the extended node in the enterprise domain knowledge graph, project the relational dependency pattern onto the argument slots of the predefined semantic framework, and label the extended node with the corresponding semantic role according to the functional attributes undertaken by the argument slots in the predefined semantic framework.

[0027] The extended nodes are used as vertices to build a graph structure. For any two nodes in the extended nodes, the semantic role annotation of the two nodes is used to retrieve the role co-occurrence constraints in the semantic dependency rule base. According to the compatibility judgment condition in the role co-occurrence constraints, constraint edges are built in the graph structure for the node pairs that meet the compatibility judgment condition. The constraint edges carry the compatibility strength value corresponding to the compatibility judgment condition, and the role constraint graph is constructed.

[0028] Initialize node selection states and constraint satisfaction on the role constraint graph. Iteratively propagate the node selection states using a constraint satisfaction propagation algorithm. Once the propagation process is stable, select nodes whose node selection states meet the activation conditions to form a node sequence that satisfies role compatibility.

[0029] Training a reward model using the user preference sample set, and then using the output of the reward model as a reinforcement learning signal to update the attention weight calculation parameters of the graph attention propagation, includes:

[0030] The query response pairs from the user preference sample set are input into the reward model to calculate the preference difference value between different response qualities. Based on the consistency loss between the preference difference value and the preference relationship labeled in the user preference sample set, the parameters of the reward model are optimized so that the score output by the reward model can reflect the user's preference for the quality of the query response.

[0031] The trained reward model scores the query results generated by graph attention propagation, and the score is used as a reward signal for reinforcement learning. A policy gradient is constructed based on the causal relationship between the reward signal and the attention weight distribution in graph attention propagation. The policy gradient determines the optimization direction and adjustment range of the attention weight calculation parameters through sensitivity analysis of the reward signal to the attention weight distribution.

[0032] The attention weight calculation parameters of the graph attention propagation are updated according to the policy gradient.

[0033] Based on the candidate analysis path accepted by the user, an associated query task is performed to obtain extended query results. The extended query results are then fused with the direct query results according to semantic relationships to generate a comprehensive analysis report, including:

[0034] The sequence of related query nodes is parsed based on the candidate analysis path accepted by the user. A multi-hop related query task is performed in the enterprise domain knowledge graph along the sequence of related query nodes. The path is traversed through the relationship edges between adjacent nodes in the sequence of related query nodes to obtain the entity attribute information of the terminal node of the sequence of related query nodes as the extended query result.

[0035] Semantic relationship identification is performed on the extended query results and the direct query results. By calculating the association pattern between the vector representations of the extended query results and the direct query results in the semantic space, the semantic relationship type between the extended query results and the direct query results is determined.

[0036] Based on the semantic relationship type, a content organization structure for the comprehensive analysis report is constructed. In the content organization structure, the extended query results and the direct query results are grouped and aggregated according to the semantic relationship type. The grouped and aggregated content is arranged according to the narrative logic corresponding to the semantic relationship type to generate a comprehensive analysis report.

[0037] A second aspect of this invention provides an AI-driven enterprise big data intelligent analysis service system, comprising:

[0038] The graph construction unit is used to build an enterprise domain knowledge graph. It receives natural language query input from users, extracts core query entities and query relationships, and obtains an initial query representation.

[0039] The multi-hop inference unit is used to map the initial query representation to the corresponding node of the enterprise domain knowledge graph, and to perform multi-hop inference based on graph attention propagation starting from the corresponding node. It calculates the semantic association strength between nodes, propagates attention weights layer by layer, and activates relevant subgraph regions to obtain activated subgraphs.

[0040] The path generation unit is used to extract extended nodes from the activated subgraph, generate multiple candidate analysis paths based on the semantic relationships of the extended nodes, execute the main query task based on the initial query representation to obtain direct query results, and convert the multiple candidate analysis paths into structured questions to present to the user.

[0041] The feedback training unit is used to receive user feedback on the structured question, record the candidate analysis paths of user acceptance and rejection and label them as positive and negative samples, construct a user preference sample set, train a reward model using the user preference sample set, and use the output of the reward model as a reinforcement learning signal to update the attention weight calculation parameters of the graph attention propagation.

[0042] The report generation unit is used to perform related query tasks based on the candidate analysis paths accepted by the user to obtain extended query results, and to fuse the extended query results with the direct query results according to semantic relationships to generate a comprehensive analysis report.

[0043] A third aspect of the present invention provides an electronic device, comprising:

[0044] processor;

[0045] Memory used to store processor-executable instructions;

[0046] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0047] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0048] By constructing an enterprise domain knowledge graph that integrates business concepts, metrics, dimensions, and semantic relationships, natural language queries are transformed into precise structured representations, enabling non-technical users to interact directly with complex enterprise data systems using everyday language. A multi-hop inference mechanism based on graph attention propagation automatically calculates the semantic relationship strength between nodes, activating subgraph regions highly relevant to the user's query intent layer by layer. This significantly improves the accuracy of search results and avoids noise interference caused by traditional keyword matching, thereby achieving efficient and in-depth data exploration in complex business scenarios.

[0049] By extracting extended nodes from the activated subgraph and generating multiple candidate analysis paths, the system can proactively reveal the correlation dimensions and anomaly patterns that users may be interested in, upgrading passive querying to intelligent assisted exploration. Presenting these structured questions to users allows them to intuitively grasp the causal relationships and evolutionary trends between data without needing data analysis skills, significantly lowering the barrier to gaining in-depth insights. By recording users' acceptance or rejection of candidate paths as positive and negative samples and constructing a preference sample set, the system can dynamically optimize the attention weight calculation parameters of the graph attention layer using reinforcement learning, making the subsequent reasoning process closer to the individual user's business understanding and decision-making habits. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the AI-driven enterprise big data intelligent analysis service method according to an embodiment of the present invention.

[0051] Figure 2 This is a flowchart illustrating the enterprise data intelligent analysis service based on knowledge graphs, as described in an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0053] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0054] Figure 1 This is a flowchart illustrating the AI-driven enterprise big data intelligent analysis service method according to an embodiment of the present invention.

[0055] AI-driven enterprise big data intelligent analysis service methods include:

[0056] Construct an enterprise domain knowledge graph, receive users' natural language query input, extract core query entities and query relationships, and obtain an initial query representation;

[0057] The initial query representation is mapped to the corresponding node of the enterprise domain knowledge graph. Multi-hop reasoning based on graph attention propagation is performed starting from the corresponding node. Attention weights are propagated layer by layer by calculating the semantic association strength between nodes and activating relevant subgraph regions to obtain activated subgraphs.

[0058] Extended nodes are extracted from the activated subgraph, and multiple candidate analysis paths are generated based on the semantic relationships of the extended nodes. The main query task is executed based on the initial query representation to obtain direct query results. At the same time, the multiple candidate analysis paths are converted into structured questions and presented to the user.

[0059] Receive user feedback on the structured question, record the candidate analysis paths of user acceptance and rejection and label them as positive and negative samples, construct a user preference sample set, train a reward model using the user preference sample set, and use the output of the reward model as a reinforcement learning signal to update the attention weight calculation parameters of the graph attention propagation.

[0060] Based on the candidate analysis path accepted by the user, an associated query task is executed to obtain extended query results. The extended query results are then fused with the direct query results according to semantic relationships to generate a comprehensive analysis report.

[0061] In one optional implementation, the initial query representation is mapped to a corresponding node in the enterprise domain knowledge graph. Multi-hop inference based on graph attention propagation is performed starting from the corresponding node. Attention weights are propagated layer by layer by calculating the semantic association strength between nodes, and relevant subgraph regions are activated. The resulting activated subgraphs include:

[0062] Discriminative semantic representations are generated by comparing the initial query representation with the counterexample query using an adversarial dual encoder. In the enterprise domain knowledge graph, the node is jointly ranked by subgraph isomorphic matching and node role similarity. The node with the highest ranking is constructed as an attention source with an initial potential field as the corresponding node.

[0063] Obtain the set of first-order neighbor nodes of the corresponding node in the enterprise domain knowledge graph connected by semantic association edges, and parse the relation type and direction attribute of each semantic association edge according to the semantic dependency tree decomposition of the edge. Based on the semantic role of the relation type and the second-order matching of the query slot in the initial query representation, generate initial attention weights for each node in the set of first-order neighbor nodes.

[0064] Nodes whose initial attention weights exceed the propagation threshold in the first-order neighbor node set are selected as activation nodes. The second-order neighbor nodes of the activation nodes are obtained. The path semantic consistency is aggregated with the initial attention weights of the activation nodes through semantic interaction modeling in the tensor product space, and a hop count decay coefficient is applied to obtain the propagation attention weights of the second-order neighbor nodes.

[0065] Adaptive step-size diffusion is performed along the information entropy gradient direction to perform filtering and weight calculation until the attention weight of all nodes is lower than the propagation threshold or reaches the preset inference hop count. All nodes in each layer whose attention weight exceeds the activation threshold and their connecting edges are then combined to form an activation subgraph.

[0066] Before mapping the initial query representation to the corresponding nodes in the enterprise domain knowledge graph, a discriminative semantic representation needs to be generated. An adversarial dual-encoder structure is adopted, where the initial query representation and manually constructed or automatically sampled negative example queries are simultaneously fed into two encoder branches sharing weights. By contrastive learning of the objective function, the semantic distance between positive sample pairs is narrowed, while the semantic distance between negative sample pairs is widened. The construction of negative example queries includes: randomly replacing core entity words in the query, using synonyms or cross-domain obfuscation to replace business indicator names, and making disruptive modifications to slots such as the query's time range and aggregation granularity. After contrastive learning training, the semantic representation output by the encoder can form a more compact and discriminative distribution in the embedding space, enabling the subsequent node matching stage to effectively resist the interference of diverse query expressions and avoid misjudging nodes with similar semantics but different business meanings as matching targets.

[0067] After obtaining the discriminative semantic representation, subgraph isomorphic matching and node role similarity ranking are jointly performed in the enterprise domain knowledge graph. Subgraph isomorphic matching aligns the structural pattern of the query with the structure of local subgraphs in the knowledge graph, identifying a set of candidate nodes that structurally match the query intent. Node role similarity measures the degree of match between the semantic role of a candidate node in its subgraph and the corresponding slot in the query intent. The subgraph isomorphic matching score and the node role similarity score are weighted and jointly ranked, and the node with the highest ranking is selected as the initial mapping node. This node is then assigned an initial potential energy field, meaning that in subsequent graph attention propagation, this node, as the attention source, has the highest initial attention potential energy, and its propagation energy gradually decreases as the graph topological distance increases. This potential energy field setting method ensures the directionality and focus of multi-hop reasoning, avoiding indiscriminate diffusion of attention in the graph.

[0068] Starting with the corresponding node, obtain the set of its first-order neighbor nodes directly connected by semantic association edges in the enterprise domain knowledge graph. For the semantic association edges connecting the corresponding node to each first-order neighbor node, use the semantic dependency tree decomposition method to perform structured parsing of the semantic relations carried by the edges, extracting the relation type and direction attribute of each edge. After obtaining the relation type and direction attribute of each edge, perform second-order matching derivation: match the semantic role corresponding to the relation type of the edge with the identified query slots in the initial query representation. Query slots include analysis target slots, analysis dimension slots, time range slots, and filter condition slots. If the semantic role of a certain first-order neighbor node can fill a query slot in the initial query representation through its connecting edge, then the node receives a higher initial attention weight; if the semantic role has no corresponding relationship with any query slot, then the node receives a lower initial attention weight. Initial Attention Weight Taking into account the relationship type of the edges, the matching score is... semantic relevance of the node itself The attention distribution on the first-order neighbor node set is obtained by normalization.

[0069] Selecting initial attention weights from the set of first-order neighbor nodes Exceeding the transmission threshold The nodes are marked as active nodes, and the second-hop inference phase begins. The second-order neighbor nodes of each active node are obtained, and semantic interaction modeling in tensor product space is performed on these second-order neighbor nodes. Specifically, the complete path semantic representation from the corresponding node through the active node to the second-order neighbor node is aggregated with the initial attention weights of the active nodes through tensor product operations. The path semantic representation is obtained by sequentially accumulating the semantic association vectors of each edge on the path, capturing the combined meaning of multi-step semantic transformations. A hop count decay coefficient is applied to the aggregation result. ,in As the attenuation base, Given the current propagation hop count, for second-order neighbor nodes... Thus, the propagation attention weights of the second-order neighbor nodes are obtained. The hop count decay mechanism ensures that nodes farther away from the initial query mapping node receive lower attention weights, thereby maintaining the convergence of semantic focus in multi-hop inference.

[0070] After the second hop, adaptive step-size diffusion is performed along the direction of the information entropy gradient. At each propagation layer, the information entropy of the attention weight distribution on the current set of activated nodes is calculated. When information entropy is high, the diffusion step size is reduced, expanding outwards in a more conservative manner to avoid including a large number of low-relevance nodes in the activation range; when information entropy is low, the diffusion step size is appropriately increased, allowing for deeper inference exploration along high-confidence paths. After each layer of diffusion is completed, the same weight calculation process as described above is performed on the newly obtained neighbor nodes, and the propagation threshold is used again. A selection process is performed, retaining only nodes whose weights exceed a threshold for next-hop propagation. This continues until the attention weights of all nodes to be propagated are below the propagation threshold. Or the propagation hop count reaches the preset maximum inference hop count. At that point, the multi-hop reasoning process terminates.

[0071] After the inference process is completed, all attention weights are collected from the propagation results at each layer, exceeding the activation threshold. The nodes, along with the semantic connections between them within the enterprise domain knowledge graph, together form the activation subgraph. This activation subgraph serves as the semantic foundation for subsequent candidate analysis path generation and related query tasks. Its nodes cover business concepts, metrics, and dimensions directly related to the user's query intent, and its edge structure preserves the complete semantic relationships between these elements, providing structured knowledge support for the generation of comprehensive analysis reports.

[0072] In one optional implementation, the propagation attention weights of the second-order neighbor nodes are obtained by aggregating the path semantic consistency with the initial attention weights of the activated nodes through semantic interaction modeling in the tensor product space and applying a hop count decay coefficient, including:

[0073] For the reasoning path from the corresponding node through the activated node to the second-order neighbor node, a relational algebraic expression of the path is constructed. By decomposing the relational algebraic expression into the minimum non-redundant relation atom set, the satisfiability of the logical implication relationship between each relation atom and the query predicate in the initial query representation is solved. The semantic consistency of the path is quantified based on the proportion of relation atoms that satisfy the constraints in the solution results.

[0074] The path semantic consistency is encoded as a diagonalized semantic operator in the tensor product space, and the initial attention weights of the activated nodes are embedded as eigenstate vectors in the tensor product space. The spectral decomposition and reconstruction of the eigenstate vectors by the diagonalized semantic operator are used to extract the dominant mode contribution of path semantics to attention distribution.

[0075] Based on the hop count of the inference path, construct an information decay Markov chain, and extract the transition probability related to the hop count from the stationary distribution of the Markov chain as the hop count decay coefficient.

[0076] The dominant mode contribution and the hop count decay coefficient are fused in Bayesian mode in the posterior distribution space using a variational inference framework to obtain the propagation attention weights of the second-order neighbor nodes.

[0077] For inference paths from the corresponding node through the activated node to the second-order neighbor node, it is necessary to first establish the relational algebraic expression of the path. Specifically, the inference path consists of a series of directed semantically related edges, each carrying a relation type label. The relation types of each edge on the path are combined sequentially to form a relational compound expression, such as a chain-like relational structure like "department → responsible indicator → related dimension". A decomposition operation is performed on this relational algebraic expression, with the goal of obtaining the minimum non-redundant relation atom set: by identifying and eliminating redundant sub-relations in the relational compound, the basic relational units that are logically indivisible and mutually independent are retained as an atomic description of the path semantics.

[0078] Satisfactionability is determined for the logical implication relationship between each relation atom and the query predicate in the initial query representation, using a first-order logic constraint satisfaction method: the query predicate is treated as the target logical formula, each relation atom is treated as a precondition, and it is determined whether the relation atom semantically implies the relation type described by the query predicate. The satisfactionability solution process is completed using resolution reasoning in propositional logic or reasoning rules based on existing axioms in the knowledge graph. The ratio of the number of relation atoms satisfying the implication constraint to the total number of relation atoms is calculated, and this ratio is used as a quantitative measure of path semantic consistency, denoted as [the ratio is missing from the original text]. ,in The higher the value, the stronger the semantic alignment between the reasoning path and the user's query intent.

[0079] Path semantic consistency The encoding is a diagonalized semantic operator in a tensor product space. The tensor product space is constructed based on the semantic embedding space of knowledge graph nodes. A high-dimensional interactive representation space is obtained by performing a tensor product operation on the two semantic embedding spaces. (Diagonalized semantic operator) It is a space in which the tensor product is The diagonal matrix constructed to control the scaling factor assigns values ​​to its diagonal elements based on the implication strength of each relation atom and query predicate in the path. This amplifies the dimensions of paths with high semantic consistency while suppressing dimensions with semantic deviation. Initial attention weights for activated nodes. Embedded as eigenstate vectors in tensor product space The components of this vector correspond to the attention distribution of the activated nodes across different semantic dimensions. This is achieved by diagonalizing the semantic operator. For eigenstate vectors Perform spectral decomposition, Expanding the operator under its eigenbase into a linear superposition of its eigencomponents, and then... The diagonal elements are used to weight and reconstruct each component, resulting in the dominant mode contribution vector of path semantics to the attention distribution. ,Right now This vector It retains attention components that are strongly related to query semantics while suppressing attention dispersion in semantic directions unrelated to the query, thereby achieving attention refinement guided by path semantics.

[0080] A Markov chain with information decay is constructed based on the hop count of the inference path. Each hop in the inference process is considered a state transition step in the Markov chain. The transition probability matrix is ​​set according to the semantic association strength between nodes, and an information decay mechanism is introduced: as the number of hops increases, the information contribution of the distant node to the current query gradually weakens, which is reflected in the spectral radius of the transition probability matrix decreasing with the number of hops. A stationary distribution is then obtained for this Markov chain. The stationary distribution describes the occupancy probability of each state after an infinite number of transitions, where the hop count is a constant. The relevant transition probability component is the hop count decay coefficient. . Specifically, Extract the corresponding i from the stationary distribution vector The component representing the hop state naturally encodes the statistical law of information loss during multi-hop propagation, which is superior to directly using a fixed attenuation base. The power form of the attenuation coefficient, based on the stationary distribution of the Markov chain, can adaptively reflect the information transmission characteristics under different path structures.

[0081] The dominant mode contribution vector is determined using a variational inference framework. With the number of jumps attenuation coefficient Perform Bayesian fusion in the posterior distribution space to obtain the propagation attention weights of the second-order neighbor nodes. The variational inference framework will Treated as a latent variable, its prior distribution is determined by the contribution vector of the dominant mode. The normalized result is given, and the likelihood function is derived from the hop count decay coefficient. Modeling the level of support for different attention values. Variational posterior distribution. Optimization is achieved by minimizing the KL divergence with the true posterior distribution, and the optimization objective is equivalent to maximizing the lower bound of evidence (ELBO). In actual computation, The scalar convergence value is denoted as The posterior expectation is estimated as ,in To normalize the allocation function, we ensure that the sum of the propagation attention weights of all second-order neighbor nodes is 1. The advantage of this Bayesian fusion mechanism is that: path semantic consistency influences the semantic direction of attention through the dominant pattern contribution vector, and the hop count decay coefficient controls the decay of attention with distance. The two work together in a probabilistic framework, avoiding the semantic information loss caused by simple linear weighting, and achieving a statistically optimal balance between semantic rationality and distance decay constraints in the propagation attention weights.

[0082] The above calculation process is executed during the second-order propagation phase of each round of multi-hop inference, and the results are... Then with activation threshold The comparison determines whether the corresponding second-order neighbor node enters the activated subgraph, thereby effectively transmitting the fine quantification results of path semantic consistency to the subsequent stages of subgraph activation and candidate analysis path generation, ensuring the layer-by-layer refinement and transmission of semantic relevance in the overall reasoning process.

[0083] In one optional implementation, extended nodes are extracted from the activation subgraph, multiple candidate analysis paths are generated based on the semantic relationships of the extended nodes, and a main query task is executed based on the initial query representation to obtain direct query results. Simultaneously, the multiple candidate analysis paths are converted into structured questions, including:

[0084] Nodes whose propagation attention weight gradient change rate exceeds the inflection point detection threshold are selected from the activation subgraph as expansion nodes. Semantic role annotation is performed on the expansion nodes. A role constraint graph is constructed based on the semantic role annotation. On the role constraint graph, the node sequence that satisfies role compatibility is enumerated through the constraint satisfaction propagation algorithm. The node sequence is used as multiple candidate analysis paths.

[0085] For each of the multiple candidate analysis paths, the topological features and semantic relation type combinations of the candidate analysis path are extracted. The topological features are mapped to the syntactic framework of the problem, and the semantic relation type combinations are mapped to the predicate slots in the syntactic framework. A structured problem is generated by assembling the syntactic framework and the predicate slots.

[0086] After completing multi-hop inference of the activation subgraph, it is necessary to identify key nodes with extended analytical value from the activation subgraph. Each node in the activation subgraph carries propagation attention weights, but not all nodes with high weights are suitable as the starting point for extended analysis. Truly valuable extension nodes often appear at the inflection points where attention weights change with topological distance—these nodes mark the transition boundary of semantic propagation from strongly related regions to weakly related regions and have the strongest analytical extension potential.

[0087] Specifically, for each node in the activation subgraph, the gradient rate of change of the propagation attention weight with respect to the propagation step size is calculated along the propagation path it belongs to. Let the node... The propagation attention weight at the location is Its gradient rate of change Defined as the ratio of the absolute value of the weight difference between adjacent propagation steps to the local weight mean, i.e. ,in For nodes The mean of the attention weights of all nodes within the local neighborhood centered on the center. Exceeding the preset inflection point detection threshold At that time, the node These nodes are marked as candidate expansion nodes. This mechanism effectively filters out ordinary propagation nodes with smoothly decreasing weights, retaining only semantic boundary nodes with abrupt weight changes, thus ensuring that the expansion path starts from the position with the most semantic extension meaning.

[0088] After completing the selection of extended nodes, semantic role annotation is performed on each extended node. The goal of semantic role annotation is to assign a functional role to each extended node in the business semantic scenario, such as predefined role types like "analysis subject," "metric," "time dimension," "regional dimension," and "comparison reference." The annotation process relies on a comprehensive judgment based on the type attributes of nodes in the enterprise domain knowledge graph, the conceptual level to which the node belongs, and the semantic distance between the node and the core entity being queried. For business concept nodes, they are preferentially labeled as "analysis subject" or "comparison reference"; for metric nodes, they are labeled as "metric"; and for dimension nodes, they are labeled with the corresponding dimension role according to their dimension type. The results of semantic role annotation provide a structured semantic constraint foundation for subsequent path generation.

[0089] Based on the semantic role annotation results, a role constraint graph is constructed. This graph is a directed graph structure where the node set represents the expanded nodes after annotation, and the edge set is defined by role compatibility rules: if the semantic role combination of two nodes satisfies a predefined role compatibility rule, a directed edge with a compatibility relationship type label is added between them. Role compatibility rules are stored in the meta-layer of the enterprise domain knowledge graph, covering common role combination patterns in business analysis scenarios, and can be dynamically expanded according to business needs.

[0090] After the role constraint graph is constructed, a constraint satisfaction propagation algorithm is used to enumerate the node sequences that satisfy role compatibility. Starting from the extension node corresponding to the query core entity, the algorithm progressively expands the path according to the compatibility relationship type labels of the directed edges in the role constraint graph. During each expansion, it checks whether the role of the newly added node has a compatibility relationship with the role of the terminal node of the current path, and simultaneously checks whether the entire path's role sequence constitutes a semantically complete analysis link. Node sequences that satisfy the above constraints are recorded as candidate analysis paths, and all enumerated node sequences together constitute a set of multiple candidate analysis paths.

[0091] While generating a set of candidate analysis paths, the main query task is executed based on the initial query representation. The main query task transforms the initial query representation into a structured query statement, directly retrieving data directly related to the user's original question from the enterprise data warehouse or knowledge graph, yielding direct query results. These direct query results serve as the core response content for the current round, ensuring that the user's original information needs are met immediately without waiting for the extended analysis process to complete.

[0092] For each of the multiple candidate analysis paths, it needs to be converted into a structured question in natural language form to present to the user for selection. The conversion process consists of four stages: topological structure feature extraction, syntactic frame mapping, semantic relation type combination extraction, and predicate slot filling. In the topological structure feature extraction stage, the geometric attributes of the candidate paths, such as the number of nodes, path depth, and branch structure, are analyzed to form the topological feature vector of the path. Different topological structures correspond to different question sentence types: linear paths correspond to single-chain query sentence types such as "What is C of B in relation to A?", star-shaped paths correspond to multi-dimensional comparison sentence types such as "What are C of A in different dimensions of B?", and chained paths correspond to causal chain sentence types such as "Analysis of the transmission path of A affecting C through B and then affecting D?". This mapping relationship from topological structure to syntactic frame is predefined in a syntactic frame template library, covering the main question sentence types in enterprise data analysis scenarios.

[0093] Once the syntactic framework is determined, the semantic relation type combinations of each edge in the candidate path are extracted. The semantic relation types of the edges are derived from the relation labels of semantically related edges in the enterprise domain knowledge graph, such as "aggregation," "comparison," "trend," "attribution," and "prediction." These semantic relation type combinations are mapped to predicate slots in the syntactic framework, meaning that the relation type of each edge in the path is filled into the predicate placeholder at the corresponding position in the syntactic framework. For example, for a linear path containing an edge with an "aggregation" relation, its syntactic framework is "How are the [indicators] of [subject] summarized by [dimension]?" The "aggregation" relation is mapped to the predicate slot "summary." The business names of each node in the path are filled into the "subject," "indicator," and "dimension" slots, respectively, and finally, a complete structured question text is generated.

[0094] The generated structured question text, along with the direct query results, is presented to the user. The structured questions are arranged in card format, with each card corresponding to a candidate analysis path. The card content includes the structured question text and a brief description of the corresponding business scenario. Users can provide feedback on each card by accepting or rejecting it. This interactive design, which presents the main query results and candidate analysis questions side by side, not only meets the user's immediate information needs but also provides a natural interactive entry point for subsequent reinforcement learning signal collection based on user preferences, forming a closed loop of data-driven analysis path recommendation.

[0095] In one optional implementation, semantic role annotation is performed on the extended nodes, and a role constraint graph is constructed based on the semantic role annotation. On the role constraint graph, a constraint satisfaction propagation algorithm is used to enumerate the sequence of nodes that satisfy role compatibility, including:

[0096] Obtain the relational dependency pattern of the extended node in the enterprise domain knowledge graph, project the relational dependency pattern onto the argument slots of the predefined semantic framework, and label the extended node with the corresponding semantic role according to the functional attributes undertaken by the argument slots in the predefined semantic framework.

[0097] The extended nodes are used as vertices to build a graph structure. For any two nodes in the extended nodes, the semantic role annotation of the two nodes is used to retrieve the role co-occurrence constraints in the semantic dependency rule base. According to the compatibility judgment condition in the role co-occurrence constraints, constraint edges are built in the graph structure for the node pairs that meet the compatibility judgment condition. The constraint edges carry the compatibility strength value corresponding to the compatibility judgment condition, and the role constraint graph is constructed.

[0098] Initialize node selection states and constraint satisfaction on the role constraint graph. Iteratively propagate the node selection states using a constraint satisfaction propagation algorithm. Once the propagation process is stable, select nodes whose node selection states meet the activation conditions to form a node sequence that satisfies role compatibility.

[0099] After obtaining the extended nodes of the activated subgraph, semantic role annotation is required for these nodes to ensure semantic consistency and business rationality in subsequent path generation. The core of semantic role annotation lies in identifying the functional role of each extended node within the enterprise business context, such as whether a node acts as a "metric subject," "time dimension," "geographical scope," or "business action." To this end, the relational dependency pattern of each extended node in the enterprise domain knowledge graph is first obtained, i.e., what types of semantic association edges connect the node to its neighbors and how it participates in the relational structure. Specifically, for extended nodes... The distribution of a node as a head entity, tail entity, or intermediate bridging entity in the knowledge graph is statistically analyzed to form the relation dependency pattern vector of that node. Each dimension of the vector corresponds to the participation frequency or weight of a predefined relation type.

[0100] Dependency Patterns When projecting onto argument slots in a predefined semantic framework, the frame-argument structure from frame semantics is used as the mapping target. The predefined semantic framework is pre-built based on enterprise big data analytics scenarios and includes several typical business analysis frameworks. Each framework has several argument slots; for example, the "sales analysis framework" includes "sales subject slot," "time slot," "product slot," "region slot," and "metric slot," etc. The projection process involves calculation... The prototype vector of each argument slot Cosine similarity between Determine the optimal matching slot ,in Iterate through all argument slots of the currently active semantic frame. The functional attributes undertaken in the predefined semantic framework will be labeled with corresponding semantic roles. Assign extended nodes Complete semantic role labeling. If the optimal matching similarity of a certain extended node is lower than the preset slot matching threshold... If the node is not defined, it will be marked as "undefined role". In the subsequent constraint graph construction, its constraint edges will be assigned a lower compatibility strength value to reduce its priority in being included in the candidate path.

[0101] After semantic role labeling is completed, all extended nodes are used as vertices to construct a graph structure, and constraint edges are created for node pairs based on role co-occurrence constraints, thereby constructing a role constraint graph. For any two extended nodes... and With its semantic role tags and The index retrieves pre-stored role co-occurrence constraint entries from the semantic dependency rule base. The semantic dependency rule base is pre-built based on enterprise business knowledge and linguistic dependency specifications. Each rule records the compatibility determination conditions and corresponding compatibility strength values ​​for a pair of roles in a specific analysis scenario. ,in The larger the value, the higher the semantic rationality of the co-occurrence of the two roles in the same analysis path.

[0102] If the corresponding rule is found and the compatibility criteria are met, then it is a node pair in the graph structure. Establish constraint edges, and The weight attribute is used as the constraint edge. If the role combination of two nodes does not have a corresponding entry in the rule base, or the compatibility judgment condition is not met, then no constraint edge is established, meaning that the semantic legality of the co-occurrence of these two nodes on the same path is not supported. After searching and judging all node pairs one by one, the role constraint graph is finally obtained. The vertex set consists of all extended nodes, and the edge set consists of the constraint edges of all compatible node pairs. Each edge carries the corresponding compatibility strength value.

[0103] In the role constraint diagram The constraint satisfaction propagation algorithm is used to select a sequence of nodes that satisfy the overall role compatibility from a large number of expanded nodes. During algorithm initialization, for each node... Set node selection status The initial value is uniformly set to This indicates that the node is in an undetermined state; simultaneously, the initial constraint satisfaction degree is calculated for each node. The calculation method is to take the weighted average of the compatibility strength values ​​of all constraint edges of the node, that is... ,in Represents a node exist The set of neighboring nodes in the middle, This represents the compatibility strength value for the corresponding constraint edge. If the node... exist If there are no constrained edges, then its initial constraint satisfaction is... Set as .

[0104] Iterative propagation phase by round This process is repeated, updating the selection state and constraint satisfaction of each node in each round. In the Wheel constraint satisfaction The update rule is determined by its current selection state and the propagated information from its neighboring nodes. ,in The self-maintaining coefficient is the proportion of the control node that retains its own historical constraint satisfaction. This corresponds to the propagation contribution from neighboring nodes. Node selection state. This is obtained by mapping the constraint satisfaction degree through the Sigmoid activation function, i.e. ,in For steepness parameters, The central threshold for constraining satisfaction is used to control the sensitive region of the activation curve.

[0105] The stability of the propagation process is determined based on the maximum change in the selection state between two adjacent rounds. At that time, it was believed that the spread had subsided, among which This sets the preset convergence precision. After convergence, the selection status of all nodes is determined. Compare with preset activation conditions, and The node is determined to be an active node, among which Activation thresholds are set for nodes. All activated nodes are topologically sorted according to their constraint edge connections in the role constraint graph, generating a node sequence that satisfies role compatibility. If multiple connected subgraphs each generate activated nodes, the activated nodes in each connected subgraph are sorted separately to generate multiple candidate node sequences. The subsequent path generation module will use these node sequences as a skeleton, combined with the path structure in the knowledge graph, to generate specific candidate analysis paths, thereby ensuring that the final path has business rationality and logical consistency at the semantic role level.

[0106] In one optional implementation, training a reward model using the user preference sample set and updating the attention weight calculation parameters of the graph attention propagation using the output of the reward model as a reinforcement learning signal includes:

[0107] The query response pairs from the user preference sample set are input into the reward model to calculate the preference difference value between different response qualities. Based on the consistency loss between the preference difference value and the preference relationship labeled in the user preference sample set, the parameters of the reward model are optimized so that the score output by the reward model can reflect the user's preference for the quality of the query response.

[0108] The trained reward model scores the query results generated by graph attention propagation, and the score is used as a reward signal for reinforcement learning. A policy gradient is constructed based on the causal relationship between the reward signal and the attention weight distribution in graph attention propagation. The policy gradient determines the optimization direction and adjustment range of the attention weight calculation parameters through sensitivity analysis of the reward signal to the attention weight distribution.

[0109] The attention weight calculation parameters of the graph attention propagation are updated according to the policy gradient.

[0110] When query-response pairs from the user preference sample set are input into the reward model, each sample consists of a query input, a response corresponding to a candidate analysis path accepted by the user (positive sample), and a response corresponding to a candidate analysis path rejected by the user (negative sample). The reward model outputs a scalar score for both positive and negative responses, denoted as the score for the positive sample response. The score for the negative sample response is The difference between the two is the preference variance value. The preference variance value reflects how well the reward model fits the actual user preference relationship: when... When the model's judgment aligns with user preferences, when... At times, the model's judgments contradict user preferences.

[0111] To ensure that the rating distribution of the reward model is consistent with the user preference annotations, a consistency loss function is constructed based on the Bradley-Terry preference probability framework. For the th in the sample set... For each sample, the probability of preference predicted by the model is defined as follows: ,in For the Sigmoid function, This represents the preference variance value corresponding to this sample. The consistency loss is expressed as the negative log-likelihood: By minimizing This allows the reward model to output higher scores on positive samples and lower scores on negative samples, thus ensuring that the score distribution accurately captures users' preferences for query response quality. The parameter updates of the reward model are performed using standard backpropagation and an adaptive gradient optimizer (such as AdamW), iterating multiple times on the user preference sample set during training until... Converge to the preset accuracy range.

[0112] After the trained reward model has its parameters frozen, it is used to score the query results generated by graph attention propagation online. Specifically, when graph attention propagation generates a batch of activation subgraphs based on the current attention weights and produces corresponding query results, these query results are input into the reward model to obtain the corresponding reward score sequence. ,in This represents the number of query results in the current batch. For the first The reward score for each query result. To reduce the impact of randomness in a single score, a weighted average is taken for the scores of multiple paths under the same query. The weights are obtained by normalizing the activation attention weights of the paths, thus obtaining the batch average reward under the current strategy. ,in For the first The normalized weights of the paths and satisfying .

[0113] When using reward scores as reinforcement learning signals, a causal relationship needs to be established between the reward signals and the distribution of attention weights in graph attention propagation. In each hop of inference, the distribution of attention weights on the neighbor set of a node can be viewed as a stochastic policy: given the semantic representation of the current node, the attention weights determine the probability distribution of the propagation path selection. The parameters for calculating the attention weights are denoted as... , No. Path in parameters The log probability of being selected is denoted as Then the estimation form of the policy gradient is:

[0114] ;

[0115] in To reinforce the learning objective function, This is an advantage estimate with the batch average reward as a baseline, used to reduce the variance of gradient estimation. When the reward score of a path is higher than the batch mean, The policy gradient direction adjusts the attention weight calculation parameters corresponding to the path in a direction that increases the probability of choosing the path; conversely, it adjusts them in a direction that decreases the probability.

[0116] Sensitivity analysis of reward signals to attention weight distribution was performed by calculating... In this implementation, at each hop of the graph attention propagation, the attention weights are calculated from the node semantic vectors through linear transformation and Softmax normalization. The linear transformation matrix and the bias term are both parameters. Components. For about The partial derivatives of each component are calculated to obtain the direction and magnitude of the influence of each parameter component on the path selection probability. For the cross-hop propagation parameters involved in the attention weight calculation, the gradient is propagated backward layer by layer along the direction of propagation hops using the time-backward expansion (BPTT) method to ensure that the attention weight calculation parameters of each layer in multi-hop inference can obtain effective gradient information from the reward signal.

[0117] After determining the optimization direction and adjustment range, proceed according to the policy gradient. Attention weight calculation parameters Perform gradient updates. The update rule employs a proximal policy optimization (PPO) mechanism with pruning to prevent excessively large single parameter updates from causing drastic shifts in the attention distribution. Specifically, the probability ratio is defined. ,in For the old strategy before the update, the first The path selection probability, and the objective function after pruning are:

[0118] ;

[0119] in This is the advantage estimate. The clipping range hyperparameter is used to constrain the probability ratio within a certain range. Within the interval. By maximizing... While ensuring stable strategy updates, the attention weight calculation parameters of graph attention propagation are continuously optimized in the direction of improving the reward model score.

[0120] After each round of reinforcement learning is completed, the updated parameters will be... The system is redeployed to the graph attention propagation module to generate new candidate analysis paths in subsequent user queries, and user feedback is collected again to expand the user preference sample set. This forms a closed-loop iterative mechanism of "user feedback - reward model training - reinforcement learning update - path generation optimization". As the number of iterations increases, the accuracy of the reward model in fitting user preferences gradually improves, and the attention weight calculation parameters of graph attention propagation also converge in a direction that better matches the actual analysis needs of users, ultimately leading to a continuous improvement in the quality of candidate analysis path generation and user acceptance.

[0121] In one optional implementation, performing an association query task based on the candidate analysis path accepted by the user to obtain extended query results, and then fusing the extended query results with the direct query results according to semantic relationships to generate a comprehensive analysis report includes:

[0122] The sequence of related query nodes is parsed based on the candidate analysis path accepted by the user. A multi-hop related query task is performed in the enterprise domain knowledge graph along the sequence of related query nodes. The path is traversed through the relationship edges between adjacent nodes in the sequence of related query nodes to obtain the entity attribute information of the terminal node of the sequence of related query nodes as the extended query result.

[0123] Semantic relationship identification is performed on the extended query results and the direct query results. By calculating the association pattern between the vector representations of the extended query results and the direct query results in the semantic space, the semantic relationship type between the extended query results and the direct query results is determined.

[0124] Based on the semantic relationship type, a content organization structure for the comprehensive analysis report is constructed. In the content organization structure, the extended query results and the direct query results are grouped and aggregated according to the semantic relationship type. The grouped and aggregated content is arranged according to the narrative logic corresponding to the semantic relationship type to generate a comprehensive analysis report.

[0125] like Figure 2 As shown, the method includes:

[0126] When executing a related query task based on the candidate analysis paths accepted by the user, the candidate analysis paths are first parsed in a structured manner, extracting the node sequences in the paths into related query node sequences according to the graph traversal order. Each candidate analysis path corresponds to an ordered chain of nodes in the enterprise domain knowledge graph, consisting of business concept nodes, indicator nodes, or dimension nodes connected sequentially. During the parsing process, the relationship edge type between adjacent nodes in the path is read one by one, confirming the semantic definition of the relationship edge in the graph, and recording the directionality and weight attributes of each relationship edge so that subsequent path traversal can perform multi-hop jumps according to the correct semantic direction. In the case of multiple user-accepted paths, the node sequences of each path are parsed independently, and the consistency of the starting point of each node sequence is checked before traversal to ensure that the starting node is semantically aligned with the initial query entity node used in the main query task.

[0127] When performing multi-hop association queries along a sequence of related query nodes in an enterprise domain knowledge graph, the process starts from the initial node of the node sequence and jumps sequentially along each relation edge to the next node until the final node of the sequence is reached. During path traversal, the validity of each relation edge is verified, i.e., it checks whether a corresponding relation edge of the same type exists between the current node and the target node in the graph. If the relation edge does not exist or has expired, a path backtracking mechanism is triggered, attempting to search for a semantically equivalent alternative edge in the graph to complete the jump. When the path traversal reaches the final node, all entity attribute information of that final node is extracted as extended query results. Entity attribute information covers the numerical attributes, text attributes, and associated metadata attached to the final node in the graph. For cases where multiple versions of attribute records exist for the same final node, the latest version of the attribute information is selected based on the principle of prioritizing data timeliness, while key numerical attributes of historical versions are retained for subsequent comparative analysis.

[0128] After obtaining the extended query results, semantic relationship identification is performed between them and the direct query results obtained from the main query task. Specifically, the extended query results and the direct query results are mapped to the vector space of a pre-trained semantic encoder to obtain their respective semantic vector representations. In the semantic vector space, the semantic relationship type is determined by calculating the association pattern between the two sets of result vectors. The calculation of the association pattern comprehensively considers the following dimensions: First, the direction cosine similarity between the two sets of vectors, which measures the overall closeness of the two types of results in terms of semantic content; second, the difference distribution of the two sets of vectors in each semantic dimension, which is used to identify in which semantic dimensions the extended query results have significant deviations from the direct query results, thereby determining whether they belong to a complementary or contrastive relationship; third, the node distance and path type of the two sets of results in the knowledge graph, which is used to help determine whether there is a causal relationship, a subordinate relationship, or a parallel relationship between them. Based on the calculation results of the above three dimensions, the semantic relationship types are classified into the following categories: causal relationship, complementary relationship, contrastive relationship, subordinate relationship, and parallel relationship.

[0129] After determining the semantic relationship type, the content organization structure of the comprehensive analysis report is constructed according to the type. Different semantic relationship types correspond to different content organization logics: when the semantic relationship is causal, the report adopts a "conclusion first - cause expansion" narrative structure, first presenting the core business phenomenon revealed by the direct query results, and then introducing the driving factors or influence paths described by the extended query results, forming an analysis chain from appearance to mechanism; when the semantic relationship is complementary, the report adopts a parallel structure of "main information - background information", placing the direct query results as the main content in the core position of the report, and embedding the extended query results as background supplementary information in the form of notes or explanations, so that readers can understand the business background while obtaining the core conclusions.

[0130] When the semantic relationship is comparative, the report adopts a "horizontal comparison - difference analysis" narrative structure, displaying direct query results and extended query results side by side within the same dimensional framework, and quantifying the numerical or trend differences between them. When the semantic relationship is subordinate, the report adopts a "general-specific" hierarchical structure, using direct query results as the overall description and extended query results as the detailed breakdown of each sub-dimension, arranging the sub-items according to business importance or numerical size. When the semantic relationship is parallel, the report adopts a parallel narrative structure, placing direct query results and extended query results at the same level, and presenting them sequentially according to time order or business logic order.

[0131] After determining the content organization structure, the extended query results and direct query results are grouped and aggregated according to semantic relationship types. The specific operations for grouping and aggregation are as follows: results belonging to the same semantic relationship type are grouped into the same content group; multiple results within the same content group are deduplicated and merged—for numerical attributes, their aggregation statistics are calculated to form summary data within the group; for textual attributes, key semantic fragments of each result are extracted and concatenated into coherent descriptive paragraphs. After grouping and aggregation, the content groups are arranged according to the narrative logic corresponding to the semantic relationship type, and the arranged content groups are sequentially filled into the corresponding positions in the content organization structure to form a comprehensive analysis report with a complete structure and logical coherence. The final output format of the comprehensive analysis report supports multiple formats such as structured text, embedded charts, and visual data dashboards to meet the differentiated needs of users for report presentation formats in different business scenarios. After the report is generated, the content organization structure and semantic relationship judgment results of this report are used as samples to feed back to the user preference sample set, providing data support for the continuous optimization of the subsequent reward model.

[0132] A second aspect of this invention provides an AI-driven enterprise big data intelligent analysis service system, comprising:

[0133] The graph construction unit is used to build an enterprise domain knowledge graph. It receives natural language query input from users, extracts core query entities and query relationships, and obtains an initial query representation.

[0134] The multi-hop inference unit is used to map the initial query representation to the corresponding node of the enterprise domain knowledge graph, and to perform multi-hop inference based on graph attention propagation starting from the corresponding node. It calculates the semantic association strength between nodes, propagates attention weights layer by layer, and activates relevant subgraph regions to obtain activated subgraphs.

[0135] The path generation unit is used to extract extended nodes from the activated subgraph, generate multiple candidate analysis paths based on the semantic relationships of the extended nodes, execute the main query task based on the initial query representation to obtain direct query results, and convert the multiple candidate analysis paths into structured questions to present to the user.

[0136] The feedback training unit is used to receive user feedback on the structured question, record the candidate analysis paths of user acceptance and rejection and label them as positive and negative samples, construct a user preference sample set, train a reward model using the user preference sample set, and use the output of the reward model as a reinforcement learning signal to update the attention weight calculation parameters of the graph attention propagation.

[0137] The report generation unit is used to perform related query tasks based on the candidate analysis paths accepted by the user to obtain extended query results, and to fuse the extended query results with the direct query results according to semantic relationships to generate a comprehensive analysis report.

[0138] A third aspect of the present invention provides an electronic device, comprising:

[0139] processor;

[0140] Memory used to store processor-executable instructions;

[0141] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0142] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0143] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0144] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-driven intelligent analysis service method for enterprise big data, characterized in that, include: Construct an enterprise domain knowledge graph, receive users' natural language query input, extract core query entities and query relationships, and obtain an initial query representation; The initial query representation is mapped to the corresponding node of the enterprise domain knowledge graph. Multi-hop reasoning based on graph attention propagation is performed starting from the corresponding node. Attention weights are propagated layer by layer by calculating the semantic association strength between nodes and activating relevant subgraph regions to obtain activated subgraphs. Extended nodes are extracted from the activated subgraph, and multiple candidate analysis paths are generated based on the semantic relationships of the extended nodes. The main query task is executed based on the initial query representation to obtain direct query results. At the same time, the multiple candidate analysis paths are converted into structured questions and presented to the user. Receive user feedback on the structured question, record the candidate analysis paths of user acceptance and rejection and label them as positive and negative samples, construct a user preference sample set, train a reward model using the user preference sample set, and use the output of the reward model as a reinforcement learning signal to update the attention weight calculation parameters of the graph attention propagation. Based on the candidate analysis path accepted by the user, an associated query task is executed to obtain extended query results. The extended query results are then fused with the direct query results according to semantic relationships to generate a comprehensive analysis report.

2. The method according to claim 1, characterized in that, The initial query representation is mapped to the corresponding node in the enterprise domain knowledge graph. Multi-hop inference based on graph attention propagation is then performed starting from the corresponding node. Attention weights are propagated layer by layer by calculating the semantic association strength between nodes, and relevant subgraph regions are activated. The resulting activated subgraphs include: Discriminative semantic representations are generated by comparing the initial query representation with the counterexample query using an adversarial dual encoder. In the enterprise domain knowledge graph, the node is jointly ranked by subgraph isomorphic matching and node role similarity. The node with the highest ranking is constructed as an attention source with an initial potential field as the corresponding node. Obtain the set of first-order neighbor nodes of the corresponding node in the enterprise domain knowledge graph connected by semantic association edges, and parse the relation type and direction attribute of each semantic association edge according to the semantic dependency tree decomposition of the edge. Based on the semantic role of the relation type and the second-order matching of the query slot in the initial query representation, generate initial attention weights for each node in the set of first-order neighbor nodes. Nodes whose initial attention weights exceed the propagation threshold in the first-order neighbor node set are selected as activation nodes. The second-order neighbor nodes of the activation nodes are obtained. The path semantic consistency is aggregated with the initial attention weights of the activation nodes through semantic interaction modeling in the tensor product space, and a hop count decay coefficient is applied to obtain the propagation attention weights of the second-order neighbor nodes. Adaptive step-size diffusion is performed along the information entropy gradient direction to perform filtering and weight calculation until the attention weight of all nodes is lower than the propagation threshold or reaches the preset inference hop count. All nodes in each layer whose attention weight exceeds the activation threshold and their connecting edges are then combined to form an activation subgraph.

3. The method according to claim 2, characterized in that, By modeling semantic interactions in tensor product space, the path semantic consistency is aggregated with the initial attention weights of the activated nodes, and a hop count decay coefficient is applied to obtain the propagation attention weights of the second-order neighbor nodes, including: For the reasoning path from the corresponding node through the activated node to the second-order neighbor node, a relational algebraic expression of the path is constructed. By decomposing the relational algebraic expression into the minimum non-redundant relation atom set, the satisfiability of the logical implication relationship between each relation atom and the query predicate in the initial query representation is solved. The semantic consistency of the path is quantified based on the proportion of relation atoms that satisfy the constraints in the solution results. The path semantic consistency is encoded as a diagonalized semantic operator in the tensor product space, and the initial attention weights of the activated nodes are embedded as eigenstate vectors in the tensor product space. The spectral decomposition and reconstruction of the eigenstate vectors by the diagonalized semantic operator are used to extract the dominant mode contribution of path semantics to attention distribution. Based on the hop count of the inference path, construct an information decay Markov chain, and extract the transition probability related to the hop count from the stationary distribution of the Markov chain as the hop count decay coefficient. The dominant mode contribution and the hop count decay coefficient are fused in Bayesian mode in the posterior distribution space using a variational inference framework to obtain the propagation attention weights of the second-order neighbor nodes.

4. The method according to claim 1, characterized in that, Extended nodes are extracted from the activated subgraph, and multiple candidate analysis paths are generated based on the semantic relationships of the extended nodes. The main query task is executed based on the initial query representation to obtain direct query results. Simultaneously, the multiple candidate analysis paths are converted into structured questions, including: Nodes whose propagation attention weight gradient change rate exceeds the inflection point detection threshold are selected from the activation subgraph as expansion nodes. Semantic role annotation is performed on the expansion nodes. A role constraint graph is constructed based on the semantic role annotation. On the role constraint graph, the node sequence that satisfies role compatibility is enumerated through the constraint satisfaction propagation algorithm. The node sequence is used as multiple candidate analysis paths. For each of the multiple candidate analysis paths, the topological features and semantic relation type combinations of the candidate analysis path are extracted. The topological features are mapped to the syntactic framework of the problem, and the semantic relation type combinations are mapped to the predicate slots in the syntactic framework. A structured problem is generated by assembling the syntactic framework and the predicate slots.

5. The method according to claim 4, characterized in that, The extended nodes are semantically labeled, and a role constraint graph is constructed based on the semantic role labels. On the role constraint graph, a constraint satisfaction propagation algorithm is used to enumerate the sequence of nodes that satisfy role compatibility, including: Obtain the relational dependency pattern of the extended node in the enterprise domain knowledge graph, project the relational dependency pattern onto the argument slots of the predefined semantic framework, and label the extended node with the corresponding semantic role according to the functional attributes undertaken by the argument slots in the predefined semantic framework. The extended nodes are used as vertices to build a graph structure. For any two nodes in the extended nodes, the semantic role annotation of the two nodes is used to retrieve the role co-occurrence constraints in the semantic dependency rule base. According to the compatibility judgment condition in the role co-occurrence constraints, constraint edges are built in the graph structure for the node pairs that meet the compatibility judgment condition. The constraint edges carry the compatibility strength value corresponding to the compatibility judgment condition, and the role constraint graph is constructed. Initialize node selection states and constraint satisfaction on the role constraint graph. Iteratively propagate the node selection states using a constraint satisfaction propagation algorithm. Once the propagation process is stable, select nodes whose node selection states meet the activation conditions to form a node sequence that satisfies role compatibility.

6. The method according to claim 1, characterized in that, Training a reward model using the user preference sample set, and then using the output of the reward model as a reinforcement learning signal to update the attention weight calculation parameters of the graph attention propagation, includes: The query response pairs from the user preference sample set are input into the reward model to calculate the preference difference value between different response qualities. Based on the consistency loss between the preference difference value and the preference relationship labeled in the user preference sample set, the parameters of the reward model are optimized so that the score output by the reward model can reflect the user's preference for the quality of the query response. The trained reward model scores the query results generated by graph attention propagation, and the score is used as a reward signal for reinforcement learning. A policy gradient is constructed based on the causal relationship between the reward signal and the attention weight distribution in graph attention propagation. The policy gradient determines the optimization direction and adjustment range of the attention weight calculation parameters through sensitivity analysis of the reward signal to the attention weight distribution. The attention weight calculation parameters of the graph attention propagation are updated according to the policy gradient.

7. The method according to claim 1, characterized in that, Based on the candidate analysis path accepted by the user, an associated query task is performed to obtain extended query results. The extended query results are then fused with the direct query results according to semantic relationships to generate a comprehensive analysis report, including: The sequence of related query nodes is parsed based on the candidate analysis path accepted by the user. A multi-hop related query task is performed in the enterprise domain knowledge graph along the sequence of related query nodes. The path is traversed through the relationship edges between adjacent nodes in the sequence of related query nodes to obtain the entity attribute information of the terminal node of the sequence of related query nodes as the extended query result. Semantic relationship identification is performed on the extended query results and the direct query results. By calculating the association pattern between the vector representations of the extended query results and the direct query results in the semantic space, the semantic relationship type between the extended query results and the direct query results is determined. Based on the semantic relationship type, a content organization structure for the comprehensive analysis report is constructed. In the content organization structure, the extended query results and the direct query results are grouped and aggregated according to the semantic relationship type. The grouped and aggregated content is arranged according to the narrative logic corresponding to the semantic relationship type to generate a comprehensive analysis report.

8. An AI-driven enterprise big data intelligent analysis service system, used to implement the method as described in any one of claims 1-7, characterized in that, include: The graph construction unit is used to build an enterprise domain knowledge graph. It receives natural language query input from users, extracts core query entities and query relationships, and obtains an initial query representation. The multi-hop inference unit is used to map the initial query representation to the corresponding node of the enterprise domain knowledge graph, and to perform multi-hop inference based on graph attention propagation starting from the corresponding node. It calculates the semantic association strength between nodes, propagates attention weights layer by layer, and activates relevant subgraph regions to obtain activated subgraphs. The path generation unit is used to extract extended nodes from the activated subgraph, generate multiple candidate analysis paths based on the semantic relationships of the extended nodes, execute the main query task based on the initial query representation to obtain direct query results, and convert the multiple candidate analysis paths into structured questions to present to the user. The feedback training unit is used to receive user feedback on the structured question, record the candidate analysis paths of user acceptance and rejection and label them as positive and negative samples, construct a user preference sample set, train a reward model using the user preference sample set, and use the output of the reward model as a reinforcement learning signal to update the attention weight calculation parameters of the graph attention propagation. The report generation unit is used to perform related query tasks based on the candidate analysis paths accepted by the user to obtain extended query results, and to fuse the extended query results with the direct query results according to semantic relationships to generate a comprehensive analysis report.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Citation Information

Patent Citations

  • Policy intelligent question and answer method and system based on knowledge graph and storage medium

    CN121326943A

  • Knowledge graph-based dynamic retrieval enhancement generation method and system, terminal and medium

    CN121327107A