Intelligent decision support method and system for digital application of government affair cloud platform

Abstract

This invention relates to the field of e-government cloud platform technology, and in particular to an intelligent decision support method and system for digital applications of e-government cloud platforms. It acquires multi-source heterogeneous data from e-government business scenarios and performs semantic parsing and feature encoding to obtain encoded feature vectors. It utilizes a dual-stream decision network architecture, employing first-stream rule-driven reasoning and second-stream data-driven learning, and constructs a fusion decision space through an adversarial verification mechanism. Within this space, it builds a decision path search tree and uses Monte Carlo tree search to expand nodes, calculate values, determine the optimal search path, extract decision action sequences to obtain the target decision scheme, generate decision instructions, issue and execute them, and collect feedback to update search parameters. This achieves intelligent and adaptive optimization of e-government decision-making, significantly improving decision accuracy and business response efficiency.

Description

Technical Field

[0001] This invention relates to the field of government cloud platform technology, and in particular to an intelligent decision support method and system for digital applications of government cloud platforms. Background Technology

[0002] In the digital application scenarios of government cloud platforms, intelligent decision support typically relies on a single data-driven model or a rule-based system. Data-driven models predict trends or recommend solutions by analyzing historical business data, such as using traditional machine learning algorithms to train structured business data to form decision classifiers or regression models. Rule-based systems, on the other hand, rely on predefined logical rules by domain experts to associate business conditions with corresponding actions to achieve compliance review or process guidance. Some solutions attempt to simply combine the two, such as using rule systems to post-process and constrain the output of data models, or using data models to provide dynamic parameters for rule matching. However, such integrations often remain at the level of serial or weighted averaging, lacking deep interaction and verification mechanisms.

[0003] Existing practices suffer from two main drawbacks. First, the reasoning results of rule systems and data models are prone to conflict in complex government scenarios. For example, rules are based on static regulations while data models reflect real-time dynamics, potentially leading to contradictory decision recommendations for the same business state. Current technology lacks effective conflict identification and fusion strategies, causing systems to tend to adopt a single source or directly interrupt the process when conflicts occur, making it difficult to guarantee the robustness and reliability of decisions. Second, decision-making schemes often employ fixed algorithms or search strategies, such as exhaustive or greedy search. Given the large state space and dynamically changing action options in government operations, these methods are inefficient and unable to adapt. More importantly, feedback data after decision execution is often ignored or used only for offline retraining, failing to form a closed-loop optimization mechanism. This prevents the decision model from being updated online based on real-time business results, leading to significant performance degradation over long-term operation. Summary of the Invention

[0004] This invention provides an intelligent decision support method and system for digital applications of government cloud platforms, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides an intelligent decision support method for digital applications of a government cloud platform, comprising:

[0006] Acquire multi-source heterogeneous data in government business scenarios, including structured business data, unstructured document data, and real-time interactive data; perform semantic parsing and feature encoding on the multi-source heterogeneous data to obtain encoded feature vectors.

[0007] A dual-stream decision network architecture is constructed. The first decision stream performs rule-driven reasoning on the encoded feature vector to obtain rule-based reasoning results, while the second decision stream performs data-driven learning on the encoded feature vector to obtain learned reasoning results. An adversarial verification mechanism is used to cross-validate the rule-based reasoning results and the learned reasoning results, identifying the consistency and conflict regions between them, and obtaining a fused decision space.

[0008] In the fusion decision space, a decision path search tree is constructed, with the current business state as the root node. The decision action nodes are expanded using the Monte Carlo tree search method, and the strategy value and state value are calculated to determine the optimal search path. The decision action sequence is extracted along the optimal search path to obtain the target decision scheme.

[0009] Decision instructions are generated based on the target decision scheme and sent to the government application module for execution. Business feedback data after execution is collected, and the parameters of the Monte Carlo tree search method are updated.

[0010] Semantic parsing and feature encoding are performed on the multi-source heterogeneous data to obtain encoded feature vectors, including:

[0011] Perform field semantic analysis on the structured business data in the multi-source heterogeneous data, identify the business attribute types and value constraint rules of the fields, and label each field with semantic tags and business meanings according to the field name and field content to obtain a structured semantic representation;

[0012] Unstructured document data is analyzed for document structure, and the heading levels, paragraph boundaries and chapter topics are identified. Each paragraph is mapped to a topic vector and cosine similarity is calculated. The paragraphs are clustered based on the cosine similarity to obtain the document semantic clustering results.

[0013] Multi-turn dialogue modeling is performed on real-time interactive data to extract the information of dereference resolution and omission completion in the dialogue. The pronouns are replaced with the entities they point to, and the omitted components are completed into complete semantic units to form a dereferenced complete dialogue semantic sequence, thus obtaining the interactive semantic sequence.

[0014] The structured semantic representation, the document semantic clustering result, and the interaction semantic sequence are respectively embedded into vectors. The embedded vectors are then propagated using a neighborhood information aggregation method. The propagated vectors are then concatenated and normalized to obtain the encoded feature vector.

[0015] The first decision stream performs rule-driven inference on the encoded feature vector to obtain a rule-driven inference result, and the second decision stream performs data-driven learning on the encoded feature vector to obtain a learning inference result, including:

[0016] The government business rules are used as nodes, the dependencies between rules are used as directed edges, and a rule reasoning graph is constructed. Feature matching is performed on the encoded feature vector to activate the initial nodes in the rule reasoning graph, thus obtaining a set of activated nodes.

[0017] Starting from the set of activated nodes, a breadth-first traversal is performed along the directed edges of the rule reasoning graph. During the traversal, it is determined whether the preconditions of each visited node are satisfied by the encoded feature vector. Nodes that satisfy the preconditions are added to the reasoning queue, and the reasoning operations corresponding to the nodes in the reasoning queue are executed in sequence to obtain the intermediate reasoning state.

[0018] The intermediate inference states are subjected to consistency verification to detect whether there are logically contradictory inference results. The inference operations that cause contradictions are undone through a backtracking mechanism, and a feasible inference path is selected again until a contradiction-free rule inference result is obtained.

[0019] The encoded feature vector is projected onto the decision embedding space, and the historical decision sample that is closest to the encoded feature vector is retrieved in the decision embedding space to obtain the neighbor sample set.

[0020] Decision patterns are extracted from the samples in the neighboring sample set, common decision features among the samples are identified, and feature fusion is performed based on the common decision features and the encoded feature vector. Learning inference results are generated by fusing the features.

[0021] An adversarial verification mechanism is used to cross-validate the rule-based reasoning results with the learned reasoning results, identifying areas of consistency and conflict between the two, thus obtaining a fused decision space, including:

[0022] The rule-based reasoning results and the learning-based reasoning results are respectively decomposed into decision features to obtain multiple rule-based decision units and multiple learning-based decision units, and a semantic correspondence between the rule-based decision units and the learning-based decision units is established.

[0023] For each pair of corresponding decision units, semantic similarity is calculated. The vector angle between decision units is measured by the cosine similarity method. Decision unit pairs with semantic similarity higher than the upper similarity bound are marked as consistent decision pairs, and decision unit pairs with semantic similarity lower than the lower similarity bound are marked as conflicting decision pairs.

[0024] The confidence fusion is performed on the consistent decision pair to obtain the rule confidence of the rule decision unit and the learning confidence of the learning decision unit. The feature vectors of the two decision units are linearly combined according to the confidence ratio to obtain a consistent fusion vector set.

[0025] An adversarial evaluation is performed on the conflict decision pair, an evaluation sample set is constructed, and decision prediction is performed using the rule decision unit and the learning decision unit in the conflict decision pair respectively. The matching degree between the prediction result and the real label is calculated, and the decision unit is selected as the effective decision based on the matching degree to obtain the conflict resolution vector set. The consistency fusion vector set and the conflict resolution vector set are spliced ​​and recombined to construct a complete decision feature space as the fusion decision space.

[0026] A decision path search tree is constructed in the fused decision space, with the current business state as the root node. Decision action nodes are expanded using the Monte Carlo tree search method, and policy value and state value are calculated to determine the optimal search path. A sequence of decision actions is extracted along the optimal search path to obtain the target decision scheme, including:

[0027] Construct a decision path search tree and map the current business state to the root node. Assign an initial access count and accumulated value to the root node. Extract a set of decision candidates for the current business state from the fused decision space. Create a child node for each decision candidate in the decision candidate set and connect it to the root node as a decision action node.

[0028] The iterative search process is executed. In each iteration, the confidence upper bound of each decision action node is calculated. The decision action node with the highest confidence upper bound is simulated and expanded. The transition state after executing the corresponding decision action is calculated through the state transition function. The transition state is input into the policy network to obtain the action selection probability distribution as the policy value. The transition state score is input into the value network to obtain the state evaluation score as the state value.

[0029] The strategy value and the state value are linearly combined to obtain a comprehensive score, and the score is updated backtracking to all nodes on the path from the root node to the selected node, updating the cumulative value and access count of each node.

[0030] Repeat the above iterative search process until the search termination condition is met. Select the path with the highest cumulative value starting from the root node as the optimal search path in the decision path search tree. Collect the decision actions of each node along the optimal search path, construct the decision action sequence and output it as the target decision scheme.

[0031] The decision action node with the highest confidence upper bound is simulated and expanded. The transition state after executing the corresponding decision action is calculated through the state transition function, and then input into the policy network to obtain the action selection probability distribution as the policy value. The state evaluation score is then input into the value network to obtain the state value, including:

[0032] Obtain the decision action identifier and current business state vector corresponding to the decision action node with the highest confidence upper bound value, and convert the decision action identifier into an action embedding vector;

[0033] The current business state vector is concatenated with the action embedding vector to form a joint state input. The joint state input is then subjected to nonlinear transformation and state evolution calculation through a state transition function to output the state vector representation of the transition state.

[0034] The state vector of the transition state is input into the encoding layer of the policy network for feature extraction. The state features are abstracted and represented by a multi-layer neural network structure. After processing by a fully connected layer and activation function, the output is sent to the action space mapping layer. In the action space mapping layer, the feature vector is mapped into a score vector for each candidate action. The score vector is normalized to obtain the action selection probability distribution as the policy value.

[0035] The state vector representation of the transition state is input into the value network for feature encoding and value prediction, and the output state evaluation score is used as the state value.

[0036] A second aspect of this invention provides an intelligent decision support system for digital applications of a government cloud platform, comprising:

[0037] The data encoding unit is used to acquire multi-source heterogeneous data in government business scenarios, including structured business data, unstructured document data and real-time interactive data, and to perform semantic parsing and feature encoding on the multi-source heterogeneous data to obtain encoded feature vectors.

[0038] A dual-stream decision unit is used to construct a dual-stream decision network architecture. The first decision stream performs rule-driven reasoning on the encoded feature vector to obtain rule-driven reasoning results, and the second decision stream performs data-driven learning on the encoded feature vector to obtain learning reasoning results. An adversarial verification mechanism is used to cross-validate the rule-driven reasoning results and the learning reasoning results, identify the consistency and conflict regions between the two, and obtain a fusion decision space.

[0039] The decision search unit is used to construct a decision path search tree in the fused decision space, take the current business state as the root node, expand the decision action nodes through the Monte Carlo tree search method, calculate the strategy value and state value to determine the optimal search path, extract the decision action sequence along the optimal search path, and obtain the target decision scheme.

[0040] The execution update unit is used to generate decision instructions based on the target decision scheme and send them to the government application module for execution, collect business feedback data after execution, and update the parameters of the Monte Carlo tree search method.

[0041] A third aspect of the present invention provides an electronic device, comprising:

[0042] processor;

[0043] Memory used to store processor-executable instructions;

[0044] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0045] 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.

[0046] This invention achieves unified representation of structured business data, unstructured document data, and real-time interactive data by semantic parsing and feature encoding of multi-source heterogeneous data, which greatly improves the utilization efficiency of fragmented information in government affairs scenarios.

[0047] The synergistic effect of rule-driven reasoning and data-driven learning in the dual-stream decision network architecture accurately identifies the consistency and conflict areas between rules and data through an adversarial verification mechanism. This constructs a fusion decision space that takes into account both prior knowledge and implicit rules, avoiding the systematic biases caused by a single reasoning path. This enables the system to both comply with policies and regulations and adapt to dynamic changes in complex government affairs.

[0048] Constructing a decision path search tree and selecting the optimal path in the fusion decision space can effectively address the high-dimensional state space and long-cycle impact in government decision-making, generate executable decision action sequences in real time, significantly improve the efficiency and quality of solution exploration, and ensure that the decision solution approaches the global optimum under limited computing resources.

[0049] The execution instructions issued based on the target decision-making scheme generate feedback data in government applications, dynamically updating the Monte Carlo tree search parameters and forming a closed-loop optimization mechanism. This continuous learning capability enables the decision model to adapt to the evolution of the business environment and adjustments in user needs, gradually accumulating domain experience, avoiding the limitations of outdated rules or static models, and maintaining high accuracy and timeliness of decision support in the long term. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the intelligent decision support method for digital applications of the government cloud platform according to an embodiment of the present invention.

[0051] Figure 2 This is a flowchart illustrating the method for generating a fusion decision space based on an adversarial verification mechanism according to 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 intelligent decision support method for digital applications of a government cloud platform according to an embodiment of the present invention. The present invention provides an intelligent decision support method for digital applications of a government cloud platform, comprising:

[0055] Acquire multi-source heterogeneous data in government business scenarios, including structured business data, unstructured document data, and real-time interactive data; perform semantic parsing and feature encoding on the multi-source heterogeneous data to obtain encoded feature vectors.

[0056] A dual-stream decision network architecture is constructed. The first decision stream performs rule-driven reasoning on the encoded feature vector to obtain rule-based reasoning results, while the second decision stream performs data-driven learning on the encoded feature vector to obtain learned reasoning results. An adversarial verification mechanism is used to cross-validate the rule-based reasoning results and the learned reasoning results, identifying the consistency and conflict regions between them, and obtaining a fused decision space.

[0057] In the fusion decision space, a decision path search tree is constructed, with the current business state as the root node. The decision action nodes are expanded using the Monte Carlo tree search method, and the strategy value and state value are calculated to determine the optimal search path. The decision action sequence is extracted along the optimal search path to obtain the target decision scheme.

[0058] Decision instructions are generated based on the target decision scheme and sent to the government application module for execution. Business feedback data after execution is collected, and the parameters of the Monte Carlo tree search method are updated.

[0059] Semantic parsing and feature encoding are performed on the multi-source heterogeneous data to obtain encoded feature vectors, including:

[0060] Perform field semantic analysis on the structured business data in the multi-source heterogeneous data, identify the business attribute types and value constraint rules of the fields, and label each field with semantic tags and business meanings according to the field name and field content to obtain a structured semantic representation;

[0061] Unstructured document data is analyzed for document structure, and the heading levels, paragraph boundaries and chapter topics are identified. Each paragraph is mapped to a topic vector and cosine similarity is calculated. The paragraphs are clustered based on the cosine similarity to obtain the document semantic clustering results.

[0062] Multi-turn dialogue modeling is performed on real-time interactive data to extract the information of dereference resolution and omission completion in the dialogue. The pronouns are replaced with the entities they point to, and the omitted components are completed into complete semantic units to form a dereferenced complete dialogue semantic sequence, thus obtaining the interactive semantic sequence.

[0063] The structured semantic representation, the document semantic clustering result, and the interaction semantic sequence are respectively embedded into vectors. The embedded vectors are then propagated using a neighborhood information aggregation method. The propagated vectors are then concatenated and normalized to obtain the encoded feature vector.

[0064] In government business scenarios, multi-source heterogeneous data includes three types of sources: structured business data, unstructured document data, and real-time interactive data. The semantic characteristics and encoding methods of each type of data are fundamentally different, requiring the adoption of targeted parsing strategies and the generation of encoded feature vectors through a unified feature fusion mechanism.

[0065] For structured business data, when performing field semantic analysis, it is necessary to extract information from each field in the database table or form, including both the field name and the field content. Field names typically carry naming conventions that reflect business meaning, such as "applicant's ID number," "approval status," and "acceptance date." Using a pre-built government domain dictionary and named entity recognition model, the field names are mapped to a predefined business attribute type system, identifying whether the field belongs to a specific type such as identity, time, status enumeration, or numerical quantification. The field content is used to infer value constraint rules. For example, enumeration constraints are identified by statistically analyzing the field's value range distribution, format constraints by regular expression pattern matching, and boundary constraints by numerical range analysis. After jointly analyzing the field names and content, each field is labeled with a semantic tag. The semantic tagging uses a three-level labeling system: "business domain - attribute type - semantic role." For example, a field might be labeled "social security business - time type - enrollment start date," along with accompanying descriptive text about its business meaning, ultimately resulting in a structured semantic representation. This annotation method enables semantic alignment of similar fields from different business systems, eliminating semantic ambiguity caused by naming differences.

[0066] For unstructured document data, when performing text structure analysis, the document's heading hierarchy is identified through layout features such as font size, indentation levels, and numbering patterns. The document is then parsed into a hierarchical structure composed of chapter trees. Paragraph boundaries are determined based on rules such as blank lines, indentation variations, and punctuation distribution. Chapter themes are obtained through keyword extraction and theme modeling of the paragraph text. The text content of each paragraph is input into a pre-trained language model to extract a dense vector representation of the paragraph, denoted as the theme vector. For any two paragraphs... and Topic Vectors and Calculate cosine similarity This is used to measure the semantic similarity between paragraphs. Based on the obtained similarity matrix, a hierarchical clustering method is used to cluster the paragraphs, grouping semantically highly related paragraphs into the same cluster. Each cluster represents a set of content in the document that revolves around a core issue, ultimately yielding the document semantic clustering results. The clustering results preserve the original hierarchical information of the paragraphs, allowing the subsequent vector embedding stage to utilize both local paragraph semantics and global document structure information.

[0067] For real-time interactive data, multi-turn dialogue modeling is a crucial step in understanding users' true intentions. Dialogues in government scenarios often exhibit numerous instances of referentialism and ellipsis. For example, a user might use "it" in a second round to refer to a specific policy name mentioned in a previous round, or omit the subject to directly express their request. Referential resolution involves constructing a candidate set of entities from the dialogue history, using an attention mechanism to calculate the association score between the current referent and each candidate entity, and selecting the entity with the highest score as the resolution target, replacing the referent with the complete expression of the corresponding entity in the dialogue text. Ellipsis completion, on the other hand, infers omitted components based on contextual semantics. For example, it completes omitted subjects in subsequent sentences based on subject information from previous dialogues, and completes missing objects based on the semantic framework of the predicate verb. After referential resolution and ellipsis completion, each round of dialogue discourse forms a semantically complete and clearly defined independent semantic unit. By sequentially piecing together the multi-turn discourse, a dereferentialized complete dialogue semantic sequence, i.e., the interactive semantic sequence, is obtained. Each semantic unit in the interactive semantic sequence corresponds to a specific operational intent or information query need in government affairs, providing accurate intent input for subsequent decision-making and reasoning.

[0068] After obtaining the structured semantic representation, document semantic clustering results, and interaction semantic sequence, vector embedding operations are performed on the three types of semantic representations to transform them into a dense vector space representation of a unified dimension. In the structured semantic representation, the semantic label of each field is used to obtain a field embedding vector by looking up the embedding matrix. In the document semantic clustering results, each cluster is used to obtain a cluster embedding vector by taking a weighted average of the paragraph vectors within the cluster. In the interaction semantic sequence, each semantic unit is used to obtain a discourse embedding vector through an encoder. Subsequently, a neighborhood information aggregation method is used to propagate the features of the above embedding vectors, constructing a heterogeneous graph structure from the embedding vectors of the three data sources. Nodes represent various semantic units, and edges are established based on semantic relevance and business relevance. The core operation of neighborhood information aggregation is to collect the vector representations of each node's neighboring nodes and then use a learnable aggregation function to weightedly fuse the neighbor information, updating the current node's vector representation. After several rounds of neighborhood information aggregation, the vector representation of each node not only contains its own local semantic information but also incorporates contextual semantic information from related nodes, thereby achieving semantic association propagation across data sources.

[0069] After feature propagation, the propagated vectors from all nodes are pooled according to their source (structured data, document data, and interaction data) to obtain three global semantic vectors. These three vectors are then concatenated to form a high-dimensional concatenated vector. Normalization is then performed on the concatenated vector to scale the values ​​of each dimension to a uniform range, eliminating feature bias caused by differences in vector scales from different data sources. This results in the encoded feature vector. This encoded feature vector comprehensively carries the field semantics of structured business data, the topic clustering semantics of unstructured document data, and the dialogue intent semantics of real-time interaction data, providing a unified and complete feature input foundation for rule-driven reasoning and data-driven learning in the subsequent two-stream decision network architecture.

[0070] The first decision stream performs rule-driven inference on the encoded feature vector to obtain a rule-driven inference result, and the second decision stream performs data-driven learning on the encoded feature vector to obtain a learning inference result, including:

[0071] The government business rules are used as nodes, the dependencies between rules are used as directed edges, and a rule reasoning graph is constructed. Feature matching is performed on the encoded feature vector to activate the initial nodes in the rule reasoning graph, thus obtaining a set of activated nodes.

[0072] Starting from the set of activated nodes, a breadth-first traversal is performed along the directed edges of the rule reasoning graph. During the traversal, it is determined whether the preconditions of each visited node are satisfied by the encoded feature vector. Nodes that satisfy the preconditions are added to the reasoning queue, and the reasoning operations corresponding to the nodes in the reasoning queue are executed in sequence to obtain the intermediate reasoning state.

[0073] The intermediate inference states are subjected to consistency verification to detect whether there are logically contradictory inference results. The inference operations that cause contradictions are undone through a backtracking mechanism, and a feasible inference path is selected again until a contradiction-free rule inference result is obtained.

[0074] The encoded feature vector is projected onto the decision embedding space, and the historical decision sample that is closest to the encoded feature vector is retrieved in the decision embedding space to obtain the neighbor sample set.

[0075] Decision patterns are extracted from the samples in the neighboring sample set, common decision features among the samples are identified, and feature fusion is performed based on the common decision features and the encoded feature vector. Learning inference results are generated by fusing the features.

[0076] In the context of intelligent decision support for government affairs, the first decision flow is responsible for rule-driven reasoning. Its core lies in organizing and activating explicit knowledge in the government domain using a graph structure. Government business rules (such as administrative approval conditions, compliance judgment standards, and policy constraints) are used as nodes, and the triggering dependencies and logical premise relationships between rules are used as directed edges to construct a rule reasoning graph. This graph structure can fully express the semantic relationships of rule chain transmission in the government domain. For example, complex dependencies such as "rule C can only be triggered if condition A and condition B are met" can be represented by directed edges. After constructing the rule reasoning graph, feature matching is performed on the encoded feature vectors. The matching score between the encoded feature vectors and the feature templates corresponding to each rule node is calculated. Nodes with scores exceeding a preset activation threshold are included in the activated node set as the starting point for subsequent reasoning traversal.

[0077] Starting from the set of activated nodes, a breadth-first traversal is performed along the directed edges of the rule inference graph. This strategy ensures that all reachable nodes at the same inference depth are visited first, avoiding the overexpansion of local paths caused by depth-first strategies. During the traversal, each visited node's preconditions are checked against the current encoded feature vector. Precondition satisfaction is determined by semantically aligning the node's precondition description with the encoded feature vector. If the alignment score reaches a threshold, the node is added to the inference queue; otherwise, it is skipped, and its corresponding inference operation is not triggered. The inference operations corresponding to each node in the inference queue are executed sequentially. Each inference operation adds a new inference conclusion to the existing state, gradually forming intermediate inference states. These intermediate states record all intermediate conclusions derived from the currently activated rule chains, providing a foundation for subsequent consistency verification.

[0078] When performing consistency verification on intermediate inference states, it is necessary to check whether there are logically contradictory inference conclusions within the intermediate states. Contradiction detection is achieved by comparing pairwise pairs of conclusions within the intermediate states. If a contradiction exists... Conclusion If two or more rules are semantically contradictory or their values ​​conflict, the reasoning path is considered to have a logical contradiction. Once a contradiction is detected, a backtracking mechanism is initiated, reversing the execution order of the reasoning operations to cancel the reasoning operation that caused the contradiction, restoring the intermediate reasoning state before the contradiction occurred. At the backtracking point, other feasible reasoning path branches are selected from the reasoning queue to continue reasoning until no logical contradictions exist in the intermediate reasoning state, ultimately yielding a contradiction-free rule reasoning result. This backtracking mechanism effectively ensures the logical self-consistency of the rule reasoning result and prevents conflicting government rule conclusions from appearing simultaneously in the final reasoning output.

[0079] The second decision stream is responsible for data-driven learning. Its reasoning process uses historical decision samples as knowledge sources, projecting encoded feature vectors onto a decision embedding space. This decision embedding space is a pre-trained low-dimensional continuous vector space where semantically similar decision scenarios are also spatially close. The projection operation is performed using a linear transformation matrix. The original encoded feature vector Mapped to decision embedding vector ,Right now ,in Pre-trained on historical decision data, in the decision embedding space, with For the query vector, efficient retrieval of the corresponding query vector using approximate nearest neighbor retrieval methods (such as index structures based on product quantization) is achieved. The nearest historical decision samples in terms of Euclidean distance constitute the neighborhood sample set. Each sample in the neighborhood sample set contains a feature vector of the historical business scenario and its corresponding decision result label, providing a data foundation for subsequent decision pattern extraction.

[0080] When extracting decision patterns from samples in a neighboring sample set, the process of identifying common decision features among samples is achieved by aggregating and statistically analyzing the sample feature vectors. Let the neighboring sample set contain a total of... There are n samples, and their feature vectors are respectively Regarding this The feature vectors are weighted by dimension, and the weights are... From each sample and the query vector The common decision feature vector is obtained by normalizing the inverse distance, thus making samples that are closer in distance contribute more to the common features. Weighted aggregation yields the common decision feature vector. , This collectively reflects recurring decision-making patterns in recent historical decision-making scenarios. It uses common decision-making feature vectors. With the current encoded feature vector Feature fusion is performed using a method that combines concatenation with linear transformation for dimensionality reduction; that is, first... and The concatenated vector is obtained by concatenating along the feature dimensions, and then a fusion transformation matrix is ​​applied. Map the concatenated vector to the fused feature vector. Fusion of feature vectors It also carries real-time semantic information from the current business scenario and common patterns extracted from historical decision-making experience, which are then processed through subsequent learning and inference network layers. Perform nonlinear transformations and classification predictions to generate learning and inference results.

[0081] In practical government applications, rule-driven reasoning and data-driven reasoning represent decision-making judgments from two dimensions: explicit rule knowledge and implicit data experience, respectively. Their synergistic integration can significantly improve the reliability and scope of decision-making. Rule-driven reasoning provides strong interpretability assurance for policy compliance judgments, while data-driven learning can capture complex business models not yet covered by the rule system. The two complement each other, jointly supporting the subsequent adversarial verification mechanism in identifying areas of consistency and conflict, thereby forming a high-quality integrated decision-making space.

[0082] Figure 2 This is a flowchart illustrating the method for generating a fusion decision space based on an adversarial verification mechanism according to an embodiment of the present invention. The adversarial verification mechanism allows the rule-based reasoning results and the learned reasoning results to be cross-validated, identifying areas of consistency and conflict between them, thus obtaining the fusion decision space, including:

[0083] The rule-based reasoning results and the learning-based reasoning results are respectively decomposed into decision features to obtain multiple rule-based decision units and multiple learning-based decision units, and a semantic correspondence between the rule-based decision units and the learning-based decision units is established.

[0084] For each pair of corresponding decision units, semantic similarity is calculated. The vector angle between decision units is measured by the cosine similarity method. Decision unit pairs with semantic similarity higher than the upper similarity bound are marked as consistent decision pairs, and decision unit pairs with semantic similarity lower than the lower similarity bound are marked as conflicting decision pairs.

[0085] The confidence fusion is performed on the consistent decision pair to obtain the rule confidence of the rule decision unit and the learning confidence of the learning decision unit. The feature vectors of the two decision units are linearly combined according to the confidence ratio to obtain a consistent fusion vector set.

[0086] An adversarial evaluation is performed on the conflict decision pair, an evaluation sample set is constructed, and decision prediction is performed using the rule decision unit and the learning decision unit in the conflict decision pair respectively. The matching degree between the prediction result and the real label is calculated, and the decision unit is selected as the effective decision based on the matching degree to obtain the conflict resolution vector set.

[0087] The consistency fusion vector set and the conflict resolution vector set are concatenated and recombined to construct a complete decision feature space as the fusion decision space.

[0088] After obtaining the results of rule-based reasoning and learning-based reasoning, the core of the adversarial verification mechanism lies in decomposing the two reasoning results into fine-grained decision-making units and establishing precise semantic correspondences at the unit level. This allows subsequent verification, fusion, and conflict resolution to be carried out within a controllable local scope, avoiding information loss caused by coarse-grained comparisons at the overall level.

[0089] When decomposing the decision features of the rule-based reasoning results, based on the reasoning chain structure output by the rule engine, the complete reasoning conclusion is broken down into several independent rule-based decision units. Each rule-based decision unit corresponds to a local judgment triggered by a specific government rule clause, carrying the applicable conditions, judgment basis, and output conclusion of that clause. When decomposing the decision features of the learning reasoning results, based on the attention distribution and hidden activation information generated by the data-driven model during the reasoning process, the overall prediction is broken down into several learning-based decision units. Each learning-based decision unit corresponds to the local decision tendency formed by the model on a specific feature dimension. After the decomposition is completed, the rule-based decision units and learning-based decision units are uniformly mapped to the same semantic vector space through semantic embedding methods. Based on the business semantics described by the unit (such as category labels for approval conditions, qualification determination, risk assessment, etc.), a semantic correspondence between the rule-based decision units and learning-based decision units is established, forming a one-to-one or one-to-many correspondence mapping table, which serves as the basis for subsequent unit-level verification.

[0090] For each pair of corresponding decision units, their semantic similarity is calculated to quantify the consistency between the two reasoning approaches in that local decision dimension. Let the semantic vector of the rule-based decision unit be... The semantic vector corresponding to the learning decision unit is The cosine similarity between the two is calculated as follows: This value reflects the cosine of the angle between the vectors of the two decision units in the semantic space, and its range is [value missing]. The closer the value is The more semantically similar the two values ​​are, the closer they are to each other. The more semantically opposed the two terms, the higher the upper bound of their similarity is set. The lower bound of similarity is ,for The decision-making unit pairs are marked as consistent decision pairs, indicating that rule-based reasoning and data-driven reasoning have reached mutually supportive judgments in this local dimension; for Decision-making unit pairs that are similar to each other are marked as conflicting decision-making pairs, indicating that there is a substantial contradiction between the two reasoning approaches in this dimension; for those with similarity between... and The decision-making units between them are considered as fuzzy areas, and a conservative strategy (leaning towards rule-based reasoning) or an aggressive strategy (leaning towards learning-based reasoning) can be selected to handle them according to the risk level of the business scenario. and The specific value can be determined through validation set optimization based on the distribution characteristics of historical government decision-making data. In general government approval scenarios... Can be set to , Can be set to .

[0091] When fusing the execution confidence scores for consistent decision-making, the rule confidence scores of each rule-making unit are obtained separately. Learning confidence with learning decision unit Rule confidence The confidence level is derived from the rule engine's score on the degree to which the conditions for the clause's applicability are met. The predicted probability outputs of the data-driven model on the corresponding feature dimension of this decision unit are derived. These two values ​​are normalized and used as a linear combination weight to obtain the normalized weights. and The feature vector of the rule-based decision unit in the corresponding consensus decision pair With the feature vector of the learning decision unit A linear combination based on weights yields the fusion vector of the consensus decision pair. The above fusion operation is performed sequentially on all consistent decision pairs, and the resulting fusion vectors are aggregated to form a consistent fusion vector set. This set preserves the complementary information of the two inference paths within the consistent region, while suppressing the noise introduced by the lower confidence side through confidence weighting.

[0092] When assessing the adversarial nature of conflict decision-making in execution, an assessment sample set is constructed by retrieving samples from historical databases that are highly correlated with the current government business scenario in terms of business type, application conditions, and processing results. Each sample in this set contains business feature inputs and verified, manually reviewed, real decision labels. The rule-based decision units and learned decision units from the conflict decision pair are used respectively. Each sample in the dataset is used for decision prediction, and the matching degree between each prediction and the true label is calculated. The matching degree is calculated using an accuracy metric, where the matching degree of the rule-based decision unit on the evaluation sample set is denoted as . The matching degree of the learning decision unit is ,Compare and The size of the decision pair is used to select the decision unit with the highest matching degree as the effective decision, and its feature vector is directly used as the resolution vector of the conflict decision pair. .when and When the difference is extremely small (below the preset tolerance threshold), the result of the rule-based decision-making unit is adopted first to ensure the compliance and interpretability of government decisions. The above adversarial assessment is performed sequentially on all conflicting decision pairs, and the resulting resolution vectors are aggregated to form a conflict resolution vector set. .

[0093] Complete the consistent fusion vector set With conflict resolution vector set After construction, the vectors in the two sets are concatenated and recombined according to the original decision-making units' arrangement order in the business semantic dimension. This allows decision features from different sources (consistency regions and conflict resolution regions) to be arranged in an orderly sequence structure, constructing a complete decision feature space as a fusion decision space. This fusion decision space retains the structured constraints of rule-based reasoning in terms of policy compliance while incorporating the adaptive expressive ability of data-driven reasoning to business complexity. It provides a high-quality feature foundation for subsequent Monte Carlo tree search in the fusion decision space, ensuring that the decision path search process can unfold within a decision space that balances compliance and flexibility.

[0094] A decision path search tree is constructed in the fused decision space, with the current business state as the root node. Decision action nodes are expanded using the Monte Carlo tree search method, and policy value and state value are calculated to determine the optimal search path. A sequence of decision actions is extracted along the optimal search path to obtain the target decision scheme, including:

[0095] Construct a decision path search tree and map the current business state to the root node. Assign an initial access count and accumulated value to the root node. Extract a set of decision candidates for the current business state from the fused decision space. Create a child node for each decision candidate in the decision candidate set and connect it to the root node as a decision action node.

[0096] The iterative search process is executed. In each iteration, the confidence upper bound of each decision action node is calculated. The decision action node with the highest confidence upper bound is simulated and expanded. The transition state after executing the corresponding decision action is calculated through the state transition function. The transition state is input into the policy network to obtain the action selection probability distribution as the policy value. The transition state score is input into the value network to obtain the state evaluation score as the state value.

[0097] The strategy value and the state value are linearly combined to obtain a comprehensive score, and the score is updated backtracking to all nodes on the path from the root node to the selected node, updating the cumulative value and access count of each node.

[0098] Repeat the above iterative search process until the search termination condition is met. Select the path with the highest cumulative value starting from the root node as the optimal search path in the decision path search tree. Collect the decision actions of each node along the optimal search path, construct the decision action sequence and output it as the target decision scheme.

[0099] After the fusion decision space is formed, a decision path search tree needs to be constructed within it, with the current business state as the root node for initialization. The current business state is a comprehensive representation of all relevant elements in the government business scenario, covering dimensions such as the type of matter to be processed, urgency, resource constraints, and historical processing records. After mapping this state to the root node, an initial access count and accumulated value are assigned to the root node. The initial access count is set to 0, and the accumulated value is set to 0, indicating that the node has not yet undergone any simulation expansion. Subsequently, a set of decision candidates for the current business state is extracted from the fusion decision space. This set consists of decision units in the fusion decision space that are semantically related to the current business state. Each decision candidate represents a feasible government processing action, such as approval, rejection of supplementary documents, transfer to higher authorities, or initiation of joint review. A corresponding child node is created for each candidate in the decision candidate set, and these child nodes are connected to the root node as decision action nodes, forming the first-level expansion structure of the search tree.

[0100] In the iterative search process, each iteration starts from the root node and selects the most worthwhile decision-making node to explore based on the upper confidence bound. The calculation of the upper confidence bound comprehensively considers both the average cumulative value of the node and the exploration reward. Let the node be... The current cumulative value is The number of visits is The total number of visits to its parent node is The policy network outputs the prior action probability for this node as follows: The exploration coefficient is Then the node upper confidence bound The calculation is as follows:

[0101]

[0102] in The first item reflects the historical average return of a node, while the second item is the exploration reward, which is the number of times a node is visited. Less exploration yields greater rewards, encouraging the search process to fully utilize known high-value paths while maintaining attention to underexplored areas. Select the upper confidence bound. After the highest node, perform a simulated unfolding operation on it.

[0103] During the simulation deployment phase, a state transition function is used to calculate the transition state reached after executing the corresponding decision action. The state transition function takes the business state represented by the current node and the corresponding decision action as input, and outputs the new state the government business process enters after executing the action. The new state reflects the actual impact of the decision action on the business process; for example, after executing an approval action, the business state transitions from "pending approval" to "approval completed, awaiting archiving." The transition states are then input into the policy network and value network for evaluation. The policy network takes the feature representation of the transition state as input and outputs the probability distribution of each possible action in that state. This distribution serves as the policy value, reflecting the relative priority of each action in the current state. The value network takes the same transition state feature representation as input and outputs a scalar evaluation score, which serves as the state value, representing the degree of contribution of the current state to the final decision objective. Both the policy network and the value network are pre-trained based on historical government decision data and their parameters are continuously updated in subsequent execution feedback.

[0104] The strategy value and state value are linearly combined to obtain a comprehensive score. Let the aggregate scalar representation of the strategy value be... (Obtained by weighted summation of the action probability distributions output by the policy network), the scalar evaluation score of the state value is: The linear combination weights are respectively and The overall score is then calculated as follows:

[0105]

[0106] in and satisfy The values ​​of both are configured according to the characteristics of government business scenarios, and should be appropriately increased in scenarios with strong rule constraints. Strengthen strategy guidance and appropriately increase [the amount of data] in scenarios with sufficient historical data. To strengthen the leading role of value assessment.

[0107] Overall Score After the calculation is complete, a backtracking update operation is performed. Starting from the currently selected node, the process traces backwards along the path to the root node, updating the cumulative value and visit count of each node on the path. Let the cumulative value of a node on the path before backtracking be... The number of visits is The updated number of visits Updated cumulative value Calculated using the incremental averaging method:

[0108]

[0109] This update method ensures that the cumulative value always reflects the average comprehensive score of all simulations conducted throughout the node's history, avoiding an excessive impact of a single high-scoring or low-scoring simulation result on the node's value assessment.

[0110] The iterative search process described above continues until the search termination condition is met. The termination condition can be a preset maximum number of iterations, a maximum search time limit, or a stability condition where the optimal path remains unchanged after a certain number of iterations. In government service scenarios, considering the timeliness requirements of decision-making, a joint constraint of the maximum number of iterations and the maximum search time is typically used as the termination condition to ensure that the decision result is output within a limited time.

[0111] After the search terminates, the path with the highest accumulated value starting from the root node is selected as the optimal search path in the decision path search tree. Specifically, starting from the root node, the path with the highest accumulated value is selected at each level. The highest-ranking child node is selected greedily layer by layer until a leaf node is reached or the path termination condition is met. The resulting complete node sequence is the optimal search path. The decision actions corresponding to each node are collected along the optimal search path and arranged in order from the root node to the leaf node to construct a decision action sequence. This sequence fully describes the complete action chain from the current business state, through a series of intermediate states, to finally achieve the decision goal. This decision action sequence is output as the target decision scheme. In addition to the action sequence itself, the target decision scheme also includes the confidence upper bound, comprehensive score, and corresponding transition state description for each action node, providing a complete decision basis for subsequent decision execution and audit traceability.

[0112] The decision action node with the highest confidence upper bound is simulated and expanded. The transition state after executing the corresponding decision action is calculated through the state transition function, and then input into the policy network to obtain the action selection probability distribution as the policy value. The state evaluation score is then input into the value network to obtain the state value, including:

[0113] Obtain the decision action identifier and current business state vector corresponding to the decision action node with the highest confidence upper bound value, and convert the decision action identifier into an action embedding vector;

[0114] The current business state vector is concatenated with the action embedding vector to form a joint state input. The joint state input is then subjected to nonlinear transformation and state evolution calculation through a state transition function to output the state vector representation of the transition state.

[0115] The state vector of the transition state is input into the encoding layer of the policy network for feature extraction. The state features are abstracted and represented by a multi-layer neural network structure. After processing by a fully connected layer and activation function, the output is sent to the action space mapping layer. In the action space mapping layer, the feature vector is mapped into a score vector for each candidate action. The score vector is normalized to obtain the action selection probability distribution as the policy value.

[0116] The state vector representation of the transition state is input into the value network for feature encoding and value prediction, and the output state evaluation score is used as the state value.

[0117] During the Monte Carlo tree search process, when simulating and expanding the decision action node with the highest confidence upper bound, it is necessary to first locate the decision action identifier and the current business state vector corresponding to that node. The decision action identifier is a discretized code for executable operations in the government affairs business scenario. For example, operations such as "approval passed," "return for supplementary materials," and "transfer to superior department" each correspond to a unique action code symbol. The current business state vector is a continuous vector representation of all key elements in the current government affairs business scenario, including numerical codes for dimensions such as the completeness of application materials, historical approval records, and business process progress.

[0118] The process of converting decision action identifiers into action embedding vectors is accomplished through an action embedding matrix. The action embedding matrix is ​​jointly optimized with the overall network during the training phase, ensuring that semantically similar actions are close together in the embedding space, while semantically different actions are spaced further apart. Let the action embedding vector corresponding to the decision action identifier be... The current business state vector is Then and By concatenating along the feature dimensions, the joint state input is obtained. ,in This represents the vector concatenation operation. The joint state input integrates the current state information and the action information to be executed into the same representation space, providing a complete input basis for subsequent state transition calculations.

[0119] State transition function for joint state input Perform nonlinear transformation and state evolution calculations, and output the state vector representation of the transition states. The state transition function consists of a multi-layer fully connected network, with each layer containing linear transformations and nonlinear activation operations. The activation function employs ReLU or GELU to capture the complex evolution of the business state after performing a specific action. The calculation of the state transition function can be expressed as follows: ,in Representing the overall mapping relationship of the state transition network, in government business scenarios, the state transition function needs to model state changes with clear semantic logic, such as "after the supplementary materials are submitted, the approval status changes from pending supplementation to pending review". Therefore, when constructing the training data, a large number of real business process flow records need to be introduced to ensure the accuracy of the state transition function.

[0120] Obtain the state vector representation of the transition state. Then, it is input into the policy network to calculate the action selection probability distribution. The encoding layer of the policy network needs to... Feature extraction is performed by abstracting and representing state features layer by layer through a multi-layer neural network structure. Each layer of the network performs nonlinear transformation on the input features, gradually extracting the high-level semantic features most relevant to the decision action selection. The feature vector output by the encoding layer is then processed by several fully connected layers and activation functions to further compress and integrate the feature information, and finally sent to the action space mapping layer.

[0121] The action space mapping layer maps the encoded feature vectors into score vectors for each candidate action. ,in Each dimension corresponds to the original score of a candidate decision action. The candidate action set consists of legal and executable actions in the current business scenario. Illegal actions are masked to negative infinity during the mapping phase, thus being automatically excluded during subsequent normalization. For the score vector... During normalization, the softmax function is used to convert the original scores into a probability distribution, resulting in the action selection probability distribution. ,in Each dimension represents the probability of the corresponding candidate action being selected, and the sum of all dimensions is 1. Action selection probability distribution. As a strategy value output, it reflects the priority ranking of each decision action in the current transition state, providing a priori probability basis for node expansion in Monte Carlo tree search.

[0122] The value network and policy network share some of the underlying feature encoding structures, but use an independent value prediction head at the top level to represent the state vectors of the transition states. After being input into the value network, the value-related features of the state are extracted through the feature encoding layer, and then the state evaluation score is output through the value prediction head. ,in Let be a scalar, representing the expected long-term cumulative return at this transition state. The output range of the value network is constrained by the tanh activation function. Within the range, a positive value indicates that the state makes a positive contribution to achieving government business objectives, while a negative value indicates that the state carries the risk of deviating from the expected objectives. State Assessment Score As a state value output, it participates in the comprehensive score calculation in Monte Carlo tree search together with the policy value, providing a quantitative basis for determining the optimal search path.

[0123] In real-world government service scenarios, both the policy network and the value network require thorough pre-training and online fine-tuning. During pre-training, supervised learning samples are constructed using historical government service data. Expert decision sequences serve as the training labels for the policy network, while the final business outcome (such as approval rate and processing timeliness) is used as the regression target for the value network. In the online fine-tuning phase, self-play data accumulated during Monte Carlo tree search is used to jointly optimize the parameters of both networks through a policy gradient method and a value regression loss. This ensures that the action selection probability distribution output by the policy network aligns with the distribution of nodes with high access frequency during the search process, and that the state evaluation score output by the value network closely matches the actual revenue of the final business outcome.

[0124] To improve the computational efficiency of simulation deployment, when processing multiple decision action nodes in batches, the joint state inputs of multiple nodes can be organized into a batch matrix. GPU parallel computing can be used to accelerate the forward inference process of state transition functions, policy networks, and value networks. In addition, for frequently occurring business states, a state caching mechanism can be established to store the calculated transition states and their corresponding policy and state values ​​in the cache, avoiding repeated calculations and further reducing the response latency of online decision-making, thus meeting the real-time requirements of government service scenarios.

[0125] A second aspect of this invention provides an intelligent decision support system for digital applications of a government cloud platform, comprising:

[0126] The data encoding unit is used to acquire multi-source heterogeneous data in government business scenarios, including structured business data, unstructured document data and real-time interactive data, and to perform semantic parsing and feature encoding on the multi-source heterogeneous data to obtain encoded feature vectors.

[0127] A dual-stream decision unit is used to construct a dual-stream decision network architecture. The first decision stream performs rule-driven reasoning on the encoded feature vector to obtain rule-driven reasoning results, and the second decision stream performs data-driven learning on the encoded feature vector to obtain learning reasoning results. An adversarial verification mechanism is used to cross-validate the rule-driven reasoning results and the learning reasoning results, identify the consistency and conflict regions between the two, and obtain a fusion decision space.

[0128] The decision search unit is used to construct a decision path search tree in the fused decision space, take the current business state as the root node, expand the decision action nodes through the Monte Carlo tree search method, calculate the strategy value and state value to determine the optimal search path, extract the decision action sequence along the optimal search path, and obtain the target decision scheme.

[0129] The execution update unit is used to generate decision instructions based on the target decision scheme and send them to the government application module for execution, collect business feedback data after execution, and update the parameters of the Monte Carlo tree search method.

[0130] A third aspect of the present invention provides an electronic device, comprising:

[0131] processor;

[0132] Memory used to store processor-executable instructions;

[0133] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0134] 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.

[0135] 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.

[0136] 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. A method for intelligent decision support for digital applications of a government cloud platform, characterized in that, include: Acquire multi-source heterogeneous data in government business scenarios, including structured business data, unstructured document data, and real-time interactive data; perform semantic parsing and feature encoding on the multi-source heterogeneous data to obtain encoded feature vectors. A dual-stream decision network architecture is constructed. The first decision stream performs rule-driven reasoning on the encoded feature vector to obtain rule-based reasoning results, while the second decision stream performs data-driven learning on the encoded feature vector to obtain learned reasoning results. An adversarial verification mechanism is used to cross-validate the rule-based reasoning results and the learned reasoning results, identifying the consistency and conflict regions between them, and obtaining a fused decision space. In the fusion decision space, a decision path search tree is constructed, with the current business state as the root node. The decision action nodes are expanded using the Monte Carlo tree search method, and the strategy value and state value are calculated to determine the optimal search path. The decision action sequence is extracted along the optimal search path to obtain the target decision scheme. Decision instructions are generated based on the target decision scheme and sent to the government application module for execution. Business feedback data after execution is collected, and the parameters of the Monte Carlo tree search method are updated.

2. The method of claim 1, wherein, Semantic parsing and feature encoding are performed on the multi-source heterogeneous data to obtain encoded feature vectors, including: Perform field semantic analysis on the structured business data in the multi-source heterogeneous data, identify the business attribute types and value constraint rules of the fields, and label each field with semantic tags and business meanings according to the field name and field content to obtain a structured semantic representation; Unstructured document data is analyzed for document structure, and the heading levels, paragraph boundaries and chapter topics are identified. Each paragraph is mapped to a topic vector and cosine similarity is calculated. The paragraphs are clustered based on the cosine similarity to obtain the document semantic clustering results. Multi-turn dialogue modeling is performed on real-time interactive data to extract the information of dereference resolution and omission completion in the dialogue. The pronouns are replaced with the entities they point to, and the omitted components are completed into complete semantic units to form a dereferenced complete dialogue semantic sequence, thus obtaining the interactive semantic sequence. The structured semantic representation, the document semantic clustering result, and the interaction semantic sequence are respectively embedded into vectors. The embedded vectors are then propagated using a neighborhood information aggregation method. The propagated vectors are then concatenated and normalized to obtain the encoded feature vector.

3. The method of claim 1, wherein, The first decision stream performs rule-driven inference on the encoded feature vector to obtain a rule-driven inference result, and the second decision stream performs data-driven learning on the encoded feature vector to obtain a learning inference result, including: The government business rules are used as nodes, the dependencies between rules are used as directed edges, and a rule reasoning graph is constructed. Feature matching is performed on the encoded feature vector to activate the initial nodes in the rule reasoning graph, thus obtaining a set of activated nodes. Starting from the set of activated nodes, a breadth-first traversal is performed along the directed edges of the rule reasoning graph. During the traversal, it is determined whether the preconditions of each visited node are satisfied by the encoded feature vector. Nodes that satisfy the preconditions are added to the reasoning queue, and the reasoning operations corresponding to the nodes in the reasoning queue are executed in sequence to obtain the intermediate reasoning state. The intermediate inference states are subjected to consistency verification to detect whether there are logically contradictory inference results. The inference operations that cause contradictions are undone through a backtracking mechanism, and a feasible inference path is selected again until a contradiction-free rule inference result is obtained. The encoded feature vector is projected onto the decision embedding space, and the historical decision sample that is closest to the encoded feature vector is retrieved in the decision embedding space to obtain the neighbor sample set. Decision patterns are extracted from the samples in the neighboring sample set, common decision features among the samples are identified, and feature fusion is performed based on the common decision features and the encoded feature vector. Learning inference results are generated by fusing the features.

4. The method of claim 1, wherein, An adversarial verification mechanism is used to cross-validate the rule-based reasoning results with the learned reasoning results, identifying areas of consistency and conflict between the two, thus obtaining a fused decision space, including: The rule-based reasoning results and the learning-based reasoning results are respectively decomposed into decision features to obtain multiple rule-based decision units and multiple learning-based decision units, and a semantic correspondence between the rule-based decision units and the learning-based decision units is established. For each pair of corresponding decision units, semantic similarity is calculated. The vector angle between decision units is measured by the cosine similarity method. Decision unit pairs with semantic similarity higher than the upper similarity bound are marked as consistent decision pairs, and decision unit pairs with semantic similarity lower than the lower similarity bound are marked as conflicting decision pairs. The confidence fusion is performed on the consistent decision pair to obtain the rule confidence of the rule decision unit and the learning confidence of the learning decision unit. The feature vectors of the two decision units are linearly combined according to the confidence ratio to obtain a consistent fusion vector set. An adversarial evaluation is performed on the conflict decision pair, an evaluation sample set is constructed, and decision prediction is performed using the rule decision unit and the learning decision unit in the conflict decision pair respectively. The matching degree between the prediction result and the real label is calculated, and the decision unit is selected as the effective decision based on the matching degree to obtain the conflict resolution vector set. The consistency fusion vector set and the conflict resolution vector set are spliced ​​and recombined to construct a complete decision feature space as the fusion decision space.

5. The method of claim 1, wherein, A decision path search tree is constructed in the fused decision space, with the current business state as the root node. Decision action nodes are expanded using the Monte Carlo tree search method, and policy value and state value are calculated to determine the optimal search path. A sequence of decision actions is extracted along the optimal search path to obtain the target decision scheme, including: Construct a decision path search tree and map the current business state to the root node. Assign an initial access count and accumulated value to the root node. Extract a set of decision candidates for the current business state from the fused decision space. Create a child node for each decision candidate in the decision candidate set and connect it to the root node as a decision action node. The iterative search process is executed. In each iteration, the confidence upper bound of each decision action node is calculated. The decision action node with the highest confidence upper bound is simulated and expanded. The transition state after executing the corresponding decision action is calculated through the state transition function. The transition state is input into the policy network to obtain the action selection probability distribution as the policy value. The transition state score is input into the value network to obtain the state evaluation score as the state value. The strategy value and the state value are linearly combined to obtain a comprehensive score, and the score is updated backtracking to all nodes on the path from the root node to the selected node, updating the cumulative value and access count of each node. Repeat the above iterative search process until the search termination condition is met. Select the path with the highest cumulative value starting from the root node as the optimal search path in the decision path search tree. Collect the decision actions of each node along the optimal search path, construct the decision action sequence and output it as the target decision scheme.

6. The method according to claim 5, characterized in that, The decision action node with the highest confidence upper bound is simulated and expanded. The transition state after executing the corresponding decision action is calculated through the state transition function, and then input into the policy network to obtain the action selection probability distribution as the policy value. The state evaluation score is then input into the value network to obtain the state value, including: Obtain the decision action identifier and current business state vector corresponding to the decision action node with the highest confidence upper bound value, and convert the decision action identifier into an action embedding vector; The current business state vector is concatenated with the action embedding vector to form a joint state input. The joint state input is then subjected to nonlinear transformation and state evolution calculation through a state transition function to output the state vector representation of the transition state. The state vector of the transition state is input into the encoding layer of the policy network for feature extraction. The state features are abstracted and represented by a multi-layer neural network structure. After processing by a fully connected layer and activation function, the output is sent to the action space mapping layer. In the action space mapping layer, the feature vector is mapped into a score vector for each candidate action. The score vector is normalized to obtain the action selection probability distribution as the policy value. The state vector representation of the transition state is input into the value network for feature encoding and value prediction, and the output state evaluation score is used as the state value.

7. An intelligent decision support system for digital applications of government cloud platforms, used to implement the method as described in any one of claims 1-6, characterized in that, include: The data encoding unit is used to acquire multi-source heterogeneous data in government business scenarios, including structured business data, unstructured document data and real-time interactive data, and to perform semantic parsing and feature encoding on the multi-source heterogeneous data to obtain encoded feature vectors. A dual-stream decision unit is used to construct a dual-stream decision network architecture. The first decision stream performs rule-driven reasoning on the encoded feature vector to obtain rule-driven reasoning results, and the second decision stream performs data-driven learning on the encoded feature vector to obtain learning reasoning results. An adversarial verification mechanism is used to cross-validate the rule-driven reasoning results and the learning reasoning results, identify the consistency and conflict regions between the two, and obtain a fusion decision space. The decision search unit is used to construct a decision path search tree in the fused decision space, take the current business state as the root node, expand the decision action nodes through the Monte Carlo tree search method, calculate the strategy value and state value to determine the optimal search path, extract the decision action sequence along the optimal search path, and obtain the target decision scheme. The execution update unit is used to generate decision instructions based on the target decision scheme and send them to the government application module for execution, collect business feedback data after execution, and update the parameters of the Monte Carlo tree search method.

8. 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 6.

9. 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 6.

Images