Large model content generation auditing method and system based on man-machine cooperation

Abstract

The invention provides a large model content generation auditing method and system based on man-machine collaboration, and relates to the technical field of large models, and the method comprises the steps: deconstructing the generation process data of to-be-audited content, constructing a content risk propagation map, mapping a multi-level auditing strategy library, and recognizing an uncertainty region needing manual intervention. And obtaining manual auditing judgment and correction operation, forming auditing knowledge precipitation and perfecting an auditing strategy library. According to the method, the content risk source can be accurately positioned, the auditing efficiency is improved, a closed-loop optimized man-machine collaborative auditing mechanism is formed, the misjudgment rate is effectively reduced, and the large model content security is improved.

Description

A Human-Machine Collaborative Large-Model Content Generation and Review Method and System Technical Field

[0001] This invention relates to large model technology, and more particularly to a method and system for generating and reviewing large model content based on human-computer collaboration. Background Technology

[0002] With the rapid development of artificial intelligence technology, large language models have been widely applied in the field of content generation, and their powerful text generation capabilities have brought convenience to various industries. However, content generated by large models may contain security risks, biases, and inappropriate information. Therefore, content moderation has become a crucial link in ensuring the quality and compliance of generated content. Traditional content moderation mainly relies on keyword filtering, rule matching, and simple machine learning models. However, with the increasing complexity of large language models and their content generation capabilities, these methods are no longer sufficient to meet the needs of moderation. Currently, the industry has begun to explore a hybrid moderation model that combines human moderation with automated systems. By leveraging the complementary advantages of human judgment and the efficient processing capabilities of machines, the accuracy and efficiency of content moderation can be improved.

[0003] Existing review systems lack a deep understanding of the internal generation process of large language models, making it impossible to effectively track the propagation path of semantic risks. This results in review results lacking interpretability and making it difficult to pinpoint the root cause of risky content. Reviewers can only see the final generated text and cannot understand how the model constructs this content step by step, making the review process a black box operation.

[0004] Existing human-machine collaborative review methods lack a refined task allocation mechanism, making it impossible to dynamically adjust the scope and depth of human intervention based on the complexity and uncertainty of content risks. Review task allocation often adopts a simplistic division method, failing to fully consider the impact and propagation effects of decision-making nodes, resulting in unreasonable human resource allocation, insufficient review of high-risk areas and excessive review of low-risk areas. Summary of the Invention

[0005] This invention provides a method and system for generating and reviewing large model content based on human-computer collaboration, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides a method for reviewing large-scale model content generation based on human-computer collaboration, comprising:

[0007] Obtain the content to be reviewed generated by the large language model, deconstruct the generation process data corresponding to the content to be reviewed, the generation process data includes attention distribution and semantic decoding path; construct a content risk propagation graph based on the attention distribution and the semantic decoding path, the content risk propagation graph represents the semantic dependency relationship and the associated propagation path of potential risk nodes in the generation process, and mark decision nodes in the content risk propagation graph;

[0008] The content risk propagation map is mapped and associated with a multi-level review strategy library to identify uncertain areas that require human intervention. Based on the influence range of the decision node in the associated propagation path, the priority and granularity of the review tasks are divided to form a review task allocation scheme.

[0009] According to the aforementioned task allocation scheme, the review tasks are imported into the manual review interface to obtain the review judgments and semantic correction operations returned by the manual reviewers for the uncertain areas; the content to be reviewed is reshaped based on the review judgments and semantic correction operations, and the manual decision-making logic in the semantic correction operations is transformed into rule-based constraints. The source node that caused the risk is located through the reverse tracing of the content risk propagation graph, thus forming a review knowledge accumulation.

[0010] The aforementioned review knowledge is integrated into the matching rules of the multi-level review strategy library, and the construction parameters of the content risk propagation graph are improved.

[0011] A content risk propagation graph is constructed based on the attention distribution and the semantic decoding path. This graph represents the semantic dependencies and associated propagation paths of potential risk nodes during the generation process. Decision nodes are marked in the content risk propagation graph, including:

[0012] The association weights between semantic units in the attention distribution are analyzed, and a directed semantic dependency network is constructed based on the association weights. The generation sequence of the generation steps in the semantic decoding path is combined with the generation time sequence identifier in each node of the directed semantic dependency network, and the causal chain of semantic evolution is determined by the generation time sequence identifier and the directionality of the edges in the directed semantic dependency network.

[0013] Traverse the nodes and their associated edges in the directed semantic dependency network, identify semantic consistency deviation regions, mark the nodes in the semantic consistency deviation regions as potential risk nodes, and construct a risk impact diffusion tree along the edge propagation direction of the directed semantic dependency network, using the potential risk nodes as the propagation source points.

[0014] A traversal analysis is performed on all nodes in the directed semantic dependency network to locate nodes with multiple outgoing edges. The semantic direction difference of each outgoing edge is calculated, and nodes whose semantic direction difference exceeds a preset difference threshold and their outgoing edges are marked as decision nodes.

[0015] Based on the risk impact diffusion tree and the decision node, the propagation path of the potential risk node in the risk impact diffusion tree is associated with the outgoing edges of the decision node to identify the decision node passed through in the risk propagation path.

[0016] The content risk propagation map is mapped and associated with a multi-level review strategy library to identify areas of uncertainty requiring manual intervention. Based on the influence range of the decision-making node in the associated propagation path, the priority and granularity of review tasks are divided to form a review task allocation scheme, including:

[0017] Explore potential risk nodes and their corresponding associated propagation paths in the content risk propagation graph, and simultaneously guide a set of review strategies that match the risk types of the potential risk nodes from the multi-level review strategy library. Through dynamic association, deeply match the strategy rules in the review strategy set with the topology of the associated propagation path.

[0018] The inherent characteristics of the deep matching results are analyzed, the path segments in the association propagation path where the policy rule coverage is missing are identified, the path segments are defined as uncertainty regions, the decision nodes contained in the uncertainty regions are characterized, and the degree of downstream influence diffusion of each decision node in the association propagation path is evaluated.

[0019] Based on the degree of downstream impact diffusion, a hierarchical structure of influence of decision-making nodes within the uncertainty area is shaped, and a priority system for review tasks is derived based on the hierarchical structure of influence.

[0020] The granularity of the review task is constructed based on the topological boundary range of the uncertain region. The granularity of the review task is three-dimensionally characterized by the node distribution density and semantic level span within the uncertain region. The priority system of the review task and the granularity of the review task are integrated into the review task allocation scheme.

[0021] The path segment is defined as an uncertainty region, the decision nodes contained within the uncertainty region are characterized, and the degree of downstream influence diffusion of each decision node in the associated propagation path is evaluated, including:

[0022] In the associated propagation path of the content risk propagation map, the associated propagation path is segmented according to the semantic relevance and content coherence of the path nodes. The path segments with semantic relevance lower than the preset relevance threshold and broken content coherence are defined as uncertain regions, and the semantic boundary contour of the uncertain regions is constructed.

[0023] Based on the semantic boundary contour, semantic association analysis is performed on the nodes in the uncertainty region. Nodes with semantic divergence higher than a preset divergence threshold and multiple semantic associations are marked as decision nodes. The influence chain of decision nodes is constructed by the semantic dependency strength between the decision nodes.

[0024] Along the propagation direction of the influence chain of the decision node, the coverage range of downstream nodes of each decision node in the associated propagation path is statistically analyzed, and the coverage range of downstream nodes is quantified as the degree of downstream influence diffusion of the decision node.

[0025] Based on the review judgment and the semantic correction operation, the content to be reviewed is reshaped, and the manual decision-making logic in the semantic correction operation is transformed into rule-based constraints. The source node that caused the risk is located through the reverse tracing of the content risk propagation graph, forming a review knowledge accumulation including:

[0026] Based on the results of the review and judgment, the sequence of adjustment instructions in the semantic correction operation is explored, a directional association mapping graph between the sequence of adjustment instructions and the semantic structure of the content to be reviewed is constructed, and the content to be reviewed is semantically reconstructed according to the hierarchical transmission relationship in the association mapping graph to form a semantic hierarchical organization with a clear correction trajectory.

[0027] A deep analysis is conducted around the semantic hierarchy organization to identify the judgment criteria and correction evolution path contained in the semantic correction operation. The judgment criteria and correction evolution path are cross-fused to construct a rule-based constraint chain. At the same time, the semantic nodes that trigger the rule-based constraint chain are depicted in the semantic hierarchy organization.

[0028] By leveraging the spatial location characteristics of the semantic nodes in the content risk propagation graph, reverse tracing is performed along the transmission path of the semantic dependency network to locate the source node group that triggers the risk. The rule-based constraint chain is then linked and integrated with the semantic feature spectrum of the source node group to form a systematic accumulation of review knowledge.

[0029] A deep analysis is conducted around the semantic hierarchy to identify the judgment criteria and correction evolution paths inherent in the semantic correction operation. The judgment criteria and correction evolution paths are then cross-integrated to construct a rule-based constraint chain, including:

[0030] The analysis focuses on the structural context of the semantic hierarchy, decomposing the semantic hierarchy into semantic association subgraphs, tracing the semantic transmission of nodes in the semantic association subgraphs, and characterizing the transmission links and affected areas activated by each semantic correction operation in the semantic association subgraphs.

[0031] At the transmission node of the transmission link, the judgment criteria included in the semantic correction operation are identified, the judgment threshold and judgment dimension in the judgment criteria are extracted, the judgment threshold and the judgment dimension are combined into a semantic evaluation vector, and the evolution trajectory of semantic correction is drawn in the influence area, and the evolution trajectory is transformed into a correction evolution path.

[0032] Based on the interaction between the semantic evaluation vector and the modified evolution path, the constraint triggering conditions and the constraint scope are marked, and a regularized constraint chain is formed through the mapping association between the constraint triggering conditions and the constraint scope.

[0033] The matching rules for integrating the aforementioned review knowledge into the multi-level review strategy library, and the construction parameters for improving the content risk propagation graph, include:

[0034] Deconstruct the rule constraint structure in the knowledge accumulation of the audit, identify the judgment conditions and correction strategies in the rule constraint structure, map the judgment conditions and correction strategies to the matching rules of the corresponding level in the multi-level audit strategy library, and construct a fusion and association network of new and old rules;

[0035] Based on the fusion association network, the supplementary coverage rate and rule conflict rate of the newly added rule constraints on the original matching rules are calculated. Rule constraints with a supplementary coverage rate exceeding a preset coverage threshold and a rule conflict rate lower than a preset conflict threshold are selected and integrated into the matching rule system of the multi-level review strategy library.

[0036] Based on the rule coverage in the fusion and association network, the structural features of the content risk propagation graph are analyzed, the weight distribution deviation on the risk propagation path is extracted, and the construction parameters of the content risk propagation graph are dynamically adjusted according to the weight distribution deviation.

[0037] A second aspect of the present invention provides a large-scale model content generation and review system based on human-machine collaboration, comprising:

[0038] The first module is used to acquire the content to be reviewed generated by the large language model, deconstruct the generation process data corresponding to the content to be reviewed, the generation process data includes attention distribution and semantic decoding path; construct a content risk propagation graph based on the attention distribution and the semantic decoding path, the content risk propagation graph represents the semantic dependency relationship and the associated propagation path of potential risk nodes in the generation process, and mark decision nodes in the content risk propagation graph;

[0039] The second module is used to map and associate the content risk propagation map with the multi-level review strategy library, identify uncertain areas that require manual intervention, and divide the priority and granularity of review tasks according to the influence range of the decision node in the associated propagation path to form a review task allocation scheme.

[0040] The third module is used to import the review tasks into the manual review interface according to the review task allocation scheme, obtain the review judgment and semantic correction operation returned by the manual reviewer for the uncertain area; reshape the content to be reviewed according to the review judgment and the semantic correction operation, transform the manual decision-making logic in the semantic correction operation into rule-based constraints, and locate the source node that caused the risk through the reverse tracing of the content risk propagation graph, forming a review knowledge accumulation.

[0041] The fourth module is used to integrate the audit knowledge into the matching rules of the multi-level audit strategy library and improve the construction parameters of the content risk propagation graph.

[0042] A third aspect of the present invention provides an electronic device, comprising:

[0043] processor;

[0044] Memory used to store processor-executable instructions;

[0045] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0046] 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.

[0047] The beneficial effects of this application are as follows:

[0048] By deconstructing the attention distribution and semantic decoding path of the content to be reviewed generated by the large language model, a content risk propagation map can be constructed. This map can intuitively present the semantic dependencies and the associated propagation paths of potential risk nodes in the generation process, enabling the review system to trace the formation mechanism of risky content from the source, rather than just focusing on the identification of surface content.

[0049] By prioritizing and granularizing review tasks based on the influence of decision nodes in the related propagation path, the review resources were rationally allocated, avoiding indiscriminate review of all content and significantly improving review efficiency. At the same time, by accurately locating uncertain areas, manual review resources were concentrated on the parts that truly needed judgment, thus improving review accuracy.

[0050] The decision-making logic of human reviewers in review and semantic correction operations is transformed into rule-based constraints. By tracing back to the source node that caused the risk through the content risk propagation map, the review knowledge is accumulated, enabling the system to continuously learn and optimize review standards, and reduce the need for manual intervention for similar issues in the future. Attached Figure Description

[0051] Figure 1 is a flowchart illustrating the large model content generation and review method based on human-computer collaboration according to an embodiment of the present invention;

[0052] Figure 2 is a flowchart of semantic knowledge accumulation and rule constraint construction analysis and processing according to an embodiment of the present invention. Detailed Implementation

[0053] 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.

[0054] 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.

[0055] Figure 1 is a flowchart illustrating the large-model content generation and review method based on human-machine collaboration according to an embodiment of the present invention. As shown in Figure 1, the method includes:

[0056] Obtain the content to be reviewed generated by the large language model, deconstruct the generation process data corresponding to the content to be reviewed, the generation process data includes attention distribution and semantic decoding path; construct a content risk propagation graph based on the attention distribution and the semantic decoding path, the content risk propagation graph represents the semantic dependency relationship and the associated propagation path of potential risk nodes in the generation process, and mark decision nodes in the content risk propagation graph;

[0057] The content risk propagation map is mapped and associated with a multi-level review strategy library to identify uncertain areas that require human intervention. Based on the influence range of the decision node in the associated propagation path, the priority and granularity of the review tasks are divided to form a review task allocation scheme.

[0058] According to the aforementioned task allocation scheme, the review tasks are imported into the manual review interface to obtain the review judgments and semantic correction operations returned by the manual reviewers for the uncertain areas; the content to be reviewed is reshaped based on the review judgments and semantic correction operations, and the manual decision-making logic in the semantic correction operations is transformed into rule-based constraints. The source node that caused the risk is located through the reverse tracing of the content risk propagation graph, thus forming a review knowledge accumulation.

[0059] The aforementioned review knowledge is integrated into the matching rules of the multi-level review strategy library, and the construction parameters of the content risk propagation graph are improved.

[0060] In one optional implementation, a content risk propagation graph is constructed based on the attention distribution and the semantic decoding path. The content risk propagation graph represents the semantic dependencies and associated propagation paths of potential risk nodes during the generation process. Decision nodes are marked in the content risk propagation graph, including:

[0061] The association weights between semantic units in the attention distribution are analyzed, and a directed semantic dependency network is constructed based on the association weights. The generation sequence of the generation steps in the semantic decoding path is combined with the generation time sequence identifier in each node of the directed semantic dependency network, and the causal chain of semantic evolution is determined by the generation time sequence identifier and the directionality of the edges in the directed semantic dependency network.

[0062] Traverse the nodes and their associated edges in the directed semantic dependency network, identify semantic consistency deviation regions, mark the nodes in the semantic consistency deviation regions as potential risk nodes, and construct a risk impact diffusion tree along the edge propagation direction of the directed semantic dependency network, using the potential risk nodes as the propagation source points.

[0063] A traversal analysis is performed on all nodes in the directed semantic dependency network to locate nodes with multiple outgoing edges. The semantic direction difference of each outgoing edge is calculated, and nodes whose semantic direction difference exceeds a preset difference threshold and their outgoing edges are marked as decision nodes.

[0064] Based on the risk impact diffusion tree and the decision node, the propagation path of the potential risk node in the risk impact diffusion tree is associated with the outgoing edges of the decision node to identify the decision node passed through in the risk propagation path.

[0065] A content risk propagation graph is constructed based on attention distribution and semantic decoding path. This graph represents the propagation path of semantic dependencies and potential risk nodes in the content generation process, and the decision nodes are marked.

[0066] When constructing a content risk propagation map, analyze the association weights between semantic units in the attention distribution, and construct a directed semantic dependence network based on these association weights. Specifically, extract the weight matrix of each attention head in the multi-head attention mechanism, and calculate the association strength between any two semantic units i and j. If the weight value exceeds a preset threshold (such as 0.6), then establish a directed edge from node i to node j in the dependence network, and the weight of the edge is the association strength value. In this way, a preliminary directed semantic dependence network is formed.

[0067] Combined with the generation step sequence in the semantic decoding path, mark the generation time sequence identifier for each node in the directed semantic dependence network. By analyzing the output sequence of the model decoder, record the timestamp when each semantic unit is generated, and add this timestamp as an attribute of the node to the directed semantic dependence network. Through the generation time sequence identifier and the directionality of the edges in the directed semantic dependence network, determine the causal chain of semantic evolution. For example, if the time sequence identifier of node A is t1, the time sequence identifier of node B is t2, and t1 < t2, and there is a directed edge from A to B at the same time, it can be determined that A has a causal impact on B, constituting a link in semantic evolution.

[0068] After completing the construction of the directed semantic dependence network, traverse the nodes and their associated edges in the network to identify semantic consistency deviation regions. The identification of semantic consistency deviation regions can be achieved by calculating the cosine similarity of semantic vectors between nodes. If the cosine similarity of the semantic vectors between a certain node and its directly associated predecessor node is lower than a specific threshold (such as 0.5), then mark this region as a semantic consistency deviation region. Mark the nodes in the semantic consistency deviation region as potential risk nodes, and construct a risk impact diffusion tree along the direction of the edges in the directed semantic dependence network with these potential risk nodes as the propagation source points.

[0069] The construction of the risk impact diffusion tree uses the breadth-first search algorithm. Starting from each potential risk node, expand along the directed edges and record the impact propagation path. To avoid loops, mark the visited nodes. At the same time, calculate the risk propagation probability according to the weight of the edge. When the cumulative risk value propagated to a certain node is lower than a specified threshold (such as 0.2), stop the expansion of this branch to ensure a reasonable boundary of the risk diffusion tree.

[0070] A traversal analysis is performed on all nodes in the directed semantic dependency network to locate nodes with multi-branch outgoing edges. A multi-branch outgoing edge is defined as two or more directed edges originating from the same node and pointing to different target nodes. For each node with multi-branch outgoing edges, the semantic direction difference of each outgoing edge is calculated. The semantic direction difference can be quantified by comparing the cosine distance between the semantic vectors of the target nodes connected by the edges. If the cosine distance between the semantic vectors of two target nodes is greater than a preset difference threshold (e.g., 0.4), the source node and its outgoing edges are labeled as decision nodes.

[0071] Based on the risk impact diffusion tree and labeled decision nodes, the propagation path of potential risk nodes in the risk impact diffusion tree is mapped to the outgoing edges of the decision nodes. Specifically, it checks whether the risk propagation path passes through labeled decision nodes and records all decision nodes on the path. This mapping relationship can be used to analyze key turning points in the risk propagation process, providing precise location data for subsequent risk intervention.

[0072] In a practical application scenario, suppose that in a generated article, attention analysis reveals a strong correlation between "policy adjustment" in the second sentence of the third paragraph and "economic impact" in the first sentence of the fifth paragraph, with a significant deviation in their semantic consistency. Marking "policy adjustment" as a potential risk node and constructing a risk impact diffusion tree with it as the source, we find that the risk propagation passes through multiple branches. One important decision node appears in the fourth paragraph, pointing to two branches with significant semantic differences: "positive prospects" and "potential challenges." This analysis allows for precise identification of the source and propagation path of content risks, providing a basis for content review and risk intervention.

[0073] The entire process of constructing the content risk propagation map integrates attention mechanism analysis, temporal semantic dependency modeling, and risk propagation simulation. It can effectively capture potential risks in complex texts and their impact diffusion patterns, providing a new technical solution for intelligent content review and risk management.

[0074] In one optional implementation, the content risk propagation map is mapped and associated with a multi-level review strategy library to identify uncertain areas requiring manual intervention. Based on the influence range of the decision node in the associated propagation path, the priority and granularity of review tasks are divided to form a review task allocation scheme, including:

[0075] Explore potential risk nodes and their corresponding associated propagation paths in the content risk propagation graph, and simultaneously guide a set of review strategies that match the risk types of the potential risk nodes from the multi-level review strategy library. Through dynamic association, deeply match the strategy rules in the review strategy set with the topology of the associated propagation path.

[0076] The inherent characteristics of the deep matching results are analyzed, the path segments in the association propagation path where the policy rule coverage is missing are identified, the path segments are defined as uncertainty regions, the decision nodes contained in the uncertainty regions are characterized, and the degree of downstream influence diffusion of each decision node in the association propagation path is evaluated.

[0077] Based on the degree of downstream impact diffusion, a hierarchical structure of influence of decision-making nodes within the uncertainty area is shaped, and a priority system for review tasks is derived based on the hierarchical structure of influence.

[0078] The granularity of the review task is constructed based on the topological boundary range of the uncertain region. The granularity of the review task is three-dimensionally characterized by the node distribution density and semantic level span within the uncertain region. The priority system of the review task and the granularity of the review task are integrated into the review task allocation scheme.

[0079] The mapping and association process between the content risk propagation graph and the multi-level review strategy library is achieved collaboratively through a graph traversal engine and a strategy matcher. The graph traversal engine uses a breadth-first search algorithm to perform a full scan of the content risk propagation graph, identifying all nodes with risk levels exceeding a preset threshold of 0.3 as potential risk nodes. Each potential risk node contains four basic attribute fields: node identifier, risk type enumeration value, risk intensity floating-point number, and location coordinates (3-tuple). The extraction of the associated propagation path starts from each potential risk node, tracing bidirectionally along directed edges with weights greater than 0.1, backtracking upwards to the source node with an in-degree of zero, and extending downwards to the termination node with an out-degree of zero, forming a complete propagation chain. The propagation path data structure includes four fields: starting node identifier, termination node identifier, intermediate node sequence, and total path weight. The total path weight is calculated by accumulating the weights of all edges.

[0080] The multi-level review strategy library is organized using a three-level tree index structure. The root node is classified into three main categories based on risk type: content security, semantic consistency, and logical rationality. Second-level nodes are divided into three levels based on risk severity: minor, moderate, and severe. Third-level leaf nodes store specific strategy rules, each containing three core components: an applicable condition expression, a processing action instruction, and a priority weight. The strategy guidance process is implemented through a type matcher. The matcher locates the corresponding strategy branch in the strategy library based on the risk type attribute of potential risk nodes and extracts all strategy rules that meet the type matching conditions to form a review strategy set.

[0081] A dynamic association mechanism establishes an adaptation relationship between policy rules and association propagation paths. The adaptation algorithm performs syntactic analysis on the applicable condition expressions in the policy rules, extracting node attributes, connection relationships, and topological features involved in the condition constraints. A topology structure matcher performs pattern matching between the structural pattern of the policy rules and the actual topology of the propagation path, calculating a structural similarity score. The similarity calculation considers three dimensions: node type consistency, edge connection direction matching degree, and subgraph isomorphism. A weighted sum is used to obtain a comprehensive matching score, which is a floating-point number ranging from 0 to 1.

[0082] The intrinsic feature analysis of deep matching results is performed through a coverage analyzer. The analyzer constructs a policy-path coverage matrix, where the row indices correspond to policy rules in the audit policy set, and the column indices correspond to nodes in the associated propagation path. Matrix element values ​​represent the degree of coverage of the corresponding policy rule for a specified node, ranging from 0 to 1. Coverage is calculated based on the matching degree between the applicability conditions of the policy rule and the node attributes. A coverage level of 1 indicates that the node fully meets all the applicability conditions of the policy rule; 0 indicates that it does not meet any conditions; and for nodes that partially meet the conditions, an intermediate value is calculated based on the proportion of conditions met.

[0083] The strategy rule identifies missing path segments through threshold filtering and connectivity analysis. The filter iterates through each column of the coverage matrix, calculating the maximum value of all elements in that column as the highest coverage of the corresponding node. When a node's highest coverage is below a set threshold of 0.7, the node is marked as a missing coverage node. The connectivity analyzer performs connected component detection on all missing coverage nodes, identifying the set of missing nodes interconnected on the propagation path, and defining each connected component as an uncertain region. The uncertain region data structure contains four fields: region identifier, list of contained nodes, set of boundary nodes, and a connectivity graph within the region.

[0084] The characterization of decision nodes within uncertain regions is based on their topological features and semantic importance. A decision node identifier calculates the out-degree value of each node within the region; nodes with an out-degree greater than 1 are marked as potential decision nodes. A semantic importance evaluator analyzes the positional weight and contextual dependency strength of nodes in the semantically generated sequence, calculating a comprehensive importance score through attention weight distribution and semantic embedding similarity. The final determination of decision nodes requires satisfying both the topological branching condition and a semantic importance threshold of 0.5.

[0085] The assessment of the downstream impact diffusion level is achieved using an impact propagation simulator. The simulator performs a breadth-first traversal starting from each decision node, counting the number of reachable downstream nodes as the impact breadth metric. The impact depth metric is obtained by calculating the path length from the decision node to the farthest reachable node. The impact strength metric considers the cumulative decay effect of all edge weights along the propagation path, using an exponential decay function to discount impacts farther from the decision node. The decay factor is set to 0.9 to ensure that the impact strength decreases by 10% for every additional layer of propagation depth.

[0086] The influence hierarchy is structured using clustering analysis to stratify decision-making nodes. A feature vector builder creates a 3-dimensional feature vector for each node, corresponding to the breadth, depth, and intensity of influence. The clustering algorithm uses the k-means method to divide the decision-making nodes into three influence levels, with cluster centers determined through iterative optimization. The high-influence level contains decision-making nodes with high values ​​in all dimensions of the feature vector; the medium-influence level contains nodes with high values ​​in some dimensions; and the low-influence level contains nodes with low values ​​in all dimensions.

[0087] The task prioritization system is directly mapped based on an influence hierarchy. High influence corresponds to urgent priority, requiring completion within 4 hours. Medium influence corresponds to important priority, requiring completion within 12 hours. Low influence corresponds to regular priority, allowing completion within 24 hours. The priority scheduler dynamically adjusts time requirements based on the current task queue load. Under high load, processing time may be appropriately extended, but not exceeding 1.5 times the original time limit.

[0088] The granularity of the audit task is constructed based on the topological boundary range analysis of uncertain regions. The boundary range calculator determines the spatial range of a region using the minimum enclosing subgraph algorithm and calculates the region diameter as a range metric. Node distribution density is calculated by the ratio of the number of nodes in the region to the region diameter. A ratio greater than 2 defines a high-density region, a ratio between 1 and 2 defines a medium-density region, and a ratio less than 1 defines a low-density region. The semantic hierarchy span analyzer examines the semantic hierarchy attributes of all nodes in the region and calculates the difference between the highest and lowest levels. A span greater than 3 levels is defined as a large-span region, a span between 1 and 3 levels is defined as a medium-span region, and a span less than 1 level is defined as a small-span region.

[0089] The 3D characterization mechanism uses node distribution density and semantic hierarchy span as the horizontal and vertical axes of a 2D coordinate system to locate each uncertain region in the granular space. High-density, small-span regions correspond to fine-grained tasks, requiring detailed node-by-node review. Medium-density, medium-span regions correspond to medium-granularity tasks, which can be reviewed in batches by sub-region. Low-density, large-span regions correspond to coarse-grained tasks, which can be processed using sampling review. The mapping relationship between granularity level and review method is stored in a configuration table and can be dynamically adjusted according to business needs.

[0090] The integration process of the task allocation scheme establishes a combination matrix of priority systems and granularity levels. This matrix contains nine different task type combinations, each corresponding to a specific review process and staffing requirements. Urgent priority tasks with fine granularity are assigned to senior review experts, equipped with auxiliary analysis tools and a rapid response channel. Important priority tasks with medium granularity are assigned to experienced reviewers, providing standardized review templates. Regular priority tasks with coarse granularity are assigned to junior reviewers, equipped with automated pre-screening tools. The allocation algorithm considers three factors: current staff workload, professional skill matching, and historical processing efficiency, selecting the optimal staffing allocation scheme through weighted scoring.

[0091] In one optional implementation, defining the path segment as an uncertainty region, characterizing the decision nodes contained within the uncertainty region, and evaluating the degree of downstream influence diffusion of each decision node in the associated propagation path includes:

[0092] In the associated propagation path of the content risk propagation map, the associated propagation path is segmented according to the semantic relevance and content coherence of the path nodes. The path segments with semantic relevance lower than the preset relevance threshold and broken content coherence are defined as uncertain regions, and the semantic boundary contour of the uncertain regions is constructed.

[0093] Based on the semantic boundary contour, semantic association analysis is performed on the nodes in the uncertainty region. Nodes with semantic divergence higher than a preset divergence threshold and multiple semantic associations are marked as decision nodes. The influence chain of decision nodes is constructed by the semantic dependency strength between the decision nodes.

[0094] Along the propagation direction of the influence chain of the decision node, the coverage range of downstream nodes of each decision node in the associated propagation path is statistically analyzed, and the coverage range of downstream nodes is quantified as the degree of downstream influence diffusion of the decision node.

[0095] In the associated propagation path of the content risk propagation graph, the propagation path is segmented to identify uncertain regions. Specifically, this is achieved by analyzing the semantic relevance and content coherence between path nodes. Semantic relevance is obtained by calculating the cosine similarity of text vectors between nodes. Word embedding is performed on the content of each node to obtain its feature vector representation, and the vector similarity between adjacent nodes is calculated. If the semantic relevance of two adjacent nodes is lower than a preset relevance threshold (e.g., 0.6), the connection between these two nodes is marked as a potential semantic breakpoint.

[0096] The content coherence between path nodes is assessed by detecting obvious topic shifts, logical breaks in argumentation, or abrupt changes in sentiment. Content coherence analysis is achieved through topic modeling techniques, extracting the core topic distribution for each node and calculating the JS divergence of topic distributions between adjacent nodes. When the divergence value exceeds a preset threshold, it is marked as a break in content coherence.

[0097] When a path segment simultaneously meets the conditions of semantic relevance being below a preset relevance threshold (e.g., 0.6) and content coherence being broken, the path segment is defined as an uncertain region. For example, in a path involving the dissemination of health information, when a node suddenly shifts from scientific medical advice to a node promoting unverified folk remedies, this transition area will be marked as an uncertain region.

[0098] After identifying the uncertain region, a semantic boundary contour of the region is constructed. This contour is achieved by analyzing the semantic features of nodes before and after the uncertain region, extracting keywords and topic vectors from stable nodes before and after the region to form the semantic features of entering and exiting the region. Through cluster analysis, these features are mapped into a high-dimensional semantic space to form a polygonal semantic contour, which represents the semantic boundary range and changing trend of the uncertain region.

[0099] Based on the constructed semantic boundary contour, semantic association analysis is performed on nodes in the uncertain region, the distance between each node in the region and the semantic boundary contour is calculated, and semantic localization features are identified. Then, a semantic association network of nodes in the region is constructed, the centrality and bridging degree of the nodes are calculated, and their positional importance in the semantic network is evaluated. Finally, the semantic diversity of the nodes is analyzed, and word sense disambiguation technology is used to identify whether the node content contains multiple semantic interpretations.

[0100] Based on the above analysis, nodes with semantic divergence exceeding a preset divergence threshold and possessing multiple semantic associations are marked as decision nodes. Semantic divergence is quantified by calculating the variance of the semantic association vector formed by the node and its surrounding nodes. When the variance of a node's semantic association with other nodes exceeds a preset threshold (e.g., 0.4), and the node can connect at least two different semantic clusters, it is marked as a decision node. For example, in a health information dissemination path, a transitional statement node that simultaneously connects medical advice and commercial marketing content would be identified as a decision node.

[0101] After decision node identification, an influence chain is constructed based on the semantic dependency strength between decision nodes. The semantic dependency strength is calculated through information flow analysis, considering the reference relationships, temporal sequence, and content similarity between nodes. For each pair of decision nodes (A, B), a conditional probability P(B|A) is calculated, representing the probability that node B will appear after node A appears, used to quantify the semantic dependency strength of A on B. When the semantic dependency strength exceeds a threshold (e.g., 0.7), directed connections are established between the decision nodes, forming the influence chain.

[0102] Along the propagation direction of the constructed decision node influence chain, the coverage of downstream nodes in the associated propagation path of each decision node is statistically analyzed. Specifically, starting from each decision node, all its downstream paths are traversed, and the number of all reachable nodes is recorded. For downstream nodes influenced by multiple decision nodes, the influence contribution is allocated by calculating influence weights, which are calculated based on the propagation path length and semantic relevance.

[0103] The coverage of downstream nodes is quantified as the degree of downstream influence diffusion of decision-making nodes. The degree of downstream influence diffusion considers not only the number of covered nodes, but also the importance of these nodes (such as node participation, activity, and attention) and the dissemination effect (such as information conversion rate and interaction rate). Through weighted calculation, a downstream influence diffusion index is obtained for each decision-making node, which reflects the key influence of the decision-making node in the process of content risk dissemination.

[0104] Taking the spread of rumors on social media as an example, in a propagation path that gradually evolves from health consultation to the promotion of fake medical products, the above methods can identify key decision-making nodes—such as the node where unverified efficacy descriptions are first introduced. Analysis shows that the downstream influence of this decision-making node reaches 85%, meaning that this node is a critical turning point in the spread of risky content. Based on this analysis, targeted intervention measures can be implemented at this decision-making node to effectively block the large-scale spread of risky content.

[0105] In one optional implementation, the content to be reviewed is reshaped based on the review judgment and the semantic correction operation. The manual decision-making logic in the semantic correction operation is transformed into rule-based constraints. The source node causing the risk is located through reverse tracing of the content risk propagation graph, forming a review knowledge accumulation, including:

[0106] Based on the results of the review and judgment, the sequence of adjustment instructions in the semantic correction operation is explored, a directional association mapping graph between the sequence of adjustment instructions and the semantic structure of the content to be reviewed is constructed, and the content to be reviewed is semantically reconstructed according to the hierarchical transmission relationship in the association mapping graph to form a semantic hierarchical organization with a clear correction trajectory.

[0107] A deep analysis is conducted around the semantic hierarchy organization to identify the judgment criteria and correction evolution path contained in the semantic correction operation. The judgment criteria and correction evolution path are cross-fused to construct a rule-based constraint chain. At the same time, the semantic nodes that trigger the rule-based constraint chain are depicted in the semantic hierarchy organization.

[0108] By leveraging the spatial location characteristics of the semantic nodes in the content risk propagation graph, reverse tracing is performed along the transmission path of the semantic dependency network to locate the source node group that triggers the risk. The rule-based constraint chain is then linked and integrated with the semantic feature spectrum of the source node group to form a systematic accumulation of review knowledge.

[0109] As shown in Figure 2, the method includes:

[0110] The semantic reconstruction process of the content to be reviewed is achieved through instruction sequence analysis. The analyzer extracts the execution records of semantic correction operations from the review judgment results. Each execution record contains four basic fields: operation type identifier, target content location, semantic state before correction, and semantic state after correction. The operation type identifier is represented by enumeration values, including four basic types: deletion, replacement, insertion, and reordering. The target content location is precisely located using a tuple of character offset and semantic level coordinates. The character offset records the start and end positions in the original text, and the semantic level coordinates identify the node path in the semantic parsing tree. The semantic states before and after correction are encoded using vectorized representations. The vector dimension is fixed at 512 dimensions, with each dimension corresponding to a specific semantic feature weight.

[0111] The construction of the directional association mapping graph is based on the correspondence between the adjustment instruction sequence and the semantic structure of the content to be reviewed. Semantic structure parsing transforms the content to be reviewed into a hierarchical semantic tree structure. The root node of the tree represents the overall semantics, and the leaf nodes correspond to specific semantic units such as words, phrases, and sentences. Each semantic node contains five attributes: node identifier, semantic type, hierarchical depth, parent-child relationship pointer, and semantic weight. The semantic type is divided into four categories: topic words, modifiers, conjunctions, and function words. The hierarchical depth ranges from 0 to a maximum of 15. The association mapping algorithm traverses each instruction in the adjustment instruction sequence, locates the corresponding semantic node in the semantic tree according to the position of the target content, and establishes a direct mapping relationship between instructions and nodes. The mapping relationship data structure contains four fields: instruction identifier, corresponding node identifier, influence strength, and propagation range. The influence strength ranges from 0.1 to 1.0, and the propagation range represents the number of affected neighboring nodes.

[0112] The identification of hierarchical transmission relationships is performed in the association mapping graph using a depth-first traversal algorithm. Transmission path tracing starts from each affected semantic node and expands bidirectionally along parent-child and sibling relationships to identify all nodes affected by the transmission. Transmission strength calculation is based on a distance decay model: the transmission strength of a directly corrected node is 1.0, the transmission strength of a neighboring node at a distance of 1 is 0.8, the transmission strength of a node at a distance of 2 is 0.6, and so on until the strength falls below the threshold of 0.1. Semantic reconstruction reorganizes the affected nodes according to the transmission strength. Nodes with a strength greater than 0.5 require a complete recalculation of their semantic representation; nodes with a strength between 0.2 and 0.5 require partial adjustment of their semantic weights; and nodes with a strength less than 0.2 retain their original semantic representation. The correction trajectory record includes five core fields: operation timestamp, list of involved nodes, change type, change magnitude, and verification status. The operation timestamp uses millisecond precision to ensure accurate sorting of the operation sequence. The magnitude of the change is calculated by the cosine similarity of the semantic vectors before and after the correction. A similarity of less than 0.7 indicates a major change, between 0.7 and 0.9 indicates a moderate change, and a similarity of more than 0.9 indicates a minor change.

[0113] The in-depth analysis of the semantic hierarchy organization involves scanning the reconstructed semantic structure using pattern recognition to identify decision criterion patterns and correction evolution path patterns. Decision criterion patterns are identified using a rule template matching algorithm. The template library contains 50 common semantic decision patterns, each defining a standardized description of triggering conditions, decision logic, and decision output. Triggering conditions describe the attribute constraints of semantic nodes using Boolean expressions. The decision logic uses a decision tree structure to represent the decision process, and the decision output includes three result types: pass, reject, and correct. Structural similarity calculation is achieved through topological matching between the semantic organization and the templates; templates with a similarity greater than 0.8 are identified as matching decision criterion patterns. Correction evolution path identification analyzes the temporal and spatial distribution characteristics of semantic correction operations. Temporal characteristics are calculated to determine the time interval distribution between correction operations, identifying three temporal patterns: continuous correction, intermittent correction, and batch correction. Spatial distribution is calculated to determine the positional distribution of correction operations in the semantic tree, identifying three spatial patterns: centralized correction, distributed correction, and hierarchical correction.

[0114] The construction of the rule-based constraint chain combines decision criteria with modified evolution paths through a cross-fusion algorithm. The fusion algorithm creates a Cartesian product of decision criteria and evolution paths, with each combination corresponding to a potential constraint rule. The constraint rule comprises four components: preconditions, triggering events, execution actions, and post-verification. Preconditions inherit the triggering conditions from the decision criteria; triggering events correspond to the initial operations of the evolution path; execution actions serialize all modified operations in the evolution path; and post-verification fuses the decision output of the decision criteria and the termination conditions of the path. Rule optimization refines the constraint rules through redundancy elimination and logical simplification, removing duplicate condition judgments and merging similar execution actions to form a compact constraint chain. Semantic node trigger feature identification extracts a 20-dimensional feature vector from each node in the semantic hierarchy. Features include five categories: semantic type, hierarchical position, connectivity, weight distribution, and contextual similarity. Semantic type features are represented using one-hot encoding, hierarchical position features are normalized to the range of 0 to 1, connectivity features calculate the sum of the in-degree and out-degree of a node, weight distribution features statistically analyze the weight changes of a node in different semantic dimensions, and context similarity features calculate the mean semantic similarity between a node and its neighboring nodes.

[0115] The reverse source tracing process analyzes the coordinates and neighborhood distribution of semantic nodes in the content risk propagation graph through spatial location feature analysis. Spatial location is achieved by projecting the high-dimensional semantic space onto a 2D plane using a graph embedding algorithm, preserving the relative distance relationships between nodes. Neighborhood feature analysis examines the distribution density and type diversity of other nodes within a radius of 3 around a given node. The propagation path tracing starts from the triggering semantic node and proceeds backward along the semantic dependency edges, tracing all propagation paths until the source node with an in-degree of 0 is found. The propagation path includes four attributes: dependency strength, propagation direction, path length, and intermediate nodes. Dependency strength is calculated based on semantic similarity and co-occurrence frequency, ranging from 0.1 to 1.0. The propagation direction records the direction of information flow. The path length counts the number of edges traversed from the source node to the target node. The intermediate node list records the identifiers of all nodes along the propagation path.

[0116] The source node clusters are located using clustering analysis to group the back-traced source nodes. The distance metric uses a weighted combination of semantic similarity and spatial distance, with semantic similarity weighted at 0.7 and spatial distance weighted at 0.3. The clustering algorithm employs density-based clustering to automatically determine the number and boundaries of clusters, forming several source node clusters. Each node cluster contains four basic attributes: a cluster identifier, a list of core nodes, a list of boundary nodes, and a cluster feature vector. The cluster feature vector is obtained by calculating a weighted average of the feature vectors of all nodes within the cluster, with the weights determined based on the node's centrality within the cluster. The semantic feature hierarchy is integrated by analyzing the semantic evolution history of each node in the source node clusters, extracting the change trajectory and distribution pattern of semantic features. The feature change trajectory records the change process of the semantic vector over time, while the distribution pattern statistically analyzes the statistical properties of each dimension of the feature vector, including the mean, variance, and skewness.

[0117] The systematic review and knowledge accumulation is structured and stored in the form of a knowledge graph, which consists of three layers: entity nodes, relationship edges, and attribute attachments. Entity nodes correspond to knowledge units such as source node groups, constraint rules, and semantic features; relationship edges represent semantic connections between knowledge units; and attribute attachments store detailed information and metadata of the knowledge units. Graph construction is completed through three steps: entity recognition, relationship extraction, and attribute filling. Entity recognition extracts knowledge entities from the semantic feature spectrum and constraint chains; relationship extraction establishes connections between entities based on co-occurrence patterns and dependencies; and attribute filling attaches detailed feature information and metadata to the corresponding entities. Version management provides version control functionality for the knowledge graph, supporting incremental updates and rollback operations to ensure the continuous accumulation and optimization of knowledge.

[0118] In one optional implementation, a deep analysis is conducted around the semantic hierarchy to identify the judgment criteria and correction evolution paths inherent in the semantic correction operation. The judgment criteria and correction evolution paths are then cross-fused to construct a rule-based constraint chain, including:

[0119] The analysis focuses on the structural context of the semantic hierarchy, decomposing the semantic hierarchy into semantic association subgraphs, tracing the semantic transmission of nodes in the semantic association subgraphs, and characterizing the transmission links and affected areas activated by each semantic correction operation in the semantic association subgraphs.

[0120] At the transmission node of the transmission link, the judgment criteria included in the semantic correction operation are identified, the judgment threshold and judgment dimension in the judgment criteria are extracted, the judgment threshold and the judgment dimension are combined into a semantic evaluation vector, and the evolution trajectory of semantic correction is drawn in the influence area, and the evolution trajectory is transformed into a correction evolution path.

[0121] Based on the interaction between the semantic evaluation vector and the modified evolution path, the constraint triggering conditions and the constraint scope are marked, and a regularized constraint chain is formed through the mapping association between the constraint triggering conditions and the constraint scope.

[0122] When conducting in-depth analysis around the aforementioned semantic hierarchy, the semantic hierarchy is decomposed into semantic association subgraphs. In specific implementation, multiple semantic association subgraphs are constructed by establishing relationships between nodes. For example, in a medical text analysis scenario, symptom descriptions and examination results in a diagnostic report can be associated with the disease diagnosis to form a semantic association subgraph of "symptom-examination-diagnosis".

[0123] Semantic transmission tracking of nodes in a semantic association subgraph can be performed using a depth-first traversal algorithm. Starting from the source node, the algorithm records the transmission process of semantic information between nodes and marks the semantic state changes of each node. During the transmission process, each edge is assigned a transmission weight to reflect the degree of attenuation or enhancement of semantic information during transmission. For example, in legal document analysis, the semantic transmission path from the "cause of action" node to the "judgment result" node can record the semantic changes of intermediate nodes such as key fact determinations and application of legal provisions.

[0124] When characterizing the transmission chain and affected area of ​​semantic correction operations, an impact propagation model needs to be established. When a semantic correction occurs at a node, the diffusion range of the correction's impact is calculated according to preset propagation rules. The propagation rules can be determined based on a combination of semantic relevance between nodes, distance, and the importance weight of the nodes themselves. The affected area can be identified by setting a threshold; that is, when the impact value is below a certain threshold, the node is considered to be outside the affected area. In financial risk analysis scenarios, if the "default risk" node undergoes semantic correction, its impact on related nodes such as "credit rating" and "credit limit" can be tracked.

[0125] To identify the judgment criteria included in the semantic correction operation at the transmission nodes of the transmission link, a semantic feature extractor needs to be constructed. By analyzing the semantic change patterns between nodes, the judgment logic can be extracted. For example, in product evaluation analysis, if a specific semantic transformation rule is found in the transmission process from "user experience" to "product rating," such as "complex operation" leading to a lower rating, then the judgment dimension of "ease of operation" and its threshold conditions affecting the rating can be extracted.

[0126] When extracting the judgment threshold and judgment dimension from the judgment criteria, a feature importance analysis method is used. First, the change range of each dimension feature before and after semantic correction is statistically analyzed. Then, the contribution of each feature to the correction result is calculated, and features with a contribution higher than a preset value are selected as judgment dimensions. The judgment threshold is determined by analyzing the correspondence between feature values ​​and correction results. Methods such as decision trees can be used to automatically extract split points as candidate thresholds. For example, in network security analysis, by analyzing multiple security incident handling records, the judgment threshold of "access frequency exceeding 100 times per minute" and the judgment dimension of "access frequency" can be extracted.

[0127] The judgment threshold and judgment dimension are combined into a semantic evaluation vector. Specifically, an N-dimensional vector space is constructed, where each dimension corresponds to a judgment dimension, and the value in each dimension corresponds to the judgment threshold or standardized feature value of that dimension. In intelligent recommendation systems, evaluation vectors containing dimensions such as "user interest similarity," "content timeliness," and "interaction frequency" can be constructed to quantify the accuracy of content recommendations.

[0128] When depicting the evolutionary trajectory of semantic correction, a temporal state recording method is employed, decomposing the correction process into multiple temporal state points, each recording the set of semantic features at that moment. By comparing the differences between adjacent state points, the direction and magnitude of semantic changes are identified, thereby constructing a complete evolutionary trajectory. In natural language processing scenarios, the migration path of semantic focus during text editing can be tracked, such as the evolution from "problem description" to "cause analysis" and then to "solution."

[0129] When transforming an evolutionary trajectory into a corrected evolutionary path, the trajectory needs to be abstracted and standardized. The continuous evolutionary trajectory is discretized into a sequence of key state nodes, and state transition rules are extracted to form a reusable path template. For example, in a customer service dialogue system, the process of "problem identification - information collection - solution recommendation - result confirmation" can be refined into a standardized corrected evolutionary path to guide the processing of similar scenarios in the future.

[0130] Based on the interaction between the semantic evaluation vector and the modified evolution path, the constraint triggering conditions and the scope of constraint effect are marked. In specific implementation, the combination of conditions that triggers a specific constraint is determined by analyzing the state changes of the evaluation vector at each node of the evolution path. The scope of constraint effect is calculated using an influence propagation model. In smart contract execution monitoring, it can be determined that when the "transaction amount" dimension exceeds a specific threshold and the "credit score" dimension is below the threshold, an "additional review" constraint is triggered, and its scope of effect includes transaction verification and fund transfer.

[0131] By mapping and associating constraint triggering conditions with their scope of application, a rule-based constraint chain is formed, establishing a many-to-many mapping relationship from triggering conditions to their scope of application, and constructing a complete constraint network. Each constraint rule includes attributes such as a description of the triggering condition, the object of application, the execution operation, and the priority. In an enterprise risk control system, a constraint chain from "abnormal transaction identification" to "account freezing" can be constructed, clearly defining the judgment criteria and execution steps for each stage, ensuring the accurate triggering and effective implementation of risk control measures.

[0132] In one optional implementation, integrating the audit knowledge into the matching rules of the multi-level audit strategy library and improving the construction parameters of the content risk propagation graph include:

[0133] Deconstruct the rule constraint structure in the knowledge accumulation of the audit, identify the judgment conditions and correction strategies in the rule constraint structure, map the judgment conditions and correction strategies to the matching rules of the corresponding level in the multi-level audit strategy library, and construct a fusion and association network of new and old rules;

[0134] Based on the fusion association network, the supplementary coverage rate and rule conflict rate of the newly added rule constraints on the original matching rules are calculated. Rule constraints with a supplementary coverage rate exceeding a preset coverage threshold and a rule conflict rate lower than a preset conflict threshold are selected and integrated into the matching rule system of the multi-level review strategy library.

[0135] Based on the rule coverage in the fusion and association network, the structural features of the content risk propagation graph are analyzed, the weight distribution deviation on the risk propagation path is extracted, and the construction parameters of the content risk propagation graph are dynamically adjusted according to the weight distribution deviation.

[0136] This process deconstructs the rule constraint structure within the accumulated review knowledge, identifying the judgment conditions and correction strategies. In practice, semantic analysis is performed on the collected review knowledge data to extract key information elements. For example, the rule constraint "containing sensitive word A and published more than 10 times / hour" is broken down into the judgment conditions "containing sensitive word A" and "publishing frequency > 10 times / hour," with the correction strategy being "restrict publishing." Natural language processing techniques are used to perform part-of-speech tagging and dependency analysis on the rule text, identifying conditional statements and result statements. Conditional statements typically contain three parts: trigger words, object words, and thresholds, while result statements contain action words and the scope of influence. These identified judgment conditions and correction strategies are mapped to corresponding levels in a multi-level review strategy library, such as the basic layer, content layer, user layer, or scenario layer. Based on semantic similarity calculation, new rules are associated and paired with existing rules, constructing a fusion and association network between the old and new rules.

[0137] The supplementary coverage rate and rule conflict rate of newly added rule constraints are calculated based on the constructed fusion and correlation network. The supplementary coverage rate is calculated as follows: for a specific risk category, the number of content samples that the new rule can cover but the original rule cannot is counted, and then divided by the total number of samples in that category. For example, if the original rule has a coverage rate of 80% for illegal advertising content, and the new rule can identify the 15% of content that was missed, then the supplementary coverage rate of the new rule is 15%. The rule conflict rate is calculated for the same content sample. When the old and new rules give different review results, it is considered a rule conflict. The number of conflicting samples is counted and divided by the total number of samples that are covered in both categories. By setting preset coverage thresholds (e.g., 10%) and preset conflict thresholds (e.g., 5%), rule constraints with high supplementary coverage rates and low conflict rates are selected. For the selected rules, their applicable scenarios and triggering conditions are further analyzed, the rule priorities and weights are adjusted, and finally they are integrated into the matching rule system of the multi-level review strategy library.

[0138] Based on the coverage after rule integration, the structural characteristics of the content risk propagation graph are analyzed. The content risk propagation graph consists of nodes (content entities) and edges (propagation relationships). Each node has risk attributes and propagation influence, while each edge has propagation intensity and timeliness. By analyzing historical review data, the risk trigger frequency and impact range of each node in the graph are calculated, identifying weight distribution deviations along the risk propagation path. For example, it was found that the propagation speed of a certain type of content on social media platforms is three times that of news platforms, but the weights of the two are similar in the original graph, indicating a weight distribution deviation.

[0139] To address identified biases in weight distribution, the construction parameters of the content risk propagation graph are dynamically adjusted. These parameters primarily include node risk weights, edge propagation strength, time-decrease factors, and path importance. Based on the deviation between the actual propagation effect and the prediction model, a backpropagation algorithm is used to optimize each parameter. Specifically, when the actual propagation speed of a certain type of content differs significantly from the predicted value, the edge strength parameter of the corresponding propagation path is adjusted; when the impact range of a certain type of risky content exceeds expectations, the risk weight of the corresponding node is increased. By iteratively adjusting these parameters, the content risk propagation graph more accurately reflects the actual risk propagation situation, thereby providing a more precise decision-making basis for the multi-level review strategy library.

[0140] In practical applications, taking short video platforms as an example, when the spread of misinformation about a trending event deviates from expectations, the risk weights and side propagation strengths of relevant nodes in the propagation graph are adjusted using the methods described above. After the adjustment, the system can identify potential risk propagation nodes earlier, improving the accuracy of review by 15% and reducing the average interception time of risky content by 30 minutes, effectively controlling the spread of misinformation.

[0141] Through the aforementioned technologies, the matching rules of the multi-level review strategy library are integrated with the accumulated review knowledge, and the construction parameters of the content risk propagation graph are improved. This enables the effective utilization of review knowledge and the continuous optimization of risk prevention and control capabilities, providing content platforms with a more accurate and efficient review guarantee mechanism.

[0142] A second aspect of the present invention provides a large-scale model content generation and review system based on human-machine collaboration, comprising:

[0143] The first module is used to acquire the content to be reviewed generated by the large language model, deconstruct the generation process data corresponding to the content to be reviewed, the generation process data includes attention distribution and semantic decoding path; construct a content risk propagation graph based on the attention distribution and the semantic decoding path, the content risk propagation graph represents the semantic dependency relationship and the associated propagation path of potential risk nodes in the generation process, and mark decision nodes in the content risk propagation graph;

[0144] The second module is used to map and associate the content risk propagation map with the multi-level review strategy library, identify uncertain areas that require manual intervention, and divide the priority and granularity of review tasks according to the influence range of the decision node in the associated propagation path to form a review task allocation scheme.

[0145] The third module is used to import the review tasks into the manual review interface according to the review task allocation scheme, obtain the review judgment and semantic correction operation returned by the manual reviewer for the uncertain area; reshape the content to be reviewed according to the review judgment and the semantic correction operation, transform the manual decision-making logic in the semantic correction operation into rule-based constraints, and locate the source node that caused the risk through the reverse tracing of the content risk propagation graph, forming a review knowledge accumulation.

[0146] The fourth module is used to integrate the audit knowledge into the matching rules of the multi-level audit strategy library and improve the construction parameters of the content risk propagation graph.

[0147] A third aspect of the present invention provides an electronic device, comprising:

[0148] processor;

[0149] Memory used to store processor-executable instructions;

[0150] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0151] 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.

[0152] 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.

[0153] 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 generating and reviewing large-scale model content based on human-computer collaboration, characterized in that: include: Obtain the content to be reviewed generated by the large language model, deconstruct the generation process data corresponding to the content to be reviewed, the generation process data includes attention distribution and semantic decoding path; construct a content risk propagation graph based on the attention distribution and the semantic decoding path, the content risk propagation graph represents the semantic dependency relationship and the associated propagation path of potential risk nodes in the generation process, and mark decision nodes in the content risk propagation graph; The content risk propagation graph is mapped and associated with a multi-level review strategy library to identify uncertain areas requiring human intervention. Based on the influence range of the decision nodes in the associated propagation path, the priority and granularity of review tasks are divided to form a review task allocation scheme. Review tasks are imported into a human review interface according to the allocation scheme to obtain the review judgments and semantic correction operations returned by human reviewers for the uncertain areas. The content to be reviewed is reshaped based on the review judgments and semantic correction operations, and the human decision-making logic in the semantic correction operations is transformed into rule-based constraints. The source node causing the risk is located through reverse tracing of the content risk propagation graph, forming a review knowledge accumulation. This review knowledge accumulation is integrated into the matching rules of the multi-level review strategy library, and the construction parameters of the content risk propagation graph are improved.

2. The method according to claim 1, characterized in that, A content risk propagation graph is constructed based on the attention distribution and the semantic decoding path. This graph represents the semantic dependencies and the associated propagation paths of potential risk nodes during the generation process. Marking decision nodes in the content risk propagation graph includes: parsing the association weights between semantic units in the attention distribution; constructing a directed semantic dependency network based on these weights; marking generation time sequence identifiers in each node of the directed semantic dependency network in conjunction with the generation step sequence in the semantic decoding path; determining the causal chain of semantic evolution through the generation time sequence identifiers and the directionality of edges in the directed semantic dependency network; and traversing the nodes and their associated edges in the directed semantic dependency network to identify semantic consistency. In the semantic consistency deviation region, nodes within the region are marked as potential risk nodes. Using these potential risk nodes as propagation sources, a risk impact diffusion tree is constructed along the edge propagation direction of the directed semantic dependency network. All nodes in the directed semantic dependency network are traversed and analyzed to locate nodes with multiple outgoing edges. The semantic direction difference of each outgoing edge is calculated, and nodes whose semantic direction difference exceeds a preset difference threshold, along with their outgoing edges, are marked as decision nodes. Based on the risk impact diffusion tree and the decision nodes, the propagation path of the potential risk nodes in the risk impact diffusion tree is associated with the outgoing edges of the decision nodes, identifying the decision nodes traversed in the risk propagation path.

3. The method according to claim 1, characterized in that, The content risk propagation map is mapped and associated with a multi-level review strategy library to identify uncertain areas requiring manual intervention. Based on the influence range of the decision nodes in the propagation path, the priority and granularity of review tasks are divided to form a review task allocation scheme. This scheme includes: exploring potential risk nodes in the content risk propagation map and their corresponding propagation paths; simultaneously guiding a set of review strategies from the multi-level review strategy library that matches the risk types of the potential risk nodes; performing deep matching between the strategy rules in the review strategy set and the topology of the propagation path through dynamic association; and analyzing the inherent characteristics of the deep matching results to identify the strategy rules in the propagation path. The process involves covering missing path segments, defining these segments as uncertainty regions, characterizing the decision nodes within these uncertainty regions, and assessing the downstream influence diffusion degree of each decision node in the associated propagation path. Based on the downstream influence diffusion degree, an influence hierarchy structure of the decision nodes within the uncertainty regions is shaped, and a priority system for review tasks is derived from this hierarchy. The granularity of the review tasks is constructed based on the topological boundary range of the uncertainty regions. This granularity is characterized by the node distribution density and semantic level span within the uncertainty regions. Finally, the priority system and the granularity of the review tasks are integrated into a review task allocation scheme.

4. The method according to claim 3, characterized in that, Defining the path segments as uncertain regions, characterizing the decision nodes contained within the uncertain regions, and evaluating the downstream influence diffusion degree of each decision node in the associated propagation path includes: in the associated propagation path of the content risk propagation map, segmenting the associated propagation path according to the semantic relevance and content coherence of the path nodes, defining the path segments with semantic relevance below a preset relevance threshold and broken content coherence as uncertain regions, and constructing the semantic boundary contour of the uncertain regions; performing semantic relevance analysis on the nodes in the uncertain regions based on the semantic boundary contour, marking the nodes with semantic divergence higher than a preset divergence threshold and multiple semantic associations as decision nodes, and constructing the decision node influence chain through the semantic dependency strength between the decision nodes; along the propagation direction of the decision node influence chain, statistically analyzing the downstream node coverage of each decision node in the associated propagation path, and quantifying the downstream node coverage as the downstream influence diffusion degree of the decision node.

5. The method according to claim 1, characterized in that, Based on the review judgment and the semantic correction operation, the content to be reviewed is reshaped, and the manual decision-making logic in the semantic correction operation is transformed into rule-based constraints. The source node causing the risk is located through reverse tracing of the content risk propagation graph, forming a review knowledge accumulation. This includes: exploring the adjustment instruction sequence in the semantic correction operation based on the review judgment result; constructing a directional association mapping graph between the adjustment instruction sequence and the semantic structure of the content to be reviewed; and semantically reconstructing the content to be reviewed based on the hierarchical transmission relationship in the association mapping graph, forming a semantic hierarchical organization with a clear correction trajectory. The semantic hierarchy is analyzed in depth to identify the judgment criteria and correction evolution paths contained in the semantic correction operation. The judgment criteria and correction evolution paths are cross-integrated to construct a rule-based constraint chain. At the same time, the semantic nodes that trigger the rule-based constraint chain are depicted in the semantic hierarchy. By using the spatial location characteristics of the semantic nodes in the content risk propagation graph, the source nodes that cause the risk are located by tracing back along the transmission path of the semantic dependency network. The semantic feature spectrum of the rule-based constraint chain and the source node group is associated and integrated to form a systematic accumulation of review knowledge.

6. The method according to claim 5, characterized in that, A deep analysis of the semantic hierarchy is conducted to identify the judgment criteria and correction evolution paths inherent in the semantic correction operations. The process of cross-integrating these criteria and paths to construct a rule-based constraint chain includes: analyzing the structural context of the semantic hierarchy, decomposing it into semantic association subgraphs, tracing semantic transmission of nodes within these subgraphs, and characterizing the transmission links and influence regions activated by each semantic correction operation within the subgraphs; identifying the judgment criteria included in the semantic correction operations at the transmission nodes of these links, extracting the judgment thresholds and dimensions from these criteria, combining them into a semantic evaluation vector, and simultaneously depicting the semantic correction evolution trajectory within the influence region, transforming this trajectory into a correction evolution path; and based on the interaction between the semantic evaluation vector and the correction evolution path, marking constraint triggering conditions and constraint scope, and forming a rule-based constraint chain through the mapping between these conditions.

7. The method according to claim 1, characterized in that, Integrating the accumulated review knowledge into the matching rules of the multi-level review strategy library and improving the construction parameters of the content risk propagation graph include: deconstructing the rule constraint structure in the accumulated review knowledge, identifying the judgment conditions and correction strategies in the rule constraint structure, mapping the judgment conditions and correction strategies to the corresponding level of matching rules in the multi-level review strategy library, and constructing a fusion and association network of new and old rules; calculating the supplementary coverage rate and rule conflict rate of the newly added rule constraints on the original matching rules based on the fusion and association network, filtering rule constraints with a supplementary coverage rate exceeding a preset coverage threshold and a rule conflict rate lower than a preset conflict threshold, and integrating the rule constraints into the matching rule system of the multi-level review strategy library; analyzing the structural characteristics of the content risk propagation graph according to the rule coverage in the fusion and association network, extracting the weight distribution deviation on the risk propagation path, and dynamically adjusting the construction parameters of the content risk propagation graph for the weight distribution deviation.

8. A large-scale model content generation and review system based on human-computer collaboration, used to implement the method of any one of claims 1-7, characterized in that, include: The first module is used to acquire the content to be reviewed generated by the large language model, deconstruct the generation process data corresponding to the content to be reviewed, the generation process data includes attention distribution and semantic decoding path; construct a content risk propagation graph based on the attention distribution and the semantic decoding path, the content risk propagation graph represents the semantic dependency relationship and the associated propagation path of potential risk nodes in the generation process, and mark decision nodes in the content risk propagation graph; The second module is used to map and associate the content risk propagation map with the multi-level review strategy library, identify uncertain areas that require manual intervention, and divide the priority and granularity of review tasks according to the influence range of the decision node in the associated propagation path to form a review task allocation scheme. The third module is used to import the review tasks into the manual review interface according to the review task allocation scheme, obtain the review judgment and semantic correction operation returned by the manual reviewer for the uncertain area; reshape the content to be reviewed according to the review judgment and the semantic correction operation, transform the manual decision-making logic in the semantic correction operation into rule-based constraints, and locate the source node that caused the risk through the reverse tracing of the content risk propagation graph, forming a review knowledge accumulation; the fourth module is used to integrate the review knowledge accumulation into the matching rules of the multi-level review strategy library, and improve the construction parameters of the content risk propagation graph.

9. An electronic device, characterized in that, include: processor; A memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to perform 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.

Images