Domain corpus intelligent generation system based on policy knowledge graph

Abstract

The invention relates to the technical field of natural language processing knowledge maps, and discloses an intelligent domain corpus generation system based on a policy knowledge map. The system comprises a policy knowledge graph analysis module, a network depth evolution module and a semantic unit deconstruction module. And the system extracts nodes and relationships in the atlas according to the target domain identifier to form an initial policy association network. Isolated nodes are removed through iterative pruning, semantic nested connection between core nodes is enhanced, and a dense semantic network with a compact structure is generated. The system identifies policy semantic clusters in the network, slices description texts of nodes in the clusters and performs cross-node logic recombination, and finally outputs a normalized policy semantic unit set. According to the method, automatic generation of corpora from a knowledge graph to a high-quality and structured field is achieved, and the generated corpora are logically coherent and high in semantic relevance.

Description

Technical Field

[0001] This invention relates to the field of natural language processing knowledge graph technology, specifically to an intelligent generation system for domain corpus based on policy knowledge graph. Background Technology

[0002] In the fields of policy research and intelligent analysis, automatically constructing high-quality domain corpora from massive and structurally complex policy documents is crucial for supporting applications such as precise analysis and intelligent question answering. Existing technologies typically employ methods based on keyword queries or finite path traversal within pre-built knowledge graphs to extract relevant nodes and their associated textual descriptions, followed by simple aggregation and splicing to form the basis of the generated corpus.

[0003] Conventional methods have drawbacks. The extracted node networks often contain a large number of sparsely connected edge nodes with low semantic contribution, leading to redundancy and significant noise in the initial data. More importantly, the deep, nested semantic relationships between nodes are completely lost during simple text aggregation, resulting in a loosely logical and incoherent corpus. Mechanically splicing text fails to deconstruct and reassemble it based on semantic relationships, failing to form cohesive, structured semantic units, making it difficult to directly apply to deep learning models that require high-quality input.

[0004] Current technology needs to address two core issues: first, how to dynamically extract a tight network structure that reflects deep semantic logic from an initial, coarse network of connections; and second, how to intelligently slice and reorganize text content scattered across different nodes according to semantic logic based on this optimized structure, ultimately generating a standardized and usable set of semantic units. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent generation system for domain corpora based on policy knowledge graphs to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this invention provides an intelligent generation system for domain corpus based on policy knowledge graphs, the system comprising: The policy knowledge graph parsing module, based on the input target policy domain identifier, performs a subgraph traversal operation in the pre-built policy knowledge graph to extract all policy nodes associated with the target policy domain identifier and the relationship paths between nodes, thereby forming an initial policy association network. The network deep evolution module performs iterative pruning and reinforcement operations on the initial policy association network, removing isolated policy nodes with connectivity below a preset threshold in the initial policy association network, while enhancing the connection weights between core policy nodes with multi-layer semantic nesting relationships, generating a denser policy semantic network. The semantic unit deconstruction module identifies multiple policy semantic clusters composed of the core policy nodes in the dense policy semantic network, and slices and reassembles the description text of all policy nodes within each policy semantic cluster to generate a set of standardized policy semantic units corresponding to each policy semantic cluster.

[0007] Preferably, the deep evolution module performs iterative pruning and reinforcement operations on the initial policy association network to generate a denser policy semantic network, including: Calculate the topological connectivity of each policy node in the initial policy association network, and mark all policy nodes whose topological connectivity is lower than the connectivity screening condition according to the preset connectivity screening condition as isolated policy nodes to be pruned; Remove all marked isolated policy nodes to be pruned and their associated edges from the initial policy association network to form a pruned policy association subnetwork; In the pruned policy association subnetwork, all node pairs containing semantic relationships exceeding a preset nesting level are identified as core policy node pairs with multi-level semantic nesting relationships; The interaction frequency and semantic association strength on all relational paths between the core policy node pairs are analyzed, and the weight values ​​of the connecting edges between the core policy node pairs are increased proportionally based on the analysis results to complete the strengthening process of the pruned policy association sub-network and output the dense policy semantic network.

[0008] Preferably, the semantic unit deconstruction module generates a set of standardized policy semantic units corresponding to each policy semantic cluster, including: In the dense policy semantic network, a community detection algorithm is applied to identify multiple sub-network regions with closely clustered nodes, and each sub-network region is defined as a policy semantic cluster. For each identified policy semantic cluster, the original descriptive text of all policy nodes belonging to the policy semantic cluster is extracted to form the original text pool of the policy semantic cluster; Perform a text slicing operation on each descriptive text in the original text pool, and divide the long text into multiple independent text segments according to the principle of semantic integrity. Based on preset semantic recombination rules, all text fragments generated by the text slicing operation are recombined and sorted to generate structured text blocks. Each structured text block is defined as a standardized policy semantic unit, and all standardized policy semantic units constitute the standardized policy semantic unit set.

[0009] Preferably, the system further includes: The corpus primitive generation module receives a set of normalized policy semantic units from the semantic unit deconstruction module, matches and aligns each normalized policy semantic unit in the set with an external domain corpus, extracts corpus fragments from the external domain corpus that match the semantics of each normalized policy semantic unit, and encapsulates the corpus fragments into corpus primitives with a standardized structure to form a corpus primitive sequence.

[0010] Preferably, the corpus primitive generation module encapsulates corpus fragments into corpus primitives with standardized structures, including: Assign a globally unique primitive identifier to each corpus segment extracted from an external domain corpus; The internal grammatical structure of each corpus segment is analyzed to extract the main semantic components and modifying semantic components of the corpus segment; The main semantic components, the modifying semantic components, and the primitive identifiers are assembled according to a preset encapsulation template to generate a complete corpus primitive with a unified field structure. All generated complete corpus primitives are sorted according to the correlation of their corresponding normalized policy semantic units, and an ordered corpus primitive sequence is output.

[0011] Preferably, the system further includes: The feature evolution trajectory construction module receives the corpus primitive sequence, performs dynamic feature extraction on the main semantic components contained in each complete corpus primitive in the corpus primitive sequence, and connects the extracted dynamic features into a feature trajectory that evolves over time or sequentially according to the sorting order of the corpus primitive sequence, which is denoted as the feature evolution trajectory.

[0012] Preferably, the feature evolution trajectory construction module connects the extracted dynamic features into a feature evolution trajectory, including: Each complete corpus primitive in the corpus primitive sequence is read sequentially, and the main semantic component fields are parsed from its standardized structure; A vectorization representation operation is performed on the core semantic component fields to transform the textual core semantic components into fixed-dimensional numerical feature vectors. According to the original order of the corpus primitive sequence, the numerical feature vectors corresponding to all complete corpus primitives are sequentially concatenated to form a multidimensional feature matrix; The multidimensional feature matrix is ​​subjected to smooth interpolation and alignment processing in the sequence dimension to ensure that each feature dimension in the multidimensional feature matrix is ​​continuous and equally spaced in the sequence direction, and the processed result is used as the feature evolution trajectory.

[0013] Preferably, the system further includes: The corpus generation space modulation module constructs a generative corpus generation space based on the feature evolution trajectory. Within the generative corpus generation space, the feature evolution trajectory is used as a guiding path to iteratively diffuse and constrain random semantic seed points, generating a continuous corpus feature distribution that conforms to the semantic evolution law represented by the feature evolution trajectory.

[0014] Preferably, the corpus generation spatial modulation module generates a continuous corpus feature distribution, including: Initialize a generative corpus generation space containing multiple dimensions, and randomly generate a point in the generative corpus generation space as a semantic seed point; The feature evolution trajectory is imported into the generative corpus generation space and mapped as a parameterized guiding path in the generative corpus generation space. In the generative corpus generation space, starting from the semantic seed point, its state is randomly diffused within the space according to the preset diffusion rules. In each diffusion step, the geometric distance between the current semantic seed point and the parameterized guidance path is calculated, and a constraint force is generated based on the geometric distance. The constraint force is used to pull the current semantic seed point toward the parameterized guidance path. After multiple iterations of diffusion and constraint transformation, the motion trajectory of the semantic seed point in the generative corpus generation space will converge and closely surround the parameterized guiding path. The spatial region covered by the motion trajectory is recorded, and the set of feature points in the spatial region constitutes the continuous corpus feature distribution.

[0015] Preferably, the system further includes: The domain corpus synthesis module selects target feature points from the continuous corpus feature distribution according to a preset sampling strategy, decodes the target feature points into natural language text fragments, and serializes and combines all the decoded natural language text fragments according to their topological structure in the continuous corpus feature distribution to output the final domain corpus text.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By performing iterative pruning and reinforcement operations on the initial policy association network, on the one hand, isolated nodes with connectivity below a preset threshold are continuously removed, gradually eliminating marginal information and noise in the network. This allows the network structure to converge from broad associations to a core focus, avoiding interference from redundant and irrelevant information in subsequent processing. On the other hand, the connection weights of node pairs with multi-layered semantic nesting relationships are enhanced. Essentially, this transforms the implicit and complex deep logical dependencies in the text into quantifiable strong connections in the network topology. This results in an evolved, denser network that is not only more tightly structured, but whose connection strength directly reflects the logical tightness and importance between policy semantics, providing a strong semantic framework for the subsequent construction of semantic units.

[0017] In the obtained dense semantic network, semantic clusters composed of strong core nodes are identified. This essentially divides policy knowledge into groups with highly consistent internal logic based on network connection strength and topology. On this basis, the original descriptive text of all nodes within a semantic cluster is sliced ​​and recombined. This means breaking down the text boundaries based on individual policy documents or nodes, and filtering, cutting, and rearranging text fragments from different nodes according to the unified semantic theme represented by the cluster. This cross-node text operation based on semantic clusters no longer generates a simple collection of original texts, but rather a collection of semantic units that are logically coherent and structurally standardized around a specific policy theme. The text fragments within each unit have a high degree of semantic correlation guaranteed by the reinforced network, thus directly producing high-quality, standardized domain corpus blocks that can be directly used for model training or deep analysis. Attached Figure Description

[0018] Figure 1 This is a timeline diagram of the domain corpus intelligent generation system based on policy knowledge graph described in this invention. Figure 2 A flowchart illustrating the deep evolution of networks and the densification of semantic networks; Figure 3 A flowchart for corpus primitive encapsulation and serialization; Figure 4 The curves showing the change in geometric distance between the semantic seed point and the guiding path under different σ parameters during the iteration process; Figure 5 A multi-dimensional comparison chart of the effects of different corpus fragment serialization and combination methods. Detailed Implementation

[0019] 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.

[0020] Please see Figure 1 This invention provides an intelligent domain corpus generation system based on policy knowledge graphs. The system includes a policy knowledge graph parsing module, a network deep evolution module, and a semantic unit deconstruction module. The policy knowledge graph parsing module performs subgraph traversal operations on a pre-constructed policy knowledge graph based on the input target policy domain identifier, thereby extracting all policy nodes associated with the target policy domain identifier and the relationship paths between nodes, forming an initial policy association network. The network deep evolution module performs iterative pruning and strengthening operations on the initial policy association network, removing isolated policy nodes with connectivity below a preset threshold, while enhancing the connection weights between core policy nodes with multi-layer semantic nesting relationships, generating a dense policy semantic network. The semantic unit deconstruction module identifies multiple policy semantic clusters composed of core policy nodes in the dense policy semantic network, and slices and reassembles the description text of all policy nodes within each policy semantic cluster to generate a standardized policy semantic unit set corresponding to each policy semantic cluster.

[0021] Example 1: See Figure 2 The deep evolution module performs iterative pruning and reinforcement operations on the initial policy association network to generate a denser policy semantic network. This includes calculating the topological connectivity of each policy node in the initial policy association network, marking all policy nodes with topological connectivity below the threshold as isolated policy nodes to be pruned, removing all marked isolated policy nodes and their associated edges from the initial policy association network to form a pruned policy association subnetwork, identifying all node pairs containing semantic relationships exceeding a preset nesting level as core policy node pairs with multi-level semantic nesting, analyzing the interaction frequency and semantic association strength on all relational paths between core policy node pairs, and proportionally increasing the weight values ​​of the connecting edges between core policy node pairs based on the analysis results, thus completing the reinforcement processing of the pruned policy association subnetwork and outputting the denser policy semantic network.

[0022] The semantic unit deconstruction module generates a set of normalized policy semantic units corresponding to each policy semantic cluster. This includes using a community detection algorithm to identify multiple sub-network regions with tightly clustered nodes in a dense policy semantic network, defining each sub-network region as a policy semantic cluster, extracting the original descriptive text of all policy nodes belonging to each policy semantic cluster for each identified policy semantic cluster, forming the original text pool of the policy semantic cluster, performing a text slicing operation on each descriptive text in the original text pool, dividing the long text into multiple independent text segments according to the principle of semantic integrity, and recombining and sorting all text segments generated by the text slicing operation according to preset semantic recombination rules to generate structured text blocks. Each structured text block is defined as a normalized policy semantic unit, and all normalized policy semantic units constitute a set of normalized policy semantic units.

[0023] In practice, the network deep evolution module performs iterative pruning and reinforcement operations on the initial policy association network. Its input is an initial policy association network output by the policy knowledge graph parsing module, which consists of policy nodes and edges connecting them. The network deep evolution module calculates the topological connectivity of each policy node in the initial policy association network. Topological connectivity is a normalized ratio of the number of adjacent edges of a policy node to the total number of edges in the network. In some embodiments, the preset connectivity screening condition can be an absolute threshold, meaning that all policy nodes with a topological connectivity below 0.05 will be marked as isolated policy nodes to be pruned. In another comparative example, the connectivity screening condition can also be set to a dynamic threshold, such as half the average topological connectivity of all nodes, marking all policy nodes with a topological connectivity below this dynamically calculated value. The network deep evolution module then removes all marked isolated policy nodes to be pruned and their associated edges from the initial policy association network, forming a pruned policy association subnetwork. In the pruned policy association subnetwork, the number of nodes is reduced, and the network structure is more compact.

[0024] In practical implementation, the network deep evolution module identifies core policy node pairs with multi-level semantic nesting relationships in the pruned policy association sub-network. The identification process involves traversing all possible relationship paths in the pruned policy association sub-network, with a preset nesting level of 3. This means the network deep evolution module will look for node pairs indirectly connected by relationship paths, with at least two intermediate nodes along the path. Optionally, multi-level semantic nesting relationships can also be defined by the depth of semantic roles along the relationship paths. The interaction frequency and semantic association strength on all relationship paths between core policy node pairs are analyzed. Interaction frequency refers to the frequency of co-occurrence in historical policy texts, and semantic association strength can be obtained by calculating the semantic similarity of node description texts using a pre-trained language model. Based on the analysis results, the weight values ​​of the connecting edges between core policy node pairs are proportionally increased to complete the strengthening process of the pruned policy association sub-network. The weight increase formula can be expressed as: in: This represents the weight value of the strengthened connection edge. This represents the original edge weights before reinforcement. This represents the frequency of interaction between core policy nodes. Represents the semantic correlation strength between core policy nodes. and This is a preset adjustment coefficient. The final output is a dense policy semantic network, in which the connections between core policy nodes are enhanced and edge nodes are removed.

[0025] In practical implementation, the semantic unit deconstruction module applies a community detection algorithm to identify policy semantic clusters within the dense policy semantic network. This algorithm can be either the Louvain algorithm or a label propagation algorithm. The community detection algorithm identifies multiple sub-network regions with tightly clustered nodes, each defined as a policy semantic cluster. For example, in the dense policy semantic network of "High-tech Enterprise Certification," the community detection algorithm might identify two main policy semantic clusters: one containing nodes such as "R&D expenses," "R&D personnel," and "R&D equipment," organized around R&D activities; and the other containing nodes such as "Intellectual property," "patents," and "scientific and technological achievements," organized around innovation output. For each identified policy semantic cluster, the semantic unit deconstruction module extracts the original descriptive text of all policy nodes belonging to that cluster, forming the original text pool for the policy semantic cluster. The text in the original text pool originates from the attribute fields of nodes in the knowledge graph. A text slicing operation is performed on each descriptive text segment in the original text pool, dividing the long text into multiple independent text fragments based on the principle of semantic integrity. The principle of semantic integrity can be based on punctuation, sentence structure, or fixed length. In some embodiments, all text fragments generated by text slicing operations are recombined and sorted according to preset semantic restructuring rules. Semantic restructuring rules can be based on clustering and arranging text fragments according to their thematic keywords, chronological order, or logical relationships. It can be understood that the recombination and sorting generate structured text blocks, each of which is defined as a standardized policy semantic unit.

[0026] Example 2: See Figure 3 The corpus primitive generation module receives a set of standardized policy semantic units from the semantic unit deconstruction module. It matches and aligns each standardized policy semantic unit in the set with an external domain corpus. It extracts corpus fragments from the external domain corpus that semantically match each standardized policy semantic unit and encapsulates these fragments into corpus primitives with standardized structures, forming a corpus primitive sequence. The corpus primitive generation module encapsulates corpus fragments into corpus primitives with standardized structures, including assigning a globally unique primitive identifier to each corpus fragment extracted from the external domain corpus, parsing the internal grammatical structure of each corpus fragment, extracting the core semantic components and modifying semantic components, assembling the core semantic components, modifying semantic components, and primitive identifiers according to a preset encapsulation template to generate a complete corpus primitive with a unified field structure, and sorting all generated complete corpus primitives according to the association relationships of their corresponding standardized policy semantic units, outputting an ordered corpus primitive sequence.

[0027] In its implementation, the system also includes a corpus primitive generation module. This module receives a set of normalized policy semantic units from the semantic unit deconstruction module. In practice, the module matches and aligns each normalized policy semantic unit in the set with an external domain corpus, which is a database containing a large amount of domain-related text, independent of the policy knowledge graph. The module assigns a globally unique primitive identifier to each corpus fragment extracted from the external domain corpus. This primitive identifier can be a string generated by combining a timestamp, sequence number, and hash value. The module then parses the internal grammatical structure of each corpus fragment. The parsing process utilizes dependency parsing tools to extract the core semantic components and modifying semantic components of the fragment. The core semantic components typically include the core subject, predicate, and object of the sentence, while the modifying semantic components include additional information such as attributives, adverbs, and complements.

[0028] In one example, for the corpus fragment "Enterprises must manage R&D expenses in a dedicated account in accordance with the provisions of the financial accounting system," the parsed core semantic component might be "enterprises manage," and the modifying semantic components include "must comply with the provisions of the financial accounting system," "R&D expenses," and "dedicated account." In some embodiments, after extracting the core and modifying semantic components of the corpus fragment, the corpus primitive generation module assembles the core semantic components, modifying semantic components, and primitive identifiers according to a preset encapsulation template. The encapsulation template is a data structure that defines fixed fields and formats, such as JSON format, containing three required fields: id, core_semantic, and modifier. This generates a complete corpus primitive with a unified field structure. It can be understood that all generated complete corpus primitives are sorted according to the association relationship of their corresponding standardized policy semantic units. The sorting rule can be based on the logical order or importance weight of the standardized policy semantic units in the policy semantic cluster, outputting an ordered sequence of corpus primitives. For example, complete corpus primitives belonging to the standardized policy semantic unit "R&D activities" are ranked before complete corpus primitives belonging to the standardized policy semantic unit "intellectual property."

[0029] In practical implementation, the corpus primitive generation module encapsulates corpus fragments into corpus primitives with standardized structures. Its core function is to transform unstructured text into structured data units. Optional operations include further cleaning and normalization of the extracted core semantic components and modifying semantic components, such as removing stop words and performing lexical restoration. The corpus primitive generation module can employ different strategies in the matching and alignment stages. In some embodiments, the matching strategy can be based on Boolean retrieval using precise keywords, while in other comparative embodiments, the matching strategy can be entirely based on semantic similarity calculation using a deep learning model. The two strategies differ in recall and precision. The keyword-based strategy has a faster recall speed but may miss semantically related fragments that do not match the keywords, while the semantic similarity-based strategy has a more comprehensive recall but higher computational cost. When encapsulating the corpus primitive generation module, a weight calculation formula can be introduced to characterize the association strength between the corpus fragment and the source standardized policy semantic unit: in: Represents the numerical value of the correlation strength. The semantic similarity score represents the difference between a corpus segment and a standardized policy semantic unit. The score represents the degree of keyword overlap between the two. and This is a preset adjustment factor. The correlation strength value can be stored in the encapsulation template as an optional field of the complete corpus primitives. It can be understood that the final output ordered sequence of corpus primitives provides a structured, semantically aligned input data sequence for the subsequent feature evolution trajectory construction module.

[0030] Example 3: The feature evolution trajectory construction module receives a sequence of corpus primitives, performs dynamic feature extraction on the core semantic components contained in each complete corpus primitive in the sequence, and connects the extracted dynamic features into a feature trajectory that evolves over time or sequentially according to the sorting order of the corpus primitive sequence. This is denoted as the feature evolution trajectory. The feature evolution trajectory construction module connects the extracted dynamic features into a feature evolution trajectory, including sequentially reading each complete corpus primitive in the sequence, parsing the core semantic component field from its standardized structure, performing vectorization representation operation on the core semantic component field to convert the text-form core semantic component into a fixed-dimensional numerical feature vector, sequentially concatenating the numerical feature vectors corresponding to all complete corpus primitives according to the original order of the corpus primitive sequence to form a multi-dimensional feature matrix, performing smooth interpolation and alignment processing on the multi-dimensional feature matrix in the sequence dimension to ensure that each dimension of the multi-dimensional feature matrix is ​​continuous and equally spaced in the sequence direction, and using the processed result as the feature evolution trajectory.

[0031] In its implementation, the system also includes a feature evolution trajectory construction module. This module receives a sequence of corpus primitives output from the corpus primitive generation module. Specifically, it sequentially reads each complete corpus primitive from the sequence, where each complete corpus primitive is a data unit with a unified field structure. The module then parses the core semantic component field from this standardized structure. Next, it performs a vectorization representation operation on the core semantic component field, transforming the textual core semantic components into fixed-dimensional numerical feature vectors. This vectorization can utilize pre-trained language models, such as the [CLS] tag output of the BERT model, or a weighted average of word vectors to obtain sentence vectors. In the example of "High-tech Enterprise Certification," the corpus primitive sequence may contain dozens of complete corpus primitives. The core semantic component of each complete corpus primitive, after vectorization, becomes a 256-dimensional numerical feature vector.

[0032] In practical implementation, the feature evolution trajectory construction module sequentially concatenates the numerical feature vectors corresponding to all complete corpus primitives according to the original order of the corpus primitive sequence, forming a multidimensional feature matrix. For example, a sequence with 50 complete corpus primitives, each generating a 256-dimensional vector, will form a 50-row, 256-column multidimensional feature matrix after concatenation. The feature evolution trajectory construction module performs smooth interpolation and alignment processing on the multidimensional feature matrix along the sequence dimension. This smooth interpolation and alignment processing is necessary because the original corpus primitive sequence may have uneven intervals in semantic evolution. The goal is to ensure that each dimension of the multidimensional feature matrix is ​​continuous and equally spaced along the sequence direction. In some embodiments, smooth interpolation can use linear interpolation to uniformly insert new data points along the sequence dimension, so that the total number of points reaches a preset standard length N. In other comparative embodiments, smooth interpolation can use cubic spline interpolation to generate smoother curves. Alignment ensures that each row in the processed multidimensional feature matrix represents a "moment" with equal time intervals or equal logical order. The result after smoothing interpolation and alignment is used as the feature evolution trajectory. The feature evolution trajectory can be mathematically represented as a matrix. Each row represents a feature state at a given moment, and each column represents the evolution of a semantic feature.

[0033] Optional operations include dimensionality reduction of the feature vectors after vectorization, such as using principal component analysis to reduce a 256-dimensional vector to 50 dimensions, thereby reducing subsequent computational complexity. When the feature evolution trajectory construction module performs smooth interpolation, the interpolated digits... Feature values ​​of each standard location It can be calculated from the original feature sequence, and the calculation formula is: in: Representative at the Synthetic feature vectors at standard interpolation positions Representing the Numerical feature vectors corresponding to primitives in the original complete corpus. The number of primitives representing the original complete corpus. Represents the first, determined by an interpolation algorithm (such as linear or cubic spline). The original vector pair of the nth... The weight coefficients at each interpolation position. It can be understood that the final generated feature evolution trajectory is a continuous, smooth, multi-dimensional sequence that captures the dynamic evolution pattern of the core semantics extracted from the corpus primitive sequence. In some embodiments, the feature evolution trajectory can be directly used for visualization analysis, showing the shift of policy semantic focus along the corpus sequence. An optional approach is to standardize the feature evolution trajectory after generation, ensuring that the feature values ​​of each dimension are on the same scale, facilitating use by subsequent modules.

[0034] Example 4: The corpus generation space modulation module constructs a generative corpus generation space based on the feature evolution trajectory. Within the generative corpus generation space, the feature evolution trajectory is used as a guiding path to iteratively diffuse and constrain random semantic seed points, generating a continuous corpus feature distribution that conforms to the semantic evolution law represented by the feature evolution trajectory. The corpus generation space modulation module generates a continuous corpus feature distribution. This includes initializing a generative corpus generation space containing multiple dimensions, randomly generating a point in the generative corpus generation space as a semantic seed point, importing the feature evolution trajectory into the generative corpus generation space, and mapping it to a parameterized guiding path in the generative corpus generation space. In the generative corpus generation space, starting from the semantic seed point, its state is randomly diffused in the space according to a preset diffusion rule. In each diffusion step, the geometric distance between the current position of the semantic seed point and the parameterized guiding path is calculated, and a constraint force is generated based on the geometric distance. The constraint force is used to pull the current semantic seed point toward the parameterized guiding path. After multiple iterations of diffusion and constraint transformation, the motion trajectory of the semantic seed point in the generative corpus generation space will converge and closely surround the parameterized guiding path. The spatial region covered by the motion trajectory is recorded, and the set of feature points in the spatial region constitutes the continuous corpus feature distribution.

[0035] In its implementation, the system also includes a corpus generation space modulation module. This module constructs a generative corpus generation space based on the feature evolution trajectory output by the feature evolution trajectory construction module. This generative corpus generation space is a multi-dimensional vector space, with the number of dimensions matching the dimension of the feature vectors in the feature evolution trajectory. For example, if each feature vector in the feature evolution trajectory is 50-dimensional, then the generative corpus generation space is a 50-dimensional Euclidean space. In practice, the corpus generation space modulation module initializes a multi-dimensional generative corpus generation space and randomly generates a point within it as a semantic seed point. The coordinates of the semantic seed point are obtained by sampling from a standard normal distribution. The feature evolution trajectory is imported into the generative corpus generation space and mapped as a parameterized guiding path within it. The mapping process involves directly treating each row of vectors in the feature evolution trajectory matrix as a point in this space and connecting these points sequentially to form a trajectory. In the example of "high-tech enterprise certification", a feature evolution trajectory containing 100 standard moments will form a parameterized guiding path consisting of 100 consecutive points in 50-dimensional space.

[0036] In practical implementation, within the generative corpus generation space, starting from a semantic seed point, its state is randomly diffused within the space according to a preset diffusion rule. This diffusion rule is typically a stochastic process. In some embodiments, the diffusion rule can be a Gaussian random walk, where the semantic seed point adds random noise sampled from a multidimensional Gaussian distribution to its current position in each iteration. In each diffusion step, the geometric distance between the current semantic seed point's position and the parameterized guiding path is calculated. The geometric distance is defined as the minimum Euclidean distance from the current point to all points on the parameterized guiding path. A constraint force is generated based on this geometric distance, pulling the current semantic seed point toward the parameterized guiding path. The magnitude and direction of the constraint force are related to the geometric distance; one calculation method is to make the constraint force proportional to the negative gradient of the distance. After multiple iterations of diffusion and constraint transformation, the trajectory of the semantic seed point in the generative corpus generation space will converge and closely surround the parameterized guiding path. It can be understood that the spatial region covered by the recorded trajectory, and the set of feature points within that region, constitutes a continuous corpus feature distribution. The continuous corpus feature distribution is essentially a high-dimensional point cloud generated near the parameterized guidance path.

[0037] In some embodiments, the noise variance in the preset diffusion rule can be a gradually decaying scheduling parameter, allowing for greater exploration initially and enhancing convergence later. In each diffusion step, the constraint force... The calculation can be performed using the formula: in: This represents the generated constraint force vector. This represents the geometric distance between the current semantic seed point location and the parameterized guided path. D , This is a scale parameter that controls the range of the constraint force. Optionally, the parameterized guide path can also be smoothed, for example, by using a moving average filter to obtain a more regular guide shape. Different initial diffusion parameters affect the convergence speed and final coverage of the semantic seed point's trajectory. The corpus generation spatial modulation module can generate continuous corpus feature distributions with different densities by adjusting parameters. See Table 1, which shows a comparison of the convergence radii of the semantic seed point's trajectory under different initial variance parameters.

[0038] Table 1: Comparison of Convergence Radius of Semantic Seed Point Trajectories under Different Initial Variance Parameters It is understandable that the data in the table above is a simulated example, and the convergence radius measures the average distance between the motion trajectory and the parameterized guided path after the trajectory stabilizes. In specific implementation, the continuous corpus feature distribution output by the corpus generation spatial modulation module after iteration is used as the input to the domain corpus synthesis module.

[0039] See Figure 4 In the iterative diffusion-constraint process of the corpus generation spatial modulation module, the dynamic change of geometric distance intuitively reflects the convergence characteristics of the semantic seed point towards the parameterized guidance path. In the figure, the vertical axis represents the minimum Euclidean distance (geometric distance) from the current semantic seed point to the guidance path, and the horizontal axis represents the number of iterations. The three curves correspond to scenarios with initial variance parameters σ = 0.5, 1.0, and 2.0, respectively, and the dashed line indicates the convergence threshold (0.01). Specifically, the semantic seed point diffuses with random noise (following a Gaussian distribution corresponding to σ) and is simultaneously pulled towards the guidance path by constraints related to geometric distance: initially, the distance fluctuates significantly (corresponding to the exploration phase), but as iterations progress, the distance generally decreases and gradually stabilizes near the convergence threshold. The differences in the curves with different σ parameters reflect the impact of the initial variance on the convergence process: the curve with σ=0.5 (blue) converges the fastest and has the smallest distance fluctuation; the curve with σ=2.0 (green) has a higher initial distance peak and more drastic fluctuations, and the distance level after convergence is also relatively larger, consistent with the rule in Table 1 that "the larger the initial variance, the larger the convergence radius". The curve fluctuates slightly around the guiding path in the later stage of iteration, indicating that the semantic seed points have formed a trajectory closely around the guiding path, and the covered area is the high-dimensional point cloud corresponding to the continuous corpus feature distribution.

[0040] Example 5: In a specific implementation, the system further includes a domain corpus synthesis module. This module receives a continuous corpus feature distribution output from the corpus generation spatial modulation module. In this implementation, the domain corpus synthesis module selects target feature points from the continuous corpus feature distribution according to a preset sampling strategy. The continuous corpus feature distribution is a set of points in a high-dimensional space. The preset sampling strategy determines which points are selected from the point set and how many points are selected as targets for subsequent decoding. For example, the preset sampling strategy could be to perform uniform random sampling within the spatial region covered by the continuous corpus feature distribution, extracting N points. In some embodiments, the preset sampling strategy could be probability density-based sampling, sampling more target feature points in areas with higher point density. In the example of the "High-tech Enterprise Certification" domain, the continuous corpus feature distribution is a point cloud generated by the corpus generation spatial modulation module around the parameterized guidance path. The domain corpus synthesis module samples 200 target feature points from this point cloud.

[0041] In practical implementation, the domain corpus synthesis module decodes target feature points into natural language text fragments. The decoding process requires mapping high-dimensional feature vectors back to the text sequence. In some embodiments, the decoding operation is performed by a pre-trained language model decoder, such as using the decoder head of a GPT series model, taking the target feature points as conditional input and generating a word sequence through autoregressive generation. In practical implementation, sampling probability... It can be used to guide the selection of target feature points from a continuous corpus feature distribution, and its calculation can be based on the normalized distance from the target feature point to the parameterized guiding path. The formula is: in: The representation is the sampling from the feature distribution of the continuous corpus up to the first... The probability of each candidate feature point being the target feature point. Representing the The Euclidean distance from each candidate feature point to the parameterized guided path. This represents the total number of candidate feature points in the feature distribution of a continuous corpus. It is a positive coefficient that controls the degree of concentration in the sampling. Sampling according to this probability distribution can make the target feature points more likely to be distributed in the region close to the guiding path. The domain corpus synthesis module inputs the sampled target feature points into the decoding model, and each target feature point is independently decoded to generate a natural language text fragment, such as a sentence or a short paragraph.

[0042] The domain corpus synthesis module sequentially combines all decoded natural language text fragments according to their topological structure in the continuous corpus feature distribution. Topological structure refers to the relative positions and proximity relationships of these target feature points in the original high-dimensional space. It's understandable that sequential combination requires determining the order of these text fragments. One approach is to project the target feature points onto a one-dimensional principal axis and then arrange the corresponding text fragments according to the order of their projected coordinates. In practice, sequential combination based on topological structure can be performed using the minimum spanning tree algorithm. All target feature points are treated as nodes, constructing a tree structure connecting all nodes. Then, a depth-first traversal is performed starting from the root node, combining the text fragments according to the traversal order. The final domain corpus text is output, which is a coherent and structured natural language document. Optional processing includes inserting appropriate conjunctions or transitional phrases between adjacent text fragments during combination to enhance the overall fluency of the corpus. In another comparative embodiment, serialization composition can be independent of geometric topology, but instead calculate the semantic coherence score between adjacent segments based on the first and last word vectors of each text segment during decoding, and dynamically determine the optimal arrangement order based on the coherence score.

[0043] See Figure 5 In the text fragment serialization and combination stage of the domain corpus synthesis module, this figure shows the scores (out of 100) of four different combination methods across three dimensions: semantic coherence, text fluency, and structure. Specifically, the semantic coherence ranking method scored the highest in semantic coherence (close to 95), with a relatively high score in text fluency (around 92) and a structure score of around 87. The minimum spanning tree traversal method scored approximately 92, 88, and 90 in the three dimensions, respectively. The one-dimensional projection ranking method scored approximately 85, 82, and 88 in the three dimensions, respectively. The random ranking method scored significantly lower than the other methods in all aspects, with approximately 60 in semantic coherence, 55 in text fluency, and 50 in structure. This result indicates that ranking methods based on semantics or topology (such as semantic coherence ranking and minimum spanning tree traversal) outperform simple projection or random ranking methods in multi-dimensional performance, effectively improving the quality of the final domain corpus.

[0044] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0045] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A domain corpus intelligent generation system based on policy knowledge graph, characterized in that, The system includes: The policy knowledge graph parsing module, based on the input target policy domain identifier, performs a subgraph traversal operation in the pre-built policy knowledge graph to extract all policy nodes associated with the target policy domain identifier and the relationship paths between nodes, thereby forming an initial policy association network. The network deep evolution module performs iterative pruning and reinforcement operations on the initial policy association network, removing isolated policy nodes with connectivity below a preset threshold in the initial policy association network, while enhancing the connection weights between core policy nodes with multi-layer semantic nesting relationships, generating a denser policy semantic network. The semantic unit deconstruction module identifies multiple policy semantic clusters composed of the core policy nodes in the dense policy semantic network, and slices and reassembles the description text of all policy nodes within each policy semantic cluster to generate a set of standardized policy semantic units corresponding to each policy semantic cluster.

2. The domain corpus intelligent generation system based on policy knowledge graph as described in claim 1, characterized in that, The deep evolution module performs iterative pruning and reinforcement operations on the initial policy association network to generate a denser policy semantic network, including: Calculate the topological connectivity of each policy node in the initial policy association network, and mark all policy nodes whose topological connectivity is lower than the connectivity screening condition according to the preset connectivity screening condition as isolated policy nodes to be pruned; Remove all marked isolated policy nodes to be pruned and their associated edges from the initial policy association network to form a pruned policy association subnetwork; In the pruned policy association subnetwork, all node pairs containing semantic relationships exceeding a preset nesting level are identified as core policy node pairs with multi-level semantic nesting relationships; The interaction frequency and semantic association strength on all relational paths between the core policy node pairs are analyzed, and the weight values ​​of the connecting edges between the core policy node pairs are increased proportionally based on the analysis results to complete the strengthening process of the pruned policy association sub-network and output the densed policy semantic network.

3. The domain corpus intelligent generation system based on policy knowledge graph as described in claim 2, characterized in that, The semantic unit deconstruction module generates a set of normalized policy semantic units corresponding to each policy semantic cluster, including: In the dense policy semantic network, a community detection algorithm is applied to identify multiple sub-network regions with closely clustered nodes, and each sub-network region is defined as a policy semantic cluster. For each identified policy semantic cluster, the original descriptive text of all policy nodes belonging to the policy semantic cluster is extracted to form the original text pool of the policy semantic cluster; Perform a text slicing operation on each descriptive text in the original text pool, and divide the long text into multiple independent text segments according to the principle of semantic integrity. Based on preset semantic recombination rules, all text fragments generated by the text slicing operation are recombined and sorted to generate structured text blocks. Each structured text block is defined as a standardized policy semantic unit, and all standardized policy semantic units constitute the standardized policy semantic unit set.

4. The domain corpus intelligent generation system based on policy knowledge graph as described in claim 3, characterized in that, The system also includes: The corpus primitive generation module receives a set of normalized policy semantic units from the semantic unit deconstruction module, matches and aligns each normalized policy semantic unit in the set with an external domain corpus, extracts corpus fragments from the external domain corpus that match the semantics of each normalized policy semantic unit, and encapsulates the corpus fragments into corpus primitives with a standardized structure to form a corpus primitive sequence.

5. The domain corpus intelligent generation system based on policy knowledge graph as described in claim 4, characterized in that, The corpus primitive generation module encapsulates corpus fragments into corpus primitives with standardized structures, including: Assign a globally unique primitive identifier to each corpus segment extracted from an external domain corpus; The internal grammatical structure of each corpus segment is analyzed to extract the main semantic components and modifying semantic components of the corpus segment; The main semantic components, the modifying semantic components, and the primitive identifiers are assembled according to a preset encapsulation template to generate a complete corpus primitive with a unified field structure. All generated complete corpus primitives are sorted according to the correlation of their corresponding normalized policy semantic units, and an ordered corpus primitive sequence is output.

6. The domain corpus intelligent generation system based on policy knowledge graph as described in claim 5, characterized in that, The system also includes: The feature evolution trajectory construction module receives the corpus primitive sequence, performs dynamic feature extraction on the main semantic components contained in each complete corpus primitive in the corpus primitive sequence, and connects the extracted dynamic features into a feature trajectory that evolves over time or sequentially according to the sorting order of the corpus primitive sequence, which is denoted as the feature evolution trajectory.

7. The domain corpus intelligent generation system based on policy knowledge graph as described in claim 6, characterized in that, The feature evolution trajectory construction module connects the extracted dynamic features into a feature evolution trajectory, including: Each complete corpus primitive in the corpus primitive sequence is read sequentially, and the main semantic component fields are parsed from its standardized structure; A vectorization representation operation is performed on the core semantic component fields to transform the textual core semantic components into fixed-dimensional numerical feature vectors. According to the original order of the corpus primitive sequence, the numerical feature vectors corresponding to all complete corpus primitives are sequentially concatenated to form a multidimensional feature matrix; The multidimensional feature matrix is ​​subjected to smooth interpolation and alignment processing in the sequence dimension to ensure that each feature dimension in the multidimensional feature matrix is ​​continuous and equally spaced in the sequence direction, and the processed result is used as the feature evolution trajectory.

8. The domain corpus intelligent generation system based on policy knowledge graph as described in claim 7, characterized in that, The system also includes: The corpus generation space modulation module constructs a generative corpus generation space based on the feature evolution trajectory. Within the generative corpus generation space, the feature evolution trajectory is used as a guiding path to iteratively diffuse and constrain random semantic seed points, generating a continuous corpus feature distribution that conforms to the semantic evolution law represented by the feature evolution trajectory.

9. The domain corpus intelligent generation system based on policy knowledge graph as described in claim 8, characterized in that, The corpus generation spatial modulation module generates a continuous corpus feature distribution, including: Initialize a generative corpus generation space containing multiple dimensions, and randomly generate a point in the generative corpus generation space as a semantic seed point; The feature evolution trajectory is imported into the generative corpus generation space and mapped as a parameterized guiding path in the generative corpus generation space. In the generative corpus generation space, starting from the semantic seed point, its state is randomly diffused within the space according to the preset diffusion rules. In each diffusion step, the geometric distance between the current semantic seed point and the parameterized guidance path is calculated, and a constraint force is generated based on the geometric distance. The constraint force is used to pull the current semantic seed point toward the parameterized guidance path. After multiple iterations of diffusion and constraint transformation, the motion trajectory of the semantic seed point in the generative corpus generation space will converge and closely surround the parameterized guiding path. The spatial region covered by the motion trajectory is recorded, and the set of feature points in the spatial region constitutes the continuous corpus feature distribution.

10. The domain corpus intelligent generation system based on policy knowledge graph as described in claim 9, characterized in that, The system also includes: The domain corpus synthesis module selects target feature points from the continuous corpus feature distribution according to a preset sampling strategy, decodes the target feature points into natural language text fragments, and serializes and combines all the decoded natural language text fragments according to their topological structure in the continuous corpus feature distribution to output the final domain corpus text.

Images