A knowledge graph construction system based on multi-source data fusion

Abstract

The application discloses a kind of knowledge graph construction systems based on multi-source data fusion, it is related to data processing technical field, including: collaborative fusion module: the statistical dependence between different source characteristic labels and information flow characteristic are analyzed, data source association topology is constructed and weight mapping matrix is generated;Cross-source matching is carried out based on weight mapping matrix to standard semantic unit to extract fusion candidate, and reorganization is primary knowledge graph structure;Conflict control module: detect the logic conflict in primary knowledge graph structure, determine conflict source node and calculate the radial propagation order of magnitude of conflict influence in graph topology space;According to radial propagation order of magnitude, the local graph topology affected by conflict is executed adaptive range locking and state backtracking, and generates stable fusion entity;The application accurately isolates logic conflict in complex knowledge association network and blocks the cascading diffusion of error chain, while guaranteeing the global stability of knowledge graph, maximizes the retention of effective data.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a knowledge graph construction system based on multi-source data fusion. Background Technology

[0002] With the rapid development of big data analytics and artificial intelligence technologies, knowledge graphs, as a core infrastructure for semantic retrieval, knowledge reasoning, and intelligent decision-making, have evolved from extraction from single data sources to multi-source data fusion. Current technologies typically employ fusion frameworks based on statistical probability or simple weighted rules. This involves named entity recognition and relation extraction from heterogeneous data from the internet, structured databases, and industry reports, using entity alignment algorithms to resolve conflicting attributes within the semantic space. These solutions, through pre-defined data source reliability scores or majority voting mechanisms, achieve, to some extent, the extraction of commonalities and deduplication of redundant information from multiple sources, thus initially constructing a knowledge storage system capable of supporting large-scale semantic association queries.

[0003] However, existing multi-source data fusion construction techniques, when dealing with deep semantic conflicts, often focus only on point-to-point resolution at the entity attribute value level, ignoring the conflict diffusion effect caused by the strong association topological characteristics of knowledge graphs. Because of the complex logical dependencies between knowledge triples, once a logical contradiction is introduced during the fusion process, the error will permeate along the relational chain to neighboring nodes, triggering a chain reaction of quality collapse in the topological structure. Existing systems lack precise control over the propagation path and extent of damage caused by conflict in the graph topology, making it difficult to implement precise isolation and adaptive maintenance of damaged local areas after a conflict is detected. This makes it impossible to achieve stable evolution of the knowledge graph while ensuring global consistency. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a knowledge graph construction system based on multi-source data fusion to solve the problem of poor global consistency and stability of the graph due to the lack of topological cascade control over the semantic conflict diffusion effect.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a knowledge graph construction system based on multi-source data fusion, comprising:

[0008] Multi-source access module: Accesses raw multi-source data and assigns source identifiers, generates standard semantic units through entity recognition and relation extraction, and attaches source feature tags to each standard semantic unit;

[0009] Collaborative Fusion Module: Analyzes the statistical dependencies and information flow characteristics between feature tags from different sources, constructs the data source association topology and generates a weight mapping matrix; performs cross-source matching on standard semantic units based on the weight mapping matrix to extract fusion candidates, and reorganizes them into a preliminary knowledge graph structure;

[0010] Conflict control module: Detects logical conflicts in the preliminary knowledge graph structure, identifies conflict source nodes, and calculates the radial propagation magnitude of the conflict's influence within the graph topology space; based on the radial propagation magnitude, it performs adaptive range locking and state backtracking on the local graph topology affected by the conflict, generating stable fused entities;

[0011] Dynamic Evolution Module: Writes stable fused entities into the graph database to form a dynamic knowledge graph, and dynamically iterates the weight mapping matrix based on the fusion feedback.

[0012] Preferably, the method of attaching source feature tags to each standard semantic unit includes:

[0013] Real-time acquisition of raw multi-source data and parsing to extract the main text content and source metadata; unique source identifier matching and allocation based on source metadata, and association with source feature tags;

[0014] The text content is segmented at the character level, and boundary prediction and category attribution are performed using a preset entity category label set to identify entity candidates; the syntactic and semantic association features of entity candidates in the text content are analyzed to extract relation candidates and attribute candidates.

[0015] Entity candidates, relation candidates, and attribute candidates are encapsulated in a structured manner to generate standard semantic units, and unique source identifiers and source feature tags are attached to the standard semantic units as attribute fields.

[0016] Preferably, the method for constructing the data source association topology and generating the weight mapping matrix includes:

[0017] All standard semantic units are aggregated, sorted by source feature labels by time, and the co-occurrence frequency of different source identifiers on the same candidate content is calculated.

[0018] Based on co-occurrence frequency, the conditional probability that the first source identifier and the second source identifier produce the same standard semantic unit is calculated as the association probability, and the statistical dependency relationship between different source identifiers is extracted based on the association probability.

[0019] By combining statistical dependencies with the chronological order of appearance of standard semantic units, and comparing the first appearance timestamps of the same standard semantic units under different source identifiers, the source and path of information propagation can be determined, and the information flow characteristics between different source identifiers can be identified.

[0020] The unique source identifier is mapped to a topological node, and the information flow characteristics and statistical dependencies are mapped to directed connections between nodes to construct a data source association topology. The independence contribution value and original confidence of each node in the data source association topology are extracted to generate a weight mapping matrix.

[0021] Preferably, the method of performing cross-source matching on standard semantic units to extract fusion candidates includes:

[0022] Extract textual information from standard semantic units and convert it into semantic feature vectors; perform bucketing on the semantic feature vectors to construct a set of matching pairs for standard semantic units;

[0023] The cosine of the angle between the semantic feature vectors in the set of pairs to be matched is measured to determine the semantic similarity score. The weight values ​​are retrieved from the weight mapping matrix based on the unique source identifier corresponding to the standard semantic unit, and the weight values ​​are weighted and mapped with the semantic similarity score to generate a fusion decision score.

[0024] The fusion decision score is compared with a preset fusion threshold, and the pairs to be matched that exceed the preset fusion threshold are extracted as fusion candidates.

[0025] Preferably, the method for obtaining the preliminary knowledge graph structure includes:

[0026] The fusion candidates are aligned with entities, and attribute candidates are aggregated based on the fusion decision score to form fused entities with attribute fields. Topological connection edges are established between fused entities based on relation candidates, and relation paths with logical contradictions are eliminated.

[0027] By combining the merged entities as graph nodes with topological connecting edges, a preliminary knowledge graph structure is generated.

[0028] Preferably, the method for determining the conflict source node includes:

[0029] Semantic consistency checks and topological violation detection are performed on the fused entities and topological connection edges in the preliminary knowledge graph structure to identify logical conflicts; semantic consistency check violations and topological conflict points are summarized to generate a set of logical conflicts marked with logical violation information;

[0030] Based on the violation records in the logical conflict set, the original standard semantic units are traced back by retrieving the source feature tags attached to the fused entity, and the starting node involved in the logical conflict is identified as the conflict source node.

[0031] Preferably, the method for calculating the radial propagation magnitude of conflict influence within the topological space includes:

[0032] Starting from the conflict source node, a breadth-first search is used to traverse the associated nodes in the preliminary knowledge graph structure to construct a subgraph topology; the number of shortest edges from each node in the subgraph topology to the conflict source node is counted to determine the topological path distance.

[0033] The initial influence score is determined by retrieving the fusion decision score corresponding to the conflict source node; and a conflict diffusion mapping is performed on the initial influence score based on the topological path distance, which monotonically decreases with increasing distance; the initial influence score obtained after the mapping process is determined as the radial propagation magnitude of each node in the subgraph topology.

[0034] Preferably, the method for generating a stable fused entity includes:

[0035] The radial propagation magnitude of each node is compared with a preset safety threshold to divide the safe area into the damaged area; the logical connection path between the damaged area and the safe area is interrupted to lock the local map topology affected by the conflict.

[0036] Historical backup data is retrieved based on the timestamp identifier in the source feature label, and the status backtracking is performed on the conflict source nodes and damaged area nodes within the locked range;

[0037] The weight of the unique source identifier corresponding to the conflict source node in the weight mapping matrix is ​​reduced, and the backtracked local graph topology is merged with the safe region to generate a stable fused entity.

[0038] Preferably, the method of forming the dynamic knowledge graph includes:

[0039] Extract the unique identifier, backtested attribute fields, and topological connections of the stable fused entities, and convert them into a graph data object collection; initiate a graph database write operation transaction and determine the physical storage location of the stable fused entities within the storage space;

[0040] Atomic update instructions are used to write a collection of graph data objects into a graph database to complete node creation, attribute mounting, and topological connection edge establishment; an index is built for the written nodes and attributes and the write operation transaction is committed to form a dynamic knowledge graph.

[0041] Preferably, the method of dynamically iterating the weight mapping matrix based on the fusion feedback includes:

[0042] The fusion success rate, conflict trigger frequency, and state backtracking number corresponding to each source identifier are statistically collected and used as fusion feedback. The fusion feedback is compared with the original weight values ​​in the weight mapping matrix to determine the offset evaluation quantity of source reliability.

[0043] The weight values ​​in the weight mapping matrix are corrected based on the offset evaluation value, and the weights of the directed connection edges in the data source association topology are updated synchronously to achieve dynamic iteration of the weight mapping matrix.

[0044] The beneficial effects of this invention are as follows: By constructing a data source association topology based on information flow, it achieves accurate quantitative evaluation of the independent contribution of each heterogeneous data source. This effectively solves the problem of inflated reliability caused by information plagiarism or repeated citation in a multi-source environment, providing a solid foundation for generating a highly reliable weight mapping matrix and improving the objectivity and accuracy of cross-source fusion decisions. Furthermore, it introduces a radial propagation magnitude calculation mechanism to accurately depict the attenuation trajectory of conflict influence within the graph topology space, achieving adaptive range locking and state backtracking for damaged local areas. This mechanism improves upon the crude "one move affects the whole" approach, enabling precise isolation of logical conflicts and blocking the cascading spread of error chains in complex knowledge association networks, ensuring the global stability of the knowledge graph while maximizing the retention of effective data. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a schematic diagram of a knowledge graph construction system based on multi-source data fusion in this invention.

[0047] Figure 2 This is a flowchart of the process for generating standard semantic units in this invention.

[0048] Figure 3 This is a flowchart for extracting fusion candidates in this invention.

[0049] Figure 4 This is a flowchart illustrating the dynamic iteration process achieved by modifying the weight mapping matrix in this invention. Detailed Implementation

[0050] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0051] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0052] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0053] Reference Figure 1 , Figure 2 , Figure 3 and Figure 4 This is one embodiment of the present invention, which provides a knowledge graph construction system based on multi-source data fusion, including the following steps:

[0054] Methods for attaching source feature tags to each standard semantic unit include:

[0055] It acquires raw multi-source data in real time and parses and extracts the main text content and source metadata; it performs unique source identification matching and allocation based on the source metadata and associates source feature tags.

[0056] It should be noted that raw multi-source data is acquired in real time through stream processing sockets or web crawler interfaces. Regular expressions and web page tag parsing tools, such as Lxml parsers, are used to extract the main text content and source metadata from the raw multi-source data. The extracted source metadata is then matched against the known source fields in the source index database (a structured database that stores predefined known data source names, domain names, organization information, and their corresponding unique source identifiers). If a match is successful, the existing unique source identifier is retrieved; if a match fails, a new unique source identifier is generated and written, completing the unique source identifier matching and allocation. Finally, using the unique source identifier as the primary key, the corresponding source feature tags are retrieved from the source feature mapping table and associated with the raw multi-source data.

[0057] The text content is segmented at the character level, and boundary prediction and category attribution are performed using a preset entity category label set to identify entity candidates; the syntactic and semantic association features of entity candidates in the text content are analyzed to extract relation candidates and attribute candidates.

[0058] Specifically, the extracted text content is traversed sequentially using a scanning pointer, and redundant spaces and newlines are removed character by character to convert the text content into a continuous character sequence. Each character in the continuous character sequence is mapped to a fixed-dimensional numerical vector. Feature vector combinations are extracted using a context window containing the current character and its preceding and following characters. The feature vector combinations are then mapped to the probability space containing entity category label sets containing labels for people, places, and organizations through a linear transformation.

[0059] Calculate the classification score for each character corresponding to a person's name, place name, or organization name tag, and combine it with the tag transformation constraint probability, which includes the logical probability of adjacent occurrence between different tags (e.g., the entity end tag cannot appear directly before the entity start tag), to calculate the overall probability value of all possible tag combination sequences; extract the optimal tag sequence with the largest overall probability value, complete boundary prediction by identifying the start and end positions of consecutive identical tags in the optimal tag sequence, and complete the category classification determination based on the tag category, and identify and extract entity candidates in the main text content;

[0060] Logical connections representing master-subordinate dominance relationships are established among the extracted entity candidates. By scanning the order and distance of entity candidates in the original sentence layer by layer, a dependency relationship tree reflecting the hierarchical structure between words is constructed. The semantic association tags on the connections in the dependency relationship tree are logically compared with the semantic role templates storing the agent, patient, and action core (e.g., a logical framework defining the subject as the action initiator and the object as the action recipient). The pointing relationship of entity candidates as action initiators or action recipients is analyzed, and relational candidates describing the connection between entities are extracted, as well as attribute candidates describing the characteristics of entity candidates.

[0061] Entity candidates, relation candidates, and attribute candidates are encapsulated in a structured manner to generate standard semantic units, and unique source identifiers and source feature tags are attached to the standard semantic units as attribute fields.

[0062] Specifically, the identified entity candidates, relation candidates, and attribute candidates are mapped to the subject field, predicate field, and object field, respectively. Field mapping and encapsulation are performed according to the head entity-relation-tail entity triple format to generate preliminary semantic entries containing subject, predicate, and object structures, forming standard semantic units.

[0063] Retrieve the unique source identifier and the retrieved source feature tag, convert the unique source identifier and source feature tag into standardized attribute key-value pairs, and attach the unique source identifier and source feature tag as new attribute fields to the metadata list of the standard semantic unit through the attribute injection interface to ensure that the standard semantic unit has traceable source attributes.

[0064] Currently, most conventional multi-source knowledge fusion methods focus only on semantic similarity calculation between entities, often neglecting the correlation and independence of the data sources themselves. In real-world scenarios, plagiarism, citation, or collaborative fraud are common among different data sources. If fusion is based solely on literal similarity, false information can easily be misjudged as truth due to "high-frequency co-occurrence," resulting in inflated credibility. Furthermore, existing conflict resolution mechanisms typically lack quantitative assessment of error propagation, failing to achieve precise damage isolation in complex interconnected networks. Therefore, this invention analyzes the statistical dependencies and information flow between sources, constructs a data source association topology, and generates a weight mapping matrix, aiming to quantify information reliability from the source. Specifically:

[0065] Methods for constructing data source association topologies and generating weight mapping matrices include:

[0066] All standard semantic units are aggregated, sorted by source feature labels by time, and the co-occurrence frequency of different source identifiers on the same candidate content is calculated.

[0067] Specifically, all standard semantic units are retrieved and the source feature tags attached to the metadata list of the standard semantic units are extracted. All standard semantic units are sorted in ascending order according to the timestamp field in the source feature tags. Standard semantic units with the same subject, predicate, and object fields are divided into the same candidate content group using grouping and aggregation operations. Each candidate content group is traversed and the different unique source identifiers covered in the group are identified. The co-occurrence frequency of different source identifiers on the same candidate content is calculated by accumulating the number of times different unique source identifiers appear simultaneously in the same candidate content group.

[0068] Based on co-occurrence frequency, the conditional probability that the first source identifier and the second source identifier produce the same standard semantic unit is calculated as the association probability, and the statistical dependency relationship between different source identifiers is extracted based on the association probability.

[0069] Specifically, the number of standard semantic units jointly produced by the first source identifier and the second source identifier in the same candidate content group is divided by the total number of standard semantic units produced by the first source identifier. The numerical proportion of the second source identifier producing the same standard semantic unit under the premise that the first source identifier produces the same standard semantic unit is used as the association probability.

[0070] By comparing the association probability with the preset dependency judgment threshold (0.8 in the example), if the association probability exceeds the preset dependency judgment threshold, it is determined that there is a reference or co-existence relationship between the first source identifier and the second source identifier, thereby extracting the statistical dependency relationship between different source identifiers;

[0071] It should be noted that the process of setting the preset dependency judgment threshold is as follows: retrieve a multi-source data sample set, manually label the citation, reprint and independent output relationships in it; calculate the association probability between each pair of sources in the sample set, and draw the probability distribution curve; identify the trough position in the distribution curve, or select the minimum probability value that can cover more than 90% of the known citation relationships as the preset dependency judgment threshold, so as to ensure that strong dependencies can be effectively identified while eliminating the interference of accidental co-occurrence.

[0072] By combining statistical dependencies with the chronological order of appearance of standard semantic units, and comparing the first appearance timestamps of the same standard semantic units under different source identifiers, the source and path of information propagation can be determined, and the information flow characteristics between different source identifiers can be identified.

[0073] Specifically, the process involves retrieving two unique source identifiers that are in the same statistical dependency relationship, and then searching for the first occurrence timestamps of each unique source identifier within the same candidate content group. Numerical comparison is used to determine the order of these first occurrence timestamps. The unique source identifier with the earlier first occurrence timestamp is identified as the source of information dissemination, while the unique source identifier with the later first occurrence timestamp is identified as the end point of information dissemination. A logical vector is established from the source to the end point of dissemination. Combined with statistical dependency relationships, the specific path of information diffusion from the source to various levels of receivers is determined, thereby identifying the information flow characteristics between different source identifiers.

[0074] The unique source identifier is mapped to a topological node, and the information flow characteristics and statistical dependencies are mapped to directed connections between nodes to construct a data source association topology. The independence contribution value and original confidence of each node in the data source association topology are extracted to generate a weight mapping matrix.

[0075] Specifically, each unique source identifier is converted into a vertex, i.e. a topology node, within the topology space; the determined information flow characteristics are converted into directed connection lines with arrows between topology nodes; and the association probability values ​​in the statistical dependency relationship are marked as the weight values ​​of the directed connection edges to construct the data source association topology.

[0076] By summing up the total number of directed connection edges received by each topology node, the originality rate of information of the topology node after excluding interference from other topology nodes (e.g., by calculating the proportion of non-repeating standard semantic units produced by the topology node to the total output) is analyzed to determine the independence contribution value.

[0077] Specifically, by retrieving the data source association topology, if the output content of the first node is later in timestamp than that of the second node and the association probability exceeds the example of 0.85, then the contribution of the first node to that content is determined to be non-original output, and a penalty deduction is applied when accumulating the independence contribution value; conversely, for nodes that are at the source of dissemination and have a high in-degree (number of citations), a higher originality weight is assigned. This weight quantification method based on temporal topology can effectively identify and eliminate the interference of inflated credibility caused by plagiarism, secondary forwarding, or collaborative fraud from the source.

[0078] The initial trust level values ​​are extracted from the source feature labels as the original trust level; the independence contribution value and the original trust level are calculated using a weighted summation formula, and the resulting comprehensive evaluation value is used as a matrix element and mapped to a two-dimensional data table space with the unique source identifier as the row and column index dimension to generate a weighted mapping matrix representing the trust relationship between sources.

[0079] Methods for performing cross-source matching of standard semantic units to extract fusion candidates include:

[0080] Extract textual information from standard semantic units and convert it into semantic feature vectors; perform bucketing on the semantic feature vectors to construct a set of matching pairs for standard semantic units.

[0081] Specifically, textual information is extracted from the subject, predicate, and object fields of the standard semantic unit. The textual information is converted into a numerical sequence using character encoding conversion rules. The multidimensional numerical vector corresponding to each word is retrieved from the pre-stored word vector lookup table (a static data table that stores the mapping relationship between textual information and multidimensional numerical vectors) through a lookup operation. A weighted average calculation is then performed to convert the standard semantic unit into a fixed-length semantic feature vector.

[0082] Using the locality-sensitive hashing method, standard semantic units are allocated to different hash buckets based on the numerical distribution characteristics of semantic feature vectors, i.e., bucketing is performed. Then, within the same hash bucket, pairwise combination operations are performed on the standard semantic units to construct a set of matching pairs of standard semantic units.

[0083] The cosine of the angle between the semantic feature vectors in the set of pairs to be matched is measured to determine the semantic similarity score. The weight values ​​are retrieved from the weight mapping matrix based on the unique source identifier corresponding to the standard semantic unit, and the weight values ​​are weighted and mapped with the semantic similarity score to generate a fusion decision score.

[0084] Specifically, the semantic feature vectors corresponding to the two standard semantic units in the set of pairs to be matched are retrieved. The dot product of the two semantic feature vectors in the multidimensional space is calculated and divided by the product of the magnitudes of the two semantic feature vectors to obtain a value representing the cosine of the angle between the two semantic feature vectors. The cosine of the angle is determined as the semantic similarity score. Based on the unique source identifier attached to each standard semantic unit in the set of pairs to be matched, the corresponding weight value is located and retrieved from the weight mapping matrix. The weight value is multiplied with the semantic similarity score to generate a fusion decision score of semantic consistency and source reliability.

[0085] The fusion decision score is compared with a preset fusion threshold, and the pairs to be matched that exceed the preset fusion threshold are extracted as fusion candidates.

[0086] Specifically, the generated fusion decision score is retrieved, and the fusion decision score is compared with the preset fusion threshold (0.85 in the example) one by one. Standard semantic unit combinations with fusion decision scores greater than or equal to the preset fusion threshold are selected, and the selected matching pairs are used as fusion candidates.

[0087] The preset process for the fusion threshold is as follows: retrieve a set of validation samples with known attribution relationships, which includes standard semantic unit pairs manually labeled as the same entity and standard semantic unit pairs labeled as different entities; input the validation sample set into the fusion matching process, and calculate the fusion decision score for each set of validation samples using the weight mapping matrix and semantic feature vector; set candidate thresholds one by one in the numerical range (e.g., from 0.5 to 0.95, with a step size of 0.01) using a step-by-step iterative algorithm, statistically analyze the classification accuracy and recall of the validation sample set under different candidate thresholds, and calculate the comprehensive evaluation score reflecting the classification performance; select the candidate threshold corresponding to the point where the comprehensive evaluation score reaches its maximum value.

[0088] The methods for obtaining the preliminary knowledge graph structure include:

[0089] The fusion candidates are aligned with entities, and attribute candidates are aggregated based on the fusion decision score to form fusion entities with attribute fields. Topological connection edges are established between fusion entities based on relation candidates, and relation paths with logical contradictions are eliminated.

[0090] Specifically, the standard semantic unit combination in the fusion candidate is traversed, entity candidates pointing to the same objective object are merged, and attribute candidates from different standard semantic units are weighted and aggregated according to the fusion decision score. The attribute content corresponding to the data source with the highest fusion decision score is selected as the dominant value to generate a fusion entity with attribute fields.

[0091] Retrieve candidate relationships corresponding to the merged entities, establish logical connection lines representing semantic relationships between different merged entities, i.e., topological connection edges; use logical verification rules to detect topological connection edges, identify and eliminate relationship paths that violate the logical constraints of numerical ranges (e.g., the time sequence logical contradiction of a product's production date being later than its shipping date) and relationship paths with bidirectional exclusive logic (e.g., two merged entities having mutually opposing membership descriptions at the same time), thereby completing the elimination of relationship paths with logical contradictions.

[0092] By combining the merged entities as graph nodes with topological connecting edges, a preliminary knowledge graph structure is generated.

[0093] Specifically, the fused entities are mapped to central nodes in the graph structure, i.e., graph nodes. Topological connection edges are used to physically associate and logically reorganize the corresponding graph nodes to construct a preliminary knowledge graph structure that reflects the complex relationship network between entities.

[0094] By deeply fusing the topological and semantic feature vectors of data source associations, this invention achieves higher-dimensional knowledge verification. Compared to traditional methods, this design not only effectively eliminates interference from non-independent duplicate information using a weighted mapping matrix, ensuring the dominant role of high-value, original data in the fusion decision, but also improves the efficiency and accuracy of cross-source matching through bucketing and weighted cosine similarity calculation. The generated preliminary knowledge graph structure significantly improves the accuracy of entity alignment and can pre-emptively eliminate temporal and exclusive contradictions through a logical verification mechanism, laying a solid and high-quality structured foundation for radial propagation calculation and state backtracking in the subsequent conflict control module.

[0095] In existing knowledge graph construction processes, the detection of logical conflicts typically only reaches the level of simple attribute verification, lacking a systematic investigation of complex topological violations (such as logical loops and isolated nodes). More seriously, when conflicts occur, existing technologies often adopt a crude approach of "complete rejection" or "manual removal," failing to quantify the penetration depth and scope of erroneous information in the interconnected network. This makes the governance process prone to the accidental deletion of valid data and hinders automated state backtracking and reliability tuning. Therefore, this invention identifies conflict sources through dual detection of semantic consistency and topological violations, and introduces radial propagation magnitude calculation to quantify the impact of conflicts, as detailed below.

[0096] Methods for determining the source node of a conflict include:

[0097] Semantic consistency checks and topological violation detections are performed on the fused entities and topological connection edges in the preliminary knowledge graph structure to identify logical conflicts; semantic consistency check violations and topological conflict points are summarized to generate a set of logical conflicts marked with logical violation information.

[0098] Specifically, all attribute fields of the fused entity are retrieved from the preliminary knowledge graph structure. The numerical information in the attribute fields is compared with the pre-stored attribute value range logic (e.g., the age attribute value of a person must be in the positive integer range of 0 to 150). Logical verification is performed based on the time information in the attribute fields using the chronological order logic (e.g., the contract signing date must be earlier than the contract termination date). Conflict points where the attribute value exceeds the reasonable range or violates the time logic rules are identified.

[0099] Simultaneously, a topology analysis program is used to perform a depth-first traversal of the topological connection edges between graph nodes in the preliminary knowledge graph structure. By identifying whether there are loop pointing paths between graph nodes that violate the preset logic, path loop detection is performed (e.g., detecting whether there is a logical loop where A belongs to B and B belongs to A). It also searches for isolated nodes that have completely lost semantic connection with other graph nodes to perform island detection, thereby completing the topology violation detection of the overall layout rationality of the preliminary knowledge graph structure.

[0100] The system summarizes the fused entities of semantic consistency verification violations and the topological connection edges that generate topological conflict points, extracts the unique identifiers of the entities involved in the violations, the conflict type descriptions, and the associated paths, and generates a set of logical conflicts marked with logical violation information.

[0101] Based on the violation records in the logical conflict set, the original standard semantic units are traced back by retrieving the source feature tags attached to the fused entity, and the starting node involved in the logical conflict is identified as the conflict source node.

[0102] Specifically, the process involves traversing the violation records in the logical conflict set, retrieving the source feature tags attached to the fusion entity during the aggregation process based on the unique identifier of the fusion entity involved in the violation, and tracing back to the corresponding original standard semantic unit through the original identifier field in the source feature tag. By comparing the original textual descriptions of the subject, predicate, and object fields in the original standard semantic unit, the initial data input item that caused the logical conflict is located, and the graph node corresponding to the data input item in the preliminary knowledge graph structure is identified as the conflict source node.

[0103] Methods for calculating the radial propagation magnitude of conflict influence within the topological space of a graph include:

[0104] Starting from the conflict source node, a breadth-first search is used to traverse the associated nodes in the preliminary knowledge graph structure to construct a subgraph topology. The number of shortest edges from each node in the subgraph topology to the conflict source node is counted to determine the topological path distance.

[0105] Specifically, starting from the conflict source node, a layer-by-layer outward expansion search mechanism is used to traverse all directly and indirectly related nodes in the preliminary knowledge graph structure that have topological connection edges with the conflict source node through breadth-first search. The local graph structure composed of the conflict source node and all its affected related nodes is extracted to construct the subgraph topology. The number of topological connection edges traversed by each related node in the subgraph topology to return to the conflict source node is counted, and the number of topological connection edges with the smallest value is selected as the shortest edge number from the related node to the conflict source node, thereby determining the topological path distance.

[0106] The initial influence score is determined by retrieving the fusion decision score corresponding to the conflict source node; and a conflict diffusion mapping is performed on the initial influence score based on the topological path distance, which monotonically decreases with increasing distance; the initial influence score obtained after the mapping process is determined as the radial propagation magnitude of each node in the subgraph topology.

[0107] Specifically, the fusion decision score corresponding to the conflict source node in the fusion stage is retrieved from the metadata of the initial knowledge graph structure, and the value of the fusion decision score is determined as the initial influence score; the topological path distance corresponding to each associated node in the subgraph topology is retrieved, and the exponential decay formula is used to perform the power operation with the natural constant as the base and the product of the topological path distance and the preset decay coefficient (0.5 in the example) as the negative exponent. The result of the operation is multiplied by the initial influence score to ensure that the initial influence score shows a non-linear monotonically decreasing trend with the increase of the topological path distance, which is fast at first and then slow, thereby completing the conflict diffusion mapping;

[0108] The numerical results calculated after performing the conflict diffusion mapping are encapsulated and mapped one-to-one according to the unique identifier of each associated node in the subgraph topology. The mapped values ​​are determined as the radial propagation magnitude of each node in the subgraph topology, which is used to characterize the degree of weakening of the logical determinism of the conflict source node on the surrounding associated nodes.

[0109] It should be noted that the process of setting the preset attenuation coefficient is as follows: a known conflict source is manually injected into the test graph structure, and the penetration degree of conflict logic under different topological distances is observed; the conflict attenuation curve is fitted by regression analysis, and the attenuation slope required for the conflict influence to attenuate to less than 10% of the initial value outside the three-layer relationship nodes is calculated; the coefficient value that minimizes the sum of squares of the fitted residuals is selected as the preset attenuation coefficient (0.5 in the example) to balance the strictness of conflict control and the data retention rate;

[0110] It should be noted that the nonlinear decay design aims to simulate the diffusion pattern of "logic pollution" in semantic networks. Nodes closer to the source of conflict (with fewer shortest edges) are more likely to be affected by erroneous logic, increasing exponentially. As the reasoning chain (topological path) lengthens, the semantic relevance of erroneous information is rapidly diluted. Through exponential decay mapping, the "logic weakening gradient" of damaged regions can be scientifically quantified, providing a quantitative guideline for subsequent precise pruning and avoiding excessive data loss caused by the traditional "one-size-fits-all" deletion of associated nodes.

[0111] Methods for generating stable fused entities include:

[0112] The radial propagation magnitude of each node is compared with a preset safety threshold to divide the safe area into the damaged area; the logical connection path between the damaged area and the safe area is interrupted to lock the local map topology affected by the conflict.

[0113] Specifically, the radial propagation magnitude of each node in the subgraph topology is retrieved, and the radial propagation magnitude is compared one by one with a preset safety threshold (0.3 in the example). Graph nodes with radial propagation magnitudes greater than or equal to the preset safety threshold are classified as damaged regions. All topological connection edges in the subgraph topology are traversed, and the starting node identifier and ending node identifier associated with each topological connection edge are extracted. By comparing the region attributes to which the starting node and ending node belong, the connection edges that cross the region boundary are identified, that is, the connection edges where the starting node belongs to the damaged region and the ending node belongs to the safe region (or vice versa).

[0114] By retrieving the edge deletion command from the graph database or setting a logical mask identifier, the aforementioned associated edge information that crosses the region boundary is removed from the adjacency list or association matrix of the graph storage structure, thereby interrupting the semantic pointing and attribute transfer path between the damaged region and the safe region at the physical storage layer or logical view layer.

[0115] Perform topological isolation verification to ensure that the damaged region forms an independent subgraph closed loop that is completely decoupled from the main graph in terms of topological structure. This locks the local graph topology affected by the conflict and structurally prevents the conflict logic from continuing to penetrate into the safe region.

[0116] It should be noted that the process of setting the preset safety threshold is as follows: statistically analyze the node logic repair costs under different magnitudes of impact in historical conflict handling records; analyze the correlation between the radial propagation magnitude and the node attribute tampering risk; select a magnitude threshold with a risk spillover probability of less than 5% as the preset safety threshold to ensure that the designated damaged area can cover all nodes with the risk of logical collapse, while the nodes in the safe area have statistically high certainty in terms of logical consistency.

[0117] Historical backup data is retrieved based on the timestamp identifier in the source feature label, and the status backtracking is performed on the conflict source nodes and damaged area nodes within the locked range.

[0118] Specifically, the unique identifiers of the conflict source nodes and each node in the damaged area are extracted. Based on the timestamp identifier in the source feature tag attached to the damaged area node, the historical version snapshot data that is closest to and earlier than the current conflict time point is retrieved in the historical backup database (such as a snapshot database or incremental log database based on time axis storage) using a time window filtering algorithm.

[0119] Retrieve the list of attribute key-value pairs (including historical weights, historical attribute descriptions, etc.) and topology connection records (including historical pointed nodes and edge types) corresponding to each damaged node identifier from the historical version snapshot data; perform attribute overwrite operation to replace the conflict attribute values ​​of damaged nodes in the current preliminary knowledge graph with stable values ​​from the historical snapshot;

[0120] Simultaneously, topology link reconstruction is performed. Based on historical relationship records, the historical connection relationship between damaged nodes and surrounding undamaged nodes (safe zone nodes) is restored. The storage state of the graph database is updated through physical write operations. This allows for state backtracking of conflict source nodes and damaged zone nodes within the locked range. The damaged node set is then restored to its logically stable state before the conflict occurred in both the attribute and topology dimensions. This triggers a second logical consistency check after backtracking, ensuring seamless compatibility between the restored local topology and the global graph.

[0121] The weight of the unique source identifier corresponding to the conflict source node in the weight mapping matrix is ​​reduced, and the backtracked local graph topology is merged with the safe region to generate a stable fused entity.

[0122] Specifically, the unique source identifier corresponding to the conflict source node that caused the logical conflict is located. The coordinate position with the row and column index of the unique source identifier is searched in the weight mapping matrix, and the corresponding weight value is reduced (for example, the original weight value is multiplied by the weight reduction coefficient, which is 0.5 in this example). Using the topology reconstruction instruction, the local graph topology after the execution state backtracking is connected to the safe area by re-establishing topology connection edges, eliminating logical violation information and generating a stable fusion entity with consistent attributes and a robust structure.

[0123] It should be explained that the process of setting the weighting coefficient is as follows: statistically analyze the impact of conflict triggering frequency on the fluctuation of fusion quality, set the reliability recovery period, and calculate the number of successful fusions required to offset the negative evaluation of a single conflict; based on the matching result between the conflict isolation efficiency requirement and the reliability recovery rate, select a value that enables the conflict source to be quickly isolated while having the ability to automatically recover reliability over a long period as the weighting coefficient, with an example value of 0.5.

[0124] Through this conflict control design based on radial propagation magnitude, this invention achieves significant local technical effects in ensuring the quality of knowledge graphs. The design utilizes exponential decay mapping to accurately characterize the spatial distribution of conflict influence, thereby scientifically delineating damaged areas within complex graph topologies and avoiding data loss caused by excessive expansion of the defense scope. Simultaneously, by interrupting association paths and performing timestamp-based state backtracking, the system not only automatically isolates logical conflicts and blocks the cascading spread of error chains, but also achieves closed-loop management of confidence feedback through dynamic weight reduction of conflict sources. The resulting stable fused entity achieves extremely high standards in both logical consistency and topological robustness, maximizing the preservation of the global effectiveness of the knowledge graph.

[0125] The ways to form dynamic knowledge graphs include:

[0126] Extract the unique identifier, backtested attribute fields, and topological connections of the stable fused entities, and convert them into a graph data object set; initiate a graph database write operation transaction and determine the physical storage location of the stable fused entities within the storage space.

[0127] Specifically, unique identifiers, backtested attribute fields, and topological connections between stable fused entities are extracted from the stable fused entities. A graph data modeling protocol is used to map the extracted unique identifiers, attribute fields, and topological connections into node objects, attribute key-value pairs, and directed connection arcs conforming to the graph database storage format, converting them into a collection of graph data objects. A write operation transaction is initiated through the graph database control interface, requesting memory locking and disk write permissions from the graph database. The unique identifiers of the stable fused entities are obtained using data extraction commands, and the unique identifiers are input into a preset hash function for numerical calculation. The expression is as follows: ;in, Is it processed up to the first The intermediate hash value when there are 1 character. It is the hash value after processing the previous character. These represent left shift and right shift operations, respectively. The first unique identifier string The numerical value corresponding to each character; by performing shift and XOR operations on the uniquely identified string sequence, a fixed-length hexadecimal value, i.e., a hash value, is generated.

[0128] The address mapping table of the graph database's physical storage space (a structured information table that records the correspondence between logical indexes and physical storage sectors) is retrieved. The generated hash value is mapped to the logical block index range covered by the address mapping table to obtain the corresponding logical storage address. By querying the underlying allocation record of the storage engine, the logical storage address is converted into a specific offset in the disk sector or memory address space, thereby calculating and determining the physical storage location of the stable fused entity within the graph database's physical storage space.

[0129] Atomic update instructions are used to write a collection of graph data objects into a graph database to complete node creation, attribute mounting, and topological connection edge establishment; an index is built for the written nodes and attributes and the write operation transaction is committed to form a dynamic knowledge graph.

[0130] Specifically, atomic update instructions are used to write the graph data object set sequentially to the physical storage location of the graph database. Node creation operations are performed in the storage layer, and standardized attribute key-value pairs are mounted to the corresponding graph nodes. At the same time, topological connection edges between graph nodes are established according to logical pointers. Full-text indexing tools are used to build inverted indexes or B-tree indexes for the written graph node identifiers and attribute fields. Memory lock permissions are released by sending write transaction commit confirmation instructions, and the temporary data in memory is synchronized to the persistent storage space to form a dynamic knowledge graph.

[0131] The methods for dynamically iterating the weight mapping matrix based on fusion feedback include:

[0132] The fusion success rate, conflict trigger frequency, and state backtracking number corresponding to each source identifier are statistically collected and used as fusion feedback. The fusion feedback is compared with the original weight values ​​in the weight mapping matrix to determine the offset evaluation quantity of source reliability.

[0133] Specifically, the system uses the access log management interface to collect statistics on the fusion success rate (e.g., the proportion of standard semantic units produced that enter stable fusion entities), conflict trigger frequency (e.g., the number of times semantic consistency verification violations are triggered), and state backtracking frequency (e.g., the number of times associated nodes perform historical backup data restoration) of each unique source identifier during the fusion process. The statistically obtained fusion success rate, conflict trigger frequency, and state backtracking frequency are encapsulated into a multi-dimensional numerical vector as fusion feedback. The system retrieves the corresponding original weight values ​​from the weight mapping matrix and uses the numerical difference calculation method to compare the fusion feedback with the original weight values ​​to calculate the fluctuation range of the reliability score, thereby determining the offset evaluation quantity of the source reliability.

[0134] The weight values ​​in the weight mapping matrix are corrected based on the offset evaluation value, and the weights of the directed connection edges in the data source association topology are updated synchronously to achieve dynamic iteration of the weight mapping matrix.

[0135] Specifically, based on the offset evaluation value, arithmetic adjustments are performed on the corresponding weight values ​​in the weight mapping matrix. For unique source identifiers with negative offset evaluation values ​​(e.g., unique source identifiers whose conflict trigger frequency exceeds a preset threshold), weight deduction is performed, and for unique source identifiers with positive offset evaluation values, weight compensation is performed. Simultaneously, the data source association topology is retrieved, and the updated values ​​in the weight mapping matrix are mapped to the weights of the directed connection edges between nodes in the data source association topology, covering the original statistical dependency values, so as to realize the dynamic iteration of the weight mapping matrix.

[0136] In summary, this invention achieves precise quantitative evaluation of the independent contribution of each heterogeneous data source by constructing a data source association topology based on information flow. This effectively solves the problem of inflated reliability caused by information plagiarism or repeated citations in multi-source environments, providing a solid foundation for generating a highly reliable weight mapping matrix and improving the objectivity and accuracy of cross-source fusion decisions. Furthermore, it introduces a radial propagation magnitude calculation mechanism to accurately characterize the attenuation trajectory of conflict influence within the graph topology space, enabling adaptive range locking and state backtracking for damaged local areas. This mechanism improves upon the crude "one move affects the whole" approach, accurately isolating logical conflicts and blocking the cascading spread of error chains in complex knowledge association networks, ensuring the global stability of the knowledge graph while maximizing the retention of effective data.

[0137] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A knowledge graph construction system based on multi-source data fusion, characterized in that, include: Multi-source access module: Accesses raw multi-source data and assigns source identifiers, generates standard semantic units through entity recognition and relation extraction, and attaches source feature tags to each standard semantic unit; Collaborative Fusion Module: Analyzes the statistical dependencies and information flow characteristics between feature tags from different sources, constructs the data source association topology, and generates a weight mapping matrix; Cross-source matching of standard semantic units is performed based on the weight mapping matrix to extract fusion candidates, which are then reorganized into a preliminary knowledge graph structure. Conflict control module: detects logical conflicts in the preliminary knowledge graph structure, identifies conflict source nodes, and calculates the radial propagation magnitude of the conflict's influence within the graph topology space; Based on the radial propagation magnitude, adaptive range locking and state backtracking are performed on the local topology affected by the conflict to generate a stable fused entity; Dynamic Evolution Module: Writes stable fused entities into the graph database to form a dynamic knowledge graph, and dynamically iterates the weight mapping matrix based on the fusion feedback.

2. The knowledge graph construction system based on multi-source data fusion as described in claim 1, characterized in that, The methods for attaching source feature tags to each standard semantic unit include: Real-time acquisition of raw multi-source data and parsing to extract the main text content and source metadata; unique source identifier matching and allocation based on source metadata, and association with source feature tags; The text content is segmented at the character level, and boundary prediction and category attribution are performed using a preset entity category label set to identify entity candidates; the syntactic and semantic association features of entity candidates in the text content are analyzed to extract relation candidates and attribute candidates. Entity candidates, relation candidates, and attribute candidates are encapsulated in a structured manner to generate standard semantic units, and unique source identifiers and source feature tags are attached to the standard semantic units as attribute fields.

3. The knowledge graph construction system based on multi-source data fusion as described in claim 2, characterized in that, The methods for constructing the data source association topology and generating the weight mapping matrix include: All standard semantic units are aggregated, sorted by source feature labels by time, and the co-occurrence frequency of different source identifiers on the same candidate content is calculated. Based on co-occurrence frequency, the conditional probability that the first source identifier and the second source identifier produce the same standard semantic unit is calculated as the association probability, and the statistical dependency relationship between different source identifiers is extracted based on the association probability. By combining statistical dependencies with the chronological order of appearance of standard semantic units, and comparing the first appearance timestamps of the same standard semantic units under different source identifiers, the source and path of information propagation can be determined, and the information flow characteristics between different source identifiers can be identified. The unique source identifier is mapped to a topological node, and the information flow characteristics and statistical dependencies are mapped to directed connections between nodes to construct a data source association topology. The independence contribution value and original confidence of each node in the data source association topology are extracted to generate a weight mapping matrix.

4. The knowledge graph construction system based on multi-source data fusion as described in claim 3, characterized in that, The method of performing cross-source matching on standard semantic units to extract fusion candidates includes: Extract textual information from standard semantic units and convert it into semantic feature vectors; perform bucketing on the semantic feature vectors to construct a set of matching pairs for standard semantic units; The cosine of the angle between the semantic feature vectors in the set of pairs to be matched is measured to determine the semantic similarity score. The weight values ​​are retrieved from the weight mapping matrix based on the unique source identifier corresponding to the standard semantic unit, and the weight values ​​are weighted and mapped with the semantic similarity score to generate a fusion decision score. The fusion decision score is compared with a preset fusion threshold, and the pairs to be matched that exceed the preset fusion threshold are extracted as fusion candidates.

5. A knowledge graph construction system based on multi-source data fusion as described in claim 4, characterized in that, The methods for obtaining the preliminary knowledge graph structure include: The fusion candidates are aligned with entities, and attribute candidates are aggregated based on the fusion decision score to form fused entities with attribute fields. Topological connection edges are established between fused entities based on relation candidates, and relation paths with logical contradictions are eliminated. By combining the merged entities as graph nodes with topological connecting edges, a preliminary knowledge graph structure is generated.

6. The knowledge graph construction system based on multi-source data fusion as described in claim 5, characterized in that, The methods for determining the conflict source node include: Semantic consistency checks and topological violation detection are performed on the fused entities and topological connection edges in the preliminary knowledge graph structure to identify logical conflicts; semantic consistency check violations and topological conflict points are summarized to generate a set of logical conflicts marked with logical violation information; Based on the violation records in the logical conflict set, the original standard semantic units are traced back by retrieving the source feature tags attached to the fused entity, and the starting node involved in the logical conflict is identified as the conflict source node.

7. A knowledge graph construction system based on multi-source data fusion as described in claim 6, characterized in that, The methods for calculating the radial propagation magnitude of conflict influence within the topological space of the graph include: Starting from the conflict source node, a breadth-first search is used to traverse the associated nodes in the preliminary knowledge graph structure to construct a subgraph topology; the number of shortest edges from each node in the subgraph topology to the conflict source node is counted to determine the topological path distance. The initial influence score is determined by retrieving the fusion decision score corresponding to the conflict source node; and a conflict diffusion mapping is performed on the initial influence score based on the topological path distance, which monotonically decreases with increasing distance; the initial influence score obtained after the mapping process is determined as the radial propagation magnitude of each node in the subgraph topology.

8. A knowledge graph construction system based on multi-source data fusion as described in claim 7, characterized in that, The methods for generating stable fused entities include: The radial propagation magnitude of each node is compared with a preset safety threshold to divide the safe area into the damaged area; the logical connection path between the damaged area and the safe area is interrupted to lock the local map topology affected by the conflict. Historical backup data is retrieved based on the timestamp identifier in the source feature label, and the status backtracking is performed on the conflict source nodes and damaged area nodes within the locked range; The weight of the unique source identifier corresponding to the conflict source node in the weight mapping matrix is ​​reduced, and the backtracked local graph topology is merged with the safe region to generate a stable fused entity.

9. A knowledge graph construction system based on multi-source data fusion as described in claim 8, characterized in that, The methods for forming dynamic knowledge graphs include: Extract the unique identifier, backtested attribute fields, and topological connections of the stable fused entities, and convert them into a graph data object collection; initiate a graph database write operation transaction and determine the physical storage location of the stable fused entities within the storage space; Atomic update instructions are used to write a collection of graph data objects into a graph database to complete node creation, attribute mounting, and topological connection edge establishment; an index is built for the written nodes and attributes and the write operation transaction is committed to form a dynamic knowledge graph.

10. A knowledge graph construction system based on multi-source data fusion as described in claim 9, characterized in that, The method of dynamically iterating the weight mapping matrix based on fusion feedback includes: The fusion success rate, conflict trigger frequency, and state backtracking number corresponding to each source identifier are statistically collected and used as fusion feedback. The fusion feedback is compared with the original weight values ​​in the weight mapping matrix to determine the offset evaluation quantity of source reliability. The weight values ​​in the weight mapping matrix are corrected based on the offset evaluation value, and the weights of the directed connection edges in the data source association topology are updated synchronously to achieve dynamic iteration of the weight mapping matrix.

Images