An agent collaborative decision optimization method and system based on multi-source heterogeneous knowledge graphs

Abstract

The application provides an agent collaborative decision optimization method and system based on a multi-source heterogeneous knowledge graph, and relates to the technical field of knowledge graph intelligent decision-making. The application extracts a time-effect feature set from entities of a multi-source heterogeneous knowledge graph, generates a knowledge new-old evaluation matrix containing time decay characteristics and credibility weights, analyzes policy document version numbers, academic literature citation relationships and news event transmission chains to obtain a dynamic time-effect vector, labels the time-effect level of the knowledge graph nodes according to the dynamic time-effect vector, iterates through nodes with different time-effect levels to identify conflict node pairs, generates a directional retrieval rule set according to the contradiction type and the dynamic time-effect vector characteristics, then schedules three types of intelligent agents for parallel retrieval, exchanges intermediate results to optimize the strategy, aggregates the retrieval results to form a cross-platform evidence chain, and further generates a knowledge node update recommendation report. Based on the multi-source heterogeneous knowledge graph, the multi-agent collaborative retrieval and analysis can optimize the knowledge node decision and generate an update recommendation.

Description

Technical Field

[0001] This application relates to the field of knowledge graph intelligent decision-making technology, and in particular to an intelligent agent collaborative decision-making optimization method and system based on multi-source heterogeneous knowledge graphs. Background Technology

[0002] In the dynamic verification of decision-making scenarios using personal knowledge bases, the information explosion has led to a surge in the size of personal knowledge bases. Knowledge sources encompass academic journals, government websites, industry white papers, news apps, and more, in formats including text, tables, and charts. The timeliness of knowledge varies significantly: cutting-edge academic papers may be supplemented by new discoveries within six months, tax policy adjustments can render old interpretations obsolete, and market research reports change quarterly. This necessitates technology that can track knowledge timeliness in real time, accurately locate outdated literature, contradictory data, and conflicts arising from policy changes, and automate the processing of massive amounts of information to reduce human intervention and ensure the accuracy of decision-making.

[0003] Among existing solutions, verification systems based on single-domain knowledge graphs and timed update mechanisms are widely used. These systems first build a domain ontology, defining entities and relationships, and mapping knowledge to graph nodes and edges. The timed update mechanism starts at a fixed period, based on preset rules, such as setting a 5-year validity period for medical literature, comparing deviation thresholds for data from different sources of the same indicator, scanning node attributes to determine if they are expired or conflicting, and marking and notifying the user of any anomalies found.

[0004] However, this approach has significant drawbacks. Single-domain knowledge maps struggle to handle cross-domain knowledge; for example, medical research involves biological data and ethical policies, and related information is easily lost during integration. Fixed update cycles make it difficult to respond promptly to unforeseen circumstances, such as emergency policy announcements or major academic breakthroughs, leading to decision-making delays. Pre-defined rules cannot identify implicit conflicts, such as logically contradictory but unrelated conclusions from two articles, or indirect contradictions arising from policy revisions, resulting in missed or incorrect judgments and impacting the reliability of the knowledge base. Summary of the Invention

[0005] The purpose of this application is to provide an intelligent agent collaborative decision optimization method and system based on multi-source heterogeneous knowledge graphs, so as to solve the problems in the existing technology of verification systems based on single-domain knowledge graphs and timed update mechanisms, which are prone to losing related information, have difficulty responding to emergencies in a timely manner, and cannot identify implicit conflicts when processing cross-domain knowledge.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides an agent collaborative decision optimization method based on multi-source heterogeneous knowledge graphs, comprising:

[0007] Extract timeliness feature sets from entities in a multi-source heterogeneous knowledge graph, and dynamically generate a knowledge age evaluation matrix that includes time decay characteristics and credibility weights based on the timeliness feature sets.

[0008] A dynamic timeliness vector is generated by analyzing the version number of policy documents, the citation relationship of academic literature and the propagation chain of news events. The timeliness level of nodes in the multi-source heterogeneous knowledge graph is marked by the dynamic timeliness vector and the knowledge newness evaluation matrix.

[0009] Traverse the nodes with different timeliness levels in the multi-source heterogeneous knowledge graph, identify contradictory statement combinations based on the knowledge newness evaluation matrix, and output conflicting node pairs, with each conflicting node pair carrying a contradiction type label.

[0010] Based on the conflict type label and the heterogeneous data structure features contained in the dynamic timeliness vector, a set of targeted retrieval rules is generated for each conflict node pair;

[0011] For the conflict nodes, the government database agent, academic knowledge base agent, and authoritative media source agent are scheduled in parallel. Based on the targeted retrieval rule set, the government policy database, academic literature database, and news media database associated with the dynamic timeliness vector are executed in parallel and intermediate retrieval results are exchanged to optimize the retrieval strategy.

[0012] The search results from the government database agent, academic knowledge base agent, and authoritative media source agent are aggregated to generate a cross-platform evidence chain, and a knowledge node update suggestion report is generated based on the cross-platform evidence chain.

[0013] Optionally, the step of aggregating the search results from the government database agent, the academic knowledge base agent, and the authoritative media source agent to generate a cross-platform evidence chain, and generating a knowledge node update suggestion report based on the cross-platform evidence chain, includes:

[0014] The search results returned by the government database agent, the academic knowledge base agent, and the authoritative media source agent are aggregated, and the search results include data records, source tags, and time tags;

[0015] All search results are sorted chronologically according to time stamps, and data records belonging to the same entity are linked in chronological order to generate a cross-platform evidence chain;

[0016] The consistency of data records in the cross-platform evidence chain is analyzed. If all records are consistent, an update suggestion to keep the nodes unchanged is generated. If there are conflicts in the records and the record time of the government policy database is the most recent, an update suggestion to adopt the government policy value is generated. If there are conflicts in the records and the record of the academic literature database has the highest credibility, an update suggestion to adopt the academic literature value is generated. If there are conflicts in the records and the news media database provides new status information, an update suggestion to adopt the new status value is generated.

[0017] Based on the update suggestions, a knowledge node update suggestion report is generated, which includes specific update content and reasons for the update.

[0018] Optionally, the step of traversing node pairs with different timeliness levels in the multi-source heterogeneous knowledge graph, identifying contradictory statement combinations based on the knowledge novelty evaluation matrix, and outputting conflicting node pairs, wherein the conflicting node pairs carry contradiction type labels, including:

[0019] Traverse all directly related node pairs in the multi-source heterogeneous knowledge graph and filter out node pairs with different timeliness levels;

[0020] For each selected node pair, the knowledge newness and oldness evaluation value of the corresponding entity is obtained from the knowledge newness and oldness evaluation matrix, and the difference between the two obtained knowledge newness and oldness evaluation values ​​is calculated.

[0021] The difference value is compared with a preset difference threshold. If the difference value exceeds the difference threshold, the corresponding node pair is identified as a combination of contradictory statements.

[0022] Query the statement content of the entity corresponding to the node in the contradictory statement combination. If the entity attribute values ​​conflict with each other, it is a node pair labeled with a numerical contradiction type label. If the entity relationship descriptions conflict with each other, it is a node pair labeled with a relationship contradiction type label. If the entity state descriptions conflict with each other, it is a node pair labeled with a state contradiction type label.

[0023] Output the node pairs labeled with the contradiction type as conflict node pairs.

[0024] Optionally, the step of generating a targeted retrieval rule set for each pair of conflict nodes based on the heterogeneous data structure features contained in the conflict type label and the dynamic timeliness vector includes:

[0025] Based on the contradiction type labels of the conflict node pairs, a priority allocation rule is determined. The priority allocation rule stipulates that numerical contradiction type labels are preferentially assigned to the government policy database, relational contradiction type labels are preferentially assigned to the academic literature database, and state contradiction type labels are preferentially assigned to the news media database.

[0026] The dynamic timeliness vector is analyzed to extract policy version sequence features, document citation sequence features, and news dissemination sequence features;

[0027] Based on the policy version sequence characteristics, a version time window constraint is set for the retrieval operation pointing to the government policy database; based on the document citation sequence characteristics, a citation time range constraint is set for the retrieval operation pointing to the academic literature database; and based on the news dissemination sequence characteristics, a dissemination time chain constraint is set for the retrieval operation pointing to the news media database.

[0028] For each pair of conflicting nodes, a set of targeted retrieval rules is generated, which includes the target database type, time constraints, and retrieval keywords, based on the priority allocation rules and corresponding time constraints.

[0029] Optionally, the parallel scheduling of government database agents, academic knowledge base agents, and authoritative media source agents for the conflict node, based on the targeted retrieval rule set, performs parallel retrieval operations on the government policy database, academic literature database, and news media database associated with the dynamic timeliness vector, and exchanges intermediate retrieval results to optimize the retrieval strategy, including:

[0030] For each pair of conflicting nodes, the government database agent, the academic knowledge base agent, and the authoritative media source agent are activated in parallel.

[0031] The government database agent accesses the government policy database according to the rule entries in the targeted retrieval rule set and performs the retrieval using the retrieval keywords and time constraints specified in the rules.

[0032] The academic knowledge base agent accesses the academic literature database according to the rule entries in the targeted retrieval rule set and performs the retrieval using the retrieval keywords and time constraints specified in the rules.

[0033] The authoritative media source agent accesses the news media database according to the corresponding rule entries in the targeted retrieval rule set, and performs the retrieval using the retrieval keywords and time constraints specified in the rules.

[0034] During the retrieval process, the time range of the retrieval is expanded or narrowed proportionally according to the magnitude of the time change rate parameter in the dynamic timeliness vector. During the retrieval process, the document identifiers, policy article numbers or news event IDs of the retrieved documents are exchanged, and the search keywords and time range are adjusted according to the exchanged information to achieve collaborative retrieval among the government policy database, academic literature database and news media database.

[0035] Optionally, the generation of a dynamic timeliness vector from parsing policy document version numbers, academic literature citation relationships, and news event propagation chains, and the use of the dynamic timeliness vector and the knowledge freshness assessment matrix to label the timeliness levels of nodes in the multi-source heterogeneous knowledge graph, includes:

[0036] Analyze the version number of policy documents, extract the release time of the version, and calculate the reciprocal of the number of days between consecutive release times as the policy time change rate; analyze the citation relationship of academic literature, extract the publication time and number of citations of the literature, and calculate the growth rate of the number of citations per unit time as the academic time change rate; analyze the propagation chain of news events, extract the first release time and propagation path time of the event, and calculate the reciprocal of the average time delay from the first release to the reprint of the news event as the news time change rate.

[0037] The policy time change rate, academic time change rate, and news time change rate are combined to generate a dynamic timeliness vector;

[0038] Using the knowledge newness and oldness assessment values ​​corresponding to each entity in the knowledge newness and oldness assessment matrix, a high timeliness level threshold and a medium timeliness level threshold are set. Simultaneously, the weighted average of the policy time change rate, academic time change rate and news time change rate in the dynamic timeliness vector is used to proportionally adjust the high timeliness level threshold and the medium timeliness level threshold.

[0039] Based on the adjusted high timeliness level threshold and medium timeliness level threshold, the knowledge newness assessment value of each entity is divided into high timeliness level, medium timeliness level or low timeliness level, and according to the division results, the corresponding timeliness level is labeled for the nodes in the multi-source heterogeneous knowledge graph.

[0040] Optionally, the step of extracting a set of time-sensitivity features from entities in a multi-source heterogeneous knowledge graph, and dynamically generating a knowledge age evaluation matrix containing time decay characteristics and credibility weights based on the set of time-sensitivity features, includes:

[0041] Extract time stamps and source stamps from each entity in the multi-source heterogeneous knowledge graph to form a set of timeliness features;

[0042] Based on the time markers in the timeliness feature set, calculate the difference between the current system time and each time marker, and apply the exponentially decreasing function e based on the difference. (-λ·Δt) Calculate the time decay value for each entity using the exponentially decreasing function e. (-λ·Δt) In this context, λ is the attenuation coefficient, and Δt is the time difference.

[0043] Based on the source markers in the timeliness feature set, query the pre-defined correspondence table between source types and credibility weight values ​​based on historical data statistical analysis to obtain the credibility weight corresponding to each source marker;

[0044] The time decay value of each entity is multiplied by the corresponding credibility weight to obtain the knowledge old and new evaluation value of the entity. Based on the knowledge old and new evaluation values ​​of all entities, a knowledge old and new evaluation matrix is ​​generated by organizing them according to the entity identifier.

[0045] Secondly, this application provides an intelligent agent collaborative decision-making optimization system based on a multi-source heterogeneous knowledge graph, comprising:

[0046] The extraction module is used to extract a set of time-sensitive features from entities in a multi-source heterogeneous knowledge graph, and dynamically generate a knowledge oldness evaluation matrix containing time decay characteristics and credibility weights based on the set of time-sensitive features.

[0047] The parsing module is used to parse policy document version numbers, academic literature citation relationships, and news event propagation chains to generate dynamic timeliness vectors. The dynamic timeliness vectors and the knowledge newness and oldness evaluation matrix are used to label the timeliness level of nodes in the multi-source heterogeneous knowledge graph.

[0048] The identification module is used to traverse the node pairs with different timeliness levels in the multi-source heterogeneous knowledge graph, identify contradictory statement combinations based on the knowledge newness evaluation matrix, and output conflicting node pairs, wherein the conflicting node pairs carry contradiction type labels.

[0049] The first generation module is used to generate a set of targeted retrieval rules for each pair of conflict nodes based on the heterogeneous data structure features contained in the conflict type label and the dynamic timeliness vector.

[0050] The scheduling module is used to schedule government database agents, academic knowledge base agents, and authoritative media source agents in parallel for the conflict nodes. Based on the targeted retrieval rule set, it performs retrieval operations in parallel on the government policy database, academic literature database, and news media database associated with the dynamic timeliness vector and exchanges intermediate retrieval results to optimize the retrieval strategy.

[0051] The second generation module is used to aggregate the search results of the government database agent, the academic knowledge base agent, and the authoritative media source agent to generate a cross-platform evidence chain, and to generate a knowledge node update suggestion report based on the cross-platform evidence chain.

[0052] Thirdly, this application provides an electronic device, comprising:

[0053] Memory, used to store computer programs;

[0054] A processor is configured to implement the steps of the agent collaborative decision optimization method based on multi-source heterogeneous knowledge graph as described in the first aspect above when executing the computer program.

[0055] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the intelligent agent collaborative decision optimization method based on multi-source heterogeneous knowledge graphs as described in the first aspect above.

[0056] The intelligent agent collaborative decision-making optimization method based on multi-source heterogeneous knowledge graphs provided in this application extracts a set of timeliness features from entities in the multi-source heterogeneous knowledge graph, dynamically generating a knowledge age evaluation matrix that includes time decay characteristics and credibility weights; it analyzes policy document version numbers, academic literature citation relationships, and news event propagation chains to generate a dynamic timeliness vector, and combines this vector with the evaluation matrix to label the timeliness level of nodes, accurately capturing the timeliness differences of knowledge from different sources, breaking through the limitations of single-domain knowledge graphs. By traversing node pairs, it identifies conflicting node pairs carrying contradiction type labels based on the evaluation matrix, effectively solving the problems of implicit conflicts and missed or misjudged cases by preset rules. Based on the heterogeneous data structure characteristics of contradiction type labels and dynamic timeliness vectors, it generates a set of targeted retrieval rules, and schedules government database intelligent agents, academic knowledge base intelligent agents, and authoritative media source intelligent agents in parallel. According to the rules, it performs parallel retrieval of related databases and exchanges intermediate results to optimize strategies, aggregating retrieval results to form a cross-platform evidence chain, and finally generating a knowledge node update suggestion report, realizing real-time response to dynamic changes in knowledge, comprehensively improving the accuracy, timeliness, and automation level of knowledge verification in personal knowledge bases, and ensuring the reliability of decision-making basis.

[0057] Furthermore, the system aggregates search results returned by three types of intelligent agents, including data records, sources, and timestamps. These results are then sorted chronologically by timestamp, and data records of the same entity are linked by time to generate a cross-platform evidence chain. The consistency of data records within the evidence chain is analyzed, and update suggestions are generated based on different scenarios: keeping nodes unchanged, adopting government policy values, academic literature values, or new status values ​​from news media. This results in a knowledge node update suggestion report with specific content and justifications. This solution forms a complete cross-platform evidence chain through standardized aggregation, sorting, and association methods. It establishes clear update rules based on the characteristics of data from different sources, making knowledge node update suggestions more targeted, reasonable, and interpretable. This further improves the accuracy and reliability of knowledge updates, providing clear and effective operational guidance for the dynamic optimization of personal knowledge bases. Attached Figure Description

[0058] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 A flowchart illustrating an agent collaborative decision optimization method based on a multi-source heterogeneous knowledge graph, provided for an embodiment of this application;

[0060] Figure 2A schematic diagram illustrating the specific implementation process of an agent collaborative decision optimization method based on a multi-source heterogeneous knowledge graph, provided for an embodiment of this application;

[0061] Figure 3 A schematic diagram illustrating the specific implementation structure of an agent collaborative decision optimization method based on a multi-source heterogeneous knowledge graph, provided in this application embodiment;

[0062] Figure 4 This is a schematic diagram of the structure of an intelligent agent collaborative decision optimization system based on a multi-source heterogeneous knowledge graph, provided in an embodiment of this application. Detailed Implementation

[0063] In dynamic verification decision-making scenarios using personal knowledge bases, existing verification systems based on single-domain knowledge graphs and timed update mechanisms have significant shortcomings. Single-domain graphs struggle to handle cross-domain knowledge, and during integration, they easily lose information linking cross-domain knowledge such as biological data and ethical policies; fixed timed update cycles make it difficult to respond promptly to emergencies such as urgent policy releases or major academic breakthroughs, leading to decision-making delays; preset rules fail to identify implicit conflicts such as logically contradictory but not directly related literature conclusions or indirect contradictions arising from policy revisions, resulting in missed or incorrect judgments and affecting the reliability of the knowledge base.

[0064] To address the aforementioned issues, this application proposes an agent-based collaborative decision-making optimization method based on a multi-source heterogeneous knowledge graph. The core of this method involves extracting timeliness features from the multi-source heterogeneous knowledge graph to generate an evaluation matrix, combining this with dynamic timeliness vectors obtained from analyzing multiple types of information to label the timeliness levels of nodes, identifying conflicting node pairs to generate a targeted retrieval rule set, scheduling multiple agents to perform cross-retrieval, and aggregating the results to generate an evidence chain and an update suggestion report. This method adapts to complex data structures using a multi-source heterogeneous knowledge graph, captures timeliness changes using dynamic timeliness vectors and evaluation matrices, and identifies implicit conflicts through multi-agent collaborative retrieval and cross-platform evidence chains. It solves the problems of lost cross-domain knowledge associations, delayed response, and weak implicit conflict identification in existing solutions, thereby improving the accuracy and timeliness of knowledge verification.

[0065] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0066] The core of this application is to provide an agent collaborative decision optimization method based on multi-source heterogeneous knowledge graphs. A flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes:

[0067] S101. Extract a set of time-sensitive features from entities in a multi-source heterogeneous knowledge graph, and dynamically generate a knowledge old / new evaluation matrix containing time decay characteristics and credibility weights based on the set of time-sensitive features.

[0068] Optionally, step S101 may specifically include the following steps:

[0069] S1011. Extract time stamps and source stamps from each entity in the multi-source heterogeneous knowledge graph to form a timeliness feature set;

[0070] S1012. Based on the time markers in the timeliness feature set, calculate the difference between the current system time and each time marker, and apply the exponentially decreasing function e based on the difference. (-λ·Δt) Calculate the time decay value for each entity using the exponentially decreasing function e. (-λ·Δt) In this context, λ is the attenuation coefficient, and Δt is the time difference.

[0071] S1013. Based on the source markers in the timeliness feature set, query the correspondence table between source types and credibility weight values ​​pre-set based on historical data statistical analysis, and obtain the credibility weight corresponding to each source marker;

[0072] S1014. Multiply the time decay value of each entity by the corresponding credibility weight to obtain the knowledge newness assessment value of the entity, and generate the knowledge newness assessment matrix based on the knowledge newness assessment values ​​of all entities and organized by entity identifier.

[0073] In the above scheme, the timeliness feature set is a set of key information extracted from entities in a multi-source heterogeneous knowledge graph. It includes time stamps and source stamps. The time stamps record the time when the entity's information was generated or updated, while the source stamps indicate the origin of the entity's information. The knowledge age assessment matrix is ​​used to evaluate the age of knowledge entities. It includes time decay characteristics and credibility weights, with each entity's knowledge age assessment value corresponding to its identifier.

[0074] In this embodiment, step S1011 first uses an entity attribute extraction algorithm to scan each entity in the multi-source heterogeneous knowledge graph, accurately locating and extracting time-related attribute information as a time stamp, and simultaneously extracting attribute information indicating the source of the information as a source stamp. These two types of stamps together constitute the timeliness feature set. For example, in a knowledge graph containing content related to "environmental protection policies," for a specific policy entity, the algorithm extracts its publication date as October 15, 2023, which is the time stamp, while its source is the official website of the Ministry of Ecology and Environment of China, which is the source stamp. Together, these two constitute the timeliness feature set of that entity.

[0075] Secondly, in step S1012, the current system time is obtained using a time difference calculation tool, and compared with the time stamp of each entity obtained in step S1011 to accurately calculate the time difference value Δt between the two. The time unit can be set to year, month, or day according to actual needs. Next, an exponentially decreasing function e is used... (-λ·Δt) As a computational model, λ is the decay coefficient. The obtained time difference Δt is substituted into this function to calculate the time decay value of each entity. Because the value of information decreases over time, the larger the time difference Δt, the smaller the calculated time decay value. For example, if the current system time is October 15, 2025, and the timestamp of a certain entity is October 15, 2024, then Δt = 1 year. Substituting this into the exponentially decreasing function, assuming λ = 0.3, the time decay value is e. (-0.3·1) ≈0.74; If the time stamp of another entity is October 15, 2021, and Δt = 4 years, the calculated time decay value is e. (-0.3·4) ≈0.3.

[0076] Next, in step S1013, based on the source tags in the timeliness feature set, a database query technique is used to retrieve data from a pre-defined table corresponding to source types and credibility weights, based on historical data statistical analysis. This table is pre-defined, taking into account factors such as the authority and professionalism of different information sources, with each source type corresponding to a fixed credibility weight. Through this retrieval, the credibility weight corresponding to each source tag can be obtained. For example, when the source tag is "core academic journal," the corresponding credibility weight retrieved from the table is 0.9; when the source tag is "ordinary commercial website," the retrieved credibility weight is 0.4.

[0077] Finally, in step S1014, numerical multiplication is used to multiply the time decay value of each entity by its corresponding credibility weight to obtain the knowledge newness assessment value of the entity. Then, matrix construction techniques are applied to arrange the knowledge newness assessment values ​​of all entities in an ordered manner according to their unique identifiers in the knowledge graph, thereby generating a knowledge newness assessment matrix. For example, an entity with a time decay value of 0.7 and a corresponding credibility weight of 0.9 has a knowledge newness assessment value of 0.7 × 0.9 = 0.63; another entity with a time decay value of 0.3 and a credibility weight of 0.4 has a knowledge newness assessment value of 0.3 × 0.4 = 0.12. Following the identifier order of these two entities in the knowledge graph, 0.63 and 0.12 are arranged together to form a knowledge newness assessment matrix containing both entities.

[0078] In practical applications, a multi-source heterogeneous knowledge graph about "new energy vehicle technology" contains two entities: Entity A is a paper on "electric vehicle battery range technology," and Entity B is an industry report on "a certain brand's new energy vehicle charging standard." First, the timestamp of Entity A is extracted as March 2024, and the source is identified as "an authoritative automotive engineering journal." The timestamp of Entity B is extracted as May 2023, and the source is identified as "an industry news website," forming a set of timeliness features for both. Next, given that the current system time is March 2025, the time difference for Entity A is calculated to be one year, which is then substituted into the exponentially decreasing function e. (-λ·Δt) (Assuming λ is 0.2) The time decay value is 0.82; the time difference for entity B is 2 years, resulting in a time decay value of 0.67. Then, based on the source tag lookup mapping table, "authoritative automotive engineering journal" corresponds to a credibility weight of 0.9, and "industry information website" corresponds to a credibility weight of 0.6. Finally, multiplying the time decay value of entity A (0.82) by the credibility weight (0.9) yields a knowledge newness / oldness assessment value of 0.738; multiplying the time decay value of entity B (0.67) by the credibility weight (0.6) yields an assessment value of 0.402. These are then arranged according to the order of the two entities' identifiers in the knowledge graph, generating a knowledge newness / oldness assessment matrix containing [0.738, 0.402].

[0079] The overall solution of S101 described above, by extracting the time and source features of entities and combining time decay calculation and credibility weight allocation, generates a knowledge newness and oldness evaluation matrix that can comprehensively reflect the timeliness and reliability of knowledge entities. This provides a scientific and reasonable basis for subsequent timeliness level labeling and conflict identification of knowledge entities, helps to more accurately judge the newness and oldness of knowledge and credibility, and improves the effectiveness and accuracy of knowledge evaluation in personal knowledge bases.

[0080] S102. Analyze the version number of policy documents, the citation relationship of academic literature and the propagation chain of news events to generate a dynamic timeliness vector, and use the dynamic timeliness vector and the knowledge newness and oldness evaluation matrix to label the timeliness level of the nodes in the multi-source heterogeneous knowledge graph.

[0081] Optionally, step S102 may specifically include the following steps:

[0082] S1021. Analyze the version number of policy documents, extract the version release time, and calculate the reciprocal of the number of days between consecutive version release times as the policy time change rate; analyze the citation relationship of academic literature, extract the publication time and citation count of the literature, and calculate the growth rate of citation count per unit time as the academic time change rate; analyze the news event propagation chain, extract the first release time and propagation path time of the event, and calculate the reciprocal of the average time delay from the first release to the reprint of the news event as the news time change rate.

[0083] S1022. Combine the policy time change rate, academic time change rate and news time change rate to generate a dynamic timeliness vector;

[0084] S1023. Using the knowledge newness and oldness assessment values ​​corresponding to each entity in the knowledge newness and oldness assessment matrix, set a high timeliness level threshold and a medium timeliness level threshold, and simultaneously use the weighted average of the policy time change rate, academic time change rate and news time change rate in the dynamic timeliness vector to proportionally adjust the high timeliness level threshold and the medium timeliness level threshold.

[0085] S1024. Based on the adjusted high timeliness level threshold and medium timeliness level threshold, the knowledge newness assessment value of each entity is divided into high timeliness level, medium timeliness level or low timeliness level, and according to the division results, the corresponding timeliness level is labeled for the nodes in the multi-source heterogeneous knowledge graph.

[0086] In the above scheme, the dynamic timeliness vector refers to a vector that comprehensively reflects the time-related characteristics of policies, academic documents, and news events. It includes the policy time change rate, academic time change rate, and news time change rate, and can be used to adjust the timeliness level threshold. The timeliness level is a classification of the timeliness of nodes in the knowledge graph, including high, medium, and low timeliness levels, used to identify the timeliness status of nodes. The policy time change rate refers to the rate of change in the release time between consecutive versions obtained by parsing the version number of policy documents, reflecting the speed of policy updates. The academic time change rate refers to the rate of change in the publication time and citation count of academic documents over time, obtained by parsing citation relationships, reflecting the changes in the influence of academic documents. The news time change rate refers to the rate of change in the dissemination time delay obtained by parsing the news event propagation chain, reflecting the timeliness characteristics of news event dissemination. The high and medium timeliness level thresholds are critical values ​​used to classify timeliness levels and are adaptively adjusted according to the dynamic timeliness vector.

[0087] In this embodiment, firstly, in step S1021, text parsing technology is used to identify the version number of the policy document, extract the release time of each consecutive version, calculate the number of days between the release times of the next version and the previous version, and take the reciprocal of this number of days as the policy time change rate; for academic literature, the publication time and the number of citations in different time periods are extracted using citation relationship analysis tools, and the growth rate of the number of citations per unit time is calculated, which is taken as the academic time change rate; for the news event propagation chain, propagation path tracking technology is used to obtain the time of the first release of the event and the timestamps of each node in the propagation process, calculate the average time delay from the first release of the news event to its reprint, and take its reciprocal as the news time change rate.

[0088] Secondly, after obtaining the three time change rates in step S1022, a vector combination technique is used to arrange the policy time change rate, academic time change rate, and news time change rate in a fixed order to form a dynamic timeliness vector containing these three values. This dynamic timeliness vector can comprehensively reflect the time change characteristics of different types of information. For example, the obtained policy time change rate of 0.67, academic time change rate of 20, and news time change rate of -2 are combined in order to generate a dynamic timeliness vector [0.67, 20, -2].

[0089] Next, after obtaining the dynamic timeliness vector through step S1023, based on the knowledge timeliness assessment values ​​of all entities in the knowledge timeliness assessment matrix, preliminary high timeliness level thresholds and medium timeliness level thresholds are determined using statistical analysis tools. Subsequently, the weighted average of the policy timeliness change rate, academic timeliness change rate, and news timeliness change rate in the dynamic timeliness vector is calculated, and the two preliminary thresholds are adjusted proportionally according to this weighted average: if the weighted average is large, it indicates that the overall information timeliness changes rapidly, and the corresponding threshold will be increased proportionally; conversely, it will be decreased. For example, if the assessment values ​​of entities in the knowledge timeliness assessment matrix are distributed between 0.2 and 0.9, the initial high threshold is set to 0.7, the medium threshold is set to 0.4, and the dynamic timeliness vector is [0.67, 20, -2]. Assuming the weighted average of the policy, academic, and news timeliness change rates is 0.15, then the high threshold is finally adjusted to 0.7 × (1 + 0.15) = 0.785, and the medium threshold is adjusted to 0.4 × (1 + 0.15) = 0.46.

[0090] Finally, in step S1024, a grading tool is used to compare the knowledge newness assessment value of each entity in the knowledge newness assessment matrix with the adjusted high-timeliness threshold and medium-timeliness threshold, respectively. Nodes corresponding to entities with assessment values ​​greater than the high threshold are labeled as high-timeliness level, nodes with assessment values ​​between the high and medium thresholds are labeled as medium-timeliness level, and nodes with assessment values ​​less than the medium threshold are labeled as low-timeliness level. This completes the timeliness level labeling of nodes in the multi-source heterogeneous knowledge graph. For example, an entity with a knowledge newness assessment value of 0.8, which is greater than the high threshold of 0.78, is labeled as high-timeliness level; an entity with an assessment value of 0.6, which is between 0.45 and 0.78, is labeled as medium-timeliness level; and an entity with an assessment value of 0.3, which is less than 0.45, is labeled as low-timeliness level.

[0091] In practical applications, within the multi-source heterogeneous knowledge graph of "Artificial Intelligence Development," there are three entities: Entity C is the "Artificial Intelligence Industry Policy" document, Entity D is the "Machine Learning Algorithm Research" paper, and Entity E is the "Artificial Intelligence Product Launch" news. First, we analyze the version number of Entity C: V1 was released in 2022, V2 in 2024, and the current time is 2025. We extract the version release time and calculate the reciprocal of the number of days between consecutive version releases as the policy time change rate, i.e., Analyze the citation relationships of entity D. Published in 2023, cited 5 times in 2023, and cited 25 times in 2024. Extract the publication date and citation count, and calculate the growth rate of citations per unit time as the rate of change over academic time. Analyze the propagation chain of entity E. It was published on February 1, 2025. The propagation was delayed by 2 hours on February 2nd and by 1 hour on February 3rd. Extract the initial publication time and propagation path time of the event, and calculate the reciprocal of the average time delay from initial publication to being reprinted as the news time change rate. The three change rates mentioned above are combined to generate a dynamic timeliness vector [0.67, 20, -1]. Then, the evaluation values ​​of entities C, D, and E in the knowledge newness assessment matrix are 0.8, 0.7, and 0.5, respectively. An initial high threshold of 0.7 and a medium threshold of 0.4 are set. The high and medium timeliness thresholds are adjusted proportionally using the weighted average of the policy time change rate, academic time change rate, and news time change rate in the dynamic timeliness vector. After adjustment, the high threshold is 0.8 and the medium threshold is 0.5. Finally, entity C, with an evaluation value of 0.8 equal to the high threshold, is labeled as high timeliness; entity D, with an evaluation value of 0.7 between 0.5 and 0.8, is labeled as medium timeliness; and entity E, with an evaluation value of 0.5 equal to the medium threshold, is labeled as medium timeliness.

[0092] The overall solution in S102 described above analyzes policy document version numbers, academic literature citation relationships, and news event propagation chains to extract the policy time change rate, academic time change rate, and news time change rate, and combines them to generate a dynamic timeliness vector. Then, it uses a knowledge newness / oldness assessment matrix to set and adjust timeliness level thresholds, classifying knowledge newness / oldness assessment values ​​into high, medium, and low timeliness levels, thereby labeling nodes in a multi-source heterogeneous knowledge graph with their timeliness levels. This process comprehensively considers the time change characteristics of multi-source information such as policy, academic, and news, making the timeliness level classification more closely aligned with the dynamic evolution of actual information. It provides accurate timeliness criteria for identifying conflicts between nodes of different timeliness levels, effectively improving the rationality and accuracy of node timeliness labeling in the knowledge graph.

[0093] S103. Traverse the node pairs with different timeliness levels in the multi-source heterogeneous knowledge graph, identify contradictory statement combinations based on the knowledge newness evaluation matrix and output conflicting node pairs, wherein the conflicting node pairs carry contradiction type labels.

[0094] Optionally, step S103 may specifically include the following steps:

[0095] S1031. Traverse all directly related node pairs in the multi-source heterogeneous knowledge graph and filter out node pairs with different timeliness levels.

[0096] S1032. For each selected node pair, obtain the knowledge newness and oldness evaluation value of the corresponding entity from the knowledge newness and oldness evaluation matrix, and calculate the difference value between the two obtained knowledge newness and oldness evaluation values.

[0097] S1033. Compare the difference value with a preset difference threshold. If the difference value exceeds the difference threshold, identify the corresponding node pair as a contradictory statement combination.

[0098] S1034. Query the statement content of the entity corresponding to the node in the contradictory statement combination. If the entity attribute values ​​conflict with each other, it is a node pair labeled with a numerical contradiction type label. If the entity relationship descriptions conflict with each other, it is a node pair labeled with a relationship contradiction type label. If the entity state descriptions conflict with each other, it is a node pair labeled with a state contradiction type label.

[0099] S1035. Output the node pairs labeled with the contradiction type as conflict node pairs.

[0100] In the above scheme, conflicting node pairs refer to combinations of nodes in a multi-source heterogeneous knowledge graph that contain contradictory statements, carrying contradiction type labels to identify conflicts in the knowledge. Node pairs with different timeliness levels refer to pairs of directly related nodes in the knowledge graph that have different timeliness levels, and can be used as key objects for contradiction identification. The difference value refers to the difference between the old and new knowledge evaluation values ​​of the two entities in the node pair, reflecting the gap in their degree of newness. The difference threshold is a preset critical value used to determine whether a node pair is a combination of contradictory statements; exceeding this value may indicate a contradiction. Contradiction type labels include numerical contradiction type labels, relational contradiction type labels, and state contradiction type labels, used to identify conflicts in entity attribute values, relation descriptions, and state descriptions, respectively.

[0101] In this embodiment, step S1031 first uses a depth-first traversal algorithm to comprehensively scan the multi-source heterogeneous knowledge graph, recording all directly related node pairs one by one. Then, a ranking comparison tool is used to check the timeliness level of the two nodes in each node pair, filtering out node pairs with different ranking levels. For example, in the "financial investment" knowledge graph, if the node "a stock's performance report" and the node "an analyst's rating of the stock" have a reference relationship, with the former having a high timeliness level and the latter a low timeliness level, this node pair will be filtered out.

[0102] Secondly, in step S1032, for each selected node pair, matrix indexing technology is used to accurately locate and extract the knowledge newness and oldness assessment values ​​of the corresponding entities of the two nodes from the knowledge newness and oldness assessment matrix. The absolute difference between the two assessment values ​​is calculated using a subtraction operation tool to obtain the difference value. For example, in the above node pair, the knowledge newness and oldness assessment value of "a stock's performance report" is 0.85, and the assessment value of "an analyst's rating of the stock" is 0.35, resulting in a difference value of 0.5.

[0103] Next, after obtaining the difference value in step S1033, a comparison tool is used to compare the difference value of each node pair with a preset difference threshold in the system based on a large amount of historical data and domain characteristics. If the difference value is greater than the difference threshold, it is determined that the node pair may be contradictory and is identified as a contradictory statement combination. For example, if the preset difference threshold is 0.3, the difference value of the above node pair, 0.5, is greater than 0.3, and therefore it is identified as a contradictory statement combination.

[0104] Then, in step S1034, natural language processing technology is used to analyze the entity statements corresponding to the two nodes in the combination, extracting key information such as attribute values, relationship descriptions, and state descriptions. If the attribute values ​​of the entities are found to be contradictory, the node pair is labeled with a numerical contradiction type; if the relationship descriptions between the entities are completely opposite, a relationship contradiction type label is added; if the state descriptions of the entities are inconsistent, a state contradiction type label is added. For example, in the above contradictory statement combination, "a stock's performance report" states "the company's net profit increased by 10%", while "an analyst's rating of the stock" states "the company's net profit decreased by 5%", indicating a conflict in attribute values, therefore a numerical contradiction type label is added.

[0105] Finally, in step S1035, the node pairs labeled with conflict type are organized using a data packaging tool to form a standardized conflict node pair data format. This format is then output through the output interface, providing a clear target for subsequent conflict resolution. For example, the node pairs labeled with numerical conflict type, such as "A Stock's Performance Report" and "An Analyst's Rating on the Stock," are output as conflict node pairs.

[0106] In practical applications, within the "New Energy Technology" knowledge graph, a depth-first traversal algorithm was used to identify a direct correlation between the node "Energy Density of a New Type of Battery" (high timeliness level) and the node "Range Calculation of a Certain Electric Vehicle" (medium timeliness level). Since their timeliness levels differ, these nodes were selected. Matrix indexing technology was used to extract the former's evaluation value (0.9) and the latter's (0.4) from the knowledge newness evaluation matrix. Subtraction yielded a difference of 0.5. This difference was then compared to a preset threshold of 0.3. Since 0.5 > 0.3, it was identified as a contradictory statement combination. Natural language processing was then used to analyze the content, revealing a conflict between the stated "Energy Density of a New Type of Battery" ("500Wh per kilogram") and the stated "Range Calculation of a Certain Electric Vehicle" ("300Wh per kilogram"). This attribute value conflict was labeled as numerical contradiction. Finally, after data packaging, this node pair was output as a conflicting node pair.

[0107] The overall solution of S103 described above, by focusing on node pairs with different timeliness levels, combining the judgment of differences in the evaluation values ​​of knowledge newness and the analysis of contradictions in entity content, can accurately identify conflicting node pairs in the knowledge graph and clarify the conflict type. This provides a clear goal for subsequent targeted resolution of knowledge conflicts, helps to improve the accuracy and consistency of knowledge in the knowledge graph, and reduces decision-making errors caused by information conflicts.

[0108] S104. Based on the heterogeneous data structure features contained in the contradiction type label and the dynamic timeliness vector, generate a set of targeted retrieval rules for each pair of conflict nodes.

[0109] Optionally, step S104 may specifically include the following steps:

[0110] S1041. Based on the contradiction type labels of the conflict node pairs, determine the priority allocation rules. The priority allocation rules stipulate that numerical contradiction type labels should be prioritized to the government policy database, relational contradiction type labels should be prioritized to the academic literature database, and state contradiction type labels should be prioritized to the news media database.

[0111] S1042. Analyze the dynamic timeliness vector and extract policy version sequence features, document citation sequence features, and news dissemination sequence features;

[0112] S1043. Based on the policy version sequence characteristics, a version time window constraint is set for the retrieval operation pointing to the government policy database; based on the document citation sequence characteristics, a citation time range constraint is set for the retrieval operation pointing to the academic literature database; and based on the news dissemination sequence characteristics, a dissemination time chain constraint is set for the retrieval operation pointing to the news media database.

[0113] S1044. For each pair of conflicting nodes, based on the priority allocation rules and corresponding time constraints, generate a set of targeted retrieval rules that includes the target database type, time constraints, and retrieval keywords.

[0114] In the above scheme, the targeted retrieval rule set is a combination of retrieval rules generated for each conflict node, including the target database type, time constraints, and search keywords, used to guide the agent to perform precise retrieval. The priority allocation rule determines the priority of the search database based on the conflict type label, specifying that different conflict types preferentially target government policy databases, academic literature databases, or news media databases. The policy version sequence feature is the sequence feature of the policy version release time parsed from the dynamic timeliness vector, which can be used to set the time window for policy database retrieval. The document citation sequence feature is the sequence feature of the academic document citation time parsed, used to set the time range for document database retrieval. The news dissemination sequence feature is the sequence feature of the news event dissemination time parsed, used to set the dissemination time chain for media database retrieval. Time constraints include version time window constraints, citation time range constraints, and dissemination time chain constraints, used to limit the retrieval time range for different databases.

[0115] In this embodiment, the conflict type label carried by the conflict node pair is first obtained through step S1041. Then, a rule matching algorithm is used to compare the label with a preset priority allocation rule. The priority allocation rule explicitly stipulates that the retrieval priority corresponding to the numerical conflict type label is the highest in the government policy database, the retrieval priority corresponding to the relational conflict type label is the highest in the academic literature database, and the retrieval priority corresponding to the state conflict type label is the highest in the news media database. The retrieval priority order of the conflict node pair is determined through this comparison. For example, when the conflict node pair carries a state conflict type label, after rule matching, its retrieval priority order is determined to be the news media database, the government policy database, and the academic literature database.

[0116] Secondly, in step S1042, the dynamic timeliness vector is analyzed using a sequence feature extraction algorithm. Time series information related to policies is separated from the dynamic timeliness vector, and the release time points of policy versions are extracted and arranged in chronological order to form policy version sequence features. At the same time, the citation time series information of academic literature is separated, and the time points when the literature is cited are extracted and organized into a sequence as literature citation sequence features. In addition, the dissemination time series information of news events is separated, and the key time nodes in the dissemination process of the event are extracted and formed into a sequence to obtain news dissemination sequence features. For example, the dynamic timeliness vector contains information about policies released in 2020, 2022, and 2024, which, after extraction, forms a policy version sequence feature [2020, 2022, 2024]; documents were cited in 2021, 2022, and 2023, forming a document citation sequence feature [2021, 2022, 2023]; and news was disseminated on May 1, 3, and 5, 2024, forming a news dissemination sequence feature [2024-05-01, 2024-05-03, 2024-05-05].

[0117] Next, after obtaining the sequence features in step S1043, for the government policy database, based on the policy version sequence features, a time window generation tool is used to extend the time range to include a certain number of past versions, using the latest version release time as the benchmark, as the version time window constraint; for the academic literature database, based on the literature citation sequence features, a time range delineation tool is used to select a relatively recent continuous time segment in the citation sequence as the citation time range constraint; for the news media database, based on the news dissemination sequence features, a dissemination chain time definition tool is used to define the range from the first release time of the event to the last dissemination node time as the dissemination time chain constraint. For example, based on the policy version sequence characteristics [2020, 2022, 2024], the version time window constraint is set to 2022-2024; based on the document citation sequence characteristics [2021, 2022, 2023], the citation time range constraint is set to 2022-2023; based on the news dissemination sequence characteristics [2024-05-01, 2024-05-03, 2024-05-05], the dissemination time chain constraint is set to May 1-5, 2024.

[0118] Finally, in step S1044, for each conflicting node pair, a rule integration tool is used to integrate the target database types sorted by priority, the time constraints corresponding to each database, and the core terms extracted from the entity descriptions of the conflicting node pairs as search keywords, forming a targeted search rule set. For example, if the priority order of a conflicting node pair is academic literature database, government policy database, and news media database, with corresponding time constraints of 2022-2023, 2022-2024, and May 1-5, 2024, and the search keyword is "application scope of a certain technology", the integrated targeted search rule set will include these contents.

[0119] In practical applications, within the "Corporate Tax Policy" knowledge graph, a conflict node pair carries a label indicating a conflicting relationship between two tax policies. Based on the rules, its retrieval priority is determined to be highest in the academic literature database. Subsequently, the dynamic timeliness vector is analyzed, extracting policy version sequence characteristics (released in 2021, 2023, and 2024), document citation sequence characteristics (multiple citations between 2022 and 2024), and news dissemination sequence characteristics (dissemination between October 1st and 5th, 2024). Then, based on the document citation sequence characteristics, the citation time range for the academic literature database is constrained to 2022-2024; based on the policy version sequence characteristics, the version time window for the government policy database is constrained to 2023-2024; and based on the news dissemination sequence characteristics, the dissemination time chain for the news media database is constrained to October 1st-5th, 2024. Finally, combining priority and time constraints, a set of targeted search rules is generated: the target database types include (academic literature database, government policy database, news media database), the time constraints include (literature database 2022-2024, policy database 2023-2024, media database October 1-5, 2024), and the search keyword is "the inclusion relationship between the two tax policies".

[0120] The overall solution of S104 above generates a targeted retrieval rule set based on the contradiction type and dynamic time-sensitivity vector characteristics of the conflict node pair. It clarifies the target database, time range and keywords for retrieval, which can guide the agent to conduct more accurate and efficient retrieval, avoid the waste of resources caused by blind retrieval, lay the foundation for quickly obtaining effective conflict resolution evidence, and improve the pertinence and efficiency of knowledge conflict resolution.

[0121] S105. For the conflict node, the government database agent, academic knowledge base agent and authoritative media source agent are scheduled in parallel. According to the targeted retrieval rule set, the government policy database, academic literature database and news media database associated with the dynamic timeliness vector are executed in parallel and intermediate retrieval results are exchanged to optimize the retrieval strategy.

[0122] Optionally, step S105 may specifically include the following steps:

[0123] S1051. For each pair of conflicting nodes, activate the government database agent, the academic knowledge base agent, and the authoritative media source agent in parallel.

[0124] S1052. The government database agent accesses the government policy database according to the rule entries in the targeted retrieval rule set and performs a retrieval using the retrieval keywords and time constraints specified in the rules.

[0125] S1053. The academic knowledge base agent accesses the academic literature database according to the rule entries in the targeted retrieval rule set and performs a retrieval using the retrieval keywords and time constraints specified in the rules.

[0126] S1054. The authoritative media source intelligent agent accesses the news media database according to the corresponding rule entries in the targeted retrieval rule set, and performs a retrieval using the retrieval keywords and time constraints specified in the rules.

[0127] S1055. During the retrieval process, the time range of the retrieval is expanded or narrowed proportionally according to the magnitude of the time change rate parameter in the dynamic timeliness vector. During the retrieval process, the document identifiers, policy article numbers or news event IDs that have been retrieved are exchanged. The retrieval keywords and time range are adjusted according to the exchanged information to achieve collaborative retrieval among the government policy database, academic literature database and news media database.

[0128] In the above scheme, the government database agent is an intelligent program used to access the government policy database and perform retrieval operations, capable of querying policy information according to rules. The academic knowledge base agent is an intelligent program used to access the academic literature database and can retrieve academic literature according to rules. The authoritative media source agent is an intelligent program used to access the news media database and can retrieve news information according to rules. Cross-referencing refers to the exchange of intermediate results between different agents during the retrieval process, enabling cross-referencing queries of information from multiple databases. The time change rate parameter is a parameter in the dynamic timeliness vector that reflects the rate of time change and can be used to dynamically adjust the time range of the retrieval. Intermediate results are partial retrieval information obtained by the agents during the retrieval process, used for mutual exchange to assist in the retrieval of other databases.

[0129] In this embodiment, step S1051 first sends a synchronization activation command for each conflicting node pair using multi-agent collaborative startup technology. The command carries the identification information of the conflicting node pair and simultaneously activates the government database agent, the academic knowledge base agent, and the authoritative media source agent, putting them into a running state to prepare for receiving retrieval rules and executing tasks. For example, for the conflicting node pair of "a certain product quality standard," the activation command carries the identifier of the node pair, and the three agents immediately start upon receiving it, waiting for the next operation.

[0130] Next, steps 1052, 1053, and 1054 are executed in parallel. Specifically, the government database agent obtains rule entries for the government policy database through a rule receiving interface, logs into the policy database after secure authentication, and uses keyword matching and time filtering technologies to filter policy documents that meet the criteria, such as searching by "product quality testing indicators" and constraints from 2023-2024. The academic knowledge base agent obtains rule entries for the academic literature database through the same interface, connects to the literature database using a dedicated protocol, and uses topic matching and time-limited technologies to search for relevant literature, such as searching by "research on product quality defects" and constraints from 2022-2023. The authoritative media source agent obtains rule entries for the news media database, accesses the media database through an API interface, and uses word segmentation matching and publication time filtering technologies to search for news, such as searching by "reports on product quality problems" and constraints from January to June 2024.

[0131] During steps 1052, 1053, and 1054, each agent establishes a connection based on a real-time data exchange protocol, extracts a time change rate parameter from the dynamic timeliness vector, and expands or narrows the search time range proportionally based on the magnitude of this parameter. Simultaneously, agents exchange retrieved document identifiers, policy article numbers, or news event IDs in real time. After receiving information from other agents, each agent performs in-depth analysis: if the received document identifier, policy article number, or news event ID is related to its current search content, it extracts key keywords, time clues, and other elements from this information. For example, if a government database agent discovers an academic document identifier associated with a policy provision number, it will extract the core research keywords from the identifier and add them to its own search keywords. Simultaneously, based on the implicit time information in the identifier, such as the publication year, it will reassess and adjust the search time range. Similarly, if an academic knowledge base agent discovers a news event ID cited in a document, it will extract the key keywords and time points of the news event, add them to its own search keywords, and adjust the search time range accordingly. This enables collaborative retrieval between the government policy database, academic document database, and news media database. For instance, if a government database agent discovers an academic document identifier associated with a policy provision number, it sends it to the academic knowledge base agent, which then optimizes its search keywords and time range to obtain more accurate documents. If the academic knowledge base agent discovers a news event ID cited in a document, it passes it to the authoritative media source agent, helping it efficiently retrieve the corresponding news.

[0132] In practical applications, when searching for conflicting nodes in the "new energy vehicle subsidy policy," the government database agent, academic knowledge base agent, and authoritative media source agent are activated in parallel for this conflicting node pair. The government database agent, based on the corresponding rule entries in the targeted retrieval rule set, uses "new energy vehicle subsidy standards" as the search keyword and, under the time constraint of 2023-2024, accesses the government policy database to perform a search and obtain partial subsidy policy information. The academic knowledge base agent, based on the corresponding rule entries in the targeted retrieval rule set, uses "research on the effectiveness of new energy vehicle subsidies" as the search keyword and, under the time constraint of 2022-2023, searches the academic literature database to obtain relevant research literature. The authoritative media source agent, based on the corresponding rule entries in the targeted retrieval rule set, uses "implementation status of new energy vehicle subsidies" as the search keyword and, under the time constraint of March-April 2024, searches the news media database to obtain relevant news reports. During the retrieval process, due to the rapid policy updates indicated by the time change rate parameter of the dynamic timeliness vector, each agent proportionally expands the retrieval time range and exchanges the retrieved policy article numbers, document identifiers, and news event IDs. The government database agent sends the newly discovered policy intermediate results from February 2024 to other agents, and the academic and media agents adjust their retrieval keywords and time range accordingly, thereby achieving collaborative retrieval among the government policy database, academic document database, and news media database.

[0133] The overall scheme of S105 described above, through parallel scheduling of government database agents, academic knowledge base agents, and authoritative media source agents, performs retrieval operations on government policy databases, academic literature databases, and news media databases associated with dynamic timeliness vectors, based on a targeted retrieval rule set. During the retrieval process, each agent exchanges intermediate retrieval results in real time, dynamically optimizing the retrieval strategy accordingly. This effectively avoids the limitations of single-database retrieval, enabling rapid and comprehensive acquisition of multi-source information. The retrieved information is not only rich and diverse but also closely related, laying a solid and interconnected information foundation for subsequent aggregation of evidence chains, significantly improving the efficiency and comprehensiveness of information retrieval.

[0134] S106. Aggregate the search results from the government database agent, academic knowledge base agent, and authoritative media source agent to generate a cross-platform evidence chain, and generate a knowledge node update suggestion report based on the cross-platform evidence chain.

[0135] Optionally, step S106 may specifically include the following steps:

[0136] S1061. Aggregate the search results returned by the government database agent, the academic knowledge base agent, and the authoritative media source agent, wherein the search results include data records, source tags, and time tags;

[0137] S1062. Sort all search results in chronological order according to time stamps, associate data records belonging to the same entity in chronological order, and generate a cross-platform evidence chain.

[0138] S1063. Analyze the consistency of data records in the cross-platform evidence chain. If all records are consistent, generate an update suggestion to keep the nodes unchanged. If there are conflicts in the records and the record time of the government policy database is the latest, generate an update suggestion to adopt the government policy value. If there are conflicts in the records and the record of the academic literature database has the highest credibility, generate an update suggestion to adopt the academic literature value. If there are conflicts in the records and the news media database provides new status information, generate an update suggestion to adopt the new status value.

[0139] S1064. Generate a knowledge node update suggestion report containing specific update content and reasons for the update based on the update suggestion.

[0140] In the above scheme, the cross-platform evidence chain is a coherent sequence of evidence aggregating results retrieved by multiple agents from different databases. It includes data records, source markers, and time stamps, used to support update recommendations for knowledge nodes. Data records are the specific information content retrieved; source markers indicate which database the information originated from; and time stamps record the time the information was generated or published. Update recommendations are suggestions for modifying knowledge nodes based on the cross-platform evidence chain analysis results, including maintaining the original state, adopting government policy values, academic literature values, or new state values. The knowledge node update recommendation report is a document containing specific update content and reasons for the update, used to guide the update operation of knowledge nodes.

[0141] In the embodiments of this application, such as Figure 2 As shown, firstly, in step S1061, distributed data aggregation technology is used to receive all search results returned by the government database agent, academic knowledge base agent, and authoritative media source agent through a unified data interface. Each result is then standardized to ensure consistency in data records, source tags, and time tags. These standardized results are then stored in a temporary database to form an aggregated dataset. For example, the source tag "a certain material hardness standard is HRC50" returned by the government agent corresponds to "Government Policy Database" and the time tag is 2024-03; the source tag "a certain material hardness standard is HRC50" returned by the academic agent corresponds to "Academic Literature Database" and the time tag is 2023-11; and the source tag "a certain material hardness standard is HRC50" returned by the media agent corresponds to "News Media Database" and the time tag is 2024-02. After standardization, these results are aggregated into the temporary database.

[0142] Secondly, after obtaining the aggregated dataset in step S1062, a time series sorting algorithm is used to sort all results in the aggregated dataset in ascending order by time stamp. Then, entity recognition and association technology is used to extract entity keywords from each result. Data records with the same entity keywords are linked together in chronological order to form a coherent cross-platform evidence chain. Each record in the chain retains its source and time stamps for traceability. For example, if the above three results are sorted by time as 2023-11 (academic), 2024-02 (media), and 2024-03 (government), since the entity keyword is "a certain material hardness standard", they are linked together to form a cross-platform evidence chain for that entity.

[0143] Next, in step S1063, natural language consistency analysis technology is used to compare the core content of all data records in the evidence chain. If the content is completely consistent, an update suggestion to keep the knowledge node unchanged is generated. If there is a conflict, the source priority determination model is used to first check whether the time of the government policy database record is the most recent. If so, a suggestion to adopt the government policy value is generated. If the government record is not the most recent, the credibility score of the academic literature database record is checked to see if it is the highest. If so, a suggestion to adopt the academic literature value is generated. If neither of the above two conditions is met, and the news media database record contains new information such as event progress and status changes, a suggestion to adopt the new status value is generated. For example, in the evidence chain, the government records "HRC50" in March 2024, the academic records "HRC48" in November 2023, and the media records "HRC50" in February 2024. The content conflicts, and the government record is the most recent. Therefore, a suggestion to adopt "HRC50" is generated.

[0144] Finally, in step S1064, the report automatic generation tool is used to extract specific modifications from the update suggestions. These modifications are then combined with supporting records from the evidence chain to form a knowledge node update suggestion report containing the updated content, source of evidence, and judgment logic. The report format is standardized to natural language descriptions to ensure readability. For example, the report content might be: "It is recommended to update the hardness standard of a certain material from HRC48 to HRC50. The reason is that the latest record in the government policy database in March 2024 shows the standard as HRC50, which is later than the record in the academic literature database in November 2023 and consistent with the record in the news media database in February 2024."

[0145] In practical application, within the knowledge graph of "environmental protection indicators for a certain industry," the search results of three agents are aggregated: the government policy database records "environmental protection indicator at level A" in 2024, with the source tag "government" and the time tag "2024.06"; the academic literature database records "environmental protection indicator at level B" in 2023, with the source tag "academic" and the time tag "2023.11"; and the news media database records "environmental protection indicator at level A" in 2024, with the source tag "media" and the time tag "2024.05." These results are then sorted by time as 2023.11 (academic), 2024.05 (media), and 2024.06 (government). Since they all belong to the "environmental protection indicators for a certain industry" entity, a cross-platform evidence chain is generated. Subsequent analysis reveals content conflicts; the government policy database record is the most recent, and an update suggestion adopting level A is generated. Finally, a report was generated, with the specific update being "updating the environmental protection indicator of a certain industry from level B to level A". The reason for the update was that "the government policy database records the indicator as level A in June 2024, which is later than the level B record in the academic literature database in November 2023, and the news media database records also support level A in May 2024".

[0146] The overall solution of S106 described above forms a cross-platform evidence chain by aggregating multi-source retrieval results and generating an update suggestion report based on the analysis of the evidence chain. It can provide clear guidance for updating knowledge nodes by comprehensively considering information from multiple aspects, ensuring the objectivity and rationality of the update suggestions. At the same time, it allows users to clearly understand the content and reasons for the update, which helps to improve the accuracy and timeliness of knowledge in the knowledge graph and provides a reliable basis for knowledge maintenance and optimization.

[0147] The following is a complete example for steps 101 to 106, such as Figure 3 As shown, in the multi-source heterogeneous knowledge graph of "Green Building Standards in a Certain Region", there are nodes J and K. Node J has a high timeliness level and an evaluation value of 0.85, which means "the energy saving rate must be ≥75%"; node K has a low timeliness level and an evaluation value of 0.3, which means "the energy saving rate must be ≥65%".

[0148] First, timeliness features are extracted from the two nodes. The time stamp of node J is 2024-01 and the source is marked as "Ministry of Housing and Urban-Rural Development document". The time stamp of node K is 2021-05 and the source is marked as "local blog". The time decay values ​​are calculated to be 0.9 and 0.6 respectively, and the credibility weights are 0.95 and 0.5 respectively. Then, the knowledge newness evaluation matrix [[0.85,0.3]] is generated.

[0149] Next, the policy version number was analyzed to obtain a policy time change rate of 0.7, the document citation relationship was analyzed to obtain an academic time change rate of 15, and the news dissemination chain was analyzed to obtain a news time change rate of -1.2. A dynamic timeliness vector [0.7, 15, -1.2] was generated, the high threshold was adjusted to 0.8 and the medium threshold was adjusted to 0.4, and node J was labeled as high timeliness level and node K as low timeliness level.

[0150] Then, the process iterates through the nodes and finds two nodes that are directly related but have different timeliness levels. The calculated difference in evaluation value is 0.55, which exceeds the threshold of 0.3. Therefore, it is identified as a combination of contradictory statements. Due to the conflicting attribute values, a numerical contradiction type label is added, and the conflicting node pair is output. Next, based on the label, the government policy database is prioritized for retrieval. The dynamic timeliness vector is parsed to extract sequence features. The time window of the policy database is set to 2023-2024, the citation range of the document database is 2022-2023, and the dissemination chain of the media database is from January to March 2024. A set of targeted retrieval rules is generated.

[0151] Subsequently, three agents were activated in parallel: the government agent retrieved the Ministry of Housing and Urban-Rural Development document from January 2024 stating "energy saving rate ≥75%", the academic agent retrieved a paper from 2023 supporting "≥75%", and the media agent retrieved a report from February 2024 confirming the new regulations. Intermediate results were shared during the retrieval process to optimize the scope.

[0152] Finally, the aggregation results were sorted by time as follows: 2023 Academic, 2024-01 Government, and 2024-02 Media, forming a chain of evidence. The analysis found that the government records were the most recent, and an update recommendation report was generated, suggesting that the "≥65%" of node K be changed to "≥75%" based on the latest government documents and academic and media support.

[0153] Figure 4 This application provides a schematic diagram of a specific implementation of an agent collaborative decision-making optimization system based on a multi-source heterogeneous knowledge graph, referring to... Figure 4 The system may include:

[0154] Extraction module 41 is used to extract a set of time-sensitive features from entities in a multi-source heterogeneous knowledge graph, and dynamically generate a knowledge old and new evaluation matrix containing time decay characteristics and credibility weights based on the set of time-sensitive features.

[0155] The parsing module 42 is used to parse the version number of policy documents, the citation relationship of academic literature and the propagation chain of news events to generate a dynamic timeliness vector, and to use the dynamic timeliness vector and the knowledge newness evaluation matrix to mark the timeliness level of the nodes in the multi-source heterogeneous knowledge graph.

[0156] The identification module 43 is used to traverse the node pairs with different timeliness levels in the multi-source heterogeneous knowledge graph, identify contradictory statement combinations based on the knowledge newness evaluation matrix, and output conflicting node pairs, wherein the conflicting node pairs carry contradiction type labels.

[0157] The first generation module 44 is used to generate a set of targeted retrieval rules for each pair of conflict nodes based on the heterogeneous data structure features contained in the conflict type label and the dynamic timeliness vector.

[0158] The scheduling module 45 is used to schedule the government database agent, academic knowledge base agent and authoritative media source agent in parallel for the conflict node, and to perform retrieval operations in parallel for the government policy database, academic literature database and news media database associated with the dynamic timeliness vector according to the targeted retrieval rule set, and exchange intermediate retrieval results to optimize the retrieval strategy.

[0159] The second generation module 46 is used to aggregate the search results of the government database agent, the academic knowledge base agent, and the authoritative media source agent to generate a cross-platform evidence chain, and to generate a knowledge node update suggestion report based on the cross-platform evidence chain.

[0160] The agent collaborative decision optimization system based on multi-source heterogeneous knowledge graphs in this application is used to implement the aforementioned agent collaborative decision optimization method based on multi-source heterogeneous knowledge graphs. Therefore, the specific implementation of the agent collaborative decision optimization system based on multi-source heterogeneous knowledge graphs can be found in the embodiment section of the agent collaborative decision optimization method based on multi-source heterogeneous knowledge graphs above. The specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0161] This application also provides an electronic device, including: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the above-described agent collaborative decision optimization method based on multi-source heterogeneous knowledge graph.

[0162] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described agent collaborative decision optimization methods based on multi-source heterogeneous knowledge graphs.

[0163] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0164] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the embodiments of the intelligent agent collaborative decision optimization method based on multi-source heterogeneous knowledge graphs.

[0165] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0166] The foregoing has provided a detailed description of an intelligent agent collaborative decision-making optimization method and system based on a multi-source heterogeneous knowledge graph, as provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A collaborative decision-making optimization method for intelligent agents based on multi-source heterogeneous knowledge graphs, characterized in that, include: Extract timeliness feature sets from entities in a multi-source heterogeneous knowledge graph, and dynamically generate a knowledge age evaluation matrix that includes time decay characteristics and credibility weights based on the timeliness feature sets. A dynamic timeliness vector is generated by analyzing the version number of policy documents, the citation relationship of academic literature and the propagation chain of news events. The timeliness level of nodes in the multi-source heterogeneous knowledge graph is marked by the dynamic timeliness vector and the knowledge newness evaluation matrix. Traverse the nodes with different timeliness levels in the multi-source heterogeneous knowledge graph, identify contradictory statement combinations based on the knowledge newness evaluation matrix, and output conflicting node pairs, with each conflicting node pair carrying a contradiction type label. Based on the conflict type label and the heterogeneous data structure features contained in the dynamic timeliness vector, a set of targeted retrieval rules is generated for each conflict node pair; For the conflict nodes, the government database agent, academic knowledge base agent, and authoritative media source agent are scheduled in parallel. Based on the targeted retrieval rule set, the government policy database, academic literature database, and news media database associated with the dynamic timeliness vector are executed in parallel and intermediate retrieval results are exchanged to optimize the retrieval strategy. The search results from the government database agent, academic knowledge base agent, and authoritative media source agent are aggregated to generate a cross-platform evidence chain, and a knowledge node update suggestion report is generated based on the cross-platform evidence chain. The process of traversing node pairs with different timeliness levels in the multi-source heterogeneous knowledge graph, identifying contradictory statement combinations based on the knowledge novelty evaluation matrix, and outputting conflicting node pairs includes: Traverse all directly related node pairs in the multi-source heterogeneous knowledge graph and filter out node pairs with different timeliness levels; For each selected node pair, the knowledge newness and oldness evaluation value of the corresponding entity is obtained from the knowledge newness and oldness evaluation matrix, and the difference between the two obtained knowledge newness and oldness evaluation values ​​is calculated. The difference value is compared with a preset difference threshold. If the difference value exceeds the difference threshold, the corresponding node pair is identified as a combination of contradictory statements. Query the statement content of the entity corresponding to the node in the contradictory statement combination. If the entity attribute values ​​conflict with each other, it is a node pair labeled with a numerical contradiction type label. If the entity relationship descriptions conflict with each other, it is a node pair labeled with a relationship contradiction type label. If the entity state descriptions conflict with each other, it is a node pair labeled with a state contradiction type label. Output the node pairs labeled with the contradiction type as conflict node pairs.

2. The method according to claim 1, characterized in that, The process aggregates the search results from the government database agent, academic knowledge base agent, and authoritative media source agent to generate a cross-platform evidence chain, and generates a knowledge node update suggestion report based on the cross-platform evidence chain, including: The search results returned by the government database agent, the academic knowledge base agent, and the authoritative media source agent are aggregated, and the search results include data records, source tags, and time tags; All search results are sorted chronologically according to time stamps, and data records belonging to the same entity are linked in chronological order to generate a cross-platform evidence chain; The consistency of data records in the cross-platform evidence chain is analyzed. If all records are consistent, an update suggestion to keep the nodes unchanged is generated. If there are conflicts in the records and the record time of the government policy database is the most recent, an update suggestion to adopt the government policy value is generated. If there are conflicts in the records and the record of the academic literature database has the highest credibility, an update suggestion to adopt the academic literature value is generated. If there are conflicts in the records and the news media database provides new status information, an update suggestion to adopt the new status value is generated. Based on the update suggestions, a knowledge node update suggestion report containing specific update content and reasons for the update is generated.

3. The method according to claim 1, characterized in that, The step of generating a targeted retrieval rule set for each pair of conflict nodes based on the heterogeneous data structure features contained in the conflict type label and the dynamic timeliness vector includes: Based on the contradiction type labels of the conflict node pairs, a priority allocation rule is determined. The priority allocation rule stipulates that numerical contradiction type labels are preferentially assigned to the government policy database, relational contradiction type labels are preferentially assigned to the academic literature database, and state contradiction type labels are preferentially assigned to the news media database. The dynamic timeliness vector is analyzed to extract policy version sequence features, document citation sequence features, and news dissemination sequence features; Based on the policy version sequence characteristics, a version time window constraint is set for the retrieval operation pointing to the government policy database; based on the document citation sequence characteristics, a citation time range constraint is set for the retrieval operation pointing to the academic literature database; and based on the news dissemination sequence characteristics, a dissemination time chain constraint is set for the retrieval operation pointing to the news media database. For each pair of conflicting nodes, a set of targeted retrieval rules is generated, which includes the target database type, time constraints, and retrieval keywords, based on the priority allocation rules and corresponding time constraints.

4. The method according to claim 1, characterized in that, The method involves parallel scheduling of government database agents, academic knowledge base agents, and authoritative media source agents for the conflict nodes. Based on the targeted retrieval rule set, parallel retrieval operations are performed on the government policy database, academic literature database, and news media database associated with the dynamic timeliness vector, and intermediate retrieval results are exchanged to optimize the retrieval strategy. This includes: For each pair of conflicting nodes, the government database agent, the academic knowledge base agent, and the authoritative media source agent are activated in parallel. The government database agent accesses the government policy database according to the rule entries in the targeted retrieval rule set and performs the retrieval using the retrieval keywords and time constraints specified in the rules. The academic knowledge base agent accesses the academic literature database according to the rule entries in the targeted retrieval rule set and performs the retrieval using the retrieval keywords and time constraints specified in the rules. The authoritative media source agent accesses the news media database according to the corresponding rule entries in the targeted retrieval rule set, and performs the retrieval using the retrieval keywords and time constraints specified in the rules. During the retrieval process, the time range of the retrieval is expanded or narrowed proportionally according to the magnitude of the time change rate parameter in the dynamic timeliness vector. During the retrieval process, the document identifiers, policy article numbers or news event IDs of the retrieved documents are exchanged, and the search keywords and time range are adjusted according to the exchanged information to achieve collaborative retrieval among the government policy database, academic literature database and news media database.

5. The method according to claim 1, characterized in that, The process involves analyzing policy document version numbers, academic literature citation relationships, and news event propagation chains to generate dynamic timeliness vectors. These dynamic timeliness vectors, along with the knowledge freshness / oldness assessment matrix, are used to label the timeliness levels of nodes in the multi-source heterogeneous knowledge graph, including: Analyze the version number of policy documents, extract the release time of the version, and calculate the reciprocal of the number of days between consecutive release times as the policy time change rate; analyze the citation relationship of academic literature, extract the publication time and number of citations of the literature, and calculate the growth rate of the number of citations per unit time as the academic time change rate; analyze the propagation chain of news events, extract the first release time and propagation path time of the event, and calculate the reciprocal of the average time delay from the first release to the reprint of the news event as the news time change rate. The policy time change rate, academic time change rate, and news time change rate are combined to generate a dynamic timeliness vector; Using the knowledge newness and oldness assessment values ​​corresponding to each entity in the knowledge newness and oldness assessment matrix, a high timeliness level threshold and a medium timeliness level threshold are set. Simultaneously, the weighted average of the policy time change rate, academic time change rate and news time change rate in the dynamic timeliness vector is used to proportionally adjust the high timeliness level threshold and the medium timeliness level threshold. Based on the adjusted high timeliness level threshold and medium timeliness level threshold, the knowledge newness assessment value of each entity is divided into high timeliness level, medium timeliness level or low timeliness level, and according to the division results, the corresponding timeliness level is labeled for the nodes in the multi-source heterogeneous knowledge graph.

6. The method according to claim 1, characterized in that, The step of extracting a set of time-sensitive features from entities in a multi-source heterogeneous knowledge graph, and dynamically generating a knowledge age evaluation matrix containing time decay characteristics and credibility weights based on the set of time-sensitive features, includes: Extract time stamps and source stamps from each entity in the multi-source heterogeneous knowledge graph to form a set of timeliness features; Based on the time markers in the timeliness feature set, calculate the difference between the current system time and each time marker, and apply an exponentially decreasing function based on the difference. The time decay value for each entity is calculated using the exponentially decreasing function. In this context, λ is the attenuation coefficient. This is the time difference; Based on the source markers in the timeliness feature set, query the pre-defined correspondence table between source types and credibility weight values ​​based on historical data statistical analysis to obtain the credibility weight corresponding to each source marker; The time decay value of each entity is multiplied by the corresponding credibility weight to obtain the knowledge old and new evaluation value of the entity. Based on the knowledge old and new evaluation values ​​of all entities, a knowledge old and new evaluation matrix is ​​generated by organizing them according to the entity identifier.

7. An intelligent agent collaborative decision-making optimization system based on multi-source heterogeneous knowledge graphs, characterized in that, include: The extraction module is used to extract a set of time-sensitive features from entities in a multi-source heterogeneous knowledge graph, and dynamically generate a knowledge oldness evaluation matrix containing time decay characteristics and credibility weights based on the set of time-sensitive features. The parsing module is used to parse policy document version numbers, academic literature citation relationships, and news event propagation chains to generate dynamic timeliness vectors. The dynamic timeliness vectors and the knowledge newness and oldness evaluation matrix are used to label the timeliness level of nodes in the multi-source heterogeneous knowledge graph. The identification module is used to traverse the node pairs with different timeliness levels in the multi-source heterogeneous knowledge graph, identify contradictory statement combinations based on the knowledge newness evaluation matrix, and output conflicting node pairs, wherein the conflicting node pairs carry contradiction type labels. The process of traversing node pairs with different timeliness levels in the multi-source heterogeneous knowledge graph, identifying contradictory statement combinations based on the knowledge novelty evaluation matrix, and outputting conflicting node pairs includes: Traverse all directly related node pairs in the multi-source heterogeneous knowledge graph and filter out node pairs with different timeliness levels; For each selected node pair, the knowledge newness and oldness evaluation value of the corresponding entity is obtained from the knowledge newness and oldness evaluation matrix, and the difference between the two obtained knowledge newness and oldness evaluation values ​​is calculated. The difference value is compared with a preset difference threshold. If the difference value exceeds the difference threshold, the corresponding node pair is identified as a combination of contradictory statements. Query the statement content of the entity corresponding to the node in the contradictory statement combination. If the entity attribute values ​​conflict with each other, it is a node pair labeled with a numerical contradiction type label. If the entity relationship descriptions conflict with each other, it is a node pair labeled with a relationship contradiction type label. If the entity state descriptions conflict with each other, it is a node pair labeled with a state contradiction type label. Output the node pairs labeled with the contradiction type as conflict node pairs; The first generation module is used to generate a set of targeted retrieval rules for each pair of conflict nodes based on the heterogeneous data structure features contained in the conflict type label and the dynamic timeliness vector. The scheduling module is used to schedule government database agents, academic knowledge base agents, and authoritative media source agents in parallel for the conflict nodes. Based on the targeted retrieval rule set, it performs retrieval operations in parallel on the government policy database, academic literature database, and news media database associated with the dynamic timeliness vector and exchanges intermediate retrieval results to optimize the retrieval strategy. The second generation module is used to aggregate the search results of the government database agent, the academic knowledge base agent, and the authoritative media source agent to generate a cross-platform evidence chain, and to generate a knowledge node update suggestion report based on the cross-platform evidence chain.

8. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the agent collaborative decision optimization method based on multi-source heterogeneous knowledge graph as described in any one of claims 1 to 6 when executing the computer program.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the agent collaborative decision optimization method based on a multi-source heterogeneous knowledge graph as described in any one of claims 1 to 6.

Images