Penetrating correlation recognition method based on multi-relation graph and fact extraction

Abstract

The present application relates to the field of information technology, in particular to a penetrating correlation identification method based on multi-relation graph and fact extraction. The method comprises: based on historical data, structured business data and fact data of a plurality of subjects involved in a plurality of candidate paths, performing strength evaluation on a plurality of relation edges to obtain a single-edge strength value of each relation edge; for each candidate path, fusing the single-edge strength value and weight value of the plurality of relation edges included in the candidate path to obtain an original strength value of the candidate path; for each candidate path, correcting the original strength value thereof based on a reliability value to obtain a comprehensive strength value of the candidate path; analyzing the relative sizes of the comprehensive strength values of the plurality of candidate paths to obtain a target path corresponding to a starting subject, and determining a plurality of subjects included in the target path as correlation subjects of the starting subject. The present application can improve the accuracy of penetrating correlation identification of multi-relation graph and fact extraction.

Description

Technical Field

[0001] This invention relates to the field of information technology, and more specifically to a penetrating association recognition method based on multi-relationship graphs and fact extraction. Background Technology

[0002] In scenarios such as corporate procurement management, supplier access, anti-commercial bribery, and conflict of interest investigation, procurement decision-makers and supplier companies often form seemingly independent but substantially related networks of interests through multi-layered equity nesting, cross-shareholding, implicit nominee shareholding, kinship, joint investment, joint senior management, joint addresses, and profit-sharing agreements. Accurately identifying the strength of these substantial connections is the foundation for procurement compliance review, conflict of interest investigation, and anti-commercial bribery.

[0003] In related technologies, a key approach is to use a knowledge graph constructed from single declaration data or single business equity data, employing K-hop traversal combined with fixed weights to perform penetrating association identification, in order to determine the strength of the association between procurement decision-makers and supplier companies. However, this method treats non-equity relationships such as kinship, joint investment, and common address relationships in the path as general connections with the same weight as equity relationships. This makes it difficult to distinguish between strong associations and weak interference, resulting in strong interest-related paths being easily missed or overwhelmed by invalid paths.

[0004] Therefore, it is evident that the accuracy of penetrating association identification based on multi-relationship graphs and fact extraction in related technologies is relatively low. Summary of the Invention

[0005] The purpose of this invention is to provide a penetrating association identification method based on multi-relation graphs and fact extraction, which solves the technical problem of low accuracy in penetrating association identification based on multi-relation graphs and fact extraction in related technologies.

[0006] In a first aspect, the present invention provides a penetrating association identification method based on multi-relationship graphs and fact extraction, the method comprising: Based on historical data, structured business data and factual data of multiple subjects involved in multiple candidate paths, the strength of multiple relationship edges is evaluated to obtain the single-edge strength value of each relationship edge. The single-edge strength value is used to characterize the association strength between the subjects connected by the corresponding relationship edge. The starting subjects of multiple candidate paths are the same. For each candidate path, the original strength value of each candidate path is obtained by fusing the single-edge strength value and weight value of the multiple relation edges it includes. The original strength value is used to characterize the original association strength of the corresponding candidate path, and the weight value is obtained by analyzing the order of the corresponding relation edge in its candidate path and the total number of relation edges included in its candidate path. For each candidate path, its original strength value is corrected based on its reliability value to obtain the comprehensive strength value of each candidate path. The comprehensive strength value is used to characterize the final association strength of the corresponding candidate path after removing interference, and the reliability value is used to characterize the credibility of the corresponding candidate path after removing interference. The reliability value is obtained based on the degree of sharing of the intermediate subjects of the corresponding candidate path, the relative strength of the terminal relation edges, the proportion of relation edges with a single edge strength value lower than the threshold, and the time effective value of each relation edge. By analyzing the relative magnitudes of the comprehensive strength values ​​of multiple candidate paths, the target path corresponding to the starting subject is obtained, and the multiple subjects included in the target path are identified as the associated subjects of the starting subject. The target path is used to characterize the path with the highest association strength with the starting subject among multiple candidate paths.

[0007] In one embodiment, the step of evaluating the strength of multiple relationship edges based on historical data, structured business data and factual data of multiple entities involved in multiple candidate paths, and obtaining a single-edge strength value for each relationship edge, includes: Based on historical data, scenario relevance statistics are performed on each relation edge involved in the path to be identified to obtain the scenario credibility value of each relation edge of the path to be identified. The path to be identified is any one of multiple candidate paths. The scenario credibility value is used to characterize the credibility of the relation type to which the corresponding relation edge belongs under the same historical business scenario. The relation types include: equity relationship, investment relationship, employment relationship, agreement relationship, shared address relationship and factual relationship. Based on the structured business data and factual data of multiple subjects involved in the path to be identified, the relative strength level of each relation edge in its relation type is analyzed to obtain the relative single-edge strength value of each relation edge in the path to be identified. The relative single-edge strength value is used to characterize the relative strength of the corresponding relation edge in its relation type. Based on the structured business data and factual data of multiple entities involved in the path to be identified, a credibility analysis is performed on the data source of each relation edge of the path to be identified to obtain the source credibility value of each relation edge of the path to be identified. The source credibility value is used to characterize the credibility of the data source of the corresponding relation edge. Based on the structured business data and factual data of multiple entities involved in the path to be identified, the validity period of each relation edge of the path to be identified is identified, and the time validity value of each relation edge of the path to be identified is obtained. The time validity value is used to characterize the validity of the corresponding relation edge. For each relation edge of the path to be identified, the single-edge strength value of each relation edge is calculated by multiplying its scene confidence value, relative single-edge strength value, source confidence value, and time validity value.

[0008] In one embodiment, the step of performing scene relevance statistics on each relation edge involved in the path to be identified based on historical data to obtain the scene confidence value of each relation edge of the path to be identified includes: For each relation edge of the path to be identified, its relation type is determined based on historical data; For each relationship type, the frequency is counted based on historical data to obtain the number of times each relationship type appears in historical high-value paths and the number of times each relationship type appears in historical low-value paths. High-value paths are used to represent paths that have been confirmed to have a high degree of correlation in the same historical business scenario, and low-value paths are used to represent paths that have been confirmed to have a low degree of correlation in the same historical business scenario. For each type of relationship, the relative frequency of each type of relationship in the high-value path is calculated by comparing the number of times it appears in the historical high-value path with the total number of times all types of relationships appear in the historical high-value path. For each type of relationship, the relative frequency of each relationship type in low-value paths is calculated by comparing the number of times it appears in historical low-value paths with the total number of times multiple relationship types appear in historical low-value paths. For each relation type, data analysis is performed based on its relative frequency in high-value paths and relative frequency in low-value paths to obtain the scenario credibility value for each relation type. The scenario credibility value of each relation edge in the path to be identified is equal to the scenario credibility value of its relation type. The scenario credibility value is positively correlated with the relative frequency in high-value paths of the corresponding relation type, and negatively correlated with the relative frequency in low-value paths of the corresponding relation type.

[0009] In one embodiment, the step of analyzing the relative strength level of each relation edge within its respective relation type based on the structured business data and factual data of multiple entities involved in the path to be identified, to obtain the relative one-sided strength value of each relation edge in the path to be identified, includes: For each relation edge, based on the structured business data and factual data of multiple subjects involved in the path to be identified, numerical extraction is performed to obtain the original value of each relation edge, where the original value is used to characterize the original strength of the corresponding relation edge; For each relation edge, a relative single-edge strength value is calculated based on its original value, the original values ​​of the multiple relation edges included in its relation type, and the total number of relation edges included in its relation type. The relative single-edge strength value is positively correlated with the original value of the corresponding relation edge, negatively correlated with the original values ​​of other relation edges in the relation type to which the corresponding relation edge belongs, and negatively correlated with the total number of relation edges included in the relation type to which the corresponding relation edge belongs.

[0010] In one embodiment, the credibility analysis of the data source of each relation edge of the path to be identified, based on the structured business data and factual data of multiple entities involved in the path to be identified, to obtain the source credibility value of each relation edge of the path to be identified, includes: Based on the structured business data and factual data of multiple entities involved in the path to be identified, the supporting data of each relation edge is identified to obtain the supporting value of each relation edge of the path to be identified. The supporting data is used to characterize the evidence of the existence of the relation edge obtained from different sources. Based on the structured business data and factual data of multiple entities involved in the path to be identified, the conflict data of each relation edge is identified to obtain the conflict value of each relation edge of the path to be identified. The conflict data is used to characterize the evidence that negates the relation edge obtained from data from different sources. For each relation edge of the path to be identified, data analysis is performed based on its support value and conflict value to obtain the source confidence value of each relation edge of the path to be identified. The source confidence value is positively correlated with the support value of the corresponding relation edge, and the source confidence value is negatively correlated with the conflict value of the corresponding relation edge.

[0011] In one embodiment, identifying the validity period of each relation edge in the path to be identified based on the structured business data and factual data of multiple entities involved in the path to be identified, and obtaining the time validity value of each relation edge in the path to be identified, includes: Based on the structured business data and factual data of multiple entities involved in the path to be identified, the expiration information of each relation edge is parsed to obtain the effective time interval of each relation edge in the path to be identified. Based on the time distance between the effective time interval of each relation edge of the path to be identified and the corresponding identification time of the path to be identified, the deviation value of each relation edge of the path to be identified is obtained. The deviation value is used to characterize the degree of deviation between the effective time interval of the corresponding relation edge and the corresponding identification time of the path to be identified. For each relation edge of the path to be identified, analyze the degree of dispersion of its deviation value and the deviation values ​​of multiple relation edges included in its relation type to obtain the effective time value of each relation edge of the path to be identified.

[0012] In one embodiment, the reliability value is jointly represented by a path quality value and an effectiveness value. The path quality value is used to characterize the reliability of the corresponding candidate path structure, and the effectiveness value is used to characterize the effectiveness of the corresponding candidate path. For each candidate path, its original strength value is corrected based on its reliability value to obtain a comprehensive strength value for each candidate path, including: For each candidate path, its original strength value is corrected based on its path quality value to obtain the initial strength value of each candidate path; For each candidate path, its initial strength value is corrected based on its effective value to obtain the comprehensive strength value of each candidate path.

[0013] In one embodiment, the path quality value is characterized by a shared value and a landing value. The shared value is used to characterize the independence of the intermediate entities of the corresponding candidate path, and the landing value is used to characterize the relative strength of the single-sided strength value of the terminal relationship edge of the corresponding candidate path. For each candidate path, the original strength value is corrected based on its path quality value to obtain the initial strength value of each candidate path, including: For each candidate path, a common value is calculated based on the number of entities connected by its intermediate entities and the maximum number of connections among all intermediate entities of the entity type to which the intermediate entities belong. The entity type includes at least one of the following: corporate entities, natural person entities, and address entities. The common value is negatively correlated with the number of entities connected by the intermediate entities of the corresponding candidate path, and positively correlated with the maximum number of connections among the entity types to which the intermediate entities of the corresponding candidate path belong. For each candidate path, the landing point value is calculated based on the single-sided strength value of its terminal relation edge and the strength values ​​of its multiple relation edges. The landing point value is positively correlated with the single-sided strength value of the terminal relation edge of the corresponding candidate path, and negatively correlated with the single-sided strength values ​​of the other relation edges of the corresponding candidate path excluding the terminal relation edge. For each candidate path, its common value, landing point value, and original intensity value are multiplied to obtain the initial intensity value of each candidate path.

[0014] In one embodiment, the effective value is obtained based on the weak edge ratio and attenuation value analysis. The weak edge ratio is used to characterize the intensity of weak edge interference in the corresponding candidate path, and the attenuation value is used to characterize the degree of weakening of the association strength of the corresponding candidate path due to the change of the temporal effective value with positional importance. For each candidate path, the initial strength value is corrected based on its effective value to obtain the comprehensive strength value of each candidate path, including: For each candidate path, the weak edge ratio of each candidate path is calculated by the ratio of the sum of the single-edge strength values ​​of the weak edges it includes to the sum of the single-edge strength values ​​of the multiple relation edges it includes. The weak edges are used to indicate relation edges whose single-edge strength values ​​are lower than the strength threshold. For each candidate path, a weighted average is calculated based on the time validity value and weight value of each relation edge to obtain the decay value of each candidate path; For each candidate path, the effective value of each candidate path is calculated based on its weak edge ratio and decay value. The effective value is negatively correlated with the weak edge ratio of the corresponding candidate path and negatively correlated with the decay value of the corresponding candidate path. For each candidate path, calculate the product of its effective value and its initial strength value to obtain the comprehensive strength value of each candidate path.

[0015] In one embodiment, the step of obtaining the weight value of each relation edge includes: Data analysis is performed based on the order of the target relation edge among the multiple relation edges included in its candidate path, and the total number of relation edges included in its candidate path, to obtain the weight value of the target relation edge. The target relation edge is any relation edge among the multiple relation edges, and the order increases sequentially from the starting body to the ending body in the candidate path. The weight value of the target relation edge is positively correlated with its order among the multiple relation edges included in its candidate path, and negatively correlated with the total number of relation edges included in its candidate path.

[0016] The present invention has the following beneficial effects: This invention assesses the strength of multiple relationship edges based on historical data, structured business data, and factual data of multiple entities involved in multiple candidate paths, obtaining a single-edge strength value for each relationship edge. Then, it fuses the single-edge strength values ​​and weight values ​​of the multiple relationship edges included in each candidate path to obtain the original strength value of each candidate path. Subsequently, it corrects the original strength value based on the reliability value of each candidate path to obtain the comprehensive strength value of each candidate path. Next, it analyzes the relative magnitude of the comprehensive strength values ​​of multiple candidate paths to obtain the target path corresponding to the starting entity, and identifies the multiple entities included in the target path as the associated entities of the starting entity. This invention quantifies the strength differences of different relationship edges through single-edge strength values, avoiding the problem of treating various heterogeneous relationship edges as having equal strength. Combining the weights of relationship edges can strengthen the influence of relationship edges at different positions (orders) on the association strength of candidate paths. Furthermore, the correction through reliability values ​​eliminates interference from common nodes, weak edges, and time-decrease interference in candidate paths. Therefore, this invention improves the accuracy of penetrating association identification based on multi-relationship graphs and fact extraction. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a penetrating association identification method based on multi-relationship graphs and fact extraction provided in an embodiment of the present invention. Detailed Implementation

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] The specific scheme of the penetrating association recognition method based on multi-relation graph and fact extraction provided by the present invention will be described in detail below with reference to the accompanying drawings.

[0020] This invention proposes a penetrating association recognition method based on multi-relation graphs and fact extraction. The method provided by this invention can be executed by a penetrating association recognition system based on multi-relation graphs and fact extraction, specifically by the controller within that system. Please refer to [link to relevant documentation]. Figure 1 The diagram illustrates a flowchart of a penetrating association identification method based on multi-relationship graphs and fact extraction, according to an embodiment of the present invention. The method includes the following steps: Step S1: Based on historical data, structured business data and factual data of multiple entities involved in multiple candidate paths, perform strength assessment on multiple relationship edges to obtain the single-edge strength value of each relationship edge.

[0021] Among them, the unilateral strength value is used to characterize the association strength between the subjects connected by the corresponding relation edge. The starting subjects of multiple candidate paths are the same, and the starting subject of each candidate path is a natural person subject. The natural person subject of the starting subject is used to instruct the procurement decision-makers. The ending subject of each candidate path is a corporate subject.

[0022] Historical data is used to characterize historical review data that are the same as the business scenarios to which multiple candidate paths belong. The business scenario can be characterized by a procurement business scenario identifier, which is used to distinguish different procurement business scenarios, that is, to uniquely locate the procurement business scenario among multiple procurement business scenarios. For example, multiple procurement business scenarios include: bulk material procurement, engineering service procurement, office supplies procurement, information technology project procurement, and consulting service procurement.

[0023] Structured business data includes: subject type, relationship type with connected subjects, and subject type of the connected subjects. Subject types include at least one of the following: corporate subject, natural person subject, and address subject. For example, subject types include: corporate subject, natural person subject, and address subject. Relationship types include: equity relationship, investment relationship, employment relationship, agreement relationship, shared address relationship, and factual relationship. Agreement relationships characterize the synergy of interests or control relationship formed between two subjects through formal or informal agreements. For example, a profit-sharing agreement indicates an agreement between the parties to share economic benefits. Factual relationships include: kinship and actual control relationships extracted from unstructured text. Kinship relationships include spousal relationships, lineal consanguineous relationships, and collateral consanguineous relationships.

[0024] For each candidate path, the controller can search for structured business data of the entities involved in each candidate path from external data sources. Subsequently, it uses a large model to extract facts from unstructured data (including bidding documents, contract texts, news reports, corporate announcements, and judgments), and performs entity identification and disambiguation on the extraction results to obtain factual data (i.e., factual evidence data). External data sources include, for example, business registration databases, bidding information release platforms, contract management systems, judicial open data, news media platforms, corporate announcement release platforms, judgment document databases, and historical procurement case databases.

[0025] Specifically, the controller can use pre-stored entity recognition algorithms (such as named entity recognition algorithms) and entity disambiguation algorithms (such as clustering-based entity disambiguation algorithms) to perform entity recognition and entity disambiguation on the extracted results to obtain factual data. In this way, the subject and object words extracted from unstructured text can be mapped to a unified subject name, ensuring that the same entity is uniformly identified and located in different texts, avoiding duplication or conflict caused by name variations (such as "Company A" and "A Technology Co., Ltd."), thereby providing a unique and traceable subject identifier for subsequent calculation of the comprehensive strength value of candidate paths.

[0026] In this embodiment of the invention, before fact extraction, the controller can acquire multiple training data sets. Each training data set includes: text samples (such as unstructured texts like bidding documents, contract texts, news reports, company announcements, and judgments), and corresponding relationship extraction tags (including subject words, object words, relationship trigger words, time words, and evidence fragment positions). Then, the controller can train a model on these multiple training data sets to obtain a large model. For example, the text sample is: Zhang and Li are husband and wife. Li is the actual controller of Company A, and Zhang is the legal representative of Company B. Company B participated in the bidding for Company A's office supplies procurement project on January 15, 2024. The relationship extraction tags are: subject word: Zhang, object word: Company A, relationship trigger word: husband and wife, time word: January 15, 2024, and evidence fragment: "Zhang and Li are husband and wife. Li is the actual controller of Company A, and Zhang is the legal representative of Company B."

[0027] Relationship edges of varying strengths differ in terms of business experience, objective load level (i.e., original strength), evidence reliability, and timeliness. The scenario credibility value reflects these differences in business experience, the relative unilateral strength value reflects the differences in objective strength, the source credibility value reflects the reliability of the data source (evidence), and the time validity value reflects timeliness. Based on this, the controller can obtain the scenario credibility value, relative unilateral strength value, source credibility value, and time validity value for each relationship edge, thereby obtaining the unilateral strength value for each relationship edge.

[0028] Specifically, the strength evaluation of multiple relationship edges based on historical data, structured business data and factual data of multiple entities involved in multiple candidate paths, to obtain the single-edge strength value of each relationship edge, includes: Based on historical data, scenario relevance statistics are performed on each relation edge involved in the path to be identified to obtain the scenario credibility value of each relation edge of the path to be identified. The path to be identified is any one of multiple candidate paths, and the scenario credibility value is used to characterize the credibility of the relation type to which the corresponding relation edge belongs under the same historical business scenario. Based on the structured business data and factual data of multiple subjects involved in the path to be identified, the relative strength level of each relation edge in its relation type is analyzed to obtain the relative single-edge strength value of each relation edge in the path to be identified. The relative single-edge strength value is used to characterize the relative strength of the corresponding relation edge in its relation type. Based on the structured business data and factual data of multiple entities involved in the path to be identified, a credibility analysis is performed on the data source of each relation edge of the path to be identified to obtain the source credibility value of each relation edge of the path to be identified. The source credibility value is used to characterize the credibility of the data source of the corresponding relation edge. Based on the structured business data and factual data of multiple entities involved in the path to be identified, the validity period of each relation edge of the path to be identified is identified, and the time validity value of each relation edge of the path to be identified is obtained. The time validity value is used to characterize the validity of the corresponding relation edge. For each relation edge of the path to be identified, the single-edge strength value of each relation edge is calculated by multiplying its scene confidence value, relative single-edge strength value, source confidence value, and time validity value.

[0029] For example, the one-sided strength value G(e) of the e-th relation edge can be expressed as: In the formula, C(r) is the scenario credibility value of the r-th relation type to which the e-th relation edge belongs, and the scenario credibility value of the e-th relation edge is equal to C(r). L(e) is the relative one-sided strength value of the e-th relation edge. E(e) is the source credibility value of the e-th relation edge. T(e) is the time validity value of the e-th relation edge.

[0030] Analysis revealed that historically high-value paths generally possess stronger information authenticity compared to historically low-value paths. Therefore, relationship types appearing on historically high-value paths are considered more reliable than those appearing on historically low-value paths. Furthermore, as the frequency of a relationship type's occurrence in historically high-value paths increases, its reliability is considered to increase; conversely, as the frequency of a relationship type's occurrence in historically low-value paths increases, its reliability is considered to decrease. Based on this, the controller can calculate the scenario reliability value of a relationship type by comparing its relative frequency of occurrence on historically high-value and historically low-value paths.

[0031] In one example, the process of performing scene relevance statistics on each relation edge involved in the path to be identified based on historical data, and obtaining the scene confidence value of each relation edge of the path to be identified, includes: For each relation edge of the path to be identified, its relation type is determined based on historical data; For each relationship type, the frequency is counted based on historical data to obtain the number of times each relationship type appears in historical high-value paths and the number of times each relationship type appears in historical low-value paths. High-value paths are used to represent paths that have been confirmed to have a high degree of correlation in the same historical business scenario, and low-value paths are used to represent paths that have been confirmed to have a low degree of correlation in the same historical business scenario. For each type of relationship, the relative frequency of each type of relationship in the high-value path is calculated by comparing the number of times it appears in the historical high-value path with the total number of times all types of relationships appear in the historical high-value path. For each type of relationship, the relative frequency of each relationship type in low-value paths is calculated by comparing the number of times it appears in historical low-value paths with the total number of times multiple relationship types appear in historical low-value paths. For each relation type, data analysis is performed based on its relative frequency in high-value paths and relative frequency in low-value paths to obtain the scenario credibility value for each relation type. The scenario credibility value of each relation edge in the path to be identified is equal to the scenario credibility value of its relation type. The scenario credibility value is positively correlated with the relative frequency in high-value paths of the corresponding relation type, and negatively correlated with the relative frequency in low-value paths of the corresponding relation type.

[0032] For example, the scenario confidence value C(r) of the r-th relation type to which the e-th relation edge belongs can be expressed as: In the formula, Let P(r) be the relative frequency of the r-th type of relation to which the e-th relation edge belongs in the high-value path, P(r) be the number of times the r-th type of relation to which the e-th relation edge belongs appears in the historical high-value path, and P(all) be the sum of the number of times all relation types appear in the historical high-value path. Let N(r) be the relative frequency of the r-th type of relation to which the e-th relation edge belongs in the low-value path, N(r) be the number of times the r-th type of relation to which the e-th relation edge belongs appears in the historical low-value path, and N(all) be the total number of times all relation types appear in the historical low-value path. A very small positive number pre-stored for the controller, for example, The value is 0.001. Among them, the historical high-value paths and historical low-value paths are all paths with substantial relevance that have been independently verified by humans and are not dependent on the unilateral strength value or comprehensive strength value calculation in this method.

[0033] For example, in the scenario of office supplies procurement, a total of 500 paths were manually verified in the past year, including 150 high-value paths (i.e., 150 historical high-value paths) and 350 low-value paths (i.e., 350 historical low-value paths). Equity relationships appeared 90 times in high-value paths and 20 times in low-value paths. The controller can obtain the relative frequency of equity relationships in high-value paths as 90 / 150 = 0.6, and the relative frequency of equity relationships in low-value paths as 20 / 350 ≈ 0.057. Therefore, the scenario confidence value for equity relationships is approximately 0.912.

[0034] Because the original values ​​(also known as original load values) of relation edges of different relation types may have different dimensions and cannot be directly compared, for each relation edge, normalization can be performed based on its original value, the original values ​​of multiple relation edges included in its relation type, and the total number of relation edges included in its relation type to obtain the relative single-edge strength value of each relation edge. This eliminates heterogeneity and achieves comparability of strength across relation types. For example, for equity relations, the original value is the shareholding ratio (e.g., 30%), while for shared address relations, the original value is the reciprocal of the number of sharing entities (e.g., 1 / 100 = 0.01, meaning 100 companies share one address). The original values ​​of the two belong to different dimensions and value ranges, making it impossible to directly compare which better reflects the association strength: 30% or 0.01.

[0035] In one example, the process of analyzing the relative strength level of each relation edge in its respective relation type based on the structured business data and factual data of multiple entities involved in the path to be identified, and obtaining the relative one-sided strength value of each relation edge in the path to be identified, includes: For each relation edge, based on the structured business data and factual data of multiple subjects involved in the path to be identified, numerical extraction is performed to obtain the original value of each relation edge, where the original value is used to characterize the original strength of the corresponding relation edge; For each relation edge, a relative single-edge strength value is calculated based on its original value, the original values ​​of the multiple relation edges included in its relation type, and the total number of relation edges included in its relation type. The relative single-edge strength value is positively correlated with the original value of the corresponding relation edge, negatively correlated with the original values ​​of other relation edges in the relation type to which the corresponding relation edge belongs, and negatively correlated with the total number of relation edges included in the relation type to which the corresponding relation edge belongs.

[0036] For example, the relative one-sided strength value L(e) of the e-th relation edge can be expressed as: In the formula, q(e) represents the order of the original value of the e-th relation edge among the original values ​​of the multiple relation edges included in the r-th relation type to which it belongs. This order is arranged in ascending order of the original values ​​from smallest to largest. This represents the total number of relation edges belonging to the r-th relation type to which the e-th relation edge belongs. A very small positive number pre-stored for the controller, for example, The value is 0.0001.

[0037] In this embodiment of the invention, the original value of equity relationships is represented by shareholding ratio or voting rights ratio. For example, 51% indicates a shareholding ratio of 51%. The original value of investment relationships is represented by the proportion of joint investment to the total investment of both parties. For example, 0.4 indicates that the joint investment project accounts for 40% of the total investment of both parties. The original value of employment relationships is represented by a job level coefficient. For example, 1.0 represents the legal representative, 0.8 represents the director, 0.6 represents the manager, and the remaining ordinary management personnel or general employees can be represented by 0.4. The original value of agreement relationships is represented by the profit distribution ratio or the strength of agreement constraints. For example, 0.7 indicates a profit distribution ratio of 70%. The shared address relationship is represented by the reciprocal of the number of entities sharing the same address. For example, 1 / 100 = 0.01 indicates that 100 companies share the address. For kinship relationships in factual relationships, the original value of spousal relationships is 1.0, the original value of direct blood relatives (such as parents and children) is 0.8, and the original value of collateral blood relatives (such as siblings, uncles and nephews) is 0.5. The original value of the factual relationship is the percentage of fields that are completely extracted from the five necessary fields (subject words, object words, relationship trigger words, time words, and evidence fragment positions). For example, 1.0 means that all five fields are completely extracted, and 0.6 means that only three of the five fields are complete.

[0038] In one example, the process of performing credibility analysis on the data source of each relation edge of the path to be identified, based on the structured business data and factual data of multiple entities involved in the path to be identified, to obtain the source credibility value of each relation edge of the path to be identified, includes: Based on the structured business data and factual data of multiple entities involved in the path to be identified, the supporting data of each relation edge is identified to obtain the supporting value of each relation edge of the path to be identified. The supporting data is used to characterize the evidence of the existence of the relation edge obtained from different sources. Based on the structured business data and factual data of multiple entities involved in the path to be identified, the conflict data of each relation edge is identified to obtain the conflict value of each relation edge of the path to be identified. The conflict data is used to characterize the evidence that negates the relation edge obtained from data from different sources. For each relation edge of the path to be identified, data analysis is performed based on its support value and conflict value to obtain the source confidence value of each relation edge of the path to be identified. The source confidence value is positively correlated with the support value of the corresponding relation edge, and the source confidence value is negatively correlated with the conflict value of the corresponding relation edge.

[0039] The source credibility value is positively correlated with the support value (i.e., the more sufficient the evidence, the more credible it is) and negatively correlated with the conflict value (i.e., the more contradictions, the less credible it is). By combining the support value and the conflict value, the reliability of the evidence can be comprehensively assessed, avoiding misjudgment based on a single dimension, thereby improving the reliability of the obtained source credibility value.

[0040] For example, the source confidence value E(e) of the e-th relation edge can be expressed as: In the formula, S(e) is the support value, which satisfies... , Let be the set of data supporting the e-th relation edge. The data set supporting this data The source level value of the k-th supporting data is used to indicate the authority of the corresponding supporting data. The higher the source level value, the more authoritative and credible the supporting data is. For example, the business registration database corresponds to a high source level value (e.g., set to 1), the bidding information release platform, contract management system, and historical procurement case database correspond to the second highest source level value (e.g., set to 0.8), the judicial disclosure data and judgment document database correspond to a medium source level value (e.g., set to 0.6), and the enterprise announcement release platform and news media platform correspond to a low source level value (e.g., set to 0.4). The data set supporting this data The evidence independence value of the k-th supporting data is used to indicate whether the corresponding supporting data was formed from an independent source. If the source entity of the k-th supporting data has not been recorded before, and its file content hash is inconsistent with existing sources, then the supporting data is determined to be independently formed. The value is 1; if the source entity of the k-th supporting data has been recorded, or the file content hash is consistent with an existing source, then the evidence is determined to be duplicated from the same source, meaning the supporting data is not independently formed. The value is 0. The source entity is used to indicate the direct publisher of the evidence. For example, if information about the shareholders of Company A is obtained from the business registration database, the source entity is the administrative agency that published the business registration information, such as a municipal market supervision bureau.

[0041] D(e) is the conflict value, which can satisfy... . Let be the set of conflict data for the e-th relation edge. The dataset for this conflict data The source level value of the kth conflict data in the dataset. The dataset for this conflict data The evidence independence value of the k-th conflicting data. In this embodiment of the invention, the implementation method for obtaining the source level value of the conflicting data is similar to the implementation method for obtaining the source level value of the supporting data, and the implementation method for obtaining the evidence independence value of the conflicting data is similar to the implementation method for obtaining the evidence independence value of the supporting data, and will not be described again here. A very small positive number pre-stored for the controller, for example, The value is 0.001.

[0042] Optionally, the controller can employ a pre-stored rule-matching-based evidence statistics method to identify supporting data for each relation edge based on structured business data and factual data of multiple entities involved in the path to be identified, thereby obtaining the supporting value for each relation edge of the path to be identified. It is understood that the controller can employ a pre-stored rule-matching-based evidence statistics method to identify conflict data for each relation edge based on structured business data and factual data of multiple entities involved in the path to be identified, thereby obtaining the conflict value for each relation edge of the path to be identified.

[0043] In one example, the process of identifying the validity period of each relation edge in the path to be identified based on the structured business data and factual data of multiple entities involved in the path to be identified, and obtaining the time validity value of each relation edge in the path to be identified, includes: Based on the structured business data and factual data of multiple entities involved in the path to be identified, the expiration information of each relation edge is parsed to obtain the effective time interval of each relation edge in the path to be identified. Based on the time distance between the effective time interval of each relation edge of the path to be identified and the corresponding identification time of the path to be identified, the deviation value of each relation edge of the path to be identified is obtained. The deviation value is used to characterize the degree of deviation between the effective time interval of the corresponding relation edge and the corresponding identification time of the path to be identified. For each relation edge of the path to be identified, analyze the degree of dispersion of its deviation value and the deviation values ​​of multiple relation edges included in its relation type to obtain the effective time value of each relation edge of the path to be identified.

[0044] Based on this, the controller can use the degree of dispersion of deviation values ​​of the same type of relationship (such as interquartile range) as a scale to convert the deviation value of each relationship edge into a relative time validity value, thereby eliminating the evaluation bias caused by the difference in time sensitivity between different relationship types and realizing the fair quantification of time validity within the same type of relationship.

[0045] For example, the time validity value T(e) of the e-th relation edge can be expressed as: In the formula, This represents the deviation value of the e-th relation edge. The controller obtains the deviation value of the e-th relation edge. The process is as follows: Given the effective time interval [T] of the e-th relation edge. start ,T end The time point T for penetrating association identification with the controller. now If T start ≤T now ≤T end If the identification time point falls within the effective time interval of the e-th relation edge, then the controller can determine the deviation value of the e-th relation edge. =0. If T now <T start That is, if the identification time is earlier than the effective time of the e-th relation edge, then the deviation value of the e-th relation edge is... =T start T now If T now >T end That is, if the identification time is later than the failure time of the e-th relation edge, then the deviation value of the e-th relation edge is... =T now T endIn this embodiment of the invention, when the deviation values ​​of multiple relation edges of the path to be identified are all 0, the controller can directly determine the effective time values ​​of all multiple relation edges of the path to be identified as 1.

[0046] I(r) is the interquartile range of multiple deviation values ​​of the r-th relation type to which the e-th relation edge belongs. I(r) is used to characterize the dispersion of deviation values ​​of all relation edges in the r-th relation type to which the e-th relation edge belongs. Furthermore, I(r) satisfies I(r) = Q3. Q1 is the first quartile, meaning 25% of relation edges have deviations less than this first quartile. Q3 is the third quartile, meaning 75% of relation edges have deviations less than this third quartile. A larger I(r) indicates a more dispersed distribution of deviations across the relation edges of that type; conversely, a smaller I(r) indicates a more concentrated distribution of deviations across the relation edges of that type. For example, suppose the r-th relation type includes 7 relation edges with deviations of 0, 0, 10, 30, 45, 80, and 200. Sorting the deviations from smallest to largest, they are 0, 0, 10, 30, 45, 80, and 200. The controller can determine that the first quartile Q1 is 0 and the third quartile Q3 is 80, thus obtaining I(r) as 80.

[0047] Step S2: For each candidate path, the original strength value of each candidate path is obtained by fusing the one-sided strength value and weight value of the multiple relation edges it includes.

[0048] The original strength value is used to characterize the original association strength of the corresponding candidate path, and the weight value is obtained based on the order of the corresponding relation edge in its candidate path and the total number of relation edges included in its candidate path.

[0049] For example, the original intensity value M(P) of the Pth candidate path can be expressed as: In the formula, B(e) is the weight value of the e-th relation edge. G(e) is the one-sided strength value of the e-th relation edge. A very small positive number pre-stored for the controller, for example, The value is 0.001. This represents the sum of the weights of the multiple relation edges included in the P-th candidate path. This represents the weighted sum of the weights of the multiple relation edges included in the P-th candidate path, divided by the individual edge strength values.

[0050] During the transmission of associations along a path, the association needs to pass through multiple relationship edges from the starting subject to the ending subject. Even if the strength value of the preceding relationship edge is high, if the strength value of the ending relationship edge is low, the association cannot be effectively transmitted to the target subject. Therefore, the closer the relationship edge is to the ending subject, the greater its decisive role in whether a candidate path can constitute a substantial association. Based on this, a weighted average of the edge strength value and weight value is used to avoid misclassifying invalid paths with "strong beginnings and weak endings" as strong association paths.

[0051] In one example, the steps for the controller to obtain the weight value of each relationship edge include: Data analysis is performed based on the order of the target relation edge among the multiple relation edges included in its candidate path, and the total number of relation edges included in its candidate path, to obtain the weight value of the target relation edge. The target relation edge is any relation edge among the multiple relation edges, and the order increases sequentially from the starting body to the ending body in the candidate path. The weight value of the target relation edge is positively correlated with its order among the multiple relation edges included in its candidate path, and negatively correlated with the total number of relation edges included in its candidate path.

[0052] For example, the weight B(e) of the e-th relation edge can be expressed as: It is understandable that the weight value of the e-th relation edge can also be expressed as... In the formula, Let be the order of the e-th relation edge among the multiple relation edges included in its candidate path. Let n be the total number of relation edges included in the candidate path to which the e-th relation edge belongs. For example, a candidate path might be ABCD. Here, A is the starting point, D is the ending point, the relation edge between A and B is the first relation edge, the relation edge between B and C is the second relation edge, and the relation edge between C and D is the third relation edge. That is, the relation edge between A and B is ranked 1st among the three relation edges included in this candidate path, the relation edge between B and C is ranked 2nd, and the relation edge between C and D is ranked 3rd. The controller can then determine that the weight of the first relation edge is... The weight of the second relation edge is... and the weight of the third relation edge is .

[0053] Step S3: For each candidate path, correct its original strength value based on its reliability value to obtain the comprehensive strength value of each candidate path.

[0054] Among them, the comprehensive strength value is used to characterize the final association strength of the corresponding candidate path after removing interference, and the reliability value is used to characterize the credibility of the corresponding candidate path after removing interference. The reliability value is obtained based on the degree of sharing of the intermediate subjects of the corresponding candidate path, the relative strength of the terminal relation edges, the proportion of relation edges with a single edge strength value lower than the threshold, and the time validity value of each relation edge.

[0055] Specifically, the reliability value is jointly represented by the path quality value and the validity value. The path quality value is used to characterize the reliability of the corresponding candidate path structure, and the validity value is used to characterize the validity of the corresponding candidate path. For each candidate path, its original strength value is corrected based on its reliability value to obtain the comprehensive strength value of each candidate path, including: For each candidate path, its original strength value is corrected based on its path quality value to obtain the initial strength value of each candidate path; For each candidate path, its initial strength value is corrected based on its effective value to obtain the comprehensive strength value of each candidate path.

[0056] The more intermediate entities (excluding the starting and ending entities) a candidate path connects to, the more unrelated entities that intermediate entities are to the candidate path. The more easily these intermediate entities are shared by different paths, the weaker their ability to distinguish specific paths, and the lower the path quality. Furthermore, the higher the single-edge strength value of the terminal relationship edge connected to the ending entity relative to the average value of the preceding relationship edges of the candidate path, the more likely the candidate path is to constitute a substantial association; that is, the stronger the terminal connection capability of the candidate path, and the greater the association strength. Based on this, the controller uses both shared value and endpoint value to characterize the path quality value. The terminal relationship edge is used to characterize the relationship edge connected to the ending entity among the multiple relationship edges included in the candidate path.

[0057] In one example, the path quality value is jointly represented by a shared value and a landing value. The shared value is used to characterize the independence of the intermediate entities of the corresponding candidate path, and the landing value is used to characterize the relative strength of the unilateral strength value of the terminal relationship edge of the corresponding candidate path. The process of correcting the original strength value of each candidate path based on its path quality value to obtain the initial strength value of each candidate path includes: For each candidate path, a common value is calculated based on the number of subjects connected by the intermediate subject and the maximum number of connections of all intermediate subjects in the subject type to which the intermediate subject belongs. The common value is negatively correlated with the number of subjects connected by the intermediate subject of the corresponding candidate path, and positively correlated with the maximum number of connections of the subject type to which the intermediate subject of the corresponding candidate path belongs. For each candidate path, the landing point value is calculated based on the single-sided strength value of its terminal relation edge and the strength values ​​of its multiple relation edges. The landing point value is positively correlated with the single-sided strength value of the terminal relation edge of the corresponding candidate path, and negatively correlated with the single-sided strength values ​​of the other relation edges of the corresponding candidate path excluding the terminal relation edge. For each candidate path, its common value, landing point value, and original intensity value are multiplied to obtain the initial intensity value of each candidate path.

[0058] For example, the initial intensity value S(P) of the Pth candidate path can be expressed as: In the formula, M(P) is the original intensity value of the Pth candidate path. U(P) is the shared value of the Pth candidate path. Z(P) is the landing value of the Pth candidate path.

[0059] For example, the shared value U(P) of the Pth candidate path can be expressed as: In the formula, H(v) is the ratio of the number of entities connected to the v-th intermediate entity in the P-th candidate path to the maximum number of connections of all intermediate entities in the entity type to which the v-th intermediate entity belongs. H(v) is used to characterize the sharing degree of the v-th intermediate entity, and Avg[H(v)] is the average level of the sharing degree of all intermediate entities in the P-th candidate path. For example, if the P-th candidate path has 3 intermediate entities, intermediate entity 1 is a corporate entity, and the number of entities it connects to is 8, and the maximum number of connections of the corporate entity to which intermediate entity 1 belongs is 100, then the sharing degree of intermediate entity 1 can be expressed as H(1) = 8 / 100 = 0.08. Similarly, H(2) = 0.15, H(3) = 0.4. Based on this, the controller can obtain the sharing value U(P) = 1 for the P-th candidate path. (0.08+0.15+0.4) / 3=0.79.

[0060] In this embodiment of the invention, the process by which the controller obtains the landing value Z(P) of the Pth candidate path is as follows: If the Pth candidate path has only one relation edge, the controller directly sets the landing value Z(P) of the Pth candidate path to 1. If the Pth candidate path has multiple relation edges, the controller can calculate the average strength of the individual edge strength values ​​of the multiple relation edges of the Pth candidate path, excluding the terminal relation edge. Subsequently, the controller can determine the landing value of the Pth candidate path. Where G(m) is the single-sided strength value of the terminal relation edge of the Pth candidate path, Avg[G pre[ ] represents the average intensity of the individual edge strength values ​​of the multiple relation edges in the P-th candidate path, excluding the terminal relation edge.

[0061] The more weak edges a candidate path has, the more it relies on weak signals (such as common address relationships) for support, and the lower its effectiveness. Furthermore, as mentioned earlier, the closer a relation edge is to the endpoint, the greater its importance. Therefore, the closer a relation edge is to the endpoint in a candidate path, the lower its temporal validity value, and the more severely it weakens the effectiveness of that candidate path. Based on this, this invention combines the proportion of weak edges (i.e., the weak edge ratio) and the degree of temporal decay (i.e., the decay value) to jointly characterize the validity value, correcting the initial strength value to obtain a comprehensive strength value for each candidate path. This effectively distinguishes between high-value paths supported by strong relationships and low-value paths supported by weak signals (or failed relationships), thereby improving the accuracy of penetrating association identification results.

[0062] In one example, the effective value is obtained based on the weak edge ratio and attenuation value analysis. The weak edge ratio is used to characterize the intensity of weak edge interference in the corresponding candidate path, and the attenuation value is used to characterize the degree of weakening of the association strength of the corresponding candidate path due to the change of the effective value over time with the importance of location. The process of correcting the initial strength value of each candidate path based on its effective value to obtain the comprehensive strength value of each candidate path includes: For each candidate path, the weak edge ratio of each candidate path is calculated by the ratio of the sum of the single-edge strength values ​​of the weak edges it includes to the sum of the single-edge strength values ​​of the multiple relation edges it includes. The weak edges are used to indicate relation edges whose single-edge strength values ​​are lower than the strength threshold. For each candidate path, a weighted average is calculated based on the time validity value and weight value of each relation edge to obtain the decay value of each candidate path; For each candidate path, the effective value of each candidate path is calculated based on its weak edge ratio and decay value. The effective value is negatively correlated with the weak edge ratio of the corresponding candidate path and negatively correlated with the decay value of the corresponding candidate path. For each candidate path, calculate the product of its effective value and its initial strength value to obtain the comprehensive strength value of each candidate path.

[0063] For example, the comprehensive strength value V(P) of the Pth candidate path can be expressed as: In the formula, S(P) is the initial strength value of the Pth candidate path. W(P) is the weak edge ratio of the Pth candidate path. R(P) is the attenuation value of the Pth candidate path.

[0064] For example, the ratio of weak edges W(P) of the Pth candidate path can be expressed as: In the formula, Let represent the sum of the individual strength values ​​of the weak edges in the P-th candidate path. For the eth w The strength value of a single edge of a weak edge. The strength threshold can be set empirically to 0.4. Based on this, the controller can identify the edges with a single edge strength value lower than 0.4 among the multiple relation edges of the Pth candidate path as weak edges in the Pth candidate path. This represents the cumulative sum of the individual strength values ​​of the multiple relation edges included in the P-th candidate path.

[0065] For example, the decay value R(P) of the Pth candidate path can be expressed as: In the formula, n is the total number of relation edges included in the Pth candidate path, which is the total number of relation edges included in the candidate path to which the eth relation edge belongs. B(e) is the weight value of the eth relation edge, and the weight value of each relation edge is used to determine the impact of its time validity on the overall decay degree. T(e) is the time validity value T(e) of the eth relation edge.

[0066] Step S4: Analyze the relative magnitudes of the comprehensive strength values ​​of multiple candidate paths to obtain the target path corresponding to the starting subject, and determine the multiple subjects included in the target path as the associated subjects of the starting subject.

[0067] Among them, the target path is used to characterize the path with the highest correlation strength with the starting subject among multiple candidate paths.

[0068] Specifically, the controller can compare the comprehensive strength values ​​of multiple candidate paths, determine the candidate path with the highest comprehensive strength value as the target path, and determine all entities (including intermediate entities and end entities) traversed by the target path as associated entities of the starting entity.

[0069] Optionally, if there are multiple candidate paths with the highest overall strength value, i.e., multiple candidate paths that are tied for the strongest, the controller can compare the initial strength values ​​of these multiple strongest candidate paths and determine the candidate path with the highest initial strength value among them as the target path. If they are still the same, the controller continues to compare the weak edge ratios and determines the candidate path with the lower weak edge ratio as the target path; if the weak edge ratios are still the same, the controller continues to compare the attenuation values ​​and determines the candidate path with the lower attenuation value as the target path.

[0070] In this embodiment of the invention, before step S1, the method further includes: based on the subject identifier of the target enterprise, obtaining enterprise subjects, natural person subjects, address subjects and various relationship edges (including equity relationship, investment relationship, employment relationship, agreement relationship, shared address relationship and factual relationship) related to the target enterprise from the above-mentioned external data source, and constructing a multi-relationship knowledge graph centered on the target enterprise; then, using a graph traversal algorithm (such as breadth-first traversal algorithm) to perform path search on the multi-relationship knowledge graph, obtaining multiple candidate paths starting from the target enterprise and ending at other subjects.

[0071] In summary, this invention assesses the strength of multiple relationship edges based on historical data, structured business data and factual data of multiple entities involved in multiple candidate paths, obtaining a single-edge strength value for each relationship edge. Then, it fuses the single-edge strength values ​​and weight values ​​of the multiple relationship edges included in each candidate path to obtain the original strength value of each candidate path. Subsequently, it corrects the original strength value based on the reliability value of each candidate path to obtain a comprehensive strength value for each candidate path. Next, it analyzes the relative magnitude of the comprehensive strength values ​​of multiple candidate paths to obtain the target path corresponding to the starting entity, and identifies the multiple entities included in the target path as the associated entities of the starting entity. This invention quantifies the strength differences of different relationship edges through single-edge strength values, avoiding the problem of treating various heterogeneous relationship edges as having equal strength. Combining the weights of relationship edges can strengthen the influence of relationship edges at different positions (orders) on the association strength of candidate paths. Furthermore, the correction through reliability values ​​eliminates interference from common nodes, weak edges, and time-decrease interference in candidate paths. Therefore, this invention improves the accuracy of penetrating association identification based on multi-relationship graphs and fact extraction.

[0072] It should be noted that the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some implementations, multitasking and parallel processing are possible or may be advantageous.

Claims

1. A penetrating association identification method based on multi-relation graphs and fact extraction, characterized in that, The method includes: Based on historical data, structured business data and factual data of multiple subjects involved in multiple candidate paths, the strength of multiple relationship edges is evaluated to obtain the single-edge strength value of each relationship edge. The single-edge strength value is used to characterize the association strength between the subjects connected by the corresponding relationship edge. The starting subjects of multiple candidate paths are the same. For each candidate path, the original strength value of each candidate path is obtained by fusing the single-edge strength value and weight value of the multiple relation edges it includes. The original strength value is used to characterize the original association strength of the corresponding candidate path, and the weight value is obtained by analyzing the order of the corresponding relation edge in its candidate path and the total number of relation edges included in its candidate path. For each candidate path, its original strength value is corrected based on its reliability value to obtain the comprehensive strength value of each candidate path. The comprehensive strength value is used to characterize the final association strength of the corresponding candidate path after removing interference, and the reliability value is used to characterize the credibility of the corresponding candidate path after removing interference. The reliability value is obtained based on the degree of sharing of the intermediate subjects of the corresponding candidate path, the relative strength of the terminal relation edges, the proportion of relation edges with a single edge strength value lower than the threshold, and the time effective value of each relation edge. By analyzing the relative magnitudes of the comprehensive strength values ​​of multiple candidate paths, the target path corresponding to the starting subject is obtained, and the multiple subjects included in the target path are identified as the associated subjects of the starting subject. The target path is used to characterize the path with the highest association strength with the starting subject among multiple candidate paths.

2. The penetrating association identification method based on multi-relation graph and fact extraction according to claim 1, characterized in that, The structured business data and factual data of multiple entities involved in multiple candidate paths, based on historical data, are used to evaluate the strength of multiple relationship edges, obtaining the single-edge strength value of each relationship edge, including: Based on historical data, scenario relevance statistics are performed on each relation edge involved in the path to be identified to obtain the scenario credibility value of each relation edge of the path to be identified. The path to be identified is any one of multiple candidate paths. The scenario credibility value is used to characterize the credibility of the relation type to which the corresponding relation edge belongs under the same historical business scenario. The relation types include: equity relationship, investment relationship, employment relationship, agreement relationship, shared address relationship and factual relationship. Based on the structured business data and factual data of multiple subjects involved in the path to be identified, the relative strength level of each relation edge in its relation type is analyzed to obtain the relative single-edge strength value of each relation edge in the path to be identified. The relative single-edge strength value is used to characterize the relative strength of the corresponding relation edge in its relation type. Based on the structured business data and factual data of multiple entities involved in the path to be identified, a credibility analysis is performed on the data source of each relation edge of the path to be identified to obtain the source credibility value of each relation edge of the path to be identified. The source credibility value is used to characterize the credibility of the data source of the corresponding relation edge. Based on the structured business data and factual data of multiple entities involved in the path to be identified, the validity period of each relation edge of the path to be identified is identified, and the time validity value of each relation edge of the path to be identified is obtained. The time validity value is used to characterize the validity of the corresponding relation edge. For each relation edge of the path to be identified, the single-edge strength value of each relation edge is calculated by multiplying its scene confidence value, relative single-edge strength value, source confidence value, and time validity value.

3. The penetrating association identification method based on multi-relation graphs and fact extraction according to claim 2, characterized in that, The process of performing scene relevance statistics on each relation edge involved in the path to be identified based on historical data to obtain the scene confidence value of each relation edge of the path to be identified includes: For each relation edge of the path to be identified, its relation type is determined based on historical data; For each relationship type, the frequency is counted based on historical data to obtain the number of times each relationship type appears in historical high-value paths and the number of times each relationship type appears in historical low-value paths. High-value paths are used to represent paths that have been confirmed to have a high degree of correlation in the same historical business scenario, and low-value paths are used to represent paths that have been confirmed to have a low degree of correlation in the same historical business scenario. For each type of relationship, the relative frequency of each type of relationship in the high-value path is calculated by comparing the number of times it appears in the historical high-value path with the total number of times all types of relationships appear in the historical high-value path. For each type of relationship, the relative frequency of each relationship type in low-value paths is calculated by comparing the number of times it appears in historical low-value paths with the total number of times multiple relationship types appear in historical low-value paths. For each relation type, data analysis is performed based on its relative frequency in high-value paths and relative frequency in low-value paths to obtain the scenario credibility value for each relation type. The scenario credibility value of each relation edge in the path to be identified is equal to the scenario credibility value of its relation type. The scenario credibility value is positively correlated with the relative frequency in high-value paths of the corresponding relation type, and negatively correlated with the relative frequency in low-value paths of the corresponding relation type.

4. The penetrating association identification method based on multi-relation graph and fact extraction according to claim 2, characterized in that, The structured business data and factual data of multiple entities involved in the path to be identified are used to analyze the relative strength level of each relation edge in its respective relation type, thereby obtaining the relative unilateral strength value of each relation edge in the path to be identified, including: For each relation edge, based on the structured business data and factual data of multiple subjects involved in the path to be identified, numerical extraction is performed to obtain the original value of each relation edge, where the original value is used to characterize the original strength of the corresponding relation edge; For each relation edge, a relative single-edge strength value is calculated based on its original value, the original values ​​of the multiple relation edges included in its relation type, and the total number of relation edges included in its relation type. The relative single-edge strength value is positively correlated with the original value of the corresponding relation edge, negatively correlated with the original values ​​of other relation edges in the relation type to which the corresponding relation edge belongs, and negatively correlated with the total number of relation edges included in the relation type to which the corresponding relation edge belongs.

5. The penetrating association identification method based on multi-relation graph and fact extraction according to claim 2, characterized in that, Based on the structured business data and factual data of multiple entities involved in the path to be identified, a credibility analysis is performed on the data source of each relation edge in the path to be identified to obtain the source credibility value of each relation edge in the path to be identified, including: Based on the structured business data and factual data of multiple entities involved in the path to be identified, the supporting data of each relation edge is identified to obtain the supporting value of each relation edge of the path to be identified. The supporting data is used to characterize the evidence of the existence of the relation edge obtained from different sources. Based on the structured business data and factual data of multiple entities involved in the path to be identified, the conflict data of each relation edge is identified to obtain the conflict value of each relation edge of the path to be identified. The conflict data is used to characterize the evidence that negates the relation edge obtained from data from different sources. For each relation edge of the path to be identified, data analysis is performed based on its support value and conflict value to obtain the source confidence value of each relation edge of the path to be identified. The source confidence value is positively correlated with the support value of the corresponding relation edge, and the source confidence value is negatively correlated with the conflict value of the corresponding relation edge.

6. The penetrating association identification method based on multi-relation graph and fact extraction according to claim 2, characterized in that, The process of identifying the validity period of each relation edge in the path to be identified, based on the structured business data and factual data of multiple entities involved in the path to be identified, and obtaining the time validity value of each relation edge in the path to be identified, includes: Based on the structured business data and factual data of multiple entities involved in the path to be identified, the expiration information of each relation edge is parsed to obtain the effective time interval of each relation edge in the path to be identified. Based on the time distance between the effective time interval of each relation edge of the path to be identified and the corresponding identification time of the path to be identified, the deviation value of each relation edge of the path to be identified is obtained. The deviation value is used to characterize the degree of deviation between the effective time interval of the corresponding relation edge and the corresponding identification time of the path to be identified. For each relation edge of the path to be identified, analyze the degree of dispersion of its deviation value and the deviation values ​​of multiple relation edges included in its relation type to obtain the effective time value of each relation edge of the path to be identified.

7. The penetrating association identification method based on multi-relation graph and fact extraction according to claim 1, characterized in that, The reliability value is jointly represented by the path quality value and the validity value. The path quality value is used to characterize the reliability of the corresponding candidate path structure, and the validity value is used to characterize the validity of the corresponding candidate path. For each candidate path, its original strength value is corrected based on its reliability value to obtain the comprehensive strength value of each candidate path, including: For each candidate path, its original strength value is corrected based on its path quality value to obtain the initial strength value of each candidate path; For each candidate path, its initial strength value is corrected based on its effective value to obtain the comprehensive strength value of each candidate path.

8. The penetrating association identification method based on multi-relation graph and fact extraction according to claim 7, characterized in that, The path quality value is characterized by a shared value and a landing value. The shared value represents the independence of the intermediate entities of the corresponding candidate path, and the landing value represents the relative strength of the single-sided strength value of the terminal relationship edge of the corresponding candidate path. For each candidate path, its original strength value is corrected based on its path quality value to obtain an initial strength value for each candidate path, including: For each candidate path, a common value is calculated based on the number of entities connected by its intermediate entities and the maximum number of connections among all intermediate entities of the entity type to which the intermediate entities belong. The entity type includes at least one of the following: corporate entities, natural person entities, and address entities. The common value is negatively correlated with the number of entities connected by the intermediate entities of the corresponding candidate path, and positively correlated with the maximum number of connections among the entity types to which the intermediate entities of the corresponding candidate path belong. For each candidate path, the landing point value is calculated based on the single-sided strength value of its terminal relation edge and the strength values ​​of its multiple relation edges. The landing point value is positively correlated with the single-sided strength value of the terminal relation edge of the corresponding candidate path, and negatively correlated with the single-sided strength values ​​of the other relation edges of the corresponding candidate path excluding the terminal relation edge. For each candidate path, its common value, landing point value, and original intensity value are multiplied to obtain the initial intensity value of each candidate path.

9. The penetrating association identification method based on multi-relation graph and fact extraction according to claim 7, characterized in that, The effective value is obtained based on the weak edge ratio and attenuation value analysis. The weak edge ratio is used to characterize the intensity of weak edge interference in the corresponding candidate path, and the attenuation value is used to characterize the degree of weakening of the association strength of the corresponding candidate path due to the change of the effective value over time with the importance of location. For each candidate path, its initial strength value is corrected based on its effective value to obtain the comprehensive strength value of each candidate path, including: For each candidate path, the weak edge ratio of each candidate path is calculated by the ratio of the sum of the single-edge strength values ​​of the weak edges it includes to the sum of the single-edge strength values ​​of the multiple relation edges it includes. The weak edges are used to indicate relation edges whose single-edge strength values ​​are lower than the strength threshold. For each candidate path, a weighted average is calculated based on the time validity value and weight value of each relation edge to obtain the decay value of each candidate path; For each candidate path, the effective value of each candidate path is calculated based on its weak edge ratio and decay value. The effective value is negatively correlated with the weak edge ratio of the corresponding candidate path and negatively correlated with the decay value of the corresponding candidate path. For each candidate path, calculate the product of its effective value and its initial strength value to obtain the comprehensive strength value of each candidate path.

10. The penetrating association identification method based on multi-relation graph and fact extraction according to claim 1, characterized in that, The steps to obtain the weight value of each relation edge include: Data analysis is performed based on the order of the target relation edge among the multiple relation edges included in its candidate path, and the total number of relation edges included in its candidate path, to obtain the weight value of the target relation edge. The target relation edge is any relation edge among the multiple relation edges, and the order increases sequentially from the starting body to the ending body in the candidate path. The weight value of the target relation edge is positively correlated with its order among the multiple relation edges included in its candidate path, and negatively correlated with the total number of relation edges included in its candidate path.

Images