A knowledge graph updating method, device and equipment and readable storage medium

Abstract

The application relates to the technical field of knowledge graphs, and discloses a knowledge graph updating method, device and equipment and a readable storage medium, which comprises the following steps: acquiring public text data related to a target enterprise; using a preset large model to analyze the public text data and generate a candidate triple set representing the target enterprise and its associated subject events; performing credibility evaluation and semantic alignment processing on the candidate triple set to obtain a high-credibility event knowledge set; performing conflict detection on the high-credibility event knowledge set based on preset knowledge graph constraints to determine a target knowledge set; and writing the target knowledge set into an enterprise risk knowledge graph to generate updated enterprise risk knowledge graph data. The application significantly improves the automation degree and accuracy of enterprise risk knowledge graph updating, reduces the cost of manual participation and rule maintenance, and provides more stable and reliable knowledge support for enterprise risk identification, early warning and analysis.

Description

Technical Field

[0001] This application relates to the field of knowledge graph technology, and in particular to a knowledge graph updating method, apparatus, device, and readable storage medium. Background Technology

[0002] Knowledge graphs are a technological system that organizes and represents knowledge using a graph structure. Nodes represent entities, and edges represent semantic relationships between entities, integrating scattered structured and unstructured data into a unified semantic network. They have been widely applied in fields such as intelligent question answering, recommendation systems, and semantic search. With the continuous increase in the scale and frequency of knowledge updates, existing technologies have gradually formed mechanisms such as dynamic updates based on information extraction, representation updates based on incremental learning, and knowledge correction based on change detection. Furthermore, large language models are being introduced to generate entity and relation triples from public texts, documents, news, and other data. Combined with certain confidence scores and consistency rules, new knowledge is integrated into existing knowledge graphs to enhance their coverage and semantic expressive power.

[0003] However, existing knowledge graph update methods based on large language models still have significant limitations. On the one hand, the knowledge generated by the model is uncertain, easily introducing illusory information or logical errors. They lack robust credibility assessment and adaptive filtering mechanisms, making it difficult to promptly remove low-quality or inconsistent candidate knowledge, which can easily lead to knowledge conflicts, redundancy, and inconsistencies. On the other hand, existing methods are insufficient in aligning old and new knowledge, perceiving fine-grained changes, and controlling semantic drift. The update granularity is coarse, and the reasoning overhead of large models is substantial, making it difficult to meet the high-frequency or near-real-time update needs of large-scale knowledge graphs. Furthermore, it is difficult to achieve efficient and reliable dynamic evolution while ensuring knowledge quality and consistency. Summary of the Invention

[0004] In view of this, embodiments of this application provide a knowledge graph updating method, apparatus, device, and readable storage medium, which can effectively solve the problems in the prior art such as difficulty in controlling the credibility of generated knowledge, difficulty in timely identification of knowledge conflicts, and low updating efficiency when dynamically updating knowledge graphs based on large language models.

[0005] In a first aspect, embodiments of this application provide a knowledge graph updating method, including: Obtain publicly available text data related to the target company; The publicly available text data is parsed using a pre-defined large model to generate a set of candidate triples representing the target enterprise and its related events; The set of candidate triples is subjected to credibility evaluation and semantic alignment to obtain a set of highly reliable event knowledge. Based on preset knowledge graph constraints, conflict detection is performed on the high-confidence event knowledge set to determine the target knowledge set; The target knowledge set is written into the enterprise risk knowledge graph to generate updated enterprise risk knowledge graph data.

[0006] In some embodiments, obtaining publicly available text data related to the target enterprise includes: Obtain the identification information representing the target enterprise; Based on the identification information, text data related to the target enterprise is retrieved from a preset public text data source to obtain a candidate public text set; The candidate public text set is subjected to relevance filtering and time range filtering to obtain the target public text data, and the target public text data is used as the public text data.

[0007] In some embodiments, the step of parsing the publicly available text data using a preset large model to generate a set of candidate triples representing the target enterprise and its associated events includes: Retrieve the preset relation template and sample triplet set; Based on the relation template and the set of example triples, construct a prompt message corresponding to the public text data; The publicly available text data and the prompt information are input into the large model to obtain an initial set of triplet results representing the target enterprise and its related events; The initial triplet result set is subjected to format validation and relation type filtering to generate a candidate triplet set, which is then used as the candidate triplet set representing the target enterprise and its related subject events.

[0008] In some embodiments, the process of performing credibility evaluation and semantic alignment on the candidate triple set to obtain a highly credible event knowledge set includes: Based on a preset evaluation strategy, the candidate triplet set is analyzed from multiple perspectives to obtain a set of multi-perspective evaluation results corresponding to each candidate triplet. The multi-perspective evaluation result set is converted into a semantic representation characterizing each candidate triple, and the evaluation index of each candidate triple is determined based on the semantic representation. The candidate triplet set is screened according to the evaluation criteria to obtain a set of highly reliable event knowledge.

[0009] In some embodiments, converting the multi-view evaluation result set into a semantic representation characterizing each candidate triplet, and determining the evaluation metric for each candidate triplet based on the semantic representation, includes: Each candidate triplet is input into the teacher model and the student model respectively to obtain the corresponding first semantic representation and second semantic representation; The distillation error of each candidate triplet is determined based on the first semantic representation and the second semantic representation; The distillation error is combined with the similarity index corresponding to the multi-view evaluation result to generate the evaluation index for each candidate triplet.

[0010] In some embodiments, the step of performing conflict detection on the high-confidence event knowledge set based on preset knowledge graph constraints to determine the target knowledge set includes: Obtain the set of historical event knowledge related to the set of highly credible event knowledge in the enterprise risk knowledge graph; Based on entity type constraints and relation constraints, the high-confidence event knowledge set and the historical event knowledge set are compared according to rules to determine the first conflict determination result. The highly reliable event knowledge set and the historical event knowledge set are mapped to a first embedding representation and a second embedding representation, respectively, and the similarity between the two is calculated based on the embedding representation to determine the second conflict determination result; The high-confidence event knowledge set is filtered based on the first conflict determination result and the second conflict determination result to obtain the target knowledge set.

[0011] In some embodiments, writing the target knowledge set into the enterprise risk knowledge graph to generate updated enterprise risk knowledge graph data includes: The target knowledge set is converted into graph entity data and relation data that conform to the enterprise risk knowledge graph storage structure; Based on the entity data and relation data of the graph, insert and / or update operations are performed on the enterprise risk knowledge graph to obtain updated enterprise risk knowledge graph data.

[0012] Secondly, embodiments of this application provide a knowledge graph updating apparatus, including: The data acquisition module is used to acquire publicly available text data related to the target company; The data processing module is used to parse the publicly available text data using a preset large model to generate a set of candidate triples representing the target enterprise and its related events. The evaluation module is used to perform credibility evaluation and semantic alignment processing on the candidate triple set to obtain a highly reliable event knowledge set. The detection module is used to perform conflict detection on the high-confidence event knowledge set based on preset knowledge graph constraints, and to determine the target knowledge set; The graph acquisition module is used to write the target knowledge set into the enterprise risk knowledge graph and generate updated enterprise risk knowledge graph data.

[0013] Thirdly, embodiments of this application provide a terminal device, the terminal device including a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the knowledge graph updating method of the first aspect described above.

[0014] Fourthly, embodiments of this application provide a computer-readable storage medium, wherein when the computer program is executed on a processor, it implements the knowledge graph updating method of the first aspect described above.

[0015] The embodiments of this application have the following beneficial effects: By acquiring publicly available text data related to the target enterprise, parsing the publicly available text data using a preset large model, generating a set of candidate triples representing events of the target enterprise and its related entities, performing credibility assessment and semantic alignment processing on the candidate triples set to obtain a set of highly credible event knowledge, performing conflict detection on the highly credible event knowledge set based on preset knowledge graph constraints, determining the target knowledge set, and writing the target knowledge set into the enterprise risk knowledge graph to generate updated enterprise risk knowledge graph data; This application can automatically complete the structured extraction, screening, and writing of enterprise risk events based on publicly available text data, reducing the workload of manual processing and rule maintenance, improving the reliability and consistency of event knowledge, reducing the interference of conflict and redundant information on the knowledge graph, and enabling the enterprise risk knowledge graph to maintain continuous updates and structural integrity in multi-source text input scenarios, thereby providing a more accurate and timely knowledge data foundation for subsequent enterprise risk identification, early warning, and analysis. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart of a knowledge graph updating method according to an embodiment of this application is shown; Figure 2 Another flowchart of the knowledge graph updating method according to an embodiment of this application is shown; Figure 3 This paper illustrates yet another flowchart of a knowledge graph updating method according to an embodiment of this application; Figure 4This illustration shows a structural diagram of a knowledge graph updating method according to an embodiment of this application. Detailed Implementation

[0018] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0019] The components of the embodiments of this application described and illustrated in the accompanying drawings can be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of this application provided in the drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0020] In the following text, the terms "comprising," "having," and their cognates, which may be used in various embodiments of this application, are intended only to indicate a particular feature, number, step, operation, element, component, or combination thereof, and should not be construed as primarily excluding the presence of one or more other features, numbers, steps, operations, elements, components, or combinations thereof, or adding the possibility of one or more combinations thereof. Furthermore, the terms "first," "second," "third," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance.

[0021] Unless otherwise specified, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of this application pertain. Terms (such as those defined in commonly used dictionaries) shall be interpreted as having the same meaning as in their contextual meaning in the relevant technical field and shall not be construed as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of this application.

[0022] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0023] Considering the shortcomings of existing knowledge graph updates based on large language models in terms of knowledge credibility control, knowledge conflict identification, and update efficiency, this application proposes a knowledge graph update method. This method performs credibility evaluation, semantic alignment, and knowledge conflict detection on candidate triples generated by the large model, and writes the selected target knowledge increments into the enterprise risk knowledge graph.

[0024] The following examples illustrate the knowledge graph updating method.

[0025] Figure 1 A flowchart of a knowledge graph updating method according to an embodiment of this application is shown. Exemplarily, the knowledge graph updating method includes the following steps: Step S100: Obtain publicly available text data related to the target company.

[0026] The "target enterprises" category indicates the enterprises whose risk event knowledge needs to be updated in the enterprise risk knowledge graph. "Public text data" refers to text information released to the public through public channels, including financial news reports, listed company announcements, and other publicly available text materials. For example, for a pre-defined set of target enterprises, text content related to each target enterprise can be collected from mainstream financial media and listed company announcements at predetermined time intervals. The collected raw text is then organized according to source, time, and enterprise identifier to form a raw public text dataset for subsequent analysis and processing.

[0027] In an optional embodiment, step S100 includes the following sub-steps: S101, Obtain the identification information representing the target enterprise.

[0028] The identification information is used to determine the range of texts related to the target company in a preset public text data source. For example, it may include the company name, a uniformly used company abbreviation, or other identification fields that can distinguish different companies in financial news and announcements.

[0029] S102, based on the identification information, retrieve text data related to the target enterprise from the preset public text data source to obtain a candidate public text set.

[0030] The preset public text data source refers to the pre-configured financial news, announcements and related text data sources. For example, the identification information can be used to perform conditional searches on the above data sources, and the retrieved texts can be collected into a candidate public text set at the chapter level.

[0031] S103, perform relevance filtering and time range filtering on the candidate public text set to obtain the target public text data, and use the target public text data as the public text data.

[0032] The relevance filtering is used to remove text that is irrelevant to the target company's risk event theme based on preset keyword matching rules or simple semantic matching strategies. The time range filtering is used to limit the text range to a preset time interval based on the publication time of candidate texts or the event time information contained in the text. For example, the title and body of each text in the candidate public text set can be matched with rules to retain only the text that is relevant to the target company and its risk event theme and is within the target time window, and the remaining text can be used to form the target public text data.

[0033] In other implementations, the candidate public text set can be subjected to simple deduplication or merging based on the text source and publication time before relevance screening, so as to reduce the impact of duplicate texts on subsequent processing steps.

[0034] Step S200: Use a pre-set large model to parse the publicly available text data and generate a set of candidate triples representing the target enterprise and its related events.

[0035] Here, the candidate triple set refers to the initial set of event knowledge generated by the large model from publicly available text data, which has not yet undergone credibility assessment and conflict detection. Exemplarily, for a specific target company, news reports and announcements related to that company can be input into a pre-defined large model, requiring the model to... The output result is in the form of, where For head entity, For tail entity, To represent the relationship labels of event types or semantic relationships, a set of candidate triples covering events of the target company and its related entities such as acquirers, acquirees, regulatory agencies, and partners is obtained.

[0036] In one alternative embodiment, such as Figure 2 As shown, step S200 includes the following sub-steps: S201, obtain the preset relation template and sample triplet set.

[0037] The relation template is used to define the types of relations that are allowed in the current knowledge graph update task and their semantic descriptions. It can include relation types such as "acquisition", "investment", "holding", "being penalized", and "signing an agreement" and their corresponding text descriptions. The example triple set is used to provide output examples for the large model. It can select several known triples such as "Company A - Acquisition - Company B" and "Company C - Being Penalized - Regulatory Agency D" as reference samples.

[0038] S202, construct a prompt message corresponding to the public text data based on the relation template and the set of example triples.

[0039] The prompt message is used to indicate the parsing target and output format of the large model. For example, a prompt text can be composed of a list of relation types, example triples, and constraints on the output format (e.g., requiring all identified entities and relations to be listed in standard triple form). Different prompt templates can be selected based on the type of publicly available text data (e.g., news, announcements, or reports). For instance, when processing listed company announcements, the prompt message could include additional explanations such as "Focus on extracting triples related to the transaction's effective status, consideration arrangements, and counterparties" to improve targeting.

[0040] In other implementations, the prompt may also include instructions to guide the large model through step-by-step reasoning, such as requiring the identification of the main events in the text first, and then extracting the subjects and relationships involved in sequence, in order to adapt to the parsing needs of complex and long texts.

[0041] S203. Input the publicly available text data and prompts into the large model to obtain an initial set of triplet results representing the target company and its related events.

[0042] The initial triplet result set refers to the set of triplet text results directly generated by the large model under the current prompt information constraints. For example, the large model can be called multiple times for the same public text data, each time with slightly different prompt expressions or sampling parameters, to obtain multiple initial triplet results, which are then organized by text line or record to cover different potential event expressions and semantic understanding results in the text.

[0043] S204. Perform format validation and relation type filtering on the initial triplet result set to generate a candidate triplet set, and use the candidate triplet set as a candidate triplet set representing the target enterprise and its related subject events.

[0044] The format validation check is used to verify whether each initial triplet meets the preset structural requirements, such as whether it contains the three elements of a header entity, relation, and tail entity, and whether it conforms to the requirements. The basic form; relation type filtering is used to compare the relation fields in the initial triples based on the relation template obtained in step S201, and remove triples that are not in the allowed relation set. For example, each record in the initial triple result set can be parsed first, and text records that do not conform to the triple format can be filtered out. Then, the relation fields in the remaining triples can be mapped to a preset relation type set, and triples that cannot match a valid relation type can be deleted. Based on this, simple deduplication and merging operations are performed to finally obtain a set of candidate triples with a standardized structure and controlled relation types.

[0045] Step S300: Perform credibility evaluation and semantic alignment on the candidate triple set to obtain a highly credible event knowledge set.

[0046] Exemplarily, multi-perspective evaluation results are constructed for each candidate triple, and based on this, semantic representations and evaluation metrics are extracted. The candidate triples determined to be reliable via a unified evaluation strategy are aggregated into a high-confidence event knowledge set for subsequent conflict detection and knowledge graph update.

[0047] In an optional embodiment, step S300 includes the following sub-steps: S301, perform multi-perspective parsing on the candidate triple set based on a preset evaluation strategy to obtain a multi-perspective evaluation result set corresponding to each candidate triple.

[0048] Among them, the preset evaluation strategy is used to indicate how the system rechecks candidate triples from different perspectives, various prompt configurations or generation modes, and the multi-perspective evaluation result set is used to refer to multiple sets of text or vector-level evaluation results obtained for the same candidate triple under this evaluation strategy. Exemplarily, for the original text segment corresponding to each candidate triple, several different prompt templates, decoding parameters or context combinations can be configured, and the large model is called multiple times to generate triple description results semantically related to this candidate triple, and the encoding model is used to encode each generation result to form a result set representing the semantic performance of this candidate triple under different generation perspectives, thereby constituting the multi-perspective evaluation result set. Specifically, for each candidate triple , it is repeatedly generated k times using multiple Prompt templates to obtain a multi-perspective generation result set , where represents the triple text representation generated for the i-th time, and the encoder is used to map each generation result to the vector space, and the following formula is used to calculate the multi- Figure 1 consistency score:

[0049] Among them, is the consistency index to be calculated, which is used to measure the aggregation degree of the embedding vectors of a group of objects; is the normalization factor; is the total number of objects participating in the calculation; is the sum over all object pairs satisfying i < j; represents the object and of the cosine similarity of the embedding vectors, where is the embedding function (mapping an object to a vector), and the cosine similarity is used to measure the directional similarity of two vectors (the value range is 1, 1], and the closer to 1, the more similar the vector directions).

[0050] For example, for the candidate triple "Company A - Acquisition - Company B", the system can generate related triples or explanatory text multiple times using different prompts such as "extracting factual events", "judging transaction status", and "listing the involved entities and relationships". The system will collect the results from each iteration and use them for the next step of semantic representation and evaluation index calculation.

[0051] S302, convert the multi-perspective evaluation result set into a semantic representation of each candidate triplet, and determine the evaluation index of each candidate triplet based on the semantic representation.

[0052] Semantic representation refers to the vector form that represents the semantic features of candidate triples within a unified vector space. Evaluation metrics measure the consistency of candidate triples across multiple perspectives and their proximity to the teacher's knowledge distribution. Exemplarily, the system first encodes the text representations of each perspective in the multi-perspective evaluation result set using a sentence vector encoding model. The resulting vectors are used to calculate the cosine similarity between perspectives, thus obtaining the first-class metric C_{h,r,t} reflecting multi-perspective consistency. Subsequently, the system combines the teacher and student models to construct a self-distillation structure, using the distillation error as the second-class metric to measure the semantic stability of the candidate triples.

[0053] In one optional implementation, step S302 includes the following sub-steps: Each candidate triple is input into the teacher model and the student model respectively to obtain the corresponding first semantic representation and second semantic representation. The teacher model... For frozen large language models or their encoding networks, used to generate candidate triples. High-dimensional semantic representation Student Model This is a lightweight knowledge discrimination network that takes the same candidate triples as input and outputs the corresponding low-dimensional semantic representation. .

[0054] The distillation error of each candidate triplet is determined based on the first semantic representation and the second semantic representation, and its loss function is defined as: ; in, This is due to distillation losses; The representation vector output by the Teacher Model; The representation vector output by the Student Model; This represents the square of the L2 norm, used to calculate the squared Euclidean distance between two vectors.

[0055] The distillation error is combined with the similarity index corresponding to the multi-view evaluation results to generate an evaluation index for each candidate triple. Specifically, the comprehensive credibility score quantifies the reliability of candidate knowledge by weighting and fusing the consistency score with the distillation error; its calculation formula is as follows:

[0056] in, Let σ(·) represent the trust probability, and let σ(·) denote the Sigmoid function, used to normalize the rating to the [0,1] interval. and These are weighting coefficients used to balance the relative contributions of consistency and distillation error; This is a consistency metric for the embedded vectors; This represents the knowledge distillation loss. The smaller the value, the better the alignment between the output representations of the student model and the teacher model.

[0057] In other implementations, normalization or threshold pruning operations may be introduced when generating evaluation metrics to maintain consistency in evaluation scales across different batches of candidate triples.

[0058] S303. Based on the evaluation criteria, the candidate triple set is screened to obtain a set of highly reliable event knowledge.

[0059] The evaluation metrics serve as the selection criteria to differentiate between candidate triples in terms of credibility and semantic consistency. The high-credibility event knowledge set refers to the set of event knowledge retained through this selection process that meets the preset credibility requirements. For example, for each candidate triple... Calculate the corresponding overall credibility score and with preset threshold Comparison, when the credibility score meets > (in When the threshold is reached, the candidate triple is added to the high-confidence knowledge set. These triples are combined into a set of highly reliable event knowledge as highly reliable event knowledge.

[0060] In other implementations, candidate triples can be classified and managed according to evaluation indicators. For example, candidate triples can be divided into different levels such as high credibility, pending manual review, and low credibility. Only the high credibility part can be directly written into the subsequent processing flow, while the part pending review can be handed over to manual review or delayed processing, so as to adapt to the knowledge quality control requirements of different business scenarios.

[0061] Step S400: Based on the preset knowledge graph constraints, perform conflict detection on the high-confidence event knowledge set to determine the target knowledge set.

[0062] The high-credibility event knowledge set refers to the set of event triples that have undergone credibility assessment and semantic alignment. The target knowledge set refers to the subset of event knowledge that is retained after comparison with the existing enterprise risk knowledge graph and meets the preset consistency requirements. The preset knowledge graph constraints are used to uniformly describe the judgment rules regarding entity type, relationship type, and factual consistency in the enterprise risk knowledge graph. Exemplarily, based on these constraints, the high-credibility event knowledge set is compared line by line with the historical event knowledge in the enterprise risk knowledge graph. Potential conflicts or contradictions are identified at both the rule and semantic levels, and based on this, the target knowledge set that meets the knowledge graph consistency requirements is selected.

[0063] In one alternative embodiment, such as Figure 3 As shown, step S400 includes the following sub-steps: S401, Obtain the set of historical event knowledge related to the set of highly credible event knowledge in the enterprise risk knowledge graph.

[0064] The enterprise risk knowledge graph stores risk event triples and their attribute information about the target enterprise and its related entities. The historical event knowledge set refers to existing event records in the knowledge graph that are associated with entities or relationships involved in the high-confidence event knowledge set. Exemplarily, the head entity, relationship, and tail entity of each triple in the high-confidence event knowledge set are first parsed. Based on these entities or relationships, queries are performed in the enterprise risk knowledge graph to retrieve existing event triples related to the same entity and the same or similar relationship types. These triples are then categorized by entity pair, relationship type, or time attribute to form the historical event knowledge set corresponding to the high-confidence event knowledge set.

[0065] For example, for the "Company A - Acquisition - Company B" triple contained in the high-credibility event knowledge set, the system can search for existing triples in the enterprise risk knowledge graph where "Company A" and "Company B" are the subject or object and the relationship type is "acquisition", "proposed acquisition", "terminated acquisition", etc., and collect the search results as the corresponding historical event knowledge records.

[0066] S402, based on entity type constraints and relation constraints, compares the high-confidence event knowledge set with the historical event knowledge set according to rules to determine the first conflict judgment result.

[0067] Entity type constraints are used to ensure that the type definitions of the same entity identifier in the knowledge graph do not contradict each other. Relationship constraints are used to limit the mutual exclusion or inclusion relationships between different relationship types, such as mutually exclusive relationships, pre- and post-relationships, or conditional relationships. The first conflict determination result is used to indicate the conflict or non-conflict conclusion obtained at the rule level. Exemplarily, based on a preset type hierarchy and relationship constraint table, the system compares each triple in the high-confidence event knowledge set with the corresponding triple in the historical event knowledge set: when the same entity is marked as an incompatible type at the same level, or when the same entity has mutually exclusive relationship descriptions within the same time frame, the system marks the triple as a rule conflict record and marks it in the first conflict determination result. In other embodiments, entity type constraints and relationship constraints may also include time consistency constraints to identify descriptions of the same event that cannot be simultaneously valid in the time dimension.

[0068] S403, map the high-credibility event knowledge set and the historical event knowledge set to the first embedding representation and the second embedding representation respectively, and calculate the similarity between the two based on the embedding representation to determine the second conflict determination result.

[0069] In this model, the first embedding representation refers to the representation obtained by vectorizing each triple in the high-confidence event knowledge set, the second embedding representation refers to the representation obtained by vectorizing each triple in the historical event knowledge set, and the second conflict determination result is used to determine whether there is a potential conflict between the two types of representations based on semantic similarity and relation attribute combination. Exemplarily, a pre-defined knowledge graph embedding model or sentence vector encoding model is used to jointly map "head entity—relation—tail entity" to a vector space, and the first embedding representation is calculated for each triple (h,r,t) in the high-confidence event knowledge set. It also retrieves the closest historical event triples within the embedding space of the enterprise risk knowledge graph. The second embedding representation Then, the maximum semantic similarity between the two is calculated based on the following formula:

[0070] in, The confidence level of the target triple (h,r,t) is used to quantify the credibility of the triple; the max operation is used to select the maximum similarity among all comparison items. This indicates traversing the "old triplet set". All triples in “” ; The cosine similarity function; These are the target triple (h, r, t) and the triple in the old set. The corresponding embedding vector.

[0071] When similarity score If the value exceeds a preset threshold and the corresponding relationship type is determined to be mutually exclusive or inconsistent in time, the triple is marked as a semantic conflict candidate, and the corresponding conflict mark is recorded in the second conflict determination result.

[0072] In other implementations, information such as the similarity of entity pairs, the directional differences of relation vectors, and the encoding of temporal attributes can be combined to make a more granular distinction of potential conflicts.

[0073] S404. Based on the first conflict determination result and the second conflict determination result, the set of highly reliable event knowledge is filtered to obtain the target knowledge set.

[0074] The first and second conflict determination results reflect the conflict judgments of candidate knowledge at the rule level and the embedded semantic level, respectively. The target knowledge set refers to the set of event knowledge that is retained under the joint constraints of the two types of determination results and meets the consistency requirements of the knowledge graph. Exemplarily, a corresponding conflict marker is established for each triple in the high-confidence event knowledge set. When a triple is marked as a rule conflict in the first conflict determination result or a semantic conflict in the second conflict determination result, the triple is removed from subsequent update candidates, and only triples not marked by either conflict determination result are retained as the target knowledge set.

[0075] For example, for a highly reliable triple "Company A - Acquisition - Company B", if the rule comparison shows that there is already a fact record of "Company A - Termination of Acquisition - Company B" within the same valid time range, and the embedding similarity calculation results show that the two are highly similar in semantic space, the system can directly regard the triple as a conflicting triple and remove it, and not include it in the target knowledge set.

[0076] In other implementations, candidate knowledge can be graded according to the type and severity of the conflict determination results. For example, some records with minor conflicts can be marked as awaiting manual review, while only records without conflicts can be directly included in the target knowledge set.

[0077] Step S500: Write the target knowledge set into the enterprise risk knowledge graph to generate updated enterprise risk knowledge graph data.

[0078] The enterprise risk knowledge graph is used to organize and store risk event entities and relationships related to the target enterprise and its associated entities in a graph structure. The updated enterprise risk knowledge graph data represents the current state of the graph after being written into the target knowledge set. For example, after conflict detection, each triplet record in the target knowledge set is traversed sequentially, converted into a data structure adapted to the enterprise risk knowledge graph, and written through a graph database interface to ensure that the entities, relationships, and attributes in the enterprise risk knowledge graph are consistent with the latest event knowledge.

[0079] In other implementations, version information or timestamp information related to the target knowledge set is recorded during the writing process to provide a data foundation for subsequent time-based risk analysis or evolution tracking.

[0080] In an optional embodiment, step S500 includes the following sub-steps: S501 converts the target knowledge set into graph entity data and relation data that conform to the enterprise risk knowledge graph storage structure.

[0081] In this framework, graph entity data represents enterprises, related entities, or event objects stored as nodes in the enterprise risk knowledge graph, while relation data represents the relationships between entities stored as edges in the knowledge graph. For example, for each triple (h, r, t) in the target knowledge set, the head entity h and tail entity t are mapped to corresponding graph entity node records, and a corresponding relation record is constructed for relation r. During the mapping process, fields such as identifiers, text descriptions, and time attributes can be added to each entity and relation to ensure consistency with the existing node and edge structure of the enterprise risk knowledge graph.

[0082] For example, when the target knowledge set contains the triple "Company A - Acquisition - Company B", entity node records for "Company A" and "Company B" can be constructed or updated in the graph entity data, and edge records representing the "acquisition" relationship can be constructed in the relation data, while attaching attribute information such as announcement release time or event occurrence time.

[0083] S502, based on the entity data and relation data of the graph, perform insertion and / or update operations on the enterprise risk knowledge graph to obtain the updated enterprise risk knowledge graph data.

[0084] The insertion operation adds nodes and edges from the entity and relation data that are not yet present in the enterprise risk knowledge graph to the graph structure, while the update operation modifies or supplements the attributes of existing nodes and edges. For example, it first determines whether each entity node in the entity data already exists in the enterprise risk knowledge graph based on the entity's unique identifier or matching rules. For non-existent entity nodes, it inserts them; for existing entity nodes, it updates their attribute fields. Then, a similar matching process is performed on each relation record in the relation data, inserting new relation edges between entity nodes or updating the time and status attributes of existing relation edges, ultimately generating the updated enterprise risk knowledge graph data.

[0085] Figure 4 A schematic diagram of a knowledge graph updating apparatus according to an embodiment of this application is shown. Exemplarily, the apparatus 100 includes: Data acquisition module 110 is used to acquire publicly available text data related to the target company; Data processing module 120 is used to parse the publicly available text data using a preset large model to generate a set of candidate triples representing the target enterprise and its related events. Evaluation module 130 is used to perform credibility evaluation and semantic alignment processing on the candidate triple set to obtain a high-credibility event knowledge set; Detection module 140 is used to perform conflict detection on the high-confidence event knowledge set based on preset knowledge graph constraints to determine the target knowledge set; The graph acquisition module 150 is used to write the target knowledge set into the enterprise risk knowledge graph and generate updated enterprise risk knowledge graph data.

[0086] It is understood that the apparatus of this embodiment corresponds to the method of the above embodiments, and the options in the above embodiments are also applicable to this embodiment, so they will not be described again here.

[0087] This application also provides a terminal device, exemplary of which includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to enable the terminal device to perform the functions of the various modules in the above-described method or apparatus.

[0088] The processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, including at least one of a Central Processing Unit (CPU), Graphics Processing Unit (GPU), Network Processor (NP), Digital Signal Processor (DSP), Application-Specific Integrated Circuit (ASIC), Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application.

[0089] The memory can be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The memory is used to store computer programs, and the processor can execute the computer programs accordingly after receiving execution instructions.

[0090] This application also provides a computer-readable storage medium for storing the computer program used in the aforementioned terminal device. For example, the computer-readable storage medium may include, but is not limited to, various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0091] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that, in alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and combinations of blocks in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0092] In addition, the functional modules or units in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0093] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a smartphone, personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0094] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A knowledge graph update method, characterized in that, The method includes: Obtain publicly available text data related to the target company; The publicly available text data is parsed using a pre-defined large model to generate a set of candidate triples representing the target enterprise and its related events; The set of candidate triples is subjected to credibility evaluation and semantic alignment to obtain a set of highly reliable event knowledge. Based on preset knowledge graph constraints, conflict detection is performed on the high-confidence event knowledge set to determine the target knowledge set; The target knowledge set is written into the enterprise risk knowledge graph to generate updated enterprise risk knowledge graph data.

2. The knowledge graph updating method according to claim 1, characterized in that, The acquisition of publicly available text data related to the target company includes: Obtain the identification information representing the target enterprise; Based on the identification information, text data related to the target enterprise is retrieved from a preset public text data source to obtain a candidate public text set; The candidate public text set is subjected to relevance filtering and time range filtering to obtain the target public text data, and the target public text data is used as the public text data.

3. The knowledge graph updating method according to claim 1, characterized in that, The step involves parsing the publicly available text data using a pre-defined large model to generate a set of candidate triples representing the target enterprise and its related events, including: Retrieve the preset relation template and sample triplet set; Based on the relation template and the set of example triples, construct a prompt message corresponding to the public text data; The publicly available text data and the prompt information are input into the large model to obtain an initial set of triplet results representing the target enterprise and its related events; The initial triplet result set is subjected to format validation and relation type filtering to generate a candidate triplet set, which is then used as the candidate triplet set representing the target enterprise and its related subject events.

4. The knowledge graph updating method according to claim 1, characterized in that, The process of performing credibility evaluation and semantic alignment on the candidate triplet set to obtain a highly reliable event knowledge set includes: Based on a preset evaluation strategy, the candidate triplet set is analyzed from multiple perspectives to obtain a set of multi-perspective evaluation results corresponding to each candidate triplet. The multi-perspective evaluation result set is converted into a semantic representation characterizing each candidate triple, and the evaluation index of each candidate triple is determined based on the semantic representation. The candidate triplet set is screened according to the evaluation criteria to obtain a set of highly reliable event knowledge.

5. The knowledge graph updating method according to claim 4, characterized in that, The step of converting the multi-view evaluation result set into a semantic representation characterizing each candidate triplet, and determining the evaluation index of each candidate triplet based on the semantic representation, includes: Each candidate triplet is input into the teacher model and the student model respectively to obtain the corresponding first semantic representation and second semantic representation; The distillation error of each candidate triplet is determined based on the first semantic representation and the second semantic representation; The distillation error is combined with the similarity index corresponding to the multi-view evaluation result to generate the evaluation index for each candidate triplet.

6. The knowledge graph updating method according to claim 1, characterized in that, The method of performing conflict detection on the high-confidence event knowledge set based on preset knowledge graph constraints to determine the target knowledge set includes: Obtain the set of historical event knowledge related to the set of highly credible event knowledge in the enterprise risk knowledge graph; Based on entity type constraints and relation constraints, the high-confidence event knowledge set and the historical event knowledge set are compared according to rules to determine the first conflict determination result. The highly reliable event knowledge set and the historical event knowledge set are mapped to a first embedding representation and a second embedding representation, respectively, and the similarity between the two is calculated based on the embedding representation to determine the second conflict determination result; The high-confidence event knowledge set is filtered based on the first conflict determination result and the second conflict determination result to obtain the target knowledge set.

7. The knowledge graph updating method according to claim 1, characterized in that, The step of writing the target knowledge set into the enterprise risk knowledge graph to generate updated enterprise risk knowledge graph data includes: The target knowledge set is converted into graph entity data and relation data that conform to the enterprise risk knowledge graph storage structure; Based on the entity data and relation data of the graph, insert and / or update operations are performed on the enterprise risk knowledge graph to obtain updated enterprise risk knowledge graph data.

8. A knowledge graph updating device, characterized in that, include: The data acquisition module is used to acquire publicly available text data related to the target company; The data processing module is used to parse the publicly available text data using a preset large model to generate a set of candidate triples representing the target enterprise and its related events. The evaluation module is used to perform credibility evaluation and semantic alignment processing on the candidate triple set to obtain a highly reliable event knowledge set. The detection module is used to perform conflict detection on the high-confidence event knowledge set based on preset knowledge graph constraints, and to determine the target knowledge set; The graph acquisition module is used to write the target knowledge set into the enterprise risk knowledge graph and generate updated enterprise risk knowledge graph data.

9. A terminal device, characterized in that, The terminal device includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the knowledge graph update method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed on a processor, implements the knowledge graph update method according to any one of claims 1-7.

Images