Dynamic knowledge graph construction method and system based on large model driving

Abstract

The invention provides a dynamic knowledge graph construction method and system based on large model driving, and relates to the technical field of dynamic knowledge graph construction, and the construction method comprises the steps: collecting multi-source data, collecting new knowledge and user search data in real time, and extracting entities and relationships through preprocessing to construct a knowledge graph. And constructing a graph attention network model optimized by an annealing moth optimization algorithm, inputting new knowledge according to a preset interval to obtain a fusion strategy, fusing the data to obtain entity update data, and adjusting the data and updating the graph according to user tendency to obtain a personalized dynamic knowledge graph. According to the method, the tendency data is extracted from the user search data, personalized adjustment is performed on the knowledge graph, so that new knowledge can be accurately fused into the knowledge graph, data conflict and redundancy are avoided, higher-quality and more personalized knowledge services are provided for users, and the application value of the knowledge graph in various fields is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of dynamic knowledge graph construction, and in particular to a method and system for constructing a dynamic knowledge graph driven by a large model. Background Art

[0002] In today's era of information explosion, knowledge graphs, as an effective knowledge representation and management tool, are widely used in many fields such as intelligent question answering, recommendation systems, and information retrieval. Knowledge graphs are designed to store and display various types of knowledge in a structured form, and to express rich semantic information through entities and the relationships between them.

[0003] Traditional knowledge graph construction methods often rely on a single or limited data source, which makes the content of the knowledge graph one-sided and limited, and it is difficult to fully cover the required knowledge field. At the same time, due to the single source of data, the data quality is also uneven, containing a lot of noise, errors or incomplete information, which brings great challenges to the subsequent entity extraction, relationship analysis and graph construction.

[0004] As time goes by and knowledge is constantly updated, new knowledge data are continuously generated. However, existing technologies face many difficulties in processing the integration of new knowledge data with existing data. On the one hand, the format, structure and semantics of new knowledge data may be quite different from those of existing data, making it difficult to directly integrate them; on the other hand, the lack of effective fusion strategies and methods makes it easy for data conflicts, redundancy or loss to occur during the fusion process, which seriously affects the accuracy and consistency of the knowledge graph.

[0005] In addition, different users have different needs and preferences when using knowledge graphs. For example, in an intelligent question-answering system, some users may be more concerned with knowledge in the technical field, while others may be more interested in common sense; in a recommendation system, users' interests and consumption habits vary. However, most existing knowledge graphs are general and lack consideration for users' personalized needs. They are unable to provide users with accurate and personalized knowledge services, which greatly limits the application effect and user experience of knowledge graphs.

[0006] In order to solve the above problems, some studies have tried to introduce large models to drive the construction of knowledge graphs. Large models have powerful language understanding and generation capabilities and can handle complex data and semantic information. This paper proposes a dynamic knowledge graph construction method and system based on large model drive. Summary of the invention

[0007] The present invention provides a method and system for constructing a dynamic knowledge graph driven by a large model, so as to solve the defects in the prior art that it is difficult to merge new data with existing data and the knowledge graph is difficult to meet the personalized needs of different users.

[0008] On the one hand, the present invention provides a method for constructing a dynamic knowledge graph driven by a large model, including: Collect multi-source data, and collect new knowledge data and user search data in real time.

[0009] Preprocess the multi-source data, extract entity data from it, analyze the relationships between entities in the entity data to generate entity pair relationships, and integrate the entity data and entity pair relationships according to the graph structure to obtain a knowledge graph.

[0010] Construct a graph attention network model, optimize the graph attention network model using the annealing moth optimization algorithm to obtain an optimized graph attention network model, input new knowledge data at preset time intervals, output a fusion strategy, and fuse the new knowledge data with the entity data according to the fusion strategy to obtain entity update data.

[0011] Extract tendency data from the user search data, adjust the entity update data according to the tendency data to obtain entity adjustment data, and update the knowledge graph according to the entity adjustment data to obtain a personalized dynamic knowledge graph.

[0012] According to the method for constructing a dynamic knowledge graph driven by a large model provided by the present invention, the steps of extracting entity data include: Convert the audio in the multi-source data into text format, and split the text format and the text data in the multi-source data into multiple tokens.

[0013] Annotate the part of speech for each token, and identify the key information in each part of speech to form text entities, and obtain processed text data by processing the text entities.

[0014] Split the video data in the multi-source data into multiple image segments, and perform histogram equalization operations on the image data and the multiple image segments in the multi-source data.

[0015] Annotate the objects in the image data and determine the categories to obtain image classifications, assign the image classifications to the corresponding scenes to obtain image scene data, and together with the processed text data, form entity data.

[0016] According to the method for constructing a dynamic knowledge graph driven by a large model provided by the present invention, the steps of generating entity pair relationships include: In the processed text data, determine the existence of relationship one by statistically analyzing the co-occurrence of entities within the same text paragraph, sentence, and a certain window range.

[0017] Determine the existence of relationship two according to the appearance of multiple objects in the same image scene in the image scene data.

[0018] According to the existence relationship one and the existence relationship two, formulate a rule template for extracting entity relationships, and apply the rule template to the entity data to extract entity pair relationships.

[0019] According to a method for constructing a dynamic knowledge graph driven by a large model provided by the present invention, the steps of integrating to obtain a knowledge graph include: Define entity types according to the domain of multi-source data, determine relationship types according to entity pair relationships, and define corresponding attributes for entity types and relationship types to obtain attribute values. Determine the graph schema from entity types, relationship types, and attribute values.

[0020] Classify and map the entity data according to entity types to obtain type mapping data, and match and map the entity pair relationships and relationship types to obtain relationship mapping data.

[0021] According to the type mapping data, create corresponding nodes in the graph to obtain multiple entity nodes. According to the relationship mapping data, create edges between the corresponding entity nodes. Add corresponding attribute information to the multiple entity nodes and edges according to the attribute values, thereby constructing a knowledge graph.

[0022] According to a method for constructing a dynamic knowledge graph driven by a large model provided by the present invention, the steps of obtaining an optimized graph attention network model include: Correspond the position vectors of the moths to the parameters of the graph attention network model, define the number of moths and the dimension of each moth, and initialize the moth position matrix.

[0023] Use the graph attention network model with the position vectors of each moth corresponding to the parameters to calculate the fitness value corresponding to each moth.

[0024] Sort the moths corresponding to multiple fitness values in ascending order, select the preset number of moths with the highest fitness value as the flames to obtain a flame position matrix.

[0025] Each moth updates its own position according to the flame position matrix.

[0026] After the moth position is updated, calculate the fitness value of the current moth and determine whether it is less than the fitness value before the position is not updated. If so, accept the updated position. Otherwise, calculate the update probability and update the updated position to obtain an excellent position.

[0027] Adjust the updated moth positions that exceed the boundary of the preset search space to the boundary of the preset search space. When the moth positions are updated to the preset number of iterations, select the parameters represented by the moth position with the highest fitness value as the parameters of the graph attention network model to obtain an optimized graph attention network model.

[0028] A method for constructing a dynamic knowledge graph driven by a large model provided by the present invention, the steps of outputting a fusion strategy include: Convert the new knowledge data into a format acceptable to the optimized graph attention network model, and perform normalization processing. According to the input requirements of the optimized graph attention network model, construct a corresponding graph data structure for the normalized new knowledge data.

[0029] Input the graph data structure into the optimized graph attention network model for forward propagation calculation, and obtain the output result through multiple layers of graph attention layers and reasoning about the relationships between nodes.

[0030] Determine the association relationship between the entities in the new knowledge data and the entities in the entity data according to the output result, and remove the conflicting parts to determine the fusion target.

[0031] Calculate the similarity between the entities in the new knowledge data and the entities in the existing data, match and merge the entities with high similarity to obtain an entity alignment method, merge the data in the new knowledge data with the existing data according to the association relationship, and update the inconsistent relationships according to the reliability to obtain a relationship fusion method.

[0032] Evaluate the confidence of the new knowledge data and the existing data, retain the data with high confidence as a conflict resolution method, and the fusion strategy is composed of the entity alignment method, the relationship fusion method, and the conflict resolution method.

[0033] A method for constructing a dynamic knowledge graph driven by a large model provided by the present invention, the steps of obtaining entity update data include: According to the fusion strategy, calculate the similarity between the entities in the knowledge data and the entities in the entity data, and determine the mutually matching entities in the new knowledge data and the entity data according to a preset similarity threshold.

[0034] Merge the attributes of the mutually matching entities, add the attributes unique to the entities in the new knowledge data to the corresponding entities in the entity data, merge the relationships between the entities in the new knowledge data that are consistent with the relationships between the corresponding entities in the entity data, update the inconsistent relationships, and add the new relationships between the entities that appear in the new knowledge data to the entity data.

[0035] A method for constructing a dynamic knowledge graph driven by a large model provided by the present invention, the steps of obtaining entity adjustment data include: Remove the noise information in the user search data, unify the format to obtain search processed data, count the frequency of the search keywords in the search processed data, and analyze the semantics and context information of the search keywords to obtain a preliminary intention.

[0036] Analyze the behavioral characteristics of the user's search time, search frequency, and search depth, and combine the association relationships between search keywords to obtain the user's potential search patterns.

[0037] Combine the preliminary intention and the user's potential search patterns to obtain the inclination data, match the search keywords with the entities in the entity update data, and analyze the relationship between the inclination data and the entity attributes to obtain the association relationship.

[0038] According to the preliminary intention and the user's potential search patterns, adjust the entity weights, attribute priorities, and relationships between entities in the entity update data, and filter out the entities that are not relevant to the user's inclination according to the inclination data for hiding to obtain the entity adjustment data.

[0039] According to a method for constructing a personalized dynamic knowledge graph driven by a large model provided by the present invention, the steps for updating to obtain a personalized dynamic knowledge graph include: Make the format of the entity adjustment data consistent with the data format of the knowledge graph, and match the entities in the entity adjustment data with the entities in the knowledge graph by using the unique identifiers and attribute information of the entities.

[0040] For the matched entities, adjust the attributes and relationships therebetween according to the entity adjustment data, and add the newly added entities in the entity adjustment data to the knowledge graph.

[0041] Combine the inclination data to perform personalized customization on the knowledge graph, and update the knowledge graph according to the preset time and user behavior to obtain a dynamic knowledge graph.

[0042] Remove the inconsistencies existing in the dynamic knowledge graph, keep the entity attributes and the relationships between entities complete, and perform performance optimization to obtain a personalized dynamic knowledge graph.

[0043] On the other hand, the present invention also provides a system for constructing a dynamic knowledge graph driven by a large model, including a data source module and an intelligent agent: The data source module is used to collect multi-source data, and collect new knowledge data and user search data in real time.

[0044] The intelligent agent includes a knowledge graph construction module, an update data acquisition module, and a knowledge graph update module; The knowledge graph construction module is used to preprocess the multi-source data, extract entity data therefrom, analyze the relationships between the entities in the entity data to generate entity pair relationships, and integrate the entity data and the entity pair relationships according to the graph structure to obtain a knowledge graph.

[0045] An updated data acquisition module, which is used to build a graph attention network model, optimize the graph attention network model using an annealing moth optimization algorithm to obtain an optimized graph attention network model, input new knowledge data at a preset time interval, output a fusion strategy, and fuse the new knowledge data with entity data according to the fusion strategy to obtain entity update data.

[0046] A knowledge graph update module, which is used to extract tendency data from user search data, adjust the entity update data according to the tendency data to obtain entity adjustment data, and update the knowledge graph according to the entity adjustment data to obtain a personalized dynamic knowledge graph.

[0047] A method and system for constructing a dynamic knowledge graph driven by a large model provided by the present invention collect multi-source data and real-time collect new knowledge data and user search data, extract entity data, analyze entity relationships, and integrate them into a knowledge graph through preprocessing, build and optimize a graph attention network model to output a fusion strategy to fuse new knowledge and entity data, and then adjust the data according to the user search tendency and update the knowledge graph, which solves the problems of data diversity and quality, the accuracy and efficiency of knowledge graph construction, personalization and dynamic adaptability, model optimization and data fusion, and achieves the beneficial effects of improving the quality of the knowledge graph, enhancing its timeliness and dynamics, improving the user experience, optimizing the model performance, and promoting knowledge integration and application. Brief Description of the Drawings

[0048] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0049] Figure 1 It is a schematic flowchart of a method for constructing a dynamic knowledge graph driven by a large model provided by an embodiment of the present invention; Figure 2 It is a schematic structural diagram of a system for constructing a dynamic knowledge graph driven by a large model provided by an embodiment of the present invention. Detailed Embodiments

[0050] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions in the present invention in conjunction with the drawings in the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments of the present invention belong to the scope of protection of the present invention.

[0051] The following will Figure 1 - Figure 2 describe a method and system for constructing a dynamic knowledge graph driven by a large model according to the present invention.

[0052] Figure 1 is a schematic flowchart of a method for constructing a dynamic knowledge graph driven by a large model provided by an embodiment of the present invention.

[0053] As Figure 1 shown, a method for constructing a dynamic knowledge graph driven by a large model provided by an embodiment of the present invention, the execution subject may be a system for constructing a dynamic knowledge graph driven by a large model, and the method includes: Collect multi-source data, and collect new knowledge data and user search data in real time. Widely and comprehensively collect multi-source data, covering data from different fields and different structural forms, such as structured database records, semi-structured web information, and unstructured text files, social media content, etc. At the same time, use advanced real-time data collection technology to continuously and uninterruptedly collect new knowledge data in real time, including but not limited to cutting-edge research results, technological innovation trends, updated information on policies and regulations in the industry, etc., to ensure the timeliness and novelty of knowledge. In addition, by monitoring and recording the search behavior of users on various platforms (such as search engines, Q&A communities, e-commerce platforms, etc.), collect user search data in real time to gain insights into users' interest preferences, information needs, and focus of attention.

[0054] Preprocess the multi-source data, extract entity data from it, analyze the relationships between entities in the entity data to generate entity pair relationships, and integrate the entity data and entity pair relationships according to the graph structure to obtain a knowledge graph.

[0055] The steps for extracting entity data include: Perform noise reduction, remove silent segments, and adjust the volume for the audio in the multi-source data, and convert it into text format. Combine the text format with the text data in the multi-source data, and segment the text into multiple tokens.

[0056] Annotate the part of speech for each token, and identify the key information in each part of speech to form text entities, and obtain processed text data by analyzing and processing text entities with the same name but different meanings through context.

[0057] Segment the video data in the multi-source data into multiple image segments, and perform histogram equalization operations on the image data and multiple image segments in the multi-source data using image enhancement technology.

[0058] Annotate the objects in the image data and determine their categories to obtain image classifications, assign the image classifications to the corresponding scenes to obtain image scene data, and together with the processed text data, form entity data.

[0059] The steps for generating entity pair relationships include: In processing text data, relationship one is determined by counting the co-occurrence of entities within the same text paragraph, sentence, and within a certain window range. When multiple entities are found to co-occur frequently within these defined ranges, based on a large amount of language knowledge and semantic understanding experience, we can reasonably infer that there is relationship one between them, and this relationship can be a causal relationship, a subordinate relationship, a parallel relationship, or other semantic associations.

[0060] Based on multiple objects appearing in the same image scene in the image scene data, relationship two is determined. When multiple objects are detected to appear in the same image scene, not only their coexistence is recorded, but also further analysis is carried out on various aspects of information such as the spatial position relationship between these objects, the occlusion situation between each other, the action postures, and the overall layout of the scene. Considering these factors comprehensively, we can determine that there is relationship two, which can include: the interaction relationship, the ownership relationship, or the collaborative relationship in a specific scene, etc., between the objects.

[0061] According to relationship one and relationship two, a rule template for extracting entity relationships is formulated, and the rule template is applied to the entity data to extract the entity pair relationships. The rule template stipulates how to accurately judge and extract the relationships between entities under different data types and scenarios. The rule template is applied to the entity data after preprocessing and feature extraction, and through an automated processing process, the entity pair relationships are extracted.

[0062] The steps for integrating to obtain a knowledge graph include: Define entity types according to the domain of multi-source data, determine relationship types according to the entity pair relationships, and define corresponding attributes for the entity types and relationship types to obtain attribute values. Determine the graph schema from the entity types, relationship types, and attribute values.

[0063] Classify and map the entity data according to the entity types to obtain type mapping data, and match and map the entity pair relationships and relationship types to obtain relationship mapping data.

[0064] According to the type mapping data, create corresponding nodes in the graph to obtain multiple entity nodes. According to the relationship mapping data, create edges between the corresponding entity nodes. Add corresponding attribute information to the multiple entity nodes and edges according to the attribute values, thereby constructing a knowledge graph.

[0065] Construct a graph attention network model, optimize the graph attention network model using the annealing moth optimization algorithm to obtain an optimized graph attention network model, input new knowledge data at preset time intervals, output a fusion strategy, and fuse the new knowledge data with the entity data according to the fusion strategy to obtain entity update data.

[0066] The steps to obtain the optimized graph attention network model include: Correspond the position vector of the moth to the parameters of the graph attention network model. Define the number of moths as M, the dimension of each moth as D, and initialize the moth position matrix.

[0067] Use the graph attention network model with the position vector of each moth corresponding to the parameters to calculate the fitness value corresponding to each moth.

[0068] Sort the moths corresponding to multiple fitness values in ascending order, select the preset number of moths with the highest fitness values as the flames, and obtain the flame position matrix.

[0069] Each moth updates its own position according to the flame position matrix, and the formula is expressed as:

[0070] In the formula, is the th dimension position of the th moth, is the th dimension position of the th flame, and are random numbers in the interval [0, 1].

[0071] After the moth position is updated, calculate the fitness value of the current moth and determine whether it is less than the fitness value before the position is updated. If so, accept the updated position; otherwise, calculate the update probability and update the updated position to obtain an excellent position. The formula for calculating the update probability is expressed as:

[0072]

[0073] In the formula, is the fitness value of the updated position, is the fitness value before the position is not updated, is the update probability, is the current iteration number, is the current temperature, is the initial temperature, is the temperature reduction coefficient.

[0074] Adjust the updated moth position that exceeds the boundary of the preset search space to the boundary of the preset search space. When the moth position is updated to the preset number of iterations, select the parameters represented by the moth position with the highest fitness value as the parameters of the graph attention network model to obtain the optimized graph attention network model.

[0075] The steps of obtaining the fusion strategy include: Convert the new knowledge data into a format acceptable to the optimized graph attention network model and perform normalization. According to the input requirements of the optimized graph attention network model, construct the corresponding graph data structure for the normalized new knowledge data. This includes determining the set of nodes, the set of edges, and the feature matrices of nodes and edges in the graph. For example, in a node classification task, each node represents an entity, the node feature is the attribute information of the entity, and the edge represents the relationship between entities.

[0076] Input the graph data structure into the optimized graph attention network model for forward propagation calculation, and obtain the output result through multiple layers of graph attention layers and reasoning about the relationships between nodes. In each layer of the graph attention layer, the graph data structure performs a linear transformation on each node of the graph data structure through a learnable weight matrix to obtain a feature vector. For each node and its neighbor nodes, concatenate the feature vector and the feature vectors of the neighbor nodes to obtain a combined vector, perform a dot product on the combined vector, and process it through an activation function to obtain the attention coefficients of all nodes. Use the softmax function to normalize the attention coefficients of all nodes to obtain the normalized attention weights, aggregate the neighbor node features according to the normalized attention weights across the country, update the feature vectors of each node to obtain multiple node features, continuously optimize the node features and the ability to capture relationships, and finally process the node features according to the specific task to obtain the output result.

[0077] Determine the association relationship between the entities in the new knowledge data and the entities in the entity data according to the output result, and remove the conflicting parts to determine the fusion target. By comparing and analyzing the features in the output result, it is found that there are partial similarities and associations in terms of algorithm principles, application scenarios, etc., and then the existing association relationship can be determined, which can be an inclusion relationship or a homogeneous relationship in a parallel relationship, etc. By referring to multiple reliable data sources and applying data verification techniques, evaluate and judge the conflicting parts, and thus remove these conflicting parts. For example, if the new knowledge data contains the latest information on entity attributes that are not in the entity data, then the fusion target is to reasonably integrate this new information into the entity data to update and improve the attribute description of the entity. If the new knowledge data reveals new relationships between entities, the fusion target is to add these new relationships to the entity data to enhance the integrity and accuracy of the entity relationships in the knowledge graph.

[0078] Calculate the similarity between the entities in the new knowledge data and the entities in the existing data, match and merge the entities with high similarity to obtain the entity alignment method, merge the data in the new knowledge data with the existing data according to the association relationship, and update the inconsistent relationships according to the reliability to obtain the relationship fusion method.

[0079] Evaluate the confidence of new knowledge data and existing data, retain the data with high confidence as the conflict resolution method, and the fusion strategy consists of entity alignment method, relationship fusion method and conflict resolution method.

[0080] The steps to obtain entity updated data include: According to the fusion strategy, calculate the similarity between the entities in the knowledge data and the entities in the entity data, and determine the mutually matching entities in the new knowledge data and the entity data according to the preset similarity threshold.

[0081] Merge the attributes of the mutually matching entities, add the unique attributes of the entities in the new knowledge data to the corresponding entities in the entity data, merge the relationships between the entities in the new knowledge data that are consistent with the relationships between the corresponding entities in the entity data, update the inconsistent relationships, and add the new relationships between the entities that appear in the new knowledge data to the entity data.

[0082] Extract the tendency data from the user search data, adjust the entity updated data according to the tendency data to obtain entity adjustment data, and update the knowledge graph according to the entity adjustment data to obtain a personalized dynamic knowledge graph.

[0083] The steps to obtain entity adjustment data include: Remove the noise information in the user search data, unify the format to obtain search processed data, count the frequency of the search keywords in the search processed data, and analyze the semantics and context information of the search keywords to obtain the preliminary intention.

[0084] Analyze the behavioral characteristics of users' search time, search frequency, and search depth, and combine the correlation relationships between search keywords to obtain potential user search patterns. Comprehensively collect users' search records at different dates and times, accurate to the hour or even minute. Through the collation and statistics of these time data, understand the distribution of users' search behaviors within a day, a week, a month, or even a longer period, so as to obtain the users' search time. Conduct a detailed grouped statistics on the users' search frequency, and divide users into different groups of high-frequency search users, medium-frequency search users, and low-frequency search users according to the level of search frequency, so as to obtain the search frequency. Based on multi-dimensional information such as the level at which users click on search results and the duration of stay on each result page, comprehensively evaluate the search depth. Perform accurate word segmentation on search keywords, extract the core stems and various inflections of keywords, and eliminate the differences in the surface forms of vocabulary to facilitate more accurate semantic analysis. Through cluster analysis, group users with similar search behavioral characteristics and keyword correlation patterns into the same category, and then deeply study the typical search patterns of each category of users. At the same time, use the decision tree algorithm to predict users' search behaviors and needs in different scenarios. After validating and interpreting the mined typical search patterns, potential user search patterns are obtained.

[0085] Combine the preliminary intention and the potential user search pattern to obtain tendency data, match the search keywords with the entities in the entity update data, and analyze the relationship between the tendency data and the entity attributes to obtain the correlation relationship.

[0086] According to the preliminary intention and the potential user search pattern, adjust the entity weights, attribute priorities, and relationships between entities in the entity update data, and filter out the entities that are not relevant to the user's tendency for hiding according to the tendency data to obtain entity adjustment data.

[0087] The steps to update and obtain the personalized dynamic knowledge graph include: Make the format of the entity adjustment data consistent with the data format of the knowledge graph, and match the entities in the entity adjustment data with the entities in the knowledge graph using the unique identifiers and attribute information of the entities.

[0088] For the matched entities, adjust the attributes and relationships between them according to the entity adjustment data, and add the newly added entities in the entity adjustment data to the knowledge graph.

[0089] Combine the tendency data to customize the knowledge graph, and update the knowledge graph according to the preset time and user behavior to obtain the dynamic knowledge graph.

[0090] Remove the inconsistencies existing in the dynamic knowledge graph, keep the entity attributes and the relationships between entities intact, and perform performance optimization to obtain a personalized dynamic knowledge graph. Performance optimization can include index optimization, storage structure adjustment, etc., to improve the query efficiency and data processing ability of the knowledge graph. For example, establish indexes for entities or relationships that are frequently queried to speed up querying.

[0091] A method for constructing a dynamic knowledge graph driven by a large model provided in this embodiment analyzes the relationships between entities in entity data to generate entity pair relationships, and integrates them according to the graph structure, making the construction of the knowledge graph more standardized and accurate. It also constructs a graph attention network model and optimizes it using the annealing moth optimization algorithm, solving the problems of model performance and parameter optimization. The optimized graph attention network model can more effectively process new knowledge data, output reasonable fusion strategies, and improve the accuracy and efficiency of knowledge graph updates. It also solves the problem of fusing new knowledge data with existing entity data by inputting new knowledge data at preset time intervals and outputting fusion strategies, and fusing the new knowledge data with entity data according to the fusion strategies, ensuring that new knowledge can be accurately integrated into the knowledge graph and avoiding data conflicts and redundancies. And it solves the problem of the dynamic adaptability of the knowledge graph by updating the knowledge graph in real time according to entity adjustments, enabling the knowledge graph to evolve continuously with the changes in new knowledge and user needs.

[0092] Based on the same general inventive concept, the present invention also protects a system for constructing a dynamic knowledge graph driven by a large model. The following describes a system for constructing a dynamic knowledge graph driven by a large model provided by the present invention. The system for constructing a dynamic knowledge graph driven by a large model described below can be correspondingly referred to with the method for constructing a dynamic knowledge graph driven by a large model described above.

[0093] Figure 2 It is a schematic structural diagram of a system for constructing a dynamic knowledge graph driven by a large model provided by an embodiment of the present invention.

[0094] As Figure 2 shown, a system for constructing a dynamic knowledge graph driven by a large model includes a data source module and an agent: The data source module is used to collect multi-source data and collect new knowledge data and user search data in real time.

[0095] The agent includes a knowledge graph construction module, an updated data acquisition module, and a knowledge graph update module: The knowledge graph construction module is used to preprocess multi-source data, extract entity data from it, analyze the relationships between entities in the entity data to generate entity pair relationships, and integrate the entity data and entity pair relationships according to the graph structure to obtain a knowledge graph.

[0096] The updated data acquisition module is used to construct a graph attention network model, optimize the graph attention network model using the annealing moth optimization algorithm to obtain an optimized graph attention network model, input new knowledge data at a preset time interval, output a fusion strategy, and fuse the new knowledge data with the entity data according to the fusion strategy to obtain entity updated data.

[0097] The knowledge graph update module is used to extract tendency data from user search data, adjust the entity updated data according to the tendency data to obtain entity adjusted data, and update the knowledge graph according to the entity adjusted data to obtain a personalized dynamic knowledge graph.

[0098] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the above technical solution, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a dynamic knowledge graph driven by a large model, characterized in that, Including: Collecting multi-source data, and real-time collecting new knowledge data and user search data; Preprocessing the multi-source data, extracting entity data therefrom, analyzing the relationships between entities in the entity data to generate entity pair relationships, and integrating the entity data and the entity pair relationships according to a graph structure to obtain a knowledge graph; Constructing a graph attention network model, optimizing the graph attention network model using an annealing moth optimization algorithm to obtain an optimized graph attention network model, inputting the new knowledge data at preset time intervals, outputting a fusion strategy, and fusing the new knowledge data with the entity data according to the fusion strategy to obtain entity update data; Extracting tendency data from the user search data, adjusting the entity update data according to the tendency data to obtain entity adjustment data, and updating the knowledge graph according to the entity adjustment data to obtain a personalized dynamic knowledge graph.

2. A method for constructing a dynamic knowledge graph driven by a large model according to claim 1, characterized in that, The steps of extracting the entity data include: Converting the audio in the multi-source data into text format, and splitting the text format and the text data in the multi-source data into multiple tokens; Annotating the part of speech for each token, and identifying the key information in each part of speech to form text entities, and obtaining processed text data by processing the text entities; Splitting the video data in the multi-source data into multiple image segments, and performing histogram equalization operations on the image data and the multiple image segments in the multi-source data; Annotating the objects in the image data and determining the categories to obtain image classifications, assigning the image classifications to corresponding scenes to obtain image scene data, and jointly constituting the entity data with the processed text data.

3. A method for constructing a dynamic knowledge graph driven by a large model according to claim 2, characterized in that, The steps of generating entity pair relationships include: In the processed text data, determining the existence of relationship one by statistically co-occurring of entities in the same text paragraph, sentence, and within a certain window range; Determining the existence of relationship two according to multiple objects in the image scene data appearing in the same image scene; According to relationship one and relationship two, formulating a rule template for extracting entity relationships, and applying the rule template to the entity data to extract the entity pair relationships.

4. A method for constructing a dynamic knowledge graph driven by a large model according to claim 1, characterized in that, The steps of integrating to obtain the knowledge graph include: Defining entity types according to the domain of the multi-source data, determining relationship types according to the entity pair relationships, and defining corresponding attributes for the entity types and the relationship types to obtain attribute values, and determining a graph pattern from the entity types, the relationship types, and the attribute values; Classifying and mapping the entity data according to the entity types to obtain type mapping data, and performing matching and mapping on the entity pair relationships and the relationship types to obtain relationship mapping data; According to the type mapping data, creating corresponding nodes in the graph to obtain multiple entity nodes, according to the relationship mapping data, creating edges between the corresponding entity nodes, and adding corresponding attribute information to the multiple entity nodes and edges according to the attribute values, thereby constructing the knowledge graph.

5. A method for constructing a dynamic knowledge graph driven by a large model according to claim 1, characterized in that, The steps of obtaining the optimized graph attention network model include: Correspond the position vector of the moth to the parameters of the graph attention network model, define the number of moths and the dimension of each moth, and initialize the moth position matrix; Use the graph attention network model with the position vector of each moth corresponding to the parameters to calculate the fitness value corresponding to each moth; Sort the moths corresponding to multiple fitness values in ascending order, and select the preset number of moths with the highest fitness value as the flames to obtain the flame position matrix; Each moth updates its own position according to the flame position matrix; After the moth position is updated, calculate the fitness value of the current moth and determine whether it is less than the fitness value before the position is updated. If so, accept the updated position. Otherwise, calculate the update probability and update the updated position to obtain an excellent position; Adjust the updated moth position that exceeds the boundary of the preset search space to the boundary of the preset search space. When the moth position is updated to the preset number of iterations, select the parameters represented by the moth position with the highest fitness value as the parameters of the graph attention network model to obtain the optimized graph attention network model.

6. A method for constructing a dynamic knowledge graph driven by a large model according to claim 1, characterized in that, The steps to obtain the fusion strategy include: Convert the new knowledge data into a format that can be accepted by the optimized graph attention network model and perform normalization processing. According to the input requirements of the optimized graph attention network model, construct the corresponding graph data structure for the normalized new knowledge data; Input the graph data structure into the optimized graph attention network model for forward propagation calculation, and obtain the output result through multiple graph attention layers and reasoning about the relationships between nodes; Determine the association relationship between the entities in the new knowledge data and the entities in the entity data according to the output result, and remove the conflicting parts to determine the fusion target; Calculate the similarity between the entities in the new knowledge data and the entities in the existing data, match and merge the entities with high similarity to obtain the entity alignment method. Merge the data in the new knowledge data with the existing data according to the association relationship. For inconsistent relationships, update according to the reliability to obtain the relationship fusion method; Evaluate the confidence of the new knowledge data and the existing data, retain the data with high confidence as the conflict resolution method, and the entity alignment method, the relationship fusion method, and the conflict resolution method constitute the fusion strategy.

7. A method for constructing a dynamic knowledge graph driven by a large model according to claim 1, characterized in that, The steps to obtain the entity update data include: According to the fusion strategy, calculate the similarity between the entities in the knowledge data and the entities in the entity data, and determine the mutually matching entities in the new knowledge data and the entity data according to the preset similarity threshold; Merge the attributes of the mutually matching entities, add the attributes unique to the entities in the new knowledge data to the corresponding entities in the entity data, merge the relationships between the entities in the new knowledge data that are consistent with the relationships between the corresponding entities in the entity data, update the inconsistent relationships, and add the new relationships between the entities that appear in the new knowledge data to the entity data.

8. A method for constructing a dynamic knowledge graph driven by a large model according to claim 7, characterized in that, The steps to obtain the entity adjustment data include: Remove the noise information in the user search data, unify the format to obtain the search processed data, count the frequency of the search keywords in the search processed data, and analyze the semantics and context information of the search keywords to obtain the preliminary intention; Analyze the behavioral characteristics of the user's search time, search times, and search depth, and combine the association relationships between the search keywords to obtain the user search potential pattern; Combine the preliminary intention and the user search potential pattern to obtain the tendency data, match the search keywords with the entities in the entity update data, and analyze the relationship between the tendency data and the entity attributes to obtain the association relationship; According to the preliminary intention and the user search potential pattern, adjust the entity weights, attribute priorities, and relationships between entities in the entity update data, and filter out the entities irrelevant to the user's tendency according to the tendency data for hiding to obtain the entity adjustment data.

9. A method for constructing a dynamic knowledge graph driven by a large model according to claim 1, characterized in that, The steps for updating to obtain the personalized dynamic knowledge graph include: Make the format of the entity adjustment data consistent with the data format of the knowledge graph, and match the entities in the entity adjustment data with the entities in the knowledge graph using the unique identifiers and attribute information of the entities; For the matched entities, adjust the attributes and relationships according to the entity adjustment data, and add the newly added entities in the entity adjustment data to the knowledge graph; Combine the tendency data to perform personalized customization on the knowledge graph, and update the knowledge graph according to the preset time and user behavior to obtain the dynamic knowledge graph; Remove the inconsistent situations existing in the dynamic knowledge graph, keep the entity attributes and the relationships between entities complete, and perform performance optimization to obtain the personalized dynamic knowledge graph.

10. A dynamic knowledge graph construction system driven by a large model, which adopts a dynamic knowledge graph construction method driven by a large model as described in any one of claims 1 to 9, characterized in that, The construction system includes a data source module and an agent: The data source module is used to collect multi-source data, and collect new knowledge data and user search data in real time; The agent includes a knowledge graph construction module, an update data acquisition module, and a knowledge graph update module; The knowledge graph construction module is used to preprocess the multi-source data, extract entity data from it, analyze the relationships between the entities in the entity data to generate entity pair relationships, and integrate the entity data and the entity pair relationships according to the graph structure to obtain the knowledge graph; The update data acquisition module is used to construct a graph attention network model, optimize the graph attention network model using the annealing moth optimization algorithm to obtain an optimized graph attention network model, input the new knowledge data at preset time intervals, output the obtained fusion strategy, and fuse the new knowledge data with the entity data according to the fusion strategy to obtain the entity update data; The knowledge graph update module is used to extract the tendency data from the user search data, adjust the entity update data according to the tendency data to obtain the entity adjustment data, and update the knowledge graph according to the entity adjustment data to obtain the personalized dynamic knowledge graph.

Images