A method and device for constructing a multi-component data intelligent entity

Abstract

The present invention discloses a method and device for constructing a multi-component data intelligent body, and the construction method specifically includes the following steps: obtaining sample data and constructing a vector knowledge base; extracting knowledge business flows with sample data, determining the ontology through knowledge representation, and forming a knowledge graph; training an initial language model based on the vector knowledge base and providing a data processing model; combining the logical tool chain and mind map of the data intelligent body, componentizing and linking the knowledge graph and the data processing model to construct a multi-component data intelligent body. The present invention constructs a business rule engine by integrating the knowledge graph, provides training data support, and enhances the accuracy of the result reasoning of the data intelligent body by modeling the thinking chain process as a graph-structured mind map, and executing the data processing model and knowledge graph associated with the workflow node, thereby expanding the knowledge boundary, especially for financial audit data, and can better extract key information and improve the work efficiency of power grid auditors.

Description

Technical Field

[0001] The present invention belongs to the technical field of big data processing, and specifically relates to a method and device for constructing a multi-component data intelligent body. Background Art

[0002] With the development of natural language models and knowledge graph technology, more and more fields have begun to develop and design related intelligent application platforms, using intelligent application platforms as "assistants" to assist professionals in transaction processing.

[0003] The development of intelligent application platforms basically involves big data processing and analysis. The current research focus is also on the construction of data processing models and other directions. For example, patent CN119089931A provides a multi-layer semantic understanding large model intelligent body construction and application method, including the following steps: S1, pre-processing input data; S2, using a large language model and conversation history to analyze the pre-processed data, identify user intent and extract deep semantic features; S3, extract key information from the intent recognition data, perform structured representation, and adjust the model configuration through dynamic hyperparameter optimization; S4, query the knowledge graph, associate and organize the extracted key information with existing data, and generate a structured knowledge representation; S5, vectorize the structured knowledge representation and store it in a vector database to support fast similarity search; S6, combine the generated unified semantic representation and large model, perform in-depth analysis of user input, generate coherent and logical personalized answers, and return them to the user. This solution can support a variety of application scenarios such as intelligent question and answer, content generation, sentiment analysis, and multimodal search.

[0004] In addition, there are also related studies on applications in specific scenarios. For example, patent CN118863044A provides a multi-scale course knowledge graph construction method and an intelligent course teaching method. The first layer of knowledge graph is established by extracting knowledge point entities and knowledge point entity relationships in the course textbook text. The second layer of knowledge graph is formed by extracting questions and their corresponding answers in the exercise materials, combining relevant knowledge points, and further summarizing the node relationships of the second layer of knowledge nodes to form the third layer of knowledge graph. Finally, the relevant nodes of each layer of knowledge graph are connected through the course objectives to form a multi-scale course knowledge graph, thereby providing learners with personalized and hierarchical knowledge recommendations based on the knowledge graph. It also provides an intelligent teaching method and teaching system based on a multi-scale course knowledge graph, which can generate learning paths according to the knowledge mastery of different learners, thereby improving learning efficiency and learning effects.

[0005] However, current intelligent application platforms and research on related models and knowledge graphs all have problems such as insufficient precision and speed in data processing and extraction of key information, poor generalization and promotion of intelligent applications, and insufficient transparency in the process and difficulty in adapting to dynamic changes.

[0006] Therefore, how to improve the accuracy of big data processing and key information extraction, and build a data intelligence entity that is suitable for big data processing and analysis, especially to realize the digitization and intelligence in the field of power grid financial auditing, is an urgent problem to be solved by technical personnel in this field. Summary of the invention

[0007] In view of the defects existing in the above-mentioned prior art, the present invention provides a method and device for constructing a multi-component data intelligent body, and the construction method specifically includes the following steps: obtaining sample data and constructing a vector knowledge base; extracting knowledge business flow with sample data, determining the ontology through knowledge representation, and forming a knowledge graph; training the initial language model based on the vector knowledge base, and giving a data processing model; combining the logical tool chain and mind map of the data intelligent body, componentizing and linking the knowledge graph and the data processing model, forming a workflow canvas between multiple components through the process starting point, intermediate node, and logical control node, and constructing a multi-component data intelligent body; wherein, the process starting point, intermediate node, and logical control node represent each subtask of the data intelligent body, and the edge formed between each point represents the dependency relationship between the subtasks. The present invention constructs a business rule engine by integrating the knowledge graph, provides training data support, and enhances the accuracy of the result reasoning of the data intelligent body by modeling the thinking chain process as a graph structured thinking map, and executing the data processing model and knowledge graph associated with the workflow node, expanding the knowledge boundary, especially for financial audit data, can better extract key information and improve the work efficiency of power grid auditors.

[0008] In a first aspect, the present invention provides a method for constructing a multi-component data agent, which specifically comprises the following steps:

[0009] Obtain sample data and build a vector knowledge base;

[0010] Extract knowledge business flows from sample data, determine the ontology through knowledge representation, and form a knowledge graph;

[0011] Based on the vector knowledge base, train the initial language model and provide a data processing model;

[0012] Combined with the logical tool chain and mind map of the data intelligent body, the knowledge graph and data processing model are componentized and linked. Through the process starting point, intermediate nodes, and logical control nodes, a workflow canvas between multiple components is formed to construct a multi-component data intelligent body.

[0013] Among them, the process starting point, intermediate nodes and logical control nodes represent the various subtasks of the data agent, and the edges formed between the points represent the dependency relationship between the subtasks.

[0014] Furthermore, sample data is obtained and a vector knowledge base is constructed, which specifically includes the following steps:

[0015] Obtain and analyze sample data, and delineate structured data and unstructured data;

[0016] Transform unstructured data into structured data and merge them into structured data;

[0017] The structured data is sliced ​​to form multiple data segments, wherein the structured data is sliced, specifically including:

[0018] Set the slice length to L, the number of sliding characters when slicing to S, and make the current slice and the previous slice form overlapping data segments;

[0019] Based on the embedding model, the data fragments are converted into high-dimensional vectors by mapping the text data to the vector space;

[0020] The similarity data between each high-dimensional vector is determined, and a vector knowledge base is constructed using the high-dimensional vectors and the similarity data.

[0021] Furthermore, based on the embedding model, the data fragments are converted into high-dimensional vectors by mapping the text data to the vector space, which can be specifically expressed as:

[0022]

[0023] Among them, E is the embedding model, W is the fragment set formed by the data fragments, and R n is an n-dimensional real space, w is a data segment, v is an n-dimensional vector, v w is the n-dimensional vector corresponding to the data segment w;

[0024] Determine the similarity data between each high-dimensional vector, which is specifically expressed as:

[0025]

[0026] Among them, S is the similarity data set between each high-dimensional vector, v1, v2, v m-1 、v m They are high-dimensional vectors numbered 1, 2, m-1, and m respectively.

[0027] Furthermore, the ontology is determined through knowledge representation to form a knowledge graph, which specifically includes the following steps:

[0028] Extract knowledge and semantic tags from sample data to form a knowledge business flow, and determine the entities and relationships for knowledge representation from the knowledge business flow;

[0029] Based on the graph database, the entities and relationships of knowledge representation are combined to form a knowledge graph.

[0030] Furthermore, based on the graph database, combined with the entities and relationships of knowledge representation, a knowledge graph is formed, which specifically includes:

[0031] Map each entity in the knowledge representation to a node in the graph database;

[0032] Based on the relationship between entities in the knowledge representation, the corresponding nodes in the graph database are connected to form edges, and the types and attributes of each edge are labeled;

[0033] Gather the nodes, edges, edge types and attributes of the graph database to form a knowledge graph;

[0034] Among them, the nodes, edges, edge types and attributes of the graph database are collected to form a knowledge graph, which is specifically expressed as:

[0035]

[0036] G is the knowledge graph set, N is the set of nodes, E is the set of edges, n i 、n j are nodes numbered i and j respectively, T ij is the type of edge connecting node i and node j, P ij is the attribute of the edge connecting node i and node j.

[0037] Furthermore, based on the vector knowledge base, the initial language model is trained and a data processing model is given, which specifically includes the following steps:

[0038] Based on the vector knowledge base, a training vector set is defined, wherein the training vector set includes a training vector and a corresponding actual result vector;

[0039] The initial language model analyzes the similarity of the training vector and gives the prediction result vector;

[0040] Based on the comparative analysis of the predicted result vector and the actual result vector, the initial language model is iterated to construct a data processing model;

[0041] Among them, the iterative initial language model is specifically expressed as:

[0042]

[0043] Among them, d is the prediction result vector given after iteration, is the initial similarity weight, s is the scaling factor, q is the training vector, and b is the bias term.

[0044] Furthermore, the similarity analysis of the training vectors by the initial language model is used to give a prediction result vector, which specifically includes the following steps:

[0045] The initial language model is used to analyze the distance between the training vector and each high-dimensional vector in the vector knowledge base, and multiple sets of prediction data are given. The distance between the training vector and each high-dimensional vector in the vector knowledge base is specifically expressed as:

[0046]

[0047] s q is the training vector q and the high-dimensional vector d in the vector knowledge base y The distance, n is the dimension of the high-dimensional vector, q x is the x-th dimension vector of the training vector q, d yx is a high-dimensional vector d y The x-th dimension vector of ;

[0048] Sort multiple groups of prediction data, determine the optimal result, and give a prediction result vector.

[0049] Furthermore, by combining the logical tool chain and mind map of the data agent, the knowledge graph and the data processing model are componentized and linked to construct a multi-component data agent, which specifically includes the following steps:

[0050] Based on the logical tool chain and mind map of the data agent, combined with the knowledge map and data processing model, define the process starting point, intermediate nodes and logical control nodes;

[0051] Based on the weight setting of the knowledge graph and data processing model, combined with the logic control node, the weight is configured at the matching process starting point and intermediate nodes, and the process starting point and intermediate nodes are linked;

[0052] Component-based orchestration of knowledge graphs and data processing models to build an initial multi-component data agent;

[0053] The initial multi-component data agent is optimized to construct a multi-component data agent.

[0054] Further, the initial multi-component data agent is optimized to construct a multi-component data agent, which specifically includes the following steps:

[0055] Define the optimization metrics for the initial multi-component data agent;

[0056] Perform optimization tests on the initial multi-component data agent and provide test values ​​corresponding to the optimization indicators;

[0057] According to the optimization strategy, the components to be optimized of the initial multi-component data agent are matched and optimized to construct a multi-component data agent;

[0058] Among them, the optimization strategy is specifically expressed as:

[0059]

[0060] Among them, P new For the optimized overall performance, P old For the overall performance before optimization, is the accuracy index weight, is the accuracy function increment, C is the component mapped to the accuracy indicator, is the completeness index weight, is the integrity function increment, R is the component relationship mapped to the integrity index, is the normative indicator weight, is the normative function increment, ICL F is the annotation parameter mapped to the normative indicator, Incremental readability function, ICL R are annotation parameters that map to readability metrics.

[0061] In a second aspect, the present invention further provides a device for constructing a multi-component data agent, characterized in that the device adopts the method for constructing a component data agent as described above, specifically comprising:

[0062] The acquisition module is used to obtain sample data and build a vector knowledge base;

[0063] The analysis module is used to extract knowledge business flows from sample data, determine the ontology through knowledge representation, and form a knowledge graph; based on the vector knowledge base, train the initial language model and provide a data processing model;

[0064] The construction module is used to combine the logical tool chain and mind map of the data intelligent body, organize and link the knowledge map and data processing model in a componentized manner, form a workflow canvas between multiple components through the process starting point, intermediate nodes, and logical control nodes, and construct a multi-component data intelligent body;

[0065] Among them, the process starting point, intermediate nodes and logical control nodes represent the various subtasks of the data agent, and the edges formed between the points represent the dependency relationship between the subtasks.

[0066] The present invention provides a method and device for constructing a multi-component data intelligent entity, which has at least the following beneficial effects:

[0067] (1) The present invention builds a business rule engine by integrating knowledge graphs and provides training data support. By modeling the thought chain process as a graph-structured thought graph and executing the data processing model and knowledge graph associated with the workflow node, the accuracy of the result reasoning of the data agent is enhanced, and the knowledge boundary is expanded. In particular, for financial audit data, key information can be better extracted, thereby improving the work efficiency of power grid auditors.

[0068] (2) Through the entities and relationships represented by knowledge, the extracted entities and relationships can be stored in the graph database to form a knowledge graph. The design, operation and function of nodes and edges are the basis for building and maintaining knowledge graphs. Through the combined use of nodes and edges, complex data analysis, pattern matching and graph algorithms can be performed.

[0069] (3) Introduce a dynamic ontology update mechanism into the knowledge graph to extract entities and relationships of knowledge representation mapped to nodes in the graph database from new sample data, thereby improving the accuracy and completeness of the knowledge graph.

[0070] (4) The vector knowledge base trains the initial language model, using the similarity between the predicted result vector and the training vector as an indicator to continuously optimize the neural network parameters of the initial language model. By modifying the calculation method of attention weights during the inference process, the initial token is kept visible to all subsequent tokens, while limiting the upper limit of the distance in the attention calculation to adapt to the relative distance that has not been seen during training, thereby achieving stable generation of ultra-long texts.

[0071] In addition, the task instructions of the training vectors will be optimized so that the data processing model can perform tasks according to the expectations of the calling tasks. When optimizing the neural network parameters iteratively, the processing capacity of the model is accelerated by setting the scaling factor and bias term, and the output results after the optimization iteration are quickly verified.

[0072] (5) The optimization of multi-component data agents is reflected in logic adjustment, parameter configuration, etc., mainly through the addition of components and the modification of the calling relationship between components, the accuracy and completeness indicators are optimized, and the standardization and readability indicators are optimized by adjusting the labeled samples (ICL) of different categories. In addition, the training and adjustment of the data processing model are carried out according to the test results to improve the accuracy of the output results of the data agent. By combining the data agent with the mind map model, not only the accuracy and comprehensiveness of risk identification are improved, but also the interpretability and traceability of the entire reasoning process are enhanced, providing a powerful and flexible analysis tool for the internal audit of the enterprise. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 A schematic diagram of a process for constructing a multi-component data intelligent body provided by the present invention;

[0074] Figure 2 A schematic diagram of a process of obtaining sample data and building a vector knowledge base in an embodiment of the present invention;

[0075] Figure 3 A schematic diagram of a process of determining an ontology through knowledge representation to form a knowledge graph in a certain embodiment of the present invention;

[0076] Figure 4 A schematic diagram of a process of forming a knowledge graph based on a graph database combined with entities and relationships represented by knowledge in an embodiment of the present invention;

[0077] Figure 5 A flowchart of a data processing model is provided for training an initial language model based on a vector knowledge base in an embodiment of the present invention;

[0078] Figure 6 A schematic diagram of a process for constructing a multi-component data intelligent entity according to a certain embodiment of the present invention;

[0079] Figure 7 An example diagram of a multi-component data agent provided by the present invention;

[0080] Figure 8 A schematic diagram of the architecture of a device for constructing a multi-component data intelligent entity provided by the present invention. DETAILED DESCRIPTION

[0081] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0082] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The singular forms "a", "said" and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings, and "multiple" generally includes at least two.

[0083] It should also be noted that the term "includes", "comprising" or any other variation thereof is intended to cover non-exclusive inclusion, so that a commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprising a ..." do not exclude the existence of other identical elements in the commodity or device including the elements.

[0084] In order to give full play to the data processing and analysis capabilities of natural language models and knowledge graphs, when building data agents, the knowledge graph and natural language model are linked and integrated. Through the complementary role of the knowledge graph, the accuracy of the result reasoning of the data agent is enhanced, which can not only expand the boundaries of knowledge, but also improve the generalization ability of the data agent.

[0085] Therefore, multiple components are formed by linking knowledge graphs and natural language models. By setting nodes and logical points, key information in the data can be better extracted to build a multi-component data intelligent body. This can better improve the generalization ability of data processing and the practicality of promotion, and be integrated into the business digital platform, such as the power grid financial audit platform, to effectively improve the work efficiency of power grid auditors and realize intelligence and digitization in the field of power grid auditing.

[0086] The constructed multi-component data agent can be considered to achieve the question-answering result reasoning task through a thought graph, that is, modeling the thought chain process as a graph structure. In the multi-dimensional model of the multi-component data agent, each node represents a step or decision point in the problem-solving process, and the connections between nodes represent different thinking paths and possible decisions. The complex task is completed by executing the language learning model (LLM) tasks associated with the workflow nodes and determining the subsequent path based on the output of the LLM in the conditional node.

[0087] like Figure 1 As shown, the present invention provides a method for constructing a multi-component data agent, which specifically includes the following steps:

[0088] Obtain sample data and build a vector knowledge base;

[0089] Extract knowledge business flows from sample data, determine the ontology through knowledge representation, and form a knowledge graph;

[0090] Based on the vector knowledge base, train the initial language model and provide a data processing model;

[0091] Combined with the logical tool chain and mind map of the data intelligent body, the knowledge graph and data processing model are componentized and linked. Through the process starting point, intermediate nodes, and logical control nodes, a workflow canvas between multiple components is formed to construct a multi-component data intelligent body.

[0092] Among them, the process starting point, intermediate nodes and logical control nodes represent the various subtasks of the data agent, and the edges formed between the points represent the dependency relationship between the subtasks.

[0093] Based on the knowledge vector library processing technology, and by integrating the knowledge graph to build a business rule engine, provide training data support, rely on data processing models, logical tool chains and mind maps and other platform architectures to build a multi-component data agent to enhance the accuracy of result reasoning using data agents.

[0094] like Figure 2 As shown, obtaining sample data and building a vector knowledge base specifically includes the following steps:

[0095] Obtain and analyze sample data, and delineate structured data and unstructured data;

[0096] Transform unstructured data into structured data and merge them into structured data;

[0097] The structured data is sliced ​​to form multiple data segments, wherein the structured data is sliced, specifically including:

[0098] Set the slice length to L, the number of sliding characters when slicing to S, and make the current slice overlap with the previous slice;

[0099] Based on the embedding model, the data fragments are converted into high-dimensional vectors by mapping the text data to the vector space;

[0100] The similarity data between each high-dimensional vector is determined, and a vector knowledge base is constructed using the high-dimensional vectors and the similarity data.

[0101] The construction of vector knowledge base mainly includes two parts: data collection and knowledge base processing.

[0102] The sample data is a multi-domain corpus, from which a multi-domain vector knowledge base can be constructed. The sample data obtained can come from a known knowledge base. For example, in the field of finance and auditing, the sample data also includes finance, law, etc., and can rely on more than 40,000 known high-quality accounting and auditing knowledge data such as policies, standards, accounting documents, and accounting textbooks. In addition, sample data can be collected incrementally, and the types of data collected are also in the policy, system, experience, case, rule, and other categories.

[0103] For unstructured data (such as audit reports, regulatory documents, etc.), useful information is extracted through text mining technology (such as named entity recognition and relationship extraction) and converted into structured data. For example, key audit findings in audit reports can be extracted; clauses related to audit processes, audit methods, and legal responsibilities can be extracted from regulatory documents to construct corresponding structured data sets.

[0104] For structured data (such as financial tables, audit intermediate tables, etc.), after data cleaning, extract key indicators (such as total assets, total liabilities, etc.) and build sub-tables that reflect financial health. After that, you can choose to slice the structured data in directory, length, and intelligent ways. Split long documents into multiple Chunk segments so that the model can process and retrieve them. In a specific application scenario, you can use a sliding window of length L, sliding S words each time to form overlapping text blocks to retain contextual information. For the segmented Chunk segments, perform semantic extraction to enrich the knowledge semantic elements and improve the recall accuracy.

[0105] The slice length L and the number of sliding characters S are adjusted according to the structural characteristics of the audit document. In the audit report, the selection of the sliding window should ensure that the slice contains complete audit findings and audit opinions. In the regulatory documents, the integrity of each system and regulation should be maintained, and the name of the regulatory document corresponding to the system should be added before each system.

[0106] Use an embedding model (such as Word2Vec, BERT, or GPT) to convert Chunk fragments into high-dimensional vectors. In this mapping conversion process, the data fragments are converted into high-dimensional vectors by mapping the text to the vector space.

[0107] Specifically, based on the embedding model, the data fragments are converted into high-dimensional vectors by mapping the text data to the vector space, which can be specifically expressed as:

[0108]

[0109] Among them, E is the embedding model, W is the fragment set formed by the data fragments, and R n is an n-dimensional real space, w is a data segment, v is an n-dimensional vector, v w is the n-dimensional vector corresponding to the data segment w.

[0110] Afterwards, the relationship between each high-dimensional vector is analyzed. Specifically, the similarity data between each high-dimensional vector is determined, which is specifically expressed as:

[0111]

[0112] Among them, S is the similarity data set between each high-dimensional vector, v1, v2, vm-1 、v m They are high-dimensional vectors numbered 1, 2, m-1, and m respectively.

[0113] The vector knowledge base is stored and constructed with high-dimensional vectors and the similarity data between each high-dimensional vector. The type of the vector knowledge base is not limited and can be types such as Milvus or FAISS. Based on the above construction scheme, the complex structured data is decomposed into multiple data fragments. The vector knowledge base supports efficient similarity search to retrieve the most relevant high-dimensional vectors and the data fragments corresponding to the high-dimensional vectors.

[0114] For a question-and-answer scenario, when a user asks a question, the question is first converted into a vector using an embedding model, and then the most similar high-dimensional vector is retrieved from the vector knowledge base, and data fragments and document blocks are given. This can be achieved by calculating the similarity between the question vector and all high-dimensional vectors in the vector knowledge base. In addition, because the slicing process uses an overlapping sliding method to retain context information and retrieve the most relevant document blocks based on the question vector, the question vector can be semantically corrected and these most relevant document blocks can be passed as context to the data processing model for result reasoning, thereby improving the accuracy of retrieval matching.

[0115] The collected sample data can be used to construct a corresponding knowledge graph, which can provide knowledge generation and planning capabilities. Figure 3 As shown in the figure, the ontology is determined through knowledge representation to form a knowledge graph, which specifically includes the following steps:

[0116] Extract knowledge and semantic tags from sample data to form a knowledge business flow, and determine the entities and relationships for knowledge representation from the knowledge business flow;

[0117] Based on the graph database, the entities and relationships of knowledge representation are combined to form a knowledge graph.

[0118] The sample data is first preprocessed, mainly including data cleaning, entity recognition and other steps. Data cleaning aims to remove erroneous, duplicate or incomplete information, and noise removal, data normalization and missing value processing are common data cleaning methods. Semantic labels are extracted from the preprocessed sample data, and professional business information including human resources, finance, materials, construction, operation and inspection, regulation, marketing, logistics, technology, information and communication, planning, transaction, economics and law, and comprehensive information are sorted out according to the semantic labels, as well as business steps, forms and fields in the process, business objects, data owners and other data structures to form a knowledge business flow.

[0119] Entity recognition relies on natural language processing technologies such as named entity recognition (NER) to identify knowledge business flows. Relationship extraction and representation rely on text analysis and data mining technologies, including rule-based methods, deep learning methods, and data processing model capabilities to extract structured knowledge. After that, mainstream knowledge representation models, such as Resource Description Framework (RDF), Web Ontology Language (OWL), and property graph models, are used to complete the core link in knowledge graph construction, namely, knowledge representation.

[0120] Choose a suitable graph database that is specially designed for processing graph data and provides efficient node and edge query and storage capabilities. For example, common graph databases such as Neo4j, ArangoDB, Apache Jena, etc., combine the entities and relationships of knowledge representation to form a knowledge graph. For example, still taking the financial audit field as an example, the knowledge graph focuses on professional content related to auditing, such as audit risks, audit findings, audit standards, and audit types (such as financial audits, compliance audits, etc.). Of course, the knowledge graph contains not only financial data and regulatory texts, but also proprietary knowledge such as audit methods and audit conclusions. The knowledge graph formed by the present invention is intended to help automate the audit process, improve audit efficiency and quality, and uncover more audit risk points.

[0121] Specifically, if Figure 4 As shown in the figure, based on the graph database, combined with the entities and relationships of knowledge representation, a knowledge graph is formed, which specifically includes:

[0122] Map each entity in the knowledge representation to a node in the graph database;

[0123] Based on the relationship between entities in the knowledge representation, the corresponding nodes in the graph database are connected to form edges, and the types and attributes of each edge are labeled;

[0124] Gather the nodes, edges, edge types and attributes of the graph database to form a knowledge graph;

[0125] Among them, the nodes, edges, edge types and attributes of the graph database are collected to form a knowledge graph, which is specifically expressed as:

[0126]

[0127] G is the knowledge graph set, N is the set of nodes, E is the set of edges, n i 、n j are nodes numbered i and j respectively, T ij is the type of edge connecting node i and node j, P ij is the attribute of the edge connecting node i and node j.

[0128] Through the entities and relationships of knowledge representation, the extracted entities and relationships can be stored in the graph database to form a knowledge graph. The above design, operation and function of nodes and edges are the basis for building and maintaining knowledge graphs. Through the combination of nodes and edges, complex data analysis, pattern matching and graph algorithms can be performed.

[0129] In addition, after forming the knowledge graph, it also includes updating the ontology, which is specifically expressed as:

[0130] Update sample data, repeat knowledge extraction and semantic label extraction, update knowledge business flow, and update entities and relationships of knowledge representation determined by knowledge business flow.

[0131] A dynamic ontology update mechanism is introduced into the knowledge graph to extract entities and relationships of knowledge representation mapped to nodes of the graph database from new sample data, thereby improving the accuracy and completeness of the knowledge graph.

[0132] The construction and training of data processing models, which are important links in the construction of data agents, need to rely on vector knowledge bases. The initial language model (e.g., large language model, LLM) is trained through the vector knowledge base to provide a data processing model suitable for building multi-component data agents.

[0133] like Figure 5 As shown, based on the vector knowledge base, the initial language model is trained and a data processing model is given, which specifically includes the following steps:

[0134] Based on the vector knowledge base, a training vector set is defined, wherein the training vector set includes a training vector and a corresponding actual result vector;

[0135] The initial language model analyzes the similarity of the training vector and gives the prediction result vector;

[0136] Based on the comparative analysis of the predicted result vector and the actual result vector, the initial language model is iterated to construct a data processing model;

[0137] Among them, the iterative initial language model is specifically expressed as:

[0138]

[0139] Among them, d is the prediction result vector given after iteration, is the initial similarity weight, s is the scaling factor, q is the training vector, and b is the bias term.

[0140] The initial language model is the Large Language Model (LLM). LLM is a technological advancement in the field of natural language processing (NLP). The core components of LLM include the Transformer architecture, self-attention mechanism, position encoding, etc. LLM can handle various complex language tasks, such as text generation, automatic question answering, text summarization, etc. The training of the initial language model by the vector knowledge base uses the similarity between the predicted result vector and the training vector as an indicator to continuously optimize the neural network parameters of the initial language model. In addition, the task instructions of the training vector will also be optimized so that the data processing model can perform tasks according to the expectations of the calling task.

[0141] When optimizing the neural network parameters iteratively, the model's processing capabilities can be accelerated by setting scaling factors and bias items, and the output results after optimization iteration can be quickly verified.

[0142] In the initial language model, the prediction of the training vector and the matching high-dimensional vector in the vector knowledge base is performed by analyzing the similarity of the training vector through the initial language model to give a prediction result vector, which specifically includes the following steps:

[0143] The initial language model is used to analyze the distance between the training vector and each high-dimensional vector in the vector knowledge base, and multiple sets of prediction data are given. The distance between the training vector and each high-dimensional vector in the vector knowledge base is specifically expressed as:

[0144]

[0145] s q is the training vector q and the high-dimensional vector d in the vector knowledge base y The distance, n is the dimension of the high-dimensional vector, q x is the x-th dimension vector of the training vector q, d yx is a high-dimensional vector d y The x-th dimension vector of ;

[0146] Sort multiple groups of prediction data, determine the optimal result, and give a prediction result vector.

[0147] In a question-and-answer scenario, the initial language model will generate a series of prediction data based on the initial retrieval of the training vector, forming multiple prediction data groups. The distance between these prediction data and the training vector meets the preset threshold and has a certain semantic relevance to the training vector. Of course, the prediction of the training vector needs to consider the information of each dimension of the training vector and comprehensively consider the features extracted from multiple dimensions.

[0148] For the training of the initial language model, by modifying the attention weight calculation method during the inference process, the initial token is kept visible to all subsequent tokens. At the same time, the upper limit of the distance in the attention calculation is limited to adapt to the relative distance that has not been seen during training, thereby achieving stable generation of very long texts.

[0149] The sorting of multiple sets of prediction data can adopt a machine learning model, such as a random forest, a gradient boosting machine or a deep learning model, to score each set of prediction data, select the prediction data with the highest score as the most relevant result, and determine it as the optimal result and the prediction result vector.

[0150] The construction of multi-component data intelligence uses logical tool chains and mind maps to orchestrate services in a componentized form using the data processing models and knowledge graphs developed in the early training. Figure 6 As shown in the figure, combined with the logical tool chain and mind map of the data agent, the knowledge graph and data processing model are componentized and linked. Through the process starting point, intermediate nodes, and logical control nodes, a workflow canvas between multiple components is formed to construct a multi-component data agent, which specifically includes the following steps:

[0151] Based on the logical tool chain and mind map of the data agent, combined with the knowledge map and data processing model, define the process starting point, intermediate nodes and logical control nodes;

[0152] Based on the weight setting of the knowledge graph and data processing model, combined with the logic control node, the weight is configured at the matching process starting point and intermediate nodes, and the process starting point and intermediate nodes are linked;

[0153] Component-based orchestration of knowledge graphs and data processing models to build an initial multi-component data agent;

[0154] The initial multi-component data agent is optimized to construct a multi-component data agent.

[0155] Using the logic tool chain and mind map as a means, through the process starting point, intermediate nodes, and logical control nodes, a workflow canvas between multiple components is formed, the input and output of different components are defined, and the flow links are formed to form a complete intelligent agent workflow. There is no specific limitation on the specific form of the logic tool chain and mind map, as long as the logical functions of the node and link settings can be realized.

[0156] In the mind map model, the reasoning process of the data agent is represented by a directed acyclic graph (DAG), where the process starting point, intermediate nodes and logical control nodes represent the subtasks of the data agent, and the edges formed between the points represent the dependencies between the subtasks. Each subtask contains specific thinking steps (for example, making assumptions, collecting information, analyzing data, etc.), and of course, specific thinking steps can be set according to specific application scenario tasks.

[0157] Combining multimodal models (knowledge graphs, etc.) with LLMs achieves a unified representation of knowledge graphs and data processing models, expands the boundaries of knowledge, and enhances the accuracy of the result reasoning of data agents.

[0158] In a specific application scenario, the componentized arrangement of knowledge graphs and data processing models can be completed by dragging components, transferring connections, and configuring component parameters to complete the preliminary data agent arrangement. In addition, the initial multi-component data agent can also import and load other types of models in a componentized manner. Other types of models include vector knowledge question and answer models, knowledge graph question and answer models, OCR models, etc.; plug-ins can also be imported into multi-component data agents. For example, plug-ins can be collected and imported according to the needs of audit intelligent arrangement, including web search, calculator, image generation, etc.; for non-model technologies, technical code completion and shared scripts can also be written, such as SQL data query execution code for querying audit data.

[0159] The multi-component data agent also includes a result generation module, which automatically generates a result report that meets the specifications based on the analysis results, risk assessment results, and pre-set templates. For example, the result generation module can use the text summary model to generate concise investigation conclusions, use the text generation model to generate natural and fluent text descriptions, and use the table generation model to generate structured table data.

[0160] like Figure 7 As shown, in an optional application scenario, such as "intelligent knowledge on the go", "intelligent assignment of tasks", "intelligent generation of portraits", etc., multi-component data agents form a complete agent workflow based on the input and output of different locations and the flow links.

[0161] For the initial multi-component data agent, it is often necessary to use a test data set to simulate real scenarios, such as simulating the financial statement data provided by the enterprise, business system data, and auditors' investigation records, etc., to evaluate the quality of the audit report generated by the agent. Based on the test results, analyze the weaknesses of the data agent and optimize it.

[0162] Specifically, optimizing the initial multi-component data agent and constructing a multi-component data agent includes the following steps:

[0163] Define the optimization indicators of the initial multi-component data agent; generally speaking, the evaluation indicators include: accuracy, that is, the accuracy of the information extracted by the data agent, the results of the analysis, the risk points identified, and the content of the generated report. Completeness, that is, whether the data agent covers all necessary procedures and risk points, and whether the report content is complete. Normativity, that is, whether the report generated by the data agent complies with relevant guidelines and specifications. Readability, that is, whether the report content generated by the data agent is logically clear, fluent in language, and easy to understand.

[0164] Perform optimization tests on the initial multi-component data agent and provide test values ​​corresponding to the optimization indicators;

[0165] According to the optimization strategy, the components to be optimized of the initial multi-component data agent are matched and optimized to construct a multi-component data agent;

[0166] Among them, the optimization strategy is specifically expressed as:

[0167]

[0168] Among them, P new For the optimized overall performance, P old For the overall performance before optimization, is the accuracy index weight, is the accuracy function increment, C is the component mapped to the accuracy indicator, is the completeness index weight, is the integrity function increment, R is the component relationship mapped to the integrity index, is the normative indicator weight, is the normative function increment, ICL F is the annotation parameter mapped to the normative indicator, Incremental readability function, ICL R are annotation parameters that map to readability metrics.

[0169] The optimization of multi-component data agents is reflected in logic adjustments, parameter configuration, etc., mainly through the addition of components, the splitting and aggregation of components, and the modification of the calling relationships between components, to optimize the accuracy and completeness indicators (i.e., coverage). By adjusting the labeled samples (ICL) of different categories, the standardization and readability indicators are optimized, and the data processing model is trained and adjusted according to the test results to improve the accuracy of the output results of the data agent.

[0170] In the mind map model, the addition of components is a generation transformation. According to the optimization strategy, components with corresponding functional optimization are generated. The decomposition of components is a refinement transformation, and the aggregation of components is a clustering transformation. The optimization strategy adopts a parallel approach, taking into account multiple independent subtasks and multiple indicators at the same time to achieve performance improvement of the data agent.

[0171] The data agent can use the refinement, generation, and clustering transformation operations of the mind map model to dynamically identify and analyze relevant content during the reasoning process. Taking the reasoning process of financial audit data as an example, refinement transformation: decompose complex audit risk tasks into smaller and more manageable subtasks. For example, "financial reporting risk" is refined into subcategories such as "revenue recognition risk" and "asset valuation risk". Generation transformation: dynamically add new audit considerations based on newly discovered information or risk factors. For example, when new regulatory changes are discovered, relevant compliance risk nodes and their connections are generated. Clustering transformation: group similar or related risk nodes together to form higher-level risk categories. For example, cluster multiple related operational risks into a broader "business process risk" category. The combined use of these transformation operations enables the mind map model to comprehensively analyze risk factors, dynamically adapt to new information, and effectively summarize major risk areas.

[0172] In the scenario of audit risk identification, the mind map will generate a series of possible risk points and associations based on the initial data agent. These risk points and relationships constitute multiple risk assessment paths, each of which represents a possible risk analysis solution. Through refinement, generation, and clustering transformations, the system can continuously optimize these risk assessment paths to ensure that all possible risk factors are fully considered. The risk identification method based on the mind map not only improves the accuracy and comprehensiveness of risk identification, but also enhances the interpretability and traceability of the entire reasoning process. Each step of the transformation operation can be recorded and tracked, which improves the transparency of the audit process and provides a powerful and flexible analysis tool for internal audits of enterprises.

[0173] In practical applications, the combination of mind maps with knowledge graphs and data processing models further improves the effect of audit risk identification. Knowledge graphs provide rich domain knowledge that can guide the reasoning process of mind maps; while data processing models can process and analyze large amounts of audit data to provide empirical support. Through the synergy of this multi-component data agent, it is possible to more comprehensively and accurately identify and assess the audit risks within the enterprise, providing strong support for enterprise management and decision-making.

[0174] like Figure 8 As shown, the present invention also provides a device for constructing a multi-component data intelligent body, characterized in that it adopts the method for constructing a component data intelligent body as described above, specifically comprising:

[0175] The acquisition module is used to obtain sample data and build a vector knowledge base;

[0176] The analysis module is used to extract knowledge business flows from sample data, determine the ontology through knowledge representation, and form a knowledge graph; based on the vector knowledge base, train the initial language model and provide a data processing model;

[0177] The construction module is used to combine the logical tool chain and mind map of the data intelligent body, organize and link the knowledge map and data processing model in a componentized manner, form a workflow canvas between multiple components through the process starting point, intermediate nodes, and logical control nodes, and construct a multi-component data intelligent body;

[0178] Among them, the process starting point, intermediate nodes and logical control nodes represent the various subtasks of the data agent, and the edges formed between the points represent the dependency relationship between the subtasks.

[0179] The present invention provides a method and device for constructing a multi-component data intelligent entity, which has at least the following beneficial effects:

[0180] (1) The present invention builds a business rule engine by integrating knowledge graphs and provides training data support. By modeling the thought chain process as a graph-structured thought graph and executing the data processing model and knowledge graph associated with the workflow node, the accuracy of the result reasoning of the data agent is enhanced, and the knowledge boundary is expanded. In particular, for financial audit data, key information can be better extracted, thereby improving the work efficiency of power grid auditors.

[0181] (2) Through the entities and relationships represented by knowledge, the extracted entities and relationships can be stored in the graph database to form a knowledge graph. The design, operation and function of nodes and edges are the basis for building and maintaining knowledge graphs. Through the combined use of nodes and edges, complex data analysis, pattern matching and graph algorithms can be performed.

[0182] (3) Introduce a dynamic ontology update mechanism into the knowledge graph to extract entities and relationships of knowledge representation mapped to nodes in the graph database from new sample data, thereby improving the accuracy and completeness of the knowledge graph.

[0183] (4) The vector knowledge base trains the initial language model, using the similarity between the predicted result vector and the training vector as an indicator to continuously optimize the neural network parameters of the initial language model. By modifying the calculation method of attention weights during the inference process, the initial token is kept visible to all subsequent tokens, while limiting the upper limit of the distance in the attention calculation to adapt to the relative distance that has not been seen during training, thereby achieving stable generation of ultra-long texts.

[0184] In addition, the task instructions of the training vectors will be optimized so that the data processing model can perform tasks according to the expectations of the calling tasks. When optimizing the neural network parameters iteratively, the processing capacity of the model is accelerated by setting the scaling factor and bias term, and the output results after the optimization iteration are quickly verified.

[0185] (5) The optimization of multi-component data agents is reflected in logic adjustment, parameter configuration, etc., mainly through the addition of components and the modification of the calling relationship between components, the accuracy and completeness indicators are optimized, and the standardization and readability indicators are optimized by adjusting the labeled samples (ICL) of different categories. In addition, the training and adjustment of the data processing model are carried out according to the test results to improve the accuracy of the output results of the data agent. By combining the data agent with the mind map model, not only the accuracy and comprehensiveness of risk identification are improved, but also the interpretability and traceability of the entire reasoning process are enhanced, providing a powerful and flexible analysis tool for the internal audit of the enterprise.

[0186] Although preferred embodiments of the present invention have been described, additional changes and modifications may be made to these embodiments by those skilled in the art once the basic inventive concepts are known. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for constructing a multi-component data agent, characterized in that: The specific steps include: Obtain sample data and build a vector knowledge base; Extract knowledge business flows from sample data, determine the ontology through knowledge representation, and form a knowledge graph; Based on the vector knowledge base, a training vector set is defined, wherein the training vector set includes a training vector and a corresponding actual result vector; The initial language model analyzes the similarity of the training vector and gives the prediction result vector; Based on the comparative analysis of the predicted result vector and the actual result vector, the initial language model is iterated to construct a data processing model; Among them, the iterative initial language model is specifically expressed as: ； Among them, d is the prediction result vector given after iteration, is the initial similarity weight, s is the scaling factor, q is the training vector, and b is the bias term; Combined with the logical tool chain and mind map of the data intelligent body, the knowledge graph and data processing model are componentized and linked. Through the process starting point, intermediate nodes, and logical control nodes, a workflow canvas between multiple components is formed to construct a multi-component data intelligent body. Among them, the process starting point, intermediate nodes and logical control nodes represent the various subtasks of the data agent, and the edges formed between the points represent the dependency relationship between the subtasks.

2. The method for constructing a multi-component data agent as claimed in claim 1, characterized in that: Obtain sample data and build a vector knowledge base, which includes the following steps: Obtain and analyze sample data, and delineate structured data and unstructured data; Transform unstructured data into structured data and merge them into structured data; The structured data is sliced ​​to form multiple data segments, wherein the structured data is sliced, specifically including: Set the slice length to L, the number of sliding characters when slicing to S, and make the current slice overlap with the previous slice; Based on the embedding model, the data fragments are converted into high-dimensional vectors by mapping the text data to the vector space; The similarity data between each high-dimensional vector is determined, and a vector knowledge base is constructed using the high-dimensional vectors and the similarity data.

3. The method for constructing a multi-component data agent as claimed in claim 2, characterized in that: Based on the embedding model, the data fragments are converted into high-dimensional vectors by mapping the text data to the vector space, which can be specifically expressed as: ； Among them, E is the embedding model, W is the fragment set formed by the data fragments, and R n n is the real number space, w is the data segment, v is the n-dimensional vector, v w is the n-dimensional vector corresponding to the data segment w; Determine the similarity data between each high-dimensional vector, which is specifically expressed as: ； Among them, S is the similarity data set between each high-dimensional vector, v1, v2, v m-1 、v m They are high-dimensional vectors numbered 1, 2, m-1, and m respectively.

4. The method for constructing a multi-component data agent as claimed in claim 1, characterized in that: Determine the ontology through knowledge representation and form a knowledge graph, which specifically includes the following steps: Extract knowledge and semantic tags from sample data to form a knowledge business flow, and determine the entities and relationships for knowledge representation from the knowledge business flow; Based on the graph database, the entities and relationships of knowledge representation are combined to form a knowledge graph.

5. The method for constructing a multi-component data agent as claimed in claim 4, characterized in that: Based on the graph database, combined with the entities and relationships of knowledge representation, a knowledge graph is formed, which includes: Map each entity in the knowledge representation to a node in the graph database; Based on the relationship between entities in the knowledge representation, the corresponding nodes in the graph database are connected to form edges, and the types and attributes of each edge are labeled; Gather the nodes, edges, edge types and attributes of the graph database to form a knowledge graph; Among them, the nodes, edges, edge types and attributes of the graph database are collected to form a knowledge graph, which is specifically expressed as: ； G is the knowledge graph set, N is the set of nodes, E is the set of edges, n i 、n j are nodes numbered i and j respectively, T ij is the type of edge connecting node i and node j, P ij is the attribute of the edge connecting node i and node j.

6. The method for constructing a multi-component data agent according to claim 1, characterized in that: The initial language model is used to analyze the similarity of the training vector to give a prediction result vector, which specifically includes the following steps: The initial language model is used to analyze the distance between the training vector and each high-dimensional vector in the vector knowledge base, and multiple sets of prediction data are given. The distance between the training vector and each high-dimensional vector in the vector knowledge base is specifically expressed as: ； s q is the training vector q and the high-dimensional vector d in the vector knowledge base y The distance, n is the dimension of the high-dimensional vector, q x is the x-th dimension vector of the training vector q, d yx is a high-dimensional vector d y The x-th dimension vector of ; Sort multiple groups of prediction data, determine the optimal result, and give a prediction result vector.

7. The method for constructing a multi-component data agent as claimed in claim 1, characterized in that: Combined with the logical tool chain and mind map of the data agent, the knowledge graph and data processing model are componentized and linked. Through the process starting point, intermediate nodes, and logical control nodes, a workflow canvas between multiple components is formed to construct a multi-component data agent. The specific steps include: Based on the logical tool chain and mind map of the data agent, combined with the knowledge map and data processing model, define the process starting point, intermediate nodes and logical control nodes; Based on the weight setting of the knowledge graph and data processing model, combined with the logic control node, the weight is configured at the matching process starting point and intermediate nodes, and the process starting point and intermediate nodes are linked; Component-based orchestration of knowledge graphs and data processing models to build an initial multi-component data agent; The initial multi-component data agent is optimized to construct a multi-component data agent.

8. The method for constructing a multi-component data agent as claimed in claim 7, characterized in that: Optimizing the initial multi-component data agent and constructing a multi-component data agent specifically includes the following steps: Define the optimization metrics for the initial multi-component data agent; Perform optimization tests on the initial multi-component data agent and provide test values ​​corresponding to the optimization indicators; According to the optimization strategy, the components to be optimized of the initial multi-component data agent are matched and optimized to construct a multi-component data agent; Among them, the optimization strategy is specifically expressed as: ； Among them, P new For the optimized overall performance, P old For the overall performance before optimization, is the accuracy index weight, is the accuracy function increment, C is the component mapped to the accuracy indicator, is the completeness index weight, is the integrity function increment, R is the component relationship mapped to the integrity index, is the normative indicator weight, is the normative function increment, ICL F is the annotation parameter mapped to the normative indicator, Incremental readability function, ICL R are annotation parameters that map to readability metrics.

9. A device for constructing a multi-component data intelligent entity, characterized in that: The method for constructing a component data intelligent entity as claimed in any one of claims 1 to 8 specifically comprises: The acquisition module is used to obtain sample data and build a vector knowledge base; The analysis module is used to extract knowledge business flows from sample data, determine the ontology through knowledge representation, and form a knowledge graph; based on the vector knowledge base, train the initial language model and provide a data processing model; The construction module is used to combine the logical tool chain and mind map of the data intelligent body, organize and link the knowledge map and data processing model in a componentized manner, form a workflow canvas between multiple components through the process starting point, intermediate nodes, and logical control nodes, and construct a multi-component data intelligent body; Among them, the process starting point, intermediate nodes and logical control nodes represent the various subtasks of the data agent, and the edges formed between the points represent the dependency relationship between the subtasks.

Images