Data marking method for problem solving process based on AI model

Abstract

The invention relates to the technical field of data marking, and discloses a data marking method for a problem solving process based on an AI model, and the method comprises the steps: obtaining a process action and operation record information thereof in historical business process data, and building a process event graph database with the process action as a node and a node dynamic relationship as an edge; the work recording agent extracts an entity by using a BERT-CRF model, generates a structured phenomenon statement and mounts the structured phenomenon statement to a corresponding node; the progress agent integrates nodes, edges and phenomenon statements, marks quality color labels, establishes a mapping relation library, and uses a GNN model to train a quality evaluation model to obtain node quality colors; the opportunity agent mines a frequent sub-graph by using a gSpan algorithm based on a second color node and an edge, obtains a high-value sub-graph through value analysis, and transmits the high-value sub-graph to the upper computer; the risk agent calculates the root cause probability of a third color node based on a DoWhy framework, and if the root cause probability exceeds a threshold value, a process patch is triggered and blockchain evidence storage is carried out; and finally, summarizing to generate a full-process optimization decision report and transmitting the report to an upper computer.

Description

TECHNICAL FIELD

[0001] The present specification relates to the technical field of data labeling, in particular to a data labeling method for problem solving process based on AI model. BACKGROUND

[0002] As the core link of model training data construction, the quality and efficiency of artificial intelligence data labeling directly affect the value of data assets and the performance of downstream models. "Problem traceability" and "quality controllable" are the key basis to ensure the standardization of labeling process and improve data reliability. However, the existing data labeling technology system still faces the following deficiencies in large-scale application: In the aspect of business process control, labeling projects often fall into a vicious cycle of "repeatedly fixing the same problem", such as "boundary box offset" in image labeling or "entity category confusion" in text labeling. These errors still occur repeatedly after multiple corrections, but there is a lack of technical means to locate the root cause of the problem, making it difficult to optimize the process, and causing serious loss of manpower and time cost; In the aspect of data management, the existing labeling system only focuses on the correct or incorrect judgment of the final result, completely missing the structured retention of the problem solving path. The system only records the data qualified state, but does not save the key traceability information such as the operation track of labeling, auditing, rework, data version evolution and problem correction node, forming a "result black box", which makes it difficult to trace the problem and optimize the process without data support; In the aspect of analysis and decision-making, project review highly depends on manual report writing, which requires a lot of time to integrate scattered records, and the conclusion is easily affected by personal experience and memory bias, resulting in strong subjectivity of review conclusion.

[0003] Therefore, there is an urgent need for a data labeling method for problem solving process based on AI model, which can realize precise optimization and quality closed-loop management of labeling process through structured process modeling, process data retention and intelligent analysis and decision-making. SUMMARY Therefore, the present application aims to provide a data labeling method for problem solving process based on AI model to solve the problems of repeated rework of the same problem, fuzzy root cause positioning of quality and low efficiency of process optimization caused by the lack of problem traceability, insufficient process data structuring and reliance on manual experience for review optimization in the existing data labeling system.

[0004] To achieve the above-mentioned purpose, the present application adopts the following technical solutions: A data labeling method for problem solving process based on AI model, the method comprising: S1, obtaining a plurality of entities in historical business process data and operation record information corresponding to the entities; the entity is a process action, and the operation record information is an operation log and screen OCR text generated in the execution process of the process action.

[0005] S2, establish a flow event graph database; construct a flow event graph with process actions as nodes and dynamic relationships between nodes as edges, and store the nodes and edges in the flow event graph database; the dynamic relationship between the nodes includes the execution order between the nodes and the data transmitted between the nodes.

[0006] S3, through the work record intelligent agent, using a process entity recognition model based on BERT-CRF architecture to extract entities from the operation record information, generate phenomenon sentences; and mount the phenomenon sentences to the nodes corresponding to the operation record information through node matching; the phenomenon sentence is a structured data containing identifier, phenomenon type, key entity and timestamp.

[0007] S4, through the progress intelligent agent, integrate all nodes, edges and phenomenon sentences in the flow event graph database to obtain node feature vectors; label quality color labels based on the node feature vectors, and establish a mapping relationship library through feature label matching method; distribute the sample data in the mapping relationship library into training data set and validation data set according to the data amount of 4:1; select GNN model as the initial training model, train the initial training model based on the training data set, input the validation data set into the trained initial training model, and when the output accuracy of the initial training model is higher than 90%, the model is used as the quality evaluation model; input the node feature vector to be evaluated into the quality evaluation model to obtain the quality color of the corresponding node; the quality color includes the first color representing no error of the node, the second color representing non-fatal error of the node and the third color representing fatal error of the node.

[0008] S5, through the opportunity intelligent agent, based on the second color node and the edge corresponding to the second color node, frequent subgraphs are mined by gSpan algorithm; based on the historical business process data, the value of the frequent subgraph is analyzed to obtain high value frequent subgraph; the high value frequent subgraph is packaged according to the preset format and transmitted to the upper computer.

[0009] S6, through the risk intelligent agent, based on the DoWhy causal judgment framework, the root cause probability of the third color node is calculated to obtain a root cause probability ranking table; if the root cause probability exceeds the preset threshold, the process patch is triggered and packaged, and the process patch is written into the blockchain through the Ethereum smart contract for storage.

[0010] S7, summarize the phenomenon sentences, quality color labels, high value frequent subgraphs and process patches in the above steps S3-S6 to generate a full process optimization decision report and transmit it to the upper computer to realize the marking and dynamic optimization of the historical business process.

[0011] The beneficial effects of the present scheme are: in the prior art, the data labeling business process generally has the problems of repeated rework of the same problem, fuzzy quality root cause positioning, and process optimization relying on manual experience summary, which restricts the continuous improvement of labeling efficiency and quality. The present application realizes the leap from manual review to data-driven intelligent tracing and optimized decision-making by constructing a process event graph database, deploying a multi-agent collaborative analysis mechanism, and based on GNN quality evaluation and causal inference root cause positioning, effectively improving the accuracy of problem positioning, the automation degree of process optimization, and the closed-loop level of quality control.

[0012] Further, the step S2 further comprises: S210, assigning a unique identifier to each node and creating a node table to store node attributes; S220, assigning a unique identifier to each edge and creating an edge table to store edge attributes; the edge attributes are used to represent the dynamic relationship between nodes; S230, structurally storing the node table and the edge table through unique identifier association; wherein the node table stores the node attributes with the node unique identifier as the primary key, and the edge table stores the edge attributes with the edge unique identifier as the primary key and establishes a foreign key association with the node table, to build a process event graph database.

[0013] By assigning unique identifiers to nodes and edges and establishing structurally associated node tables and edge tables for storage, standardized management and efficient organization of process event data are achieved. This design ensures accurate tracing and rapid retrieval of process actions and their dynamic relationships, providing a solid and reliable data foundation for subsequent multi-agent analysis and full-process optimization.

[0014] Further, the step S3 further comprises: S310, cleaning, segmenting and normalizing the operation record information.

[0015] S320, inputting the preprocessed operation record information into a BERT model to generate word-level vectors; predicting the entity type label of each word based on the word-level vectors through a CRF layer; and merging the entity type label of each word to obtain a complete entity.

[0016] S330, generating the phenomenon sentence based on the extracted entity according to a preset template, and matching the node unique identifier in the phenomenon sentence with the corresponding node.

[0017] S340, store the matched phenomenon statement as an extended attribute into the node table and the edge table in the process event graph database. By cleaning and structuring the operation record information, and generating standardized phenomenon statements based on a preset template, accurate extraction and efficient integration of unstructured process information are realized. This method converts scattered logs and texts into structured data with clear semantics and correlation, significantly improving the readability and analyzability of operation records, and providing high-quality and standardized data input for subsequent quality evaluation and root cause analysis.

[0018] Further, in the step S340, the matched phenomenon statement is stored in the process event graph database in JSON format. By storing the structured phenomenon statement in JSON format, efficient organization and flexible expansion of complex process data are realized. JSON format has good readability and cross-platform compatibility, which not only facilitates direct parsing and query by the system, but also supports subsequent addition of fields without modifying the database structure, effectively improving the adaptability and maintainability of the system to various business data.

[0019] Further, in the step S5, when mining frequent subgraphs by gSpan algorithm, the support of the second color node is calculated based on its frequency of occurrence, and the support of the edge corresponding to the second color node is calculated based on its weight; the calculation logic of the weight is as follows: ; In the formula, is the weight of the edge; is used to represent the node is followed by the node The total number of times this happens; is the total number of times the node occurs. The refined definition of edge weight effectively improves the accuracy and business interpretability of frequent subgraph mining, and can more accurately identify truly high-frequency and stable non-fatal error patterns in the process.

[0020] Further, the step S6 further includes: S610, taking the third color node as the result variable Y, taking the nodes, edges upstream of the third color node and the phenomenon statements mounted thereon as potential cause variables X, and building a causal graph based on the DoWhy framework; S620, identify and estimate the causal effect based on the causal graph, and calculate the average causal effect value of each potential cause variable X on the result variable Y; S630, ranking all potential cause variables X according to the average causal effect value, and generating a root cause probability ranking table. By taking the fatal error node in the process as the result variable, and systematically quantifying the average causal effect of each factor upstream based on the causal inference framework, the precise positioning and quantitative sorting of the root cause of the problem are realized. This method overcomes the limitations of traditional root cause analysis, which relies on manual experience and is highly subjective, and can automatically and objectively identify the most critical influencing factors from complex process associations, providing reliable data support for subsequent development of accurate and effective process patches.

[0021] Further, in step S610, the causal graph further includes a confounding variable Z; the confounding variable Z is another variable that has a direct causal path with both the potential cause variable X and the result variable Y. By introducing the confounding variable Z to construct a more complete causal graph, the interference of other potential factors on causal analysis is effectively excluded, significantly improving the accuracy and reliability of the average causal effect estimation, making the root cause positioning result more objective and reliable.

[0022] Further, in step S620, the estimation of the average causal effect value is realized based on a double machine learning model, and the core calculation logic is as follows: ; In the formula, is the average causal effect value; is a result regression model; is a propensity score model; is a result variable; is a cause variable. By using a double machine learning model to estimate the causal effect, the model setting bias in controlling confounding variables in traditional methods is effectively overcome. This model can flexibly handle high-dimensional confounding variables, and accurately strip the confounding factor influence through orthogonalization calculation, thereby obtaining a more reliable average causal effect value. BRIEF DESCRIPTION OF DRAWINGS

[0023] The present specification will be further illustrated in the form of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting, and in these embodiments, the same numbers represent the same structures, wherein: Figure 1 is an exemplary flowchart of a data labeling method for a problem solving process based on an AI model; Figure 2 is an exemplary flowchart of constructing a process event graph database; Figure 3 is an exemplary flowchart of entity extraction and information mounting; Figure 4 is an exemplary flowchart of a causal inference method for root cause analysis. Detailed Implementation

[0024] To more clearly illustrate the technical solutions of the embodiments in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some examples or embodiments of this specification. For those skilled in the art, these drawings can be applied to other similar scenarios without creative effort. Unless obvious from the context or otherwise specified, the same reference numerals in the drawings represent the same structures or operations.

[0025] As indicated in this specification and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of expressly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0026] Flowcharts are used in this specification to illustrate the operations performed by the system according to embodiments of this specification. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, the steps can be processed in reverse order or simultaneously. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0027] The following detailed explanation illustrates the specific implementation methods: Example 1: Figure 1 This is an exemplary flowchart of a data labeling method for problem-solving processes based on AI models, such as... Figure 1 As shown, a data labeling method for problem-solving processes based on an AI model, executed by a processor, includes: Step S1: Obtain several entities from historical business process data and the corresponding operation record information for each entity. The entity is a process action, and the operation record information consists of operation logs and screen OCR text generated during the execution of the process action.

[0028] Historical business process data refers to a complete set of operation records from task distribution, annotation execution, quality review to problem rework in completed data annotation projects. In this embodiment, historical business process data can be obtained in various ways, such as by retrieving data creation, deletion, modification, and query operation records from a database in a host computer.

[0029] Operation logs are system-generated running logs and operation trace logs recorded during user operations. Examples include user operation records and system logs.

[0030] In this embodiment, operation logs can be obtained in various ways. For example, they can be obtained from the database in the host computer, recording data creation, deletion, modification, and query operations.

[0031] Screen OCR text is text information extracted from screenshots of the operation interface corresponding to the execution of a process action using OCR technology. Examples include button names, pop-up prompts, and data field values.

[0032] In this embodiment, screen OCR text can be obtained in various ways. For example, screen recording technology can be used to capture screenshots of the interface during the operation process, and OCR processing can be performed on the captured screenshots to extract the interface text information; another example is to obtain the text content and status information of interface controls through UI automation testing tools.

[0033] Step S2: Establish a process event graph database; construct a process event graph with process actions as nodes and dynamic relationships between nodes as edges, and store the nodes and edges in the process event graph database.

[0034] A process event graph database is a database used to store, manage, and query events and their complex relationships within a business process. In this embodiment, the process event graph database may include graph databases such as Neo4j and Nebula Graph.

[0035] A node is the smallest unit of operation in the process of execution. In this embodiment, an edge specifically represents each process action in the data annotation task flow. For example, a node may include process actions such as task reception, image pre-annotation, label classification, initial quality review, and result submission.

[0036] An edge represents a directed relationship connecting adjacent nodes. In this embodiment, an edge specifically represents the dynamic relationship between nodes; the dynamic relationship between nodes includes the execution order between nodes and the data transmitted between nodes.

[0037] Furthermore, such as Figure 2 As shown, storing nodes and edges in the process event graph database in step S2 specifically includes: Step S210: Assign a unique identifier to each node and create a node table to store node attributes.

[0038] An identifier is a string or numeric code used to uniquely identify and locate each node within a database.

[0039] The node table is a structured data table used to centrally store the core attributes of all nodes. In this embodiment, the node table uses the node's unique identifier as the primary key.

[0040] The node attribute refers to a set of metadata describing the characteristics and state of the node. In the embodiment, the node attribute can include fixed fields such as node type, timestamp, execution state, and reserved extension fields for storing dynamic attributes.

[0041] In step S220, a unique identifier is assigned to each edge, and an edge table is created to store edge attributes; the edge attributes are used to represent the dynamic relationship between nodes.

[0042] The edge table is a structured data table used to store and manage all connection relationships in the process event graph.

[0043] In the embodiment, the edge attributes can include timing information, starting node identifier, ending node identifier, etc.

[0044] In step S230, the node table and the edge table are stored in a structured manner through unique identifier association; wherein the node table stores node attributes with node unique identifier as the primary key, the edge table stores edge attributes with edge unique identifier as the primary key and establishes foreign key association with the node table, and a process event graph database is constructed.

[0045] In the embodiment, by assigning unique identifiers to nodes and edges and constructing a structured and associated node table and edge table, efficient organization and accurate representation of dynamic relationships of business process data are achieved, which not only guarantees independent management of node attributes and flexible update of dynamic relationships of edges, but also improves the query efficiency and data consistency of the graph database through the primary key-foreign key association mechanism, providing a reliable data foundation for subsequent frequent pattern mining, causal reasoning and quality evaluation based on graph structure.

[0046] In step S3, the work record agent is used to extract entities from the operation record information by using a process entity recognition model based on the BERT-CRF architecture, and generate phenomenon sentences; and the phenomenon sentences are hung on the nodes corresponding to the operation record information through node matching.

[0047] The work record agent is an agent used to realize automatic analysis and structured extraction of operation record information. In the embodiment, the agent can automatically identify and extract key entities by natural language processing of the operation record information, and convert them into standardized phenomenon sentences, thereby converting unstructured operation records into structured data that can be queried and analyzed.

[0048] The BERT-CRF architecture is a sequence labeling model framework for accurately extracting structured information from text. BERT is a pre-trained language model that deeply understands the true meaning of each word in a specific context. CRF is used to ensure the legitimacy of the labels. The word-level vector output by BERT is input into the CRF model, and CRF can output the best label sequence for all words in the entire sentence.

[0049] A phenomenon statement is a structured data containing an identifier, a phenomenon type, an entity, and a timestamp. As an example, the phenomenon statement can be: {node identifier: "001", phenomenon type: "error prompt", entity: ["data format error", "version v2.1"], timestamp: "2023-10-01T14:30:00"}.

[0050] Further, as shown in Figure 3 Step S3 further includes: Step S310, cleaning, sentence segmentation, and normalization preprocessing of the operation record information.

[0051] Step S320, inputting the preprocessed operation record information into the BERT model to generate word-level vectors; predicting the entity type label of each word based on the word-level vectors through the CRF layer; and merging the entity type label of each word to obtain a complete entity.

[0052] Step S330, generating a phenomenon statement based on a preset template according to the extracted entity, and matching the phenomenon statement with the corresponding node through the node unique identifier in the phenomenon statement.

[0053] The preset template is set according to prior experience; when matching the node unique identifier with the corresponding node, it is also necessary to ensure that the timestamp of the phenomenon statement overlaps with the node execution time range, otherwise it is a mounting failure.

[0054] Step S340, storing the matched phenomenon statement as an extended attribute in the node table and the edge table in the process event graph database.

[0055] Preferably, in step S340, the matched phenomenon statement is stored in the process event graph database in JSON format. By using the JSON format to store the structured phenomenon statement, efficient organization and flexible expansion of complex process data are achieved. JSON format has good readability and cross-platform compatibility, which not only facilitates direct parsing and querying by the system, but also supports subsequent addition of fields without modifying the database structure, effectively improving the adaptability and maintainability of the system to various business data.

[0056] In this embodiment, after the storage is completed, 10% of the mounting results are randomly extracted, and the accuracy of entity extraction and phenomenon sentence generation is verified by manual verification.

[0057] In this embodiment, entity extraction is performed by combining BERT and CRF, which improves the entity recognition accuracy and boundary rationality; phenomenon sentences are generated based on a preset template and accurately matched to process action nodes through unique identifiers, ensuring strong association between semantic structured expression and business entities; finally, the phenomenon sentences are stored as extended attributes in the node table and edge table of the graph database, enhancing the multi-dimensional information carrying capacity of the process event graph.

[0058] Step S4, by the progress agent, all nodes, edges and phenomenon sentences in the process event graph database are integrated and processed to obtain node feature vectors; quality color labels are labeled based on the node feature vectors, and a mapping relationship library is established through feature label matching; sample data in the mapping relationship library are distributed according to a data amount ratio of 4:1 into a training data set and a verification data set; a GNN model is selected as an initial training model, and the initial training model is trained based on the training data set; the verification data set is input into the trained initial training model, and when the output accuracy of the initial training model is higher than 90%, the model is used as a quality evaluation model; the node feature vector to be evaluated is input into the quality evaluation model to obtain the quality color of the corresponding node; the quality color includes a first color representing no error of the node, a second color representing non-fatal error of the node, and a third color representing fatal error of the node.

[0059] Among them, the progress agent is an agent for realizing automatic evaluation of the quality of the process action node.

[0060] Step S5, by the opportunity agent, based on the second color node and the edge corresponding to the second color node, a frequent subgraph is mined by a gSpan algorithm; a high-value frequent subgraph is obtained by value analysis of the frequent subgraph based on historical business process data; and the high-value frequent subgraph is packaged in a preset format and transmitted to an upper computer.

[0061] Among them, the opportunity agent is an agent for mining non-fatal error patterns and identifying process optimization opportunities.

[0062] Further, in the step S5, when the frequent subgraph is mined by the gSpan algorithm, the support of the second color node is calculated based on its frequency. Specifically, the number of instances of the node appearing in the historical process event graph set is counted, and the ratio of the number to the total number of process instances is taken as the support.

[0063] The support of the edge corresponding to the second color node is calculated based on its weight, and the calculation logic is as follows: ; wherein, is the weight of the edge; is used to represent the node is followed by the node the total number of times this case occurs; is the node the total number of times it appears.

[0064] In this embodiment, the weight of the edge is calculated by the formula, and then the conditional probability of the transition between the nodes can be quantified, reflecting the frequency of state transition in the process, so that the representative process path is more accurately identified in the frequent subgraph mining, and the frequent subgraph structure with more reliable support is obtained.

[0065] Step S6, through the risk agent, the root cause probability of the third color node is calculated based on the DoWhy causal judgment framework, and a root cause probability ranking table is obtained; if the root cause probability exceeds a preset threshold, a process patch is triggered and the process patch is packaged, and the process patch is written into the blockchain through the Ethereum smart contract for storage.

[0066] The risk agent is an intelligent agent for identifying fatal error root causes and automatically generating risk prevention and control measures.

[0067] The root cause probability ranking table refers to a list formed by arranging in descending order the causal probability values calculated for each potential cause variable that leads to the occurrence of the third color node; it is used to represent the influence of each suspicious factor on the occurrence of the problem, and provides data basis for accurate implementation of the process patch.

[0068] Further, as Figure 4 indicated, for the calculation of the root cause probability ranking table, step S6 further includes: Step S610, taking the third color node as the result variable Y, the nodes, edges and phenomenon statements mounted thereon upstream of the third color node as the potential cause variable X, and building a causal graph based on the DoWhy framework.

[0069] Further, in step S610, the causal graph further includes a confounding variable Z; the confounding variable Z is another variable that has a direct causal path with both the potential cause variable X and the result variable Y. By introducing the confounding variable Z in the causal graph, external interference factors that simultaneously affect the cause variable and the result variable are effectively identified and controlled, significantly improving the accuracy of causal effect estimation and the reliability of conclusions.

[0070] Step S620, based on the causal graph, the causal effect is identified and estimated, and the average causal effect value of each potential cause variable X on the result variable Y is calculated.

[0071] Further, in step S620, the calculation of the average causal effect value is realized based on a double machine learning model, and the core calculation logic is as follows: ; In the formula, is the average causal effect value; is a result regression model; is a propensity score model; is a result variable; is a cause variable.

[0072] In this embodiment, by calculating the average causal effect value, the interference of the confounding variable Z on the cause variable X and the result variable Y can be effectively stripped, so that the net causal effect strength of the cause variable X on the result variable Y can be accurately quantified.

[0073] In step S630, all potential cause variables X are sorted according to the average causal effect value, and a root cause probability sorting table is generated.

[0074] In this embodiment, by taking the fatal error node in the process as the result variable and systematically quantifying the average causal effect of each factor upstream based on the causal inference framework, the precise positioning and quantification of the root cause of the problem are realized. This method overcomes the limitations of traditional root cause analysis, which relies on manual experience and is highly subjective, and can automatically and objectively identify the most critical influencing factors from complex process associations, providing reliable data support for subsequent development of accurate and effective process patches.

[0075] In step S7, the phenomenon sentences, quality color labels, high-value frequent subgraphs, and process patches in steps S3-S6 are summarized to generate a full-process optimization decision report and transmit it to the upper computer to realize the labeling and dynamic optimization of historical business processes.

[0076] In this embodiment, by constructing a process event graph database and deploying four types of intelligent agents for work records, progress, opportunities, and risks to work together, the whole intelligent processing from structured extraction of original operation records, automatic evaluation of node quality, non-fatal error pattern mining to fatal error root cause analysis is realized. This method changes the traditional discrete problem tracing relying on manual review into a data-driven optimization mode based on graph neural networks and causal inference, significantly improving the problem positioning accuracy, process analysis efficiency, and automation level of optimization decisions.

[0077] Having described the basic concepts, it is obvious that the above detailed disclosure is merely intended for purposes of illustration and is not intended to limit the present specification. Although not explicitly described, those skilled in the art can make various modifications, improvements and corrections to the present specification. Such modifications, improvements and corrections are suggested in the present specification, and thus still belong to the spirit and scope of the exemplary embodiments of the present specification.

[0078] Further, the order of the processing elements and sequences described in this specification are not intended to be construed as a limitation, unless specifically specified in the claims. Although some of the above-disclosed exemplary methods embody the present specification, it is understood that one or more of the steps can be carried out in a manner that is different than what is explicitly described. For example, although the system components described above can be implemented by hardware devices, they can also be implemented by software solutions only, such as installing the described system on an existing server or mobile device.

[0079] Similarly, it is to be noted that the description of the embodiments of the present specification in the foregoing description can sometimes refer to various features or aspects of the present specification in terms of being incorporated into one embodiment, figure or description of a particular aspect provided in a single embodiment. However, the description of the embodiments of the present specification in the foregoing description does not mean that a greater number of features of the embodiments of the present specification are required than those recited in the claims. In fact, the features of the embodiments are less than the total features of the single embodiment disclosed above.

[0080] Some embodiments use numerical values to describe components, quantities of attributes, and it is understood that such numerical values used in the description of the embodiments are in some examples modified by the adjectives "about", "approximately", or "substantially". Unless otherwise stated, "about", "approximately", or "substantially" indicate that the described numerical value allows for a ±20% variation. Accordingly, in some embodiments, numerical values used in the specification and claims are approximations that can vary depending on the desired characteristics of the individual embodiments. In some embodiments, numerical values should be considered to be defined with the specified significant digits and carried out with the normal rounding procedures. Although the numerical ranges and parameters setting forth the broadest scope of the embodiments of the present specification are approximations, in specific embodiments, the numerical values set forth are intended to be used in a precise manner.

[0081] Each patent, patent application, patent publication, and other material cited in this specification is hereby incorporated by reference in its entirety herein for the teachings relevant to the sentence and / or paragraph in which the reference is presented. Document histories, to the extent not inconsistent with the pertinent U.S. patent application file history, are also incorporated by reference herein. To the extent that material incorporated by reference contradicts or contradicts any portion of this specification, including definition, the portion of the material incorporated by reference prevails. Note, however, that in the event of inconsistencies between any such material and the present specification, including definitions, the present specification, including definitions, will control.

[0082] Finally, it should be understood that the embodiments described herein are merely exemplary of the principles of the present description. Other embodiments can be devised without departing from the scope of the present description. Accordingly, the embodiments described herein are not intended to limit the scope of the present description, but rather are intended to be exemplary thereof.

Claims

1. A data labeling method for problem-solving processes based on an AI model, characterized in that, The method includes: S1. Obtain several entities from historical business process data and the corresponding operation record information of the entities; the entities are process actions, and the operation record information is the operation log and screen OCR text generated during the execution of the process actions; S2. Establish a process event graph database; construct a process event graph with process actions as nodes and dynamic relationships between nodes as edges, and store the nodes and edges in the process event graph database; the dynamic relationships between nodes include the execution order between nodes and the data transmitted between nodes; S3. Using the work record intelligent agent, the process entity recognition model based on the BERT-CRF architecture is used to extract entities from the operation record information and generate phenomenon statements; and the phenomenon statements are attached to the nodes corresponding to the operation record information through node matching; the phenomenon statements are a type of structured data containing identifiers, phenomenon types, key entities and timestamps. S4. Through the progress agent, all nodes, edges, and phenomenon statements in the process event graph database are integrated and processed to obtain node feature vectors; quality color labels are marked based on the node feature vectors, and a mapping relationship library is established through feature label matching; the sample data in the mapping relationship library is allocated into training dataset and validation dataset at a data volume ratio of 4:1; a GNN model is selected as the initial training model, and the initial training model is trained based on the training dataset. The validation dataset is input into the trained initial training model. When the output accuracy of the initial training model is higher than 90%, the model is used as the quality evaluation model; the feature vector of the node to be evaluated is input into the quality evaluation model to obtain the quality color of the corresponding node; the quality color includes a first color representing that the node has no errors, a second color representing that the node has non-fatal errors, and a third color representing that the node has fatal errors. S5. Using an opportunistic agent, based on the second color node and the edge corresponding to the second color node, a frequent subgraph is mined using the gSpan algorithm; based on the historical business process data, the frequent subgraph is value-analyzed to obtain a high-value frequent subgraph; the high-value frequent subgraph is packaged in a preset format and transmitted to the host computer. S6. Through the risk intelligence agent, the root cause probability of the third color node is calculated based on the DoWhy causal judgment framework to obtain the root cause probability ranking table; if the root cause probability exceeds the preset threshold, the process patch is triggered and the process patch is encapsulated and written to the blockchain for evidence storage through the Ethereum smart contract. S7. Summarize the phenomenon statements, quality color labels, high-value frequent subgraphs, and process patches described in steps S3-S6 above, generate a full-process optimization decision report, and transmit it to the host computer to realize the marking and dynamic optimization of historical business processes.

2. The data labeling method for problem-solving processes based on an AI model according to claim 1, characterized in that, Step S2 further includes: S210. Assign a unique identifier to each node and create a node table to store node attributes; S220. Assign a unique identifier to each edge and create an edge table to store edge attributes; the edge attributes are used to characterize the dynamic relationships between nodes. S230. The node table and the edge table are structured and stored by association with unique identifiers; wherein, the node table stores the node attributes with the node unique identifier as the primary key, and the edge table stores the edge attributes with the edge unique identifier as the primary key and establishes a foreign key association with the node table to store the edge attributes, thereby constructing a process event graph database.

3. The data labeling method for problem-solving processes based on an AI model according to claim 1, characterized in that, Step S3 further includes: S310. The operation record information is cleaned, segmented, and normalized preprocessed. S320. Input the preprocessed operation record information into the BERT model to generate word-level vectors; predict the entity type label of each word through the CRF layer based on the word-level vectors; merge the entity type labels of each word to obtain the complete entity; S330. Based on the extracted entities, generate the phenomenon statement according to the preset template, and match the corresponding node with the node unique identifier in the phenomenon statement; S340. The matched phenomenon statements are stored as extended attributes in the node table and edge table of the process event graph database, respectively.

4. The data labeling method for problem-solving processes based on an AI model according to claim 3, characterized in that, In step S340, the matched phenomenon statement is stored in the process event graph database in JSON format.

5. The data labeling method for problem-solving processes based on an AI model according to claim 1, characterized in that, In step S5, when mining frequent subgraphs using the gSpan algorithm, the support of the second-colored node is calculated based on its frequency of occurrence, and the support of the edge corresponding to the second-colored node is calculated based on its weight. The calculation logic for the weight is as follows: ； In the formula, Let be the weight of the edge; Used to characterize nodes Then comes the node. The total number of times this occurred; For nodes Total number of occurrences.

6. The data labeling method for problem-solving processes based on an AI model according to claim 1, characterized in that, Step S6 further includes: S610. Using the third color node as the result variable Y, and the upstream nodes, edges, and the phenomenon statements attached to the third color node as potential cause variables X, construct a causal graph based on the DoWhy framework. S620. Based on the causal diagram, identify and estimate causal effects, and calculate the average causal effect value of each potential cause variable X on the outcome variable Y. S630. Sort all potential causal variables X according to the average causal effect value to generate a root cause probability ranking table.

7. A data labeling method for problem-solving processes based on an AI model according to claim 6, characterized in that, In step S610, the causal graph further includes a confounding variable Z; the confounding variable Z is other variables that have a direct causal path with both the potential cause variable X and the outcome variable Y.

8. A data labeling method for problem-solving processes based on an AI model according to claim 7, characterized in that, In step S620, the estimation of the average causal effect value is based on a dual machine learning model, the core calculation logic of which is as follows: ； In the formula, This represents the average causal effect value. For the outcome regression model; For propensity score model; For the outcome variable; This is the causal variable.

Images