Workflow orchestration methods, devices, computer equipment, storage media and products

By acquiring natural language requirements and historical interaction information, extracting entity information, and using intent recognition dimensions for feature extraction and orchestration, the problem of inaccurate response to requirements in traditional workflow orchestration solutions is solved, achieving more accurate and targeted workflow orchestration.

CN123117752APending Publication Date: 2026-11-17CHINA CONSTRUCTION BANK +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611464636.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Traditional automated workflow orchestration solutions struggle to adapt to users' flexible demands, resulting in inaccurate responses to those demands.

Method used

By acquiring the natural language requirements and historical interaction information of the target object, entity information and relationships are extracted, features are extracted using preset intent recognition dimensions, target intents that meet the intent recognition conditions are selected, and workflow orchestration is performed based on the target intents and entity information.

Benefits of technology

It improves the accuracy and relevance of workflow response to requirements, ensuring that workflow nodes, sequences, and data processing logic are consistent with the actual needs of the target object, and reducing the possibility of misjudgment due to insufficient information.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117752A_ABST
    Figure CN123117752A_ABST
Patent Text Reader

Abstract

This application relates to the field of intelligent process control technology, and discloses a workflow orchestration method, apparatus, computer equipment, storage medium, and computer program product. The method includes: acquiring fusion information corresponding to a target object; extracting entity information from natural language demand information, the entity information including multiple key entities and the relationships between these key entities; extracting features from the fusion information according to multiple preset intent recognition dimensions to obtain the dimension recognition features corresponding to the target object under each intent recognition dimension; selecting target intents that meet the intent recognition conditions from multiple pre-constructed candidate intents based on the dimension recognition features; and orchestrating the workflow based on the target intent and entity information to obtain a target workflow for responding to natural language demand information. This method can improve the accuracy and relevance of demand responses.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of intelligent process control technology, and in particular to a workflow orchestration method, apparatus, computer equipment, storage medium and computer program product. Background Technology

[0002] With the rapid development of fintech and artificial intelligence, financial institutions are increasingly diversifying their business types and service scenarios. Customer needs are becoming more complex, diverse, personalized, and dynamic, and cross-departmental and cross-system business processes are gradually increasing. Traditional manual workflow orchestration methods are no longer sufficient to meet the requirements for efficient, accurate, and flexible responses. Against this backdrop, automated workflow orchestration has become a key element in improving the efficiency of financial business processing, reducing the cost of manual intervention, and enhancing the intelligence level of the system.

[0003] Currently, automated workflow orchestration typically employs rule-driven, template-based, or static process configuration approaches. These approaches rely on predefined rule bases or process templates, and the construction of rules and templates largely depends on human experience, resulting in limited coverage and difficulty in encompassing complex and ever-changing natural language expressions. Therefore, current automated workflow orchestration solutions struggle to adapt to flexible user requirements, leading to inaccurate responses to demands. Summary of the Invention

[0004] Therefore, it is necessary to provide a workflow orchestration method, apparatus, computer equipment, computer-readable storage medium, and computer program product that can improve the accuracy and relevance of response to the above-mentioned technical problems.

[0005] Firstly, this application provides a workflow orchestration method. The method includes:

[0006] Obtain the fusion information corresponding to the target object; the fusion information is obtained by fusing the natural language requirement information input to the target object and the historical interaction information of the target object;

[0007] Entity information is extracted from the natural language demand information, and the entity information includes multiple key entities and the relationships between the key entities.

[0008] According to multiple preset intent recognition dimensions, feature extraction is performed on the fused information to obtain the dimension recognition features of the target object under each intent recognition dimension.

[0009] Based on the identification features of each dimension, target intents that meet the intent identification conditions are selected from a pre-constructed set of candidate intents; the candidate intents are obtained by combining the candidate intent categories corresponding to each of the intent identification dimensions.

[0010] Based on the target intent and the entity information, workflow orchestration is performed to obtain a target workflow for responding to the natural language requirement information.

[0011] In one embodiment, the step of identifying target intents that meet the intent recognition conditions from a pre-constructed set of candidate intents based on the features identified in each dimension includes:

[0012] Based on the identification features of each dimension and the multiple candidate intent categories contained in each of the pre-constructed multiple sets of candidate intents, the intent matching degree of each candidate intent is determined;

[0013] In the case where there are multiple selected intentions with an intent matching degree greater than or equal to a preset matching degree threshold, the selected intention with the highest intent matching degree among the selected intentions is determined as the target intent that meets the intent recognition conditions.

[0014] In one embodiment, the method further includes:

[0015] If the matching degree of each intent is less than the preset matching degree threshold, an intent supplement prompt page is generated; the intent supplement prompt page contains intent supplement prompt information determined according to the matching degree of each intent.

[0016] In response to an input operation on the intent supplement prompt page, obtain the intent supplement information of the target object;

[0017] The fused information is supplemented based on the intent supplement information to obtain updated fused information;

[0018] The updated fusion information is used to return and execute the step of extracting features from the fusion information according to multiple preset intent recognition dimensions, until there is a selected intent with an intent matching degree greater than or equal to the preset matching degree threshold.

[0019] In one embodiment, determining the intent matching degree of each candidate intent based on the features identified in each dimension and the multiple candidate intent categories contained in each of the pre-constructed multiple sets of candidate intents includes:

[0020] For each candidate intent category of each candidate intent, the category feature of the candidate intent category is matched with the dimension recognition feature of the intent recognition dimension to which the candidate intent category belongs, to obtain the matching similarity of the candidate intent category.

[0021] Based on the respective dimensional weights of each intent recognition dimension, the matching similarity of each candidate intent category is weighted and summed to obtain the intent matching degree of the candidate intent.

[0022] In one embodiment, the workflow orchestration based on the target intent and the entity information to obtain a target workflow for responding to the natural language requirement information includes:

[0023] If the target object has a historical intent, the target intent is matched with the historical intent to obtain a matching result;

[0024] If the matching result indicates that there is an intent offset between the target intent and the historical intent, then the intent offset information is determined based on the entity information;

[0025] Based on the intent offset information, the historical workflow of the historical intent is corrected to obtain the target workflow that responds to the natural language requirement information.

[0026] In one embodiment, the workflow orchestration based on the target intent and the entity information to obtain a target workflow for responding to the natural language requirement information includes:

[0027] The large language model is invoked to parse the target intent and the entity information to generate a preliminary workflow framework;

[0028] Each work step in the preliminary workflow framework is decomposed into corresponding sub-tasks, resulting in multiple sub-tasks;

[0029] For each subtask, a preset mapping relationship is searched according to the task type of the subtask to determine the task execution model that matches the task type; the preset mapping relationship is used to characterize the correspondence between each task type and each task execution model;

[0030] Each subtask is sent to its corresponding task execution model, and task execution information is received from each task execution model.

[0031] The initial workflow framework is refined based on the task execution information to generate a target workflow that responds to the natural language requirement information.

[0032] In one embodiment, the method further includes:

[0033] Obtain the execution result of the target workflow;

[0034] The target workflow is evaluated based on the execution results using preset evaluation metrics to obtain the evaluation result of the target workflow;

[0035] If the evaluation result indicates that the target workflow does not meet the execution criteria, the target workflow is modified according to the evaluation result to obtain the modified target workflow;

[0036] Based on the revised target workflow, return to the step of obtaining the execution result of the target workflow until the target workflow meets the execution criteria.

[0037] In one embodiment, the method further includes:

[0038] The key entities are normalized to obtain their standard names in the financial field.

[0039] For each of the aforementioned key entities, the corresponding graph node is searched in the financial domain knowledge graph based on the standard name of the key entity.

[0040] By traversing the financial knowledge graph with the graph node as the center, the graph association information of the key entity is obtained;

[0041] Based on the association information of each of the aforementioned maps, the fusion information is supplemented to obtain the updated fusion information;

[0042] Using the updated fusion information, perform the step of extracting features from the fusion information according to multiple preset intent recognition dimensions.

[0043] Secondly, this application also provides a workflow orchestration apparatus. The apparatus includes:

[0044] The information acquisition module is used to acquire fusion information corresponding to the target object; the fusion information is obtained by fusing natural language requirement information input to the target object and the target object's historical interaction information;

[0045] An information extraction module is used to extract entity information from the natural language demand information. The entity information includes multiple key entities and the relationships between the key entities.

[0046] The feature extraction module is used to extract features from the fused information according to multiple preset intent recognition dimensions, so as to obtain the dimension recognition features of the target object under each intent recognition dimension.

[0047] The target intent determination module is used to identify features based on each of the aforementioned dimensions and filter out target intents that meet the intent recognition conditions from a pre-constructed set of candidate intents; the candidate intents are obtained by combining the candidate intent categories corresponding to each of the aforementioned intent recognition dimensions.

[0048] The workflow orchestration module is used to orchestrate workflows based on the target intent and the entity information to obtain a target workflow for responding to the natural language requirement information.

[0049] In one embodiment, the target intent determination module is used to: determine the intent matching degree of each of the candidate intents based on the identification features of each dimension and the multiple candidate intent categories contained in each of the pre-constructed multiple sets of candidate intents; and when there are multiple selected intents with intent matching degrees greater than or equal to a preset matching degree threshold, determine the selected intent with the highest intent matching degree among the selected intents as the target intent that meets the intent recognition conditions.

[0050] In one embodiment, the device further includes:

[0051] The matching threshold comparison module is used to generate an intent supplement prompt page when the matching degree of each intent is less than the preset matching degree threshold; the intent supplement prompt page includes intent supplement prompt information determined according to the matching degree of each intent.

[0052] An operation response module is used to respond to an input operation on the intent supplement prompt page and obtain intent supplement information of the target object;

[0053] An information supplementation module is used to supplement the fused information based on the intent supplementation information to obtain updated fused information;

[0054] The loop execution module is used to return to the execution step of extracting features from the fusion information according to multiple preset intent recognition dimensions, using the updated fusion information, until there is a selected intent with an intent matching degree greater than or equal to the preset matching degree threshold.

[0055] In one embodiment, the target intent determination module is configured to: for each candidate intent category of each candidate intent, perform similarity matching between the category features of the candidate intent category and the dimensional recognition features of the intent recognition dimension to which the candidate intent category belongs, to obtain the matching similarity of the candidate intent category; and perform weighted summation of the matching similarities of each candidate intent category according to the dimensional weight of each intent recognition dimension, to obtain the intent matching degree of the candidate intent.

[0056] In one embodiment, the workflow orchestration module is configured to: match the target intent with the historical intent when the target object has a historical intent, and obtain a matching result; if the matching result indicates that there is an intent offset between the target intent and the historical intent, determine intent offset information based on the entity information; and correct the historical workflow of the historical intent based on the intent offset information to obtain a target workflow that responds to the natural language requirement information.

[0057] In one embodiment, the workflow orchestration module is configured to: invoke a large language model to parse the target intent and the entity information to generate a preliminary workflow framework; decompose each work step in the preliminary workflow framework into corresponding subtasks to obtain multiple subtasks; for each subtask, according to the task type of the subtask, search a preset mapping relationship to determine a task execution model that matches the task type; the preset mapping relationship is used to characterize the correspondence between each task type and each task execution model; send each subtask to the corresponding task execution model and receive task execution information fed back by each task execution model; refine the preliminary workflow framework according to the task execution information to generate a target workflow that responds to the natural language requirement information.

[0058] In one embodiment, the device further includes:

[0059] The execution result feedback module is used to obtain the execution result of the target workflow;

[0060] The result evaluation module is used to evaluate the target workflow based on the execution result using preset evaluation indicators, and obtain the evaluation result of the target workflow;

[0061] The workflow correction module is used to correct the target workflow according to the evaluation results when the evaluation results indicate that the target workflow does not meet the execution criteria, thereby obtaining a corrected workflow.

[0062] In one embodiment, the device further includes:

[0063] The normalization processing module is used to normalize each of the key entities to obtain the standard names of each key entity in the financial field;

[0064] The graph lookup module is used to find the corresponding graph node in the financial domain knowledge graph for each of the key entities, based on the standard name of the key entity.

[0065] The graph association module is used to traverse the financial domain knowledge graph with the graph node as the center to obtain the graph association information of the key entity;

[0066] The graph association and supplementation module is used to supplement the fusion information according to the graph association information to obtain the updated fusion information;

[0067] The feature extraction execution module is used to perform a step of extracting features from the fused information according to multiple preset intent recognition dimensions using the updated fused information.

[0068] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the above-described method.

[0069] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

[0070] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the steps of the above-described method.

[0071] The aforementioned workflow orchestration method, apparatus, computer equipment, storage medium, and computer program product, by fusing the natural language requirement information input by the target object with the target object's historical interaction information, obtains fused information. This enables intent understanding to obtain more sufficient contextual basis, thereby reducing the possibility of misjudgment due to insufficient information. Simultaneously, multiple key entities and the relationships between these key entities are extracted from the natural language requirement information, providing a clear data object and dependency foundation for subsequent workflow orchestration. Subsequently, feature extraction is performed on the fused information according to multiple preset intent recognition dimensions, obtaining the target object's corresponding dimensional recognition features under each intent recognition dimension. Based on these dimensional recognition features, the target intent that meets the intent recognition conditions is selected from multiple sets of candidate intents obtained by combining candidate intent categories corresponding to each intent recognition dimension. Since the candidate intents cover multiple dimensional combinations, the determination of the target intent is more global and accurate. Finally, workflow orchestration is performed based on the target intent and entity information to obtain the target workflow for responding to the natural language requirement information. This ensures that the workflow nodes, sequence, and data processing logic are consistent with the target object's actual needs and entity relationships, thereby improving the accuracy and relevance of responding to the target object's needs. Attached Figure Description

[0072] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0073] Figure 1 This is a diagram illustrating the application environment of a workflow orchestration method in one embodiment.

[0074] Figure 2This is a flowchart illustrating a workflow orchestration method in one embodiment;

[0075] Figure 3 This is a flowchart illustrating the workflow orchestration method in another embodiment;

[0076] Figure 4 This is a flowchart illustrating a process of orchestrating a workflow based on target intent and entity information to obtain a target workflow for responding to natural language request information in one embodiment.

[0077] Figure 5 In another embodiment, a flowchart is shown to illustrate the process of orchestrating a target workflow based on the target intent and entity information to obtain a target workflow for responding to natural language demand information.

[0078] Figure 6 This is a flowchart illustrating the workflow orchestration method in yet another embodiment;

[0079] Figure 7 Here is a flowchart illustrating the workflow orchestration method in one embodiment;

[0080] Figure 8 This is a flowchart illustrating the workflow orchestration method in another embodiment;

[0081] Figure 9 This is a structural block diagram of a workflow orchestration device in one embodiment;

[0082] Figure 10 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0083] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. It should be noted that existing industry solutions such as software, components, and models may be mentioned in the embodiments of this application. These should be considered exemplary and are intended only to illustrate the feasibility of implementing the technical solutions of this application, but do not imply that the applicant has already used or necessarily used such solutions.

[0084] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with relevant regulations. The acquisition, storage, use and processing of data in the technical solution of this application all comply with the relevant provisions of national laws and regulations.

[0085] The terms "first," "second," etc., used in this application may be used to describe various elements, but these elements are not limited by these terms. These terms are used only to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0086] The workflow orchestration method provided in this application embodiment can be applied to, for example, Figure 1 In the application environment shown, business system 102 communicates with user terminal 104 via a network. A data storage system can store the data that business system 102 needs to process. The data storage system can be integrated into business system 102 or placed in the cloud or on other network servers. Business system 102 integrates an orchestration tool 1021. The orchestration tool 1021 can obtain natural language demand information input by the target object through user terminal 104 via business system 102, fuse the natural language demand information with the target object's historical interaction information to obtain fused information corresponding to the target object, and extract entity information from the natural language demand information. The entity information includes multiple key entities and the relationships between these key entities. Subsequently, according to multiple preset intent recognition dimensions, feature extraction is performed on the fused information to obtain the dimension recognition features corresponding to each intent recognition dimension for the target object. Based on the dimension recognition features, target intents that meet the intent recognition conditions are selected from multiple pre-constructed candidate intents. The candidate intents are obtained by combining the candidate intent categories corresponding to each intent recognition dimension. Finally, based on the target intent and entity information, the workflow is orchestrated to obtain the target workflow used to respond to natural language requirements.

[0087] The user terminal 104 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. Portable wearable devices can be smartwatches, smart bracelets, head-mounted devices, etc. The business system 102 can be integrated on the user terminal 104 or on a separate server. The server can be a separate server or a server cluster consisting of multiple servers.

[0088] In one embodiment, such as Figure 2 As shown, a workflow orchestration method is provided, which can be applied to... Figure 1 Taking the 1021 layout tool as an example, the explanation includes the following steps:

[0089] S202, Obtain the fusion information corresponding to the target object. The fusion information is obtained by fusing the natural language requirement information input by the target object and the historical interaction information of the target object.

[0090] The target object refers to the user object that makes a request and accepts the system's processing. In the financial business scenario, the target object can refer to the user or customer who inputs natural language request information.

[0091] Among them, fused information is a unified information identifier formed by merging two information sources. It can be obtained by fusing natural language demand information input by the target object and the target object's historical interaction information.

[0092] Natural language requirement information refers to the target audience's description of their needs expressed in natural language; that is, unstructured requirements input by the target audience in the form of text or voice. For example, a user could input that they want to conduct risk assessments for recently registered corporate clients and then recommend financial products. This input constitutes natural language requirement information.

[0093] Among them, historical interaction information refers to the interaction information generated when the target object interacted with the business system before this dialogue, such as the input (natural language text), output (corresponding workflow orchestration results), execution status, user feedback and other recorded information generated when the target object interacted with the system in the past.

[0094] In one embodiment, a time-series database can be used to store historical interaction information of the target object. Each historical record in the historical interaction information can be represented as a tuple (input_i, workflow_i, status_i, feedback_i), where i represents the number of interactions between the target user and the business system.

[0095] For example, the orchestration tool can obtain fusion information corresponding to the target object. This fusion information is obtained by fusing the natural language requirement information input by the target object with the target object's historical interaction information. By fusing natural language requirement information and historical interaction information, the orchestration tool can understand the target object's target intent based on context. For instance, if the target object has previously conducted a risk assessment for a general customer, when it inputs "conduct a risk assessment" again, the orchestration tool can determine, based on the context, whether it is targeting the same type of customer or whether the assessment parameters need to be adjusted.

[0096] In one embodiment, the fusion information may include a fusion vector. The orchestration tool can acquire natural language requirement information input by the target object, and simultaneously acquire the target object's historical interaction information based on the target object's object identifier. The natural language requirement information is then vectorized to obtain a first vector corresponding to the natural language requirement information. The historical interaction information is then vectorized to obtain a second vector corresponding to the historical interaction information. Finally, the first vector and the second vector are weighted and fused according to preset fusion weights to obtain the fusion vector.

[0097] In one embodiment, the orchestration tool may also use vector concatenation to merge the first vector and the second vector.

[0098] For example, the first vector corresponding to natural language demand information is vector x, and the second vector corresponding to historical interaction information is vector c. Vectors x and c can be fused by concatenation or weighted summation to obtain a fused vector Z = [x; c] or Z = α × x + (1 - α) × c. Here, α is a preset fusion weight, a learnable weight parameter that can be obtained through training.

[0099] S204 Extract entity information from natural language demand information. The entity information includes multiple key entities and the relationships between them.

[0100] Entity information consists of business objects and their relationships extracted from natural language demand information. Entity information can include multiple key entities and the relationships between them. Key entities are business-meaning objects in the natural language demand information, i.e., words or phrases that have a substantial impact on business processing. For example, key entities may include corporate customers, risk assessment, and financial products. The relationships between key entities are used to characterize the semantic or business connections between them. For example, a corporate customer is the object of risk assessment, and the risk assessment result is the basis for recommending financial products.

[0101] For example, the orchestration tool can extract entity information from natural language requirement information, which includes multiple key entities and the relationships between them.

[0102] In one embodiment, the orchestration tool may be configured with a large language model, which uses named entity recognition (NER) technology to extract key entities from natural language requirement information. Then, semantic role labeling is used to clarify the relationships between entities, clarify the specific entity steps and processes, and obtain entity information.

[0103] S206, according to multiple preset intent recognition dimensions, feature extraction is performed on the fused information to obtain the corresponding dimension recognition features of the target object under each intent recognition dimension.

[0104] Among them, the intent recognition dimension is a classification dimension used to analyze the intent of the target object from different perspectives. For example, the intent recognition dimension may include, but is not limited to, the business scenario dimension, the operation behavior dimension, and the operation object dimension.

[0105] Feature extraction is the process of encoding fused information according to the intent recognition dimension to obtain corresponding dimensional recognition features. Dimensional recognition features are the feature identifiers of fused information under the corresponding intent recognition dimension; that is, the feature vectors obtained after feature extraction of fused information according to the corresponding intent recognition dimension. For example, under the business scenario dimension, the dimensional recognition features extracted from fused information are biased towards customers and products. Under the operational behavior dimension, the dimensional recognition features extracted from fused information will be biased towards processing and recommendation.

[0106] For example, the orchestration tool can extract features from the fused information according to multiple preset intent recognition dimensions to obtain the corresponding dimension recognition features of the target object under each intent recognition dimension.

[0107] S208: Based on the identification features of each dimension, select the target intent that meets the intent identification conditions from multiple pre-built candidate intents.

[0108] Among them, candidate intents are obtained by combining the candidate intent categories corresponding to each intent recognition dimension. The possible intent option orchestration tool needs to select the target intent that best matches the needs of the target object from multiple groups of candidate intents based on the recognition features of each dimension.

[0109] Candidate intent categories are category options under the corresponding intent recognition dimension. Understandably, each intent recognition dimension has predefined intent categories that can be combined. For example, candidate intent categories for the business scenario dimension may include, but are not limited to, customers, products, activities, and knowledge. Candidate intent categories for the operational behavior dimension may include, but are not limited to, question and answer, query, processing, modification, and recommendation. Candidate intent categories for the operational object dimension may include, but are not limited to, accounts, contracts, reports, customers, and financial products.

[0110] Among them, the intent recognition condition is a preset judgment condition used to determine whether the candidate intent category can represent the needs of the target object, and is used to filter out the target intent from multiple groups of candidate intents.

[0111] In one embodiment, the intent recognition condition can be that the overall matching score of the candidate intent is greater than a first preset score threshold, and the candidate intent category score corresponding to each dimension is not lower than a second preset score threshold.

[0112] For example, the orchestration tool can identify target intents that meet the intent recognition criteria from a pre-built set of candidate intents based on features identified in various dimensions.

[0113] S210, based on the target intent and entity information, the workflow is orchestrated to obtain the target workflow for responding to natural language demand information.

[0114] Workflow orchestration refers to the process of determining the nodes, node order, dependencies, inputs and outputs, and model allocation required for the business process corresponding to the needs of the target object, based on the target intent and entity information. The target workflow is the final executable process generated.

[0115] For example, an orchestration tool can orchestrate workflows based on target intent and entity information to obtain a target workflow for responding to natural language demand information.

[0116] In the aforementioned workflow orchestration method, fused information is obtained by fusing the natural language requirement information input by the target object with its historical interaction information. This provides more sufficient contextual basis for intent understanding, thereby reducing the possibility of misjudgment due to insufficient information. Simultaneously, multiple key entities and their interrelationships are extracted from the natural language requirement information, providing a clear data object and dependency foundation for subsequent workflow orchestration. Subsequently, feature extraction is performed on the fused information according to multiple preset intent recognition dimensions, obtaining the target object's corresponding dimensional recognition features for each dimension. Based on these dimensional recognition features, the target intent that meets the intent recognition conditions is selected from multiple candidate intents obtained by combining candidate intent categories corresponding to each intent recognition dimension. Since the candidate intents cover multiple dimensional combinations, the determination of the target intent is more global and accurate. Finally, workflow orchestration is performed based on the target intent and entity information to obtain the target workflow for responding to the natural language requirement information. This ensures that the workflow nodes, sequence, and data processing logic are consistent with the target object's actual needs and entity relationships, thereby improving the accuracy and relevance of the response to the target object's needs.

[0117] In one embodiment, S208, based on the identification features of each dimension, the target intent that meets the intent identification conditions is selected from a pre-constructed set of candidate intents, including: determining the intent matching degree of each candidate intent based on the identification features of each dimension and the multiple candidate intent categories contained in each of the pre-constructed set of candidate intents. If there are multiple selected intents with intent matching degrees greater than or equal to a preset matching degree threshold, the selected intent with the highest intent matching degree among the selected intents is determined as the target intent that meets the intent identification conditions.

[0118] Intent matching degree is a metric used to characterize the degree of similarity between candidate intent and the true intent of the target object. Understandably, a higher intent matching degree indicates a greater degree of similarity between the candidate intent and the true intent of the target object, and vice versa.

[0119] The preset matching threshold is a numerical limit used to determine whether the intent matching degree of a candidate intent meets the requirements. Understandably, if the intent matching degree is greater than or equal to the preset matching threshold, it means that the candidate intent matches the user's true intent to the required degree, meets the intent recognition conditions, and can be selected as an intent. Conversely, if the intent matching degree is less than the preset matching threshold, it means that the candidate intent does not match the user's true intent to the required degree, does not meet the intent recognition conditions, cannot be selected as an intent, and must be excluded.

[0120] Among them, the selected intent is the candidate intent filtered by the threshold. If the intent matching degree of the candidate intent is greater than or equal to the preset matching degree threshold, the candidate intent can be determined as the selected intent.

[0121] For example, the orchestration tool can determine the intent matching degree of each candidate intent based on the identification features of each dimension and the multiple candidate intent categories contained in each of the pre-built multiple sets of candidate intents. Then, it compares the intent matching degree of each candidate intent with a preset matching degree threshold, and identifies the candidate intent with an intent matching degree greater than or equal to the preset matching degree threshold as the selected intent. In the case of multiple selected intents, the selected intent with the highest intent matching degree is identified as the target intent that meets the intent recognition criteria. For example, the orchestration tool can sort the intent matching degrees of each selected intent in descending order, and identify the selected intent corresponding to the first intent matching degree as the target intent that meets the intent recognition criteria.

[0122] In one embodiment, if there is only one selected intent, that is, only one candidate intent with an intent matching degree greater than or equal to a preset matching degree threshold, then the selected intent can be directly determined as the target intent of the target object.

[0123] In the above embodiments, by matching the identification features of each dimension with multiple sets of candidate intent categories, the degree of consistency between each candidate intent and the user's true intent can be quantified. Threshold filtering eliminates low-matching candidates, ensuring basic recognition quality. Selecting the highest-matching candidate from multiple qualified candidates as the target intent can effectively reduce misjudgments caused by a single dimension or a single candidate, thereby improving the accuracy and reliability of intent recognition.

[0124] In one embodiment, such as Figure 3 As shown, workflow orchestration methods may also include:

[0125] S302, if the matching degree of each intent is less than the preset matching degree threshold, generate an intent supplement prompt page.

[0126] The intent supplementation prompt page is an interactive interface used to guide the target user to provide supplementary intent information. This page can contain intent supplementation prompts determined based on the intent matching degree. These prompts specifically indicate to the target user what aspects of the content they can supplement. Understandably, the intent supplementation prompts are determined based on the intent matching degree.

[0127] In one embodiment, when it is determined that the matching degree of each intent is less than a preset matching degree threshold, the candidate intents can be sorted in descending order according to their matching degrees to obtain a candidate intent sequence. The first and second candidate intents in the sequence are then compared in terms of dimensional differences to determine the difference in intent identification dimensions between them. Finally, supplementary intent prompts are determined based on these difference in intent identification dimensions. For example, if the difference in intent identification dimensions are the business scenario dimension and the operational behavior dimension, supplementary intent prompts such as "Do you want to conduct customer business or recommend financial products?" can be generated.

[0128] For example, when the orchestration tool determines that the matching degree of each intent is less than the preset matching degree threshold, it can generate intent supplementation prompts and generate corresponding intent supplementation prompt pages based on the intent supplementation prompts.

[0129] S304, in response to an input operation on the intent supplement prompt page, obtain intent supplement information for the target object.

[0130] Input operations are interactive actions performed by the target object in response to the intent supplementation prompt page, such as clicking, selecting, or inputting information. Intent supplementation information is supplemented by the target object to improve intent recognition.

[0131] For example, the orchestration tool can display an intent supplementation prompt page on the user terminal and, in response to the target object's input operation on the intent supplementation prompt page, obtain the target object's intent supplementation information.

[0132] S306, Supplement the fused information based on the intent supplement information to obtain updated fused information.

[0133] For example, the orchestration tool can supplement the fused information based on intent supplement information to obtain updated fused information.

[0134] In one embodiment, the orchestration tool can vectorize the intent supplement information to obtain the intent supplement vector, and then fuse the intent supplement vector with the fusion vector to obtain the updated fusion information, i.e., the updated fusion vector.

[0135] S308, using the updated fusion information, return and perform the step of extracting features from the fusion information according to multiple preset intent recognition dimensions, until there is a selected intent with an intent matching degree greater than or equal to a preset matching degree threshold.

[0136] For example, the orchestration tool can use the updated fusion information to return and perform the step of extracting features from the fusion information according to multiple preset intent recognition dimensions, until there is a selected intent with an intent matching degree greater than or equal to a preset matching degree threshold.

[0137] In the above embodiments, when the matching degree of each intent is less than the preset threshold, it indicates that the current fused information is insufficient to distinguish candidate intents. Supplementary prompt information is determined by the matching degree of each intent, which can locate the dimension where the information is missing or ambiguous. After the user supplements the information, the fused information is updated and feature extraction and matching are re-executed. This enables intent recognition to gradually converge under the drive of supplementary information until a qualified intent is selected, thereby reducing the probability of direct recognition failure due to insufficient information and improving the success rate of intent recognition.

[0138] In one embodiment, based on the identification features of each dimension and the multiple candidate intent categories contained in each of the pre-constructed multiple sets of candidate intents, the intent matching degree of each candidate intent is determined, including: for each candidate intent category of each candidate intent, performing similarity matching between the category features of the candidate intent category and the dimensional identification features of the intent identification dimension to which the candidate intent category belongs, to obtain the matching similarity of the candidate intent category; and performing a weighted summation of the matching similarities of each candidate intent category according to the dimensional weight of each intent identification dimension, to obtain the intent matching degree of the candidate intent.

[0139] Among them, category features are feature information used to represent the semantics of candidate intent categories, which can be generated in advance for each candidate intent category.

[0140] Similarity matching between categorical features and dimensional recognition features refers to the process of comparing categorical features and dimensional recognition features and calculating their similarity. Matching similarity is a numerical value used to characterize the degree of similarity between two features. Understandably, the higher the matching similarity, the more similar the two features are. Conversely, the lower the matching similarity, the lower the degree of similarity between the two features.

[0141] In one embodiment, the orchestration tool may use cosine similarity or Euclidean distance to determine the matching similarity between two features.

[0142] Among them, the dimension weight is a weight coefficient used to characterize the importance of each intent recognition dimension in the comprehensive calculation. The higher the dimension weight, the higher the importance of the corresponding intent recognition dimension in the intent matching degree calculation, and the greater the impact on the final intent recognition result. Conversely, the lower the dimension weight, the lower the importance of the corresponding intent recognition dimension in the intent matching degree calculation, and the smaller the impact on the final intent recognition result.

[0143] For example, for each candidate intent category of each candidate intent, the orchestration tool can perform similarity matching between the category features of the candidate intent category and the dimensional recognition features of the intent recognition dimension to which the candidate intent category belongs, thereby obtaining the matching similarity of the candidate intent category. After obtaining the matching similarity between all candidate intent categories and their corresponding dimensional recognition features in the candidate intent, the orchestration tool can perform weighted summation of each matching similarity according to the dimensional weight of each intent recognition dimension to obtain the intent matching degree of the corresponding candidate intent. Based on this method, the intent matching degree of each candidate intent can be obtained.

[0144] For example, the dimension weights for the business scenario dimension and the operational behavior dimension are β1 and β3 respectively, and β1 + β2 + β3 = 1. The candidate intents include categories such as "customer" (business scenario dimension), "recommendation" (operation behavior dimension), and "financial product" (operation object dimension). The matching similarity for "customer" is 0.9, for "recommendation" it is 0.85, and for "financial product" it is 0.82. Therefore, the intent matching score for this candidate intent can be determined as score = β1 × 0.9 + β2 × 0.85 + β3 × 0.82.

[0145] In one embodiment, the dimensional weight of each intent recognition dimension can be dynamically determined according to the business scenario. For example, during a marketing campaign, the weight of the "campaign" business category and the "participation" operation category can be automatically increased, enabling the system to quickly adapt to the shift in business focus.

[0146] In the above embodiments, by calculating the matching similarity between the candidate intent category and the identification features of the corresponding dimension, the local matching degree of each dimension can be quantified in a fine-grained manner. Then, the weighted summation according to the weight of each dimension is performed to obtain the overall intent matching degree of the candidate intent by combining the importance of different dimensions. This makes the matching result more in line with the user's true intent and provides an accurate quantitative basis for subsequent target intent screening.

[0147] In one embodiment, such as Figure 4 As shown, S210, based on the target intent and entity information, the workflow is orchestrated to obtain the target workflow for responding to natural language request information, including:

[0148] S402, if the target object has a historical intent, perform intent matching between the target intent and the historical intent to obtain the matching result.

[0149] Historical intent refers to the intent expressed by the target object in previous interactions. Intent matching between target intent and historical intent is the process of comparing the similarity between the two intents, which can be used to determine whether the historical intent and the target intent are consistent or different.

[0150] For example, an orchestration tool can determine whether a target object has a historical intent based on its object identifier. If so, it matches the target intent with the historical intent to obtain a matching result. Understandably, the matching result can be the intent similarity between the target intent and the historical intent.

[0151] In one embodiment, the orchestration tool can determine whether intent shift has occurred by calculating the semantic similarity between the target intent and historical intents. For example, it can perform intent matching by calculating the cosine similarity between the target intent and historical intents. The cosine similarity formula is as follows:

[0152]

[0153] Where 'a' represents the semantic feature vector of the target intent, and 'b' represents the semantic feature vector of the historical intent. Represents the vector dot product. The norm of a vector.

[0154] S404 If the matching result indicates that there is an intention offset between the target intention and the historical intention, then the intention offset information is determined based on the entity information.

[0155] Among them, intention deviation refers to the situation where the target intention is inconsistent with the historical intention. It can be that the target intention is completely inconsistent with the historical intention, or the target intention has changed in terms of direction, object, operation, etc., based on the historical intention.

[0156] Among them, the intent offset information is descriptive information used to describe the specific situation of intent offset, which can be used to explain the intent offset type, offset dimension and offset content, etc.

[0157] For example, when the orchestration tool determines that there is an intent offset between the target intent represented by the matching result and the historical intent, it can determine the intent offset information based on the entity information.

[0158] In one embodiment, the orchestration tool can compare the intent similarity between the target intent and the historical intent with a preset similarity threshold. If the intent similarity is lower than the preset similarity threshold, the matching result indicates that there is an intent offset between the target intent and the historical intent.

[0159] In one embodiment, the orchestration tool can compare entity information with historical entity information corresponding to historical intents to determine entity changes and, based on these changes, determine intent offset information. For example, historical entity information includes enterprise customers, risk assessment, and risk rating. Current entity information includes risk rating, portfolio, and adjustments. Therefore, the entity changes can be deduced as retaining the risk rating, adding a new portfolio, making adjustments, and removing enterprise customers and risk assessment from the key entities. This further determines the offset type in the intent offset information as a continuation offset, meaning the current intent does not completely negate the historical intent but rather further adjusts the portfolio based on the historical risk assessment results. The offset dimensions are the action dimension and the object dimension. The offset content includes adding a new portfolio and making adjustments to the two entities, retaining the risk rating entity, while enterprise customers and risk assessment are no longer key entities.

[0160] S406, Based on the intent offset information, the historical workflow of the historical intent is corrected to obtain the target workflow that responds to the natural language requirement information.

[0161] Among them, the historical workflow is a workflow obtained by orchestrating the historical intent of the target object during the historical interaction process.

[0162] Correcting historical workflows refers to the process of adjusting historical workflows according to their offsets. For example, based on the intended offset information, you can retain nodes and their results that have not been offset in the historical workflow, replace or delete nodes that have been offset, add nodes corresponding to the offset content, and adjust dependencies and parameters.

[0163] For example, the orchestration tool can modify the historical workflow of historical intents based on intent offset information to obtain the target workflow that responds to natural language demand information.

[0164] In the above embodiments, when historical intents exist, it can be determined whether an offset has occurred by matching the target intent with the historical intent. When the matching result indicates an offset, the intent offset information is determined based on entity information, thereby accurately locating the change. Accordingly, only the offset portion in the historical workflow is corrected, reusing the unoffset nodes and their results, reducing the workload that may be caused by a full re-orchestration, and improving workflow generation efficiency and the accuracy of responding to natural language requirements.

[0165] In one embodiment, such as Figure 5 As shown, S210, based on the target intent and entity information, the workflow is orchestrated to obtain the target workflow for responding to natural language request information, including:

[0166] S502 calls the large language model to parse the target intent and entity information, generating a preliminary workflow framework.

[0167] Among them, the large language model is a model trained on a large scale of text that can understand and generate natural language. It is used to parse target intent and entity information to generate a preliminary workflow framework.

[0168] Parsing the target intent and entity information refers to the process of extracting the steps and relationships required for the workflow based on the target intent and entity information. The preliminary workflow framework is a process framework that includes multiple work steps, their approximate order, and dependencies.

[0169] For example, when the target object lacks a historical intent, the orchestration tool can invoke a large language model to parse the target intent and entity information, generating a preliminary workflow framework. For instance, it can determine the main steps that need to be executed sequentially, such as "customer screening," "risk assessment," and "financial product recommendation," and clarify the logical order and dependencies between each step.

[0170] S504 breaks down each work step in the initial workflow framework into corresponding subtasks, resulting in multiple subtasks.

[0171] In this context, a work step is a processing stage within the initial workflow framework; that is, a business operation step that needs to be performed within the initial workflow framework. For example, "customer screening" is a work step, and "risk assessment" is also a work step.

[0172] In this context, the sub-tasks corresponding to the work steps refer to the specific execution tasks after the work steps are broken down, which can be handled by the task execution model. For example, "risk assessment" can be broken down into three sub-tasks: "collecting customer financial data", "calculating risk indicators", and "generating risk ratings".

[0173] For example, the large language model in the orchestration tool can decompose each work step in the initial workflow framework into corresponding subtasks, resulting in multiple subtasks.

[0174] S506. For each subtask, based on the task type of the subtask, find the preset mapping relationship and determine the task execution model that matches the task type.

[0175] The task type is a category identifier used to classify subtasks, distinguishing the business or processing categories to which different subtasks belong. For example, the task type for "calculate risk indicators" can be "risk calculation," and the task type for "recommend financial products" can be "product recommendation."

[0176] The preset mapping relationship is a pre-established correspondence table that represents the relationship between each task type and each task execution model. The task execution model is responsible for executing specific sub-tasks; for example, task execution models may include, but are not limited to, risk assessment mini-models, robo-advisory mini-models, and customer screening mini-models. Each task execution model pre-contains corresponding task execution logic, which can execute the sub-tasks of the corresponding task type. For example, the risk assessment mini-model can calculate risk perception based on customer financial data using algorithms such as Logistic Regression and Random Forest; the robo-advisory mini-model can combine risk ratings and market data to generate financial product recommendation strategies.

[0177] For example, for each subtask, the orchestration tool can find a preset mapping relationship based on the task type of the subtask and determine the task execution model that matches the task type.

[0178] S508 sends each subtask to its corresponding task execution model and receives task execution information from each task execution model.

[0179] Among them, task execution information refers to the result information returned by the task execution model after processing the sub-task, which may include processing results and status data. For example, the risk assessment mini-model can provide feedback such as "Customer A is low risk, Customer B is medium risk, Customer C is high risk".

[0180] For example, the large language model can send each subtask to the corresponding task execution model and receive task execution information from each task execution model.

[0181] In one embodiment, an interaction interface and communication protocol are pre-configured between the large language model and each task execution model to ensure accurate and efficient data and information transmission. After completing the sub-task processing, each task execution model can feed back the results to the large model.

[0182] In one embodiment, the unified data interface format between the large language model and each task execution model can adopt a data exchange format such as JSON or Protocol Buffers. The large language model transmits task parameters and input data to each task execution model, such as customer asset information and transaction records required by the risk assessment sub-model. The risk assessment sub-model returns processing results to the large language model, such as risk assessment scores and investment recommendations.

[0183] S510 refines the initial workflow framework based on the execution information of each task, and generates a target workflow that responds to natural language requirements.

[0184] Refining the initial workflow framework refers to supplementing and adjusting the initial workflow framework based on the execution information of each task, in order to determine the specific calling order, parameter configuration, and data flow.

[0185] For example, the large language model can refine the initial workflow framework based on the execution information of each task, and generate a target workflow that responds to the natural language requirement information.

[0186] The above embodiments utilize a large language model to parse target intent and entity information to generate a preliminary framework, leveraging its macro-planning capabilities to decompose work steps into sub-tasks and find preset mapping relationships based on task types. This ensures that each sub-task accurately matches the corresponding task execution model, enabling specialized tasks to be handled by specialized models. Furthermore, the framework is refined based on task execution information feedback from each model, allowing the final target workflow to incorporate specific execution results, thus improving the accuracy and executability of workflow generation.

[0187] In one embodiment, such as Figure 6 As shown, workflow orchestration methods may also include:

[0188] S602, obtain the execution result of the target workflow.

[0189] The execution result of the target workflow refers to the output data generated after the target workflow runs. It can include the processing results returned by each node after the target workflow executes each node in the orchestration order, as well as the final output.

[0190] For example, after the target workflow is completed, the orchestration tool can obtain the execution result of the target workflow.

[0191] S604 uses preset evaluation metrics to evaluate the target workflow based on the execution results, and obtains the evaluation results of the target workflow.

[0192] Among them, the preset evaluation indicators are pre-defined quantitative standards used to measure the effectiveness of the target workflow. For example, preset evaluation indicators may include risk assessment accuracy, customer acceptance of financial product recommendations, workflow execution time, etc.

[0193] The evaluation of the target workflow refers to the process of determining whether the target workflow meets the execution standards based on the execution results. The evaluation results are the data that characterize whether the target workflow meets the execution standards.

[0194] For example, the orchestration tool can use preset evaluation metrics to evaluate the target workflow based on the execution results, thereby obtaining the evaluation results of the target workflow. For instance, the execution results are compared with the preset evaluation metrics to determine whether the target workflow meets the execution criteria.

[0195] S606: If the evaluation results indicate that the target workflow does not meet the execution criteria, the target workflow is modified based on the evaluation results to obtain the modified target workflow.

[0196] The execution criteria are the standards for judging whether the target workflow is qualified, and they are usually composed of preset evaluation indicators and their thresholds. For example, the execution criteria may include a risk assessment accuracy rate of not less than 90%, or a customer acceptance rate of not less than 80% for financial product recommendations.

[0197] In one embodiment, the orchestration tool can determine the index value corresponding to the preset evaluation index based on the evaluation results, and then compare the index value with the index threshold corresponding to the preset evaluation index to determine whether the target workflow meets the execution standard. When there are multiple preset evaluation indices, the target workflow can only be determined to meet the execution standard if the evaluation result of each preset evaluation index indicates that the target workflow meets the execution standard; otherwise, the target workflow is determined not to meet the execution standard.

[0198] The revised target workflow is a new workflow obtained by modifying the original target workflow. The orchestration tool can modify the nodes, sequence, parameters, or task assignments in the target workflow based on the non-compliance points identified in the evaluation results, thus obtaining the revised workflow. For example, if the risk assessment accuracy is low, the parameters of the risk assessment mini-model are adjusted. If the recommendation acceptance is low, the strategy of the robo-advisor mini-model is adjusted.

[0199] For example, the orchestration tool can be used to modify the target workflow based on the evaluation results when the evaluation results indicate that the target workflow does not meet the execution criteria, thus obtaining a modified target workflow.

[0200] S608, based on the modified target workflow, returns to the steps for obtaining the execution results of the target workflow until the target workflow meets the execution criteria.

[0201] For example, the orchestration tool can, based on the modified target workflow, return to the steps of obtaining the execution results of the target workflow until the target workflow meets the execution criteria.

[0202] In the above embodiments, by obtaining the execution results and quantifying them with preset evaluation indicators, it is possible to objectively determine whether the target workflow meets the standards. When the execution standards are not met, targeted corrections are made based on the evaluation results, and execution is returned to obtain the execution results again, forming a closed-loop iteration of "execution-evaluation-correction-re-execution", so that the workflow gradually approaches and ultimately meets the execution standards, thereby improving its accuracy and reliability.

[0203] In one embodiment, such as Figure 7 As shown, workflow orchestration methods may also include:

[0204] S702 normalizes each key entity to obtain its standard name in the financial field.

[0205] Normalization is a process of standardizing key entities, eliminating differences such as synonyms, colloquial expressions, and abbreviations, and making them correspond to standard names in the financial field. Standard names refer to the standardized expressions of key entities in the financial sector. For example, if the key entity is "corporate customer," it can be normalized to "enterprise customer."

[0206] For example, programming tools can normalize the key entities to obtain standard names for each key entity in the financial field.

[0207] S704: For each key entity, search for the corresponding graph node in the financial domain knowledge graph based on the standard name of the key entity.

[0208] In the financial domain, a knowledge graph is a structured identifier of financial business relationships. It consists of nodes and edges, where nodes represent financial business objects and edges identify the relationships between them. A node corresponding to a key entity represents a business object within the financial knowledge graph, such as a corporate customer node or a risk assessment node. Understandably, a financial knowledge graph can be pre-designed by engineers based on at least one data source, such as financial business documents, product catalogs, customer information, business process descriptions, business rules, risk indicator definitions, and regulatory rules, through entity identification, relationship extraction, entity linking, and knowledge fusion. Because a financial knowledge graph integrates professional financial knowledge, entity relationships, and business rules, and associates key entities with nodes in the graph, it effectively solves the problem of semantic ambiguity in complex financial scenarios using traditional methods.

[0209] For example, the programming tool can look up the corresponding graph node in the financial domain knowledge graph for each key entity based on the standard name of the key entity.

[0210] S706 traverses the financial knowledge graph centered on the graph nodes to obtain graph association information of key entities.

[0211] Among them, graph association information refers to the information of neighboring nodes and relational edges related to a knowledge graph determined after traversing the knowledge graph with the graph node as the center.

[0212] For example, programming tools can traverse the financial knowledge graph centered on graph nodes to obtain the graph association information of each key entity.

[0213] In one embodiment, the programming tool can use a graph neural network algorithm to perform information propagation and feature learning on a financial knowledge graph centered on graph nodes, thereby obtaining graph association information.

[0214] S708, based on the correlation information of each map, supplement the fusion information to obtain the updated fusion information.

[0215] For example, the programming tool can supplement the fusion information based on the association information of each graph to obtain the updated fusion information.

[0216] In one embodiment, the programming tool can vectorize each graph association information to obtain a graph association vector, and then concatenate or weightedly fuse the graph association vector with the fusion vector to supplement information and obtain updated fusion information, i.e., updated fusion vector.

[0217] S710 uses the updated fusion information to perform a step of extracting features from the fusion information according to multiple preset intent recognition dimensions.

[0218] For example, when a programming tool performs intent recognition according to multiple intent recognition dimensions, it can use the updated fusion information to perform the corresponding intent recognition steps, that is, return to the step of extracting features from the fusion information according to the preset multiple intent recognition dimensions.

[0219] In the above embodiments, normalizing key entities can eliminate differences in synonymous and colloquial expressions, providing an accurate entry point for graph lookup. Furthermore, searching for graph nodes by standard names and traversing them to obtain graph association information can supplement unspoken financial business relationships, thereby updating the fusion information and providing a more complete business semantic basis for subsequent multi-dimensional feature extraction, thus improving the accuracy of feature expression and intent recognition.

[0220] In one embodiment, such as Figure 8 As shown, a workflow orchestration method is provided, which may specifically include the following steps:

[0221] S801: Obtain the natural language requirement information input by the user and the user's historical interaction information, and then fuse them to obtain the initial fused information.

[0222] S802, extract entity information from natural language demand information. Entity information includes multiple key entities and the relationships between them.

[0223] For example, users can input their financial service needs in natural language through the user interaction page, such as "conduct risk assessments for recently registered corporate clients, and then recommend suitable wealth management products based on the assessment results." The large language model in the orchestration tool uses Named Entity Recognition (NER) technology to extract key entities from the user's needs, such as "recently registered corporate clients," "risk assessment," and "wealth management products." Semantic role labeling clarifies the relationships between these entities; for example, "client" is the object of "risk assessment," and "risk assessment results" are the basis for "wealth management product recommendations." Specific entity steps and processes are clearly defined.

[0224] In one embodiment, the large language model is a model tool obtained by fine-tuning a general large language model. Designers can collect a large amount of natural language text data in the financial field, including business requirement documents, customer consultation records, business process descriptions, etc., and fine-tune the general large language model. During the fine-tuning process, the intent categories, key entities, and semantic relationships in the text are labeled, enabling the model to learn financial terminology and business logic, thereby improving the accuracy of intent understanding.

[0225] Manually annotate the text with intent categories (e.g., risk assessment intent, robo-advisor intent), key entities (customer names, financial product names), and semantic relationships (e.g., the object-operation relationship between "customer" and "risk assessment"). Convert the annotated data into feature vectors. For text data, bag-of-words and TF-IDF models can be used for vectorization.

[0226] S803 links each key entity to a knowledge graph in the financial field, obtaining the graph association information for each key entity.

[0227] S804, use the map association information to supplement the initial fusion information to obtain the fusion information.

[0228] S805 extracts features from the fused information according to multiple preset intent recognition dimensions to obtain the corresponding dimension recognition features of the user under each intent recognition dimension.

[0229] For example, designers can design multiple small model subtasks according to financial scenarios, using multiple intent dimensions. For instance, they can categorize them into four core categories based on business scenarios: customers, products, activities, and knowledge; further subdivide them into five types based on user actions: question and answer, query, processing, modification, and recommendation; and clearly define the specific objects involved in the operation, such as accounts, contracts, and reports. In practical use, semantic features can be extracted from the fused information based on the pre-designed hierarchical relationship of the multi-dimensional semantic feature library, obtaining the corresponding dimensional recognition features for each dimension, thus forming a multi-dimensional feature representation.

[0230] S806 determines the intent matching degree of each candidate intent based on the identification features of each dimension and the multiple candidate intent categories contained in each of the pre-constructed multiple sets of candidate intents.

[0231] S807, determine whether there is a selected intent with an intent matching degree greater than or equal to the preset matching degree threshold. If not, execute S808; if yes, execute S809.

[0232] S808: Obtain the user's intent supplementary information, supplement the fusion information based on the intent supplementary information, obtain the updated fusion information, and return to execute S805.

[0233] S809, determine whether there is only a single selected intent. If yes, execute S810; otherwise, execute S811.

[0234] S810, the selected intent is determined as the user's target intent, and S812 is executed.

[0235] S811, determine the selected intent with the highest intent matching degree among all selected intents as the user's target intent, and execute S812.

[0236] S812, determine if the user has a history of intent. If not, execute S813; if so, execute S818.

[0237] S813 calls the large language model to parse the target intent and entity information, generating a preliminary workflow framework.

[0238] S814 breaks down each work step in the initial workflow framework into corresponding subtasks, resulting in multiple subtasks.

[0239] S815: For each subtask, based on the task type of the subtask, find the preset mapping relationship and determine the task execution model that matches the task type.

[0240] S816 sends each subtask to its corresponding task execution model and receives task execution information from each task execution model.

[0241] S817, refine the initial workflow framework based on the execution information of each task, and generate the target workflow that responds to the natural language requirement information. Execute S822.

[0242] S818, perform intent matching between the target intent and historical intents to obtain the matching result.

[0243] S819, determine if there is an intention offset. If not, execute S820; if so, execute S821.

[0244] S820: Determine the historical workflow corresponding to the historical intent as the target workflow for responding to user needs. Execute S822.

[0245] S821, based on the intent offset information determined by the entity information, correct the historical workflow to obtain the target workflow that responds to the user's needs.

[0246] For example, if an intent shift is detected, the large and small models can be invoked again to orchestrate the workflow based on the new intent and context information. The large language model generates a new orchestration logic framework based on the updated intent, while the small model adjusts specific modules, such as reselecting risk assessment indicators or adjusting the robo-advisor strategy, to generate and execute a new workflow.

[0247] In one embodiment, during the execution of the target workflow, the orchestration tool's execution layer also monitors the workflow's running status in real time, collecting execution data and feedback information from each module. If an anomaly occurs, such as a small model timeout or data format mismatch, the orchestration tool will immediately trigger an exception handling mechanism. Based on the anomaly type, the workflow is automatically adjusted, such as re-calling modules, skipping the problematic module, or inserting a manual intervention node, and the anomaly information is fed back to the large language model for subsequent optimization.

[0248] S822, obtain the execution result of the target workflow, and evaluate the target workflow based on the execution result using preset evaluation indicators.

[0249] S823: Determine if the target workflow meets the execution criteria. If it does, terminate. If it does not, execute S824.

[0250] S824, Based on the evaluation results, revise the target workflow to obtain the revised target workflow. Return to execute S822.

[0251] For example, after a specific task is processed, the result is returned to the large language model. If the evaluation result is unsatisfactory, the large language model uses reinforcement learning algorithms to automatically adjust the calling order and parameter configuration of modules in the workflow based on the evaluation feedback, or reassigns the task to a more suitable small model. At the same time, the execution data and optimization information are fed back to the basic service layer for further training and optimization of the large and small models, improving the accuracy and efficiency of subsequent workflow orchestration and execution, forming a complete closed loop of "intent understanding - workflow orchestration - execution feedback - model optimization" until a satisfactory workflow orchestration result is obtained.

[0252] In one embodiment, the large language model can calculate a composite reward based on the evaluation results, then update the planning strategy in the large language model through the PPO algorithm, calculate the weight adjustment coefficient based on the composite reward, update the dimension weights corresponding to each intent recognition dimension based on the weight adjustment coefficient, and then perform intent recognition based on the updated dimension weights to form an optimization closed loop.

[0253] The aforementioned workflow orchestration method addresses the pain points of traditional financial business processes by integrating technologies such as context awareness, knowledge graphs, multi-dimensional intent design, and reinforcement learning. Compared to methods that simply rely on large-scale text understanding models, it has the following significant advantages:

[0254] First, context-awareness and knowledge graph-driven approaches significantly improve feature representation accuracy. Traditional large-scale model text understanding relies solely on semantic analysis of the input text, lacking in-depth mining of textual background and professional knowledge. This makes it susceptible to issues such as polysemy and ambiguous referents, leading to biased intent recognition. This solution introduces context-awareness and knowledge graph technologies to construct a more accurate feature representation system. Context-awareness technology forms a dynamic context model by recording user history and business process states. When a user inputs a new command, the system combines historical operations with the current business scenario to understand the semantics. For example, in continuous financial transactions, it automatically associates previous transaction information, avoiding repeated queries and misjudgments. The knowledge graph integrates professional knowledge, entity relationships, and business rules in the financial field, associating entities in the text with nodes in the graph, effectively solving the problem of ambiguous semantic understanding in complex financial scenarios using traditional methods.

[0255] Secondly, multi-dimensional intent design and dynamic parameter optimization enhance the flexibility and adaptability of intent recognition. Simple large-scale text understanding models typically calculate the matching score between text and intent category from only a single dimension, lacking consideration for the complexity of user needs, and with fixed parameters, it is difficult to adapt to changes in different business scenarios. This solution innovatively adopts multi-dimensional intent design, classifying and matching intents from multiple dimensions such as business type, operation type, and object type, enabling a more detailed capture of various aspects of user needs. Specifically, this solution distinguishes categories such as customers, products, and activities in the business type dimension; covers operations such as question-and-answer, query, and processing in the operation type dimension; and clearly defines specific objects such as accounts and contracts in the object type dimension. This multi-dimensional cross-analysis makes intent recognition more accurate. Simultaneously, dynamic learnable parameters are introduced, and reinforcement learning algorithms automatically adjust the weights of each dimension based on historical matching accuracy and business feedback. For example, during marketing campaigns, the weights of the "activity" business type and the "participation" operation type are automatically increased, allowing the system to quickly adapt to shifts in business priorities.

[0256] Third, reinforcement learning is used for closed-loop optimization to achieve continuous intelligent evolution of workflow planning. Traditional solutions often lack feedback optimization mechanisms in workflow planning. Once the model is trained, the workflow strategy is difficult to adjust based on actual execution results, failing to adapt to the dynamic changes in financial business. This solution introduces reinforcement learning into the entire workflow planning process, constructing a complete closed loop of "semantic matching - process planning - execution feedback - model optimization". After each workflow execution, the system collects feedback data on execution efficiency, task completion quality, user satisfaction, and other aspects. A carefully designed composite reward function is used to evaluate the workflow planning strategy and semantic matching weights. If the execution result does not meet expectations, the reinforcement learning algorithm automatically adjusts the workflow planning strategy of the large model, such as optimizing the module call order, adjusting parameter configurations, and updating the weights of multi-dimensional intent matching, enabling the system to learn and improve from failures.

[0257] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0258] Based on the same inventive concept, this application also provides a workflow orchestration apparatus for implementing the workflow orchestration method described above. The solution provided by this apparatus is similar to the implementation described in the above method; therefore, the specific limitations in one or more workflow orchestration apparatus embodiments provided below can be found in the limitations of the workflow orchestration method described above, and will not be repeated here.

[0259] In one embodiment, such as Figure 9 As shown, a workflow orchestration device 900 is provided, including: an information acquisition module 901, an information extraction module 902, a feature extraction module 903, a target intent determination module 904, and a workflow orchestration module 905, wherein:

[0260] The information acquisition module 901 is used to acquire the fusion information corresponding to the target object; the fusion information is obtained by fusing the natural language requirement information input by the target object and the historical interaction information of the target object.

[0261] The information extraction module 902 is used to extract entity information from natural language demand information. The entity information includes multiple key entities and the relationships between them.

[0262] The feature extraction module 903 is used to extract features from the fused information according to multiple preset intent recognition dimensions, so as to obtain the dimension recognition features of the target object under each intent recognition dimension.

[0263] The target intent determination module 904 is used to select target intents that meet the intent recognition conditions from a pre-built set of candidate intents based on the recognition features of each dimension; the candidate intents are obtained by combining the candidate intent categories corresponding to each intent recognition dimension.

[0264] Workflow orchestration module 905 is used to orchestrate workflows based on target intent and entity information to obtain target workflows for responding to natural language requirements.

[0265] In one embodiment, the target intent determination module 904 is used to: determine the intent matching degree of each candidate intent based on the identification features of each dimension and the multiple candidate intent categories contained in each of the pre-constructed multiple sets of candidate intents; and in the case where there are multiple selected intents with intent matching degrees greater than or equal to a preset matching degree threshold, determine the selected intent with the highest intent matching degree among the selected intents as the target intent that meets the intent recognition conditions.

[0266] In one embodiment, the workflow orchestration apparatus 900 may further include:

[0267] The matching threshold comparison module is used to generate an intent supplement prompt page when the matching degree of each intent is less than the preset matching degree threshold; the intent supplement prompt page contains intent supplement prompt information determined according to the matching degree of each intent.

[0268] The operation response module is used to respond to input operations on the intent supplement prompt page and obtain intent supplement information of the target object;

[0269] The information supplementation module is used to supplement the fused information based on the intent supplementation information to obtain the updated fused information;

[0270] The loop execution module is used to return the updated fusion information and perform the steps of feature extraction on the fusion information according to multiple preset intent recognition dimensions, until there is a selected intent with an intent matching degree greater than or equal to a preset matching degree threshold.

[0271] In one embodiment, the target intent determination module 904 is configured to: for each candidate intent category of each candidate intent, perform similarity matching between the category features of the candidate intent category and the dimension recognition features of the intent recognition dimension to which the candidate intent category belongs, to obtain the matching similarity of the candidate intent category; and perform weighted summation of the matching similarities of each candidate intent category according to the respective dimension weights of each intent recognition dimension, to obtain the intent matching degree of the candidate intent.

[0272] In one embodiment, the workflow orchestration module 905 is used to: match the target intent with the historical intent when the target object has a historical intent, and obtain a matching result; if the matching result indicates that there is an intent offset between the target intent and the historical intent, determine the intent offset information based on the entity information; and correct the historical workflow of the historical intent based on the intent offset information to obtain the target workflow that responds to the natural language requirement information.

[0273] In one embodiment, the workflow orchestration module 905 is used to: invoke a large language model to parse the target intent and entity information to generate a preliminary workflow framework; decompose each work step in the preliminary workflow framework into corresponding subtasks to obtain multiple subtasks; for each subtask, according to the task type of the subtask, search for a preset mapping relationship to determine the task execution model that matches the task type; the preset mapping relationship is used to characterize the correspondence between each task type and each task execution model; send each subtask to the corresponding task execution model and receive the task execution information fed back by each task execution model; refine the preliminary workflow framework according to the task execution information to generate a target workflow that responds to natural language requirement information.

[0274] In one embodiment, the target intent determination module 904 is used for:

[0275] The execution result feedback module is used to obtain the execution results of the target workflow.

[0276] The results evaluation module is used to evaluate the target workflow based on the execution results using preset evaluation indicators, and obtain the evaluation results of the target workflow.

[0277] The workflow correction module is used to correct the target workflow based on the evaluation results when the evaluation results indicate that the target workflow does not meet the execution criteria, thus obtaining a corrected workflow.

[0278] In one embodiment, the target intent determination module 904 is used for:

[0279] The normalization module is used to normalize each key entity to obtain the standard name of each key entity in the financial field.

[0280] The graph search module is used to find the corresponding graph node in the financial knowledge graph for each key entity based on the standard name of the key entity.

[0281] The graph association module is used to traverse the financial knowledge graph with graph nodes as the center to obtain the graph association information of key entities.

[0282] The graph association and supplementation module is used to supplement the fusion information based on the association information of each graph, so as to obtain the updated fusion information.

[0283] The feature extraction execution module is used to perform feature extraction on the fused information according to multiple preset intent recognition dimensions using the updated fused information.

[0284] Each module in the aforementioned workflow orchestration device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0285] In one embodiment, a computer device is provided, which may be a server integrated with a business system, and its internal structure diagram may be as follows: Figure 10 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data related to workflow orchestration methods. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a workflow orchestration method.

[0286] Those skilled in the art will understand that Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0287] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the specific steps of the method embodiment described above.

[0288] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the specific steps of the above-described method embodiment.

[0289] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the specific steps of the method embodiment described above.

[0290] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the acquisition, storage, processing, and transmission of the data all comply with relevant laws and regulations.

[0291] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0292] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0293] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A workflow orchestration method, characterized in that, The method includes: Obtain the fusion information corresponding to the target object; the fusion information is obtained by fusing the natural language requirement information input to the target object and the historical interaction information of the target object; Entity information is extracted from the natural language demand information, and the entity information includes multiple key entities and the relationships between the key entities. According to multiple preset intent recognition dimensions, feature extraction is performed on the fused information to obtain the dimension recognition features of the target object under each intent recognition dimension. Based on the identification features of each dimension, target intents that meet the intent identification conditions are selected from a pre-constructed set of candidate intents; the candidate intents are obtained by combining the candidate intent categories corresponding to each of the intent identification dimensions. Based on the target intent and the entity information, workflow orchestration is performed to obtain a target workflow for responding to the natural language requirement information.

2. The method according to claim 1, characterized in that, The step of identifying target intents that meet the intent recognition conditions from a pre-constructed set of candidate intents based on the features identified in each dimension includes: Based on the identification features of each dimension and the multiple candidate intent categories contained in each of the pre-constructed multiple sets of candidate intents, the intent matching degree of each candidate intent is determined; In the case where there are multiple selected intentions with an intent matching degree greater than or equal to a preset matching degree threshold, the selected intention with the highest intent matching degree among the selected intentions is determined as the target intent that meets the intent recognition conditions.

3. The method according to claim 2, characterized in that, The method further includes: If the matching degree of each intent is less than the preset matching degree threshold, an intent supplement prompt page is generated; the intent supplement prompt page contains intent supplement prompt information determined according to the matching degree of each intent. In response to an input operation on the intent supplement prompt page, obtain the intent supplement information of the target object; The fused information is supplemented based on the intent supplement information to obtain updated fused information; The updated fusion information is used to return and execute the step of extracting features from the fusion information according to multiple preset intent recognition dimensions, until there is a selected intent with an intent matching degree greater than or equal to the preset matching degree threshold.

4. The method according to claim 2 or 3, characterized in that, The step of determining the intent matching degree of each candidate intent based on the features identified in each dimension and the multiple candidate intent categories contained in each of the pre-constructed multiple sets of candidate intents includes: For each candidate intent category of each candidate intent, the category feature of the candidate intent category is matched with the dimension recognition feature of the intent recognition dimension to which the candidate intent category belongs, to obtain the matching similarity of the candidate intent category. Based on the respective dimensional weights of each intent recognition dimension, the matching similarity of each candidate intent category is weighted and summed to obtain the intent matching degree of the candidate intent.

5. The method according to any one of claims 1 to 3, characterized in that, The process of orchestrating workflows based on the target intent and the entity information to obtain a target workflow for responding to the natural language requirement information includes: If the target object has a historical intent, the target intent is matched with the historical intent to obtain a matching result; If the matching result indicates that there is an intent offset between the target intent and the historical intent, then the intent offset information is determined based on the entity information; Based on the intent offset information, the historical workflow of the historical intent is corrected to obtain the target workflow that responds to the natural language requirement information.

6. The method according to any one of claims 1 to 3, characterized in that, The process of orchestrating workflows based on the target intent and the entity information to obtain a target workflow for responding to the natural language requirement information includes: The large language model is invoked to parse the target intent and the entity information to generate a preliminary workflow framework; Each work step in the preliminary workflow framework is decomposed into corresponding sub-tasks, resulting in multiple sub-tasks; For each subtask, a preset mapping relationship is searched according to the task type of the subtask to determine the task execution model that matches the task type; the preset mapping relationship is used to characterize the correspondence between each task type and each task execution model; Each subtask is sent to its corresponding task execution model, and task execution information is received from each task execution model. The initial workflow framework is refined based on the task execution information to generate a target workflow that responds to the natural language requirement information.

7. The method according to any one of claims 1 to 3, characterized in that, The method further includes: Obtain the execution result of the target workflow; The target workflow is evaluated based on the execution results using preset evaluation metrics to obtain the evaluation result of the target workflow; If the evaluation result indicates that the target workflow does not meet the execution criteria, the target workflow is modified according to the evaluation result to obtain the modified target workflow; Based on the revised target workflow, return to the step of obtaining the execution result of the target workflow until the target workflow meets the execution criteria.

8. The method according to any one of claims 1 to 3, characterized in that, The method further includes: The key entities are normalized to obtain their standard names in the financial field. For each of the aforementioned key entities, the corresponding graph node is searched in the financial domain knowledge graph based on the standard name of the key entity. By traversing the financial knowledge graph with the graph node as the center, the graph association information of the key entity is obtained; Based on the association information of each of the aforementioned maps, the fusion information is supplemented to obtain the updated fusion information; Using the updated fusion information, perform the step of extracting features from the fusion information according to multiple preset intent recognition dimensions.

9. A workflow orchestration device, characterized in that, The device includes: The information acquisition module is used to acquire fusion information corresponding to the target object; the fusion information is obtained by fusing natural language requirement information input to the target object and the target object's historical interaction information; An information extraction module is used to extract entity information from the natural language demand information. The entity information includes multiple key entities and the relationships between the key entities. The feature extraction module is used to extract features from the fused information according to multiple preset intent recognition dimensions, so as to obtain the dimension recognition features of the target object under each intent recognition dimension. The target intent determination module is used to identify features based on each of the aforementioned dimensions and filter out target intents that meet the intent recognition conditions from a pre-constructed set of candidate intents; the candidate intents are obtained by combining the candidate intent categories corresponding to each of the aforementioned intent recognition dimensions. The workflow orchestration module is used to orchestrate workflows based on the target intent and the entity information to obtain a target workflow for responding to the natural language requirement information.

10. The apparatus according to claim 9, characterized in that, The target intent determination module is used to: identify the features of each dimension and determine the intent matching degree of each of the pre-constructed multiple candidate intents, each containing multiple candidate intent categories; and when there are multiple selected intents with intent matching degrees greater than or equal to a preset matching degree threshold, determine the selected intent with the highest intent matching degree among the selected intents as the target intent that meets the intent recognition conditions.

11. The apparatus according to claim 10, characterized in that, The device further includes: The matching threshold comparison module is used to generate an intent supplement prompt page when the matching degree of each intent is less than the preset matching degree threshold; the intent supplement prompt page includes intent supplement prompt information determined according to the matching degree of each intent. An operation response module is used to respond to an input operation on the intent supplement prompt page and obtain intent supplement information of the target object; An information supplementation module is used to supplement the fused information based on the intent supplementation information to obtain updated fused information; The loop execution module is used to return to the execution step of extracting features from the fusion information according to multiple preset intent recognition dimensions, using the updated fusion information, until there is a selected intent with an intent matching degree greater than or equal to the preset matching degree threshold.

12. The apparatus according to claim 10 or 11, characterized in that, The target intent determination module is used to: for each candidate intent category of each candidate intent, perform similarity matching between the category features of the candidate intent category and the dimensional recognition features of the intent recognition dimension to which the candidate intent category belongs, to obtain the matching similarity of the candidate intent category; and perform weighted summation of the matching similarities of each candidate intent category according to the dimensional weight of each intent recognition dimension, to obtain the intent matching degree of the candidate intent.

13. The apparatus according to any one of claims 9 to 11, characterized in that, The workflow orchestration module is used to: match the target intent with the historical intent when the target object has historical intent, and obtain a matching result; if the matching result indicates that there is an intent offset between the target intent and the historical intent, then determine the intent offset information based on the entity information; Based on the intent offset information, the historical workflow of the historical intent is corrected to obtain the target workflow that responds to the natural language requirement information.

14. The apparatus according to any one of claims 9 to 11, characterized in that, The workflow orchestration module is used to: call a large language model to parse the target intent and the entity information, and generate a preliminary workflow framework; Each work step in the preliminary workflow framework is decomposed into corresponding sub-tasks, resulting in multiple sub-tasks; For each subtask, a preset mapping relationship is searched according to the task type of the subtask to determine the task execution model that matches the task type; the preset mapping relationship is used to characterize the correspondence between each task type and each task execution model; each subtask is sent to the corresponding task execution model, and task execution information is received from each task execution model; the preliminary workflow framework is refined according to the task execution information to generate a target workflow that responds to the natural language requirement information.

15. The apparatus according to any one of claims 9 to 11, characterized in that, The device further includes: The execution result feedback module is used to obtain the execution result of the target workflow; The result evaluation module is used to evaluate the target workflow based on the execution result using preset evaluation indicators, and obtain the evaluation result of the target workflow; The workflow correction module is used to correct the target workflow according to the evaluation results when the evaluation results indicate that the target workflow does not meet the execution criteria, thereby obtaining a corrected workflow.

16. The apparatus according to any one of claims 9 to 11, characterized in that, The device further includes: The normalization processing module is used to normalize each of the key entities to obtain the standard names of each key entity in the financial field; The graph lookup module is used to find the corresponding graph node in the financial domain knowledge graph for each of the key entities, based on the standard name of the key entity. The graph association module is used to traverse the financial domain knowledge graph with the graph node as the center to obtain the graph association information of the key entity; The graph association and supplementation module is used to supplement the fusion information according to the graph association information to obtain the updated fusion information; The feature extraction execution module is used to perform a step of extracting features from the fused information according to multiple preset intent recognition dimensions using the updated fused information.

17. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 9.

18. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 9.

19. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 9.