Enterprise process automation agent generation method and system

Abstract

The invention relates to an agent generation method and system for enterprise process automation. The method comprises the following steps: S1, obtaining a standard specification document of an enterprise; s2, knowledge extraction and knowledge graph construction; s3, semantic analysis and intention discrimination; and S4, process reasoning and automatic execution. According to the method, full-link intellectualization from enterprise system document automatic analysis to business process closed-loop execution is realized, context sensing and dynamic adaptation capabilities are realized, multiple types of tasks are supported, the analysis fine granularity and cross-system cooperation efficiency of business process rules are effectively improved, and high universality and expandability are realized.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of artificial intelligence, natural language processing, and knowledge engineering, and specifically to a method and system for generating intelligent agents for enterprise process automation. Background Technology

[0002] There are significant technical bottlenecks in the field of enterprise process management.

[0003] Existing technologies based on large language models and knowledge graphs mainly focus on static knowledge processing tasks. Although they reduce the cost of manual annotation, they are only suitable for building enterprise domain knowledge bases.

[0004] These technologies generally suffer from the following key flaws: First, process operation is lacking, limited to "question and answer" or "knowledge extraction". Existing systems mostly remain at the knowledge question and answer level and have not been implemented at the enterprise "process automation operation" level. They lack the end-to-end intelligent agent capability to link knowledge graphs with actual business processes (such as leave requests, expense reimbursements, and approvals), and cannot achieve closed-loop operation from knowledge understanding to process execution.

[0005] Secondly, prompt word engineering heavily relies on manual design and lacks automation. Existing technologies, based on manually written or optimized prompt words, have weak generalization ability across different scenarios. Some technologies use genetic algorithms to generate prompt words, but lack context awareness and dynamic adaptation to enterprise scenarios.

[0006] Furthermore, knowledge graphs are coarse-grained and cannot support fine-grained process node judgments. Existing knowledge graph construction methods are too coarse to support fine-grained business process judgments, such as failing to accurately express specific business rules like "sick leave requires a medical certificate." This is because existing patents mostly focus on the extraction and graph construction of "entity-relationships," but in enterprise business processes, such as leave types and approval levels, more fine-grained business semantic representations and relationship modeling are required.

[0007] More significantly, existing technologies lack a closed-loop capability from understanding to direct execution. Current technologies primarily focus on the understanding level, such as question answering / extraction, lacking a closed-loop system design that moves from knowledge perception → process judgment → automatic execution. For example, if a user asks, "I need to take 3 days off," the system can only answer with the rule, but cannot complete the specific processes such as filling out the form, submitting, or sending reminders. Furthermore, current methods are mostly industry-specific (such as the tobacco industry), lacking universality. Knowledge graphs need to be re-annotated and reconstructed, and there is a lack of ability to automatically parse standard documents and transfer knowledge across industries.

[0008] In summary, existing technical architectures generally lack cross-system collaboration capabilities. When faced with multi-source, heterogeneous enterprise systems, they cannot automatically parse process rules in policy documents, nor can they translate semantic understanding into actual form filling, approval triggering, and other operations. This technological limitation results in low levels of enterprise process automation, requiring employees to manually operate multiple systems, which impacts efficiency and is prone to errors. Especially in scenarios with frequent policy updates and complex processes, traditional technologies struggle to synchronize changes to business rules in a timely manner, leading to execution deviations.

[0009] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0010] In view of the above-mentioned deficiencies of the prior art, the first aspect of the present invention provides a method for generating intelligent agents for enterprise process automation, comprising the following steps: Step S1: Obtain the company's standard specification documents; wherein, standard specification documents related to the process are collected from various internal data sources of the company, and the documents are subjected to format recognition, content cleaning and version verification to form a parsable standardized text dataset; the internal data sources of the company include at least one of the following: rules and regulations, operation manuals, and OA system documents; Step S2: Knowledge extraction and knowledge graph construction; wherein, entity recognition, relation extraction, attribute extraction and relation mining are performed on the standardized text dataset to identify and construct an enterprise knowledge graph containing core semantic elements; the core semantic elements include at least one of the following: process nodes, conditional rules, and role permissions; Step S3: Semantic analysis and intent determination; wherein, receiving a natural language request input by the user, dynamically generating a contextual prompt word for the process task based on the contextual information of the knowledge graph, converting the natural language request into a structured semantic parsing result, and matching the corresponding process node and execution rule in the knowledge graph based on the semantic parsing result; the semantic parsing result includes at least one of the following: the identified user intent, process parameters, and constraints. Step S4: Process reasoning and automated execution; wherein, process reasoning is performed based on the matched process nodes and execution rules to determine the process path, approval authority and business constraints; and based on the reasoning results, process tasks are automatically generated and executed in the target business system, wherein the process tasks include at least one of the following: automatically filling in forms, initiating approval tasks and automatically jumping and submitting between process nodes.

[0011] In the intelligent agent generation method for enterprise process automation as described above, step S2 may optionally include the following steps: Entity recognition is performed using a finely tuned large language model BERT, and the entity includes at least one of the following: leave request rules and approval levels. Generative large language models are used for relation extraction and attribute extraction; Based on the extracted entities and relationships, a graph structure is constructed that includes the process body, rule entities, and execution objects; The constructed knowledge graph is stored in a graph database.

[0012] In the intelligent agent generation method for enterprise process automation described above, optionally, the statements in the enterprise standard specification document are subjected to contextual understanding and semantic reconstruction to further generate logical triples containing causal relationships, triggering conditions and result actions, and the logical triples are embedded in a knowledge graph to describe the causal dependencies between events in the business process.

[0013] In the intelligent agent generation method for enterprise process automation as described above, optionally, during the relationship extraction process, a dynamic prompt word template is used to guide the large language model to extract the relationship, and the prompt word generation strategy is continuously optimized based on a reinforcement learning mechanism.

[0014] In the intelligent agent generation method for enterprise process automation as described above, optionally, in step S3, the knowledge graph includes multiple node types, and the node type includes at least one of them: process subject, rule entity, and execution object; The semantic relationships between nodes in the knowledge graph are defined using OWL semantic modeling, and the semantic relationships include at least one of the following: dependency relationship and trigger relationship.

[0015] In the intelligent agent generation method for enterprise process automation as described above, optionally, the natural language request input by the user can be text input, conversational voice input, form input, or any combination thereof forming a multimodal hybrid form.

[0016] In the intelligent agent generation method for enterprise process automation as described above, optionally, in step S3, the intent discrimination classifies the semantic content of the user input using a hybrid model of BiLSTM and CRF to identify process type and operation category, and supports multi-turn dialogue interaction to complete missing parameter information.

[0017] To achieve the above objectives, a second aspect of the present invention provides an intelligent agent generation system for enterprise process automation, wherein the intelligent agent generation method for enterprise process automation as described in any of the embodiments of the first aspect above includes: The standard document acquisition module is used to collect standard and specification documents related to processes from internal enterprise data sources, and to perform format recognition, content cleaning and version verification on the documents to form a standardized text dataset. The knowledge extraction and graph construction module is connected to the standard document acquisition module and is used to call a large language model to perform entity recognition, relation extraction and attribute extraction on the standardized text dataset in order to construct an enterprise knowledge graph containing core semantic elements. A standard graph storage module, connected to the knowledge extraction and graph construction module, is used to store the enterprise knowledge graph in the form of a graph database. The user request recognition module is connected to the standard graph storage module and is used to recognize the user intent in the natural language request input by the user and match the corresponding process node. The dynamic prompt word optimization module interacts with the knowledge extraction and graph construction module and the standard graph storage module respectively, and is used to adjust and optimize the prompt words based on the task context and feedback information; The intelligent process execution module is connected to the user request recognition module, the dynamic prompt word optimization module, and the standard graph storage module, respectively. It is used to perform process reasoning based on the matched process nodes and execution rules, and drive the virtual intelligent agent to complete the automated execution of process tasks.

[0018] To achieve the above objectives, a third aspect of the present invention provides a terminal device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when running the program, implements an agent generation method for automating enterprise processes as described in any one of the first aspects above.

[0019] To achieve the above objectives, a fourth aspect of the present invention provides a computer-readable storage medium storing computer-executable instructions or a computer program that, when executed by a processor, implements an intelligent agent generation method for enterprise process automation as described in any of the first aspects above.

[0020] The present invention provides a method and system for generating intelligent agents for enterprise process automation. By constructing a dual technical path of automatic graphing of standard documents and semantic-driven process execution, it achieves automatic parsing and structured graphing of enterprise policy documents, improving the intelligence level of knowledge extraction and specification modeling. On the other hand, through context-aware prompt word generation and intelligent agent control, it completes the closed-loop linkage from user intent recognition to process execution.

[0021] This invention can dynamically adapt to complex and ever-changing enterprise process specifications, support multiple types of tasks such as approval, leave application, and repair, and has high versatility, scalability, and intelligent interaction capabilities. Ultimately, it realizes the core value of a closed loop from knowledge understanding to execution, fine-grained business rule parsing, and enhanced cross-system collaboration.

[0022] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of a process of generating an intelligent agent for enterprise process automation provided by the present invention; Figure 2 This is a schematic diagram of the system architecture of an embodiment of an intelligent agent generation system for enterprise process automation provided by the present invention. Detailed Implementation

[0024] To make the technical means, inventive features, objectives, and effects of the invention readily understandable, the invention is further illustrated below with reference to specific figures. However, the invention is not limited to the embodiments described below.

[0025] It should be noted that the structures, proportions, sizes, etc., illustrated in the accompanying drawings of this specification are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0026] Terms such as “comprising” and “including” indicate that, in addition to the components that are directly and explicitly stated in the specification and claims, the technical solution of the present invention does not exclude the presence of other components that are not directly or explicitly stated.

[0027] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0028] In traditional enterprise process automation (ERP) technologies, systems based on large language models and knowledge graphs often focus on static knowledge extraction and question-and-answer interactions, lacking the ability to semantically analyze multi-source heterogeneous policy documents and thus failing to construct knowledge graph structures with fine-grained business rule expressions. Such systems struggle to map user natural language requests to executable process nodes, resulting in insufficient accuracy in process path reasoning and an inability to drive automated operations across business systems. Because the granularity of knowledge representation is limited to the entity relationship level, it cannot cover dynamic business elements such as conditional rules and role permissions, making the system prone to logical breaks in scenarios such as approval process transitions and automatic form filling.

[0029] For example, in cross-departmental leave approval processes, company policy documents are scattered across HTML forms in the OA system, PDF user manuals, and internal regulations databases. Current technology uses fixed templates to extract leave rule entities, but it cannot recognize nested conditional rules such as "sick leave exceeding 3 days requires an additional medical certificate." When a user submits a voice request for "5 days of sick leave," the system only returns the policy terms and cannot automatically trigger the medical certificate upload step. Due to the lack of fine-grained rule reasoning, the approval process still relies on manual verification of form fields and policy versions, causing the process to stall at the department manager approval node. The relationship between leave types and approval levels in the knowledge graph is not linked to the ERP system interface, preventing automatic redirection to the finance system to deduct the remaining leave balance.

[0030] If the above issues are not addressed, business process automation will remain at the rule query level, failing to form an end-to-end execution loop. Manual intervention leads to increased process execution delays and error rates, especially in multi-system collaboration scenarios where version differences can cause rule conflicts and process deadlocks. Coarse-grained knowledge representation necessitates the re-annotation of numerous industry-specific entities when migrating the system to new business domains, increasing implementation costs and timelines. The lack of an adaptive mechanism during dynamic business rule changes will cause the knowledge graph to become disconnected from real business processes, reducing system reliability and maintenance efficiency.

[0031] In response, this application proposes a method for generating intelligent agents for enterprise process automation, wherein, for example... Figure 1 As shown, the following steps may be included: Step S1: Obtain the company's standard specifications documents.

[0032] In step S1, standard and specification documents related to the process are collected from the internal data sources of each enterprise. The documents are then subjected to format recognition, content cleaning, and version verification to form a parsable standardized text dataset.

[0033] Standard and specification documents refer to institutional documents related to the company's business processes. Internal data sources can include rules and regulations, operation manuals, OA system documents, etc. Format recognition refers to the structuring of documents in different formats, which can be achieved using OCR technology or document parsing tools to convert unstructured text into parsable standardized data. Content cleaning refers to removing redundant information from documents, which can be achieved using stop word filtering and regular expression matching in natural language processing technologies. In this embodiment, invalid documents (such as historical version files) can be automatically filtered, and document uniqueness is ensured through MD5 hash verification, improving the accuracy of subsequent knowledge extraction. Version verification ensures the timeliness of documents, which can be achieved using timestamp comparison or version number recognition algorithms to avoid using expired process rules.

[0034] Step S2: Knowledge extraction and knowledge graph construction.

[0035] In step S2, entity recognition, relation extraction, attribute extraction, and relation mining are performed on the standardized text dataset to identify and construct an enterprise knowledge graph containing core semantic elements.

[0036] The core semantic elements can include: process nodes, conditional rules, role permissions, etc. Entity recognition refers to extracting business elements from text, while relation extraction refers to establishing business logic relationships between entities. Knowledge graph construction refers to forming an executable semantic network.

[0037] In an optional embodiment, step S2 further includes the following steps: Entity recognition is performed using a fine-tuned BERT large language model, with entity types including but not limited to process entities, rule entities, and execution objects. For example, based on the "Enterprise Employee Attendance Management System" document, entities such as "Attendance Management System," "Leave Rules," "Attendance System," and "HR System" are identified.

[0038] Generative large language models are used to extract relations and attributes from document semantics, and construct structured triples such as "Attendance Management System, applicable to formal employees", "Sick Leave, requires medical certificate", and "Number of late arrivals, 3 times / month, performance deduction", to form a semantic description of conditions, triggers and dependencies in business processes.

[0039] Based on the extracted entities and relationships, a knowledge graph structure is constructed that includes core semantic elements such as process nodes, conditional rules, and role permissions. The node types include process subjects (such as "leave application process"), rule entities (such as "annual leave calculation standard"), and execution objects (such as "HR system"). The relationship types adopt OWL semantic modeling and define attributes such as "requires" (dependency relationship) and "triggers" (trigger relationship).

[0040] The constructed knowledge graph can be stored in graph databases such as Neo4j or TigerGraph, supporting efficient retrieval of enterprise process paths and rule constraints through the Cypher query language.

[0041] Among them, the fine-tuned large language model BERT achieves domain adaptation by adjusting intermediate layer parameters. The proportion of labeled business terms in the training data is controlled between 15% and 30% to balance the accuracy of general semantic understanding and business entity recognition. The generative large language model adopts a sequence-to-sequence architecture. In the relation extraction stage, it captures the logical connections across sentences through an attention mechanism. For example, in the leave application rule, "continuous leave exceeding 5 days requires department head approval" is parsed as a dependency relationship between triggering conditions and approval levels. The process entity is defined as the initiator or executor of the operation in the business process, the rule entity is modeled as condition constraints or operation specifications, and the execution object is mapped to a specific business form or system interface. The three are connected by directed edges to form a traversable semantic network. The graph database uses an attribute graph model to store entities and relations. Node attributes include entity type, effective time, and version identifier, while edge attributes record relation strength and confidence. The indexing strategy adopts a hybrid partitioning to improve the efficiency of multi-hop queries.

[0042] Specifically, in the entity recognition stage, the fine-tuned BERT model analyzes the context of business terms in standardized text to identify the "annual leave day calculation benchmark" in leave application rules and the "cross-departmental collaborative approval" in approval levels as independent entities, improving accuracy to over 92%. During relation extraction, the generative model generates implicit logical relationships based on the semantic distance between entities, for example, associating "department manager approval" with "amount exceeding 5000 yuan" as a conditional trigger relationship. When constructing the graph structure, the process entity is instantiated as employee roles or system accounts, rule entities are personified as approval thresholds or permission levels, and execution objects are bound to electronic forms or API endpoints, forming an executable operation chain. When stored in the graph database, entity attributes and relationship attributes are persisted in key-value pairs, supporting real-time traversal and dynamic updates based on the Cypher language, reducing the response time of process path reasoning to milliseconds. Thus, through the collaborative operation of domain-adaptive entity recognition, context-aware relation extraction, and structured graph storage, precise mapping and efficient execution of fine-grained rules in enterprise business processes are achieved.

[0043] Through the above technical solutions, this application improves the accuracy of entity recognition, particularly in its ability to identify enterprise-specific domain terms. The application of generative models enhances the contextual understanding of relation extraction, capturing more implicit semantic relationships. The constructed graph structure clearly distinguishes between process entities, rule entities, and execution objects, providing effective support for fine-grained business rules. The use of a graph database optimizes knowledge storage and retrieval efficiency. These improvements collectively enhance the quality and efficiency of knowledge graph construction, making it better suited to the execution needs of enterprise process automation.

[0044] Furthermore, in an optional embodiment, the statements in the enterprise standard specification document are subjected to contextual understanding and semantic reconstruction to generate logical triples containing causal relationships, triggering conditions, and result actions. These logical triples are then embedded in a knowledge graph to describe the causal dependencies between events in the business process.

[0045] Contextual understanding and semantic reconstruction are achieved through a generative large language model, which is trained to recognize the implicit causal logic in sentences. For example, when parsing "an employee needs to upload a medical certificate after submitting a sick leave application," the model analyzes the temporal relationship between "submit" and "need to upload" to identify the association between the triggering condition and the resulting action. When generating logical triples, dynamic prompt word templates guide the model to output structured data, such as mapping the input sentence to a triple form of "sick leave application → trigger → upload medical certificate." The storage of logical triples adopts the extended attribute mechanism of a graph database, attaching causal relationships as edge attributes to the entity relationship edges, making the original graph structure compatible with business rule expressions.

[0046] Specifically, when processing process description statements in enterprise standard documents, the first step is to call a finely tuned large language model to perform dependency parsing and semantic role labeling on the statements. For example, when parsing "Leave exceeding 3 days requires department head approval," the model identifies "exceeding 3 days" as the trigger condition and "department head approval" as the result action. Subsequently, the semantic roles are transformed into triple structures using a preset causal logic template, where the trigger condition is encoded as a condition parameter; for example, the threshold parameter "3 days" is associated with the entity "leave days." During the knowledge graph embedding stage, logical triples are transformed into rule nodes with time constraints and dynamically associated with the main process nodes: logical triples are added as new edges or nodes in the knowledge graph, such as creating new edge types like "trigger" and "condition" to connect related event nodes. Alternatively, the entire logical triple can be added to the graph as a composite node and connected to related entities. In this way, the knowledge graph not only includes static relationships between entities but also expresses dynamic causal dependencies between events in the business process. For example, in the leave application flowchart, the "Sick Leave Application" node is connected to the "Upload Medical Certificate" node via a causal edge. The edge attributes record the triggering condition and execution priority parameters. By storing logical triples and entity relationships together, the process reasoning engine can perform multi-hop reasoning based on the causal chains in the graph. For instance, when a user initiates a sick leave application, the requirement to upload a medical certificate is automatically triggered, and the process terminates if the condition is not met. This method overcomes the limitations of traditional entity relationship extraction, achieving fine-grained support for automated process execution using knowledge graphs by structurally expressing causal dependencies in business rules.

[0047] Through the above technical solution, this application can achieve accurate capture and representation of complex business logic in enterprise standard specification documents. By generating logical triples and embedding them into a knowledge graph, the knowledge graph's ability to express causal dependencies in business processes can be enhanced. This enhanced semantic expression capability provides a more accurate and richer information foundation for subsequent process node judgment and rule matching, thereby supporting more granular and accurate automated execution of business processes.

[0048] Furthermore, in an optional embodiment, during the relation extraction process, a dynamic prompt word template can be used to guide the large language model to extract the relation, and the prompt word generation strategy can be continuously optimized based on a reinforcement learning mechanism.

[0049] The dynamic prompt word template generation logic is configured to dynamically adjust the template structure based on the semantic features, entity type distribution, and business rule constraints of the current document. For example, when processing leave application documents, the template automatically embeds entity tags such as "leave type" and "approval level," and associates attribute fields such as "trigger conditions" and "required materials." The reinforcement learning mechanism quantifies the relationship extraction effect by setting a reward function. The reward function parameters can include entity relationship matching accuracy, graph node coverage, and business rule conflict rate, with the accuracy weight set to 0.6-0.8 and the conflict rate penalty coefficient set to 0.2-0.4. The strategy optimization module updates the prompt word generation strategy every 100 documents processed, using the PPO algorithm for gradient updates.

[0050] Specifically, during the knowledge graph construction phase, when the large language model performs relation extraction, the dynamic prompt word template first parses the semantic features of the current document. For example, it extracts high-frequency keywords using the TF-IDF algorithm and combines this with the entity type distribution output by the entity recognition module to generate prompt word content containing context-appropriate instruction statements and example structures. For instance, for a text paragraph containing "sick leave application requires approval from direct supervisor," the template automatically generates an instruction containing "extract approval conditions and triggering rules," along with an example "if the leave type is sick leave, then the approval level is direct supervisor." Subsequently, the reinforcement learning agent calculates reward values ​​based on the relation extraction results. If the extracted "sick leave-approval level" relationship matches existing rules in the knowledge graph, a positive reward is given; if a conflict with existing rules is detected, a penalty mechanism is triggered. The policy network iteratively optimizes and gradually adjusts the semantic strength of the instructions, the number of examples, and the priority of entity types in the template. After multiple rounds of training, the prompt word template can adapt to the characteristics of different companies' institutional documents. For example, it can strengthen the instructions related to "risk control rules" in the financial industry process and add example guidance on "production safety standards" in the manufacturing process.

[0051] In this embodiment, the system records the results and accuracy of each extraction, and feeds the accuracy rate as a reward signal to the reinforcement learning model. The model adjusts its prompt generation strategy based on the feedback, such as increasing or decreasing the frequency of certain keywords or adjusting the structure of the prompts. Thus, the prompt generation strategy can be iteratively optimized as the task is executed. As one implementation method, the reinforcement learning model can employ a Deep Q-Network (DQN) algorithm. The state space includes information such as current document features and extracted entity relationships. The action space is for selecting different prompt templates. The reward function is designed based on the extraction accuracy; the higher the accuracy, the greater the reward. Through multiple rounds of training, the model gradually learns the optimal prompt generation strategy.

[0052] Through the above technical solutions, this application improves the accuracy and generalization of relation extraction. Dynamic prompt word templates enhance the model's adaptability to diverse business scenarios and reduce reliance on fixed templates or manual coding. The reinforcement learning mechanism enables adaptive optimization of the prompt word generation strategy, forming a closed loop of continuous improvement. This not only reduces the cost of manual intervention but also enhances the robustness of relation extraction through dynamic adaptation and continuous learning mechanisms, providing reliable support for fine-grained semantic modeling of knowledge graphs.

[0053] Step S3: Semantic analysis and intent determination.

[0054] In step S3, a natural language request input from the user is received. A contextual prompt word for the process task is dynamically generated based on the contextual information of the knowledge graph. The natural language request is then converted into a structured semantic parsing result. Based on the semantic parsing result, the corresponding process node and execution rule are matched in the knowledge graph. The semantic parsing result may include: identified user intent, process parameters, and constraints, etc.

[0055] Dynamically generating contextual prompts refers to adjusting the semantic parsing strategy according to task requirements. Structured semantic parsing results refer to converting natural language into machine-understandable instructions.

[0056] Optionally, in this embodiment, the knowledge graph may include various node types, such as process entities, rule entities, and execution objects. The semantic relationships between nodes in the knowledge graph are defined using OWL semantic modeling, and these semantic relationships may include dependencies, triggering relationships, etc.

[0057] In this framework, the process entity corresponds to the core object that initiates or carries the process, such as the applicant entity in an approval process; the rule entity corresponds to business constraints, such as leave day limit rules; and the execution object corresponds to specific operational goals, such as forms or approval system entities. The division of node types clarifies the responsibilities of each entity through functional role separation. When semantic relationships are defined using OWL semantic modeling, dependencies represent the logical association between approval levels and leave days, while triggering relationships define the operational conditions for automatically jumping to the next node after form submission. For example, when the leave days exceed a threshold, the dependency relationship forcibly triggers multi-level approval rules; when the user completes the form, the triggering relationship drives the system to automatically submit to the next processing node. The combination of node types and semantic relationships enables the knowledge graph to dynamically reflect the entity interaction logic in the business process.

[0058] Specifically, in the semantic analysis and intent determination process, the node type classification and semantic relationship modeling of the knowledge graph work synergistically. The separation of the main process node and the execution object node avoids the permission confusion caused by traditional single node types; for example, the applicant entity is only associated with form filling permissions, while the approval system entity is associated with form processing permissions. Rule entity nodes define dependencies using the OWL semantic model; for example, binding the leave day threshold attribute to the approval level rule entity forms a reasonable logical chain. When a user request triggers a process, the semantic parsing results determine the process jump conditions and execution order by matching the trigger relationships in the knowledge graph. For example, after a user submits a leave application, the trigger relationship automatically associates the form submission node and the approval node, and the dependency relationship verifies whether the leave days meet the threshold defined by the rule entity, thereby dynamically generating the approval path. Compared to traditional graph construction methods based on simple relationships, this solution enhances the interpretability and execution accuracy of process reasoning through node classification and formal semantic modeling.

[0059] Through the above technical solutions, this application achieves a refined division of knowledge graph node types, improving the ability to express different entity roles in business processes. OWL semantic modeling is used to define relationships between nodes, enhancing the formal description capability of complex business logic. Therefore, the accuracy of the knowledge graph's expression of enterprise process semantics is improved, providing a reliable semantic foundation for subsequent process reasoning and automated execution.

[0060] In an optional embodiment, intent discrimination can classify the semantic content of user input using a hybrid model of BiLSTM and CRF to identify process type and operation category, and support multi-turn dialogue interaction to complete missing parameter information.

[0061] The BiLSTM employs a bidirectional long short-term memory network structure. The input layer receives the text sequence of user requests, and the hidden layer dimension is set to 256. Forward and backward LSTMs extract contextual features. For example, in a request for "3 days of leave," the forward LSTM captures the temporal correlation between "leave" and "3 days," while the backward LSTM identifies the potential connection between "3 days" and the leave type. The CRF layer receives the feature vector output by the BiLSTM and constrains the intent label sequence using a pre-defined label transition matrix. The transition matrix parameters are trained based on historical process data. For example, in a composite intent scenario of "initiating approval and notifying the supervisor," the CRF layer ensures that the transition probabilities of the "approval" and "notification" labels conform to the business logic order. The multi-turn dialogue interaction mechanism is triggered by pre-defined missing parameter detection rules. When a key parameter is detected as unmet, the dialogue management module is invoked to generate follow-up questions. For example, when a user inputs "apply for reimbursement," the system automatically detects missing amounts and invoice types and generates a follow-up request: "Please enter the reimbursement amount and invoice type."

[0062] Specifically, the user's natural language request is first converted into a word vector sequence and input into the BiLSTM model. Bidirectional feature extraction obtains the contextual semantic representation of each word; for example, the word "leave" is associated with the medical certificate parameter in the context of "sick leave requires a diagnostic certificate." The CRF layer calculates the globally optimal label sequence based on the features output by the BiLSTM, avoiding logical conflicts caused by single-label classification. For example, in complex intent scenarios, the "approval" label is enforced to precede the "notification" label. When missing parameters are detected in the semantic parsing results, the system calls the multi-turn dialogue module to generate a parameter completion request and re-inputs the user's supplemented input into the model for iterative parsing. For example, after the user adds "amount 500 yuan," the model updates the process parameters and re-matches the reimbursement rules in the knowledge graph. Through the combination of the hybrid model and the interaction mechanism, the accuracy of user request intent classification is improved to 92.7%, and the parameter completion response time is shortened to 1.2 seconds, thus ensuring the complete information input required for process reasoning and ultimately improving the reliability of automated process task execution.

[0063] Through the above technical solutions, this application improves the accuracy and robustness of intent discrimination. BiLSTM effectively captures long-distance dependencies, and CRF achieves the globally optimal label sequence, avoiding errors that may be caused by local decision-making. The multi-turn dialogue interaction mechanism completes the missing parameters, ensuring the complete information input required for process execution. This method adapts to complex and ever-changing user input scenarios, improving the execution efficiency and reliability of process automation.

[0064] In optional embodiments, the natural language request input by the user can be text input, conversational voice input, form input, or any combination thereof forming a multimodal hybrid form. Text input receives text data input by the user through a natural language processing interface, covering scenarios requiring text description; conversational voice input converts audio signals into text sequences through a voice recognition engine, suitable for mobile devices or on-site operation scenarios; form input extracts preset field values ​​through a structured data parsing module, compatible with data calls from existing enterprise business systems; multimodal hybrid input uses a data fusion layer to perform timestamp alignment and semantic association on input content of different formats. For example, when a user uploads a leave application form and simultaneously submits a voice explanation to supplement the reason for sick leave, the system needs to jointly parse the number of leave days in the form with the description of the illness in the voice. The above input modalities are dynamically matched with process rules in the knowledge graph through a context-aware module. For example, when the keyword "urgent" is detected in the voice input, the retrieval of the expedited approval path in the knowledge graph is automatically triggered.

[0065] Specifically, text input is converted into word vector sequences by a character encoder and then fed into the semantic analysis module. Voice input undergoes dual verification using both acoustic and language models to generate a text transcription result with confidence. Form input binds table fields to knowledge graph entity attributes through predefined template mapping. In multimodal scenarios, the data fusion layer employs an attention mechanism to calculate the semantic association weights between different input modalities. For example, when a user submits both a reimbursement form and a voice description, the system performs cross-modal feature fusion of the amount field in the form and the reason described in the voice, generating a parsing result containing complete business semantics. This parsing result is further matched with approval rule nodes in the knowledge graph. For instance, by combining the reimbursement type in the form with the project number mentioned in the voice, the system automatically associates the corresponding financial review process. By supporting multimodal input and its combinations, the system can adapt to the interaction needs of different business scenarios. For example, in an equipment repair scenario, on-site personnel describe the fault phenomenon via voice while simultaneously uploading an equipment status form. The system integrates these two types of input data and accurately matches the repair process nodes, thereby improving the completeness of information acquisition and the accuracy of process triggering in complex scenarios.

[0066] Through the above technical solutions, this application achieves support for multiple input formats, improving the system's adaptability and user-friendliness. Text input meets routine operational needs, voice input facilitates use in mobile scenarios, form input is compatible with existing business systems, and multimodal input further enhances the completeness and accuracy of information. This flexible input mechanism enables the system to adapt to different user habits and various work scenarios, effectively solving the interaction limitations caused by the single input format in traditional process automation systems.

[0067] Step S4: Process reasoning and automated execution.

[0068] In step S4, process reasoning is performed based on the matched process nodes and execution rules to determine the process path, approval permissions, and business constraints. Simultaneously, based on the reasoning results, process tasks are automatically generated and executed in the target business system. These process tasks may include: automatically filling out forms, initiating approval tasks, and automatically navigating and submitting between process nodes. This can be implemented using API calls or RPA robot technology to achieve a closed-loop operation of form filling and task submission.

[0069] The synergistic action of each step ultimately achieves intelligent parsing of corporate policy documents, automatic execution of business rules, and cross-system process collaboration. Through this solution, this application realizes end-to-end intelligent processing from policy document parsing to automatic process execution. It overcomes the problems of low efficiency and high error rate associated with manual operation in traditional process management. By constructing a fine-grained business knowledge graph, the expressive power and matching accuracy of process rules are improved. The dynamic prompt generation mechanism enhances the system's ability to understand diverse user requests. The process reasoning and automated execution stages achieve cross-system business collaboration, reducing manual intervention and improving process execution efficiency. This solution has strong versatility, can adapt to the business scenarios of different enterprises, and reduces system migration and maintenance costs.

[0070] To achieve the above objectives, the present invention also provides an intelligent agent generation system for enterprise process automation, including a standard document acquisition module, a knowledge extraction and graph construction module, a standard graph storage module, a user request recognition module, a dynamic prompt word optimization module, and an intelligent process execution module.

[0071] The system comprises several modules: a standard document acquisition module, a knowledge graph storage module, and a user graph storage module. The standard document acquisition module collects relevant standard and specification documents from internal enterprise data sources, performs format recognition, content cleaning, and version verification to create a standardized text dataset. A knowledge extraction and graph construction module, connected to the standard document acquisition module, uses a large language model to perform entity recognition, relation extraction, and attribute extraction on the standardized text dataset to construct an enterprise knowledge graph containing core semantic elements. A standard graph storage module, also connected to the knowledge extraction and graph construction module, stores the enterprise knowledge graph in a graph database format. A user request recognition module, connected to the standard graph storage module, identifies user intent in natural language requests and matches corresponding process nodes. A dynamic prompt word optimization module interacts with both the knowledge extraction and graph construction module and the standard graph storage module, adjusting and optimizing prompt words based on task context and feedback information. An intelligent process execution module, connected to the user request recognition module, the dynamic prompt word optimization module, and the standard graph storage module, performs process reasoning based on matched process nodes and execution rules, and drives a virtual agent to automate the execution of process tasks.

[0072] Furthermore, in such Figure 2 In the illustrated embodiment, this system adopts a layered architecture design, which can be further subdivided into a data support layer, a flowchart layer, an intelligent parsing layer, a user interaction layer, an execution control layer, and a self-learning and optimization layer. The functional modules and connections of each layer are as follows: The data support layer includes: The standard document acquisition module is used to collect standard and specification documents related to processes from internal enterprise data sources, and to perform format recognition, content cleaning and version verification on the documents to form a standardized text dataset. The standard graph storage module is used to store enterprise knowledge graphs in the form of a graph database; The log and version management database is used to store process execution logs and system version information.

[0073] The flowchart layer includes: The knowledge extraction and graph construction module is connected to the standard document acquisition module. It is used to call large language models to perform entity recognition, relation extraction and attribute extraction on standardized text datasets in order to construct an enterprise knowledge graph containing core semantic elements. The entity relation extraction and disambiguation module is integrated into the knowledge extraction and graph construction module. It is driven by a large language model and is responsible for semantic disambiguation and logical triple generation.

[0074] The intelligent parsing layer includes: The dynamic prompt word optimization module interacts with the knowledge extraction and graph construction module and the standard graph storage module respectively, and is used to adjust and optimize the prompt words based on the task context and feedback information; The intent recognition and parameter extraction module, based on a large language model, is used to parse user requests.

[0075] The user interaction layer includes: The user request recognition module, connected to the standard graph storage module, is used to recognize the user intent in the natural language request input by the user and match the corresponding process node. The process status query and feedback display module is used to show users the process execution status and results.

[0076] The execution control layer includes: The intelligent process execution module is connected to the user request recognition module, the dynamic prompt word optimization module, and the standard graph storage module, respectively. It is used to perform process reasoning based on the matched process nodes and execution rules, and drive the virtual intelligent agent to complete the automated execution of process tasks. The form autofill and submission module and the OA / ERP system interface module are integrated into the intelligent process execution module, respectively responsible for form operations and external system integration.

[0077] The self-learning and optimization layer includes: The process execution log analysis module, user behavior feedback recording module, and knowledge graph and dynamic prompt word (Prompt) optimization module together constitute the system's closed-loop optimization mechanism, which is used to continuously optimize the knowledge graph and prompt word strategy based on logs and feedback.

[0078] The specific implementation methods have been described in detail above and will not be repeated here.

[0079] To achieve the above objectives, the present invention also provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor runs the program, it can implement the steps of the intelligent agent generation method for enterprise process automation as described in any of the foregoing embodiments.

[0080] Processors and memory can be configured separately or integrated together, for example, integrated on a system-on-chip (SOC) in a terminal device.

[0081] To achieve the above objectives, the present invention also provides a computer-readable storage medium storing computer-executable instructions or computer programs, which, when processed and executed, implement the intelligent agent generation method for enterprise process automation as described above.

[0082] The computer-readable storage medium is, for example, memory. Memory can be volatile or non-volatile, or it can include both volatile and non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which serves as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct rambus RAM (DRRAM).

[0083] If the integrated units in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in the aforementioned computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause one or more computer devices (which may be personal computers, servers, or network devices, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0084] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. An agent generation method for enterprise process automation, characterized by, Comprise the following steps: Step S1: Obtain the standard specification documents of the enterprise; wherein, collect the standard specification documents related to the process from various internal data sources of the enterprise, perform format recognition, content cleaning and version checking on the documents, and form a set of standardized text data that can be parsed; the internal data sources of the enterprise at least include one of the following: rules and regulations, operation manual, OA system documents; Step S2: Knowledge extraction and knowledge graph construction; wherein, perform entity recognition, relationship extraction, attribute extraction and relationship mining on the standardized text data set to identify and construct an enterprise knowledge graph containing core semantic elements; the core semantic elements at least include one of the following: process nodes, condition rules, role permissions; Step S3: Semantic analysis and intent discrimination; wherein, receive a natural language request input by a user, dynamically generate a context prompt word Prompt for a process task based on the context information of the knowledge graph, convert the natural language request into a structured semantic analysis result, and match the corresponding process nodes and execution rules in the knowledge graph based on the semantic analysis result; the semantic analysis result at least includes one of the following: identified user intent, process parameters and constraint conditions; Step S4: Process reasoning and automatic execution; wherein, according to the matched process nodes and execution rules, perform process reasoning to determine the process path, approval authority and business constraints; and based on the reasoning result, automatically generate and execute the process task in the target business system, the process task at least includes one of the following: automatic form filling, initiating an approval task, and at least one of automatically jumping and submitting between process nodes.

2. The method of claim 1, wherein the enterprise process automation agent is generated by: The step S2 further comprises the following steps: Use fine-tuned large language model BERT for entity recognition, the entity at least includes one of the following: leave rules, approval levels; Use a generative large language model for relationship extraction and attribute extraction; Based on the extracted entities and relationships, construct a graph structure containing process subjects, rule entities and execution objects; Store the constructed knowledge graph in a graph database.

3. The method of claim 2, wherein the method further comprises: Contextual understanding and semantic reconstruction of sentences in the enterprise standard specification documents are further generated to generate logical triples containing causal relationships, trigger conditions and result actions, and the logical triples are embedded in the knowledge graph to describe the causal dependency relationship between events in the business process.

4. The method of claim 2, wherein the method further comprises: In the process of relationship extraction, a dynamic prompt word template is used to guide the large language model to extract, and the prompt word generation strategy is continuously optimized based on the reinforcement learning mechanism.

5. The method of claim 1, wherein the enterprise process automation agent is generated by: In the step S3, the knowledge graph includes multiple node types, and the node types at least include one of the following: process subjects, rule entities, execution objects; Wherein, the semantic relationship between nodes in the knowledge graph is defined by OWL semantic modeling, and the semantic relationship at least includes one of the following: dependency relationship, trigger relationship.

6. The method of claim 1, wherein the enterprise process automation agent is generated by: The natural language request input by the user is a text input, a dialogue voice input, a form input, or a multi-modal mixed form formed by any combination thereof.

7. The method of claim 1, wherein the enterprise process automation agent is generated by: In the step S3, the intent discrimination classifies the semantic content of the user input through a hybrid model of BiLSTM and CRF, for identifying the process type and operation category, and supporting multi-round dialogue interaction to complete the missing parameter information.

8. An agent generation system for enterprise process automation, characterized by The embodiment of the application provides an agent generation method for enterprise process automation, and the method comprises the following steps: a standard document collection module, configured to collect standard specification documents related to the process from an internal data source of an enterprise, and perform format recognition, content cleaning and version checking on the documents to form a standardized text dataset; a knowledge extraction and graph construction module, connected with the standard document collection module, configured to call a large language model to perform entity recognition, relationship extraction and attribute extraction on the standardized text dataset to construct an enterprise knowledge graph containing core semantic elements; a standard graph storage module, connected with the knowledge extraction and graph construction module, configured to store the enterprise knowledge graph in the form of a graph database; a user request recognition module, connected with the standard graph storage module, configured to recognize the user intent in the natural language request input by a user and match a corresponding process node; a dynamic prompt word optimization module, respectively connected with the knowledge extraction and graph construction module and the standard graph storage module, configured to adjust and optimize the prompt words based on task context and feedback information; an intelligent process execution module, connected with the user request recognition module, the dynamic prompt word optimization module and the standard graph storage module, configured to perform process reasoning according to the matched process node and execution rules, and drive a virtual intelligent agent to complete the automatic execution of a process task.

9. A terminal device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor runs the program to implement the agent generation method for enterprise process automation according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer executable instructions or computer programs, and the computer executable instructions or computer programs are processed and executed by the processor to implement the agent generation method for enterprise process automation according to any one of claims 1 to 7.

Images