LLM-based approval process intelligent information management method and device

Abstract

The invention relates to the field of information management systems, and discloses an approval process intelligent information management method and device based on LLM, and the method comprises the steps: constructing an approval process description vector library; the user inputs an approval demand, and meanwhile, the approval demand is vectorized to obtain a user approval demand vector; inputting the user approval demand vector into an approval process description vector library for matching to obtain an approval process entry with the highest matching degree; and the user enters the approval process entrance with the highest matching degree, and performs approval process circulation according to a preset template. According to the invention, the intensity of the manual approval process is obviously reduced, and the approval efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information management systems, in particular to an intelligent information management method and device for approval process based on LLM. BACKGROUND

[0002] The approval process management system is an integrated platform for whole life cycle management and control of multi-dimensional information of process initiation subject, approval decision subject and system management subject through informationization means. Its typical operation paradigm can be abstracted as the following four-order closed-loop process (taking leave application as an example): 1. Process initialization stage: after the initiator (such as the leave applicant) logs in the system through identity verification, the target approval portal is located through the process mapping engine, the multi-modal form data is entered (mandatory fields: leave classification, time interval, reason for leave; non-mandatory fields: supporting documents, work handover person, supplementary explanation, etc.), and the submission operation is performed; 2. Process routing stage: the system automatically executes approval path analysis according to the preset rule engine (Rule Engine), persistently stores the structured application data to the distributed database, and asynchronously pushes it to the corresponding approval node through the message queue (Message Queue); 3. Decision processing stage: the approval subject reviews the application metadata (including text, attachments and other multi-source heterogeneous data) in the visual workbench, completes the decision information annotation (mandatory fields: approval status; non-mandatory fields: approval notes, etc.) based on business rules, generates a digital signature and submits it; 4. Process closed-loop stage: the system persistently stores the decision result, triggers the state change notification to the initiator through the event-driven architecture (Event-Driven Architecture), and updates the process instance state machine.

[0003] The prior art mainly relies on traditional information management systems to realize linear control of the approval process, but as the complexity of the process matrix increases and the multi-state process variants derive, the existing system presents the following technical limitations: 1. Low efficiency of process entry addressing: heterogeneous processes have high similarity in naming rules, access conditions and form structure dimensions, and the current keyword-based search mechanism faces the problem of semantic ambiguity, resulting in cognitive load for new users in the process selection stage; 2. Insufficient fault tolerance of form filling: when the field space dimension of the form is too large, the traditional linear filling mode is easy to cause field omission (especially mandatory items), resulting in operation entropy increase of process rollback and repeated submission; 3. Weak support for approval decision: when facing unstructured data (such as PDF reports, Excel tables), the approval personnel need to manually perform pattern matching and compliance checking, which has the risk of decision delay and misjudgment; 4. Data value mining is missing: massive historical approval data is only stored as archives, lacking time series analysis and process pattern mining capabilities based on machine learning, making it difficult to achieve process self-optimization. SUMMARY

[0004] The purpose of the present application is to provide an LLM-based intelligent information management method and device for approval process, which solves the technical problems of low processing efficiency and poor fault tolerance of the existing process approval information management system.

[0005] Specifically, the present application provides an LLM-based intelligent information management method for approval process, which comprises the following: S1, constructing an approval process description vector library; S2, the user inputs the approval requirement, and the approval requirement is vectorized to obtain a user approval requirement vector; S3, inputting the user approval requirement vector into the approval process description vector library for matching to obtain the approval process entry with the highest matching degree; S4, the user enters the approval process entry with the highest matching degree and performs the approval process flow according to the preset template.

[0006] A storage medium stores instructions and data for implementing an LLM-based intelligent information management method for approval process.

[0007] An LLM-based intelligent information management device for approval process comprises a processor and a storage medium; the processor loads and executes the instructions and data in the storage medium to implement an LLM-based intelligent information management method for approval process.

[0008] The present application provides the beneficial effects that: by deeply integrating LLM and process automation technology, an intelligent approval system with cognitive reasoning ability is constructed. The present application adopts a hierarchical fusion architecture design, realizes a voice-text dual-modal adaptive input channel at the interaction level, relies on a multi-modal semantic analysis engine to dynamically extract and structure the unstructured information; a rule engine and an LLM collaborative verification mechanism are constructed at the logic processing level, a conflict detection chain is constructed by using a prompt engineering to realize form compliance verification and business logic consistency review of the declaration content. By using the dialog state tracking technology, the progressive form filling is realized, and the context perception module is combined to dynamically recommend the associated fields, which significantly reduces the artificial review intensity. BRIEF DESCRIPTION OF DRAWINGS

[0009] Figure 1 is a simple process schematic diagram of the method of the present application; Figure 2 is a schematic diagram of the LLM-based approval process of the present application; Figure 3It is a schematic diagram of the working of the hardware device of an embodiment of the present invention. DETAILED DESCRIPTION

[0010] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0011] Before formally explaining the present invention, the scheme of the present invention is first generally explained for easy understanding.

[0012] The technical terms involved in this invention are as follows: LLM: Large Language Model (LLM) refers to a deep learning model trained with large amounts of text data, which enables the model to generate natural language text or understand the meaning of language text. These models can provide in-depth knowledge and language production on various topics by training on large datasets. [1] The core idea is to learn the patterns and structures of natural language through large-scale unsupervised training, and to simulate the human language cognition and generation process to a certain extent.

[0013] Sentence-Embedding models: This is a technology that converts sentences into vector representations, primarily used for matching and calculating text semantic similarity. Sentence-Embedding models convert sentences into points in a high-dimensional vector space, enabling computers to understand and process the semantic content of sentences. This technology captures the semantic and structural relationships between data by mapping discrete symbolic data (such as words and sentences) into a continuous vector space. Specifically, Sentence-BERT (Sentence Embedding using BERT) is a commonly used sentence embedding model that uses the BERT model to encode sentences and generate fixed-length vector representations.

[0014] Please refer to Figure 1 The present invention provides an LLM-based intelligent information management method for approval processes, comprising: S1. Build an approval process description vector library; It should be noted that step S1 is specifically as follows: collecting all approval process entry URLs and corresponding process descriptions, vectorizing the process descriptions, and generating an approval process description vector library.

[0015] It should be noted that the process description is generated with the assistance of LLM.

[0016] It should be noted that the process description vectorization is performed using a Sentence-Embedding model.

[0017] S2, the user inputs the approval requirement, and the approval requirement is vectorized to obtain a user approval requirement vector; It should be noted that when the user inputs the approval requirement, the user generally has the approval requirement, however, there are many different types of approvals, such as leave approval or financial reimbursement approval, and therefore, the corresponding entry is different for different approval types or requirements, and in this case, the user generally needs to search for the approval entry.

[0018] In the present application, two methods are provided, one is that the user explicitly inputs the requirement, and the other is that the user does not clearly know the requirement or is not familiar with the approval information management system, and in this case, the LLM process approval intention module is used for identification, and the user can complete the submission of the approval process by conversing with the LLM.

[0019] Regardless of which method is used, when the user's exact approval requirement is obtained, the Sentence-Embedding type model is used for vectorization to obtain the corresponding user approval requirement vector.

[0020] S3, inputting the user approval requirement vector into an approval process description vector library for matching to obtain an approval process entry with the highest matching degree; Specifically, the user approval requirement description vector is used to search the vector library constructed in step S1 to retrieve the approval process entry with the highest matching degree and return top-N results for the user to select. The user can determine the final approval process entry according to the user's own situation.

[0021] S4, the user enters the approval process entry with the highest matching degree and performs the approval process flow according to a preset template.

[0022] It should be noted that step S4 is as follows: S41, the user enters the approval process entry with the highest matching degree; S42, the corresponding template corresponding to the approval process entry is called from the template library; As an embodiment, the information in the approval template can be as follows: Name, title, basic information, detailed information, direct supervisor approval, HR department approval, high-level approval, etc.

[0023] Further, the basic information can include department, position, application date, etc. Further, the detailed information can include leave type, leave time, leave reason, sick leave certificate, work handover person, work handover content, emergency contact person, emergency contact person's phone number, etc. Further, the direct supervisor approval can further include: the approver, the approval time, the approval opinion, the approval reason, the supplementary opinion, etc. Further, the HR department approval can further include: the approver, the approval time, the holiday balance, the attendance system mark, etc. Further, the high-level approval can further include: the approver, the approval time, whether to approve, the note, etc. The above is only an approval template of a leave process, and other types of leave templates can be set according to actual conditions, and the present application is only illustrative and not limited.

[0024] S43, the LLM refers to the corresponding template to guide the user to fill in the required and optional information; S44, the LLM performs completeness confirmation, if the filled information is not complete, then returns to step S43, otherwise enters step S45; Whether the filled information is complete or not is judged according to whether the set required information is filled in.

[0025] S45, the LLM displays the filled information and guides the user to confirm the submission of the application; S46, the information is stored in the database, and the user application is transferred to each approval link for approval in turn.

[0026] It should be noted that after the user performs the process approval, it further performs item-by-item approval at each approval link. For example, the LLM refers to the template to check the application information and generates an abstract for the approver, and simultaneously prompts the important items and the check conclusion.

[0027] Further, the LLM refers to the template to guide the approver to gradually fill in the required information and the optional information.

[0028] Further, the LLM can adopt a dialogue mode to guide and assist the approver to confirm the filled information, and finally, the complete approval process is displayed to the applicant.

[0029] As another embodiment, when the user demand is not too clear in step S2, or the user is not familiar with the approval information management system, the LLM process approval intention module is used to identify, and the user can complete the submission of the approval process by dialoguing with the LLM.

[0030] Specifically, when the user initiates a question, the LLM enters the approval process intention recognition state, and if it is confirmed that the user needs to initiate the approval process, the user is guided to fill in the corresponding approval information based on the filling process template library. On the other hand, the LLM extracts the necessary information required in the template from the user's reply and checks whether it has replied to the mandatory items. If the mandatory items are completed, it is considered to meet the inspection requirements, the approval process information is stored in the database, and the background flow is entered at the same time, and the approver is notified. The LLM generates an abstract based on the declaration information filled in by the declarer, highlights the approval concerns, and guides the approver to complete the approval.

[0031] In the process of guiding the approver to complete the approval, the key information can still be extracted from the reply of the approver in the form of dialogue, and whether the approver has filled in the mandatory items is checked. Finally, the approval process ends, and the LLM notifies the declarer and displays the approval result.

[0032] Please refer to Figure 3 , Figure 3 is a hardware device working schematic diagram of an embodiment of the application. The hardware device specifically comprises: an LLM-based approval process intelligent information management device 401, a processor 402, and a storage medium 403.

[0033] The LLM-based approval process intelligent information management device 401: the LLM-based approval process intelligent information management device 401 realizes the LLM-based approval process intelligent information management method.

[0034] The processor 402: the processor 402 loads and executes the instructions and data in the storage medium 403 to realize the LLM-based approval process intelligent information management method.

[0035] The storage medium 403: the storage medium 403 stores instructions and data; the storage medium 403 is used to realize the LLM-based approval process intelligent information management method.

[0036] In the application, the approval process template library architecture design adopts a structured data format such as JSON / XML to construct an extensible process template library, and the following elements are defined: a. Metadata attributes of each process node (mandatory item / optional item label) b. Logical check rules of the dependency relationship between data items (including conditional constraints and cascading trigger mechanisms) c. Process flow control parameters based on state machines, and a Schema verification mechanism is used to ensure the semantic integrity and logical consistency of the template.

[0037] On the other hand, the B / S architecture is used to evolve the LLM dialogue approval system: Constructing a multi-modal interaction system: the front end realizes a natural language interaction interface (NLUI), integrating the semantic analysis capabilities of large language models; the back end retains traditional back-end management modules to maintain process supervision and data analysis functions At the same time, some core functions have been refactored, including: Intelligent auxiliary reporting: automatically complete form information through entity recognition and context understanding of LLM; dynamic verification engine: real-time data integrity check and logic conflict warning based on template rules; Supporting multi-modal input (text / speech / document) unstructured data normalization conversion.

[0038] As some extendable embodiments, a historical approval process database (including structured process logs and unstructured approval opinions) can also be used to build a small-scale specialized model with strong field adaptability through parameter tuning (P-tuning).

[0039] As some extendable embodiments, for large-size base models such as GPT-4o, knowledge-enhanced prompt engineering (including Few-shot Learning and Chain-of-Thought strategy) and retrieval-augmented generation (RAG) technology can be used to integrate process knowledge base.

[0040] As some extendable embodiments, when identifying the intentions of the declarer, a LLM-based process selection classifier or an integrated semantic similarity calculation module (such as BERT-Whitening) can be established to deal with the diversity of expressions.

[0041] As some extendable embodiments, based on the approval process template, multi-round dialogue management can be adopted to dynamically bind template verification rules to dialogue state, ensuring process compliance.

[0042] As some extendable embodiments, the invention can also implement a gradual information collection strategy based on the mandatory guidance mechanism of template mandatory items, context-related recommendations for optional fields, and real-time correction suggestions for abnormal inputs.

[0043] In summary, the beneficial effects of the invention are: by deeply integrating LLM and process automation technology, an intelligent approval system with cognitive reasoning ability is constructed. The system adopts a hierarchical fusion architecture design, realizes voice-text dual-modal adaptive input channel at the interaction level, relies on a multi-modal semantic analysis engine to dynamically extract and structure unstructured information; at the logic processing level, a rule engine and LLM collaborative verification mechanism are constructed, and a conflict detection chain is constructed using prompt engineering to verify the form compliance of the declaration content and the business logic consistency. Through dialogue state tracking technology, progressive form guidance and reporting are realized, and related fields are dynamically recommended by combining context awareness modules, significantly reducing the intensity of manual review.

[0044] The above description is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An intelligent information management method for approval process based on LLM, characterized by: The following steps are involved: S1. Build an approval process description vector library; S2. The user inputs the approval requirement, and the approval requirement is vectorized to obtain the user approval requirement vector; S3. Input the user approval requirement vector into the approval process description vector library for matching, and obtain the approval process entry with the highest matching degree; S4. The user enters the approval process entrance with the highest matching degree and proceeds through the approval process according to the preset template.

2. The LLM-based intelligent information management method for approval process according to claim 1, characterized in that: Step S1 is specifically as follows: collect all approval process entry URLs and corresponding process descriptions, vectorize the process descriptions, and generate an approval process description vector library.

3. The LLM-based intelligent information management method for approval process according to claim 2, characterized in that: The process description is generated with the assistance of LLM.

4. The LLM-based intelligent information management method for approval process according to claim 2, characterized in that: The process description vectorization is performed using a Sentence-Embedding model.

5. The LLM-based intelligent information management method for approval process according to claim 4, characterized in that: The approval requirement vectorization in step S2 is performed using the same Sentence-Embedding model.

6. The LLM-based intelligent information management method for approval process according to claim 1, characterized in that: Step S4 is specifically as follows: S41. The user enters the approval process entrance with the highest matching degree; S42. Retrieving a template corresponding to the entry point of the approval process from the template library; S43, LLM guides users to fill in required and optional information by referring to the corresponding template; S44: LLM confirms the completeness of the information submitted. If the submitted information is incomplete, it returns to step S43; otherwise, it proceeds to step S45. S45, LLM displays the filled information and guides the user to confirm and submit the application; S46. Information is stored in the database, and the user's application is transferred to each approval link for approval in turn.

7. The LLM-based intelligent information management method for approval process according to claim 1, characterized in that: In step S2, the user inputs the approval requirement by actively inputting it or communicating with the LLM. After the LLM recognizes the user's intention, it assists in generating the user's approval requirement.

8. A storage medium, characterized in that: The storage medium stores instructions and data for implementing an LLM-based intelligent information management system for approval processes as described in any one of claims 1 to 7.

9. An intelligent information management device for approval process based on LLM, characterized by: include: Processor and storage medium; the processor loads and executes instructions and data in the storage medium to implement an LLM-based approval process intelligent information management system as described in any one of claims 1 to 7.

Images