Knowledge enhanced process approval system and method based on large model assistance

Abstract

The invention belongs to the technical field of process approval, and particularly relates to a knowledge-enhanced process approval system and method based on large model assistance. Comprising a data acquisition module, a knowledge base construction module, a large model integration module, a knowledge reasoning engine, an interactive interface and a log and feedback module. The data acquisition module collects and preprocesses various data in the approval process in the production management field and management files related to enterprise production management. The knowledge base construction module establishes a comprehensive multi-mode knowledge base which comprises various types of knowledge resources and carries out tagging processing on the resources in the knowledge base. The method has the beneficial effects that the key information in the approval process can be automatically analyzed, and the pre-trained large model is called to carry out related knowledge pushing, knowledge reasoning and decision suggestion, so that the approver is assisted to make a judgment quickly and accurately, and the approval efficiency of the process can be improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of process approval, and particularly relates to a knowledge enhancement type process approval system and method based on large model assistance. BACKGROUND

[0002] With the expansion of enterprise scale and the increase of business complexity, process approval has become an indispensable part of enterprise management. Traditional process approval systems mainly rely on manual judgment and fixed rules. In the field of production management, it is often difficult to provide comprehensive and accurate knowledge support when facing some complex business scenarios, resulting in low efficiency or decision-making errors. SUMMARY

[0003] The purpose of the present application is to provide a knowledge enhancement type process approval system and method based on large model assistance, which can automatically analyze the key information in the production management field approval process, call the pre-trained large model for related knowledge pushing, knowledge reasoning and decision-making suggestions, so as to assist the approval personnel to make judgments quickly and accurately.

[0004] The technical solution of the present application is as follows: a knowledge enhancement type process approval system based on large model assistance, comprising a data acquisition module, a knowledge base construction module, a large model integration module, a knowledge reasoning engine, an interactive interface and a log and feedback module;

[0005] The data acquisition module collects and pre-processes various data in the production management field approval process and enterprise production management related management files;

[0006] The knowledge base construction module establishes a comprehensive multi-modal knowledge base, including various types of knowledge resources, and tags the resources in the knowledge base;

[0007] The large model integration module integrates pre-trained domain large models, dynamically selects and loads the corresponding models according to the application type;

[0008] The knowledge reasoning engine analyzes the approval data using the large model, identifies key information points, performs knowledge matching and reasoning, and generates approval suggestions or risk warnings;

[0009] At the same time, it can recommend related resources according to the current task type, recommend suitable resources according to the user role, and also make personalized recommendations combined with historical behavior patterns.

[0010] The interactive interface provides an intuitive user interface to display approval details, large model suggestions and reasons, and related knowledge recommendations;

[0011] The log and feedback module records the workflow approval process and the performance of the large model each time, and collects user feedback.

[0012] The various types of data in the data collection module include applicant information, application form category, application form details, and historical approval records.

[0013] The knowledge reasoning engine recommends relevant resources according to the current task type, recommends suitable resources according to the user role, and makes personalized recommendations based on historical behavior patterns.

[0014] A knowledge enhancement type process approval method based on large model assistance, comprising the following steps:

[0015] Step 1: Data collection

[0016] Collect and analyze the data of each approval form in the historical approval workflow in the production management field, including structured data, unstructured data, and associated relationship data; collect documents, pictures, graphics, and audio and video related to the enterprise production management field as a large model training data set;

[0017] Step 2: Data cleaning and preprocessing

[0018] Clean and preprocess the collected data, including removing stop words, standardizing named entities, desensitizing, content analysis, extraction, deduplication, and quality inspection;

[0019] For the collected data, fill in the default values of the missing fields in the application form by rule filling and model prediction, and check for outliers in the application form by Isolation Forest;

[0020] For document resources, analyze the content of the resources through document layout analysis, table recognition and structured analysis, format conversion, text recognition editing and modification, and image processing; structure the document data by data information annotation and intelligent extraction of key elements of the document;

[0021] For audio and video resources, extract key frames, extract subtitle text information, and store video information in a structured manner with metadata indexing; transcribe, label, and classify audio and video resources through ASR;

[0022] Use a graph database to establish a correlation network of "application form-approval opinion-approver-regulation clause", replace synonyms or back-translation for approval opinions, and generate diversified training samples;

[0023] Step 3: Knowledge base construction and storage

[0024] Store unstructured data through MinIO, store graph relationship data through Neo4j graph database, and use Elasticsearch as a search engine to support full-text search; identify knowledge entities, then extract entity relationships using rule engines and deep learning, and finally construct a knowledge graph through ontology mapping and entity linking.

[0025] Step 4: Large model integration

[0026] Develop large models using a L0-L2 three-layer architecture. Based on the L0 level open source large model, adaptively pre-train different domains of data in the knowledge base, and then perform supervised fine-tuning.

[0027] Step 5: Knowledge reasoning and decision-making suggestions:

[0028] Convert the approval system into DRL rules, dynamically load the rules, build a rule base, efficiently match the rules through the Rete algorithm, solve rule conflicts according to priority and recency, traverse the knowledge graph based on Cypher language, and perform knowledge reasoning.

[0029] Use large model reasoning chain and multi-modal understanding to make process decisions, combine Peter-Clark algorithm to find causal relationships between variables from historical data, and give corresponding application approval data in historical approval. All associated rule data, knowledge base data, and historical related approval data are given corresponding provenance information.

[0030] Step 6: User interaction

[0031] Display approval details and model suggestions through the interface, and provide enterprise-related knowledge recommendations corresponding to model suggestions.

[0032] Step 7: Log recording and feedback collection

[0033] Record the entire approval process, score and comment on large model recommendations, and iterate and optimize the model by collecting user feedback.

[0034] The step 4 includes the following:

[0035] Data annotation;

[0036] Output label;

[0037] Parameter efficient fine-tuning: add LoRA adapter to the attention matrix layer of Transformer, inject LoRA weight during inference, and combine the weight;

[0038] Multi-task learning: set different output layers for each task, assign weights to each task when calculating the loss function, and use dynamic task weighting during task training to automatically adjust the weights according to the task training difficulty;

[0039] Fine-tune learning production management different categories of special data to obtain L1 level large model;

[0040] Based on retrieval enhancement (RAG), the L1 level large model is optimized to obtain the L2 level large model, which is a large model with multiple different domains;

[0041] Vector library construction: The cleaned and pre-processed approval rules and historical case data are stored in the FAISS vector database after being vectorized into text, created into vector indexes, and mapped to metadata;

[0042] Real-time retrieval: When a user asks a question, retrieve Top-K relevant fragments and inject them into the prompt;

[0043] Decision generation: The large model combines search results to generate suggestions.

[0044] The beneficial effects of this invention are as follows: This invention can automatically analyze key information in the approval process, call pre-trained large models to push relevant knowledge, reason about knowledge and make decision suggestions, thereby assisting approvers to make judgments quickly and accurately, which can improve the efficiency of the approval process, enhance the accuracy of decision-making, and effectively promote the transfer and utilization of enterprise knowledge. Attached Figure Description

[0045] Figure 1 The flowchart of a knowledge-enhanced process approval method based on a large model provided by the present invention is shown. Detailed Implementation

[0046] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0047] A knowledge-enhanced workflow approval system based on a large model includes a data acquisition module, a knowledge base construction module, a large model integration module, a knowledge reasoning engine, an interactive interface, and a log and feedback module.

[0048] The data acquisition module primarily handles the unified acquisition and preprocessing of multi-source heterogeneous data. Utilizing ETL (Extract-Transform-Load) data pipeline technology, it extracts various structured data from production management approval processes in systems such as ERP, OA, and EAM via API gateways and database connectors. This includes, but is not limited to, process definition data (flowchart files, approval process nodes, process variables, etc.) and historical approval data (applicant information, application type, application details, historical approval records, etc.). Unstructured data related to enterprise production management is extracted from the ECM system, including relevant national standards, industry standards, enterprise standards, company production management procedures / operation guidelines / case studies and experience feedback, and other related documents. For process definition data, the topology of approval nodes is extracted by parsing BPMN format process files, and a node relationship matrix is ​​constructed. For historical approval data, a time-series database is built, recording the approval event sequence E = {e1, e2, ..., e...}. n}, each event e i It includes a four-tuple containing (timestamp, operator, operation type, and approval comment). For unstructured data, document parsing engines and OCR recognition technologies are used to convert the unstructured data into XML format text information.

[0049] The knowledge base construction module includes the construction and storage of multimodal knowledge, encompassing knowledge graphs, vector libraries, and rule bases. It constructs the knowledge graph through entity recognition, relation extraction, and graph storage technologies; uses a dynamic segmentation algorithm based on semantic boundaries to adaptively adjust block size; employs the Sentence-BERT model to generate document vectors; and utilizes multi-classifier fusion to achieve automated annotation of knowledge resources.

[0050] The large model integration module is mainly based on commercial or open-source base models and uses techniques such as dynamic masking pre-training and domain vocabulary expansion to train domain large models. It calculates the model matching degree based on application single features and dynamically selects and loads the corresponding models.

[0051] Knowledge Reasoning Engine: Utilizes large-scale models to perform in-depth analysis of approval data, identify key information points, perform knowledge matching and reasoning, and generate approval suggestions or risk warnings. It can also recommend relevant resources based on the current task type, suitable resources based on user roles, and personalized recommendations based on historical behavior patterns.

[0052] Interactive Interface: Provides an intuitive user interface that displays approval details, large model suggestions and reasons, relevant knowledge recommendations, and supports manual adjustments and confirmation of approval results by approvers.

[0053] Log and Feedback Module: Records each workflow approval process and the performance of the large model, collects user feedback, and is used for continuous model optimization.

[0054] A knowledge-enhanced workflow approval method based on a large model includes the following steps:

[0055] Step 1: Data Collection

[0056] Data acquisition involves both structured and unstructured data collection. Structured data primarily includes process definition data (flowchart files, approval process nodes, process variables, etc.) and historical approval data (applicant information, application type, application details, historical approval records, etc.). Unstructured data mainly includes relevant documents related to enterprise production management, such as national standards, industry standards, enterprise standards, company production management procedures / operation guidelines / case studies and experience feedback. All acquired data will be used as the training dataset for a large model.

[0057] Step 2: Data Cleaning and Preprocessing

[0058] The collected data is cleaned and preprocessed, including stop word removal, standardized named entities, desensitization, content parsing, extraction, deduplication, and quality inspection. High-quality content is then used to build a vector knowledge base.

[0059] For process definition data, the approval node topology is extracted by parsing BPMN format process files, and a node relationship matrix is ​​constructed. For collected historical approval data, default values ​​for missing fields in application forms are automatically filled using rule-based filling and model prediction methods. An isolation forest is used to check for outliers in application forms. A time-series database is constructed to record the approval event sequence E = {e1, e2, ..., e...}. n}, each event e i It contains a quadruple of (timestamp, operator, operation type, approval opinion).

[0060] For document resources, the content is parsed through document layout analysis (using Sentence-BERT to generate embeddings), table recognition (using TabTransformer to encode structured fields), structured analysis, format conversion, text recognition, editing and modification, garbled character detection, and image processing. Document data is structured through data annotation and intelligent extraction of key document elements.

[0061] For audio and video resources, ASR transcription, tagging, and classification are performed through keyframe extraction, subtitle text information extraction, and structured storage of video information and metadata indexing.

[0062] Using the graph database (Neo4j), a network of associations is built, consisting of "timestamp-operator-operation type-approval opinion". Synonym replacement (NLTK) or back translation (Google Translate API) is performed on the approval opinions to generate diverse training samples.

[0063] The large-scale model-based conflict detection system automatically detects conflicting entries in the knowledge base (such as contradictions between old and new policies) and recommends supplementary knowledge base resources (using web crawling technology to crawl the latest policy and regulatory documents from the Internet and add them to the knowledge base).

[0064] Step 3: Knowledge Base Construction and Storage

[0065] Unstructured data is stored using MinIO, graph relation data is stored using Neo4j graph database, and Elasticsearch serves as the search engine to support full-text search. Knowledge entity recognition is performed using technologies such as spaCy and Faster R-CNN, then entity relations are extracted using a rule engine and deep learning, and finally a knowledge graph is constructed through ontology mapping and entity linking.

[0066] Step 4: Large Model Integration

[0067] A three-layer (L0-L2) architecture was adopted for large-scale model development. Based on the L0-level open-source large model (Deepseek), adaptive pre-training was performed on data from different domains within the knowledge base. Through techniques such as dynamic masking and domain vocabulary expansion, the large model was able to understand the professional terminology and rule logic of the approval domain. Supervised fine-tuning (STF) was then performed, with the following steps:

[0068] Data annotation: During the annotation phase, a structured rule base has been built. The standard process generates corresponding approval decisions and justifications based on matching the input data with the rules in the rule base.

[0069] The template for the rule is as follows:

[0070] IF<condition set> THEN<decision> WITH<reason template>

[0071] Condition types include the following:

[0072] ● Numerical conditions: Field OP threshold (OP: >, <, =, ≥, ≤)

[0073] ●State condition: State ∈ {set of allowed states}

[0074] ● Relationship Conditions: The entity relationship satisfies a specific schema.

[0075] ●Complete condition: A logical combination of the above conditions

[0076] The annotation process is as follows:

[0077] 1) Data parsing: Based on natural language processing techniques, the text is segmented into sentences, analyzed for parts of speech, identified as entities, and extracted numerical values. Then, entity relationships are extracted based on dependency parsing, and finally, structured entity information is obtained.

[0078] 2) Rule Retrieval: The rule base is retrieved based on different application scenarios such as acceptance approval, procurement approval, and quality approval. The extracted data is compared with the rule conditions to calculate the rule matching degree.

[0079] Rule matching degree = ∑(condition weight × condition satisfaction degree)

[0080] The condition satisfaction degree is defined as follows:

[0081] ●Exact match: 1.0

[0082] ●Partial match: 0-1 based on the degree of match.

[0083] ●Mismatch: 0

[0084] The rule matching process is as follows: For each rule, iterate through its conditions, calculate the satisfaction of each condition, and sum the rule matching score; filter out rules with matching scores below the threshold; sort the matched rules by priority, specificity (number of conditions), matching score, and update date; select the top 1 rule as the final rule; and calculate the confidence score.

[0085] Confidence score calculation method: rule matching degree × rule reliability × quality factor

[0086] When completing rule matching, it is necessary to record the decision conditions and confidence levels for use in the label generation process.

[0087] 3) Tag Generation: The final output tags include decision, reason, confidence, and supporting evidence. Decision, confidence, and supporting evidence are derived from the rule retrieval stage. The reason is generated based on the defined reason template. The reason template contains fixed text, variable interpolation, and conditional logic. Variable interpolation uses the tag "{variable name}", and conditional logic uses the tag "{?conditional expression}true value text{:false value text}". These tags are replaced during tag generation to obtain the final tags.

[0088] The following example illustrates the processing procedure.

[0089] Input: {"text":"All Category 1 items are closed, 80% of Category 2 items are closed, Applicant: Zhang San, Department: Debugging Department"}

[0090] The processing procedure is as follows:

[0091] 1) Entity extraction: {Category 1 item:{Status:"All Closed",Close Rate:100%}, Category 2 item:{Close Rate:

[0092] 80%

[0093] 2) Rule matching: Matched the rule "Debugging and Acceptance Standard 3.2"

[0094] 3) Condition verification:

[0095] Condition 1: Closure rate of Category 1 items = 100% → Satisfied

[0096] Condition 2: Closure rate of Category II items ≥ 75% → Satisfied

[0097] 4) Tag generation:

[0098] Define the following reason template

[0099]

[0100] The final generated tags are as follows:

[0101]

[0102] Parameter-efficient fine-tuning (LoRA, Low-Rank Adaptation): A hierarchical learning rate strategy is designed, with a conservative learning rate of 1e-5 for the base model parameters and an aggressive learning rate of 5e-4 for the LoRA adapter parameters, resulting in a 50-fold difference. A LoRA adapter is added to the Transformer's attention matrix (Query / Value) layer, injecting LoRA weights and merging them during inference. This technique of reducing the number of trainable parameters through low-rank matrix factorization achieves results comparable to full-parameter fine-tuning of a large model with extremely low resource consumption (e.g., a difference of <2% in approval decision accuracy).

[0103] Multi-task learning: By employing a gradient coordination algorithm, this approach balances knowledge sharing with task specificity in multi-task learning, achieving collaborative optimization across tasks such as approval decision-making, reason generation, and risk assessment. For different types of tasks like approval decision-making, opinion generation, and risk prediction, the underlying model shares most of the parameters from the first few layers of the Transformer, learning general feature representations, but sets different output layers for each task.

[0104] The approval decision-making process uses a linear layer with a softmax classification head, opinion generation uses a generation head, and risk prediction uses a regression head. When calculating the loss function, weights are assigned to each task to balance the scale and importance of different tasks. During task training, dynamic task weighting (GradNorm algorithm) is used, automatically adjusting weights based on task training difficulty to avoid simple tasks dominating training. Multi-task learning improves the model's generalization ability and efficiency.

[0105] After fine-tuning the learning of different categories of specialized data for production management, a large L1-level model was obtained.

[0106] The L1-level large model is optimized using the Retrieval-Augmented Generation (RAG) technique to obtain the L2-level large model, which consists of multiple large models from different domains.

[0107] The main steps required for RAG are as follows:

[0108] Vector library construction: The cleaned and pre-processed approval rules and historical case data are stored in the FAISS vector database after being vectorized into text, created into vector indexes, and mapped to metadata.

[0109] Time-based retrieval: When a user asks a question, retrieve Top-K relevant fragments and inject them into the Prompt.

[0110] Decision generation: The large model combines search results to generate suggestions.

[0111] Step 5: Knowledge Reasoning and Decision Recommendations:

[0112] Symbolic reasoning (rule-driven) and neural reasoning (data-driven) work together to assist in knowledge reasoning and decision-making recommendations.

[0113] The approval system is transformed into DRL (Drools Rule Language) rules, supporting dynamic rule loading and rule base construction. The Rete algorithm efficiently matches rules, resolving rule conflicts based on priority (salience) and recency. Furthermore, knowledge reasoning can be performed by traversing the knowledge graph using the Cypher language.

[0114] The system employs large-scale model inference chains and multimodal understanding to assist in process decision-making. It combines the Peter-Clark algorithm to discover causal relationships between variables in historical data, providing corresponding application approval data from past approval processes. All rule data, knowledge base data, and historical approval data retrieved through related queries are accompanied by source information for further user review.

[0115] Step 6: User Interaction

[0116] The interface displays approval details and model suggestions, and provides relevant enterprise knowledge recommendations corresponding to the model suggestions, so that users can promptly access the relevant knowledge to assist in decision-making. Users can make approval decisions based on the suggestions or choose the approval conclusion themselves.

[0117] Step 7: Log Recording and Feedback Collection

[0118] The entire approval process is recorded, and users can rate and comment on the large model recommendations. By collecting user feedback in this way, the model can be continuously iterated and optimized.

Claims

1. A knowledge-enhanced workflow approval system based on a large model, characterized in that: It includes a data acquisition module, a knowledge base construction module, a large model integration module, a knowledge reasoning engine, an interactive interface, and a log and feedback module; The data acquisition module collects and preprocesses various types of data in the approval process of production management and enterprise production management-related management documents. The knowledge base construction module establishes a comprehensive knowledge base, including various types of knowledge resources, and tags the resources in the knowledge base. The large model integration module integrates pre-trained domain-specific large models and dynamically selects and loads the corresponding model according to the application type. The knowledge reasoning engine uses a large model to analyze approval data, identify key information points, perform knowledge matching and reasoning, and generate approval suggestions or risk warnings. It can also recommend relevant resources based on the current task type, recommend suitable resources based on the user's role, and make personalized recommendations based on historical behavior patterns. The interactive interface provides an intuitive user interface that displays approval details, large model suggestions and reasons, and recommendations for relevant knowledge; The log and feedback module records each workflow approval process and the performance of the large model, and collects user feedback.

2. The knowledge-enhanced workflow approval system based on a large model as described in claim 1, characterized in that: The data acquisition module includes various types of data such as applicant information, application type, application details, and historical approval records.

3. The knowledge-enhanced workflow approval system based on a large model as described in claim 1, characterized in that: The knowledge reasoning engine recommends relevant resources based on the current task type, recommends suitable resources based on the user role, and makes personalized recommendations by combining historical behavior patterns.

4. A knowledge-enhanced workflow approval method based on a large model, characterized in that, Includes the following steps: Step 1: Data Collection Collect and analyze data from various approval forms in historical approval workflows in the production management field, including structured data, unstructured data, and relational data; Collect documents, images, graphics, and audio / video materials related to enterprise production management to serve as a training dataset for a large model; Step 2: Data Cleaning and Preprocessing Step 3: Knowledge Base Construction and Storage Unstructured data is stored using MinIO, graph relation data is stored using Neo4j graph database, and full-text search is supported using Elasticsearch as the search engine. Knowledge entity recognition is performed, and then entity relations are extracted using a rule engine and deep learning. Finally, a knowledge graph is constructed through ontology mapping and entity linking. Step 4: Large Model Integration We adopt a three-layer architecture of L0-L2 for large model development. Based on the L0-level open-source large model, we perform adaptive pre-training for different domains of data in the knowledge base, followed by supervised fine-tuning. Step 5: Knowledge Reasoning and Decision Recommendations: Step 6: User Interaction The interface displays approval details and model suggestions, and provides relevant enterprise knowledge recommendations corresponding to the model suggestions. Step 7: Log Recording and Feedback Collection The entire approval process is recorded, the large-scale model recommendations are scored and commented on, and user feedback is collected to iterate and optimize the model.

5. The knowledge-enhanced process approval method based on a large model as described in claim 4, characterized in that, Step 4 includes the following: Data annotation; Output labels; Efficient parameter fine-tuning: Add a LoRA adapter to the attention matrix layer of the Transformer, inject LoRA weights during inference and merge the weights; Multi-task learning: Different output layers are set up for each task, and weights are assigned to each task when calculating the loss function. During task training, dynamic task weighting is used, and the weights are automatically adjusted according to the training difficulty of the task. After fine-tuning the learning of different categories of specialized data for production management, an L1-level large model was obtained. Based on retrieval enhancement generation, the L1 level large model is optimized to obtain the L2 level large model, which is a large model of multiple different domains. Vector library construction: The cleaned and pre-processed approval rules and historical case data are stored in the FAISS vector database after being vectorized into text, vector indexes are created, and metadata is mapped. Real-time retrieval: When a user asks a question, retrieve Top-K relevant fragments and inject them into the prompt; Decision generation: The large model combines search results to generate suggestions.

6. The knowledge-enhanced process approval method based on a large model as described in claim 4, characterized in that: Step 2 involves cleaning and preprocessing the collected data, including removing stop words, standardizing named entities, desensitizing, parsing content, extracting, deduplication, and quality inspection. For the collected data, the default values ​​of missing fields in the application form are automatically filled in through rule-based filling and model prediction, and outlier checks are performed on the application form through an isolated forest. For document resources, the content is parsed through document layout analysis, table recognition and structured analysis, format conversion, text recognition editing and modification, garbled character detection and image processing. The document data is structured through data information annotation and intelligent extraction of key document elements. For audio and video resources, ASR transcription, tagging and classification are performed on the audio and video resources through keyframe extraction, subtitle text information extraction and structured storage of video information and metadata indexing; Using a graph database, a network of connections is established between "application form - approval opinion - approver - policy clauses". Synonym replacement or back-translation is performed on the approval opinion to generate diverse training samples.

7. The knowledge-enhanced process approval method based on a large model as described in claim 4, characterized in that: Step 5 transforms the approval system into DRL rules, dynamically loads the rules, builds a rule base, efficiently matches rules using the Rete algorithm, resolves rule conflicts based on priority and proximity, and performs knowledge reasoning by traversing the knowledge graph using the Cypher language. The process decision-making is carried out using large model inference chains and multimodal understanding. The Peter-Clark algorithm is combined to discover causal relationships between variables from historical data and provide the corresponding application approval data in the historical approval process. All the rule data, knowledge base data and historical approval data retrieved through related queries are given corresponding source information.

Images