Business process rapid modeling method based on large language model

Abstract

The invention discloses a business process rapid modeling method based on a large language model. The method comprises the following steps: preprocessing an input business process description text, and splitting complex sentences and composite sentences into simple sentences by utilizing a large language model; and extracting the fact types in the simple sentences in a chained manner, wherein the fact types comprise entities, events and event types. And giving a plurality of event relation sets with possible prediction according to the extracted event sets through a large language model, and combining the event relation sets with the event relation sets extracted in a chain manner to form an event relation pool. And using the event relationship as a gene by adopting a genetic algorithm. After the evolution process of the genetic algorithm, an event relation set with the highest fitness is filtered out from an event relation pool. And constructing a logical relationship between the fact type and the business process elements, and converting the extracted information into corresponding json data. The method comprises the following steps of: realizing conversion among structured data through a template, and converting json data into a business process execution language (BPEL) and a language for describing a graphic element of a BPMN (Business Process Execution Language). And finally, inputting the BPEL and a language for describing the BPMN graphic elements into business flow chart modeling software to generate a visual business flow chart. Through the above mode, the modeling efficiency can be improved, the modeling threshold can be reduced, and further customization and editing can be carried out in a professional modeling tool to meet specific requirements.

Description

Technical Field

[0001] The present invention relates to the technical field of large language models and business processes, belonging to business process modeling technology, and in particular to a method for rapid business process modeling based on a large language model. Background Art

[0002] With the development of information technology, business process modeling has become an important means of enterprise management and optimization. Business process modeling can help enterprises understand and analyze business processes, identify problems and optimize and improve them. The life cycle of business process management includes six steps: process modeling, discovery, analysis, redesign, implementation and process monitoring and control. Among them, process modeling can be considered one of the most important and complex steps. Traditional manual business process modeling methods usually require professionals to spend a lot of time and energy on analysis and design, and have defects such as high threshold, low efficiency and non-customization. Some scholars have developed different methods for business process modeling automation using NLP technology, but they are often accompanied by various defects, such as restrictions on input text and incomplete BPMN diagrams.

[0003] In the new information age, with the continuous evolution of technology, enterprises need more efficient, customizable and user-friendly business process modeling methods. The emergence of large language models provides new ideas and methods for business process modeling. Large language models use neural networks to train large-scale text corpora, which can learn the grammar and semantic rules of natural language and generate text similar to the original text. The core advantage of large language models (such as the GPT series) lies in their powerful natural language understanding capabilities. Relying on the deep understanding of semantics by large language models, they can accurately and efficiently complete tasks such as the extraction of entities, events and event relationships, and the prediction of event relationships.

[0004] The business process rapid modeling method based on large language model can provide a faster, automated, accurate and user-friendly business process modeling method, which has broad application prospects and can improve the efficiency of process modeling in various industries and fields. Therefore, it is necessary to design a business process rapid modeling method based on large language model to solve the above technical problems. Summary of the invention

[0005] In view of the problems existing in the prior art, the main purpose of the present invention is to provide a rapid modeling technology, which uses a large language model to extract fact types from business process description text, maps them into relevant business modeling elements, and then puts them into business process modeling software for modeling, aiming to solve the structural limitations, grammatical limitations and other problems of existing automated business process modeling methods.

[0006] Technical solution of the present invention: A method for rapid business process modeling based on a large language model comprises the following steps: Step S1, extracting fact types existing in the business process description text; Fact types include entities, events, and event relations; The entity represents a specific object or participant involved in the business process and is the subject or object of the process; The event represents a specific action, operation or task that occurs in a business process; The event relationship represents the logical connection or sequence between different events, and is used to describe the sequence, causal relationship, and dependency relationship of events. Step S2: construct the logical relationship between fact type and business process element; Business process elements include flow objects, connection objects, and swimlanes; the details are as follows: The flow object is the main graphical element for defining a business process, and includes three types: events, activities, and gateways. An event refers to something that happens during the operation of a business process. An activity represents a specific operation, task, or step in a business process. A gateway represents a branch, merge, or decision point in a process, and is used to control the flow of the process. The connection object represents the relationship or data flow between different elements in the process; The swimlanes indicate which participants are responsible for tasks or activities in the process; Step S3: convert the business process elements into corresponding business process json data; Use the large language model to resolve synonyms of business process elements and generate corresponding JSON code; Step S4, converting the json data into BPEL; Both json data and BPEL are formatted data. The conversion between json data and BPEL data is completed through the template that implements the formatted data conversion. Step S5, completing the visual business process diagram modeling; Use business process diagram modeling software to input json data and BPEL data to generate a visual business process diagram.

[0007] Beneficial Effects 1. Improve modeling efficiency: This invention combines a large language model for automation, which can effectively improve modeling efficiency compared to the traditional business process modeling method that requires professionals to manually perform complex process analysis and design; 2. Generate a complete business process diagram: Existing automated business process modeling technology has structural limitations and cannot effectively process business process elements such as intermediate events. It can only generate incomplete BPMN diagrams, and the modeling effect in most situations is not ideal. Based on the traditional NLP rules, the present invention constructs a new set of mapping rules from fact types to business process elements based on a large language model, and successfully represents a variety of business process elements. At the same time, using the json structure as an intermediate representation, more diverse information can be stored to support more complex business process modeling.

[0008] 3. Grammatical complexity processing: The present invention realizes the processing of texts containing complex grammar based on the large language model. Compared with the traditional NLP method, the large language model has stronger context understanding ability, model generalization ability, and automatic learning feature representation ability when processing complex sentences. It can further improve the performance through task-oriented fine-tuning, thus playing an important role in automated business process modeling.

[0009] 4. Synonym resolution: In business process modeling, performing synonym resolution can help improve the accuracy of the model, avoid ambiguity and information loss, and also help improve the effectiveness of subsequent tasks. The present invention constructs a synonym resolution method based on a large language model. Since the large language model learns a wide range of language contexts during the training process, it can better understand the context and relationships in the text. This enables them to accurately perform synonym resolution, especially in complex contexts. Since it is extremely difficult to perform synonym resolution using NLP technology, existing automated business process modeling technologies have not yet completed synonym resolution. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a flowchart of the rapid modeling of business processes based on a large language model of the present invention; Figure 2 is a flow chart of extracting fact types of the present invention; Figure 3 It is a text preprocessing flow chart of the present invention; Figure 4 is a flow chart of fact type extraction of the present invention; Figure 5 It is a genetic algorithm flow chart of the present invention; Figure 6 It is the logical mapping relationship between the fact type and the business process element of the present invention; Figure 7 The business process diagram generated by embodiment 1 of the present invention; Figure 8 The BPMN diagram generated by Example 1 of the present invention. DETAILED DESCRIPTION

[0011] In order to make those skilled in the art better understand the technical solutions in this application, the present invention is further described in detail below. However, it should not be understood that the scope of the above subject matter of the present invention is limited to the following embodiments, and all technologies realized based on the content of the present invention belong to the scope of the present invention.

[0012] like Figure 1 As shown, a method for rapid business process modeling based on a large language model of the present invention comprises the following steps: Step S1, extracting fact types existing in the business process description text; Fact types include entities, events, and event relations; The entity represents a specific object or participant involved in a business process, usually a subject or object in the process; an entity can be a person, organization, equipment, system or object, etc.; The event represents a specific action, operation or task that occurs in a business process, usually an action that occurs at a certain point in time or during a process, such as "order created", "payment completed", etc. The event relationship represents the logical connection or sequence between different events, and is usually used to describe the sequence of events, cause-effect relationships, dependency relationships, etc.

[0013] Step S2: construct the logical relationship between fact type and business process element; Business process elements include flow objects, connection objects, and swimlanes; the details are as follows: The flow object is the main graphical element for defining a business process, and includes three types: events, activities, and gateways. An event refers to something that happens during the operation of a business process. An activity represents a specific operation, task, or step in a business process. A gateway represents a branch, merge, or decision point in a process, and is used to control the flow of the process. The connection object represents the relationship or data flow between different elements in the process; The swimlanes indicate which participants are responsible for tasks or activities in the process.

[0014] Step S3: convert the business process elements into corresponding business process json data; Use the large language model to resolve synonyms of business process elements and generate corresponding JSON code; Step S4, converting the json data into BPEL; Both json data and BPEL are formatted data. The conversion between json data and BPEL data is completed through the template that implements the formatted data conversion. Step S5, completing the visual business process diagram modeling; Use business process diagram modeling software to input json data and BPEL data to generate a visual business process diagram.

[0015] like Figure 2 As shown, step S1 specifically includes: S1-1, preprocessing the input business process description text, using the large language model to split the complex sentences and compound sentences in the business process description text into separate simple sentences, and storing them as a simple sentence list; S1-2, extracting fact types from the preprocessing results; S1-3. Build an event relationship pool and use genetic algorithm to select the event relationship set with the highest fitness.

[0016] like Figure 3 As shown, step S1-1 specifically includes: S1-1-1, split the business process description text into single sentences; S1-1-2, input the single sentence into the large language model to determine whether it is a simple sentence, requiring only "yes" or "no" as the answer; S1-1-3. Construct a simple sentence list, divided into two cases: a) If the sentence is judged to be a simple sentence in S1-1-2, it is directly added to the simple sentence list; b) If the sentence is not a simple sentence, follow-up questions are required. The large language model is required to give the split simple sentence results, and steps S1-1-2 to S1-1-3 are repeated for each simple sentence obtained. In this round of questioning, it is necessary to emphasize that the sentence may not be a simple sentence, and only "yes" or "no" is required as the answer.

[0017] Divide the text into simple sentences and non-simple sentences for different question-answering processes, because simple sentences are easier to judge correctly and do not require subsequent prompts, while text judgments containing multiple simple sentences are more prone to errors and require further careful inspection and verification. Use redundant prompts to induce uncertainty when appropriate to encourage negative answers, which allows the model to reanalyze the text instead of reinforcing previous answers. Enforce a strict "yes / no" answer format to reduce uncertainty and make automation easier.

[0018] like Figure 4 As shown, step S1-2 specifically includes: Step S1-2-1, using the large language model to extract the entity set existing in each simple sentence in the simple sentence list obtained in S1-1; Step S1-2-2, the entity set extracted in step S1-2-1 is combined with the template as input again, and the event set existing in the text is chain-extracted by the large language model; Step S1-2-3, the event set extracted in step S1-2-2 is combined with the template as input again, and the event relationship set existing in the text is chain-extracted by the large language model; the extracted event relationships include sequential relationships, conditional relationships, parallel relationships, cyclic relationships, etc.

[0019] The template is a set of predefined prompt words designed manually, and is used to guide the large language model to further extract an event set or an event relationship set from the extracted entity set or event set.

[0020] The template consists of: 1) A description of the task to be completed; 2) Specification of specific output formats that large language models should adhere to; 3) A small number of input-output pair examples; 4) The entity set or event set extracted in the previous step.

[0021] Among them, 1)-3) are fixed in each input during chain extraction.

[0022] Step S1-2 decomposes the extraction task into three steps, and uses the information obtained in the previous step as the input of the data extraction in the next step, which helps to filter out non-existent element types, reduces the search space and computational complexity, and is conducive to extracting information.

[0023] like Figure 5 As shown, step S1-3 specifically includes: Step S1-3-1, initial population generation: construct an event relationship pool, use a large language model to predict several possible event relationship sets based on the input event set, combine them with the chain-extracted event relationship sets into an event relationship pool, and use event relationships as genes; Step S1-3-2, Adaptation Evaluation: Using an adaptation function that measures accuracy and comprehensiveness to evaluate each event relationship set helps to further evaluate and verify the results of large language model predictions and reduce the bias caused by hallucination problems.

[0024] Preferably, in order to enhance the reliability of the evaluation, the idea of ​​multi-agent is introduced, that is, the accuracy of the event relationship set is verified by another large language model. The model will make logical and semantic judgments on each event relationship and calculate the accuracy of the event relationship set as the adaptation function. The accuracy is specifically the ratio of the number of correct event relationships to the total number of event relationships; Step S1-3-3, selection: according to the evaluation result of the fitness function, select the individual with higher fitness as the parent solution and use it to generate the next generation solution; Step S1-3-4, crossover: Perform a genetic crossover operation on the newly generated solution to increase the availability of the solution space and introduce randomness based on the distribution probability of different types of event relations in the data set. Specifically, we adjust certain relations in the event relation set according to the frequency of different event relation types in the data set; Step S1-3-5, updating the population: replacing or merging the newly generated individuals with the parent population to form an updated pool of event relationship sets; Step S1-3-6, repeat steps S1-3-1 to S1-3-6, repeatedly perform selection and crossover operations, set a maximum number of iterations, and terminate the evolution process when the number of iterations is reached.

[0025] After the evolution process of the genetic algorithm, the event relationship set with the highest fitness is filtered out in the event relationship pool, and the obtained event relationship set has a high fitness value and quality for the summary of the logical relationship and key information of the business process.

[0026] Step S2: Construct the logical relationship between fact types and business process elements.

[0027] The present invention designs a set of logical relationships based on the extraction results of the large language model, thereby constructing mapping rules from predefined fact types to business process elements. The construction of the rules adopts semantic mapping rules, that is, mapping the semantic information in the natural language text to the rules in the structured model or system.

[0028] like Figure 6 As shown, the logical relationship of step S2 specifically includes: Activity: refers to the specific tasks, operations or steps performed in the entire business process. It is highly related to the extracted events and can be directly mapped to the events. The activity name corresponds to the event type name; Event: refers to a specific time point or moment that occurs during process execution. In a business process diagram, there is usually one start event and one or more end events. Set the start event before the first activity and set the end event after each activity branch ends; Swimlane: used to assign related activities or tasks to specific participants or responsible parties. It is highly related to the executor of a single event and can be mapped to the event. The name of the swimlane corresponds to the name of the event executor. Connection object: used to describe the relationship and process sequence between different activities, events and gateways in the process. It is highly related to the business process sequence, and the event relationship between events and adjacent events is mapped; Gateway, used to control the branching, merging and decision-making of the process. It is highly related to the conditional relationship and can be mapped with the event relationship.

[0029] Step S3: Convert the business process elements into corresponding business process json data.

[0030] Step S3 specifically includes: Step S3-1: for the attribute "ETSubject" of the event (Event) obtained in step S1, use the large language model to resolve the data; the specific operation is as follows: Step S3-1-1, for each event extracted in step S1-2-2, a resolution queue is established according to the attribute "ETSubject", where the resolution queue represents a queue waiting for same-referencing resolution; the length of each resolution queue is 1; the resolution queues corresponding to all events constitute a resolution queue list; Step S3-1-2, traverse the list of resolution queues starting from the second queue, and perform same-reference resolution with each of the previous queues in turn. The previous queue is regarded as a target queue during same-reference resolution, and is used to determine whether it belongs to the same entity as the current resolution queue. The judgment method is as follows: input the full text of the business process description in natural language, the content of the current resolution queue (the queue length is 1), and the content of the target queue into the large language model, and use the large language model's understanding of the text and semantic information to determine whether the current resolution queue and the target queue belong to the same entity.

[0031] Step S3-1-3: If the current resolution queue and the target queue belong to the same entity, add the elements contained in the current resolution queue to the end of the target queue and delete the current resolution queue; if the current resolution queue and the target queue do not belong to the same entity, repeat step S3-1-2 until the same-pointer resolution is completed.

[0032] Step S3-1-4, finally there are several queues left, the name of the "ETSubject" at the head of each queue replaces the name of the "ETSubject" of all events under this queue, and the name of the "ETSubject" at the head of each queue is used as the name of a lane.

[0033] Step S3-2: Generate json codes corresponding to business process elements. There are four types in total, including events, lanes, connection objects, and gateways. The details are as follows: Event: record the event name and event argument. Event argument consists of elements such as "event subject" and "event object". The format is as follows: "Event":[ { "ETId": "", " ETName " : " " , "ETSubject": " ", "ETObject": "", "ETTime" : " ", "ETPlace": "", } ] Among them, ETid records the event code; ETName records the event name, which is a combination of the event type and the event object; ETSubject records the event subject, which is the entity extracted in step 1; ETObject records the event object, which is the entity extracted in step 1; ETTime records the time when the event occurs. If it does not exist, it is represented by "NULL"; ETPlace records the location of the event, and if it does not exist, it is represented by "NULL"; Lane: Records the name of the event subject "ETSubject" in all events "Event". The format is as follows: "Swimlane": { " SLId " : " " , “ SLName ” : “ ” } ] Among them, SLId records the lane code; SLName records the lane name, which is consistent with the name of ETSubject in the corresponding Event; Connection object: records the name of the event or gateway that has a logical relationship. The format is as follows: "Connection":[ { "COId": " ", " COSource " : " " " COTarget " : " " } ] Among them, COId records the connection object code; COSource records the source endpoint of the connection object. If the source endpoint is an event, COSource is consistent with the name of ETId in the corresponding Event. If the source endpoint is a gateway, COSource is consistent with the name of GWId in the corresponding Gateway. COTarget records the target endpoint of the connection object. If the source endpoint is an event, COTarget is consistent with the name of ETId in the corresponding Event. If the source endpoint is a gateway, COTarget is consistent with the name of GWId in the corresponding Gateway.

[0034] Gateway: records the event relationships between events that are not sequential or cyclic. The format is as follows: “Gateway”: { "GWId": "", “GWType” : “ ” } ] Among them, GWId records the gateway code; GWType records the gateway type.

[0035] Step S4: Convert the json data into BPEL.

[0036] Step S5: Complete the visual business process diagram modeling; Two key elements are needed to complete the visual business process diagram modeling: a standardized language for defining business process behaviors and a graphical language for defining graphical symbols and their relationships. The former is the business process execution language BPEL, which is obtained through step S4. The latter is preferably obtained by using Camunda Modeler business process modeling software to convert json data into BPMN Diagram Interchange (bpmndi) format, which is used to describe the visual attributes of BPMN graphical elements (such as events, activities, gateways, etc.), such as position, size, and bounding box. After BPEL and bpmndi are input into the business process diagram modeling software, a visual business process diagram can be generated.

[0037] Example 1 Using the method of the present invention, the only business process public data set PET ( https: / / huggingface.co / datasets / patriziobellan / PET / viewer) Documentation 1.1 to be processed.

[0038] Document 1.1 reads as follows: A small company makes custom bicycles. Whenever the sales department receives an order, A new process instance is created. Members of the sales department can then reject or accept the order for the custom bike. In the latter case, the warehouse and engineering are notified. The warehouse processes the order immediately. The parts list is created and the quantity required for each part is checked. If the part is available in-house, it is reserved. If not, it is deferred This process is repeated for each item on the parts list. Meanwhile, the engineering department prepares the ordered bicycles. If the warehouse successfully orders or backorders every item on the parts list and the preparation activities are complete, Once the product is complete, the engineering department assembles the bicycle. After that, the sales department ships the bicycle to the customer and completes the process instance. ” (Translated from English to Chinese using translation software for demonstration purposes) Step S1, using the big model to extract the fact types existing in the business process description text, and obtain the corresponding entity set, event set, and event relationship set; The entity set is [Sales Department, Sales Department, Sales Department, Sales Department, Warehouse, Warehouse, Warehouse, Warehouse, Engineering Department, Engineering Department, Sales Department, Sales Department]; The set of events is [order received, order rejected, order accepted, warehouse and engineering notified, parts list processed, parts reserved, backordered, repeat for each item on parts list, prepare bike for assembly, assemble bike, ship bike to customer, process instance completed]; The set of event relations is [sequential, exclusive, sequential, sequential, sequential, exclusive, sequential, sequential, sequential, sequential, sequential, sequential].

[0039] 1) The template for the chain extraction of event sets in step S1-2-2 is as follows: “ ###Description of the task to be completed You are an expert in business process modeling and are familiar with common process structures such as exclusive choices, redo loops, etc. Your task is to analyze the textual description of the process and capture the business process elements in it. You will now be provided with a business process description. The entity set consists of the text and each entity, and you need to extract the events corresponding to each entity.

[0040] The entity represents the specific object or participant involved in the business process, usually the subject or object in the process. Entity; an entity can be a person, organization, equipment, system or object, etc. The event represents a specific action, operation or task that occurs in a business process, usually at a certain point in time or Actions that occur during the process, such as "order created", "payment completed", etc. ###Specification of specific output formats You need to return the model as an array, like this: [event1, event2, ..., eventn].

[0041] Do not reply with anything else.

[0042] The events provided need to correspond one-to-one with entities at the same location.

[0043] ### Input-Output Pair Example Here are some examples to learn from: Text: Customer brings in a defective computer, CRS checks the defect and sends back the repair cost calculation. If the customer If the cost is determined to be acceptable, the process continues, otherwise she will take the computer home intact. Repairs being done include Two activities, which can be performed in any order. The first activity is to check and repair the hardware, while the second activity is to check and configure After each activity, proper system functionality is tested. If an error is detected, another optional repair is performed. otherwise the repair is complete.

[0044] Entity Set: [customer,CRS,CRS,customer,customer,CRS,CRS,CRS,CRS,CRS] Reply: [Bring defective computer, check failure, distribute repair cost calculation, continue processing, bring computer back Check and repair hardware, check and configure software, test proper system functionality, perform any other repair activity, repair [Complete] ###Target text and entity set obtained in the previous step Now give the business process text description as follows: Target text: Entity Set: ” 2) The template for chain extraction of event relationship sets in step S1-2-3 is as follows: “ ###Description of the task to be completed You are an expert in business process modeling and are familiar with common process structures such as exclusive choices, redo loops, etc. Your task is to analyze the textual description of the process and capture the business process elements in it. You will now be provided with a business process description. The entity set consists of the text and each entity and event. You need to extract the event relationship between each two events. The entity represents the specific object or participant involved in the business process, usually the subject or object in the process. Entity; an entity can be a person, organization, equipment, system or object, etc. The event represents a specific action, operation or task that occurs in a business process, usually at a certain point in time or Actions that occur during the process, such as "order created", "payment completed", etc. The event relationship represents the logical connection or sequence between different events, usually used to describe the precedence of events. Post-order, causality, dependency, etc.

[0045] ###Specification of specific output formats You need to return the model in the form of an array, like this: [event-relationship1, event-relationship2,..., event-relationship n-1]。

[0046] Do not reply with anything else.

[0047] The event relationship provided represents the relationship between an event at the same location and the next event.

[0048] ### Input-Output Pair Example Here are some examples to learn from: Text: Customer brings in a defective computer, CRS checks the defect and sends back the repair cost calculation. If the customer If the cost is determined to be acceptable, the process continues, otherwise she will take the computer home intact. Repairs being done include Two activities, which can be performed in any order. The first activity is to check and repair the hardware, while the second activity is to check and configure After each activity, proper system functionality is tested. If an error is detected, another optional repair is performed. otherwise the repair is complete.

[0049] Entity Set: [customer,CRS,CRS,customer,customer,CRS,CRS,CRS,CRS,CRS] Event set: [Bring defective computer, check failure, distribute repair cost calculation, continue processing, bring computer Go home, check and repair hardware, check and configure software, test proper system functionality, perform any other repair activities, Repair Completed] Reply: [sequence, sequence, sequence, exclusive, sequence, sequence, sequence, sequence, exclusive] ###Target text and entity set obtained in the previous step Now give the business process text description as follows: Target text: Entity Set: Event Set: ” Step S2, constructing the logical relationship between fact type and business process element; The logical relationship between fact types and business process elements is predefined and does not change with different tasks.

[0050] Step S3 converts the business process elements into corresponding business process json data, and the json data corresponding to the above content can be obtained as follows: event: "Event":[ { "ETId": "E1", " ETName " : " Order Received " , “ETSubject”: “Sales Department”, "ETObject": "Order", "ETTime": "NULL", "ETPlace": "NULL", }, { "ETId": "E2", "ETName":"Reject Order", “ETSubject”: “Sales Department”, "ETObject": "Order", "ETTime": "NULL", "ETPlace": "NULL", }, { "ETId": "E3", "ETName":"Accept Order", “ETSubject”: “Sales Department”, "ETObject": "Order", "ETTime": "NULL", "ETPlace": "NULL", }, { "ETId": "E4", "ETName": "Notify warehouse and engineering department", “ETSubject”: “Sales Department”, "ETObject": "Warehouse and Engineering Department", "ETTime": "NULL", "ETPlace": "NULL", }, { "ETId": "E5", "ETName":"Process Parts List", "ETSubject": "warehouse", "ETObject": "Parts List", "ETTime": "NULL", "ETPlace": "NULL", }, { "ETId": "E6", "ETName": "Retain Part", "ETSubject": "warehouse", "ETObject": "Part", "ETTime": "NULL", "ETPlace": "NULL", }, { "ETId": "E7", " ETName " : " Backorder Parts " , "ETSubject": "warehouse", "ETObject": "Part", "ETTime": "NULL", "ETPlace": "NULL", }, { "ETId": "E8", "ETName": "Repeat for each item on the parts list", "ETSubject": "warehouse", "ETObject": "Each item on the parts list", "ETTime": "NULL", "ETPlace": "NULL", }, { "ETId": "E9", " ETName " : " Ready to assemble bicycle " , “ETSubject”: “Engineering Department”, "ETObject": "Bicycle", "ETTime": "NULL", "ETPlace": "NULL", }, { "ETId": "E10", " ETName " : " Assemble bicycle " , “ETSubject”: “Engineering Department”, "ETObject": "Bicycle", "ETTime": "NULL", "ETPlace": "NULL", }, { "ETId": "E11", " ETName " : " Deliver bicycles to customers " , “ETSubject”: “Sales Department”, "ETObject": "Customer", "ETTime": "NULL", "ETPlace": "NULL", }, { "ETId": "E12", "ETName":"Complete process instance", “ETSubject”: “Sales Department”, "ETObject": "NULL", "ETTime": "NULL", "ETPlace": "NULL", } ] Lanes: "Swimlane": { " SLId " : " S1 " , “ SLName ” : “ Sales Department ” }, { " SLId " : " S2 " , “ SLName ” : “ Warehouse” }, { " SLId " : " S3 " , “ SLName ” : “ Engineering Department” } ] Connection object: "Connection":[ { "COId": "C1", “ COSource ” : “ E1 ” “ COTarget ” : “ G1 ” }, { "COId": "C2", " COSource " : " G1 " “ COTarget ” : “ E2 ” }, { "COId": "C3", " COSource " : " G1 " “ COTarget ” : “ E3 ” }, { "COId": "C4", " COSource " : " E3 " “ COTarget ” : “ E4 ” }, { "COId": "C5", " COSource " : " E4 " “ COTarget ” : “ E5 ” }, { “COId”: “C6”, “COSource” : “E5” “COTarget” : “G2” }, { “COId” : “C7”, “COSource” : “G2” “COTarget” : “E6” }, { “COId” : “C8”, “COSource” : “G2” “COTarget” : “E7” }, { “COId” : “C9”, “COSource” : “E7” “COTarget” : “G3” }, { “COId”: “C10”, “COSource” : “E8” “COTarget” : “G3” }, { “COId”: “C11”, “COSource” : “G3” “COTarget” : “E9” }, { “COId”: “C12”, “COSource” : “E9” " COTarget " : " E10 " }, { “COId”: “C13”, " COSource " : " E10 " "COTarget":"E11" }, { "COId": "C14", " COSource " : " E11 " "COTarget":"E12" }, ] Gateway “Gateway”: { "GWId": "G1", "GWType": "Exclusive" }, { "GWId": "G2", "GWType": "Exclusive" }, { "GWId": "G3", "GWType": "Exclusive" } ] Step S4, converting the json data into BPEL; Step S5: Complete the visual business process diagram modeling.

[0051] Figure 7 This is a business process diagram drawn according to document 1.1.

[0052] Figure 8 This is the actual BPMN diagram corresponding to document 1.1.

[0053] The recall of connected objects and events is calculated by referring to the evaluation indicators used in the 2024 paper "Large Language Models Can Accomplish Business Process Management Tasks" (https: / / doi.org / 10.1007 / 978-3-031-50974-2_34). The manually annotated business process elements on the dataset PET are used as groundtruth to calculate the recall of connected objects and events.

[0054] The recall of the connected object is calculated as follows: The recall of an event is calculated as follows: The recall rates of the method of the present invention and the other six methods are compared as shown in the following table: Among them, the My_LLM column is the result of the present invention, with a connection object recall rate of 73.7% and an event recall rate of 100%. The present invention compares two popular methods in the field of business processes. The column named Friedrich is selected from the author Friedrich's paper "Process Model Generation from Natural Language Text". This method is based on traditional NLP and uses techniques such as syntactic analysis, anaphora resolution, and semantic role labeling to identify the relationship between these entities and construct a BPMN diagram; the five columns named Michael are selected from the author Michael's paper "Large Language Models CanAccomplish Business Process Management Tasks". This method is based on the large language model prompt project, and the predefined prompt words are input into the large language model to obtain business process elements and construct a BPMN diagram. Michael 1, Michael 2, and Michael 3 are the results obtained by repeatedly inputting the same prompt word. Michael 4 and Michael 5 are the results obtained by inputting different prompt words.

[0055] From the comparison data, it can be seen that after comparison, the business process element extraction recall rate of the present invention is higher and has better performance.

[0056] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions only describe the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention, and these changes and improvements fall within the scope of the present invention claimed.

Claims

1. A method for rapid business process modeling based on a large language model, characterized in that: The following steps are involved: Step S1, extracting fact types existing in the business process description text; Fact types include entities, events, and event relations; The entity represents a specific object or participant involved in the business process and is the subject or object of the process; The event represents a specific action, operation or task that occurs in a business process; The event relationship represents the logical connection or sequence between different events, and is used to describe the sequence, causal relationship, and dependency relationship of events. Step S2: construct the logical relationship between fact type and business process element; Business process elements include flow objects, connection objects, and swimlanes; the details are as follows: The flow object is the main graphical element for defining a business process, and includes three types: events, activities, and gateways. An event refers to something that happens during the operation of a business process. An activity represents a specific operation, task, or step in a business process. A gateway represents a branch, merge, or decision point in a process, and is used to control the flow of the process. The connection object represents the relationship or data flow between different elements in the process; The swimlanes indicate which participants are responsible for tasks or activities in the process; Step S3: convert the business process elements into corresponding business process json data; Use the large language model to resolve the same meanings of the business process elements and generate the JSON code corresponding to the business process elements; Step S4, converting the json data into BPEL; Both json data and BPEL are formatted data. The conversion between json data and BPEL data is completed through the template that implements the formatted data conversion. Step S5, completing the visual business process diagram modeling; Use business process diagram modeling software to input json data and BPEL data to generate a visual business process diagram.

2. A method for rapid business process modeling based on a large language model as claimed in claim 1, characterized in that: Step S1 specifically includes: S1-1, preprocessing the input business process description text, using the large language model to split the complex sentences and compound sentences in the business process description text into separate simple sentences, and storing them as a simple sentence list; S1-2, extracting fact types from the preprocessing results; S1-3. Build an event relationship pool and use genetic algorithm to select the event relationship set with the highest fitness.

3. A method for rapid business process modeling based on a large language model as claimed in claim 2, characterized in that: Step S1-1 specifically includes: S1-1-1, split the business process description text into single sentences; S1-1-2, input the single sentence into the large language model to determine whether it is a simple sentence, requiring only "yes" or "no" as the answer; S1-1-3. Construct a simple sentence list, divided into two cases: a) If the sentence is judged to be a simple sentence in S1-1-2, it is directly added to the simple sentence list; b) If the sentence is judged not to be a simple sentence, follow-up questions are required. The large language model is required to give the split simple sentence results, and steps S1-1-2 to S1-1-3 are repeated for each simple sentence obtained; in this round of questioning, it is necessary to emphasize that the sentence may not be a simple sentence, and only "yes" or "no" is required as the answer.

4. A method for rapid business process modeling based on a large language model as claimed in claim 2, characterized in that: Step S1-2 specifically includes: Step S1-2-1, using the large language model to extract the entity set existing in each simple sentence in the simple sentence list obtained in S1-1; Step S1-2-2, the entity set extracted in step S1-2-1 is combined with the template as input again, and the event set existing in the text is chain-extracted by the large language model; Step S1-2-3, the event set extracted in step S1-2-2 is combined with the template as input again, and the event relationship set existing in the text is chain-extracted by the large language model; the extracted event relationships include sequential relationships, conditional relationships, parallel relationships, and cyclic relationships.

5. A method for rapid business process modeling based on a large language model as claimed in claim 4, characterized in that: The template is a set of predefined prompt words designed manually, which is used to guide the large language model to further extract an event set or an event relationship set from the extracted entity set or event set; The template consists of: 1) A description of the task to be completed; 2) Specification of specific output formats that large language models should adhere to; 3) A small number of examples of input-output pairs; 4) The entity set or event set extracted in the previous step; Among them, 1)-3) are fixed in each input during chain extraction.

6. A method for rapid business process modeling based on a large language model as claimed in claim 2, characterized in that: Step S1-3 specifically includes: Step S1-3-1, initial population generation: construct an event relationship pool, use a large language model to predict several possible event relationship sets based on the input event set, combine them with the chain-extracted event relationship sets into an event relationship pool, and use event relationships as genes; Step S1-3-2, Adaptation Evaluation: Using an adaptation function that measures accuracy and comprehensiveness to evaluate each event relationship set helps to further evaluate and verify the results predicted by the large language model and reduce the bias caused by the hallucination problem; Step S1-3-3, selection: according to the evaluation result of the fitness function, select the individual with higher fitness as the parent solution and use it to generate the next generation solution; Step S1-3-4, Crossover: Perform a genetic crossover operation on the newly generated solution to increase the availability of the solution space by introducing randomness; Perform a genetic crossover operation on the newly generated solution to increase the availability of the solution space, and introduce randomness based on the distribution probability of different types of event relations in the data set. Specifically, we adjust certain relations in the event relation set according to the frequency of different event relation types in the data set; Step S1-3-5, updating the population: replacing or merging the newly generated individuals with the parent population to form an updated pool of event relationship sets; Step S1-3-6, repeating steps S1-3-1 to S1-3-6, repeatedly performing selection and crossover operations until the termination condition is met; After the evolution process of the genetic algorithm, the event relationship set with the highest fitness is filtered out in the event relationship pool, and the obtained event relationship set has a high fitness value and quality for the summary of the logical relationship and key information of the business process.

7. A method for rapid business process modeling based on a large language model as claimed in claim 6, characterized in that: In step S1-3-2, the idea of ​​multi-agent is introduced, that is, the accuracy of the event relationship set is verified through another large language model. The model makes logical and semantic judgments on each event relationship and calculates the accuracy of the event relationship set as an adaptation function; the accuracy is specifically the ratio of the number of correct event relationships to the total number of event relationships.

8. A method for rapid business process modeling based on a large language model as claimed in claim 1, characterized in that: Step S2: Constructing the logical relationship between fact types and business process elements is as follows: designing a set of logical relationships based on the extraction results of the large language model, thereby constructing the mapping rules of predefined fact types to business process elements; the construction of the rules adopts semantic mapping rules, that is, mapping the semantic information in the natural language text to the rules in the structured model or system; The logical relationship specifically includes: Activity: refers to the specific tasks, operations or steps performed in the entire business process; it is highly related to the extracted events and can be directly mapped to the events. The activity name corresponds to the event type name; Event: refers to a specific time point or moment that occurs during process execution. In a business process diagram, a start event is set before the first activity, and an end event is set after each activity branch ends. Swimlane: used to assign related activities or tasks to specific participants or responsible parties; highly related to the executor of a single event, can be mapped with the event, and the name of the swimlane corresponds to the name of the event executor; Connection object: used to describe the relationship and process sequence between different activities, events and gateways in the process; highly related to the business process sequence, and mapping of event relationships between events and adjacent events; Gateway, used to control the branching, merging and decision-making of the process. It is highly related to the conditional relationship and can be mapped with the event relationship.

9. A method for rapid business process modeling based on a large language model as claimed in claim 1, characterized in that: Step S3 specifically includes: Step S3-1: for the attribute "event subject" ETSubject of the event Event obtained in step S1, use the large language model to perform homonymy resolution on the data; the specific operations are as follows: Step S3-1-1, for each event extracted in step S1-2-2, a resolution queue is established according to the attribute "ETSubject", where the resolution queue represents a queue waiting for same-referencing resolution; the length of each resolution queue is 1; the resolution queues corresponding to all events constitute a resolution queue list; Step S3-1-2, starting from the second queue, traverse the list of resolution queues, and perform same-reference resolution on the current resolution queue and each of the previous queues in turn. The previous queue is regarded as a target queue during same-reference resolution, and is used to determine whether the current resolution queue belongs to the same entity; the determination method is as follows: input the full text of the business process description in natural language, the content contained in the current resolution queue, and the content contained in the target queue into the large language model, and use the large language model's understanding of the text and semantic information to determine whether the current resolution queue and the target queue belong to the same entity; Step S3-1-3, if the current resolution queue and the target queue belong to the same entity, add the elements contained in the current resolution queue to the end of the target queue and delete the current resolution queue; if the current resolution queue and the target queue do not belong to the same entity, repeat step S3-1-2 until the same-reference resolution is completed; Step S3-1-4, finally there are several queues left, the name of "ETSubject" at the head of each queue replaces the name of "ETSubject" of all events under this queue, and the name of "ETSubject" at the head of each queue is used as the name of a lane; Step S3-2: Generate json codes corresponding to business process elements, which are four types: events, lanes, connection objects, and gateways.

10. A method for rapid business process modeling based on a large language model as claimed in claim 9, characterized in that: The json code corresponding to the business process element is as follows: Event: record the event name and event argument; event argument consists of "event subject" and "event object" and other elements; the format is as follows: "Event":[ { "ETId": "", " ETName " : " " , "ETSubject" : "" , "ETObject" : "" , "ETTime" : "" , "ETPlace" : "" , } ] Among them, ETid records the event code; ETName records the event name, which is a combination of the event type and the event object; ETSubject records the event subject, which is the entity extracted in step 1; ETObject records the event object, which is the entity extracted in step 1; ETTime records the time when the event occurs. If it does not exist, it is represented by "NULL"; ETPlace records the location of the event. If it does not exist, it will be represented by "NULL"; Lane: Records the name of the event subject "ETSubject" in all events "Event"; the format is as follows: "Swimlane": { " SLId " : " " , " SLName " : " " } ] Among them, SLId records the lane code; SLName records the lane name, which is consistent with the name of ETSubject in the corresponding Event; Connection object: records the name of the event or gateway that has a logical relationship; the format is as follows: "Connection":[ { "COId" : "" , " COSource " : " " " COTarget " : " " } ] Among them, COId records the connection object code; COSource records the source endpoint of the connection object. If the source endpoint is an event, COSource is consistent with the name of ETId in the corresponding Event; if the source endpoint is a gateway, COSource is consistent with the name of GWId in the corresponding Gateway; COTarget records the target endpoint of the connection object. If the source endpoint is an event, COTarget is consistent with the name of ETId in the corresponding Event; if the source endpoint is a gateway, COTarget is consistent with the name of GWId in the corresponding Gateway. Gateway: records the event relationships between events that are not sequential or cyclical; the format is as follows: "Gateway":[ { " GWId " : " " , "GWType" : "" } ] Among them, GWId records the gateway code; GWType records the gateway type.

Images