An RPA process scheduling method based on large model fine tuning

Abstract

The application discloses an RPA process scheduling method based on large model fine tuning, comprising the following steps: constructing an RPA process knowledge base, fine tuning an intention classification model and an entity recognition model, collecting user instructions, inputting the user instructions into the fine-tuned intention classification model for classification, obtaining corresponding intentions, inputting the corresponding intentions into the fine-tuned entity recognition model and the RPA process knowledge base respectively, extracting relevant entities and matching corresponding processes to obtain entity extraction results and process matching results; and inputting the entity extraction results and the process matching results into the RPA to execute the RPA process. Through the organic combination of the double model, the RPA knowledge base architecture and the RPA process, the application realizes accurate recognition of user instructions and flexible scheduling of the RPA process.

Description

Technical Field

[0001] This invention relates to the field of RPA process scheduling, and in particular to an RPA process scheduling method based on large model fine-tuning. Background Technology

[0002] Typical LLMs, such as generative pre-trained transform models (GPT series) and bidirectional encoder representation models (BERT series), utilize their core multi-head self-attention mechanism to process and model long-range dependencies between any two elements in an input sequence in parallel. These models typically follow a "pre-training + fine-tuning" paradigm. First, they are pre-trained on a large-scale unlabeled corpus with a general language modeling objective (such as autoregression or autoencoder) to form a general language representation foundation. Then, for specific downstream tasks (such as text classification, question answering, and summary generation), the model parameters are fine-tuned using labeled task-related data to adapt them to the needs of specific application scenarios.

[0003] Robotic Process Automation (RPA) is an application software technology that automates numerous, repetitive, rule-based business processes by configuring software "robots" to simulate and integrate human user interactions on digital system interfaces. These RPA robots can capture data, trigger responses, operate applications, process transactions, and communicate across multiple software systems, functioning like a virtual workforce. Its core value lies in achieving seamless automation at the user interface (UI) layer without requiring deep API integration or underlying code modifications to existing enterprise information systems (such as ERP, CRM, and OA), effectively improving the efficiency, accuracy, and compliance of business processes.

[0004] However, traditional RPA technology also has inherent limitations. Its automation capabilities heavily rely on predefined, structured rules and input data. When faced with unstructured data (such as free text, emails, and complex documents), abnormal situations requiring subjective judgment, or process changes, traditional RPA robots lack sufficient flexibility and intelligent decision-making capabilities, often requiring human intervention and causing the automated process to be interrupted. In other words, traditional RPA excels at performing "hand" tasks but lacks the cognitive and judgmental functions of a "brain." Therefore, how to combine the efficient execution capabilities of RPA with the capabilities of large models to handle complex, unstructured business processes and achieve end-to-end intelligent automation has become an important technological development direction in this field. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings and deficiencies of the prior art and provide an RPA process scheduling method based on large model fine-tuning.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] An RPA process scheduling method based on large model fine-tuning includes the following steps:

[0008] S1. Based on the developed and packaged RPA process and the RPA process storage information, construct an RPA process knowledge base;

[0009] S2. Fine-tune the intent classification model and entity recognition model. The fine-tuning process consists of four steps: data collection, data augmentation, data annotation, and model fine-tuning. Data collection uses data generation models A, B, and C. Data generation model A generates initial user command data, while data generation models B and C rewrite and optimize the initial user command data, respectively. Model fine-tuning uses a base model to fine-tune the intent classification model and entity recognition model. Finally, the fine-tuned intent classification model and entity recognition model are obtained.

[0010] S3. Collect user commands, input the user commands into the fine-tuned intent classification model for classification, obtain the corresponding intent, input the corresponding intent into the fine-tuned entity recognition model and RPA process knowledge base respectively, extract relevant entities and match the corresponding process, obtain entity extraction results and process matching results; input the entity extraction results and process matching results into RPA, and execute the RPA process.

[0011] Step S1 specifically involves: developing and encapsulating a series of parameterized RPA processes to complete specific tasks. Each RPA process has a unique identifier and defines the input parameters required for the RPA process. The storage information of each RPA process includes the process name, process function, process ID, and input parameters. The input parameters include the parameter name, parameter description, and parameter type. Based on the developed and encapsulated RPA processes and their storage information, an RPA process knowledge base is constructed. In this knowledge base, each record corresponds to a user intent, and each record maps a user intent to one or more processes. The entity to be extracted under the user intent is defined as the set of input parameters for all RPA processes under the user intent.

[0012] In step S2, the fine-tuning of the intent classification model and entity recognition model is divided into four parts: data collection, data augmentation, data annotation, and model fine-tuning.

[0013] The data collection method involves using training data from the same batch of user command data for both the intent classification model and the entity recognition model. The generation of user instruction data adopts a multi-model collaborative generation strategy: three large-scale parameter models are used, namely data generation models A, B, and C. Data generation model A generates the initial user instruction data, while data generation models B and C rewrite and optimize the initial user instruction data, respectively. Finally, the highest quality data is retained through consistency screening. The retained highest quality data is packaged into a specific generation task using an instruction template, and structured JSON is output to obtain the user instruction data.

[0014] In step S2, the data enhancement involves synonym replacement, entity replacement, and back translation of the user command data. Synonym replacement replaces verbs and nouns in the user command data without changing the core semantics of the sentence and increases the diversity of language expression. Entity replacement maintains the original sentence structure and replaces the entity values ​​in the user command data. Back translation adopts a bidirectional translation process of "Chinese-English-Chinese", translating the original Chinese command data into English and then translating the English translation back into Chinese, ultimately generating enhanced user command data with different sentence structures but completely consistent semantics with the original text.

[0015] In step S2, the data annotation is based on the functional differences between the intent classification model and the entity recognition model, and different labels are used to annotate the enhanced user command data;

[0016] In the intent classification model, each augmented user command data point is labeled as ,in, This represents a single enhanced user command data. It is a discrete intent label. ,in It is a predefined finite set of tags Each enhanced user instruction data has one and only one label, resulting in user instruction data for the intent classification model;

[0017] The entity recognition model labels the slots explicitly provided in the enhanced user command data as individual enhanced user command data. For entity values, slots not explicitly provided in the enhanced user instruction data are marked as default values ​​to obtain the entity recognition model user instruction data.

[0018] The user command data of the intent classification model and entity recognition model are labeled and validated using a large-scale parameter data annotation model combined with command templates. The validation includes consistency verification and manual sampling. Consistency verification involves labeling the same input user command data of the intent classification model and entity recognition model three times, requiring the results to be completely consistent. Manual sampling involves randomly selecting 5% of the labeled user command data of the intent classification model and entity recognition model for manual verification. Finally, the labeled user command data of the intent classification model and entity recognition model are obtained.

[0019] In step S2, the model fine-tuning implementation steps are as follows: the intention classification model and the entity recognition model are fine-tuned using labeled user command data of the intention classification model and the entity recognition model respectively, and an open-source autoregressive language model with small parameters is selected as the base model; LoRA is used as the core technology for lightweight fine-tuning to complete model adaptation with extremely low parameter overhead.

[0020] Linear layer weights in the base model The forward propagation process changes to:

[0021] ;

[0022] in , , and rank Where h is the output of the base model, x is the labeled user command data from the intent classification model and entity recognition model, ΔW is the weight update during fine-tuning, k represents the dimension of the linear layer input feature vector x, d represents the dimension of the linear layer output feature vector h, and r is the shared dimension of the low-rank matrices B and A; during training, the linear layer weights... The data is frozen and does not participate in the update; the parameter values ​​of the low-rank matrices A and B are updated.

[0023] The intent classification model is fine-tuned by adding a classification head after the base model. The classification head is a trainable module attached to the output of the base model. When fine-tuning the intent classification model, the classification head inputs the labeled user command data of the intent classification model into the intent classification model through forward propagation. It extracts the hidden state at the end of the sequence of labeled user command data of the intent classification model as the overall semantic representation. The probability distribution of the intent label is calculated by the classification layer. The difference between the loss function and the true label is calculated. The parameters of the base model and the classification head are optimized simultaneously using the backpropagation algorithm to obtain the fine-tuned intent classification model.

[0024] Use standard cross-entropy loss as the loss function for the intent classification model:

[0025] ;

[0026] Where N is the number of labeled user command data for the intent classification model, and y i For real labels, To predict probabilities;

[0027] The entity recognition model is fine-tuned using a base model. Given user-instructed data labeled with the entity recognition model as input, the target JSON sequence is generated autoregressively. ,in, The total number of tokens in the target JSON sequence. Represents the first in the target JSON sequence One token; use a causal language to model the loss function and calculate the cross-entropy over the target JSON sequence of tokens:

[0028] ;

[0029] in, For the target JSON sequence Total number of tokens For the first JSON sequence in the target JSON sequence The real token at each location, and the Prompt is the user instruction data for the input labeled entity recognition model. Represents the position All previous tokens, This represents all trainable parameters in the entity recognition model. For entity recognition models in parameters Below, based on the given conditions Calculate the correct token The conditional probabilities of occurrence are used to obtain the fine-tuned entity recognition model.

[0030] Meanwhile, this invention provides:

[0031] A server includes a processor and a memory, the memory storing at least one program that is loaded and executed by the processor to implement the RPA process scheduling method for large-scale model fine-tuning described above.

[0032] A computer-readable storage medium storing at least one program, which is loaded and executed by a processor to implement the above-described RPA process scheduling method based on large model fine-tuning.

[0033] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0034] 1. This invention employs a dual-model structure to decouple the complex natural language instruction understanding process into two sub-tasks: intent classification and entity recognition. This avoids mutual interference from multi-task learning, allowing each model to focus on optimizing its own task, significantly improving the accuracy and reliability of the final results.

[0035] 2. This invention uses instruction templates for fine-tuning, enabling the model to directly output structured JSON objects. This method not only allows the model to handle complex semantic reasoning and missing value imputation, but also eliminates tedious post-processing steps, achieving end-to-end structured information generation and greatly improving the system's integrability and robustness.

[0036] 3. This invention employs LoRA (Low-Rank Adaptation) fine-tuning technology to fine-tune the model. By freezing most of the parameters of the original model and training only the injected very low-rank matrix, the number of trainable parameters is compressed to less than 1% of the original model. This technology makes it possible to quickly complete domain adaptation on consumer-grade GPUs, greatly lowering the barrier to entry for this invention.

[0037] 4. This invention, through the organic combination of dual-model + knowledge base architecture innovation and fine-tuning technology innovation, successfully injects the cognitive capabilities of large models into traditional RPA systems in a low-cost, high-efficiency, and high-security manner in limited scenarios, providing a complete and feasible technical solution for achieving truly intelligent, natural language interactive enterprise process automation.

[0038] 5. This invention offers high flexibility and low computational cost. It eliminates the need to write RPA scripts for every new instruction; simply training the LLM to understand new intents and entities allows it to invoke existing processes, significantly expanding the application boundaries of RPA. By employing a large model with a small number of parameters and combining it with fine-tuning, the hardware requirements for training and inference are greatly reduced. Enterprises do not need to purchase expensive AI computing infrastructure and can complete model customization and deployment locally.

[0039] 6. This invention boasts low maintenance costs and convenient scenario expansion. Its modular design allows for independent updates to the model, knowledge base, and RPA scripts. When business processes change, only the process knowledge base or RPA process needs updating; retraining the model is unnecessary, making maintenance simple and efficient. When adding new automation scenarios, the mapping relationship between new intents and RPA processes is registered in the process knowledge base, and the model is fine-tuned for the new scenario without altering the core system architecture. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating a system application example.

[0041] Figure 2 Flowchart for fine-tuning the model. Detailed Implementation

[0042] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0043] like Figure 1 As shown, an RPA process scheduling method based on large model fine-tuning includes the following steps:

[0044] S1. Based on the developed and packaged RPA process and the RPA process storage information, construct an RPA process knowledge base;

[0045] S2. Fine-tune the intent classification model and entity recognition model. The fine-tuning process consists of four steps: data collection, data augmentation, data annotation, and model fine-tuning. Data collection uses data generation models A, B, and C. Data generation model A generates initial user command data, while data generation models B and C rewrite and optimize the initial user command data, respectively. Model fine-tuning uses a base model to fine-tune the intent classification model and entity recognition model. Finally, the fine-tuned intent classification model and entity recognition model are obtained.

[0046] S3. Collect user commands, input the user commands into the fine-tuned intent classification model for classification, obtain the corresponding intent, input the corresponding intent into the fine-tuned entity recognition model and RPA process knowledge base respectively, extract relevant entities and match the corresponding process, obtain entity extraction results and process matching results; input the entity extraction results and process matching results into RPA, and execute the RPA process.

[0047] Step S1 specifically involves: developing and encapsulating a series of parameterized RPA processes to complete specific tasks. Each RPA process has a unique identifier (ID) and explicitly defines the input parameters (param) required for the RPA process. The storage information of each RPA process must include the process name (name), process function (description), process ID (ID), and input parameters (param). The input parameters (param) include the parameter name, parameter description, and parameter type. Based on the developed and encapsulated RPA processes and their storage information, an RPA process knowledge base is constructed. In this knowledge base, each record corresponds to a user intent, and each record maps a user intent (Intent) to one or more processes. The entity (Entity) to be extracted under the user intent is defined as the set of input parameters for all RPA processes under the user intent.

[0048] In step S2, such as Figure 2 As shown, the intent classification model ( ) and entity recognition model ( The fine-tuning section is divided into four parts: data collection, data augmentation, data annotation, and model fine-tuning.

[0049] The data collection method involves using training data from the same batch of user command data for both the intent classification model and the entity recognition model. The generation of user instruction data adopts a multi-model collaborative generation strategy: three large-scale parameter models are used, such as any common large-scale parameter models on the market, such as Doubao, Deepseek, Wenxin, and Hunyuan, which are data generation models A, B, and C, respectively. Among them, data generation model A generates the initial user instruction data, and data generation models B and C rewrite and optimize the initial user instruction data, respectively. Finally, the highest quality data is retained through consistency screening to obtain the user instruction data. The highest quality data is packaged into a specific generation task using an instruction template (Prompt Template) and output structured JSON to obtain the user instruction data.

[0050] In step S2, the data augmentation involves performing synonym replacement, entity replacement, and back translation on the user command data. Synonym replacement replaces verbs and nouns in the user command data without altering the core semantics of the sentence and increases the diversity of language expression. Entity replacement maintains the original sentence structure and replaces the entity values ​​in the user command data, enabling the entity recognition model to learn entity types rather than memorizing specific values. Back translation employs a bidirectional "Chinese-English-Chinese" translation process, translating the original Chinese command data into English and then translating the English translation back into Chinese, ultimately generating augmented user command data with different sentence structures but completely identical semantics to the original text.

[0051] In step S2, the data annotation is based on the functional differences between the intent classification model and the entity recognition model, and different labels are used to annotate the enhanced user command data;

[0052] In the intent classification model, each augmented user command data is labeled as ,in, This represents a single enhanced user command data. It is a discrete intent label. ,in It is a predefined finite set of tags Each enhanced user instruction data has one and only one label, resulting in user instruction data for the intent classification model;

[0053] The slot labels explicitly provided in the input data of the entity recognition model are single enhanced user command data. Entity value; Output data It is a strictly normalized JSON string that fully contains all predefined entity slots and entity values; for slots not explicitly provided in the enhanced user instruction data, they are marked as default values ​​(e.g., "not provided"), thus obtaining the entity recognition model user instruction data;

[0054] Large-scale parameter data annotation models, i.e., large models on the market, such as Qianwen 72B, with any parameter scale of 30B or more, are used to annotate and verify user command data of the intent classification model and entity recognition model in conjunction with command templates. The verification includes consistency verification and manual sampling. Consistency verification involves annotating the same input user command data of the intent classification model and entity recognition model three times, requiring the results to be completely consistent. Manual sampling involves randomly selecting 5% of the annotated user command data of the intent classification model and entity recognition model for manual verification. Finally, the annotated user command data of the intent classification model and entity recognition model are obtained.

[0055] In step S2, the model fine-tuning process involves fine-tuning the intent classification model and the entity recognition model using labeled user command data. Small-scale open-source autoregressive language models, such as Qwen2.5-1.5B, Qwen3-1.7B, and Qwen2.5-4B, are selected as the base model. LoRA (Low-Rank Adaptation) is used as the core technology for lightweight fine-tuning, enabling model adaptation with extremely low parameter overhead.

[0056] Base model Linear layer weights The forward propagation process changes to:

[0057] ;

[0058] in , , and rank Where h is the output of the base model, x is the labeled user command data from the intent classification model and entity recognition model, ΔW is the weight update during fine-tuning, k represents the dimension of the linear layer input feature vector x, d represents the dimension of the linear layer output feature vector h, and r is the shared dimension of the low-rank matrices B and A; during training, the linear layer weights... The data is frozen and does not participate in the update. The injected low-rank matrices A and B are optimized, and the parameter values ​​of low-rank matrices A and B are updated.

[0059] Intent classification model Fine-tuning was performed on the base model. A classification head is then added. This classification head is a trainable module attached to the output of the base model, mapping the text semantic representation generated by the base model to a specific intent category space. When fine-tuning the intent classification model, the classification head inputs the labeled user command data into the intent classification model via forward propagation. It extracts the hidden state at the end of the labeled user command data sequence as the overall semantic representation. The probability distribution of the intent labels is calculated by the classification layer, and the difference between the result and the true labels is calculated using a loss function. Finally, the parameters of both the base model and the classification head are optimized simultaneously using a backpropagation algorithm, resulting in the fine-tuned intent classification model.

[0060] Use standard cross-entropy loss as the loss function for the intent classification model:

[0061] ;

[0062] Where N is the number of labeled user command data for the intent classification model, and y i For real labels, To predict probabilities;

[0063] The entity recognition model is fine-tuned using a base model. Given user-instructed data labeled with the entity recognition model as input, the target JSON sequence is generated autoregressively. ,in, The total number of tokens in the target JSON sequence. Represents the first in the target JSON sequence One token; use a causal language to model the loss function and calculate the cross-entropy over the target JSON sequence of tokens:

[0064] ;

[0065] in:

[0066] : target JSON sequence Total length (number of tokens);

[0067] The first JSON sequence in the target JSON sequence The real token in each location;

[0068] Prompt: The input command Prompt;

[0069] : Represents the position All previous tokens, i.e. This is the meaning of "causality": the model predicts... At that time, it was impossible to see And subsequent information;

[0070] : Represents all trainable parameters in the model, namely the adapter parameters injected through LoRA and the weights of the language model head;

[0071] Model in parameters Below, based on the given conditions Calculate the correct token The conditional probability of occurrence;

[0072] The final entity recognition model was obtained after fine-tuning.

[0073] This invention supports adding new automation scenarios. New intents and their mapping relationships to the RPA process knowledge base are registered, and the intent classification and entity recognition models are fine-tuned for the new scenario without altering the core system architecture. Furthermore, when adding a new scenario, new scenario data is added to the existing dataset (old scenario) as a new dataset. The model fine-tuned using the new dataset (old scenario combined with new scenario) still supports the old scenario.

[0074] Meanwhile, this invention provides:

[0075] A server includes a processor and a memory, the memory storing at least one program that is loaded and executed by the processor to implement the RPA process scheduling method for large-scale model fine-tuning described above.

[0076] A computer-readable storage medium storing at least one program, which is loaded and executed by a processor to implement the above-described RPA process scheduling method based on large model fine-tuning.

[0077] The specific implementation of the present invention will be described in detail below using examples from Tencent Meeting.

[0078] S1: System Preset and Preparation

[0079] S11: RPA Process Development: Developed RPA processes such as Tencent Meeting_Quick Meeting, Tencent Meeting_Schedule Meeting, and Tencent Meeting_Join Meeting;

[0080] S12: Define Intents and Entities: Defines multiple intents such as Quick Meeting (create a new quick meeting), Schedule a Meeting (schedule a meeting for a future time), and Join a Meeting (join an existing meeting), and defines entity fields for different intents, such as: Quick Meeting (meeting name, invitees), Schedule a Meeting (meeting name, start date, start time, meeting duration, meeting frequency, etc.), and Join a Meeting (meeting name, meeting order).

[0081] S13: Based on the RPA process developed above and the defined intents and entity information, construct a process knowledge base. Each record in the knowledge base corresponds to an RPA process and contains complete process attribute information, as follows:

[0082] The "Quick Meeting" process is named "Quick Meeting" and its function is described as "creating a new quick meeting". The unique process ID assigned is 000000000001. The input parameters (param) of this process include two items: one is "Meeting Name", which is described as "the name of the quick meeting to be created" and is of string type; the other is "Invitees", which is described as "the names of the people to be invited to the meeting" and is of list type.

[0083] The "Schedule a Meeting" process is named "Schedule a Meeting," described as "Schedule a meeting for a future time," and assigned a unique process ID of 000000000002. The process's input parameters (param) include five items: "Meeting Name" (description: the name of the meeting to be scheduled, type: string), "Start Date" (description: the date the meeting starts, type: string), "Start Time" (description: the time the meeting starts, type: string), "Meeting Duration" (description: the duration of the meeting, type: string), and "Meeting Frequency" (description: the frequency at which the created meeting is repeated, type: string).

[0084] The “Join Meeting” process is named “Join Meeting” and its function is described as “Join an existing meeting”. The unique process ID assigned is 000000000003. The input parameters (param) of this process include two items: “Meeting Name” (description: the name of the meeting to be joined, type: string) and “Meeting Order” (description: the order of the meetings to be joined in the list, type: integer).

[0085] S2: Model Fine-tuning

[0086] S21: Data Collection: Use Doubao to generate raw user data, such as "Schedule a meeting for tomorrow morning", and use DeepSeek-V3 and Qwen2.5-72B to rewrite and optimize the data to get "Schedule a sales report meeting for me at 8 am tomorrow".

[0087] S22: Data Augmentation: Augment the original data, such as by replacing it with synonyms and entities, for example: "Schedule a product exchange meeting for me at 6 pm the day after tomorrow";

[0088] S23: Data Annotation: For the above data, use Qwen2.5-72B to apply the instructions "Extract user intent from user input, optional value: {process name}, function of each process is: {process function}" and "Extract parameters of the required scheduled meeting information from user input according to the following rules. Required field is {parameter name}, unprovided field value is 'not provided'", output in JSON format to obtain the annotated data;

[0089] S24: Fine-tuning: Fine-tuning the intent classification model and entity recognition model with the labeled data to obtain the fine-tuned intent classification model and entity recognition model;

[0090] S3: A system consisting of a fine-tuned intent classification model, a fine-tuned entity recognition model, a process knowledge base, and RPA processes.

[0091] S31: Input the user's command "Create a meeting for me now and invite Xiaoming" into the system;

[0092] S32: Input the user command into the fine-tuned intent classification model and obtain the output of the intent classification model as a quick meeting;

[0093] S33: Input the output of the user instruction and intent classification model into the fine-tuned entity recognition model, and obtain the entity recognition model output as: "{"Invited Person":"Xiaoming","Meeting Name":"Exchange Meeting"}";

[0094] S34: Match the “Quick Meeting” process in the process knowledge base based on the output of the intent classification model;

[0095] S35: Use the results of entity recognition model output combined with the results of process knowledge base matching as the execution parameters of RPA process, create a new quick meeting named Exchange Meeting and invite Xiaoming to join the meeting.

[0096] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. An RPA process scheduling method based on large model fine-tuning, characterized in that, Includes the following steps: S1. Based on the developed and packaged RPA process and the RPA process storage information, construct an RPA process knowledge base; S2. Fine-tune the intent classification model and entity recognition model. The fine-tuning process consists of four steps: data collection, data augmentation, data annotation, and model fine-tuning. Data collection uses data generation models A, B, and C. Data generation model A generates initial user command data, while data generation models B and C rewrite and optimize the initial user command data, respectively. Model fine-tuning uses a base model to fine-tune the intent classification model and entity recognition model. Finally, the fine-tuned intent classification model and entity recognition model are obtained. The implementation steps of the model fine-tuning are as follows: fine-tuning the intent classification model and the entity recognition model using labeled user command data, respectively, and selecting an open-source autoregressive language model with small parameters as the base model; using LoRA as the core technology for lightweight fine-tuning to complete model adaptation with extremely low parameter overhead. Linear layer weights in the base model The forward propagation process changes to: ； in , , and rank Where h is the output of the base model, x is the labeled user command data from the intent classification model and entity recognition model, ΔW is the weight update during fine-tuning, k represents the dimension of the linear layer input feature vector x, d represents the dimension of the linear layer output feature vector h, and r is the shared dimension of the low-rank matrices B and A; during training, the linear layer weights... The data is frozen and does not participate in the update; the parameter values ​​of the low-rank matrices A and B are updated. The intent classification model is fine-tuned by adding a classification head after the base model. The classification head is a trainable module attached to the output of the base model. When fine-tuning the intent classification model, the classification head inputs the labeled user command data of the intent classification model into the intent classification model through forward propagation. It extracts the hidden state at the end of the sequence of labeled user command data of the intent classification model as the overall semantic representation. The probability distribution of the intent label is calculated by the classification layer. The difference between the loss function and the true label is calculated. The parameters of the base model and the classification head are optimized simultaneously using the backpropagation algorithm to obtain the fine-tuned intent classification model. Use standard cross-entropy loss as the loss function for the intent classification model: ； Where N is the number of labeled user command data for the intent classification model, and y i For real labels, To predict probabilities; The entity recognition model is fine-tuned using a base model. Given user-instructed data labeled with the entity recognition model as input, the target JSON sequence is generated autoregressively. ,in, The total number of tokens in the target JSON sequence. Represents the first in the target JSON sequence One token; use a causal language to model the loss function and calculate the cross-entropy over the target JSON sequence of tokens: ； in, The total number of tokens in the target JSON sequence S. For the first JSON sequence in the target JSON sequence The real token in each location To input user instruction data for the labeled entity recognition model, Represents the position All previous tokens, This represents all trainable parameters in the entity recognition model. For entity recognition models in parameters Under the given conditions Calculate the correct token The conditional probabilities of occurrence are used to obtain the fine-tuned entity recognition model. S3. Collect user commands, input the user commands into the fine-tuned intent classification model for classification, obtain the corresponding intent, input the corresponding intent into the fine-tuned entity recognition model and RPA process knowledge base respectively, extract relevant entities and match the corresponding process, obtain entity extraction results and process matching results; input the entity extraction results and process matching results into RPA, and execute the RPA process.

2. The RPA process scheduling method based on large model fine-tuning according to claim 1, characterized in that, Step S1 specifically involves: developing and encapsulating a series of parameterized RPA processes to complete specific tasks. Each RPA process has a unique identifier and defines the input parameters required for the RPA process. The storage information of each RPA process includes the process name, process function, process ID, and input parameters. The input parameters include the parameter name, parameter description, and parameter type. Based on the developed and encapsulated RPA processes and their storage information, an RPA process knowledge base is constructed. In this knowledge base, each record corresponds to a user intent, and each record maps a user intent to one or more processes. The entity to be extracted under the user intent is defined as the set of input parameters for all RPA processes under the user intent.

3. The RPA process scheduling method based on large model fine-tuning according to claim 1, characterized in that, In step S2, the data collection involves the training data for both the intent classification model and the entity recognition model originating from the same batch of user instruction data. The generation of user instruction data adopts a multi-model collaborative generation strategy: three large-scale parameter models are used, namely data generation models A, B, and C. Data generation model A generates the initial user instruction data, while data generation models B and C rewrite and optimize the initial user instruction data, respectively. Finally, the highest quality data is retained through consistency screening. The highest quality data is packaged into a specific generation task using an instruction template, and structured JSON is output to obtain the user instruction data. The data augmentation involves synonym replacement, entity replacement, and back translation of user command data. Synonym replacement replaces verbs and nouns in the user command data without changing the core semantics of the sentence and increases the diversity of language expression. Entity replacement maintains the original sentence structure and replaces the entity values ​​in the user command data. Back translation adopts a bidirectional translation process of "Chinese-English-Chinese", translating the original Chinese command data into English and then translating the English translation back into Chinese, ultimately generating augmented user command data with different sentence structures but completely consistent semantics with the original text.

4. The RPA process scheduling method based on large model fine-tuning according to claim 1, characterized in that, In step S2, the data annotation is based on the functional differences between the intent classification model and the entity recognition model, and different labels are used to annotate the enhanced user command data; In the intent classification model, each augmented user command data point is labeled as ,in, This represents a single enhanced user command data. It is a discrete intent label. ,in It is a predefined finite set of tags Each enhanced user instruction data has one and only one label, resulting in user instruction data for the intent classification model; The entity recognition model labels the slots explicitly provided in the enhanced user command data as individual enhanced user command data. For entity values, slots not explicitly provided in the enhanced user instruction data are marked as default values ​​to obtain the entity recognition model user instruction data. The user command data of the intent classification model and entity recognition model are labeled and validated using a large-scale parameter data annotation model combined with command templates. The validation includes consistency verification and manual sampling. Consistency verification involves labeling the same input user command data of the intent classification model and entity recognition model three times, requiring the results to be completely consistent. Manual sampling involves randomly selecting 5% of the labeled user command data of the intent classification model and entity recognition model for manual verification. Finally, the labeled user command data of the intent classification model and entity recognition model are obtained.

5. A server, characterized in that, The server includes a processor and a memory, the memory storing at least one program, which is loaded and executed by the processor to implement the RPA process scheduling method based on large model fine-tuning as described in any one of claims 1 to 4.

6. A computer-readable storage medium, characterized in that, The storage medium stores at least one program, which is loaded and executed by a processor to implement the RPA process scheduling method based on large model fine-tuning as described in any one of claims 1 to 4.

Images