A method and system for intent acquisition and business processing in charging scenarios based on a large model.
By fine-tuning the basic large language model and optimizing the dataset, the problem of accuracy in recognizing user intent in highway toll scenarios was solved, and the efficiency of business processing under special circumstances was improved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING MICROVIDEO TECH
- Filing Date
- 2024-05-31
- Publication Date
- 2026-07-31
AI Technical Summary
In highway toll collection scenarios, existing technologies struggle to accurately identify user intent, especially in special circumstances (such as when there is no toll card or the card is faulty), leading to low business processing efficiency.
By fine-tuning the basic large language model, utilizing the business dataset of highway toll stations and specific tools, the model is optimized to identify user intent, and combined with the configuration of prompt words, more scenario-appropriate responses are generated.
It achieves accurate user intent recognition in highway toll scenarios, especially under special circumstances, thereby improving business processing efficiency and toll collection speed.
Smart Images
Figure CN118708683B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the technical field of natural language processing, specifically to a method and system for obtaining intent and processing business in a charging scenario based on a large model. Background Technology
[0002] With the acceleration of global economic integration and urbanization, transportation systems, as a crucial component of national infrastructure, play a vital role in connecting cities, driving economic development, and facilitating social exchange. However, the transportation industry faces increasingly severe challenges, primarily including traffic congestion, environmental pollution, energy consumption, traffic safety, and service quality improvement. Therefore, exploring and applying advanced information technologies, aiming for intelligence, greening, and efficiency, has become a significant trend in the current development of the transportation industry. Intelligent Transportation Systems (ITS) integrate information processing, communication, sensor, and automatic control technologies to achieve comprehensive monitoring, management, and optimization of the transportation system. ITS encompasses traffic information collection, processing, and dissemination, as well as traffic flow control, traveler information services, and public transportation management, aiming to improve road efficiency, reduce traffic accidents, lessen environmental burden, and provide the public with a safer, more convenient, and comfortable travel experience. The application of big data and artificial intelligence enables traffic management departments to collect and analyze massive amounts of traffic data, including vehicle operation data, passenger travel data, and traffic incident data, processing them efficiently to provide data support for traffic planning, operation management, and accident prevention. Therefore, in the transportation sector, it is necessary to address the special circumstances of highway toll stations and propose intent recognition methods to obtain the user's true and accurate intent.
[0003] In the field of artificial intelligence, the rise of Large Language Models (LLMs) represents a major leap forward in natural language processing technology, demonstrating unprecedented capabilities and potential in numerous application scenarios. Large models, especially large-scale pre-trained language models, have become a hot topic in AI research and application. By learning from the vast amount of text data available on the internet, they not only acquire rich linguistic knowledge but also capture subtle differences and complex structures in language, thus exhibiting near-human or even surpassing performance in tasks such as understanding, generation, and dialogue. The core technical framework of large models is typically based on a "pre-training + fine-tuning" paradigm. First, the model undergoes self-supervised learning on massive amounts of unlabeled text data, pre-training by predicting words in a mask, the next sentence's content, or sentence continuity. The goal of this stage is to allow the model to learn the general rules and structures of language. Subsequently, for specific tasks such as question answering, translation, and sentiment analysis, the pre-trained model is fine-tuned, quickly achieving high performance using only a small amount of labeled data, greatly improving the model's generalization ability and adaptability. The development of large-scale models is not only a technological breakthrough, but also a reshaping of the overall research paradigm and application ecosystem in the field of artificial intelligence, leading the future trend of natural language processing and even the entire AI industry. General-purpose large-scale models cannot handle the special business of the transportation sector; therefore, a business model specifically designed for handling toll collection scenarios is needed.
[0004] In summary, the in-depth development of the transportation sector, combined with large-scale model technology, addresses the handling of toll collection, particularly in the acquisition of user intent in toll collection scenarios under special circumstances, thereby improving the efficiency of business processing. Summary of the Invention
[0005] To achieve a more accurate understanding of user intent and accelerate the efficiency of toll collection in highway tolling scenarios, this application provides a method and system for intent acquisition and business processing in tolling scenarios based on a large model.
[0006] Firstly, this application provides a method for obtaining intent and processing business in a charging scenario based on a large model, including:
[0007] Based on the toll collection scenarios at highway toll stations, the business types are organized to determine the business of highway toll stations, including normal toll collection scenarios and special toll collection scenarios.
[0008] During the business processing of various tolling scenarios at highway toll stations, the corpus of manual business processes is collected, organized, and a fine-tuning dataset is generated. The tools to be called are defined in the fine-tuning dataset.
[0009] The basic large language model was selected as the prototype model for fine-tuning. The basic large language model was fine-tuned using the fine-tuning dataset to obtain the fine-tuned basic large language model.
[0010] After fine-tuning, prompt words are configured on the fine-tuned basic large language model to complete the optimization of the fine-tuned basic large language model and obtain the optimized basic large language model.
[0011] The system acquires the questions and answers from each round of dialogue and adds them to the history list of the optimized basic large language model. It then analyzes the contents of the history list using the optimized basic large language model to obtain dialogue information and the results obtained by calling tools based on the dialogue information. Finally, it combines the dialogue information and the results obtained by calling tools based on the dialogue information to identify user intent and output business responses based on user intent.
[0012] By adopting the above scheme, effective data consisting of dialogue data from various business scenarios at highway toll stations and the result data obtained by calling tool functions based on the dialogue data is generated to fine-tune the pre-trained basic language model. Combined with the configuration of prompt words, the basic large language model is further optimized, so that the optimized basic large language model can better recognize user intent in various business scenarios at highway toll stations and output more scenario-appropriate answers.
[0013] Preferred options also include:
[0014] The question and answer of each round of dialogue are obtained as a dialogue data and stored in the fine-tuning dataset. Whenever the number of new dialogue data entries in the fine-tuning dataset reaches the preset number, the basic large language model is fine-tuned again using the latest fine-tuning dataset to obtain the fine-tuned basic language model.
[0015] Obtain and compare the satisfaction evaluation results of the business response content output by the re-fine-tuned base language model and the fine-tuned base language model. If the satisfaction of the business response content output by the re-fine-tuned base language model is higher than that of the fine-tuned base language model, then the re-fine-tuned base language model will be used to replace the fine-tuned base language model.
[0016] By adopting the above scheme, the dialogue information generated in each round is obtained to continuously expand and fine-tune the dataset, thereby enabling continuous fine-tuning and optimization of the basic large language model to identify more accurate user intent.
[0017] Preferably, the special circumstances include: no payment card or a faulty payment card;
[0018] The tools to be called in the fine-tuning dataset include: tools for querying and recognizing license plate images for special toll scenarios, tools for querying vehicle information corresponding to license plate images for special toll scenarios, and tools for calculating vehicle toll information corresponding to vehicle information for special toll scenarios.
[0019] By adopting the above scheme, and considering the handling of special situations in highway tolling scenarios, such as when there is no card or the card is faulty, it often involves processing methods such as obtaining license plate information and billing based on license plate information. Corresponding calling methods are set to supplement the fine-tuning dataset and improve the comprehensiveness of the fine-tuning dataset.
[0020] Preferably, the specific process of fine-tuning the basic large language model using the fine-tuning dataset includes:
[0021] Set and initialize the fine-tuning parameters, which include: the path to the base large language model, the path and name of the fine-tuning dataset, the save path of the base large language model after fine-tuning, the fine-tuning strategy, the batch size, the maximum length of tokens, the gradient accumulation steps, the learning rate, the learning scheduler, the number of training rounds, and the maximum number of samples.
[0022] The model is trained based on the set LORA fine-tuning strategy and fine-tuning dataset; among them, the LORA fine-tuning technique is used to perform a low-rank approximation of the weight matrix of the large model, and the fine-tuning update formula includes:
[0023] W′=W+U·V T
[0024] In the formula, W∈R m×n Let U and V represent the weight matrix, where U and V represent the transposes of E and F, respectively, and E∈R. m×k and F∈R k×n For the two smaller matrices introduced, k << min(m,n) denotes a lower rank;
[0025] Determine whether the loss function value of the basic large language model has reached the preset loss value. If it has, then select to end the fine-tuning. The loss function value of the basic large language model is a weighted sum of the loss function values of the training set and the validation set.
[0026] By adopting the above scheme, the LORA fine-tuning strategy is selected to fine-tune the large model to save storage and computing resources; the fine-tuning update is completed by transposing two low-rank matrices to better conform to the matrix multiplication rules.
[0027] Preferably, after the fine-tuning is completed, the prompt word configuration for the fine-tuned basic large language model includes: identity configuration, scene configuration, business type processing method, speech recognition result correction method, province name abbreviation table, license plate color judgment method, and final output result example.
[0028] By adopting the above approach, the large model is optimized using prompt word configuration to generate more accurate responses, enhance the progression of the dialogue, and achieve faster recognition of user intent.
[0029] Preferably, the number of dialogue data entries recorded in the history list of the optimized basic large language model can be set according to the business type of the highway toll station; wherein, the number of matches set for normal tolling scenarios is less than the number of matches set for special case handling scenarios.
[0030] By adopting the above solution, considering that there will be more rounds of dialogue with users in special circumstances compared to normal circumstances, the recording capacity of the history list is adjusted accordingly to accommodate different scenarios.
[0031] Preferred options also include:
[0032] If there are conflicting dialogue data between the questions and results of multiple rounds of question-and-answer sessions, the question and result of the latest round of questions will be used as the standard, and the conflicting content between the question and result of the latest round of questions will be deleted from the history list of the optimized basic large language model.
[0033] By adopting the above solution, considering that users may make mistakes in their answers during the dialogue, leading to contradictions between the questions and results, contradictory data is deleted from the history list to prevent affecting the accuracy of subsequent user intent recognition.
[0034] Secondly, this application provides a system for obtaining and processing payment scenarios based on a large model, including:
[0035] The business determination module for highway toll stations is used to organize business types based on the tolling scenarios of highway toll stations and determine the business of highway toll stations, including normal tolling scenario business and special tolling scenario business.
[0036] The business fine-tuning dataset acquisition module for highway toll stations is used to collect and organize the corpus of manual business processes during the business processing of various tolling scenarios at highway toll stations, generate fine-tuning datasets, and define the tools to be called in the fine-tuning datasets.
[0037] The module for obtaining the fine-tuned basic large language model is used to select the basic large language model as the fine-tuning prototype model, use the fine-tuning dataset to fine-tune the basic large language model, and obtain the fine-tuned basic large language model.
[0038] The optimized basic large language model acquisition module is used to configure prompt words on the fine-tuned basic large language model after fine-tuning, thereby completing the optimization of the fine-tuned basic large language model and obtaining the optimized basic large language model.
[0039] The user intent and business response acquisition module is used to acquire the questions and answers of each round of dialogue and add them to the history list of the optimized basic large language model. The optimized basic large language model is used to analyze the contents of the history list to obtain dialogue information and the results information obtained by calling tools based on the dialogue information. The module combines the dialogue information and the results information obtained by calling tools based on the dialogue information to identify user intent and output business responses based on user intent.
[0040] By adopting the above approach, effective data from toll station scenarios can be collected, and the data can be fully utilized for model training and optimization. This results in an optimized model with excellent response speed to identify user intent and quickly generate accurate answers to process toll collection.
[0041] Thirdly, this application provides a computer-readable storage medium including a stored computer program, wherein the computer program, when running, controls the device where the computer-readable storage medium is located to perform the method described above.
[0042] Fourthly, this application provides a computer device, the computer device including a memory, a processor and a program stored in the memory and executable thereon, the program being executed by the processor to implement the steps of the method described above.
[0043] In summary, this application has the following beneficial effects:
[0044] 1. The basic language model is fine-tuned and trained using a fine-tuning dataset with access tools. Combined with configured prompts, the fine-tuned basic language model is further optimized to obtain an optimized basic language model. This optimized model is then used to accurately identify user intent and provide professional responses from the model in two major business scenarios: normal and special toll collection on highways. This accelerates toll collection and improves toll collection speed.
[0045] 2. Supplement the fine-tuning dataset with dialogue data completed through the optimized base language model to achieve continuous fine-tuning of the base language model. Attached Figure Description
[0046] Figure 1 This is a flowchart of the method for obtaining and processing the charging scenario based on a large model, as described in a specific embodiment.
[0047] Figure 2 This is a flowchart illustrating the actual usage process of the method for obtaining and processing the charging scenario based on the large model, as described in a specific embodiment.
[0048] Figure 3 This is a schematic diagram of the system for obtaining and processing payment scenarios based on a large model, as described in a specific embodiment. Detailed Implementation
[0049] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0050] like Figure 1 As shown in the figure, this application discloses a method for obtaining intent and processing business in a charging scenario based on a large model. The specific steps are as follows:
[0051] S1. Organize the business types according to the toll collection scenarios of highway toll stations to determine the business types of highway toll stations.
[0052] Specifically, based on the toll collection scenarios at highway toll stations, the corresponding business is categorized, defining the business types as two categories: normal toll collection scenarios and special circumstances toll collection scenarios. Special circumstances include: no toll card or a faulty toll card (damaged card). No toll card generally refers to a lost card or failure to collect the card at the entrance, while a faulty card refers to a toll card that cannot be recognized.
[0053] For special charging scenarios, the business processing generally follows a standard process. If the card is lost after being taken at the entrance, the lost card will be recorded; if the card is damaged, the card information will be recorded; finally, the license plate recognition method is called to query the vehicle entrance information, the billing method is called, and the actual cost incurred is queried.
[0054] S2. Obtain dialogue data during the business processing of various toll scenarios at highway toll stations and organize it to generate a fine-tuning dataset.
[0055] Specifically, during the business processing of various tolling scenarios at highway toll stations, the corpus of human business processes is collected and organized. That is, the dialogue data of human toll collectors during business processing in specific tolling scenarios is collected and organized, and a fine-tuning dataset is generated as a training dataset according to the question-and-answer correspondence.
[0056] Some parts of the fine-tuning dataset require the use of specific tools. Therefore, the tools to be used need to be defined within the fine-tuning dataset. Each tool has a corresponding function, and the result of that function is stored in the fine-tuning dataset with the key "observation". For example, if a question requires the use of a specific method, the answer is set as "observation". The answer is then obtained based on the observation and historical messages. For instance, using the query function corresponding to a tool for querying vehicle information corresponding to a license plate image, the query function and historical vehicle information data can be used to retrieve the specific vehicle information (license plate number) corresponding to the current image. Specifically, defining the tools to be used in the fine-tuning dataset includes: defining tools for license plate image query and recognition for special toll scenarios; defining tools for querying vehicle information corresponding to license plate images for special toll scenarios; and defining tools for calculating vehicle toll information corresponding to vehicle information for special toll scenarios.
[0057] S3. Select the basic large language model as the fine-tuning prototype model, and use the fine-tuning dataset to fine-tune the basic large language model to obtain the fine-tuned basic large language model.
[0058] Specifically, the process of fine-tuning the basic large language model using the fine-tuning dataset includes:
[0059] The fine-tuning parameters are set and initialized, including: the path to the base large language model, the path and name of the fine-tuning dataset, the save path to the base large language model after fine-tuning, the fine-tuning strategy, the batch size, the maximum length of tokens, the gradient accumulation steps, the learning rate, the learning scheduler, the number of logging steps, the weight saving interval steps, the number of training epochs, the maximum number of samples, and the quantization bit width. In this embodiment, the LORA fine-tuning strategy is selected, the target layer of the fine-tuning application is W_pack, the maximum length of tokens is 1024, the gradient accumulation steps are 8, the learning rate scheduler is cosine, the logging steps are 10, and the weight saving interval steps are 100. The total number of training epochs is 500, the maximum number of samples is 3000, and the quantization bit width is 4 bits.
[0060] The model is trained based on the set LORA fine-tuning strategy and fine-tuning dataset. Specifically, LORA fine-tuning is used to approximate the weight matrix of the large model with a low rank. Instead of directly updating the entire matrix, LORA introduces two smaller matrices. The update formula during fine-tuning includes:
[0061] W ′ =W + U·V T
[0062] In the formula, W∈Rm×n Let U and V represent the weight matrix, where U and V represent the transposes of E and F, respectively, and E∈R. m×k and F∈R k×n Let U and V be two smaller matrices, k << min(m,n) representing a low rank. In practical applications, U and V are small parameter matrices learned through backpropagation, used to fit task-related weight changes. In the context of LoRA, U and V are subject to regularization constraints to guarantee their low-rank properties, and they typically maintain a small rank k during training, enabling efficient and resource-saving fine-tuning.
[0063] Determining whether the fine-tuned model has reached its optimal state during fine-tuning requires monitoring multiple metrics during training and observing how the loss function changes with the number of iterations or epochs. Ideally, the training loss should gradually decrease as training progresses, indicating that the model is learning and adapting to the training data. Overfitting should be avoided; if the training loss continues to decrease while the validation loss begins to rise or stabilizes, this may be an indication of overfitting. Specifically, it is determined whether the loss function value of the base language model has reached the preset loss value. If so, fine-tuning is terminated. The loss function (Loss) value of the base language model is a weighted average of the training set loss function (Train Loss) value and the validation set loss function (Validation Loss) value.
[0064] After fine-tuning, the model is deployed locally, using an NVIDIA L40S graphics card with 48GB of video memory to load the model, and corresponding interface services are developed to continuously monitor performance and maintain operations.
[0065] S4. After fine-tuning, configure prompt words for the fine-tuned basic large language model to complete the optimization of the fine-tuned basic large language model and obtain the optimized basic large language model.
[0066] Specifically, the prompt word configuration includes: identity configuration, scenario configuration, business type processing method, speech recognition result correction method, province name abbreviation table, license plate number rule acquisition method, license plate color judgment method, and final output result sample.
[0067] S5. For business dialogue content, use the optimized basic large language model to identify user intent and output business responses based on user intent.
[0068] Specifically, in actual use, the questions and answers of each round of dialogue are acquired and added to the history list of the optimized basic large language model. The optimized basic large language model is then used to analyze the history content to obtain dialogue information and the results obtained by calling tools based on the dialogue information. User intent is identified by combining the dialogue information and the results obtained by calling tools based on the dialogue information, and a business answer is output according to the user intent. In the first round of dialogue, the answer is generated by using the optimized basic large language model to identify the user intent of the question and outputting a business answer according to the user intent. In each subsequent round, the optimized basic large language model is used to analyze the history content to generate the answer.
[0069] The optimized basic large language model's history list records the number of questions and answers for each round of dialogue, and the matching quantity can be set according to the business type of the highway toll station. Considering that the special case handling scenario is more complex and involves more rounds of dialogue, the matching quantity set for the normal tolling scenario is less than the matching quantity set for the special case handling scenario.
[0070] like Figure 2 As shown, the method described in this application enables user intent recognition and generates business responses for special charging scenarios during actual use, thereby accelerating business processing in special charging scenarios, as detailed below:
[0071] Speech recognition is performed on dialogues in special charging scenarios to obtain dialogue content; business processing records are collected to fine-tune the basic large language model;
[0072] Set system prompts, define the identity of the large model, set the dialogue scenario, and define the business logic that the large model needs to process. Create a license plate province abbreviation table, the final answer template, and obtain the optimized basic large language model.
[0073] By analyzing the dialogue content of the speech recognition results using the optimized basic large language model, dialogue information can be obtained. If a tool is available, the dialogue information and the result information (license plate number) obtained by calling the tool based on the dialogue information can be obtained. In addition, based on the dialogue information, the license plate information described in the dialogue content can be corrected using the configured speech recognition result correction method, such as correcting "hook" to "J". According to the license plate number matching rules of regular expressions, the license plate contains a combination of province abbreviation, letters, numbers, and special characters arranged according to the rules. The corrected license plate information is then combined to obtain the correct license plate number.
[0074] Using the optimized basic large language model to comprehensively judge and analyze the license plate information and conversation information obtained from multi-round Q&A, complete the recognition of the user's business intention, and generate corresponding business processing answers to complete the business processing of special situation charging scenarios.
[0075] In a specific embodiment, considering the continuous optimization of the large prediction model, in order to improve the accuracy of model intention recognition, the method further includes:
[0076] Obtain the questions and answers of each round of conversation as a piece of conversation data and store them in the fine-tuning dataset; whenever the number of newly added conversation data in the fine-tuning dataset reaches the preset number, use the latest fine-tuning dataset to retrain the basic large language model to obtain the retrained basic language model; in this embodiment, the preset number is set to 500.
[0077] Obtain and compare the satisfaction evaluation results of the business answers output by the retrained basic language model and the fine-tuned basic language model. If the satisfaction of the business answers output by the retrained basic language model is higher than that of the fine-tuned basic language model, use the retrained basic language model to replace the fine-tuned basic language model to generate responses for the next round of conversation.
[0078] In a specific embodiment, considering that there may be contradictory content in the process of the user providing information, which affects the accuracy of subsequent intention understanding and answer content generation, in order to improve the accuracy of model intention recognition, the method further includes:
[0079] Traverse the questions and results of multi-round Q&A stored in the history list of the optimized basic large language model. If there are contradictory data in the conversation data corresponding to the questions and results of multi-round Q&A, such as: the voice content in the first round of conversation is: license plate number "Beijing F·Lxxx8", and the second round of conversation content is "Beijing F·Lxxx0", then take the questions and results of the latest round of questions as the standard, and delete the content in the history list of the optimized basic large language model that conflicts with the questions and results of the latest round of questions.
[0080] As Figure 3 shown, an embodiment of the present application discloses a system for obtaining intention and business processing in a charging scenario based on a large model, including:
[0081] The business determination module 101 of the highway toll station is used to sort out the business types according to the charging scenarios of the highway toll station, and determine the business of the highway toll station, including normal charging scenario business and special situation charging scenario business;
[0082] The business fine-tuning dataset acquisition module 102 of the highway toll station is used to collect the data of manual business processes during the business processing of various toll scenarios at the highway toll station, organize and generate fine-tuning datasets, and define the tools to be called in the fine-tuning datasets.
[0083] The fine-tuned basic large language model acquisition module 103 is used to select the basic large language model as the fine-tuning prototype model, use the fine-tuning dataset to fine-tune the basic large language model, and obtain the fine-tuned basic large language model.
[0084] The optimized basic large language model acquisition module 104 is used to configure prompt words for the fine-tuned basic large language model after fine-tuning, thereby completing the optimization of the fine-tuned basic large language model and obtaining the optimized basic large language model. The user intent and business response acquisition module 105 is used to acquire the questions and answers of each round of dialogue and add them to the history list of the optimized basic large language model. The optimized basic large language model is used to analyze the content of the history to obtain dialogue information and the result information obtained by calling the tool based on the dialogue information. The user intent is identified by combining the dialogue information and the result information obtained by calling the tool based on the dialogue information, and the business response is output according to the user intent.
[0085] The fine-tuned basic large language model acquisition module 103 is also used to acquire the questions and answers of each round of dialogue as a dialogue data and store it in the fine-tuning dataset; whenever the number of new dialogue data entries in the fine-tuning dataset reaches a preset number, the basic large language model is fine-tuned again using the latest fine-tuning dataset to obtain the fine-tuned basic language model; the satisfaction evaluation results of the business answer content output by the fine-tuned basic language model and the business answer content output by the fine-tuned basic language model are acquired and compared. If the satisfaction of the business answer content output by the fine-tuned basic language model is higher than that of the business answer content output by the fine-tuned basic language model, the fine-tuned basic language model is used to replace the fine-tuned basic language model.
[0086] This application also discloses a computer-readable storage medium.
[0087] Specifically, the computer-readable storage medium stores a computer program that can be loaded by a processor and executed as described above regarding the method for obtaining intent and processing business in a charging scenario based on a large model. The computer-readable storage medium includes, for example, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0088] This application also discloses a computer device.
[0089] Specifically, the computer device includes a memory and a processor. The memory stores computer programs that can be loaded by the processor and executed to perform the aforementioned methods for obtaining intent and processing business in a charging scenario based on a large model.
[0090] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.
Claims
1. A method for intent acquisition and business processing in a charging scenario based on a large model, characterized in that, include: Based on the toll collection scenarios at highway toll stations, the business types are organized to determine that highway toll station businesses include normal toll collection scenarios and special toll collection scenarios. During the business processing of various tolling scenarios at highway toll stations, the corpus of manual business processes is collected, organized, and a fine-tuning dataset is generated. The tools to be called are defined in the fine-tuning dataset. The basic large language model was selected as the prototype model for fine-tuning. The basic large language model was fine-tuned using the fine-tuning dataset to obtain the fine-tuned basic large language model. After fine-tuning, prompt words are configured on the fine-tuned basic large language model to complete the optimization of the fine-tuned basic large language model and obtain the optimized basic large language model. The questions and answers of each round of dialogue are obtained and added to the history list of the optimized basic large language model. The optimized basic large language model is used to analyze the contents of the history list to obtain dialogue information and the results obtained by calling tools based on the dialogue information. By combining dialogue information and the results obtained from calling tools based on the dialogue information, user intent is identified, and business responses are output according to the user intent.
2. The method for obtaining intent and processing business in a charging scenario based on a large model according to claim 1, characterized in that, Also includes: The question and answer for each round of dialogue are obtained as a dialogue data point and stored in the fine-tuning dataset; Whenever the number of new dialogue data entries in the fine-tuning dataset reaches a preset number, the basic large language model is fine-tuned again using the latest fine-tuning dataset to obtain the fine-tuned basic language model. Obtain and compare the satisfaction evaluation results of the business response content output by the re-fine-tuned base language model and the fine-tuned base language model. If the satisfaction of the business response content output by the re-fine-tuned base language model is higher than that of the fine-tuned base language model, then the re-fine-tuned base language model will be used to replace the fine-tuned base language model.
3. The method for obtaining intent and processing business in a charging scenario based on a large model according to claim 1, characterized in that, The special circumstances include: no payment card or a faulty payment card; The tools to be called in the fine-tuning dataset include: tools for querying and recognizing license plate images for special toll scenarios, tools for querying vehicle information corresponding to license plate images for special toll scenarios, and tools for calculating vehicle toll information corresponding to vehicle information for special toll scenarios.
4. The method for obtaining intent and processing business in a charging scenario based on a large model according to claim 1, characterized in that, The specific process of fine-tuning the basic large language model using the fine-tuning dataset includes: Set and initialize the fine-tuning parameters, which include: the path to the base large language model, the path and name of the fine-tuning dataset, the save path of the base large language model after fine-tuning, the fine-tuning strategy, the batch size, the maximum length of tokens, the gradient accumulation steps, the learning rate, the learning scheduler, the number of training rounds, and the maximum number of samples. The model is trained based on the set LORA fine-tuning strategy and fine-tuning dataset; among them, the LORA fine-tuning technique is used to perform a low-rank approximation of the weight matrix of the large model, and the update formula during fine-tuning includes: W′=W+U·V T In the formula, W∈R m×n Let U and V represent the weight matrix, where U and V represent the transposes of E and F, respectively, and E∈R. m×k and F∈R k×n For the two smaller matrices introduced, k << min(m,n) denotes a lower rank; Determine whether the loss function value of the basic large language model has reached the preset loss value. If it has, then select to end the fine-tuning. The loss function value of the basic large language model is a weighted sum of the loss function values of the training set and the validation set.
5. The method for obtaining intent and processing business in a charging scenario based on a large model according to claim 1, characterized in that, After fine-tuning, the basic large language model after fine-tuning is configured with prompt words, including: identity configuration, scene configuration, business type processing method, speech recognition result correction method, province name abbreviation table, license plate color judgment method, and final output result sample.
6. The method for obtaining intent and processing business in a charging scenario based on a large model according to claim 2, characterized in that, The number of dialogue data entries recorded in the history list of the optimized basic large language model can be set according to the business type of the highway toll station; the number of matches set for normal tolling scenarios is less than the number of matches set for special case handling scenarios.
7. The method for obtaining and processing payment scenarios based on a large model according to claim 1, characterized in that, Also includes: If there are conflicting dialogue data between the questions and results of multiple rounds of question-and-answer sessions, the question and result of the latest round of questions will be used as the standard, and the conflicting content between the question and result of the latest round of questions will be deleted from the history list of the optimized basic large language model.
8. A system for acquiring and processing payment scenarios based on a large model, characterized in that, include: The business determination module for highway toll stations is used to organize business types based on the tolling scenarios of highway toll stations, and to determine whether the business of highway toll stations includes normal tolling scenarios and special tolling scenarios. The business fine-tuning dataset acquisition module for highway toll stations is used to collect and organize the corpus of manual business processes during the business processing of various tolling scenarios at highway toll stations, generate fine-tuning datasets, and define the tools to be called in the fine-tuning datasets. The module for obtaining the fine-tuned basic large language model is used to select the basic large language model as the fine-tuning prototype model, use the fine-tuning dataset to fine-tune the basic large language model, and obtain the fine-tuned basic large language model. The optimized basic large language model acquisition module is used to configure prompt words on the fine-tuned basic large language model after fine-tuning, thereby completing the optimization of the fine-tuned basic large language model and obtaining the optimized basic large language model. The user intent and business response acquisition module is used to acquire the questions and answers of each round of dialogue and add them to the history list of the optimized basic large language model. The optimized basic large language model is used to analyze the contents of the history list to obtain dialogue information and the result information obtained by calling tools based on the dialogue information. By combining dialogue information and the results obtained from calling tools based on the dialogue information, user intent is identified, and business responses are output according to the user intent.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the method as described in any one of claims 1 to 7.
10. A computer device, characterized in that, The computer device includes a memory, a processor, and a program stored in and executable on the memory, the program being executed by the processor to implement the steps of the method as described in any one of claims 1 to 7.