A large model-based tax appeal analysis method and system

Abstract

The present application relates to the tax information technology field, especially relates to a tax appeal analysis method and system based on a large model. The tax appeal analysis method based on a large model converts appeal dialogue voice into dialogue text, and pre-processes to obtain appeal text. Tax exclusive business data is collected in advance, a large model suitable for tax appeal analysis is obtained through multi-round adaptive training, the appeal text is automatically classified and labeled, the single appeal text is analyzed and content is condensed in combination with the Prompt engineering, and key information is automatically extracted. The data is uniformly collected and entered into a vector library. Cluster analysis forms a typical cause summary of high-frequency hot issues. The tax appeal analysis method and system based on a large model greatly reduce labor costs, improve hotline processing efficiency, reduce human error, facilitate quick retrieval, statistics and traceability, automatically mine hot issues and common causes, and provide reliable data support for tax service optimization, policy adjustment and process improvement.

Description

Technical Field

[0001] This invention relates to the field of tax information technology, and in particular to a method and system for analyzing tax claims based on a large model. Background Technology

[0002] Current methods and systems for handling tax hotline complaints generally employ a traditional approach, primarily relying on manual intervention supplemented by tools. During calls between tax hotline operators and taxpayers, operators manually record key points of the complaint while answering the call. After the call, they categorize the complaints according to their experience, extracting the content, causes, and suggested solutions for each complaint. Finally, the complaint data is manually compiled at fixed intervals, hot issues are identified, and their causes are summarized. The entire process is highly dependent on manual operation.

[0003] The current method of handling tax hotline requests has the following technical flaws: 1) Voice communication relies on manual transcription into text, which is prone to problems such as omissions, errors, and untimely recording, and the original information is easily distorted.

[0004] 2) The classification of demands relies on subjective judgment based on human experience, which has problems such as inconsistent standards, low efficiency and easy errors; at the same time, the general large model has not been trained with tax professional data, so the classification is inaccurate and the understanding is biased in tax scenarios, which cannot meet the requirements of high-precision analysis of tax business.

[0005] 3) The summary of the content, analysis of the causes and extraction of solutions for each request all rely on manual processing, which is labor-intensive, time-consuming and labor-intensive, and lacks the ability to process on a large scale.

[0006] 4) The collection of hot issues and their causes requires manual statistical screening, which results in problems such as long analysis cycles, coarse data granularity, insufficient objectivity of results, and poor timeliness.

[0007] To address the problems of high manual processing costs, low overall efficiency, poor consistency of processing results, inability to achieve real-time structured and intelligent analysis of demand data, and delayed discovery of hot issues in existing technologies, which make it difficult to support rapid decision-making and continuous optimization of tax services, this invention proposes a tax demand analysis method and system based on a large model. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies, this invention provides a simple and efficient method and system for analyzing tax claims based on a large model.

[0009] This invention is achieved through the following technical solution: A tax claim analysis method based on a large model includes the following steps: Step S1: Real-time voice acquisition and transcription The system collects real-time audio of conversations between taxpayers and tax hotline operators, converts the audio stream into real-time dialogue text with speaker role annotations through a real-time audio transcription module, and preprocesses it to obtain a request text that conforms to the custom specifications. In step S1, the dialogue text transcribed by the real-time speech transcription module is cleaned to remove redundant interjections and pause words, and retain the text that meets the custom specifications.

[0010] Step S2: Fine-tuning the large-scale tax claim model Tax-specific business data is collected in advance, processed to obtain a large model training dataset in Alpaca format, and the general large model is subjected to multiple rounds of domain-specific adaptation training to obtain a special large model suitable for tax claim analysis. In step S2, the general large model is trained by supervised fine-tuning (SFT) and combined with LoRA (Low-Rank Adaptation) fine-tuning to perform domain-specific multi-round adaptation training, thereby obtaining a large model specifically for tax claim analysis.

[0011] Step S3: Automatic classification and labeling of dimensional requirements Input the request text that conforms to the custom specifications into the fine-tuned large-scale model for tax request analysis, and the large-scale model for tax request analysis will automatically classify and label the request text. In step S3, the dedicated large model for tax complaint analysis automatically classifies and labels the complaint text in three dimensions: complaint type, tax type, and business matter.

[0012] Step S4: Analysis and Information Extraction of a Single Request Based on a dedicated large model for tax demand analysis, combined with the Prompt project, the text of a single demand is parsed and its content is condensed, automatically extracting three key pieces of information: the taxpayer's core demand, the reason for the demand, and the solution to the demand. Step S5: Gathering Demands and Uncovering the Causes of Hot Topics Taxpayer complaint data within a defined period is collected and entered into a vector library. A pre-trained large model is used to perform cluster analysis on the collected data to extract high-frequency hot issues within the corresponding time period. Then, vector retrieval is used to match and associate the corresponding causes of the complaints. The data is further summarized and refined by a dedicated large model for tax complaint analysis to form a summary of the typical causes of each hot issue. Finally, a structured analysis result that can support tax management decisions is output.

[0013] A tax complaint analysis system based on a large model consists of a real-time speech transcription module, a large model processing module, a multi-dimensional classification module, a single complaint parsing module, and a complaint collection and hotspot extraction module. The modules communicate with each other and work together. The real-time speech transcription module is responsible for collecting the speech of taxpayers and tax hotline operators in real time, converting the speech stream into speech text with speaker role labels in real time, and preprocessing it to obtain the request text that conforms to the custom specifications. The real-time speech transcription module is responsible for cleaning the transcribed dialogue text, removing redundant interjections and pause words, and retaining the text that meets the custom specifications.

[0014] The large model processing module is responsible for pre-collecting tax-specific business data, processing it to obtain a large model training dataset in Alpaca format, performing multi-round domain adaptation training on the general large model, and obtaining a special large model suitable for tax claim analysis. The large model processing module uses SFT (Supervised Fine-Tuning) for supervised fine-tuning training, combined with LoRA (Low-Rank Adaptation) fine-tuning, to perform multi-round adaptation training on the general large model to obtain a large model specifically for tax claim analysis.

[0015] The multi-dimensional classification module is responsible for inputting the request text that conforms to the custom specifications into the fine-tuned large-scale model for tax request analysis, and calling the large-scale model for tax request analysis to automatically classify and label the request text. The dedicated large-scale model for analyzing tax-related requests automatically classifies and labels request texts across three dimensions: request type, tax / fee category, and business matter.

[0016] The single-claim parsing module is responsible for parsing and refining the text of a single claim based on the dedicated large model for tax claim analysis and in conjunction with the Prompt project, and automatically extracting three key pieces of information: the taxpayer's core claim, the reason for the claim, and the solution to the claim. The request collection and hotspot extraction module is responsible for uniformly collecting taxpayer request data within a custom period and entering it into a vector library; performing cluster analysis on the collected data through a pre-trained large model to extract high-frequency hot issues within the corresponding time period; then matching and associating the corresponding causes of the requests using vector retrieval; further summarizing and refining the data through a dedicated large model for tax request analysis to form a summary of the typical causes of each hot issue; and finally outputting structured analysis results that can support tax management decisions.

[0017] A tax claim analysis device based on a large model includes a memory and a processor; the memory is used to store a computer program, and the processor is used to implement the above-described method when executing the computer program.

[0018] A readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0019] The beneficial effects of this invention are as follows: This tax complaint analysis method and system based on a large model not only enables fully automated processing of tax complaints, including automated collection, accurate analysis, standardized classification, and intelligent aggregation, significantly reducing labor costs and improving hotline processing efficiency; it also performs supervised fine-tuning of the general large model using SFT based on tax-specific data, making the model accurately adaptable to tax scenarios, improving understanding and classification accuracy, ensuring standardized and consistent processing throughout the process, and reducing human error; at the same time, the real-time structured complaint data facilitates rapid retrieval, statistics, and traceability, and can automatically uncover hot issues and common causes, providing reliable data support for tax service optimization, policy adjustments, and process improvements. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Appendix Figure 1 This is a schematic diagram of the tax claim analysis method based on a large model according to the present invention. Detailed Implementation

[0022] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions in the embodiments of this invention will be clearly and completely described below in conjunction with the embodiments of this invention. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0023] This tax claim analysis method based on a large model includes the following steps: Step S1: Real-time voice acquisition and transcription The system collects real-time audio of conversations between taxpayers and tax hotline operators, converts the audio stream into real-time dialogue text with speaker role annotations through a real-time audio transcription module, and preprocesses it to obtain a request text that conforms to the custom specifications. In step S1, the dialogue text transcribed by the real-time speech transcription module is cleaned to remove redundant interjections and pause words, and retain the text that meets the custom specifications.

[0024] Step S2: Fine-tuning the large-scale tax claim model Tax-specific business data is collected in advance, processed to obtain a large model training dataset in Alpaca format, and the general large model is subjected to multiple rounds of domain-specific adaptation training to obtain a special large model suitable for tax claim analysis. In step S2, the general large model is trained by supervised fine-tuning (SFT) and combined with LoRA (Low-Rank Adaptation) fine-tuning to perform domain-specific multi-round adaptation training, thereby obtaining a large model specifically for tax claim analysis.

[0025] Step S3: Automatic classification and labeling of dimensional requirements Input the request text that conforms to the custom specifications into the fine-tuned large-scale model for tax request analysis, and the large-scale model for tax request analysis will automatically classify and label the request text. In step S3, the dedicated large model for tax complaint analysis automatically classifies and labels the complaint text in three dimensions: complaint type, tax type, and business matter.

[0026] Step S4: Analysis and Information Extraction of a Single Request Based on a dedicated large model for tax demand analysis, combined with the Prompt project, the text of a single demand is parsed and its content is condensed, automatically extracting three key pieces of information: the taxpayer's core demand, the reason for the demand, and the solution to the demand. Step S5: Gathering Demands and Uncovering the Causes of Hot Topics Taxpayer complaint data within a defined period is collected and entered into a vector library. A pre-trained large model is used to perform cluster analysis on the collected data to extract high-frequency hot issues within the corresponding time period. Then, vector retrieval is used to match and associate the corresponding causes of the complaints. The data is further summarized and refined by a dedicated large model for tax complaint analysis to form a summary of the typical causes of each hot issue. Finally, a structured analysis result that can support tax management decisions is output.

[0027] This tax complaint analysis system based on a large model consists of a real-time speech transcription module, a large model processing module, a multi-dimensional classification module, a single complaint parsing module, and a complaint collection and hotspot extraction module. The modules communicate with each other and work together. The real-time speech transcription module is responsible for collecting the speech of taxpayers and tax hotline operators in real time, converting the speech stream into speech text with speaker role labels in real time, and preprocessing it to obtain the request text that conforms to the custom specifications. The real-time speech transcription module is responsible for cleaning the transcribed dialogue text, removing redundant interjections and pause words, and retaining the text that meets the custom specifications.

[0028] The large model processing module is responsible for pre-collecting tax-specific business data, processing it to obtain a large model training dataset in Alpaca format, performing multi-round domain adaptation training on the general large model, and obtaining a special large model suitable for tax claim analysis. The large model processing module uses SFT (Supervised Fine-Tuning) for supervised fine-tuning training, combined with LoRA (Low-Rank Adaptation) fine-tuning, to perform multi-round adaptation training on the general large model to obtain a large model specifically for tax claim analysis.

[0029] The multi-dimensional classification module is responsible for inputting the request text that conforms to the custom specifications into the fine-tuned large-scale model for tax request analysis, and calling the large-scale model for tax request analysis to automatically classify and label the request text. The dedicated large-scale model for analyzing tax-related requests automatically classifies and labels request texts across three dimensions: request type, tax / fee category, and business matter.

[0030] The single-claim parsing module is responsible for parsing and refining the text of a single claim based on the dedicated large model for tax claim analysis and in conjunction with the Prompt project, and automatically extracting three key pieces of information: the taxpayer's core claim, the reason for the claim, and the solution to the claim. The request collection and hotspot extraction module is responsible for uniformly collecting taxpayer request data within a custom period and entering it into a vector library; performing cluster analysis on the collected data through a pre-trained large model to extract high-frequency hot issues within the corresponding time period; then matching and associating the corresponding causes of the requests using vector retrieval; further summarizing and refining the data through a dedicated large model for tax request analysis to form a summary of the typical causes of each hot issue; and finally outputting structured analysis results that can support tax management decisions.

[0031] This large-model-based tax claim analysis device includes a memory and a processor; the memory is used to store computer programs, and the processor is used to execute the methods implemented when the computer programs are executed.

[0032] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor can be a microprocessor or any conventional processor.

[0033] Memory is used to store computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. Memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, at least one application program required for a function, etc.; the data storage area can store data created based on the use of the terminal, etc. In addition, memory can also include high-speed random access memory, and can also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart memory cards (SMC), secure digital cards (SD cards), flash memory cards, at least one disk storage device, flash memory devices, or other volatile solid-state storage devices.

[0034] The readable storage medium stores a computer program, the method of which is implemented when the computer program is executed by a processor.

[0035] Compared with existing technologies, this tax claim analysis method and system based on a large model has the following characteristics: It achieves fully automated processing of real-time voice transcription, request classification, request information summary and hotspot analysis, which greatly reduces labor costs and improves hotline processing efficiency. Supervised fine-tuning of the general model using tax-specific data through SFT enables the model to accurately adapt to tax scenarios, improves understanding and classification accuracy, ensures standardization and consistency throughout the process, and reduces human error. Meanwhile, the real-time structured data of the demands facilitates rapid retrieval, statistics and traceability, and can automatically discover hot issues and common causes, providing reliable data support for tax service optimization, policy adjustment and process improvement.

[0036] The foregoing has provided a detailed description of a tax claim analysis method and system based on a large model, as exemplified by this invention. This section uses specific examples to illustrate the principles and implementation methods of the invention. These examples are merely for the purpose of helping to understand the core ideas of this invention. All other embodiments obtained by those skilled in the art without creative effort, without departing from the principles of this invention, should fall within the scope of protection of this invention.

Claims

1. A method for analyzing tax claims based on a large model, characterized by: Includes the following steps: Step S1: Real-time voice acquisition and transcription The system collects real-time audio of conversations between taxpayers and tax hotline operators, converts the audio stream into real-time dialogue text with speaker role annotations through a real-time audio transcription module, and preprocesses it to obtain a request text that conforms to the custom specifications. Step S2: Fine-tuning the large-scale tax claim model Tax-specific business data is collected in advance, processed to obtain a large model training dataset in Alpaca format, and the general large model is subjected to multiple rounds of domain-specific adaptation training to obtain a special large model suitable for tax claim analysis. Step S3: Automatic classification and labeling of dimensional requirements Input the request text that conforms to the custom specifications into the fine-tuned large-scale model for tax request analysis, and the large-scale model for tax request analysis will automatically classify and label the request text. Step S4: Analysis and Information Extraction of a Single Request Based on a dedicated large model for tax demand analysis, combined with the Prompt project, the text of a single demand is parsed and its content is condensed, automatically extracting three key pieces of information: the taxpayer's core demand, the reason for the demand, and the solution to the demand. Step S5: Gathering Demands and Uncovering the Causes of Hot Topics Taxpayer complaint data within a defined period are collected and entered into a vector library. A pre-trained large model is used to perform cluster analysis on the collected data to extract high-frequency hot issues within the corresponding time period. Then, vector retrieval is used to match and associate the corresponding causes of the complaints. The data is further summarized and refined by a dedicated large model for tax complaint analysis to form a summary of the typical causes of each hot issue. Finally, a structured analysis result that can support tax administration decisions is output.

2. The tax claim analysis method based on a large model according to claim 1, characterized in that: In step S1, the dialogue text transcribed by the real-time speech transcription module is cleaned to remove redundant interjections and pause words, and retain the text that meets the custom specifications.

3. The tax claim analysis method based on a large model according to claim 1, characterized in that: In step S2, the general large model is trained in multiple rounds of domain adaptation by supervising the training with SFT and combining it with the LoRA low-rank adaptation fine-tuning method to obtain a special large model for tax claim analysis.

4. The tax claim analysis method based on a large model according to claim 1, characterized in that: In step S3, the dedicated large model for tax complaint analysis automatically classifies and labels the complaint text in three dimensions: complaint type, tax type, and business matter.

5. A tax claim analysis system based on a large model, characterized in that: It consists of a real-time speech transcription module, a large model processing module, a multi-dimensional classification module, a single request parsing module, and a request collection and hotspot extraction module. The modules communicate with each other and work together. The real-time speech transcription module is responsible for collecting the speech of taxpayers and tax hotline operators in real time, converting the speech stream into speech text with speaker role labels in real time, and preprocessing it to obtain the request text that conforms to the custom specifications. The large model processing module is responsible for pre-collecting tax-specific business data, processing it to obtain a large model training dataset in Alpaca format, performing multi-round domain adaptation training on the general large model, and obtaining a special large model suitable for tax claim analysis. The multi-dimensional classification module is responsible for inputting the request text that conforms to the custom specifications into the fine-tuned large-scale model for tax request analysis, and calling the large-scale model for tax request analysis to automatically classify and label the request text. The single-claim parsing module is responsible for parsing and refining the text of a single claim based on the dedicated large model for tax claim analysis and in conjunction with the Prompt project, and automatically extracting three key pieces of information: the taxpayer's core claim, the reason for the claim, and the solution to the claim. The request collection and hotspot extraction module is responsible for uniformly collecting taxpayer request data within a custom period and entering it into a vector library; performing cluster analysis on the collected data through a pre-trained large model to extract high-frequency hot issues within the corresponding time period; then matching and associating the corresponding causes of the requests using vector retrieval; further summarizing and refining the data through a dedicated large model for tax request analysis to form a summary of the typical causes of each hot issue; and finally outputting structured analysis results that can support tax management decisions.

6. The tax claim analysis system based on a large model according to claim 5, characterized in that: The real-time speech transcription module is responsible for cleaning the transcribed dialogue text, removing redundant interjections and pause words, and retaining the text that meets the custom specifications.

7. The tax claim analysis system based on a large model according to claim 5, characterized in that: The large model processing module performs domain-specific large model training through SFT supervised fine-tuning training and LoRA low-rank adaptation fine-tuning, resulting in a large model for tax claim analysis.

8. The tax claim analysis system based on a large model according to claim 5, characterized in that: The dedicated large-scale model for analyzing tax-related requests automatically classifies and labels request texts across three dimensions: request type, tax / fee category, and business matter.

9. A tax claim analysis device based on a large model, characterized in that: It includes a memory and a processor; the memory is used to store a computer program, and the processor is used to execute the computer program to implement the method as described in any one of claims 1 to 4.

10. A readable storage medium, characterized in that: The readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 4.

Images