Government affair large model fine tuning data processing method and device, electronic equipment and storage medium

Abstract

The invention discloses a government affair large model fine tuning data processing method and device, electronic equipment and a storage medium. The method comprises the steps that S1, QA pairs are extracted from sessions in an artificial customer service session data library and then classified to obtain government affair policy consultation class QA pairs and item search query class QA pairs; s2, according to the dialogue round number and theme number of government affair policy consultation and item search query class QA pairs, questions and answers are combined and assembled to form OA pairs, and the OA pairs are added into a QA pair data table; s3, customizing the standardized government affair service knowledge in a grading manner to form direct QA pairs, and storing the direct QA pairs in a QA pair data table; and S4, performing fine adjustment on the QA pair pair basic large model in the QA pair data table obtained in the two modes to obtain a government affair large model. According to the method, it is ensured that the fine-tuned government affair large model can quickly understand the user questions, accurate and professional answers are given, the service efficiency is remarkably improved, and the performance and competitiveness of the government affair large model are kept.

Description

Technical Field

[0001] The present invention relates to the field of government service information technology, and in particular, to a method, device, electronic device and storage medium for fine-tuning data processing of a large government model. Background Art

[0002] In the government service field, in addition to standardized government service knowledge, a large part of knowledge is implicit in data in real scenarios, such as mailbox data of various government departments and long-term manual customer service databases of various units. This type of data has the characteristics of large data volume, wide data coverage, uneven data quality, and high processing difficulty. In particular, the process of processing it into knowledge often requires a large amount of human resources. However, this part of data plays an important role in enhancing the generalization and understanding ability of the big model in the vertical field of government services, mainly reflected in the following aspects: (1) This part of data reflects the real intentions of the people. Based on this type of data, it can well enhance the ability of the big model to understand the expressions of the people in real scenarios and accurately match them with official expressions. However, the data in real scenarios only reflects the needs of some users themselves, and such needs are often difficult to fully cover all the questions and answers involved in government service knowledge; (2) Although standardized government service knowledge has a wide coverage after artificial pre-design and can greatly enrich the government service knowledge base, it is often relatively empty, which may be quite different from the answers that the people actually want to get.

[0003] Therefore, how to reasonably organize standardized government service knowledge and manual customer service data, establish the connection and complementarity between manual customer service colloquial expressions and standardized government service knowledge, and then enhance the large model in the government service field through model fine-tuning, is the core issue that this application urgently needs to solve. Summary of the invention

[0004] On the one hand, the present application provides a method for fine-tuning data processing for a large government service model to solve the current technical problem of requiring a large amount of human resources investment, high cost, and great difficulty in enhancing the large model in the field of government services by establishing the association and complementarity between the spoken expressions of manual customer service and standardized government service knowledge, and then fine-tuning the model.

[0005] This application is implemented through the following scheme:

[0006] A method for fine-tuning data processing of a large government affairs model, comprising the steps of:

[0007] S1. Extract QA pairs from the conversations in the manual customer service conversation data database and classify them into government policy consultation QA pairs and matter search query QA pairs;

[0008] S2. According to the number of dialogue rounds and the number of topics of the government policy consultation and matter search and query QA pairs, the questions and answers are merged and assembled respectively, and the merged questions and assembled answers are added to the QA pair data table as OA pairs;

[0009] S3. For standardized government service knowledge, direct QA pairs are customized in a hierarchical manner and stored in the QA pair data table;

[0010] S4. Fine-tune the basic large model with the QA pairs in the data table obtained by the above two methods to obtain a large model in the field of government services.

[0011] Furthermore, the step S1 specifically includes the steps of:

[0012] S11. Use a general large model to extract complete conversation data from the manual customer service original conversation database to form initial training data, ensuring that each record contains the customer service consultation question and the operator's answer to form a QA pair to build a common question and answer database, including single-round QA pairs or multi-round QA pairs;

[0013] S12. Classify the QA pairs in the constructed FAQ database using the general large model, including: system operation consultation QA pairs, personal service consultation QA pairs, government policy consultation QA pairs, matter search and query QA pairs, chat QA pairs, and complaint QA pairs;

[0014] S13. Extract QA pairs for government policy consultation and QA pairs for matter search and query.

[0015] Furthermore, the step S2 specifically includes the steps of:

[0016] S21. When the government policy consultation QA pairs and the matter search query QA pairs are single-topic single-round dialogues,

[0017] Use the general large model to judge the validity of the QA pair conversation, including whether it is complete and whether the QA pair is completely matched. Add two columns of labels "whether the conversation is complete" and "whether the QA pair is completely matched" to the data table for storage;

[0018] S22, extracting complete single-topic single-round dialogues with complete matching of questions and answers;

[0019] S23. Use the general large model to process personal descriptions or sensitive information in the questions that are not related to the answers, simplify and summarize the core questions, and add a new column "summarized questions" in the data table for storage;

[0020] S24. Add the simplified summary of the core questions and the original answers to the QA data table.

[0021] Furthermore, the step S2 specifically includes the steps of:

[0022] S21. When the QA pairs for government policy consultation and the QA pairs for item search and query are single-topic multi-round dialogues, extract the user's initial question, the customer service's follow-up question, the user's supplementary answer, and the customer service's answer, and add four columns to the data table;

[0023] S22, based on the newly added user initial question, customer service follow-up question content and user supplementary answer, merge to form a complete and specific question, and add a new column "complete question" in the data table for storage;

[0024] S23. After removing the customer service follow-up questions, the customer service answers are assembled to form a complete answer content, and stored as a column of "complete answer" in the data table;

[0025] S24. Add the merged questions and assembled answers to the QA data table.

[0026] Furthermore, the step S2 comprises the steps of:

[0027] S21. When the QA pairs for government policy consultation and the QA pairs for item search and query are multi-topic and multi-round conversations, use the general large model to decompose the topic of a conversation. If the decomposed single-topic conversation data still contains multi-round conversations, extract the user's initial question, the customer service's follow-up question, the user's supplementary answer, and the customer service's answer for each single-topic conversation data, and add four columns to the data table;

[0028] S22, based on the newly added user initial question, customer service follow-up question content and user supplementary answer, merge to form a complete and specific question, and add a new column "complete question" in the data table for storage;

[0029] S23. After removing the customer service follow-up questions, the customer service answers are assembled to form a complete answer content, and stored as a column of "complete answer" in the data table;

[0030] S24. Add the merged questions and assembled answers to the QA data table.

[0031] Furthermore, the step S2 further comprises the steps of:

[0032] S25. When the QA pairs in the QA pair data table are government policy consulting QA pairs, the complete questions formed by merging are processed through the prompt engineering of the large model to remove redundant information;

[0033] S26. Using the prompting engineering of the large model, judging whether the combined complete question is consistent with the original question from the perspective of language understanding; using the large model to standardize the assembled complete answer, judging whether the standardized complete answer is consistent with the original recorded answer from the perspective of language understanding;

[0034] S27. If they are inconsistent, delete the original record; if they are consistent, retain the original record and add a new record of the question and answer combination standardized by the large model to the QA pair data table.

[0035] Furthermore, the step S3 specifically includes the steps of:

[0036] S31. Construct a complete item description system based on the set elements, wherein the set elements include: item name description, term explanation, value range description, item handling situation, acceptance conditions, and legal basis;

[0037] S32, extracting the constructed item description system into QA pairs, and grading the questions of each QA pair of the item description system, including first-level QA pairs, second-level QA pairs, and third-level QA pairs, wherein the first-level QA questions only convert the item name description into colloquial expressions, the second-level QA questions include simple condition description + item name description converted into colloquial expressions, and the third-level QA questions include complex condition description + simple reasoning required;

[0038] S33. Store the direct QA pairs formed after classification into the QA pair data table.

[0039] On the other hand, the present application also provides a device for fine-tuning data processing of a large government model, including:

[0040] The conversation QA pair extraction and classification module is used to extract QA pairs from the conversations in the manual customer service conversation data database and then classify them into government policy consultation QA pairs and matter search query QA pairs;

[0041] The conversational QA pair merging and assembly module merges and assembles questions and answers according to the number of conversation rounds and the number of topics of government policy consultation and matter search query QA pairs, and adds the merged questions and assembled answers into OA pairs and adds them to the QA pair data table;

[0042] The QA pair grading module is used to customize the standardized government service knowledge in a graded manner to form direct QA pairs and store them in the QA pair data table;

[0043] The large model fine-tuning module is used to fine-tune the basic large model of the QA pairs in the QA pair data table obtained by the above two methods to obtain a large model in the field of government services.

[0044] On the other hand, the present application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method for fine-tuning data processing of the government affairs big model when executing the computer program.

[0045] On the other hand, the present application also provides a storage medium, which includes a stored program, and when the program is running, controls the device where the storage medium is located to execute the steps of the government affairs big model fine-tuning data processing method.

[0046] Compared with the prior art, this application has the following beneficial effects:

[0047] The present application provides a method, device, electronic device and storage medium for fine-tuning data processing of a government affairs big model. The method for fine-tuning data processing of the government affairs big model classifies QA pairs extracted from conversations in a manual customer service conversation data database, merges and assembles them according to the number of conversation rounds and the number of topics to obtain OA pairs and adds them to the QA pair data table, and customizes standardized government service knowledge in a hierarchical manner to form direct QA pairs and stores them in the QA pair data table, thereby ensuring that the obtained QA pairs used for fine-tuning the big model have both the accuracy and authenticity of actual manual customer service conversation data and the comprehensive coverage of OA pairs obtained from standardized data, ensuring that the fine-tuned big model can quickly understand user problems and give accurate and professional answers, significantly improving service efficiency, and promoting continuous iteration and optimization of the big model with the continuous accumulation and updating of government service data to maintain its performance and competitiveness.

[0048] In addition to the above-described purposes, features and advantages, the present application also has other purposes, features and advantages. The present application will be further described in detail with reference to the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application.

[0050] Figure 1 It is a flow chart of the method for fine-tuning data processing of the government affairs big model according to the preferred embodiment of the present application.

[0051] Figure 2 It is a schematic diagram of the sub-step flow of step S1 of the preferred embodiment of the present application.

[0052] Figure 3 It is a schematic diagram of the sub-step flow of step S2 of the preferred embodiment of the present application.

[0053] Figure 4It is a schematic diagram of the sub-step flow of step S2 of another preferred embodiment of the present application.

[0054] Figure 5 It is a schematic diagram of the sub-step flow of step S2 of another preferred embodiment of the present application.

[0055] Figure 6 It is a schematic diagram of the sub-step flow of step S2 of another preferred embodiment of the present application.

[0056] Figure 7 It is a schematic diagram of the sub-step flow of step S3 of the preferred embodiment of the present application.

[0057] Figure 8 It is a schematic diagram of the module of the government affairs big model fine-tuning data processing device of the preferred embodiment of the present application.

[0058] Fig. 9 It is a schematic block diagram of an electronic device entity of a preferred embodiment of the present application.

[0059] Fig.10 It is a diagram of the internal structure of a computer device of a preferred embodiment of the present application. DETAILED DESCRIPTION

[0060] The embodiments of the present application are described in detail below in conjunction with the accompanying drawings, but the present application can be implemented in a variety of different ways defined and covered below.

[0061] like Figure 1 As shown, the preferred embodiment of the present application provides a method for fine-tuning data processing of a large government affairs model, comprising the steps of:

[0062] S1. Extract QA pairs from the conversations in the manual customer service conversation data database and classify them into government policy consultation QA pairs and matter search query QA pairs;

[0063] S2. According to the number of dialogue rounds and the number of topics of the government policy consultation and matter search and query QA pairs, the questions and answers are merged and assembled respectively, and the merged questions and assembled answers are added to the QA pair data table as OA pairs;

[0064] S3. For standardized government service knowledge, direct QA pairs are customized in a hierarchical manner and stored in the QA pair data table;

[0065] S4. Fine-tune the basic large model with the QA pairs in the data table obtained by the above two methods to obtain a large model in the field of government services.

[0066] The present embodiment provides a method for fine-tuning data processing of a government affairs big model, which classifies QA pairs extracted from conversations in a manual customer service conversation data database, merges and assembles them according to the number of conversation rounds and the number of topics to obtain OA pairs, and adds them to a QA pair data table, and customizes standardized government service knowledge in a hierarchical manner to form direct QA pairs and stores them in the QA pair data table, thereby ensuring that the obtained QA pairs used for fine-tuning the big model have both the accuracy and authenticity of actual manual customer service conversation data and the comprehensive coverage of OA pairs obtained from standardized data, ensuring that the fine-tuned big model can quickly understand user problems, has the ability to understand the expressions of ordinary people in real scenarios, accurately matches them with official expressions, and gives accurate and professional answers, thereby significantly improving service efficiency. With the continuous accumulation and updating of government service data, the big model is promoted to be continuously iterated and optimized to maintain its performance and competitiveness.

[0067] like Figure 2 As shown, in a preferred embodiment of the present application, step S1 specifically includes the steps of:

[0068] S11. Use a general large model to extract complete conversation data from the manual customer service original conversation database to form initial training data, ensuring that each record contains the customer service consultation question and the operator's answer to form a QA pair to build a common question and answer database, including single-round QA pairs or multi-round QA pairs;

[0069] S12. Classify the QA pairs in the constructed FAQ database using the general large model, including: system operation consultation QA pairs, personal service consultation QA pairs, government policy consultation QA pairs, matter search and query QA pairs, chat QA pairs, and complaint QA pairs;

[0070] S13. Extract QA pairs for government policy consultation and QA pairs for matter search and query.

[0071] A complete conversation is a logical, structured, and purposeful communication process. This embodiment uses the general big model's ability to understand semantics to complete the integrity judgment of each record in the database, complete the process of extracting complete conversation data from the manual customer service original conversation database to form initial training data, and ensure that each record contains customer service consultation questions and operator answers (single or multiple rounds of QA pairs). Then, the QA pairs in the database are classified into knowledge categories for the conversation records using the general big model, that is, to determine whether the conversation is one of the following subdivisions. The knowledge entering the big model fine-tuning training data set should be government policy consultation QA pairs and matter search query QA pairs, so the government policy consultation QA pairs and matter search query QA pairs are further extracted to meet the knowledge needs of the big model fine-tuning training data set.

[0072] like Figure 3 In the preferred embodiment of the present application, step S2 specifically includes the following steps:

[0073] S21. When the QA pairs of government policy consultation and the QA pairs of matter search and query are single-topic single-round dialogues, the general large model is used to judge the validity of the QA pairs, including whether they are complete and whether the QA pairs are completely matched. Two columns of labels "whether the conversation is complete" and "whether the QA pairs are completely matched" are added to the data table for storage;

[0074] S22, extracting complete single-topic single-round dialogues with complete matching of questions and answers;

[0075] S23. Use the general large model to process personal descriptions or sensitive information in the questions that are not related to the answers, simplify and summarize the core questions, and add a new column "summarized questions" in the data table for storage;

[0076] S24. Add the simplified summary of the core questions and the original answers to the QA data table.

[0077] This embodiment first performs validity judgment on single-topic single-round conversations, and further extracts single-topic single-round conversations on the premise that the conversation is complete and fully matched. Finally, sensitive information in the question that is not related to the answer is processed to ensure the completeness, accuracy and security of the question and the answer, and avoid leakage of personal sensitive information.

[0078] In real consultation scenarios, users’ complete questions are often gradually formed after multiple rounds of dialogue and follow-up questions with customer service. In a complete conversation, the overall structure of the input question includes the user’s initial question, the customer service’s follow-up questions, and the user’s supplementary answers. Therefore, by polling each record in the database, it is determined whether the record is a multi-round dialogue. If it is a multi-round dialogue, then Figure 4 As shown, the step S2 specifically includes the steps of:

[0079] S21. When the QA pairs for government policy consultation and the QA pairs for item search and query are single-topic multi-round dialogues, extract the user's initial question, the customer service's follow-up question, the user's supplementary answer, and the customer service's answer, and add four columns to the data table;

[0080] S22, based on the newly added user initial question, customer service follow-up question content and user supplementary answer, merge to form a complete and specific question, and add a new column "complete question" in the data table for storage;

[0081] For example, when a user is handling the housing fund withdrawal business, in the initial input, he may only describe "I didn't have enough money when I bought a house, and I want to withdraw the housing fund. What materials do I need to prepare?" At this time, the customer service will guide the user to clarify the more specific withdrawal situation by asking "Are you buying a pre-sale house or a second-hand house?", so as to give a more accurate answer. At this time, the complete question should be "I didn't have enough money when I bought a pre-sale house, and I want to withdraw the housing fund. What materials do I need to prepare?" or "I didn't have enough money when I bought a second-hand house, and I want to withdraw the housing fund. What materials do I need to prepare?"

[0082] S23. After removing the customer service follow-up questions, the customer service answers are assembled to form a complete answer content, and stored as a column of "complete answer" in the data table;

[0083] S24. Add the merged questions and assembled answers to the QA data table.

[0084] In this embodiment, when the QA pairs of government policy consultation and the QA pairs of matter search and query are single-topic multi-round dialogues, the questions of the single-topic multi-round dialogues are merged to form complete questions, the answers are assembled to complete the complete answers, and finally the merged questions and assembled answers are added to the QA pair data table. The purpose and benefit of this are: to ensure the colloquial nature of the questions while eliminating the redundancy of the answers, and to ensure the accuracy and non-ambiguity of the answers.

[0085] like Figure 5 In a preferred embodiment of the present application, step S2 includes the following steps:

[0086] S21. When the QA pairs for government policy consultation and the QA pairs for item search and query are multi-topic and multi-round conversations, use the general large model to decompose the topic of a conversation. If the decomposed single-topic conversation data still contains multi-round conversations, extract the user's initial question, the customer service's follow-up question, the user's supplementary answer, and the customer service's answer for each single-topic conversation data, and add four columns to the data table;

[0087] S22, based on the newly added user initial question, customer service follow-up question content and user supplementary answer, merge to form a complete and specific question, and add a new column "complete question" in the data table for storage;

[0088] S23. After removing the customer service follow-up questions, the customer service answers are assembled to form a complete answer content, and stored as a column of "complete answer" in the data table;

[0089] S24. Add the merged questions and assembled answers to the QA data table.

[0090] Different from the above-mentioned embodiment, this embodiment polls each record in the database to determine whether the record is a cross-topic multi-round dialogue. When there may be consultations on multiple topics in a conversation, this embodiment first disassembles the content of the multiple rounds of dialogues, and each consultation content on a different topic can be used as an independent QA pair. Then, the multi-round dialogues on each topic are handled in the manner of the above-mentioned single-topic multi-round conversation embodiment, and the questions are merged to form complete questions, and the answers are assembled to complete the complete answers. Finally, the merged questions and assembled answers are added to the QA pair data table. Since this embodiment takes into account the multi-topic decomposition in actual conversations, it is more in line with the actual conversation scenarios between customer service and users, greatly expanding the scope of application of this solution.

[0091] like Figure 6 In a preferred embodiment of the present application, step S2 further comprises the following steps:

[0092] S25. When the QA pairs in the QA pair data table are government policy consulting QA pairs, the complete questions formed by merging are processed through the prompt engineering of the large model to remove redundant information;

[0093] S26. Using the prompting engineering of the large model, judging whether the combined complete question is consistent with the original question from the perspective of language understanding; using the large model to standardize the assembled complete answer, judging whether the standardized complete answer is consistent with the original recorded answer from the perspective of language understanding;

[0094] S27. If they are inconsistent, delete the original record; if they are consistent, retain the original record and add a new record of the question and answer combination standardized by the large model to the QA pair data table.

[0095] In this embodiment, no matter it is conversation data of multi-topic multi-round dialogue or single-topic multi-round dialogue, after the merged questions and assembled answers are added to the QA pair data table, there may still be redundancy and consistency problems of QA pairs, which will affect the fine-tuning of the government affairs big model. Therefore, in order to ensure the streamlining and consistency of the QA pairs in the data table, this embodiment uses the prompt project of the general big model to further process the merged questions and assembled answers in the data table to remove redundant information, judge whether the merged complete question is consistent with the original question, judge whether the standardized complete answer is consistent with the original record answer, and delete the original record if it is inconsistent. If it is consistent, the original record is retained while adding a new record of the question and answer combination standardized by the big model to the QA pair data table, thereby ensuring that the semantics are not ambiguous while achieving the diversity of QA pairs.

[0096] like Figure 7As shown, in a preferred embodiment of the present application, step S3 specifically includes the steps of:

[0097] S31. Construct a complete item description system based on the set elements, wherein the set elements include: item name description, term explanation, value range description, item handling situation, acceptance conditions, and legal basis;

[0098] S32. Extract the constructed matter description system into QA pairs, and classify the problems of each QA pair of the matter description system, including problems of first-level QA pairs, problems of second-level QA pairs, and problems of third-level QA pairs, wherein:

[0099] The first-level QA questions are to convert only the name of the item into colloquial expressions, such as "I want to handle business-to-public business" - "housing provident fund loan business - bank housing loan to provident fund loan";

[0100] The second-level QA questions include simple condition description + item name description, which is converted into colloquial expressions, such as "I am a flexible employment person and need to pay provident fund" - "housing provident fund payment business - flexible employment personnel payment";

[0101] The third-level QA questions include those with complex condition descriptions + simple reasoning, such as "I used my husband's housing provident fund loan before, can it be converted into my own housing provident fund loan?" - "Housing Provident Fund Loan Business - Change of Main Borrower";

[0102] S33. Store the direct QA pairs formed after classification into the QA pair data table.

[0103] In this embodiment, for standardized government service knowledge, a complete matter description system is first constructed by setting elements, and then the constructed matter description system is extracted into QA pairs. The problems of each QA pair of the matter description system are graded and then stored in the QA pair data table. The purposes and benefits include: more comprehensive coverage of government service knowledge and enhanced capabilities of the fine-tuned government service big model.

[0104] like Figure 8 As shown, another preferred embodiment of the present application further provides a device for fine-tuning data processing of a government affairs large model, comprising:

[0105] The conversation QA pair extraction and classification module is used to extract QA pairs from the conversations in the manual customer service conversation data database and then classify them into government policy consultation QA pairs and matter search query QA pairs;

[0106] The conversational QA pair merging and assembly module merges and assembles questions and answers according to the number of conversation rounds and the number of topics of government policy consultation and matter search query QA pairs, and adds the merged questions and assembled answers into OA pairs and adds them to the QA pair data table;

[0107] The QA pair grading module is used to customize the standardized government service knowledge in a graded manner to form direct QA pairs and store them in the QA pair data table;

[0108] The large model fine-tuning module is used to fine-tune the basic large model of the QA pairs in the QA pair data table obtained by the above two methods to obtain a large model in the field of government services.

[0109] like Fig. 9 As shown, a preferred embodiment of the present application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method for fine-tuning data processing of the government affairs big model in the above-mentioned embodiment when executing the computer program.

[0110] like Fig.10 As shown, the preferred embodiment of the present application also provides a computer device, which can be a terminal or a liveness detection server, and its internal structure diagram can be as shown in Fig.10 As shown. The computer device includes a processor, a memory and a network interface connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with other external computer devices via a network connection. When the computer program is executed by the processor, the steps of the above-mentioned government affairs large model fine-tuning data processing method are implemented.

[0111] Those skilled in the art will understand that Fig.10 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0112] A preferred embodiment of the present application also provides a storage medium, which includes a stored program, and when the program is running, controls the device where the storage medium is located to execute the steps of the government affairs big model fine-tuning data processing method in the above embodiment.

[0113] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0114] If the functions described in the method of this embodiment are implemented in the form of software functional units and sold or used as independent products, they can be stored in one or more computing devices readable storage media. Based on this understanding, the part of the embodiment of the present application that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions to enable a computing device (which can be a personal computer, server, mobile computing device or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program code.

[0115] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0116] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0117] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1A function specified in one or more boxes.

[0118] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0119] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0120] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for fine-tuning data processing of a large government model, characterized in that: Includes steps: S1. Extract QA pairs from the conversations in the manual customer service conversation data database and classify them into government policy consultation QA pairs and matter search query QA pairs; S2. According to the number of dialogue rounds and the number of topics of the government policy consultation and matter search and query QA pairs, the questions and answers are merged and assembled respectively, and the merged questions and assembled answers are added to the QA pair data table as OA pairs; S3. For standardized government service knowledge, direct QA pairs are customized in a hierarchical manner and stored in the QA pair data table; S4. Fine-tune the basic large model with the QA pairs in the data table obtained by the above two methods to obtain a large model in the field of government services.

2. The method for fine-tuning data for a government affairs big model according to claim 1 is characterized in that: The step S1 specifically includes the following steps: S11. Use a general large model to extract complete conversation data from the manual customer service original conversation database to form initial training data, ensuring that each record contains the customer service consultation question and the operator's answer to form a QA pair to build a common question and answer database, including single-round QA pairs or multi-round QA pairs; S12. Classify the QA pairs in the constructed FAQ database using the general large model, including: system operation consultation QA pairs, personal service consultation QA pairs, government policy consultation QA pairs, matter search and query QA pairs, chat QA pairs, and complaint QA pairs; S13. Extract QA pairs for government policy consultation and QA pairs for matter search and query.

3. The method for fine-tuning data for a government affairs big model according to claim 1 is characterized in that: The step S2 specifically includes the following steps: S21. When the government policy consultation QA pairs and the matter search query QA pairs are single-topic single-round dialogues, Use the general large model to judge the validity of the QA pair conversation, including whether it is complete and whether the QA pair is completely matched. Add two columns of labels "whether the conversation is complete" and "whether the QA pair is completely matched" to the data table for storage; S22, extracting complete single-topic single-round dialogues with complete matching of questions and answers; S23. Use the general large model to process personal descriptions or sensitive information in the questions that are not related to the answers, simplify and summarize the core questions, and add a new column "summarized questions" in the data table for storage; S24. Add the simplified summary of the core questions and the original answers to the QA data table.

4. The method for fine-tuning data for a government affairs big model according to claim 1 is characterized in that: The step S2 specifically includes the following steps: S21. When the QA pairs for government policy consultation and the QA pairs for item search and query are single-topic multi-round dialogues, extract the user's initial question, the customer service's follow-up question, the user's supplementary answer, and the customer service's answer, and add four columns to the data table; S22, based on the newly added user initial question, customer service follow-up question content and user supplementary answer, merge to form a complete and specific question, and add a new column "complete question" in the data table for storage; S23. After removing the customer service follow-up questions, the customer service answers are assembled to form a complete answer content, and stored as a column "complete answer" in the data table; S24. Add the merged questions and assembled answers to the QA data table.

5. The method for fine-tuning data for a government affairs big model according to claim 1 is characterized in that: The step S2 comprises the steps of: S21. When the QA pairs for government policy consultation and the QA pairs for item search and query are multi-topic and multi-round conversations, use the general large model to decompose the topic of a conversation. If the decomposed single-topic conversation data still contains multi-round conversations, extract the user's initial question, the customer service's follow-up question, the user's supplementary answer, and the customer service's answer for each single-topic conversation data, and add four columns to the data table; S22, based on the newly added user initial question, customer service follow-up question content and user supplementary answer, merge to form a complete and specific question, and add a new column "complete question" in the data table for storage; S23. After removing the customer service follow-up questions, the customer service answers are assembled to form a complete answer content, and stored as a column "complete answer" in the data table; S24. Add the merged questions and assembled answers to the QA data table.

6. The method for fine-tuning data of a government affairs big model according to claim 4 or 5, characterized in that: The step S2 further comprises the steps of: S25. When the QA pairs in the QA pair data table are government policy consulting QA pairs, through the big model The prompt engineering processes the merged complete questions to remove redundant information; S26. Using the prompting engineering of the large model, judging whether the combined complete question is consistent with the original question from the perspective of language understanding; using the large model to standardize the assembled complete answer, judging whether the standardized complete answer is consistent with the original recorded answer from the perspective of language understanding; S27. If they are inconsistent, delete the original record; if they are consistent, retain the original record and add a new record of the question and answer combination standardized by the large model to the QA pair data table.

7. The method for fine-tuning data for a government affairs big model according to claim 1 is characterized in that: The step S3 specifically comprises the following steps: S31. Construct a complete item description system for standardized government service knowledge based on set elements, wherein the set elements include: item name description, term explanation, value range description, item handling situation, acceptance conditions, and legal basis; S32, extracting the constructed item description system into QA pairs, and grading the questions of each QA pair of the item description system, including first-level QA pairs, second-level QA pairs, and third-level QA pairs, wherein the first-level QA questions only convert the item name description into colloquial expressions, the second-level QA questions include simple condition description + item name description converted into colloquial expressions, and the third-level QA questions include complex condition description + simple reasoning required; S33. Store the direct QA pairs formed after classification into the QA pair data table.

8. A device for fine-tuning data processing of a large government model, characterized in that: include: The conversation QA pair extraction and classification module is used to extract QA pairs from the conversations in the manual customer service conversation data database and then classify them into government policy consultation QA pairs and matter search query QA pairs; The conversational QA pair merging and assembly module merges and assembles questions and answers according to the number of conversation rounds and the number of topics of government policy consultation and matter search query QA pairs, and adds the merged questions and assembled answers into OA pairs and adds them to the QA pair data table; The QA pair grading module is used to customize the standardized government service knowledge in a graded manner to form direct QA pairs and store them in the QA pair data table; The large model fine-tuning module is used to compare the QA pairs in the QA pair data table obtained by the above two methods with the base The basic large model is fine-tuned to obtain a large model in the field of government services.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the steps of the method for fine-tuning data processing of the government affairs big model as described in any one of claims 1 to 7.

10. A storage medium, comprising a stored program, which controls a device where the storage medium is located to execute the steps of the method for fine-tuning data processing of a government affairs big model as described in any one of claims 1 to 7 when the program is running.

Images