Government service vertical domain large model construction method and system based on knowledge graph

Abstract

The application provides a knowledge graph-based Xiyang government service vertical domain large model construction method and system, belonging to the technical field of artificial intelligence and government service. The method comprises the following steps: S100, analyzing the original query demand of a user to obtain a plurality of query keywords; S200, determining a benchmark large model; S300, the benchmark large model outputs a plurality of extended query demands, and the extended query demand is a visual knowledge graph with the query keyword as a root node; S400, in response to the selection of the user on the extended query demand, at least one vertical domain corpus is determined, and the vertical domain corpus comprises a plurality of different modal government service corpora; S500, the vertical domain corpus is loaded, the benchmark large model is trained, and a trained vertical domain large model is obtained. The system comprises a plurality of candidate benchmark large models and a plurality of candidate vertical domain corpora, and is used for realizing the method. The application can quickly and effectively construct a government service vertical domain large model and is easy to use.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of artificial intelligence and government service, and particularly relates to a government service vertical domain large model construction method and system based on a knowledge graph, a computer readable storage medium, a computer program product and an electronic device for implementing the method. BACKGROUND

[0002] The vertical domain large model is also called an industry large model, a professional / field large model, and is a short form of a vertical field large model. It is relative to a general large model and is a large model application dedicated to a certain field or industry. The general large model is usually trained on a wide range of public data sets, which enables it to learn and solve various common problems, but these data sets cannot completely cover all professional knowledge and technical details in certain specific fields, which leads to the fact that although the general large model has extensive general knowledge, it lacks sufficient depth of knowledge to meet the complex needs of certain specific fields. Therefore, it is particularly important to construct a vertical field large model according to the needs of a specific industry.

[0003] According to the search, Chinese invention patent CN112100344B discloses a financial field knowledge question and answer method based on a knowledge graph. The method effectively improves the accuracy of the question and answer system through neural network training model and similarity comparison adaptive parameter adjustment. Related literature (“Customization of Vertical Domain Large Models: Theoretical Basis and Key Technologies”, Chen Haoli, et al., Data Collection and Processing, 2024, 39(03)) also has some introduction and reports on related theories and technologies. However, there are still few reports on the specific use of vertical domain large models in government services.

[0004] According to the search, related literature (“Application Status and Opportunity Challenges of Government Large Models”, Li Wei, et al., Communication Enterprise Management, 2024, (06): 42-44.) all introduce the basic status and principles of the use of general large models in government services, and do not give a clear technical solution to how to train and construct vertical domain large models for government services. SUMMARY

[0005] In view of the above technical problems, the present application provides a government service vertical domain large model construction method and system based on a knowledge graph, a computer readable storage medium, a computer program product and an electronic device for implementing the method.

[0006] In the first aspect of the present application, a government service vertical domain large model construction method based on a knowledge graph is provided, which comprises the following steps:

[0007] S100: analyzing original query requirements of users to obtain a plurality of query keywords;

[0008] The original query requirement includes a text query requirement, a voice query requirement, and an image query requirement

[0009] S200: determining a benchmark large model based on the query keyword;

[0010] S300: inputting the query keyword into the benchmark large model, and outputting a plurality of extended query requirements by the benchmark large model;

[0011] S400: determining at least one vertical corpus in response to user selection of the extended query requirement;

[0012] S500: loading the vertical corpus, training the benchmark large model, and obtaining a trained vertical large model;

[0013] The vertical corpus includes a plurality of different modal government service corpora, and the different modalities include one of text, static image, video frame, voice sequence, or any combination thereof.

[0014] The plurality of extended query requirements are a visual knowledge graph with the query keyword as a root node.

[0015] In the actual application of the method, after step S500, the method further includes:

[0016] S600: inputting the original query requirement of the user into the vertical large model, and outputting a government service result in at least one modal form according to the current attribute of the user terminal by the vertical large model.

[0017] The benchmark large model is an industry general large model, and the industry general large model includes a text-to-text model, a prompt word class text-to-graph model, and a graph-to-graph model.

[0018] The step S500 loads the vertical corpus, trains the benchmark large model, and obtains a trained vertical large model, and specifically includes:

[0019] The vertical corpus is used as a supplementary training library and an external query library of the benchmark large model by using a retrieval enhancement generation method;

[0020] The training parameters of the benchmark large model are adjusted based on the supplementary training library;

[0021] The trained vertical large model only performs query and retrieval in the external query library.

[0022] The step S600 inputs the original query requirement of the user into the vertical large model, and outputs a government service result in at least one modal form according to the current attribute of the user terminal by the vertical large model, and specifically includes:

[0023] S610: generating a plurality of field prompt words based on the original query requirement;

[0024] S620: integrating the field prompt words into the original query requirement as additional context information and inputting the original query requirement and the additional context information into the vertical domain large model as input;

[0025] S630: the vertical domain large model performs retrieval in the vertical domain corpus;

[0026] S640: when the retrieval result contains multiple modalities, outputting a result of a government service in at least one modality according to the current attributes of the user terminal.

[0027] Corresponding to the method for constructing a knowledge graph-based government service vertical domain large model according to the first aspect, in the second aspect of the present application, a system for constructing a knowledge graph-based government service vertical domain large model is provided, the system comprising a plurality of candidate benchmark large models and a plurality of candidate vertical domain corpora, the plurality of candidate benchmark large models comprising a text-to-text model, a prompt word-based text-to-graph model, and a graph-to-graph model; the plurality of candidate vertical domain corpora comprising a plurality of government service corpora in different modalities, the different modalities comprising one of text, static graph, video frame, and speech sequence or any combination thereof;

[0028] The system further comprises:

[0029] an analysis unit configured to analyze the original query requirement of the user to obtain a plurality of query keywords;

[0030] a benchmark large model determination unit configured to determine a benchmark large model based on the query keywords;

[0031] a knowledge graph expansion unit configured to input the query keywords into the benchmark large model, the benchmark large model outputting a plurality of expanded query requirements; the plurality of expanded query requirements being a visual knowledge graph with the query keywords as root nodes;

[0032] a vertical domain corpus determination unit configured to determine at least one vertical domain corpus in response to a user's selection of the expanded query requirements;

[0033] a training unit configured to load the vertical domain corpus and train the benchmark large model to obtain a trained vertical domain large model.

[0034] In specific applications, the system further comprises:

[0035] an input unit configured to input the original query requirement of the user through the input unit, the original query requirement comprising a text query requirement, a speech query requirement, and an image query requirement.

[0036] The result output unit inputs the original query requirement of the user into the vertical domain large model, and the vertical domain large model outputs a result of government service in at least one modal form according to the current attribute of the user terminal.

[0037] The aforementioned government service vertical domain large model construction method based on a knowledge graph can be realized automatically through computer program instructions via various forms of electronic devices, and the computer program instructions can be stored in different forms of storage media and loaded into a computer electronic device for execution.

[0038] Therefore, in the third aspect of the present application, a computer readable storage medium is further provided for storing computer instructions, which, when executed on an electronic device, cause the electronic device to perform all or part of the steps of the aforementioned government service vertical domain large model construction method based on a knowledge graph.

[0039] In the fourth aspect of the present application, a computer device is further provided, which includes a processor and a memory, the memory being used to store instructions, and the processor being used to call the instructions in the memory, so that the computer device performs the aforementioned government service vertical domain large model construction method based on a knowledge graph.

[0040] In the fifth aspect of the present application, a computer program product is further provided, which includes a computer program, when the computer program is executed, all or part of the steps of the aforementioned government service vertical domain large model construction method based on a knowledge graph are realized.

[0041] The present application firstly determines a benchmark large model through a query keyword, and then outputs a plurality of extended query requirements through the benchmark large model; then, in response to the selection of the user on the extended query requirement, at least one vertical domain corpus is determined, and then the vertical domain corpus is loaded to train the benchmark large model, so that a trained vertical domain large model is obtained, and the government service vertical domain large model can be quickly and effectively constructed and is easy to use.

[0042] The further advantages of the present application will be further embodied in detail in the specific embodiment part combined with the drawings of the specification. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0044] Figure 1is the main flow chart of the knowledge graph-based government service vertical domain large model construction method of an embodiment of the present application.

[0045] Figure 2 is the main flow chart of the knowledge graph-based government service vertical domain large model construction method of another embodiment of the present application.

[0046] Figure 3 is the data analysis schematic diagram of analyzing the original query demand of the user to obtain a plurality of query keywords.

[0047] Figure 4 is the knowledge graph schematic diagram of outputting a plurality of expanded query demands.

[0048] Figure 5 is the visual knowledge graph schematic diagram with the query keyword as the root node under a specific application example of the present application.

[0049] Figure 6 is the structural schematic diagram of the knowledge graph-based government service vertical domain large model construction system of an embodiment of the present application.

[0050] Figure 7 is the partial interaction structural schematic diagram of the input unit and the output unit of the knowledge graph-based government service vertical domain large model construction system. DETAILED DESCRIPTION

[0051] The embodiments of this part are in a progressive manner. First, the application status and implementation effect of the government service vertical domain large model proposed by the applicant are introduced, and then the basic construction principle of the vertical domain large model is introduced, that is, the related method embodiments and the related schematic principles of applying the method; finally, the related system embodiments corresponding to the method embodiments are introduced.

[0052] It can be understood that the method embodiments and the system embodiments should be corresponding in the execution process, the execution principle, the corresponding functional modules and the method steps. Therefore, after the method embodiments are introduced in detail, the system embodiments do not have to be introduced repeatedly.

[0053] The technical solution of the present application comes from the key research and development plan project undertaken by the related units of the applicant, that is, the research and demonstration of the construction technology based on the government service large language model. Based on the "large model" formed by the project research and development, it has successfully passed the generative artificial intelligence (large language model) and deep synthesis service algorithm "double filing", which integrates advanced AI capabilities such as multi-round dialogue, planning suggestion, content generation, and provides customized solutions for government services, housing accumulation fund, education and other fields.

[0054] Corresponding to the above-mentioned industry model, the present application proposes the related technical solutions of the corresponding vertical domain large model construction method and system as follows.

[0055] Referring first to Figure 1 , Figure 1 is a subject flowchart of a knowledge graph-based government service vertical domain large model construction method of an embodiment of the present application.

[0056] In Figure 1 , the method comprises steps S100-S500, each of which is implemented as follows:

[0057] S100: analyze the user's original query requirement to obtain a plurality of query keywords;

[0058] S200: determine a benchmark large model based on the query keywords;

[0059] S300: input the query keywords into the benchmark large model, and the benchmark large model outputs a plurality of expanded query requirements;

[0060] S400: in response to the user's selection of the expanded query requirement, determine at least one vertical domain corpus;

[0061] S500: load the vertical domain corpus, train the benchmark large model, and obtain a trained vertical domain large model.

[0062] After the vertical domain large model is constructed in step S500, the method can be specifically applied to the current query, that is, the vertical domain large model of the present embodiment realizes query-on-demand and dynamic construction of each query and each call.

[0063] Specifically, referring to the embodiment of Figure 2 , after step S500, the method further comprises:

[0064] S600: input the user's original query requirement into the vertical domain large model, and the vertical domain large model outputs at least one modality form of government service result according to the current attributes of the user terminal.

[0065] In the above embodiment, the vertical domain corpus comprises a plurality of different modalities of government service corpus, and the different modalities comprise one of text, static picture, video frame, speech sequence or any combination thereof;

[0066] Next, the above method embodiment of each step will be specifically introduced in combination with Figures 3-5 . Figure 1 or Figure 2 .

[0067] Step S100: analyze the user's original query requirement to obtain a plurality of query keywords.

[0068] Figure 3The data analysis schematic diagram for analyzing the original query requirement of the user to obtain a plurality of query keywords is provided.

[0069] In a specific application, the original query requirement includes a text query requirement, a voice query requirement, and an image query requirement.

[0070] As Figure 3 Illustratively, examples of the original query requirement include:

[0071] Receiving a voice input of the user: "process and proof materials of XXX";

[0072] Receiving a file scan uploaded by the user, i.e., the user uploads a scan of a filled XXX form through a terminal device, and the user needs to know the next submission window;

[0073] Receiving a text description word directly input by the user: process and proof materials of XXX.

[0074] After receiving the original query requirement of the user, semantic analysis is performed thereon, including voice semantic analysis, text semantic analysis, and image semantic analysis, etc., to obtain a plurality of query keywords.

[0075] Figure 3 In the example, only one keyword XXX obtained by analysis is shown, and it can be understood that other keywords can also be obtained based on the above example, such as "text, picture, window, and proof". Figure 3 In the example, only one XXX is shown.

[0076] Next, step S200 is entered: determining a benchmark large model based on the query keyword;

[0077] The benchmark large model is an industry general large model, and the industry general large model includes a text-to-text industry general large model, a prompt word class text-to-picture industry general large model, and a picture-to-picture industry general large model.

[0078] Industry general large models can be generative artificial intelligence engines well known to those skilled in the art. Generative artificial intelligence can use techniques such as neural networks, deep learning, etc. to identify patterns and structures in existing data to generate new original content. One of the breakthroughs of generative AI models is the ability to train using different learning methods, including unsupervised or semi-supervised learning. This allows organizations to more easily and quickly leverage large amounts of unlabeled data to create base models. As the name implies, base models can serve as the foundation for AI systems that can perform multiple tasks. Examples of base models include GPT-3 and Stable Diffusion, which allow users to harness the power of language. For example, popular applications such as ChatGPT (which is based on GPT-3) allow users to generate an article based on a short text request. On the other hand, Stable Diffusion allows users to generate realistic images based on text input.

[0079] A typical example of a text-to-text industry general large model is the GPT (including GPT1-GPT3) class of technology-enabled artificial intelligence engines, the most typical example being ChatGPT (released by OpenAI on November 3, 2022, and continuously evolving).

[0080] Prompt-based image generation industry general large models use user-initiated text as a prompt to intelligently generate image output results. The most typical example is the Midjourney tool released by David Holz in March 2022. In addition, the DALL E2 tool released by OpenAI in January 2022 also has similar functions.

[0081] Image-to-image industry general large models use partial image description text and an image as a prompt input, intelligently generating preliminary design images. The most typical example is the Stable Diffusion tool released by Stability AI in August 2022, and the FUgenerator tool released by Tongji CAUP in March 2023.

[0082] It should be noted that the three different types of industry general large models described above are continuously evolving, and later-stage industry general large models typically have the functions of earlier-stage industry general large models, but have different focuses. For example, natural language-driven text-to-text industry general large models (such as the evolving ChatGPT) are mainly good at text-to-text, but also have basic text-to-image capabilities; prompt-based image generation industry general large models have both text-to-text and text-to-image capabilities; image-to-image industry general large models can also have text-to-text, text-to-image, and image-to-image capabilities.

[0083] In various embodiments of the present application, when a general industry model is involved, the improved function part thereof is mainly used, for example, for the prompt word class text-to-image general industry model, the text-to-image function thereof is mainly used, and for the image-to-image general industry model, the part function thereof for image-to-image is mainly used, that is, the advantages of each different type of general industry model are fully utilized.

[0084] In step S200, at least one industry large model is determined as a benchmark large model based on the query keyword;

[0085] For example, if the query keyword involves picture input or output, the prompt word class text-to-image model or the image-to-image model is preferentially selected as the benchmark large model; if the query keyword only involves or the user terminal attribute only supports text input and output, only the natural language driven class text-to-text model can be selected as the benchmark large model.

[0086] The benchmark large model can be an open source large model or a commercial paid large model. Preferably, the open source large model is selected in the embodiment to facilitate the subsequent training process. Of course, the use of the open source large model also needs to comply with the corresponding open source rules.

[0087] Step S300: inputting the query keyword into the benchmark large model, and the benchmark large model outputs a plurality of extended query requirements.

[0088] The plurality of extended query requirements is a visual knowledge graph with the query keyword as a root node.

[0089] Specifically, referring to Figure 4 , Figure 4 is a knowledge graph schematic diagram of the present application for outputting a plurality of extended query requirements.

[0090] Taking the natural language driven class text-to-text model as the benchmark large model as an example, the query keyword xxx is input into the benchmark large model, and the benchmark large model outputs a plurality of extended query requirements, including:

[0091] XXX-AXXX; XXX-BXXX; XXX-CXXX; XXX-DXXX-dXX1;

[0092] In order to more intuitively describe the above process, jump to Figure 5 , Figure 5 The visual knowledge graph schematic diagram with the query keyword as a root node in one specific application example of the present application.

[0093] In Figure 5 , the user query keyword includes "education", and after inputting it into the benchmark large model, a plurality of extended query requirements are obtained, including:

[0094] Education - Preschool education - Private schools

[0095] Education - Private education

[0096] Education - Junior high school education

[0097] Education - Preschool education - Kindergarten

[0098] …

[0099] It can be seen that multiple extended query requirements are visual knowledge graphs with the query keyword as the root node.

[0100] Next, step S400 determines at least one vertical domain corpus in response to the user's selection of the extended query requirement.

[0101] The vertical domain corpus includes a plurality of government service corpora of different modalities, including one of text, static image, video frame, speech sequence, or any combination thereof.

[0102] Continuing to refer to Figure 5 The vertical domain corpus composed of a plurality of government service corpora includes:

[0103] Special column corpus, service corpus, convenient query corpus, interactive knowledge base, policy and regulation library, etc.

[0104] In Figure 5 , assuming that the user selects the extended query requirement as the path "Education - Preschool education - Kindergarten", the determined at least one vertical database is "interactive knowledge base".

[0105] Next, step S500 loads the vertical domain corpus, trains the benchmark large model, and obtains the trained vertical large model.

[0106] It can be understood that after determining the training corpus and the query database, the training of the vertical large model and the training of the ordinary (general) large model are basically similar, therefore, the specific training process can be referred to the prior art, and this embodiment will not be further expanded.

[0107] As part of different from the prior art, the embodiment is further preferred as follows:

[0108] Specifically, the step S500 loads the vertical domain corpus, trains the benchmark large model, and obtains the trained vertical large model, specifically including:

[0109] Using a retrieval enhancement generation method, the vertical domain corpus is used as a supplementary training library and an external query library for the benchmark large model;

[0110] adjusting training parameters of the base large model based on a supplementary training corpus;

[0111] The trained vertical large model only queries and retrieves in the plug-in query corpus.

[0112] Correspondingly, the step S600 inputs the original query requirement of the user into the vertical large model, and the vertical large model outputs at least one modal form of government service result according to the current attribute of the user terminal, specifically including:

[0113] S610: generating a plurality of field prompt words based on the original query requirement;

[0114] For example, the field prompt word of preschool education can also be kindergarten, childcare, etc.

[0115] S620: integrating the field prompt word into the original query requirement as additional context information and original query requirement together as the input of the vertical large model;

[0116] S630: the vertical large model retrieves in the vertical corpus;

[0117] S640: when the retrieval result contains multiple modalities, output at least one modal form of government service result according to the current attribute of the user terminal.

[0118] The retrieval result containing multiple modalities in the step S640 means that there are multiple optional retrieval results, for example, video modal output result, picture modal output result and text modal output result.

[0119] However, the current attribute of the user terminal may not support all modalities. For example, the current available graphics card resource and GPU resource of the user terminal are insufficient, so that the video cannot be played, therefore, if the video modal data result is pushed to the user terminal, it may cause lag, and the user may have a poor human-computer interaction experience; at this time, the output result of the modal can be output according to the current attribute of the user terminal; and when the related graphics card resource and GPU resource are sufficient, the output result of the video modal and the output result of the picture modal are pushed again.

[0120] Figure 6 is a structural schematic diagram of a government service vertical large model construction system based on a knowledge graph according to an embodiment of the present application.

[0121] In Figure 6In the specific implementation, the system comprises a plurality of candidate benchmark large models and a plurality of candidate vertical domain corpora, the plurality of candidate benchmark large models comprise a natural language driven text generation model, a prompt word driven text-to-image generation model and an image-to-image generation model, and the plurality of candidate vertical domain corpora comprise a plurality of different modal government service corpora, the different modalities comprising one of text, static image, video frame and speech sequence or any combination thereof.

[0122] The system further comprises:

[0123] The analysis unit analyzes the original query requirement of the user to obtain a plurality of query keywords.

[0124] The benchmark large model determination unit determines a benchmark large model based on the query keywords.

[0125] The knowledge graph expansion unit inputs the query keywords into the benchmark large model, and the benchmark large model outputs a plurality of expanded query requirements; the plurality of expanded query requirements are visual knowledge graphs with the query keywords as root nodes.

[0126] The vertical domain corpus determination unit determines at least one vertical domain corpus in response to user selection of the expanded query requirements.

[0127] The training unit loads the vertical domain corpus and trains the benchmark large model to obtain a trained vertical domain large model.

[0128] Figure 7 FIG. 1 is a schematic diagram of part of the interaction structure of the input unit and the output unit of the knowledge graph-based government service vertical domain large model construction system.

[0129] In the specific implementation, Figure 7 In the specific implementation, the system further comprises:

[0130] The input unit: the user inputs the original query requirement through the input unit, and the original query requirement comprises a text query requirement, a speech query requirement and an image query requirement.

[0131] The result output unit: the original query requirement of the user is input into the vertical domain large model, and the vertical domain large model outputs a government service result in at least one modal form according to the current attribute of the user terminal.

[0132] Specifically, the result output unit is further connected to a user terminal attribute detection unit, the user terminal attribute detection unit is used to detect the resource type currently available to the user terminal, and determine the output modal of the government service result based on the resource type currently available to the user terminal, the output modal comprising one of text output, interactive animation output or picture / video output, so as to meet the user's expectation and improve the user experience.

[0133] Other principles or implementation functions of the system embodiments can refer to the method embodiments, and will not be described again.

[0134] Other technologies, principles, algorithms or models not described in detail in the present application can refer to prior art.

[0135] In the foregoing embodiment section, the present application gives a plurality of embodiments, each of which can constitute an independent technical solution and maycontribute to the prior art and solve the corresponding technical problems.

[0136] However, it should be pointed out that different embodiments can be combined with each other without violating the logic; at the same time, each embodiment can solve at least one technical problem, but it does not require each individual embodiment to solve multiple or all technical problems.

[0137] At the same time, in the specific embodiments of the present application, if user-related data is involved, when the embodiments of the present application are applied to specific products or technologies, the user's permission or consent needs to be obtained, and the collection, use and processing of relevant data need to comply with relevant national and regional laws, regulations and standards.

[0138] It can be seen that the present application first determines the benchmark large model by querying the keyword, and then outputs a plurality of extended query requirements through the benchmark large model; then, in response to the user's selection of the extended query requirement, at least one vertical domain corpus is determined, and then the vertical domain corpus is loaded to train the benchmark large model, so as to obtain the trained vertical large model, which can quickly and effectively construct the government service vertical large model and is easy to use.

[0139] The method embodiments and systems of the present application have been shown and described, but for those skilled in the art, it can be understood that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A government service vertical domain large model construction method based on a knowledge graph, characterized in that, The method comprises the following steps: S100: analyzing the original query requirement of the user to obtain a plurality of query keywords; S200: determining a benchmark large model based on the query keywords; the benchmark large model is an industry general large model, and the industry general large model comprises a text-to-text model, a prompt word class text-to-graph model and a graph-to-graph model; S300: inputting the query keywords into the benchmark large model, and outputting a plurality of expanded query requirements by the benchmark large model; S400: determining at least one vertical domain corpus in response to the selection of the expanded query requirement by the user; S500: loading the vertical domain corpus, and using a retrieval enhancement generation mode to take the vertical domain corpus as a supplementary training library and an external query library of the benchmark large model; training the benchmark large model, adjusting the training parameters of the benchmark large model based on the supplementary training library; obtaining a trained vertical domain large model; the trained vertical domain large model only performs query and retrieval in the external query library; S600: inputting the original query requirement of the user into the vertical domain large model, and outputting a government service result in at least one modal form according to the current attribute of the user terminal by the vertical domain large model; the vertical domain corpus comprises a plurality of government service corpora in different modalities, and the different modalities comprise one of text, static graph, video frame and speech sequence or any combination thereof; the plurality of expanded query requirements are a visual knowledge graph with the query keywords as root nodes.

2. The government service vertical domain large model construction method based on a knowledge graph according to claim 1, characterized in that, In the step S600, the original query requirement of the user is input into the vertical domain large model, and a government service result in at least one modal form is output according to the current attribute of the user terminal by the vertical domain large model, and specifically comprises: S610: generating a plurality of field prompt words based on the original query requirement; S620: integrating the field prompt words into the original query requirement as additional context information and original query requirement together as the input of the vertical domain large model; S630: the vertical domain large model performs retrieval in the vertical domain corpus; S640: when the retrieval result contains a plurality of modalities, a government service result in at least one modal form is output according to the current attribute of the user terminal.

3. The government service vertical domain large model construction method based on a knowledge graph according to any one of claims 1-2, characterized in that, The original query requirement comprises a text query requirement, a voice query requirement and an image query requirement.

4. A government service vertical domain large model construction system based on a knowledge graph, the system comprising a plurality of candidate benchmark large models and a plurality of candidate vertical domain corpora, the plurality of candidate benchmark large models comprising a text-to-text model, a prompt word class text-to-graph model and a graph-to-graph model, and the plurality of candidate vertical domain corpora comprising a plurality of government service corpora in different modalities, and the different modalities comprising one of text, static graph, video frame and speech sequence or any combination thereof; characterized in that the system further comprises: an analysis unit for analyzing the original query requirement of the user to obtain a plurality of query keywords; a benchmark large model determination unit for determining a benchmark large model based on the query keywords; The knowledge graph expansion unit inputs the query keyword into the benchmark large model, and the benchmark large model outputs a plurality of expanded query requirements; the plurality of expanded query requirements are visual knowledge graphs with the query keyword as a root node; The vertical domain corpus determination unit determines at least one vertical domain corpus in response to user selection of the expanded query requirement; The training unit loads the vertical domain corpus and uses a retrieval enhancement generation method to use the vertical domain corpus as a supplementary training library and an external query library of the benchmark large model; The benchmark large model is trained, and the training parameters of the benchmark large model are adjusted based on the supplementary training library; A trained vertical domain large model is obtained; the trained vertical domain large model only performs query and retrieval in the external query library; The result output unit inputs the original query requirement of the user into the vertical domain large model, and the vertical domain large model outputs a government service result in at least one modal form according to the current attributes of the user terminal.

5. The government service vertical domain large model construction system based on a knowledge graph according to claim 4, characterized in that, The system further comprises: The input unit: the user inputs the original query requirement through the input unit, and the original query requirement includes a text query requirement, a voice query requirement, and an image query requirement.

6. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer executable instructions, and the computer executable instructions are executed by the processor to implement the government service vertical domain large model construction method based on the knowledge graph in any one of claims 1 to 3.

Images