Question and answer processing method, device and system

Abstract

The invention discloses a question and answer processing method, device and system, and the method comprises the steps: inputting question information inputted by a user and identity information of the user into a large model, querying at least one piece of candidate knowledge data used for answering the question information from a private domain knowledge base matched with the query intention of the question information, and sending the at least one piece of candidate knowledge data to the user; and processing the at least one piece of candidate knowledge data through the large model, generating target reply information corresponding to the identity information, and outputting the target reply information.

Description

Technical Field

[0001] The present application relates to the field of large model applications, and in particular to a question-answering processing method, device and system. Background Art

[0002] In recent years, large models (LM) have achieved remarkable results in various question-answering tasks. They can accurately understand and respond to human instructions, enabling them to correctly understand and answer complex questions, and have been widely used in various intelligent question-answering scenarios.

[0003] However, for the same question information raised by different users, the current question-answering processing system based on large models outputs the same answer information, which cannot guarantee that the answer information meets the actual needs of all users, affecting the accuracy and reliability of the question-answering. Summary of the invention

[0004] In view of the above problems, this application provides the following solutions:

[0005] The first aspect of the present application provides a question-answering processing method, the method comprising:

[0006] Obtaining question information input by the user and identity information of the user;

[0007] Input the identity information and the question information into a large model, and query at least one candidate knowledge data for answering the question information from a private domain knowledge base that matches the query intent of the question information;

[0008] Processing the at least one piece of candidate knowledge data through the large model to generate target response information corresponding to the identity information;

[0009] Output the target reply information.

[0010] In a possible implementation, inputting the identity information and the question information into a large model, and querying at least one candidate knowledge data for answering the question information from a private domain knowledge base that matches the query intent of the question information, includes:

[0011] Obtaining task prompt information for guiding the large model to process the problem information;

[0012] Inputting the identity information, the question information and the task prompt information into a large model, and guiding the large model to perform intent recognition on the question information according to a first task prompt word included in the task prompt information, to obtain at least one query intent;

[0013] If it is determined that the at least one query intention matches the first query intention for the target business scenario, at least one candidate knowledge data for answering the first query intention is queried from the private domain knowledge base for the target business scenario.

[0014] In a possible implementation, the processing of the at least one candidate knowledge data by the large model to generate target response information corresponding to the identity information includes:

[0015] According to the second task prompt word included in the task prompt information, guiding the large model to select at least one target knowledge data corresponding to the identity information from the at least one candidate knowledge data;

[0016] If the target knowledge data includes multiple pieces, batch-input the multiple pieces of target knowledge data into the large model according to the third task prompt word included in the task prompt information;

[0017] According to the second task prompt word, guide the large model to process each of the target knowledge data to generate target response information corresponding to the identity information.

[0018] In a possible implementation, guiding the large model to process the target knowledge data based on the second task prompt word to generate target response information corresponding to the identity information includes:

[0019] According to the second task prompt, the large model is guided to convert each of the target knowledge data in accordance with the reply method indicated by the identity information, and correspondingly generates first reply information that conforms to the reply method;

[0020] If there are multiple first reply messages, the multiple first reply messages are reorganized to generate target reply messages that conform to the reply method.

[0021] In a possible implementation, after inputting the identity information and the question information into the large model, the method further includes:

[0022] If it is determined that the at least one query intent also includes a second query intent, and the second query intent cannot match the first query intent for the target business scenario, at least one second reply information for replying to the second query intent is generated using the large model;

[0023] The processing of the at least one target knowledge data to generate target response information corresponding to the identity information includes:

[0024] According to the third task prompt word included in the task prompt information, batch input each of the target knowledge data and each of the second reply information into the large model;

[0025] According to the second task prompt word, the large model is guided to process each of the target knowledge data and each of the second reply information at the same time to generate target reply information corresponding to the identity information.

[0026] In a possible implementation, querying at least one candidate knowledge data for answering the first query intention from the private domain knowledge base for the target business scenario includes:

[0027] Querying, from a private domain knowledge base for the target business scenario, multiple pieces of knowledge data matching the query term of the first query intent and their respective knowledge matching degrees;

[0028] At least one candidate knowledge data is screened from the multiple pieces of knowledge data that have been queried according to the knowledge matching degree.

[0029] In a possible implementation, guiding the large model to perform intent recognition on the question information based on the first task prompt word included in the task prompt information to obtain at least one query intent includes:

[0030] According to the first task prompt word included in the task prompt information, guiding the large model to perform intent recognition on the question information, and outputting at least one question information of the query intent;

[0031] Obtaining supplementary information of the question input by the user in response to the question information;

[0032] The at least one query intention is obtained based on the question supplementary information and the question information.

[0033] In a possible implementation, the large model is obtained by training the artificial intelligence control platform on a question-answering task for a target business scenario;

[0034] The process of training the large model for the question-answering task of the target business scenario by the artificial intelligence control platform includes:

[0035] The artificial intelligence control platform obtains question-answering training data;

[0036] Using the question-answering training data, training a large language model to obtain a pre-trained model;

[0037] Obtain task instruction fine-tuning data based on task prompt information for various question-answering processing tasks in target business scenarios;

[0038] The task instructions are used to fine-tune the data and adjust the model parameters of the pre-trained model to obtain the large model.

[0039] In a possible implementation, using the task instruction to fine-tune the data, adjusting the model parameters of the pre-trained model, and obtaining the large model includes:

[0040] Using the task instruction fine-tuning data, adjusting the model parameters of the pre-trained model to obtain a fine-tuning model suitable for the multiple question-answering processing tasks;

[0041] Using the fine-tuned model and the visual encoder, a multimodal large model is constructed;

[0042] Acquire multimodal sample data from a private domain knowledge base for a target business scenario; the private domain knowledge base includes knowledge data of at least two modalities of text, image, audio, and video generated in the target business scenario corresponding to the corresponding private domain;

[0043] According to the task prompt information, the multimodal large model is trained using the multimodal sample data to obtain the large model, which is then deployed to the question and answer processing system in the private domain where the target business scenario is located.

[0044] The second aspect of the present application further provides a question-answer processing device, the device comprising:

[0045] An information acquisition module, used to acquire question information input by a user and identity information of the user;

[0046] A query module, used to input the identity information and the question information into a large model, and query at least one candidate knowledge data for answering the question information from a private domain knowledge base that matches the query intent of the question information;

[0047] A generating module, configured to process the at least one candidate knowledge data through the large model to generate target reply information corresponding to the identity information;

[0048] An output module is used to output the target response information.

[0049] The third aspect of the present application also provides a question-answering processing system, the system comprising:

[0050] Private domain knowledge base, used to store knowledge data for target business scenarios;

[0051] A storage device, used to store an authentication description table, wherein the authentication description table indicates the corresponding relationship between the identity information of different users and each piece of the knowledge data in the private domain knowledge base;

[0052] A large model of electronic equipment is deployed to implement the question-answering processing method as described above. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. Throughout the accompanying drawings, the same or similar reference numerals represent the same or similar elements. It should be understood that the drawings are schematic and the originals and elements are not necessarily drawn to scale.

[0054] Figure 1 A flowchart of a question-and-answer processing method proposed in Example 1 of the present application;

[0055] Figure 2 A flowchart of a question-and-answer processing method proposed in Embodiment 2 of the present application;

[0056] Figure 3 A flowchart of a question-and-answer processing method proposed in Embodiment 3 of the present application;

[0057] Figure 4 A flowchart of a question-and-answer processing method proposed in Embodiment 4 of the present application;

[0058] Figure 5 A flowchart of a question-and-answer processing method proposed in Embodiment 5 of the present application;

[0059] Figure 6 A flowchart of a question-and-answer processing method proposed in Example 6 of the present application;

[0060] Figure 7 A flowchart of a question-and-answer processing method proposed in Embodiment 7 of the present application;

[0061] Figure 8 A schematic diagram of the structure of a question-answer processing device provided in an embodiment of the present application;

[0062] Fig. 9 A structural diagram of a question-answering processing system provided in an embodiment of the present application;

[0063] Fig.10 A scenario diagram of a question-answering processing system provided in an embodiment of the present application;

[0064] Fig.11 A schematic diagram of the hardware structure of an electronic device suitable for the question and answer processing method provided in an embodiment of the present application. DETAILED DESCRIPTION

[0065] The embodiments of the present application are described below in conjunction with the drawings in the embodiments of the present application. The terms used in the implementation mode of the present application are only used to explain the specific embodiments of the present application and are not intended to limit the present application. The embodiments of the present application are described below in conjunction with the drawings. It is known to those of ordinary skill in the art that with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0066] The term "first", "second" etc. that the full text of the application relates to is used to distinguish similar objects, and need not be used to describe a specific order or sequential order. It should be understood that the terms used in this way can be interchangeable in appropriate circumstances, and this is only to describe the distinction mode adopted by the objects of the same attributes when describing in the embodiments of the application. In addition, the terms "including" and "having" and any of their deformations are intended to cover non-exclusive inclusions, so that the process, method, system, product or equipment comprising a series of units need not be limited to those units, but may include other units that are not clearly listed or inherent to these processes, methods, products or equipment.

[0067] It should be understood that before using the technical solutions disclosed in the embodiments of this application, the type, scope of use, usage scenarios, etc. of personal information (such as user identity information) involved in this application should be informed to the user and the user's authorization should be obtained in an appropriate manner in accordance with relevant laws and regulations.

[0068] In order to solve the technical problems described in the background technology section, the embodiment of the present application provides a question-answering processing method. In the process of using a large model to answer the question information input by the user, if the question information needs to be answered using the knowledge data in a private domain knowledge base, the present application will combine the user's identity information to generate target answer information that conforms to the user's identity. In this way, for users with different identity information who input the same question information, different target answer information for different user identities can be generated, so that the output target answer information is more in line with user expectations, and the accuracy and reliability of the question-answering of the large model are improved. Among them, the large model can be trained through an artificial intelligence control (AI master) platform. The question-answering processing method of the embodiment of the present application is described in detail below in conjunction with the accompanying drawings.

[0069] Reference Figure 1 , is a flowchart of a question-answering processing method proposed in the first embodiment of the present application. The question-answering processing method can be applied to an electronic device, which can be a terminal device or a server, such as an application with a question-answering processing function running on the terminal device or a cloud service provided by a cloud server. Figure 1 As shown, the question-answer processing method provided in this embodiment may include but is not limited to the following steps:

[0070] Step S101, obtaining question information input by the user and the identity information of the user;

[0071] In an embodiment of the present application, the question information may be a question input by a user using an input component of a terminal device, and may be one or more combinations of text information, image information, video information, or audio information. For different types of question information, corresponding types of input components may be used to complete the input operation. The present application does not impose any restrictions on the input methods of various types of question information.

[0072] For example, when a user activates the voice assistant of a terminal device, he or she may input questions such as "What's the weather like in Beijing? What attractions can I visit?" and "When is the promotion activity in Hangzhou scheduled?" etc. It can be seen that the query intent contained in the question information input by the user is not limited, and may include one or more query intents. Multiple query intents may be related query intents or independent and unrelated intents, etc. This application does not limit the content of the question information input by the user.

[0073] In addition, the present application can also perform input operations (such as touch operations) on the touch screen or accessories of the terminal device through other input components such as fingers, stylus, mouse keys or joysticks, such as performing the input operation on the question-answering interactive interface displayed on the touch screen, responding to the input operation, and obtaining the question information input by the user. According to actual needs, during the question input process, you can also enter content such as picture or video addresses or other types of file addresses to form the question information raised this time. This application does not elaborate on the input methods of various question information, which can be flexibly determined according to the question input environment.

[0074] In combination with the above description of the technical concept of the present application, the present application proposes to guide the large model to automatically generate target reply information for the question information input by the user according to the identities of different users. Therefore, the present embodiment can obtain the identity information of the user who inputs the question information, such as user account information, user position information in a specified private domain (such as a company), user's professional information or a unique user identifier (such as an ID number, practice number, etc.). The present application does not restrict the content of the user's identity information, which may be determined according to the circumstances.

[0075] It should be understood that, in the case where the electronic device is a server, the terminal device may transmit the question information to the server or send a question-answering request containing the question information to the server after obtaining the question information input by the user according to the method described above but not limited to the method described above. The user's identity information may be obtained by the terminal device and then sent to the server synchronously, or the user's identity information may be obtained when it is determined that a user-leveled reply is required based on the user's identity information based on the question information input by the user.

[0076] Step S102, inputting the identity information and the question information into the big model, and searching for at least one candidate knowledge data for answering the question information from the private domain knowledge base that matches the query intent of the question information;

[0077] Following the above analysis, the electronic device calls the big model to identify the intent of the question information input by the user, and determines that the user is asking questions in a specified private domain (such as a specified field / enterprise, etc.), that is, it determines that one or more query intentions of the question information are query intentions for specified private domain knowledge data, such as a query intention for a certain enterprise information, indicating that the query intention of the question information matches the private domain knowledge base of the specified private domain, and the knowledge data contained in the private domain knowledge base needs to be used to answer the question information. Based on this, the big model can match the question information or its query intention with the knowledge data contained in the private domain knowledge base, and determine the various candidate knowledge data that can be used to answer the question information. This application does not limit the implementation method of the big model querying the candidate knowledge data.

[0078] Among them, the private domain knowledge base can be a system or database built for knowledge data of a specified private domain scenario through large models and artificial intelligence technology. The knowledge data contained in it includes but is not limited to various documents, professional knowledge, customer data, market research, and intellectual property resources in the private domain. It can also be configured not to be disclosed to the outside world (i.e., outside the private domain) as needed to ensure the security and confidentiality of the knowledge data. This application does not limit the content and construction method of the private domain knowledge base corresponding to each private domain.

[0079] Step S103, processing at least one piece of candidate knowledge data through the large model to generate target response information corresponding to the identity information;

[0080] Step S104, outputting the target reply information.

[0081] Based on the above description of the technical concept of the present application, the big model used for question and answer processing in the present application has multi-user hierarchical (i.e., different access rights are assigned to users with different identities) question and answer processing capabilities, so that the big model can answer the same question information raised by users of different levels using appropriate knowledge data within the knowledge range that the user identity allows access to, and may generate different reply information (the differences here include different semantic content of the reply information, or different expression / structure of the reply information, etc.). Compared with the same question raised by users of any identity, the big model can only generate a fixed reply information. The present application generates reply information that matches the user's actual identity through the big model, thereby improving the flexibility and accuracy of question and answer processing and enhancing the user's question and answer experience.

[0082] For example, for users in different departments / levels (i.e., identity information) within the same organization / enterprise, since they are allowed to access different knowledge data of the organization / enterprise, or their accustomed / expected ways of answering are different, when calling the big model to understand and analyze the same question information input by users in different departments / levels, the knowledge data that users in different departments / levels are allowed to access are used, and different response information is generated for users in different departments / levels, thereby ensuring that the response information better meets the actual needs of users in the corresponding departments / levels.

[0083] Afterwards, the generated target reply information corresponding to the user identity can be fed back to the user. If the electronic device is a terminal device, the target reply information can be output through one or more output devices such as a display screen, an audio player or a video player. If the electronic device is a server, the target reply information can be fed back to the terminal device that sent the question information input by the user, and the target reply information is output by at least one output device of the terminal device to inform the user. Optionally, the server can also transmit the target reply information to a pre-configured target device (such as a terminal device, a server or a database, etc.). This application does not limit the output method of the target reply information, which can be determined based on factors such as the type of electronic device and the processing requirements in the actual question-and-answer environment. This application does not give detailed examples one by one.

[0084] To summarize, in the scenario of using a big model for question and answer processing, after the big model queries at least one candidate knowledge data for answering the question information from the private domain knowledge base that matches the query intent of the question information input by the user, the big model will process the queried candidate knowledge data in combination with the user's identity information to generate target reply information corresponding to the identity information. In this way, for the same question information raised by multiple users, the big model will generate different target reply information due to the different identity information of the multiple users. Compared with the problem processing method of only understanding the same question information and generating fixed reply information, the question and answer processing method of the present application can reliably ensure that the generated target reply information is more in line with the actual needs of users with corresponding identity information, thereby improving the accuracy and reliability of the big model question and answer processing.

[0085] Reference Figure 2 , is a flow chart of a question-answering processing method proposed in Embodiment 2 of the present application. This embodiment may be an optional refinement of the question-answering processing method proposed above, such as Figure 2 As shown, the refinement implementation method may include:

[0086] Step S201, obtaining question information input by the user and the identity information of the user;

[0087] Step S202, obtaining task prompt information for guiding the large model to process the problem information;

[0088] In the embodiment of the present application, the task prompt information can help the big model better understand the user's intention, formulate prompt words for each processing task of answering the user's question according to the requirements of the present application for generating answers to the user's question, so as to guide the big model to perform the corresponding processing tasks for the user's question, standardize the content and grammar of the answer to the user's question, and generate more accurate answers to the user's question. Therefore, the task prompt information of the present application may include formulated task prompt words corresponding to each processing task of answering the user's question, that is, prompt words that describe or indicate the processing rules of the corresponding processing task, such as several keywords or sentences, etc. The present application does not limit the content and method of the task prompt information.

[0089] In actual applications, the task prompt information of the present application can be an information set consisting of multiple task prompt words, or a set of multiple prompt word templates including corresponding task prompt words. In addition to containing task prompt words that guide the large model to perform corresponding processing tasks, the prompt template can also include a question field to supplement the user's question information or its query intention or its keywords, so as to form task prompt information that conforms to the input data structure under the prompt mode of the large model, but is not limited to the task prompt information composition method described in this application.

[0090] It should be understood that in the business scenarios / fields corresponding to different private domains, the specification requirements for answering user questions are different, the various processing tasks that the big model needs to perform may be different, and the task prompt information content formulated for this is often different.

[0091] For example, in text generation scenarios such as assisted writing and content creation, the first task prompt information is used to guide the big model to generate articles, stories, reference design drawings or other text content that meet the requirements, thereby improving the efficiency and diversity of text generation; in sentiment analysis scenarios, the second task prompt information is used to guide the big model to more accurately identify the sentiment tendency of user input information; in information extraction scenarios, the third task prompt information is used to guide the big model to more accurately identify and extract the data required by the user, etc. The personalized task prompt information content formulated for different business scenarios is not the same.

[0092] It should be noted that for the same processing requirements in the same type of business scenarios, the same task prompt information can also be formulated, and this application does not limit the method for obtaining the task prompt information.

[0093] Step S203, inputting the user's identity information, the question information and the task prompt information into the big model, and guiding the big model to perform intent recognition on the question information according to the first task prompt word included in the task prompt information, to obtain at least one query intent;

[0094] Combined with the above description of the task prompt information, the task prompt information contains task prompt words corresponding to different processing tasks, and different role templates with processing capabilities corresponding to different processing tasks are formulated for the big model, so that the big model can play (i.e., play or simulate) different roles according to different task prompt words, so that the big model has the task processing capabilities of the corresponding actual roles, that is, build a role agent (Agent, i.e., intelligent body) of the corresponding role based on the big model, autonomously execute behaviors that meet the characteristics and goals of the role, and efficiently and accurately generate content that meets the task requirements. This application does not elaborate on how to use the big model as the agent kernel, build the corresponding role agent according to the formulated task prompt words (i.e., the big model plays the corresponding role), that is, the big model plays multiple roles and collaboratively completes the question-and-answer task.

[0095] Based on this, it is assumed that the roles involved in the formulation of the question-and-answer task include reviewers, dispatchers, and authenticators, that is, based on the big model as the agent core, the implementation process of constructing a review agent group (ReviewAgent) containing these three role agents for the question-and-answer task is not described in detail in this application. Among them, the reviewer role agent is responsible for disassembling user questions (Query), and the dispatcher role agent is responsible for screening, filtering, and guiding the answers of the big model; the dispatcher role agent can be responsible for parallel organization of user problem points, batch input of the big model to speed up the answer efficiency, etc. The processing task content of the question-and-answer scenario formulated by different role agents can be determined according to the actual situation, including but not limited to the formulation content described in this embodiment.

[0096] In the above example, this application can record the prompt words formulated for the three role agents of reviewer, dispatcher and authenticator as the first task prompt words, the second task prompt words and the third task prompt words in sequence, so that the big model executes the processing tasks formulated for the reviewer according to the first task prompt words, that is, guides the big model as the reviewer role agent to identify the intent of the question information input by the user, and determine one or more query intents (i.e., intent query) contained in the question information. This application does not elaborate on the implementation process of how the big model identifies the query intent of the question information input by the user.

[0097] In a possible implementation, the present application can construct a problem disassembly template (a task prompt template) for the question-and-answer task processing capability formulated for the role of the reviewer, such as the user problem disassembly capability, to implement the formulated capability of the reviewer role, and the problem disassembly template is used to represent the disassembly method / rule for the user's problem. In this way, the large model can call the pre-formulated problem disassembly template based on the first task prompt word for the reviewer role, disassemble the question information input by the user, and then identify the intent of each sub-question obtained by the disassembly, that is, predict the intent of each sub-question, so as to accurately and quickly determine the query intents contained in the question information input by the user. It can be seen that the first task prompt word in this embodiment can be a prompt message that guides the large model to call the problem disassembly template to disassemble the question information input by the user. The present application does not limit the content of the prompt message for implementing the problem disassembly task.

[0098] Step S204: if it is determined that the at least one query intent matches the first query intent for the target business scenario, query at least one candidate knowledge data for answering the first query intent from the private domain knowledge base for the target business scenario;

[0099] Following the above analysis, the reviewer role established in this application is also responsible for assessing whether the question information input by the user has a query intention for the business information in the target business scenario (i.e., the knowledge data in the private domain knowledge base of the target business scenario). If so, it is necessary to query the matching knowledge data from the private domain knowledge base for each corresponding query intention, that is, to query the candidate knowledge data used to answer the query intention; if not, to directly guide the big model to understand each query intention and generate target reply information, which is equivalent to directly calling the big model to identify the intent of the question information input by the user, and generate target reply information that meets the query intention and feed it back to the user.

[0100] In an embodiment of the present application, if the business information query intention for the target business scenario (i.e., the knowledge data query intention for the private domain knowledge base of the target business scenario) is recorded as the first query intention, an application scenario in which each query intention of the question information input by the user matches the first query intention is taken as an example for explanation. At this time, each sub-question obtained from the question information is a question related to the target business scenario. In order to answer each sub-question, it is necessary to query at least one candidate knowledge data for answering each sub-question from the private domain knowledge base for the target business scenario. The present application does not limit the query implementation method of the candidate knowledge data.

[0101] Step S205, guiding the large model to select at least one target knowledge data corresponding to the identity information of the user from at least one candidate knowledge data according to the second task prompt word included in the task prompt information;

[0102] Following the above analysis, the second task prompt word of the embodiment of the present application can be a prompt word formulated for the role agent of the authenticator. According to the question and answer task processing capabilities formulated for the authenticator role above, such as screening, filtering and guiding the answers of the large model, a corresponding knowledge screening prompt template can be constructed. Through this knowledge screening prompt template, it is instructed to read the authentication description table that allows the user identity to access the knowledge content, so as to implement the screening rules for further screening of the candidate knowledge data. This application does not limit the method of constructing the knowledge screening prompt template.

[0103] Among them, the authentication description table can be a description of the identity classification (such as different departments or job levels, etc.) requirements and access rights of users in the private domain (such as employees within an organization / enterprise), indicating which knowledge data in the corresponding private domain knowledge base users with what identity information are allowed to access, and answering the user's questions accordingly. This application does not restrict the output method of the authentication description table. The correspondence between the identity information of users with different identities represented by it and the description of allowed access to knowledge data can be recorded through a table, relational expression or other output method.

[0104] According to the description of the second task prompt word above, after the big model obtains each candidate knowledge data used to answer the user's query intention for the target business scenario knowledge, it can simulate the authenticator role according to the second task prompt word, use the big model as the role agent to read the authentication description table, and further screen each candidate knowledge data according to the permission description corresponding to the current user's identity information, and determine the candidate knowledge data that matches the user's identity as the target knowledge data for answering the user's question with this identity information.

[0105] For example, ordinary employees and department managers in the product sales department of an enterprise have different job levels, and the scope of knowledge they are allowed to access within the enterprise is different. The scope of knowledge that ordinary employees can access is often smaller than the scope of knowledge that some managers can access. At this time, if the department manager and ordinary employee A ask the same question about the enterprise's product sales knowledge, such as "How are the product sales in the past six months?", since the authentication description table indicates that employee A is allowed to view the sales of the products he is responsible for (such as recording the identification of each product that employee A is responsible for), the department manager is allowed to view the sales of all products that each employee in the department is responsible for. The big model is used to query the sales data of each product generated in the past six months from the enterprise knowledge base (that is, a private domain knowledge base), that is, candidate knowledge data. The big model acts as an authenticator role agent, and based on the authentication description table, it selects the sales data of each product that employee A is responsible for as the target knowledge data for answering user A's question; for the department manager, the sales data of each product that the product sales department is responsible for will be selected as the target knowledge data for answering some managers' questions.

[0106] Similarly, for other users such as employees, managers, or business owners, when they enter the question "How are the product sales in the past six months?", since different identity users set in the authentication description table are allowed to access different scopes of corporate knowledge, the large model is used to simulate the reviewer role and the authenticator role. According to the question-and-answer processing method described above, the target knowledge data screened to answer the same questions of users with different identities can be different, which improves the diversity and accuracy of the knowledge data screened and generated by the large model to answer the questions.

[0107] Step S206, processing at least one piece of target knowledge data to generate target reply information corresponding to the user's identity information;

[0108] Step S207, output the target reply information.

[0109] According to the above description of the second task prompt words, the processing tasks assigned to the authenticator simulated by the large model include answering the guided large model, that is, generating reply information. After the target knowledge data is obtained through the above screening operation, it is also possible to construct a task prompt template for processing the target knowledge data, such as a reply angle conversion template or a viewpoint reorganization template, etc., so as to continue to use the authenticator role agent constructed by the large model to process each target knowledge data according to the corresponding task prompt template after screening out the target knowledge data that meets the identity of the current user, and generate the target reply information corresponding to the identity information of the current user, and feed it back to the user to ensure that the target reply information meets the question and answer needs of the user with this identity.

[0110] In some embodiments, in order to improve the efficiency of question and answer processing, the big model can also simulate the dispatcher, and based on the third task prompt words (prompt words formulated for the role of the dispatcher to guide the big model to play the role of the dispatcher and realize the ability to process the question and answer tasks formulated by the dispatcher, that is, parallel processing), the screened multiple target knowledge data batches (the size, that is, the number of inputs each time, can be determined according to the situation) are input into the big model, thereby guiding the big model to process each target knowledge data at the same time according to the second task prompt words to generate target reply information.

[0111] It can be seen that in the embodiment of the present application, in order to enable the big model to play different roles in the private domain (that is, as agents of different roles), according to the prompt templates (that is, task processing strategies / rules) formulated for the processing tasks that the corresponding roles are responsible for, the question and answer tasks proposed by users with different identities are collaboratively completed, and the user identity information is combined to accurately screen the target knowledge data used to answer the user, and autonomously generate the target answer information corresponding to the identity information. Compared with the question and answer processing method in which the big model generates fixed answer information for the same question information input by users of different identities, the big model of the present application has the ability to grade and answer users of multiple identities based on the task prompts of different roles, and the diverse answers to users of different identities improve the efficiency and accuracy of question and answer, and enhance the question and answer experience of users of different identities.

[0112] In some embodiments, combined with the above description of the big model playing different roles and collaboratively completing question-and-answer tasks, for the above-mentioned first task prompt words, second task prompt words and third task prompt words, in addition to the method described in the above embodiments, the task prompt words used to guide the big model to play different roles are included in the task prompt information, and input together with the question information input by the user and the user's identity information into the big model to guide the big model to perform corresponding processing tasks one by one. In addition to the implementation method, the electronic device can also input the corresponding task prompt words and required data into the big model according to the execution order of each role's responsibility for processing tasks, execute the corresponding processing tasks, and generate target reply information corresponding to the user's identity information.

[0113] Based on this, refer to Figure 3 FIG. 1 is a flow chart of a question-answering processing method proposed in Embodiment 3 of the present application. Figure 3 As shown, the method may include:

[0114] Step S301, obtaining question information and a first task prompt word input by a user;

[0115] Step S302, inputting the question information and the first task prompt word into the big model, guiding the big model to query at least one candidate knowledge data for answering the question information from the private domain knowledge base that matches the query intent of the question information;

[0116] In combination with the above description of the prompt words for the first task, in a possible implementation, the present application can be combined with the question-and-answer task processing capabilities developed by the corresponding role (such as the above-mentioned reviewer), develop a corresponding prompt word template containing the prompt words for the first task, fill the question information input by the user into the corresponding field of the prompt word template, and obtain the question prompt word prompt that the big model can recognize, recorded as the first question prompt information. After inputting the big model, the big model extracts the query intent of the question information based on the first question prompt information, queries the private domain knowledge base that matches it, and obtains candidate knowledge data for answering the corresponding query intent. The query implementation process can refer to the description of the corresponding part of the above embodiment.

[0117] Of course, as described in the above steps of this embodiment, the question information and the first task prompt word input by the user can also be directly input into the large model, so that it can play the corresponding role (that is, act as an agent of the corresponding role), independently complete the processing tasks formulated for the role, and obtain various candidate knowledge data for answering the question information.

[0118] Among them, in the above implementation process, the electronic device can implement the above steps by calling the API (Application Programming Interface); or the formulated task prompt words can be embedded in the electronic device system, so that when calling the big model to perform the question and answer task scenario, it can obtain appropriate task prompt words, guide the called big model to perform the corresponding task processing (that is, collaboratively implement a subtask of the question and answer task) according to the task processing rules / strategies indicated by the task prompt words, etc. This application does not impose any restrictions on the management method of the various task prompt words used for question and answer tasks, which can be determined according to the circumstances.

[0119] Step S303, obtaining the user's identity information and the second task prompt word;

[0120] Step S304, inputting the queried candidate knowledge data, the identity information and the second task prompt word into the big model, guiding the big model to screen multiple target knowledge data corresponding to the user's identity information from the candidate knowledge data;

[0121] Step S305, obtaining a third task prompt word;

[0122] Step S306, according to the third task prompt word, batch-input the selected multiple target knowledge data into the large model, and according to the second task prompt word, guide the large model to process the multiple target knowledge data at the same time to generate target reply information corresponding to the user's identity information;

[0123] Step S307, output the target reply information.

[0124] In combination with the above description of the implementation method of step S302, in a possible implementation, the queried candidate knowledge data, the user's identity information and the second task prompt word can be used to obtain the second question prompt information. For example, based on the prompt word project of the big model, the candidate knowledge data and the user's identity information are filled into the corresponding field of the task prompt template containing the second task prompt word to obtain the second question prompt information that can be recognized by the big model, and then input into the big model so that the big model simulates the corresponding role (such as authenticator), executes the processing task formulated for the role, and filters at least one target knowledge data corresponding to the user's identity information from the candidate knowledge data, that is, the candidate knowledge data that the identity user is allowed to access.

[0125] In the implementation process of step S306, since the third task prompt for the dispatcher guides the big model to adopt a batch processing method (also known as a parallel processing method, that is, a batch of multiple target knowledge data are input into the big model each time), compared with sequentially processing each target knowledge data (that is, inputting one target knowledge data into the big model for processing each time), the processing efficiency is greatly improved. Therefore, after the authenticator transmits the screened target knowledge data to the dispatcher, these target knowledge data can be batch-input into the big model according to the third task prompt. Afterwards, the big model still simulates the role of the authenticator and processes multiple target knowledge data in parallel according to the second task prompt to generate target reply information corresponding to the user identity information. The implementation process is not described in detail in this application.

[0126] It can be seen that this application uses a large model as the agent kernel to construct three role agents such as reviewer, authenticator and dispatcher, each of which performs the processing tasks assigned to the corresponding role, and uses the description or instructions of how to perform each processing task (such as task processing strategy / rule description or instructions) as the corresponding prompt words, such as Figure 4 As shown, after users with different identities input question information about the business in the private domain, the big model will be called according to the execution order of each role agent based on the corresponding task prompt word, and the target knowledge data that the identity user can access to the private domain knowledge base and can answer the input question information will be automatically screened, and the target answer information of the corresponding user identity will be generated quickly and accurately, meeting the different answer requirements of different identity information for the same question, and improving the diversity of the big model's answers to the same question input by users with different identities.

[0127] Reference Figure 5 , is a flowchart of a question-answering processing method proposed in Example 5 of the present application. This embodiment can provide a detailed description of the implementation process of different roles in a large model simulation to perform a question-answering task, but is not limited to the detailed implementation method described in this embodiment. Figure 5 As shown, the question-answering processing method proposed in this embodiment may include:

[0128] Step S501, obtaining question information input by the user and the identity information of the user;

[0129] Step S502, obtaining task prompt information for guiding the large model to process the problem information;

[0130] Step S503, inputting the user's identity information, the question information and the task prompt information into the big model, and guiding the big model to perform intent recognition on the question information according to the first task prompt word contained in the task prompt information, to obtain multiple query intents;

[0131] Regarding the implementation method of step S501-step S503, you can refer to the description of the corresponding part of the above embodiment, such as the description content of the corresponding part of step S201-step S203, or the description content of the corresponding part of step S301 and step S302, etc. The embodiments of the present application will not be repeated here.

[0132] Step S504, determining whether the multiple query intentions match the first query intention for the target business scenario; if at least one query intention matches the first query intention, proceeding to step S505; if at least one second query intention does not match the first query intention, executing step S509;

[0133] In the embodiment of the present application, the first query intention refers to the query intention for the target business scenario, that is, the user inputs question information for the information in the target business scenario (that is, the knowledge data in the corresponding private domain, such as information in the enterprise), and there is a query intention for the relevant information in the target business scenario. The content of the intention can be determined based on the type of the target business scenario and the information content it contains. For the convenience of description, the query intention that cannot be matched with the first query intention is recorded as the second query intention. In other words, the second query intention is a query intention that is unrelated to the information in the target business scenario, and the user expects the answer to the question to be irrelevant to the target business scenario. This application does not restrict the content of the first query intention and the second query intention.

[0134] Exemplarily, assuming that the target business scenario refers to a business scenario supported by a certain enterprise / organization, the production and sales scenario of Class A products is taken as an example. The question information input by the user is "What is the production situation of Product 1 in July in the past three years? What fun places are there in Hangzhou?" The big model identifies the intent of the question information in the above manner, and the query intention for the production data of Product 1 in July of each year in the past three years matches the first query intention, and the query intention can be recorded as the first query intention; the query intention for Hangzhou attractions determined from the question information is irrelevant to the target business scenario, and does not match the first query intention defined in this application, and can be recorded as the second query intention.

[0135] Step S505, searching for multiple pieces of knowledge data matching the query term of the first query intent and their respective knowledge matching degrees from the private domain knowledge base for the target business scenario;

[0136] Step S506, screening multiple candidate knowledge data from the multiple pieces of queried knowledge data according to the knowledge matching degree;

[0137] In the candidate knowledge data screening method proposed in the embodiment of the present application, for each first query intention (i.e., intent query) identified from the question information, the matched knowledge data can be first queried from the corresponding private domain knowledge base. Since the higher the knowledge matching degree between the knowledge data and the corresponding first query intention, the higher the accuracy of the knowledge data in answering the first query intention, and the higher the probability of meeting the user's answer expectations, the present application can select several knowledge data with higher knowledge matching degrees as candidate knowledge data for answering the first query intention based on the knowledge matching degrees corresponding to the queried knowledge data.

[0138] Optionally, the present application may select the first number (the present application does not limit its value, because the larger the value, the greater the subsequent calculation amount, but the lower the probability of missing the actual required knowledge data, the higher the accuracy of the question answering, the present application may determine the value of the first number based on the actual requirements for question-answering efficiency, real-time and accuracy) of candidate knowledge data with high knowledge matching degrees according to the knowledge matching degree sorting results of each piece of knowledge data, starting from the highest knowledge matching degree. In another possible implementation, the knowledge data corresponding to the knowledge matching degree greater than the matching threshold may also be directly determined as candidate knowledge data, etc. The present application does not limit the implementation method of step S506.

[0139] Step S507, guiding the large model to screen multiple target knowledge data corresponding to the user's identity information from multiple candidate knowledge data according to the second task prompt word included in the task prompt information;

[0140] Combined with the above analysis, the big model acts as an agent of the reviewer under the guidance of the first task prompt word. After independently screening out the candidate knowledge data of each first query intention used to answer the question information from the private domain knowledge base, it can send it and the user identity information to the authenticator. According to the permission description corresponding to the user's identity information read from the authentication description table (such as the description of the scope of knowledge allowed to be accessed by the identity user), the received candidate knowledge data is further screened to obtain the target knowledge data corresponding to the user's identity information, that is, the candidate knowledge data allowed to be accessed by the user with the identity information is screened as the target knowledge data, and the target knowledge data meets the permission description of the identity user in the authentication description table.

[0141] Step S508, according to the third task prompt word included in the task prompt information, a plurality of target knowledge data are batch-inputted into the large model, and according to the second task prompt word, the large model is guided to convert each target knowledge data according to the reply method indicated by the user's identity information, and correspondingly generates at least one first reply information that conforms to the reply method;

[0142] In some embodiments, when the query intent of the question information matches the first query intent, the screened target knowledge data can be directly processed through the big model to generate target response information for the question information input by the user.

[0143] In order to ensure that the reply information fed back to the user meets the user's reply expectations, the large model of this application will play the role of the dispatcher according to the third task prompt, that is, as a role agent of the dispatcher, autonomously execute the processing tasks pre-set for the dispatcher role. Based on this, the reply method indicated by the user's identity information in this application may include but is not limited to the sentence structure of the reply information, the information format (such as tables, pictures or videos, etc.), the expression method of the reply content that the identity user is accustomed to / expected, the description angle, or the pre-learned language style (such as formal style, informal style, academic style, technical style, humorous style, expression style and relaxed style) / attitude that the identity user likes / is accustomed to, and other dialogue methods, so as to improve the diversity of questions and answers and meet the needs of users with different identities.

[0144] For reply methods with different identity information indications, corresponding task prompt templates can be pre-formulated, such as a question reply angle conversion template. In this way, after using the large model to act as an authenticator to screen out each target knowledge data from the private domain knowledge data, that is, the knowledge data that matches the user's identity and can answer the user's first query intention of the question information input, the target knowledge data can be reorganized from the perspective of the identity user's concern according to the pre-formulated reply angle conversion template, and the first reply information described from the perspective of the identity user's concern can be obtained.

[0145] Optionally, the big model can also generate candidate response information for answering the corresponding first query intention based on the target knowledge data, and then reorganize the viewpoints of the candidate response information to generate the first response information in a response method that conforms to the user's identity. This application does not limit the implementation method of how the big model, acting as a dispatcher, generates the first response information in a response method that conforms to the user's identity based on the target knowledge data.

[0146] In order to improve the efficiency of question and answer processing, if the big model acts as an authenticator to screen out multiple target knowledge data and generates multiple first reply messages, the third task prompt can be used to guide the big model to act as a dispatcher role agent, and input multiple target knowledge data (multiple first reply messages) into the big model in batches (that is, organized in parallel into batches for acceleration, and this application does not limit the size of the batch) according to the pre-established parallelization template. Then, based on the second task prompt, guide the big model to continue to perform parallel conversion processing on the multiple target knowledge data to quickly generate the corresponding multiple first reply messages.

[0147] It can be seen that the big model serves as the base for the three roles of reviewer, authenticator and dispatcher. Through the logic code of the corresponding role, the big model is called to implement the processing tasks assigned to the corresponding role. In the above implementation process, after the authenticator screens out all the target knowledge data used to reply to the first query intention, in order to improve the subsequent processing efficiency, it is sent to the dispatcher. According to the parallel organization capability, these target knowledge data are input into the big model in batches, and then the authenticator is triggered to continue to guide the big model to execute the subsequent reply generation steps. Compared with the processing method in which the authenticator sequentially inputs each screened target knowledge data into the big model to generate the corresponding first reply information, the parallel processing method of this embodiment improves the question and answer processing efficiency.

[0148] It should be noted that the intent recognition is performed on the question information input by the user. If the question is simple, a query intent may be recognized. In this case, it is determined that the query intent matches the first query intent, and the first reply information used to reply to the first query intent can be generated through the implementation method of steps S505-S508. However, it is not limited to the query method for candidate knowledge data and the method for generating the first reply information described in the above steps, and can be adaptively adjusted according to actual conditions. The embodiments of the present application will not be described in detail here.

[0149] Among them, if each query intention of the question information input by the user matches the first query intention, according to the implementation method described above, after generating multiple first reply information, the multiple first reply information can be directly reorganized (the dispatcher can input multiple first reply information into a large model in batches, and then reorganize all the first reply information) to generate target reply information that conforms to the corresponding reply method, such as reorganizing the question answer points represented by each first reply information (such as summarizing / extracting the main idea of ​​each first reply information), generating reply information with smooth context sentences and concise and concise, etc. This application does not impose any restrictions on the reorganization implementation method.

[0150] Step S509, using the large model, generating at least one second reply information for replying to the second query intention;

[0151] In actual applications, if the question information is a complex question, the big model acts as a reviewer to disassemble the question information, and multiple query intentions (i.e., intent queries) will be obtained. These multiple query intentions may partially or completely match the first query intention. In this case, the subsequent processing steps can refer to the description of the corresponding part of the above embodiment; these multiple query intentions may also include at least one second query intention, that is, one or more intent queries that cannot match the first query intention. The answer to the second query intention does not involve the knowledge data in the private domain knowledge base for the target business scenario. Therefore, for each determined second query intention, it can be input into the big model to directly generate at least one second reply information to reply to the second query intention, or the question information can be directly input into the big model to generate the target reply information. This processing process is directly implemented using the basic question and answer processing capabilities of the big model, and the implementation process is not described in detail in this application.

[0152] According to the above analysis, it can be understood that the reviewer performs intent recognition on the question information input by the user, and the various query intents obtained cannot match the first query intent, that is, they are all second query intents. The large model does not need to act as an authenticator to perform processing tasks. It can use the original question-and-answer processing capabilities of the large model to generate second reply information to answer each second query intent. After the second reply information is reorganized, the target reply information is generated. Preferably, the large model can adopt a batch processing method to generate second reply information corresponding to multiple second query intents. After repeating the viewpoints, the target reply information that conforms to the reply method indicated by the user's identity information is quickly generated. The reply information generated to answer each query intent can be one or more. It depends on the situation.

[0153] Step S510, according to the second task prompt word, guiding the large model to reorganize the at least one first reply information and the at least one second reply information to generate target reply information that conforms to the above reply method;

[0154] Combined with the above description of the processing tasks formulated for the big model simulation dispatcher, in the embodiment of the present application, according to the method described above, the reply information corresponding to each query intention contained in the question information input by the user is input into the big model, and the big model acts as the role agent of the dispatcher, and reorganizes multiple reply information according to the viewpoint reorganization template formulated for the user identity, and quickly generates the target reply information, ensuring that the content of the target reply information is contextually coherent, semantically smooth, and consistent with the reply method of the user of this identity. It should be understood that for the reply methods of different contents listed above, the processing method of each reply information is different, and the processing implementation process is not described in detail in this application.

[0155] In one possible implementation, the corresponding second reply information generated by the big model for each second query intention, and the target knowledge data screened by the big model simulation authenticator for each first query intention, can be input into the big model, such as being sent to the dispatcher, and these data can be input into the big model in batches according to the third task prompt. At this time, the big model will continue to simulate the reply method indicated by the authenticator according to the user's identity information, batch reorganize each target knowledge data and each second reply information, and generate target reply information that conforms to the reply method.

[0156] Step S511, output the target reply information.

[0157] For the target reply information generated by the large model simulation dispatcher, format conversion processing can also be performed according to the output type requirements. For example, if the user requires a voice reply, a text reply, or an image / video reply, target reply information of the corresponding output type (such as one or more data types such as audio, text, image or video) can be generated and sent to the user terminal device for output. This application does not describe the format conversion processing process in detail.

[0158] In summary, in the embodiments of the present application, for the three roles of reviewer, authenticator and dispatcher who perform question-and-answer tasks, corresponding task processing rules / strategies are formulated for the subtasks (i.e., processing tasks) contained in the question-and-answer tasks assigned to them, as well as the required task prompt templates, and task prompt words describing the corresponding processing tasks are configured, so as to use the same large model to simulate the corresponding roles to perform question-and-answer processing, that is, after the reviewer obtains the question information input by the user, the question information is input into the large model for disassembly, that is, the large model is guided to simulate the reviewer to identify the intent of the question information, and obtain the query intents in the question information. If the reviewer confirms It is determined that at least one query intention matches the first query intention for the target business scenario. The candidate knowledge data with a high knowledge match degree with each first query intention will be queried from the corresponding private domain knowledge base and sent to the authenticator. The target knowledge data that matches the user identity will be screened and input into the big model in batches through the dispatcher. Alternatively, after the authenticator converts and processes each target knowledge data and obtains the first reply information in a reply method that matches the user identity, it will be input into the big model in batches through the dispatcher. The authenticator uses the big model to reorganize the viewpoints of each reply information, generates the final target reply information, and feeds it back to the user.

[0159] It can be seen that in the process of using the big model to simulate the three roles of reviewer, authenticator and dispatcher to collaboratively complete the question-and-answer task, the reviewer determines that the user has the intention to query the knowledge data in the target business scenario. After screening out the matching candidate knowledge data from the corresponding private domain knowledge base, the authenticator will combine the user's identity information to further screen the target knowledge data that meets the knowledge access rights of the user's identity (that is, the knowledge scope allowed to be accessed by the user corresponding to the identity information). Compared with the method of generating the same reply information for users of different identities based on the candidate knowledge data, this application will generate different target reply information corresponding to the user's identity based on the different target knowledge data corresponding to users of different identities, thereby improving the diversity and accuracy of the big model's questions and answers, and enhancing the question-and-answer experience of users of different identities. In addition, this application combines the batch processing method provided by the dispatcher to generate target reply information that meets the reply method of the user's identity, thereby improving the efficiency of question-and-answering and better meeting the reply expectations of users of different identities.

[0160] In some embodiments, during the execution of the question-answering processing method proposed in any of the above embodiments, the accuracy of intent recognition can be improved through interactive question-answering, that is, the large model can determine that the user needs to complete the question based on the intent recognition result of the question information input by the user, and can output the corresponding question information for the user to input supplementary information based on the question, and then obtain the user's query intent through intent recognition. Based on this, the above large model simulates the reviewer to perform intent recognition on the question information, and the implementation process of obtaining at least one query intent can include but is not limited to the following implementation methods:

[0161] Based on the first task prompt word included in the task prompt information, guide the large model to identify the intent of the question information, and output question information of at least one query intention, such as displaying it on the question and answer interface of the terminal device used by the user, or outputting the question information by voice, etc., so that the user can input supplementary information of the question accordingly. In response to the user input operation, obtain the supplementary information of the question input by the user for the question information, and thus obtain at least one query intention based on the supplementary information and the question information, such as guiding the large model simulation reviewer to re-identify the intent of the supplementary information and the question information based on the first task prompt word, or optimizing the obtained intent identification results based on the supplementary information of the question, so as to reliably obtain the user's true query intention. The intention identification process is not described in detail in this application.

[0162] The large model in the above question-answering processing method embodiments can be obtained by training the question-answering task of the target business scenario through the artificial intelligence control platform (AIMaster platform), referring to Figure 6 As shown in the flowchart of the question-answering processing method, the implementation process of training a large model on an artificial intelligence control platform may include but is not limited to the following steps:

[0163] Step S601, the artificial intelligence control platform obtains question-answering training data;

[0164] In the embodiments of the present application, in order to ensure that the large model is applicable to the question-answering processing method proposed in the embodiments of the present application, the question-answering training data may include dialogue data in the question-answering format, which may be obtained from a third-party platform or an open source database or historical question-answering data of the question-answering processing system in various scenarios. The present application does not restrict the source of the question-answering training data. In order to improve the robustness of the model, question-answering training data for different question-answering tasks in various scenarios may be obtained, and the present application does not restrict the data content.

[0165] Step S602, using the question-answer training data to train a large language model to obtain a pre-trained model;

[0166] During the model pre-training stage, the open source large language model (LLM) can be trained using question and answer training data in text format. For example, the question and answer training data including question training data is input into the LLM to generate corresponding predicted reply data, and the prediction loss between the predicted reply data and the reply training data included in the question and answer training data is obtained. The model parameters of the large language model are adjusted accordingly, and a pre-trained model suitable for question and answer scenarios is obtained through training. The present application does not impose any restrictions on the network structure of the pre-trained model. The model pre-training implementation process of step S602 includes but is not limited to the implementation method described in the present application.

[0167] Step S603, obtaining task instruction fine-tuning data according to task prompt information for various question-answering processing tasks in target business scenarios;

[0168] Step S604, using the task instructions to fine-tune the data, adjust the model parameters of the pre-trained model, and obtain a large model.

[0169] In order to obtain a large model suitable for the question-and-answer processing method proposed in the embodiment of this application, the pre-trained model can be fine-tuned by fine-tuning the instructions. The task instruction fine-tuning data used to train the large model to simulate the corresponding role can be obtained based on the task prompt words contained in the task prompt information. The pre-trained model can be fine-tuned accordingly, and a large model (i.e., a generative AI model) that can simulate different roles such as reviewers, authenticators, and dispatchers to collaboratively complete the question-and-answer task can be trained. Regarding the content of the task prompt information and its formulation / construction process, please refer to the description of the corresponding part of the above embodiment, which will not be described in detail in this embodiment.

[0170] After the AI ​​Master platform trains the big model according to but not limited to the method described above, it can perform a question-and-answer test on the big model according to the private domain knowledge base of one or more business scenarios and the corresponding question-and-answer test data in accordance with the question-and-answer processing method proposed in this application. After determining that the test passes, the big model is deployed to an electronic device so that the electronic device supports the question-and-answer processing method proposed in this application. In this way, for the question-and-answer tasks of the target business scenarios, task prompt words can be pre-formulated to guide the big model to simulate the different roles mentioned above to perform corresponding processing tasks, as well as various task prompt templates required to perform corresponding processing tasks, that is, role agents (intelligent agents) of different roles are constructed based on the big model, and they are associated and stored with the application program (the application corresponding to the question-and-answer processing system) that executes the question-and-answer processing method in the electronic device. In this way, after receiving the question information input by the user, the deployed big model can be retrieved by calling an API or the like, the question-and-answer processing method proposed in this application is executed, and the generated target reply information is fed back to the user.

[0171] Reference Figure 7 , is a flow chart of a question-answering processing method proposed in Embodiment 7 of the present application. In order to meet the multimodal question-answering processing requirements, the large model applicable to the question-answering processing method of the present application may be a multimodal large model. Based on this, Figure 7 As shown in the figure, the implementation process of AI Master platform training multimodal large models can include:

[0172] Step S701, the artificial intelligence control platform obtains question-answering training data;

[0173] Step S702, using the question-answer training data to train a large language model to obtain a pre-trained model;

[0174] Step S703, obtaining task instruction fine-tuning data according to task prompt information for various question-answering processing tasks in target business scenarios;

[0175] Step S704, using the task instruction to fine-tune the data, adjusting the model parameters of the pre-trained model to obtain a fine-tuned model suitable for a variety of question-answering processing tasks;

[0176] Regarding the pre-training and instruction fine-tuning of the large language model, the implementation process of training to obtain the fine-tuning model can be described in the corresponding part of the above embodiment, such as the description of the corresponding part of step S601-step S604, and this embodiment will not be repeated. Among them, in the case where the question-answering training data is a text type data trained to obtain a fine-tuning model, the various question-answering processing tasks applicable to the fine-tuning model can be text question-answering tasks, which include user input text question information, or a question-answering scenario in which the audio text information input by the user is converted into text information.

[0177] Step S705, using the fine-tuning model and the visual encoder to build a multimodal large model;

[0178] Step S706, obtaining multimodal sample data from a private domain knowledge base for the target business scenario; the private domain knowledge base includes knowledge data of at least two modalities of text, image, audio and video generated in the target business scenario corresponding to the corresponding private domain;

[0179] Step S707, based on the task prompt information, use the multimodal sample data to train the multimodal large model to obtain the large model, so as to deploy it to the question and answer processing system in the private domain where the target business scenario is located.

[0180] In order to meet the needs of multimodal question answering, the fine-tuning model can be connected to the visual encoder through a connector to form a multimodal large model, which can understand the question information composed of one or more modalities among the text, pictures, audio or video input by the user and generate corresponding reply information.

[0181] Afterwards, combined with the above description of the technical concept of this application, this application hopes that the multimodal big model can simulate different roles, consider the knowledge access rights corresponding to the user's actual identity, understand the user's questions, and generate target response information that conforms to the user's identity, so that the big model has multi-user graded question and answer capabilities. Therefore, multimodal sample data will be obtained from the private domain knowledge base of the target business scenario, and the multimodal big model will be fine-tuned according to the task prompts of each role to obtain a big model for the question and answer processing method proposed in this application.

[0182] Among them, the multimodal training data can use the training data in the question-answer format composed of at least two modalities listed above, such as dialogues composed of pictures and texts, etc., to train the constructed multimodal large model, and deploy it to the electronic device after reasoning test, so that the electronic device supports the question-answer processing method proposed in this application. This application does not elaborate on the testing and deployment implementation process of the large model. It can be seen that the large model of the present application is a multimodal large model that has been injected with private domain knowledge (such as corporate knowledge, etc.) and fine-tuned by instructions, so that it can automatically screen the appropriate target knowledge data within the knowledge range of the corresponding identity user for the same question information raised by users of different identities, and generate different target response information corresponding to users of different identities. The implementation process can refer to the description of the corresponding part of the above method embodiment.

[0183] The above introduces a question and answer processing method provided by an embodiment of the present application. The following will introduce a device for executing the above question and answer processing method.

[0184] Reference Figure 8 , is a schematic diagram of the structure of a question-answering processing device provided in an embodiment of the present application. Figure 8 As shown, the question-answer processing device may include:

[0185] The information acquisition module 810 is used to acquire the question information input by the user and the identity information of the user;

[0186] A query module 820 is used to input the identity information and the question information into a large model, and query at least one candidate knowledge data for answering the question information from a private domain knowledge base that matches the query intent of the question information;

[0187] A generating module 830, configured to process the at least one candidate knowledge data through the large model to generate target response information corresponding to the identity information;

[0188] The output module 840 is used to output the target response information.

[0189] In a possible implementation, the query module 820 may include:

[0190] A task prompt information obtaining unit, used to obtain task prompt information for guiding the large model to process the problem information;

[0191] An intention recognition unit, used to input the identity information, the question information and the task prompt information into a large model, and guide the large model to perform intention recognition on the question information according to a first task prompt word included in the task prompt information, so as to obtain at least one query intention;

[0192] The candidate knowledge data query unit is used to query at least one candidate knowledge data for answering the first query intention for the target business scenario from the private domain knowledge base for the target business scenario if it is determined that the at least one query intention matches the first query intention for the target business scenario.

[0193] In a possible implementation, the generating module 830 may include:

[0194] a target knowledge data screening unit, configured to guide the large model to screen at least one target knowledge data corresponding to the identity information from the at least one candidate knowledge data according to the second task prompt word included in the task prompt information;

[0195] A first batch input unit, configured to batch input the plurality of target knowledge data into the large model according to the third task prompt word included in the task prompt information if the target knowledge data includes a plurality of pieces;

[0196] The target response information generating unit is used to guide the large model to process each of the target knowledge data according to the second task prompt word, and generate target response information corresponding to the identity information.

[0197] Optionally, the target reply information generating unit may include:

[0198] a conversion processing unit, configured to guide the large model to convert each of the target knowledge data according to the reply mode indicated by the identity information based on the second task prompt, and to generate first reply information that conforms to the reply mode;

[0199] A reorganization unit is used to reorganize the multiple first reply messages if there are multiple first reply messages, so as to generate target reply messages that conform to the reply method.

[0200] In some embodiments, the question and answer processing device may further include:

[0201] A second reply information generating module, configured to generate at least one second reply information for replying to the second query intent by using the large model if it is determined that the at least one query intent also includes a second query intent, and the second query intent cannot match the first query intent for the target business scenario;

[0202] Based on this, the target reply information generating unit may include:

[0203] A second batch input unit, used for batch inputting each of the target knowledge data and each of the second reply information into a large model according to the third task prompt word included in the task prompt information;

[0204] The first processing unit is used to guide the large model to process each of the target knowledge data and each of the second reply information simultaneously according to the second task prompt word, so as to generate target reply information corresponding to the identity information.

[0205] In a possible implementation, the candidate knowledge data query unit may include:

[0206] A matching unit, configured to query a plurality of pieces of knowledge data matching the query term of the first query intention and their respective knowledge matching degrees from a private domain knowledge base for the target business scenario;

[0207] The screening unit is used to screen at least one candidate knowledge data from the multiple pieces of knowledge data that have been queried according to the knowledge matching degree.

[0208] For the question-and-answer processing apparatus described in the above embodiments, the intention recognition unit may include:

[0209] A question information output unit, configured to guide the large model to perform intent recognition on the question information according to the first task prompt word included in the task prompt information, and output question information of at least one query intent;

[0210] A question supplementary information acquisition unit, used to acquire the question supplementary information input by the user in response to the question information;

[0211] The query intention obtaining unit is used to obtain the at least one query intention based on the question supplementary information and the question information.

[0212] In view of the question-answer processing method and device described in the above embodiments, a system for executing the above question-answer processing method will be introduced below.

[0213] Reference Fig. 9 , is a schematic diagram of the structure of a question-answering processing system provided in an embodiment of the present application. Fig. 9 As shown, the question-answer processing system may include: a private domain knowledge base 910, a storage device 920, and an electronic device 930, wherein:

[0214] The private domain knowledge base 910 is used to store knowledge data for target business scenarios, such as the enterprise knowledge base of any organization / enterprise. It can be a database deployed in the enterprise cloud or the enterprise's local device, which stores knowledge data within the enterprise (i.e., the corresponding private domain), such as various data and management data generated by the operations of various departments of the enterprise. This application does not restrict the content of the private domain knowledge base.

[0215] The storage device 920 is used to store an authentication description table, which indicates the correspondence between the identity information of different users and the various pieces of knowledge data in the private domain knowledge base. It can represent the accessible knowledge scope of the private domain knowledge base for users with different identities. Combined with the description of the corresponding part of the above method embodiment, the higher the level of users with different identities in the private domain, the greater their access rights to the private domain knowledge base, and the corresponding accessible knowledge scope is larger, but it is not limited to this. The authentication description table can be pre-configured according to the actual situation in the private domain. This application does not limit the representation form of the authentication description table and its configuration implementation method.

[0216] The electronic device 930 deploying the large model can implement the question-answering processing method proposed in each embodiment of the present application. Fig.10 The question-and-answer processing scenario diagram shown in the figure can formulate three role templates of assessor, authenticator and dispatcher for the big model, that is, build the corresponding intelligent agent, i.e., role agent, based on the big model. In the question-and-answer processing process, according to the configured corresponding Prompt task prompt, the big model is used to simulate the corresponding assessor, authenticator or dispatcher, execute the processing task assigned to the role, and generate the target reply information for the question information input by the user. The implementation process can refer to the description process of the method embodiment above, and this embodiment will not be repeated here.

[0217] In order to use the large model to simulate the reviewer, authenticator and dispatcher and autonomously perform the corresponding processing tasks, the corresponding task prompt templates can be configured for different stages of the question-answering process, such as Fig.10 The problem decomposition template, knowledge filtering template, answer angle conversion, viewpoint parallelization and viewpoint reorganization shown are used for the corresponding role agent to call the appropriate task prompt template to execute the processing task assigned to the role. The implementation process is not described in detail in this application.

[0218] In practical applications, the storage device 920 and the electronic device 930 can be the same device or can be independent different devices. The storage device 920 can be a terminal device with storage function, a database, or a storage medium such as a floppy disk, a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk. This application does not limit the type of the storage device 920 and its data storage method. Optionally, the above-mentioned task prompt templates can also be stored in the storage device 920, or stored in the electronic device 930, to meet the retrieval requirements when the large model simulates different roles.

[0219] Optionally, the electronic device 930 may be a terminal device or a server, and the terminal device may include but is not limited to a mobile phone, a tablet computer, a wearable device, a vehicle-mounted device, an augmented reality (AR) / virtual reality (VR) device, a laptop computer, a netbook, a robot or a business terminal, etc. The server may include one or more physical servers, or a cloud server, etc. In the actual application of this application, for example Fig.11 As shown, the electronic device 930 may include but is not limited to: at least one communication port 931, at least one memory 932 and at least one processor 933, wherein:

[0220] At least one communication port 931, at least one memory 932 and at least one processor 933 can communicate with each other via a bus 934. The bus 934 can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Fig.11 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0221] The communication port 931 can be used to obtain question information input by the user, realize communication between the electronic device 930 and the private domain knowledge base 910 and the storage device 920, so as to transmit the data or instructions required in the execution of the question and answer processing method, such as the screened candidate knowledge data, the reading instruction of the authentication specification, etc.

[0222] The communication port 931 may include communication elements corresponding to wireless communication methods such as WiFi, Bluetooth and / or near-field channel methods, so that the electronic device can use the communication element to realize data transmission with other devices (such as the above-mentioned private domain knowledge base 910, storage device 920, etc.). The present application does not limit the composition structure of the communication port 931 to realize this function and its corresponding communication transmission mechanism, which can be determined according to the circumstances.

[0223] In one possible implementation, the communication port 931 may also include one or more interfaces that support wired communication, such as a general-purpose input / output (GPIO) interface, a USB interface, a universal asynchronous receiver / transmitter (UART) interface, etc., to achieve data transmission between various components within the electronic device.

[0224] The memory 932 can be used to store computer-readable instructions for implementing the question and answer processing method proposed in the embodiment of the present application; the processor 934 can load and execute the computer-readable instructions stored in the memory 932 to implement the various steps of the question and answer processing method proposed in the embodiment of the present application. The implementation process is not described in detail in this embodiment.

[0225] In the embodiment of the present application, the memory 932 may include the storage media listed above. The processor 130 may include any one or more of a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor (MP), a digital signal processor (DSP), an application specific integrated circuit (ASIC), or a field-programmable gate array (FPGA).

[0226] It should be understood that Fig.11The structure of the electronic device shown does not constitute a limitation on the electronic device in the embodiment of the present application. In practical applications, the electronic device may include Fig.11 This application does not provide detailed descriptions of more or fewer components shown, or combinations of certain components.

[0227] An embodiment of the present application also provides a computer program product, including computer program instructions. When the computer program instructions are executed on an electronic device, the electronic device implements any one of the question and answer processing methods provided in the embodiments of the present application.

[0228] Wherein, when loading and executing computer program instructions on an electronic device, the process or function described in the embodiment of the present application is generated in whole or in part. The electronic device may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer program instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer program instructions may be transmitted from a website site, a computer, a training device, or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, training device, or data center. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device, a data center, etc. that includes one or more available media integrations. Available media may be magnetic media, (e.g., floppy disk, hard disk, tape), optical media (e.g., DVD), or semiconductor media (e.g., solid-state drive (SSD)), etc.

[0229] A computer-readable storage medium is also provided in an embodiment of the present application. The storage medium carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any question and answer processing method provided in the embodiment of the present application.

[0230] It should also be noted that the device embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed over multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, in the drawings of the device embodiments provided by the present application, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines.

[0231] Through the description of the above implementation methods, those skilled in the art can clearly understand that the present application can be implemented by means of software plus necessary general hardware, and of course can also be implemented by dedicated hardware including dedicated integrated circuits, dedicated CPUs, dedicated memories, dedicated components, etc. It can be seen that in the above embodiments, all or part of the implementation can be achieved by software, hardware, firmware or any combination thereof. When implemented by software, it can be implemented in whole or in part in the form of a computer program product.

[0232] In this specification, each embodiment is described in a progressive or parallel manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the devices, systems, and electronic devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method part.

[0233] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A question-answering processing method, the method comprising: Obtaining question information input by the user and identity information of the user; Input the identity information and the question information into a large model, and query at least one candidate knowledge data for answering the question information from a private domain knowledge base that matches the query intent of the question information; Processing the at least one piece of candidate knowledge data through the large model to generate target response information corresponding to the identity information; Output the target reply information.

2. The method according to claim 1, wherein the inputting the identity information and the question information into a large model, and querying at least one candidate knowledge data for answering the question information from a private domain knowledge base that matches the query intent of the question information, comprises: Obtaining task prompt information for guiding the large model to process the problem information; Inputting the identity information, the question information and the task prompt information into a large model, and guiding the large model to perform intent recognition on the question information according to a first task prompt word included in the task prompt information, to obtain at least one query intent; If it is determined that the at least one query intention matches the first query intention for the target business scenario, at least one candidate knowledge data for answering the first query intention is queried from the private domain knowledge base for the target business scenario.

3. The method according to claim 2, wherein the step of processing the at least one candidate knowledge data by the large model to generate target response information corresponding to the identity information comprises: According to the second task prompt word included in the task prompt information, guiding the large model to select at least one target knowledge data corresponding to the identity information from the at least one candidate knowledge data; If the target knowledge data includes multiple pieces, batch-input the multiple pieces of target knowledge data into the large model according to the third task prompt word included in the task prompt information; According to the second task prompt word, guide the large model to process each of the target knowledge data to generate target response information corresponding to the identity information.

4. The method according to claim 3, wherein the step of guiding the large model to process the target knowledge data and generate target response information corresponding to the identity information based on the second task prompt word comprises: According to the second task prompt, the large model is guided to convert each of the target knowledge data in accordance with the reply method indicated by the identity information, and correspondingly generates first reply information that conforms to the reply method; If there are multiple first reply messages, the multiple first reply messages are reorganized to generate target reply messages that conform to the reply method.

5. The method according to claim 3, after inputting the identity information and the question information into the large model, further comprising: If it is determined that the at least one query intent also includes a second query intent, and the second query intent cannot match the first query intent for the target business scenario, at least one second reply information for replying to the second query intent is generated using the large model; The processing of the at least one target knowledge data to generate target response information corresponding to the identity information includes: According to the third task prompt word included in the task prompt information, batch input each of the target knowledge data and each of the second reply information into the large model; According to the second task prompt word, the large model is guided to process each of the target knowledge data and each of the second reply information simultaneously to generate target reply information corresponding to the identity information.

6. The method according to claim 2, wherein querying at least one candidate knowledge data for answering the first query intention from the private domain knowledge base for the target business scenario comprises: Querying, from a private domain knowledge base for the target business scenario, multiple pieces of knowledge data matching the query term of the first query intent and their respective knowledge matching degrees; At least one candidate knowledge data is screened from the multiple pieces of knowledge data that have been queried according to the knowledge matching degree.

7. The method according to claim 2, wherein the step of guiding the large model to identify the intent of the question information based on the first task prompt word contained in the task prompt information to obtain at least one query intent comprises: According to the first task prompt word included in the task prompt information, guiding the large model to perform intent recognition on the question information, and outputting at least one question information of the query intent; Obtaining supplementary information of the question input by the user in response to the question information; The at least one query intention is obtained based on the question supplementary information and the question information.

8. According to the method described in any one of claims 1 to 7, the large model is obtained by training the question-answering task of the artificial intelligence control platform for the target business scenario; in, The process of training the large model for the question-answering task of the target business scenario by the artificial intelligence control platform includes: The artificial intelligence control platform obtains question-answering training data; Using the question-answering training data, training a large language model to obtain a pre-trained model; Obtain task instruction fine-tuning data based on task prompt information for various question-answering processing tasks in target business scenarios; The task instructions are used to fine-tune the data and adjust the model parameters of the pre-trained model to obtain the large model.

9. A question-answer processing device, comprising: An information acquisition module, used to acquire question information input by a user and identity information of the user; A query module, used to input the identity information and the question information into a large model, and query at least one candidate knowledge data for answering the question information from a private domain knowledge base that matches the query intent of the question information; A generating module, configured to process the at least one candidate knowledge data through the large model to generate target reply information corresponding to the identity information; An output module is used to output the target response information.

10. A question-answering processing system, the system comprising: Private domain knowledge base, used to store knowledge data for target business scenarios; A storage device, used to store an authentication description table, wherein the authentication description table indicates the corresponding relationship between the identity information of different users and each piece of the knowledge data in the private domain knowledge base; An electronic device deploying a large model is used to implement the question-answering processing method as described in any one of claims 1 to 8.

Images