Large model-based answer acquisition method and related apparatus

Abstract

Embodiments of the present application provide a large model-based answer acquisition method and a related apparatus. The method comprises: desensitizing sensitive information in first information by means of a first algorithm to obtain second information, wherein the first information comprises a first question and first corpus data related to the first question, the second information comprises a second question and second corpus data, the second question comprises a question obtained after desensitizing the first question, and the second corpus data comprises corpus data obtained after desensitizing the first corpus data; sending the second information to a first device; receiving a first answer sent by the first device, wherein the first answer is an answer to the second question analyzed from the second corpus data; and restoring sensitive information in the first answer by means of the first algorithm to obtain a second answer to the first question. By means of the embodiments of the present application, the computational overhead during answer acquisition on a client side can be reduced while guaranteeing the privacy security of the client side.

Description

A method for obtaining answers based on a large model and related devices

[0001] This application claims priority to the Chinese patent application filed with the State Intellectual Property Office of China on March 30, 2024, with application number 202410389232.4, and invention name “A method for obtaining answers based on a large model and related devices”, the entire contents of which are incorporated by reference into this application. Technical Field

[0002] The present application relates to the field of artificial intelligence (AI) technology, and in particular to a method for obtaining answers based on a large model and related devices. Background Art

[0003] With the release of ChatGPT, large language models have become a representative technology for the emergence of intelligence, attracting widespread attention from academia and industry. However, "serious nonsense" is a common problem with large models, particularly GPT4, and it severely restricts their widespread application in a wide range of fields. While language models cannot guarantee 100% accuracy, limiting their application at some physical levels, industrial applications still require reliable authentication mechanisms to ensure that the answers provided by large models can be safely deployed. Large model verification technology has been proposed to address this issue. This approach typically involves using an external knowledge base to check and verify key information in question-and-answer pairs using existing knowledge and logical rules. Applications can then supplement and maintain the knowledge base rules to ensure that verified answers meet their application expectations, thereby ensuring the reliable application of large model question-and-answer results. Currently, the main approach is to install a verification knowledge base on a cloud server as a supplementary component to the large model. This knowledge base corrects basic logical errors by comparing keywords. This approach typically requires integrating relevant business knowledge into the verification knowledge base.

[0004] In specific applications, whether it is building a verification knowledge base or users asking questions to a large model, it will pose a challenge to user data security. Summary of the Invention

[0005] The embodiments of the present application disclose a method and related apparatus for obtaining answers based on a large model, which can reduce the computational overhead of obtaining answers on the end side while ensuring the privacy security of the end side.

[0006] In a first aspect, an embodiment of the present application provides a method for obtaining an answer based on a large model, the method comprising:

[0007] Desensitizing sensitive information in the first information using a first algorithm to obtain second information, wherein the first information includes a first question and a first corpus related to the first question; the second information includes a second question and a second corpus, the second question includes the desensitized version of the first question, and the second corpus includes the desensitized version of the first corpus;

[0008] sending the second information to the first device;

[0009] receiving a first answer sent by the first device, where the first answer is an answer to the second question analyzed from the second corpus;

[0010] The sensitive information in the first answer is restored by using the first algorithm to obtain a second answer to the first question.

[0011] Through the above method, the third device desensitizes the first question and the first corpus related to the first question and then sends it to the first device deployed with a large model in the cloud. The first device generates a preliminary answer based on the desensitized information, such as the first answer. The third device then restores the sensitive information of the first answer to obtain a second answer that can be understood locally. In this process, the first device provides the third device with answer analysis capabilities, but does not have access to sensitive information on the third device, fully protecting the privacy and security of the terminal. Because the first device in the cloud provides answer analysis capabilities, the computational overhead of obtaining answers on the terminal side is reduced, and the terminal-side functions are avoided from being overly complex.

[0012] In conjunction with the first aspect, in a possible implementation of the first aspect, before desensitizing the sensitive information in the first information using the first algorithm to obtain the second information, the method further includes:

[0013] receiving the first question sent by the second device;

[0014] Determining a first corpus related to the first question from a pre-stored corpus;

[0015] After restoring the sensitive information in the first answer by using the first algorithm to obtain a second answer to the first question, the method further includes:

[0016] A second answer to the first question is sent to the second device.

[0017] In this implementation, the end-edge cloud architecture is clearly constructed by the first device, the second device and the third device. The second device initiates the question request, the first device performs the main calculation, and the third device performs the transfer and auxiliary calculation. It can provide efficient question-and-answer services for the second device without bringing large computing overhead to the second device.

[0018] In combination with the first aspect, or any of the foregoing possible implementations of the first aspect, in another possible implementation of the first aspect, determining the first corpus related to the first question from a pre-stored corpus includes:

[0019] Searching a pre-stored corpus for initial corpus related to the first question;

[0020] The initial corpus is segmented, compressed, and concatenated in sequence to obtain a first corpus with a smaller data volume.

[0021] In this implementation, the first corpus obtained by processing the initial corpus related to the first question has a higher density of useful information, which can reduce the computational overhead during model training without losing too much useful information.

[0022] In combination with the first aspect, or any one of the above-mentioned possible implementations of the first aspect, in another possible implementation of the first aspect, the method is applied to a third device, the third device and the second device are both deployed in a first local area network, and the first device is deployed in a network outside the first local area network.

[0023] It can be understood that the third device and the second device are both deployed in the first local area network. Therefore, during the entire process, the original corpus information is only in the first local area network, which builds a security barrier for the third device and the second device, further improving the security of the end side.

[0024] In combination with the first aspect, or any of the foregoing possible implementations of the first aspect, in another possible implementation of the first aspect, restoring the sensitive information in the first answer using the first algorithm to obtain a second answer to the first question includes:

[0025] restoring the sensitive information in the first answer by the first algorithm to obtain a third answer containing the sensitive information;

[0026] A key information verification is performed on the third answer including the sensitive information to obtain a second answer to the first question.

[0027] In combination with the first aspect, or any of the foregoing possible implementations of the first aspect, in another possible implementation of the first aspect, performing key information verification on the third answer including sensitive information to obtain a second answer to the first question includes:

[0028] Perform key information verification on the third answer containing sensitive information;

[0029] If the key information of the third answer including the sensitive information is incorrect, correcting the incorrect key information, wherein the corrected third answer is the second answer to the first question;

[0030] If the key information of the third answer including the sensitive information is correct, the third answer including the sensitive information is the second answer to the first question.

[0031] In this implementation, the third device will also verify and correct the answer provided by the first device, which can effectively improve the accuracy and rationality of the answer and alleviate the situation where the answer is wrong or has low relevance due to the misunderstanding error of the large model of the first device.

[0032] In combination with the first aspect, or any of the foregoing possible implementations of the first aspect, in another possible implementation of the first aspect, correcting incorrect key information includes:

[0033] Correct incorrect key information using stored information.

[0034] In combination with the first aspect, or any of the foregoing possible implementations of the first aspect, in another possible implementation of the first aspect, the key information includes keywords and / or key logic.

[0035] In combination with the first aspect, or any of the foregoing possible implementations of the first aspect, in another possible implementation of the first aspect,

[0036] Desensitizing the sensitive information in the first information by using the first algorithm to obtain the second information includes:

[0037] Calling a hash mapping table to determine desensitized information corresponding to the sensitive information in the first information;

[0038] Replacing the sensitive information in the first information with the desensitized information to obtain second information;

[0039] Restoring the sensitive information in the first answer by using the first algorithm to obtain a second answer to the first question includes:

[0040] Calling the hash map to determine the sensitive information corresponding to the desensitized information in the first answer;

[0041] The desensitized information in the first answer is replaced with the sensitive information to obtain a second answer to the first question.

[0042] In a second aspect, an embodiment of the present application provides a method for obtaining an answer based on a large model, the method comprising:

[0043] A third device desensitizes sensitive information in the first information using a first algorithm to obtain second information, wherein the first information includes a first question and a first corpus related to the first question; the second information includes a second question and a second corpus, the second question includes the desensitized version of the first question, and the second corpus includes the desensitized version of the first corpus;

[0044] The third device sends the second information to the first device;

[0045] The first device analyzes the second corpus to obtain a first answer to the second question;

[0046] The first device sends the first answer to the third device;

[0047] receiving a first answer sent by the first device;

[0048] The sensitive information in the first answer is restored by using the first algorithm to obtain a second answer to the first question.

[0049] Through the above method, the third device desensitizes the first question and the first corpus related to the first question and then sends it to the first device deployed with a large model in the cloud. The first device generates a preliminary answer based on the desensitized information, such as the first answer. The third device then restores the sensitive information of the first answer to obtain a second answer that can be understood locally. In this process, the first device provides the third device with answer analysis capabilities, but does not have access to sensitive information on the third device, fully protecting the privacy and security of the terminal. Because the first device in the cloud provides answer analysis capabilities, the computational overhead of obtaining answers on the terminal side is reduced, and the terminal-side functions are avoided from being overly complex.

[0050] In a third aspect, an embodiment of the present application provides a large model-based answer acquisition device, comprising:

[0051] a processing unit configured to desensitize sensitive information in the first information using a first algorithm to obtain second information, wherein the first information includes a first question and a first corpus related to the first question; the second information includes a second question and a second corpus, the second question includes the desensitized version of the first question, and the second corpus includes the desensitized version of the first corpus;

[0052] a communication unit, configured to send the second information to the first device;

[0053] The communication unit is further configured to receive a first answer sent by the first device, where the first answer is an answer to the second question analyzed from the second corpus;

[0054] The processing unit is further configured to restore the sensitive information in the first answer using the first algorithm to obtain a second answer to the first question.

[0055] Through the above method, the third device desensitizes the first question and the first corpus related to the first question and then sends it to the first device deployed with a large model in the cloud. The first device generates a preliminary answer based on the desensitized information, such as the first answer. The third device then restores the sensitive information of the first answer to obtain a second answer that can be understood locally. In this process, the first device provides the third device with answer analysis capabilities, but does not have access to sensitive information on the third device, fully protecting the privacy and security of the terminal. Because the first device in the cloud provides answer analysis capabilities, the computational overhead of obtaining answers on the terminal side is reduced, and the terminal-side functions are avoided from being overly complex.

[0056] In conjunction with the second aspect, in another possible implementation of the second aspect

[0057] The communication unit is further configured to receive the first question sent by the second device;

[0058] The processing unit is further configured to determine a first corpus related to the first question from a pre-stored corpus;

[0059] The communication unit is further configured to send a second answer to the first question to the second device.

[0060] In this implementation, the end-edge cloud architecture is clearly constructed by the first device, the second device and the third device. The second device initiates the question request, the first device performs the main calculation, and the third device performs the transfer and auxiliary calculation. It can provide efficient question-and-answer services for the second device without bringing large computing overhead to the second device.

[0061] In combination with the second aspect, or any of the foregoing possible implementations of the second aspect, in yet another possible implementation of the second aspect, the determining of the first corpus aspect related to the first question from a pre-stored corpus, the processing unit is specifically configured to:

[0062] Searching a pre-stored corpus for initial corpus related to the first question;

[0063] The initial corpus is segmented, compressed, and concatenated in sequence to obtain a first corpus with a smaller data volume.

[0064] In this implementation, the first corpus obtained by processing the initial corpus related to the first question has a higher density of useful information, which can reduce the computational overhead during model training without losing too much useful information.

[0065] In combination with the second aspect, or any one of the above-mentioned possible implementations of the second aspect, in another possible implementation of the second aspect, the apparatus is applied to a third device, the third device and the second device are both deployed in a first local area network, and the first device is deployed in a network outside the first local area network.

[0066] It can be understood that the third device and the second device are both deployed in the first local area network. Therefore, during the entire process, the original corpus information is only in the first local area network, which builds a security barrier for the third device and the second device, further improving the security of the end side.

[0067] In combination with the second aspect, or any of the foregoing possible implementations of the second aspect, in yet another possible implementation of the second aspect, in terms of restoring the sensitive information in the first answer using the first algorithm to obtain the second answer to the first question, the processing unit is specifically configured to:

[0068] restoring the sensitive information in the first answer by the first algorithm to obtain a third answer containing the sensitive information;

[0069] A key information verification is performed on the third answer including the sensitive information to obtain a second answer to the first question.

[0070] In combination with the second aspect, or any of the foregoing possible implementations of the second aspect, in yet another possible implementation of the second aspect, in the aspect of performing key information verification on the third answer including sensitive information to obtain the second answer to the first question, the processing unit is specifically configured to:

[0071] Perform key information verification on the third answer containing sensitive information;

[0072] If the key information of the third answer including the sensitive information is incorrect, correcting the incorrect key information, wherein the corrected third answer is the second answer to the first question;

[0073] If the key information of the third answer including the sensitive information is correct, the third answer including the sensitive information is the second answer to the first question.

[0074] In this implementation, the third device will also verify and correct the answer provided by the first device, which can effectively improve the accuracy and rationality of the answer and alleviate the situation where the answer is wrong or has low relevance due to the misunderstanding error of the large model of the first device.

[0075] In combination with the second aspect, or any of the foregoing possible implementations of the second aspect, in another possible implementation of the second aspect, in terms of correcting incorrect key information, the processing unit is specifically configured to:

[0076] Correct incorrect key information using stored information.

[0077] In combination with the second aspect, or any of the foregoing possible implementations of the second aspect, in yet another possible implementation of the second aspect, the key information includes keywords and / or key logic.

[0078] In combination with the second aspect, or any of the foregoing possible implementations of the second aspect, in another possible implementation of the second aspect,

[0079] In terms of desensitizing the sensitive information in the first information by using the first algorithm to obtain the second information, the processing unit is specifically configured to:

[0080] Calling a hash mapping table to determine desensitized information corresponding to the sensitive information in the first information;

[0081] Replacing the sensitive information in the first information with the desensitized information to obtain second information;

[0082] In terms of restoring the sensitive information in the first answer by using the first algorithm to obtain the second answer to the first question, the processing unit is specifically configured to:

[0083] Calling the hash map to determine the sensitive information corresponding to the desensitized information in the first answer;

[0084] The desensitized information in the first answer is replaced with the sensitive information to obtain a second answer to the first question.

[0085] In a fourth aspect, an embodiment of the present application provides a large model-based answer acquisition system, the system comprising a first device, a second device, and a third device, wherein:

[0086] A third device desensitizes sensitive information in the first information using a first algorithm to obtain second information, wherein the first information includes a first question and a first corpus related to the first question; the second information includes a second question and a second corpus, the second question includes the desensitized version of the first question, and the second corpus includes the desensitized version of the first corpus;

[0087] The third device sends the second information to the first device;

[0088] The first device analyzes the second corpus to obtain a first answer to the second question;

[0089] The first device sends the first answer to the third device;

[0090] receiving a first answer sent by the first device;

[0091] The sensitive information in the first answer is restored by using the first algorithm to obtain a second answer to the first question.

[0092] In a fifth aspect, an embodiment of the present application provides a device, comprising a processor, a transceiver, and a memory, wherein:

[0093] The memory is used to store a computer program, the transceiver is used to perform data reception and transmission, and the processor is used to call the computer program to implement the method described in the first aspect or any possible implementation manner of the first aspect.

[0094] In the sixth aspect, an embodiment of the present application provides a chip system, which includes at least one processor, a memory and an interface circuit, the memory, the interface circuit and the at least one processor are interconnected through lines, and a computer program is stored in the at least one memory; when the computer program is executed by the processor, the method described in the first aspect or any possible implementation method of the first aspect is implemented.

[0095] In a seventh aspect, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer-readable storage medium runs on a processor, it implements the method described in the first aspect or any possible implementation of the first aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] The following is an introduction to the drawings used in the embodiments of this application.

[0097] FIG1 is a flow chart of a method for obtaining answers to questions based on a large model according to an embodiment of the present application;

[0098] FIG2 is a schematic diagram of a large model service architecture provided by an embodiment of the present application;

[0099] FIG3 is a schematic diagram of a large model service architecture provided by an embodiment of the present application;

[0100] FIG4 is a flow chart of a method for obtaining an answer based on a large model according to an embodiment of the present application;

[0101] FIG5 is a flow chart of a corpus processing method provided in an embodiment of the present application;

[0102] FIG6 is a schematic structural diagram of a large model-based answer acquisition device provided in an embodiment of the present application;

[0103] FIG7 is a schematic structural diagram of a large-model-based answer acquisition device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0104] The embodiments of the present application are described below in conjunction with the drawings in the embodiments of the present application.

[0105] In cloud computing, the user side is a very lightweight application that performs the most basic, predefined calculations, with information processing largely taking place in the cloud. However, as application scenarios expand, user tiers are becoming increasingly refined. Business users need to manage the usage behavior of individual employees, necessitating company-side management of relevant information. Therefore, within a business company, users are often divided into different domains. Users within a domain generally belong to the same company or organization, enabling trust and information sharing. However, information barriers must be established between domains to prevent the leakage of important information. Combining cloud computing with big models is currently a key research area for big model service models, as illustrated below.

[0106] Solution 1: A technical model for question-answering services based on a unified large model:

[0107] The application asks a question to the cloud-based large model service provider and provides information that guides the answer as a prompt. The large model service provider gives an answer based on the information received in the cloud and pushes the answer to the application, as shown in Figure 1. The cloud-based service provider collects data to build a knowledge base for verifying the answer. The verified answer is fed back to the user for application. This model is the main way that large models are currently widely used. Large models such as GPT4 are all accessible to the public based on this model. This model is often suitable for scenarios where data privacy is not strong. Users often do not upload important data to ask questions. The application side often mainly judges the reliability of the answer based on their own experience. Professional answer guarantee mechanism. This model has the following defects: (1) It cannot meet the privacy needs of the application side; (2) The application side lacks its own means of automatic verification, has no truly trustworthy guarantee means, and cannot meet the personalized verification needs of the application side.

[0108] Solution 2: Technical model for fine-tuning model deployment within the domain:

[0109] A cloud-based large-scale model service provider provides the initial large-scale model and computing resources, and the application side provides data (OpenAI). The cloud-based large-scale model fine-tunes the initial large-scale model based on the data provided by the application side, resulting in a large-scale model tailored to the application side. Alternatively, the application side can directly access the model and fine-tune it (Limma fine-tuning). Currently, OpenAI's GPT model is not open source for commercial reasons. It provides computing power and a platform for users to fine-tune the model. However, this model requires users to upload corpus data. While OpenAI maintains a commitment to data security, it is difficult to fine-tune commercial data that truly requires confidentiality. Large-scale models such as Limma have been made open source, and many organizations have fine-tuned Limma models to train their own large-scale models for application. However, the overall performance of models trained using this model will inevitably lag behind that of large-scale models like GPT4. This model has the following drawbacks: the application side requires a certain amount of computing resources, data resources, and algorithm resources. The training data, servers for fine-tuning the large-scale model, and the development of algorithms for optimizing the fine-tuning large model all incur significant costs, making them prohibitively expensive for relatively lightweight applications and small-scale users.

[0110] In view of the problems existing in the current big model service model, the embodiment of the present application adopts an end-edge-cloud collaborative architecture to support big model service providers in providing services to various corporate business users. This architecture is a relatively generalized big model service architecture.

[0111] As shown in Figure 2, Figure 2 is a schematic diagram of a large model service architecture provided by an embodiment of the present application. The architecture includes collaborative processing of the cloud 201 (where the large model is deployed) and the local area network 202. There can be multiple local area networks. After a question is generated in each local area network, the question will be fuzzified and sent to the cloud 201. After the large model of the cloud 201 generates the answer to the question, the answer to the question is sent to the corresponding local area network 202. The corresponding local area network 202 parses the answer and reconstructs the answer to restore the key information in the answer.

[0112] As shown in Figure 3, Figure 3 is a schematic diagram of a large model service architecture provided by an embodiment of the present application. The different local area networks 202 in Figure 2 can belong to different user domains (such as different enterprises). Figure 3 uses three user domains as an example for illustration. In actual application, there can be more or fewer user domains. Among them, the cloud 201 includes a first device 301, on which a large model is deployed. Each local area network 202 includes a second device 302 and a third device 303. Among them, the first device 301 has strong computing power. For example, the first device 301 can be a server or a server cluster composed of multiple servers. It can train a large model based on a large amount of training data and provide parsing capabilities for other devices based on the trained large model. The third device 303 also has certain computing power. For example, the third device 303 can be a server or a server cluster composed of multiple servers. It can further process the parsing results (such as the generated answers) returned by the first device 301 (such as verification, etc.), and then send the final results to the second device 302. The second device 302 is a user device, such as a computer, a handheld device (for example, a mobile phone, a tablet computer, a PDA, etc.), a vehicle-mounted device (for example, a car, a bicycle, an electric car, an airplane, a ship, etc.), a wearable device (for example, a smart watch (such as iWatch, etc.), a smart bracelet, a pedometer, etc.), a smart home device (for example, a refrigerator, a television, an air conditioner, an electric meter, etc.), an intelligent robot, a workshop equipment, and the like.

[0113] In the end-edge-cloud architecture, the second device corresponds to the end (device), the third device corresponds to the edge (edge), and the first device corresponds to the cloud (cloud).

[0114] Please refer to FIG4 , which is a flowchart of a method for obtaining an answer based on a large model provided in an embodiment of the present application. The method can be implemented based on the architecture shown in FIG2 or FIG3 , and includes but is not limited to the following steps:

[0115] Step S401: The second device sends a first question to the third device.

[0116] Specifically, the second device, as a user device, can generate corresponding questions according to the needs of the user. For the sake of convenience of description, it can be called the first question. For example, when the user encounters a problem when handling a certain business or performing a certain task and hopes to seek an answer, the user can ask the question through the second device. For example, the user inputs the question to the second device through text input, voice input, touch selection, mouse selection, etc., and then sends the first question to the third device to wait for feedback.

[0117] It should be noted that the second device and the third device, and the third device and the first device, can both establish communication connections in advance. For example, the second device sends a request to the third device (such as the company domain server) to trigger the establishment of a communication connection with the large model in the cloud, and then the third device sends a communication connection request to the first device in the cloud. Correspondingly, the first device in the cloud feeds back a connection confirmation to the third device in the company domain, and the third device in the company domain then feeds back a connection confirmation to the second device. At this point, the three-party session connection based on the end-edge cloud architecture is established. Of course, in addition to being pre-established, the three-party session based on the end-edge cloud architecture can also be established instantly when there is a communication need, which is not limited here.

[0118] Step S402: The third device receives the first question sent by the second device.

[0119] Step S403: The third device searches a pre-stored corpus for a first corpus related to the first question.

[0120] In an embodiment of the present application, a third device maintains a corpus containing a large amount of corpus, such as a business manual corpus, so that guidance answers to questions (such as business-related questions) can be found from the corpus. The corpus can be deployed on the third device or deployed as a plug-in on another device that can be called by the third device. Optionally, the other device and the third device belong to the same local area network.

[0121] Step S404: The third device desensitizes the sensitive information in the first information using the first algorithm to obtain second information.

[0122] Among them, the first information includes a first question and a first corpus related to the first question; the second information includes a second question and a second corpus, the second question includes the question after the first question is desensitized, and the second corpus includes the corpus after the first corpus is desensitized; for example, the first algorithm can be a hash mapping algorithm, which performs hash encoding, desensitization and encryption on these sensitive words by calling a hash table; in addition, the third device maintains a sensitive information list containing a large amount of sensitive information that needs to be protected. If the above-mentioned first information contains information in the sensitive information list, it is desensitized by the first algorithm, which can also be called fuzzy processing. For ease of understanding, an example is given below.

[0123] Case 1:

[0124] The first question is "How to complete the deployment of parameter A?"

[0125] The first corpus is "Parameter A is..., parameter A is used for..., parameter A has..., in the deployment process, first call..., then..., and finally start... to complete the deployment."

[0126] The sensitive information table contains sensitive information such as "parameter A", "parameter B", and "parameter C".

[0127] Then, after desensitizing the sensitive information in the first information, the second information is obtained, which includes:

[0128] The second question is "How to complete the deployment of SS?".

[0129] The second corpus is "SS is..., SS is used for..., SS has..., in the deployment process first call..., then..., and finally start... to complete the deployment."

[0130] It can be seen that during the desensitization process, the sensitive information "parameter A" is mapped to the desensitized information "SS".

[0131] Step S405: The third device sends the second information to the first device.

[0132] Specifically, the second information can be regarded as a prompt to ask questions to the large model. In addition to the second question and the second corpus mentioned above, the second information can also include other information that can play a role in generating answers.

[0133] Step S406: The first device receives the second information sent by the third device.

[0134] Step S407: The first device analyzes the first answer to the second question from the second corpus included in the second information.

[0135] Specifically, the first device deploys a large model and has strong understanding and analysis capabilities, so the first device can analyze the answer to the second question from the second corpus, which is called the first answer for ease of understanding.

[0136] Step S408: The first device sends the first answer to the third device.

[0137] Specifically, when the first device sends the first answer again, it will carry information related to the second question, such as the question identifier or the second question itself, to indicate that the first answer is the answer to the second question. Optionally, it may also carry other information related to the first answer.

[0138] Step S409: The third device receives the first answer sent by the first device.

[0139] Step S410: The third device restores the sensitive information in the first answer using the first algorithm to obtain a second answer to the first question.

[0140] Specifically, the first algorithm has been introduced above. The previous step is to perform desensitization processing using the first algorithm. This step is to restore the sensitive information, which is a reverse process of the previous desensitization processing. Optionally, the desensitized information contained in the first answer can be decoded and restored by calling a hash mapping table to obtain a second answer containing sensitive information. For example, the previous desensitization processing is to map the sensitive information "parameter A" to the desensitized information "SS". In this step, the sensitive information is restored by mapping the desensitized information "SS" to the sensitive information "parameter A". That is, all the desensitized information "SS" in the first answer will be replaced with the sensitive information "parameter A". If there is other desensitized information, it will also be mapped to the sensitive information corresponding to the other desensitized information. This process is the process of reconstructing the answer to the first question. It can be understood that the second answer containing sensitive information can be understood by the third device and the second device.

[0141] In an optional solution, restoring the sensitive information in the first answer using the first algorithm to obtain the second answer to the first question may include: restoring the sensitive information in the first answer using the first algorithm to obtain an answer containing the sensitive information; and performing key information verification on the answer containing the sensitive information to obtain the second answer to the first question. That is, after restoring the sensitive information in the first answer, key information verification is also performed on the restored answer to obtain the second answer to the first question. There are many ways to verify key information. For ease of understanding, the following examples are provided:

[0142] The third answer including sensitive information is verified for key information, and the key information includes preset keywords and / or key logic. It can be understood that the third device maintains a proofreading knowledge base based on business knowledge, which provides word usage patterns or characteristics in specific scenarios or semantic environments, as well as the patterns and characteristics of language logic. Therefore, based on the proofreading knowledge base, it is possible to verify whether the key information in the third answer is correct, and if it is incorrect, it can also provide a basis and reference for correction.

[0143] If the key information of the third answer including sensitive information is incorrect, the incorrect key information is corrected, wherein the corrected third answer is the second answer to the first question. If the key information of the third answer including sensitive information is correct, the third answer including sensitive information is the second answer to the first question. Optionally, it is also possible that the key information of the third answer is incorrect, but the proofreading knowledge base of the third device cannot provide a correction reference. In this case, the third answer is unavailable, so the third device can re-acquire the first corpus related to the first question (generally different from the first corpus acquired last time, such as new corpus or new information (such as incorrect key information) is supplemented on the basis of the first corpus acquired last time), and then re-execute steps S401-S410.

[0144] Step S411: The third device sends a second answer to the first question to the second device.

[0145] Specifically, if the second device is the initiator of the first question, the third device sends the second answer to the first question to the second device.

[0146] In an optional implementation of the present application, the third device and the second device are both deployed in a first local area network (such as a local area network of an enterprise), and the first device is deployed in a network outside the first local area network.

[0147] It should be noted that the third device in the embodiment of the present application maintains a small model. Based on the small model, the third device performs the above-mentioned operations of searching for the first corpus corresponding to the first question, desensitizing (or fuzzifying) sensitive information, adding sensitive information to the first answer fed back by the first device, and performing key information verification on the third answer with added sensitive information.

[0148] In an optional implementation method used in the present application, steps S401, S402 and S411 may be omitted, that is, there is no participation of the second device, the first question is directly generated on the third device, and the second answer to the first question is finally obtained and used on the third device.

[0149] In the optional implementation mode used in the present application, the above-mentioned step S403 searches for the corpus related to the first question from the pre-stored corpus. If it is a corpus with a relatively large amount of information, the corpus can be compressed first, and the compressed corpus is the first corpus. In the embodiment of the present application, taking into account that the paragraph text is too long, it is difficult for the large model to accept it at one time (there is an upper limit on the number of input words), so the relatively large corpus is compressed first. For example, as shown in Figure 5, the larger corpus is first divided into individual sentences, wherein the length of each sentence, the number of sentences and the semantic coherence of each clause need to be considered. After the segmentation, each segment is compressed in sections, and after compression, each compressed segment is merged into a relatively streamlined (or smaller data volume) first corpus, wherein each segment connection can be continued to connect different segments into complete paragraphs.

[0150] Optionally, when segmenting sentences, an original sentence query index can be constructed for each sentence. This index can be either some keywords or a model embedding of the sentence (the advantage of embedding over keywords is that it contains contextual information of the sentence). After constructing the above index, in the process of verifying the key information of the above third answer, if some key information is identified from the above third answer, then the target position of the key information in the larger corpus before compression can be found based on the established index, and then the meaning of the sensitive information at the target position can be compared with the meaning of the sensitive information in the third answer. If the two meanings are the same, the third answer does not need to be corrected or revised. If the two meanings are different, correction or revision is required. This is the process of verifying the third answer.

[0151] It should be noted that the above description describes a scenario in which the third device retrieves a larger corpus, compresses it, and then sends it to the first device for analysis to generate an answer. However, in actual applications, the third device may also retrieve a larger corpus and directly send it to the first device, which then compresses and analyzes it to generate an answer. Optionally, whether a corpus belongs to a larger corpus can be determined by setting corresponding criteria or analyzing it using a corresponding model. The specific details are not elaborated here.

[0152] In the method described in Figure 4, the third device desensitizes the first question and the first corpus related to the first question before sending them to the first device on the cloud, which has a large model deployed. The first device then generates a preliminary answer, such as the first answer, based on the desensitized information. The third device then restores the sensitive information of the first answer to obtain a second answer that can be understood locally. In this process, the first device provides the third device with answer analysis capabilities, but does not have access to sensitive information on the third device, fully protecting the privacy and security of the end-side. Since the first device on the cloud provides answer analysis capabilities, the computational overhead of obtaining answers on the end-side is reduced, and the end-side functions are avoided from being overly complex.

[0153] The above describes in detail the method of the embodiment of the present application, and the following provides an apparatus of the embodiment of the present application.

[0154] Please refer to Figure 6, which is a structural diagram of a large model-based answer acquisition device provided in an embodiment of the present application. The device 60 may include a processing unit 601 and a communication unit 602, wherein each unit is described in detail as follows.

[0155] Processing unit 601 is configured to desensitize sensitive information in first information using a first algorithm to obtain second information, wherein the first information includes a first question and first corpus related to the first question; the second information includes a second question and second corpus, the second question includes the desensitized version of the first question, and the second corpus includes the desensitized version of the first corpus;

[0156] The communication unit 602 is configured to send the second information to the first device;

[0157] The communication unit is further configured to receive a first answer sent by the first device, where the first answer is an answer to the second question analyzed from the second corpus;

[0158] The processing unit 601 is further configured to restore the sensitive information in the first answer using the first algorithm to obtain a second answer to the first question.

[0159] Through the above method, the third device desensitizes the first question and the first corpus related to the first question and then sends it to the first device deployed with a large model in the cloud. The first device generates a preliminary answer based on the desensitized information, such as the first answer. The third device then restores the sensitive information of the first answer to obtain a second answer that can be understood locally. In this process, the first device provides the third device with answer analysis capabilities, but does not have access to sensitive information on the third device, fully protecting the privacy and security of the terminal. Because the first device in the cloud provides answer analysis capabilities, the computational overhead of obtaining answers on the terminal side is reduced, and the terminal-side functions are avoided from being overly complex.

[0160] In another possible implementation:

[0161] The communication unit 602 is further configured to receive the first question sent by the second device;

[0162] The processing unit 601 is further configured to determine a first corpus related to the first question from a pre-stored corpus;

[0163] The communication unit 602 is further configured to send a second answer to the first question to the second device.

[0164] In this implementation, the end-edge cloud architecture is clearly constructed by the first device, the second device and the third device. The second device initiates the question request, the first device performs the main calculation, and the third device performs the transfer and auxiliary calculation. It can provide efficient question-and-answer services for the second device without bringing large computing overhead to the second device.

[0165] In yet another possible implementation, in determining the first corpus aspect related to the first question from a pre-stored corpus, the processing unit 601 is specifically configured to:

[0166] Searching a pre-stored corpus for initial corpus related to the first question;

[0167] The initial corpus is segmented, compressed, and concatenated in sequence to obtain a first corpus with a smaller data volume.

[0168] In this implementation, the first corpus obtained by processing the initial corpus related to the first question has a higher density of useful information, which can reduce the computational overhead during model training without losing too much useful information.

[0169] In another possible implementation, the apparatus is applied to a third device, the third device and the second device are both deployed in a first local area network, and the first device is deployed in a network outside the first local area network.

[0170] It can be understood that the third device and the second device are both deployed in the first local area network. Therefore, during the entire process, the original corpus information is only in the first local area network, which builds a security barrier for the third device and the second device, further improving the security of the end side.

[0171] In yet another possible implementation, in terms of restoring the sensitive information in the first answer using the first algorithm to obtain the second answer to the first question, the processing unit is specifically configured to:

[0172] restoring the sensitive information in the first answer by the first algorithm to obtain a third answer containing the sensitive information;

[0173] A key information verification is performed on the third answer including the sensitive information to obtain a second answer to the first question.

[0174] In yet another possible implementation, in terms of performing key information verification on the third answer including sensitive information to obtain the second answer to the first question, the processing unit is specifically configured to:

[0175] Perform key information verification on the third answer containing sensitive information;

[0176] If the key information of the third answer including the sensitive information is incorrect, correcting the incorrect key information, wherein the corrected third answer is the second answer to the first question;

[0177] If the key information of the third answer including the sensitive information is correct, the third answer including the sensitive information is the second answer to the first question.

[0178] In this implementation, the third device will also verify and correct the answer provided by the first device, which can effectively improve the accuracy and rationality of the answer and alleviate the situation where the answer is wrong or has low relevance due to the misunderstanding error of the large model of the first device.

[0179] In yet another possible implementation, in terms of correcting incorrect key information, the processing unit is specifically configured to:

[0180] Correct incorrect key information using stored information.

[0181] In yet another possible implementation, the key information includes keywords and / or key logic.

[0182] In another possible implementation,

[0183] In terms of desensitizing the sensitive information in the first information by using the first algorithm to obtain the second information, the processing unit is specifically configured to:

[0184] Calling a hash mapping table to determine desensitized information corresponding to the sensitive information in the first information;

[0185] Replacing the sensitive information in the first information with the desensitized information to obtain second information;

[0186] In terms of restoring the sensitive information in the first answer by using the first algorithm to obtain the second answer to the first question, the processing unit is specifically configured to:

[0187] Calling the hash map to determine the sensitive information corresponding to the desensitized information in the first answer;

[0188] The desensitized information in the first answer is replaced with the sensitive information to obtain a second answer to the first question.

[0189] It should be noted that the implementation of each unit may also correspond to the corresponding description of the method embodiment shown in FIG. 4 .

[0190] Please refer to Figure 7, which is a large model-based answer acquisition device 70 provided in an embodiment of the present application. The device 70 includes a processor 701, a memory 702 and a communication interface 703, and the processor 701, memory 702 and communication interface 703 are interconnected through a bus.

[0191] Memory 702 includes, but is not limited to, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), or compact disc read-only memory (CD-ROM). Memory 702 is used for storing computer programs and data. Communication interface 703 is used to receive and send data.

[0192] The processor 701 may be one or more central processing units (CPUs). In the case where the processor 701 is a CPU, the CPU may be a single-core CPU or a multi-core CPU.

[0193] The processor 701 in the device 70 is configured to read the computer program code stored in the memory 702 and perform the following operations:

[0194] Desensitizing sensitive information in the first information using a first algorithm to obtain second information, wherein the first information includes a first question and a first corpus related to the first question; the second information includes a second question and a second corpus, the second question includes the desensitized version of the first question, and the second corpus includes the desensitized version of the first corpus;

[0195] sending the second information to the first device through the communication interface 703;

[0196] receiving, via the communication interface 703, a first answer sent by the first device, where the first answer is an answer to the second question analyzed from the second corpus;

[0197] The sensitive information in the first answer is restored by using the first algorithm to obtain a second answer to the first question.

[0198] Through the above method, the third device desensitizes the first question and the first corpus related to the first question and then sends it to the first device deployed with a large model in the cloud. The first device generates a preliminary answer based on the desensitized information, such as the first answer. The third device then restores the sensitive information of the first answer to obtain a second answer that can be understood locally. In this process, the first device provides the third device with answer analysis capabilities, but does not have access to sensitive information on the third device, fully protecting the privacy and security of the terminal. Because the first device in the cloud provides answer analysis capabilities, the computational overhead of obtaining answers on the terminal side is reduced, and the terminal-side functions are avoided from being overly complex.

[0199] In one possible implementation, before desensitizing the sensitive information in the first information using the first algorithm to obtain the second information, the method further includes:

[0200] receiving the first question sent by the second device through the communication interface 703;

[0201] Determining a first corpus related to the first question from a pre-stored corpus;

[0202] After restoring the sensitive information in the first answer by using the first algorithm to obtain a second answer to the first question, the method further includes:

[0203] A second answer to the first question is sent to the second device through the communication interface 703 .

[0204] In this implementation, the end-edge cloud architecture is clearly constructed by the first device, the second device and the third device. The second device initiates the question request, the first device performs the main calculation, and the third device performs the transfer and auxiliary calculation. It can provide efficient question-and-answer services for the second device without bringing large computing overhead to the second device.

[0205] In another possible implementation, determining the first corpus related to the first question from a pre-stored corpus includes:

[0206] Searching a pre-stored corpus for initial corpus related to the first question;

[0207] The initial corpus is segmented, compressed, and concatenated in sequence to obtain a first corpus with a smaller data volume.

[0208] In this implementation, the first corpus obtained by processing the initial corpus related to the first question has a higher density of useful information, which can reduce the computational overhead during model training without losing too much useful information.

[0209] In yet another possible implementation, the method is applied to a third device, the third device and the second device are both deployed in a first local area network, and the first device is deployed in a network outside the first local area network.

[0210] It can be understood that the third device and the second device are both deployed in the first local area network. Therefore, during the entire process, the original corpus information is only in the first local area network, which builds a security barrier for the third device and the second device, further improving the security of the end side.

[0211] In yet another possible implementation, restoring the sensitive information in the first answer using the first algorithm to obtain a second answer to the first question includes:

[0212] restoring the sensitive information in the first answer by the first algorithm to obtain a third answer containing the sensitive information;

[0213] A key information verification is performed on the third answer including the sensitive information to obtain a second answer to the first question.

[0214] In yet another possible implementation, performing key information verification on the third answer including sensitive information to obtain the second answer to the first question includes:

[0215] Perform key information verification on the third answer containing sensitive information;

[0216] If the key information of the third answer including the sensitive information is incorrect, correcting the incorrect key information, wherein the corrected third answer is the second answer to the first question;

[0217] If the key information of the third answer including the sensitive information is correct, the third answer including the sensitive information is the second answer to the first question.

[0218] In this implementation, the third device will also verify and correct the answer provided by the first device, which can effectively improve the accuracy and rationality of the answer and alleviate the situation where the answer is wrong or has low relevance due to the misunderstanding error of the large model of the first device.

[0219] In another possible implementation, the correcting of incorrect key information includes:

[0220] Correct incorrect key information using stored information.

[0221] In yet another possible implementation, the key information includes keywords and / or key logic.

[0222] In another possible implementation, desensitizing the sensitive information in the first information by using the first algorithm to obtain the second information includes:

[0223] Calling a hash mapping table to determine desensitized information corresponding to the sensitive information in the first information;

[0224] Replacing the sensitive information in the first information with the desensitized information to obtain second information;

[0225] Restoring the sensitive information in the first answer by using the first algorithm to obtain a second answer to the first question includes:

[0226] Calling the hash map to determine the sensitive information corresponding to the desensitized information in the first answer;

[0227] The desensitized information in the first answer is replaced with the sensitive information to obtain a second answer to the first question.

[0228] It should be noted that the implementation of each operation may also correspond to the corresponding description of the method embodiment shown in FIG. 4 .

[0229] An embodiment of the present application also provides a chip system, which includes at least one processor, a memory and an interface circuit, wherein the memory, the transceiver and the at least one processor are interconnected through lines, and a computer program is stored in the at least one memory; when the computer program is executed by the processor, the method flow shown in Figure 4 is implemented.

[0230] An embodiment of the present application further provides a computer-readable storage medium, in which a computer program is stored. When the computer-readable storage medium is executed on a processor, the method flow shown in FIG4 is implemented.

[0231] An embodiment of the present application further provides a computer program product, which, when executed on a processor, implements the method flow shown in FIG4 .

[0232] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by a computer program or computer program-related hardware. The computer program can be stored in a computer-readable storage medium. When executed, the computer program can include the processes in the above-described method embodiments. The aforementioned storage medium includes various media capable of storing computer program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for obtaining answers based on a large model, characterized in that: include: Desensitizing sensitive information in the first information using a first algorithm to obtain second information, wherein the first information includes a first question and a first corpus related to the first question; the second information includes a second question and a second corpus, the second question includes the desensitized version of the first question, and the second corpus includes the desensitized version of the first corpus; sending the second information to the first device; receiving a first answer sent by the first device, where the first answer is an answer to the second question analyzed from the second corpus; The sensitive information in the first answer is restored by using the first algorithm to obtain a second answer to the first question.

2. The method according to claim 1, characterized in that Before desensitizing the sensitive information in the first information by using the first algorithm to obtain the second information, the method further includes: receiving the first question sent by the second device; Determining a first corpus related to the first question from a pre-stored corpus; After restoring the sensitive information in the first answer by using the first algorithm to obtain a second answer to the first question, the method further includes: A second answer to the first question is sent to the second device.

3. The method according to claim 2, characterized in that The determining of a first corpus related to the first question from a pre-stored corpus includes: Searching a pre-stored corpus for initial corpus related to the first question; The initial corpus is segmented, compressed, and concatenated in sequence to obtain a first corpus with a smaller data volume.

4. The method according to claim 2 or 3, characterized in that The method is applied to a third device, where the third device and the second device are both deployed in a first local area network, and the first device is deployed in a network outside the first local area network.

5. The method according to any one of claims 1 to 4, characterized in that Restoring the sensitive information in the first answer by using the first algorithm to obtain a second answer to the first question includes: restoring the sensitive information in the first answer by the first algorithm to obtain a third answer containing the sensitive information; A key information verification is performed on the third answer including the sensitive information to obtain a second answer to the first question.

6. The method according to claim 5, characterized in that The performing key information verification on the third answer including the sensitive information to obtain a second answer to the first question includes: Perform key information verification on the third answer containing sensitive information; If the key information of the third answer including the sensitive information is incorrect, correcting the incorrect key information, wherein the corrected third answer is the second answer to the first question; If the key information of the third answer including the sensitive information is correct, the third answer including the sensitive information is the second answer to the first question.

7. The method according to claim 6, characterized in that The correction of incorrect key information includes: Correct incorrect key information using stored information.

8. The method according to any one of claims 1 to 7, characterized in that Desensitizing the sensitive information in the first information by using the first algorithm to obtain the second information includes: Calling a hash mapping table to determine desensitized information corresponding to the sensitive information in the first information; Replacing the sensitive information in the first information with the desensitized information to obtain second information; Restoring the sensitive information in the first answer by using the first algorithm to obtain a second answer to the first question includes: Calling the hash map to determine the sensitive information corresponding to the desensitized information in the first answer; The desensitized information in the first answer is replaced with the sensitive information to obtain a second answer to the first question.

9. A method for obtaining answers based on a large model, characterized in that: include: A third device desensitizes sensitive information in the first information using a first algorithm to obtain second information, wherein the first information includes a first question and a first corpus related to the first question; the second information includes a second question and a second corpus, the second question includes the desensitized version of the first question, and the second corpus includes the desensitized version of the first corpus; The third device sends the second information to the first device; The first device analyzes the second corpus to obtain a first answer to the second question; The first device sends the first answer to the third device; receiving a first answer sent by the first device; The sensitive information in the first answer is restored by using the first algorithm to obtain a second answer to the first question.

10. A device for obtaining an answer based on a large model, characterized in that: include: a processing unit configured to desensitize sensitive information in the first information using a first algorithm to obtain second information, wherein the first information includes a first question and a first corpus related to the first question; the second information includes a second question and a second corpus, the second question includes the desensitized version of the first question, and the second corpus includes the desensitized version of the first corpus; a communication unit, configured to send the second information to the first device; The communication unit is further configured to receive a first answer sent by the first device, where the first answer is an answer to the second question analyzed from the second corpus; The processing unit is further configured to restore the sensitive information in the first answer using the first algorithm to obtain a second answer to the first question.

11. The device according to claim 10, characterized in that: The communication unit is further configured to receive the first question sent by the second device; The processing unit is further configured to determine a first corpus related to the first question from a pre-stored corpus; The communication unit is further configured to send a second answer to the first question to the second device.

12. The device according to claim 11, characterized in that The determining of the first corpus aspect related to the first question from the pre-stored corpus, the processing unit is specifically configured to: Searching a pre-stored corpus for initial corpus related to the first question; The initial corpus is segmented, compressed, and concatenated in sequence to obtain a first corpus with a smaller data volume.

13. The device according to claim 11 or 12, characterized in that The apparatus is applied to a third device, the third device and the second device are both deployed in a first local area network, and the first device is deployed in a network outside the first local area network.

14. The device according to any one of claims 10 to 13, characterized in that In terms of restoring the sensitive information in the first answer by using the first algorithm to obtain the second answer to the first question, the processing unit is specifically configured to: restoring the sensitive information in the first answer by the first algorithm to obtain a third answer containing the sensitive information; A key information verification is performed on the third answer including the sensitive information to obtain a second answer to the first question.

15. The device according to claim 14, characterized in that In terms of performing key information verification on the third answer including sensitive information to obtain the second answer to the first question, the processing unit is specifically configured to: Perform key information verification on the third answer containing sensitive information; If the key information of the third answer including the sensitive information is incorrect, correcting the incorrect key information, wherein the corrected third answer is the second answer to the first question; If the key information of the third answer including the sensitive information is correct, the third answer including the sensitive information is the second answer to the first question.

16. The device according to claim 15, characterized in that In terms of correcting incorrect key information, the processing unit is specifically configured to: Correct incorrect key information using stored information.

17. The device according to any one of claims 10 to 16, characterized in that In terms of desensitizing the sensitive information in the first information by using the first algorithm to obtain the second information, the processing unit is specifically configured to: Calling a hash mapping table to determine desensitized information corresponding to the sensitive information in the first information; Replacing the sensitive information in the first information with the desensitized information to obtain second information; In terms of restoring the sensitive information in the first answer by using the first algorithm to obtain the second answer to the first question, the processing unit is specifically configured to: Calling the hash map to determine the sensitive information corresponding to the desensitized information in the first answer; The desensitized information in the first answer is replaced with the sensitive information to obtain a second answer to the first question.

18. A large model-based answer acquisition system, characterized in that: The system comprises a first device, a second device, and a third device, wherein: A third device desensitizes sensitive information in the first information using a first algorithm to obtain second information, wherein the first information includes a first question and a first corpus related to the first question; the second information includes a second question and a second corpus, the second question includes the desensitized version of the first question, and the second corpus includes the desensitized version of the first corpus; The third device sends the second information to the first device; The first device analyzes the second corpus to obtain a first answer to the second question; The first device sends the first answer to the third device; receiving a first answer sent by the first device; The sensitive information in the first answer is restored by using the first algorithm to obtain a second answer to the first question.

19. A device, characterized in that Includes a processor, a transceiver, and a memory, wherein: The memory is used to store a computer program, the transceiver is used to receive and send data, and the processor is used to call the computer program to implement the method according to any one of claims 1 to 8.

20. A chip system, characterized in that: The chip system includes at least one processor, a memory and an interface circuit, the memory, the interface circuit and the at least one processor are interconnected by lines, and a computer program is stored in the at least one memory; when the computer program is executed by the processor, the method described in any one of claims 1 to 8 is implemented.

Images