Model correction system and method based on AI computing device

Abstract

The invention discloses a model correction system and method based on an AI computing device, and the system comprises a question input module which is used for a user to input a question needing to be known; the local preprocessing module is used for preprocessing questions input by a user, preprocessing comprises keyword desensitization, question classification and question and file format optimization, and the data receiving and transmitting module is used for transmitting the optimized questions and files to the selected large model and receiving original answers fed back by the large model; and the local verification module is used for verifying the original answer fed back by the large model. According to the method, sensitive information is prevented from being uploaded to the cloud, meanwhile, due to the portability of the AI computing device, a user can conveniently ask questions and solve the questions anytime and anywhere, and the calling cost of a large model can be reduced; the question and answer gap period of the AI calculation device is fully utilized, and under the condition that the predicted question is the same as the secondary question asking question of the user, the response time of the user can be greatly saved, so that the question and answer delay is reduced, and the use experience is improved.

Description

Technical Field

[0001] The present invention relates to the field of AI computing devices, and more specifically, to a model correction system and method based on an AI computing device. Background Art

[0002] A model correction system based on an AI computing device is an intelligent system that uses artificial intelligence technology to dynamically optimize, error-fix, or improve the performance of existing models. Such systems typically combine real-time data feedback, algorithm optimization, and hardware acceleration to achieve the adaptability and robustness of the model in complex environments.

[0003] However, the current AI computing device has a low degree of optimization for problems and a poor scheduling method for large models, resulting in unsatisfactory reply results of large models and an inability to fully utilize large models, affecting the user experience. Summary of the Invention

[0004] Aiming at the problems existing in the prior art, the purpose of the present invention is to provide a model correction system and method based on an AI computing device.

[0005] To solve the above problems, the present invention adopts the following technical solutions:

[0006] A model correction system based on an AI computing device, comprising:

[0007] A problem input module, configured to receive the problem content that the user needs to understand, wherein the input methods include text entry, voice entry, picture entry, and video entry;

[0008] A local preprocessing module, configured to preprocess the problem input by the user, and the preprocessing includes keyword desensitization, problem classification, and optimization of the problem and file format;

[0009] A data transceiver module, configured to transmit the optimized problem and file to the selected large model and receive the original answer fed back by the large model;

[0010] A local verification module, configured to verify the original answer fed back by the large model;

[0011] A model correction module, configured to generate a correction strategy for the original answer that fails to pass the verification by the local verification module;

[0012] A model optimization module, configured to re-call the large model through the data transceiver module according to the correction strategy of the model correction module;

[0013] A result output module, configured to output and display the processed problem result;

[0014] A keyword addition module, which is used to add keywords of the same category as the question to the question input by the user to form a predicted question, and send it to the large model through the data transceiver module;

[0015] A secondary question comparison module, which is used to compare the coincidence degree between the user's secondary question and the predicted question.

[0016] As a further description of the above technical solution:

[0017] The question input module is the input unit of the AI computing device, and this input unit can be one of an entity button, a touch button, voice input, and data input.

[0018] As a further description of the above technical solution:

[0019] The local preprocessing module includes a privacy desensitization module, a routing decision engine, and a question optimization module. The privacy desensitization module is used to identify and delete sensitive information in the question input by the user; the routing decision engine is used to analyze the true intention of the user's question and select a model with the highest confidence weight in the large model API cluster; the question optimization module is used to convert the text formats of voice, pictures, and video files and optimize the specifications and formats of the text.

[0020] As a further description of the above technical solution:

[0021] The local verification module includes a knowledge graph verification unit, a rule detection unit, and a template matching unit. The knowledge graph verification unit is used to fact-check the answer feedback by the large model; the rule detection unit is used to detect the logical order of the answer feedback by the large model; the template matching unit is used to determine whether the answer feedback by the large model conforms to the format specification.

[0022] As a further description of the above technical solution:

[0023] It also includes a delayed question module, which is used to receive the question from the keyword addition module, ask the target large model through the data transceiver module after the user's independent question obtains the result output module, and cache the feedback answer.

[0024] As a further description of the above technical solution:

[0025] The delayed question module includes a delayed output unit, a question termination instruction sending module, and an answer caching unit. The delayed output unit is used to delay the output of the question to the data transceiver module; the question termination instruction sending module immediately sends a question termination instruction through the data transceiver module after obtaining the comparison failure information of the secondary question comparison module before caching the answer in the answer caching unit; the answer caching unit is used to cache the reply content of the large model for the predicted question.

[0026] The present invention also adopts:

[0027] A model correction method based on an AI computing device, and the specific correction steps are as follows:

[0028] Step 1: The user asks a question and inputs the question through the question input module of the AI computing device;

[0029] Step 2: The local preprocessing module performs preprocessing work on the input question, such as keyword desensitization, question classification, and optimizing the question and file format, to obtain a refined question that meets the format requirements;

[0030] Step 3: Send the refined question to the large model selected by the routing decision engine through the data transceiver module and receive the answer feedback from the large model;

[0031] Step 4: The local verification module verifies the answer feedback received from the large model. If the verification passes, directly enter Step 6; if the verification fails, enter the next step;

[0032] Step 5: For the question with the original answer that fails the verification, generate a correction strategy through the model correction module, and then the model optimization module calls the large model again through the data transceiver module according to the correction strategy of the model correction module, and enter Step 3;

[0033] Step 6: Output the answer from the large model that meets the requirements.

[0034] Compared with the prior art, the advantages of the present invention are:

[0035] In this solution, an AI computing device is used as a bridge between the user and the large model, and data is processed locally, avoiding uploading sensitive information to the cloud. At the same time, due to the portability of the AI computing device, it is convenient for users to ask questions and get answers anytime and anywhere, which can reduce the calling cost of the large model;

[0036] Make full use of the question-and-answer gap period of the AI computing device. When the predicted question is the same as the question asked by the user for the second time, it can greatly save the response time of the user, thereby reducing the question-and-answer delay and improving the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of the principle of the present invention;

[0038] Figure 2 It is a schematic diagram of the principle of the local preprocessing module of the present invention;

[0039] Figure 3 It is a schematic diagram of the principle of the local verification module of the present invention;

[0040] Figure 4 It is a schematic diagram of the principle of the predicted question of the present invention;

[0041] Figure 5 This is a schematic flow chart of the present invention.

[0042] Explanation of the reference numerals in the figure:

[0043] 1. Problem input module; 2. Local preprocessing module; 21. Privacy desensitization module; 22. Routing decision engine; 23. Problem optimization module; 3. Data transceiver module; 4. Local verification module; 41. Knowledge graph verification unit; 42. Rule detection unit; 43. Template matching unit; 5. Model correction module; 6. Model optimization module; 7. Result output module; 8. Keyword addition module; 9. Secondary problem comparison module; 10. Delayed question module; 101. Delayed output unit; 102. Question termination instruction issuing module; 103. Answer cache unit. Detailed implementation manners

[0044] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention;

[0045] Please refer to Figures 1-4 , the present invention provides Embodiment 1:

[0046] A model correction system based on an AI computing device, comprising:

[0047] A problem input module 1, configured to receive the problem content that the user needs to understand, wherein the input methods include text input, voice input, picture input, and video input; wherein the problem input module 1 is an input unit of the AI computing device, and the input unit can be one of an entity button, a touch button, voice input, and data input. The input method can select a scanning pen technology to scan text, input the voice data of the user, or import audio files, picture files, video files, and document files through wired and wireless methods, so as to provide various ways of asking questions to meet the user's question-asking needs;

[0048] A local preprocessing module 2, configured to preprocess the problem input by the user, and the preprocessing includes keyword desensitization, problem classification, and optimization of the problem and file format;

[0049] The data sending and receiving module 3 is used to send the optimized questions and files to the selected large model and receive the original answers fed back by the large model; the local preprocessing module 2 includes a privacy desensitization module 21, a routing decision engine 22 and a question optimization module 23. The privacy desensitization module 21 is used to identify and delete sensitive information in the questions input by users; the routing decision engine 22 is used to analyze the true intention of the user's questions and select a model with the highest confidence weight in the large model API cluster; the question optimization module 23 is used to convert the text formats of voice, pictures and video files and optimize the specifications and formats of the texts;

[0050] The local verification module 4 verifies the original answers fed back by the large model; the local verification module 4 includes a knowledge graph verification unit 41, a rule detection unit 42 and a template matching unit 43. The knowledge graph verification unit 41 is used to conduct fact-checking on the answers fed back by the large model; the rule detection unit 42 is used to detect the logical order of the answers fed back by the large model; the template matching unit 43 is used to determine whether the answers fed back by the large model conform to the format specifications.

[0051] The model correction module 5 generates correction strategies for the original answers that fail to pass the verification by the local verification module 4;

[0052] The model optimization module 6 re-calls the large model through the data sending and receiving module 3 according to the correction strategies of the model correction module 5;

[0053] The result output module 7 is used to output and display the processed question results;

[0054] Using an AI computing device as a bridge between the user and the large model can optimize the Q&A content between the user and the large model, perform local data processing, avoid uploading sensitive information to the cloud. At the same time, due to the portability of the AI computing device, it is convenient for users to ask questions and get answers anytime and anywhere. Meanwhile, optimizing the questions and selecting a suitable large model for answering can reduce the calling cost of the large model.

[0055] Please refer to Figures 1-4 , on the basis of Embodiment 1, the present invention also provides Embodiment 2:

[0056] The keyword adding module 8 is used to add keywords of the same category as the question to the questions input by users to form predicted questions, and send them to the large model through the data sending and receiving module 3;

[0057] The secondary question comparison module 9 is used to compare the coincidence degree between the user's secondary questions and the predicted questions.

[0058] The delayed question asking module 10 is used to receive the questions of the keyword adding module 8, ask questions to the target large model through the data sending and receiving module 3 after the user independently asks questions and obtains the result output module 7, and cache the fed-back answers;

[0059] The delayed question module 10 includes a delayed output unit 101, a question termination instruction issuing module 102, and an answer caching unit 103. The delayed output unit 101 is used to delay the output of questions to the data transceiver module 3. The question termination instruction issuing module 102 immediately sends a question termination instruction through the data transceiver module 3 after obtaining the comparison failure information of the secondary question comparison module 9 before caching the answer in the answer caching unit 103. The answer caching unit 103 is used to cache the large model reply content of the predicted question.

[0060] Among them, the keyword addition module 8 is based on the question raised by the user, and forms a new question or refines the question by adding keywords of the same category, and then through the delayed question module 10, after the user obtains the reply to the preliminary question, when the user reads the reply to the preliminary question, asks the large model, and buffers the reply of the large model through the answer caching unit 103. When the user makes a secondary question, the secondary question comparison module 9 is used to compare the question content. When the true intentions of the two questions are the same, the answer in the answer caching unit 103 is directly output, so as to reduce the response time obtained by the user, improve the user experience, and save the Q&A time. When the true intentions of the two questions are different, the reply in the answer caching unit 103 is deleted, and the user question is used for Q&A;

[0061] Among them, if before caching the answer in the answer caching unit 103, after obtaining the comparison failure information of the secondary question comparison module 9, immediately send a question termination instruction through the data transceiver module 3 to avoid the problem of occupying the large model;

[0062] Make full use of the Q&A gap period of the AI computing device. When the predicted question is the same as the user's secondary question, it can greatly save the response time obtained by the user, thereby reducing the Q&A delay and improving the user experience.

[0063] Please refer to Figures 1-5 , on the basis of Embodiment 1 and Embodiment 2, the present invention further provides Embodiment 3:

[0064] A model correction method based on an AI computing device, and the specific correction steps are as follows:

[0065] Step 1: The user asks a question, and the question is input through the question input module 1 of the AI computing device;

[0066] Step 2: The local preprocessing module 2 performs preprocessing work on the input question, such as keyword desensitization, question classification, and optimizing the question and file format, to obtain a refined question that meets the format requirements;

[0067] Step 3: Send the streamlined problem to the large model selected by the routing decision engine 22 through the data transceiver module 3, and receive the answer feedback from the large model;

[0068] Step 4: The local verification module 4 verifies the answer feedback received from the large model. If the verification passes, directly proceed to Step 6; if the verification fails, proceed to the next step;

[0069] Step 5: For the problem with the original answer that fails the verification, generate a correction strategy through the model correction module 5. Then, according to the correction strategy of the model correction module 5, the model optimization module 6 re - invokes the large model through the data transceiver module 3, and enter Step 3;

[0070] Step 6: Output the answer from the large model that meets the requirements.

[0071] As described above, it is only a preferred specific embodiment of the present invention; however, the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution of the present invention and its improved concept, makes equivalent substitution or change, and should be covered by the protection scope of the present invention.

Claims

1. An AI computing device-based model correction system, characterized in that, Including: A question input module (1) for a user to input the content of the question to be understood, where the input methods include text input, voice input, picture input, and video input; A local preprocessing module (2) for preprocessing the question input by the user, and the preprocessing includes keyword desensitization, question classification, and optimizing the question and file format; A data sending and receiving module (3) for sending the optimized question and file to the selected large model and receiving the original answer fed back by the large model; A local verification module (4) for verifying the original answer fed back by the large model; A model correction module (5) for generating a correction strategy for the original answer that fails to pass the verification by the local verification module (4); A model optimization module (6) for re-calling the large model through the data sending and receiving module (3) according to the correction strategy of the model correction module (5); A result output module (7) for outputting and displaying the processed question result; A keyword adding module (8) for adding keywords of the same question category to the question input by the user to form a predicted question and sending it to the large model through the data sending and receiving module (3); A secondary question comparison module (9) for comparing the coincidence degree between the user's secondary question and the predicted question.

2. The model correction system based on an AI computing device according to claim 1, wherein: The question input module (1) is an input unit of the AI computing device, and this input unit can be one of an entity button, a touch button, voice input, and data input.

3. The model correction system based on the AI computing device according to claim 1, wherein: The local preprocessing module (2) includes a privacy desensitization module (21), a routing decision engine (22), and a question optimization module (23). The privacy desensitization module (21) is used to identify and delete sensitive information in the question input by the user; The routing decision engine (22) is used to analyze the true intention of the user's question and select a model with the highest confidence weight in the large model API cluster; the question optimization module (23) is used to convert voice, picture, and video files into text format and optimize the specifications and formats of the text.

4. The model correction system based on an AI computing device according to claim 1, wherein: The local verification module (4) includes a knowledge graph verification unit (41), a rule detection unit (42), and a template matching unit (43). The knowledge graph verification unit (41) is used to verify the facts of the answer fed back by the large model; the rule detection unit (42) is used to detect the logical order of the answer fed back by the large model; the template matching unit (43) is used to determine whether the answer fed back by the large model conforms to the format specification.

5. The model correction system based on an AI computing device according to claim 1, wherein: It further includes a delayed question module (10). The delayed question module (10) is used to receive the question from the keyword adding module (8), ask the target large model through the data sending and receiving module (3) after the user independently asks a question and obtains the result output module (7), and cache the fed-back answer.

6. The model correction system based on the AI computing device according to claim 5, characterized in that: The delay question module (10) includes a delay output unit (101), a question termination instruction issuing module (102) and an answer cache unit (103). The delay output unit (101) is used to output questions to the data transceiver module (3) with a delay; the question termination instruction issuing module (102) immediately sends a question termination instruction through the data transceiver module (3) after obtaining the information that the comparison by the secondary question comparison module (9) fails before the answer cache unit (103) caches the answer; the answer cache unit (103) is used to cache the large model reply content of the predicted questions.

7. A correction method applied to the model correction system of the AI computing device according to any one of claims 1-6, characterized in that: The specific correction steps are as follows: Step 1: The user asks a question and inputs the question through the question input module (1) of the AI computing device. Step 2: The local preprocessing module (2) performs preprocessing work on the input question, including keyword desensitization, question classification, and optimizing the question and file format, to obtain a refined question that meets the format requirements. Step 3: The refined question is sent to the large model selected by the routing decision engine (22) through the data transceiver module (3), and the large model feedback answer is received. Step 4: The local verification module (4) verifies the large model feedback answer received. If the verification passes, it directly enters Step 6; if the verification fails, it enters the next step. Step 5: For the question with the original answer that fails the verification, the model correction module (5) generates a corresponding correction strategy, and then the model optimization module (6) calls the large model again through the data transceiver module (3) according to the correction strategy of the model correction module (5), and enters Step 3. Step 6: Output the large model reply that meets the requirements.

Images