AI Question Answering Method and System for a Large Language Model Knowledge Base Based on Speech Recognition

Abstract

The present invention provides a large language model knowledge base AI question-answering method and system based on speech recognition, including: initially recognizing the question voice issued by the user to determine the user's question purpose, retrieving relevant response knowledge bases, constructing a response model based on the response knowledge bases, training the response model according to the question purpose to obtain several response texts, retrieving a corresponding virtual response character according to the question scenario information where the user is located, controlling the virtual response character to conduct a response with the user, obtaining response information, self-evaluating the current round of response according to the response information, generating information improvement feedback and knowledge base update feedback according to the self-evaluation result, adjusting the model construction method according to the information improvement feedback, adjusting the knowledge base according to the knowledge base update feedback, driving a three-dimensional human face through speech synthesis and speech recognition technologies, enabling the digital human to interact with the user, and realizing a dialogue scenario similar to that between humans.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence question - answering, and particularly to a large - language model knowledge - base AI question - answering method and system based on speech recognition. Background Art

[0002] Artificial intelligence has deeply penetrated into various fields, including healthcare, finance, education, transportation, entertainment, etc. In the medical field, AI is used for disease diagnosis, prediction, and even assisting doctors during surgeries. In the financial field, AI is used for risk management, stock trading analysis, and customer service. This expansion has had a profound impact on the lives of ordinary people.

[0003] Currently, the application of artificial intelligence has penetrated into various industries. Internet companies, financial institutions, the medical industry, etc. are all actively exploring how to use artificial intelligence technology to provide better services. However, current artificial intelligence can only initiate interactions with users online, and users lack a sense of experience.

[0004] Therefore, the present invention provides a large - language model knowledge - base AI question - answering method and system based on speech recognition. Summary of the Invention

[0005] The large - language model knowledge - base AI question - answering method based on speech recognition of the present invention drives a three - dimensional face through speech synthesis and speech recognition technologies, enabling a digital human to interact with a user and realizing a dialogue scenario similar to that between humans.

[0006] The present invention provides a large - language model knowledge - base AI question - answering method based on speech recognition, including:

[0007] Step 1: Initially recognize the question speech issued by the user, determine the purpose of the user's question, and retrieve the relevant Q&A knowledge base.

[0008] Step 2: Construct a Q&A model according to the Q&A knowledge base, train the Q&A model according to the purpose of the question, and obtain several Q&A texts.

[0009] Step 3: Retrieve the corresponding virtual Q&A character according to the question - asking scenario information where the user is located, and control the virtual Q&A character to conduct a Q&A with the user.

[0010] Step 4: Obtain Q&A information, conduct self - evaluation on the current round of Q&A according to the Q&A information, and generate information improvement feedback and knowledge - base update feedback according to the self - evaluation result.

[0011] Step 5: Adjust the model construction method according to the information improvement feedback, and adjust the knowledge base according to the knowledge - base update feedback.

[0012] In an implementable manner,

[0013] Step 1 includes:

[0014] Step 11: Obtain the question voice issued by the user, establish the speech rhythm of the question voice according to the prosodic features of the question voice, divide the question voice into several rhythm segments according to the speech rhythm, and establish the recognition frequency corresponding to each rhythm segment according to the rhythm strength feature corresponding to each rhythm segment;

[0015] Step 12: Based on the recognition frequency and using a preset speech recognition technology, perform language recognition on the corresponding rhythm segment to obtain the semantic information corresponding to each rhythm segment, and reorganize the semantic information to obtain the question text of the user;

[0016] Step 13: Perform semantic analysis on the question text to obtain several question purposes of the user, respectively establish a knowledge index corresponding to each question purpose, and use the knowledge index to search for an answer knowledge base with a relevance greater than a preset relevance threshold related to the corresponding question purpose.

[0017] In an implementable manner,

[0018] Step 2 includes:

[0019] Step 21: Obtain several answer knowledge bases corresponding to the question voice, construct a knowledge structure network corresponding to each answer knowledge base according to the knowledge points included in each answer knowledge base, and analyze the knowledge logic between different answer knowledge bases according to the knowledge structure network;

[0020] Step 22: Construct logical docking nodes for each answer knowledge base according to the knowledge logic, and use the docking directions corresponding to the logical docking nodes to arrange the logical libraries in a logical distribution to generate an answer model;

[0021] Step 23: Sort out the knowledge of the question purpose to obtain several relevant knowledge points included in the question purpose, search for each relevant knowledge point in the answer model, and construct a knowledge structure corresponding to each relevant knowledge point according to the search result;

[0022] Step 24: Establish an answer format according to the question purpose, and respectively perform knowledge matching on each knowledge structure according to the question purpose, and reorganize the matched knowledge according to the answer format to obtain several answer texts.

[0023] In an implementable manner,

[0024] Step 3 includes:

[0025] Step 31: Obtain the question-asking scenario information of the user, construct a conceptual scenario based on the question-asking scenario information, and perform dynamic analysis and static analysis on the conceptual scenario to obtain the dynamic scenario features and static scenario features included in the conceptual scenario;

[0026] Step 32: Use the dynamic scenario features to construct a conceptual time sequence diagram of the conceptual scenario, use the static scenario features to construct a static background diagram of the conceptual scenario, map the conceptual time sequence diagram to the static background diagram, and obtain the dynamic change rules corresponding to each scenario area in the conceptual scenario;

[0027] Step 33: Construct the overall change rule of the conceptual scenario according to the dynamic change rule corresponding to each scenario area, use the overall change rule to retrieve the corresponding character appearance information, construct the corresponding character activity information according to the scenario outside of the conceptual scenario, and construct the response volume of the character corresponding to each scenario area according to the dynamic change rule corresponding to each scenario area and the character activity information;

[0028] Step 34: Construct a virtual response character according to the character appearance information, and construct the activity trajectory of the virtual response character and the response volume of the virtual response character during the activity according to the character activity information and the response volume;

[0029] In an implementable manner,

[0030] Step 3 further includes:

[0031] Step 35: Retrieve the corresponding response atmosphere according to the conceptual scenario, obtain the response text, add the corresponding atmosphere tone to the response text according to the response atmosphere, and generate an atmosphere response text;

[0032] Step 36: Control the virtual response character to conduct a response with the user using the atmosphere response text.

[0033] In an implementable manner,

[0034] Step 4 includes:

[0035] Step 41: Obtain the response information between the user and the response model, perform question-answer separation on the response information to obtain the question information corresponding to the user and the answer information of the response model;

[0036] Step 42: Establish the change characteristics of the user's question-asking tone according to the question information, construct the satisfaction characteristics of the user for the current round of question and answer according to the change characteristics of the question-asking tone, establish the question asked by the user according to the question information, and construct the question answered by the response model according to the answer information;

[0037] Step 43: Analyze the matching degree between the asked question and the answered question, and establish a matching fit feature for this round of Q&A according to the matching degree;

[0038] Step 44: Establish a self-evaluation result for this round of Q&A according to the satisfaction feature and the matching fit feature, and construct information improvement feedback and knowledge base update feedback according to the self-evaluation result.

[0039] In an implementable manner,

[0040] The said step 5 includes:

[0041] Step 51: Determine the model defect feature of the dialogue model according to the information improvement feedback, and determine the knowledge to be supplemented feature of the corresponding dialogue knowledge base according to the knowledge update feedback;

[0042] Step 52: Establish the defect position of the dialogue model according to the model defect feature, and adjust the model structure of the dialogue model according to the defect position;

[0043] Step 53: Retrieve the knowledge to be supplemented in the knowledge base update feedback, determine the knowledge association method between the knowledge to be supplemented and the existing knowledge based on the knowledge to be supplemented adjustment, and supplement the knowledge to be supplemented into the corresponding dialogue knowledge base by using the knowledge association method.

[0044] In an implementable manner,

[0045] It further includes:

[0046] Generate an updated dialogue model after completing the model structure adjustment, and use the updated dialogue model to replace the corresponding dialogue model;

[0047] Use the updated dialogue model to conduct a dialogue with the user.

[0048] The present invention provides a large language model knowledge base AI Q&A system based on speech recognition, including:

[0049] An information preprocessing module, used for initially recognizing the question voice sent by the user, determining the asking purpose of the user, and retrieving the relevant dialogue knowledge base;

[0050] A model construction and usage module, used for constructing a dialogue model according to the dialogue knowledge base, training the dialogue model according to the asking purpose, and obtaining a number of dialogue texts;

[0051] A virtual character dialogue module, used for retrieving the corresponding virtual dialogue character according to the asking scenario information where the user is located, and controlling the virtual dialogue character to conduct a dialogue with the user;

[0052] The self-evaluation analysis module is used to obtain the response information, conduct self-evaluation on the current round of response according to the response information, and generate information improvement feedback and knowledge base update feedback according to the self-evaluation results;

[0053] The update and reorganization module is used to adjust the model construction method according to the information improvement feedback, and adjust the knowledge base according to the knowledge base update feedback.

[0054] In an implementable manner,

[0055] The information preprocessing module includes:

[0056] The first preprocessing unit is used to obtain the question voice issued by the user, establish the speech rhythm of the question voice according to the prosodic features of the question voice, divide the question voice into several rhythm segments according to the speech rhythm, and establish the recognition frequency of the corresponding rhythm segment according to the rhythm strength feature corresponding to each rhythm segment;

[0057] The second preprocessing unit is used to perform language recognition on the corresponding rhythm segment based on the recognition frequency and using a preset speech recognition technology, obtain the semantic information corresponding to each rhythm segment, and reorganize the semantic information to obtain the user's question text;

[0058] The third preprocessing unit is used to perform semantic analysis on the question text to obtain several question purposes of the user, respectively establish a knowledge index corresponding to each question purpose, and use the knowledge index to search for a response knowledge base with a relevance greater than a preset relevance threshold related to the corresponding question purpose.

[0059] The beneficial effects that the present invention can achieve are as follows: In order to construct a beautiful and realistic virtual character to communicate with the user, when the user issues a question, relevant response knowledge bases are retrieved according to the user's question purpose, and then a response model is established, thereby constructing several response texts related to the question purpose. Then, a virtual response character is established according to the question scenario where the user is located, and then the virtual response character is used to communicate with the user. In order to continuously improve and optimize the model, self-evaluation is performed after each round of question and answer, and the model and the knowledge base are adjusted according to the evaluation results. In this way, through the AI large language model interaction service, the thresholds for consulting, handling affairs, and learning are reduced, facilitating users to obtain the required information and services; the platform can also access the external network, realizing an upgrade of the innovative interaction experience, meeting diverse needs, and solving problems such as insufficient manpower and high personnel pressure, playing a role in cost reduction and efficiency improvement, promoting digital construction, and assisting in the layout of the metaverse. Additionally, through virtual digital human anchors, merchants can display product and service features, attract more target customers, increase the exposure opportunities of merchants, and improve popularity and influence. At the same time, intelligent interaction can meet user needs, improve user conversion rates and repurchase rates.

[0060] Other features and advantages of the present invention will be set forth in the following description, and in part will be obvious from the description, or may be learned by practice of the present invention. The objectives and other advantages of the present invention may be realized and attained by the structure particularly pointed out in the written description, claims, as well as the drawings.

[0061] The technical solutions of the present invention will be further described in detail below through the drawings and embodiments. Description of the Drawings

[0062] The drawings are used to provide a further understanding of the present invention, and constitute a part of the description. They are used together with the embodiments of the present invention to explain the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0063] Figure 1 It is a schematic diagram of the working process of the large language model knowledge base AI question and answer method based on speech recognition in the embodiment of the present invention;

[0064] Figure 2 It is a schematic diagram of the composition of the large language model knowledge base AI question and answer system based on speech recognition in the embodiment of the present invention. Detailed Embodiments

[0065] The following describes the preferred embodiments of the present invention with reference to the drawings. It should be understood that the preferred embodiments described herein are only for the purpose of illustrating and explaining the present invention, and are not used to limit the present invention.

[0066] Embodiment 1

[0067] This embodiment provides a large language model knowledge base AI question and answer method based on speech recognition, as Figure 1 shown, including:

[0068] Step 1: Initially recognize the question voice issued by the user, determine the purpose of the user's question, and retrieve the relevant Q&A knowledge base;

[0069] Step 2: Construct a Q&A model according to the Q&A knowledge base, train the Q&A model according to the purpose of the question, and obtain several Q&A texts;

[0070] Step 3: Retrieve the corresponding virtual Q&A character according to the question scenario information where the user is located, and control the virtual Q&A character to conduct a Q&A with the user;

[0071] Step 4: Obtain the Q&A information, conduct a self-evaluation of the current round of Q&A according to the Q&A information, and generate information improvement feedback and knowledge base update feedback according to the self-evaluation result;

[0072] Step 5: Improve the feedback adjustment model construction method according to the information, and update the knowledge base according to the knowledge base for feedback adjustment.

[0073] In this example, the initial recognition represents the text obtained by processing and conversion, and operations such as semantic analysis and entity recognition are performed to understand the user's intentions and needs. This involves knowledge related to natural language processing, machine learning, etc.;

[0074] In this example, the Q&A knowledge base represents the knowledge base related to the purpose of asking questions;

[0075] In this example, the Q&A model represents an artificial intelligence model that processes and analyzes the text input by the user to generate corresponding answers;

[0076] In this example, when establishing a virtual Q&A character, virtual reality and augmented reality technologies, character modeling and animation technologies are used to design and implement the appearance and actions of the digital human.

[0077] The working principle and beneficial effects of the above technical solution: In order to build a beautiful and realistic virtual character to communicate with the user, when the user asks a question, relevant Q&A knowledge bases are retrieved according to the purpose of the user's question, and then a Q&A model is established to construct several Q&A texts related to the purpose of the question. Then, a virtual Q&A character is established according to the question scenario where the user is located, and then the virtual Q&A character is used to communicate with the user. In order to continuously improve and optimize the model, self-evaluation is performed after each round of Q&A, and the model and knowledge base are adjusted according to the evaluation results. In this way, through the AI large language model interaction service, the thresholds for consultation, handling affairs, and learning are reduced, making it convenient for users to obtain the required information and services; the platform can also access the external network to achieve an upgrade of the innovative interaction experience, meet diverse needs, and solve problems such as insufficient manpower and high personnel pressure, playing a role in reducing costs and increasing efficiency, promoting digital construction, and assisting in the metaverse layout. Additionally, through virtual digital human anchors, merchants can display product and service features, attract more target customers, increase merchant exposure opportunities, and improve popularity and influence. At the same time, intelligent interaction can meet user needs, improve user conversion rates and repurchase rates.

[0078] Embodiment 2

[0079] Based on Embodiment 1, for the large language model knowledge base AI Q&A method based on speech recognition, Step 1 includes:

[0080] Step 11: Obtain the question voice issued by the user, establish the speech rhythm of the question voice according to the prosodic features of the question voice, divide the question voice into several rhythm segments according to the speech rhythm, and establish the recognition frequency of the corresponding rhythm segment according to the rhythm strength feature corresponding to each rhythm segment;

[0081] Step 12: Based on the recognition frequency and using a preset speech recognition technology, perform language recognition on the corresponding rhythm segments to obtain semantic information corresponding to each rhythm segment, and reorganize the semantic information to obtain the user's question text;

[0082] Step 13: Perform semantic analysis on the question text to obtain several question purposes of the user, respectively establish a knowledge index corresponding to each question purpose, and use the knowledge index to search for an answer knowledge base with a relevance greater than a preset relevance threshold related to the corresponding question purpose.

[0083] In this example, one rhythm segment corresponds to one recognition frequency, and the recognition frequency represents the frequency strength of the rhythm segment;

[0084] In this example, semantic information represents the information presented by voice in one rhythm segment;

[0085] In this example, a knowledge index represents a primer used to search for a question purpose;

[0086] In this example, the preset relevance threshold is 60%.

[0087] The working principle and beneficial effects of the above technical solution: In order to timely solve the problems raised by users, the question voice issued by the user is divided into several rhythm segments according to the speech rhythm, and the recognition frequency corresponding to each rhythm segment is determined. Then, different rhythm segments with different frequencies are recognized to obtain the semantic information corresponding to each rhythm segment. Further, the semantic information is reorganized to obtain the user's question text, and the question purpose of the user is determined. Then, a knowledge index is established for each question purpose, and an answer knowledge base with a high relevance is retrieved. In this way, each rhythm segment can be accurately analyzed to determine the semantics therein, improving the effectiveness of semantic analysis and avoiding the omission of the answer knowledge base.

[0088] Embodiment 3

[0089] Based on Embodiment 1, in the large language model knowledge base AI question-answering method based on speech recognition, Step 2 includes:

[0090] Step 21: Obtain several answer knowledge bases corresponding to the question voice, construct a knowledge structure network corresponding to each answer knowledge base according to the knowledge points included in each answer knowledge base, and analyze the knowledge logic between different answer knowledge bases according to the knowledge structure network;

[0091] Step 22: Construct logical docking nodes for each answer knowledge base according to the knowledge logic, and use the docking directions corresponding to the logical docking nodes to arrange the logical libraries in a logical distribution to generate an answer model;

[0092] Step 23: Conduct knowledge collation on the questioning purpose to obtain several related knowledge points included in the questioning purpose, search for each of the related knowledge points in the Q&A model, and construct a knowledge structure corresponding to each of the related knowledge points according to the search results;

[0093] Step 24: Establish a Q&A format according to the questioning purpose, and perform knowledge matching on each of the knowledge structures according to the questioning purpose. Reorganize the formatted knowledge according to the Q&A format to obtain several Q&A texts.

[0094] In this example, the knowledge structure network represents the relationship between the knowledge points included in a Q&A knowledge base;

[0095] In this example, the knowledge logic represents the logical relationship between different Q&A knowledge bases;

[0096] In this example, the logical docking node represents a flag with a logical relationship pointing;

[0097] In this example, the related knowledge points represent the knowledge points with a relevance greater than 60% to the questioning purpose;

[0098] In this example, the Q&A text represents the statement text used to answer the questions raised by the user.

[0099] The working principle and beneficial effects of the above technical solution: In order to achieve effective communication, construct the knowledge structure network of the Q&A knowledge base according to the knowledge points included in the retrieved Q&A knowledge base, so as to determine the knowledge logic between different knowledge structure networks, then establish logical docking nodes for the corresponding Q&A knowledge bases, and then arrange them according to the logical relationship to generate a Q&A model. Further determine several related knowledge points included in the questioning purpose, and then search for each related knowledge point in the Q&A model to construct a knowledge structure. In order to bring a better experience to the user, establish a Q&A format according to the questioning purpose, so as to perform knowledge matching on each knowledge structure and generate several Q&A texts, laying a foundation for subsequent interactive Q&A.

[0100] Example 4

[0101] Based on Example 1, the method for AI Q&A of the large language model knowledge base based on speech recognition, step 3 includes:

[0102] Step 31: Obtain the questioning scenario information where the user is located, construct a conceptual scenario according to the questioning scenario information, and perform dynamic analysis and static analysis on the conceptual scenario to obtain the dynamic scenario features and static scenario features included in the conceptual scenario;

[0103] Step 32: Construct a conceptual time-series graph of the conceptual scene using the dynamic scene features, construct a static background graph of the conceptual scene using the static scene features, and map the conceptual time-series graph into the static background graph to obtain the dynamic change rules corresponding to each scene area in the conceptual scene;

[0104] Step 33: Construct the overall change rule of the conceptual scene according to the dynamic change rule corresponding to each scene area, retrieve the corresponding character appearance information using the overall change rule, construct the corresponding character activity information according to the scene outside of the conceptual scene, and construct the response volume corresponding to the character in each scene area according to the dynamic change rule corresponding to each scene area and the character activity information;

[0105] Step 34: Construct a virtual response character according to the character appearance information, and construct the activity trajectory of the virtual response character and the response volume of the virtual response character during the activity according to the character activity information and the response volume;

[0106] In this example, the conceptual scene represents a way to represent the scene where the user is located in the form of virtual numbers;

[0107] In this example, the conceptual time-series graph represents a graph used to display data of the dynamic information in the conceptual scene changing over time;

[0108] In this example, the static background graph represents the background in the conceptual scene;

[0109] In this example, the scene area represents a part of the conceptual scene;

[0110] In this example, the character appearance information includes the gender, age, and clothing matching of the virtual response character;

[0111] In this example, the character activity information represents the activity trajectory of the virtual response character in the conceptual scene;

[0112] The working principle and beneficial effects of the above technical solution: In order to construct a beautiful and intelligent virtual response character, first construct a conceptual scene according to the question scene information where the user is located, and then perform dynamic and static analysis on the conceptual scene to determine the dynamic scene features and static scene features contained therein, so as to construct a conceptual time-series graph and a static background graph, determine the dynamic change rules of each scene area through mapping, and construct the overall change rule of the conceptual scene according to the dynamic change rules, so as to retrieve the corresponding character appearance information, character activity information, and response volume, and then construct a virtual response character and guide it to move and respond according to the specified trajectory.

[0113] Embodiment 5

[0114] Based on Embodiment 4, in the AI question-answering method of the large language model knowledge base based on speech recognition, Step 3 further includes:

[0115] Step 35: Retrieve the corresponding response atmosphere according to the concept scenario, obtain the response text, and add the corresponding atmosphere tone to the response text according to the response atmosphere to generate an atmosphere response text;

[0116] Step 36: Control the virtual response character to conduct a response with the user using the atmosphere response text.

[0117] The working principle and beneficial effects of the above technical solution: In order to provide the user with an enhanced experience and increase the user's interest in continuing to ask questions, the corresponding response atmosphere is retrieved according to the concept scenario, and the response atmosphere is used to render the response text, thus narrowing the distance between the user and the system.

[0118] Embodiment 6

[0119] Based on Embodiment 1, in the AI question-answering method of the large language model knowledge base based on speech recognition, Step 4 includes:

[0120] Step 41: Obtain the response information between the user and the response model, separate the questions and answers in the response information to obtain the question information corresponding to the user and the answer information of the response model;

[0121] Step 42: Establish the changing characteristics of the user's questioning tone according to the question information, construct the satisfaction characteristics of the user for this round of question and answer according to the changing characteristics of the questioning tone, establish the user's questioning questions according to the question information, and construct the answering questions of the response model according to the answer information;

[0122] Step 43: Analyze the matching degree between the questioning questions and the answering questions, and establish the matching fit characteristics for this round of question and answer according to the matching degree;

[0123] Step 44: Establish the self-evaluation result of this round of question and answer according to the satisfaction characteristics and the matching fit characteristics, and construct the information improvement feedback and the knowledge base update feedback according to the self-evaluation result.

[0124] Working principle and beneficial effects of the above technical solution: In order to train and update the response model and knowledge base in real time, the question information of the user and the answer information of the response model are constructed based on the conversation information between the user and the response model. Then, the changing characteristics of the user's questioning tone are established based on the question information, so as to analyze the satisfaction characteristics of the user with the current round of question and answer. And the matching degree between the response model and the user is constructed based on the answer information, so as to determine the matching degree between the two. Thus, the self-evaluation result of the original question and answer is established, and the improvement feedback and knowledge base update feedback are constructed. In this way, training can be carried out once after each question and answer, and the degree of cooperation with the user can be gradually improved during the question and answer process, achieving the purpose of accurate cooperation.

[0125] Example 7

[0126] Based on Example 1, the AI question and answer method for the large language model knowledge base based on speech recognition, step 5 includes:

[0127] Step 51: Determine the model defect characteristics of the response model according to the information improvement feedback, and determine the knowledge to be supplemented characteristics of the corresponding response knowledge base according to the knowledge update feedback;

[0128] Step 52: Establish the defect location of the response model according to the model defect characteristics, and adjust the model structure of the response model according to the defect location;

[0129] Step 53: Retrieve the knowledge to be supplemented in the knowledge base update feedback, determine the knowledge association method between the knowledge to be supplemented and the existing knowledge based on the knowledge to be supplemented adjustment, and use the knowledge association method to supplement the knowledge to be supplemented into the corresponding response knowledge base.

[0130] Working principle and beneficial effects of the above technical solution: Using the information improvement feedback to adjust the defects of the response model, and using the knowledge update feedback to fill the knowledge base. On the one hand, it can improve the intelligence of the response model, and on the other hand, it can supplement the knowledge base to build a complete knowledge base.

[0131] Example 8

[0132] Based on Example 7, the AI question and answer method for the large language model knowledge base based on speech recognition further includes:

[0133] After completing the model structure adjustment, generate an updated response model, and use the updated response model to replace the corresponding response model;

[0134] Use the updated response model to conduct a conversation with the user.

[0135] Example 9

[0136] This embodiment provides a large language model knowledge base AI Q&A system based on speech recognition, as Figure 2 shown, including:

[0137] An information preprocessing module for initially recognizing the question speech issued by the user, determining the purpose of the user's question, and retrieving relevant Q&A knowledge bases;

[0138] A model construction and usage module for constructing a Q&A model based on the Q&A knowledge base, training the Q&A model according to the purpose of the question, and obtaining several Q&A texts;

[0139] A virtual character Q&A module for retrieving a corresponding virtual Q&A character according to the question scenario information where the user is located, and controlling the virtual Q&A character to conduct a Q&A with the user;

[0140] A self-evaluation and analysis module for obtaining Q&A information, self-evaluating the current round of Q&A according to the Q&A information, and generating information improvement feedback and knowledge base update feedback according to the self-evaluation results;

[0141] An update and reorganization module for adjusting the model construction method according to the information improvement feedback, and adjusting the knowledge base according to the knowledge base update feedback.

[0142] In this example, the initial recognition refers to the text obtained by processing and conversion, and operations such as semantic analysis and entity recognition are performed to understand the user's intentions and needs. This involves relevant knowledge such as natural language processing and machine learning;

[0143] In this example, the Q&A knowledge base refers to the knowledge base related to the purpose of the question;

[0144] In this example, the Q&A model refers to an artificial intelligence model that processes and analyzes the text input by the user to generate corresponding answers;

[0145] In this example, when establishing a virtual Q&A character, virtual reality and augmented reality technologies, character modeling and animation technologies are used to design and implement the appearance and actions of the digital human.

[0146] Working principle and beneficial effects of the above technical solution: In order to build a beautiful and realistic virtual character to communicate with users, when a user asks a question, relevant Q&A knowledge bases are retrieved according to the purpose of the user's question, and then a Q&A model is established to construct several Q&A texts related to the purpose of the question. Then, a virtual Q&A character is established according to the question scenario where the user is located, and the virtual Q&A character is used to communicate with the user. In order to continuously improve and optimize the model, self-evaluation is carried out after each round of Q&A, and the model and knowledge base are adjusted according to the evaluation results. In this way, through the AI large language model interaction service, the thresholds for consultation, handling affairs, and learning are reduced, facilitating users to obtain the required information and services; the platform can also access the external network to achieve an upgrade of the innovative interaction experience, meet diverse needs, and solve problems such as insufficient manpower and high personnel pressure, playing a role in cost reduction and efficiency improvement, promoting digital construction, and assisting in the layout of the metaverse. Additionally, through virtual digital human anchors, merchants can display product and service features, attract more target customers, increase merchant exposure opportunities, and enhance popularity and influence. At the same time, intelligent interaction can meet user needs and improve user conversion and repurchase rates.

[0147] Example 10

[0148] Based on Example 9, for the large language model knowledge base AI Q&A system based on speech recognition, the information preprocessing module includes:

[0149] The first preprocessing unit is used to obtain the question voice issued by the user, establish the speech rhythm of the question voice according to the prosody features of the question voice, divide the question voice into several rhythm segments according to the speech rhythm, and establish the recognition frequency corresponding to each rhythm segment according to the rhythm strength feature corresponding to each rhythm segment;

[0150] The second preprocessing unit is used to perform language recognition on the corresponding rhythm segments based on the recognition frequency and using a preset speech recognition technology to obtain the semantic information corresponding to each rhythm segment, and reorganize the semantic information to obtain the question text of the user;

[0151] The third preprocessing unit is used to perform semantic analysis on the question text to obtain several question purposes of the user, respectively establish a knowledge index corresponding to each question purpose, and use the knowledge index to search for a Q&A knowledge base whose relevance to the corresponding question purpose is greater than a preset relevance threshold.

[0152] In this example, one rhythm segment corresponds to one recognition frequency, and the recognition frequency represents the frequency strength of the rhythm segment;

[0153] In this example, semantic information represents the information presented in a rhythm segment in the form of speech;

[0154] In this example, the knowledge index is used to find the lead-in for the purpose of the question.

[0155] In this example, the preset relevance threshold is 60%.

[0156] The working principle and beneficial effects of the above technical solution are as follows: In order to promptly solve the problems raised by users, the question voice emitted by the user is divided into several rhythm segments according to the speech rhythm, and the recognition frequency corresponding to each rhythm segment is determined. Then, different rhythm segments with different frequencies are recognized to obtain the semantic information corresponding to each rhythm segment. Further, the semantic information is recombined to obtain the user's question text, and the purpose of the user's question is determined. Then, a knowledge index is established for each question purpose, and the answer knowledge base with high relevance is retrieved. In this way, each rhythm segment can be accurately analyzed to determine the semantics therein, improving the effectiveness of semantic analysis and avoiding omission of the answer knowledge base.

[0157] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention also intends to include these modifications and variations.

Claims

1. A large language model knowledge base AI question answering method based on speech recognition, characterized in that: include: Step 1: Initially recognize the question voice sent by the user, determine the purpose of the user's question, and retrieve the relevant answer knowledge base; Step 2: constructing a dialogue model according to the dialogue knowledge base, training the dialogue model according to the purpose of the question, and obtaining a plurality of dialogue texts; Step 3: Retrieve the corresponding virtual answering character according to the questioning scene information of the user, and control the virtual answering character to answer the user; Step 4: Obtain the answer information, conduct self-evaluation on this round of answering based on the answer information, and generate information improvement feedback and knowledge base update feedback based on the self-evaluation results; Step 5: adjusting the model building method according to the information improvement feedback, and adjusting the knowledge base according to the knowledge base update feedback; The step 1 comprises: Step 11: obtaining a question voice issued by the user, establishing a voice rhythm of the question voice according to the rhythmic features of the question voice, dividing the question voice into a plurality of rhythm segments according to the voice rhythm, and establishing a recognition frequency of a corresponding rhythm segment according to the rhythm strength features corresponding to each rhythm segment; Step 12: Based on the recognition frequency and using a preset speech recognition technology, language recognition is performed on the corresponding rhythm segment to obtain semantic information corresponding to each rhythm segment, and the semantic information is reorganized to obtain the user's question text; Step 13: semantically analyze the question text to obtain several question purposes of the user, respectively establish a knowledge index corresponding to each question purpose, and use the knowledge index to search for a knowledge base of answers with a relevance greater than a preset relevance threshold to the corresponding question purpose; The step 2 comprises: Step 21: obtaining a plurality of answer knowledge bases corresponding to the question speech, constructing a knowledge structure network corresponding to the answer knowledge base according to the knowledge points contained in each answer knowledge base, and analyzing the knowledge logic between different answer knowledge bases according to the knowledge structure network; Step 22: constructing a logical docking node for each of the dialogue knowledge bases according to the knowledge logic construction, and logically distributing and arranging the logical bases using the docking points corresponding to the logical docking nodes to generate a dialogue model; Step 23: sorting out the question purpose to obtain a number of relevant knowledge points contained in the question purpose, searching for each of the relevant knowledge points in the dialogue model, and constructing a knowledge structure corresponding to each of the relevant knowledge points according to the search results; Step 24: Establish an answer format according to the purpose of the question, and perform knowledge matching on each of the knowledge structures according to the purpose of the question, and reorganize the matched knowledge according to the answer format to obtain a plurality of answer texts.

2. The AI ​​question-answering method based on a large language model knowledge base for speech recognition according to claim 1, characterized in that: The step 3 comprises: Step 31: Acquire the questioning scenario information of the user, construct a conceptual scenario according to the questioning scenario information, perform dynamic analysis and static analysis on the conceptual scenario, and obtain dynamic scenario features and static scenario features contained in the conceptual scenario; Step 32: construct a concept sequence diagram of the concept scene using the dynamic scene features, construct a static background map of the concept scene using the static scene features, map the concept sequence diagram to the static background map, and obtain a dynamic change rule corresponding to each scene area in the concept scene; Step 33: constructing the overall change law of the concept scene according to the dynamic change law corresponding to each of the scene areas, using the overall change law to retrieve the corresponding character appearance information, respectively constructing the corresponding character activity information according to the scene outside of the concept scene, and constructing the character's corresponding answering volume in each of the scene areas according to the dynamic change law corresponding to each of the scene areas and the character activity information; Step 34: construct a virtual talking character according to the character appearance information, and construct an activity track of the virtual talking character and the talking volume of the virtual talking character during the activity according to the character activity information and the talking volume.

3. The AI ​​question-answering method based on a large language model knowledge base for speech recognition according to claim 2, characterized in that: The step 3 further includes: Step 35: Retrieve the corresponding dialogue atmosphere according to the concept scenario, obtain the dialogue text, add the corresponding atmosphere tone to the dialogue text according to the dialogue atmosphere, and generate the atmosphere dialogue text; Step 36: Control the virtual dialogue character to communicate with the user using the atmosphere dialogue text.

4. The AI ​​question-answering method based on a large language model knowledge base for speech recognition according to claim 1, characterized in that: The step 4 comprises: Step 41: obtaining the dialogue information between the user and the dialogue model, performing question-answer separation on the dialogue information, and obtaining the question information corresponding to the user and the answer information of the dialogue model; Step 42: Establish the question tone change feature of the user according to the question information, construct the user's satisfaction feature for this round of question and answer according to the question tone change feature, establish the user's question according to the question information, and construct the answer question of the question and answer model according to the answer information; Step 43: Analyze the matching degree between the question and the answer, and establish matching features for matching this round of questions and answers according to the matching degree; Step 44: Establish a self-evaluation result of this round of question and answer based on the satisfactory features and the matching features, and construct information improvement feedback and knowledge base update feedback based on the self-evaluation results.

5. The AI ​​question-answering method based on large language model knowledge base of speech recognition according to claim 1, characterized in that: The step 5 comprises: Step 51: determining the model defect features of the answer model according to the information improvement feedback, and determining the knowledge to-be-complemented features of the answer knowledge base according to the knowledge update feedback; Step 52: establishing a defect position of the dialogue model according to the model defect feature, and adjusting the model structure of the dialogue model according to the defect position; Step 53: retrieve the knowledge to be supplemented in the knowledge base update feedback, determine the knowledge association method between the knowledge to be supplemented and the existing knowledge based on the knowledge to be supplemented features, and use the knowledge association method to supplement the knowledge to be supplemented into the corresponding answer knowledge base.

6. The AI ​​question-answering method based on a large language model knowledge base for speech recognition according to claim 5, characterized in that: Also includes: After the model structure adjustment is completed, an updated dialogue model is generated, and the updated dialogue model is used to replace the corresponding dialogue model; The updated dialogue model is used to dialogue with the user.

7. AI question-answering system based on large language model knowledge base for speech recognition, characterized by: include: An information preprocessing module is used to initially recognize the question voice issued by the user, determine the purpose of the user's question, and retrieve the relevant answer knowledge base; A model building and using module, used to build a dialogue model according to the dialogue knowledge base, train the dialogue model according to the purpose of the question, and obtain a plurality of dialogue texts; A virtual character answering module, used to retrieve a corresponding virtual answering character according to the questioning scene information of the user, and control the virtual answering character to answer the user; A self-evaluation analysis module is used to obtain the answer information, conduct self-evaluation on the current round of answers based on the answer information, and generate information improvement feedback and knowledge base update feedback based on the self-evaluation results; An updating and restructuring module, used to improve the feedback adjustment model building method according to the information, and adjust the knowledge base according to the knowledge base update feedback; The information preprocessing module comprises: A first preprocessing unit is used to obtain a question voice issued by a user, establish a voice rhythm of the question voice according to the rhythmic features of the question voice, divide the question voice into a plurality of rhythm segments according to the voice rhythm, and establish a recognition frequency of a corresponding rhythm segment according to a rhythm strength feature corresponding to each rhythm segment; A second preprocessing unit is used to perform language recognition on the corresponding rhythm segment based on the recognition frequency and using a preset speech recognition technology to obtain semantic information corresponding to each rhythm segment, and reorganize the semantic information to obtain the user's question text; A third preprocessing unit is used to perform semantic analysis on the question text to obtain a plurality of question purposes of the user, respectively establish a knowledge index corresponding to each question purpose, and use the knowledge index to search for a knowledge base of answers with a relevance greater than a preset relevance threshold to the corresponding question purpose; The model building uses modules including: A first construction unit is used to obtain a plurality of answer knowledge bases corresponding to the question speech, construct a knowledge structure network corresponding to the answer knowledge base according to the knowledge points contained in each of the answer knowledge bases, and analyze the knowledge logic between different answer knowledge bases according to the knowledge structure network; The second construction unit is used to construct a logical docking node for each of the dialogue knowledge bases according to the knowledge logic construction, and to arrange the logic bases in a logical distribution by using the docking points corresponding to the logical docking nodes to generate a dialogue model; A third construction unit is used to sort out the question purpose to obtain a number of relevant knowledge points contained in the question purpose, search for each of the relevant knowledge points in the dialogue model, and construct a knowledge structure corresponding to each of the relevant knowledge points according to the search result; The fourth construction unit is used to establish a dialogue format according to the purpose of the question, and to perform knowledge matching on each of the knowledge structures according to the purpose of the question, and to reorganize the matched knowledge according to the dialogue format to obtain a plurality of dialogue texts.

Images