Generative model architecture system for reducing loyalty illusion problem and construction method thereof

Abstract

The invention discloses a generative model architecture system for reducing loyalty illusion problems, comprising a data information input unit for inputting modal data information; the information preprocessing unit is used for preprocessing the input modal data information according to target information to obtain data feature information; the model construction unit is used for constructing an initial model according to the preprocessed data feature information; the model training unit is used for training the constructed initial model; the information processing unit is used for processing the data feature information again by utilizing a training model to obtain final data features and outputting the final data features; the output unit is used for reading out the obtained data feature information; the evaluation unit is used for judging the difference between the data information generated by the training model and target information features according to the output data information; through an improved self-attention mechanism, a context awareness mechanism and a source context enhancement technology, the problem of loyalty illusion of a large model during content generation can be effectively solved, and the quality and reliability of the generated content are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a generative model architecture system and a construction method thereof for reducing the fidelity illusion problem. Background Art

[0002] In programming languages, a token usually refers to a separate, meaningful unit in the source code. It can be a keyword, identifier, operator, delimiter, string literal, and so on. In compilers and interpreters, tokens are often used in lexical analysis, syntax analysis, and program parsing. In identity authentication, a token usually refers to a string generated by the server, which is used as a token for the client to make requests. When a user logs in for the first time, the server generates a token and returns it to the client. In the future, the client only needs to bring this token to request data, without having to bring the username and password again. The purpose of this is to reduce the pressure on the server, reduce frequent database queries, and make the server more robust.

[0003] In existing model construction, fidelity illusion is a common problem, and the content generated by the model may be inconsistent with the user's instructions or context. This is usually because the model does not pay enough attention to tokens at different positions when processing long sequences, resulting in inaccurate understanding of the context. To this end, this application uses improved self-attention mechanisms, context-aware mechanisms, and source context enhancement technologies to effectively solve the problem of fidelity illusion when large models generate content, thereby improving the quality and reliability of generated content. Summary of the invention

[0004] The embodiments of the present application provide a generation model architecture system and a construction method for reducing the fidelity illusion problem. Through improved self-attention mechanism, context-aware mechanism and source context enhancement technology, it is possible to effectively solve the fidelity illusion problem of large models when generating content and improve the quality and reliability of generated content.

[0005] The embodiment of the present application provides a generative model architecture system for reducing the fidelity illusion problem, including a data information input unit, which inputs modal data information, and the input modal data information includes text, images, audio, etc.;

[0006] The information preprocessing unit cleans and standardizes the input modal data information according to the target information, and then quantifies, sorts and encodes the information features to obtain the preprocessed data feature information;

[0007] A model building unit builds an initial model based on the preprocessed data feature information;

[0008] The model training unit trains the constructed initial model so that the model training is performed multiple times according to the target information;

[0009] The information processing unit uses the training model to decode and sort the data feature information again, and quantize the data features to obtain data features similar or close to the target information for output;

[0010] An output unit, which uses speech synthesis technology to read out the acquired data feature information;

[0011] The evaluation unit determines the difference between the data information generated by the training model and the target information characteristics based on the output data information.

[0012] Further, it includes an information regulation and allocation unit, which distinguishes the context data sequence of the modal data information and focuses on analyzing and regulating the data feature information to ensure the accuracy of the initial model construction;

[0013] The information monitoring unit monitors the data information features in real time during the model training process and compares them with the target information multiple times so as to correct the errors in the data information features in a timely manner;

[0014] The information adjustment unit modifies the constructed model based on the errors in the monitored data information features combined with the context data sequence.

[0015] Further information preprocessing units include text information processing and language information processing;

[0016] Text information processing analyzes and extracts text information from modal data information, generates text records, extracts text features and converts them into data sequences;

[0017] Language information processing analyzes and extracts the voice information in the modal data information, uses speech recognition technology to obtain the text information in the voice information, extracts the voice features and converts them into data sequences.

[0018] Further information regulation allocation units include information comparison, information regulation;

[0019] Information comparison compares the modal data information with the target information, and comprehensively analyzes the upper and lower data sequences;

[0020] Information regulation is based on a comprehensive analysis of the contextual data sequences, and can identify subtle differences in user input instructions and adjust the model's generation behavior accordingly to generate content that is more consistent with user instructions and context.

[0021] The method for constructing a generative model architecture to reduce the problem of faithfulness illusion includes the following steps:

[0022] (1) Input modal data information, such as text, language, image, etc., into the architecture system;

[0023] (2) Clean and standardize the input data information, distinguish the context data and make adjustments, then quantify and sort each feature to obtain the preprocessed information;

[0024] (3) constructing an initial model using the preprocessed information;

[0025] (4) Training the constructed initial model, and monitoring the data information characteristics in real time during the model training process, so as to correct the errors in the data information characteristics in a timely manner;

[0026] (5) Using the training model to generate each feature, decode and sort it, and quantify it numerically to obtain new information features for output;

[0027] (6) Based on the output data information, determine the difference between the data information generated by the training model and the target information characteristics.

[0028] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: through improved self-attention mechanism, context-aware mechanism and source context enhancement technology, it is possible to effectively solve the problem of fidelity illusion of large models when generating content, and improve the quality and reliability of generated content. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a schematic diagram of the system structure of the generative model architecture used in this application to reduce the fidelity illusion problem;

[0030] Figure 2 Flowchart of the method for constructing a generative model architecture for reducing the fidelity illusion problem in this application. DETAILED DESCRIPTION

[0031] In order to facilitate the understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. The preferred embodiments of the present invention are given in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thoroughly understood.

[0032] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0033] Embodiment 1

[0034] See also Figure 1-2 , a generative model architecture system for reducing the fidelity illusion problem, including a data information input unit, inputting modal data information, wherein the input modal data information includes text, image, audio, etc.;

[0035] The user inputs instruction information into the architecture system in the form of text, image or audio.

[0036] The information preprocessing unit cleans and standardizes the input modal data information according to the target information, and then quantifies, sorts and encodes the information features to obtain the preprocessed data feature information;

[0037] The information preprocessing unit includes text information processing and language information processing;

[0038] Text information processing analyzes and extracts text information from modal data information, generates text records, extracts text features and converts them into data sequences;

[0039] Language information processing analyzes and extracts the voice information in the modal data information, uses speech recognition technology to obtain the text information in the voice information, extracts the voice features and converts them into data sequences.

[0040] A model building unit builds an initial model based on the preprocessed data feature information;

[0041] The decoder and encoder are used to numerically quantify the information features of the modal data, and then sort and encode them to construct text or language sequences.

[0042] The model training unit trains the constructed initial model so that the model training is performed multiple times according to the target information.

[0043] The information processing unit uses the training model to decode and sort the data feature information again, and quantize the data features to obtain data features similar or close to the target information for output;

[0044] The decoder and encoder are used to decode, sort and numerically quantize the modal data information after multiple trainings to obtain modal data information output with similar characteristics to the target data.

[0045] An output unit, which uses speech synthesis technology to read out the acquired data feature information;

[0046] Speech recognition is to obtain the literal text of the speech to analyze the speech features of the target data information, such as pitch features, speaking speed features, intonation features and other speech parameters; and read and output the final processed modal data information.

[0047] The evaluation unit determines the difference between the data information generated by the training model and the target information characteristics based on the output data information.

[0048] The construction system also includes an information regulation and allocation unit, which distinguishes the context data sequence of the modal data information and focuses on analyzing and regulating the data feature information to ensure the accuracy of the initial model construction;

[0049] The information regulation and allocation unit includes information comparison and information regulation;

[0050] Information comparison compares the modal data information with the target information, and comprehensively analyzes the upper and lower data sequences;

[0051] Information regulation is based on a comprehensive analysis of the contextual data sequences, and can identify subtle differences in user input instructions and adjust the model's generation behavior accordingly to generate content that is more consistent with user instructions and context.

[0052] The information monitoring unit monitors the data information features in real time during the model training process and compares them with the target information multiple times so as to correct the errors in the data information features in a timely manner.

[0053] The information adjustment unit modifies the constructed model based on the errors in the monitored data information features combined with the context data sequence.

[0054] The system constructed in this application is operated by a server in combination with a terminal device, and data is transmitted between the server and the terminal device via a wireless network or a wired network;

[0055] The server can be a server cluster composed of multiple computer hosts, which respectively provide database, cloud computing, data processing functions (data processing algorithms), cloud storage, network services, communication input and output, security services, etc.; the server is used to provide pre-training services for the initial model to obtain a target pre-training model; it can also provide text generation training services for the target pre-training model to generate a target text generation model; or it can also receive speech submitted by the user, perform speech processing on it, and obtain the corresponding input text;

[0056] Terminal devices include intelligent physical devices such as desktop computers, smart phones, and tablet computers, which are used to receive input text, voice, etc. submitted by users and transmit them to the server to build the modal model.

[0057] The method for constructing a generative model architecture to reduce the problem of faithfulness illusion includes the following steps:

[0058] (1) Input modal data information, such as text, language, image, etc., into the architecture system;

[0059] (2) Clean and standardize the input data information, distinguish the context data and make adjustments, then quantify and sort each feature to obtain the preprocessed information;

[0060] (3) constructing an initial model using the preprocessed information;

[0061] (4) Training the constructed initial model, and monitoring the data information characteristics in real time during the model training process, so as to correct the errors in the data information characteristics in a timely manner;

[0062] (5) Using the training model to generate each feature, decode and sort it, and quantify it numerically to obtain new information features for output;

[0063] (6) Based on the output data information, determine the difference between the data information generated by the training model and the target information characteristics.

[0064] The technical solutions in the above-mentioned embodiments of the present application have at least the following technical effects or advantages: through improved self-attention mechanism, context-aware mechanism and source context enhancement technology, it can effectively solve the problem of fidelity illusion of large models when generating content, and improve the quality and reliability of generated content.

[0065] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A generative model architecture system for reducing the fidelity illusion problem, characterized by: A data information input unit is included to input modal data information, where the input modal data information includes text, image, audio, etc.; The information preprocessing unit cleans and standardizes the input modal data information according to the target information, and then quantifies, sorts and encodes the information features to obtain the preprocessed data feature information; A model building unit builds an initial model based on the preprocessed data feature information; The model training unit trains the constructed initial model so that the model training is performed multiple times according to the target information; The information processing unit uses the training model to decode and sort the data feature information again, and quantize the data features to obtain data features similar or close to the target information for output; An output unit, which uses speech synthesis technology to read out the acquired data feature information; The evaluation unit determines the difference between the data information generated by the training model and the target information characteristics based on the output data information.

2. The generative model architecture system for reducing the fidelity illusion problem according to claim 1, characterized in that: It also includes an information regulation and allocation unit, which distinguishes the context data sequence of the modal data information and focuses on analyzing and regulating the data feature information to ensure the accuracy of the initial model construction; The information monitoring unit monitors the data information features in real time during the model training process and compares them with the target information multiple times so as to correct the errors in the data information features in a timely manner; The information adjustment unit modifies the constructed model based on the errors in the monitored data information features combined with the context data sequence.

3. The generative model architecture system for reducing the fidelity illusion problem according to claim 1, characterized in that: The information preprocessing unit includes text information processing and language information processing; Text information processing analyzes and extracts text information from modal data information, generates text records, extracts text features and converts them into data sequences; Language information processing analyzes and extracts the voice information in the modal data information, uses speech recognition technology to obtain the text information in the voice information, extracts the voice features and converts them into data sequences.

4. The generative model architecture system for reducing the fidelity illusion problem according to claim 1, characterized in that: The information regulation and allocation unit includes information comparison and information regulation; Information comparison compares the modal data information with the target information, and comprehensively analyzes the upper and lower data sequences; Information regulation is based on a comprehensive analysis of the contextual data sequences, and can identify subtle differences in user input instructions and adjust the model's generation behavior accordingly to generate content that is more consistent with user instructions and context.

5. A method for constructing a generative model architecture for reducing the problem of fidelity illusion, characterized in that: The specific process includes the following: (1) Input modal data information, such as text, language, image, etc., into the architecture system; (2) Clean and standardize the input data information, distinguish the context data and make adjustments, then quantify and sort each feature to obtain the preprocessed information; (3) constructing an initial model using the preprocessed information; (4) Training the constructed initial model, and monitoring the data information characteristics in real time during the model training process, so as to correct the errors in the data information characteristics in a timely manner; (5) Using the training model to generate each feature, decode and sort it, and quantify it numerically to obtain new information features for output; (6) Based on the output data information, determine the difference between the data information generated by the training model and the target information characteristics.

Images