User permission management-oriented large model attribution and right confirmation method, device and equipment

Abstract

The application relates to a large model attribution right confirmation method, device and equipment for user permission management and control. The method comprises the following steps: classifying large model attribution according to an attribution source type, setting specific metrics and general metrics of each type of attribution, and constructing a large model attribution evaluation system from two aspects of the specific metrics and the general metrics of each type of attribution; determining an attribution source by adopting a large model internal knowledge attribution method based on a reward model and feedback according to the large model attribution evaluation system; and determining the most right permission by adopting a fuzzy logic-based attribution source-permission matching method on the attribution source. The method realizes internal knowledge attribution, reduces large model attribution illusion, overcomes the uncertainty problem caused by attribution errors in permission determination, improves the accuracy and reliability of permission determination, improves the efficiency of user permission management and control, and reduces the security risks caused by permission mismatch.

Description

Technical Field

[0001] The present application relates to the field of attribution and rights confirmation technology, and in particular to a large-scale attribution and rights confirmation method, apparatus, and device for user rights management. Background Art

[0002] Permission control management is very important in an enterprise environment. Permission control in an enterprise means controlling users' read and write permissions to system resources, limiting access to key resources, and preventing intrusion by illegal users or loss of confidentiality caused by careless operations by illegal users.

[0003] Large models are trained using self-supervised learning methods on large amounts of unlabeled text. After pre-training and fine-tuning, the model forms a "memory" within the neural network. Because the training text contains different permission levels and types, the model's memory includes relevant information from various training data. This allows users with different permission types to easily obtain information that does not match their permissions by asking questions, creating security risks. Therefore, it is necessary to research large model attribution and ownership confirmation solutions for user permission management, and to implement measures to manage and control the content output by large models based on user permissions.

[0004] Automatic attribution and confirmation of rights for large models refers to the process of automatically providing attribution sources supporting output content through large models and determining the corresponding permissions for the output content based on the attribution sources. This is an important means of large-scale model permission management. Existing large-scale model automatic attribution methods are still in the early stages of research. Although they can provide reference literature or text paragraphs for output content through design instructions and the addition of intermediate modules, problems such as unclear attribution, incorrect attribution, and insufficient attribution still exist, seriously affecting the accuracy and credibility of permission determination. Therefore, there is an urgent need to explore large-scale automatic attribution and confirmation solutions for user permission management. Summary of the Invention

[0005] Based on this, it is necessary to provide a large-model attribution and rights confirmation method, device and equipment for user authority management to address the above technical problems.

[0006] A large-model attribution and rights confirmation device for user rights management, comprising:

[0007] The large-model attribution evaluation system construction module is used to classify large-model attribution according to the attribution source type, set specific metrics and general attribution metrics for each type of attribution, and construct a large-model attribution evaluation system from the two perspectives of specific metrics and general metrics for each type of attribution.

[0008] The attribution source confirmation module is used to determine the attribution source according to the large model attribution evaluation system using the large model intrinsic knowledge attribution method based on the reward model and feedback.

[0009] The attribution module is configured to determine the most right permission by adopting a fuzzy logic-based attribution source-permission matching method for the attribution source.

[0010] A large model attribution right confirmation method for user permission management, the method comprising:

[0011] According to the type of attribution source, the large model attribution is classified, specific metrics and general metrics of each type of attribution are set, and a large model attribution evaluation system is constructed from two aspects of specific metrics and general metrics of each type of attribution.

[0012] According to the large model attribution evaluation system, a large model internal knowledge attribution method based on reward model and feedback is adopted to determine the attribution source.

[0013] The attribution module is configured to determine the most right permission by adopting a fuzzy logic-based attribution source-permission matching method for the attribution source.

[0014] In one embodiment, according to the type of attribution source, the large model attribution is classified, specific metrics and general metrics of each type of attribution are set, and a large model attribution evaluation system is constructed from two aspects of specific metrics and general metrics of each type of attribution, comprising:

[0015] According to the type of attribution source, the large model attribution is classified into internal knowledge attribution, pre-training data attribution, and input question attribution.

[0016] A attribution effect measurement framework is constructed from two aspects of specific metrics and general metrics of each type of attribution.

[0017] According to the attribution effect measurement framework, an attribution evaluation method is designed.

[0018] According to the attribution classification result, the attribution effect measurement framework, and the attribution evaluation method, a large model attribution evaluation system is obtained.

[0019] In one embodiment, a attribution effect measurement framework is constructed from two aspects of specific metrics and general metrics of each type of attribution, comprising:

[0020] For attribution effect measurement, general metrics and specific metrics of attribution are determined, the general metrics of attribution include accuracy, sufficiency, and relevance, the specific metrics of internal knowledge attribution are matching / implication of attribution source and output text, the specific metrics of pre-training data attribution are matching of attribution source and training data, and the specific metrics of input question attribution are contribution of attribution source to output.

[0021] In one embodiment, the attribution evaluation method comprises a human evaluation method, a classification-based evaluation method, and a quantitative evaluation method.

[0022] In one of the embodiments, the attribution source is determined by the large model internal knowledge attribution method based on reward model and feedback according to the large model attribution evaluation system, which includes:

[0023] The design instruction guides the large model to automatically generate reference literature / source text according to the internal knowledge.

[0024] The attribution quality reward model is constructed and trained; and the large model attribution is scored by using the attribution quality reward model.

[0025] In the attribution feedback stage, the score given by the attribution quality reward model is normalized, and the instruction executable by the large model is given according to the average score of the attribution quality reward model on all references.

[0026] In one of the embodiments, the training process of the attribution quality reward model includes:

[0027] The training data set in the form of <question, document, attribution> is designed,

[0028] The poor quality attribution data set is constructed by randomly mixing or deleting part of the references through the initial attribution.

[0029] The initial attribution data set and the poor quality attribution data set are combined to form the reference quality data set.

[0030] The initial attribution quality reward model is trained by using the reference quality data set to obtain the trained attribution quality reward model; and the training loss is:

[0031]

[0032] Wherein, L represents the training loss, R represents the attribution quality reward model, represents the data in the attribution quality data set.

[0033] In one of the embodiments, the attribution source is determined by the attribution source-privilege matching method based on fuzzy logic, and the most privileged is determined, which includes:

[0034] The Internet public data is obtained and labeled to form a <text, privilege> database.

[0035] For each document Y in the database, the attribution document X generated by the large model is matched with the document Y, and the text matching similarity is calculated.

[0036] The attribution document X is matched with the document Y in fuzzy semantic matching, and the fuzzy semantic matching similarity is calculated.

[0037] According to the text matching similarity and the fuzzy semantic matching similarity, the matching texts higher than the preset similarity threshold in the database are selected to form a permission matching library of the attribution document X.

[0038] The permission determination system is constructed based on expert knowledge.

[0039] According to the fuzzy rules set in the fuzzy permission determination system, the final determination of the permission of the attribution document X is realized.

[0040] In one of the embodiments, the attribution document X is matched with the document Y in fuzzy semantics, and the fuzzy semantic matching similarity is calculated, including:

[0041] The attribution document X and the document Y are processed through a word-level convolution layer and a sentence-level convolution layer to obtain the word-level semantic representation of the attribution document X and the document Y and the sentence-level semantic representation of the attribution document X and the document Y.

[0042] The fuzzy features of the word-level semantic representation and the sentence-level semantic representation of the attribution document X and the document Y are extracted through a fuzzy membership function to obtain the fuzzy semantic representation of the attribution document X and the document Y at the word level and the sentence level, and then the dimensions of the fuzzy semantic representation are adjusted through a fully connected neural network,

[0043] The fuzzy semantic representations of the attribution document X at the word level and the sentence level after the dimension adjustment are merged through a fuzzy aggregation operator to obtain the fuzzy representation of the attribution document X.

[0044] The fuzzy semantic representations of the document Y at the word level and the sentence level after the dimension adjustment are merged through a fuzzy aggregation operator to obtain the fuzzy representation of the document Y.

[0045] According to the fuzzy representation of the attribution document X and the fuzzy representation of the document Y, the fuzzy semantic matching similarity is obtained as:

[0046]

[0047] wherein, the fuzzy semantic matching similarity of the two is calculated by using a fuzzy distance, and are the fuzzy semantic representations of the attribution document X and the document Y after the fuzzy aggregation, respectively, and are the membership values of and , respectively.

[0048] A computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0049] The above-mentioned large-model attribution rights confirmation method, device, and equipment for user authority management include: classifying large-model attribution according to attribution source type, setting specific metrics and universal attribution metrics for each type of attribution, and constructing a large-model attribution evaluation system from the two perspectives of specific metrics and universal metrics for each type of attribution; determining the attribution source based on the large-model attribution evaluation system using a large-model intrinsic knowledge attribution method based on a reward model and feedback; and determining the final authority using an attribution source-authority matching method based on fuzzy logic. This method implements intrinsic knowledge attribution, alleviates the large-model attribution illusion, overcomes the uncertainty problem caused by attribution errors in authority determination, improves the accuracy and credibility of authority determination, improves the efficiency of user authority management, and reduces security risks caused by authority mismatches. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A flowchart of a large-model attribution and confirmation method for user rights management in one embodiment;

[0051] Figure 2 A schematic diagram of the technical concept of automatic attribution and confirmation of ownership of a large model in one embodiment;

[0052] Figure 3 A schematic diagram of a process for constructing a large-scale model attribution indicator and evaluation system in one embodiment;

[0053] Figure 4 A schematic diagram of a flow chart of a large-model attribution mode based on a reward model and feedback in another embodiment;

[0054] Figure 5 2 is a flowchart of an attribution source-authority matching method based on fuzzy logic in another embodiment;

[0055] Figure 6 is a flow chart of a fuzzy logic module in a fuzzy semantic matching method in another embodiment;

[0056] Figure 7 This is a structural block diagram of a large-model attribution and rights confirmation device for user rights management in one embodiment;

[0057] Figure 8 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0059] The method aims to address the challenges of large model attribution illusion, attribution error, and permission determination. It studies large model attribution classification, metric indicators, and evaluation methods, and establishes a large model attribution evaluation system. On this basis, it studies large model automatic attribution schemes based on prompts, explores reward model construction and large model attribution feedback methods. Finally, it studies the attribution source-right matching method based on fuzzy logic to realize the association of attribution sources and rights under fuzzy information. The main steps include: (1) Large model attribution evaluation system construction: To address the problems of non-standard large model attribution quality measurement indicators and lack of unified evaluation system, the attribution measurement indicators of three types of attribution, including pre-training data attribution, internal knowledge attribution, and input question attribution, are studied, and a unified large model attribution evaluation framework is constructed to guide the subsequent attribution method research. (2) Large model internal knowledge attribution method based on "prompt-feedback-correction": To address the challenges of attribution error and insufficient attribution, the instruction design method for large model automatic attribution is studied, and on this basis, the construction and training method of the attribution quality reward model is proposed, the feedback mechanism of the reward model to the attribution effect of the large model is designed, and the attribution accuracy is enhanced. (3) Attribution source-right matching and determination technology based on fuzzy logic: After obtaining the attribution source, the matching algorithm of the attribution source and the standard <document, right> library is studied, the fuzziness of natural language is fully considered, and the attribution source-right matching model based on fuzzy logic is constructed. Considering the case of matching multiple attribution sources at the same time, the fuzzy right determination method is explored to realize the determination of the most right.

[0060] In one embodiment, as shown in Figure 1 a large model attribution right determination method for user right management is provided, which includes the following steps:

[0061] Step 100: Classify large model attribution according to attribution source type, set specific metrics and general metrics for each type of attribution, and construct a large model attribution evaluation system from the perspectives of specific metrics and general metrics for each type of attribution.

[0062] Specifically, the large model attribution evaluation system is constructed to guide the large model to automatically generate accurate and sufficient attribution and establish the matching relationship between attribution and right, providing key support for large model right management. The technical idea of large model automatic attribution right determination is shown in Figure 2 .

[0063] To address the problems of lack of unified standards and non-uniform evaluation indicators in large model attribution evaluation, it is necessary to consider various types of large model attribution, establish a large model attribution classification method, set attribution indicators and their measurement methods, and construct a large model attribution indicator and evaluation system.

[0064] According to the type of attribution source, the large model attribution method is divided into three types, as shown in Figure 3The first type is internal knowledge attribution, which refers to the large model using internal knowledge to give attribution sources by itself. The attribution source can be a reference or a reference paragraph. The second type is pre-training data attribution, which is different from the previous attribution method. In this case, the available pre-training corpus is matched with the attribution source generated by the large model to provide ideas for which data sources in the pre-training corpus have a significant contribution to the output of the large model. The third type of attribution is input question attribution, which refers to attributing the output content of the large model to the input question to give the contribution of each word (phrase) in the input question to the output content.

[0065] Step 102: Determine the attribution source using the reward model-based and feedback-based large model internal knowledge attribution method according to the large model attribution evaluation system.

[0066] Specifically, considering that large model internal knowledge attribution is prone to hallucinations, including attribution that is not related to the question and generated content, and the accuracy of attribution itself is insufficient, the large model internal knowledge attribution algorithm is studied, the instruction is designed to effectively guide the large model to repeat the reference literature, and the reward model is constructed to realize the feedback of the large model attribution.

[0067] The reward model-based and feedback-based large model attribution method is as shown in Figure 4 The present application plans to complete the large model internal knowledge attribution method in three steps. First, in the large model instruction design step, the instruction is designed to guide the large model to automatically generate reference literature / source text based on internal knowledge. Then in the reward model construction step, the attribution quality reward model is constructed and trained, so that the model can identify the defects of the large model attribution and provide negative feedback for poor attribution and positive feedback for good attribution; finally, in the attribution feedback stage, the reward model is used to feedback the attribution of the large model to reduce the hallucination of the large model attribution.

[0068] Step 104: Determine the most authoritative authority using the fuzzy logic-based attribution source-authority matching method.

[0069] Specifically, in the attribution source-authority matching method and authority determination technology, considering the fuzziness caused by the uncertainty of the meaning of words in natural language and different contexts, the fuzziness of the attribution source itself, and the fuzziness of the attribution source and the corresponding authority, a fuzzy logic-based attribution source-authority matching method is explored to realize the final determination of the authority.

[0070] The attribution source-authority matching method based on fuzzy logic is to use and improve the existing text matching and fuzzy semantic matching methods to obtain the text matching similarity and semantic matching similarity between the attribution source X and each text in the authority library, form the authority matching library of the attribution source X, and then build a fuzzy system based on expert knowledge to achieve the final authority determination. The flowchart of the attribution source-authority matching method based on fuzzy logic is as follows: Figure 5 shown.

[0071] In a specific application case, the target unit has N personnel, , and the corresponding permissions are , that is, each person has one authority. The unit deployed a large model The model is based on the ChatGLM-6B model and fine-tuned using the data of this unit. Personnel Using large models Ask questions, the question is , the output of the large model is ,in As mentioned above, the output of the large model is May exceed personnel Permissions This method first constructs a large model attribution evaluation system, and based on this, uses the large model internal knowledge attribution method to generate the large model output answer. Attribution source . Then, build the M <text, permissions> database of data, each data in the database contains text And the corresponding permissions , , and by calculating and The text matching similarity and fuzzy semantic matching similarity are obtained. The permission matching library, each data in the library is Matched text And the corresponding permissions Finally, a fuzzy authority determination system is designed based on expert knowledge to obtain The ultimate authority .like and If they are consistent, the correct answer of the large model itself is output , such as: "Based on your question, the answers are as follows: Otherwise, the corrected model output is: "The question you asked is not within the scope of your access rights. Please change the content of your question."

[0072] In the above user-oriented permission management and control large model attribution method, the method comprises: classifying large model attribution according to attribution source types, setting specific metrics and general metrics of each type of attribution, and constructing a large model attribution evaluation system from two aspects of specific metrics and general metrics of each type of attribution; determining the attribution source according to the large model attribution evaluation system using a large model internal knowledge attribution method based on reward model and feedback; and determining the most right permission by using a fuzzy logic-based attribution source-permission matching method on the attribution source. This method realizes internal knowledge attribution, reduces large model attribution illusion, overcomes the uncertainty problem caused by attribution error in permission determination, improves the accuracy and credibility of permission determination, improves user permission management efficiency, and reduces security risks caused by permission mismatch.

[0073] In one of the embodiments, step 100 comprises: dividing the large model attribution into internal knowledge attribution, pre-training data attribution, and input question attribution according to the attribution source types; constructing an attribution effect measurement framework from two aspects of specific metrics and general metrics of each type of attribution; designing an attribution evaluation method according to the attribution effect measurement framework; and obtaining a large model attribution evaluation system according to the attribution classification results, the attribution effect measurement framework, and the attribution evaluation method.

[0074] In one of the embodiments, constructing an attribution effect measurement framework from two aspects of specific metrics and general metrics of each type of attribution comprises: determining attribution general metrics and specific metrics for attribution effect measurement, the general metrics of attribution include: accuracy, sufficiency, and relevance; the specific metrics of internal knowledge attribution are attribution source matching / implication with output text, the specific metrics of pre-training data attribution are attribution source matching with training data, and the specific metrics of input question attribution are attribution source contribution to output.

[0075] Specifically, after reasonably classifying the large model attribution methods, the present application intends to construct an attribution effect measurement framework for each type of attribution method after classification, comprehensively considering each dimension of attribution effect measurement from the perspectives of specific measurement and general measurement of each type of attribution. Firstly, for the general attribution measurement effect, high-quality attribution means that: (1) the attribution is accurate, that is, the attribution source can support the output content; (2) the attribution is sufficient, that is, each necessary part of the output content has a corresponding attribution source; (3) the attribution is relevant, that is, the attribution source needs to be closely related to the question and the answer output by the large model. In addition, in addition to the above general attribution measurement, the three types of attribution methods also need to meet the specific measurement of their respective characteristics: for pre-training data attribution, the attribution source needs to match the data in the pre-training corpus, including text matching (symbol level matching) and semantic matching (semantic level matching); for internal knowledge attribution, the attribution source needs to support the output content of the large model, including text matching (symbol level matching), semantic matching (semantic level matching) and text entailment (according to the input question and the output content of the large model, part of the output content can be logically inferred from the attribution source); for input question attribution, the attribution source needs to provide the contribution of each word (phrase) in the input question to the output content of the large model, including loss contribution and output contribution.

[0076] Finally, the present application intends to design an attribution evaluation method based on the attribution effect measurement framework: (1) artificial evaluation method. Since artificial evaluation can reflect the thinking and judgment of experts, it is the most commonly used method for large model attribution evaluation at present, but it has the defects of being expensive and time-consuming. (2) Classification-based evaluation method, that is, by converting attribution evaluation into other natural language processing tasks (such as natural language reasoning, that is, determining whether the answer output by the large model is supported by the attribution source), thereby providing binary classification evaluation (the attribution source supports the answer, the attribution source does not support the answer) or ternary classification evaluation (the attribution source completely supports the answer, the attribution source partially supports the answer, the attribution source does not support the answer) and the like. (3) Quantitative evaluation method, that is, according to the attribution effect measurement framework, the closeness of attribution to each index is measured from different angles. Specifically, for the general measurement evaluation method: ① attribution accuracy, the alignment degree of the attribution source and the true source is used for evaluation; ② attribution sufficiency, the proportion of output fragments with effective attribution sources in the entire output answer is used for evaluation; ③ attribution relevance, the proportion of attribution sources related to the input question and the output answer among all attribution sources is used for evaluation. For the evaluation method of specific measurement, the present application intends to improve the existing text matching and semantic matching methods by comprehensively considering the fuzziness of natural language and the fuzziness of matching; for the evaluation of contribution related to, the present application intends to evaluate the loss contribution and output contribution through the input disturbance method.

[0077] In one of the embodiments, the attribution evaluation method includes a human evaluation method, a classification-based evaluation method, and a quantitative evaluation method.

[0078] In one of the embodiments, step 102 includes: designing instructions to guide the large model to automatically generate reference literature / source text according to the internal knowledge; constructing and training an attribution quality reward model; scoring the attribution of the large model using the attribution quality reward model; in the attribution feedback stage, normalizing the score given by the attribution quality reward model, and giving instructions executable by the large model according to the average score of the attribution quality reward model on all references.

[0079] In one of the embodiments, the training process of the attribution quality reward model includes: designing a training data set in the form of <question, document, attribution>; constructing a poor quality attribution data set by randomly mixing or deleting part of the references using the initial attribution; combining the initial attribution data set and the poor quality attribution data set to form a reference quality data set; training the initial attribution quality reward model using the reference quality data set to obtain the trained attribution quality reward model; and the training loss is as follows:

[0080] (1)

[0081] wherein, L represents the training loss, R represents the attribution quality reward model, represents the data in the attribution quality data set.

[0082] Specifically, in the large model instruction design step, in order to design instructions to guide the large model to automatically generate reference literature / source text, the present application proposes to generate an initial attribution based on a prompt, and the prompt is designed as follows: “use the form of [index] to cite references, cite the most relevant content to the question and answer, and do not cite irrelevant content, and each sentence in the answer contains at least one reference”.

[0083] Then, the attribution quality reward model (hereinafter referred to as the reward model) is constructed R The attribution of the large model is scored. The present application proposes to design a training data set in the form of <question, document, attribution>, and to construct a poor quality attribution data set by randomly mixing or deleting part of the references using the initial attribution generated by the large model. And combine the initial attribution data set and the poor quality attribution data set to form a reference quality data set. The present application proposes to use GPT-J-6B as the initial reward model, and to train it using the constructed reference quality data set, using one initial attribution and three poor quality attributions in each round. The training target aims to make the reward model give lower scores to poor quality attributions, and the loss is shown in the following formula (1).

[0084] Finally, the trained reward model is used R Feedback and guidance are provided to the initial attribution of the large model to improve the quality of the citations of the large model and reduce the attribution illusion. Specifically, the scores given by the reward model are normalized, and instructions executable by the large model are given according to the average scores of the reward model on all citations, for example, “you have cited appropriate literature, but please provide specific citations for each output sentence”. Thus, the large model is helped to correct its own attribution.

[0085] In one of the embodiments, step 104 includes: obtaining Internet public data and labeling to form a <text, authority> database; for each document Y in the database, performing text matching between the attribution document X generated by the large model and the document Y to calculate a text matching similarity; performing fuzzy semantic matching between the attribution document X and the document Y to calculate a fuzzy semantic matching similarity; selecting matching texts higher than a preset similarity threshold from the database according to the text matching similarity and the fuzzy semantic matching similarity to form an authority matching library of the attribution document X; constructing an authority judgment system based on expert knowledge; and realizing final judgment of the authority of the attribution document X according to fuzzy rules set in the fuzzy authority judgment system.

[0086] In one of the embodiments, the fuzzy semantic matching between the attribution document X and the document Y is performed to calculate the fuzzy semantic matching similarity, which includes: processing the attribution document X and the document Y through a word-level convolution layer and a sentence-level convolution layer to obtain word-level semantic representations of the attribution document X and the document Y and sentence-level semantic representations of the attribution document X and the document Y; extracting fuzzy features of the word-level semantic representations and the sentence-level semantic representations of the attribution document X and the document Y through a fuzzy membership function to obtain fuzzy semantic representations of the word-level and the sentence-level of the attribution document X and the document Y, and then adjusting dimensions of the fuzzy semantic representations through a full connection neural network; merging the fuzzy semantic representations of the word-level and the sentence-level of the attribution document X after the dimension adjustment through a fuzzy aggregation operator to obtain a fuzzy representation of the attribution document X; merging the fuzzy semantic representations of the word-level and the sentence-level of the document Y after the dimension adjustment through the fuzzy aggregation operator to obtain a fuzzy representation of the document Y; and obtaining the fuzzy semantic matching similarity according to the fuzzy representation of the attribution document X and the fuzzy representation of the document Y as follows:

[0087] (2)

[0088] wherein, the fuzzy semantic matching similarity between the attribution document X and the document Y is calculated by using a fuzzy distance, and are fuzzy semantic representations of the attribution document X and the document Y after the fuzzy aggregation, respectively, and are membership values of and , respectively.

[0089] Specifically, the fuzzy logic-based attribution source-privilege matching method is divided into the following steps:

[0090] (1) By scraping Internet public data, combining existing label system for annotation, using metadata, GPT-4 machine annotation and expert annotation, the scale reaches 10K+, forming a <text, privilege> database.

[0091] (2) For each document Y, the fuzzy semantic matching model is constructed by considering the fuzzy information, and the fuzzy semantic matching of the attribution source X and Y is performed, and the matching similarity is calculated . The fuzzy semantic matching is realized by introducing a fuzzy logic module into the semantic matching convolutional neural network. The process of the fuzzy logic module in the fuzzy semantic matching method is shown in . Figure 6

[0092] This module first receives the semantic representation of the word-level and sentence-level convolutional layer output, extracts the fuzzy features of the word-level and sentence-level representation through the membership function, measures their uncertainty and fuzziness, and converts the semantic representation into fuzzy semantic representation. Then, through the fully connected neural network, the fuzzy semantic representations of different levels are made to have the same dimension, so as to facilitate the subsequent fuzzy aggregation operation. Finally, the fuzzy aggregation operator is used to combine the fuzzy semantic representations of the word-level and sentence-level into the final representation. Let the two be and , respectively, and the fuzzy aggregation operation is as follows:

[0093] (3)

[0094] wherein is a trainable parameter initialized to a random value in the interval [0, 1], represents the fuzzy multiplication operation:

[0095] (4)

[0096] After obtaining the fuzzy representation, the matching similarity of the documents X and Y is calculated by the fuzzy similarity calculation method. Let and be the fuzzy semantic representations of the documents X and Y after fuzzy aggregation, respectively, and the fuzzy distance is used to calculate the semantic matching similarity between the two, and the semantic matching similarity is shown in equation (2).

[0097] (3) According to the text matching similarity​ and semantic similarity degree If the similarity degree is higher than a similarity threshold, the matched text is selected to form a permission matching library of the attribution source X.

[0098] (4) Considering the diversity of the matching results, a fuzzy permission judgment system is constructed based on expert knowledge to realize the judgment of the most appropriate permission. For example, a rule in the fuzzy permission judgment system is as follows:

[0099] If the text similarity degree of X and Y is high, the semantic similarity degree is moderate, and the permission level of Y is high, then the degree of the same permission of X and Y is high.

[0100] According to the fuzzy rules set by experts in the fuzzy permission judgment system, the final judgment of the permission of the document X can be realized.

[0101] It should be understood that, although Figure 1 the steps in the flowchart of the method are displayed in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other orders. Moreover, Figure 1 at least part of the steps in may include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or sub-steps or stages of other steps.

[0102] Figure 7 In one embodiment, as shown in , a large model attribution right confirmation device for user permission management and control is provided, comprising: a large model attribution evaluation system construction module, an attribution source confirmation module, and a right confirmation module, wherein:

[0103] The large model attribution evaluation system construction module is used to classify the large model attribution according to the attribution source type, set specific metrics and general metrics of each type of attribution, and construct a large model attribution evaluation system from two aspects of specific metrics and general metrics of each type of attribution.

[0104] The attribution source confirmation module is used to determine the attribution source by using a large model internal knowledge attribution method based on a reward model and feedback according to the large model attribution evaluation system.

[0105] The right confirmation module is used to determine the most appropriate permission by using a fuzzy logic-based attribution source-permission matching method for the attribution source.

[0106] In one of the embodiments, the large model attribution evaluation system construction module is further configured to divide the large model attribution into internal knowledge attribution, pre-training data attribution, and input question attribution according to attribution source types; construct an attribution effect measurement framework from two perspectives of specific measurement and general measurement of each type of attribution; design an attribution evaluation method according to the attribution effect measurement framework; and obtain the large model attribution evaluation system according to the attribution classification result, the attribution effect measurement framework, and the attribution evaluation method.

[0107] In one of the embodiments, the large model attribution evaluation system construction module is further configured to determine general measurement and specific measurement of attribution effect, wherein the general measurement of attribution includes accuracy, sufficiency, and relevance; the specific measurement of internal knowledge attribution is that the attribution source matches / implies the output text; the specific measurement of pre-training data attribution is that the attribution source matches the training data; and the specific measurement of input question attribution is that the attribution source contributes to the output.

[0108] In one of the embodiments, the attribution evaluation method in the large model attribution evaluation system construction module includes a human evaluation method, a classification-based evaluation method, and a quantitative evaluation method.

[0109] In one of the embodiments, the attribution source confirmation module is further configured to design an instruction to guide the large model to automatically generate reference texts / source texts according to internal knowledge; construct and train an attribution quality reward model; score the large model attribution by using the attribution quality reward model; normalize the score given by the attribution quality reward model in the attribution feedback stage, and give an executable instruction to the large model according to the average score of the attribution quality reward model on all references.

[0110] In one of the embodiments, the training process of the attribution quality reward model in the attribution source confirmation module includes designing a training data set in the form of <question, document, attribution>; constructing a poor quality attribution data set by randomly mixing or deleting part of the reference texts of the initial attribution; combining the initial attribution data set and the poor quality attribution data set to form a reference quality data set; training the initial attribution quality reward model by using the reference quality data set to obtain the trained attribution quality reward model; and the training loss is shown in formula (1).

[0111] In one embodiment, the rights confirmation module is further configured to obtain and annotate public data on the Internet to form a <text, rights> database; for each document Y in the database, text matching is performed between the attributed document X generated by the large model and document Y to calculate text matching similarity; fuzzy semantic matching is performed between the attributed document X and document Y to calculate fuzzy semantic matching similarity; based on the text matching similarity and the fuzzy semantic matching similarity, matching texts with a similarity higher than a preset threshold are selected from the database to form a rights matching library for the attributed document X; a rights determination system is constructed based on expert knowledge; and a final determination of the rights of the attributed document X is made based on fuzzy rules set in the fuzzy rights determination system.

[0112] In one embodiment, the right confirmation module is further configured to process the attributed document X and document Y through a word-level convolutional layer and a sentence-level convolutional layer to obtain word-level semantic representations of the attributed document X and document Y and sentence-level semantic representations of the attributed document X and document Y; extract fuzzy features of the word-level semantic representations and sentence-level semantic representations of the attributed document X and document Y through a fuzzy membership function to obtain word-level and sentence-level fuzzy semantic representations of the attributed document X and document Y, and then adjust the dimensions of the fuzzy semantic representations through a fully connected neural network; merge the word-level and sentence-level fuzzy language representations of the attributed document X after dimension adjustment through a fuzzy aggregation operator to obtain a fuzzy representation of the attributed document X; merge the word-level and sentence-level fuzzy language representations of the document Y after dimension adjustment through a fuzzy aggregation operator to obtain a fuzzy representation of document Y; and obtain the fuzzy semantic matching similarity as shown in formula (2) based on the fuzzy representation of the attributed document X and the fuzzy representation of document Y.

[0113] For the specific limitations of the large-model attribution confirmation device for user rights control, please refer to the limitations of the large-model attribution confirmation method for user rights control above, which will not be repeated here. The various modules in the above-mentioned large-model attribution confirmation device for user rights control can be implemented in whole or in part through software, hardware and their combination. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0114] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 8As shown in the figure. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium, an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The network interface of the computer device is used to communicate with the external terminal through the network connection. The computer program is executed by the processor to realize a large model attribution right management method for user authority control. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0115] Those skilled in the art can understand that, Figure 8 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0116] In one embodiment, a computer device is provided, including a memory and a processor, the memory stores a computer program, and the processor executes the computer program to realize the steps in the above method embodiments.

[0117] The technical features of the above embodiments can be combined in any way. In order to make the description simple, not all possible combinations of the technical features in the above embodiments are described, but as long as the combination of the technical features does not exist contradictory, it should be artificially within the scope of the present application.

[0118] The above-described embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the present application. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of protection of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A large model attribution and right protection device oriented to user permission management, characterized in that, The device comprises: The large model attribution evaluation system construction module is configured to classify large model attribution according to attribution source types, set specific metrics and general metrics of each type of attribution, and construct a large model attribution evaluation system from two aspects of specific metrics and general metrics of each type of attribution; specifically, the large model attribution is divided into internal knowledge attribution, pre-training data attribution, and input question attribution according to the attribution source types; an attribution effect measurement framework is constructed from two aspects of specific metrics and general metrics of each type of attribution; an attribution evaluation method is designed according to the attribution effect measurement framework; a large model attribution evaluation system is obtained according to the attribution classification results, the attribution effect measurement framework, and the attribution evaluation method; the attribution effect measurement framework includes general metrics of attribution and specific metrics of internal knowledge attribution; the general metrics of attribution include accuracy, sufficiency, and relevance; the specific metrics of internal knowledge attribution include specific metrics of internal knowledge attribution, pre-training data attribution, and input question attribution, wherein the specific metrics of internal knowledge attribution are that the attribution source matches or implies the output text, the specific metrics of pre-training data attribution are that the attribution source matches the training data, and the specific metrics of input question attribution are that the attribution source contributes to the output; The attribution source confirmation module is configured to determine the attribution source by using a large model internal knowledge attribution method based on a reward model and feedback according to the large model attribution evaluation system; specifically, an instruction is designed to guide the large model to automatically generate reference literature or source text according to internal knowledge; an attribution quality reward model is constructed and trained; the attribution quality reward model is used to score the large model attribution; in the attribution feedback stage, the score given by the attribution quality reward model is normalized, and an executable instruction of the large model is given according to the average score of the attribution quality reward model on all citations; The right confirmation module is configured to determine the most right authority by using a fuzzy logic-based attribution source-authority matching method for the attribution source.

2. A large model attribution right method for user permission management, characterized in that, The method comprises: According to the attribution source type, the large model attribution is classified, specific metrics and general metrics of each type of attribution are set, and a large model attribution evaluation system is constructed from the two aspects of specific metrics and general metrics of each type of attribution; Specifically, according to the attribution source type, the large model attribution is divided into internal knowledge attribution, pre-training data attribution and input question attribution; A attribution effect measurement framework is constructed from the two aspects of specific metrics and general metrics of each type of attribution; An attribution evaluation method is designed according to the attribution effect measurement framework; A large model attribution evaluation system is obtained according to the attribution classification result, the attribution effect measurement framework and the attribution evaluation method; Wherein the attribution effect measurement framework includes general metrics of attribution and specific metrics of internal knowledge attribution; Wherein the general metrics of attribution include accuracy, sufficiency and correlation; The specific metrics of internal knowledge attribution include the specific metrics of internal knowledge attribution, pre-training data attribution and input question attribution, wherein the specific metrics of internal knowledge attribution are that the attribution source matches or implies the output text, the specific metrics of pre-training data attribution are that the attribution source matches the training data, and the specific metrics of input question attribution are that the attribution source contributes to the output; According to the large model attribution evaluation system, a large model internal knowledge attribution method based on reward model and feedback is used to determine the attribution source; Specifically, an instruction is designed to guide the large model to automatically generate a reference document or a source text according to internal knowledge; A attribution quality reward model is constructed and trained; The attribution quality reward model is used to score the large model attribution; In the attribution feedback stage, the score given by the attribution quality reward model is normalized, and an executable instruction for the large model is given according to the average score of the attribution quality reward model on all citations; The attribution source is matched with the authority based on fuzzy logic to determine the most authority.

3. The user permission-oriented large model attribution method according to claim 2, characterized in that, The attribution evaluation method includes a human evaluation method, a classification-based evaluation method, and a quantitative evaluation method.

4. The user permission-oriented large model attribution method according to claim 2, characterized in that, The training process of the attribution quality reward model includes: Designing a training data set in the form of <question, document, attribution>; Using the initial attribution to construct a poor quality attribution data set by randomly mixing or deleting part of the reference documents; Combining the initial attribution data set and the poor quality attribution data set to form a citation quality data set; The citation quality data set is used to train the initial attribution quality reward model to obtain a trained attribution quality reward model; The training loss is: wherein, L denotes a training loss, R denotes an attribution quality reward model, denotes data in an attribution quality dataset.

5. The user rights-oriented large model attribution method of claim 2, wherein, The attribution source is matched with the authority based on fuzzy logic to determine the most authority, including: Obtain Internet public data and label them to form a <text, authority> database; For each document Y in the database, the attribution document X generated by the large model is matched with the document Y to calculate the text matching similarity; The attribution document X and the document Y are matched with the fuzzy semantic matching to calculate the fuzzy semantic matching similarity; According to the text matching similarity and the fuzzy semantic matching similarity, select the matching texts higher than the preset similarity threshold in the database to form the authority matching library of the attribution document X; Construct an authority judgment system based on expert knowledge; According to the fuzzy rule set in the fuzzy permission judgment system, the final judgment of the attribution document X permission is realized.

6. The user rights-oriented large model attribution method of claim 5, wherein, The attribution document X is matched with the document Y in fuzzy semantic matching, and the fuzzy semantic matching similarity is calculated, including: The attribution document X and the document Y are processed through the word-level convolution layer and the sentence-level convolution layer to obtain the word-level semantic representation of the attribution document X and the document Y and the sentence-level semantic representation of the attribution document X and the document Y; The fuzzy features of the word-level semantic representation and the sentence-level semantic representation of the attribution document X and the document Y are extracted through the fuzzy membership function to obtain the fuzzy semantic representation of the attribution document X and the document Y at the word level and the sentence level, and then the dimensions of the fuzzy semantic representation are adjusted through the full connection neural network, The fuzzy language representations of the attribution document X at the word level and the sentence level after the dimension adjustment are merged through the fuzzy aggregation operator to obtain the fuzzy representation of the attribution document X; The fuzzy language representations of the document Y at the word level and the sentence level after the dimension adjustment are merged through the fuzzy aggregation operator to obtain the fuzzy representation of the document Y; According to the fuzzy representation of the attribution document X and the fuzzy representation of the document Y, the fuzzy semantic matching similarity is obtained. 7.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-6 when the computer program is executed by the processor. The processor executes the computer program to realize the large model attribution right method for user permission management and control in any one of claims 2 to 6.

Images