Large model attribution right confirmation method, device and equipment oriented to user right management and control

Abstract

The invention relates to a large model attribution right confirmation method, device and equipment oriented to user right control. The method comprises the following steps: classifying large model attributions according to attribution source types, setting specific metrics and attribution general metrics of all types of attributions, and constructing a large model attribution evaluation system from two aspects of the specific metrics and the general metrics of all types of attributions; according to the large model attribution evaluation system, determining an attribution source by adopting a large model internal knowledge attribution method based on a reward model and feedback; and determining a final permission for the attribution source by adopting an attribution source-permission matching method based on fuzzy logic. According to the method, internal knowledge attribution is realized, large model attribution illusion is relieved, the problem of uncertainty of permission determination caused by attribution errors is solved, the accuracy and credibility of permission determination are improved, the user permission control efficiency is improved, and potential safety hazards caused by permission mismatching are reduced.

Description

Technical Field

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

[0002] Permission control management is very important in the enterprise environment. Permission control in the 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] The large model uses a self-supervised learning method to train with a large amount of unlabeled text. After pre-training and fine-tuning, the model will form a "memory" in the neural network. Since the training text contains different permission levels and types, the model's memory includes relevant information from various training data, allowing users with different permission types to easily obtain information that does not match their permissions by asking questions, thereby causing security risks. Therefore, it is necessary to study the large model attribution and confirmation scheme for user permission control, and take measures to manage and control the content output by the large model according to the user's permissions.

[0004] Automatic attribution and confirmation of rights for large models refers to automatically providing the attribution source supporting the output content through the large model and determining the corresponding permissions of the output content based on the attribution source. It is an important means of large model permission control. The existing large model automatic attribution method is still in the early stages of research. Although it can provide reference literature or text paragraphs for output content by designing instructions and adding intermediate modules, problems such as unclear attribution, incorrect attribution, and insufficient attribution still exist, which seriously affects the accuracy and credibility of permission determination. Therefore, it is urgent to explore large model automatic attribution and confirmation schemes for user permission control. 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 in response to the above-mentioned technical problems.

[0006] A large-model attribution and confirmation device for user rights management, the device comprising: 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 attribution metrics for each type of attribution, and construct the large model attribution evaluation system from the two perspectives of specific metrics and general metrics for each type of attribution.

[0007] The attribution source confirmation module is used to determine the attribution source according to the big model attribution evaluation system using the big model intrinsic knowledge attribution method based on the reward model and feedback.

[0008] The rights confirmation module is used to determine the final authority by using the attribution source-authority matching method based on fuzzy logic.

[0009] A large-model attribution and confirmation method for user rights management and control, the method comprising: The big model attribution is classified according to the attribution source type, and specific metrics and general attribution metrics are set for each type of attribution. The big model attribution evaluation system is constructed from the two perspectives of specific metrics and general metrics for each type of attribution.

[0010] According to the big model attribution evaluation system, the attribution source is determined by using the big model intrinsic knowledge attribution method based on reward model and feedback.

[0011] The attribution source-authority matching method based on fuzzy logic is used to determine the final authority.

[0012] In one embodiment, the large model attribution is classified according to the attribution source type, and specific metrics and general attribution metrics of each type of attribution are set. A large model attribution evaluation system is constructed from two perspectives: the specific metrics and general metrics of each type of attribution, including: According to the attribution source type, large model attribution is divided into intrinsic knowledge attribution, pre-trained data attribution, and input question attribution.

[0013] The attribution effect measurement framework is constructed from two perspectives: various attribution-specific metrics and general metrics.

[0014] Design an attribution evaluation method based on the attribution effect measurement framework.

[0015] According to the attribution classification results, attribution effect measurement framework and attribution evaluation method, a large-model attribution evaluation system is obtained.

[0016] In one embodiment, an attribution effect measurement framework is constructed from two perspectives: various attribution specific metrics and general metrics, including: For the attribution measurement effect, the general and specific attribution metrics are determined. The general metrics of attribution include accuracy, sufficiency, and relevance. The specific metric for intrinsic knowledge attribution is that the attribution source matches / implies the output text, the specific metric for pre-trained data attribution is that the attribution source matches the training data, and the specific metric for input question attribution is that the attribution source contributes to the output.

[0017] In one embodiment, the attribution evaluation method includes: a human evaluation method, a classification-based evaluation method, and a quantitative evaluation method.

[0018] In one embodiment, the attribution source is determined according to the big model attribution evaluation system using a big model intrinsic knowledge attribution method based on a reward model and feedback, including: Design instructions guide the large model to automatically generate references / source texts based on inherent knowledge.

[0019] Build and train an attribution quality reward model; use the attribution quality reward model to score large model attribution.

[0020] In the attribution feedback phase, the scores given by the attribution quality reward model are normalized, and executable instructions for the large model are given based on the average score of the attribution quality reward model over all citations.

[0021] In one embodiment, the training process of the attributed quality reward model includes: Design a training dataset in the form of <question, document, attribution>, Initial attribution is used to construct a poor-quality attribution dataset by randomly mixing or deleting some references.

[0022] The initial attribution dataset and the poor-quality attribution dataset are combined to form a citation-quality dataset.

[0023] The initial attribution quality reward model is trained using the citation quality dataset to obtain a trained attribution quality reward model; the training loss is:

[0024] in, L represents the training loss, R represents the attributed quality reward model, Represents data from the Attribution Quality dataset.

[0025] In one embodiment, a fuzzy logic-based attribution source-authority matching method is used for the attribution source to determine the final authority, including: Obtain and annotate public data from the Internet to form a <text, permissions> database.

[0026] For each document Y in the database, the attributed document X generated by the large model is text-matched with document Y to calculate the text matching similarity.

[0027] Perform fuzzy semantic matching on the attributed document X and document Y, and calculate the fuzzy semantic matching similarity.

[0028] According to 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 permission matching library of the attributed document X.

[0029] Construct an authority determination system based on expert knowledge.

[0030] According to the fuzzy rules set in the fuzzy permission determination system, the final determination of the permission of the attributed document X is achieved.

[0031] In one embodiment, fuzzy semantic matching is performed on the attributed document X and the document Y, and the fuzzy semantic matching similarity is calculated, including: The attributed document X and document Y are processed 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.

[0032] The fuzzy features of the lexical-level semantic representation and sentence-level semantic representation of the attributed document X and document Y are extracted through the fuzzy membership function, and the lexical-level and sentence-level fuzzy semantic representation of the attributed document X and document Y are obtained. Then, the dimension of the fuzzy semantic representation is adjusted through a fully connected neural network. The fuzzy representation of the attributed document X is obtained by merging the lexical-level and sentence-level fuzzy linguistic representations of the dimensionally adjusted attributed document X through the fuzzy aggregation operator.

[0033] The fuzzy representation of document Y is obtained by merging the lexical-level and sentence-level fuzzy linguistic representations of the dimensionally adjusted document Y through the fuzzy aggregation operator.

[0034] According to the fuzzy representation of the attributed document X and the fuzzy representation of the document Y, the fuzzy semantic matching similarity is obtained as:

[0035] in, In order to use fuzzy distance to calculate the fuzzy semantic matching similarity between the two, and are the fuzzy semantic representations of document X and document Y after fuzzy aggregation, and They are and The membership value of .

[0036] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0037] The above-mentioned large-model attribution and confirmation method, device and equipment for user authority management include: classifying large-model attribution according to the attribution source type, setting specific metrics and general attribution metrics for each type of attribution, and constructing a large-model attribution evaluation system from the two perspectives of specific metrics and general metrics for each type of attribution; determining the attribution source based on the large-model attribution evaluation system using the large-model intrinsic knowledge attribution method based on the reward model and feedback; and determining the final authority using the attribution source-authority matching method based on fuzzy logic. This method realizes intrinsic knowledge attribution, reduces the illusion of large-model attribution, 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 the security risks caused by authority mismatch. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A flowchart of a large-model attribution and confirmation method for user rights management in one embodiment; Figure 2 A schematic diagram of the technical idea of ​​automatic attribution and confirmation of ownership of a large model in an embodiment; Figure 3 A schematic diagram of a large model attribution indicator and evaluation system construction process in an embodiment; Figure 4 It is a flowchart of a large model attribution mode based on a reward model and feedback in another embodiment; Figure 5 A flowchart of an attribution source-authority matching method based on fuzzy logic in another embodiment; Figure 6 is a flow chart of a fuzzy logic module in a fuzzy semantic matching method in another embodiment; Figure 7 It is a structural block diagram of a large-model attribution and confirmation device for user rights management in one embodiment; Figure 8 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0040] This method aims to address the challenges of big model attribution illusion and attribution errors in authority determination, and studies the classification, measurement indicators and evaluation methods of big model attribution, and establishes a big model attribution evaluation system. On this basis, the prompt-based big model automatic attribution scheme is studied, and the reward model construction and big model attribution feedback methods are explored. Finally, the attribution source-authority matching method based on fuzzy logic is studied to achieve the correlation and correspondence between attribution sources and permissions under fuzzy information. It mainly includes the following three method steps: (1) Construction of a big model attribution evaluation system. In order to address the problems of non-standard attribution quality measurement indicators and lack of a unified evaluation system for big model attribution, the attribution measurement indicators of three attribution types, including pre-trained data attribution, intrinsic knowledge attribution and input question attribution, are studied to build a unified big model attribution evaluation framework to guide the subsequent research on attribution methods. (2) Based on the "prompt-feedback-correction" method of attributing the intrinsic knowledge of large models, in order to address challenges such as attribution errors and insufficient attribution, we study the instruction design method for automatic attribution of large models. On this basis, we propose the construction and training method of the attribution quality reward model, and design a feedback mechanism for the reward model on the attribution effect of the large model to enhance the attribution accuracy. (3) Based on the attribution source-authority matching and judgment technology of fuzzy logic, after obtaining the attribution source, we study the matching algorithm between the attribution source and the standard <document, authority> library, fully consider the fuzziness of natural language, and construct an attribution source-authority matching model that introduces fuzzy logic. Considering the situation of matching multiple attribution sources at the same time, we explore the fuzzy authority judgment method to achieve the final authority determination.

[0041] In one embodiment, Figure 1 As shown, a large-model attribution and confirmation method for user rights management is provided, and the method includes the following steps: Step 100: Classify the big model attribution according to the attribution source type, set specific metrics for each type of attribution and general attribution metrics, and build a big model attribution evaluation system from the two perspectives of specific metrics and general metrics for each type of attribution.

[0042] Specifically, we build a large model attribution evaluation system to guide the large model to automatically generate accurate and sufficient attribution and establish a matching relationship between attribution and authority, providing key support for large model authority management. Figure 2 shown.

[0043] In view of the lack of unified standards and inconsistent evaluation indicators in the attribution evaluation of large models, it is necessary to comprehensively consider the various attribution types of large models, establish a large model attribution classification method, set attribution indicators and their measurement methods, and construct a large model attribution indicator and evaluation system.

[0044] The large model attribution methods are divided into three categories according to the attribution source type, such as Figure 3As shown. The first category is intrinsic knowledge attribution, which means that the large model uses internal knowledge to give the attribution source by itself. The attribution source may be a reference document or a reference paragraph. The second category is pre-training data attribution. The difference from the previous attribution method is that here, for the case where pre-training corpus is available, it refers to matching the attribution source generated by the large model with the pre-training corpus, providing ideas for which data sources in the pre-training corpus have significant contributions to the output of the large model. The third category of attribution is input question attribution, which refers to attributing the output content of the large model to the input question, and giving the contribution of each word (phrase) in the input question to the output content.

[0045] Step 102: Determine the attribution source according to the big model attribution evaluation system using the big model intrinsic knowledge attribution method based on the reward model and feedback.

[0046] Specifically, considering that the intrinsic knowledge attribution of large models is prone to hallucinations, including the attribution being irrelevant to the questions asked and the generated content, as well as the lack of accuracy of the attribution itself, we study the intrinsic knowledge attribution algorithm for large models, design instructions to effectively guide the large model to repeat the references, and construct a reward model to provide feedback on the attribution of the large model.

[0047] Large model attribution methods based on reward models and feedback such as Figure 4 As shown, the present application intends to complete the large model intrinsic knowledge attribution method through three steps. First, in the large model instruction design step, the design instructions guide the large model to automatically generate references / source texts based on intrinsic 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 provide feedback on the attribution of the large model to reduce the large model attribution illusion.

[0048] Step 104: A fuzzy logic-based attribution source-authority matching method is used for the attribution source to determine the final authority.

[0049] Specifically, in the attribution source-authority matching method and authority determination technology, considering the uncertainty of the meaning of words themselves and the ambiguity brought by different contexts in natural language, the ambiguity of the attribution source itself, the ambiguity of the correspondence between the attribution source and the authority, etc., explore the attribution source-authority matching method based on fuzzy logic to achieve the final determination of authority.

[0050] The attribution source-authority matching method based on fuzzy logic means: using and improving 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 permission library, forming a permission matching library for the attribution source X, and then building 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.

[0051] In a specific application case, the target unit has N Personnel, 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 unit data. Set the authority to Personnel Using large models Ask a question, 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 , thus causing safety hazards. This method first constructs a large model attribution evaluation system, and based on this, uses the large model intrinsic 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 piece of data in the library is Matching text And the corresponding permissions Finally, a fuzzy authority determination system is designed based on expert knowledge, and we get The final authority .like and If they are consistent, the answer of the large model itself is correctly output , such as: "Based on your question, the answers are as follows: Otherwise, the output of the modified model is: "The question you asked is not within the scope of your access rights. Please change the content of your question." In the above-mentioned large-model attribution and confirmation method for user authority control, the method includes: classifying the large-model attribution according to the attribution source type, setting specific metrics and general attribution metrics for each type of attribution, and constructing a large-model attribution evaluation system from the two perspectives of specific metrics and general metrics for each type of attribution; determining the attribution source based on the large-model attribution evaluation system using the large-model intrinsic knowledge attribution method based on the reward model and feedback; and determining the final authority using the attribution source-authority matching method based on fuzzy logic. This method realizes intrinsic knowledge attribution, reduces the illusion of large-model attribution, 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 control, and reduces the security risks caused by authority mismatch.

[0052] In one embodiment, step 100 includes: dividing the large model attribution into intrinsic knowledge attribution, pre-trained data attribution and input question attribution according to the attribution source type; constructing an attribution effect measurement framework from two perspectives: specific metrics and general metrics of each type of attribution; designing an attribution evaluation method based on the attribution effect measurement framework; and obtaining a large model attribution evaluation system based on the attribution classification results, the attribution effect measurement framework and the attribution evaluation method.

[0053] In one embodiment, an attribution effect measurement framework is constructed from two perspectives: various attribution specific metrics and universal metrics, including: for attribution measurement effect, universal and specific metrics of attribution are determined, and the universal metrics of attribution include: accuracy, sufficiency, and relevance; the specific metric for intrinsic knowledge attribution is that the attribution source matches / implies the output text, the specific metric for pre-trained data attribution is that the attribution source matches the training data, and the specific metric for input question attribution is that the attribution source contributes to the output.

[0054] Specifically, after rationally classifying the large model attribution methods, the present invention intends to comprehensively consider various dimensions of attribution effect measurement for each type of attribution method after classification, and construct an attribution effect measurement framework from two perspectives: specific attribution measurement and general measurement. First, for general attribution measurement effects, high-quality attribution means: (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 asked, and also needs to be closely related to the answer output by the large model. In addition to the above-mentioned general attribution metrics, the three types of attribution methods also require specific metrics that meet 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 intrinsic knowledge attribution, the attribution source needs to support and verify the output content of the large model, including text matching (symbol-level matching), semantic matching (semantic-level matching) and text implications between the attribution source and the model output (based on 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 the contribution at the loss level and the contribution at the output level.

[0055] Finally, the present invention intends to design an attribution evaluation method based on the attribution effect measurement framework: (1) Human evaluation method. Since human evaluation can reflect the ideas and judgments of experts, it is currently the most commonly used method for large-scale model attribution evaluation, but human evaluation has the disadvantages of being expensive and time-consuming. (2) Classification-based evaluation method, that is, by converting the 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 a two-category evaluation (the attribution source supports the answer, the attribution source does not support the answer) or a three-category evaluation (the attribution source fully supports the answer, the attribution source partially supports the answer, the attribution source does not support the answer), etc. (3) Quantitative evaluation method, that is, based on the attribution effect measurement framework, measure the degree of proximity between attribution and various indicators from different angles. Specifically, for the evaluation method of general measurement: ① Attribution accuracy, use the degree of alignment between the attribution source and the true source for evaluation; ② Attribution sufficiency, use the proportion of output fragments with valid attribution sources to the entire output answer for evaluation; ③ Attribution relevance, use the proportion of attribution sources related to the input question and the output answer to all attribution sources for evaluation. For the evaluation method of specific measurement, for the evaluation related to matching, the present invention intends to improve the existing text matching and semantic matching methods by comprehensively considering the ambiguity of natural language and the ambiguity of matching; for the evaluation related to contribution, the present invention intends to evaluate the loss contribution and output contribution through the input perturbation method.

[0056] In one embodiment, the attribution evaluation method includes: a human evaluation method, a classification-based evaluation method, and a quantitative evaluation method.

[0057] In one embodiment, step 102 includes: designing instructions to guide the big model to automatically generate references / source texts based on intrinsic knowledge; constructing and training an attribution quality reward model; using the attribution quality reward model to score the attribution of the big model; in the attribution feedback stage, normalizing the score given by the attribution quality reward model, and giving executable instructions for the big model based on the average score of the attribution quality reward model on all citations.

[0058] In one embodiment, 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 using the initial attribution by randomly mixing or deleting some references; merging the initial attribution data set and the poor quality attribution data set to form a citation quality data set; training the initial attribution quality reward model using the citation quality data set to obtain a trained attribution quality reward model; the training loss is: (1) in, L represents the training loss, R represents the attributed quality reward model, Represents data from the Attribution Quality dataset.

[0059] Specifically, in the big model instruction design step, in order to design instructions to guide the big model to automatically generate references / source texts, the present invention intends to generate initial attribution based on prompts, and the prompts are designed as follows: "Use the form of [index] to cite references, cite the content that is most relevant to the question and answer, do not cite irrelevant content, and each sentence in the answer contains at least one reference."

[0060] Then build the attribution quality reward model (referred to as reward model, the same below) R Score the attribution of the large model. The present invention intends to design a training data set in the form of <question, document, attribution>, and use the initial attribution generated by the large model in the previous step to construct a poor quality attribution data set by randomly mixing or deleting some references. The initial attribution data set is combined with the poor quality attribution data set to form a citation quality data set. The present invention intends to use GPT-J-6B as the initial reward model, and use the constructed citation quality data set for training, using one initial attribution and three poor quality attributions in each round. The training goal is to make the reward model give a lower score to poor quality attributions, and the loss is shown in the following formula (1).

[0061] Finally, using the trained reward model RProvide feedback and guidance on the initial attribution of the big model to improve the quality of the big model's citations and reduce attribution illusions. Specifically, normalize the scores given by the reward model, and give the big model executable instructions based on the average score of the reward model on all citations, such as "You have cited appropriate literature, but please provide specific citations for each output statement." This helps the big model correct its own attribution.

[0062] In one embodiment, step 104 includes: obtaining public data on the Internet and annotating it to form a <text, authority> database; for each document Y in the database, performing text matching between the attributed document X generated by the large model and the document Y, and calculating the text matching similarity; performing fuzzy semantic matching between the attributed document X and the document Y, and calculating the fuzzy semantic matching similarity; based on the text matching similarity and the fuzzy semantic matching similarity, selecting matching texts higher than a preset similarity threshold in the database to form a permission matching library of the attributed document X; constructing a permission determination system based on expert knowledge; and achieving a final determination of the authority of the attributed document X based on the fuzzy rules set in the fuzzy permission determination system.

[0063] In one embodiment, fuzzy semantic matching is performed on the attributed document X and the document Y, and the fuzzy semantic matching similarity is calculated, including: processing the attributed 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 attributed document X and the document Y and sentence-level semantic representations of the attributed document X and the document Y; extracting fuzzy features of the word-level semantic representations and the sentence-level semantic representations of the attributed document X and the document Y through a fuzzy membership function to obtain word-level and sentence-level fuzzy semantic representations of the attributed document X and the document Y, and then adjusting the dimension of the fuzzy semantic representation through a fully connected neural network; merging 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; merging 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 the document Y; according to the fuzzy representation of the attributed document X and the fuzzy representation of the document Y, the fuzzy semantic matching similarity is obtained as follows: (2) in, In order to use fuzzy distance to calculate the fuzzy semantic matching similarity between the two, and are the fuzzy semantic representations of document X and document Y after fuzzy aggregation, and They are and The membership value of .

[0064] Specifically, the attribution source-authority matching method based on fuzzy logic is divided into the following steps: (1) We crawled public data from the Internet, annotated it with the existing label system, and used a combination of metadata, GPT-4 machine annotation, and expert annotation to create a <text, permissions> database with a scale of over 10K.

[0065] (2) For each document Y, perform text matching between the attribution source X generated by the large model and Y, and calculate the matching similarity ; Considering the fuzzy information, a fuzzy semantic matching model is constructed to perform fuzzy semantic matching on the attribution sources X and Y and calculate the matching similarity The realization of fuzzy semantic matching is proposed 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 as follows: Figure 6 shown.

[0066] This module first receives the semantic representation output by the convolutional layers at the word level and sentence level, extracts the fuzzy features of the word level and sentence level representations through the membership function, measures their uncertainty and fuzziness, and converts the semantic representations into fuzzy semantic representations. Then, through the fully connected neural network, the dimensions of the fuzzy semantic representations at different levels are made the same for subsequent fuzzy aggregation operations. Finally, the fuzzy aggregation operator is used to merge the fuzzy semantic representations at the word level and sentence level into the final representation. Let them be and , the fuzzy aggregation operation is as follows: (3) in, is a trainable parameter, initialized to a random value in the interval [0, 1]. Represents a fuzzy multiplication operation: (4) After obtaining the fuzzy representation, the document is calculated by the fuzzy similarity calculation method. X and Y The matching similarity of . and They are the fuzzy semantic representations of document X and document Y after fuzzy aggregation, and the fuzzy distance is used to calculate the semantic matching similarity between the two. , the semantic matching similarity is shown in formula (2).

[0067] (3) Based on text matching similarity and semantic matching similarity , select matching texts above the similarity threshold to form the authority matching library of attribution source X.

[0068] (4) Considering the diversity of matching results, a fuzzy permission determination system is constructed based on expert knowledge to achieve the final permission determination. For example, a rule in the fuzzy permission determination system is: If X and Y have high textual similarity and medium semantic similarity, and Y has a high authority level, then X and Y have a high degree of similar authority.

[0069] According to the fuzzy rules set by experts in the fuzzy authority determination system, the final determination of the authority of document X can be achieved.

[0070] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0071] In one embodiment, Figure 7 As shown, a large model attribution and confirmation device for user authority management is provided, including: a large model attribution evaluation system construction module, an attribution source confirmation module and a confirmation module, wherein: 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 attribution metrics for each type of attribution, and construct the large model attribution evaluation system from the two perspectives of specific metrics and general metrics for each type of attribution.

[0072] The attribution source confirmation module is used to determine the attribution source according to the big model attribution evaluation system using the big model intrinsic knowledge attribution method based on the reward model and feedback.

[0073] The rights confirmation module is used to determine the final authority by using the attribution source-authority matching method based on fuzzy logic.

[0074] In one of the embodiments, the large model attribution evaluation system construction module is also used to divide the large model attribution into intrinsic knowledge attribution, pre-trained data attribution and input question attribution according to the attribution source type; construct an attribution effect measurement framework from the two perspectives of various attribution specific metrics and general metrics; design an attribution evaluation method based on the attribution effect measurement framework; and obtain a large model attribution evaluation system based on the attribution classification results, the attribution effect measurement framework and the attribution evaluation method.

[0075] In one of the embodiments, the large model attribution evaluation system construction module is also used to determine the general and specific attribution metrics for the attribution measurement effect. The general attribution metrics include: accuracy, sufficiency, and relevance; the specific metric for intrinsic knowledge attribution is that the attribution source matches / implies the output text, the specific metric for pre-trained data attribution is that the attribution source matches the training data, and the specific metric for input question attribution is that the attribution source contributes to the output.

[0076] In one of the embodiments, the attribution evaluation methods in the large model attribution evaluation system construction module include: manual evaluation method, classification-based evaluation method and quantitative evaluation method.

[0077] In one of the embodiments, the attribution source confirmation module is also used to design instructions to guide the large model to automatically generate references / source texts based on intrinsic knowledge; build and train an attribution quality reward model; use the attribution quality reward model to score the attribution of the large model; in the attribution feedback stage, normalize the score given by the attribution quality reward model, and give executable instructions for the large model based on the average score of the attribution quality reward model on all citations.

[0078] In one embodiment, 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 using the initial attribution by randomly mixing or deleting some references; merging the initial attribution data set and the poor quality attribution data set to form a citation quality data set; training the initial attribution quality reward model using the citation quality data set to obtain a trained attribution quality reward model; the training loss is shown in formula (1).

[0079] In one of the embodiments, the right confirmation module is also used to obtain public data on the Internet and annotate it to form a <text, authority> database; for each document Y in the database, the attributed document X generated by the large model is text-matched with the document Y to calculate the text matching similarity; the attributed document X is fuzzy-semantically matched with the document Y to calculate the fuzzy semantic matching similarity; according to the text matching similarity and the fuzzy semantic matching similarity, matching texts higher than a preset similarity threshold are selected from the database to form a permission matching library of the attributed document X; a permission determination system is constructed based on expert knowledge; and according to the fuzzy rules set in the fuzzy permission determination system, the final determination of the authority of the attributed document X is achieved.

[0080] In one embodiment, the right confirmation module is further used to process the attributed document X and document Y through a word-level convolution layer and a sentence-level convolution layer to obtain the word-level semantic representation of the attributed document X and document Y and the sentence-level semantic representation of the attributed document X and document Y; extract the fuzzy features of the word-level semantic representation and the sentence-level semantic representation of the attributed document X and document Y through a fuzzy membership function to obtain the word-level and sentence-level fuzzy semantic representation of the attributed document X and document Y, and then adjust the dimension of the fuzzy semantic representation through a fully connected neural network; merge the word-level and sentence-level fuzzy language representation of the attributed document X after the dimension adjustment through a fuzzy aggregation operator to obtain the fuzzy representation of the attributed document X; merge the word-level and sentence-level fuzzy language representation of the document Y after the dimension adjustment through a fuzzy aggregation operator to obtain the 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.

[0081] 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 mentioned above, which will not be repeated here. Each module 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 a combination thereof. 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.

[0082] 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 8 As shown. 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 and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a large model attribution and confirmation method for user authority management is implemented. 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 covered on the display screen, or a button, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse.

[0083] Those skilled in the art will understand that Figure 8The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0084] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps in the above method embodiment when executing the computer program.

[0085] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0086] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A large-model attribution and confirmation device for user rights management, characterized in that: The device comprises: 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 attribution metrics for each type of attribution, and build a large model attribution evaluation system from the two perspectives of specific metrics and general metrics for each type of attribution; An attribution source confirmation module is used to determine the attribution source according to the large model attribution evaluation system using a large model intrinsic knowledge attribution method based on a reward model and feedback; The right confirmation module is used to determine the final authority by using the attribution source-authority matching method based on fuzzy logic for the attribution source.

2. A large-model attribution and confirmation method for user rights management, characterized in that: The method comprises: Classify the big model attribution according to the attribution source type, set specific metrics and general attribution metrics for each type of attribution, and build a big model attribution evaluation system from two perspectives: specific metrics and general metrics for each type of attribution; According to the large model attribution evaluation system, the attribution source is determined by using the large model intrinsic knowledge attribution method based on the reward model and feedback; The attribution source-authority matching method based on fuzzy logic is adopted for the attribution source to determine the final authority.

3. The large-model attribution and confirmation method for user rights management according to claim 2 is characterized in that: Classify the big model attribution according to the attribution source type, set specific metrics and general attribution metrics for each type of attribution, and build a big model attribution evaluation system from two perspectives: specific metrics and general metrics for each type of attribution, including: According to the attribution source type, the large model attribution is divided into intrinsic knowledge attribution, pre-training data attribution, and input question attribution; Construct an attribution effect measurement framework from two perspectives: specific attribution metrics and general metrics. Designing an attribution evaluation method based on the attribution effect measurement framework; According to the attribution classification results, attribution effect measurement framework and attribution evaluation method, a large-model attribution evaluation system is obtained.

4. The large-model attribution and confirmation method for user rights management according to claim 3 is characterized in that: The attribution effect measurement framework is constructed from two perspectives: various attribution-specific metrics and general metrics, including: For the attribution measurement effect, the general and specific attribution metrics are determined. The general metrics of attribution include accuracy, sufficiency, and relevance. The specific metric for intrinsic knowledge attribution is that the attribution source matches / implies the output text, the specific metric for pre-trained data attribution is that the attribution source matches the training data, and the specific metric for input question attribution is that the attribution source contributes to the output.

5. The large-model attribution and confirmation method for user rights management according to claim 3 is characterized in that: The attribution evaluation method includes: a manual evaluation method, a classification-based evaluation method and a quantitative evaluation method.

6. The large-model attribution and confirmation method for user rights management according to claim 2 is characterized in that: According to the large model attribution evaluation system, the attribution source is determined by using the large model intrinsic knowledge attribution method based on the reward model and feedback, including: Design instructions guide the large model to automatically generate references / source texts based on inherent knowledge; Constructing and training an attribution quality reward model; using the attribution quality reward model to score the large model attribution; In the attribution feedback stage, the scores given by the attribution quality reward model are normalized, and instructions executable by the large model are given according to the average score of the attribution quality reward model on all references.

7. The large-model attribution and confirmation method for user rights management according to claim 6 is characterized in that: The training process of the attributed quality reward model includes: Design a training dataset in the form of <question, document, attribution>; Using initial attribution to construct a poor quality attribution dataset by randomly mixing or deleting some references; The initial attribution dataset and the poor-quality attribution dataset are combined to form a citation-quality dataset; The initial attribution quality reward model is trained using the citation quality dataset to obtain a trained attribution quality reward model; the training loss is: in, L represents the training loss, R represents the attributed quality reward model, Represents data from the Attribution Quality dataset.

8. The large-model attribution and confirmation method for user rights management according to claim 2 is characterized in that: The attribution source is matched with the authority based on fuzzy logic to determine the final authority, including: Obtain and annotate public data on the Internet to form a <text, permissions> database; For each document Y in the database, perform text matching between the attributed document X generated by the large model and document Y, and calculate the text matching similarity; Perform fuzzy semantic matching on the attributed document X and document Y, and calculate the fuzzy semantic matching similarity; According to 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 permission matching library of the attributed document X; Build an authority determination system based on expert knowledge; According to the fuzzy rules set in the fuzzy permission determination system, the final determination of the permission of the attributed document X is achieved.

9. The large-model attribution and confirmation method for user rights management according to claim 8 is characterized in that: Perform fuzzy semantic matching on the attributed document X and document Y, and calculate the fuzzy semantic matching similarity, including: Process the attributed document X and document Y through a word-level convolutional layer and a sentence-level convolutional layer to obtain a word-level semantic representation of the attributed document X and document Y and a sentence-level semantic representation of the attributed document X and document Y; The fuzzy features of the lexical level semantic representation and the sentence level semantic representation of the attributed document X and document Y are extracted through the fuzzy membership function to obtain the lexical level and sentence level fuzzy semantic representation of the attributed document X and document Y, and then the dimension of the fuzzy semantic representation is adjusted through a fully connected neural network. The fuzzy representation of the attributed document X is obtained by merging the lexical-level and sentence-level fuzzy language representations of the dimensionally adjusted attributed document X through the fuzzy aggregation operator; The fuzzy representation of document Y is obtained by merging the lexical-level and sentence-level fuzzy language representations of the dimensionally adjusted document Y through the fuzzy aggregation operator; According to the fuzzy representation of the attributed document X and the fuzzy representation of the document Y, the fuzzy semantic matching similarity is obtained as follows: in, In order to use fuzzy distance to calculate the fuzzy semantic matching similarity between the two, and are the fuzzy semantic representations of document X and document Y after fuzzy aggregation, and They are and The membership value of .

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the large-model attribution and confirmation method for user rights management as described in any one of claims 2 to 9.

Images