A multi-stage human-machine attribute value alignment method and system based on large model

Abstract

The invention discloses a multi-stage human-machine attribute value alignment method and system based on a large model, the method comprising: using multiple large models to obtain the decision attribute weights of the #imgabs1# stage according to the situation information of the #imgabs0# stage, and then using weighted factors to fuse to obtain the machine intelligence weights of each attribute of the #imgabs2# stage; using a group decision method to obtain the human decision weights of each attribute of the #imgabs3# stage; based on the machine intelligence weights and human decision weights of each attribute of the #imgabs4# stage, a loss function is established; by calculating the partial derivative of the loss function with respect to the large model weight factor, using the gradient descent method and updating the weighted factors of each large language model on each attribute in the #imgabs5# stage based on the partial derivative calculation value, for the attribute weights output by the fusion of each large language model in the #imgabs6# stage; looping the above process to finally realize the gradual alignment of human-machine attribute weights. The invention can reduce decision-making risks and effectively improve the accuracy and reliability of artificial intelligence in complex decision support.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing for emergency decision-making, and in particular relates to a multi-stage human-machine attribute value alignment method and system based on a large model. Background Art

[0002] With the rapid development of artificial intelligence technology, machine learning algorithms, especially large models, have shown significant advantages in dealing with complex decision-making problems. However, these algorithms and models have a certain degree of instability. Once the model deviates or fails, its continued use will bring considerable risks to decision-making. . The value of decision attributes reflects the importance of evaluation criteria and the orientation of decision goals. The distribution of attribute values ​​directly affects the evaluation of options and the aggregation of final opinion preferences. It is a basic and key research content that cannot be avoided in decision-making problems. Therefore, it is urgent to guide machine learning algorithms to draw wisdom from human experience in the decision-making process to ensure the alignment of machine and human in the value of decision attributes, thereby reducing decision risks and improving the reliability of decisions. Traditional attribute alignment methods focus more on reaching consensus among humans, usually through group decision-making and other methods. Although these methods can ensure the value consistency among human decision makers to a certain extent, they cannot solve the cognitive differences between machine learning models and human decision makers. In the prior art, although some studies have attempted to combine machine learning algorithms with human decision attribute values, most of them focus on simply calculating the weight distribution results from decision preference data. , which is not suitable for scenarios where humans and machines make mixed decisions. Especially in the multi-stage decision-making process, how to dynamically adjust and optimize the value matching between humans and machines remains a technical problem to be solved.

[0003] Traditional attribute value alignment methods only focus on the unity of value within humans , but ignores the difference in value cognition between machines and humans in the process of human-machine collaborative decision-making. Therefore, there is a lack of systematic methods to integrate machine learning algorithms and human attribute values, making it difficult to effectively apply the value alignment process in complex human-machine collaborative decision-making scenarios. At the same time, existing technologies focus more on a single decision-making scenario and cannot achieve human-machine attribute value matching in a multi-stage decision-making process.

[0004] Since human decision-making developed earlier than human-machine collaborative decision-making, past research has focused more on the human decision-making process alone. , rather than in-depth exploration of the mechanism of human-machine value coordination. At the same time, the neglect of dynamics makes the existing technology unable to adapt to the continuous changes in the decision-making environment, and lacks a mechanism that can effectively adjust value matching at different decision-making stages. Summary of the invention

[0005] The present invention provides a multi-stage human-machine attribute value alignment method and system based on a large model, aiming to overcome the limitations of the existing technology and achieve deeper value collaboration between machine learning models and human decision makers.

[0006] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0007] A multi-stage human-machine attribute value alignment method based on a big model, including:

[0008] Step 1, use A large language model, based on events The situation information of the stage is used to obtain the event decision problem in The attribute weight of the stage; where the decision problem includes multiple target factors, each target factor is recorded as a decision attribute of the decision problem;

[0009] Step 2: For each decision attribute On, use The weighting factor of the large model of the stage Bringing it all together The attribute weights output by the large language model are obtained The machine intelligence attribute weights of the stage; among them, for Phase Large language model In Properties The weighting factor on ;

[0010] Step 3: Using the group decision-making method, based on the evaluation matrix of human decision makers on the decision problem under multi-dimensional attributes, we can obtain attribute weights of human decision makers at each stage;

[0011] Step 4: The value allocation of each attribute in the stage is based on the attribute weights of the machine intelligence and the attribute weights of the human decision maker in that stage, and the loss function is established;

[0012] Step 5: Calculate the loss function with respect to the large model weighting factor The partial derivative is obtained by using the gradient descent method and updating based on the partial derivative calculation value Phase A large language model in attributes The weighting factor on ;

[0013] Step 6, After back propagation of each attribute in the stage, we get The set of weighting factors for the stage , used for In the stage, the attribute weights output by each language model are integrated to obtain The machine intelligence attribute weights of the stage; where N is the attribute set of the event decision problem The number of attributes in .

[0014] Furthermore, step 1 is divided into two steps to obtain the decision problem in Attribute weights of the phases:

[0015] Step 1.1, use Large language model Prompt content based on the given model role and Specifications , output new events in Analysis results of the stage :

[0016]

[0017] Among them, the model role prompt content Includes: You will play an excellent decision-making problem Data analysis experts, please combine the decision-making problem Expertise in the field, Stage situation information The results will be collated and analyzed to provide reference for decision-making;

[0018] Step 1.2, Large language model Combined analysis results , prompt content based on the given model role and Specifications , output decision attribute set Machine intelligence attribute weights :

[0019]

[0020] in, For the Large language model Attributes exist Recommended weights for stages; model role hints Including: (1) Decision-making problem: Now there is a decision-making problem ; (2) Confirm the background: Please output the decision-making issues (3) Tasks: In order to provide Choose the best response plan and have a set of alternative plans , and from the attribute set Evaluate the alternative plans; (4) Please analyze the results based on the current situation data of stage t , combined with the decision problem Expertise in the field, given the current stage t for the attribute set Attribute weight distribution.

[0021] Furthermore, step 2 is expressed using the formula:

[0022]

[0023] in, For the Large language model Attributes exist The recommended weights for the stages, For integration A large language model obtains machine intelligence attributes exist The weight of the stage, for The weight set of machine intelligence attributes at each stage.

[0024] Furthermore, step 3 is specifically as follows:

[0025] Step 3.1, construct an evaluation matrix of the P alternative solutions for the decision problem by human decision makers; wherein, M evaluation matrices are obtained corresponding to M human decision makers, and the rows and columns of each evaluation matrix are the P solutions and the N attributes of the attribute set, respectively, and the element value is the evaluation value of the corresponding attribute of the corresponding solution by the human decision maker, and the attribute evaluation value is obtained by the human decision maker expressing his preference through a language term set;

[0026] Step 3.2, clustering the M evaluation matrices according to the cosine similarity between every two evaluation matrices to obtain several evaluation matrix centers;

[0027] Step 3.3: Perform weighted summation of all evaluation matrix centers to obtain the evaluation matrix of the human decision maker ;

[0028] Step 3.4, Evaluation Matrix Based on Human Decision Makers , the entropy weight of each attribute is calculated as the attribute weight of the human decision maker.

[0029] Furthermore, the loss function established in step 4 is:

[0030]

[0031] in, represents the loss function, Represents the human decision maker's decision on the attribute exist The weight of the stage is expressed as, Representing large language model pairs exist The recommended weight for the stage.

[0032] Furthermore, step 5 is specifically as follows:

[0033] Step 5.1, calculate the loss function with respect to the proposed weights Partial derivative of:

[0034]

[0035] Step 5.2, calculate the recommended weight About weighting factors Partial derivative:

[0036]

[0037] Step 5.3, use the chain rule to get the loss function with respect to the weighting factor Partial derivative of:

[0038]

[0039] In the formula, Indicates The large model is The learning rate under each attribute is used to control the speed of alignment of human and machine attribute values;

[0040] Step 5.4, update the weighting factor using gradient descent :

[0041] .

[0042] A multi-stage human-machine attribute value alignment system based on a large model, including:

[0043] Large model module, including A large language model, each of which is used to: The situation information of the stage is used to obtain the event decision problem in The attribute weight of the stage; where the decision problem includes multiple target factors, each target factor is recorded as an attribute of the decision problem;

[0044] The weight fusion module is used to: On, use The weighting factor of the large model of the stage Bringing it all together The attribute weights output by the large language model are obtained The machine intelligence attribute weights of the stage; among them, for Phase Large language model In Properties The weighting factor on ;

[0045] The group decision module is used to: adopt the group decision method, based on the evaluation matrix of human decision makers on the decision problem under multi-dimensional attributes, obtain attribute weights of human decision makers at each stage;

[0046] Loss function building block for: The value allocation of each decision attribute in the stage is based on the attribute weights of the machine intelligence and the attribute weights of the human decision maker at that stage, and the loss function is established;

[0047] Weighted factor update module, used to calculate the loss function with respect to the large model weighted factor The partial derivative is obtained by using the gradient descent method and updating based on the partial derivative calculation value Phase A large language model in attributes The weighting factor on ;in, For weight fusion module The attribute weights output by each language model are integrated during the phase.

[0048] Furthermore, the weight fusion module is expressed using the formula:

[0049]

[0050] in, For the Large language model Attributes exist The recommended weights for the stages, For integration A large language model obtains machine intelligence attributes exist The weight of the stage, for The machine intelligence attribute weight set of the stage, N is the attribute set of the event decision problem The number of attributes in .

[0051] Furthermore, the group decision-making module is specifically used to:

[0052] (1) Construct an evaluation matrix for the P alternative solutions of the decision problem by human decision makers; where M human decision makers correspond to M evaluation matrices, and the rows and columns of each evaluation matrix are the P solutions and the N attributes of the attribute set, respectively. The element value is the evaluation value of the corresponding attribute of the corresponding solution by the human decision maker, and the evaluation value is obtained by the human decision maker expressing his preference through a language term set;

[0053] (2) Cluster the M evaluation matrices according to the cosine similarity between every two evaluation matrices to obtain several evaluation matrix centers;

[0054] (3) Perform weighted summation on the centers of all evaluation matrices to obtain the evaluation matrix of the human decision maker: ;

[0055] (4) Evaluation Matrix Based on Human Decision Makers , the entropy weight of each attribute is calculated as the attribute weight of the human decision maker.

[0056] Furthermore, the loss function established by the loss function building module is:

[0057]

[0058] in, represents the loss function, Represents the human decision maker's decision on the attribute exist The weight of the stage is expressed as, Representing large language model pairs exist The recommended weight of the stage;

[0059] The weighting factor update module is specifically used for:

[0060] (1) Calculate the loss function with respect to the recommended weights Partial derivative of:

[0061]

[0062] Where N is the attribute set of the event decision problem The number of attributes in ;

[0063] (2) Calculate the recommended weight About weighting factors Partial derivative of:

[0064]

[0065] (3) Use the chain rule to get the loss function with respect to the weighting factor Partial derivative of:

[0066]

[0067] In the formula, Indicates The large model is The learning rate under each attribute is used to control the speed of alignment of human and machine attribute values;

[0068] (4) Update the weighting factor using the gradient descent method:

[0069] .

[0070] Compared with the prior art, the beneficial effects of the present invention are: through a multi-stage iterative process, a deep collaboration between the machine learning model and the human decision maker on the value of decision attributes is achieved, thereby reducing decision risks and effectively improving the accuracy and reliability of artificial intelligence in complex decision support. In addition, the method of the present invention can dynamically adjust and optimize the value matching between humans and machines, adapt to the ever-changing decision-making environment, and provide effective support for multi-stage decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is the basic framework of multi-stage decision making in the embodiment of the present application;

[0072] Figure 2 This is a method for aligning the value of human-machine attributes based on a large model in an embodiment of the present application. DETAILED DESCRIPTION

[0073] The following is a detailed description of an embodiment of the present invention. This embodiment is based on the technical solution of the present invention, and provides a detailed implementation method and a specific operation process to further explain the technical solution of the present invention.

[0074] The decision-making process is based on the decision-making environment information The continuous change of Situation, at each decision node t, from the decision alternatives set At each stage of decision-making, the collective wisdom of humans and the reasoning of machines are integrated to obtain the decision results of the current stage, and the difference in the value of humans and machines in multiple attribute dimensions is evaluated. Then, based on the collective wisdom of humans, the machine's emphasis on data calculation and knowledge reasoning in the next stage is optimized.

[0075] Human Groups and machines All are decision-makers. Specifically, human decision makers refer to experts in the field who participate in specific decision-making problems. They express their opinions on decision-making problems based on their own experience and knowledge. Machine intelligence refers to algorithms or intelligent agents that display the decision results of problems based on real-time data and algorithmic logic. The decision-making process is a progressive multi-stage decision-making for the same problem, and t is the decision-making stage. Environmental information for decision-making It changes over time and is different at different decision nodes, often indicating the state of the decision event. Experts and machines express their opinions independently at each decision stage, and the solution for execution at the current stage is solved by integrating the preferences of both parties. The decision-making process is shown as follows: Figure 1 shown.

[0076] In the present invention, attribute value alignment means that in the continuous decision-making iteration, humans and machine intelligence can reach a consensus on the importance of the multi-objective factors of the decision-making problem. The attribute set Q refers to the evaluation dimension of the alternatives, that is, the multi-objective factors for evaluating each alternative, and the attribute value refers to the multi-objective importance distribution of the scheme evaluation, that is, the weight vector W corresponding to the attribute set Q.

[0077] Multidimensional attribute sets for decision-making problems between human groups and machines , with their own attribute value distribution, humans are , the machine is , both weight vectors conform to: and , and Specifically, the present invention designs different models to calculate the human decision maker and machine intelligence The value of each decision attribute. Human decision makers use language preference to express the evaluation value of each attribute, and obtain their preference matrix , and then calculate the attribute value distribution of human decision makers based on the group decision method Machine intelligence calculates the final machine attribute value distribution by integrating the outputs of multiple large language models (LLMs). . Back propagation algorithms are applied in the multi-stage decision-making process to adjust the results of these models to make them closer to human decision-making results. This approach effectively achieves value alignment between machine learning models and human decision makers, and improves the accuracy and reliability of artificial intelligence in complex decision support. The specific model framework is as follows Figure 2 shown.

[0078] In the process of human-machine collaborative decision-making, through the feedback of human collective wisdom, machine intelligence learns the value model of human decision-making, and gradually aligns the cognition of human and machine attribute values ​​in the gradual iteration. Specifically, establish a value distribution model from the attribute of human decision makers. Assigning value to machine attributes The back-propagation algorithm gradually optimizes the machine attribute value allocation in the multi-stage decision-making process and realizes the human-machine alignment of attribute value allocation.

[0079] This embodiment is based on a multi-stage human-machine attribute value alignment method based on a large model. Based on the above definition, the detailed process is as follows:

[0080] Step 1, use A large language model, based on events The situation information of the stage is used to obtain the event decision problem in The attribute weight of the stage.

[0081] The LLM large language model stores the learned knowledge in the huge Transformer model parameters through large-scale pre-training. It can not only master the surface knowledge such as lexical and syntactic, but also has the ability to understand and reason about factual and conceptual knowledge. This step obtains attribute weights through the large language model, including analyzing the development of events and determining attribute weights, and sets large model role prompts and specifications. .

[0082] Step 1.1: Analyze the development of events. Specifically, use a large language model to prompt content based on the given model role. and Specifications , output new events in Analysis results of the stage :

[0083] .

[0084] in, Indicates A large language model, the role prompt content of the large language model in step 1.1 and norms As shown in Table 1.

[0085]

[0086] Step 1.2, Large language model Combined analysis results , prompt content based on the given model role and Specifications , output decision attribute set Machine intelligence attribute weights :

[0087]

[0088] In the formula, Indicates The large language model output decision problem is The attribute weight of the stage.

[0089] Among them, the role prompt content of the large language model in step 1.2 and norms As shown in Table 2.

[0090]

[0091] Step 2: For each decision attribute On, use The weighting factor of the large model of the stage Bringing it all together The attribute weights output by the large language model are obtained The machine intelligence attribute weight of the stage. It is expressed as:

[0092]

[0093] in, For the Large language model Attributes exist The recommended weights for the stages, for Phase Large language model In Properties The weighting factor on For integration A large language model obtains machine intelligence attributes exist The weight of the stage, for The weight set of machine intelligence attributes at each stage.

[0094] Step 3: Using the group decision-making method, based on the evaluation matrix of human decision makers on the decision problem under multi-dimensional attributes, we can obtain The attribute weights of human decision makers at each stage. Including:

[0095] Step 3.1, construct an evaluation matrix of P alternative plans for the decision problem for human decision makers; wherein, M evaluation matrices are obtained corresponding to M human decision makers, and the rows and columns of each evaluation matrix are P plans and N attributes of the attribute set, respectively, and the element value is the evaluation value of the corresponding attribute of the corresponding plan by the human decision maker, and the evaluation value is obtained by expressing the preference of the human decision maker through a language term set; the language term set adopts seven-value language terms: very good 6, good 5, good 4, average 3, poor 2, poor 1, and very poor 0.

[0096] Step 3.2, clustering the M evaluation matrices according to the cosine similarity between every two evaluation matrices to obtain several evaluation matrix centers;

[0097] Step 3.3: Perform weighted summation of all evaluation matrix centers to obtain the evaluation matrix of the human decision maker ;

[0098] Step 3.4, Evaluation Matrix Based on Human Decision Makers , calculate the entropy weight of each attribute as the attribute weight of the human decision maker. exist The weight of the stage is expressed as .

[0099] Step 4: The value distribution of each attribute in the stage is based on the attribute weights of the machine intelligence and the attribute weights of the human decision maker at that stage, and a loss function is established ,Right now:

[0100]

[0101] in, represents the loss function, Represents the human decision maker's decision on the attribute exist The weight of the stage is expressed as, Representing large language model pairs exist The recommended weight for the stage.

[0102] Step 5: Calculate the loss function with respect to the weighted factor of the large language model The partial derivative is obtained by using the gradient descent method and updating based on the partial derivative calculation value Phase A large language model in attributes The weighting factor on .

[0103] Step 5.1, calculate the loss function with respect to the proposed weights Partial derivative of:

[0104]

[0105] Step 5.2, calculate the recommended weight About weighting factors Partial derivative of:

[0106]

[0107] Step 5.3, use the chain rule to get the loss function with respect to the weighting factor Partial derivative of:

[0108]

[0109] In the formula, Indicates The large model is The learning rate under each attribute is used to control the speed of alignment of human and machine attribute values;

[0110] Step 5.4, update the weighting factor using gradient descent :

[0111] .

[0112] Step 6, After back propagation of each attribute in the stage, The set of weighting factors for the stage , used for In the stage, the attribute weights output by each language model are integrated to obtain The machine intelligence attribute weights of the stage. Machine intelligence learns the value preferences of human attribute weights and gradually aligns the attribute weights of humans and machines.

[0113] The above embodiments are preferred embodiments of the present application. Ordinary technicians in this field can also make various changes or improvements on this basis. Without departing from the overall concept of the present application, these changes or improvements should fall within the scope of protection required by the present application.

[0114] References:

[0115] [1] Balasubramanian N, Ye Y, Xu M. Substituting Human Decision-Making with Machine Learning: Implications for Organizational Learning[J]. Academy of Management Review, 2022, 47(3): 448-465.

[0116] [2] Haesevoets T, De Cremer D, Dierckx K, et al. Human-machine collaboration in managerial decision making[J]. Computers in Human Behavior, 2021, 119: 106730.

[0117] [3] Sturm T, Gerlach JP, Pumplun L, et al. Coordinating Human and Machine Learning for Effective Organizational Learning[J]. MIS Quarterly, 2021, 45(3b): 1581-1602.

[0118] [4] Turel O, Kalhan S. Prejudiced against the Machine? ImplicitAssociations and the Transience of Algorithm Aversion[J]. MIS Quarterly, 2023, 47(4): 1369-1394.

[0119] [5] Dombi J, Jónás T. Consensus measures based on a fuzzy concept[J]. European Journal of Operational Research, 2024, 315(2): 642-653.

[0120] [6] Zi Yingping, Liu Xinwang, Han Shilian. Probabilistic linguistic multi-attribute group decision-making method based on trust network based on quantum cognition. Systems Engineering - Theory and Practice.

[0121] [7] Xu XH, Zhao CV, He JS, et al. Multi-attribute decision making method in associated variable weight space with multi-type heterogeneous data[J]. Systems Engineering - Theory & Practice, 2020, 40(7): 1895-1905.

[0122] [8] Xu Xuanhua, Xiao Ting, Chen Xiaohong. Consensus model for emergency decision-making of large groups based on group intelligence knowledge fusion in social network environment[J]. Chinese Journal of Management Science, 2024, 32(02): 285-297.

[0123] [9] Xu Xuanhua, Chen Xiaohong. Research on a multi-attribute and multi-option large group decision-making method[J]. Journal of Systems Engineering, 2008, 23(2): 137-141.

[0124]

[10] Xu Xuanhua, Liu Ying, Chen Xiaohong. Large group risk emergency decision-making method based on group pressure[J]. Journal of Systems Engineering, 2022, 37(4): 460-476.

Claims

1. A multi-stage human-machine attribute value alignment method based on a large model, characterized in that: include: Step 1, use A large language model, based on events The situation information of the stage is used to obtain the event decision problem in The attribute weight of the stage; where the decision problem includes multiple target factors, each target factor is recorded as a decision attribute of the decision problem; Step 2: For each decision attribute On, use The weighting factor of the large model of the stage Bringing it all together The attribute weights output by the large language model are obtained The machine intelligence attribute weights of the stage; among them, for Phase Large language model In Properties The weighting factor on ; Step 3: Using the group decision-making method, based on the evaluation matrix of human decision makers on the decision problem under multi-dimensional attributes, we can obtain attribute weights of human decision makers at each stage; Step 4: The value allocation of each attribute in the stage is based on the attribute weights of the machine intelligence and the attribute weights of the human decision maker in that stage, and the loss function is established; Step 5: Calculate the loss function with respect to the large model weighting factor The partial derivative is obtained by using the gradient descent method and updating based on the partial derivative calculation value Phase A large language model in attributes The weighting factor on ; Step 6, After back propagation of each attribute in the stage, The set of weighting factors for the stage , used for In the stage, the attribute weights output by each language model are integrated to obtain The machine intelligence attribute weights of the stage; where N is the attribute set of the event decision problem The number of attributes in .

2. The multi-stage human-machine attribute value alignment method based on a large model according to claim 1 is characterized in that: Step 1: Get the decision problem in two steps Attribute weights of the phases: Step 1.1, use Large language model Prompt content based on the given model role and Specifications , output new events in Analysis results of the stage : ； Among them, the model role prompt content Includes: You will play an excellent decision-making problem Data analysis experts, please combine the decision-making problem Expertise in the field, Stage situation information The results will be collated and analyzed to provide reference for decision-making; Step 1.2, Large language model Combined analysis results , prompt content based on the given model role and Specifications , output decision attribute set Machine intelligence attribute weights : ； in, For the Large language model Attributes exist Recommended weights for stages; model role hints Including: (1) Decision-making problem: Now there is a decision-making problem ; (2) Confirm the background: Please output the decision-making issues (3) Tasks: In order to provide Choose the best response plan and have a set of alternative plans , and from the attribute set Evaluate the alternative plans; (4) Please analyze the results based on the current situation data of stage t , combined with the decision problem Expertise in the field, given the current stage t for the attribute set Attribute weight distribution.

3. The multi-stage human-machine attribute value alignment method based on a large model according to claim 1 is characterized in that: Step 2 is expressed using the formula: ； in, For the Large language model Attributes exist The recommended weights for the stages, For integration A large language model obtains machine intelligence attributes exist The weight of the stage, for The weight set of machine intelligence attributes at each stage.

4. The multi-stage human-machine attribute value alignment method based on a large model according to claim 1 is characterized in that: Step 3 is as follows: Step 3.1, construct an evaluation matrix of the P alternative solutions for the decision problem by human decision makers; wherein, M evaluation matrices are obtained corresponding to M human decision makers, and the rows and columns of each evaluation matrix are the P solutions and the N attributes of the attribute set, respectively, and the element value is the evaluation value of the corresponding attribute of the corresponding solution by the human decision maker, and the evaluation value is obtained by the human decision maker expressing his preference through a language term set; Step 3.2, clustering the M evaluation matrices according to the cosine similarity between every two evaluation matrices to obtain several evaluation matrix centers; Step 3.3: Perform weighted summation of all evaluation matrix centers to obtain the evaluation matrix of the human decision maker ; Step 3.4, Evaluation Matrix Based on Human Decision Makers , the entropy weight of each attribute is calculated as the attribute weight of the human decision maker.

5. The multi-stage human-machine attribute value alignment method based on a large model according to claim 1 is characterized in that: The loss function established in step 4 is: ； in, represents the loss function, Represents the human decision maker's decision on the attribute exist The weight of the stage is expressed as, Representing large language model pairs exist The recommended weight for the stage.

6. The multi-stage human-machine attribute value alignment method based on a large model according to claim 5 is characterized in that: Step 5 is as follows: Step 5.1, calculate the loss function with respect to the proposed weights Partial derivative of: ； Step 5.2, calculate the recommended weight About weighting factors Partial derivative: ； Step 5.3, use the chain rule to get the loss function with respect to the weighting factor Partial derivative of: ； In the formula, Indicates The large model is The learning rate under each attribute is used to control the speed of alignment of human and machine attribute values; Step 5.4, update the weighting factor using gradient descent : 。 7. A multi-stage human-machine attribute value alignment system based on a large model, characterized in that: include: Large model module, including A large language model, each of which is used to: The situation information of the stage is used to obtain the event decision problem in The attribute weight of the stage; where the decision problem includes multiple target factors, each target factor is recorded as an attribute of the decision problem; The weight fusion module is used to: On, use The weighting factor of the large model of the stage Bringing it all together The attribute weights output by the large language model are obtained The machine intelligence attribute weights of the stage; among them, for Phase Large language model In Properties The weighting factor on ; The group decision module is used to: adopt the group decision method, based on the evaluation matrix of human decision makers on decision problems under multiple attributes, obtain attribute weights of human decision makers at each stage; Loss function building block for: The value allocation of each decision attribute in the stage is based on the attribute weights of the machine intelligence and the attribute weights of the human decision maker at that stage, and the loss function is established; Weighted factor update module, used to calculate the loss function with respect to the large model weighted factor The partial derivative is obtained by using the gradient descent method and updating based on the partial derivative calculation value Phase A large language model in attributes The weighting factor on ;in, For weight fusion module The attribute weights output by each language model are integrated during the phase.

8. The multi-stage human-machine attribute value alignment system based on a large model according to claim 7 is characterized in that: The weight fusion module is expressed using the formula: ； in, For the Large language model Attributes exist The recommended weights for the stages, For integration A large language model obtains machine intelligence attributes exist The weight of the stage, for The machine intelligence attribute weight set of the stage, N is the attribute set of the event decision problem The number of attributes in .

9. The multi-stage human-machine attribute value alignment system based on a large model according to claim 7 is characterized in that: Group decision-making module, specifically used for: (1) Construct an evaluation matrix of the P alternative solutions for the decision problem. The M human decision makers are matched with M attribute evaluation matrices. The rows and columns of each attribute evaluation matrix are the P solutions and the N attributes of the attribute set. The element value is the evaluation value of the corresponding attribute of the corresponding solution by the human decision maker. The attribute evaluation value is obtained by expressing the preference of the human decision maker through a language term set. (2) According to the cosine similarity between every two attribute evaluation matrices, the M attribute evaluation matrices are clustered to obtain several attribute evaluation matrix centers; (3) Perform weighted summation on the centers of all attribute evaluation matrices to obtain the attribute evaluation matrix of the human decision maker: ; (4) Attribute evaluation matrix based on human decision makers , the entropy weight of each attribute is calculated as the attribute weight of the human decision maker.

10. The multi-stage human-machine attribute value alignment system based on a large model according to claim 7 is characterized in that: The loss function established by the loss function building module is: ； in, represents the loss function, Represents the human decision maker's decision on the attribute exist The weight of the stage is expressed as, Representing large language model pairs exist The recommended weight of the stage; The weighting factor update module is specifically used for: (1) Calculate the loss function with respect to the recommended weights Partial derivative of: ； Where N is the attribute set of the event decision problem The number of attributes in ; (2) Calculate the recommended weight About weighting factors Partial derivative of: ； (3) Use the chain rule to get the loss function with respect to the weighting factor Partial derivative of: ； In the formula, Indicates The large model is The learning rate under each attribute is used to control the speed of alignment of human and machine attribute values; (4) Update the weighting factor using the gradient descent method: 。

Images