Method and device for eliminating large model illusion in structured prediction task and medium
By selecting the large model set in the structured prediction task and initializing the discriminant filtering engine, the output results of the large model are classified and reversely disassembled, the problem of inaccurate prediction results caused by the large model hallucination phenomenon is solved, and higher execution stability and accuracy are achieved.
Patent Information
- Application Number
- CN202411708934.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-27
- Publication Date
- 2025-03-18
AI Technical Summary
In structured prediction tasks, the phenomenon of large-model hallucinations leads to inaccurate prediction results, and existing methods are difficult to completely avoid hallucinations, especially in multi-type structured prediction tasks, where real-time and generalization capabilities are insufficient.
By selecting the large model set, the weight parameters of each major model are obtained and the discriminative filtering engine is initialized, the preset large model processing instruction prompt words are constructed, the initial state and output of the large model are standardized, the pending target tasks are input to the large model set for processing, and the output results of each major model are input to the discriminative filtering engine for classification. The reliable metric values are obtained based on the classification results, and the contribution of each classification category and/or each model is reversely disassembled to obtain the contribution of each classification category and/or each model, and the execution results of the target task are determined and output.
Effectively eliminate the illusion of big models, improve the execution stability and robustness of structured prediction tasks, improve the accuracy and efficiency of task execution, and enhance the scientificity and accuracy of automated judgment and screening of credible prediction results.
Smart Images
Figure CN119647592A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of computer technology applications, and specifically to a method, device and medium for eliminating large model hallucinations in structured prediction tasks. Background Art
[0002] Large language models obtain model parameters by learning statistical features in training corpus, so large models may mistake statistical relationships for logical relationships, thereby generating fictitious or inaccurate content. This is the large model hallucination problem. This phenomenon is common in various natural language processing fields and is one of the important challenges faced by the development and application of generative models, especially for structured prediction tasks with closed spaces such as sequence segmentation, sequence labeling, entity extraction, and semantic role labeling. The accuracy of the prediction results is very high, and uncontrollable hallucinations will bring very serious consequences to the tasks. There are many reasons for the occurrence of hallucinations. First, the data distribution is inconsistent, and there are large differences between training data and application data; second, the model training freedom is too high or overfitting occurs, especially the generation of text content has no strict constraints, resulting in falsely high indicators; third, the noise pollution of the training data causes the model to learn wrong knowledge.
[0003] At present, scholars have proposed some solutions from the model training stage and model generation stage. The methods in the training stage mainly include data quality improvement, introduction of contrastive learning and multi-task learning, etc. These methods focus on the optimization of the intrinsic parameters of the model; the prevention in the generation stage mainly includes retrieval enhanced generation (RAG), evidence-based generation, uncertainty quantification, etc. These methods focus on improving the verifiability and reliability of the output by enhancing the knowledge base of the model.
[0004] However, these methods are still unable to completely avoid hallucinations, especially in multi-type structured prediction tasks (text classification, sequence labeling, etc.). Existing methods lack real-time and generalization capabilities and are difficult to use quickly in practical applications. Retraining the model can only cover and solve the hallucinations currently discovered by the trainer. Due to the high training cost, this approach cannot sustainably improve the hallucination phenomenon. The method of introducing external knowledge often limits the analytical reasoning ability of large models, sacrificing flexibility in exchange for the accuracy of the answer. Moreover, this type of method also has high requirements for additional data. Summary of the invention
[0005] Based on this, this application proposes a solution to eliminate the large model hallucination in structured prediction tasks to address the above problems, aiming to eliminate the large model hallucination in structured prediction tasks while improving the stability and robustness of task execution.
[0006] On the one hand, the present application provides a method for eliminating large model hallucination in a structured prediction task, the method comprising:
[0007] Select a large model set according to the type of target task, obtain the weight parameters of each large model in the set and initialize the discrimination and screening engine;
[0008] Construct preset large model processing instruction prompt words to standardize the initial state and output of the large model;
[0009] Input the target tasks to be processed into a large model set for processing, and input the output results of each large model into the discrimination and screening engine for classification;
[0010] Obtain the trustworthy measurement value of the current target task execution based on the classification result;
[0011] If the credibility metric value meets the preset conditions, reverse decomposition is performed based on the credibility metric value to obtain the contribution of each classification category and / or each model;
[0012] Based on the contribution of each classification category and / or each model, the execution result of the target task is determined and output.
[0013] Preferably, the step of obtaining the weight parameters of each major model in the set and initializing the discrimination and screening engine includes:
[0014] If the selected large model has executed the target task type, the weight parameters of the corresponding large model are retrieved from the historical target task discrimination and screening engine;
[0015] Otherwise, the weight parameters of the selected large model are cold-started;
[0016] The weight parameters of each major model in the obtained set are input into the discrimination and screening engine for initialization.
[0017] Preferably, the credibility metric value satisfies:
[0018]
[0019] Among them, T is the trust metric value of the current target task processing, M is the number of categories classified by each model, and p(c i ) is the probability of belonging to the i-th category, is the weight of the i-th category, satisfying:
[0020]
[0021] Among them, ω j is the jth large model weight, N is the number of selected large models, l(x j =c i ) is a piecewise function: xj is the execution result of the jth large model, c i is the result of the i-th category.
[0022] Furthermore, the method further comprises:
[0023] Compare the trustworthiness metric value T of the current target task execution with the preset threshold value T0 of the current task;
[0024] If T>T0, it is determined that the processing results of the various models in the current round are too dispersed, the credibility metric value does not meet the preset conditions, and the current processing data is marked as data to be tested;
[0025] If T≦T0, it is determined that the credibility metric value meets the preset condition.
[0026] Furthermore, the method further comprises:
[0027] Back-decomposition of the weighted self-information of each category and / or the weight of each model;
[0028] Sort the weighted self-information of each category of the classification results;
[0029] If there is only one category C with minimum weighted self-information minI’ , then C minI’ The corresponding result will be used as the execution result of the target task.
[0030] Preferably, the method further comprises:
[0031] If there is more than one minimum weighted self-information, sort the weight factors of each model in the minimum weighted self-information category;
[0032] Traverse each weight factor from large to small, and obtain the execution result corresponding to the first large model with different weight factors as the execution result of the target task.
[0033] Furthermore, the method further comprises:
[0034] According to the results of each round of prediction target task, the prediction accuracy of each model involved in the task is recorded;
[0035] After running a preset scale round, the judgment results are integrated and optimized with the weight parameters of the current engine, and the optimized parameters are updated to the judgment and screening engine before the execution of a new round of tasks.
[0036] A second aspect of the present application also provides a device for eliminating large model hallucinations in structured prediction tasks, the device comprising.
[0037] The engine initialization unit is used to select a large model set according to the type of target task and obtain the weight parameters of each large model in the set to initialize the discrimination and screening engine;
[0038] The prompt word processing unit is used to construct the instruction prompt words for natural language processing of the large model and standardize the initial state and output of the large model;
[0039] The large model processing unit is used to input the target text corpus to be processed into the large model set for processing, and input the output results of each large model into the discrimination and screening engine for classification;
[0040] A computing processing unit, used to obtain a trustworthy metric value of the current target task processing based on the classification result; if the trustworthy metric value meets the preset conditions, reverse decomposition is performed based on the trustworthy metric value to obtain the contribution of each classification category and / or each model;
[0041] The result output unit determines and outputs the execution result of the target task based on the contribution of each classification category and / or each model.
[0042] The third aspect of the present application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of any one of the above methods.
[0043] A fourth aspect of the present application provides a computer terminal device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of any one of the above methods.
[0044] The solution for eliminating large model hallucinations in structured prediction tasks provided in the present application configures a model set parallel processing mechanism according to the task attribute category, merges the multi-channel large model results into the discrimination screening engine for quantitative evaluation, and controls the output expectations by adjusting the engine parameters and preset thresholds, so that the execution results of the task are more in line with actual needs. By utilizing the randomness of the large model hallucination itself, in the task results where hallucinations occur, especially in closed set structured prediction tasks, this dispersion can be effectively identified, forming an automated error correction capability for large model hallucinations, and reducing the adverse effects of hallucinations on prediction tasks.
[0045] Furthermore, a trust metric calculation and result screening optimization scheme based on information entropy is provided. Information entropy can be used to better screen out features related to the target problem and improve the efficiency of task execution. After the result set obtained in each round of tasks, the historical execution status of the model of each task and the execution results of the current round are organically quantified and integrated, providing a quality evaluation basis for the execution results of each round of tasks, eliminating the influence of hallucinations that may be caused by some models, and improving the accuracy of task execution results. Furthermore, the scheme of the present application can be used to evaluate the concentration and dispersion of the result set, improve the scientificity and accuracy of automated judgment and screening of trustworthy prediction results, and enhance the influence of high-quality models on the discrimination and screening engine by separately evaluating the prediction performance of the model in different tasks, thereby improving the stability and robustness of task execution, thereby forming a sustainable hallucination elimination capability. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0047] in:
[0048] Figure 1 A flowchart of a method for eliminating large model hallucinations in a structured prediction task in one embodiment;
[0049] Figure 2 A schematic diagram of the overall architecture of a method for eliminating large model hallucinations in a structured prediction task in one embodiment;
[0050] Figure 3 A schematic diagram of an architecture for updating the parameters of a discrimination and screening engine based on the target task execution result in one embodiment;
[0051] Figure 4 A structural block diagram of a device for eliminating large model hallucinations in a structured prediction task in one embodiment;
[0052] Figure 5 FIG. 4 is a structural block diagram of a computer device in one embodiment. DETAILED DESCRIPTION
[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0054] The terms "include", "comprising" and "having" and any variations thereof in the specification and claims of the present application and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device comprising a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or devices. In the claims, specification and drawings of the present application, relational terms such as "first" and "second" are merely used to distinguish one entity / operation / object from another entity / operation / object, and do not necessarily require or imply any such actual relationship or order between these entities / operations / objects.
[0055] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The occurrence of the phrase at various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.
[0056] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.
[0057] The embodiment of the present application provides a method for eliminating large model hallucinations in a closed set structured prediction task, which is performed by an electronic device, which may be a server or a terminal device. The server may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. The terminal device may be a smart phone, a tablet computer, a laptop computer, a desktop computer, etc., but is not limited thereto. The server and the terminal device may be directly or indirectly connected via wired or wireless communication, which is not limited in the embodiment of the present application.
[0058] In one embodiment, if Figure 1The flowchart of the method for eliminating large model hallucination in a structured prediction task of the present application is shown as follows: Figure 2 A schematic diagram of the overall architecture of a method for eliminating large model hallucinations in a structured prediction task in one embodiment of the present application, the method comprising:
[0059] S10. Select a large model set according to the type of target task, obtain the weight parameters of each large model in the set and initialize the discrimination and screening engine.
[0060] Specifically, a large model set of a corresponding type is selected according to the type of the target task to be processed, and the large model set includes multiple large models that can process the target type task. After the corresponding large model set is selected, the weight parameters of each large model in the set are obtained and the discrimination and screening engine is initialized. Preferably, the method of obtaining the weight parameters of each large model in the set and initializing the discrimination and screening engine includes:
[0061] S101. If the selected large model has executed the target task type, the weight parameters of the corresponding large model are retrieved from the historical target task discrimination and screening engine.
[0062] S102. Otherwise, cold start assignment is performed on the weight parameters of the selected large model.
[0063] S103, input the weight parameters of each major model in the obtained set into the discrimination and screening engine for initialization.
[0064] For example, in one embodiment, the structured prediction task is to segment classical Chinese texts. Most ancient Chinese books do not have punctuation marks, and segmentation and punctuation are important tasks in proofreading ancient books. Correct segmentation helps spread and popularize excellent ancient books, while incorrect segmentation may affect reading and even lead to a misunderstanding of the author's original intention. Therefore, using a large model to segment classical Chinese texts requires a high accuracy rate.
[0065] N models are selected to participate in the task, one of which (assuming it is the Nth one) has never performed the classical Chinese sentence segmentation task, and the remaining N-1 have performed the classical Chinese sentence segmentation task. The weights of models 1 to N-1 are obtained from the historical execution data saved in the current task discrimination and screening engine, and are recorded as ω1,...,ω N-1 There are many ways to cold-start the weight of model N. In this embodiment, the average of the weights of the first N-1 models is taken as an example, that is, These weight parameters are then input into the discriminant screening engine for initialization.
[0066] S11. Construct instruction prompts for large model processing and standardize the initial state and output of the large model.
[0067] Specifically, prompts are a method specifically designed to optimize language models. They aim to guide these models to generate more accurate and targeted output results by designing and adjusting input prompts.
[0068] Taking the above-mentioned classical Chinese sentence segmentation as an example, according to the specific requirements of the classical Chinese sentence segmentation task, a corresponding large model prompt word is constructed, and the output format is limited to facilitate the summary and analysis of the output results. For example, the output is "use commas as sentence segmentation markers."
[0069] S12. Input the target task to be processed into the large model set for processing, and input the output results of each large model into the discrimination and screening engine for classification.
[0070] Specifically, the target task to be processed is input into each major model of the large model set and processed in combination with the above-mentioned prompt words, and the execution results of each major model are collected, and the output results of each major model are input into the discrimination and screening engine for classification. In the discrimination and screening engine, the result set of the model output is first integrated and then automatically clustered using a preset algorithm, and models with consistent output results are classified into one category to obtain multiple category results.
[0071] For example, in the above-mentioned sentence segmentation task, N large models will output N replies of the sentence segmentation results, recorded as A1, ..., AN. These N reply results are collected or recorded by automated means, and the output results of each large model are input into the discrimination and screening engine for classification. The collected N replies A1, ..., AN are clustered, and the completely identical ones are classified into the same category, resulting in M categories.
[0072] S13. Obtain a trustworthy measurement value of the current target task execution based on the classification result.
[0073] Specifically, the trustworthiness metric of the current target task execution can be measured and characterized based on different physical parameters and information parameters. Preferably, in order to better eliminate the influence of hallucinations that may be produced by some models, improve the accuracy of task execution results, and better screen out features related to the target problem, and improve the efficiency of task execution, in an example of the present application, the trustworthiness metric for measuring the execution of the current target task is obtained based on information entropy calculation, and the trustworthiness metric includes:
[0074]
[0075] Among them, T is the trust metric value of the current target task processing, M is the number of categories classified by each model, and p(c i ) is the probability of belonging to the i-th category, is the weight of the i-th category, satisfying:
[0076]
[0077] Among them, ω j is the jth large model weight, N is the number of selected large models, l(x j =c i ) is a piecewise function: x j is the execution result of the jth large model, c i is the result of the i-th category.
[0078] The above-mentioned implementation method of a trustworthy metric value provided in the present application uses the performance of the model in historical tasks as a weight factor for calculating the self-information of the category, which can effectively take into account the relationship between the model performance and the category of the task in this round, so that the historical execution status of the model of each task and the execution results of the current round are organically quantified and integrated, providing a quality evaluation basis for the execution results of each round of tasks, eliminating the influence of hallucinations that may be caused by some models, and improving the accuracy of task execution results.
[0079] S14: If the credibility metric value meets the preset conditions, reverse decomposition is performed based on the credibility metric value to obtain the contribution of each classification category and / or each model.
[0080] Specifically, after the result set output by the model is integrated and automatically classified, the trust metric value of the task is obtained according to the classification result, and the trust metric value is compared with the preset threshold value. According to the comparison result, it is determined whether the trust metric value of the current target task execution meets the preset conditions and whether it meets the expectations, which specifically includes:
[0081] S141, comparing the credibility metric value T of the current target task execution with the current task preset threshold value T0;
[0082] S142, if T>T0, it is determined that the processing results of the various models in the current round are too dispersed, the credibility metric value does not meet the preset conditions, and the current processing data is marked as data to be tested;
[0083] S143: If T≦T0, it is determined that the credibility metric value meets the preset conditions, and the contribution of each category and model to the credibility metric value is reversely decomposed according to the classification result.
[0084] The above scheme of this application compares the trust metric value T calculated from the current target task execution result with the preset threshold value T0 of this task. When T>T0, it means that the results of the responses of the various models in this round of tasks are too scattered, and the task execution result is not ideal. The target task data of this round is marked as the data to be tested and executed in the next round or after the engine parameters are optimized, and the program continues to execute the next target task data.
[0085] When T<=T0, it is judged that the trust metric value meets the preset conditions, indicating that the result of this round of task execution may be able to obtain the final result, and the calculation parameters of the trust metric value T of the current target task execution are output to the category screening engine, and further according to the classification results, the respective contributions to the trust metric value are reversely disassembled according to the category and model, and the current task execution result is further determined by the reverse disassembly of the respective contributions to the trust metric value.
[0086] S15. Based on the contribution of each classification category and / or each model, determine and output the execution result of the target task.
[0087] Specifically, the weight of each category in the present application is composed of the weight of a large model whose output result is the same as ci. Therefore, in one embodiment of the present application, the execution result of the current target task is judged based on the weighted self-information of each category and / or the weight of each model as the contribution, specifically including:
[0088] S151. Reversely decompose the weighted self-information of each category and / or the weight of each model.
[0089] Specifically, according to the calculation formula of the credibility metric value T, the weighted self-information of each category is reversely decomposed And further reversely decompose the weight ω corresponding to each model j .
[0090] S152. Sort the weighted self-information of each category of the classification results.
[0091] Specifically, after calculating the weighted information of each category according to the weighted self-information formula, the weighted self-information of each category (such as the M categories) of the current task result classification result is sorted from large to small, or from small to large.
[0092] S153. If there is only one category C with minimum weighted self-information minI’ , then C minI’ The corresponding result will be used as the execution result of the target task.
[0093] S154. If there is more than one minimum weighted self-information, sort the weight factors of each model in the minimum weighted self-information category, traverse each weight factor from large to small, and obtain the execution result corresponding to the first large model with a different weight factor as the execution result of the target task.
[0094] Specifically, if there is more than one minimum weighted self-information, the model weight factors whose output results are these minimum weighted self-information categories are sorted. If there is one and only one weight factor among these models that is the largest, then the result of the model with the largest weight factor or the result of its category is used as the execution result of this round of tasks; otherwise, continue to traverse and if there is one and only one weight factor among all models that is the second largest, then the result of the model with the second largest weight factor or the result of its category is used as the execution result of this round of tasks, and so on.
[0095] Furthermore, in one embodiment, the solution of the present application further includes updating the discrimination and screening engine according to each prediction target task. Specifically, updating the discrimination and screening engine further includes:
[0096] S16. Record the prediction accuracy of each model involved in the task execution based on the results of each round of prediction target task;
[0097] S17. After the preset scale round runs, the judgment results are integrated and optimized with the weight parameters of the current engine, and the optimized parameters are updated to the judgment and screening engine before the execution of a new round of tasks.
[0098] In one embodiment, Figure 3 The figure shows a schematic diagram of the architecture for updating the parameters of the discrimination and screening engine based on the target task execution results of the present application, wherein the judgment results are integrated with the weight parameters of the current engine, and the optimized parameters are updated to the discrimination and screening engine before the execution of a new round of tasks, including:
[0099] Define reward and penalty factors according to task attributes, accumulate the number of successful judgments made by each model in the update cycle, weight them with reward factors, and get the reward coefficient of each model; accumulate the number of wrong judgments made by each model in the update cycle, weight them with penalty factors, and get the penalty coefficient of each model. Recalculate the engine weight parameters based on the reward coefficient and penalty coefficient of each model in the update cycle and the weight parameters of the corresponding model in the engine.
[0100] Specifically, after the above steps, the difference between the execution results of each round of each model in the model set and the task results can be obtained. After the task is run for P rounds, the results are recorded as a matrix R∈R^(P×N), where consistent records are 1 and inconsistent records are -1. The reward factor is defined as λ Y , the penalty factor is λ w , then the weight factor of the model can be updated as follows:
[0101]
[0102] The above-mentioned embodiment of the present application provides an implementation method for determining and filtering engine parameter updates, defines reward and penalty factors according to task attributes, and jointly considers the model within the update cycle and historical performance. This can more reasonably evaluate the performance of each model in the large model set, and adjust the engine parameters in a timely manner to improve the quality of task execution.
[0103] In one embodiment, Figure 4 As shown, it is a structural block diagram of a device for eliminating large model hallucination in a structured prediction task provided by the present application, and the device includes.
[0104] The engine initialization unit is used to select a large model set according to the type of target task and obtain the weight parameters of each large model in the set to initialize the discrimination and screening engine;
[0105] The prompt word processing unit is used to construct the instruction prompt words for natural language processing of the large model and standardize the initial state and output of the large model;
[0106] The large model processing unit is used to input the target text corpus to be processed into the large model set for processing, and input the output results of each large model into the discrimination and screening engine for classification;
[0107] A computing processing unit, used to obtain a trustworthy metric value of the current target task processing based on the classification result; if the trustworthy metric value meets the preset conditions, reverse decomposition is performed based on the trustworthy metric value to obtain the contribution of each classification category and / or each model;
[0108] The result output unit determines and outputs the execution result of the target task based on the contribution of each classification category and / or each model.
[0109] In one embodiment, the present application further provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor performs the following steps:
[0110] Select a large model set according to the type of target task, obtain the weight parameters of each large model in the set and initialize the discrimination and screening engine;
[0111] Construct preset large model processing instruction prompt words to standardize the initial state and output of the large model;
[0112] Input the target tasks to be processed into a large model set for processing, and input the output results of each large model into the discrimination and screening engine for classification;
[0113] Obtain the trustworthy measurement value of the current target task execution based on the classification result;
[0114] If the credibility metric value meets the preset conditions, reverse decomposition is performed based on the credibility metric value to obtain the contribution of each classification category and / or each model;
[0115] Based on the contribution of each classification category and / or each model, the execution result of the target task is determined and output.
[0116] In one embodiment, Figure 5 The present application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following steps:
[0117] Select a large model set according to the type of target task, obtain the weight parameters of each large model in the set and initialize the discrimination and screening engine;
[0118] Construct preset large model processing instruction prompt words to standardize the initial state and output of the large model;
[0119] Input the target tasks to be processed into a large model set for processing, and input the output results of each large model into the discrimination and screening engine for classification;
[0120] Obtain the trustworthy measurement value of the current target task execution based on the classification result;
[0121] If the credibility metric value meets the preset conditions, reverse decomposition is performed based on the credibility metric value to obtain the contribution of each classification category and / or each model;
[0122] Based on the contribution of each classification category and / or each model, the execution result of the target task is determined and output.
[0123] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).
[0124] The technical features of the above embodiments may be combined arbitrarily. 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.
[0125] The above-mentioned 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 method for eliminating large model hallucination in structured prediction tasks, characterized in that: The method comprises: Select a large model set according to the type of target task, obtain the weight parameters of each large model in the set and initialize the discrimination and screening engine; Construct preset large model processing instruction prompt words to standardize the initial state and output of the large model; Input the target tasks to be processed into a large model set for processing, and input the output results of each large model into the discrimination and screening engine for classification; Obtain the trustworthy measurement value of the current target task execution based on the classification result; If the credibility metric value meets the preset conditions, reverse decomposition is performed based on the credibility metric value to obtain the contribution of each classification category and / or each model; Based on the contribution of each classification category and / or each model, the execution result of the target task is determined and output.
2. The method according to claim 1, characterized in that The step of obtaining the weight parameters of each major model in the set and initializing the discrimination and screening engine includes: If the selected large model has executed the target task type, the weight parameters of the corresponding large model are retrieved from the historical target task discrimination and screening engine; Otherwise, the weight parameters of the selected large model are cold-started; The weight parameters of each major model in the obtained set are input into the discrimination and screening engine for initialization.
3. The method according to claim 1, characterized in that The trustworthiness metric satisfies: Among them, T is the trust metric value of the current target task processing, M is the number of categories classified by each model, and p(c i ) is the probability of belonging to the i-th category, is the weight of the i-th category, satisfying: Among them, ω j is the weight of the jth large model, N is the number of selected large models, l(x j =c i ) is a piecewise function: x j is the execution result of the jth large model, c i is the result of the i-th category.
4. The method according to claim 1, characterized in that: The method further comprises: Compare the trustworthiness metric value T of the current target task execution with the preset threshold value T0 of the current task; If T>T0, it is determined that the processing results of the various models in the current round are too dispersed, the credibility metric value does not meet the preset conditions, and the current processing data is marked as data to be tested; If T≦T0, it is determined that the credibility metric value meets the preset condition.
5. The method according to claim 4, characterized in that The method further comprises: Back-decomposition of the weighted self-information of each category and / or the weight of each model; Sort the weighted self-information of each category of the classification results; If there is only one category C with minimum weighted self-information minI’ , then C minI’ The corresponding result will be used as the execution result of the target task.
6. The method according to claim 5, characterized in that The method further comprises: If there is more than one minimum weighted self-information, sort the weight factors of each model in the minimum weighted self-information category; Traverse each weight factor from large to small, and obtain the execution result corresponding to the first large model with different weight factors as the execution result of the target task.
7. The method according to any one of claims 1 to 6, characterized in that: The method further comprises: According to the results of each round of prediction target task, the prediction accuracy of each model involved in the task is recorded; After running a preset scale round, the judgment results are integrated and optimized with the weight parameters of the current engine, and the optimized parameters are updated to the judgment and screening engine before the execution of a new round of tasks.
8. A device for eliminating large model hallucination in structured prediction tasks, characterized in that: The device comprises: The engine initialization unit is used to select a large model set according to the type of target task, obtain the weight parameters of each large model in the set and initialize the discrimination and screening engine; A prompt word processing unit, used to construct a preset large model processing instruction prompt word, and standardize the initial state and output of the large model; The large model processing unit is used to input the target task to be processed into the large model set for processing, and input the output results of each large model into the discrimination and screening engine for classification; A computing processing unit, used to obtain a trustworthy metric value of the current target task processing based on the classification result; if the trustworthy metric value meets the preset conditions, reverse decomposition is performed based on the trustworthy metric value to obtain the contribution of each classification category and / or each model; The result output unit determines and outputs the execution result of the target task based on the contribution of each classification category and / or each model.
9. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 7.
10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 7.