Intelligent correcting method and system for correcting paper homework

Abstract

The invention discloses an intelligent correction method and system for correcting paper homework, and the method comprises the steps: constructing a knowledge base according to the examination line content and course knowledge, and obtaining the scanning data of the to-be-corrected paper homework from the outside; performing hyper-parameter fine tuning on the plurality of large models, adding role setting of a corrector to the large models in training, and using an RAG retrieval enhancement technology according to the constructed knowledge base; correcting the paper homework scanning data by using the fine-tuned large models, checking the consistency of a plurality of large models for generated answers by using an LLM confrontation communication method, and performing multiple rounds of challenges to obtain a correction result meeting a consistency requirement; and integrating a final correction result according to the correction results of the plurality of large models, and feeding back the final correction result to the paper job through the printer. According to the method, based on specially designed LLM countermeasure communication, RAG related technologies are integrated, and the model discrimination accuracy is greatly improved; and during exporting, a hardware printing exporting mode is adopted, the method is simple and direct, and examination paper marking correction can be realized.

Description

Technical Field

[0001] The present invention relates to the field of educational technology, and in particular to an intelligent marking method and system for marking paper homework. Background Art

[0002] With the continuous advancement of educational informatization, intelligent homework grading systems have gradually become one of the key tools for teachers to assist teaching. Such systems aim to improve the efficiency of homework grading through automation and reduce teachers' repetitive work. However, existing intelligent grading systems still have many limitations in practical applications and are difficult to meet the needs of modern education for efficient, accurate and diversified teaching feedback.

[0003] The current intelligent homework grading systems are mainly divided into two categories: platform-based grading systems and OCR (optical character recognition) + question bank-based grading systems. Platform-based systems rely on users to upload photos or texts of questions and rely on the answers of other users in the community or platform to complete homework grading. This type of system has certain application value in some open education scenarios, but it cannot operate efficiently in daily large-scale teaching scenarios. OCR+question bank-based grading systems are more widely used in primary and secondary education. This type of system uses OCR technology to identify paper homework submitted by students and provide answers based on the cloud-based question bank. However, the current OCR+question bank model has significant limitations:

[0004] Recognition accuracy and scenario limitations: Although OCR technology has a certain application effect when recognizing standardized text, its recognition accuracy drops significantly when encountering complex question types (such as graphic questions, formula questions, handwritten answers, etc.). This not only affects the accuracy of the system, but also limits its application scenarios. In particular, for subjects such as mathematics and physics that contain a large number of symbols and graphics, OCR recognition often makes mistakes, resulting in unreliable grading results.

[0005] Limited coverage of question banks: Although the cloud question bank has a large number of questions, it is difficult to cover all self-made test papers and local teaching materials due to differences in teaching content between schools, regions and even classes. Especially in primary and secondary education, there are significant differences in the question banks used by different regions and schools, resulting in some questions being unable to be matched by the automated system, and teachers still need to manually correct them, which to some extent offsets the efficiency improvement of the system.

[0006] Disadvantages of relying on question banks: Current intelligent grading systems rely heavily on the existence of question banks. However, the content in question banks usually involves copyright issues of supplementary teaching materials or test questions. With the increasingly stringent protection of intellectual property rights, systems that rely too much on question banks may face copyright restrictions in the future and will not be able to continue to expand or update. Copyright disputes will not only limit the functionality of the system, but also have an adverse impact on the intelligent development of the entire education industry.

[0007] Lack of intelligent feedback and data analysis functions: The existing intelligent grading system focuses on automated grading, but lacks in-depth analysis and feedback mechanisms for students' learning situations. Traditional grading systems usually only provide grading results, and fail to deeply analyze students' error types, knowledge blind spots, etc., so teachers cannot make teaching adjustments based on grading results. In addition, the lack of integration with learning management systems (LMS) or other educational data systems means that the grading data cannot provide effective support for teaching.

[0008] Insufficient system scalability and adaptability: Existing systems are often designed only for standardized exams and assignments, and cannot cope with the diverse needs encountered by teachers in the actual teaching process. For example, for open-ended questions or innovative test questions, the existing OCR+question bank system cannot provide efficient automatic grading, and the system has poor scalability and cannot adapt to the complex requirements of different subjects and different teaching environments.

[0009] Based on the above problems, the current intelligent homework grading system still has a lot of room for optimization in terms of efficiency and functionality. With the rapid development of large-scale pre-trained language models (LLM) and multimodal models, how to apply these emerging technologies to the field of education, especially to solve complex homework grading problems, has become a hot topic of research. The new generation of intelligent grading systems should not only be able to handle more complex question types, but also integrate data analysis and feedback functions to support teachers to efficiently grasp students' learning dynamics and further optimize teaching strategies. Therefore, designing an intelligent grading system based on large models and multi-model optimization technology will provide a more efficient, accurate and comprehensive solution for the field of education, which has important practical significance and application prospects. Summary of the invention

[0010] The present invention aims to address the deficiencies of the prior art and to provide an intelligent marking method and system for marking paper assignments.

[0011] The objective of the present invention is achieved through the following technical solution: an intelligent marking method for marking paper homework, comprising the following steps:

[0012] S1. Build a knowledge base based on the syllabus content and course knowledge, and obtain scanned data of paper homework to be corrected from the outside world;

[0013] S2. Fine-tune the hyperparameters of several large models, add the role of reviewer to the large models during training, and use RAG retrieval enhancement technology based on the constructed knowledge base;

[0014] S3. Use the fine-tuned large model to correct the scanned data of paper homework, use the LLM adversarial communication method to check the consistency of several large models for the generated answers, and conduct multiple rounds of questioning to obtain the correction results that meet the consistency requirements;

[0015] S4. Integrate the final grading results based on the grading results of several large models and feed them back to the paper work through the printer.

[0016] Furthermore, the knowledge base is a special knowledge base covering the syllabus content and curriculum knowledge of primary and secondary schools, including the existing primary and secondary school question banks, and the data in the knowledge base is vectorized and stored in a database, so that relevant knowledge can be retrieved when users ask questions.

[0017] Furthermore, the hyperparameter fine-tuning includes: using hyperparameter tuning techniques such as grid search and Bayesian optimization, combined with experiments to determine the optimal value of the hyperparameter; this process can be expressed by the following formula:

[0018] λ * = arg min E[L(θ;λ)]

[0019] The loss function of the model is L(θ), where θ represents the parameters of the model; the loss function L is minimized by optimizing the hyperparameters; the optimal hyperparameter set λ is found through grid search and Bayesian optimization methods. * .

[0020] Furthermore, the adding of the role setting of the reviewer to the large model includes:

[0021] In the training, the role of the large model is explicitly specified as the teacher to improve the accuracy of the generated judgment; the role-specific loss term is introduced into the objective function of the model training to ensure that the model output is accurate and meets the needs of the education role. The formula is as follows:

[0022]

[0023] Where L(θ) is the loss function, D is the training dataset, y is the correct feedback, CrossEntropy is the cross entropy loss used to measure prediction accuracy, R(θ) is the regularization term, and λ is the adjustment factor.

[0024] Further, the RAG retrieval enhancement technology includes: enhancing the ability of the generation model by extracting relevant knowledge from the retrieval knowledge base, thereby improving the accuracy and relevance of the generated content;

[0025]

[0026] Among them, α k is retrieved from kK represents the total number of relevant contents retrieved from the knowledge base, and k is an index variable used to represent the kth content item {Retrieved}_k obtained from the retrieval base.

[0027] Furthermore, the generated answer is checked for consistency of several large models using the LLM adversarial communication method, including:

[0028] After each single model generates an answer, the answer to the same question from each individual model is obtained through multi-threading technology;

[0029] The LLM adversarial communication system is used to verify and optimize the answers: the answers are compared and judged through a joint check of semantic and logical consistency or a logical consistency check to obtain a consistency score for the generated answers.

[0030] Feedback and re-questioning: If the consistency score is lower than the preset threshold, the server opens a new thread for each model, sends the answers of all models including itself to each model through the new thread, and conducts a second question in a new thread. The questions include: scoring these answers and trying to generate better answers based on all the information currently obtained;

[0031] Repeat verification: The new answers of the model are checked for consistency again; if the consistency score still does not meet the preset standard, the feedback and re-questioning steps are repeated until the consistency score meets the requirement.

[0032] Furthermore, the logical consistency check includes: inputting all the correction results generated by the large model into the large model for questioning, obtaining the logical consistency score of the correction result output of each model relative to the correction result output of other models, and taking the average of all logical consistency scores as the consistency score of this round of questioning.

[0033] Furthermore, the semantic and logical consistency joint check includes: inputting all the correction results generated by the big model into the big model for questioning, obtaining the logical consistency score of the correction result output of each model relative to the correction result output of other models, and calculating the average of all logical consistency scores as the logical consistency score of this round of questioning;

[0034] The semantic consistency of each correction result relative to the correction results of other models is calculated, and the average of the semantic consistency scores is calculated as the semantic consistency score of this round of questioning; the consistency score of this round is constructed based on the logical consistency score, the semantic consistency score and the preset weights.

[0035] Furthermore, the calculation of the semantic consistency score is specifically to calculate the cosine similarity between the semantics, and the formula is as follows:

[0036]

[0037] Among them, A and B are the vector representations of the outputs of the two models, A·B represents the dot product of the two vectors, ||A|| and ||B|| represent the norms of vectors A and B respectively; if the similarity is higher than the set threshold, the semantics are considered to be consistent; check whether there are synonyms or similar expressions in the answer; verify whether the answer is consistent with the question background by matching key context information.

[0038] According to another aspect of the specification, there is also provided an intelligent grading system for grading paper assignments, the system comprising: a large model preprocessing module, a grading processing module, a data processing module and a communication module;

[0039] The large model preprocessing module fine-tunes the hyperparameters of several large models, adds the role setting of reviewers to the large models during training, and uses RAG retrieval enhancement technology based on the constructed knowledge base;

[0040] The correction processing module corrects the paper homework scan data using the fine-tuned large model;

[0041] The data processing module coordinates the answers from different models, uses the LLM adversarial communication method to check the consistency of several large models for the generated answers, and conducts multiple rounds of questioning to obtain the correction results that meet the consistency requirements;

[0042] The communication module is used to obtain scans of paper assignments to be corrected from the outside, and to integrate final correction results according to correction results of several large models, and to feed back the final correction results to the paper assignments through the printer.

[0043] Beneficial effects of the present invention:

[0044] It allows the use of automatic batch reading hardware devices similar to printers as a means of reading student answers, without the need for a large number of repeated shooting and importing operations; the current multimodal large models can recognize pictures, and the introduction of these models allows the system to correct questions with pictures, expanding the scope of automatic correction; the present invention makes full use of large model single model and multi-model optimization technology, based on the specially designed LLM adversarial communication system, and integrates RAG related technologies, which greatly improves the model discrimination accuracy; when exporting, the hardware direct printing export method is adopted, the method is simple and direct, and the test paper can be corrected with traces. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A flow chart of an intelligent marking method for marking paper assignments provided by an embodiment of the present invention;

[0046] Figure 2 A schematic diagram of an intelligent grading device for grading paper assignments provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the patent of the present invention to clearly and completely describe the technical solutions in the embodiments of the patent of the present invention. Obviously, the described embodiments are only part of the embodiments of the patent of the present invention, not all of them. Based on the embodiments in the patent 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 patent of the present invention.

[0048] An intelligent marking method for marking paper homework includes the following steps:

[0049] S1. Build a knowledge base based on the syllabus content and course knowledge, and obtain scanned data of paper homework to be corrected from the outside world;

[0050] This method prepares a special knowledge base covering the syllabus content and curriculum knowledge of primary and secondary schools, including the existing primary and secondary school question bank, and vectorizes the data in the knowledge base and stores it in the database, and retrieves relevant knowledge when the user asks a question.

[0051] The paper work scan data can be completed by scanning with a printer, taking photos with a camera, etc. In this embodiment, the test paper is placed in the paper tray, and the scanning function of the printer is used to automatically start scanning the test paper.

[0052] S2. Fine-tune the hyperparameters of several large models, add the role of reviewer to the large models during training, and use RAG retrieval enhancement technology based on the constructed knowledge base;

[0053] In this embodiment, the model uses the language large model Llama-3-70b-Groq, Claude-3.5, and the multimodal large model GPT-4o, Gemini.

[0054] Each model uses a single-model system optimization approach that includes:

[0055] Fine-tune hyperparameters. When learning a small amount of pre-trained large models, use hyperparameter tuning techniques such as grid search and Bayesian optimization, combined with experiments to determine the best values ​​of hyperparameters. Let the loss function of the model be L(θ), where θ represents the parameters of the model. Minimize the loss function L by optimizing the hyperparameters. Find the optimal hyperparameter set λ through methods such as grid search and Bayesian optimization * , the hyperparameter set λ * The main parameters include learning rate η, number of iterations T, and temperature τ. This process can be expressed by the following formula:

[0056] λ * =argminE[L(θ;λ)]

[0057] In this system, the general learning rate is at the 1e-6 level, the number of iterations is about 500, and the temperature value is 0.3; even if adjusted to other parameter combinations, the system is still within the protection range.

[0058] GPT role setting. In training, the large model role is explicitly specified as the teacher to improve the accuracy of generated judgments. Role-specific loss terms are introduced into the objective function of model training to ensure that the model output is accurate and meets the needs of the education role. The formula is as follows:

[0059]

[0060] Where L(θ) is the loss function, D is the training dataset, y is the correct feedback, CrossEntropy is the cross entropy loss used to measure prediction accuracy, R(θ) is the regularization term, and λ is the adjustment factor;

[0061] RAG retrieval enhancement technology. This technology is a machine learning method that combines retrieval and generation. Its main function is to enhance the ability of the generation model by retrieving relevant information, thereby improving the accuracy and relevance of the generated content. It is particularly suitable for applications that need to rely on specific knowledge base information. It can effectively extract relevant knowledge from the knowledge base to assist in homework correction. The formula is as follows:

[0062]

[0063] Among them, α k is retrieved from k The conversion weight to the generated content, K represents the total number of relevant content retrieved from the knowledge base, and k is an index variable used to represent the kth content item {Retrieved}_k obtained from the retrieval base.

[0064] To improve the accuracy of single model judgment. Parallel processing uses thread pool technology to speed up parallel processing and manage threads reasonably.

[0065] S3. Use the fine-tuned large model to correct the scanned data of paper homework, use the LLM adversarial communication method to check the consistency of several large models for the generated answers, and conduct multiple rounds of questioning to obtain the correction results that meet the consistency requirements;

[0066] Specifically, after each single model generates an answer, we coordinate the answers from different models and use a specially designed LLM adversarial communication system to improve the accuracy of the generated content, including:

[0067] S3.1. After each single model generates an answer, the answer of each individual model to the same question is obtained through multi-threading technology;

[0068] S3.2. Use the LLM adversarial communication system to verify and optimize the answer: compare and judge the answer through the joint check of semantic and logical consistency or logical consistency check to obtain the consistency score of the generated answer;

[0069] The logical consistency check includes: inputting all the correction results generated by the big model into the big model for questioning, obtaining the logical consistency score of the correction result output of each model relative to the correction result output of other models, and taking the average of all logical consistency scores as the consistency score of this round of questioning. If the consistency score is higher than the preset threshold (for example, 90 points), the answer is considered correct and directly adopted.

[0070] The semantic and logical consistency joint check includes: inputting all the correction results generated by the big model into the big model for questioning, obtaining the logical consistency score of the correction result output of each model relative to the correction result output of other models, and calculating the average of all logical consistency scores as the logical consistency score of this round of questioning;

[0071] The semantic consistency of each correction result relative to the correction results of other models is calculated, and the average of the semantic consistency scores is calculated as the semantic consistency score of this round of questioning; the consistency score of this round is constructed based on the logical consistency score, the semantic consistency score and the preset weight. If the consistency score is higher than the preset threshold (for example, 90 points), the answer is considered correct and is directly adopted.

[0072] The semantic consistency score uses a word vector or sentence vector model to calculate the semantic similarity between answers. The method for determining semantic similarity is cosine similarity, and the formula is as follows:

[0073]

[0074] Where A and B are the vector representations (word vectors or sentence vectors) output by the two models, A·B represents the dot product of the two vectors, and ||A|| and ||B|| represent the norms of vectors A and B (i.e., the lengths of the vectors). If the similarity is higher than the set threshold, the semantics are considered consistent. Check whether there are synonyms or similar expressions in the answer. Verify whether the answer is consistent with the question context by matching key context information.

[0075] S3.3, Feedback and re-questioning: If the consistency score is lower than the preset threshold, the server opens a new thread for each model, sends the answers of all models including itself to each model through the new thread, and conducts a second question in a new thread. The questions include: scoring these answers and trying to generate better answers based on all the information currently obtained;

[0076] The mathematical representation of the above process is as follows: a new thread is opened for each model, and the answer f is summarized summary (x) and the outputs of other models f1(x),f2(x),…,f n (x) Feedback is given to each model for secondary query. The input content received by each model is:

[0077]

[0078] That is, each model sees the answers of other models, as well as the combined answers, and then regenerates a new output f i new (x).

[0079] S3.4. Repeat verification: Check the consistency of the new answers of the model again; if the consistency score still does not meet the preset standard, repeat the feedback and re-questioning steps until the consistency score meets the requirement.

[0080] If the score is higher than the scoring threshold, the answer is adopted and the model is deemed to have reached a consensus. If it is still impossible to judge, the query operation is repeated. In this method, it is performed three times in total to achieve a balance between efficiency and accuracy. If it is still impossible to judge after three times, the highest scoring result is adopted by default.

[0081] S4. Integrate the final grading results based on the grading results of several large models and feed them back to the paper work through the printer.

[0082] The revision results are sent back to the client. The revision results are provided in two forms:

[0083] 1. A pdf version with only marking traces (such as hooks and crosses. If the school has special needs, the system can manually change the marking traces). This version is suitable for direct printing on student test papers to achieve marking traces.

[0084] 2. A PDF with scanned student papers and corrections. This version is suitable for some special needs, such as retaining the original papers. These two versions can be sent to teachers as electronic versions or sent directly to a printer for batch printing. Teachers can choose which mode to use for correction, thus achieving customizable correction with traces.

[0085] Corresponding to the aforementioned embodiment of an intelligent grading method for grading paper assignments, the present invention also provides an embodiment of an intelligent grading system for grading paper assignments.

[0086] An intelligent marking system for marking paper homework, comprising: a large model preprocessing module, a marking processing module, a data processing module and a communication module;

[0087] The large model preprocessing module fine-tunes the hyperparameters of several large models, adds the role setting of reviewers to the large models during training, and uses RAG retrieval enhancement technology based on the constructed knowledge base;

[0088] The correction processing module corrects the paper homework scan data using the fine-tuned large model;

[0089] The data processing module coordinates the answers from different models, uses the LLM adversarial communication method to check the consistency of several large models for the generated answers, and conducts multiple rounds of questioning to obtain the correction results that meet the consistency requirements;

[0090] The communication module is used to obtain scans of paper assignments to be corrected from the outside, and to integrate final correction results according to correction results of several large models, and to feed back the final correction results to the paper assignments through the printer.

[0091] This system is deployed through AWS cloud services. EC2 instances are used on AWS for main application code deployment, computing, and running services; S3 is used to store static resources and data; Amazon RDS is used for database hosting in the cloud; and lambda is used for serverless computing to achieve automated tasks and event-driven function execution.

[0092] The high-definition image data obtained by scanning is received in the AWS cloud. A parallel processing method based on JAVA multi-threading technology is used to enable multiple models to process the same question and generate answers at the same time by starting multiple threads and sending requests to multiple models in parallel, thereby improving efficiency. Here we use thread pool technology to facilitate thread management and reuse. Even if the thread pool technology in this system is replaced with other parallel processing technologies, its functions and effects are still within the scope of protection.

[0093] In summary, the core of the present invention is to develop a full-process automation method and system that integrates automatic scanning, large-scale model correction, result generation and direct printing. At present, there is no similar system on the market that can realize the full automation of the entire correction process. The advantage of this system is that it can process test papers in batches. Users only need to put the test papers into the device, and the system can automatically complete scanning, correction and printing, and integrate the correction process efficiently and smoothly.

[0094] Corresponding to the aforementioned embodiment of an intelligent marking method for marking paper assignments, the present invention also provides an embodiment of an intelligent marking device for marking paper assignments.

[0095] See also Figure 2An embodiment of the present invention provides an intelligent grading device for grading paper assignments, including a memory and one or more processors. The memory stores executable codes. When the processor executes the executable codes, it is used to implement an intelligent grading method for grading paper assignments in the above embodiment.

[0096] An embodiment of an intelligent grading device for grading paper assignments provided by the present invention can be applied to any device with data processing capabilities, and the device with data processing capabilities can be a device or apparatus such as a computer. The device embodiment can be implemented through software, or through hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it is formed by the processor of any device with data processing capabilities in which it is located reading the corresponding computer program instructions in the non-volatile memory into the memory for execution. From a hardware perspective, if Figure 2 As shown in FIG. 1 , a hardware structure diagram of an intelligent marking device for marking paper assignments provided by the present invention is provided in any device with data processing capability, except Figure 2 In addition to the processor, memory, network interface, and non-volatile memory shown, any device with data processing capabilities in which the apparatus in the embodiments is located may also include other hardware, generally based on the actual functions of the device with data processing capabilities, which will not be described in detail.

[0097] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0098] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of the present invention. Ordinary technicians in this field can understand and implement it without paying creative work.

[0099] An embodiment of the present invention further provides a computer-readable storage medium on which a program is stored. When the program is executed by a processor, an intelligent marking method for marking paper assignments in the above embodiment is implemented.

[0100] The computer-readable storage medium may be an internal storage unit of any device with data processing capability described in any of the aforementioned embodiments, such as a hard disk or a memory. The computer-readable storage medium may also be an external storage device of any device with data processing capability, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. equipped on the device. Furthermore, the computer-readable storage medium may also include both an internal storage unit and an external storage device of any device with data processing capability. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capability, and may also be used to temporarily store data that has been output or is to be output.

[0101] The present invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, the intelligent marking method for marking paper assignments is implemented.

[0102] Those skilled in the art will readily appreciate other embodiments of the present application after considering the description and practicing the contents disclosed herein. The present application is intended to cover any modification, use or adaptation of the present application, which follows the general principles of the present application and includes common knowledge or customary techniques in the art that are not disclosed in the present application. The description and examples are intended to be exemplary only, and the true scope and spirit of the present application are indicated by the claims.

[0103] It should be understood that the above general description and the detailed description below are only exemplary and explanatory and cannot limit the present application. The present application is not limited to the precise structure described above and shown in the drawings, and various modifications and changes can be made without departing from the scope thereof. The scope of the present application is limited only by the attached claims.

Claims

1. An intelligent marking method for marking paper assignments, characterized in that: The following steps are involved: S1. Build a knowledge base based on the syllabus content and course knowledge, and obtain scanned data of paper homework to be corrected from the outside world; S2. Fine-tune the hyperparameters of several large models, add the role of reviewer to the large models during training, and use RAG retrieval enhancement technology based on the constructed knowledge base; S3. Use the fine-tuned large model to correct the scanned data of paper homework, use the LLM adversarial communication method to check the consistency of several large models for the generated answers, and conduct multiple rounds of questioning to obtain the correction results that meet the consistency requirements; S4. Integrate the final grading results based on the grading results of several large models and feed them back to the paper work through the printer.

2. The intelligent marking method for marking paper assignments according to claim 1, characterized in that: The knowledge base is a special knowledge base covering the syllabus content and curriculum knowledge of primary and secondary schools, including the existing primary and secondary school question banks, and the data in the knowledge base is vectorized and stored in the database, so that relevant knowledge can be retrieved when users ask questions.

3. The intelligent marking method for marking paper assignments according to claim 1, characterized in that: The hyperparameter fine-tuning includes: using hyperparameter tuning technology and combining experiments to determine the optimal value of the hyperparameter; this process can be expressed by the following formula: l * =argminE[L(θ;λ)] The loss function of the model is L(θ), where θ represents the parameters of the model; the loss function L is minimized by optimizing the hyperparameters; the optimal hyperparameter set λ is found through grid search and Bayesian optimization methods. * .

4. The intelligent marking method for marking paper assignments according to claim 1, characterized in that: The role setting of adding a reviewer to the large model includes: In the training, the role of the large model is explicitly specified as the teacher to improve the accuracy of the generated judgment; the role-specific loss term is introduced into the objective function of the model training to ensure that the model output is accurate and meets the needs of the education role. The formula is as follows: Where L(θ) is the loss function, D is the training dataset, y is the correct feedback, CrossEntropy is the cross entropy loss used to measure prediction accuracy, R(θ) is the regularization term, and λ is the adjustment factor.

5. The intelligent marking method for marking paper assignments according to claim 1, characterized in that: The RAG retrieval enhancement technology includes: enhancing the ability of the generation model by extracting relevant knowledge from the retrieval knowledge base, thereby improving the accuracy and relevance of the generated content; Among them, α k is the content retrieved from k The conversion weight to the generated content, K represents the total number of relevant content retrieved from the knowledge base, and k is an index variable used to represent the kth content item {Retrieved}_k obtained from the retrieval base.

6. The intelligent marking method for marking paper assignments according to claim 1, characterized in that: The generated answers are checked for consistency using the LLM adversarial communication method for several large models, including: After each single model generates an answer, the answer to the same question from each individual model is obtained through multi-threading technology; The LLM adversarial communication system is used to verify and optimize the answers: the answers are compared and judged through a joint check of semantic and logical consistency or a logical consistency check to obtain a consistency score for the generated answers. Feedback and re-questioning: If the consistency score is lower than the preset threshold, the server opens a new thread for each model, sends the answers of all models including itself to each model through the new thread, and conducts a second question in a new thread. The questions include: scoring these answers and trying to generate better answers based on all the information currently obtained; Repeat verification: The new answers of the model are checked for consistency again; if the consistency score still does not meet the preset standard, the feedback and re-questioning steps are repeated until the consistency score meets the requirement.

7. The intelligent marking method for marking paper assignments according to claim 6, characterized in that: The logical consistency check includes: inputting all the correction results generated by the large model into the large model for questioning, obtaining the logical consistency score of the correction result output of each model relative to the correction result output of other models, and taking the average of all logical consistency scores as the consistency score of this round of questioning.

8. The intelligent marking method for marking paper assignments according to claim 6, characterized in that: The semantic and logical consistency joint check includes: inputting all the correction results generated by the big model into the big model for questioning, obtaining the logical consistency score of the correction result output of each model relative to the correction result output of other models, and calculating the average of all logical consistency scores as the logical consistency score of this round of questioning; The semantic consistency of each correction result relative to the correction results of other models is calculated, and the average of the semantic consistency scores is calculated as the semantic consistency score of this round of questioning; the consistency score of this round is constructed based on the logical consistency score, the semantic consistency score and the preset weights.

9. The intelligent marking method for marking paper assignments according to claim 8, characterized in that: The calculation of the semantic consistency score is specifically to calculate the cosine similarity between semantics, and the formula is as follows: Among them, A and B are the vector representations of the outputs of the two models, A·B represents the dot product of the two vectors, ∥A∥ and ∥B∥ represent the norms of vectors A and B respectively; if the similarity is higher than the set threshold, the semantics are considered to be consistent; check whether there are synonyms or similar expressions in the answer; verify whether the answer is consistent with the question background by matching key context information.

10. An intelligent marking system for marking paper assignments, characterized in that: The system includes: a large model preprocessing module, a batch correction processing module, a data processing module and a communication module; The large model preprocessing module fine-tunes the hyperparameters of several large models, adds the role setting of reviewers to the large models during training, and uses RAG retrieval enhancement technology based on the constructed knowledge base; The correction processing module corrects the paper homework scan data using the fine-tuned large model; The data processing module coordinates the answers from different models, uses the LLM adversarial communication method to check the consistency of several large models for the generated answers, and conducts multiple rounds of questioning to obtain the correction results that meet the consistency requirements; The communication module is used to obtain scans of paper assignments to be corrected from the outside, and to integrate final correction results according to correction results of several large models, and to feed back the final correction results to the paper assignments through the printer.

Images