Intelligent question correcting method based on multi-modal large model

Abstract

The invention discloses an intelligent question correction method based on a multi-modal large model, and relates to the technical field of intelligent education, and the method mainly comprises the steps: firstly, carrying out the multi-modal recognition and region segmentation of an image containing a question and a handwritten answer, and extracting structured question data and answer data; inputting the structured data into a multi-modal large model trained by knowledge in the education field, and identifying knowledge points and scoring requirements of question examination and logic correctness and normalization of answers; based on an identification result, a personalized correction rule for mapping the knowledge points and the answer logic into specific score points is dynamically generated; automatically correcting the answers according to the rule and generating score details; and finally, outputting personalized feedback information containing error positioning and associated correction suggestions based on a correction result. According to the method, a complete technical chain from multi-modal identification understanding to dynamic rule generation and execution is constructed, so that efficient and accurate execution of the intelligent correction system is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent education technology, specifically to an intelligent question grading method based on a multimodal large model. Background Technology

[0002] In the current development of intelligent education technology, automatic question grading has become a key link in improving teaching efficiency. However, existing intelligent grading systems still have a series of core problems that urgently need to be solved. First, these systems generally lack the comprehensive processing capability for multimodal questions, making it difficult to effectively handle complex questions that integrate text, mathematical formulas, graphs, and handwriting. This results in insufficient grading coverage across all K-12 subjects, and often requires the development of dedicated modules for different subjects, leading to poor system reusability and high expansion costs. Second, the accuracy and depth of grading are seriously insufficient: in science questions, the system can usually only judge the correctness of the final result, but cannot assign tiered scores based on the completeness of the solution steps, the coherence of the logic, and the standardization of formula writing, resulting in missed and incorrect judgments; in humanities subjective questions, it relies too much on mechanical methods such as keyword matching, ignoring the internal logic of the expression, the depth of the argument, and the correlation between text and graphs, resulting in a significant deviation between the grading results and the teacher's professional judgment. More critically, the existing system suffers from a fragmented grading logic. Its optical character recognition, image recognition, and knowledge point understanding modules operate independently, failing to effectively link the multimodal elements in the question stem (such as graphic annotations and formula parameters) with the student's answers. This makes it impossible to pinpoint deeper errors, such as "incorrect parameter references." Furthermore, system feedback is often limited to simple right / wrong judgments or score presentations, failing to provide specific error type analysis, correction suggestions, and related weak knowledge points. This severely limits its educational value in student self-correction and teacher-targeted instruction. Finally, the existing system has poor large-scale adaptability; the configuration of grading rules heavily relies on manual processes, which is time-consuming, labor-intensive, and slow, failing to meet the urgent needs for real-time and efficient grading in daily teaching and large-scale examinations. Summary of the Invention

[0003] To address the problems existing in current intelligent education technologies, this invention proposes an intelligent question grading method based on a multimodal large model, comprising the following steps: S1: Perform multimodal recognition and region segmentation on the image containing questions and handwritten answers to extract structured question data and answer data. The multimodal recognition includes at least the recognition of text, formulas and graphics. S2: Input the structured question data and answer data into a multimodal big model trained with knowledge in the education field to identify the knowledge points and scoring requirements examined in the question data, as well as the logical correctness and multimodal element standardization of the answer data; S3: Based on the recognition results, dynamically generate personalized grading rules for the current question, and map the question's knowledge points and answer logic into specific scoring points and scoring standards; S4: Automatically grade the answer data based on the generated personalized grading rules and generate grading results containing detailed scores for each grading point; S5: Generate and output feedback information based on the grading results, including error location and suggestions for correcting weak knowledge points.

[0004] Furthermore, in step S1, the multimodal recognition specifically includes: Different recognition models are used to process printed text, handwritten text, printed formulas, handwritten formulas, and graphic elements, and the recognized formulas are converted into a standard format.

[0005] Furthermore, step S1 also includes: After the region is segmented, the relationship between the question region, the answer region, and the auxiliary region is established to generate the structured question data and answer data.

[0006] Furthermore, in step S2, the knowledge points and scoring requirements examined by the question data are specifically identified in the following way: The core test points of the questions are extracted based on the subject knowledge graph, and a hierarchical scoring requirement is defined, which includes at least one dimension of step completeness, logical coherence and multimodal element standardization.

[0007] Furthermore, in step S2, the logical correctness and multimodal element normalization of the response data are specifically identified in the following ways: When answering science questions, analyze the logical coherence between the problem-solving steps and the matching between formulas and steps; And / or, for humanities-related answers, analyze the logical structure and depth of argumentation in the expression; At the same time, the standardization of formula symbols and graphic annotations in the answer data is verified.

[0008] Furthermore, in step S4, the automated grading includes hierarchical scoring, specifically as follows: For science questions, they are graded and scored step by step; for humanities questions, they are scored separately based on multiple dimensions, including key points, logical coherence, and the connection between text and graphics.

[0009] Furthermore, in step S3, mapping the question's knowledge points and answer logic to specific scoring points and scoring criteria specifically involves: The multimodal elements identified in the question data are matched with the corresponding multimodal elements in the answer data, and the first type of scoring points and scoring criteria are defined based on the matching results. The logical requirements extracted from the question data are analyzed in correspondence with the logical structure analyzed from the answer data, and the second type of scoring points and scoring criteria are defined based on the analysis results.

[0010] Furthermore, in step S5, the feedback information also includes a class learning statistics report for the teacher, which includes at least error rate data calculated by knowledge point or error type.

[0011] Furthermore, it also includes iterative optimization steps: S6: Collect user feedback data on grading results or feedback information, and use the feedback data to optimize the multimodal large model or grading rule generation logic.

[0012] Compared with the prior art, the present invention has at least the following beneficial effects: (1) The present invention proposes an intelligent question grading method based on a multimodal large model, which realizes end-to-end intelligent grading of multimodal questions that integrate text, formulas, graphics and handwriting. It fundamentally overcomes the shortcomings of existing technologies, such as strong subject coverage and poor module reusability, and significantly broadens the scope of application of the system. (2) By using a multimodal model trained with knowledge in the field of education to deeply analyze the examination intent of the questions and the students’ answering logic, and dynamically generating personalized grading rules that are highly adapted to specific questions, it realizes the transformation from single result judgment to precise tiered scoring in multiple dimensions such as completeness of steps, logical coherence and standardization of multimodal elements, which greatly improves the depth and accuracy of grading. (3) It generates personalized feedback that includes specific error location, correction suggestions and links to weak knowledge points, which breaks through the limitation of traditional systems that only output correct and wrong scores. It provides support for students to correct their own mistakes and for teachers to understand the learning situation, and gives intelligent grading educational guidance value. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the steps of an intelligent question grading method based on a multimodal large model. Detailed Implementation

[0014] The following are specific embodiments of the present invention, which are described in conjunction with the accompanying drawings. However, the present invention is not limited to these embodiments.

[0015] While significant progress has been made in the development of intelligent question grading systems in the current educational technology field, their practical application still faces considerable limitations. Traditional systems generally lack the comprehensive processing capabilities for multimodal questions, making it difficult to effectively handle complex question scenarios that integrate text, mathematical formulas, graphics, and handwriting. This results in a significantly insufficient actual coverage across all K-12 subjects. Specifically, existing technologies lack depth in understanding key elements such as graphical elements and formula derivations in science questions, and the connection between text and graphics and logical reasoning in humanities questions. They often only perform superficial result comparisons or mechanical keyword matching, failing to conduct in-depth analysis and tiered evaluation of the completeness of solution steps, logical coherence, and the standardization of expression. This superficial approach leads to significant discrepancies between grading results and the evaluation standards of professional teachers, making it difficult to accurately reflect students' true knowledge mastery. Furthermore, the lack of an effective collaborative mechanism between the various recognition and understanding modules in the system causes the inherent multimodal correlation between the question stem information and the student's answer to be overlooked, making it impossible to pinpoint the root cause of errors. Furthermore, the feedback information provided by existing systems is too simplistic, typically limited to right / wrong judgments or score presentations. It fails to offer students specific error analysis and correction guidance, and also struggles to help teachers gain valuable insights into classroom learning, thus severely limiting its practical value in promoting mutual learning between teachers and students. Therefore, such as... Figure 1 As shown, this invention proposes an intelligent question grading method based on a multimodal large model, including the following steps: S1: Perform multimodal recognition and region segmentation on the image containing questions and handwritten answers to extract structured question data and answer data. The multimodal recognition includes at least the recognition of text, formulas and graphics. S2: Input the structured question data and answer data into a multimodal big model trained with knowledge in the education field to identify the knowledge points and scoring requirements examined in the question data, as well as the logical correctness and multimodal element standardization of the answer data; S3: Based on the recognition results, dynamically generate personalized grading rules for the current question, and map the question's knowledge points and answer logic into specific scoring points and scoring standards; S4: Automatically grade the answer data based on the generated personalized grading rules and generate grading results containing detailed scores for each grading point; S5: Generate and output feedback information based on the grading results, including error location and suggestions for correcting weak knowledge points; S6: Collect user feedback data on grading results or feedback information, and use the feedback data to optimize the multimodal large model or grading rule generation logic.

[0016] The implementation of this invention begins with the processing of the original image containing the questions and students' handwritten answers. This original image can be a scanned copy of a paper exam paper, a screenshot of an electronic assignment, or a direct photograph of a handwritten answer sheet. The intelligent question grading system first executes a multimodal input processing flow, the core objective of which is to transform unstructured image information into structured data that is machine-understandable and reasonable. The system performs a series of preprocessing operations on the input image, including but not limited to using digital image processing algorithms for noise reduction to remove paper stains and scanning noise, using Hough transform or similar techniques for tilt correction to ensure image alignment, and performing resolution normalization. Based on this, the system enhances and optimizes the image according to different subject characteristics; for example, for mathematical geometry problems, it enhances the sharpness and continuity of lines; for chemical experiment images, it enhances the contrast between different color regions to facilitate subsequent recognition and segmentation.

[0017] After preprocessing, the system enters the multimodal element recognition stage. In this stage, the intelligent question grading system does not use a single model for brute-force recognition, but rather employs a series of dedicated, collaborative recognition models to process different types of modal information separately. For the text portion, the system's built-in optical character recognition model can clearly distinguish and recognize the printed text in the question stem from the handwritten text in the student's answer area, and output the text content, its coordinate position in the image, and the text type (printed or handwritten). For the formula portion, whether it's a printed formula in the question stem or a formula handwritten by the student, the system activates a dedicated formula recognition engine that integrates optimized algorithms (such as an improved algorithm based on Mathpix). This engine can not only accurately recognize the visual structure of the formula but also convert it into standard LaTeX format, for example, converting a handwritten Pythagorean theorem into... And attempt to semantically annotate key symbols in the formula, such as annotating " "Represents the length of the hypotenuse." For graphical elements, the system invokes deep learning-based object detection and semantic segmentation algorithms for specialized recognition across different disciplines: for mathematical geometry diagrams, it can identify vertices, edges, angles, and auxiliary lines; for physical circuit diagrams, it can identify components such as resistors and power supplies and their connection methods; for illustrations in the humanities, it can identify scenes, people, and annotation text within the diagrams. All identified multimodal elements—including text blocks, LaTeX formula strings, and graphical elements and their attributes—are uniformly integrated to generate a structured multimodal element library.

[0018] Next, the intelligent question grading system performs question area segmentation. This step is to solve the problem of area confusion caused by multiple questions mixed together and text and images intersecting in actual questions. Based on the aforementioned recognition results, the system uses spatial layout analysis and content clustering algorithms to accurately locate and segment three key areas: the first is the "whole question area," which is located by recognizing the question number (such as "1.", "2.") and combining it with the clustering analysis of printed text, and includes the complete question stem text, formulas and graphs in the question stem; the second is the "answer area," which is located by judging that the proportion of handwritten content exceeds a certain threshold (such as 80%), its spatial position is adjacent to the whole question area (usually located below or to the right of it), and by referring to common answer markings (such as underlines and boxes), and encompasses all the student's handwriting, drawn graphs and formulas; the third is the "auxiliary area," which is located by recognizing printed prompts such as "full marks 5 points" and "unit: centimeters." The system then establishes a one-to-one correspondence between the "question area - answer area - auxiliary area" and generates an area association table, thus logically separating and linking the various questions and their answers on a complex layout. Finally, the system integrates the multimodal element library with the area association table, outputting a complete "question-answer multimodal structured data". This data exists in a structured format such as tables or JSON, clearly recording the whole question information, the corresponding answer information, and the multimodal elements associated with each question, preparing for subsequent core reasoning.

[0019] After completing the structured preparation of the data, the intelligent question grading system enters the core multimodal large-scale model reasoning stage. The multimodal large-scale model relied upon in this stage is derived from a general visual-language large-scale model, deeply fine-tuned using massive amounts of educational domain data (such as textbooks, exercise sets, lesson plans, and subject knowledge graphs), giving it profound prior knowledge in the educational domain. After the structured data is input into the model, it performs two parallel parsing tasks: deep understanding of the question stem and parsing of the answer content.

[0020] In its in-depth understanding of the question stem, the model automatically extracts the core knowledge points tested in the question based on its built-in subject knowledge graph. For example, in a math problem, it identifies "the application of the Pythagorean theorem," and in a Chinese reading comprehension question, it identifies "the analysis of metaphorical rhetoric and the summarization of the main idea." Then, the model automatically defines tiered scoring dimensions based on the knowledge points and question requirements. For example, for a math proof question, it defines a step score of "correct formula writing (2 points) → correct parameter substitution (2 points) → correct calculation result (1 point)," or for a historical material analysis question, it defines a logical score of "complete key points (3 points) → rigorous logical argumentation (2 points)." Simultaneously, the model analyzes the implicit requirements of the question stem for multimodal elements, such as "auxiliary lines need to be drawn in the diagram" or "analysis needs to be combined with the content of the image."

[0021] In the analysis of student responses, the model conducts in-depth analysis. For science responses, the model reconstructs the logical chain of the solution, analyzing the coherence and rationality between steps. For example, it determines whether the student correctly identified the triangle as a right triangle before applying the Pythagorean theorem and checks whether the formula and steps match. For humanities responses, the model focuses on analyzing the logical structure of the expression, such as the completeness of the "viewpoint-evidence-conclusion" structure, and the depth of the argumentation, such as whether the information provided in the question stem is effectively integrated. Furthermore, the model verifies the standardization of multimodal elements in the responses, including checking whether the writing of formula symbols is standard (e.g., the square symbol). The model checks for clarity and accuracy, as well as the precision of annotations in the diagrams and the consistency of units with the question. Ultimately, the model outputs a detailed "Answer Analysis Report," summarizing the key points, logical flow, and compliance issues of the answer.

[0022] Next, the intelligent question grading system enters the dynamic grading rule generation stage, which is crucial for achieving personalized and accurate grading. The multimodal big data model correlates and matches the results of the "deep understanding of the question stem" with the "answer analysis report." The matching process mainly operates on two levels: first, consistency matching of multimodal elements, such as comparing "AC=3cm" marked on the geometric diagram in the question stem with "AC=4cm" in the student's answer formula; second, analysis of the compliance with logical requirements, such as checking whether the student's text mentions the image scene if the question stem requires "analysis combined with the image."

[0023] Based on this matching result, the model dynamically generates a personalized grading rule table specifically for the current question and the current answer. This rule table includes specific scoring points, the score for each point, deduction conditions, and important error tolerance instructions. For example, for the parameter error mentioned above, the rule "1 point for correct AC parameter reference, deduct 1 point for error" will be generated. For non-principle errors such as differences in the order of steps, error tolerance rules will be set, resulting in no deduction or a small deduction, thus greatly improving the reasonableness and humanization of the grading.

[0024] Subsequently, the intelligent question grading system enters the intelligent grading execution phase. This phase automates the grading of answers based on the aforementioned dynamically generated personalized grading rule table. Grading is divided into two paths: objective questions and subjective questions. For objective questions, such as multiple-choice and fill-in-the-blank questions, the system matches filled-in marks or handwritten options with the standard answers, and performs comprehensive verification of numerical values, units, and symbols for fill-in-the-blank questions containing formulas and units. For subjective questions, the system strictly implements a tiered scoring logic: for science questions, each step is graded and scored according to the rule table; for humanities questions, scores are given based on multiple dimensions such as content points, logical coherence, and the relevance of text and graphics; additionally, special scores are assigned to multimodal elements in the answer, such as graphic drawing and formula writing. All grading results, including the total score, detailed scores for each scoring point, and specific error markers, are compiled into a final grading report. Before outputting the report, the system will also perform a verification procedure, such as checking for contradictory situations such as "all steps are correct but the result is wrong" or "all key points are correct but the logic score is extremely low". If such anomalies are found, the system will trigger a manual review reminder, thereby further ensuring the reliability of the correction results.

[0025] After grading, the intelligent question grading system initiates a personalized feedback phase, transforming the grading results into educationally valuable behavioral guidance. For students, the system generates feedback information combining text and images. This feedback first clearly identifies the error location, not only indicating the error type (e.g., "incorrect symbol in the Pythagorean theorem formula"), but also linking it to the specific answer area on the image (e.g., highlighting the incorrect formula line in the answer). Then, it provides specific correction suggestions, such as showing the standard way to write the formula, or suggesting supplementary analysis of a detail in the image. More importantly, the feedback is linked to the corresponding weak knowledge points and intelligently recommends 1-2 practice questions targeting that knowledge point, thus forming a closed learning loop of "grading-location-correction-consolidation." For teachers, the system generates a macro-level learning statistics report. This report presents the class's error rate ranking across various error types and knowledge points in the form of data charts and reports, and automatically generates teaching focus suggestions based on this data. It also stratifies students according to their mastery level, providing teachers with intuitive data support for differentiated instruction and precise tutoring.

[0026] Finally, to ensure the long-term effectiveness and evolutionary capability of the intelligent question grading system, this invention also designs a system adaptation and iterative optimization process. The system incorporates a comprehensive subject-specific adaptation layer, which stores dedicated recognition rules, grading rule templates, and knowledge point graphs for each subject. When processing questions, the system automatically identifies the subject type and calls the corresponding resource library, thus achieving flexible adaptation of a single system to all K-12 subjects. Furthermore, the system features a feedback iteration loop. It continuously collects student feedback on the clarity of correction suggestions, as well as teacher feedback on the accuracy of grading and the practicality of teaching suggestions. This feedback data is used to drive continuous system optimization: if the grading deviation rate for a certain type of question is found to be high, then supplementary data for that type of question is added to fine-tune the multimodal large model; if feedback indicates that some feedback information is not specific enough, then the logic template for generating feedback is optimized; simultaneously, based on performance data from large-scale grading, the region segmentation algorithm and model inference speed are continuously optimized. For example, through model quantization technology, the average processing time per question is reduced from seconds to sub-seconds, thereby ensuring the system's efficient and stable operation in large-scale concurrent scenarios.

[0027] In summary, the intelligent question grading method based on a multimodal large model provided by this invention achieves end-to-end intelligent grading of multimodal questions that integrate text, formulas, graphics, and handwriting by constructing a complete technical chain from multimodal recognition and understanding to dynamic rule generation and execution. This fundamentally overcomes the shortcomings of existing technologies, such as limited subject coverage and poor module reusability, significantly broadening the system's applicability. Furthermore, by utilizing a multimodal large model trained with knowledge from the education field to deeply analyze the question's examination intent and the student's answer logic, and dynamically generating personalized grading rules highly adapted to specific questions, it achieves precise, tiered scoring across multiple dimensions, from single-result judgment to step completeness, logical coherence, and multimodal element standardization, greatly improving the depth and accuracy of grading. In addition, by generating personalized feedback that includes specific error location, correction suggestions, and links to weak knowledge points, it breaks through the limitations of traditional systems that only output correct / incorrect scores, providing strong support for students' self-correction and teachers' understanding of learning progress, thus giving intelligent grading true educational guidance value. Finally, the entire process is highly automated, from the dynamic generation of batching rules to efficient batching execution, significantly reducing the reliance on manually pre-configured rules and ensuring the system's efficiency and versatility in large-scale, multi-disciplinary application scenarios. At the same time, its built-in iterative optimization mechanism ensures the continuous evolution of system performance and long-term applicability.

[0028] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0029] Furthermore, in this invention, descriptions involving terms such as "first," "second," and "a" are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0030] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0031] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

Claims

1. A multi-modal large model-based intelligent question grading method, characterized in that, The method comprises the following steps: S1: multi-modal recognition and region segmentation are performed on an image containing a question and a handwritten answer, and structured question data and answer data are extracted, wherein the multi-modal recognition at least includes recognition of text, formula and graphics; S2: the structured question data and answer data are input into a multi-modal large model trained with knowledge in the education field, knowledge points and scoring requirements examined by the question data, and logical correctness and multi-modal element standardization of the answer data are identified; S3: personalized correction rules for the current question are dynamically generated based on the identification results, and the question knowledge points and answer logic are mapped to specific scoring points and grading standards; S4: the answer data is automatically corrected according to the generated personalized correction rules, and a correction result containing scoring details of each scoring point is generated; S5: feedback information including error positioning and weak knowledge point association correction suggestions is generated and output based on the correction result. 2.The intelligent question grading method based on a multi-modal large model of claim 1, wherein, In the S1 step, the multi-modal recognition specifically includes: Different recognition models are used to process printed text, handwritten text, printed formula, handwritten formula and graphic elements, and the recognized formula is converted into a standard format. 3.The intelligent question grading method based on a multi-modal large model of claim 1, wherein, The S1 step further includes: After region segmentation, an association relationship between the question region, the answer region and the auxiliary region is established, and the structured question data and answer data are generated. 4.The intelligent question grading method based on a multi-modal large model of claim 1, wherein, In the S2 step, the knowledge points and scoring requirements examined by the question data are specifically identified by the following method: Based on a subject knowledge graph, the core examination points of the question are extracted, and hierarchical scoring requirements including at least one of the following dimensions: completeness of steps, logical coherence and multi-modal element standardization are defined. 5.The intelligent question grading method based on a multi-modal large model of claim 1, wherein, In the S2 step, the logical correctness and multi-modal element standardization of the answer data are specifically identified by the following method: For scientific answers, the logical coherence between steps and the matching of formulas and steps are analyzed; And / or, for liberal arts answers, the logical structure of the expression and the depth of the argument are analyzed; At the same time, the standardization of formula symbols and graphic annotations in the answer data is verified. 6.The intelligent question grading method based on a multi-modal large model of claim 5, wherein, In the S4 step, the automatic correction includes hierarchical grading, specifically: For scientific questions, step-by-step correction and grading are performed; for liberal arts questions, multiple dimensions such as content points, logical coherence and graphic-text correlation are graded separately. 7.The intelligent question grading method based on a multi-modal large model of claim 1, wherein, In the S3 step, the mapping of question knowledge points and answer logic to specific scoring points and grading standards specifically includes: The multi-modal elements in the question data are matched with the corresponding multi-modal elements in the answer data for consistency, and the first type of scoring points and grading standards are defined according to the matching results; The logical requirements in the question data are analyzed for corresponding relationship with the logical structure in the answer data, and the second type of scoring points and grading standards are defined according to the analysis results. 8.The intelligent question grading method based on a multi-modal large model of claim 1, wherein, In the S5 step, the feedback information further includes a class learning situation statistical report for a teacher end, and the class learning situation statistical report at least includes error rate data counted by knowledge points or error types. 9.The intelligent question grading method based on a multi-modal large model of claim 1, wherein, The method further comprises an iterative optimization step: S6: Collect feedback data of the user on the correction result or feedback information, and optimize the multi-modal large model or the correction rule generation logic using the feedback data.

Images