Composition multi-dimensional personalized correction method based on multi-modal large model

Abstract

The invention relates to a composition multi-dimensional personalized correction method based on a multi-modal large model, and the method converts a conventional manual work document correction process into an automatic correction process, thereby fundamentally solving the problem of low manual correction efficiency. According to the method, multi-modal analysis and feature extraction are carried out on composition content and selectable associated modal information, and multi-modal information identification of characters, pictures and materials is realized in combination with a multi-modal evaluation model, so that the problems of single dimension and shallow feedback layer of traditional composition correction are solved, and automatic correction of multiple types of compositions is realized; and finally, objective scoring is ensured based on a multi-dimensional evaluation component table, personalized feedback is provided through targeted batch and stylized polishing, and the contradiction between high scoring subjectivity and non-uniform feedback is solved.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and educational technology, and more specifically, to a multi-dimensional personalized essay correction method based on a multimodal large model. Background Technology

[0002] With the rapid development of information technology, the education sector has gradually introduced various intelligent tools to improve teaching efficiency and student learning outcomes. Especially in essay correction and writing skills enhancement, traditional manual correction methods are not only time-consuming and labor-intensive, but also struggle to provide comprehensive and personalized feedback. With advancements in machine learning, particularly deep learning, models trained using large-scale text data can more accurately understand essay content and provide more detailed scoring and feedback.

[0003] Existing essay grading systems can only detect superficial issues such as grammatical errors and word count, and cannot cover core dimensions such as theme depth, structural logic, and emotional expression. For multimodal types such as picture-based essays and material-based essays, existing systems cannot analyze dimensions such as the relationship between the picture and the text and the fit of the material, and the grading coverage rate is less than 40%. Summary of the Invention

[0004] The technical problem this invention aims to solve is how to provide a multi-dimensional personalized grading method that can cover multiple types of essays and achieve accurate scoring.

[0005] This invention provides a multi-dimensional personalized essay correction method based on a multimodal large model, including: Step 1: Obtain essay data containing essay content and optional associated modal information; Step 2: Perform multimodal analysis and feature extraction on the essay data to generate structured data containing textual features and multimodal correlation features; Step 3: Obtain the multi-dimensional scoring scale corresponding to the essay data. The multi-dimensional scoring scale includes a general set of dimensions and a specific set of dimensions that are dynamically configured based on the subject and essay type. Each dimension in the general set of dimensions and the specific set of dimensions has a preset weight and a corresponding scoring standard. Step 4: Construct and pre-train a multimodal scoring model based on transform. Input structured data, multidimensional scoring scales and essay data into the multimodal evaluation model. The multimodal evaluation model generates multidimensional scores, targeted side notes and stylistic polishing suggestions for the essay data based on the structured data and multidimensional scoring scales, and outputs them in the form of personalized feedback.

[0006] Compared with existing technologies, this application has the following advantages: This application transforms the traditional manual essay grading process into an automated grading process, fundamentally solving the problem of low efficiency in manual grading; it performs multimodal analysis and feature extraction on the essay content and optional related modal information, and combines it with a multimodal evaluation model to realize the recognition of multimodal information of text, images, and materials, solving the problem of single dimension and shallow feedback in traditional essay grading, and realizing automated grading of multiple types of essays; finally, it ensures the objectivity of scoring based on a multidimensional evaluation subscale, and provides personalized feedback through targeted side comments and stylized polishing, resolving the contradiction between strong subjectivity in scoring and uniform feedback.

[0007] In one possible implementation, step 2 specifically includes: Step 201: Determine the type of the essay data. The types include at least picture-based essays, material-based essays, application essays, and topic-based essays. Step 202: Based on the determined type of essay data, perform the corresponding multimodal association feature extraction operation; Step 203: Perform semantic analysis on the text content and extract text features; Step 204: Integrate textual features and multimodal association features to form structured data.

[0008] By automatically identifying the essay type and executing corresponding extraction strategies, it can accurately handle various question types such as picture-based, material-based, and application-based questions, thus solving the key deficiency of existing technologies in multimodal essay processing capabilities.

[0009] In one possible implementation, the optional associated modal information includes images and background materials; step 201, specifically determining the type of the essay data, includes: Step 201A: Determine whether the essay data contains an image. If yes, the essay type is an image-based essay; otherwise, proceed to step 201B. Step 201B: Determine whether the essay data contains background material. If yes, the essay type is a material-based essay; otherwise, proceed to step 201C. Step 201C: Determine whether the format of the text content conforms to the format of an application document. If yes, the essay type is an application document; otherwise, the essay type is a topic-based essay.

[0010] In one possible implementation, step 202, which involves performing the corresponding multimodal correlation feature extraction operation, specifically includes: When the essay type is a picture-based essay, then: The visual language model is used to generate textual descriptions of the images; and the object detection model is used to identify the core element set and scene labels in the images. A sentence embedding model is used to convert text descriptions and essay content into text description vectors and essay vectors; and the semantic relevance between text and images is calculated based on the text description vectors and essay vectors. Extract an entity list from the essay content, and calculate the core element coverage rate based on the entity list and the core element set. The semantic relevance of the image and text, the coverage of core elements, and the undescribed core entity elements are output as multimodal relevance features; When the essay type is a material-based essay, then: A text summarization model is used to extract core thematic sentences from background materials; Using advanced information extraction technology, a list of core arguments is extracted from the background material; A sentence embedding model is used to convert the core theme statement and the essay content into theme statement vectors and essay vectors, respectively, and the theme fit is calculated based on the theme statement vectors and essay vectors. Extract a list of arguments from the essay content; calculate the citation rate of the arguments based on the list of arguments and the list of core arguments; The main idea relevance, citation rate of arguments, list of arguments, and list of missing arguments are output as multimodal association features; When the essay type is practical writing, then: Predefine application document tags for different application document types to form a format rule library; The position of application text tags in the essay content is detected based on the format rule base, the format conformity is calculated and output as a multimodal association feature; When the essay type is a topic-based essay, keyword extraction and main idea analysis are performed on the essay title to obtain the main idea vector; The essay content is converted into an essay vector. The relevance is calculated based on the essay vector and the topic theme vector, and then output as a multimodal association feature.

[0011] In one possible implementation, step 203, which involves semantic parsing of the text content and extracting text features, specifically includes: First, the obtained essay content is cleaned to remove garbled characters, characteristic characters, and redundant formatting characters; Secondly, a natural language processing model is used to extract core keywords from the cleaned essay content, and the importance score of the core keywords is calculated based on attention weight. The N words with the highest scores are selected as the topic keywords. A sequence labeling model is used to identify the logical relationships between paragraphs in an essay and output a sequence of logical relationship labels. A sentiment analysis model was used to calculate the sentiment positivity score in the essay content, and the sentiment positivity score was converted into classification labels; Finally, the extracted thematic keywords, logical relationship tag sequences, and emotional positivity are integrated to form textual features.

[0012] In one possible implementation, step 4 specifically includes: Step 401: Construct and pre-train a multimodal scoring model based on transform; User profiles are generated based on historical essay correction data from different authors. Based on these user profiles, ability levels are determined, and different polishing strategies are applied to authors at different ability levels. For basic users, the focus is on grammar correction and sentence simplification; For intermediate-level users, the focus is on logical connections and supplementary details; For high-level users, the focus is on rhetorical devices and innovative expression; Structured data, multidimensional rating scales, and essay data are input into the multimodal evaluation model; Step 402: The multimodal evaluation model is based on structured data, obtains the set of dimensions and weights of each dimension in the multidimensional rating scale, calculates or assists in the calculation of each dimension in the multidimensional evaluation subscale for independent evaluation based on the quantitative indicators in the correlation features, generates multidimensional scores, and finally calculates the total score of the essay and outputs the multidimensional score and total score of the essay. Step 403: Based on the multi-dimensional evaluation results, locate the text unit that caused the dimension deduction. The text unit is a sentence or paragraph. For each deducted text unit, identify the scoring dimension and specific problem type. Based on the standard of the corresponding dimension in the scoring scale, generate improvement suggestions containing specific modification strategies. Finally, associate the problem diagnosis with the improvement suggestions to generate targeted side notes located in the specific text unit. Step 404: Obtain the ability level associated with the author of the current essay, and select a suitable polishing strategy from the predefined polishing strategy library based on the ability level; For the text units associated with the targeted side notes, the selected polishing strategy is applied to generate one or more polished candidate sentences; Output the polished candidate sentences as stylistic polishing suggestions.

[0013] In one possible implementation, the set of general dimensions includes thematic relevance, structural integrity, and language fluency. The multi-dimensional scores calculated in step 402 specifically include: The score for the topic relevance dimension is calculated based on the correlation features corresponding to the essay type; where: When the essay type is a picture-based essay, the score for the theme relevance dimension is calculated based on the semantic relevance between the picture and the text. When the essay type is a material-based essay, the score for the theme relevance dimension is calculated based on the theme relevance. When the essay type is an application document, the score for the theme relevance dimension is calculated based on the format conformity. When the essay type is a topic-based essay, the score for the topic relevance dimension is calculated based on the relevance to the topic. The score for the structural integrity dimension is calculated based on the logical relationship label sequence; The score for the fluency dimension is calculated based on the number of grammatical errors and misspellings detected in the essay content, and with reference to the predefined number of misspellings and their corresponding relationships in the fluency dimension.

[0014] By transforming subjective factors such as the relationship between text and images and the relevance of the main theme into objective values ​​such as "text-image coverage" and "theme relevance," subsequent scoring becomes more data-driven, greatly improving the consistency and interpretability of the scoring.

[0015] In one possible implementation, the multi-dimensional personalized essay correction method further includes: The feedback from the reviewers is collected, and the multimodal evaluation model is iteratively optimized based on the collected feedback.

[0016] In one possible implementation, the output format of the personalized feedback includes: The multi-dimensional scores are visualized in the form of radar charts. The targeted annotations are located and marked next to the corresponding sentences or paragraphs in the original text; Provide a comparison between the original text and the revised text with editing suggestions. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the process of this application. Detailed Implementation

[0018] First, those skilled in the art should understand that these embodiments are merely used to explain the technical principles of the embodiments of this application and are not intended to limit the scope of protection of the embodiments of this application. Those skilled in the art can make adjustments as needed to adapt to specific application scenarios.

[0019] In the description of the embodiments of this application, it should be noted that, unless otherwise explicitly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in the embodiments of this application based on the specific circumstances.

[0020] In the embodiments of this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0021] The present application will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0022] See Figure 1 As shown in the embodiments of this application, the present invention discloses a method for multi-dimensional personalized essay correction based on a multimodal large model, including: Step 1: Obtain essay data containing essay content and optional associated modal information; the optional associated modal information includes images and background materials; Step 2 involves multimodal parsing and feature extraction of the essay data to generate structured data containing textual features and multimodal correlation features. The goal of this step is to transform unstructured essay input (text, images, materials) into machine-understandable and computable structured feature data. Specifically, this includes: Step 201: Determine the type of the essay data. The types include at least picture-based essays, material-based essays, application essays, and topic-based essays. Specifically, the essay type (picture-based essay, material-based essay, application essay, topic-based essay) is determined by combining rules with a lightweight text classification model (such as FastText). This includes: Step 201A: Determine whether the essay data contains an image. If yes, the essay type is an image-based essay; otherwise, proceed to step 201B. Step 201B: Determine whether the essay data contains background material. If yes, the essay type is a material-based essay; otherwise, proceed to step 201C. Step 201C: Determine whether the format of the text content conforms to the format of an application document. If yes, the essay type is an application document; otherwise, the essay type is a topic-based essay.

[0023] Step 202: Based on the determined type of essay data, perform the corresponding multimodal association feature extraction operation; specifically including: When the essay type is a picture-based essay, then: Use a visual language model (such as CLIP or GPT-4V) to generate textual descriptions of the images; and use an object detection model (such as YOLO or Faster R-CNN) to identify the core set of elements in the images. and scene tags ; A sentence embedding model is used to convert text descriptions and essay content into text description vectors. and essay vector ; and based on text description vectors and essay vector Calculate the semantic relevance between text and images ; Extract entity list from essay content Based on entity list and core element set Calculate core element coverage ; semantic correlation between images and text Core element coverage and undescribed core entity elements As a multimodal correlation feature output; When the essay type is a material-based essay, then: Text summarization models (such as BART) are used to extract core thematic statements from background material. ; Employing advanced information extraction techniques (such as BERT-based sequence labeling), a list of core arguments is extracted from the background material. ; Using a sentence embedding model, the core theme statement and the essay content are converted into theme statement vectors, respectively. and essay vector Based on the main idea statement vector and essay vector Calculate the degree of relevance to the main theme ; Extract a list of arguments from the essay content. Based on the list of arguments and a list of core arguments Calculate the citation rate of arguments ; The degree of relevance to the main theme Citation rate of arguments List of arguments The list of missing arguments is output as a multimodal association feature; When the essay type is practical writing, then: Predefine application document tags for different document types (letters, notices, diaries, etc.) to form a format rule library. It includes the necessary structural units (such as salutation, body, signature, and date); Based on format rule library The positions of application document tags in the essay content are detected, and the format conformity is calculated using rule-based or sequence labeling models. And it is output as a multimodal association feature, calculated as follows: ; When the essay type is a given topic, then keywords are extracted from the essay title. Based on the main idea analysis, we obtain the main idea vector of the question. ; Convert essay content into essay vectors Based on essay vectors and the main idea vector of the question Calculate the relevance of the topic And output as a multimodal association feature, the calculation formula is: .

[0024] Step 203: Perform semantic analysis on the text content and extract text features; specifically including: First, the obtained essay content is cleaned to remove garbled characters, characteristic characters, and redundant formatting characters; Secondly, natural language processing models (such as BERT or TF-IDF) are used to extract core keywords from the cleaned essay content, and importance scores of the core keywords are calculated based on attention weights. The N words with the highest scores are selected as topic keywords. For each core keyword... The formula for calculating its importance score is as follows: Sequence labeling models (such as Bi-LSTM+CRF) are used to identify the logical relationships between paragraphs in the essay content and output a sequence of logical relationship labels. Sentiment analysis models (such as BERT-based sentiment classifiers) are used to calculate the sentiment positivity score in the essay content, and the sentiment positivity score is converted into classification labels; The score is denoted as -1, which represents extreme negativity, and 1 represents extreme positivity. This score can be converted into category labels (such as "positive" or "negative").

[0025] Finally, the extracted thematic keywords, logical relationship tag sequences, and emotional positivity are integrated to form textual features.

[0026] Step 204: Integrate textual features and multimodal association features to form structured data.

[0027] Step 3: Obtain the multi-dimensional scoring scale corresponding to the essay data. The multi-dimensional scoring scale includes a general dimension set and a specific dimension set dynamically configured based on the subject and essay type. Each dimension in the general dimension set and the specific dimension set has a preset weight and corresponding scoring criteria. The general dimension set includes the theme relevance dimension, structural integrity dimension, and language fluency dimension. Step 4: Construct and pre-train a multimodal scoring model based on transform. Input structured data, multidimensional scoring scales, and essay data into the multimodal evaluation model. The multimodal evaluation model generates multidimensional scores, targeted side notes, and stylistic polishing suggestions for the essay data based on the structured data and multidimensional scoring scales, and outputs them in the form of personalized feedback. Specifically, this includes: Step 401: Construct and pre-train a multimodal scoring model based on transform; User profiles are generated based on historical essay correction data from different authors. Based on these user profiles, ability levels are determined, and different polishing strategies are applied to authors at different ability levels. For basic users, the focus is on grammar correction and sentence simplification; For intermediate-level users, the focus is on logical connections and supplementary details; For high-level users, the focus is on rhetorical devices and innovative expression; Structured data, multidimensional rating scales, and essay data are input into the multimodal evaluation model; Step 402: The multimodal evaluation model, based on structured data, obtains the set of dimensions and weights in the multi-dimensional rating scale. Based on the quantitative indicators in the correlation features, it calculates or assists in the calculation of each dimension in the multi-dimensional evaluation subscale for independent evaluation, generating a multi-dimensional score. Finally, it calculates the total score of the essay and outputs the multi-dimensional score and total score of the essay. Specifically, this includes: The score for the topic relevance dimension is calculated based on the correlation features corresponding to the essay type; where: When the essay type is a picture-based essay, the score for the theme relevance dimension is calculated based on the semantic relevance between the picture and the text. When the essay type is a material-based essay, the score for the theme relevance dimension is calculated based on the theme relevance. When the essay type is an application document, the score for the theme relevance dimension is calculated based on the format conformity. When the essay type is a topic-based essay, the score for the topic relevance dimension is calculated based on the relevance to the topic. The score for the structural integrity dimension is calculated based on the logical relationship label sequence; The score for the fluency dimension is calculated based on the number of grammatical errors and misspellings detected in the essay content, and with reference to the predefined number of misspellings and their corresponding relationships in the fluency dimension.

[0028] Step 403: Based on the multi-dimensional evaluation results, locate the text unit that caused the dimension deduction. The text unit is a sentence or paragraph. For each deducted text unit, identify the scoring dimension and specific problem type. Based on the standard of the corresponding dimension in the scoring scale, generate improvement suggestions containing specific modification strategies. Finally, associate the problem diagnosis with the improvement suggestions to generate targeted side notes located in the specific text unit. Step 404: Obtain the ability level associated with the author of the current essay, and select a suitable polishing strategy from the predefined polishing strategy library based on the ability level; For the text units associated with the targeted side notes, the selected polishing strategy is applied to generate one or more polished candidate sentences; Output the polished candidate sentences as stylistic polishing suggestions.

[0029] And output will be in the form of personalized feedback; specifically including: The multi-dimensional scores are visualized in the form of radar charts. The targeted annotations are located and marked next to the corresponding sentences or paragraphs in the original text; Provide a comparison between the original text and the revised text with editing suggestions.

[0030] Step 5: Collect feedback from the grading staff and iteratively optimize the multimodal evaluation model based on the collected feedback. In this embodiment, the multimodal evaluation model is fine-tuned based on manual grading samples and LoRA technology.

[0031] In the description of the embodiments of this application, it should be noted that the terms "inner" and "outer" and other terms indicating direction or positional relationship are based on the direction or positional relationship shown in the drawings. This is only for the convenience of description and does not indicate or imply that the device or component must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this application.

[0032] In the description of this application, the references to terms such as "an embodiment," "some embodiments," "in this embodiment," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0033] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multi-dimensional personalized essay correction method based on a multimodal large model, characterized in that, include: Step 1: Obtain essay data containing essay content and optional associated modal information; Step 2: Perform multimodal analysis and feature extraction on the essay data to generate structured data containing textual features and multimodal correlation features; Step 3: Obtain the multi-dimensional scoring scale corresponding to the essay data. The multi-dimensional scoring scale includes a general set of dimensions and a specific set of dimensions that are dynamically configured based on the subject and essay type. Each dimension in the general set of dimensions and the specific set of dimensions has a preset weight and a corresponding scoring standard. Step 4: Construct and pre-train a multimodal scoring model based on transform. Input structured data, multidimensional scoring scales and essay data into the multimodal evaluation model. The multimodal evaluation model generates multidimensional scores, targeted side notes and stylistic polishing suggestions for the essay data based on the structured data and multidimensional scoring scales, and outputs them in the form of personalized feedback.

2. The method for multi-dimensional personalized essay correction based on a multimodal large model according to claim 1, characterized in that, Step 2 specifically includes: Step 201: Determine the type of the essay data. The types include at least picture-based essays, material-based essays, application essays, and topic-based essays. Step 202: Based on the determined type of essay data, perform the corresponding multimodal association feature extraction operation; Step 203: Perform semantic analysis on the text content and extract text features; Step 204: Integrate textual features and multimodal association features to form structured data.

3. The method for multi-dimensional personalized essay correction based on a multimodal large model according to claim 2, characterized in that, The optional associated modal information includes images and background materials; Step 201, which involves determining the type of the essay data, specifically includes: Step 201A: Determine whether the essay data contains an image. If yes, the essay type is an image-based essay; otherwise, proceed to step 201B. Step 201B: Determine whether the essay data contains background material. If yes, the essay type is a material-based essay; otherwise, proceed to step 201C. Step 201C: Determine whether the format of the text content conforms to the format of an application document. If yes, the essay type is an application document; otherwise, the essay type is a topic-based essay.

4. The method for multi-dimensional personalized essay correction based on a multimodal large model according to claim 3, characterized in that, Step 202 involves performing the corresponding multimodal association feature extraction operation, specifically including: When the essay type is a picture-based essay, then: The visual language model is used to generate textual descriptions of the images; and the object detection model is used to identify the core element set and scene labels in the images. A sentence embedding model is used to convert text descriptions and essay content into text description vectors and essay vectors; and the semantic relevance between text and images is calculated based on the text description vectors and essay vectors. Extract an entity list from the essay content, and calculate the core element coverage rate based on the entity list and the core element set. The semantic relevance of the image and text, the coverage of core elements, and the undescribed core entity elements are output as multimodal relevance features; When the essay type is a material-based essay, then: A text summarization model is used to extract core thematic sentences from background materials; Using advanced information extraction technology, a list of core arguments is extracted from the background material; A sentence embedding model is used to convert the core theme statement and the essay content into theme statement vectors and essay vectors, respectively, and the theme fit is calculated based on the theme statement vectors and essay vectors. Extract a list of arguments from the essay content; calculate the citation rate of the arguments based on the list of arguments and the list of core arguments; The main idea relevance, citation rate of arguments, list of arguments, and list of missing arguments are output as multimodal association features; When the essay type is practical writing, then: Predefine application document tags for different application document types to form a format rule library; The position of application text tags in the essay content is detected based on the format rule base, the format conformity is calculated and output as a multimodal association feature; When the essay type is a topic-based essay, keyword extraction and main idea analysis are performed on the essay title to obtain the main idea vector; The essay content is converted into an essay vector. The relevance is calculated based on the essay vector and the topic theme vector, and then output as a multimodal association feature.

5. The method for multi-dimensional personalized essay correction based on a multimodal large model according to claim 4, characterized in that, Step 203 involves semantic parsing of the text content and extracting text features, specifically including: First, the obtained essay content is cleaned to remove garbled characters, characteristic characters, and redundant formatting characters; Secondly, a natural language processing model is used to extract core keywords from the cleaned essay content, and the importance score of the core keywords is calculated based on attention weight. The N words with the highest scores are selected as the topic keywords. A sequence labeling model is used to identify the logical relationships between paragraphs in an essay and output a sequence of logical relationship labels. A sentiment analysis model was used to calculate the sentiment positivity score in the essay content, and the sentiment positivity score was converted into classification labels; Finally, the extracted thematic keywords, logical relationship tag sequences, and emotional positivity are integrated to form textual features.

6. The method for multi-dimensional personalized essay correction based on a multimodal large model according to claim 5, characterized in that, Step 4 specifically includes: Step 401: Construct and pre-train a multimodal scoring model based on transform; User profiles are generated based on historical essay correction data from different authors. Based on these user profiles, ability levels are determined, and different polishing strategies are applied to authors at different ability levels. For basic users, the focus is on grammar correction and sentence simplification; For intermediate-level users, the focus is on logical connections and supplementary details; For high-level users, the focus is on rhetorical devices and innovative expression; Structured data, multidimensional rating scales, and essay data are input into the multimodal evaluation model; Step 402: The multimodal evaluation model is based on structured data, obtains the set of dimensions and weights of each dimension in the multidimensional rating scale, calculates or assists in the calculation of each dimension in the multidimensional evaluation subscale for independent evaluation based on the quantitative indicators in the correlation features, generates multidimensional scores, and finally calculates the total score of the essay and outputs the multidimensional score and total score of the essay. Step 403: Based on the multi-dimensional evaluation results, locate the text unit that caused the dimension deduction. The text unit is a sentence or paragraph. For each deducted text unit, identify the scoring dimension and specific problem type. Based on the standard of the corresponding dimension in the scoring scale, generate improvement suggestions containing specific modification strategies. Finally, associate the problem diagnosis with the improvement suggestions to generate targeted side notes located in the specific text unit. Step 404: Obtain the ability level associated with the author of the current essay, and select a suitable polishing strategy from the predefined polishing strategy library based on the ability level; For the text unit associated with the targeted side note, the selected polishing strategy is applied to generate one or more polished candidate sentences; Output the polished candidate sentences as stylistic polishing suggestions.

7. The multi-dimensional personalized essay correction method based on a multimodal large model as described in claim 6, characterized in that, The general set of dimensions includes thematic relevance, structural integrity, and language fluency. The multi-dimensional scores calculated in step 402 specifically include: The score for the topic relevance dimension is calculated based on the correlation features corresponding to the essay type; where: When the essay type is a picture-based essay, the score for the theme relevance dimension is calculated based on the semantic relevance between the picture and the text. When the essay type is a material-based essay, the score for the theme relevance dimension is calculated based on the theme relevance. When the essay type is an application document, the score for the theme relevance dimension is calculated based on the format conformity. When the essay type is a topic-based essay, the score for the topic relevance dimension is calculated based on the relevance to the topic. The score for the structural integrity dimension is calculated based on the logical relationship label sequence; The score for the fluency dimension is calculated based on the number of grammatical errors and misspellings detected in the essay content, and with reference to the predefined number of misspellings and their corresponding relationships in the fluency dimension.

8. The method for multi-dimensional personalized essay correction based on a multimodal large model according to claim 1, characterized in that, The multi-dimensional personalized essay correction method also includes: The feedback from the reviewers is collected, and the multimodal evaluation model is iteratively optimized based on the collected feedback.

9. The method for multi-dimensional personalized essay correction based on a multimodal large model according to claim 1, characterized in that, The output formats of the personalized feedback include: The multi-dimensional scores are visualized in the form of radar charts. The targeted annotations are located and marked next to the corresponding sentences or paragraphs in the original text; Provide a comparison between the original text and the revised text with editing suggestions.

Images