A method, device and storage medium for correcting test papers and compositions

Abstract

The present application relates to a test paper composition grading method, device and storage medium provided by the present application, and belongs to the field of artificial intelligence technology. The present application obtains the test paper composition image to be graded, the test paper composition title text, the writing text to be graded, and the composition scoring standard text; the test paper composition image, the test paper composition title text, the writing text, and the composition scoring standard text are input into a multimodal large model in combination with prompt information, and the evaluation result of the test paper composition to be graded is output. The present application not only has the plain text of the test paper composition as input, but also sets an evaluation dimension for the overall content of the test paper for the test paper composition image, so as to avoid certain limitations on the accuracy of the scoring result. In addition, an alignment mechanism for the test paper composition scoring process is added to alleviate the problem of hallucinations in the existing test paper composition grading method based on a large model, thereby greatly improving the overall effect of the test paper composition grading task.

Description

Technical Field

[0001] The present application belongs to the field of artificial intelligence technology, and specifically relates to a test paper composition grading method, device and storage medium. Background Art

[0002] In the field of education, grading test papers is a tedious and important task. Since grading test papers is highly subjective, the automatic grading system for test papers based on deep learning models has been difficult to implement.

[0003] With the rapid development and application of artificial intelligence technology, especially the application of large models, its powerful understanding of text and images and cross-domain capabilities have made the automatic examination and composition grading system based on deep learning models a research hotspot.

[0004] The existing invention patent with publication number CN118428340A proposes a composition correction method, device, electronic device and storage medium. The specific steps are as follows:

[0005] S1. Obtaining the writing category of the composition to be corrected and the composition information of the composition to be corrected;

[0006] S2. According to the writing category, determine the correction dimension corresponding to the writing category;

[0007] S3, obtaining the prompt template corresponding to the correction dimension;

[0008] S4, based on the composition information and the prompt template, generating a prompt text, the prompt text is used to guide the pre-trained composition correction model to generate a corresponding correction result;

[0009] S5, inputting the prompt text into the composition correction model, and the composition correction model corrects the composition to be corrected according to the prompt text and outputs the correction result corresponding to the correction dimension;

[0010] S6. Obtain the correction results output by the composition correction model and display the correction results.

[0011] The above method enables the large model to automatically grade test papers and essays, but there is a certain illusion in the evaluation results and the evaluation accuracy is limited. Summary of the invention

[0012] In view of the above analysis, the embodiments of the present invention aim to provide a test paper composition marking method, device and storage medium to solve the problem of low evaluation accuracy in the existing automated test paper composition marking technology.

[0013] In a first aspect of the present application, a method for correcting test papers and compositions is provided, comprising:

[0014] Obtain the test paper composition image to be corrected, the test paper composition title text, the writing text to be corrected, and the composition scoring standard text;

[0015] Inputting the test paper composition image, the test paper composition title text, the writing text, the composition scoring standard text into the multimodal large model in combination with the prompt information, and outputting the evaluation result of the test paper composition to be corrected;

[0016] The multimodal large model is pre-fine-tuned and trained, and the multimodal large model fine-tuning training process includes:

[0017] Creating a basic test paper composition correction training data set; the basic test paper composition correction training data set includes multiple composition samples, each composition sample has a test paper composition image, a test paper composition title text, a writing text to be corrected, a composition scoring standard text, and overall paper content scoring data, scoring data of each scoring dimension obtained according to the composition scoring standard, and total scoring data;

[0018] Use the multimodal large model to fine-tune the basic test paper composition correction ability on the created basic test paper composition correction training dataset;

[0019] Creating an alignment training data set for the test paper composition scoring process; the alignment training data set for the test paper composition scoring process maps the original fine-grained scoring data to the coarse-grained segmented interval according to the numerical range, and the obtained segmented interval is added to the output field;

[0020] Based on the multimodal large model that has been fine-tuned for basic test essay grading capabilities, the essay scoring process alignment and fine-tuning are completed on the test essay scoring process alignment training dataset.

[0021] Optionally, the step of creating a basic test paper composition correction training data set includes:

[0022] Collect large-scale test essay samples;

[0023] Cleaning and standardizing the text data in the test paper composition sample, and performing image preprocessing on the test paper composition image;

[0024] Obtaining manually annotated data of the test paper composition sample, the annotated data including content grade, content comment, vocabulary and grammar grade, vocabulary and grammar comment, chapter structure grade, chapter structure comment, overall paper content grade, total score, and total comment;

[0025] The labeled data is integrated into a structured database file for use in model training.

[0026] Optionally, mapping the original fine-grained scoring data to the coarse-grained segmentation interval according to the numerical range includes:

[0027] Traverse every sample in the data set;

[0028] For each sample’s fine-grained score, the corresponding coarse-grained interval is obtained using the mapping rule;

[0029] The resulting segment intervals are added as new fields to the output fields of the sample.

[0030] Optionally, completing the composition scoring process alignment fine-tuning on the test paper composition scoring process alignment training data set includes:

[0031] The model is trained simultaneously using fine-grained scoring data of each scoring dimension and coarse-grained segmentation intervals to achieve process alignment and consistency.

[0032] Optionally, the output data of the test paper composition scoring process aligned with the training data set includes:

[0033] Content level, content comments, vocabulary and grammar level, vocabulary and grammar comments, paragraph structure level, paragraph structure comments, overall paper content level, total score, total comments, and segmentation ranges.

[0034] Optionally, the test composition scoring process is aligned with the training data set and output in JSON format.

[0035] Optionally, completing the composition scoring process alignment fine-tuning on the test paper composition scoring process alignment training data set includes:

[0036] The original loss function of the multimodal large model is denoted as L caption After introducing the alignment mechanism of the test paper composition scoring process, the final loss function is recorded as L final , defined as follows:

[0037] L final

[0038] =α*L regression +(1-α)

[0039] *L caption

[0040] Among them, L regression It represents the alignment loss function term in the test essay scoring process. The total score data and the corresponding field content of the segmented interval in the output data are taken to calculate the loss. α represents the adjustable parameter of the loss function term.

[0041] Optionally, the value of α is 0.6.

[0042] In a second aspect of the present application, a test paper composition marking device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the test paper composition marking method according to any one of the above-mentioned methods is implemented.

[0043] According to a third aspect of the present application, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the method for grading test papers and compositions described above is implemented.

[0044] The test composition correction method provided in this application obtains the test composition image to be corrected, the test composition title text, the writing text to be corrected, and the composition scoring standard text; the test composition image, the test composition title text, the writing text, and the composition scoring standard text are input into the multimodal large model in combination with the prompt information, and the evaluation result of the test composition to be corrected is output. This application not only has the pure text of the test composition as input, but also sets an evaluation dimension for the overall content of the test paper for the test composition image, so as to avoid certain limitations on the accuracy of the scoring result. In addition, while the multimodal large model generates the composition evaluation result, the reasoning process description from the coarse-grained segmented interval to the fine-grained specific score is performed for the scoring result, that is, the test composition scoring process alignment mechanism is added, which allows the multimodal large model to further complete the scoring process alignment based on the evaluation composition scoring result, alleviates the problem of hallucination in the existing test composition correction method based on the large model, thereby greatly improving the overall effect of the test composition correction task.

[0045] In addition, the present application also provides an examination paper composition grading device and storage medium having the above-mentioned technical effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of this specification. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0047] Figure 1 A flowchart of a specific implementation method of the test paper composition correction method provided by this application;

[0048] Figure 2 Flowchart of the training process for fine-tuning a large multimodal model;

[0049] Figure 3 This is a structural diagram of the test paper composition marking equipment provided by this application. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. It should be noted that, in the absence of conflict, the embodiments in the present disclosure and the features in the embodiments can be combined, separated, interchanged and / or rearranged with each other. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0051] The terms used here are for the purpose of describing specific embodiments, and are not intended to be restrictive. As used here, unless the context clearly indicates otherwise, the singular forms "one (kind, person)" and "said (the)" are also intended to include plural forms. In addition, when the terms "comprise" and / or "include" and their variations are used in this specification, it is explained that there are stated features, integral bodies, steps, operations, parts, assemblies and / or their groups, but it is not excluded that there are or add one or more other features, integral bodies, steps, operations, parts, assemblies and / or their groups. It should also be noted that, as used here, the terms "substantially", "approximately" and other similar terms are used as approximate terms and not as degree terms, so that they are used to explain the inherent deviations of the measured values, calculated values ​​and / or the values ​​provided that will be recognized by those of ordinary skill in the art.

[0052] A flowchart of a specific implementation method of the test paper composition correction method provided in this application is as follows Figure 1 As shown, the method specifically includes:

[0053] S101: Obtaining an image of a test paper composition to be corrected, a text of a test paper composition title, a writing text to be corrected, and a text of a composition scoring standard.

[0054] Specifically, obtain the image of the test paper composition to be corrected, recorded as I; the test paper composition title text, recorded as T; the writing text to be corrected, recorded as A; and the composition scoring standard text, recorded as S.

[0055] To obtain the image of the test paper to be corrected, a scanner can be used to convert the test paper text into a digital image. After obtaining the digital image, the image can be preprocessed, including but not limited to adjusting the image size, cropping, removing noise, and improving contrast, so as to display the text content more clearly.

[0056] Extract text data such as test paper composition title text, writing text to be corrected, and composition scoring standard text. For text data, the text data can be pre-processed, including but not limited to removing extra spaces, punctuation correction, and grammar checking.

[0057] S102: Input the test paper composition image, the test paper composition title text, the writing text, the composition scoring standard text into the multimodal large model in combination with prompt information, and output the test paper composition evaluation result to be corrected.

[0058] The test composition image I, the test composition title text T, the writing text to be corrected A, and the composition scoring standard text S are input into the multimodal large model in combination with the prompt words to obtain the evaluation result of the test composition to be corrected, which is recorded as R.

[0059] Unlike traditional essay grading models, the model used in this application is a multimodal large model that can input not only text data but also image data, which allows the model to consider the image data of the test paper when evaluating the essay. This means that the model must not only score based on the text content, but also conduct a comprehensive evaluation based on the quality of the paper, typesetting, etc. For example, whether the handwriting is clear, whether the text is standardized, etc., these factors cannot be directly processed by traditional language models, but through image input, the model can analyze the overall quality of the test paper from a global perspective.

[0060] Among them, the multimodal large model described in this embodiment is pre-fine-tuned and trained, and the model can be deployed on a cloud server. The multimodal large model can use existing models in the industry, such as CogVLM, and can also use other multimodal large models, such as Qwen-VL, GPT-4V, etc., which can also achieve the same effect as the present invention.

[0061] like Figure 2 As shown in the flowchart of the multimodal large model fine-tuning training process, the multimodal large model fine-tuning training process may specifically include the following steps:

[0062] S201: Create a basic test paper composition correction training data set.

[0063] The basic test paper composition correction training data set includes multiple composition samples, each composition sample has a test paper composition image, a test paper composition title text, a writing text to be corrected, a composition scoring standard text, as well as overall paper content scoring data, scoring data of each scoring dimension obtained according to the composition scoring standard, and total scoring data.

[0064] The basic test paper composition correction training data set of this embodiment is derived from manual annotation. Creating a basic test paper composition correction training data set includes: collecting large-scale test paper composition samples; cleaning and standardizing the text data in the test paper composition samples, and performing image preprocessing on the test paper composition images; obtaining the annotation data of the test paper composition samples manually annotated, and the annotation data includes content level, content comment, vocabulary and grammar level, vocabulary and grammar comment, chapter structure level, chapter structure comment, overall paper content level, total score, and total comment; integrating the annotation data into a structured database file for model training.

[0065] The process specifically involves collecting multimodal data including test paper composition images, question texts, writing texts, and scoring standard texts. Each sample is scored at a fine-grained level, including but not limited to content, grammar, structure, and overall score. The text is cleaned and standardized, and the image is preprocessed as necessary, such as resizing and denoising.

[0066] Select a pre-trained model suitable for multimodal input, such as a Transformer-based model, and perform fine-tuning training. During the fine-tuning process, by setting hyperparameters such as learning rate and batch size, start training the model on the basic dataset so that it learns how to output the corresponding score based on the given input (image + text).

[0067] This application uses a multimodal large model architecture, which can give an overall paper content evaluation from a global perspective based on the test paper composition image, which is lacking in the large language model solution. Therefore, on the basis of conventional scoring standards, this application provides comprehensive input data for the model so that the model can comprehensively consider various dimensions of the test paper composition (such as content, vocabulary and grammar, chapter structure, etc.) for scoring, and also adds the overall paper content evaluation dimension. Based on the evaluation and analysis of the test paper composition image to be corrected, the evaluation results of the overall paper content can be obtained, simulating the influence of subjective factors of humans in the marking process, so that the evaluation results are closer to reality.

[0068] As a specific implementation method, the multimodal large model of the present application analyzes the visual information in the test paper composition image and evaluates the neatness, writing clarity, format standardization, etc. of the test paper according to the set scoring criteria to obtain the overall test paper content scoring data. The deep learning model is used for training so that it can predict the test paper content score based on image features. The model can automatically process the input test paper image and output the test paper content scoring result.

[0069] As a specific example, the training data set format of this application is as follows:

[0070] {"input":" Prompt{T,A,S}","target":"Answer"}

[0071] Among them, "input" means the complete input, including Prompt, input image I, test essay title text T, candidate writing text A, test essay scoring standard text S, and "target" means the test essay evaluation result R.

[0072] The specific prompt example is as follows:

[0073] #Quest Background

[0074] -You are grading essays.

[0075] -I will give you the [essay topic], [candidate writing] and [grading criteria].

[0076] #Task Description

[0077] The first step is that you need to first read the short essay in the [Essay Topic], then read the opening words of the two paragraphs that need to be continued, and finally read the writing content of [Candidate Writing].

[0078] The second step is to refer to the content-related standards in the [Evaluation Criteria] to judge the content level of the two continued paragraphs and give comments.

[0079] The third step is to refer to the vocabulary and grammar related standards in the [Evaluation Criteria] to judge the vocabulary and grammar level of the two continued paragraphs and give comments.

[0080] The fourth step is to refer to the relevant standards of paragraph structure in the [Evaluation Criteria] to judge the paragraph structure level of the two continued paragraphs and give comments.

[0081] Step 5. Based on the information obtained in the previous steps, determine the total score of the two continued paragraphs and give an overall evaluation.

[0082] #Output format

[0083] Content level

[0084] Content Comments

[0085] Vocabulary and grammar level

[0086] Lexical and grammatical comments

[0087] Chapter structure level

[0088] Comments on chapter structure

[0089] Overall content level:

[0090] Total score

[0091] Overall

[0092]

Essay title

[0093] {T}

[0094]

Evaluation Criteria

[0095] {S}

[0096]

Candidate Writing

[0097] {A}

[0098] Output analysis and scoring results

[0099] The input of the model includes: img (image data of the test paper composition), which enables the model to evaluate the content of the paper from a global perspective (such as handwriting, typesetting, format, etc.); T: the title text of the test paper composition, which provides the theme or writing requirements of the composition. A: The writing text to be corrected, the model will score based on this text. S: The scoring standard text defines the specific standards for evaluation, including scoring requirements for various dimensions such as content, vocabulary and grammar, and paragraph structure. Prompt: Prompts to guide the model on how to interpret the input and score. Usually contains a description of the task and specific requirements.

[0100] Target: The evaluation result output by the model, which includes multi-dimensional scores, such as content level, content comments, vocabulary and grammar level, vocabulary and grammar comments, text structure level, text structure comments, overall paper content level (evaluation of paper content) and total score, etc.

[0101] S202: Use the multimodal large model to fine-tune the basic test paper composition grading capability on the created basic test paper composition grading training data set.

[0102] Use the multimodal big model to complete the fine-tuning of the basic test paper composition correction ability on the basic test paper composition correction training data set created above, and enhance the basic test paper composition correction ability of the multimodal big model. The fine-tuning method can use the mainstream big model fine-tuning method, such as full parameter fine-tuning, LoRA, etc.

[0103] S203: Create a test essay scoring process alignment training data set.

[0104] The test paper composition scoring process aligns the training data set to map the original fine-grained scoring data to the coarse-grained segmented interval according to the numerical range, and the obtained segmented interval is added to the output field.

[0105] The main differences between creating a test essay scoring process alignment training dataset and creating a basic test essay correction training dataset are the granularity of data processing, the adjustment of the output format, and the addition of a new dimension of "segmentation interval". Based on the above basic test essay correction training dataset, the fine-grained scoring results are mapped to the coarse-grained segmentation interval (such as mapping the specific score range to the grade interval of 1-5), and the fine-grained scoring results and the coarse-grained segmentation interval are output at the same time. The purpose of this process is to ensure the consistency of the reasoning process, that is, the model should not only give accurate scores, but also give a wider range of scoring intervals in the same process, so as to make a higher-level summary and understanding of the scoring results.

[0106] For example, if the fine-grained score is 85 points, it may be mapped to the coarse-grained segmentation interval "4" or "4-5" interval (assuming the segmentation interval is 1-5). This mapping can help the model provide a more intuitive understanding and explanation when processing scores, avoiding over-reliance on a single score.

[0107] Mapping the original fine-grained scoring data to the coarse-grained segmentation interval according to the numerical range includes: traversing each sample in the data set; for the fine-grained score of each sample, using the mapping rule to obtain the corresponding coarse-grained interval; and adding the obtained segmentation interval as a new field to the output field of the sample.

[0108] The model is trained simultaneously using fine-grained scoring data of each scoring dimension and coarse-grained segmentation intervals to achieve process alignment and consistency.

[0109] The output data of the test paper composition scoring process aligned with the training data set includes: content level, content comment, vocabulary and grammar level, vocabulary and grammar comment, chapter structure level, chapter structure comment, overall paper content level, total score, total comment, and segment interval.

[0110] The test paper composition scoring process is aligned with the training data set and output in JSON format.

[0111] The test composition scoring process alignment training data template of the embodiment of the present application is processed based on the above-mentioned basic test composition correction training data set, that is, the original fine-grained scoring result is mapped to N coarse-grained segmentation intervals according to the numerical range, and N can be 5, so that the model outputs the final fine-grained scoring result while also outputting the coarse-grained segmentation interval to ensure the consistency of the reasoning process. The test composition scoring process alignment training data template of the embodiment of the present application is consistent with the above-mentioned basic test composition correction training data set, and the output format of the model required in the Prompt is adjusted to the JSON format, and the segmentation interval process supervision information is added.

[0112] The following is an example of the adjusted prompt:

[0113] #Quest Background

[0114] -You are grading essays.

[0115] -I will give you the [essay topic], [candidate writing] and [grading criteria].

[0116] #Task Description

[0117] The first step is that you need to first read the short essay in the [Essay Topic], then read the opening words of the two paragraphs that need to be continued, and finally read the writing content of [Candidate Writing].

[0118] The second step is to refer to the content-related standards in the [Evaluation Criteria] to judge the content level of the two continued paragraphs and give comments.

[0119] The third step is to refer to the vocabulary and grammar related standards in the [Evaluation Criteria] to judge the vocabulary and grammar level of the two continued paragraphs and give comments.

[0120] The fourth step is to refer to the relevant standards of paragraph structure in the [Evaluation Criteria] to judge the paragraph structure level of the two continued paragraphs and give comments.

[0121] Step 5. Based on the information obtained in the previous steps, determine the total score of the two continued paragraphs and give an overall evaluation.

[0122] #Output format

[0123] *Strictly output in JSON format

[0124] {

[0125] "Content Level"

[0126] "Content Comments"

[0127] "Lexical and Grammatical Level"

[0128] "Lexical and grammatical comments"

[0129] "Section Structure Level"

[0130] "Chapter Structure Comments"

[0131] "Overall content level"

[0132] "Total score"

[0133] "Overall Review"

[0134] "Segment interval"

[0135] }

[0136]

Essay title

[0137] {T}

[0138]

Evaluation Criteria

[0139] {S}

[0140]

Candidate Writing

[0141] {A}

[0142] Output analysis and scoring results:

[0143] S204: Based on the multimodal large model that has been fine-tuned for basic test essay grading capabilities, the essay scoring process alignment fine-tuning is completed on the test essay scoring process alignment training data set.

[0144] Based on the above multimodal big model that has been fine-tuned for basic test essay grading capabilities, the test essay scoring process is aligned and fine-tuned. Specifically, the multimodal big model is used to align the training data set in the test essay scoring process created above to complete fine-tuning. The fine-tuning method can use mainstream big model fine-tuning methods, such as full parameter fine-tuning, LoRA, etc.

[0145] The original loss function of the multimodal large model is denoted as L caption , we can use cross entropy, introduce the alignment mechanism in the test paper composition scoring process, and the final loss function is recorded as L final , defined as follows:

[0146] L final

[0147] =α*L regression +(1-α)

[0148] *L caption

[0149] Among them, L regression It represents the alignment loss function item of the test paper composition scoring process. Cross entropy can be used to calculate the loss by taking the total score data and the corresponding field content of the segmented interval in the output data. α represents the adjustable parameter of the loss function item. The larger the value, the higher the dominance of the corresponding test paper composition scoring process alignment task item on the overall loss function. As a specific implementation method, the value of α can be selected as 0.6.

[0150] The existing large-model-based test paper composition grading method only relies on the plain text extracted by the OCR algorithm or other document parsing tools as input, lacks the evaluation dimension of the overall paper content, and thus limits the accuracy of the scoring results. In view of the defect of the prior art that lacks the subjective global paper evaluation dimension, this application adds the evaluation dimension of the overall paper content, has a more comprehensive understanding of the scoring criteria and logic, and alleviates the illusion problem of the large model.

[0151] The existing large-model-based test composition grading method directly outputs the evaluation results, including the scoring results and comment results, based on the test composition text to be graded and the prompt words. The scoring results lack further process alignment, and the understanding of the scoring standards and logic is insufficient, which has certain illusions. In view of the defect of the lack of direct supervision in the prior art, this application introduces a test composition scoring process alignment mechanism, so that the model output includes not only fine-grained scores, but also coarse-grained score intervals to ensure the consistency of the reasoning process.

[0152] In addition, the present application also provides a test paper composition correction device, such as Figure 3 As shown in the structural block diagram of the test paper composition marking device provided in this application, the device specifically includes a memory 31 and a processor 32, and the memory 31 stores a computer program. When the computer program is executed by the processor 32, the test paper composition marking method according to any one of the above items is implemented.

[0153] In addition, the present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the test paper composition marking method according to any one of the above-mentioned methods is implemented.

[0154] This application not only has the plain text of the test composition as input, but also sets an evaluation dimension for the overall content of the test paper for the test composition image to avoid certain limitations on the accuracy of the scoring results. In addition, while the multimodal large model generates the composition evaluation results, it describes the reasoning process from the coarse-grained segmented interval to the fine-grained specific score for the scoring results, that is, it adds a test composition scoring process alignment mechanism. This mechanism allows the multimodal large model to further complete the scoring process alignment based on the evaluation of the composition scoring results, alleviating the problem of hallucinations in the existing large-model-based test composition grading methods, thereby greatly improving the overall effect of the test composition grading task.

[0155] Computer readable storage media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0156] The professionals should also be further aware that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented with electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0157] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0158] The specific implementation methods described above further illustrate the purpose, technical solutions and beneficial effects of the present application in detail. It should be understood that the above description is only the specific implementation method of the present application and is not intended to limit the scope of protection of the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A method for correcting test papers and compositions, characterized in that: include: Obtain the test paper composition image to be corrected, the test paper composition title text, the writing text to be corrected, and the composition scoring standard text; Inputting the test paper composition image, the test paper composition title text, the writing text, the composition scoring standard text into the multimodal large model in combination with the prompt information, and outputting the evaluation result of the test paper composition to be corrected; The multimodal large model is pre-fine-tuned and trained, and the multimodal large model fine-tuning training process includes: Creating a basic test paper composition correction training data set; the basic test paper composition correction training data set includes multiple composition samples, each composition sample has a test paper composition image, a test paper composition title text, a writing text to be corrected, a composition scoring standard text, and overall paper content scoring data, scoring data of each scoring dimension obtained according to the composition scoring standard, and total scoring data; Use the multimodal large model to fine-tune the basic test paper composition correction ability on the created basic test paper composition correction training dataset; Creating an alignment training data set for the test paper composition scoring process; the alignment training data set for the test paper composition scoring process maps the original fine-grained scoring data to the coarse-grained segmented interval according to the numerical range, and the obtained segmented interval is added to the output field; Based on the multimodal large model that has been fine-tuned for basic test essay grading capabilities, the essay scoring process alignment and fine-tuning are completed on the test essay scoring process alignment training dataset.

2. The method for correcting test papers and compositions according to claim 1, characterized in that: The creation of a basic test paper composition correction training data set includes: Collect large-scale test essay samples; Cleaning and standardizing the text data in the test paper composition sample, and performing image preprocessing on the test paper composition image; Obtaining manually annotated data of the test paper composition sample, the annotated data including content grade, content comment, vocabulary and grammar grade, vocabulary and grammar comment, chapter structure grade, chapter structure comment, overall paper content grade, total score, and total comment; The labeled data is integrated into a structured database file for use in model training.

3. The method for correcting test papers and compositions according to claim 2, characterized in that: Mapping the original fine-grained scoring data to coarse-grained segmentation intervals according to the numerical range includes: Traverse every sample in the data set; For each sample’s fine-grained score, the corresponding coarse-grained interval is obtained using the mapping rule; The resulting segment intervals are added as new fields to the output fields of the sample.

4. The method for correcting test papers and compositions according to claim 3, characterized in that: The method of completing the composition scoring process alignment fine-tuning on the test paper composition scoring process alignment training data set includes: The model is trained simultaneously using fine-grained scoring data of each scoring dimension and coarse-grained segmentation intervals to achieve process alignment and consistency.

5. The method for correcting test papers and compositions according to claim 3, characterized in that: The output data of the test paper composition scoring process alignment training data set includes: Content level, content comments, vocabulary and grammar level, vocabulary and grammar comments, paragraph structure level, paragraph structure comments, overall paper content level, total score, total comments, and segmentation ranges.

6. The method for correcting test papers and compositions according to claim 4, characterized in that: The test paper composition scoring process is aligned with the training data set and output in JSON format.

7. The method for correcting test papers and compositions according to any one of claims 1 to 6, characterized in that: The method of completing the composition scoring process alignment fine-tuning on the test paper composition scoring process alignment training data set includes: The original loss function of the multimodal large model is denoted as L caption After introducing the alignment mechanism of the test paper composition scoring process, the final loss function is recorded as L final , defined as follows: L final =α*L regression +(1-a)*L caption Among them, L regression It represents the alignment loss function term in the test essay scoring process. The total score data and the corresponding field content of the segmented interval in the output data are taken to calculate the loss. α represents the adjustable parameter of the loss function term.

8. The method for correcting test papers and compositions according to claim 7, characterized in that: The value of α is 0.

6.

9. A test paper composition marking device, characterized in that: It comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the test paper composition marking method according to any one of claims 1-8 is implemented.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed by a processor, the method for correcting test papers and compositions according to any one of claims 1-8 is implemented.

Images