An automatic correction method, device, system and readable storage medium

Abstract

An automatic correction method, device, system and readable storage medium are applied to the field of multi-modal technology. A pre-training data set is constructed; a layoutlmv2 pre-training model is constructed based on the pre-training data set; a correction model is constructed based on the pre-training model; and the output result of the correction model is post-processed. The application constructs a correction model, compared with a traditional correction scheme based on rules or pure text, introduces student answer image information into the model, reduces the dependence of the correction model on an OCR module, can correct fault-tolerant samples according to image information when the OCR module misses or incorrectly recognizes, and greatly improves the accuracy of automatic correction. Moreover, the application also adds a post-processing link, which can greatly reduce the samples that are misjudged and mistaken due to non-standard writing, effectively improving the correction accuracy of the model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of multi-modal technology, and in particular to an automatic correction method, device, system and readable storage medium. BACKGROUND

[0002] With the continuous development of Internet technology, users use information means to automatically correct homework, thereby improving the efficiency of homework correction, which has gradually become a technical trend. Among them, the automatic correction of fill-in-the-blank questions is based on the picture of the student's answer and the text of the standard answer or the stem information, and the model automatically corrects the right or wrong.

[0003] Automatic correction mainly includes rule-based methods and deep learning-based methods. Among them, the rule-based method needs to first identify the text content in the student's answer picture according to OCR, then calculate the edit distance between the answer text and the standard answer, and set a threshold value, and then determine the right or wrong according to the relationship between the edit distance and the threshold value; the deep learning-based method takes the stem information, student answer and standard answer as input, trains a deep neural network, and outputs the similarity score between the answer and the standard answer, thereby realizing the automatic correction of fill-in-the-blank questions.

[0004] However, when the existing automatic correction method misses and misidentifies the student's writing, and when the student's answer contains correction or sequence-changing symbols, the recognized student's correct answer will be inconsistent with the standard answer. Moreover, OCR recognition cannot recognize incorrectly written text and will only output the correct text closest to the incorrectly written text, which will result in the recognized student's incorrect answer being consistent with the standard answer. Therefore, how to improve the accuracy of automatic correction has become a technical problem to be solved. SUMMARY

[0005] The present application provides an automatic correction method, device, system and readable storage medium, which can improve the accuracy of automatic correction.

[0006] The present application discloses the following technical solutions:

[0007] In a first aspect, the present application discloses an automatic correction method, comprising:

[0008] constructing a pre-training data set;

[0009] pre-training a layoutlmv2 model based on the pre-training data set;

[0010] constructing an automatic correction model based on the layoutlmv2 model;

[0011] inputting the answer information into the automatic correction model to obtain the output result of the automatic correction model.

[0012] Preferably, the pre-training layoutlmv2 model comprises:

[0013] The mask language model pre-training task, the text-image alignment pre-training task, and the text-image matching pre-training task are used to pre-train the layoutlmv2 model.

[0014] Preferably, the method further comprises post-processing the output result of the automatic marking model.

[0015] Preferably, the automatic marking model is constructed by:

[0016] The model input of the layoutlmv2 model is modified to "student answer picture + stem information + student answer text + standard answer", and the automatic marking model is constructed.

[0017] In a second aspect, the present application discloses an automatic marking device, which comprises a first construction module, a pre-training module, a second construction module, and an automatic marking module.

[0018] The first construction module is configured to construct a pre-training data set.

[0019] The pre-training module is configured to pre-train a layoutlmv2 model based on the pre-training data set.

[0020] The second construction module is configured to construct an automatic marking model based on the layoutlmv2 model.

[0021] The automatic marking module is configured to input answer information into the automatic marking model and obtain an output result of the automatic marking model.

[0022] Preferably, the pre-training module is specifically configured to construct a mask language model pre-training task, a text-image alignment pre-training task, and a text-image matching pre-training task, and pre-train the layoutlmv2 model.

[0023] Preferably, the device further comprises a post-processing module.

[0024] The post-processing module is configured to post-process the output result of the automatic marking model.

[0025] Preferably, the second construction module is specifically configured to modify the model input of the layoutlmv2 model to "student answer picture + stem information + student answer text + standard answer", and construct an automatic marking model.

[0026] In a third aspect, the present application discloses an automatic marking device, which comprises a memory and a processor.

[0027] The memory is configured to store a program.

[0028] The processor is configured to execute the program to implement each step of the method of the first aspect.

[0029] In a fourth aspect, the present application discloses a computer storage medium, which stores a computer program, and the computer program is executed by a processor to implement each step of the method of the first aspect.

[0030] Compared with the prior art, the present application has the following beneficial effects:

[0031] The present application provides an automatic correction method, device, system and readable storage medium. Compared with the traditional correction scheme based on rules or pure text, the student answer image information is introduced into the model, the dependence of the correction model on the OCR module is reduced, the samples missed by the OCR or incorrectly recognized can be corrected according to the image information, and the accuracy of automatic correction is greatly improved. Moreover, the present application also adds a post-processing link, which can greatly reduce the samples that are misjudged or wrong due to non-standard writing, and effectively improve the accuracy of model correction. BRIEF DESCRIPTION OF DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0033] Figure 1 A schematic diagram of defects of a first automatic correction method provided by the embodiment of the present application;

[0034] Figure 2 A schematic diagram of defects of a second automatic correction method provided by the embodiment of the present application;

[0035] Figure 3 A schematic diagram of defects of a third automatic correction method provided by the embodiment of the present application;

[0036] Figure 4 A schematic diagram of defects of a fourth automatic correction method provided by the embodiment of the present application;

[0037] Figure 5 A flowchart of an automatic correction method provided by the embodiment of the present application;

[0038] Figure 6 A comparison diagram of pure text correction and multi-modal correction effects provided by the embodiment of the present application;

[0039] Figure 7 Schematic diagram of an automatic marking device provided by an embodiment of this application. Specific implementation manners

[0040] First, the technical terms involved in this application will be introduced below.

[0041] Multimodal refers to information of multiple modalities, including: text, image, video, audio, etc., and collaborative reasoning of multiple heterogeneous modality data. In this application, multimodal refers to two modalities of output, namely text and image, specifically the student's answer result and the student's answer picture.

[0042] OCR (optical character recognition) recognition refers to the process in which an electronic device (such as a scanner or a digital camera) examines the characters printed on paper and then translates the shape into computer text using character recognition methods. That is, the process of scanning text materials and then analyzing and processing the image file to obtain text and layout information.

[0043] Automatic marking is mainly divided into rule-based methods and deep learning-based methods. However, there are many defects in existing automatic marking methods. The specific defects are as follows:

[0044] First, the automatic marking methods of the prior art, that is, rule-based and deep learning-based methods, are strongly dependent on OCR detection. When OCR misses detection or misdetects, resulting in the recognition result being inconsistent with the standard answer, the marking result will be misjudged as incorrect. Refer to Figure 1 , this figure is a schematic diagram of the defect of the first automatic marking method provided by an embodiment of this application. When OCR recognizes the character "驾" is missed, the recognition result is "a small boat", which is different from the standard answer "驾一叶之扁舟". However, the answer written by the student is exactly the standard answer "驾一叶之扁舟", that is, the accuracy of the existing automatic marking method is low.

[0045] Second, the automatic marking methods of the prior art are strongly dependent on OCR recognition. OCR recognition usually misrecognizes characters with irregular writing as similar characters. When OCR recognition is incorrect, the marking result is misjudged as incorrect. Refer to Figure 2 , this figure is a schematic diagram of the defect of the second automatic marking method provided by an embodiment of this application. OCR recognizes "典" as "共", and the recognition result is "开过大共", which is different from the standard answer "开国大典". However, the answer written by the student is exactly the standard answer "开国大典", that is, the accuracy of the existing automatic marking method is low.

[0046] Third, the automatic marking methods of the prior art are likely to misrecognize the errors in the answer details and judge them as correct. That is, OCR recognition cannot recognize the part with writing errors (i.e., wrong characters), and will output the correct text closest to the writing wrong character, resulting in the marking result being misjudged as correct. Refer toFigure 3 The figure is a schematic diagram of the third automatic correction method defect provided by the embodiment of the application. The OCR recognition will identify the wrong word as "Han". The answer written by the student is a wrong word, but the OCR automatic correction will correct it as the standard answer. That is, the existing automatic correction method has a low accuracy.

[0047] Fourth, the existing automatic correction method is sensitive to the answer order. The conventional reading order of the OCR recognition result is usually returned in the order from top to bottom and from left to right. When the answer contains a correction or a change order symbol, the recognition result is different from the standard answer, resulting in a correction error. See Figure 4 The figure is a schematic diagram of the fourth automatic correction method defect provided by the embodiment of the application. The OCR recognition result is "Yi Yan Zai", which is different from the standard answer "Yi Yan Zai". However, the answer written by the student is the standard answer "Yi Yan Zai". That is, the existing automatic correction method has a low accuracy.

[0048] In view of the above defects, the present application provides an automatic correction method, device, system and readable storage medium. A correction model is constructed. Compared with the traditional rule-based or pure text correction scheme, the student answer image information is introduced into the model, the dependence of the correction model on the OCR module is reduced, the samples missed by the OCR or incorrectly recognized can be corrected according to the image information, and the accuracy of the automatic correction is greatly improved. Moreover, the present application also adds a post-processing link, which can greatly reduce the samples that are misjudged or wrong due to non-standard writing, and effectively improve the correction accuracy of the model.

[0049] In order for those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0050] Referring to Figure 5 The figure is a flowchart of an automatic correction method provided by the embodiment of the present application.

[0051] S101: Construct a pre-training data set.

[0052] First, a pre-training data set is constructed. The process of constructing the pre-training data set is to extract the relevant data of the fill-in-the-blank questions and their answers from various test papers and question banks, remove duplicate data and invalid data, thereby constructing a fill-in-the-blank question correction pre-training data set for subsequent model pre-training. In an embodiment, the test paper can be a Chinese test paper, a mathematics test paper, etc., which is not limited by the present application.

[0053] Therefore, the pre-training task corresponding to the education field is designed, that is, a large amount of fill-in-the-blank data in the education scene is collected for pre-training, a large-scale fill-in-the-blank database is constructed, and the field-related knowledge is integrated into the model. The large-scale pre-training of the data can effectively improve the performance of the downstream task in the related scene. Through the combination of "large-scale pre-training + downstream task fine-tuning", the accuracy and recall rate of correction can be effectively improved.

[0054] S102: Based on the pre-training data set, a layoutlmv2 pre-training model is constructed.

[0055] Based on the layoutlmv2 model, three pre-training tasks are constructed for the education correction scene, that is, the pre-training data set constructed in S101, and the layoutlmv2 model is pre-trained. The pre-training model can store the knowledge implicitly in the model parameters by designing specific tasks and large-scale field-related data driving. The model with prior knowledge can effectively improve the accuracy of the downstream task.

[0056] The first is the Masked Visual-Language Model (MVLM) pre-training task, that is, the masked language model pre-training task. The masked language model randomly covers some words in the document, and the pre-training model restores these covered words. Through cross-modal clues, the model learns better on the language side.

[0057] In some possible implementations, the picture answered by the student can be recognized as answer text through OCR. The masked language model randomly covers the words in the answer text, and the pre-training model output layer restores the covered words. The OCR (optical character recognition) recognition refers to the process in which an electronic device (such as a scanner or a digital camera) checks the characters printed on paper, and then translates the shapes into computer text by character recognition method. That is, the text material is scanned, and then the image file is analyzed and processed to obtain the text and layout information. In this way, the pre-training task is designed, so that the pre-training model constructed can learn more language information.

[0058] The second is the Text-ImageAlignment (TIA) pre-training task, that is, the text-image alignment pre-training task. After the text-image alignment pre-training task is detected by OCR, the image is covered according to the character granularity, and the image corresponding to the character on the language side is predicted whether it is covered or not, which is used for fine-grained alignment between the image side and the language side.

[0059] In some possible implementations, the text-image alignment pre-training task can be covered according to the word level. See Figure 6Fig. 2 is a coverage diagram of an automatic correction method according to an embodiment of the present application. In the phrase "So I teach, impart knowledge and dispel ignorance", the word "industry" is randomly covered, and the language side predicts whether the word "industry" is covered, so that the pre-training model learns more language information.

[0060] Thirdly, the Text-Image Matching (TIM) pre-training task, i.e., the text-image matching pre-training task. The text-image matching pre-training task, i.e., the image-text matching, is used to predict whether the picture and the text correspond. In the text-image matching pre-training task, the document picture and the corresponding text form a positive sample pair, and the text in the document and the randomly replaced picture form a negative sample pair. The positive sample is a sample in which the picture and the corresponding text match, and the negative sample is a sample in which the picture and the text do not match after random replacement.

[0061] In some possible embodiments, the model predicts whether the image and the text match by binary classification, in which 0 represents an incorrect answer and 1 represents a correct answer. Other classification methods can also be used to predict whether the image and the text match, which is not limited in the present application.

[0062] In some possible embodiments, since the incorrect answers of students are usually a small part of errors, the negative sample construction can adopt three ways: first, randomly replaced with other answer pictures; second, deleting or adding some answer words; and third, maintaining a replacement word table to replace some answer words with similar words. After the negative sample is constructed, whether the image and the text match is predicted at the CLS. The CLS is a special marker placed at the beginning of the input for classification, i.e., a clock signal.

[0063] It can be seen that, compared with the traditional rule-based or pure text-based correction scheme, the present application introduces the stem information and the student answer image information into the model on the basis of pure text correction, reduces the dependence of the correction model on the OCR module, can correct the samples missed by OCR or with recognition errors according to the image information, and thus effectively improves the generalization and the correction accuracy of the model. Specifically, after the stem information is introduced into the model input, the model can assist in correction on the basis of understanding the stem, and effectively improve the correction accuracy. Moreover, the student answer picture is coded, and the coded stem information, the student answer text, and the correct answer text are input into the model, the model has the fault tolerance for part of the OCR detection and recognition by integrating the image features, and reduces the accumulation of errors of the OCR to the subsequent correction module.

[0064] S103: Constructing an automatic correction model based on the pre-training model.

[0065] On the basis of constructing the pre-training model in S102, fine-tuning is performed on the downstream correction task. The model input of the correction model is "student answer picture + stem information + student answer text + standard answer", that is, the model is adjusted to be able to judge whether the information in the student answer picture is correct. The model output of the correction model is a binary classification at [CLS], predicting answer (1) or (0), where 0 represents an incorrect answer and 1 represents a correct answer in the binary classification method.

[0066] S104: input the answer information into the automatic correction model to obtain the output result of the automatic correction model.

[0067] S105: post-processing the output result of the automatic correction model.

[0068] After correcting the student's answer on the basis of constructing the correction model, post-processing is performed. For the answer that is incorrect according to the model correction result, the first few prediction results of each step are returned according to the OCR recognition, the first few candidate recognition results are returned using a specific algorithm, and the first few candidate results are compared with the standard answer through a string. If one of the candidate results hits the standard answer, the correction result is corrected to be correct. In some possible embodiments, the first few prediction results can be the first five, the first six, etc., and the specific algorithm can be a beam-search algorithm, which is not limited in the present application.

[0069] In some possible embodiments, referring to Figure 2 , the figure is a schematic diagram of the second automatic correction method defect provided by the embodiment of the present application. The answer written by the student is the correct answer "the founding ceremony", but the OCR recognition is incorrect, and "the founding ceremony" is recognized as "the founding common", and the student's writing is judged to be incorrect. However, after introducing the post-processing link of the present application, the first five prediction results can be returned according to the OCR recognition using the beam-search algorithm, and the first 5 recognition results returned by the OCR recognition are "the founding common", "the founding common", "the founding ceremony", "the founding ceremony", and "the founding ceremony" in descending order of confidence. As long as one of the five recognition results hits "the founding ceremony", the answer is correct, which improves the accuracy of the correction model in judging the test paper.

[0070] In some possible embodiments, in the primary and secondary school scene, the student's answer is usually written in a careless and non-standard manner, which brings great challenges to OCR accurate recognition. Without context information, OCR recognition is difficult to correctly recognize some characters, but tends to recognize other similar characters, resulting in incorrect correction. The present application adds a post-processing module to the samples with incorrect model prediction results, which can greatly reduce the misjudgment of samples caused by non-standard writing, effectively improve the model correction accuracy, and keep the user's correction result as consistent as possible.

[0071] According to the above description, see Figure 6 , which is a pure text grading and multi-modal grading effect comparison diagram provided by the embodiment of the application. In the first question, the student's answer is "Therefore, to teach, to learn and to enlighten", if the user manually grades, the wrong characters can be distinguished, and if the existing technology of text OCR recognition is used, the text OCR recognition only passes the recognized result and the standard answer into the model to obtain the semantic similarity between the two, here there is only a small difference, the model judges the similarity at the semantic level, resulting in the output grading result being correct. However, using multi-modal grading can distinguish the wrong characters, thereby improving the accuracy. In the second question, the student intended to answer "Old trees and crows are gloomy", but there is a wrong character, if the user manually grades, the wrong character can be distinguished, and if the pure text grading using OCR recognition is used, it will be automatically recognized as the correct "crow", that is, the judgment is wrong, and using multi-modal grading can distinguish the wrong character, thereby improving the accuracy. Similarly, in the third question, the student intended to answer "Eight hundred miles under the command of the general", but there is a wrong character, if the user manually grades, the wrong character can be distinguished, and if the pure text grading using OCR recognition is used, it will be automatically recognized as the correct "fry", that is, the judgment is wrong, and using multi-modal grading can distinguish the wrong character, thereby improving the accuracy.

[0072] The application provides an automatic grading method, compared with the traditional rule-based or pure text grading scheme, the student answer image information is introduced into the model, the dependence of the grading model on the OCR module is reduced, the samples missed by the OCR or the samples with recognition errors can be graded and fault-tolerant according to the image information, and the accuracy of automatic grading is greatly improved. Moreover, the application adds a specific field pre-training link, constructs a large-scale fill-in-the-blank question data, integrates the field-related knowledge into the model, and effectively improves the grading accuracy and recall rate through the combination of "large-scale pre-training + downstream task fine-tuning". The application also adds a post-processing link, which can greatly reduce the samples that are misjudged due to non-standard writing, and effectively improve the model grading accuracy.

[0073] Although the application describes each operation in a specific order, this should not be understood as requiring these operations to be performed in the specific order shown or in a sequential order. In certain circumstances, multitasking and parallel processing can be advantageous.

[0074] It should be understood that each step described in the method embodiments of the present disclosure can be performed in a different order and / or in parallel. In addition, the method embodiments can include additional steps and / or omit the steps shown. The scope of the present disclosure is not limited in this respect.

[0075] See Figure 7The figure is a schematic diagram of an automatic correction device provided by an embodiment of the present application. The present application discloses an automatic correction device 700, comprising: a first construction module 701, a pre-training module 702, a second construction module 703, and an automatic correction module 704.

[0076] The first construction module 701 is configured to construct a pre-training data set.

[0077] The pre-training module 702 is configured to pre-train a layoutlmv2 model based on the pre-training data set.

[0078] The second construction module 703 is configured to construct an automatic correction model based on the layoutlmv2 model.

[0079] The automatic correction module 704 is configured to input answer information into the automatic correction model and obtain an output result of the automatic correction model.

[0080] In some possible implementation manners, the pre-training module 702 is specifically configured to construct a mask language model pre-training task, a text-image alignment pre-training task, and a text-image matching pre-training task, and pre-train the layoutlmv2 model.

[0081] In some possible implementation manners, the automatic correction device 700 further comprises a post-processing module 705 configured to post-process the output result of the automatic correction model.

[0082] The present application constructs a correction model. Compared with a traditional correction scheme based on rules or pure text, the present application introduces student answer image information into the model, reduces the dependence of the correction model on an OCR module, can correct samples that are missed or misrecognized by the OCR according to image information, and greatly improves the accuracy of automatic correction. Moreover, the present application adds a specific domain pre-training link, constructs a large-scale fill-in-the-blank question data set, integrates domain-related knowledge into the model, effectively improves the correction accuracy and recall rate by means of a combination of "large-scale pre-training + downstream task fine-tuning", and greatly reduces the number of samples that are misjudged or misrecognized due to non-standard writing, thereby effectively improving the correction accuracy of the model.

[0083] The modules described in the embodiments of the present disclosure can be implemented in the form of software, or can be implemented in the form of hardware. In some cases, the name of the module does not constitute a limitation on the module itself, for example, the first construction module can also be described as a "module for constructing a pre-training data set". The functions described above can be at least partially executed by one or more hardware logic components. For example, non-limiting examples of exemplary types of hardware logic components that can be used include field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), system on a chip (SOCs), complex programmable logic devices (CPLDs), and the like.

[0084] The embodiments of the present application also provide a corresponding generation device and a computer storage medium for implementing the schemes provided by the embodiments of the present application.

[0085] The device includes a memory and a processor, the memory is used to store instructions or codes, and the processor is used to execute the instructions or codes to enable the device to execute the light control method of any embodiment of the present application.

[0086] The computer storage medium stores codes, and when the codes are executed, the device executing the codes implements the method described in any embodiment of the present application.

[0087] The "first", "second" in the names mentioned in the embodiments of the present application are only used for name identification, and do not represent the first and second in order.

[0088] From the above description of the embodiments, those skilled in the art can clearly understand that all or part of the steps of the above-mentioned embodiment methods can be implemented by means of software plus a general hardware platform. Based on this understanding, the technical solutions of the present application can be embodied in the form of a software product. The computer software product can be stored in a storage medium, such as a read-only memory (English: read-only memory, ROM) / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network communication device such as a router) to execute the methods described in the various embodiments or some parts of the embodiments of the present application.

[0089] It should be noted that each of the embodiments described in the specification of the present application adopts a progressive mode for description, and the same or similar parts between the embodiments can be mutually referred to. Each of the embodiments focuses on the differences from other embodiments. In particular, the device and system embodiments are described more simply because they are basically similar to the method embodiments, and the relevant parts can be referred to the part of the method embodiments. The above-described device and system embodiments are only illustrative, and the units described as separate components can or can not be physically separated, and the components indicated as units can or can not be physical units, i.e., they can be located in one place or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiments according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0090] The above describes only one specific embodiment of the present application, but the protection scope of the present application is not limited to this. Any skilled person in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An automatic grading method, characterized by, The method comprises the following steps: constructing a pre-training data set comprising multiple sets of fill-in-the-blank data and answer data of the fill-in-the-blank data; pre-training a LayoutLMv2 model based on the pre-training data set, the pre-training comprising a text-image alignment pre-training task and a text-image matching pre-training task, the text-image alignment pre-training task being used to cover an image at a character granularity and train the LayoutLMv2 model to predict an image covering state at a corresponding character on a language side, and the text-image matching pre-training task being used to train the LayoutLMv2 model to judge a text-image matching relationship by constructing a positive and negative sample pair; constructing an automatic correction model based on the LayoutLMv2 model; inputting answer information into the automatic correction model to obtain an output result of the automatic correction model; if the output result of the automatic correction model indicates that the answer information is incorrect, obtaining multiple candidate output results corresponding to the answer information output by the automatic correction model; if the candidate output result is consistent with a standard answer corresponding to the answer information, determining that the answer information is correct.

2. The method of claim 1, wherein, The method of constructing the automatic correction model comprises the following steps: modifying a model input of the LayoutLMv2 model to "student answer picture + stem information + student answer text + standard answer" to construct the automatic correction model.

3. An automatic correction device characterized by comprising: The method comprises the following steps: a first construction module, a pre-training module, a second construction module, an automatic correction module, and a post-processing module; the first construction module is configured to construct a pre-training data set comprising multiple sets of fill-in-the-blank data and answer data of the fill-in-the-blank data; the pre-training module is configured to pre-train a LayoutLMv2 model based on the pre-training data set, the pre-training comprising a text-image alignment pre-training task and a text-image matching pre-training task, the text-image alignment pre-training task being used to cover an image at a character granularity and train the LayoutLMv2 model to predict an image covering state at a corresponding character on a language side, and the text-image matching pre-training task being used to train the LayoutLMv2 model to judge a text-image matching relationship by constructing a positive and negative sample pair; the second construction module is configured to construct an automatic correction model based on the LayoutLMv2 model; the automatic correction module is configured to input answer information into the automatic correction model to obtain an output result of the automatic correction model; the post-processing module is configured to, if the output result of the automatic correction model indicates that the answer information is incorrect, obtain multiple candidate output results corresponding to the answer information output by the automatic correction model, and determine that the answer information is correct if the candidate output result is consistent with a standard answer corresponding to the answer information.

4. The apparatus of claim 3, wherein, The second construction module is specifically configured to modify a model input of the LayoutLMv2 model to "student answer picture + stem information + student answer text + standard answer" to construct the automatic correction model.

5. An automatic correction device characterized by comprising: The method comprises the following steps: a memory and a processor; The memory is configured to store a program. The processor is configured to execute the program to implement each step of the method according to claim 1 or 2.

6. A computer storage medium having stored thereon a computer program, characterized in that The computer program, when executed by a processor, implements each step of the method according to claim 1 or 2.

Images