Text correction method, electronic equipment and storage medium

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a text correction method, electronic equipment and a storage medium, and the method comprises the steps: predicting a plurality of correction labels of to-be-corrected elements in homework or test paper through a trained multi-label classification model; the method comprises the following steps of: correcting tags such as each empty statement smoothness, rhetorical use, associated word use, color word use, paradigm, wrongly written characters and the like of each question, calling different correction operators corresponding to different correction tags, scoring answers answered by answers in each to-be-corrected element in homework or test paper, and integrating different scores corresponding to different correction tags, so as to complete the correction of the questions or the test paper. And obtaining the final score of the answer answered by the answering person in each to-be-corrected element in the homework or the test paper. Therefore, various question types can be corrected, and the question type coverage rate and the generalization of the intelligent correction method can be improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a text correction method, electronic equipment and storage medium. Background Art

[0002] With the continuous advancement of education, a large number of homework and test papers need to be graded. The traditional manual grading method is not only prone to human errors, but also inefficient and cannot meet the needs of modern education.

[0003] In this regard, some solutions have proposed intelligent grading methods. Existing intelligent grading methods are mainly divided into two categories: rule-based intelligent grading methods and machine learning-based intelligent grading methods. Among them, the rule-based intelligent grading method is to grade homework and test papers according to pre-set rules. For example, the standard answers to multiple-choice questions are set in advance to grade whether the answers to the multiple-choice questions answered by the test takers in the homework or test papers are correct. The machine learning-based intelligent grading method is to achieve automatic grading through a neural network model. For example, the correction model is obtained by training a neural network model to automatically grade the composition.

[0004] However, existing intelligent grading methods cover limited question types and have poor generalization. Summary of the invention

[0005] In order to solve the problem that existing intelligent correction methods have limited coverage of question types and poor generalization, embodiments of the present application provide a text correction method, an electronic device, and a storage medium.

[0006] A first aspect of an embodiment of the present application provides a text correction method for an electronic device, comprising: obtaining a text to be corrected; determining a first set of correction labels corresponding to each element to be corrected in the text to be corrected; determining an element score of each element to be corrected based on the first set of correction labels; and obtaining a text score of the text to be corrected based on the element score of each element to be corrected.

[0007] Based on the above scheme, by predicting multiple correction labels corresponding to the elements to be corrected, the question type coverage can be improved, and the generalization of the correction method can be improved. In addition, by correcting the elements to be corrected based on multiple correction labels, the correction accuracy of the text to be corrected can be improved.

[0008] It can be understood that the text to be corrected can be a homework or a test paper, and the elements to be corrected can include each blank of each question in the homework or test paper, each sentence of each paragraph, each option of each multiple-choice question, etc. Correction tags can include sentence coherence, rhetoric use, conjunction use, color word use, paradigm, typos, etc., where correction tags can also be called correction angles.

[0009] In a possible implementation of the first aspect, determining a first batch of revision label sets corresponding to each element to be revised in the text to be revised includes: obtaining a first batch of revision label sets corresponding to each revision element predicted by the first model after the text to be revised is input into the first model.

[0010] In some optional instances, the first model may be a label prediction model, also known as a multi-label classification model. The label prediction model may be a correction model that can be used to predict multiple correction labels for each blank of each question in the correction text by training a neural network model based on a large number of sample correction texts that are labeled with multiple correction labels for each blank of each question.

[0011] For example, for the fill-in-the-blank question in the text to be corrected, the title is: Follow the example and write words that represent colors.

[0012] Example: Deep blue and dark black

[0013] Green red Bright Black

[0014] Example: Golden

[0015] living Lush green Dark Blue

[0016] In some optional examples, multiple correction labels for each blank in the question can be predicted. For example, three correction labels can be predicted for the first blank (i.e., the blank corresponding to emerald green), the second blank (i.e., the blank corresponding to vermilion), and the third blank (i.e., the blank corresponding to bright black): color words, words, and wrong characters, and four correction labels can be predicted for the fourth, fifth, and sixth blanks: color words, words, paradigms, and wrong characters.

[0017] In the embodiment of the present application, by using the label prediction model to directly predict multiple correction labels for each blank of each question in the text to be corrected for the existing questions, the existing correction labels can be combined for unknown questions to obtain multiple correction labels for each blank of the unknown question type, which can improve the question type coverage and the generalization of the correction method. In addition, by predicting multiple correction labels for each blank of each question in the text to be corrected through the label prediction model, each blank can be corrected from multiple correction angles, which can improve the correction accuracy and correction efficiency of the text to be corrected.

[0018] In a possible implementation of the first aspect above, determining a first set of revision labels corresponding to each element to be revised in the text to be revised includes: obtaining each revision element in the text to be revised output by the second model after the text to be revised is input into the second model; obtaining a first set of revision labels corresponding to each revision element predicted by the first model after the each revision element is input into the first model.

[0019] In some optional instances, the second model may be a layout recognition model, which may be a division model that can be used to divide each question in a test paper into a part by training a neural network model based on a large number of sample test papers in which each question is divided into a part.

[0020] In the embodiment of the present application, by using the layout recognition model to divide the text to be corrected, the accuracy of dividing the text to be corrected can be improved, and then the accuracy of correction can be improved.

[0021] In a possible implementation of the first aspect above, determining a first batch of revision label sets corresponding to each element to be revised in the text to be revised includes: obtaining first feature information of each element to be revised in the text to be revised, the first feature information including at least one of type, difficulty, and knowledge point; obtaining a first batch of revision label sets corresponding to each revision element predicted by the first model after inputting the first feature information of each element to be revised into the first model.

[0022] In some optional examples, the type may include selection, filling in the blanks, etc., and the difficulty may be a difficulty coefficient determined based on the historical correctness of the elements to be corrected. The lower the historical correctness of the elements to be corrected, the higher the difficulty coefficient, and the higher the difficulty. The higher the historical correctness of the elements to be corrected, the lower the difficulty coefficient, and the lower the difficulty. Knowledge points may include geometry, arithmetic, etc.

[0023] In the embodiment of the present application, by predicting multiple correction labels of the element to be corrected based on the feature information of the element to be corrected, the accuracy of the predicted multiple correction labels of the element to be corrected can be improved, thereby improving the correction accuracy.

[0024] In a possible implementation of the first aspect above, determining a first batch of correction label sets corresponding to each element to be corrected in the text to be corrected includes: obtaining type information of each element to be corrected in the text to be corrected; based on the type information of each element to be corrected, retrieving a first batch of correction label sets corresponding to each element to be corrected from a question library, the question library including multiple reference types and a reference correction label set corresponding to each reference type in the multiple reference types.

[0025] It can be understood that the question library may include multiple marked correction tags corresponding to various types of correction elements. After obtaining the type information of the correction element, the type information of the correction element can be directly used to search for multiple correction tags corresponding to the correction element.

[0026] In a possible implementation of the first aspect above, the elements to be corrected include but are not limited to each blank of each question, each sentence of each paragraph, and each option of each multiple-choice question.

[0027] In a possible implementation of the first aspect, determining the element score of each element to be revised based on the first set of revision labels includes: calculating a sub-score corresponding to each revision label in the first set of revision labels corresponding to each element to be revised; and determining the element score of each element to be revised based on the sub-score.

[0028] For example, based on the examples listed above, the answer "turquoise" given by the first air respondent to this question can be scored 1 point based on the correction label "color word", the answer "turquoise" given by the first air respondent to this question can be scored 1 point based on the correction label "typo", the answer "turquoise" given by the first air respondent to this question can be scored 1 point based on the correction label "number of words", the answer "turquoise" given by the first air respondent to this question can be scored 1 point based on the correction label "paper surface", and the scores corresponding to all labels are combined to obtain the final score of 1 point for the answer "turquoise" given by the first air respondent to this question.

[0029] In a possible implementation of the first aspect above, the method further includes: returning the first set of revision labels to the user; and using the revision labels selected by the user from the first set of revision labels, and / or the added revision labels, as the second set of revision labels corresponding to each element to be revised in the text to be revised.

[0030] In the embodiment of the present application, by dynamically adding or deleting multiple correction labels of some or all elements to be corrected, loose or strict adjustment of the correction intensity can be achieved, thereby adapting to different correction scenarios.

[0031] In a possible implementation of the first aspect above, the method also includes: based on the correction strictness parameter sent by the user, adding and deleting the first set of correction labels corresponding to each element to be corrected in the text to be corrected, and obtaining a second set of correction labels corresponding to each element to be corrected in the text to be corrected, wherein when the correction strictness parameter is greater than the correction strictness threshold, the number of correction labels in the second set of correction labels is less than the number of correction labels in the first set of correction labels, and when the correction strictness parameter is less than or equal to the correction strictness threshold, the number of correction labels in the second set of correction labels is greater than or equal to the number of correction labels in the first set of correction labels.

[0032] It can be understood that the grading strictness parameter can also be called a grading parameter. When the grading strictness parameter is set to 0.5, it can represent a standard grading mode, when the grading strictness parameter is greater than 0.5, it can represent a relatively strict grading mode, and when the grading strictness is less than 0.5, it can represent a relatively loose grading mode. In a relatively loose grading mode, the purpose of encouraging the test takers can be achieved, and in a relatively strict grading mode, the purpose of appropriately increasing the test takers' learning pressure can be achieved.

[0033] In a possible implementation of the first aspect above, based on the correction strictness parameter sent by the user, a first set of correction labels corresponding to each element to be corrected in the text to be corrected is added or deleted to obtain a second set of correction labels corresponding to each element to be corrected in the text to be corrected, including: based on the correction strictness parameter and the priority of each correction label in the first set of correction labels, the correction labels in the first set of correction labels are added or deleted to obtain a second set of correction labels.

[0034] For example, for the second blank in the fill-in-the-blank question listed above, the label prediction model can generate three correction labels: "color word", "word", and "typo". When the correction strictness parameter is 0.2, only one correction label, the color word, is predicted for the blank. Among them, the correction label "color word" can have a higher priority than the correction labels "word" and "typo".

[0035] In a possible implementation of the first aspect above, based on the correction strictness parameter sent by the user, a first set of correction labels corresponding to each element to be corrected in the text to be corrected is added or deleted to obtain a second set of correction labels corresponding to each element to be corrected in the text to be corrected, including: based on the correction strictness parameter, the correction labels in the first set of correction labels are randomly added or deleted to obtain a second set of correction labels.

[0036] For example, for the second blank in the fill-in-the-blank question listed above, the label prediction model can generate three correction labels: "color word", "word", and "typo". When the correction strictness parameter is 0.2, only one correction label, the color word, is predicted for the blank. The correction label "color word" can also be randomly selected from the three correction labels "color word", "word", and "typo".

[0037] In a possible implementation of the first aspect, determining the element score of each element to be revised based on the first revision label set includes: calculating a sub-score corresponding to each revision label in the second revision label set corresponding to each element to be revised, and determining the element score of each element to be revised based on the sub-score.

[0038] In a possible implementation of the first aspect above, the first set of modification labels includes the first set of modification labels, and the sub-score corresponding to each modification label in the first set of modification labels corresponding to each element to be modified is calculated, including: when the matching degree between the element to be modified and the first set of modification labels is greater than a matching degree threshold, allocating the sub-score corresponding to the first set of modification labels to the element to be modified.

[0039] In a possible implementation of the first aspect above, the method further includes: based on the element score of each element to be corrected, feeding back score details to the user.

[0040] In the embodiment of the present application, by providing feedback on the score details of each blank of each question in the homework or test paper, the score details can be clarified, thereby improving the confidence and interpretability of the grading results.

[0041] In a second aspect, the present application provides a training method for a label prediction model, including: obtaining a first sample text and a first annotated label set; using the first sample text as an input of a first neural network model to obtain a first predicted label set output by the first neural network model; and adjusting the parameters of the first neural network model based on the first annotated label set and the first predicted label set to obtain a first model.

[0042] In a third aspect, the present application provides a training method for a layout recognition model, comprising: obtaining a second sample text and a first set of annotated elements to be corrected; using the second sample text as input to a second neural network model to obtain a first set of predicted elements to be corrected output by the second neural network model; adjusting the parameters of the second neural network model based on the first set of annotated elements to be corrected and the first set of predicted elements to be corrected to obtain a second model.

[0043] In a fourth aspect, the present application provides an electronic device, comprising: a memory for storing instructions executed by one or more processors of the electronic device, and a processor, which is one of the one or more processors of the electronic device, for executing the text correction method mentioned in the present application.

[0044] In a fifth aspect, the present application provides a readable storage medium having instructions stored thereon, which, when executed on an electronic device, enables the electronic device to execute the text correction method mentioned in the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1a According to some examples of the present application, a schematic diagram of an application scenario is shown;

[0046] Figure 1b According to some examples of the present application, another application scenario schematic diagram is shown;

[0047] Figure 2According to some examples of the present application, a schematic diagram of the structure of an intelligent correction system is shown;

[0048] Figure 3 According to some examples of the present application, a structural schematic diagram of another intelligent correction system is shown;

[0049] Figure 4 According to some examples of the present application, a flowchart of an intelligent correction method is shown;

[0050] Figure 5 According to some examples of the present application, a flowchart of a method for implementing intelligent correction from the interaction between the terminal side and the cloud side is shown;

[0051] Figure 6 According to some examples of the present application, a flowchart of a text correction method is shown;

[0052] Figure 7 According to some examples of the present application, a flowchart of a method for implementing text correction from the interaction between the terminal side and the cloud side is shown;

[0053] Figure 8 According to some examples of the present application, a schematic diagram of a first display interface of an electronic device is shown;

[0054] Fig. 9 According to some examples of the present application, a schematic diagram of a second display interface of an electronic device is shown;

[0055] Fig.10 According to some examples of the present application, a schematic diagram of a third display interface of an electronic device is shown;

[0056] Fig.11 According to some examples of the present application, a flowchart of a method for implementing text correction from the interaction between the terminal side and the cloud side is shown;

[0057] Fig.12 According to some examples of the present application, a schematic diagram of a fourth display interface of an electronic device is shown;

[0058] Fig.13 According to some examples of the present application, a schematic diagram showing comparison of each blank correction label predicted for the same question under a relatively loose correction mode and a relatively strict correction mode is shown;

[0059] Fig.14 According to some examples of the present application, a schematic diagram showing a comparison of the credible grading results predicted for the same question under a relatively loose grading mode and a relatively strict grading mode is shown;

[0060] Fig.15According to some examples of the present application, a hardware structure of an electronic device is shown. DETAILED DESCRIPTION

[0061] The illustrative embodiments of the present application include, but are not limited to, a text correction method, an electronic device, and a storage medium.

[0062] It can be understood that the text correction method provided in the embodiment of the present application can be used in any implementable electronic device such as a smart phone, a tablet computer or a personal digital assistant.

[0063] like Figure 1a As shown, in some specific implementations, the first electronic device 110 can send the text to be corrected 111 to the second electronic device 120, so that the second electronic device 120 can correct the text to be corrected 111 based on the first neural network model 121 and the second neural network model 122.

[0064] like Figure 1b As shown, in some specific implementations, the second electronic device 120 can send the first neural network model 121 and the second neural network model 122 to the first electronic device 110, so that the first electronic device 110 can correct the text to be corrected based on the first neural network model 121 and the second neural network model 122.

[0065] In some specific implementations, the first electronic device 110 may be a cloud server, and the second electronic device 120 may be any implementable electronic device such as a smart phone, a tablet computer, a laptop computer, a desktop computer, or a personal digital assistant. The first neural network model 111 may be any implementable model that divides each question in the text to be corrected into a part, such as a deep convolutional neural network model. The second neural network model 122 may be any implementable model that predicts multiple correction labels for each blank of each question in the text to be corrected, such as a multimodal neural network model.

[0066] In some specific implementations, the text to be corrected may be input into the first neural network model 111, and the first neural network model 111 may output the text to be corrected by dividing each question in the text to be corrected into a part.

[0067] In some specific implementations, the text to be corrected can be input into the second neural network model 122, and based on the second neural network model 122, multiple correction labels for each blank of each question in the text to be corrected can be output, such as correction labels such as sentence coherence, use of rhetoric, use of conjunctions, use of color words, paradigm, typos, etc.

[0068] The intelligent marking system mentioned in some embodiments is introduced below.

[0069] Figure 2 A schematic diagram of the structure of an intelligent grading system is shown. The intelligent grading system can be used to grade structured questions and answers based on pre-set rules, such as grading multiple-choice questions.

[0070] The intelligent grading system may include a question parsing module 210 , a rule and template scoring module 220 , and a scoring and feedback module 230 .

[0071] The question parsing module 210 can be used to divide the text to be corrected (such as homework or test paper) into parts, and divide each question in the text to be corrected into a part.

[0072] The rule and template scoring module 220 can be used to grade the answers given by the examinee in each question of the assignment or test paper based on the pre-set rules and templates. For example, rules such as assigning a certain score when the examinee's answer is consistent with the standard answer and assigning a certain score when the examinee's answer is inconsistent with the standard answer can be set in advance, and these rules can be used as templates.

[0073] The scoring and feedback module 230 can be used to calculate the score corresponding to the answer to each question, obtain the final score, and generate learning suggestions based on the final score. For example, when the question is a tense multiple-choice question and the final score is lower than the score threshold, a learning suggestion of "learning tense grammar" can be generated.

[0074] Figure 3 A structural schematic diagram of another intelligent grading system is shown. The intelligent grading system can train a neural network model based on questions and answers of a specific question type, thereby obtaining a grading model that can automatically grade the specific question type.

[0075] Among them, the intelligent grading system may include a data preparation module 310 , a grading model training module 320 , and a grading and feedback module 330 .

[0076] The data preparation module 310 can be used to collect data such as model essays of a specified topic and their corresponding scores for a specific question type, such as composition, and perform data cleaning, deduplication, and standardization on the model essays and their corresponding scores to obtain a sample data set.

[0077] The correction model training module 320 can be used to select a neural network model that is suitable for the data characteristics in the sample data set, such as a convolutional neural network model, a recurrent neural network model, etc., and continuously adjust the parameters of the neural network model based on the sample data set, so that the neural network model can learn the mapping relationship between the model essay and the score, thereby realizing the training of the neural network model and obtaining the correction model.

[0078] The marking and feedback module 330 can be used to automatically mark the essays in the homework or test paper using the trained marking model, and generate scores and personalized feedback information. For example, when the essay topic in the homework or test paper requires a narrative essay, and the score is lower than the score threshold, feedback information of "read more narrative essays" can be generated.

[0079] In the above two intelligent grading systems, only structured questions and answers can be graded according to pre-set rules, such as grading multiple-choice questions, or a grading model is obtained by collecting sample essays on specified topics to train a neural network model to score essays. Both systems cover limited question types and have poor generalization.

[0080] In order to solve the above problems, the embodiment of the present application provides an intelligent grading method, which predicts multiple grading labels of each element to be graded in the homework or test paper through a trained multi-label classification model, such as sentence coherence, rhetoric use, conjunction use, color word use, paradigm, typos and other grading labels, and calls different grading operators corresponding to different grading labels (i.e., the grading method for each blank of each question in the homework or test paper), scores the answers given by the respondents in each element to be graded in the homework or test paper, and comprehensively calculates the different scores corresponding to different grading labels to obtain the final score of the answers given by the respondents in each element to be graded in the homework or test paper. In this way, multiple types of questions can be graded, which can improve the coverage of question types and the generalization of the intelligent grading method.

[0081] It can be understood that the elements to be corrected may include each blank of each question, each sentence of each paragraph, each option of each multiple-choice question, etc.

[0082] For example, the title of a fill-in-the-blank question in a homework or test paper is: Follow the example and write words that represent colors.

[0083] Example: Deep blue and dark black

[0084] Green Bright Black Black

[0085] The multi-label classification model can predict that the first blank correction label of this question is: color words, wrong characters, number of words, and paper surface. Based on the correction label "color words", the answer "green" given by the first air answerer of this question is scored 1 point, based on the correction label "wrong characters", the answer "green" given by the first air answerer of this question is scored 1 point, based on the correction label "number of words", the answer "green" given by the first air answerer of this question is scored 1 point, and based on the correction label "paper surface", the answer "green" given by the first air answerer of this question is scored 1 point, and the scores corresponding to all labels are combined to obtain the final score of 1 point for the answer "green" given by the first air answerer of this question.

[0086] The multi-label classification model can predict that the second blank correction label of this question is: color words, wrong characters, number of words, and paper surface. Based on the correction label "color words", the answer "Minghei" given by the first air answerer of this question is scored 0 points, based on the correction label "wrong characters", the answer "Minghei" given by the first air answerer of this question is scored 1 point, based on the correction label "number of words", the answer "Minghei" given by the first air answerer of this question is scored 1 point, based on the correction label "paper surface", the answer "Minghei" given by the first air answerer of this question is scored 0 points, and the scores corresponding to all labels are combined to obtain the final score of 0.5 points for the answer "Minghei" given by the first air answerer of this question.

[0087] The multi-label classification model can predict that the third blank correction label of this question is: color words, wrong characters, number of words, and paper surface. Based on the correction label "color words", the answer "name black" given by the first air answerer of this question is scored 0 points, based on the correction label "wrong characters", the answer "name black" given by the first air answerer of this question is scored 0 points, based on the correction label "number of words", the answer "name black" given by the first air answerer of this question is scored 1 point, based on the correction label "paper surface", the answer "name black" given by the first air answerer of this question is scored 0 points, and the scores corresponding to all labels are combined to obtain the final score of 0 points for the answer "name black" given by the first air answerer of this question.

[0088] It can be understood that the score details of each blank of each question in the homework or test paper can be fed back based on the different scores corresponding to different grading labels. In this way, the score details can be clarified and the confidence and interpretability of the grading results can be improved.

[0089] It can be understood that after the trained multi-label classification model predicts multiple correction labels for each blank of each question in the homework or test paper, the multiple correction labels for each blank of each question in the homework or test paper can be updated. In actual applications, due to different learning progress and different correction scenarios, the correction requirements for each blank of each question in the homework or test paper will be different. For example, in the homework practice scenario, delete the multiple correction labels for some or all blanks of some or all questions in the homework, and in the test paper evaluation scenario, add multiple correction labels for some or all blanks of some or all questions in the test paper.

[0090] In some optional instances, after predicting multiple correction labels for each blank of each question in the homework or test paper through the trained multi-label classification model, multiple correction labels for each blank of each question in the homework or test paper can be fed back for the user to freely choose partial or all blank correction labels for some or all questions in the homework or test paper, and the update of partial or all blank correction labels for some or all questions in the homework or test paper can be realized.

[0091] In some other optional instances, the user can set the grading parameters for each blank of each question in the homework or test paper, such as the grading strictness parameter, so that based on the grading parameters for each blank of each question in the homework or test paper, the multiple correction labels for each blank of each question in the homework or test paper predicted by the trained multi-label classification model can be updated. For example, when the grading strictness parameter is set to be less than the grading strictness parameter threshold, based on the grading strictness parameters for each blank of each question in the homework or test paper, the multiple correction labels for some or all of the blanks predicted by the trained multi-label classification model for some or all of the questions in the homework or test paper are deleted.

[0092] The following is an introduction to the intelligent correction method mentioned in the embodiments of the present application. Figure 4 A flowchart of an intelligent correction method is shown.

[0093] like Figure 4 As shown, first, the text to be corrected can be parsed for test question entry. For example, pictures uploaded by an image acquisition device and / or texts stored in a cloud server (such as assessment exercises) can be entered into an intelligent grading system that executes the intelligent grading method, and the text to be corrected can be parsed into the format required by the intelligent grading system so that the text to be corrected can be further processed later.

[0094] In some optional instances, while the intelligent grading system is parsing the text to be graded, it can extract characteristic information such as the type (such as multiple choice, fill-in-the-blank), difficulty (for example, a difficulty coefficient determined based on the historical accuracy rate, the lower the historical accuracy rate, the higher the difficulty coefficient), and knowledge points (such as geometry, arithmetic) of each question in the text to be graded.

[0095] In some optional examples, an optical character recognition (OCR) method may be used to translate the image into computer text to implement parsing of the text to be corrected.

[0096] Then, the text to be corrected can be subjected to multi-label classification. In some optional instances, the feature information of the text to be corrected can be input into a multi-label classification model, and the multi-label classification model can be used to predict (multiple) correction labels for the text to be corrected based on the feature information of the text to be corrected, so as to obtain a multi-label annotation result, that is, to predict multiple correction labels for each blank of each question in the text to be corrected, also known as multiple correction angles. In other optional instances, the text to be corrected can also be directly input into a multi-label classification model, and the multi-label classification model can be used to predict (multiple) correction labels for the text to be corrected, so as to obtain a multi-label annotation result.

[0097] In some optional instances, the multi-label classification model can be a correction model that can be used to predict multiple correction labels for each blank of each question in the correction text, obtained by training a neural network model based on a large number of sample correction texts with multiple correction labels for each blank of each question labeled.

[0098] Then, the text to be corrected can be directly corrected and feedback can be given intelligently to obtain reliable correction results, such as obtaining the score details for each blank in each question in the text to be corrected.

[0099] However, in actual applications, due to different learning progress and different grading scenarios, the requirements for grading each blank of each question in the text to be graded will be different. For example, in the homework practice scenario, the grading of each blank of each question in the grading text is relatively loose, and in the test paper evaluation scenario, the grading of each blank of each question in the grading text is relatively strict. For another example, for multiple blanks of each question in the grading text, the grading of the first blank is relatively strict, and the grading of the latter blank is relatively loose. At this time, the multi-label annotation results can be dynamically adjusted, that is, the multiple correction labels of some or all blanks of the questions can be dynamically adjusted (such as adding or deleting) to obtain the final correction label of each blank of each question in the text to be graded, which is also called the multi-label annotation result.

[0100] Finally, for each blank final correction label of each question in the text to be corrected, the correction operator corresponding to each correction label can be found from the correction operator set to perform intelligent correction on the text to be corrected. Since each blank final correction label of each question in the text to be corrected contains multiple correction labels, each blank of each question in the text to be corrected will be corrected by multiple correction operators to obtain scores from multiple correction angles. By summarizing these scores, we can obtain the score details of each blank of each question in the text to be corrected, so as to obtain a reliable correction result.

[0101] In some optional examples, the correction operator set can be a set of correction methods formed by the historical correction methods based on the historical correction texts, and the correction methods set for each correction tag can also be called a correction operator library. In the specific implementation process, the correction operator library can be continuously updated as the types of questions continue to increase and teachers and engineers continue to summarize and summarize.

[0102] Taking test marking as an example, the intelligent marking method mentioned above is introduced in detail. Figure 5 A flowchart of a method for implementing intelligent correction through interaction between the terminal side and the cloud side is shown.

[0103] The test taker (such as a student) or examiner (such as a teacher) on the client side can take a photo of the test paper using an image acquisition device and upload it to a cloud server on the cloud side.

[0104] The cloud server on the cloud side can divide each question in the test paper into a part through the layout recognition model, and can translate each question into computer text through OCR analysis. The layout recognition model can be a division model that can be used to divide each question in the test paper into a part by training a neural network model based on a large number of sample test papers that divide each question into a part.

[0105] Then, for the test paper that has been divided into each question through the layout recognition model, the cloud server can search for questions based on the question library, that is, it can search whether there are multiple marked correction labels corresponding to each blank in each question in the question library.

[0106] In some optional examples, if there are multiple marked correction labels corresponding to each blank in each question in the question database, each blank in each question can be corrected directly based on the correction operator corresponding to each correction label in the multiple marked correction labels to obtain the score details of each blank in each question. Then, the cloud server can feedback the correction angle and score suggestions to the examiner (such as a teacher) on the terminal side based on the multiple correction labels.

[0107] In some optional instances, if some or all of the blanks in some or all of the questions in the question library do not have corresponding multiple marked correction labels, the label prediction model, i.e., the multi-label classification model mentioned above, can be used to predict the correction labels for each blank in each question, that is, predict multiple correction labels for each blank in each question, also called the correction angle of "blank". Then, based on the correction operator corresponding to each correction label in the multiple correction labels predicted for each blank in each question, each question can be corrected blank by blank, that is, each blank in each question can be corrected. Then, the cloud server can feedback the correction angle and score suggestions to the examiner (such as a teacher) on the end side based on the multiple correction labels.

[0108] In some optional examples, for each question divided by the layout recognition model, after determining multiple correction labels for each blank in each question, the multiple correction labels for each blank in each question can be returned to the examiner (such as a teacher) on the terminal side.

[0109] The examiner on the end side (such as a teacher) can use the multiple correction labels for each blank in each question as a reference for manual correction. The examiner on the end side (such as a teacher) can also adjust the correction labels for some or all of the blank correction labels for some or all of the questions. For example, according to the specific scenario, the multiple correction labels for some or all of the blank correction labels for some or all of the questions can be dynamically added or deleted, and the adjusted multiple correction labels for each blank in each question can be returned to the cloud server on the cloud side.

[0110] The cloud server on the cloud side can call the correction operator corresponding to each correction label from the correction operator set based on the multiple correction labels of each blank in each question after adjustment, and correct each blank in each question, that is, correct each question blank by blank. For each blank in each question, the correction results of multiple correction operators can be obtained, and then the correction results of multiple correction operators can be summarized to obtain the total score of each blank in each question and the total score of each question. At the same time, the cloud server can provide feedback on deduction points, that is, list the score details of each blank in each question, and return it to the examiner on the end side (such as a teacher) for further adjustment.

[0111] The examiner on the client side (such as a teacher) can adjust the score details of some or all of the blanks for some or all of the questions to obtain the correction results.

[0112] The intelligent correction method mentioned in the embodiment of the present application is introduced in detail below. Figure 6 A flow chart of a text correction method is shown, and the text correction method can be executed by an electronic device. Figure 6 As shown, the text correction method may include:

[0113] 601: Obtain the text to be corrected, perform label prediction on the text to be corrected based on the label prediction model, and obtain multiple correction labels for each blank of each question in the text to be corrected.

[0114] It can be understood that the text to be corrected can be homework, test paper, etc. After obtaining the text to be corrected, the text to be corrected can be entered into a question list, and each question in the text to be corrected can be divided into a part for subsequent processing.

[0115] In some optional instances, each question in the paper answer sheet can be photographed and uploaded through an image acquisition device, or each question in the electronic answer sheet can be pasted, or the text to be corrected can be divided based on a layout recognition model, and each question in the text to be corrected can be divided into a part.

[0116] In some optional examples, the text to be corrected can be parsed for test entry, for example, each question in the text to be corrected can be parsed into the format required by the intelligent correction system. While parsing the text to be corrected, characteristic information such as the type (such as multiple choice, fill-in-the-blank), difficulty (such as a difficulty coefficient determined based on the historical accuracy rate, the lower the historical accuracy rate, the higher the difficulty coefficient), and knowledge points (such as geometry, arithmetic) of each question in the text to be corrected can be extracted.

[0117] It can be understood that the correction label prediction can be performed for each blank in each question based on the label prediction model, that is, multiple correction labels for each blank in each question are predicted, also known as the correction angle of "blank". Among them, the label prediction model can be a correction model that can be used to predict multiple correction labels for each blank in each question in the correction text by training a neural network model based on a large number of sample correction texts with multiple correction labels for each blank in each question.

[0118] In some optional instances, after predicting each blank multiple correction label for each question in the text to be corrected based on the label prediction model, the correction labels of some or all of the blank multiple correction labels of some or all of the questions in the text to be corrected can be adjusted, that is, according to the specific scenario, some or all of the blank multiple correction labels of some or all of the questions can be added or deleted.

[0119] In some other optional examples, before predicting multiple correction labels for each blank of each question in the text to be corrected based on the label prediction model, a correction mode can be set for each question in the text to be corrected (such as setting correction parameters for each question in the text to be corrected), so that for different correction modes, the multiple correction labels for each blank of each question predicted by the label prediction model are different. Among them, the correction parameter can also be called a correction strictness parameter. When the correction strictness parameter is set to 0.5, it can represent a standard correction mode. When the correction strictness parameter is greater than 0.5, it can represent a relatively strict correction mode. When the correction strictness is less than 0.5, it can represent a relatively loose correction mode.

[0120] In the embodiment of the present application, the label prediction model can directly predict multiple correction labels for each blank of each question in the text to be corrected for existing questions, and can combine the existing correction labels for unknown questions to obtain multiple correction labels for each blank of unknown question types, thereby improving the coverage of question types and the generalization of the correction method. In addition, by predicting multiple correction labels for each blank of each question in the text to be corrected through the label prediction model, each blank can be corrected from multiple correction angles, which can improve the correction accuracy of the text to be corrected.

[0121] 602: For each question in the text to be corrected, based on the correction operators corresponding to the multiple correction labels for each blank in each question, correct each question blank by blank to obtain multiple correction results for each blank in each question.

[0122] In some optional examples, based on multiple correction labels for each blank in each question in the text to be corrected, correction operators corresponding to each correction label in multiple correction labels can be called from the correction operator set to perform correction on each blank, that is, to correct each blank in each question.

[0123] In some optional examples, for each blank in each question, the grading results of multiple grading operators can be obtained, and then the grading results of the multiple grading operators can be aggregated to obtain the total score of each blank in each question and the total score of each question.

[0124] 603: Summarize and process multiple correction results for each blank in each question in the correction text to obtain a reliable correction result.

[0125] It is understandable that after obtaining multiple grading results for each blank in each question in the text to be graded, the score details for each blank in each question can be listed. The credible grading results can include the reasons for the deduction of the answers given by each blank answerer in each question, and can also include the deduction details of the answers given by each blank answerer in each question.

[0126] In the embodiment of the present application, by summarizing multiple correction results for each blank to obtain a credible correction result, the respondent can see the points where the points were lost, which can not only enhance the credibility of the correction result, but also enable the respondent to clearly identify the direction for improvement.

[0127] The text correction method mentioned above is described in detail below with a specific embodiment. Figure 7 A flowchart of a text correction method implemented from the interaction between the terminal side and the cloud side is shown. The text correction method includes three steps: correction tag prediction 710, automatic correction 720 and trusted correction feedback 730.

[0128] In the step of grading label prediction 710, the examiner (such as a teacher) or the examiner (such as a student) on the end side can enter the question list for the text to be graded (such as a paper answer paper), for example, by taking a photo or pasting the text to be graded to the cloud server on the cloud side. The cloud server on the cloud side can use the layout recognition model to divide each question in the text to be graded, and use OCR to parse each question, such as converting the paper answer paper into an electronic format. Then, the cloud server on the cloud side can use the label prediction model, that is, the multi-label classification model mentioned above, to predict the grading label for each blank in each question, obtain multiple grading labels for each blank in each question, and feedback to the examiner (such as a teacher) on the end side.

[0129] For example, for the fill-in-the-blank question in the text to be corrected, the title is: Follow the example and write words that represent colors.

[0130] Example: Deep blue and dark black

[0131] Green red Bright Black

[0132] Example: Golden

[0133] living Lush green Dark Blue

[0134] In some optional instances, the cloud server on the cloud side can use the label prediction model to predict multiple correction labels for each blank in the question. For example, three correction labels can be predicted for the first blank (i.e. the blank corresponding to emerald green), the second blank (i.e. the blank corresponding to vermilion), and the third blank (i.e. the blank corresponding to bright black): color words, words, and typos. In addition, four correction labels can be predicted for the fourth, fifth, and sixth blanks: color words, words, paradigms, and typos. Furthermore, feedback can be provided to the electronic device held by the teacher on the terminal side. The electronic device held by the teacher on the terminal side can display the following information: Figure 8 The first display interface 800 shown in the figure can display multiple correction label components (such as 810-1 to 810-3) for each blank in the question and a first correction component 820 on the first display interface 800. The correction label component can display the three correction labels for the first, second and third blanks mentioned above: color words, words, and wrong characters, and the four correction labels for the fourth, fifth and sixth blanks: color words, words, paradigms, and wrong characters, etc.

[0135] In the automated grading step 720, the examiner on the terminal side (such as a teacher) can adjust the multiple correction labels that are partially or completely blank in some or all of the questions, that is, according to the specific scenario, add or delete the multiple correction labels that are partially or completely blank in some or all of the questions. In some optional instances, the examiner on the terminal side (such as a teacher) can add or delete the multiple correction labels that are partially or completely blank in some or all of the questions by clicking on the multiple correction labels that are partially or completely blank in some or all of the questions. For example, the examiner on the terminal side (such as a teacher) can delete the correction label "typo" for the first blank, the second blank, and the third blank in the fill-in-the-blank question. At this time, the electronic device held by the teacher on the terminal side can be switched to the following mode: Fig. 9 The second display interface 900 shown can display multiple correction label components (such as 910 - 1 and 910 - 2 ) for each blank in the question and a second correction component 920 on the second display interface 900 .

[0136] In the embodiment of the present application, by the examiner (such as a teacher) dynamically adding or deleting multiple empty correction labels for some or all questions, loose and strict adjustment of the correction intensity can be achieved, thereby adapting to different correction scenarios.

[0137] In some optional examples, the cloud server on the cloud side can directly call the correction operator corresponding to each correction label in the multiple correction labels from the correction operator set based on the multiple correction labels for each blank in each question, and perform correction on a blank-by-blank basis, that is, correct each blank in each question. The cloud server on the cloud side can also call the correction operator corresponding to each correction label in the multiple correction labels from the correction operator set based on the adjusted multiple correction labels for each blank in each question, and perform correction on a blank-by-blank basis, that is, correct each blank in each question.

[0138] For example, when a marker (such as a teacher) on the client side is detected, Figure 8 The first batch of modified components 820 on the first display interface 800 shown in FIG. Fig. 9 When the second correction component 920 on the second display interface 900 is shown, the correction operator corresponding to each correction tag in the plurality of correction tags may be called to perform blank-by-blank correction.

[0139] In the trusted correction feedback step 730, the cloud server on the cloud side can summarize the correction results of each blank correction in the automatic correction step 720, obtain a trusted correction result and return it to the examiner (such as a teacher) on the client side. The trusted correction result may include the reason for the deduction of the answer given by each air test respondent in each question, and may also include the deduction details of the answer given by each air test respondent in each question.

[0140] For example, for the fill-in-the-blank questions listed above, when the cloud server on the cloud side returns the reliable correction results to the examiner (such as a teacher) on the terminal side, the electronic device held by the teacher on the terminal side can display the following information: Fig.10 The third display interface 1000 shown in the figure can display the reliable correction results of each blank in the question on the third display interface 1000. Fig.10 As shown, the reasons for deductions are given for the answers to the third, fourth and sixth blanks in the fill-in-the-blank question. Specifically, for the answer "bright and black" given by the answerer in the third blank of the fill-in-the-blank question, the reasons for deductions are given as "there is no such color" and "there is no such word", and the deduction details are given as "1 point deducted (-1 point) because of "there is no such color", "1 point deducted (-1 point) because of "there is no such word", and "1 point worth, 0 points". The score details are given for the answers to the first, second and fifth blanks in the fill-in-the-blank question.

[0141] The text correction method submitted above is described in detail below with another specific embodiment. Fig.11A flowchart of a text correction method implemented through interaction between the terminal side and the cloud side is shown. The text correction method includes three steps: question entry analysis 1110 , correction mode setting 1120 , and trusted correction feedback 1130 .

[0142] In the question entry and analysis step 1110, the examiner (such as a teacher) or the examinee (such as a student) on the end side can enter the question into the text to be marked (such as a paper answer paper), for example, by taking a photo or pasting the text to be marked to the cloud server on the cloud side. The cloud server on the cloud side can use the layout recognition model to divide each question in the text to be marked, and use OCR to analyze each question, such as converting the paper answer paper into an electronic format, and return the question analysis result to the examiner (such as a teacher) or the examinee (such as a student) on the end side.

[0143] For example, for the fill-in-the-blank question in the text to be corrected, the title is: Follow the example and write words that represent colors.

[0144] Example: Deep blue and dark black

[0145] Green Vermillion Bright Black

[0146] Example: Golden

[0147] living Lush green Dark Blue

[0148] In the grading mode setting step 1120, the examiner on the terminal side (such as a teacher) can set the grading strictness for each question and upload it to the cloud server on the cloud side for grading parameter setting. Among them, the grading parameter can also be called the grading strictness parameter. When the grading strictness parameter is set to 0.5, it can represent a standard grading mode. When the grading strictness parameter is greater than 0.5, it can represent a relatively strict grading mode. When the grading strictness is less than 0.5, it can represent a relatively loose grading mode. In a relatively loose grading mode, the purpose of encouraging the answerer can be achieved. In a relatively strict grading mode, the purpose of appropriately increasing the learning pressure of the answerer can be achieved.

[0149] In some optional implementations, the electronic device held by the teacher on the terminal side may display the following information: Fig.12 The fourth display interface 1200 shown may display a grading strictness component 1210 of the fill-in-the-blank questions listed above.

[0150] In some optional instances, the cloud server on the cloud side can use the label prediction model to predict multiple grading labels for each blank in each question based on the grading strictness parameters uploaded by the examiner on the end side (such as a teacher), and feed back the grading labels in a specific mode to the examiner on the end side (such as a teacher). Among them, for different grading modes, the multiple grading labels for each blank in each question predicted by the label prediction model are different. For example, for the fill-in-the-blank questions listed above, Fig.13 As shown in a, under the relatively loose grading mode, it can be predicted that the second blank in this question has only one grading label, "color word", and the fourth blank has two grading labels, "paradigm" and "color word". Fig.13 As shown in b, under a relatively strict grading mode, it can be predicted that the second blank in this question has four grading labels: "color words", "number of words", "typos", and "paper surface", and the fourth blank has four labels: "paradigm", "color words", "words", and "typos".

[0151] In some optional examples, the label prediction model can predict multiple correction labels for each blank in each question based on the correction strictness parameter and the priority of each correction label in the generated multiple correction labels. It can also randomly predict multiple correction labels for each blank in each question based on the correction strictness parameter.

[0152] For example, for the second blank in the fill-in-the-blank question listed above, the label prediction model can generate three correction labels: "color word", "word", and "typo". When the correction strictness parameter is 0.2, only one correction label, the color word, is predicted for the blank. Among them, the priority of the correction label "color word" can be higher than the correction labels "word" and "typo", and the correction label "color word" can also be randomly selected from the three correction labels "color word", "word", and "typo".

[0153] In the trusted correction feedback step 1130, the cloud server on the cloud side can directly call the correction operator corresponding to each correction label in the multiple correction labels from the correction operator set based on the multiple correction labels for each blank of each question, and perform correction on a blank-by-blank basis, that is, correct each blank in each question. The examiner on the end side (such as a teacher) can adjust the correction labels for the multiple correction labels that are partially or completely blank in some or all questions, that is, add or delete the multiple correction labels that are partially or completely blank in some or all questions according to the specific scenario. In some optional instances, the examiner on the end side (such as a teacher) can add or delete the multiple correction labels that are partially or completely blank in some or all questions by clicking on the multiple correction labels that are partially or completely blank in some or all questions.

[0154] In some optional instances, the cloud server on the cloud side may also call the correction operator corresponding to each correction label in the correction operator set based on the adjusted multiple correction labels for each blank in each question, and perform correction on each blank, that is, correct each blank in each question.

[0155] In some optional examples, the cloud server on the cloud side can summarize the grading results of each blank grading, obtain a reliable grading result and return it to the examiner on the client side (such as a teacher). The reliable grading result can include the reason for the deduction of the answer given by each blank test taker in each question, and can also include the deduction details of the answer given by each blank test taker in each question.

[0156] For example, for the fill-in-the-blank questions listed above, when the cloud server on the cloud side returns the reliable correction results to the examiner (such as a teacher) on the terminal side, the electronic device held by the teacher on the terminal side can display the following information: Fig.14 The reliable correction results for each blank in the question are shown.

[0157] like Fig.14 As shown in a, under the relatively loose grading mode, reasons for deductions were given for the answers to the third, fourth and sixth blanks in the fill-in-the-blank question. Specifically, for the answer "bright black" in the third blank of the fill-in-the-blank question, the reason for deduction was "there is no such color", and the deduction details were given as "1 point deducted (-1 point)" because "there is no such color" and "1 point worth, 0 points scored".

[0158] like Fig.14 As shown in Figure b, under the relatively strict grading mode, the reasons for deductions were given for the answers to the third, fourth and sixth blanks in the fill-in-the-blank question. Specifically, for the answer "vermilion" given by the answerer in the second blank of the fill-in-the-blank question, the reason for deduction was "the number of words does not meet the requirements", and the deduction details were given as "1 point (-1 point)" because "the number of words does not meet the requirements" and "1 point, 0 points". For the answer "bright black" given by the answerer in the third blank of the fill-in-the-blank question, the reasons for deductions were "there is no such color" and "there is no such word", and the deduction details were given as "1 point (-1 point)" because "there is no such color", "1 point (-1 point)" because "there is no such word", and "1 point, 0 points".

[0159] The following is an introduction to the hardware structure of electronic equipment. Fig.15As shown, the electronic device 1500 may include a processor 1510, an external memory interface 1520, an internal memory 1521, a universal serial bus (USB) interface 1530, a charging management module 1540, a power management module 1541, a battery 1542, an antenna 1, an antenna 2, a mobile communication module 1550, a wireless communication module 1560, an audio module 1570, a speaker 1570A, a receiver 1570B, a microphone 1570C, an earphone interface 1570D, a sensor module 1580, a button 1590, a motor 1591, an indicator 1592, a camera 1593, a display screen 1594, and a subscriber identification module (SIM) card interface 1595, etc. The sensor module 1580 may include a pressure sensor 1580A, a gyroscope sensor 1580B, an air pressure sensor 1580C, a magnetic sensor 1580D, an acceleration sensor 1580E, a distance sensor 1580F, a proximity light sensor 1580G, a fingerprint sensor 1580H, a temperature sensor 1580J, a touch sensor 1580K, an ambient light sensor 1580L, a bone conduction sensor 1580M, and the like.

[0160] It is to be understood that the structure illustrated in the embodiment of the present invention does not constitute a specific limitation on the electronic device. In other embodiments of the present application, the electronic device 1500 may include more or fewer components than shown in the figure, or combine some components, or split some components, or arrange the components differently. The components shown in the figure may be implemented in hardware, software, or a combination of software and hardware.

[0161] Processor 1510 may include one or more processing units, for example: processor 1510 may include an application processor (application processor, AP), a modem processor, a graphics processor (graphics processing unit, GPU), an image signal processor (image signal processor, ISP), a controller, a video codec, a digital signal processor (digital signal processor, DSP), a baseband processor, and / or a neural-network processing unit (neural-network processing unit, NPU), etc. Among them, different processing units can be independent devices or integrated in one or more processors. For example, processor 1510 can execute the text correction method mentioned in the embodiment of the present application.

[0162] The controller can generate operation control signals according to the instruction operation code and timing signal to complete the control of instruction fetching and execution.

[0163] A memory may also be provided in the processor 1510 for storing instructions and data. In some embodiments, the memory in the processor 1510 is a cache memory. The memory may store instructions or data that the processor 1510 has just used or circulated. If the processor 1510 needs to use the instruction or data again, it may be directly called from the memory. Repeated access is avoided, the waiting time of the processor 1510 is reduced, and the efficiency of the system is improved. For example, the memory may store the text correction method mentioned in the embodiment of the present application.

[0164] The electronic device 1500 implements the display function through a GPU, a display screen 1594, and an application processor. The GPU is a microprocessor for image processing, which connects the display screen 1594 and the application processor. The GPU is used to perform mathematical and geometric calculations for graphics rendering. The processor 1510 may include one or more GPUs that execute program instructions to generate or change display information.

[0165] The display screen 1594 is used to display images, videos, etc. The display screen 1594 includes a display panel. The display panel can display the first display interface, the second display interface, the third display interface, the fourth display interface, or the fifth display interface described above.

[0166] The pressure sensor 1580A is used to sense the pressure signal and can convert the pressure signal into an electrical signal. In some embodiments, the pressure sensor 1580A can be set on the display screen 1594. There are many types of pressure sensors 1580A, such as resistive pressure sensors, inductive pressure sensors, capacitive pressure sensors, etc. A capacitive pressure sensor can be a parallel plate including at least two conductive materials. When a force acts on the pressure sensor 1580A, the capacitance between the electrodes changes. The electronic device 1500 determines the intensity of the pressure based on the change in capacitance. When a touch operation acts on the display screen 1594, the electronic device 1500 detects the touch operation intensity according to the pressure sensor 1580A. The electronic device 1500 can also calculate the touch position according to the detection signal of the pressure sensor 1580A. In some embodiments, touch operations acting on the same touch position but with different touch operation intensities can correspond to different operation instructions. For example: when a touch operation with a touch operation intensity less than the first pressure threshold acts on the short message application icon, an instruction to view the short message is executed. When a touch operation with a touch operation intensity greater than or equal to a first pressure threshold acts on the short message application icon, an instruction to create a new short message is executed.

[0167] The touch sensor 1580K is also called a "touch control device". The touch sensor 1580K can be set on the display screen 1594, and the touch sensor 1580K and the display screen 1594 form a touch screen, also called a "touch control screen". The touch sensor 1580K is used to detect touch operations acting on or near it. The touch sensor can pass the detected touch operation to the application processor to determine the type of touch event. Visual output related to the touch operation can be provided through the display screen 1594. In other embodiments, the touch sensor 1580K can also be set on the surface of the electronic device 1500, which is different from the position of the display screen 1594.

[0168] The key 1590 includes a power key, a volume key, etc. The key 1590 may be a mechanical key or a touch key. The electronic device 1500 may receive key input and generate key signal input related to user settings and function control of the electronic device 1500.

[0169] The above describes the hardware structure that the electronic device may have. It is understood that the structure illustrated in the embodiments of the present application does not constitute a specific limitation on the electronic device. In other embodiments of the present application, the electronic device may include more or fewer components than shown in the figure, or combine certain components, or split certain components, or arrange the components differently. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0170] In the accompanying drawings, some structural or method features may be shown in a specific arrangement and / or order. However, it should be understood that such a specific arrangement and / or order may not be required. Instead, in some embodiments, these features may be arranged in a manner and / or order different from that shown in the illustrative drawings. In addition, the inclusion of structural or method features in a particular figure does not mean that such features are required in all embodiments, and in some embodiments, these features may not be included or may be combined with other features.

[0171] The various embodiments of the mechanism disclosed in the present application can be implemented in hardware, software, firmware or a combination of these implementation methods. The embodiments of the present application can be implemented as a computer program or program code executed on a programmable system, which includes at least one processor, a storage system (including volatile and non-volatile memory and / or storage elements), at least one input device and at least one output device.

[0172] Program code can be applied to input instructions to perform the functions described in this application and generate output information. The output information can be applied to one or more output devices in a known manner. For the purposes of this application, a processing system includes any system having a processor such as, for example, a digital signal processor (DSP), a microcontroller, an application specific integrated circuit, or a microprocessor.

[0173] Program code can be implemented with high-level programming language or object-oriented programming language to communicate with the processing system. When necessary, program code can also be implemented with assembly language or machine language. In fact, the mechanism described in this application is not limited to the scope of any specific programming language. In either case, the language can be a compiled language or an interpreted language.

[0174] In some cases, the disclosed embodiments may be implemented in hardware, firmware, software, or any combination thereof. The disclosed embodiments may also be implemented as instructions carried or stored on one or more temporary or non-temporary machine-readable (e.g., computer-readable) storage media, which may be read and executed by one or more processors. For example, instructions may be distributed over a network or through other computer-readable media. Therefore, machine-readable media may include any mechanism for storing or transmitting information in a machine (e.g., computer) readable form, including, but not limited to, floppy disks, optical disks, optical disks, read-only memories (CD-ROMs), magneto-optical disks, read-only memories (ROMs), random access memories (RAMs), erasable programmable read-only memories (EPROMs), electrically erasable programmable read-only memories (EEPROMs), magnetic or optical cards, flash memory, or a tangible machine-readable memory for transmitting information (e.g., carrier waves, infrared signals, digital signals, etc.) using the Internet in electrical, optical, acoustic, or other forms of propagation signals. Therefore, machine-readable media include any type of machine-readable media suitable for storing or transmitting electronic instructions or information in a machine (e.g., computer) readable form.

[0175] It should be noted that the units / modules mentioned in the various device embodiments of the present application are all logical units / modules. Physically, a logical unit / module can be a physical unit / module, or a part of a physical unit / module, or can be implemented as a combination of multiple physical units / modules. The physical implementation method of these logical units / modules themselves is not the most important. The combination of functions implemented by these logical units / modules is the key to solving the technical problems proposed by the present application. In addition, in order to highlight the innovative part of the present application, the above-mentioned device embodiments of the present application do not introduce units / modules that are not closely related to solving the technical problems proposed by the present application, which does not mean that there are no other units / modules in the above-mentioned device embodiments.

[0176] It should be noted that in the examples and description of this patent, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including one" do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0177] Although the present application has been illustrated and described with reference to certain preferred embodiments thereof, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the scope of the present application.

Claims

1. A text correction method, used in electronic equipment, characterized in that: include: Get the text to be revised; Determine a first set of revision labels corresponding to each element to be revised in the text to be revised; Determining an element score of each of the elements to be revised based on the first set of revision labels; Based on the element scores of the elements to be corrected, a text score of the text to be corrected is obtained.

2. The method according to claim 1, characterized in that The step of determining a first set of revision labels corresponding to each element to be revised in the text to be revised includes: After the text to be revised is input into the first model, a first set of revision labels corresponding to the revision elements predicted by the first model is obtained.

3. The method according to claim 1, characterized in that The step of determining a first set of revision labels corresponding to each element to be revised in the text to be revised includes: After the text to be revised is input into the second model, each revision element in the text to be revised output by the second model is obtained; After the batch elements are input into the first model, a first batch of revision label sets corresponding to the batch elements predicted by the first model are obtained.

4. The method according to claim 1, characterized in that The step of determining a first set of revision labels corresponding to each element to be revised in the text to be revised includes: Acquire first feature information of each element to be corrected in the text to be corrected, wherein the first feature information includes at least one of type, difficulty, and knowledge point; After inputting the first feature information of the elements to be revised into the first model, a first batch of revision label sets corresponding to the elements to be revised predicted by the first model are obtained.

5. The method according to claim 1, characterized in that The step of determining a first set of revision labels corresponding to each element to be revised in the text to be revised includes: Obtaining type information of each element to be revised in the text to be revised; Based on the type information of each element to be corrected, a first set of correction labels corresponding to each element to be corrected is retrieved from a topic library, wherein the topic library includes multiple reference types and a reference correction label set corresponding to each reference type in the multiple reference types.

6. The method according to any one of claims 1 to 5, characterized in that: The elements to be corrected include but are not limited to each blank in each question, each sentence in each paragraph, and each option in each multiple-choice question.

7. The method according to any one of claims 1 to 5, characterized in that: The determining the element score of each of the elements to be revised based on the first set of revision labels includes: Calculate the sub-score corresponding to each correction label in the first set of correction labels corresponding to each element to be corrected; Based on the sub-scores, an element score of each of the elements to be corrected is determined.

8. The method according to claim 1, characterized in that The method further comprises: Returning the first set of changed labels to the user; The revision tags selected by the user from the first revision tag set and / or the revision tags added are used as a second revision tag set corresponding to each element to be revised in the text to be revised.

9. The method according to claim 1, characterized in that: The method further comprises: Based on the correction strictness parameter sent by the user, the first set of correction labels corresponding to each element to be corrected in the text to be corrected is added and deleted to obtain the second set of correction labels corresponding to each element to be corrected in the text to be corrected, Among them, when the correction strictness parameter is greater than the correction strictness threshold, the number of correction labels in the second correction label set is less than the number of correction labels in the first correction label set; when the correction strictness parameter is less than or equal to the correction strictness threshold, the number of correction labels in the second correction label set is greater than or equal to the number of correction labels in the first correction label set.

10. The method according to claim 9, characterized in that The method of adding and deleting the first set of revision labels corresponding to each element to be revised in the text to be revised based on the revision strictness parameter sent by the user to obtain the second set of revision labels corresponding to each element to be revised in the text to be revised includes: Based on the revision strictness parameter and the priority of each revision label in the first revision label set, the revision labels in the first revision label set are added and deleted to obtain the second revision label.

11. The method according to claim 9, characterized in that The method of adding and deleting the first set of revision labels corresponding to each element to be revised in the text to be revised based on the revision strictness parameter sent by the user to obtain the second set of revision labels corresponding to each element to be revised in the text to be revised includes: Based on the revision strictness parameter, the revision labels in the first revision label set are randomly added or deleted to obtain the second revision labels.

12. The method according to any one of claims 8 to 11, characterized in that The determining the element score of each of the elements to be revised based on the first set of revision labels includes: Calculate the sub-score corresponding to each correction label in the second correction label set corresponding to each element to be corrected; Based on the sub-scores, an element score of each of the elements to be corrected is determined.

13. The method according to claim 7, characterized in that The first set of modified labels includes the first set of modified labels. The calculating of the sub-score corresponding to each correction label in the first set of correction labels corresponding to each element to be corrected includes: When the matching degree between the element to be revised and the first batch of revision labels is greater than a matching degree threshold, a sub-score corresponding to the first batch of revision labels is allocated to the element to be revised.

14. The method according to claim 1, characterized in that The method further comprises: Based on the element score of each element to be corrected, the score details are fed back to the user.

15. A method for training a label prediction model, characterized in that: include: Obtain a first sample text and a first annotation label set; Using the first sample text as input of a first neural network model, and obtaining a first prediction label set output by the first neural network model; Based on the first annotation label set and the first prediction label set, the parameters of the first neural network model are adjusted to obtain a first model.

16. A method for training a layout recognition model, characterized in that: include: Obtaining a second sample text and a first set of elements to be marked for correction; Using the second sample text as input of a second neural network model, and obtaining a first predicted set of elements to be corrected output by the second neural network model; Based on the first set of labeled elements to be corrected and the first set of predicted elements to be corrected, the parameters of the second neural network model are adjusted to obtain a second model.

17. An electronic device, characterized in that: include: A memory for storing instructions executed by one or more processors of the electronic device, and a processor, which is one of the one or more processors of the electronic device, for executing the text correction method described in any one of claims 1 to 14, or executing the training method of the label prediction model described in claim 15, or executing the training method of the layout recognition model described in claim 16.

18. A readable storage medium, characterized in that: The readable medium stores instructions, which, when executed on an electronic device, enable the electronic device to execute the text correction method described in any one of claims 1 to 14, or execute the training method of the label prediction model described in claim 15, or execute the training method of the layout recognition model described in claim 16.

Images