CN111008594B

CN111008594B - Error-correction question review method, related device and readable storage medium

Info

Publication number: CN111008594B
Application number: CN201911227742.7A
Authority: CN
Inventors: 王根; 何春江; 曾金舟; 马皓
Original assignee: iFlytek Co Ltd
Current assignee: iFlytek Co Ltd
Priority date: 2019-12-04
Filing date: 2019-12-04
Publication date: 2024-04-19
Anticipated expiration: 2039-12-04
Also published as: CN111008594A

Abstract

The application discloses a correction question review method, related equipment and a readable storage medium, which are characterized in that after an overall answer area image of a correction question to be reviewed is obtained, semantic segmentation processing is carried out on the image to obtain segmentation components in the image, and modification positions in the overall answer area image are determined based on the segmentation components in the overall answer area image; because the segmentation component obtained by carrying out semantic segmentation processing on the whole answer region image is not changed because the whole answer region image is not completely corresponding to the image of the answer template, the modification position can be accurately determined based on the segmentation component in the whole answer region image, and the accurate evaluation result of the correction problem can be obtained by analyzing the determined modification position.

Description

Error-correction question review method, related device and readable storage medium

Technical Field

The present application relates to the field of computer image processing technology, and more particularly, to a method for evaluating a correction question, a related device, and a readable storage medium.

Background

The error correction is a comprehensive test question integrating grammar knowledge and language capability, and mainly aims at examining the capability of students to identify errors and correct errors and comprehensively utilizing the capability of corresponding languages in the language. The answering machine answers questions by handwriting to add, delete or modify the characters in the questions.

In recent years, with rapid development of computer technology and information technology, particularly rapid advancement of artificial intelligence technology, use of artificial intelligence instead of traditional manual work has become a hotspot direction of various industries. The examination paper reading mode of the test education is also gradually changed from traditional pure manual examination to automatic examination by replacing part of manual work with artificial intelligence. Currently, in a large-scale examination (such as college entrance examination, middle school examination, and meeting examination), part of subjects and part of subjects (such as objective subjects, composition, blank filling subjects, etc.) can be automatically reviewed.

However, for the correction questions, symbols such as handwritten inserts and deletions appear due to the specificity of the questions, and symbol writing inaccuracy and non-uniform answer areas appear in answer of the answer questions, so that correction is difficult. So the correction questions still need to be reviewed manually so far. Particularly, when large-scale examination is carried out, the manual review mode is time-consuming and labor-consuming, great working pressure is brought to the reviewer, and in addition, the manual review is easily subjectivity-affected by the reviewer, so that the review result is not fair.

Therefore, how to provide an automatic review scheme for the error correction problem becomes a technical problem to be solved by those skilled in the art.

Disclosure of Invention

The present application has been made in view of the above-mentioned problems, and provides a debug review method, a related apparatus, and a readable storage medium. The specific scheme is as follows:

a method for error-correction question review, comprising:

acquiring an overall answer area image of a to-be-reviewed and corrected question;

Carrying out semantic segmentation processing on the integral answer region image to obtain segmentation components in the integral answer region image;

determining a modification position in the whole answer area image based on the segmentation component in the whole answer area image;

and analyzing the modification part in the integral answer area image to obtain the evaluation result of the correction questions.

Optionally, the semantic segmentation processing is performed on the whole answer area image to obtain a segmentation component in the whole answer area image, including:

Inputting the whole answer region image into a pre-trained semantic segmentation model to obtain segmentation components in the whole answer region image, wherein the semantic segmentation model is obtained by training a preset model by using a sample image set;

Wherein each set of sample images in the set of sample images comprises: and the original image of the error-correction question integral answer area and the reference image generated after labeling the category of each pixel in the original image.

Optionally, the determining, based on the segmentation component in the overall answer area image, a modification in the overall answer area image includes:

acquiring the category of each segmentation component in the integral answer area image;

determining target segmentation components from all the segmentation components according to the category to which each segmentation component belongs, wherein the category to which each target segmentation component belongs is a category for indicating a answer to modify content;

acquiring a word to be modified corresponding to the target segmentation component, wherein the word to be modified is a print word;

and determining the modification position in the whole answer area image according to the word to be modified.

Optionally, the determining, according to the word to be modified, a modification part in the overall answer area image includes:

Determining the areas corresponding to the preset number of print words on the left side of the word to be modified, determining the areas corresponding to the preset number of print words on the right side of the word to be modified, and determining the areas corresponding to the word to be modified as a modification place.

Optionally, the analyzing the modification part in the whole answer area image to obtain the review result of the correction questions includes:

generating descriptive information at each of the modifications;

Acquiring target standard answer description information corresponding to the description information of each modification place;

And comparing the description information of each modification position with the description information of the target standard answer to obtain a review result of the error correction questions.

Optionally, the generating the description information at each modification includes:

Determining a word to be modified, modification attribute information and a modified word at each modification position, wherein the modified word is a pure handwriting word;

And generating description information of each modification place according to the word to be modified, the modification attribute information and the modified word.

Optionally, the generating the description information of each modification place according to the word to be modified, the modification attribute information and the modified word includes:

Acquiring a preset number of printed words on the left side of the word to be modified and a preset number of printed words on the right side of the word to be modified;

And combining the word to be modified, the modification attribute information, the modified word, a preset number of print words on the left side of the word to be modified and a preset number of print words on the right side of the word to be modified to generate description information of each modification place.

Optionally, comparing the description information of each modification with the description information of the target standard answer to obtain a review result of the error correction question, including:

Calculating the similarity between the description information of each modification place and the description information of the target standard answer;

Obtaining a review result of each modification place according to the similarity between the description information of each modification place and the description information of the target standard answer;

And obtaining the review result of the error-correcting questions based on the review result of each modification place.

Optionally, the calculating the similarity between the description information of each modification and the description information of the target standard answer includes:

acquiring the total number of words contained in the target standard answer description information;

Determining that the descriptive information at each modification is less recognized than the target standard answer descriptive information and that the erroneous word is recognized by comparing the descriptive information at each modification with the target standard answer descriptive information;

acquiring the total number of the words with little recognition and the total number of the words with wrong recognition;

And calculating the similarity between the descriptive information of each modification and the descriptive information of the target standard answer according to the total number of words contained in the descriptive information of the target standard answer, the total number of words with little recognition and the total number of words with wrong recognition.

Optionally, the obtaining the review result of each modification site according to the similarity between the description information of each modification site and the description information of the target standard answer includes:

Acquiring a preset similarity threshold;

When the similarity between the description information of each modification site and the description information of the target standard answer is greater than or equal to the similarity threshold value, the modification at each modification site is reviewed to be correct;

And when the similarity between the description information of each modification and the description information of the target standard answer is smaller than the similarity threshold, reviewing modification errors of each modification.

An error-correction question review device, comprising:

the acquisition unit is used for acquiring an overall answer area image of the questions to be reviewed and corrected;

the semantic segmentation unit is used for carrying out semantic segmentation processing on the integral answer region image to obtain segmentation components in the integral answer region image;

A modification position determining unit, configured to determine a modification position in the overall answer area image based on the segmentation component in the overall answer area image;

And the review unit is used for analyzing the modification part in the integral answer area image to obtain the review result of the correction questions.

Optionally, the semantic segmentation unit includes:

The input unit is used for inputting the whole answer region image into a pre-trained semantic segmentation model to obtain segmentation components in the whole answer region image, and the semantic segmentation model is obtained by training a preset model by using a sample image set;

Optionally, the modification position determining unit includes:

The category acquisition unit is used for acquiring the category of each segmentation component in the integral answer area image;

A target segmentation component determining unit, configured to determine a target segmentation component from all segmentation components according to a category to which each segmentation component belongs, where each category to which each target segmentation component belongs is a category for indicating a answer to modify content;

The word to be modified obtaining unit is used for obtaining a word to be modified corresponding to the target segmentation component, wherein the word to be modified is a print word;

and the modification part determination subunit is used for determining the modification part in the whole answer area image according to the word to be modified.

Optionally, the modifying part determines the subunit, including:

the first determining subunit is configured to determine an area corresponding to a preset number of print words on the left side of the word to be modified, an area corresponding to a preset number of print words on the right side of the word to be modified, and the area corresponding to the word to be modified is a modification place.

Optionally, the evaluation unit includes:

a description information generating unit at the modification place for generating description information at each modification place;

a standard answer description information acquisition unit, configured to acquire target standard answer description information corresponding to the description information at each modification site;

And the comparison unit is used for comparing the description information of each modification position with the description information of the target standard answer to obtain a review result of the error correction questions.

Optionally, the description information generating unit at the modification includes:

A second determining subunit, configured to determine a word to be modified, modification attribute information, and a modified word at each modification location, where the modified word is a pure handwritten word;

And the first generation subunit is used for generating the description information of each modification place according to the word to be modified, the modification attribute information and the modified word.

Optionally, the first generating subunit includes:

the first acquisition subunit is used for acquiring a preset number of printed words on the left side of the word to be modified and a preset number of printed words on the right side of the word to be modified;

And the combination subunit is used for combining the word to be modified, the modification attribute information, the modified word, the preset number of print words on the left side of the word to be modified and the preset number of print words on the right side of the word to be modified to generate description information of each modification place.

Optionally, the comparing unit includes:

a similarity calculating unit, configured to calculate a similarity between the description information at each modification location and the description information of the target standard answer;

the first review result obtaining unit is used for obtaining the review result of each modification place according to the similarity between the description information of each modification place and the description information of the target standard answer;

and the second evaluation result obtaining unit is used for obtaining the evaluation result of the error-correction questions based on the evaluation result of each modification place.

Optionally, the similarity calculating unit is specifically configured to:

Optionally, the first review result obtaining unit is specifically configured to:

Acquiring a preset similarity threshold;

A system for correcting questions comprises a memory and a processor;

the memory is used for storing programs;

The processor is used for executing the program to realize each step of the error-correcting question review method.

A readable storage medium having stored thereon a computer program which, when executed by a processor, implements the steps of the debug review method as described above.

By means of the technical scheme, the application discloses a correction question review method, related equipment and a readable storage medium, after an overall answer area image of a correction question to be reviewed is obtained, semantic segmentation processing is carried out on the image to obtain segmentation components in the image, and modification positions in the overall answer area image are determined based on the segmentation components in the overall answer area image; because the segmentation component obtained by carrying out semantic segmentation processing on the whole answer region image is not changed because the whole answer region image is not completely corresponding to the image of the answer template, the modification position can be accurately determined based on the segmentation component in the whole answer region image, and the accurate evaluation result of the correction problem can be obtained by analyzing the determined modification position.

Drawings

Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are only for purposes of illustrating the preferred embodiments and are not to be construed as limiting the application. Also, like reference numerals are used to designate like parts throughout the figures. In the drawings:

FIG. 1 is a flow chart of a method for evaluating debug questions according to an embodiment of the present application;

FIG. 2 is a schematic diagram of a segmentation component in an overall answer region image according to an embodiment of the present application;

FIG. 3 is a schematic diagram of a connected domain calculation according to an embodiment of the present application;

fig. 4 is a flowchart of a method for determining a modification position in an overall answer area image based on a segmentation component in the overall answer area image according to an embodiment of the present application;

FIG. 5 is a schematic diagram of a modification of the present disclosure;

FIG. 6 is a schematic diagram of another modification of the disclosure of the embodiment of the present application;

FIG. 7 is a flow chart of a method for analyzing a modification in an overall answer question area image to obtain a review result of an error correction question according to an embodiment of the present application;

FIG. 8 is a diagram of a determination of a word to be modified in accordance with an embodiment of the present application;

fig. 9 is a schematic diagram of rendering description information of a modification of a standard answer into a picture according to an embodiment of the present application;

FIG. 10 is a schematic diagram of a device for evaluating questions of an error correction according to an embodiment of the present application;

FIG. 11 is a block diagram of a hardware architecture of a system for debugging and reviewing questions according to an embodiment of the present application.

Detailed Description

The following description of the embodiments of the present application will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present application, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the application without making any inventive effort, are intended to be within the scope of the application.

In the existing intelligent paper reading technology, a template of a to-be-read test question needs to be generated in advance, the template of the to-be-read test question comprises a single-point answer area preset according to a correct answer, an answer image corresponding to an answer written by an answer person on the to-be-read test paper needs to be obtained, and then the intelligent paper reading is realized based on the template of the to-be-read test question and the answer image. Specifically, firstly, segmenting answer images according to each single-point answer region in a template of a test question to be reviewed to obtain single-point answer images, and carrying out the following treatment on each single-point answer image: and carrying out image recognition on the single-point answer image to obtain a recognition result, and finally obtaining the review result of the single-point answer image by comparing the recognition result of the single-point answer image with the corresponding single-point answer region. After all the single-point answer images are processed, obtaining the answer review result written by the answer sheet on the test paper to be reviewed.

Based on the existing intelligent paper marking technology, in order to realize automatic review of error correction questions, the inventor performs research, and the initial thought is as follows:

The method comprises the steps of generating a template of the to-be-reviewed and correct questions in advance, wherein the template of the to-be-reviewed and correct questions comprises preset single-point answer areas according to correct answers, and each single-point answer area corresponds to one modification of the to-be-reviewed and correct questions. And obtaining answer images corresponding to the error-correction questions written by the answer questions of the test paper to be reviewed, and then realizing intelligent examination paper based on templates of the test paper to be reviewed and the answer images. Specifically, firstly, segmenting answer images according to each single-point answer region in a template of a test question to be reviewed to obtain single-point answer images, and carrying out the following treatment on each single-point answer image: and carrying out image recognition on the single-point answer image to obtain a recognition result, and finally obtaining the review result of the single-point answer image by comparing the recognition result of the single-point answer image with the corresponding single-point answer region. After all the single-point answer images are processed, obtaining the review result of the correct answer written by the answer sheet on the test paper to be reviewed.

However, at present, answer images corresponding to the correction questions written by the answer sheet on the test paper to be reviewed are obtained by photographing, shooting, scanning or directly importing pre-shot image files, and the generated single-point answer images have deviation due to the difference of the placement positions of the test paper or the deviation of photographing, shooting, scanning, and the like.

In view of the problems of the thought, the inventor conducts intensive research, and finally provides an error-correction question review method which can accurately identify error-correction questions and further improve the accuracy of automatically reviewing the error-correction questions. Next, the error-correction question review method provided by the present application will be described by way of the following examples.

It should be noted that, the method for evaluating the debug questions provided in the present application is suitable for automatically evaluating the type of the language debug questions (including the english debug questions), and the method can be applied to an intelligent device or system with an image acquisition and recognition function, which is not particularly limited in the embodiment of the present application.

Referring to fig. 1, fig. 1 is a flow chart of a method for evaluating a debug problem according to an embodiment of the present application, where the method may include:

S101: and obtaining an overall answer area image of the questions to be reviewed and corrected.

In this embodiment, the questions to be reviewed may be language questions (including english questions) that the examinee answers in a large-scale examination (e.g., college entrance examination, middle examination, and conference examination). The whole answer area of the correction questions to be reviewed can be an answer area set for the correction questions on the test paper or the answer sheet, wherein the answer area comprises the stem of the whole correction questions, and the examinee answers in a handwriting mode. The whole answer area image of the questions to be reviewed and corrected can be obtained through photographing, shooting, scanning and other modes.

S102: and carrying out semantic segmentation processing on the integral answer region image to obtain segmentation components in the integral answer region image.

In this embodiment, the whole answer region image may be preprocessed first to obtain the preprocessed whole answer region image, and then the whole answer region image after preprocessing may be subjected to semantic segmentation processing to obtain the segmentation component in the whole answer region image.

The pretreatment method can be various. As an implementation manner, the whole answer area image may be normalized, for example, the long side of the whole answer area image is normalized to 1024, and the short side is scaled according to the corresponding proportion, so as to ensure that the aspect ratio of the whole answer area image is unchanged. As still another embodiment, the handwritten symbols in the entire answer area image may be subjected to a normalization process, for example, the head and tail areas of the hand-written deleter having a long head and tail may be erased, and the entire answer area image may be subjected to an image enhancement process or the like, which is not limited to this application.

It should be noted that, there are various ways of performing semantic segmentation processing on the whole answer region image, but no matter what semantic segmentation processing way is adopted, the segmentation component in the whole answer region image can be obtained. In this embodiment, the overall answer area image may include multiple types of segmentation components, and the different types of segmentation components may identify whether the corresponding content is the stem content of the correction question or the modified content of the answer.

For ease of understanding, referring to fig. 2, fig. 2 is a schematic diagram of a segmentation component in an overall answer area image according to an embodiment of the present application, and as shown in fig. 2, a part of each rectangular frame is each segmentation component in the overall answer area image, and different types of segmentation components may be identified by using different types of rectangular frames, for example, may be identified by using rectangular frames with different colors.

S103: and determining a modification part in the integral answer area image based on the segmentation part in the integral answer area image.

In this embodiment, since the segmentation component in the overall answer area image may identify whether the content corresponding to the segmentation component is the content of the correction question stem or the content of the modification of the answer, the modification position in the overall answer area image may be determined based on the segmentation component in the overall answer area image.

It should be noted that, in the existing scheme, the position information of the modification part intercepted on the answer template is used to determine the modification part from the integral answer area image, but the integral answer area image may not completely correspond to the answer template due to various reasons (such as scanner reasons, non-uniform placement positions of test papers or answer cards, etc.), in this case, deviation may occur between the determined modification part and the modification part intercepted on the answer template, which affects the accuracy of reading.

In this embodiment, the modification part is determined based on the segmentation part in the whole answer area image, and the segmentation part in the whole answer area image obtained by performing semantic segmentation processing on the whole answer area image is not changed due to incomplete correspondence between the whole answer area image and the image of the answer template, so that the modification part can be accurately determined based on the segmentation part in the whole answer area image, and the accuracy of the modification part lays a foundation for accurate subsequent review.

S104: and analyzing the modification part in the integral answer area image to obtain the evaluation result of the correction questions.

In this embodiment, the modification position in the whole answer area image may be analyzed in various manners, and since the determined modification position can be ensured to be accurate, the accuracy of the review result of the obtained error correction question may be improved no matter whether an existing scheme such as image recognition or an analysis scheme subsequently proposed by the present application is adopted.

The application also discloses an implementation mode for carrying out semantic segmentation processing on the whole answer region image, and specifically, the whole answer region image can be input into a pre-trained semantic segmentation model to obtain segmentation components in the whole answer region image.

It should be noted that, the semantic segmentation model is obtained by training a preset model by using a sample image set. The preset model may specifically be a network structure model such as DeepLabv3+ for example, and of course, other network structure models are also within the protection scope of the present application.

In the present application, a set of sample images for training a semantic segmentation model includes a plurality of sets of sample images, each set of sample images including: and the original image of the error-correction question integral answer area and the reference image generated after labeling the category of each pixel in the original image. Wherein, all sample images can cover different areas, different schools, different examination rooms, different writing habits and the like. For example, the sample image may include a forged sample image, and the forging manner may be various, for example, the real sample image may be generated into different print images by using different fonts, or based on the real sample image, the scanned image may be obtained by writing different handwritten words, handwritten inserted symbols, and alternative symbols multiple times, or the handwritten words and the print words may be judged in similarity according to the edit distance of the words, so that the sample image may be generated in a large scale.

In the application, the category marked for each pixel in the original image can comprise a category used for indicating the content of the error-correction stem and a category used for indicating the content of the modification of the answer, wherein the category used for indicating the content of the error-correction stem can comprise a printed word, a background class and the like, and the category used for indicating the content of the modification of the answer can comprise a printed word with a deleter, a purely handwritten word, an insert, a replacement, a change position symbol and the like. Before labeling, the identification of each category can be predetermined, and when labeling, the identification is directly adopted for labeling.

For example, each identity and corresponding category may be as shown in the following table:

Identification mark	Category(s)
		0	Print words
1	Printed word with deleter
		2	Pure hand written word
3	Replacing symbol
		4	Insert symbol
5	Background class

In the application, when the sample image set is adopted to train the semantic segmentation model, the original image is taken as output, and the reference image is taken as a training target. When the whole answer region image is input into a pre-trained semantic segmentation model to obtain segmentation components in the whole answer region image, 8 neighborhood computing connected domains are needed for the pixels of the same category, so that the pixels of the same category are gathered together. When the connected domain is calculated, as shown in fig. 3, the island phenomenon exists, and if the white points are surrounded by the black points and the number of the white points is small, the classification of the surrounded black points is directly adopted to replace the white points. Therefore, in the present application, the threshold of island area can be set to be smaller than 10, and if smaller than 10, the category of the pixel point is considered to be replaced by the surrounding category.

In the present application, a specific implementation manner of determining a modification position in an overall answer area image based on a segmentation component in the overall answer area image is further disclosed, referring to fig. 4 specifically, fig. 4 is a schematic flow diagram of a method for determining a modification position in an overall answer area image based on a segmentation component in the overall answer area image disclosed in the embodiment of the present application, where the method includes the following steps:

s401: and obtaining the category of each segmentation component in the integral answer area image.

In the application, the result output by the semantic segmentation model comprises the category of each segmentation component in the whole answer region image, so that the category of each segmentation component in the whole answer region image can be obtained from the result output by the semantic segmentation model.

S402: and determining a target segmentation part from all segmentation parts according to the category of each segmentation part.

In the application, each target segmentation component belongs to a category for indicating an answer to modify content. The segmentation components for indicating the category of modified content of the answer may identify that these segmentation components are newly added modification parts compared to the original stem, which may then be considered as a modification of the wrong question by the answer. Therefore, in the present application, it is necessary to determine, from all the segmentation components, a segmentation component belonging to a category (e.g., a printed word with a deleter, an insert, a replacement, a transpose, etc.) for instructing an answer person to modify content as a target segmentation component, so as to determine a modification in the overall answer area image based on the determined target segmentation component.

S403: and acquiring a word to be modified corresponding to the target segmentation component, wherein the word to be modified is a print word.

In the application, after the target segmentation component is determined, a word to be modified corresponding to the target segmentation component can be acquired, wherein the word to be modified is a print word.

The ways of determining the corresponding word to be modified by the target segmentation components of different categories are different, and the method can be specifically as follows:

When the type of the target segmentation component is a print word with a deleter, determining that the word to be modified corresponding to the target segmentation component is a print word with a deleter symbol.

When the type of the target segmentation component is an insert, determining that the word to be modified corresponding to the target segmentation component is a print word on the left side of the sharp corner of the insert. It should be noted that, because there is some answering person writing out of specification, there is the condition that the closed angle is not closed, under this condition, can carry out the straight line and detect, fit out the contained angle of two straight lines, and then regard the word on the left side of contained angle as the word of waiting to modify.

When the type of the target segmentation component is a replacement, it is determined that the word to be modified corresponding to the target segmentation component is a print word above the replacement Fu Zheng. In addition, since there is some irregular writing of a partial answer, there is a case where a substitution symbol overlaps with a plurality of print words, and in this case, a print word having the largest overlap is preferable as a word to be modified.

S404: and determining the modification position in the whole answer area image according to the word to be modified.

In the correction questions, the standard answer area of the answer questions is a blank area below the word to be corrected, and in the application, one implementation of determining the correction place in the whole answer area image according to the word to be corrected can be as follows: the blank area below the word to be modified can be determined to be a modification place in the whole answer area image. Fig. 5 is a schematic diagram of a modification disclosed in the embodiment of the present application, and the area in the black frame in fig. 5 is a modification.

However, since the writing of partial answer is not standard, the answer is performed in the area outside the standard answer area, and if only the blank area below the word to be modified is used as the modification place, the modification place may not contain the answer of the answer.

To solve such a problem, in the present application, another embodiment of determining the modification point in the overall answer area image according to the word to be modified may be: determining the areas corresponding to the preset number of print words on the left side of the word to be modified, determining the areas corresponding to the preset number of print words on the right side of the word to be modified, and determining the areas corresponding to the word to be modified as a modification place. In the present application, the region corresponding to the word may specifically be a region corresponding to the position information of the word. The location information of the word may be obtained from the semantic segmentation result.

For example, an area corresponding to 3 print words on the left side of the word to be modified may be determined, an area corresponding to 3 print words on the right side of the word to be modified may be determined, and the area corresponding to the word to be modified is a modification place. Fig. 6 is a schematic diagram of another modification disclosed in the embodiment of the present application, and the area in the black frame in fig. 6 is a modification.

Although the accuracy of the evaluation result of the correction question can be improved by adopting the existing modification analysis mode, such as image recognition, but still not optimized, the application also discloses an implementation mode for analyzing the modification position in the whole answer area image to obtain the evaluation result of the correction question, and particularly referring to fig. 7, fig. 7 is a flow diagram of a method for analyzing the modification position in the whole answer area image to obtain the evaluation result of the correction question, which comprises the following steps:

S701: and generating description information of each modification.

In the present application, the description information of each modification station at least needs to include the word to be modified, the modification attribute information and the modified word of the modification station, so in the present application, one implementation manner of generating the description information of each modification station is as follows: and determining the word to be modified, the modification attribute information and the modified word at each modification position, and combining the word to be modified, the modification attribute information and the modified word at each modification position to generate the description information at the modification position.

Specifically, the manner of determining the word to be modified, the modification attribute information, and the modified word at each modification site may be as follows:

The modification symbol of each modification site is determined, the modification symbol is identified, and the modification attribute information of each modification site is determined, for example, the modification symbol is a delete, the modification attribute information is delete, the modification symbol is a replace, the modification attribute information is repalce (replace), the modification symbol is an insert, the modification attribute information is insert (insert), and the modification attribute information is repins (pending) for the modification symbol of this case due to the presence of the replace and insert Fu Nian connection.

And determining a word to be modified at each modification position according to the modification symbol at each modification position, for example, determining a print word with a deleter as the word to be modified when the modification symbol is a deleter, determining a print word above the replacement Fu Zheng as the word to be modified when the modification symbol is a replacement symbol, and determining the print word at the left side of the sharp corner of the interposer as the word to be modified when the modification symbol is an interposer. In addition, since there is some irregular writing of a partial answer, there is a case where a substitution symbol overlaps with a plurality of print words, and in this case, a print word having the largest overlap is preferable as a word to be modified. Because of the fact that partial answer persons write irregularly and the fact that sharp angles are not closed, under the circumstance, straight line detection can be conducted, an included angle of two straight lines is fitted, and then words on the left side of the included angle are used as words to be modified. And (3) modifying the symbol to be timed, and determining the printed word with the largest overlap ratio with the modified word as the word to be modified. As shown in FIG. 8, the handwritten word "exam" has no modified symbol assignment because the modified symbol is written on the printed word, while "exam" coincides with the printed word "exams" and "but" and the modified content is generally at the bottom right of the modified word in combination with the general writing habit. So "exams" is selected as the word to be modified.

It should be noted that, the modified word at each modification is selected from the handwritten words, so that the modified word at each modification may be determined by identifying the handwritten word at each modification, and specifically, the handwritten word at each modification may be identified as the modified word at each modification using an identification model (e.g. cnn+gru). If no handwritten word is present after recognition at a modification, the modified word at the modification is set to be blank, and typically only handwritten words are present upon insertion and replacement, so that the modified word is blank for deletion.

The principle of selecting the modified word from the handwritten word at each modification is that the word under the modification symbol and having the highest overlap with the modification symbol is selected.

It should be noted that the word to be modified, the modification attribute information, and the modified word at each modification site may be combined using, for example, "[ C ]/modification attribute information [ F ]", where [ C ] is the word to be modified at each modification site, and [ F ] is the modified word at each modification site.

After the inventor generates the description information of each modification place in the mode and obtains the review result of the error correction questions based on the description information, the consistency rate of the automatic review result and the manual review result is not very high through statistics. In order to solve the problem, the inventor proposes that when the description information of each modification place is generated, the content on the left side of the word to be modified and the content on the right side of the word to be modified are comprehensively considered, so that the accuracy can be improved to a certain extent.

Therefore, the application also discloses another embodiment for generating the description information of each modification, in particular: determining a word to be modified, modification attribute information and a modified word at each modification position, wherein the modified word is a pure handwriting word; acquiring a preset number of printed words on the left side of the word to be modified and a preset number of printed words on the right side of the word to be modified; and combining the word to be modified, the modification attribute information, the modified word, the preset number of print words on the left side of the word to be modified and the preset number of print words on the right side of the word to be modified to generate description information of each modification place.

It should be noted that under normal conditions, the replacement symbol and the replaced content will be right below the word to be modified, but if the memory of the replaced content is long, the content will shift left and right, and when the manual review is performed, the information of the left and right words will be referred to for verification, but two places are more modified, and cannot be expanded more, so the inventor finds that the consistency ratio of the automatic review result and the manual review result can be highest by considering the 2 printed words on the left side of the word to be modified and the 2 printed words on the right side of the word to be modified through multiple experimental demonstration. The test results are shown in the following table:

Method of	Correcting man-machine consistency (%)
		Containing only modified words	89.99
Modified word + right one print word	92.72
		Modified word + one print word to the left	92.36
Modified word + left and right word each	95.97
		Modified word + left and right two words each	97.12
Modified word + left and right three words each	96.95

Thus, as a preferred embodiment, the preset number of print words on the left side of the word to be modified and the preset number of print words on the right side of the word to be modified in the present application may be 2 print words on the left side of the word to be modified and 2 print words on the right side of the word to be modified.

In the present application, the description information of each modification site may be generated by combining a to-be-modified word, modification attribute information, a modified word, 2 print words on the left side of the to-be-modified word, and 2 print words on the right side of the to-be-modified word, using, for example, "AB [ C ]/modification attribute information [ F ] D E", where A, B is the 2 print words on the left side of the to-be-modified word, [ C ] is the to-be-modified word of each modification site, [ F ] is the modified word of each modification site, and D, E is the 2 print words on the right side of the to-be-modified word.

Since there are a plurality of lines of the correction question stem, if the correction is performed at the beginning of a certain line or the correction is performed at the end of a certain line, there may be no 2 print words on the left side of the word to be corrected or no 2 print words on the right side of the word to be corrected, and in this case, zero or one may be directly fetched. Examples are: "One [ hundreds ]/reproduction [ hundred ] andtwenty".

S702: and acquiring target standard answer description information corresponding to the description information of each modification.

In the present application, description information of the modification site of the standard answer may be generated in advance, each of the standard answer description information including line information. It should be noted that, the description information at the modification of the standard answer generated in advance is also described in a form such as "[ C ]/modification attribute information [ F ]" or "AB [ C ]/modification attribute information [ F ] DE", and for convenience of understanding, fig. 9 is a schematic diagram showing rendering of the description information at the modification of the standard answer into a picture. The portion in each box in fig. 9 corresponds to a modification of the standard answer, and the description information of the modification is generated in advance.

In the present application, standard answer description information corresponding to the line information at each modification may be determined as target standard answer description information. The line information at each modification may be obtained from the semantic segmentation result.

S703: and comparing the description information of each modification position with the description information of the target standard answer to obtain a review result of the error correction questions.

In the application, the description information of each modification position can be compared with the description information of the target standard answer in a plurality of modes to obtain the evaluation result of the error correction questions.

As an implementation manner, the similarity between the description information of each modification and the description information of the target standard answer can be calculated; obtaining a review result of each modification place according to the similarity between the description information of each modification place and the description information of the target standard answer; and obtaining the review result of the error-correcting questions based on the review result of each modification place. Based on the evaluation result of each modification, the specific implementation mode of the evaluation result of the error-correction question is as follows: after the review result of each modification place is obtained, the number of modification places in the standard answer is combined, so that the review result of the whole error-correcting question is obtained, and the review result can be specifically a score.

Wherein the calculating the similarity between the description information of each modification and the description information of the target standard answer comprises the following steps: acquiring the total number of words contained in the target standard answer description information; determining that the descriptive information at each modification is less recognized than the target standard answer descriptive information and that the erroneous word is recognized by comparing the descriptive information at each modification with the target standard answer descriptive information; acquiring the total number of the words with little recognition and the total number of the words with wrong recognition; and calculating the similarity between the descriptive information of each modification and the descriptive information of the target standard answer according to the total number of words contained in the descriptive information of the target standard answer, the total number of words with little recognition and the total number of words with wrong recognition.

Examples are: the degree of similarity is described as

N is the total number of words contained in the target standard answer description information; d is the total number of words with little recognition; s is the total number of words in which errors are recognized.

For example, the standard answer is "One [ hundreds ]/replacement [ threaded ] AND TWENTY", and the examinee responds to "One [ hundreds ]/repins [ threaded ] AND TWENTY", where n=6, s=1, d=0, so similar=83%, when the modification attribute information repins is modified to replacement, similar=100%. When modification attribute information repins is modified to insert, similar=83%.

And obtaining the reading result of each modification according to the similarity between the description information of each modification and the description information of the target standard answer, wherein the reading result comprises the following steps: acquiring a preset similarity threshold; when the similarity between the description information of each modification site and the description information of the target standard answer is greater than or equal to the similarity threshold value, the modification at each modification site is reviewed to be correct; and when the similarity between the description information of each modification and the description information of the target standard answer is smaller than the similarity threshold, reviewing modification errors of each modification.

Through multiple experiments, the inventor finds that the consistency rate of the automatic evaluation result and the manual evaluation result is highest when the similarity threshold value is 95%. The test results are shown in the following table:

similarity threshold (%)	Man-machine consistency (%)
		80	85.6
85	86.7
		90	95.43
95	97.12
		97	96.53

Therefore, as a preferred embodiment, the preset similarity threshold in the present application may be 95%.

It should be noted that, in the present application, when the similarity between the description information at each modification and the description information of the target standard answer is smaller than the similarity threshold, the method further includes: and generating prompt information, wherein the prompt information is used for prompting manual processing.

However, if the similarity is smaller than the similarity threshold value, possibly because of modification errors, if the manual processing is prompted every time such a situation occurs, the review efficiency is reduced, and therefore, whether the word with the wrong recognition is a word for indicating modification attributes can be judged first; if so, the prompt information is regenerated, manual processing is prompted, and the error rate is reduced under the condition of not wasting manpower.

The error-correction-problem review device disclosed by the embodiment of the application is described below, and the error-correction-problem review device described below and the error-correction-problem review method described above can be correspondingly referred to each other.

Referring to fig. 10, fig. 10 is a schematic structural diagram of an error-correction question review device according to an embodiment of the present application. As shown in fig. 10, the error-correction-question review apparatus may include:

An obtaining unit 11, configured to obtain an overall answer area image of a question to be reviewed and corrected;

the semantic segmentation unit 12 is used for performing semantic segmentation processing on the whole answer region image to obtain segmentation components in the whole answer region image;

A modification position determining unit 13, configured to determine a modification position in the overall answer area image based on the segmentation component in the overall answer area image;

and the review unit 14 is used for analyzing the modification part in the whole answer area image to obtain the review result of the correction questions.

As an embodiment, the semantic segmentation unit includes:

As an embodiment, the modification position determining unit includes:

As an embodiment, the modification determining subunit includes:

As an embodiment, the evaluation unit includes:

As an embodiment, the description information generating unit at the modification includes:

As an embodiment, the first generating subunit includes:

As an embodiment, the comparing unit includes:

As an embodiment, the similarity calculation unit is specifically configured to:

As an embodiment, the first review result obtaining unit is specifically configured to:

Acquiring a preset similarity threshold;

It should be noted that, the specific functional implementation of each unit is described in detail above, and this embodiment is not repeated.

FIG. 11 is a block diagram of a hardware structure of an error-correction-question review system according to an embodiment of the present application, and referring to FIG. 11, the hardware structure of the error-correction-question review system may include: at least one processor 1, at least one communication interface 2, at least one memory 3 and at least one communication bus 4;

In the embodiment of the application, the number of the processor 1, the communication interface 2, the memory 3 and the communication bus 4 is at least one, and the processor 1, the communication interface 2 and the memory 3 complete the communication with each other through the communication bus 4;

The processor 1 may be a central processing unit CPU, or an Application-specific integrated Circuit ASIC (Application SPECIFIC INTEGRATED Circuit), or one or more integrated circuits configured to implement embodiments of the present invention, etc.;

the memory 3 may comprise a high-speed RAM memory, and may further comprise a non-volatile memory (non-volatile memory) or the like, such as at least one magnetic disk memory;

wherein the memory stores a program, the processor is operable to invoke the program stored in the memory, the program operable to:

Alternatively, the refinement function and the extension function of the program may be described with reference to the above.

The embodiment of the present application also provides a storage medium storing a program adapted to be executed by a processor, the program being configured to:

Finally, it is further noted that relational terms such as first and second, and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element.

In the present specification, each embodiment is described in a progressive manner, and each embodiment is mainly described in a different point from other embodiments, and identical and similar parts between the embodiments are all enough to refer to each other.

The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for evaluating a debug question, comprising:

Carrying out semantic segmentation processing on the integral answer region image to obtain segmentation components in the integral answer region image, wherein the segmentation components in the integral answer region image comprise multiple categories, and the segmentation components of different categories are used for marking that the corresponding content is the question stem content of the correction question or the modification content of the answer;

Analyzing the modification part in the integral answer area image to obtain the evaluation result of the correction questions;

The determining, based on the segmentation component in the overall answer area image, a modification in the overall answer area image includes:

2. The method according to claim 1, wherein the semantic segmentation of the whole answer area image to obtain segmentation components in the whole answer area image comprises:

3. The method of claim 1, wherein said determining a modification in the overall answer area image from the word to be modified comprises:

4. The method of claim 1, wherein analyzing the modification in the overall answer area image to obtain a review of the correction questions comprises:

generating descriptive information at each of the modifications;

5. The method of claim 4, wherein said generating descriptive information at each of said modifications comprises:

6. The method of claim 5, wherein generating the description information for each modification according to the word to be modified, the modification attribute information, and the modified word comprises:

7. The method of claim 4, wherein comparing the description information of each modification with the description information of the target standard answer to obtain a review result of the error correction question comprises:

8. The method of claim 7, wherein said calculating the similarity of the descriptive information at each modification to the descriptive information of the target standard answer comprises:

9. The method of claim 8, wherein the obtaining the review result for each modification according to the similarity between the description information of each modification and the description information of the target standard answer comprises:

Acquiring a preset similarity threshold;

10. An error correction question review device, comprising:

The semantic segmentation unit is used for carrying out semantic segmentation processing on the integral answer region image to obtain segmentation components in the integral answer region image, wherein the segmentation components in the integral answer region image comprise multiple categories, and the segmentation components of different categories are used for marking that the corresponding content is the question stem content of the correction questions or the correction content of the answer questions;

the review unit is used for analyzing the modification positions in the integral answer area image to obtain review results of the correction questions;

The modification position determining unit is specifically configured to:

11. An error-correcting question review system is characterized by comprising a memory and a processor;

the memory is used for storing programs;

the processor is configured to execute the program to implement the respective steps of the debug review method as claimed in any one of claims 1 to 9.

12. A readable storage medium having stored thereon a computer program, which, when executed by a processor, implements the steps of the debug review method of any of claims 1 to 9.