Text sample recognition, analysis and labeling method and system and storage medium

Abstract

The invention belongs to the field of text identification analysis, and relates to a text sample identification, analysis and labeling method and system and a storage medium. According to the method, high-precision content recognition is realized by combining the OCR technology and the dynamic positioning of the answering area and calculating the relative distance and the correlation degree between the nearby text blocks and the answering area; personalized handwriting features are extracted through a neural network to construct a writer handwriting model, and answer content of a text sample is converted into standard characters, so that the recognition accuracy of scribbling handwriting is remarkably improved; deep structured analysis and knowledge point association analysis of answer content are realized through a LaTex analysis tree and a subject knowledge graph; scientific and comprehensive scoring and error marking are realized through three dimensions of a structure reduction rate, an operation process coincidence rate and a knowledge point association accuracy rate. The method not only improves the intelligence level of test paper text processing, but also provides efficient and accurate technical support for education evaluation and personalized teaching.

Description

Technical Field

[0001] The present invention belongs to the field of text identification analysis, and relates to a text sample recognition analysis and annotation method, system and storage medium. Background Art

[0002] In recent years, with the development of educational informatization and automated review technology, intelligent recognition and analysis technology of text samples has gradually become a research hotspot. Especially in the fields of teaching evaluation and automatic examination scoring, how to efficiently and accurately recognize handwritten answers, analyze their structure and analyze the association of knowledge points has always been a difficult technical problem.

[0003] At present, the existing technology mainly relies on optical character recognition (OCR) technology to extract text content, and combines rules or template matching to determine the answer. For example, China Patent Publication No. CN113971805A is an intelligent scoring method that combines machine vision and semantic analysis, including: handwriting recognition stage; scaling the input test paper scan image; preprocessing the scaled test paper scan image to obtain the connected domain of the handwritten text in the scaled test paper image; obtaining single-character text based on the handwritten text in the connected domain; verifying and correcting all single-character texts to obtain the test paper answer text; intelligent scoring stage; obtaining the test paper answer text vector and the standard answer text vector based on the test paper answer text and the standard answer text; calculating the similarity between the test paper answer text vector and the standard answer text vector to obtain the matching degree between the test paper answer text and the standard answer text, and the matching degree is the final question answer score.

[0004] For example, the deep learning-based intelligent marking implementation method, system, and computer program disclosed in Chinese patent publication number CN110110585B scan students' test papers, convert paper test paper information into image information and store it on the local computer hard disk; through optical character recognition, the natural scene text detection model is used to detect the position of the examinee's answer to the subjective question, and the text information of the examinee's answer is recognized using the text recognition model and the connectionist time classification model; through natural language processing, it helps the examiner to score the subjective questions, performs similarity matching through word segmentation, and then finds spelling and grammatical errors through semantic analysis of the context, assisting the examiner in assigning a score to the test paper.

[0005] However, this type of method has the following significant problems: 1. Limitations of handwritten text recognition and content positioning: The recognition accuracy of personalized handwritten text is low, especially in the distinction between the answer area and the non-answer area. It is easy to misjudge the content because the answer content exceeds the answer area. It is difficult to accurately extract the complete answer information.

[0006] 2. Insufficient neatness processing: Handwritten texts often have problems such as illegible handwriting and irregular character writing. Existing technologies make it difficult to convert them into a structured standard format, resulting in low efficiency in subsequent parsing and comparison. Although some studies have attempted to achieve standardization through template matching, they have not combined the personalized handwriting characteristics of the writer and are difficult to adapt to the writing habits of different users.

[0007] 3. Single scoring dimension: The existing scoring only focuses on the final correctness of the answer, while ignoring the structure, process and knowledge relevance of the answer. For example, if the calculation steps in a math problem are wrong but the result is correct, the existing method may not be able to effectively determine the process error, which reduces the scientific nature and guiding value of the scoring. Summary of the invention

[0008] In view of this, in order to solve the problems raised in the above background technology, a text sample recognition, analysis and annotation method, system and storage medium that integrate handwriting modeling, structured analysis and multi-dimensional correlation analysis are proposed.

[0009] According to a first aspect of the present invention, a method for identifying, analyzing and annotating text samples is provided, comprising the following steps: S1, locating and distinguishing the content of the text sample to obtain the answer content of the text sample.

[0010] S2. Based on the handwriting habits of the writer of the text sample, infer the neat answer content of the text sample.

[0011] S3. Use LaTex parse tree to parse the neat answer content of the text sample, obtain the parse tree node structure of the neat answer content, and determine whether the parse tree node structure is completely consistent with the standard answer.

[0012] S4. When the data is incomplete, structural comparison and knowledge point association are performed based on the parse tree node structure, and the structural restoration rate of the parse tree node structure, the compliance rate of the operation process and the accuracy rate of knowledge point association are analyzed.

[0013] S5. Analyze the score of neat answers based on the structure restoration rate, calculation process compliance rate and knowledge point association accuracy, and mark them. At the same time, mark the wrong structure, wrong calculation process and knowledge points not involved.

[0014] According to a second aspect of the present invention, a text sample recognition, analysis and annotation system is provided, comprising: a content positioning and differentiation module, which locates and differentiates the content of the text sample to obtain the answer content of the text sample.

[0015] The neat answer inference module combines the usual handwriting of the writer of the text sample to infer the neat answer content of the text sample.

[0016] The neat answer content parsing module uses the LaTex parse tree to parse the neat answer content of the text sample, obtains the parse tree node structure of the neat answer content, and determines whether the parse tree node structure is completely consistent with the standard answer.

[0017] The content association analysis module performs structural comparison and knowledge point association based on the parse tree node structure when it is incomplete, and analyzes the structural restoration rate of the parse tree node structure, the compliance rate of the calculation process and the accuracy rate of knowledge point association.

[0018] The content scoring and annotation module analyzes the score of neat answer content based on the restoration rate, calculation process compliance rate and knowledge point association accuracy, and annotates it, while also marking incorrect structures, incorrect calculation processes and knowledge points not involved.

[0019] According to a third aspect of the present invention, there is provided a storage medium having a computer program burned therein, and when the computer program is executed in a memory of a server, the method for identifying, analyzing and annotating text samples is implemented.

[0020] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention combines OCR technology with dynamic positioning of the answer area, and by calculating the relative distance and correlation between the nearby text blocks and the answer area, it effectively solves the problem of content misjudgment when the answer content exceeds the answer area, thereby accurately extracting the complete answer content.

[0021] (2) The present invention extracts personalized handwriting features (such as stroke shape, structural proportion, and texture features) through a neural network, constructs a handwriting model of the writer, and converts the answer content of the text sample into standard characters, significantly improving the recognition accuracy of cursive handwriting and adapting to the writing habits of different users.

[0022] (3) The present invention uses the LaTex parse tree to perform hierarchical analysis of the neat answer content (such as text nodes, mathematical formula nodes). By comparing the parse tree node structure of the standard answer, the deep structural restoration of the answer content is achieved, thereby improving the accuracy of the answer content analysis and the structural processing capability.

[0023] (4) The present invention combines the structural restoration rate, the calculation process compliance rate and the knowledge point association accuracy rate to analyze the scoring of neat answer content, which not only improves the comprehensiveness and scientificity of the scoring, but also improves the scoring efficiency, adapts to multi-disciplinary application scenarios, and promotes teaching improvement and personalized teaching, and has significant educational application value.

[0024] (5) The present invention can accurately identify errors and provide targeted feedback by synchronously marking incorrect structural nodes, incorrect calculation processes and uninvolved knowledge points during the scoring process, helping students quickly locate weak links and supporting in-depth error tracing and learning improvement. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0026] Figure 1 It is a schematic diagram for implementing the method steps of the present invention.

[0027] Figure 2 It is a schematic diagram of the structured tree corresponding to the neat answer content of the present invention.

[0028] Figure 3 It is a schematic diagram of the connection of each module of the system of the present invention.

[0029] Figure 4 A block diagram of a storage medium for implementing the present invention. DETAILED DESCRIPTION

[0030] Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that the relative arrangement of components and steps, numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present invention unless otherwise specifically stated.

[0031] At the same time, it should be understood that for the convenience of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship.

[0032] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the invention, its application, or uses.

[0033] Technologies, methods, and equipment known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods, and equipment should be considered as part of the specification.

[0034] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0035] It should be noted that like reference numerals and letters refer to similar items in the following figures, and therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0036] At present, with the development of educational informatization and intelligence, the automatic recognition and annotation technology of test paper texts has gradually become an important research direction in the field of educational assessment. The traditional manual marking method is not only time-consuming and labor-intensive, but also easily affected by subjective factors, and it is difficult to meet the needs of large-scale examinations and personalized teaching. Therefore, how to achieve automatic recognition, structured analysis and intelligent scoring of test paper texts through technical means has become a hot topic in current research. Existing test paper text recognition and annotation technologies have significant deficiencies in handwriting recognition, structured analysis, multi-dimensional scoring and knowledge association analysis. The present invention solves the above problems through technological innovation and provides a high-precision, multi-dimensional and traceable solution for educational intelligence.

[0037] In view of this, the present invention proposes a text sample recognition, analysis and annotation method, which realizes high-precision content recognition through OCR technology and dynamic positioning of the answer area; realizes deep structured analysis of the answer content and knowledge point association analysis through LaTex parsing tree and subject knowledge graph; realizes scientific and comprehensive scoring and error annotation through three dimensions: structure restoration rate, operation process compliance rate, and knowledge point association accuracy rate. This method not only improves the intelligent level of test paper text processing, but also provides efficient and accurate technical support for education evaluation and personalized teaching.

[0038] like Figure 1 As shown, a text sample recognition, analysis and annotation method includes steps S1-S5.

[0039] S1. Locate and differentiate the content of the text sample to obtain the answer content of the text sample.

[0040] The specific content of step S1 is as follows: S11, using OCR technology to perform text recognition on the text sample to obtain the recognized text content. The OCR technology is an existing mature technology and will not be described in detail here.

[0041] S12. Perform structural feature recognition on the text content to locate the question stem area and its corresponding answer question area in the text content.

[0042] S13, identifying the text block in the answer area and the nearby text blocks outside the answer area, and calculating the relative distance between the nearby text blocks and the answer area and the correlation between the nearby text blocks and the text blocks in the answer area.

[0043] S14. Determine the answer content according to the relative distance and the correlation, and then obtain the answer content of the text sample.

[0044] In a specific embodiment, the structured feature of the text content is a question number feature. The question number feature can be obtained by extracting all question numbers from the standard typesetting content of the text sample, taking the question stem area after all question numbers as the question stem area, and extracting the answer area corresponding to all question numbers based on the standard typesetting content of the text sample.

[0045] The specific contents of step S13 are as follows: S131. Identify the text within the boundary of the answer area as a text block within the answer area, and record the text content of the text block.

[0046] S132. Based on the boundary of the answering area, a distance threshold is set, and the text within the distance range threshold is determined to be a nearby text block outside the answering area, and its text content is recorded.

[0047] S133. Use the coordinate comparison method to analyze the relative distance between the nearby text blocks and the answer area.

[0048] S134. Analyze the correlation between the nearby text block and the text block in the answer area based on the text content of the text block and the nearby text blocks.

[0049] In a specific embodiment, the determination of the relative distance between the nearby text block and the answer area specifically includes: establishing a two-dimensional coordinate system with the lower left corner of the text sample as the origin (0, 0), the horizontal axis represents the horizontal direction (from left to right is the positive direction), and the vertical axis represents the vertical direction (from bottom to top is the positive direction). The coordinates of the center point of the answer area are located in the two-dimensional coordinate system as follows: , the coordinates of the center point of the nearby text block are , using the Euclidean distance formula Get the relative distance between the nearby text block and the answer area.

[0050] In a specific embodiment, the specific content of step S134 is as follows: S134-1. Obtain the standard answer in the question stem area in the text content and build a vocabulary containing all keywords.

[0051] S134-2. Count the frequency of occurrence of each keyword in the text content of the text block in the vocabulary, convert it into vector form, and obtain the text keyword vector of the text block. , and similarly get the text keyword vector of the nearby text block .

[0052] The keywords can be words, numbers, letter combinations and comprehensive combinations.

[0053] Words refer to the words that appear in the text explanation during the answering process.

[0054] A numerical value refers to a number of numbers arranged or combined in a certain order or rule. For example, a mathematical expression ,in All are numerical values.

[0055] Letter combination refers to the arrangement or combination of several letters in a certain order or rule. , trigonometric functions ,function wait.

[0056] A comprehensive combination refers to the arrangement or combination of multiple numbers and letters, such as wait.

[0057] S134-3. Calculate the content similarity between the nearby text blocks and the text block in the answer area through cosine similarity.

[0058] S134-4. Extract the first and last sentence texts from the text content of the text block and the nearby text blocks, and compare to obtain the logical similarity between the nearby text block and the text block in the answer area.

[0059] Furthermore, the last sentence text of the text block is compared with the first sentence text of the nearby text block, the number of first identical keywords between the nearby text block and the text block is counted, and the number of keywords in the last sentence text of the text block and the number of keywords in the first sentence text of the nearby text block are obtained, the smallest number of keywords is screened out as the reference number of keywords, and the ratio of the first number of identical keywords to the reference number of keywords is taken as the first logical similarity; similarly, the first sentence text of the text block is compared with the last sentence text of the nearby text block to obtain the second logical similarity, and the maximum similarity is screened out from the first logical similarity and the second logical similarity as the logical similarity between the nearby text block and the text block in the answer area.

[0060] S134-5. Analyze the relevance between the nearby text blocks and the text blocks in the answer area based on content similarity and logical similarity.

[0061] ,in is the correlation, is the content similarity, is the logical similarity, is the content similarity weight, is the logical similarity weight. Considering content similarity and logical similarity, the correlation between the nearby text block and the text block in the answer area is evaluated. Weights can be assigned to different correlation factors, and then the comprehensive correlation is calculated. For example, the weight of content similarity is 0.6, and the weight of logical similarity is 0.4. In the text sample, content similarity directly reflects the degree of consistency between the nearby text block and the text block in the answer area in terms of specific information and facts. Because the accuracy of answering questions depends largely on the grasp of these core information, the weight of content similarity is relatively high (0.6). Logical similarity is mainly used to measure the similarity between the nearby text block and the text block in the answer area in terms of logical reasoning, causal relationship, hierarchical structure, etc. However, compared with the direct reflection of content similarity on factual information, logical similarity is a deeper relationship, so its weight is relatively low (0.4) to highlight the importance of content similarity in providing direct information.

[0062] In a specific embodiment, the distance threshold is set based on experience or preliminary analysis of text layout, and a distance of 50 pixels or a plurality of character widths is selected.

[0063] It should be noted that the present invention combines OCR technology with dynamic positioning of the answer area, and effectively solves the problem of content misjudgment when the answer content exceeds the answer area by calculating the relative distance and correlation between the nearby text blocks and the answer area, thereby accurately extracting the complete answer content.

[0064] S2. Based on the handwriting habits of the writer of the text sample, infer the neat answer content of the text sample.

[0065] The specific contents of step S2 are as follows: S21, collecting various texts written daily by the writers of the text samples to form a text set of the writers' daily handwriting. The various texts written daily by the writers include but are not limited to notes, homework and test answers, and all kinds of texts are entered into the teaching system by the writers themselves or by the teachers themselves.

[0066] S22. Use a neural network to extract multidimensional features from a daily handwriting text set to obtain a reference multidimensional feature vector for each character in the daily handwriting text set.

[0067] Furthermore, the neural network usually adopts a convolutional neural network, and the convolution layer of the convolutional neural network can automatically learn the local features of the characters, such as the shape and thickness of the strokes, etc.; the pooling layer reduces the dimension of the local features, extracts the main features, and reduces the amount of calculation. Through multi-layer convolution and pooling operations, the characters are gradually converted into multi-dimensional feature vectors. These multi-dimensional feature vectors contain information about the characters in terms of shape, structure, texture, etc., and can effectively characterize the handwriting characteristics of the writer.

[0068] In a specific embodiment, the step S22 includes: S221, locating and marking each character in the daily writing handwriting text set.

[0069] S222, input all the marked characters into the neural network for feature extraction, obtain the multi-dimensional features of all the characters in each text, and construct the multi-dimensional feature vectors of all the characters in each text.

[0070] S223, using cosine similarity to calculate the multi-dimensional feature similarity of all characters in each text and other remaining texts.

[0071] S224. If the multidimensional feature similarity of a certain character in a certain text and other remaining texts is greater than the set similarity threshold, the multidimensional feature similarity of the character in the text is used as the reference multidimensional feature vector of the character; otherwise, the majority multidimensional feature vector of the character in all texts is taken as the reference multidimensional feature vector of the character, and the reference multidimensional feature vectors of each character are counted.

[0072] Furthermore, the majority multidimensional feature vector refers to a multidimensional feature vector having the largest number of identical characters in all texts.

[0073] Furthermore, multidimensional features include stroke features, structural features, and texture features. Stroke features include stroke shape, stroke thickness, and inclination angle. Stroke shape refers to the classification of strokes of characters, such as straight lines, curves (including quadratic curves, cubic curves, etc.), broken lines, arcs, etc. Stroke thickness refers to extracting the thickness variation characteristics of strokes, sampling at equal intervals along the stroke direction, calculating the pixel width of each sampling point, and counting the mean, variance, maximum, minimum, and other statistics of the stroke width as features. The inclination angle refers to calculating the direction angle of the stroke, usually based on the horizontal direction, and determining the angle by calculating the coordinate difference between the starting point and the end point of the stroke. The direction angles of all strokes are counted, and the mean and standard deviation of the angle are calculated as features.

[0074] The structural features are the proportion of character components and the center of gravity of the character. The proportion of character components refers to the calculation of the proportional relationship of each component in the horizontal and vertical directions for characters with obvious component structures (such as left-right structure, top-bottom structure, etc.). For example, for left-right structure characters, the width ratio of the left and right components is calculated; for top-bottom structure characters, the height ratio of the top-bottom components is calculated. The center of gravity of a character refers to the calculation of the center of gravity position of the character. The center of gravity is the equivalent point of gravity acting on the character, which can be obtained by calculating the weighted average position of all pixel points of the character. The position of the center of gravity can reflect the balance and layout characteristics of the character when it is written.

[0075] Texture features are gray-level co-occurrence matrix and local binary pattern. Gray-level co-occurrence matrix refers to the calculation of gray-level co-occurrence matrix of character images in different directions (such as 0°, 45°, 90°, 135°) and at different distances, and the extraction of matrix energy, contrast, correlation, entropy and other features. Local binary pattern refers to an operator used to describe the local texture features of an image. For each pixel of each character image, its gray value relationship with the surrounding pixels is compared to obtain a binary sequence, which is converted into a decimal number as the LBP value of the pixel.

[0076] S23. Construct a handwriting model of the writer based on the reference multi-dimensional feature vector of each character.

[0077] S24, extracting multidimensional feature vectors of all characters from the answer content of the text sample, substituting them into the writer's handwriting model, and matching standard characters corresponding to all characters. The standard characters are characters in standard printing, such as Kaiti characters.

[0078] S25. Based on the standard characters corresponding to all characters, a neat answer content of the text sample is constructed.

[0079] It should be noted that the present invention extracts personalized handwriting features (such as stroke shape, structural proportion, and texture features) through a neural network, constructs a handwriting model of the writer, converts the answer content of the text sample into standard characters, significantly improves the recognition accuracy of cursive handwriting, and adapts to the writing habits of different users.

[0080] S3. Use LaTex parse tree to parse the neat answer content of the text sample, obtain the parse tree node structure of the neat answer content, and determine whether the parse tree node structure is completely consistent with the standard answer.

[0081] The specific contents of step S3 are as follows: S31, pre-processing the neat answer content of the text sample.

[0082] S32, using LaTex parse tree to convert the pre-processed neat answer content into a structured tree, and obtain the parse tree node structure of the neat answer content according to the structured tree, such as Figure 2 shown.

[0083] In a specific embodiment, the step S32 includes: S321, using a parsing tool based on the LaTex parsing tree to load the pre-processed neat answer content, scanning all sentences and expressions word by word and character by character in turn, and marking all words and expression components.

[0084] S322. Start building a parse tree with the pre-processed neat answer content as the root node, all sentences as text nodes, all expressions as mathematical formula nodes, all words in the sentences as child nodes, and all expression components in the expressions as child nodes in sequence according to character types.

[0085] S323. Organize all text nodes and their child nodes and all math formula nodes and their child nodes into a structured tree according to the hierarchical relationship.

[0086] Furthermore, the preprocessing of the neat answer content is text cleaning, sentence segmentation and format normalization. Text cleaning refers to the removal of irrelevant characters in the answer content, such as some redundant punctuation marks (such as multiple consecutive identical punctuation marks), special symbols (such as unknown placeholders that may be brought in by copying and pasting), blank characters (including spaces, tabs, and redundant parts of line breaks), etc. For example, if there are three consecutive spaces in the text, reduce them to one space to make the text more compact and concise; for some full-width punctuation marks that are mistakenly entered as half-width punctuation marks, perform corresponding conversions to ensure that punctuation is used in a standardized manner. Sentence segmentation refers to the use of a lexical analyzer to sentence the answer content after text cleaning, and decompose the answer content into individual sentences. Format normalization refers to converting expressions into LaTex format. Such as \frac (fraction), \sqrt (square root), \int (integral), \sum (sum), etc.

[0087] For example, an expression in the neat answer content after preprocessing is , converted to LaTex format The expression components are , , , , , , , -, and 3. Character types include variables, operators, constants, special symbols, and functions. and is a variable, , and - are operators, 1 and 3 are constants, For special symbols, is a power function.

[0088] S33. Compare the node structure of the parse tree of the neat answer content with the standard answer to determine whether the node structure of the parse tree of the neat answer content is completely consistent with the standard answer.

[0089] It should be noted that the present invention uses the LaTex parse tree to perform hierarchical analysis of neat answer content (such as text nodes, mathematical formula nodes), and by comparing the parse tree node structure of the standard answer, it achieves deep structured restoration of the answer content, thereby improving the accuracy of answer content analysis and structured processing capabilities.

[0090] S4. When the data is incomplete, structural comparison and knowledge point association are performed based on the parse tree node structure, and the structural restoration rate of the parse tree node structure, the compliance rate of the operation process and the accuracy rate of knowledge point association are analyzed.

[0091] The specific implementation content of step S4 is as follows: S41. Compare the parse tree node structure of the neat answer content with the parse tree node structure of the standard answer to obtain the matching nodes of the parse tree node structure, where the matching nodes are the number of text node matches and the number of mathematical formula node matches, and analyze the structural restoration rate of the parse tree node structure.

[0092] The analysis content of step S41 is as follows: S411, match the neat answer content with all child nodes of the text node in the parse tree node structure of the standard answer, count the number of child node matches of all text nodes, and analyze the child node matching rate of the text node.

[0093] S412, filtering text nodes whose sub-node matching rates are greater than a set matching rate, and counting the number of matching text nodes.

[0094] It should be noted that the sub-node matching rate of the text node is the ratio of the number of sub-node matching of the text node to the total number of sub-nodes of the corresponding text node.

[0095] The match rate can be set to .

[0096] S413, comparing the neat answer content with all the child nodes of the math formula node in the parse tree node structure of the standard answer, screening out math formula nodes with the same child nodes, and counting the number of math formula node matches.

[0097] S414. Analyze the structural restoration rate of the parse tree node structure by the number of text node matches and the number of mathematical formula node matches .

[0098] The formula for calculating the structural reduction rate is: , It is the structural restoration rate, which is used to measure the comprehensive restoration degree of text and mathematical formula node matching. is the number of text node matches, is the number of math formula node matches, is the total number of all text nodes, is the total number of all math formula nodes. Indicates the matching ratio of text nodes. represents the matching ratio of the math formula node, and and They are the weights of text nodes and math formula nodes in the total nodes (the sum of text nodes and math formula nodes). Adding the two according to their weights and multiplying by 100% gives the result restoration rate, which presents the overall matching situation in percentage form. When dealing with answer content containing text and math formulas, it is important to consider the matching of text and formulas separately. Different types of nodes have their own contributions to the completeness and accuracy of the overall content, so combining them in a weighted manner to calculate the restoration rate can more comprehensively evaluate the restoration effect after content processing. The following lists the structural restoration rate analysis of the parse tree node structure using multiple sets of data, as shown in Table 1.

[0099] Table 1: Results of multiple data analysis

[0100]

[0101] From these data, we can see that different matching numbers and total numbers of text nodes and math formula nodes will lead to different result restoration rates, and the higher the matching ratio of text nodes and math formula nodes, the higher the structural restoration rate.

[0102] S42. Map all nodes in the parse tree node structure with the associated knowledge points in the subject knowledge graph to which the text sample belongs, and based on the mapping, analyze whether there is a reasonable knowledge association path between the parse tree node and the standard answer parse tree node, and analyze the knowledge point association accuracy of the parse tree node structure.

[0103] Furthermore, the analysis content of step S42 is as follows: S421. Construct a subject knowledge graph for the subject to which the text sample belongs.

[0104] S422. Using all child nodes of the node in the parse tree node structure as clues, search for matching nodes in the subject knowledge graph to obtain a preliminary mapping of all nodes and associated knowledge points in the subject knowledge graph.

[0105] S423. Following the established mapping relationship, trace the associated knowledge points of the corresponding nodes of the standard answer parse tree in the subject knowledge graph. If the same associated knowledge points can be traced, then there is a reasonable knowledge association path between the parse tree node and the standard answer parse tree node. Count the number of nodes with reasonable knowledge association paths.

[0106] S424. The ratio of the number of nodes with reasonable knowledge association paths to the number of all nodes in the parse tree node structure is used as the knowledge point association accuracy of the parse tree node structure.

[0107] For example, a mathematical formula node that uses Newton's second law to calculate acceleration appears in the parse tree of the text sample. The related force analysis node should be found in the parse tree of the standard answer. And through the knowledge graph, we can know that Newton's second law is based on the derivation of acceleration based on force analysis. There is a reasonable association path between the two, indicating that the knowledge points are accurately associated. On the contrary, if such a path cannot be found, there is a disconnection between the knowledge points.

[0108] The subject knowledge graph refers to the collection of comprehensive and authoritative knowledge materials based on the subject areas involved, covering textbooks, academic papers, professional knowledge bases, etc., and constructing a detailed knowledge graph based on the causal, progressive, and parallel relationships between knowledge points. For example, in the field of mathematics, knowledge points are sorted out for branches such as algebra, geometry, and statistics, and the concepts related to functions (domain, range, monotonicity, parity, etc.), formulas (linear functions, quadratic functions, trigonometric function formulas, etc.) and their mutual relationships are clearly sorted out to construct a mathematical knowledge graph with a hierarchical structure. The nodes represent each knowledge point, and the edges represent the logical connections between knowledge points, such as causal, derivation, inclusion, and other relationships.

[0109] S43. Analyze all mathematical formula nodes in the parse tree node structure using the operation priority rule, obtain the number of nodes with correct operation logic, and analyze the compliance rate of the operation process of the parse tree node structure.

[0110] Furthermore, the specific contents of step S43 include: S431, using the operation priority rule to analyze whether all mathematical formula nodes in the parse tree node structure comply with the operation priority rule, and counting the mathematical formula nodes that comply with the operation priority rule.

[0111] S432. Based on the operation priority rule, the operation process and operation results of all mathematical formula nodes are verified by mathematical software tools, and the mathematical formula nodes whose operation process and operation results are consistent are obtained.

[0112] S433. The mathematical formula nodes that meet the operation priority rule, operation process and operation result are regarded as correct operation logic nodes, and the number of correct operation logic nodes is counted.

[0113] S434, analyzing the computational process compliance rate of the parse tree node structure , , is the compliance rate of the operation process, is the number of correct nodes for the operation logic, The number of all math formula nodes in the parse tree node structure.

[0114] Furthermore, the operation process of the mathematical formula node is consistent with the operation process verified by a mathematical software tool and the operation process of the corresponding node in the parse tree node structure.

[0115] The calculation result of the mathematical formula node is consistent with the calculation result of the corresponding node in the parse tree node result verified by the mathematical software tool.

[0116] Furthermore, the mathematical operation priority rules in mathematics, that is, brackets first, then exponents, then multiplication and division, and finally addition and subtraction, analyze mathematical formulas. " Take the following as an example, first look at the " ", then calculate the index " ", then perform the multiplication " ", and finally do addition" During the verification process, check whether the calculation steps of the mathematical formula are strictly performed in this order. If an error occurs such as calculating addition first and then the numbers in brackets, the calculation priority rule is not met.

[0117] For complex expressions containing multiple layers of brackets, such as " ", calculate from the innermost brackets outwards, ensuring that the calculations in each layer of brackets are accurate and follow the overall calculation priority.

[0118] S5. Analyze the score of neat answers based on the structure restoration rate, calculation process compliance rate and knowledge point association accuracy, and mark them. At the same time, mark the wrong structure, wrong calculation process and knowledge points not involved.

[0119] The scoring analysis formula for the neat answer content , To score the content of the neat answer, Set the total score for the text answer content. is the structural reduction rate, is the weight of the text structure restoration rate, is the knowledge point association accuracy, is the weight of the accuracy of the text knowledge point association, is the compliance rate of the operation process, The weight of the text calculation process compliance rate. The structural restoration rate ensures that the answers are organized and easy to understand; the knowledge point association accuracy rate ensures that the answers are correct and closely related to the theme; the calculation process compliance rate ensures the scientificity and standardization of the problem-solving process in questions involving calculations. These three parameters comprehensively evaluate the quality of the answers from different dimensions and can give a more objective and accurate score for the neat answers.

[0120] It should be noted that the weight of the text structure restoration rate measures the importance of the degree of conformity of the answer content with the standard structure. When answering questions, a reasonable and clear structure helps the examiner to quickly understand the answer ideas. For example, when answering math proof questions, answer the questions according to the structure of theorem citation, reasoning steps, and conclusion statements. A high structural restoration rate can reflect the orderliness of the answer.

[0121] The weight of the accuracy of the knowledge point association in the text: This weight reflects the importance of the accuracy of the association between the knowledge points involved in the answer and the knowledge points required by the question. Accurately associating knowledge points is the key to answering questions correctly. For example, in historical essay questions, the accuracy of knowledge point association is very important to accurately use relevant historical events and background knowledge to prove the point of view.

[0122] The weight of the text calculation process compliance rate: This weight indicates the importance of the degree of compliance of the calculation steps in the answering process with the standard calculation rules and steps. In mathematical subjects, the correct calculation process is the guarantee of obtaining the correct result. Even if the final answer is wrong, if the calculation process conforms to the standard, it can also illustrate the answerer's problem-solving ideas and ability. In a specific embodiment, , , .

[0123] It should be noted that the present invention combines the structural restoration rate, the calculation process compliance rate and the knowledge point association accuracy rate to analyze the scoring of neat answer content, which not only improves the comprehensiveness and scientificity of the scoring, improves the scoring efficiency, adapts to multi-disciplinary application scenarios, and promotes teaching improvement and personalized teaching, and has significant educational application value.

[0124] In a specific embodiment, the error structure, error calculation process and the identification process of uninvolved knowledge points are as follows: screen the text nodes whose sub-node matching rate is less than or equal to the set matching rate and the mathematical formula nodes with different sub-nodes, collectively refer to them as error structure nodes, and take the error structure nodes corresponding to the answer question content as the error structure.

[0125] If a node in the subject knowledge graph cannot be traced back to the same associated knowledge point of the corresponding node in the standard answer parse tree, then there is no reasonable knowledge association path between the node and the node in the standard answer parse tree, and the content of the answer to the question corresponding to the node will be regarded as an uninvolved knowledge point.

[0126] Filter out mathematical formula nodes that do not comply with the operation priority rules, operation processes or operation results, and treat them as incorrect operation nodes, and treat the corresponding answering content of the incorrect operation nodes as incorrect operation processes.

[0127] It should be noted that the present invention can accurately identify errors, provide targeted feedback, help students quickly locate weak links, and support deep-level error tracing and learning improvement by synchronously marking erroneous structural nodes, erroneous calculation processes and uninvolved knowledge points during the scoring process.

[0128] like Figure 3 As shown, a text sample recognition, analysis and annotation system includes a content positioning and distinguishing module, a neat answer inference module, a neat answer content parsing module, a content association analysis module and a content scoring and annotation module. The connection relationship between the modules is: the content positioning and distinguishing module is connected to the neat answer inference module, the neat answer inference module is connected to the neat answer content parsing module, the neat answer content parsing module is connected to the content association analysis module, and the content association analysis module is connected to the content scoring and annotation module.

[0129] The content location and differentiation module locates and differentiates the content of the text sample to obtain the answer content of the text sample.

[0130] The neat answer inference module combines the usual handwriting of the writer of the text sample to infer the neat answer content of the text sample.

[0131] The neat answer content parsing module uses the LaTex parse tree to parse the neat answer content of the text sample, obtains the parse tree node structure of the neat answer content, and determines whether the parse tree node structure is completely consistent with the standard answer.

[0132] The content association analysis module performs structural comparison and knowledge point association based on the parse tree node structure when it is incomplete, and analyzes the structural restoration rate of the parse tree node structure, the compliance rate of the calculation process and the accuracy rate of knowledge point association.

[0133] The content scoring and annotation module analyzes the score of neat answer content based on the restoration rate, calculation process compliance rate and knowledge point association accuracy, and annotates it, while also marking incorrect structures, incorrect calculation processes and knowledge points not involved.

[0134] Figure 4 is a block diagram of a computer storage medium for implementing some embodiments of the present invention. Figure 4 As shown, the storage medium 100 may be in the form of a general-purpose computing device. The storage medium 100 includes a memory 110, a processor 120, and a bus 130 that connects different system components.

[0135] The memory 110 may include, for example, a system memory, a non-volatile storage medium, etc. The system memory may store, for example, an operating system, an application program, a boot loader, and other programs. The system memory may include a volatile storage medium, such as a random access memory (RAM) and / or a cache memory. The non-volatile storage medium may store, for example, instructions for executing at least one of the corresponding embodiments of the method for improving the torsional stiffness of the vehicle body. The non-volatile storage medium includes, but is not limited to, a disk memory, an optical memory, a flash memory, etc.

[0136] The processor 120 may be implemented by a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistors, and other discrete hardware components. Accordingly, each module such as the manufacturing module, the defect detection module, the structure supplementation module, the topology test module, the torsional stiffness enhancement module, and the automation processing module may be implemented by a central processing unit (CPU) running instructions in a memory that execute corresponding steps, or by a dedicated circuit that executes corresponding steps.

[0137] The bus 130 may use any of a variety of bus architectures, including, but not limited to, an Industry Standard Architecture (ISA) bus, a Micro Channel Architecture (MCA) bus, and a Peripheral Component Interconnect (PCI) bus.

[0138] The storage medium 100 may also include an input / output interface 140, a network interface 150, a storage interface 160, etc. These interfaces 140, 150, 160 and the memory 110 and the processor 120 may be connected via a bus 130. The input / output interface 140 may provide a connection interface for input / output devices such as a display, a mouse, and a keyboard. The network interface 950 may provide a connection interface for various networked devices. The storage interface 160 may provide a connection interface for external storage devices such as a floppy disk, a USB flash drive, and an SD card.

[0139] Here, various aspects of the present invention are described with reference to the flowchart and / or block diagram of the method, system and storage medium according to the embodiment of the present invention. It should be understood that each frame of the flowchart and / or block diagram and the combination of each frame can be implemented by computer-readable program instructions.

[0140] These computer-readable program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer or other programmable device to produce a machine, so that the processor executes the instructions to produce a device that implements the functions specified in one or more blocks in the flowchart and / or block diagram.

[0141] These computer-readable program instructions may also be stored in a computer-readable memory, which cause the computer to work in a specific manner to produce an article of manufacture, including instructions for implementing the functions specified in one or more blocks in the flowchart and / or block diagram.

[0142] The present invention can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects.

[0143] The above contents are merely examples and explanations of the concept of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, they shall all fall within the protection scope of the present invention.

Claims

1. A text sample recognition, analysis and annotation method, characterized in that: The steps include: S1. Locate and distinguish the content of the text sample to obtain the answer content of the text sample; S2. Infer the neat answer content of the text sample based on the handwriting habits of the writer of the text sample; S3, using LaTex parse tree to parse the neat answer content of the text sample, obtain the parse tree node structure of the neat answer content, and determine whether the parse tree node structure is completely consistent with the standard answer; S4, when incomplete, perform structural comparison and knowledge point association based on the parse tree node structure, and analyze the structural restoration rate of the parse tree node structure, the calculation process compliance rate, and the knowledge point association accuracy rate; S5. Analyze the score of neat answers based on the structure restoration rate, calculation process compliance rate and knowledge point association accuracy, and mark them. At the same time, mark the wrong structure, wrong calculation process and knowledge points not involved.

2. A text sample recognition, analysis and annotation method according to claim 1, characterized in that: The specific contents of step S1 are as follows: S11, using OCR technology to perform text recognition on the text sample to obtain the recognized text content; S12, performing structural feature recognition on the text content, and locating the question stem area and its corresponding answer question area in the text content; S13, identifying the text block in the answer area and the nearby text block outside the answer area, and calculating the relative distance between the nearby text block and the answer area and the correlation between the nearby text block and the text block in the answer area; S14. Determine the answer content according to the relative distance and the correlation, and then obtain the answer content of the text sample.

3. The text sample recognition, analysis and annotation method according to claim 2, characterized in that: The specific contents of step S13 are as follows: S131, identifying the text within the boundary of the answer area as a text block within the answer area, and recording the text content of the text block; S132, setting a distance threshold based on the boundary of the answering area, determining that the text within the distance range threshold is a nearby text block outside the answering area, and recording its text content; S133, using coordinate comparison method to analyze the relative distance between the nearby text blocks and the answer area; S134. Analyze the correlation between the nearby text block and the text block in the answer area based on the text content of the text block and the nearby text blocks.

4. The text sample recognition, analysis and annotation method according to claim 1, characterized in that: The specific contents of step S2 are as follows: S21, collecting various texts written daily by the writers of the text samples to form a text set of the writers' daily writing; S22, using a neural network to perform multidimensional feature extraction on a daily handwriting text set to obtain a reference multidimensional feature vector of each character in the daily handwriting text set; S23, constructing a handwriting model of the writer based on the reference multidimensional feature vector of each character; S24, extracting multidimensional feature vectors of all characters from the answer content of the text sample, substituting the multidimensional feature vectors into the writer's handwriting model, and matching the standard characters corresponding to all characters; S25. Based on the standard characters corresponding to all characters, a neat answer content of the text sample is constructed.

5. The text sample recognition, analysis and annotation method according to claim 1, characterized in that: The specific contents of step S3 are as follows: S31, pre-processing the neat answer content of the text sample; S32, using LaTex parse tree to convert the pre-processed neat answer content into a structured tree, and obtaining a parse tree node structure of the neat answer content according to the structured tree; S33. Compare the node structure of the parse tree of the neat answer content with the standard answer to determine whether the node structure of the parse tree of the neat answer content is completely consistent with the standard answer.

6. The text sample recognition, analysis and annotation method according to claim 1, characterized in that: The specific implementation content of step S4 is as follows: S41, comparing the parse tree node structure of the neat answer content with the parse tree node structure of the standard answer, obtaining matching nodes of the parse tree node structure, where the matching nodes are the number of text node matches and the number of mathematical formula node matches, and analyzing the structural restoration rate of the parse tree node structure; S42, mapping all nodes in the parse tree node structure with associated knowledge points in the subject knowledge graph to which the text sample belongs, and analyzing whether there is a reasonable knowledge association path between the parse tree node and the standard answer parse tree node based on the mapping, and analyzing the knowledge point association accuracy of the parse tree node structure; S43. Analyze all mathematical formula nodes in the parse tree node structure using the operation priority rule, obtain the number of nodes with correct operation logic, and analyze the compliance rate of the operation process of the parse tree node structure.

7. A text sample recognition, analysis and annotation method according to claim 6, characterized in that: The analysis content of step S41 is as follows: S411, matching the neat answer content with all subnodes of the text node in the parse tree node structure of the standard answer, counting the number of subnode matches of all text nodes, and analyzing the subnode matching rate of the text node; S412, filtering text nodes whose sub-node matching rates are greater than a set matching rate, and counting the number of matching text nodes; S413, comparing the neat answer content with all child nodes of the mathematical formula node in the parse tree node structure of the standard answer, selecting mathematical formula nodes with the same child nodes, and counting the number of matching mathematical formula nodes; S414. Analyze the structural restoration rate of the parse tree node structure by the number of text node matches and the number of mathematical formula node matches .

8. The text sample recognition, analysis and annotation method according to claim 7, characterized in that: The scoring analysis formula for the neat answer content , To score the content of the neat answer, Set the total score for the text answer content. is the structural reduction rate, is the weight of the text structure restoration rate, is the knowledge point association accuracy, is the weight of the accuracy of the text knowledge point association, is the compliance rate of the operation process, It is the weight of the conformity rate in the text operation process.

9. A text sample recognition, analysis and annotation system, characterized in that: include: The content location and differentiation module locates and differentiates the content of the text sample to obtain the answer content of the text sample; The neat answer inference module infers the neat answer content of the text sample based on the idiomatic handwriting of the writer of the text sample; The neat answer content parsing module uses the LaTex parsing tree to parse the neat answer content of the text sample, obtains the parsing tree node structure of the neat answer content, and determines whether the parsing tree node structure is completely consistent with the standard answer; The content association analysis module compares the structure and associates the knowledge points based on the parse tree node structure when it is incomplete, and analyzes the structural restoration rate of the parse tree node structure, the calculation process compliance rate and the knowledge point association accuracy rate; The content scoring and annotation module analyzes the score of neat answer content based on the restoration rate, calculation process compliance rate and knowledge point association accuracy, and annotates it, while also marking incorrect structures, incorrect calculation processes and knowledge points not involved.

10. A storage medium, characterized in that: The storage medium is burned with a computer program, and when the computer program is executed in the memory of the server, the method for identifying, analyzing and annotating text samples described in any one of claims 1 to 8 is implemented.

Images