Writing evaluation model training method and device and writing evaluation method and device

Abstract

The embodiment of the invention provides a writing evaluation model training method and device and a writing evaluation method and device. The writing evaluation model training method comprises the steps that a writing sample image is input into a character unit recognition model and an altering unit recognition model; extracting character unit identification information in the writing sample image through the character unit identification model, and extracting altering unit identification information in the writing sample image through the altering unit identification model; constructing writing character unit characteristics according to the character unit identification information, and constructing writing altering unit characteristics according to the altering unit identification information; and training a writing evaluation model by using the writing character unit features and the writing altering unit features until a target writing evaluation model meeting a training stop condition is obtained.

Description

TECHNICAL FIELD

[0001] Embodiments of the present specification relate to the technical field of machine learning, and particularly relate to a writing evaluation model training method and device, and a writing evaluation method and device. BACKGROUND

[0002] With the development of computer and Internet technology, the marking of papers gradually tends to be automated. Through recognition technology, the answer results in the papers can be accurately recognized, and the marking can be automatically completed to save the time of teachers in marking papers. However, most of the automatic marking processes in the prior art only focus on the correctness of the answer results, resulting in that the marking results of composition questions lack the evaluation of writing quality. In English, Chinese and other subjects, the answer content of composition questions is relatively important, but the final score also needs to refer to the writing quality of students. In specific implementation, the evaluation of writing quality mostly still needs to rely on manual completion, and manual review inevitably has subjectivity and low efficiency. Therefore, an effective solution is urgently needed to solve the above problems. SUMMARY

[0003] Therefore, the embodiments of the present specification provide a writing evaluation model training method. One or more embodiments of the present specification also relate to a writing evaluation model training device, a writing evaluation method, a writing evaluation device, a computing device, a computer-readable storage medium, and a computer program product to solve the technical defects in the prior art.

[0004] According to a first aspect of the embodiments of the present specification, a writing evaluation model training method is provided, including: inputting a writing sample image into a character unit recognition model and a correction unit recognition model; extracting character unit recognition information in the writing sample image by the character unit recognition model, and extracting correction unit recognition information in the writing sample image by the correction unit recognition model; constructing writing character unit features according to the character unit recognition information, and constructing writing correction unit features according to the correction unit recognition information; training a writing evaluation model using the writing character unit features and the writing correction unit features until a target writing evaluation model satisfying a training stop condition is obtained.

[0005] According to a second aspect of the embodiments of the present specification, a writing evaluation method is provided, including: obtaining a target writing image uploaded by a client for a target text; predict a writing score of the target writing image by using a target writing evaluation model, and predict a text score of the target text by using a text evaluation model, the target writing evaluation model being determined by the method; calculate a global score corresponding to the target text based on the writing score and the text score, and feed back the global score to the client.

[0006] According to a third aspect of the embodiments of the present specification, a writing evaluation model training apparatus is provided, comprising: an input module configured to input a writing sample image to a character unit recognition model and a correction unit recognition model; an extraction module configured to extract character unit recognition information in the writing sample image by the character unit recognition model, and extract correction unit recognition information in the writing sample image by the correction unit recognition model; a construction module configured to construct writing character unit features according to the character unit recognition information, and construct writing correction unit features according to the correction unit recognition information; a training module configured to train a writing evaluation model by using the writing character unit features and the writing correction unit features until a target writing evaluation model meeting a training stop condition is obtained.

[0007] According to a fourth aspect of the embodiments of the present specification, a writing evaluation apparatus is provided, comprising: an acquisition module configured to acquire a target writing image uploaded by a client for a target text; a prediction module configured to predict a writing score of the target writing image by using a target writing evaluation model, and predict a text score of the target text by using a text evaluation model, the target writing evaluation model being determined by the method; a feedback module configured to calculate a global score corresponding to the target text based on the writing score and the text score, and feed back the global score to the client.

[0008] According to a fifth aspect of the embodiments of the present specification, a computing device is provided, comprising: a memory and a processor; the memory is configured to store computer executable instructions, and the processor is configured to execute the computer executable instructions, the computer executable instructions being executed by the processor to implement the steps of the above writing evaluation model training method or writing evaluation method.

[0009] According to a sixth aspect of the embodiments of the present specification, a computer-readable storage medium is provided, which stores computer executable instructions, and the instructions are executed by a processor to implement the steps of the writing evaluation model training method or the writing evaluation method.

[0010] According to a seventh aspect of the embodiments of the present specification, a computer program product is provided, which includes a computer program or instructions, and the computer program or instructions are executed by a processor to implement the steps of the writing evaluation model training method or the writing evaluation method.

[0011] The writing evaluation model training method provided by the embodiments can objectively evaluate the writing quality of composition type questions and improve the writing evaluation efficiency. The writing evaluation model for writing evaluation scoring can be trained. In the model training process, in order to enable the writing evaluation model to have strong prediction ability and consider multi-dimensional writing evaluation influence, the writing sample image can be first input to the character unit recognition model and the correction unit recognition model. The character unit recognition information is extracted from the writing sample image by the character unit recognition model, and the correction unit recognition information is extracted from the writing sample image by the correction unit recognition model. The writing quality information collection in the writing dimension and the correction dimension is realized. Subsequently, the writing character unit features can be constructed according to the character unit recognition information, and the writing correction unit features can be constructed according to the correction unit recognition information. On this basis, the writing character unit features and the writing correction unit features can be fused to train the writing evaluation model until the target writing evaluation model meeting the training stop condition is obtained. When the model for evaluating the writing quality is trained, the multi-dimensional information can be combined to construct the features for model training. When the model evaluates the writing quality, the multi-dimensional evaluation ability can be comprehensively considered for scoring. The objectivity of the writing quality evaluation can be ensured, and the evaluation efficiency can be improved. Therefore, the human resources are saved, and a faster processing mode is provided for online or offline marking. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 FIG. 1 is a flowchart of a writing evaluation model training method according to an embodiment of the present specification; Figure 2 FIG. 2 is a schematic diagram of model training in a writing evaluation model training method according to an embodiment of the present specification; Figure 3 FIG. 3 is a flowchart of a writing evaluation method according to an embodiment of the present specification; Figure 4 FIG. 4 is a process flowchart of a writing evaluation method according to an embodiment of the present specification; Figure 5 FIG. 5 is a structural schematic diagram of a writing evaluation model training device according to an embodiment of the present specification; Figure 6 is a structural schematic diagram of a writing evaluation device provided by an embodiment of the present specification; Figure 7 is a structural block diagram of a computing device provided by an embodiment of the present specification. DETAILED DESCRIPTION

[0013] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present specification. However, the present specification can be practiced without the specific details, other than in the examples described herein. Those of ordinary skill in the art, and others, can readily ascertain many ways of making and using the present specification without departing from the scope of the present specification, and the present specification is not limited by the following examples.

[0014] The terminology used in this description of one or more embodiments of the present specification is for the purpose of describing particular embodiments only and is not intended to be limiting of one or more embodiments of the present specification. As used in this description of one or more embodiments of the present specification and the accompanying claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will also be understood that the term “and / or” as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0015] It will be understood that, although the terms first, second, etc. can be used herein to describe various information, these terms are not intended to denote a temporal or chronological order. Rather, these terms are used solely to distinguish one from another only. For example, without departing from the scope of one or more embodiments of the present specification, first can be termed second, and similarly, second can be termed first. Depending on the context, the word “if’ as used herein can be interpreted to mean “when” or “in response to determining” or “in response to a determination.”

[0016] In addition, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in one or more embodiments of the present specification are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation portal for user to choose authorization or refusal.

[0017] First, the terms involved in one or more embodiments of the present specification are explained.

[0018] OCR (Optical Character Recognition) technology is a technology that converts printed or handwritten text in images or documents into machine-readable text by scanning them.

[0019] Bayesian classifier: uses Bayes' theorem to calculate the probability of each class, thus determining the class of the data item. Common methods include Naive Bayes.

[0020] Support Vector Machine (SVM): based on the principle of maximizing the interval between samples, different classes are separated by constructing a hyperplane. SVM is suitable for high-dimensional space classification problems.

[0021] Gradient Boosting Tree (GBT): a boosting technique that improves classification performance by constructing multiple decision tree models and combining their results. GBT performs well in solving complex classification problems and can handle non-linear data relationships.

[0022] In the present specification, a writing evaluation model training method is provided. One or more embodiments of the present specification also relate to a writing evaluation model training device, a writing evaluation method, a writing evaluation device, a computing device, a computer-readable storage medium, and a computer program product, which are described in detail in the following embodiments.

[0023] In practical applications, the correction of composition-type questions not only needs to focus on the content, but also needs to focus on the evaluation of writing quality. In the prior art, there are mainly two kinds of technologies for evaluating the writing quality of compositions: rule-based scoring method and artificial scoring method. The rule-based scoring method generally scores the writing quality of compositions directly through pre-set rules. For example, this method can score according to the number of words in the composition, the consistency of line spacing, and other simple indicators. The advantage of this scoring method is that it is easy to implement and has high efficiency. However, the rule-based scoring method also has obvious defects: limited coverage: the pre-set rules can only cover part of the scene, such as simple indicators like word count and line spacing, and it is difficult to comprehensively evaluate other important aspects of writing quality, such as handwriting aesthetics and stroke clarity. Inaccurate: due to the rough scoring rules, it is difficult to carefully identify and evaluate the details of writing quality, and the scoring results are often not accurate enough, and there is a large difference from the subjective scoring standard.

[0024] The scoring method based on artificial scoring is to score the composition by skilled reviewers one by one, and give the score according to the subjective scoring standard. The advantages of this method are: accurate: skilled reviewers can accurately evaluate the writing quality of each composition according to experience and standards, and the scoring results have high credibility. Flexible: artificial scoring can consider various complex situations and details, and give reasonable judgment to different writing styles. However, the scoring method based on artificial scoring also has obvious limitations: time-consuming and labor-intensive: artificial review requires a lot of time and manpower, especially when facing a large number of compositions, the review work becomes time-consuming and labor-intensive, and the efficiency is low. It is difficult to realize large-scale application: due to the limitations of artificial review, it is difficult to realize large-scale application, and the automation and popularization of paper volume scoring face great challenges. Therefore, an effective solution is needed to solve the above problems.

[0025] The writing evaluation model training method provided by the embodiment can objectively evaluate the writing quality of composition type questions and improve the writing evaluation efficiency. The writing evaluation model for writing evaluation scoring can be trained. In the model training process, in order to enable the writing evaluation model to have strong prediction ability and consider multi-dimensional writing evaluation influence, the writing sample image can be first input to the character unit recognition model and the correction unit recognition model. The character unit recognition model extracts character unit recognition information in the writing sample image, and the correction unit recognition model extracts correction unit recognition information in the writing sample image. The writing quality information collection in the writing dimension and the correction dimension is realized. Subsequently, the writing character unit features can be constructed according to the character unit recognition information, and the writing correction unit features can be constructed according to the correction unit recognition information. On this basis, the writing character unit features and the writing correction unit features can be fused to train the writing evaluation model until a target writing evaluation model meeting the training stop condition is obtained. When training the model for evaluating the writing quality, the multi-dimensional information can be combined to construct the features for model training, so that the model can score by comprehensively considering the multi-dimensional evaluation ability when evaluating the writing quality. Not only the objectivity of writing quality evaluation can be ensured, but also the evaluation efficiency can be improved, so as to save human resources and provide a faster processing mode for online or offline review and correction.

[0026] Referring to Figure 1 , Figure 1 A flowchart of a writing evaluation model training method according to one embodiment of the present specification is shown, which specifically includes the following steps.

[0027] Step S102, inputting the writing sample image to the character unit recognition model and the correction unit recognition model.

[0028] The writing evaluation model training method provided by the embodiment can be applied to any scene of evaluating writing quality, such as a Chinese composition correction scene, an English composition correction scene, a calligraphy correction scene, a character practice correction scene, and the like, to evaluate the writing quality in a user-uploaded image, thereby saving human resources while improving correction efficiency.

[0029] The embodiment takes the application of the writing evaluation model training method in the Chinese composition correction scene as an example for description, and the description of writing quality evaluation in other scenes can be referred to the same or corresponding description in the embodiment, which will not be described in detail herein.

[0030] Specifically, the writing sample image specifically refers to a sample image containing writing content collected when training the writing evaluation model, and the writing content contained in the image is written by a user, so that the model has strong writing evaluation capability. The character unit recognition model specifically refers to a model for detecting and recognizing the writing character unit contained in the writing sample image, the input of which is the writing sample image, and the output is character unit recognition information. The character unit recognition model can be obtained by training the OCR recognition model separately, so that it has character unit detection and recognition capabilities. Correspondingly, the correction unit recognition model specifically refers to a model for recognizing the correction content contained in the writing sample image, the input of which is the writing sample image, and the output is correction unit recognition information.

[0031] Therefore, in order to objectively evaluate the writing quality of the composition type question and improve the writing evaluation efficiency, a writing evaluation model for writing evaluation scoring can be trained. In the model training process, in order to enable the writing evaluation model to have strong prediction capability and consider multi-dimensional writing evaluation influence, the writing sample image can be first input to the character unit recognition model and the correction unit recognition model. The character unit recognition information is extracted from the writing sample image by the character unit recognition model, and the correction unit recognition information is extracted from the writing sample image by the correction unit recognition model. The writing quality information collection is realized in the writing dimension and the correction dimension. Subsequently, the writing character unit features can be constructed according to the character unit recognition information, and the writing correction unit features can be constructed according to the correction unit recognition information. On this basis, the writing evaluation model can be trained by fusing the writing character unit features and the writing correction unit features until the target writing evaluation model meeting the training stop condition is obtained. When training the model for evaluating the writing quality, the features can be constructed by combining multi-dimensional information for model training, so that the model can score by comprehensively considering multi-dimensional evaluation capability when evaluating the writing quality. Not only the objectivity of writing quality evaluation can be ensured, but also the evaluation efficiency can be improved, thereby saving human resources while providing a faster processing mode for online or offline review and correction.

[0032] Step S104, extracting word unit recognition information in the writing sample image by the word unit recognition model, and extracting erasing unit recognition information in the writing sample image by the erasing unit recognition model.

[0033] Specifically, after obtaining the writing sample image and inputting it into the word unit recognition model and the erasing unit recognition model, in order to facilitate subsequent training of the writing evaluation model combined with multi-dimensional features, the word unit recognition information in the writing sample image can be extracted by the word unit recognition model, and the erasing unit recognition information in the writing sample image can be extracted by the erasing unit recognition model, so as to represent the writing information corresponding to the writing word unit in the writing sample image by the word unit recognition information, and represent the erasing information corresponding to the erasing unit in the writing sample image by the erasing unit recognition information.

[0034] Specifically, the word unit recognition information refers to the information extracted after the word unit recognition model recognizes the word unit contained in the writing sample image, including but not limited to the content of the word unit, the text box corresponding to the word unit, the height of the word unit, etc. The erasing unit recognition information refers to the information extracted after the erasing unit recognition model recognizes the erasing unit contained in the writing sample image, including but not limited to the number of erasing units, the frequency, etc.

[0035] Correspondingly, the writing word unit specifically refers to each word unit written by the user in the writing sample image, such as Chinese characters in a Chinese composition and letters in an English composition. Correspondingly, the erasing unit specifically refers to the mark of the user's erasing in the writing sample image, such as a deletion line, a black dot, an insertion symbol, etc., which is used to represent the erasing mark of the text.

[0036] Further, when predicting the word unit recognition information by the word unit recognition model, the text detection unit and the text recognition unit can be used to achieve this. In this embodiment, the word unit recognition model includes a text detection unit and a text recognition unit, and the word unit recognition information in the writing sample image is extracted by the word unit recognition model, including: The text detection unit adds a text box to the word unit contained in the writing sample image to obtain a to-be-recognized image containing a text box; the text recognition unit recognizes the word unit corresponding to the text box of the to-be-recognized image to obtain word unit recognition information; wherein the word unit recognition information includes word unit recognition results and text box annotation results.

[0037] Specifically, the text box specifically refers to the text box added to the character unit contained in the writing sample image, each text box is used to frame a character unit, the image to be recognized is the writing sample image containing the text box, which is used for subsequent recognition of the character unit in the text box. Correspondingly, the character unit recognition result specifically refers to the result obtained after recognizing the character unit in the writing sample image, and the text box labeling result specifically refers to the result obtained after labeling the text box for the character unit in the writing sample image.

[0038] Based on this, in order to improve the recognition accuracy of the character unit recognition information, the character unit detection and recognition processing can be completed by cooperating the text detection unit and the text recognition unit. In this process, the text detection unit can add a text box to the character unit contained in the writing sample image to obtain an image to be recognized containing the text box. The text box is used to frame each character unit. Then, the text recognition unit can recognize the character unit corresponding to the text box in the image to be recognized, and further obtain the character unit recognition information. The character unit recognition information includes the character unit recognition result and the text box labeling result.

[0039] In summary, by accurately positioning the character unit in the writing sample and adding a text box through the text detection unit, and by efficiently extracting the character unit information through the text recognition unit, the synchronous output of the text box labeling and the recognition result is realized. This method optimizes the traditional OCR process, improves the collaborative efficiency of text positioning and recognition, reduces false positives and false negatives, supports structured processing of complex formats, enhances adaptability to multiple fonts and multiple scenarios, and significantly improves the accuracy and automation level of character recognition.

[0040] On this basis, in order to enable the text detection unit to have strong detection capability and be used in writing evaluation scenarios, the text detection unit can be trained in the following manner. In the present embodiment, the training of the text detection unit comprises: obtaining a sample image, inputting the sample image into an initial text detection unit; processing the sample image through a first detection unit and a second detection unit in the initial text detection unit to obtain a prediction probability map and a prediction threshold map; constructing a prediction binary map according to the prediction probability map and the prediction threshold map; determining a label probability map and a label threshold map corresponding to the sample image, and taking the label probability map as a label binary map corresponding to the prediction binary map; based on the prediction probability map, the prediction threshold map, the prediction binary map, the label probability map, the label threshold map and the label binary map, optimizing the initial text detection unit until the text detection unit satisfying the optimization stop condition is obtained.

[0041] Specifically, the sample image specifically refers to an image containing a written character unit, the initial text detection unit specifically refers to an untrained text detection unit, which can be implemented by DBNet (Differentiable Binarization). The first detection unit and the second detection unit refer to two heads in the initial text detection unit, the predicted probability map specifically refers to a probability map output by the text detection model, the predicted threshold map specifically refers to a threshold map output by the text detection unit, the predicted binary map specifically refers to a binary map generated by combining the predicted threshold map and the predicted probability map, and the label binary map, the label probability map and the label threshold map are labels used in the model training stage. The optimization stop condition specifically refers to the condition for stopping training the text detection unit, including but not limited to the loss value comparison condition, the iteration number condition, the validation set verification condition, which is not limited in this embodiment.

[0042] Based on this, when training the text detection unit, the sample image can be obtained first, at this time the sample image can be input to the initial text detection unit. Then the first detection unit and the second detection unit in the initial text detection unit can be used to process the sample image to obtain the predicted probability map and the predicted threshold map according to the processing result. On this basis, in order to enable the text detection unit to have strong character unit detection capability, the predicted binary map can be constructed according to the predicted probability map and the predicted threshold map. After that, the label probability map and the label threshold map corresponding to the sample image can be determined, and the label probability map is taken as the label binary map corresponding to the predicted binary map. Then the initial text detection unit can be optimized based on the predicted probability map, the predicted threshold map, the predicted binary map, the label probability map, the label threshold map and the label binary map, and in the optimization process, new sample images can be continuously obtained for optimization until the text detection unit satisfying the optimization stop condition is obtained. It can be deployed to the writing quality evaluation scene for use, which can effectively improve the character unit detection capability.

[0043] In practical applications, when selecting writing sample images, you can choose written texts such as Chinese and English compositions. Chinese compositions can also include gridded paper, and English compositions can include lined paper, thereby more accurately simulating real-world situations. Furthermore, the text detection unit in the character unit recognition model can be implemented using the DBNet model. The DBNet backbone is a FPN (feature pyramid) structure, with two heads that can output a probability map and a threshold map, respectively. The image size can be upsampled from 1 / 4 (H, W) back to the original image size. The two maps are then processed through DB operations to obtain the final binary map, which can be used as the basis for model training.

[0044] During model training, the labels corresponding to sample images can be divided into two parts: the probability map and the threshold map. The binary map and probability map can share the same groundtruth. The probability map does not simply set the annotated text domain to 1, but rather sets the shrink of the text domain to 1, which is equivalent to a small kernel. This effectively prevents multiple lines of text from being detected as a single line. Regarding the threshold map: When generating the probability map annotation, the "shrink" of the text domain is called Gs, and the "dilated" text domain is generated as Gd. The gap between Gs and Gd is called the boundary domain of the text domain. The threshold map is determined by the distance from the points in the boundary domain to the true boundary line G. Model training based on this allows for precise selection of text based on the annotation, eliminating background artifacts.

[0045] In addition, the text recognition model can be implemented using the CRNN model, in which CNN, RNN and CTC are used for feature extraction, sequence modeling and decoding prediction respectively to ensure accurate recognition of word units in the text box. At the same time, the confidence level corresponding to the word unit can be obtained for subsequent training of the writing evaluation model.

[0046] In summary, by constructing a joint optimization mechanism for the prediction probability map and the label binary map, we achieve dynamic and adaptive threshold adjustment of the text detection unit, significantly improving the accuracy of text region segmentation. Based on the probability-threshold dual-map collaborative training strategy, we effectively reduce post-processing complexity and enhance the model's robustness to blurred, tilted, and multi-scale text. This end-to-end optimization framework improves detection performance and increases text localization accuracy, making it suitable for efficient text detection in documents with complex layouts.

[0047] Further, the erasing unit recognition model extracts erasing unit recognition information in the writing sample image, including: The erasing unit recognition model identifies the erasing unit in the writing sample image to obtain erasing unit recognition information; wherein the erasing unit is determined according to at least one of the deletion sub-unit, the symbol sub-unit, and the correction sub-unit.

[0048] Specifically, the deletion sub-unit specifically refers to the marks of the erasing line, erasing dot, or double oblique line in the writing sample image that remove the word units, the symbol sub-unit specifically refers to the added symbols in the writing sample image, such as the insertion symbol, the word order changing symbol, the reference symbol, and the correction sub-unit specifically refers to the marks of the word units in the writing sample image that are modified, such as the long word modified to the short word.

[0049] Based on this, the erasing unit recognition model can identify the erasing unit in the writing sample image, and then obtain the erasing unit recognition information according to the identification result; and the erasing unit can be determined according to at least one of the deletion sub-unit, the symbol sub-unit, and the correction sub-unit, which is used in subsequent writing quality evaluation.

[0050] In specific implementation, since the separate erasing recognition is important for subsequent judgment of writing evaluation results, and the conventional OCR recognition model ignores the erasing information, it is necessary to separately train the erasing recognition model, which can be trained by yolov10 or pre-trained by a large language model to have the ability to recognize common writing erasing marks such as erasing lines, frames, insertion symbols, and replacement symbols.

[0051] For example, when training a model for evaluating the writing quality of Chinese composition, a large number of writing sample images corresponding to the Chinese composition can be obtained, which can be input into the word unit recognition model. The text detection unit in the word unit recognition model detects each Chinese character in the writing sample image to obtain a to-be-recognized image for framing each Chinese character, and then inputs the to-be-recognized image into the text recognition unit in the word unit recognition model to obtain the word unit in each text frame. At the same time, the writing sample image can be input into the erasing unit recognition model to obtain marks such as erasing lines, frames, replacement symbols, and reference symbols in the composition for subsequent training of the writing evaluation model.

[0052] In summary, through the multi-granularity classification recognition mechanism, the erasing behaviors such as deletion, symbol, and correction are accurately positioned and distinguished, and the semantic analysis ability of erasing unit detection is improved.

[0053] In step S106, a writing character unit feature is constructed according to the character unit recognition information, and a writing modification unit feature is constructed according to the modification unit recognition information.

[0054] Specifically, the writing character unit feature specifically refers to a vector expression of the character unit in the writing dimension, which can be generated in combination with multiple dimensions of writing quality, such as the character shape of the character unit, the font size of the overall text, the line spacing between the overall text character units, and the like, for ensuring that the writing character unit feature covers more writing information. The writing modification unit feature specifically refers to a vector expression of the modification unit in the writing temperature, which can be constructed in combination with information having modification marks in the writing sample image, such as the modification area, the modification times, and the like.

[0055] Based on this, in order to cover more extensive writing information, the writing character unit feature can be constructed according to the character unit recognition information obtained above, and the writing modification unit feature can be constructed according to the modification unit recognition information, so that the features containing more abundant writing information can be used for subsequent model training, improving the prediction accuracy and robustness of the model.

[0056] Further, in order to ensure that the writing character unit feature and the writing modification unit feature can cover more information in the writing dimension, the writing character unit feature and the writing modification unit feature can be constructed in combination with multi-dimensional features. In this embodiment, the writing character unit feature constructed according to the character unit recognition information includes: character unit recognition confidence features, character unit area features, character unit height features, and character unit frequency features are constructed according to the character unit recognition information; and the writing character unit feature is constructed based on the character unit recognition confidence features, the character unit area features, the character unit height features, and the character unit frequency features. The writing modification unit feature is constructed according to the modification unit recognition information, including: modification unit times features and modification unit frequency features are constructed according to the modification unit recognition information; and the writing modification unit feature is constructed based on the modification unit times features and the modification unit frequency features.

[0057] On this basis, the character unit recognition confidence features are constructed according to the recognition rate determined based on the character unit recognition information, for representing the writing standardization degree of the character unit; the character unit area features are constructed according to the mean and variance of the character frame area determined based on the character unit recognition information, for representing the writing uniformity of the character shape of the character unit; the character unit height features are constructed according to the mean and variance of the character frame height determined based on the character unit recognition information, for representing the writing uniformity of the line spacing of the character unit; and the character unit frequency features are constructed according to the character unit frequency determined based on the character unit recognition information, for representing the global uniformity of the character unit.

[0058] Based on this, in order to be able to guarantee subsequent use of multi-dimensional features for model training, when constructing the writing character unit feature and the writing erasing unit feature, features can be selected from multiple dimensions for fusion. The character unit recognition confidence feature, the character unit area feature, the character unit height feature, and the character unit frequency feature can be constructed according to the character unit recognition information; thereby realizing construction of the writing character unit feature based on the character unit recognition confidence feature, the character unit area feature, the character unit height feature, and the character unit frequency feature. Meanwhile, the erasing unit frequency feature and the erasing unit frequency feature can be constructed according to the erasing unit recognition information; thereby realizing construction of the writing erasing unit feature based on the erasing unit frequency feature and the erasing unit frequency feature.

[0059] In actual application, the character unit confidence can be used to represent the writing standard degree of the character unit, and the higher the score is, the more standardized the writing is. The character unit area can represent the uniformity of the character shape of the whole text, and the smaller the variance is, the more consistent the font size is. The character unit height can represent the line spacing of the character unit, and the smaller the variance is, the more similar the line spacing is. In specific implementation, assuming that the confidence of the character unit is c n , the mean μ c of the confidence of the character unit can be calculated by the following formula (1): (1) And the variance σ c 2 of the confidence can represent the distribution of the confidence, and can be calculated by the following formula (2): (2) Further, the variance and the mean of the confidence of the text box can also be calculated by the above formula (1) and (2).

[0060] In addition, the mean and the variance of the character area can be obtained by the following formula (3) and (4): (3) (4) Wherein, A n represents the character unit area, μ A represents the mean, and σ A 2 represents the variance.

[0061] In addition, in addition to the above features, the variance and the mean of the text box can also be calculated according to the character unit with the highest frequency of occurrence, so as to reflect whether the user has stable writing habits when writing a single character unit, and thus reflect the overall uniformity.

[0062] With the above example, after obtaining the word unit recognition information and the correction unit recognition information corresponding to the writing sample image, the word unit recognition confidence feature, the word unit area feature, the word unit height feature, and the word unit frequency feature can be constructed based on the word unit recognition information. Meanwhile, the correction unit frequency feature and the correction unit frequency feature can be constructed according to the correction unit recognition information. Thereafter, by merging the word unit recognition confidence feature, the word unit area feature, the word unit height feature, the word unit frequency feature, the correction unit frequency feature, and the correction unit frequency feature, the fusion feature expressing the writing attributes comprehensively can be obtained, and the model training can be performed in this way, so that the model has strong prediction ability.

[0063] In summary, by constructing a writing quality quantitative evaluation system through multi-dimensional feature fusion, joint modeling of character shape specification (confidence), character shape alignment (area / height mean variance), global consistency (frequency distribution), and correction behavior characteristics (number / frequency) is realized. This method breaks through the limitations of traditional single index evaluation, improves the dimension coverage and feature relevance of writing quality analysis, and enhances the recognition robustness of illegible handwriting and abnormal correction behaviors.

[0064] In step S108, the writing evaluation model is trained using the writing word unit features and the writing correction unit features until a target writing evaluation model satisfying a training stop condition is obtained.

[0065] Specifically, after obtaining the writing word unit features and the writing correction unit features, the writing evaluation model can be trained using the writing word unit features and the writing correction unit features until a target writing evaluation model satisfying a training stop condition is obtained, so that the trained writing evaluation model can have writing evaluation capability, i.e., the input of the model is the writing word unit features and the writing correction unit features corresponding to the writing image, and the output is a writing score. The training stop condition specifically refers to the condition for stopping training the writing evaluation model, including but not limited to a loss value comparison condition, an iteration number condition, or a validation set validation condition, which is not limited in this embodiment.

[0066] Further, in order to improve the evaluation accuracy of the model, the validation set can be used to verify the model during the training phase. In this embodiment, the writing evaluation model is trained using the writing word unit features and the writing correction unit features until a target writing evaluation model satisfying a training stop condition is obtained, which includes: merge the writing character unit features and the writing correction unit features to obtain writing merged features, and input the writing merged features into a writing evaluation model for processing to obtain a predicted writing score; train the writing evaluation model based on the predicted writing score and a target writing score corresponding to the writing sample image; verify whether the trained writing evaluation model meets a training stop condition using a verification set; if not, obtain a candidate writing sample image, use the candidate writing sample image as the writing sample image, and perform the step of inputting the writing sample image into the character unit recognition model and the correction unit recognition model; and if yes, use the trained writing evaluation model as a target writing evaluation model.

[0067] Specifically, the writing merged features specifically refer to vector expressions obtained after merging the writing character unit features and the writing correction unit features. The predicted writing score specifically refers to a predicted score output by the model, the target writing score specifically refers to a sample label, the verification set specifically refers to a set for verifying the model, and the candidate writing sample image specifically refers to a new writing sample image selected from a sample set.

[0068] Based on this, after obtaining the writing character unit features and the writing correction unit features, the writing character unit features and the writing correction unit features can be merged to obtain writing merged features according to the feature merging result. Thereafter, the writing merged features can be input into a writing evaluation model for processing to output a predicted writing score through the writing evaluation model. Then, the writing evaluation model can be trained based on the predicted writing score and a target writing score corresponding to the writing sample image. In order to ensure the model training accuracy, the trained writing evaluation model can be verified using a verification set to determine whether the trained writing evaluation model meets a training stop condition. If not, a candidate writing sample image can be obtained, the candidate writing sample image can be used as the writing sample image, and the step of inputting the writing sample image into the character unit recognition model and the correction unit recognition model can be performed. This process is iterated until a certain training period determines that the model meets the training stop condition. At this time, the writing evaluation model trained in this training stage can be used as a target writing evaluation model.

[0069] In summary, a writing quality evaluation system is constructed by multi-source feature fusion, combining character shape specification and correction behavior features to significantly improve the representation ability of the writing scoring model. Based on a dynamic iterative training mechanism, the model generalization performance is optimized to improve the evaluation accuracy and the abnormal writing recognition rate, effectively supporting intelligent quality control in education evaluation, judicial documents and other scenarios, reducing the cost of manual review, and realizing the standardization and automation of writing quality evaluation.

[0070] In addition, in order to improve the robustness of the model, and avoid overfitting, a sample set can be constructed first, and then the sample set is filtered and used for model training. In the embodiment, after the step of merging the writing character unit features and the writing correction unit features to obtain writing merged features is performed, the following steps are further included: candidate sample pairs based on the writing merged features and the target writing scores, and stores the candidate sample pairs into a sample set; pre-processes the sample pairs contained in the sample set, determines a target sample set according to a pre-processing result; cleanses the target sample set according to a preset set cleansing strategy, and trains the writing evaluation model using the cleansed target sample set until a target writing evaluation model satisfying a training stop condition is obtained.

[0071] Specifically, the candidate sample pair specifically refers to a sample pair composed of writing merged features and target writing scores, the sample set is a set of stored sample pairs, the pre-processing specifically refers to filtering processing of the sample pairs contained in the sample set, the target sample set is a filtered sample set, and the set cleansing rule specifically refers to a rule for cleansing the sample pairs contained in the set, for ensuring that the sample pairs contained in the final obtained sample set have higher quality.

[0072] Therefore, in order to sufficiently train the model, improve the robustness of the model, and avoid overfitting of the model, the model can be trained by constructing a sample set. In this process, candidate sample pairs can be constructed based on writing merged features and target writing scores, and the candidate sample pairs can be stored in a sample set. After obtaining the sample set, in order to improve the quality of the samples contained in the sample set, the sample pairs contained in the sample set can be pre-processed, and a target sample set can be determined according to a pre-processing result. Then, the target sample set can be cleansed according to a preset set cleansing strategy, so as to ensure that the samples contained in the cleansed sample set are high-quality sample pairs. Therefore, the writing evaluation model can be trained using the cleansed target sample set until a target writing evaluation model satisfying a training stop condition is obtained.

[0073] In specific implementation, referring to the schematic diagram shown in Figure 2 When training the writing model, it can be divided into five steps of data preprocessing, feature selection, data set splitting, model training, and model evaluation. Among them, data preprocessing includes data cleansing and standardization. In this process, the strategy of discarding can be used to ignore abnormal data, that is, when the merged features are NA, it means that there is a problem with this data, which can be discarded so as to not participate in subsequent model training. Standardization is to convert the data into a distribution with a mean of 0 and a standard deviation of 1, so that the model can converge reasonably, and will not be affected by the large feature value domain to affect the model training.

[0074] The standardization process can be calculated by the following formula (5): (5) Where μ is the mean of the feature, and σ is the standard deviation of the feature.

[0075] Further, feature selection refers to selecting the most influential features for model performance from the sample set, thereby improving the accuracy of the model. This process can be implemented using principal component analysis (PCA), which is a dimensionality reduction technique that can transform the original feature space to a new feature space through linear transformation. The new features are linear combinations of the original features, and the new features are mutually independent. PCA mainly realizes this transformation by solving the eigenvalues and eigenvectors of the covariance matrix of the original data.

[0076] Further, data set splitting is mainly to split the sample set and the validation set. This process can use cross-validation to evaluate model performance, so the data set can be split into k subsets and trained and evaluated k times, thereby reducing the dependence of the model on the data set division and improving the reliability of the model evaluation effect. In specific implementation, k-fold cross-validation can be used to achieve this, to reduce model overfitting.

[0077] On this basis, the writing evaluation model can use the gradient boosting decision tree model. This model can be based on the idea of ensemble learning, by continuously stacking weak learners (decision trees), and through gradient descent method to continuously optimize the model, to improve the prediction performance of the model. Each tree is trained based on the residual of the previous tree to minimize the overall error.

[0078] The writing evaluation model training method provided by the embodiment can objectively evaluate the writing quality of composition type questions and improve the writing evaluation efficiency. The writing evaluation model for writing evaluation scoring can be trained. In the model training process, in order to enable the writing evaluation model to have strong prediction ability and consider multi-dimensional writing evaluation influence, the writing sample image can be first input to a character unit recognition model and a correction unit recognition model. The character unit recognition information is extracted from the writing sample image by the character unit recognition model, and the correction unit recognition information is extracted from the writing sample image by the correction unit recognition model. The writing quality information collection in the writing dimension and the correction dimension is realized. Subsequently, the writing character unit features can be constructed according to the character unit recognition information, and the writing correction unit features can be constructed according to the correction unit recognition information. On this basis, the writing evaluation model can be trained by fusing the writing character unit features and the writing correction unit features until the target writing evaluation model meeting the training stop condition is obtained. When the model for evaluating the writing quality is trained, the multi-dimensional information can be combined to construct the features for model training. When the writing quality is evaluated by the model, the multi-dimensional evaluation ability can be comprehensively used for scoring. The objectivity of the writing quality evaluation can be ensured, and the evaluation efficiency can be improved, thereby saving human resources and providing a faster processing mode for online or offline marking.

[0079] Referring to Figure 3 , Figure 3 A flowchart of a writing evaluation method according to one embodiment of the present specification is shown, which specifically includes the following steps.

[0080] In step S302, a target writing image uploaded by a client for a target text is acquired.

[0081] In step S304, a writing score of the target writing image is predicted by using a target writing evaluation model, and a text score of the target text is predicted by using a text evaluation model. The target writing evaluation model is determined by the above method.

[0082] In step S306, a global score corresponding to the target text is calculated based on the writing score and the text score, and the global score is fed back to the client.

[0083] The writing evaluation method provided by the embodiment, wherein the contents not described in detail can be referred to the same or corresponding description in the above embodiments. The embodiment will not be described in detail here.

[0084] ​The client specifically refers to a user terminal held by a user with a text correction demand, including but not limited to a mobile phone, a computer, a tablet, a smart wearable device, and the like. Correspondingly, the target text is a written text photographed by the user, such as a Chinese composition, an English composition, and the like, and the target writing image is a photographed image corresponding to the target text. The text evaluation model specifically refers to a model for evaluating the content of the target text, which can be implemented by using a large language model and is fine-tuned to have a text evaluation capability. Correspondingly, the text score specifically refers to a score corresponding to the content of the text, and the writing score specifically refers to a score corresponding to the writing quality. The global score is a score corresponding to the target text obtained by adding or weightedly averaging the writing score and the text score. In general, the global score will not exceed a preset score threshold.

[0085] Based on this, after obtaining the target writing image uploaded by the client for the target text, in order to score the target text by comprehensively considering the text content and the writing quality, the writing score of the target writing image can be predicted by using the target writing evaluation model, and the text score of the target text can be predicted by using the text evaluation model. Thereafter, the global score corresponding to the target text can be calculated based on the writing score and the text score, and the global score can be fed back to the client.

[0086] The following describes the writing evaluation method in conjunction with the accompanying drawings Figure 4 Taking the application of the writing evaluation method provided in the specification in the English composition evaluation scene as an example, the writing evaluation method is further described. Among them, Figure 4 FIG. 1 shows a processing process flowchart of a writing evaluation method provided in an embodiment of the specification, specifically including the following steps.

[0087] Step S402, inputting the writing sample image into a character unit recognition model and a correction unit recognition model, the character unit recognition model including a text detection unit and a text recognition unit.

[0088] Step S404, adding a text box to the character unit contained in the writing sample image by the text detection unit to obtain a to-be-recognized image containing the text box.

[0089] Step S406, recognizing the character unit corresponding to the text box of the to-be-recognized image by the text recognition unit to obtain character unit recognition information.

[0090] Step S408, recognizing the correction unit in the writing sample image by the correction unit recognition model to obtain correction unit recognition information.

[0091] Step S410, constructing character unit recognition confidence features, character unit area features, character unit height features, and character unit frequency features according to the character unit recognition information.

[0092] Step S412, based on the word unit recognition confidence feature, the word unit area feature, the word unit height feature and the word unit frequency feature, a writing word unit feature is constructed.

[0093] The word unit recognition confidence feature is constructed according to the recognition rate determined by the word unit recognition information, and is used to represent the writing standardization of the word unit; the word unit area feature is constructed according to the mean and variance of the word frame area determined by the word unit recognition information, and is used to represent the writing uniformity of the character shape of the word unit; the word unit height feature is constructed according to the mean and variance of the word frame height determined by the word unit recognition information, and is used to represent the writing uniformity of the line spacing of the word unit; and the word unit frequency feature is constructed according to the word unit frequency determined by the word unit recognition information, and is used to represent the global uniformity of the word unit.

[0094] Step S414, a correction unit frequency feature and a correction unit frequency feature are constructed according to the correction unit recognition information, and based on the correction unit frequency feature and the correction unit frequency feature, a writing correction unit feature is constructed.

[0095] Step S416, the writing word unit feature and the writing correction unit feature are merged to obtain a writing merged feature, and the writing merged feature is input into the writing evaluation model for processing to obtain a predicted writing score.

[0096] Step S418, based on the predicted writing score and the target writing score corresponding to the writing sample image, the writing evaluation model is trained until a target writing evaluation model satisfying a training stop condition is obtained.

[0097] Step S420, a target writing image uploaded by the client for the target text is obtained.

[0098] Step S422, the writing score of the target writing image is predicted by using the target writing evaluation model, and the text score of the target text is predicted by using the text evaluation model.

[0099] Step S424, based on the writing score and the text score, a global score corresponding to the target text is calculated, and the global score is fed back to the client.

[0100] In summary, in order to objectively evaluate the writing quality of composition type questions and improve the efficiency of writing evaluation, a writing evaluation model for writing evaluation scoring can be trained. In the model training process, in order to enable the writing evaluation model to have strong prediction ability and consider the influence of multi-dimensional writing evaluation, the writing sample image can be first input to a character unit recognition model and a correction unit recognition model. The character unit recognition information is extracted from the writing sample image by the character unit recognition model, and the correction unit recognition information is extracted from the writing sample image by the correction unit recognition model. The writing quality information collection is realized in the writing dimension and the correction dimension. Subsequently, the writing character unit features can be constructed according to the character unit recognition information, and the writing correction unit features can be constructed according to the correction unit recognition information. On this basis, the writing evaluation model can be trained by fusing the writing character unit features and the writing correction unit features until the target writing evaluation model meeting the training stop condition is obtained. When training the model for writing quality evaluation, the features can be constructed by combining multi-dimensional information for model training. When the model evaluates the writing quality, it can score by comprehensively considering the multi-dimensional evaluation ability. The objectivity of writing quality evaluation can be ensured, and the evaluation efficiency can be improved, thereby saving human resources and providing a faster processing method for online or offline marking.

[0101] Corresponding to the method embodiments described above, the present specification also provides writing evaluation model training device embodiments, Figure 5 A structure diagram of a writing evaluation model training device provided by one embodiment of the present specification is shown. As Figure 5 shown, the device includes: The input module 502 is configured to input the writing sample image to the character unit recognition model and the correction unit recognition model. The extraction module 504 is configured to extract character unit recognition information from the writing sample image by the character unit recognition model, and extract correction unit recognition information from the writing sample image by the correction unit recognition model. The construction module 506 is configured to construct writing character unit features according to the character unit recognition information, and construct writing correction unit features according to the correction unit recognition information. The training module 508 is configured to train the writing evaluation model by using the writing character unit features and the writing correction unit features until the target writing evaluation model meeting the training stop condition is obtained.

[0102] In an optional embodiment, the character unit recognition model includes a text detection unit and a text recognition unit, and the character unit recognition information is extracted from the writing sample image by the character unit recognition model, including: The text detection unit adds a text box to a character unit contained in the writing sample image to obtain a to-be-recognized image containing a text box; the text recognition unit recognizes a character unit corresponding to the text box in the to-be-recognized image to obtain character unit recognition information; wherein the character unit recognition information includes a character unit recognition result and a text box labeling result.

[0103] In an optional embodiment, the extracting, by the modification unit recognition model, modification unit recognition information in the writing sample image includes: The modification unit recognition model recognizes a modification unit in the writing sample image to obtain modification unit recognition information; wherein the modification unit is determined according to at least one of a deletion subunit, a symbol subunit, and a correction subunit.

[0104] In an optional embodiment, the training of the text detection unit includes: A sample image is obtained, and the sample image is input to an initial text detection unit; a first detection unit and a second detection unit in the initial text detection unit process the sample image to obtain a prediction probability map and a prediction threshold map; a prediction binary map is constructed according to the prediction probability map and the prediction threshold map; a label probability map and a label threshold map corresponding to the sample image are determined, and the label probability map is taken as a label binary map corresponding to the prediction binary map; the initial text detection unit is optimized based on the prediction probability map, the prediction threshold map, the prediction binary map, the label probability map, the label threshold map, and the label binary map until the text detection unit that meets the optimization stop condition is obtained.

[0105] In an optional embodiment, the constructing, according to the character unit recognition information, writing character unit features includes: A character unit recognition confidence feature, a character unit area feature, a character unit height feature, and a character unit frequency feature are constructed according to the character unit recognition information; writing character unit features are constructed based on the character unit recognition confidence feature, the character unit area feature, the character unit height feature, and the character unit frequency feature; wherein the constructing, according to the modification unit recognition information, writing modification unit features includes: constructing a modification unit frequency feature and a modification unit frequency feature according to the modification unit recognition information; and constructing writing modification unit features based on the modification unit frequency feature and the modification unit frequency feature.

[0106] In an optional embodiment, the word unit recognition confidence feature is constructed according to a recognition rate determined based on the word unit recognition information, and is used to represent the writing standardization degree of the word unit; the word unit area feature is constructed according to a mean value and a variance of the word frame area determined based on the word unit recognition information, and is used to represent the writing alignment degree of the word form of the word unit; the word unit height feature is constructed according to a mean value and a variance of the word frame height determined based on the word unit recognition information, and is used to represent the writing alignment degree of the line spacing of the word unit; and the word unit frequency feature is constructed according to the word unit frequency determined based on the word unit recognition information, and is used to represent the global alignment degree of the word unit.

[0107] In an optional embodiment, the training of the writing evaluation model by using the writing word unit features and the writing correction unit features until a target writing evaluation model satisfying a training stop condition is obtained comprises: merging the writing word unit features and the writing correction unit features to obtain writing merged features, inputting the writing merged features into the writing evaluation model for processing to obtain a predicted writing score, training the writing evaluation model based on the predicted writing score and a target writing score corresponding to the writing sample image, verifying whether the trained writing evaluation model satisfies the training stop condition by using a verification set, if not, obtaining a candidate writing sample image, taking the candidate writing sample image as the writing sample image, and performing the step of inputting the writing sample image into the word unit recognition model and the correction unit recognition model, and if yes, taking the trained writing evaluation model as the target writing evaluation model.

[0108] In an optional embodiment, after the step of merging the writing word unit features and the writing correction unit features to obtain writing merged features, the method further comprises: constructing a candidate sample pair based on the writing merged features and the target writing score, storing the candidate sample pair into a sample set, pre-processing the sample pairs contained in the sample set, determining a target sample set according to a pre-processing result, cleaning the target sample set according to a pre-set set cleaning strategy, and training the writing evaluation model by using the cleaned target sample set until a target writing evaluation model satisfying a training stop condition is obtained.

[0109] To be able to objectively evaluate the writing quality of composition type questions and improve the writing evaluation efficiency, a writing evaluation model for writing evaluation scoring can be trained. In the model training process, in order to enable the writing evaluation model to have strong prediction ability and consider multi-dimensional writing evaluation influence, the writing sample image can be first input to a character unit recognition model and a correction unit recognition model. The character unit recognition information is extracted from the writing sample image by the character unit recognition model, and the correction unit recognition information is extracted from the writing sample image by the correction unit recognition model. Comprehensive writing quality information collection in the writing dimension and the correction dimension is realized. Subsequently, the writing character unit features can be constructed according to the character unit recognition information, and the writing correction unit features can be constructed according to the correction unit recognition information. On this basis, the writing evaluation model can be trained by fusing the writing character unit features and the writing correction unit features until the target writing evaluation model meeting the training stop condition is obtained. When training the model for writing quality evaluation, the features can be constructed by combining multi-dimensional information for model training, so that the model can score by comprehensively considering multi-dimensional evaluation ability when evaluating the writing quality. The objectivity of writing quality evaluation can be ensured, and the evaluation efficiency can be improved, thereby saving human resources and providing a faster processing method for online or offline marking.

[0110] The above is a schematic scheme of a writing evaluation model training device of the embodiment. It should be noted that the technical scheme of the writing evaluation model training device belongs to the same concept as the technical scheme of the writing evaluation model training method described above. The technical scheme of the writing evaluation model training device, which is not described in detail, can be referred to the description of the technical scheme of the writing evaluation model training method.

[0111] Corresponding to the above method embodiment, the present specification also provides a writing evaluation device embodiment, Figure 6 A structural schematic diagram of a writing evaluation device provided by an embodiment of the present specification is shown. As shown in the figure, Figure 6 The device includes: The acquisition module 602 is configured to acquire a target writing image uploaded by a client for a target text; The prediction module 604 is configured to predict a writing score of the target writing image by using a target writing evaluation model, and predict a text score of the target text by using a text evaluation model, wherein the target writing evaluation model is determined by the above method; The feedback module 606 is configured to calculate a global score corresponding to the target text based on the writing score and the text score, and feed back the global score to the client.

[0112] The above is a schematic scheme of the writing evaluation device according to the embodiment. It should be noted that the technical scheme of the writing evaluation device belongs to the same concept as the technical scheme of the writing evaluation method described above, and the details of the technical scheme of the writing evaluation device that are not described in detail can be referred to the description of the technical scheme of the writing evaluation method.

[0113] Figure 7 A structural block diagram of a computing device 700 according to one embodiment of the present specification is shown. The components of the computing device 700 include, but are not limited to, a memory 710 and a processor 720. The processor 720 is connected to the memory 710 through a bus 730, and a database 750 is used to save data.

[0114] The computing device 700 also includes an access device 740, which enables the computing device 700 to communicate via one or more networks 760. Examples of these networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. The access device 740 can include one or more of any type of network interface (e.g., a network interface card (NIC)), wired or wireless, such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a Worldwide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, a near field communication (NFC).

[0115] In one embodiment of the present specification, the above-mentioned components of the computing device 700 and other components not shown in the Figure 7 may be connected to each other, for example, through a bus. It should be understood that Figure 7 The structural block diagram of the computing device shown is only for the purpose of example, and is not a limitation on the scope of the present specification. Those skilled in the art can add or replace other components as needed.

[0116] The computing device 700 can be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, etc.), a mobile phone (e.g., a smartphone), a wearable computing device (e.g., a smart watch, smart glasses, etc.), or other type of mobile device, or a stationary computing device such as a desktop computer or a personal computer (PC). The computing device 700 can also be a mobile or stationary server.

[0117] The processor 720 is configured to execute computer-executable instructions to implement the steps of the writing evaluation model training method or the writing evaluation method.

[0118] The above is a schematic solution of the computing device of the embodiment. It should be noted that the technical solution of the computing device belongs to the same concept as the technical solution of the writing evaluation model training method or the writing evaluation method, and details of the technical solution of the computing device that are not described in detail can be referred to the description of the technical solution of the writing evaluation model training method or the writing evaluation method.

[0119] An embodiment of the present specification further provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the writing evaluation model training method or the writing evaluation method.

[0120] The above is a schematic solution of the computer-readable storage medium of the embodiment. It should be noted that the technical solution of the storage medium belongs to the same concept as the technical solution of the writing evaluation model training method or the writing evaluation method, and details of the technical solution of the storage medium that are not described in detail can be referred to the description of the technical solution of the writing evaluation model training method or the writing evaluation method.

[0121] An embodiment of the present specification further provides a computer program product comprising a computer program or instructions, which, when executed by a processor, implement the steps of the writing evaluation model training method or the writing evaluation method.

[0122] The above is a schematic solution of the computer program product of the embodiment. It should be noted that the technical solution of the computer program product belongs to the same concept as the technical solution of the writing evaluation model training method or the writing evaluation method, and details of the technical solution of the computer program product that are not described in detail can be referred to the description of the technical solution of the writing evaluation model training method or the writing evaluation method.

[0123] The above describes particular embodiments of the present specification. Other embodiments are within the scope of the appended claims. In some cases, the acts or steps recited in the claims can be performed in a different order than those described in the embodiments and still achieve desirable results. In addition, the processes depicted in the figures do not necessarily require the particular order shown or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing can be advantageous or possible.

[0124] The computer readable medium can include any entity or apparatus capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the computer readable medium can include appropriate additions or subtractions according to the requirements of patent practice. For example, according to the patent practice in some regions, the computer readable medium does not include electrical carrier signals and telecommunication signals.

[0125] It should be noted that for the foregoing method embodiments, in order to facilitate description, they are all expressed as a combination of a series of acts, but those skilled in the art should know that the embodiments of the present specification are not limited by the order of the described acts, because according to the embodiments of the present specification, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification are all preferred embodiments, and the acts and modules involved are not necessarily necessary for the embodiments of the present specification.

[0126] In the above embodiments, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0127] The preferred embodiments of the present specification disclosed above are only used to help explain the present specification. The alternative embodiments do not describe all the details and do not limit the invention to the specific embodiments described. Obviously, according to the content of the embodiments of the present specification, many modifications and changes can be made. The present specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the embodiments of the present specification, so that those skilled in the art can well understand and use the present specification.

Claims

1. A writing assessment model training method, characterized in that: include: Inputting the writing sample image into the character unit recognition model and the erasure unit recognition model; extracting character unit recognition information from the writing sample image using the character unit recognition model, and extracting alteration unit recognition information from the writing sample image using the alteration unit recognition model; Constructing a written character unit feature based on the character unit identification information, and constructing a written erasure unit feature based on the erasure unit identification information; The writing evaluation model is trained using the written character unit features and the written alteration unit features until a target writing evaluation model that meets the training stop condition is obtained.

2. The writing assessment model training method according to claim 1, characterized in that: The character unit recognition model includes a text detection unit and a text recognition unit, and extracting character unit recognition information from the writing sample image by using the character unit recognition model includes: adding a text frame to the character unit contained in the writing sample image by the text detection unit to obtain an image to be recognized containing the text frame; Recognizing the character unit corresponding to the text box in the image to be recognized by the text recognition unit to obtain character unit recognition information; The character unit recognition information includes a character unit recognition result and a text box marking result.

3. The writing assessment model training method according to claim 1, characterized in that: Extracting the correction unit identification information from the writing sample image using the correction unit identification model includes: Identifying the altered units in the writing sample image using the altered unit identification model to obtain altered unit identification information; The modification unit is determined based on at least one of a deletion subunit, a symbol subunit, and a correction subunit.

4. The writing assessment model training method according to claim 2, characterized in that: The training of the text detection unit includes: Acquire a sample image, and input the sample image into an initial text detection unit; Processing the sample image by the first detection unit and the second detection unit in the initial text detection unit to obtain a prediction probability map and a prediction threshold map; Constructing a prediction binary map according to the prediction probability map and the prediction threshold map; Determine a label probability map and a label threshold map corresponding to the sample image, and use the label probability map as the label binary map corresponding to the predicted binary map; Based on the prediction probability map, the prediction threshold map, the prediction binary map, the label probability map, the label threshold map and the label binary map, the initial text detection unit is optimized until the text detection unit that meets the optimization stop condition is obtained.

5. The writing assessment model training method according to claim 1, characterized in that: The step of constructing a written character unit feature according to the character unit identification information includes: Constructing a character unit recognition confidence feature, a character unit area feature, a character unit height feature, and a character unit frequency feature according to the character unit recognition information; Constructing a written character unit feature based on the character unit recognition confidence feature, the character unit area feature, the character unit height feature, and the character unit frequency feature; The step of constructing a writing correction unit feature based on the correction unit identification information includes: Constructing a feature of the number of times a correction unit occurs and a feature of the frequency of the correction unit according to the correction unit identification information; Based on the number of correction unit features and the frequency of correction unit features, a writing correction unit feature is constructed.

6. The writing assessment model training method according to claim 5, characterized in that: The character unit recognition confidence feature is constructed according to the recognition rate determined by the character unit recognition information, and is used to represent the writing standardization of the character unit; The character unit area feature is constructed based on the mean and variance of the character frame area determined by the character unit identification information, and is used to represent the writing regularity of the character unit shape; The character unit height feature is constructed based on the mean and variance of the character frame height determined by the character unit identification information, and is used to represent the writing regularity of the line spacing of the character unit; The character unit frequency feature is constructed according to the character unit frequency determined by the character unit identification information, and is used to represent the global regularity of the character units.

7. The writing assessment model training method according to any one of claims 1 to 6, characterized in that: The method of training the writing evaluation model using the written character unit features and the written correction unit features until a target writing evaluation model that satisfies a training stop condition is obtained includes: Merging the written character unit feature and the written correction unit feature to obtain a writing merge feature, and inputting the writing merge feature into a writing evaluation model for processing to obtain a predicted writing score; training the handwriting evaluation model based on the predicted handwriting score and the target handwriting score corresponding to the handwriting sample image; Use the validation set to verify whether the trained writing evaluation model meets the training stop conditions; If not, obtaining a candidate writing sample image, using the candidate writing sample image as the writing sample image, and performing the step of inputting the writing sample image into a character unit recognition model and an alteration unit recognition model; If so, the trained writing evaluation model is used as the target writing evaluation model.

8. The writing assessment model training method according to claim 7, characterized in that: After the step of merging the written character unit feature and the written alteration unit feature to obtain a written merge feature is performed, the method further includes: Constructing candidate sample pairs based on the handwriting merging feature and the target handwriting score, and storing the candidate sample pairs in a sample set; Preprocessing the sample pairs included in the sample set, and determining a target sample set based on the preprocessing results; The target sample set is cleaned according to a preset set cleaning strategy, and the writing evaluation model is trained using the cleaned target sample set until a target writing evaluation model that meets the training stop condition is obtained.

9. A writing assessment method, characterized in that: include: Obtain the target writing image uploaded by the client for the target text; Predicting a handwriting score of the target handwriting image using a target handwriting evaluation model, and predicting a text score of the target text using a text evaluation model, wherein the target handwriting evaluation model is determined by the method of any one of claims 1 to 8; A global score corresponding to the target text is calculated based on the writing score and the text score, and the global score is fed back to the client.

10. A writing assessment model training device, characterized in that: include: An input module configured to input the writing sample image into the character unit recognition model and the erasure unit recognition model; an extraction module configured to extract character unit recognition information from the writing sample image using the character unit recognition model, and to extract alteration unit recognition information from the writing sample image using the alteration unit recognition model; A construction module configured to construct a written character unit feature according to the character unit identification information, and to construct a written erasure unit feature according to the erasure unit identification information; The training module is configured to train the writing evaluation model using the written character unit features and the written correction unit features until a target writing evaluation model that meets the training stop condition is obtained.

11. A writing assessment device, characterized in that: include: An acquisition module is configured to acquire a target writing image uploaded by a client for a target text; a prediction module configured to predict a handwriting score of the target handwriting image using a target handwriting evaluation model, and to predict a text score of the target text using a text evaluation model, wherein the target handwriting evaluation model is determined by the method of any one of claims 1 to 8; The feedback module is configured to calculate a global score corresponding to the target text based on the writing score and the text score, and feed back the global score to the client.

12. A computing device, characterized in that include: memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the method according to any one of claims 1 to 9 are implemented.

13. A computer-readable storage medium, characterized in that It stores computer-executable instructions, which implement the steps of the method according to any one of claims 1 to 9 when executed by a processor.

14. A computer program product, characterized in that The method comprises a computer program or an instruction, which implements the steps of the method according to any one of claims 1 to 9 when the computer program or the instruction is executed by a processor.

Images