An education resource scene-based comprehensive error correction method based on deep learning

Abstract

The application discloses a kind of based on deep learning's educational resource scene-based comprehensive error correction method, comprising the following steps: S1, the request information of resource requester is obtained;S2, obtain scene type effective information and record into request type table;S3, dispatch data information from educational resource pool;S4, enter automatic matching error correction model: if text type then enter text data error correction, if image type then enter image error correction, obtain quantized similarity;S5, if similarity is higher than the minimum threshold set then the data information is directly sent to requester, otherwise execute S6;S6, record the predicted scene type and data information of current data, as negative sample is added to educational resource pool, return step S3.The application can carry out comprehensive error correction to two kinds of data types of text and image, can provide data quality guarantee for cloud service, improve the identification accuracy of error data, collect error data, and correctly write data to scene again.

Description

Technical Field

[0001] This invention relates to agile educational data quality management, specifically to a scenario-based comprehensive error correction method for educational resources based on deep learning. Background Technology

[0002] With the promulgation and advancement of the Ministry of Education's "Education Informatization 2.0 Action Plan," my country's education informatization has moved beyond simple data collection and cleaning to pursuing higher levels and broader educational openness. Therefore, driven by cloud services, how to rapidly identify and correct high-quality educational resource big data in diverse scenarios, and provide scenario-based, end-to-end integrated data correction cloud services, has become a crucial challenge in achieving automated education quality.

[0003] Text detection and recognition in various scenarios play a crucial role in enabling large data pools to be accurately applied in cloud services, providing users with accurate data, meeting the applicability requirements of the user process, and protecting important data from erroneous leakage in different scenarios.

[0004] Existing error correction methods mostly rely on traditional classification and filtering techniques. They extract information from the acquired data, obtaining entity relationship attributes, time data, entity words, and domain words. Positive samples are obtained based on entity words and domain words, and negative samples are obtained by processing the positive samples, forming a dataset. For long texts to be corrected, word segmentation, entity extraction, and semantic analysis are performed to obtain hyponyms and hypernyms and entity relationships, followed by graph retrieval. The positive and negative datasets are used to train the text correction model, providing the final correct suggestion. However, this method is not suitable for big data management in cloud service architectures. The massive amounts of educational resources have diverse data types and content, making it impossible to rely solely on manually customized attributes and error detection and correction.

[0005] Taking the above method as an example, the error correction methods commonly used in the past, based on rules and statistical machine learning, mainly rely on lexicons or statistical language models. They lack the dependence on context and diverse scenarios when correcting errors, which can easily lead to low-level false alarms. However, if they are used in a big data education resource environment, the number of false alarms will increase significantly. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a deep learning-based scenario-based comprehensive error correction method for educational resources that can comprehensively correct both text and image data types, provide data quality assurance for cloud services, improve the accuracy of erroneous data identification, collect erroneous data, and rewrite correct data back into the scene.

[0007] The objective of this invention is achieved through the following technical solution: a deep learning-based scenario-based comprehensive error correction method for educational resources, comprising the following steps:

[0008] S1. Obtain the request information from the resource requester;

[0009] S2. Parse the request content, obtain valid information about the scenario type, and record it in the request type table;

[0010] S3. Data information is retrieved from the educational resource pool and entered into the self-learning global data error correction system;

[0011] S4. Determine the type of the output scheduling information and enter the automatic matching and error correction model: if it is text type, enter text data error correction; if it is image type, enter image error correction and obtain the quantized similarity.

[0012] S5. Perform discrimination processing on the quantified similarity. If the similarity is higher than the set minimum threshold, send the data information directly to the requester and end the error correction service; otherwise, proceed to step S6.

[0013] S6. Record the predicted scenario type and data information of the current data, add it as a negative sample to the education resource pool, and return to step S3.

[0014] The information of the resource requester in step S1 includes the scenario type to which the requested data belongs and the data request details.

[0015] The scenario type of the requested data is selected from the data scenario type table updated by self-learning in the resource pool, and a unique request ID is created for the requester.

[0016] The data request details include the data type and the corresponding number of requested data.

[0017] The specific implementation method of step S2 is as follows: using the existing scene types in the current system as rows, the requested scene type as columns, and the information content as the record request ID, a scene matching list is established.

[0018] The similarity result after quantization in step S4:

[0019]

[0020] Among them, SimValue i,j Acc represents the similarity predicted between the i-th text classification and the j-th scene; j This represents the predicted similarity for the j-th scenario, which is defined by the system in the model according to requirements.

[0021] Acc i,j The calculation method is as follows:

[0022]

[0023] Among them, TPd i,j 1 represents the correct prediction of the data in the classification model for the i-th text classification and the j-th scenario; 0 < i ≤ I, 0 < j ≤ J, where i and j are controlled by the number of texts divided by I and the number of predefined scenarios J; v represents the number of predictions, V represents the total number of predictions, and N represents the number of classification decisions.

[0024] Step S6 includes the following sub-steps:

[0025] S61. Based on the current scenario requirements, define the categories of data in the education resource pool; collect and statistically analyze information from the entire scenario data pool, and record the correct scenario type name of each data point in the form of a table, recording them one by one from coarse to fine.

[0026] S62. Dataset creation: Randomly sample the latest educational resource pool data, classify and recombine the sampled data to form initial training positive and negative samples;

[0027] S63. Obtain the educational resources contained in the full-domain scene data pool, and enter the automatic matching and error correction model according to the data type; the process of the automatic matching and error correction model is as follows: determine the type of data. If it is text data, proceed to text error correction and proceed to step S631; if it is image data, proceed to image error correction and proceed to step S632.

[0028] S631. The text detection method CTPN based on connected preselection box network is used as the text detection module to perform text detection, find the target region for text recognition in S633, and enter S633.

[0029] S632. Treat the image data as continuous vertical pixel frames, perform frame cutting on the video image to obtain image information, then remove the image information for data cleaning (remove punctuation marks, long strings of numbers and other text content), and proceed to S633.

[0030] S633. Perform content recognition and classification to identify whether the scene type of the text or image content is close to the real scene type; the content recognition and classification adopts a dual encoder architecture: encoder 1 provides global information; encoder 2 acts as a local feature extractor, and the output of encoder 2 is input to the classifier;

[0031] Encoder 1 is based on BiLSTM, and a spatial attention mechanism is added before the output layer of BiLSTM to perform fine-grained semantic representation; the attention mechanism is used to mine more valuable word information and improve similarity.

[0032] a_v t,i,j =tanh(W s s t +W f f i,j )

[0033] a_v t =softmax i,j (a_v t,i,j )

[0034] a_v t,i,j The s represents the attention mechanism mask information. t Indicates an intermediate state; f i,j Represents image or text content; W f W represents the weight; s a_v represents the weight of an image or text s in a document. t This represents the normalized attention weights; t represents time t; i represents the i-th row and j represents the j-th column;

[0035] Encoder 2 uses a DRNN network as a parallel detail feature mining module and fuses the obtained feature information with the feature information obtained by encoder 1.

[0036] h t =Conv(g t ,x t )

[0037]

[0038] Where h t g represents the feature generated by the window filter in encoder 2. t Represented as true value feature information, x t The input is feature information; the feature information is fused using Concat and denoted as H. t ; obtain fusion feature H t Then, the classification prediction result is obtained by normalization through the classification layer. The classification result is then compared with the scene value. If it is correctly predicted, it is represented as 1; otherwise, it is represented as 0. Then, it is substituted into the similarity result calculation formula.

[0039] The beneficial effects of this invention are:

[0040] 1. This invention proposes a comprehensive scenario-based error correction strategy based on deep learning for multi-scenario data error correction and recognition. It can comprehensively correct errors in both text and image data types, provide data quality assurance for cloud services, improve the accuracy of erroneous data identification, collect erroneous data, and rewrite correct data back into the scenario.

[0041] 2. Traditional text classification primarily utilizes n-grams from Natural Language Processing (NLP) to extract text features, then updates the feature weights and inputs the extracted feature information into traditional classifiers such as Logistic Regression or SVM for training. However, this approach can be severely detrimental when encountering issues like data sparsity or dimensionality explosion. To address this problem, this invention employs a novel dual-encoder architecture. Encoder 1 provides global information, while Encoder 2 acts as a local feature extractor, directly inputting the information into the classifier for classification, thereby improving classification accuracy. Attached Figure Description

[0042] Figure 1 The flowchart shows the deep learning-based scenario-based comprehensive error correction method for educational resources according to the present invention.

[0043] Figure 2 This is a schematic diagram of the spatial attention mechanism.

[0044] Figure 3 This is a schematic diagram of encoder 2. Detailed Implementation

[0045] This invention proposes an effective data error correction and scene recognition technology method based on deep learning. It studies the automation of the multi-domain educational data scene error correction process, combines it with a comprehensive educational big data knowledge base, and realizes integrated error correction services for data across multiple scenarios. This achieves scene recognition and data content error correction, and leverages existing comprehensive labeled data and application scenario types to learn the correlation between data and scenarios, thereby improving error correction capabilities. It can not only identify whether scene data is correct, but also handle various file types (text and images), assisting in the unified processing, automatic scheduling, and unified collection of erroneous data during the open use of massive educational data resources. The technical solution of this invention is further illustrated below with reference to the accompanying drawings.

[0046] like Figure 1 As shown, the present invention provides a scenario-based comprehensive error correction method for educational resources based on deep learning, comprising the following steps:

[0047] S1. Obtain the request information from the resource requester; the resource requester's information includes the scenario type to which the requested data belongs and the data request details;

[0048] The scenario type of the requested data is selected from the data scenario type table updated by self-learning in the resource pool, and a unique request ID is created for the requester.

[0049] The data request details include the data type (text / image) and the corresponding amount of requested data.

[0050] In this embodiment, the roles are: the requester (network request scenario type S), the accessed resource (network data L), the text type error correction model WT, and the image type error correction model VT.

[0051] Network A has complete request information and sends a request to obtain 100 resources for the scientific and technological knowledge scenario of operation resource L, namely 50 text resources and 50 video resources.

[0052] S2. Parse the request content, obtain valid scenario type information, and record it in the request type table. Specifically, use existing scenario types in the current system as rows, the requested scenario type as columns, and the information content as the request ID to create a scenario matching list. The error correction system identifies the request scenario type and the number of requests, generates a unique request ID, records it in the request type table, and performs scenario matching. If the scenario matching is successful, proceed to step S3.

[0053] S3. Data information is retrieved from the educational resource pool and entered into the self-learning global data error correction system;

[0054] S4. Determine the type of the output scheduling information and proceed to the automatic matching and error correction model: if it is text, proceed to text data error correction; if it is image, proceed to image error correction, and obtain the quantized similarity; quantized similarity results:

[0055]

[0056] Among them, SimValue i,j Acc represents the similarity predicted between the i-th text classification and the j-th scene; j This represents the predicted similarity of the j-th scenario, which is customized by the system in the model according to requirements. If a lower error tolerance is desired, the similarity value of the j-th scenario will be increased.

[0057] Acc i,j The calculation method is as follows:

[0058]

[0059] Among them, TPd i,j 1 represents the correct prediction of the data in the classification model for the i-th text classification and the j-th scenario; 0 < i ≤ I, 0 < j ≤ J, where i and j are controlled by the number of texts divided by I and the number of predefined scenarios J; v represents the number of predictions, V represents the total number of predictions, and N represents the number of classification decisions.

[0060] S5. Perform discrimination processing on the quantified similarity. If the similarity is higher than the set minimum threshold, send the data information directly to the requester and end the error correction service; otherwise, proceed to step S6.

[0061] When resource L outputs data of type file, it enters text correction mode, obtains the quantified scene similarity, and if the similarity is less than the set threshold, it is recorded as erroneous data, the output process exits, new data is reprocessed, the amount of output data is 0, and the erroneous data is used as a negative sample and re-input into the classifier for optimization training. Conversely, if the similarity is greater than the set threshold, it is recorded as correct data, and the amount of output data is incremented by one.

[0062] When the output data type of resource L is video, the video image is first sliced ​​into frames, treating the input image data as continuous vertical pixel frames. Image frame preprocessing is then performed, such as removing punctuation marks, long strings of numbers, and other text content to clarify the data. Then, image error correction begins. The quantized scene similarity is obtained. If the similarity is less than a set threshold, it is recorded as erroneous data, the output process exits, new data is reprocessed, the amount of output data is reduced to 0, and this erroneous data is used as a negative sample and re-input into the classifier for optimization training. Conversely, if the similarity is greater than the set threshold, it is recorded as correct data, and the amount of output data is incremented by one.

[0063] S6. Record the predicted scenario type and data information of the current data, add it as a negative sample to the educational resource pool, and return to step S3; including the following sub-steps:

[0064] S61. Based on the current scenario requirements, define the categories of data in the education resource pool; collect and statistically analyze information from the entire scenario data pool, and record the correct scenario type name of each data item in the form of a table, recording them one by one from coarse to fine, such as the major category being sports, and the minor category being basketball; the major category being electronic text, and the minor category being history, etc.

[0065] S62. Dataset Creation: Before creating the positive and negative sample datasets, in order to provide effective self-learning information for the subsequent data correction deep learning, it is necessary to randomly sample the latest educational resource pool data and classify and recombine the sampled data to form initial training positive and negative samples. The label format for positive samples is key:value, such as the real scene name: real data information; the label for negative samples is fake scene name: fake data information.

[0066] S63. Obtain the educational resources contained in the full-domain scene data pool, and enter the automatic matching and error correction model according to the data type; the process of the automatic matching and error correction model is as follows: determine the type of data. If it is text data, proceed to text error correction and proceed to step S631; if it is image data, proceed to image error correction and proceed to step S632.

[0067] S631. The resource pool contains abundant data information, but during text detection, there are many noise factors, such as insufficient clarity and tilted angles. This invention uses the CTPN text detection method based on a connected pre-selection box network as the text detection module to perform text detection, find the target region for text recognition in S633, and proceed to S633.

[0068] S632. Treat the image data as continuous vertical pixel frames, perform frame cutting on the video image to obtain image information, then remove the image information for data cleaning (remove punctuation marks, long strings of numbers and other text content), and proceed to S633.

[0069] S633. Perform content recognition and classification to identify whether the scene type of the text or image content is similar to the real scene type; the present invention aims to provide scene theme binary classification results, identifying whether the scene type of the text content is similar to the real scene type. Based on this requirement, the content recognition and classification of the present invention adopts a dual encoder architecture: encoder 1 provides global information; encoder 2 acts as a local feature extractor, and the output of encoder 2 is input to the classifier;

[0070] Encoder 1 is based on BiLSTM, and a spatial attention mechanism is added before the output layer of BiLSTM to perform fine-grained semantic representation; the attention mechanism is used to mine more valuable word information and improve similarity; the attention mechanism is as follows: Figure 2 As shown in the figure, Spatial Attention is a spatial attention mechanism, Encoder1 represents encoder 1, the dashed box on the left represents the input feature information and initial weights, after passing through the spatial attention mechanism, new weight values ​​are obtained, and then after normalization processing, the output value shown in the dashed box in the upper right corner is obtained.

[0071] a_v t,i,j =tanh(W s s t +W f f i,j )

[0072] a_v t =softmax i,j (a_v t,i,j )

[0073] a_v t,i,j The s represents the attention mechanism mask information. t This represents the intermediate state (i.e., the information generated during the model's processing of text content and classification of real-world scenes, originating from the last hidden layer of Encoder 1 network); f i,j Represents image or text content; W fW represents the weights (the initial global alignment weights assigned in Encoder 1 network, which are updated in subsequent iterations based on the frequency and importance of words); s a_v represents the weight of an image or text s in a document. t This represents the normalized attention weights; t represents time t; i represents the i-th row and j represents the j-th column;

[0074] Encoder 2 uses a DRNN network as a parallel detail feature mining module, and fuses the obtained feature information with the feature information obtained by encoder 1; for example Figure 3 As shown in the figure, Encoder1 on the left represents the output data of encoder 1, which serves as both the input for feature extraction by encoder 2 and the input for feature fusion by encoder 2.

[0075] h t =Conv(g t ,x t )

[0076]

[0077] Where h t g represents the feature generated by the window filter in encoder 2. t Represented as true value feature information, x t The input is feature information; a sentence contains k words, which are accumulated starting from j=1. The feature information is then fused using Concat and denoted as H. t H t H is used to guide the entire network by fusing feature information. t It contains information about the past and the future. This yields the fusion feature H. t Then, the classification prediction result is obtained by normalization through the classification layer. The classification result is then compared with the scene value. If it is correctly predicted, it is represented as 1; otherwise, it is represented as 0. Then, it is substituted into the similarity result calculation formula.

[0078] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A deep learning-based educational resource scene-based comprehensive error correction method, characterized in that, Includes the following steps: S1. Obtain the request information from the resource requester; S2. Parse the request content, obtain valid information about the scenario type, and record it in the request type table; S3. Data information is retrieved from the educational resource pool and entered into the self-learning global data error correction system; S4. Determine the type of the output scheduling information and enter the automatic matching and error correction model: if it is text type, enter text data error correction; if it is image type, enter image error correction and obtain the quantized similarity. S5. Perform discrimination processing on the quantified similarity. If the similarity is higher than the set minimum threshold, send the data information directly to the requester and end the error correction service. Otherwise, proceed to step S6; S6. Record the predicted scenario type and data information of the current data, add it as a negative sample to the education resource pool, and return to step S3; Includes the following sub-steps: S61. Based on the current scenario requirements, define the categories of data in the education resource pool; collect and statistically analyze information from the entire scenario data pool, and record the correct scenario type name of each data point in the form of a table, recording them one by one from coarse to fine. S62. Dataset creation: Randomly sample the latest educational resource pool data, classify and recombine the sampled data to form initial training positive and negative samples; S63. Obtain the educational resources contained in the full-domain scene data pool, and enter the automatic matching and error correction model according to the data type; the process of the automatic matching and error correction model is as follows: determine the type of data. If it is text data, proceed to text error correction and proceed to step S631; if it is image data, proceed to image error correction and proceed to step S632. S631. The text detection method CTPN based on connected preselection box network is used as the text detection module to perform text detection, find the target region for text recognition in S633, and enter S633. S632. Treat the image data as continuous vertical pixel frames, perform frame cutting on the video image to obtain image information, then remove the image information for data cleaning, and proceed to S633. S633. Perform content recognition and classification to identify whether the scene type of the text or image content is close to the real scene type; the content recognition and classification adopts a dual encoder architecture: encoder 1 provides global information; encoder 2 acts as a local feature extractor, and the output of encoder 2 is input to the classifier; Encoder 1 is based on BiLSTM, and a spatial attention mechanism is added before the output layer of BiLSTM to perform fine-grained semantic representation; the attention mechanism is used to mine more valuable word information and improve similarity. ； ； The information in the middle represents the mask information for the attention mechanism. Indicates an intermediate state; Represents image or text content; Indicates weight; Indicates the weight of an image or text s in a document; This represents the normalized attention weights; t represents time t; i represents the i-th row and j represents the j-th column; The encoder 2 selects a DRNN network as a parallel detail feature mining module, and fuses the obtained feature information with the feature information obtained by the encoder 1, and the feature information after fusion is denoted as ; and fusion features are obtained. Then, the classification prediction result is obtained through the classification layer normalization.

2. The deep learning-based educational resource contextualization and error correction method of claim 1, wherein, The information of the resource requester in step S1 includes the scenario type to which the requested data belongs and the data request details. The scenario type of the requested data is selected from the data scenario type table updated by self-learning in the resource pool, and a unique request ID is created for the requester. The data request details include the data type and the corresponding number of requested data. 3.The deep learning-based educational resource contextualization and comprehensive error correction method of claim 1, wherein, The step S2 specifically realizes the method as follows: taking the existing scene types in the current system as rows, the requested scene types as columns, and the information content as the record request ID, a scene matching list is established. 4.The deep learning-based educational resource contextualization and comprehensive error correction method of claim 1, wherein, The step S4 quantized similarity result: ； wherein, represents the predicted similarity of the i-th text classification to the j-th scenario; represents the predicted similarity of the j-th scenario, defined in the model by the system according to the requirements; The calculation method is as follows: ； wherein, represents the i-th text classification and the j-th scene, the data in the classification model is correctly predicted, recorded as 1; 0 < i ≤ I, 0 < j ≤ J, i and j are controlled by the number I of text division and the number J of pre-defined scenes; v represents the prediction times, V represents the total number of predictions, and N represents the number of classification decisions.

Images