Intelligent information processing method for family-school cooperative cultivation
By collecting multimodal data in the home-school collaboration platform, preprocessing and fusion across modalities, generating student learning semantic vectors, and constructing a knowledge graph, the problems of information fragmentation and low interaction efficiency are solved, and efficient information processing and resource recommendation are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-20
- Publication Date
- 2026-03-27
AI Technical Summary
Existing home-school collaboration platforms lack a unified integration mechanism in multimodal data processing, resulting in information fragmentation, low efficiency in cross-modal semantic alignment, low interaction efficiency, and inaccurate student learning profiles that fail to truly reflect the degree of knowledge mastery and learning habits.
Multimodal data is collected, preprocessed, and then cross-modal alignment and fusion are performed using an attention mechanism and an adaptive weight algorithm to generate student learning semantic vectors, construct a knowledge graph, generate personalized ability vectors, calculate vector similarity, push educational resources, and predict learning problems through an LSTM time series model to achieve intelligent semantic interaction.
It improved information processing efficiency, reduced information loss and redundancy, optimized the accuracy of information retrieval and resource recommendation, enabled proactive intervention, simplified the interaction process, and improved information processing efficiency.
Smart Images
Figure CN121746128A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, and in particular to an intelligent information processing method for collaborative education between home and school. Background Technology
[0002] With the in-depth application of information technology in the field of education, home-school collaboration has become an important way to improve the quality of education. At present, home-school collaboration platforms mainly realize information interaction between teachers, parents and students through multimodal data such as text messages, homework pictures, and voice messages. However, the existing technology has low efficiency in intelligent information processing in home-school collaboration scenarios, mainly reflected in the following aspects: (1) Existing systems usually process different types of data such as text, images and voice independently, lacking a unified multimodal data integration mechanism. It is difficult to effectively associate handwriting information in homework pictures, semantic content in voice feedback and text comments, resulting in information fragmentation, redundant data processing and low efficiency; (2) Existing methods mostly use simple rule matching or single-modal feature extraction, failing to achieve effective alignment of cross-modal semantics. The student's homework completion status, classroom performance and parent feedback cannot be integrated through dynamic weighting mechanism, making the generated student learning profile inaccurate and difficult to truly reflect their knowledge mastery and learning habits; (3) Existing home-school collaboration platforms usually use keyword matching or fixed templates to respond to the needs of parents and teachers, making it difficult to understand complex semantic intentions, resulting in low interaction efficiency and low information processing efficiency. Summary of the Invention
[0003] This invention provides an intelligent information processing method for collaborative education between home and school. The method includes:
[0004] Collect multimodal data in home-school scenarios, including text data, homework image data, and voice data;
[0005] The multimodal data is preprocessed, including contour extraction, noise reduction, and format unification.
[0006] Extract unimodal features from preprocessed multimodal data, perform cross-modal alignment and fusion of the unimodal features through an attention mechanism and an adaptive weight algorithm to generate student learning semantic vectors, and calculate the fusion similarity between the student learning semantic vectors and preset standard semantic vectors;
[0007] Based on the student learning semantic vector, a knowledge graph containing three types of entities—knowledge points, types of wrong questions, and learning habits—is constructed. A weighted graph embedding algorithm is used in conjunction with the student's historical learning data to generate personalized ability vectors.
[0008] Calculate the matching similarity between the personalized ability vector and the educational resource vector, select the educational resources with the highest matching degree, and push them to the corresponding terminal; predict the student's learning problems in the future within a preset period using the LSTM time series model, and generate intervention strategies in combination with the knowledge graph;
[0009] The system receives the needs information from home and school users, identifies the needs and key entities through joint identification of intent and entity, and generates targeted response information based on the personalized capability vector to achieve intelligent semantic interaction.
[0010] Access control is applied to the preprocessed multimodal data and student learning semantic vectors based on preset user permission allocation rules. A dynamic desensitization algorithm is used to process sensitive data. If the fusion similarity is greater than the first-level preset similarity, the information processing process is deemed to meet the preset standard, and the current processing parameters remain unchanged. If the fusion similarity is less than or equal to the first-level preset similarity but greater than or equal to the second-level preset similarity, the preset standard is deemed not met, and the fusion weight is adjusted and recalculated. If the fusion similarity is less than the second-level preset similarity, the preset standard is deemed not met.
[0011] Furthermore:
[0012] When preprocessing the multimodal data,
[0013] If the multimodal data is homework image data, the improved joint model of target detection and text recognition selects image feature points for contour extraction and removes background noise. The image feature points include edge points of the question area and contour points of incorrect question marks.
[0014] If the multimodal data is text data, a semantic understanding model fine-tuned in the education field is used to extract text semantic features. The text data includes student grade data and home-school notification data.
[0015] If the multimodal data is speech data, the speech is converted into text data through an education-specific speech-to-text model, and emotional features in the speech are extracted. The emotional features include three categories: anxiety, calmness, and positivity, which are used to reflect the emotional state of parents or students.
[0016] Furthermore:
[0017] When performing cross-modal alignment fusion on the single-modal features, different fusion weights are assigned to the text semantic features, the homework image recognition features, and the speech emotion features, respectively. The fusion weight of the homework image recognition features is k higher than that of the text semantic features and the speech emotion features, where 25% < k < 45%.
[0018] The calculation method for the fusion similarity is the cosine similarity between the student's learning semantic vector and the preset standard semantic vector; the preset standard semantic vector is generated based on the average learning data of students in the same grade over the past 3 months, and includes the typical knowledge point mastery characteristics and common learning habit characteristics of students in that grade.
[0019] Furthermore:
[0020] When constructing the knowledge graph, a joint algorithm of rule matching and semantic recognition is used to extract entities from the student's learning semantic vector;
[0021] The rule matching algorithm pre-sets association rules between error types and knowledge points. The association rules include associating geometric proof errors with spatial imagination knowledge points and algebraic calculation errors with numerical computation knowledge points.
[0022] The semantic recognition joint algorithm uses a fine-tuned entity recognition model to extract learning habit entities from the text data and the text data after speech conversion. The learning habit entities include procrastination in completing homework, actively asking teachers questions, and insufficient time spent with parents accompanying students in their studies.
[0023] Furthermore:
[0024] When the weighted graph embedding algorithm generates the personalized ability vector, it uses the student's historical learning data as weighting factors, including the rate of the most recent three similar questions, the duration of knowledge point mastery, and the frequency of learning habits. The time range of the student's historical learning data is the past month.
[0025] The dimension of the personalized ability vector is consistent with the number of knowledge point entities in the knowledge graph. Each dimension value represents the mastery level of the corresponding knowledge point, and the value range is 0-1.
[0026] Furthermore:
[0027] When calculating the matching similarity, a cosine similarity algorithm is used to filter educational resources with a matching similarity greater than Vxs. The top 5 educational resources are selected and pushed to the corresponding terminals, which include parent terminals and teacher terminals. The educational resources pushed to the parent terminals are accompanied by family guidance instructions. Wherein, 0.5 < Vxs < 0.8.
[0028] When the LSTM time series model predicts the future for a preset period, if it predicts a preset learning problem category, which includes knowledge gaps and learning habit degradation, then it generates an intervention strategy based on the association rules in the knowledge graph.
[0029] Furthermore, the preset user permission allocation rules include:
[0030] If the user is a parent, they can only access the personalized ability vector, the pushed educational resources, and the intervention strategies of the associated student, and cannot access any data of other students;
[0031] If the user is a teacher, they can access the student learning semantic vectors, knowledge graphs, and overall class learning statistics of the students in their class.
[0032] When the dynamic desensitization algorithm processes the sensitive data, it masks the sensitive data. When parents view the data, only the student's name and the last four digits of the mobile phone number are displayed. When teachers view the data, the student's name and the complete student ID are displayed. All desensitized data cannot be downloaded.
[0033] The sensitive data includes at least the student's name, parents' mobile phone number, and home address.
[0034] Furthermore:
[0035] When the information processing process does not meet the preset standard, the difference between the secondary preset similarity and the fused similarity is calculated as the similarity difference. The specific reason is determined based on the similarity difference, specifically:
[0036] If the similarity difference is less than Vcz1, the reason is that the single-modal feature extraction is incomplete, and the number of feature points extracted from the job image data needs to be increased by Bs times the original number.
[0037] If the similarity difference is greater than or equal to Vcz1 and less than Vcz2, the reason is determined to be improper cross-modal fusion weight. The fusion weight of the image recognition feature of the task needs to be reduced to Br times the original weight, while the fusion weight of the text semantic feature needs to be increased to Bw times the original weight.
[0038] If the similarity difference is greater than or equal to Vcz2, the reason is determined to be excessive noise reduction in the preprocessing process, and the noise reduction intensity needs to be reduced to Bz times the original value.
[0039] Wherein, Vcz1 is the first similarity difference threshold, Vcz2 is the second similarity difference threshold, 1.2 < Bs < 1.7, 0.68 < Br < 0.85, 1.1 < Bw < 1.3, 0.4 < Bz < 0.6.
[0040] Furthermore:
[0041] After the preset time for pushing the educational resources, feedback data on students' completion of resource training is collected. The feedback data includes the error correction rate, training completion time, and parent evaluation. The similarity between the feedback data and the intervention strategy is calculated. The similarity is calculated by combining the error correction rate, the training completion time target rate, and the parent evaluation score.
[0042] If the adaptation similarity is greater than Dxs1, then the content and execution frequency of the original intervention strategy are maintained.
[0043] If the adaptation similarity is less than or equal to Dxs1 and greater than or equal to Dxs2, then the frequency of pushing the educational resources is increased by Bp times.
[0044] If the adaptation similarity is less than Dxs2, then educational resources are reselected based on the knowledge graph, a new intervention strategy is generated and pushed to the relevant authorities, and the execution cycle of the new intervention strategy is shortened to 1 week in order to quickly evaluate the adjustment effect.
[0045] Where Dxs1 is the first adaptation similarity threshold, Dxs2 is the second adaptation similarity threshold, and 1.5 < Bp < 1.8.
[0046] This invention comprehensively collects text, images, and audio data to reduce information gaps and improve information coverage and completeness; it aligns multimodal features using attention mechanisms and adaptive weighting algorithms to efficiently integrate key information, reduce redundancy, and improve feature representation efficiency; it structures data into entities such as knowledge points and error types to facilitate rapid querying and reasoning, optimizing information retrieval efficiency; it generates personalized vectors based on historical data to make resource recommendations more accurate and reduce search and matching time; it automatically filters educational resources by calculating vector similarity, reducing manual intervention and improving response speed; it predicts future learning problems to achieve proactive intervention, prevent problem accumulation, and improve problem handling efficiency; and it quickly generates responses through intent and entity recognition, simplifying the interaction process and improving information processing efficiency.
[0047] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description
[0048] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the scope of the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:
[0049] Figure 1 A flowchart of an intelligent information processing method for home-school collaborative education according to an embodiment of the present invention is shown. Detailed Implementation
[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0051] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.
[0052] Figure 1 A flowchart of an intelligent information processing method for home-school collaborative education according to an embodiment of the present invention is shown. The method includes:
[0053] S101, Collect multimodal data in the home-school scenario, the multimodal data including text data, homework image data and voice data;
[0054] S102, preprocess the multimodal data, the preprocessing including contour extraction, noise reduction and format unification;
[0055] S103, extract the single-modal features of the preprocessed multimodal data, perform cross-modal alignment and fusion of the single-modal features through an attention mechanism and an adaptive weight algorithm to generate student learning semantic vectors, and calculate the fusion similarity between the student learning semantic vectors and the preset standard semantic vectors;
[0056] S104, Based on the student learning semantic vector, construct a knowledge graph containing three types of entities: knowledge points, types of wrong questions, and learning habits, and use a weighted graph embedding algorithm to combine the student's historical learning data to generate a personalized ability vector.
[0057] S105, calculate the matching similarity between the personalized ability vector and the educational resource vector, select the educational resources with the highest matching degree, and push them to the corresponding terminal; predict the student's learning problems in the future within a preset period using the LSTM time series model, and generate an intervention strategy in combination with the knowledge graph;
[0058] S106: Receive the demand information from home and school users, identify the demand intention and key entities through joint identification of intent and entity, and generate targeted response information based on the personalized capability vector to achieve intelligent semantic interaction;
[0059] S107, based on preset user permission allocation rules, access control is performed on the preprocessed multimodal data and student learning semantic vectors. A dynamic desensitization algorithm is used to process sensitive data. If the fusion similarity is greater than the first-level preset similarity, it is determined that the information processing process meets the preset standard, and the current processing parameters remain unchanged. If the fusion similarity is less than or equal to the first-level preset similarity and greater than or equal to the second-level preset similarity, it is determined that the preset standard is not met, and the fusion weight is adjusted and recalculated. If the fusion similarity is less than the second-level preset similarity, it is determined that the preset standard is not met.
[0060] In some embodiments, when preprocessing the multimodal data, if the multimodal data is homework image data, an improved joint model of object detection and text recognition is used to select image feature points for contour extraction and remove background noise. The image feature points include edge points of the question area and contour points of incorrect question marks. If the multimodal data is text data, a semantic understanding model finely tuned in the education field is used to extract text semantic features. The text data includes student grade data and home-school notification data. If the multimodal data is speech data, a speech-to-text model specifically designed for education is used to convert speech into text data and extract emotional features from the speech. The emotional features include three categories: anxiety, calmness, and positivity, which are used to reflect the emotional state of parents or students. According to embodiments of the present invention, key feature points are automatically extracted through a joint model of object detection and text recognition, background noise is removed, irrelevant information processing is reduced, core content is focused, and image analysis efficiency is improved. By adopting a semantic understanding model finely tuned in the education field, text semantic features can be extracted quickly and accurately, avoiding the incompatibility of general models and optimizing the text processing flow. Speech is converted into text through a specialized speech-to-text model, which facilitates unified processing. At the same time, emotional features are extracted, sentiment analysis is automated, manual intervention is reduced, information integration is accelerated, and thus information processing efficiency is improved.
[0061] For example, for homework image data, a math homework image (1920×1080 pixels) was collected and processed using an improved object detection and text recognition joint model. The model automatically detected three question regions in the image, and extracted 12 edge points for each question region (a total of 36 question region edge points). It also detected two wrong question marking regions, and extracted 5 contour points for each marking (a total of 10 wrong question marking contour points). After contour extraction, the background noise of the image was removed.
[0062] For text data, a home-school notification text (150 characters in length) is collected and preprocessed using a semantic understanding model finely tuned in the education field. The model extracts the semantic features of the text and generates a 300-dimensional feature vector. The student grade records in the text data are parsed into structured data.
[0063] For the voice data, a parent's voice feedback was collected (8 seconds long, 16kHz sampling rate), which was converted into text data through an education-specific speech-to-text model. The converted text was 20 characters long. At the same time, the emotional features in the voice were extracted, and the emotional probability distribution was output as follows: anxiety score 0.7, calm score 0.2, and positive score 0.1.
[0064] In some embodiments, when performing cross-modal alignment fusion on the single-modal features, different fusion weights are assigned to the text semantic features, homework image recognition features, and speech emotion features, respectively. The fusion weight of the homework image recognition features is k higher than that of the text semantic features and speech emotion features, where 25% < k < 45%. The fusion similarity is calculated as the cosine similarity between the student's learning semantic vector and the preset standard semantic vector. The preset standard semantic vector is generated based on the average learning data of students in the same grade over the past three months, including typical knowledge point mastery features and common learning habit features of students in that grade. According to the embodiments of the present invention, the differentiated fusion weights highlight the importance of homework image recognition features, reduce noise interference from text and speech features, optimize the feature fusion process, and reduce computational complexity. The cosine similarity calculation is simple and efficient, enabling rapid evaluation of vector similarity and improving information matching speed. The preset standard semantic vector is pre-generated based on historical data, avoiding the overhead of real-time standard vector calculation, reducing processing latency, and thus improving information processing efficiency.
[0065] For example, when performing cross-modal alignment and fusion on single-modal features, assuming the text semantic feature vector is [0.8, 0.2], the homework image recognition feature vector is [0.9, 0.1], and the speech emotion feature vector is [0.3, 0.7], and setting k=30%, the weights are calculated as follows: text semantic feature weight w text =(1-0.3) / 3≈0.2333, speech emotion feature weight w audio =w text ≈0.2333, Feature weight w for image recognition in assignment image =w text +0.3≈0.5333; The calculation process for the student learning semantic vector can be: x component = 0.2333×0.8 + 0.5333×0.9 + 0.2333×0.3≈0.7366, y component = 0.2333×0.2 + 0.5333×0.1 + 0.2333×0.7≈0.2633, therefore the student learning semantic vector is [0.7366, 0.2633]; The preset standard semantic vector is generated based on the average learning data of students in the same grade over the past 3 months, assumed to be [0.5, 0.5]. The cosine similarity calculation process for the fusion similarity can be: vector dot product = 0.7366×0.5 + 0.2633×0.5≈0.49995, the magnitude of the student learning semantic vector is approximately ≈0.7822, preset standard semantic vector magnitude ≈ The cosine similarity is approximately 0.7071, and the cosine similarity is approximately 0.49995 / (0.7822×0.7071)≈0.904.
[0066] In some embodiments, when constructing the knowledge graph, a joint algorithm of rule matching and semantic recognition is used to extract entities from the student learning semantic vector. The rule matching algorithm pre-sets association rules between error types and knowledge points, including association rules between geometric proof errors and spatial imagination ability knowledge points, and association rules between algebraic calculation errors and numerical operation ability knowledge points. The joint semantic recognition algorithm uses a fine-tuned entity recognition model to extract learning habit entities from the text data and the text data after speech conversion. The learning habit entities include procrastination in completing homework, actively asking teachers questions, and insufficient parental involvement in learning. According to embodiments of the present invention, the rule matching algorithm directly maps error types and knowledge points through predefined association rules, reducing the amount of real-time analysis computation and improving processing speed; the semantic recognition algorithm uses a fine-tuned model to optimize the accuracy of entity extraction, quickly identifying learning habit entities for text and speech data, reducing the error rate; the combined use of rule matching and semantic recognition complementarily covers structured and unstructured data, achieving automated entity extraction, reducing manual intervention, and thus improving information processing efficiency.
[0067] In some embodiments, when the weighted graph embedding algorithm generates the personalized ability vector, it uses the student's historical learning data as weighting factors, including the rate of similar incorrect answers in the last three attempts, the duration of knowledge point mastery, and the frequency of learning habits. The time range of the student's historical learning data is the past month. The dimension of the personalized ability vector is consistent with the number of knowledge point entities in the knowledge graph, and each dimension value represents the mastery level of the corresponding knowledge point, with a value range of 0-1. According to the embodiments of the present invention, limiting the data time range reduces the amount of data processing and lowers the computational complexity; generating a fixed-dimensional personalized ability vector simplifies data representation, making subsequent operations more efficient, thereby improving information processing efficiency.
[0068] In some embodiments, when calculating the matching similarity, a cosine similarity algorithm is used to filter educational resources with a matching similarity greater than Vxs, and the top 5 educational resources are selected and pushed to the corresponding terminals. The corresponding terminals include parent terminals and teacher terminals, and the educational resources pushed to the parent terminals are accompanied by family guidance instructions; wherein, 0.5 < Vxs < 0.8; when the LSTM time series model predicts the future preset period, if a preset learning problem category is predicted, the preset learning problem category includes knowledge point gaps and learning habit degradation, then an intervention strategy is generated based on the association rules in the knowledge graph. According to embodiments of the present invention, the matching similarity is calculated using a cosine similarity algorithm, which is highly efficient, suitable for high-dimensional data, and allows for rapid filtering of relevant resources. By setting a similarity threshold, low-similarity resources are filtered out, reducing data processing volume and improving filtering efficiency. By selecting only the top 5 educational resources for push notifications, the amount of information is limited, information overload is avoided, and information delivery efficiency is improved. Targeted push notifications to parent and teacher terminals, along with accompanying family guidance instructions for parent terminals, reduce user processing time and improve resource utilization efficiency. By using an LSTM time series model to predict learning problems and identify knowledge gaps in advance, preventative intervention is achieved, reducing subsequent processing costs and thus improving information processing efficiency.
[0069] For example, suppose a knowledge graph contains 5 knowledge entity points (K1, K2, K3, K4, K5), and the personalized ability vector is a 5-dimensional vector, where each dimension value represents the mastery level of the corresponding knowledge point. The personalized ability vector generated based on the student's historical learning data is A=[0.6,0.7,0.5,0.8,0.4], where the mastery level of K1 is 0.6, K2 is 0.7, K3 is 0.5, K4 is 0.8, and K5 is 0. 4; The educational resource database contains 10 educational resources (R1 to R10). Each resource is represented by a 5-dimensional vector indicating its relevance to the knowledge point. The vector for R1 is B1=[0.8,0.6,0.7,0.9,0.5], the vector for R2 is B2=[0.5,0.8,0.6,0.7,0.9], the vector for R3 is B3=[0.9,0.7,0.8,0.6,0.5], and the vector for R4 is B4=[0.7,0.5,0.6,0.5]. The vectors for R5 are B5=[0.4,0.6,0.5,0.7,0.8], R6=B6=[0.6,0.9,0.8,0.5,0.7], R7=B7=[0.8,0.4,0.5,0.6,0.9], R8=B8=[0.5,0.5,0.4,0.4,0.6], and R9=B9=[0.7,0.8,0.9,0.5,0.6]. The vector of R10 is B{10}=[0.6,0.7,0.5,0.8,0.7]; the cosine similarity algorithm is used to calculate the matching similarity between the personalized ability vector A and each educational resource vector, A×B1=(0.6×0.8)+(0.7×0.6)+(0.5×0.7)+(0.8×0.9)+(0.4×0.5)=0.48+0.42+0.35+0.72+0.20=2.17, ‖A‖=(0.6 2 +0.7 2 +0.5 2 +0.8 2 +0.4 2 ) 1 / 2 =1.378, ||B1||=(0.8 2 +0.6 2 +0.7 2 +0.9 2 +0.5 2 ) 1 / 2 =1.597, similarity = 2.17 / (1.378×1.597)==0.986, and so on. The similarity calculation results for all resources (rounded to two decimal places) are R1:0.99, R2:0.95, R3:0.98, R4:0.97, R5:0.89, R6:0.94, R7:0.91, R8:0.82, R9:0.96, R10:0.99;
[0070] A threshold Vxs=0.6 was set to filter resources with a similarity greater than 0.6. All resources met this requirement. Therefore, the top 5 resources {R1(0.99), R10(0.99), R3(0.98), R4(0.97), R9(0.96)} were pushed to the parents' terminals based on similarity ranking. Each resource was accompanied by family guidance instructions. For example, the instructions for R1 were "30 minutes of daily math practice, focusing on strengthening K1 and K4 knowledge points". The same 5 resources were pushed to the teachers' terminals, but without accompanying family guidance instructions.
[0071] In some embodiments, the preset user permission allocation rules include: if the user is a parent, they can only access the personalized ability vector of the associated student, the pushed educational resources, and the intervention strategy, and cannot access any data of other students; if the user is a teacher, they can access the student learning semantic vector, knowledge graph, and overall class learning statistics of the students in their class; when the dynamic desensitization algorithm processes the sensitive data, it masks the sensitive data, so that only the student's name + the last 4 digits of the mobile phone number are displayed when the parent views it, and the student's name + the complete student ID number is displayed when the teacher views it, and all desensitized data cannot be downloaded; wherein, the sensitive data includes at least the student's name, the parent's mobile phone number, and the home address. According to the embodiments of the present invention, by restricting the scope of data access based on roles through user permission allocation rules, the amount of data that the system needs to retrieve and process is reduced, thereby reducing the database query load and network transmission overhead; by processing sensitive data through masking using the dynamic desensitization algorithm, the data display efficiency is optimized and full data transmission is avoided; at the same time, the data not being downloadable prevents batch requests and reduces server pressure; the permission rules simplify the data query logic, and the system can quickly locate authorized data, thereby improving information processing efficiency.
[0072] In some embodiments, when the information processing process does not meet the preset standard, the difference between the secondary preset similarity and the fusion similarity is calculated as the similarity difference. The specific reason is determined based on the similarity difference, specifically: if the similarity difference is less than Vcz1, the reason is determined to be incomplete single-modal feature extraction, requiring an increase in the number of feature points extracted from the task image data by Bs times; if the similarity difference is greater than or equal to Vcz1 and less than Vcz2, the reason is determined to be improper cross-modal fusion weights, requiring an adjustment of the... The fusion weight of the image recognition features is reduced to Br times the original weight, while the fusion weight of the text semantic features is increased to Bw times the original weight. If the similarity difference is greater than or equal to Vcz2, the cause is determined to be excessive noise reduction in the preprocessing process, and the noise reduction intensity needs to be reduced to Bz times the original weight. Here, Vcz1 is the first similarity difference threshold, Vcz2 is the second similarity difference threshold, and 1.2 < Bs < 1.7, 0.68 < Br < 0.85, 1.1 < Bw < 1.3, and 0.4 < Bz < 0.6. According to this embodiment, by dynamically diagnosing specific bottlenecks in the information processing process through similarity differences, adaptive adjustments are achieved, avoiding inefficiency caused by fixed parameters. Targeted optimization measures are taken for different causes to accurately improve the efficiency of key processes. By adjusting parameters within a reasonable range, computational resources and processing quality are balanced, unnecessary computational overhead is reduced, thereby improving information processing efficiency.
[0073] In some embodiments, after the educational resources are pushed for a preset time, feedback data on students' completion of resource training is collected. The feedback data includes the error correction rate, training completion time, and parent evaluation. The similarity between the feedback data and the intervention strategy is calculated. The similarity is calculated by combining the error correction rate, the training completion time target rate, and the parent evaluation score. If the similarity is greater than Dxs1, the content and execution frequency of the original intervention strategy are maintained. If the similarity is less than or equal to Dxs1 and greater than or equal to Dxs2, the frequency of pushing the educational resources is increased by Bp times. If the similarity is less than Dxs2, educational resources are reselected based on the knowledge graph, a new intervention strategy is generated and pushed, and the execution cycle of the new intervention strategy is shortened to 1 week to quickly evaluate the adjustment effect. Here, Dxs1 is the first similarity threshold, Dxs2 is the second similarity threshold, and 1.5 < Bp < 1.8. According to embodiments of the present invention, by collecting feedback data such as error correction rate and training completion time, the system can evaluate the actual effect of intervention strategies in real time and dynamically optimize based on fit similarity, avoiding the lag of static strategies and reducing ineffective processing; by comparing fit similarity with thresholds, the system triggers different operations in layers, concentrating computing resources on the most needed adjustments and reducing redundant computing overhead; by comprehensively calculating fit similarity from multiple sources such as error correction rate, completion time, and parent evaluations, a more comprehensive effect evaluation is formed, reducing erroneous adjustments caused by deviations of a single indicator and improving processing accuracy.
[0074] It should be understood that the various processes described above can be used to rearrange, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.
[0075] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.
Claims
1. A smart information processing method for collaborative education between home and school, characterized in that, include: Collect multimodal data in home-school scenarios, including text data, homework image data, and voice data; The multimodal data is preprocessed, including contour extraction, noise reduction, and format unification. Extract unimodal features from preprocessed multimodal data, perform cross-modal alignment and fusion of the unimodal features through an attention mechanism and an adaptive weight algorithm to generate student learning semantic vectors, and calculate the fusion similarity between the student learning semantic vectors and preset standard semantic vectors; Based on the student learning semantic vector, a knowledge graph containing three types of entities—knowledge points, types of wrong questions, and learning habits—is constructed. A weighted graph embedding algorithm is used in conjunction with the student's historical learning data to generate personalized ability vectors. Calculate the matching similarity between the personalized ability vector and the educational resource vector, select the educational resources with the highest matching degree, and push them to the corresponding terminals; The LSTM time series model is used to predict students’ learning problems in a future preset period, and the knowledge graph is combined to generate intervention strategies. The system receives the needs information from home and school users, identifies the needs and key entities through joint identification of intent and entity, and generates targeted response information based on the personalized capability vector to achieve intelligent semantic interaction. Access control is applied to the preprocessed multimodal data and student learning semantic vectors based on preset user permission allocation rules. A dynamic desensitization algorithm is used to process sensitive data. If the fusion similarity is greater than the first-level preset similarity, the information processing process is determined to meet the preset standard, and the current processing parameters remain unchanged. If the fusion similarity is less than or equal to the first-level preset similarity and greater than or equal to the second-level preset similarity, it is determined that the preset standard is not met and the fusion weight is adjusted and recalculated; if the fusion similarity is less than the second-level preset similarity, it is determined that the preset standard is not met.
2. The intelligent information processing method for collaborative education between home and school as described in claim 1, characterized in that: When preprocessing the multimodal data, If the multimodal data is homework image data, the improved joint model of target detection and text recognition selects image feature points for contour extraction and removes background noise. The image feature points include edge points of the question area and contour points of incorrect question marks. If the multimodal data is text data, a semantic understanding model fine-tuned in the education field is used to extract text semantic features. The text data includes student grade data and home-school notification data. If the multimodal data is speech data, the speech is converted into text data through an education-specific speech-to-text model, and emotional features in the speech are extracted. The emotional features include three categories: anxiety, calmness, and positivity, which are used to reflect the emotional state of parents or students.
3. The intelligent information processing method for home-school collaborative education according to claim 2, characterized in that: When performing cross-modal alignment fusion on the single-modal features, different fusion weights are assigned to the text semantic features, the homework image recognition features, and the speech emotion features, respectively. The fusion weight of the homework image recognition features is k higher than that of the text semantic features and the speech emotion features, where 25% < k < 45%. The calculation method for the fusion similarity is the cosine similarity between the student's learning semantic vector and the preset standard semantic vector; the preset standard semantic vector is generated based on the average learning data of students in the same grade over the past 3 months, and includes the typical knowledge point mastery characteristics and common learning habit characteristics of students in that grade.
4. The intelligent information processing method for home-school collaborative education according to claim 3, characterized in that: When constructing the knowledge graph, a joint algorithm of rule matching and semantic recognition is used to extract entities from the student's learning semantic vector; The rule matching algorithm pre-sets association rules between error types and knowledge points. The association rules include associating geometric proof errors with spatial imagination knowledge points and algebraic calculation errors with numerical computation knowledge points. The semantic recognition joint algorithm uses a fine-tuned entity recognition model to extract learning habit entities from the text data and the text data after speech conversion. The learning habit entities include procrastination in completing homework, actively asking teachers questions, and insufficient time spent with parents accompanying students in their studies.
5. The intelligent information processing method for home-school collaborative education according to claim 4, characterized in that: When the weighted graph embedding algorithm generates the personalized ability vector, it uses the student's historical learning data as weighting factors, including the rate of the most recent three similar questions, the duration of knowledge point mastery, and the frequency of learning habits. The time range of the student's historical learning data is the past month. The dimension of the personalized ability vector is consistent with the number of knowledge point entities in the knowledge graph. Each dimension value represents the mastery level of the corresponding knowledge point, and the value range is 0-1.
6. The intelligent information processing method for home-school collaborative education according to claim 5, characterized in that: When calculating the matching similarity, a cosine similarity algorithm is used to filter educational resources with a matching similarity greater than Vxs. The top 5 educational resources are selected and pushed to the corresponding terminals, which include parent terminals and teacher terminals. The educational resources pushed to the parent terminals are accompanied by family guidance instructions. Wherein, 0.5 < Vxs < 0.
8. When the LSTM time series model predicts the future for a preset period, if it predicts a preset learning problem category, which includes knowledge gaps and learning habit degradation, then it generates an intervention strategy based on the association rules in the knowledge graph.
7. The intelligent information processing method for collaborative education between home and school according to claim 6, characterized in that, The preset user permission allocation rules include: If the user is a parent, they can only access the personalized ability vector, the pushed educational resources, and the intervention strategies of the associated student, and cannot access any data of other students; If the user is a teacher, they can access the student learning semantic vectors, knowledge graphs, and overall class learning statistics of the students in their class. When the dynamic desensitization algorithm processes the sensitive data, it masks the sensitive data. When parents view the data, only the student's name and the last four digits of the mobile phone number are displayed. When teachers view the data, the student's name and the complete student ID are displayed. All desensitized data cannot be downloaded. The sensitive data includes at least the student's name, parents' mobile phone number, and home address.
8. The intelligent information processing method for home-school collaborative education according to claim 7, characterized in that: When the information processing process does not meet the preset standard, the difference between the secondary preset similarity and the fused similarity is calculated as the similarity difference. The specific reason is determined based on the similarity difference, specifically: If the similarity difference is less than Vcz1, the reason is that the single-modal feature extraction is incomplete, and the number of feature points extracted from the job image data needs to be increased by Bs times the original number. If the similarity difference is greater than or equal to Vcz1 and less than Vcz2, the reason is determined to be improper cross-modal fusion weight. The fusion weight of the image recognition feature of the task needs to be reduced to Br times the original weight, while the fusion weight of the text semantic feature needs to be increased to Bw times the original weight. If the similarity difference is greater than or equal to Vcz2, the reason is determined to be excessive noise reduction in the preprocessing process, and the noise reduction intensity needs to be reduced to Bz times the original value. Wherein, Vcz1 is the first similarity difference threshold, Vcz2 is the second similarity difference threshold, 1.2 < Bs < 1.7, 0.68 < Br < 0.85, 1.1 < Bw < 1.3, 0.4 < Bz < 0.
6.
9. The intelligent information processing method for home-school collaborative education according to claim 8, characterized in that: After the preset time for pushing the educational resources, feedback data on students' completion of resource training is collected. The feedback data includes the error correction rate, training completion time, and parent evaluation. The similarity between the feedback data and the intervention strategy is calculated. The similarity is calculated by combining the error correction rate, the training completion time target rate, and the parent evaluation score. If the adaptation similarity is greater than Dxs1, then the content and execution frequency of the original intervention strategy are maintained. If the adaptation similarity is less than or equal to Dxs1 and greater than or equal to Dxs2, then the frequency of pushing the educational resources is increased by Bp times. If the adaptation similarity is less than Dxs2, then educational resources are reselected based on the knowledge graph, a new intervention strategy is generated and pushed to the relevant authorities, and the execution cycle of the new intervention strategy is shortened to 1 week in order to quickly evaluate the adjustment effect. Where Dxs1 is the first adaptation similarity threshold, Dxs2 is the second adaptation similarity threshold, and 1.5 < Bp < 1.8.