Method and system for predicting children's reading level based on picture book reading records

Abstract

The application discloses a children's reading level prediction method and system based on picture book reading records, comprising the following steps: collecting ecological behavior data in the process of parent-child reading, wherein the ecological behavior data at least includes picture book-child interaction data, parent-child interaction data and child-reading content feedback data; preprocessing the ecological behavior data, extracting multiple features and standardizing coding, wherein the features at least include key illustration touch rate of picture book-child interaction dimension, page turning force entropy value and confusion expression proportion, the application collects picture book-child interaction data, parent-child interaction data and child-reading content feedback data, constructs a multi-modal ecological behavior data system, comprehensively captures physical interaction details, language interaction quality and child cognitive feedback in the process of parent-child reading, and avoids one-sidedness caused by relying on only explicit data such as reading duration and frequency.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of intelligent education, and particularly relates to a children's reading level prediction method and system based on picture book reading records. BACKGROUND

[0002] The existing children's reading level prediction related technology mainly focuses on the collection and analysis of picture book reading behavior data, and various prediction schemes based on reading records have been formed. In terms of data collection, the commonly used methods include collecting explicit data such as reading time, reading frequency, picture book completion rate, and text recognition accuracy through intelligent picture books, reading APPs, or simple sensors. Some technologies supplement the voice interaction records in the parent-child reading process, such as the number of parent questions and the number of child responses. In terms of modeling and prediction, traditional machine learning algorithms such as logistic regression and support vector machine are used to output the grade division of children's reading level (such as primary, intermediate, and advanced) through single-dimensional data or simple feature combinations. In terms of system architecture, it usually includes data collection terminals, data transmission modules, background analysis modules, and result display modules, which can be applied to family parent-child reading, kindergarten reading teaching, and other scenarios, providing a basic reference for parents and educators on children's reading situation.

[0003] The existing technology has obvious limitations in practical application, and it is difficult to meet the needs of accurate prediction and effective improvement of children's reading level. The data collection dimension is single, only focusing on explicit result data, lacking the capture of deep behavior data such as picture book-children physical interaction details, children's micro-expression and cognitive feedback, and parent guidance quality, resulting in the inability to truly reflect the actual cognitive state of children during reading. The modeling logic lacks fine design, does not consider the cognitive development differences of children of different ages, and does not solve the problem of insufficient amount of error behavior samples, the prediction results are mostly general grades, without subdivision ability dimension and future trend prediction, and the practicality is limited. At the same time, the existing technology generally lacks a closed-loop design of "prediction-intervention-re-prediction", can only output reading level results, cannot provide targeted ability improvement schemes and parent guidance suggestions, and is difficult to solve the core pain point of "companionship ≠ effectiveness"; some schemes do not pay attention to children's data privacy protection, there is a risk of sensitive data leakage, and the collection terminal needs to modify the picture book or operate complexly, which lacks adaptability and landing feasibility. SUMMARY

[0004] In order to overcome the above-mentioned defects of the existing technology, the present application provides a children's reading level prediction method and system based on picture book reading records, which solves the problems of narrow data dimension, extensive modeling, lack of closed-loop intervention, insufficient privacy protection, and scene adaptability in the prior art.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0006] The method and system for predicting children's reading level based on picture book reading records, comprising the following steps:

[0007] S1. Collecting ecological behavior data in the process of parent-child reading, the ecological behavior data at least including picture book-child interaction data, parent-child interaction data, and child-reading content feedback data;

[0008] S2. Preprocessing the ecological behavior data, extracting multiple features and standardizing coding, the features at least including key illustration touch rate of picture book-child interaction dimension, page turning force entropy value, confusion expression proportion, open-ended question proportion of parent-child interaction dimension, expansion response rate, and guiding action frequency, and picture book-child interaction dimension, and child-reading content feedback dimension, including doodling content matching degree, logic error frequency, and character identification accuracy;

[0009] S3. Constructing a cross-subject feature association model based on attention mechanism and graph neural network, taking the standardized coded features as nodes, and the linkage relationship between features as edges, to generate an ecological behavior association graph;

[0010] S4. Using a hybrid model of graph neural network and long short-term memory network, inputting the feature vectors of the ecological behavior association graph and the longitudinal time series data of a preset time length (such as 3 months), and outputting the current value and the predicted value of a future preset period (such as 1 month) of children's reading level, the reading level at least including picture-text integration ability, narrative logic ability, metaphor understanding ability, empathy cognitive ability, and content transfer and creation ability;

[0011] S5. Based on the predicted value, generating ability improvement tasks for children and guiding optimization schemes for parents, and updating model parameters according to the ecological behavior data after intervention, forming a closed loop of prediction-intervention-re-prediction.

[0012] Preferably, the children's reading level prediction system based on picture book reading records comprises an ecological behavior perception layer, a feature association modeling layer, a dynamic prediction layer, a bidirectional intervention layer, and a privacy protection module;

[0013] The ecological behavior perception layer is used to collect ecological behavior data in the process of parent-child reading;

[0014] The feature association modeling layer is used to process the ecological behavior data and construct an ecological behavior association graph;

[0015] The dynamic prediction layer is used to output the current value and the predicted value of children's reading level;

[0016] The bidirectional intervention layer is used to generate ability improvement tasks and guiding optimization schemes;

[0017] The privacy protection module is used to protect data security.

[0018] Preferably, in step S1, the picture book-child interaction data is collected through a flexible smart reading mat and a miniature perception glasses; the flexible smart reading mat integrates a pressure sensor and a capacitive touch array for collecting page turning force, page pressing duration, fingertip touch position and touch frequency; the miniature perception glasses have an infrared micro camera and an eye movement tracking module built-in for extracting children's micro-expression features and graphic-text area gaze heat maps, while a directional sound pickup is used to collect children's unconscious sound data; the collection frequency band of the directional sound pickup is 200-3000 Hz, only the voiceprint feature parameters of children's unconscious sound are extracted, and no original audio data is stored.

[0019] Preferably, in step S1, the parent-child interaction data is collected through a dual-channel Bluetooth microphone and a parent guiding action sensor; the dual-channel Bluetooth microphone distinguishes parent and child voices through voiceprint recognition, records turn-taking timing and converts offline to text, and automatically marks question types, response lengths and expansion methods; the parent guiding action sensor is used to identify the parent's pointing to the illustration gesture and page turning assistance behavior, and is associated with the child's interaction reaction data.

[0020] Preferably, in step S1, the child-reading content feedback data is collected through a smart scribble board and a picture book recognition sticker; the smart scribble board has an image scanning module built-in for extracting the line smoothness of the scribble and the matching degree of the content with the picture book characters or plot, and generating a feature vector locally; the picture book recognition sticker includes an RFID page number label and an image recognition sticker, which are used to record the page turning order to identify logical error and capture the accuracy of children's character identification to identify cognitive error, respectively.

[0021] Preferably, in step S2, for the error behavior data in the child-reading content feedback data, an error data augmentation algorithm is used for sample expansion, which solves the problem of insufficient error sample quantity by simulating the characteristic variation of similar cognitive errors.

[0022] Preferably, in step S4, the mixed model adjusts the feature weight according to the child's age, for children aged 3-4, the feature weight of the picture book-child interaction dimension is set to 0.35; for children aged 5-6, the feature weight of the child-reading content feedback dimension is increased to 0.3.

[0023] Preferably, the ecological behavior perception layer at least includes a picture book-child interaction perception unit, a parent-child interaction perception unit and a child-reading content feedback perception unit; the picture book-child interaction perception unit includes a flexible smart reading mat and a miniature perception glasses, the flexible smart reading mat is laid under the picture book to adapt to ordinary picture books, and the miniature perception glasses are lightweight and only extract children's micro-expression feature values.

[0024] Preferably, the feature association modeling layer includes a feature extraction module and an association graph construction module; the feature extraction module decomposes the ecological behavior data and outputs standardized feature vectors; the association graph construction module quantifies the linkage relationship of different features through an attention mechanism to generate an ecological behavior association graph; the privacy protection module uses local edge computing to extract features from the original data, achieves model parameter sharing through federated learning, and splits the feature data into at least 3 fragments for fragmented storage.

[0025] Preferably, the two-way intervention layer includes a child-side interaction module and a parent-side guidance module; the child-side interaction module transforms the reading level dimension into a challenge game, with tasks directly related to ecological behavioral characteristics; the parent-side guidance module generates a parent-child reading guidance manual, and outputs actionable guidance language and suggestions based on specific ecological behaviors.

[0026] The technical effects and advantages of the children's reading level prediction method and system based on picture book reading records of this invention are as follows:

[0027] 1. This invention constructs a multimodal ecological behavioral data system by collecting picture book-child interaction data, parent-child interaction data, and child-reading content feedback data. It comprehensively captures the physical interaction details, language interaction quality, and children's cognitive feedback during parent-child reading, avoiding the one-sidedness caused by relying solely on explicit data such as reading time and frequency, and truly reflecting the actual cognitive state and parent-child interaction effect during the reading process.

[0028] 2. This invention constructs a cross-subject feature association model based on attention mechanism and graph neural network, quantifies the linkage relationship of different behavioral features, and combines graph neural network and long short-term memory network hybrid model to achieve deep integration of feature dimensions and mining of temporal patterns; at the same time, through age-adapted weight adjustment and error data augmentation algorithm optimization, it adapts to the cognitive development characteristics of children of different ages, solves the problem of insufficient error sample size, and significantly improves the accuracy of reading level prediction and model generalization ability.

[0029] 3. This invention can output scores for detailed reading level dimensions such as text-image integration ability and narrative logic ability, as well as predicted values ​​for future preset periods. It breaks through the limitation of traditional methods that can only output general levels, providing a clear basis for accurately identifying children's reading weaknesses and providing scientific guidance for subsequent personalized intervention.

[0030] 4. This invention constructs a closed-loop system of prediction-intervention-re-prediction, generates appropriate ability improvement tasks for children's reading weaknesses, outputs operable parent guidance optimization plans based on the characteristics of parent-child interaction, and dynamically updates model parameters based on behavioral data after intervention, so as to achieve two-way promotion of children's ability improvement and parent guidance optimization, significantly improving the effectiveness of parent-child reading.

[0031] 5. This invention avoids the storage and transmission of sensitive data such as original images and audio by extracting features through local edge computing, storing data in fragmented form, and sharing model parameters through federated learning. Without affecting model performance, it completely solves the risk of privacy leakage during the collection of children's data and provides security for its promotion and application in family scenarios.

[0032] 6. This invention uses lightweight data collection terminals such as flexible smart reading pads and picture book recognition stickers, which can be directly adapted to ordinary picture books without additional modifications, thus lowering the barrier to entry. At the same time, it takes into account the usage habits of young children and the ease of operation for parents, and can be widely used in parent-child reading scenarios such as families and kindergartens, with significant practical application value and promotion prospects. Attached Figure Description

[0033] Fig. 1 This is a system flowchart of the method and system for predicting children's reading level based on picture book reading records proposed in this invention;

[0034] Fig. 2 This is a flowchart of the method and system for predicting children's reading level based on picture book reading records proposed in this invention. Detailed Implementation

[0035] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0036] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include," "contain," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "includes..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0037] Example 1

[0038] refer to Figs. 1-2This embodiment provides a method and system for predicting children's reading levels based on picture book reading records, for basic implementation throughout the entire process. Specific implementation content includes:

[0039] Purpose of implementation:

[0040] This study verifies the collaborative effect of the entire ecological behavior data collection, feature processing, correlation modeling, dynamic prediction, and two-way intervention chain, and fully presents the practical application scenarios of the methods and systems.

[0041] System Architecture:

[0042] It includes an ecological behavior perception layer, a feature association modeling layer, a dynamic prediction layer, a two-way intervention layer, and a privacy protection module. The specific configurations of each layer are as follows:

[0043] Ecological Behavior Perception Layer: Used to collect ecological behavior data during parent-child reading, including:

[0044] Flexible smart reading pad: A flexible thin-film pressure sensor array with a thickness of 0.5mm (accuracy 0.1N), paired with a 256×256 resolution capacitive touch array, with a sampling rate of 10Hz, is laid under ordinary picture books and is compatible with A4 and 24-page picture books;

[0045] Miniature sensing glasses: weighing ≤30g, with built-in OV5640 infrared micro camera (15fps) and SMI low-power eye-tracking module (error ±0.3°), and equipped with a directional microphone (model MAX9814, sampling frequency band 200-3000Hz).

[0046] Dual-channel Bluetooth microphone: uses CSR8675 chip, voiceprint recognition accuracy ≥95%, supports offline speech-to-text based on WhisperTiny model;

[0047] Parental guidance motion sensor: MPU6050 posture sensor (clamped in parent's sleeve), sampling rate 50Hz, pointing gesture recognition response time ≤100ms;

[0048] Smart drawing board: 10-inch LCD erasable screen, built-in STM32H743 processor and OV2640 image scanning module (scanning resolution 300dpi).

[0049] Picture book identification sticker: Image recognition sticker containing 13.56MHz RFID page number tags and printed QR codes (associated with picture book character information).

[0050] Feature association modeling layer: used to process ecological behavior data and construct ecological behavior association graph. It adopts NVIDIA Jetson Nano edge computing module, extracts features through Python OpenCV and librosa library, and implements attention mechanism and graph neural network (2-layer GCN, 128 hidden units) based on PyTorch framework.

[0051] Dynamic prediction layer: Used to output the current and predicted values ​​of children's reading level. It adopts a hybrid model of graph neural network and long short-term memory network (built on PyTorch). The graph neural network outputs 64 feature dimensions, and the long short-term memory network has 3 hidden layers (256 units per layer). The training iterations are 50 rounds, and the learning rate is 0.001.

[0052] Two-way intervention layer: used to generate ability enhancement tasks and guidance optimization schemes. The children's adventure game is developed based on Unity, and the parent's guide is generated through the web (supports PDF export).

[0053] Privacy protection module: Used to protect data security. The original image / audio is directly converted into feature vectors (17-dimensional micro-expression features and 32-dimensional voiceprint features) in the edge computing module. The data is split into 3 segments and stored on the local SD card. FedAvg algorithm is used to implement federated learning, and the parameter update cycle is 7 days.

[0054] Implementation steps:

[0055] S1. Collecting Ecological Behavioral Data: 30 children aged 4-5 and their parents were selected to read five picture books together, including "Rainbow Fish" and "Little Bear Gets a Tooth Pulled" (each book lasting 15-20 minutes). Three types of data were collected simultaneously:

[0056] Picture book-child interaction data: Flexible reading pads record page-turning force (mean 0.3-0.8N), fingertip touch location (mean touch rate of key illustrations 45%), miniature sensory glasses extract the percentage of confused expressions (mean 22%), and the percentage of time spent looking at text areas (mean 18%).

[0057] Parent-child interaction data: Dual-channel microphones recorded turn-taking, with an average of 38% of questions being open-ended and an average of 29% being extended responses;

[0058] Children's reading content feedback data: The smart drawing board collects doodle content (average role matching rate 63%), and the picture book recognition sticker records the frequency of logical errors (average 1.2 times / book).

[0059] S2. Preprocessing and Feature Encoding: The collected data is standardized by Min-Max (mapped to the 0-1 range), and features such as key illustration touch rate, page turning force entropy value, and confused expression ratio in the picture book-child interaction dimension; open-ended question ratio, extended response rate, and guiding action frequency in the parent-child interaction dimension; and doodle content matching degree, logical error frequency, and character identification accuracy in the child-reading content feedback dimension are extracted to generate a 64-dimensional feature vector.

[0060] S3. Construct a cross-subject feature association model: Quantify the feature linkage relationship through the attention mechanism (such as the linkage weight of key illustration touch rate + open question ratio of 0.32), use standardized features as nodes and the linkage relationship between features as edges to generate an ecological behavior association graph;

[0061] S4. Dynamic Prediction: Input 3 months of longitudinal time-series data (reading records once a week), and the hybrid model outputs 5 reading level scores (out of 10): text and image integration ability 7.2 points, narrative logic ability 6.8 points, metaphor comprehension ability 5.9 points, empathy cognition ability 7.5 points, content transfer and creation ability 6.3 points. The predicted value for the next month is expected to improve by 0.8-1.2 points compared to the current value.

[0062] S5. Forming a prediction-intervention-reprevention closed loop: Based on the predicted values, a role-matching challenge game is pushed to the child's end, and a guidance manual containing open-ended questions such as why the rainbow fish shares its scales is generated on the parent's end; data is collected again after one week of intervention, the model parameters are updated, and the closed loop is completed.

[0063] Implementation results:

[0064] The average accuracy rate of the five reading level predictions was 89%, the adoption rate of the parent-guided optimization plan was 76%, and the completion rate of the children's intervention tasks was 83%. The reading level scores improved by 15%-20% after the intervention, which verified the effectiveness of the whole-link program.

[0065] Example 2

[0066] This embodiment provides a method and system for predicting children's reading levels based on picture book reading records, used for age-appropriate weight optimization. Specific implementation details include:

[0067] Purpose of implementation:

[0068] The study validates the effectiveness of the hybrid model in adjusting feature weights based on children's age, adapting it to the cognitive development characteristics of children at different ages.

[0069] System Architecture:

[0070] Consistent with Example 1, only the feature weight configuration of the dynamic prediction layer is different:

[0071] For children aged 3-4: the weight of the child interaction dimension of picture books is 0.35, the weight of the parent-child interaction dimension is 0.3, and the weight of the child-reading content feedback dimension is 0.25;

[0072] For children aged 5-6: the weight of the child-reading content feedback dimension is 0.3, the weight of the picture book-child interaction dimension is 0.3, and the weight of the parent-child interaction dimension is 0.25.

[0073] Implementation steps:

[0074] S1. Data collection by groups: 20 children aged 3-4 years (Group 1) and 20 children aged 5-6 years (Group 2) were selected to read the same set of picture books with varying cognitive difficulty (3 books with low cognitive load + 2 books with high cognitive load) together, and ecological behavior data were collected.

[0075] S2-S4. Feature processing, association modeling and prediction: The process is the same as in Example 1, except that the dynamic prediction layer adjusts the feature weights according to age.

[0076] S5. Intervention and Verification: Implement interventions based on the prediction results and compare the prediction accuracy before and after weight adjustment.

[0077] Implementation results:

[0078] Group 1 (3-4 years old): Prediction accuracy rate 87%, prediction deviation for text-image integration ability and empathic cognition ability ≤0.5 points, a 40% reduction in deviation compared to the unadjusted weights;

[0079] Group 2 (5-6 years old): Prediction accuracy rate 91%, prediction deviation of narrative logic ability, content transfer and creative ability ≤0.4 points, a 35% reduction in deviation compared to the unadjusted weights;

[0080] The results show that age-adaptation weights can optimize prediction accuracy in a targeted manner, demonstrating the age adaptability of the scheme.

[0081] Example 3

[0082] This embodiment provides a method and system for predicting children's reading levels based on picture book reading records, used for implementing error data augmentation algorithms. Specific implementation details include:

[0083] Purpose of implementation:

[0084] To verify the effectiveness of error data augmentation algorithms in addressing the problem of insufficient sample size for cognitive errors and to improve the model's generalization ability.

[0085] System Architecture:

[0086] Consistent with Example 1, the feature association modeling layer adds an error data enhancement module to expand the sample of erroneous behaviors in the child-reading content feedback data.

[0087] Implementation steps:

[0088] S1. Collect initial error data: Obtain 86 pieces of children's cognitive error data (such as role misidentification, prop function confusion);

[0089] S2. Error Data Enhancement:

[0090] Classification: The erroneous data is divided into three categories: character recognition error, prop function error, and plot logic error. Five core features are extracted for each category (such as the confusion degree of appearance features and the similarity of color features in character errors).

[0091] Extension: Gaussian perturbation (standard deviation 0.05) is applied to the core features of each type of error to generate derivative samples of the same type of error;

[0092] Screening: Remove derivative samples with a duplication rate of ≥80% with the original samples, and finally expand the number of erroneous samples to 430 (5 times).

[0093] S3-S5. Feature modeling, prediction and intervention: The model is trained based on the expanded samples, and the rest of the process is the same as in Example 1.

[0094] Implementation results:

[0095] The model improved the accuracy of identifying cognitive errors from 68% to 85%, and the accuracy of predicting narrative logic ability based on error data from 75% to 88%, verifying the effectiveness of the error data augmentation algorithm.

[0096] Example 4

[0097] This embodiment provides a method and system for predicting children's reading levels based on picture book reading records, used to enhance the implementation of the privacy protection module. Specific implementation details include:

[0098] Purpose of implementation:

[0099] Verify that the privacy protection module ensures data security without affecting model performance.

[0100] System Architecture:

[0101] The privacy protection module is the key focus, while the remaining layers are the same as in Example 1:

[0102] The privacy protection module employs local edge computing (RK3588 edge computing board), fragmented data storage, and federated learning technology.

[0103] Local edge computing: The original image (micro-expression video) is immediately deleted after 17 feature points are extracted by MTCNN, and the audio data is converted into text, retaining only 128-dimensional word vectors, while the original audio is automatically destroyed;

[0104] Data fragmented storage: The 64-dimensional feature vector is split into 3 segments according to the picture book-child interaction features (24-dimensional) + parent-child interaction features (20-dimensional) + child-reading content feedback features (20-dimensional), which are stored locally on the device, on the parent's mobile phone, and on the cloud encrypted server, and are only temporarily spliced ​​together with a key during prediction;

[0105] Federated learning: 10 families act as participating nodes. After training the model locally, they only upload parameters (≤500KB) and use a federated averaging algorithm to aggregate the parameters.

[0106] Implementation steps:

[0107] S1. Data Collection and Privacy Processing: Collect ecological behavior data from 10 families and process the raw data simultaneously through a privacy protection module;

[0108] S2-S4. Feature Modeling and Prediction: A model is built based on fragmented storage of feature data and trained using federated learning. The remaining process is the same as in Example 1.

[0109] S5. Security and Performance Verification: Detect data leakage risks and compare the accuracy of models trained in federated and centralized training.

[0110] Implementation results:

[0111] Data security: No original image or audio data was detected being transmitted outside the system; cracking fragmented feature segments individually is meaningless.

[0112] Model performance: The prediction accuracy after federated training is 89.5%, which is very close to that of centralized training (90%), proving that the privacy protection scheme does not affect the technical performance.

[0113] Example 5

[0114] This embodiment provides a method and system for predicting children's reading levels based on picture book reading records, used for the implementation of a two-way intervention closed-loop optimization. Specific implementation details include:

[0115] Purpose of implementation:

[0116] To verify the practical application effect of the prediction-intervention-repreneurial closed loop, and to refine the intervention plans for children and parents.

[0117] System Architecture:

[0118] The focus of optimization is on the bidirectional intervention layer; the remaining layers are the same as in Example 1.

[0119] The two-way intervention layer includes a child-side interaction module and a parent-side guidance module:

[0120] The interactive module for children transforms the five reading levels into five challenge roles: Explorer, Storyteller, Observer, Empathizer, and Creator. Each role corresponds to three tasks (e.g., the Storyteller task: collage picture book illustrations in plot order).

[0121] Parent-side guidance module: Generates customized guidance manuals based on predicted ability weaknesses (e.g., if metaphor comprehension is weak, provide explanations of metaphorical language in picture books using real-life examples; if logical errors are frequent, provide guidance actions by asking questions in chronological order).

[0122] Implementation steps:

[0123] S1-S4. Data collection, feature processing, modeling and prediction: 15 children were selected, and the average score of the five abilities was predicted to be 6.2 points in the first prediction.

[0124] S5. Closed-loop intervention:

[0125] Intervention implementation: Continue for 2 weeks, with children spending an average of 15 minutes per day completing the challenge tasks, and parents using guiding language 3-5 times per day on average;

[0126] Re-prediction: After intervention, data is collected again and predictions are made to assess the effectiveness of the capability improvement.

[0127] Implementation results:

[0128] After the intervention, the average score of the five abilities improved to 7.5, with the narrative logic ability and metaphor comprehension ability showing the most significant improvement (22%), verifying the effectiveness of the two-way intervention loop.

[0129] Comparative Example 1

[0130] This comparison provides a traditional single-dimensional prediction scheme, including the following:

[0131] Purpose of implementation:

[0132] The technical advantages of this invention are highlighted by comparison with existing technologies.

[0133] Solution configuration:

[0134] Data collection: Only three single-dimensional data points were collected: reading time, reading frequency, and text recognition accuracy. No multimodal data on ecological behavior were collected.

[0135] Modeling and prediction: The logistic regression model is used, which only outputs the reading level level (level 1-3) without specific ability dimension scores or future predictions;

[0136] Intervention method: No targeted intervention, only recommendations of picture books that match the level.

[0137] Implementation steps:

[0138] Thirty children, the same as in Example 1, were selected and the above-described scheme was used for prediction and intervention. The reading time and duration were the same as in Example 1.

[0139] Implementation results:

[0140] Compared with Example 1, the traditional approach had a core competency dimension prediction accuracy of 65% (Example 1 had 89%), a competency prediction deviation of ≥1.5 points in the next month (Example 1 had ≤0.6 points), a competency improvement of only 5%-8% after intervention (Example 1 had 15%-20%), and a parental adoption rate of 32% (Example 1 had 76%).

[0141] Traditional solutions rely on single explicit data, resulting in low prediction accuracy, lack of targeted intervention, and inability to address the pain point that companionship does not necessarily equate to effectiveness. In contrast, this invention significantly improves prediction accuracy and application value through multimodal ecological behavior data fusion, cross-subject correlation modeling, and a two-way intervention closed loop.

[0142] Compared with Examples 1-5 and Comparative Example 1, the core differences between Examples 1-5 and Comparative Example 1 are concentrated in five key aspects: data collection dimensions, modeling logic, predictive ability, intervention effect, and privacy protection. The comparison results fully highlight the technical advantages and application value of the present invention.

[0143] In terms of data collection, Comparative Example 1 relies solely on three single explicit data points: reading time, frequency, and text recognition accuracy, lacking the capture of interactive details and children's genuine cognitive feedback during parent-child reading. In contrast, Examples 1-5 construct a multimodal ecological behavioral data collection system, integrating picture book-child interaction data (page turning force, touch position, micro-expressions, etc.), parent-child interaction data (talking type, guiding actions, etc.), and child-reading content feedback data (doodle matching degree, incorrect behavior, etc.), achieving coverage from single result data to full-process behavioral data, providing rich data support for accurate prediction.

[0144] In terms of modeling logic, Comparative Example 1 uses a simple logistic regression model, which can only output a general reading level level without any detailed ability dimensions or future trend predictions. Examples 1-5 construct cross-subject feature association models based on attention mechanisms and graph neural networks. They combine graph neural networks and long short-term memory networks into a hybrid model. Furthermore, through age-adaptive weight adjustment in Example 2 and error data augmentation algorithm optimization in Example 3, they achieve feature linkage relationship quantification and model generalization ability improvement, and can accurately output scores for five detailed abilities such as text-image integration and narrative logic, as well as prediction values ​​for the next month.

[0145] In terms of prediction performance, Comparative Example 1 showed a core competency prediction accuracy of only 65% ​​and a future competency prediction deviation of ≥1.5 points, failing to reflect children's true reading level. Example 1 achieved a full-process prediction accuracy of 89%. Example 2 reduced the prediction deviation of key competencies for 3-4 year olds to ≤0.5 points and for 5-6 year olds to ≤0.4 points through age adaptation. Example 3 improved the accuracy of cognitive error recognition from 68% to 85% by using error data augmentation. The prediction accuracy and subdivision dimensions far exceeded those of the comparative example.

[0146] In terms of intervention effectiveness, Comparative Example 1 had no targeted intervention plan, only recommended general picture books, with a parent adoption rate of 32% and children's ability improvement of only 5%-8%. Examples 1-5 constructed a prediction-intervention-re-prediction closed loop. Example 5, through a children's challenge game and a customized guidance manual for parents, achieved a parent guidance adoption rate of 76%, a children's task completion rate of 83%, and an ability improvement of 15%-20% after intervention. Among them, the improvement in narrative logic ability and metaphor comprehension ability was the most significant, effectively solving the pain point that traditional solutions do not necessarily guarantee effectiveness.

[0147] In terms of privacy protection, Comparative Example 1 did not involve data security design, and there was a risk of leakage of original data; Example 4, through local edge computing, data fragmentation storage and federated learning technology, achieved the instant destruction of original image and audio data and the secure storage of feature data without affecting model performance (prediction accuracy of 89.5% is very similar to that of centralized training), filling the privacy protection gap of traditional solutions.

[0148] In summary, Comparative Example 1, due to its one-sided data, simplistic modeling, and weak intervention, cannot meet the needs of accurate prediction and effective improvement of children's reading level in parent-child reading scenarios. In contrast, this invention comprehensively improves prediction accuracy, intervention effect, and application security through multimodal data fusion, refined modeling optimization, closed-loop intervention design, and enhanced privacy protection, demonstrating significant technological innovation and practical application value.

[0149] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of protection of the claims.

[0150] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting children's reading level based on picture book reading records, characterized in that, Includes the following steps: S1. Collect ecological behavior data during parent-child reading, including at least picture book-child interaction data, parent-child interaction data, and child-reading content feedback data; S2. Preprocess the ecological behavior data, extract multiple features and standardize and encode them. The features include at least the key illustration touch rate, page turning force entropy value, and confused expression ratio in the picture book-child interaction dimension; the open-ended question ratio, extended response rate, and guiding action frequency in the parent-child interaction dimension; and the graffiti content matching degree, logical error frequency, and character identification accuracy in the child-reading content feedback dimension. S3. Construct a cross-subject feature association model based on attention mechanism and graph neural network, using the standardized encoded features as nodes and the linkage relationship between features as edges to generate an ecological behavior association graph; S4. Using a hybrid model of graph neural network and long short-term memory network, the feature vector of the ecological behavior association graph and longitudinal time series data of a preset duration are input, and the current value of the child's reading level and the predicted value of the future preset period are output. The reading level includes at least the ability to integrate text and graphics, narrative logic, metaphor understanding, empathic cognition, content transfer and creation. S5. Based on the predicted values, generate ability enhancement tasks for children and guidance optimization plans for parents, and update model parameters according to the ecological behavior data after intervention, forming a closed loop of prediction-intervention-re-prediction.

2. The method for predicting children's reading level based on picture book reading records as described in claim 1, characterized in that, In step S1, the picture book-child interaction data is collected through a flexible smart reading pad and miniature sensing glasses. The flexible smart reading pad integrates a pressure sensor and a capacitive touch array to collect page-turning force, page pressing duration, fingertip touch position, and touch frequency. The miniature sensing glasses have a built-in infrared micro-camera and eye-tracking module to extract children's micro-expression features and gaze heatmaps of text and image areas. They are also used with a directional microphone to collect children's unconscious vocalization data. The directional microphone has a sampling frequency band of 200-3000Hz and only extracts the voiceprint feature parameters of children's unconscious vocalizations, without storing the original audio data.

3. The method for predicting children's reading level based on picture book reading records as described in claim 1, characterized in that, In step S1, the parent-child interaction data is collected through a dual-channel Bluetooth microphone and a parent-guided motion sensor. The dual-channel Bluetooth microphone distinguishes between the parent's and child's voices through voiceprint recognition, records the turn-taking sequence and converts it to text offline, and automatically marks the question type, response length and extension method. The parent-guided motion sensor is used to recognize the parent's gestures pointing to illustrations and page-turning assistance behavior, and associates them with the child's interactive response data.

4. The method for predicting children's reading level based on picture book reading records as described in claim 1, characterized in that, In step S1, the child-reading content feedback data is collected through a smart drawing board and picture book recognition stickers. The smart drawing board has a built-in image scanning module, which is used to extract the smoothness of the lines of the drawing and the matching degree between the content and the picture book characters or plots, and generate feature vectors locally. The picture book recognition stickers include RFID page number tags and image recognition stickers, which are used to record the page turning order to identify logical errors and capture the accuracy of children's character identification to identify cognitive errors, respectively.

5. The method for predicting children's reading level based on picture book reading records as described in claim 1, characterized in that, In step S2, the error behavior data in the child-reading content feedback data is expanded using an error data augmentation algorithm. The error data augmentation algorithm solves the problem of insufficient error sample size by simulating the feature variation of similar cognitive errors.

6. The method for predicting children's reading level based on picture book reading records as described in claim 1, characterized in that, In step S4, the hybrid model adjusts the feature weights according to the child's age. For children aged 3-4, the feature weight of the picture book-child interaction dimension is set to 0.35; for children aged 5-6, the feature weight of the child-reading content feedback dimension is increased to 0.

3.

7. A system for implementing the children's reading level prediction method based on picture book reading records as described in any one of claims 1-6, characterized in that, It includes an ecological behavior perception layer, a feature association modeling layer, a dynamic prediction layer, a two-way intervention layer, and a privacy protection module; The ecological behavior perception layer is used to collect ecological behavior data during parent-child reading. The feature association modeling layer is used to process the ecological behavior data and construct an ecological behavior association graph; The dynamic prediction layer is used to output the current and predicted values ​​of children's reading levels; The two-way intervention layer is used to generate capability enhancement tasks and guide optimization solutions; The privacy protection module is used to protect data security.

8. The children's reading level prediction system based on picture book reading records as described in claim 7, characterized in that, The ecological behavior perception layer includes at least a picture book-child interaction perception unit, a parent-child interaction perception unit, and a child-reading content feedback perception unit; the picture book-child interaction perception unit includes a flexible smart reading pad and miniature sensing glasses. The flexible smart reading pad is placed under the picture book to fit ordinary picture books, and the miniature sensing glasses are lightweight and only extract children's micro-expression feature values.

9. The children's reading level prediction system based on picture book reading records as described in claim 7, characterized in that, The feature association modeling layer includes a feature extraction module and an association graph construction module. The feature extraction module decomposes the ecological behavior data and outputs standardized feature vectors. The association graph construction module quantifies the linkage relationship of different features through an attention mechanism to generate an ecological behavior association graph. The privacy protection module uses local edge computing to extract features from the original data, achieves model parameter sharing through federated learning, and splits the feature data into at least three fragments for fragmented storage.

10. The children's reading level prediction system based on picture book reading records as described in claim 7, characterized in that, The two-way intervention layer includes a child-side interactive module and a parent-side guidance module; the child-side interactive module transforms the reading level dimension into a challenge game, with tasks directly related to ecological behavioral characteristics. The parent-side guidance module generates a parent-child reading guidance manual, and outputs actionable guidance scripts and suggestions based on specific ecological behaviors.

Images