A child reading interest AI identification and personalized recommendation method

Abstract

This invention discloses an AI-based method for identifying and personally recommending children's reading interests, specifically relating to the fields of data processing and intelligent recommendation technology. The method includes: collecting multimodal data from children's reading and performing spatiotemporal alignment preprocessing; using a deep learning model to perform multimodal fusion identification of emotional pleasure, focus index, and interest focus; constructing a knowledge graph of reading materials and a graph of children's developmental stages; performing similarity retrieval based on interest focus; calculating the "zone of proximal development deviation value" to filter a preliminary recommendation list; using interactive trial reading verification; and generating a final personalized recommendation list. This invention accurately captures children's latent interests through multimodal emotion computing, combines dual knowledge graphs to match cognitive development stages, and forms a closed loop through interactive verification. It solves the technical problem that traditional recommendation methods are unable to adapt to the dynamic reading needs of young children, significantly improving the accuracy and personalization of recommendations.

Description

Technical Field

[0001] This invention relates to the field of data processing and intelligent recommendation technology, and more specifically, to a method for AI-based identification and personalized recommendation of children's reading interests. Background Technology

[0002] Children's reading is a crucial element in promoting children's cognitive development, language skills, emotional intelligence, and value formation. Choosing appropriate reading materials for children has long been a challenge for parents, educators, and the publishing industry.

[0003] Traditional methods of recommending children's books mainly rely on human experience, publisher classification, age group division, or simple user reading history. While these methods are effective for adult readers, they often lack personalization and dynamism for children whose cognitive and expressive abilities are still developing. They struggle to accurately capture children's ever-changing reading interests and cognitive development levels. Young children often cannot accurately describe their interests and preferences, and the interest information implied in their reading behavior is difficult to capture and quantify using traditional data collection methods. This results in recommendations that do not match children's actual needs, potentially leading to a decrease in reading interest or even aversion, which is detrimental to cultivating long-term reading habits. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for AI-based identification and personalized recommendation of children's reading interests, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for AI-based identification and personalized recommendation of children's reading interests, specifically including the following steps:

[0006] S1. Simultaneously collect multimodal data of children during the reading process using front-end sensing devices. The multimodal data includes facial video streams, eye movement trajectory data, voice interaction data, and body movement data. Preprocess the collected raw data and perform spatiotemporal alignment processing on the processed modal data to generate a multidimensional raw dataset of reading behavior.

[0007] The preprocessing includes extracting key points from the facial video stream to construct an emotion feature vector, mapping eye movement trajectory data to generate a reading scanning path map, extracting acoustic features from the voice interaction data, and identifying skeletal key points from the limb movement data to generate a posture feature vector.

[0008] S2. Perform deep learning model analysis and fusion on the multi-dimensional reading behavior raw dataset obtained in step S1, including real-time emotional pleasure calculation, reading focus index and interest focus determination, so as to accurately identify the child's real-time interest and focus status on the current reading content.

[0009] S3. Construct a dual knowledge graph, including a reading material knowledge graph and a child development stage graph. Using the interest focus extracted in step S2 as input, locate the candidate book set in the reading material knowledge graph through similarity retrieval. Use a mapping function to align the difficulty weight of the candidate books with the cognitive understanding threshold in the child development stage graph. Calculate the "zone of proximal development deviation value" of the candidate books relative to the child's current cognitive level. Select books associated with the interest focus identified in step S2 from the candidate book set to form a preliminary recommendation list.

[0010] S4. For the preliminary recommendation list selected in step S3, an interactive trial reading method is used to confirm the recommendations. The multimodal feedback data of children during the trial reading is monitored, and the matching degree weight between the recommended content and the children's instantaneous interests is calculated. The preliminary recommendation list is rearranged and debiased according to the matching degree weight to generate the final personalized reading recommendation list and push it to the reading device.

[0011] Preferably, as a preferred embodiment of the AI-based method for identifying and personally recommending children's reading interests according to the present invention, the preprocessing includes extracting key points from a facial video stream to construct an emotional feature vector, and further includes:

[0012] The collected facial videos Face detection and alignment are performed frame-by-frame. For each frame of the video stream, a deep learning-based face detection algorithm is used to identify all face regions in the image. For the identified child face region Geometric transformations are performed to achieve standardized alignment of the face, resulting in an aligned face image. ;

[0013] Aligned face image The data is input into a pre-trained deep learning keypoint detection model to locate the coordinates of facial keypoints, including specific points on the contours of the eyes, eyebrows, nose, mouth, and chin, forming a set of keypoint coordinates. Where N is the number of key points;

[0014] Facial motion units are extracted based on the located facial key point coordinates L, including the degree of mouth corner slant, eyelid opening and closing, and eyebrow tilt angle. For the mouth area, the vertical distance between the mouth corner key point 301 and the mouth center point 302 is calculated. Horizontal distance The ratio of the two values ​​gives the degree of upturn of the corners of the mouth. ; Extract key points for the eye contour area (200). The eyelid opening angle is obtained by calculating the ratio of the distance between pairs of points in the vertical direction to the distance between pairs of points in the horizontal direction, and is expressed as: The eyebrow tilt angle is determined by locating the inner corner point 101 and the outer corner point 102 of the eyebrow, and constructing the angle between the line connecting the two points and the horizontal baseline. The extracted facial motion unit parameters are combined to form a motion unit feature vector. ,in, Let the coordinates of the corner of the mouth be... The coordinates of the center point of the mouth are calculated from the midpoints of the upper and lower lips, and ε is a minimal constant to prevent division by zero. to The coordinates of six key points around the eye are shown. The larger the ratio, the greater the degree to which the eyes are open. Let these be the coordinates of the inner corner of the eyebrow. The coordinates of the outer corner of the eyebrow;

[0015] Extracted facial action unit feature vectors The input is fed into an emotion encoder, which is a deep neural network that performs feature transformation through fully connected layers and nonlinear activation functions to generate a three-dimensional emotion feature vector representing pleasure, confusion, and surprise, as shown below. Each dimension represents the intensity value of the child's current emotional state on the corresponding emotional dimension, normalized to... between.

[0016] Preferably, as a preferred embodiment of the AI-based method for identifying and personally recommending children's reading interests according to the present invention, the method further includes mapping eye-tracking data to generate a reading scanning path map by coordinate mapping:

[0017] The gaze point coordinates in the original eye movement trajectory data are mapped from the device coordinate system to the reading medium plane coordinate system to obtain the mapped gaze point sequence. Where M is the number of identified gaze points. Let J be the mapped coordinates of the j-th gaze point on the reading medium plane. For the corresponding gaze duration, the device coordinate system is the pixel coordinate system built into the eye-tracking device, and the reading medium plane coordinate system is the two-dimensional plane coordinate system of the physical reading material;

[0018] For the mapped gaze sequence Cluster analysis was performed to identify areas of interest that children focused on during reading. A density-based spatiotemporal clustering algorithm was used to group fixations based on spatial distance and temporal proximity. For each cluster... Calculate its center coordinates Total gaze duration and the total gaze duration identified Exceeding the preset threshold Clusters are used as effective regions of interest to form a set of regions of interest. ,in, For clusters The number of gaze points included, where K is the number of valid regions of interest identified;

[0019] Spatial location of the identified region of interest fixation sequence and fixation duration Visual encoding is performed to generate a time-series reading scan path map, where nodes in the path map correspond to the identified regions of interest. The spatial position of the node on the two-dimensional plane is determined based on the center coordinates of the region of interest, and the color depth of the path node represents the gaze duration. The longer the gaze duration, the darker the color of the node. The lines in the path diagram connect adjacent interest area nodes in sequence according to the gaze duration. The arrows on the lines indicate the direction of the child's gaze.

[0020] Preferably, as a preferred embodiment of the AI-based method for identifying and personally recommending children's reading interests according to the present invention, it includes extracting acoustic features from voice interaction data and identifying skeletal key points from limb movement data to generate a posture feature vector, further comprising:

[0021] The voice interaction data is pre-emphasized, framed, and windowed to extract acoustic feature parameters, including fundamental frequency, energy, formants, and Mel-frequency cepstral coefficients. An endpoint detection algorithm is used to remove silent segments and non-speech noise, retaining the interactive voice segments produced by the child's own speech. The extracted acoustic feature parameters are then input into a voice emotion recognition model, outputting a voice emotion tendency score vector. ,in, These respectively characterize the level of excitement, questioning, and calmness in children's speech;

[0022] The captured video stream of body movements Human detection was performed to identify children in video frames. For the detected child regions, a pose estimation algorithm was used to extract the 3D spatial coordinates of 15 key skeletal points, including the head, neck, left and right shoulders, left and right elbows, left and right wrists, left and right hips, left and right knees, and left and right ankles, forming a set of key point coordinates. ;

[0023] Based on the extracted 3D spatial coordinates of the skeletal key points, the relative angles and displacement velocities between each key point are calculated to generate a temporalized posture feature vector representing the amplitude of body swaying, gesture activity, and forward / backward tilting postures. The amplitude of body swaying is obtained by calculating the standard deviation of the displacement of the head key points, and is expressed as follows: The gesture activity level is obtained by calculating the degree of pixel change in the hand region, and is represented as... The forward and backward tilt postures are obtained by calculating the Z-axis coordinate difference between key points of the head and key points of the hips, and are expressed as follows: ,in, A positive value indicates leaning forward. A negative value indicates a backward tilt, and the absolute value indicates the degree of tilt. and The Z-axis coordinates of the left and right hip key points are shown below. For head key points in frame b and frame 6 The displacement distance between frames, This is the average displacement distance. and These are the velocity vectors of the left and right wrist key points, respectively, and B is the total number of frames within the statistical window.

[0024] The posture feature vector is then normalized to eliminate data bias caused by differences in different acquisition devices and individual child body shapes, resulting in a normalized posture feature vector. .

[0025] Preferably, as a preferred embodiment of the AI-based method for identifying and personally recommending children's reading interests according to the present invention, step S2 includes performing deep learning model analysis and fusion on the multi-dimensional raw dataset of reading behavior obtained in step S1, including real-time emotional pleasure calculation, reading focus index, and interest focus determination, in order to accurately identify the child's real-time interest and focus state on the current reading content. Specifically, this includes the following:

[0026] The instant emotional pleasure calculation involves fusing the emotional feature vector, voice emotional tendency score vector, and posture feature vector obtained in step S1 at the feature level to construct a multimodal emotional feature tensor. The fused multimodal emotional feature tensor is then input into a preset emotional calculation model to generate an index that varies with reading time. The changing instantaneous emotional pleasure curve, specifically the mathematical formula is as follows: ,in, As a time index of the reading process, the curve reflects the continuous trajectory of children's emotional changes during reading. , , These represent the time indexes respectively. The level of curiosity, confusion, and joy at any given moment;

[0027] The reading focus index is calculated by combining the posture feature vector and the eye-tracking path map to determine the percentage of time the gaze lingers in the current reading area. And the frequency of rereading, to obtain the reading focus index. ,in, The preset weighting coefficients, and These represent the normalized body sway amplitude and forward lean posture values, respectively. It is the readback frequency, which represents the number of readbacks per unit of time;

[0028] The interest focus determination is based on identifying the persistence residual of the gaze in the reading material image area. When the persistence residual exceeds a preset threshold... Furthermore, if the rate of change of the instantaneous emotional pleasure curve is greater than zero within this time window, the region is determined to be a valid focus of interest, and its spatial coordinates are extracted.

[0029] Preferably, as a preferred embodiment of the AI-based method for identifying and personally recommending children's reading interests according to the present invention, the method includes, in step S3, constructing a dual knowledge graph, including a book knowledge graph and a child development stage graph. The interest focus extracted in step S2 is used as input. A candidate book set in the book knowledge graph is located through similarity retrieval. A mapping function is used to align the difficulty weights of the candidate books with the cognitive comprehension thresholds in the child development stage graph. The "zone of proximal development deviation" of the candidate books relative to the child's current cognitive level is calculated. Books associated with the interest focus identified in step S2 are then selected from the candidate book set to form a preliminary recommendation list. Specifically, this includes the following:

[0030] Constructing a knowledge graph of reading materials composed of book difficulty and thematic features. , where the set of nodes This includes the physical book and its attributes, including the book's unique identifier and difficulty level. Theme classification; edge set Representing semantic relationships between books; constructing a child development stage map based on child cognitive theory and big data statistics. , where the set of nodes Includes age group entities, edge set This indicates a progressive relationship between stages of development;

[0031] Using the focus of interest extracted in S2 as the query input, the similarity search is performed on the knowledge graph of the book. Candidate book collection for positioning For each book in the candidate book set, extract its difficulty level. Based on the child's user ID Historical reading data and current age Using deep learning-based models Child Development Stage Map The optimal developmental stage for the child is determined. And obtain the corresponding cognitive understanding threshold vector. ;

[0032] Construct mapping function Book difficulty level Mapped to By using a cognitive threshold space of the same dimension, we can obtain the cognitive demand vector corresponding to the book. The "zone of proximal development deviation value" of candidate books relative to the child's current cognitive level is calculated as follows: And filter out those in the positive deviation range The selected books will be scored according to their similarity to the focus of interest. Sort the books in descending order and select the top Q books to form a preliminary recommendation list.

[0033] Preferably, as a preferred embodiment of the AI-based method for identifying and personally recommending children's reading interests according to the present invention, the method includes, in step S4, using an interactive trial reading method to confirm the recommendations based on the preliminary recommendation list selected in step S3, monitoring the multimodal feedback data of children during the trial reading, calculating the matching degree weight between the recommended content and the children's instantaneous interests; and rearranging and de-biasing the preliminary recommendation list according to the matching degree weight to generate the final personalized reading recommendation list and push it to the reading device. Specifically, this includes the following:

[0034] The preliminary recommendation list generated in step S3 is presented in the form of a preview summary. Front-end sensing devices are used to monitor the child's instantaneous multimodal response to each summary item, and a facial expression recognition model is used to calculate the instantaneous emotional pleasure score. Calculate the reading focus index on the preview page based on eye movement data. The system monitors children's response delay to voice-guided commands and their tendency to lean forward, generating an interaction initiative coefficient. The three indicators mentioned above are combined to calculate the normalized instantaneous interest matching weight for each candidate book. ,in, is the normalization factor, and a, g, and c are the contribution coefficients of different modes;

[0035] Extract the original similarity score of the i-th book in the initial recommendation list. Calculate the deviation between the original score and the real-time matching weight for each book. When the deviation value exceeds the preset threshold: If the initial recommendation list is found to deviate significantly from the child's current instantaneous interests, a debiasing and reordering algorithm is initiated. An exponential perturbation function is used to correct the original scores, and a final personalized score is calculated for each book. ,in, The disturbance intensity coefficient is... The sign function indicates the direction of the deviation: The rearranged scores are then probability-mapped to generate a final personalized reading list, which is then pushed to users.

[0036] On the other hand, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements the steps of a method for AI-based identification and personalized recommendation of children's reading interests as described above.

[0037] On the other hand, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements a method for AI-based identification and personalized recommendation of children's reading interests as described above.

[0038] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0039] By integrating multimodal behavioral data of children during the reading process, including facial expressions, eye movements, voice interaction, and body movements, this method can accurately and in real-time identify children's immediate emotional pleasure, reading focus index, and interest focus. Combined with a cognitive bias assessment mechanism constructed from dual knowledge graphs, it can dynamically match the difficulty of books with children's cognitive development stage, ensuring that the recommended content is always within the child's "zone of proximal development." Through interactive trial reading confirmation and intelligent bias control, this method achieves highly personalized and adaptive reading recommendations. It not only significantly improves the accuracy and effectiveness of recommendations but also stimulates children's reading interest, cultivates reading habits, comprehensively promotes children's cognitive and emotional development, and optimizes the overall reading experience and learning outcomes. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0041] Figure 1 This is a flowchart of the method of the present invention.

[0042] Figure 2 This diagram illustrates facial feature extraction. In the diagram, 101 represents the inner corner of the eyebrow, 102 the outer corner, 200 the eye contour area, P2 and P3 the key points of the upper eyelid, P5 and P6 the key points of the lower eyelid, P1 and P4 the key points of the inner and outer corners of the eyes, 301 the key point of the corner of the mouth, and 302 the center point of the mouth. The angle of the eyebrows and These are the horizontal and vertical distances between the corners of the mouth and the center point of the mouth, respectively.

[0043] Table 1 is a data recording table of the simulation experiment of this invention. Detailed Implementation

[0044] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0045] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0046] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0047] Example 1

[0048] This embodiment provides, for example Figure 1 The method for AI-based identification and personalized recommendation of children's reading interests, as shown, includes the following steps:

[0049] S1. Simultaneously collect multimodal data of children during the reading process using front-end sensing devices. The multimodal data includes facial video streams, eye movement trajectory data, voice interaction data, and body movement data. Preprocess the collected raw data and perform spatiotemporal alignment processing on the processed modal data to generate a multidimensional raw dataset of reading behavior.

[0050] The preprocessing includes extracting key points from the facial video stream to construct an emotion feature vector, mapping eye movement trajectory data to generate a reading scanning path map, extracting acoustic features from the voice interaction data, and identifying skeletal key points from the limb movement data to generate a posture feature vector.

[0051] S2. Perform deep learning model analysis and fusion on the multi-dimensional reading behavior raw dataset obtained in step S1, including real-time emotional pleasure calculation, reading focus index and interest focus determination, so as to accurately identify the child's real-time interest and focus status on the current reading content.

[0052] S3. Construct a dual knowledge graph, including a reading material knowledge graph and a child development stage graph. Using the interest focus extracted in step S2 as input, locate the candidate book set in the reading material knowledge graph through similarity retrieval. Use a mapping function to align the difficulty weight of the candidate books with the cognitive understanding threshold in the child development stage graph. Calculate the "zone of proximal development deviation value" of the candidate books relative to the child's current cognitive level. Select books associated with the interest focus identified in step S2 from the candidate book set to form a preliminary recommendation list.

[0053] S4. For the preliminary recommendation list selected in step S3, an interactive trial reading method is used to confirm the recommendations. The multimodal feedback data of children during the trial reading is monitored, and the matching degree weight between the recommended content and the children's instantaneous interests is calculated. The preliminary recommendation list is rearranged and debiased according to the matching degree weight to generate the final personalized reading recommendation list and push it to the reading device.

[0054] Preferably, in step S1, the preprocessing involves extracting key points from the facial video stream to construct an emotion feature vector, further including:

[0055] For the captured facial video stream Face detection and alignment are performed frame-by-frame. For each frame of the video stream, a deep learning-based face detection algorithm is used to identify all face regions in the image. For the identified child face region Geometric transformations are performed to achieve standardized alignment of the face, resulting in an aligned face image. ;

[0056] Aligned face image The data is input into a pre-trained deep learning keypoint detection model to locate the coordinates of facial keypoints, including specific points on the contours of the eyes, eyebrows, nose, mouth, and chin, forming a set of keypoint coordinates. Where N is the number of key points;

[0057] Facial motion units are extracted based on the located facial key point coordinates L, including the degree of mouth corner slant, eyelid opening and closing, and eyebrow tilt angle. For the mouth area, the vertical distance between the mouth corner key point 301 and the mouth center point 302 is calculated. Horizontal distance The ratio of the two values ​​gives the degree of upturn of the corners of the mouth. ; Extract key points for the eye contour area (200). The eyelid opening angle is obtained by calculating the ratio of the distance between pairs of points in the vertical direction to the distance between pairs of points in the horizontal direction, and is expressed as: The eyebrow tilt angle is determined by locating the inner corner point 101 and the outer corner point 102 of the eyebrow, and constructing the angle between the line connecting the two points and the horizontal baseline. The extracted facial motion unit parameters are combined to form a motion unit feature vector. ,in, Let the coordinates of the corner of the mouth be... The coordinates of the center point of the mouth are calculated from the midpoints of the upper and lower lips, and ε is a minimal constant to prevent division by zero. to The coordinates of six key points around the eye are shown. The larger the ratio, the greater the degree to which the eyes are open. Let these be the coordinates of the inner corner of the eyebrow. The coordinates of the outer corner of the eyebrow are as follows: Figure 2 As shown;

[0058] Extracted facial action unit feature vectors The input is fed into an emotion encoder, which is a deep neural network that performs feature transformation through fully connected layers and nonlinear activation functions to generate a three-dimensional emotion feature vector representing pleasure, confusion, and surprise, as shown below. Each dimension represents the intensity value of the child's current emotional state on the corresponding emotional dimension, normalized to... between;

[0059] The step of mapping eye-tracking data to generate a reading scan path map further includes:

[0060] The gaze point coordinates in the original eye movement trajectory data are mapped from the device coordinate system to the reading medium plane coordinate system to obtain the mapped gaze point sequence. Where M is the number of identified gaze points. Let J be the mapped coordinates of the j-th gaze point on the reading medium plane. For the corresponding gaze duration, the device coordinate system is the pixel coordinate system built into the eye-tracking device, and the reading medium plane coordinate system is the two-dimensional plane coordinate system of the physical reading material;

[0061] For the mapped gaze sequence Cluster analysis was performed to identify areas of interest that children focused on during reading. A density-based spatiotemporal clustering algorithm was used to group fixations based on spatial distance and temporal proximity. For each cluster... Calculate its center coordinates Total gaze duration and the total gaze duration identified Exceeding the preset threshold Clusters are used as effective regions of interest to form a set of regions of interest. ,in, For clusters The number of gaze points included, where K is the number of valid regions of interest identified;

[0062] Spatial location of the identified region of interest fixation sequence and fixation duration Visual encoding is performed to generate a time-series reading scan path map, where nodes in the path map correspond to the identified regions of interest. The spatial position of the node on the two-dimensional plane is determined based on the center coordinates of the region of interest, and the color depth of the path node represents the gaze duration. The longer the gaze duration, the darker the color of the node. The lines in the path diagram connect adjacent interest area nodes in the order of gaze duration. The direction of the arrows on the lines indicates the direction of the child's gaze.

[0063] Acoustic feature extraction is performed on voice interaction data, and skeletal key point recognition is performed on limb movement data to generate posture feature vectors, further including:

[0064] The voice interaction data undergoes pre-emphasis, frame segmentation, and windowing processing to extract acoustic feature parameters, including fundamental frequency, energy, formants, and Mel-frequency cepstral coefficients. An endpoint detection algorithm is used to remove silent segments and non-speech noise, retaining the interactive voice segments produced by the child's own speech. The extracted acoustic feature parameters are then input into a voice emotion recognition model, which is a Long Short-Term Memory (LSTM) network that captures emotional changes in speech through temporal modeling and outputs a voice emotion tendency score vector. ,in, These respectively characterize the level of excitement, questioning, and calmness in children's speech;

[0065] The captured video stream of body movements Human detection was performed to identify children in video frames. For the detected child regions, a pose estimation algorithm was used to extract the 3D spatial coordinates of 15 key skeletal points, including the head, neck, left and right shoulders, left and right elbows, left and right wrists, left and right hips, left and right knees, and left and right ankles, forming a set of key point coordinates. ;

[0066] Based on the extracted 3D spatial coordinates of the skeletal key points, the relative angles and displacement velocities between each key point are calculated to generate a temporalized posture feature vector representing the amplitude of body swaying, gesture activity, and forward / backward tilting postures. The amplitude of body swaying is obtained by calculating the standard deviation of the displacement of the head key points, and is expressed as follows: The gesture activity level is obtained by calculating the degree of pixel change in the hand region, and is represented as... The forward and backward tilt postures are obtained by calculating the Z-axis coordinate difference between key points of the head and key points of the hips, and are expressed as follows: ,in, A positive value indicates leaning forward. A negative value indicates a backward tilt, and the absolute value indicates the degree of tilt. and The Z-axis coordinates of the left and right hip key points are shown below. For head key points in frame b and frame 6 The displacement distance between frames, This is the average displacement distance. and These are the velocity vectors of the left and right wrist key points, respectively, and B is the total number of frames within the statistical window.

[0067] The posture feature vector is then normalized to eliminate data bias caused by differences in different acquisition devices and individual child body shapes, resulting in a normalized posture feature vector. .

[0068] Preferably, in step S2, the multi-dimensional reading behavior raw dataset obtained in step S1 is analyzed and fused using a deep learning model, including real-time emotional pleasure calculation, reading focus index, and interest focus determination, in order to accurately identify the child's real-time interest and focus state on the current reading content. Specifically, this includes the following:

[0069] The instant emotional pleasure calculation involves fusing the emotional feature vector, voice emotional tendency score vector, and posture feature vector obtained in step S1 at the feature level to construct a multimodal emotional feature tensor. This fused multimodal emotional feature tensor is then input into a preset emotional calculation model. The emotional calculation model is a multimodal fusion network based on an attention mechanism, including a cross-modal attention layer and a temporal modeling layer, generating an index that varies with reading time. The changing instantaneous emotional pleasure curve, specifically the mathematical formula is as follows: ,in, As a time index of the reading process, the curve reflects the continuous trajectory of children's emotional changes during reading. , , These represent the time indexes respectively. The level of curiosity, confusion, and joy at any given moment;

[0070] The reading focus index is calculated by combining the posture feature vector and the eye-tracking path map to determine the percentage of time the gaze lingers in the current reading area. And the frequency of rereading, to obtain the reading focus index. ,in, The preset weighting coefficients, and These represent the normalized body sway amplitude and forward lean posture values, respectively. It is the readback frequency, which represents the number of readbacks per unit of time;

[0071] The interest focus determination involves identifying the persistence residual of the gaze in the image area of ​​the reading material. The persistence residual is defined as the difference between the actual gaze duration in a specific image area and the expected gaze duration in that area based on the complexity of the image content. When the persistence residual exceeds a preset threshold... Furthermore, if the rate of change of the instantaneous emotional pleasure curve is greater than zero within this time window, the region is determined to be a valid focus of interest, and its spatial coordinates are extracted.

[0072] Preferably, in step S3, a dual knowledge graph is constructed, including a book knowledge graph and a child development stage graph. The focus of interest extracted in step S2 is used as input. A candidate book set in the book knowledge graph is located through similarity retrieval. A mapping function is used to align the difficulty weights of the candidate books with the cognitive comprehension thresholds in the child development stage graph. The "zone of proximal development deviation" of the candidate books relative to the child's current cognitive level is calculated. Books associated with the focus of interest identified in step S2 are then selected from the candidate book set to form a preliminary recommendation list, specifically including the following:

[0073] Constructing a knowledge graph of reading materials composed of book difficulty and thematic features. , where the set of nodes This includes the physical book and its attributes, including the book's unique identifier and difficulty level. Theme classification; edge set It represents semantic relationships between books, including "same series," "similar themes," and "same author"; and constructs a child development stage map based on child cognitive theory and big data statistics. , where the set of nodes Includes age group entities, edge set This indicates a progressive relationship between stages of development;

[0074] Using the focus of interest extracted in S2 as the query input, the similarity search is performed on the knowledge graph of the book. Candidate book collection for positioning For each book in the candidate book set, extract its difficulty level. Based on the child's user ID Historical reading data and current age Using deep learning-based models Child Development Stage Map The optimal developmental stage for the child is determined. And obtain the corresponding cognitive understanding threshold vector. ;

[0075] Construct mapping function Book difficulty level Mapped to By using a cognitive threshold space of the same dimension, we can obtain the cognitive demand vector corresponding to the book. The "zone of proximal development deviation value" of candidate books relative to the child's current cognitive level is calculated as follows: And filter out those in the positive deviation range The selected books will be scored according to their similarity to the focus of interest. Sort the books in descending order and select the top Q books to form a preliminary recommendation list.

[0076] Preferably, in step S4, for the preliminary recommendation list selected in step S3, an interactive trial reading method is used to confirm the recommendations, monitor the child's multimodal feedback data during the trial reading, and calculate the matching degree weight between the recommended content and the child's instantaneous interests; based on the matching degree weight, the preliminary recommendation list is rearranged and debiased to generate the final personalized reading recommendation list, which is then pushed to the reading device. Specifically, this includes the following:

[0077] The preliminary recommendation list generated in step S3 is presented in the form of a preview summary. Front-end sensing devices are used to monitor the child's instantaneous multimodal response to each summary item, and a facial expression recognition model is used to calculate the instantaneous emotional pleasure score. Calculate the reading focus index on the preview page based on eye movement data. The system monitors children's response delay to voice-guided commands and their tendency to lean forward, generating an interaction initiative coefficient. The three indicators mentioned above are combined to calculate the normalized instantaneous interest matching weight for each candidate book. ,in, is the normalization factor, and a, g, and c are the contribution coefficients of different modes;

[0078] Extract the original similarity score of the i-th book in the initial recommendation list. Calculate the deviation between the original score and the real-time matching weight for each book. When the deviation value exceeds the preset threshold: If the initial recommendation list is found to deviate significantly from the child's current instantaneous interests, a debiasing and reordering algorithm is initiated. An exponential perturbation function is used to correct the original scores, and a final personalized score is calculated for each book. ,in, The disturbance intensity coefficient is... The sign function indicates the direction of the deviation: The rearranged scores are then probability-mapped to generate a final personalized reading list, which is then pushed to users.

[0079] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0080] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of implementing an AI-based method for identifying and personally recommending children's reading interests as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0081] Example 2

[0082] The following is another embodiment of the present invention, which provides a method for AI recognition and personalized recommendation of children's reading interests. In order to verify the beneficial effects of the present invention, a simulation experiment is conducted for scientific demonstration.

[0083] This experiment aims to verify the effectiveness of an AI-based method for identifying and personally recommending children's reading interests. Through multimodal data collection, deep learning model fusion analysis, dual knowledge graph construction, and interactive trial reading confirmation, the accuracy of identifying children's reading interests and the relevance of personalized recommendations are improved. The experiment uses simulated and real-world children's reading behavior data, including multidimensional features such as facial expressions, eye movements, voice interaction, and body movements. By analyzing the consistency between the book list generated by the recommendation model and children's actual reading preference tags, the accuracy and robustness of this invention in identifying reading interests, determining attention levels, and making precise recommendations across the "zone of proximal development" are verified.

[0084] The simulation experiment steps are implemented according to the content of the AI ​​recognition and personalized recommendation method for children's reading interests provided in Example 1, and the specific steps include:

[0085] Multimodal data of children during reading was collected and preprocessed, with the collection frequency set to multiple times per second to capture instantaneous behavioral changes;

[0086] By using deep learning models to fuse and analyze preprocessed data, the system calculates immediate emotional pleasure and reading focus index, determines the focus of interest, and accurately locates the child's current reading excitement and fatigue state.

[0087] We construct a knowledge graph of reading materials and a graph of children's developmental stages, input the focus of interest into the graph for similarity retrieval, and use a mapping function to calculate the "zone of proximal development deviation value" between candidate books and children's cognitive level, thus filtering out a preliminary list that is both interesting and cognitively challenging.

[0088] During the trial reading phase, children's real-time multimodal feedback is monitored, instantaneous interest matching weights are calculated, and the preliminary list is rearranged and debiased based on the feedback results to generate and push the final personalized recommended book list.

[0089] The specific data from the above simulation experiment are as follows:

[0090]

[0091] Table 1

[0092] Experimental Analysis:

[0093] By comparing the consistency between the model-generated recommended book list and the actual interest tags shown by children during the trial reading, the effectiveness of the system in capturing instantaneous interest and matching cognitive stages was verified. The experiment shows that this method can adjust the recommendation strategy in real time based on the multimodal dynamic feedback during children's reading process. It can trigger interactive guidance when attention decreases, adaptively reduce the difficulty when cognitive dissonance occurs, and perform immediate reordering when interest is high, and finally push personalized book lists with bias-free characteristics. This can effectively improve children's reading experience, promote the cultivation of their reading interest and the spiral improvement of their reading ability, and ultimately achieve dynamic management and healthy growth support for children's reading habits throughout their entire life cycle.

[0094] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for AI-based identification and personalized recommendation of children's reading interests, characterized in that: Specifically, the following steps are included: S1. Simultaneously collect multimodal data of children during the reading process using front-end sensing devices. The multimodal data includes facial video streams, eye movement trajectory data, voice interaction data, and body movement data. Preprocess the collected raw data and perform spatiotemporal alignment processing on the processed modal data to generate a multidimensional raw dataset of reading behavior. The preprocessing includes extracting key points from the facial video stream to construct an emotion feature vector, mapping eye movement trajectory data to generate a reading scanning path map, extracting acoustic features from the voice interaction data, and identifying skeletal key points from the limb movement data to generate a posture feature vector. S2. Perform deep learning model analysis and fusion on the multi-dimensional reading behavior raw dataset obtained in step S1, including real-time emotional pleasure calculation, reading focus index and interest focus determination, so as to accurately identify the child's real-time interest and focus status on the current reading content. S3. Construct a dual knowledge graph, including a book knowledge graph and a child development stage graph. Using the interest focus extracted in step S2 as input, locate the candidate book set in the book knowledge graph through similarity retrieval. Use a mapping function to align the difficulty weight of the candidate books with the cognitive understanding threshold in the child development stage graph. Calculate the "zone of proximal development deviation value" of the candidate books relative to the child's current cognitive level. Select books associated with the interest focus identified in step S2 from the candidate book set to form a preliminary recommendation list. S4. For the preliminary recommendation list selected in step S3, an interactive trial reading method is used to confirm the recommendations. The multimodal feedback data of children during the trial reading is monitored, and the matching degree weight between the recommended content and the children's instantaneous interests is calculated. The preliminary recommendation list is rearranged and debiased according to the matching degree weight to generate the final personalized reading recommendation list and push it to the reading device.

2. The method for AI-based identification and personalized recommendation of children's reading interests according to claim 1, characterized in that: In step S1, the preprocessing involves extracting key points from the facial video stream to construct an emotion feature vector, and further includes: For the captured facial video stream Face detection and alignment are performed frame-by-frame. For each frame of the video stream, a deep learning-based face detection algorithm is used to identify all face regions in the image. For the identified child face region Geometric transformations are performed to achieve standardized alignment of the face, resulting in an aligned face image. ; Aligned face image The data is input into a pre-trained deep learning keypoint detection model to locate the coordinates of facial keypoints, including specific points on the contours of the eyes, eyebrows, nose, mouth, and chin, forming a set of keypoint coordinates. Where N is the number of key points; Facial motion units are extracted based on the located facial key point coordinates L, including the degree of mouth corner shrug, eyelid opening and closing, and eyebrow tilt angle. For the mouth area, the vertical distance between the mouth corner key point 301 and the mouth center point 302 is calculated. Horizontal distance The ratio of the two values ​​gives the degree of upturn of the corners of the mouth. ; Extract key points for the eye contour area (200). The eyelid opening angle is obtained by calculating the ratio of the distance between pairs of points in the vertical direction to the distance between pairs of points in the horizontal direction, and is expressed as: The eyebrow tilt angle is determined by locating the inner corner point 101 and the outer corner point 102 of the eyebrow, and constructing the angle between the line connecting the two points and the horizontal baseline. The extracted facial motion unit parameters are combined to form a motion unit feature vector. ,in, Let the coordinates of the corner of the mouth be... The coordinates of the center point of the mouth are calculated from the midpoints of the upper and lower lips, and ε is a minimal constant to prevent division by zero. to The coordinates of six key points around the eye are given. Let these be the coordinates of the inner corner of the eyebrow. The coordinates of the outer corner of the eyebrow; Extracted facial action unit feature vectors The input is fed into an emotion encoder, which is a deep neural network that performs feature transformation through fully connected layers and nonlinear activation functions to generate a three-dimensional emotion feature vector representing pleasure, confusion, and surprise, as shown below. Each dimension represents the intensity value of the child's current emotional state on the corresponding emotional dimension, normalized to... between.

3. The method for AI-based identification and personalized recommendation of children's reading interests according to claim 1, characterized in that: In step S1, the step of performing coordinate mapping on the eye movement trajectory data to generate a reading scan path map further includes: The gaze point coordinates in the original eye movement trajectory data are mapped from the device coordinate system to the reading medium plane coordinate system to obtain the mapped gaze point sequence. Where M is the number of identified gaze points. Let J be the mapped coordinates of the j-th gaze point on the reading medium plane. This corresponds to the duration of the gaze; For the mapped gaze sequence Cluster analysis was performed to identify areas of interest that children focused on during reading. A density-based spatiotemporal clustering algorithm was used to group fixations based on spatial distance and temporal proximity. For each cluster... Calculate its center coordinates Total gaze duration and the total gaze duration identified Exceeding the preset threshold Clusters are used as effective regions of interest to form a set of regions of interest. ,in, For clusters The number of gaze points included, where K is the number of valid regions of interest identified; Spatial location of the identified region of interest fixation sequence and fixation duration Visual encoding is performed to generate a time-series reading scan path map, where nodes in the path map correspond to the identified regions of interest. The spatial position of the node on the two-dimensional plane is determined based on the center coordinates of the region of interest, and the color depth of the path node represents the gaze duration. In the path diagram, the lines connect adjacent interest area nodes in sequence according to the time of gaze, and the arrows on the lines indicate the direction of the child's gaze jump.

4. The method for AI-based identification and personalized recommendation of children's reading interests according to claim 1, characterized in that: In step S1, the step of extracting acoustic features from voice interaction data and identifying skeletal key points from limb movement data to generate a posture feature vector further includes: The voice interaction data is pre-emphasized, framed, and windowed to extract acoustic feature parameters, including fundamental frequency, energy, formants, and Mel-frequency cepstral coefficients. An endpoint detection algorithm is used to remove silent segments and non-speech noise, retaining the interactive voice segments produced by the child's own speech. The extracted acoustic feature parameters are then input into a voice emotion recognition model, outputting a voice emotion tendency score vector. ,in, These respectively characterize the level of excitement, questioning, and calmness in children's speech; The captured video stream of body movements Human detection was performed to identify children in video frames. For the detected child regions, a pose estimation algorithm was used to extract the 3D spatial coordinates of 15 key skeletal points, including the head, neck, left and right shoulders, left and right elbows, left and right wrists, left and right hips, left and right knees, and left and right ankles, forming a set of key point coordinates. ; Based on the extracted 3D spatial coordinates of the skeletal key points, the relative angles and displacement velocities between each key point are calculated to generate a temporalized posture feature vector representing the amplitude of body swaying, gesture activity, and forward / backward tilting postures. The amplitude of body swaying is obtained by calculating the standard deviation of the displacement of the head key points, and is expressed as follows: The gesture activity level is obtained by calculating the degree of pixel change in the hand area, and is represented as... The forward and backward tilt postures are obtained by calculating the Z-axis coordinate difference between key points of the head and key points of the hips, and are expressed as follows: ,in, A positive value indicates leaning forward. A negative value indicates a backward tilt, and the absolute value indicates the degree of tilt. and The Z-axis coordinates of the left and right hip key points are shown below. For head key points in frame b and frame 6 The displacement distance between frames, This is the average displacement distance. and These are the velocity vectors of the left and right wrist key points, respectively, and B is the total number of frames within the statistical window. The posture feature vector is then normalized to eliminate data bias caused by differences in different acquisition devices and individual child body shapes, resulting in a normalized posture feature vector. .

5. The method for AI-based identification and personalized recommendation of children's reading interests according to claim 1, characterized in that: In step S2, the instant emotional pleasure calculation involves fusing the emotional feature vector, voice emotional tendency score vector, and posture feature vector obtained in step S1 at the feature level to construct a multimodal emotional feature tensor. The fused multimodal emotional feature tensor is then input into a preset emotional calculation model to generate an index that varies with reading time. The changing instantaneous emotional pleasure curve, specifically the mathematical formula is as follows: ,in, As a time index of the reading process, the curve reflects the continuous trajectory of children's emotional changes during reading. , , These represent the time indexes respectively. The level of curiosity, confusion, and joy at any given moment; The reading focus index is calculated by combining the posture feature vector and the eye-tracking path map to determine the percentage of time the gaze lingers in the current reading area. And the frequency of rereading, to obtain the reading focus index. ,in, The preset weighting coefficients, and These represent the normalized body sway amplitude and forward lean posture values, respectively. It is the readback frequency, which represents the number of readbacks per unit of time; The interest focus determination is based on identifying the persistence residual of the gaze in the reading material image area. When the persistence residual exceeds a preset threshold... Furthermore, if the rate of change of the instantaneous emotional pleasure curve is greater than zero within this time window, the region is determined to be a valid focus of interest, and its spatial coordinates are extracted.

6. The method for AI-based identification and personalized recommendation of children's reading interests according to claim 1, characterized in that: Specifically, S3 includes the following: Constructing a knowledge graph of reading materials composed of book difficulty and thematic features. , where the set of nodes This includes the physical book and its attributes, including the book's unique identifier and difficulty level. Theme classification; edge set Representing semantic relationships between books; constructing a child development stage map based on child cognitive theory and big data statistics. , where the set of nodes Includes age group entities, edge set This indicates a progressive relationship between stages of development; Using the focus of interest extracted in S2 as the query input, the similarity search is performed on the knowledge graph of the book. Candidate book collection for positioning For each book in the candidate book set, extract its difficulty level. Based on the child's user ID Historical reading data and current age Using deep learning-based models Child Development Stage Map The optimal developmental stage for the child is determined. And obtain the corresponding cognitive understanding threshold vector. ; Construct mapping function Book difficulty level Mapping to By using a cognitive threshold space of the same dimension, we can obtain the cognitive demand vector corresponding to the book. The "zone of proximal development deviation" of candidate books relative to the child's current cognitive level is calculated as follows: And filter out those in the positive deviation range The selected books will be scored according to their similarity to the focus of interest. Sort the books in descending order and select the top Q books to form a preliminary recommendation list.

7. The method for AI-based identification and personalized recommendation of children's reading interests according to claim 1, characterized in that: Specifically, S4 includes the following: The preliminary recommendation list generated in step S3 is presented in the form of a preview summary. Front-end sensing devices are used to monitor the child's instantaneous multimodal response to each summary item, and a facial expression recognition model is used to calculate the instantaneous emotional pleasure score. Calculate the reading focus index on the preview page based on eye movement data. The system monitors children's response delay to voice-guided commands and their tendency to lean forward, generating an interaction initiative coefficient. The three indicators mentioned above are combined to calculate the normalized instantaneous interest matching weight for each candidate book. ,in, is the normalization factor, and a, g, and c are the contribution coefficients of different modes; Extract the original similarity score of the i-th book in the initial recommendation list. Calculate the deviation between the original score and the real-time matching weight for each book. When the deviation value exceeds the preset threshold: The de-biased ranking algorithm is then activated, and the original scores are corrected using an exponential perturbation function to calculate the final personalized score for each book. ,in, The disturbance intensity coefficient is... The sign function indicates the direction of the deviation: The rearranged scores are then probability-mapped to generate a final personalized reading list, which is then pushed to users.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the AI-based method for identifying and personally recommending children's reading interests as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the AI-based method for identifying and personally recommending children's reading interests as described in any one of claims 1-7.

Images