A home-school communication interaction system based on AI data empowerment

Abstract

The application discloses a kind of based on AI data empowerment's home-school communication interaction system, belong to artificial intelligence and data processing technical field, including: student data acquisition module, acquisition multidimensional original data, multimodal data preprocessing module, generate standardized data, student data portrait construction module, adopt Transform network to build student comprehensive portrait, communication interaction engine, based on large language model generates personalized communication content, integrates intention recognition, sentiment analysis and push timing optimization, man-machine interface module, provide parent end and teacher end interaction, feedback learning module, through feedback data incremental training realizes self-adapting optimization.The application improves the pertinence and efficiency of home-school communication by multidimensional portrait and intelligent engine, solves the technical problem that low reading rate is caused by improper information push timing.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and data processing technology, and in particular to a home-school communication and interaction system based on AI data empowerment. Background Technology

[0002] With the development of educational informatization, home-school communication has become an important part of modern education management. Traditional home-school communication mainly relies on parent meetings, telephone communication, instant messaging software (such as WeChat and QQ), and platforms such as Home-School Communication System.

[0003] However, the existing home-school communication system still has the following technical problems:

[0004] 1. Fragmented information and low communication efficiency. The existing system lacks the ability to systematically integrate multi-dimensional data on students at school. Information such as academic performance, classroom performance, homework completion, and attendance records are scattered across different functional modules. Parents need to actively browse multiple interfaces to obtain complete information, resulting in low information retrieval efficiency.

[0005] 2. Lack of a personalized communication content generation mechanism. Existing systems typically use template-based notifications to push student information to parents, failing to generate targeted communication content based on individual student differences and parents' preferences, resulting in poor communication effectiveness.

[0006] 3. Lack of timing optimization for information push. Existing systems lack intelligent judgment on the timing of information push, often pushing information in batches at inappropriate times, resulting in information being ignored or having low read rates;

[0007] 4. Misunderstandings in communication between home and school. Parents' questions are often ambiguous and lack clear intent, making it difficult for existing systems to accurately understand parents' true intentions and provide precise answers.

[0008] Existing technologies already include solutions related to home-school communication. For example, XunJie Artificial Intelligence Technology (Henan) Co., Ltd. disclosed an AI-based intelligent classroom management system (CN120013718A), whose home-school communication module primarily uses one-way information push, lacking personalized response mechanisms and optimized push timing. Blue Origin Smart Technology (Hangzhou) Co., Ltd. disclosed a home-school interactive classroom note analysis system based on OCR and a large language model (CN121280188A), but it does not integrate parent intent recognition, sentiment analysis, and personalized response generation functions. Beijing Guorentong Education Technology Co., Ltd. disclosed a family education assistance method and interaction system based on an intelligent agent model (CN121504695A), whose communication method focuses on transmitting educational intervention instructions and does not design intent recognition and push timing optimization for parent consultation scenarios. It is evident that existing systems still have shortcomings in terms of information fragmentation, lack of personalized communication content, inappropriate push timing, and inaccurate intent understanding.

[0009] Therefore, a home-school communication system is needed that can effectively integrate students' multi-dimensional data, generate personalized communication content, intelligently optimize push timing, and accurately understand parents' intentions to solve the problems existing in the current technology. Summary of the Invention

[0010] The present invention aims to at least partially solve one of the technical problems in the above-mentioned technologies.

[0011] To achieve the above objectives, the first aspect of this invention proposes an AI-enabled home-school communication and interaction system, comprising a student data collection module, a multimodal data preprocessing module, a student data profile construction module, a communication and interaction engine, a human-computer interaction interface module, and a feedback learning module. The student data collection module is configured to connect to the school's academic affairs system, classroom management system, and attendance system, and automatically collect multi-dimensional raw data of students according to a preset cycle. This multi-dimensional raw data includes academic performance data, classroom performance data, homework completion data, attendance data, extracurricular activity participation data, and mental health assessment data. The multimodal data preprocessing module is communicatively connected to the student data collection module and is configured to perform data cleaning, missing value imputation, format standardization, and outlier detection on the collected multi-dimensional raw data to generate standardized structured academic data. The student data profile construction module is communicatively connected to the multimodal data preprocessing module and is configured to... A student data multidimensional representation learning network based on the Transformer architecture is employed to extract features from the standardized structured academic data, generating personalized feature vectors for students in four dimensions: academic level, learning behavior, comprehensive quality, and mental health. These personalized feature vectors are then fused to construct a comprehensive student profile. A communication interaction engine is connected to the student profile construction module and configured to process the comprehensive student profile data based on a large language model, generating interactive content for communication with parents. A human-computer interaction interface module is connected to the communication interaction engine and configured to provide both parent-side and teacher-side interactive interfaces. A feedback learning module is connected to both the human-computer interaction interface module and the student profile construction module, and configured to collect parent feedback behavior data on the interactive content. This feedback behavior data is then used as a supervisory signal input to the student data multidimensional representation learning network for incremental training.

[0012] In addition, the AI ​​data-enabled home-school communication and interaction system proposed above according to the present invention may also have the following additional technical features:

[0013] Furthermore, the communication and interaction engine includes: a content generation submodule, an intent recognition submodule, a sentiment analysis submodule, a response generation submodule, and a push timing optimization submodule. The content generation submodule is configured to receive interaction requests from parents, combine the student's comprehensive profile data, and use a large language model to generate personalized communication content adapted to the individual characteristics of the current student. The intent recognition submodule is configured to perform natural language processing on the text or voice information input by parents, and identify the type of communication intent of parents through an intent classification model. The intent types include academic consultation, behavioral performance, homework status, and school safety. The sentiment analysis submodule is configured to perform sentiment tendency analysis on the text or voice information input by parents and output sentiment tags. The response generation submodule is configured to generate targeted response text based on the communication intent and the sentiment tags, combined with the corresponding dimensional features in the student's comprehensive profile data. The push timing optimization submodule is configured to collect parents' interaction behavior data, analyze parents' information viewing habits and active time windows based on a time-series prediction model, and output optimal push timing parameters.

[0014] Furthermore, the student data profiling construction module employs a Transformer-based multidimensional representation learning network for student data, comprising: an input layer, a multi-head self-attention layer, a feedforward neural network layer, and an output layer. The multi-head self-attention layer uses an 8-head attention mechanism to calculate the correlation weights between elements in the input feature sequence. The output layer is used to generate academic level feature vectors, learning behavior feature vectors, comprehensive quality feature vectors, and mental health feature vectors.

[0015] Furthermore, the intent recognition submodule adopts a BERT-based intent classification model, which adds a fully connected classification layer on the basis of the pre-trained BERT model and is fine-tuned through the cross-entropy loss function.

[0016] Furthermore, the sentiment analysis submodule adopts an LSTM-Attention-based sentiment analysis model, which includes a word embedding layer, a bidirectional LSTM layer, and an attention layer.

[0017] Furthermore, the push timing optimization submodule constructs a time series of parent activity, statistically analyzing the average information viewing probability of parents in different time periods with hourly granularity, and applying a push scoring function: Calculate the push score and select the time period with the highest push score as the optimal push timing parameter; where... Time period The probability of viewing information. Time period The probability of a response. These are the weighting coefficients.

[0018] Furthermore, the feedback learning module converts the collected feedback behavior data into feedback reward values: The feedback reward value is used as the reward signal for reinforcement learning, and the PPO algorithm is used to incrementally train the student data multidimensional representation learning network.

[0019] Furthermore, the parent-side interactive interface provided by the human-computer interaction interface module includes an information viewing unit, a question consultation unit, and a feedback and evaluation unit, while the teacher-side interactive interface provided by the human-computer interaction interface module includes a data entry unit, a communication record viewing unit, and an intervention suggestion sending unit.

[0020] The second aspect of this invention proposes an AI-powered data-driven method for home-school communication and interaction, comprising the following steps:

[0021] S1: Collect multi-dimensional raw data of students through the student data collection module; S2: Preprocess the multi-dimensional raw data through the multimodal data preprocessing module to generate standardized structured academic data; S3: Construct comprehensive student profile data based on the standardized structured academic data through the student data profile construction module; S4: Respond to the interaction request from the parent's end, generate communication interaction content for parents through the communication interaction engine based on the comprehensive student profile data; S5: Push the communication interaction content to the parent's end through the human-computer interaction interface module; S6: Collect feedback behavior data of parents on the interaction content through the feedback learning module, and incrementally train the student data profile construction module based on the feedback behavior data.

[0022] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method proposed in the second aspect above.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] 1. By constructing multi-dimensional student data profiles, the systematization of student information was achieved, solving the technical problem of low communication efficiency caused by information fragmentation in existing technologies;

[0025] 2. By adopting a personalized communication content generation mechanism based on a large language model, targeted communication content can be generated according to the individual characteristics of students and the attention preferences of parents, which significantly improves the pertinence and effectiveness of home-school communication;

[0026] 3. By optimizing the push timing submodule, the optimal push timing is analyzed based on parent behavior data, which solves the technical problem of information being ignored due to inappropriate push timing, and effectively improves the readability and communication effect of information;

[0027] 4. Integrating intent recognition and sentiment analysis functions, it can accurately understand parents' communication intentions and emotional state, generate more targeted responses, and reduce information comprehension bias;

[0028] 5. The feedback learning module enables dynamic updates of student profiles and continuous optimization of communication strategies, giving the system adaptive learning capabilities and continuously improving communication effectiveness as usage time increases. Attached Figure Description

[0029] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0030] Figure 1 This is an overall architecture block diagram of a home-school communication and interaction system based on AI data empowerment according to an embodiment of the present invention;

[0031] Figure 2 This is a schematic block diagram of the internal module structure of a communication interaction engine in a home-school communication interaction system based on AI data empowerment, according to an embodiment of the present invention.

[0032] Figure 3 This is a flowchart of the feature extraction process for a student data profile construction module in an AI-enabled home-school communication and interaction system according to an embodiment of the present invention.

[0033] Figure 4 This is a flowchart of the push timing optimization submodule of a home-school communication and interaction system based on AI data empowerment, according to an embodiment of the present invention.

[0034] As shown in the figure:

[0035] 1. Student Data Acquisition Module; 2. Multimodal Data Preprocessing Module; 3. Student Data Profile Construction Module; 31. Input Layer; 32. Multi-head Self-Attention Layer; 33. Feedforward Neural Network Layer; 34. Output Layer; 4. Communication Interaction Engine; 41. Content Generation Submodule; 42. Intent Recognition Submodule; 43. Sentiment Analysis Submodule; 44. Response Generation Submodule; 45. Push Timing Optimization Submodule; 5. Human-Computer Interaction Interface Module; 6. Feedback Learning Module. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to embodiments.

[0037] Example 1: System Overall Architecture

[0038] like Figure 1 As shown, this embodiment provides a home-school communication and interaction system based on AI data empowerment, including: a student data collection module 1, a multimodal data preprocessing module 2, a student data profile construction module 3, a communication and interaction engine 4, a human-computer interaction interface module 5, and a feedback learning module 6.

[0039] Among them, the student data acquisition module 1 is connected to the multimodal data preprocessing module 2, the multimodal data preprocessing module 2 is connected to the student data profile construction module 3, the student data profile construction module 3 is connected to the communication and interaction engine 4, the communication and interaction engine 4 is connected to the human-computer interaction interface module 5, and the feedback learning module 6 is connected to both the human-computer interaction interface module 5 and the student data profile construction module 3.

[0040] It should be noted that the specific configuration of student data collection module 1 is as follows: It connects to the school's academic affairs system through a RESTful API interface, adopts the OAuth2.0 authentication mechanism, and automatically collects multi-dimensional raw data of students according to a preset cycle of, for example, once every 6 hours. The collected data is encapsulated in JSON format and includes the following fields: student unique identifier (student_id), collection timestamp, academic performance data (including scores of each subject, class ranking, and grade ranking), classroom performance data (including number of times hands were raised, accuracy rate of answering questions, and attention score), homework completion data (including submission status, score, and number of late submissions), attendance data (including records of lateness, early departure, leave, and absence), extracurricular activity participation data (including activity type, number of participations, and awards), and mental health assessment data (including scale scores and emotional state codes). The collection module has a built-in data cache queue to temporarily store data in the event of a network failure and automatically resume transmission after recovery.

[0041] It should be noted that the specific configuration of the multimodal data preprocessing module 2 is as follows: First, the original JSON data is parsed to identify missing value fields. For numerical missing values ​​(such as missing scores from a certain exam), the average of the student's historical scores in the same subject is used to fill them in. For categorical missing values ​​(such as missing extracurricular activity records), the "unknown" label is used to fill them in. Then, outlier detection is performed: the academic performance threshold range is set to, for example, 0~100 points. Values ​​outside the range are marked as outliers and removed. Abnormal attendance (such as being late more than 5 times in a day) are considered invalid data. Next, the data from different sources is formatted uniformly: the date and time are uniformly formatted to ISO8601 (YYYY-MM-DD HH:MM:SS), the scores are uniformly converted to a percentage system, and the classroom performance scores are normalized to the 0~1 range. Finally, standardized structured academic data is generated and stored as a wide table in CSV format, with each row corresponding to all dimensions of data for a student within a collection period.

[0042] It should be noted that the specific configuration of the human-computer interaction interface module 5 is as follows: The parent-side interaction interface is presented in the form of a WeChat mini program or a standalone APP, including an information viewing unit (displaying a summary of the student's comprehensive profile, recent performance highlights, and issues to be addressed), a question consultation unit (providing text input boxes and voice input buttons, supporting preset quick options for common questions), and a feedback evaluation unit (each push message has a "useful / useless" button and a star rating below it). The teacher-side interaction interface is a web management backend, including a data entry unit (teachers can manually supplement student performance records, such as "actively helped classmates today"), a communication record viewing unit (displaying all historical communication content and parent feedback by student dimension), and an intervention suggestion sending unit (teachers can fill in intervention suggestions for specific students, and the system will automatically push them to the corresponding parents).

[0043] It should be noted that the specific implementation of the student data profile construction module 3 is shown in Example 2, the specific implementation of the communication and interaction engine 4 is shown in Example 3, and the specific implementation of the feedback learning module 6 is shown in Example 4.

[0044] Example 2: Student Data Profiling Module 3

[0045] This embodiment provides a detailed description of the student data profiling module 3.

[0046] The student data profile building module 3 adopts a student data multidimensional representation learning network based on the Transformer architecture, which includes an input layer 31, a multi-head self-attention layer 32, a feedforward neural network layer 33, and an output layer 34.

[0047] It should be noted that input layer 31 receives standardized structured academic data processed by multimodal data preprocessing module 2. This data is organized into a feature sequence with a sequence length of L (L equals the number of feature dimensions, for example, L=48). Each time step corresponds to one dimension of academic feature. The feature sequence first passes through a linear embedding layer, mapping each feature to, for example, a 64-dimensional embedding vector, to obtain an embedding matrix. At the same time, learnable positional encoding is added to preserve the order information of each feature dimension.

[0048] It should be noted that the multi-head self-attention layer 32 employs an 8-head attention mechanism, where each attention head independently calculates the correlation weights between elements in the input feature sequence. Calculate the query, key, and value matrix: ,in, , , according to The attention output is calculated, and the outputs of the 8 heads are concatenated and then passed through a linear transformation layer to obtain a multi-attention fusion representation.

[0049] It should be noted that feedforward neural network layer 33: employs a two-layer fully connected network with the ReLU activation function in between. The first layer maps the input dimension from, for example, 512 dimensions to 2048 dimensions, and the second layer maps it back to 512 dimensions. Residual connections and layer normalization are added between layers.

[0050] It should be noted that output layer 34: The 512-dimensional feature vector output from feedforward neural network layer 33 is passed through four parallel fully connected sub-networks to generate personalized feature vectors in four dimensions respectively.

[0051] Academic performance feature vector: Through a fully connected layer of, for example, 128 dimensions, a 128-dimensional vector is output. The first 40 dimensions (optional) represent the current scores of each subject, the middle 40 dimensions (optional) represent the trend of score fluctuation (the slope of change of three consecutive tests), and the last 48 dimensions (optional) represent the mastery of knowledge points (based on the analysis of homework and test mistakes).

[0052] Learning behavior feature vector: Through a fully connected layer of, for example, 96 dimensions, a 96-dimensional vector is output, in which 32 dimensions (optional) represent classroom participation (including frequency of raising hands, accuracy of answering questions, and attention duration), 32 dimensions (optional) represent the quality of homework completion (timeliness of submission, score rate, and number of corrections), and 32 dimensions (optional) represent self-directed learning behavior (extracurricular learning time and frequency of resource access).

[0053] Comprehensive literacy feature vector: Through a fully connected layer of, for example, 64 dimensions, a 64-dimensional vector is output, of which 24 dimensions (optional) represent the degree of participation in extracurricular activities (richness of activity types, frequency of participation), 20 dimensions (optional) represent the ability of teamwork (peer evaluation scores in project cooperation), and 20 dimensions (optional) represent the ability of innovation (awards for creative works, scores for open-ended tasks).

[0054] Mental health feature vector: Through a fully connected layer of, for example, 64 dimensions, a 64-dimensional vector is output, in which 24 dimensions (optional) represent emotional state (based on scale scores and sentiment analysis of daily texts), 20 dimensions (optional) represent stress level (based on workload, exam frequency and self-report), and 20 dimensions (optional) represent social adaptability (peer relationship evaluation, teacher-student relationship score).

[0055] It should be noted that the above four feature vectors, after being L2 normalized, are then... The data is then stitched together and merged to generate the final comprehensive student profile data. For example, the dimensions are 352. This profile data is updated every 24 hours and is driven by incremental training triggered by the feedback learning module.

[0056] Specifically, network training: This representation learning network adopts a self-supervised pre-training method, using historical data accumulated by the school, such as two years of anonymous student data (containing, for example, 100,000 student records), for pre-training on a mask prediction task. After pre-training, it is fine-tuned on a small amount of labeled data (for example, about 2,000 records) using contrastive learning, so that the portraits of the same student at different time slices are close together in the embedding space, while the portraits of different students are far apart from each other.

[0057] Example 3: Communication and Interaction Engine 4

[0058] like Figure 2 As shown, this embodiment provides a detailed description of the communication and interaction engine 4.

[0059] The communication and interaction engine 4 includes: content generation submodule 41, intent recognition submodule 42, sentiment analysis submodule 43, response generation submodule 44, and push timing optimization submodule 45.

[0060] It should be noted that the content generation submodule 41 is configured to receive interaction requests from parents (automatically triggered when entering the dialogue for the first time, or triggered when the parent clicks the "Generate Weekly Report" button), and, in conjunction with the student's comprehensive profile data P, call the large language model to generate personalized communication content adapted to the current individual characteristics of the student.

[0061] Specifically, the submodule maintains a prompt word template library. Different templates are selected based on the communication scenario (daily report, abnormal warning, teacher feedback). For example, the template for the daily report scenario is: "{{Student Name}} has made significant progress in {{Subject}} this week, with the score improving from {{Last Score}} to {{This Score}}. In terms of classroom performance, the student raised their hand {{Number of Hand Raises}} times and answered questions with a accuracy rate of {{Accuracy Rate}}%, which suggests that you...". The placeholders in the template are filled with values ​​of the corresponding dimensions extracted from the student profile data. For proactive pushes without explicit requests, the content generation submodule 41 automatically generates a "Student Comprehensive Weekly Report" every Monday at 8:00 AM, which includes an academic summary, behavioral highlights, areas for improvement, and suggestions for family education.

[0062] It should be noted that the intent recognition submodule 42 employs a BERT-based intent classification model. The model input is text input by parents (if it is speech, it is first converted to text by Tencent Cloud or iFlytek's speech recognition API), and the output is the probability distribution of four intent types. The model structure uses a pre-trained Chinese BERT-base model (12 layers, 768-dimensional hidden layers, 12 attention heads), with a fully connected classification layer added on top of its [CLS] labeled output, resulting in an output dimension of 4. The training data consists of 5000 manually labeled dialogues from real home-school communication records, with category distributions such as 45% for academic consultation, 30% for behavioral performance, 15% for homework status, and 10% for school safety. The cross-entropy loss function is used during training. The AdamW optimizer is used, with a learning rate of 2e-5, a batch size of 32, and training for 3 epochs. A confidence threshold of 0.7 is set during recognition. If the probability of all categories is lower than 0.7, the category is classified as "other" and processed manually.

[0063] It should be noted that the sentiment analysis submodule 43 employs an LSTM-Attention-based sentiment analysis model. The model structure involves first mapping each word in the input text to a vector using pre-trained Chinese word vectors (e.g., 300-dimensional), resulting in a word vector sequence. Then, a bidirectional LSTM layer is used, with the hidden state dimensions of both the forward and backward LSTMs being, for example, 128 dimensions. The forward and backward states at each time step are concatenated to obtain, for example, a 256-dimensional context representation. Next, an attention layer is introduced: the attention weights at each time step are calculated. Where W and u are learnable parameters, representing the values ​​at each time step. The sentiment feature vector of the entire sentence is obtained by weighted summation. Finally, a fully connected layer with softmax is used to output the probabilities of the three sentiment categories. The training data consists of 8,000 labeled entries from home-school communication corpora, with approximately 2,667 entries in each of the three categories. Category balanced sampling is used. When the model outputs the sentiment label, it takes the category with the highest probability and outputs the probability value as the confidence score. When the confidence score of negative sentiment is higher than, for example, 0.8, the system alert is triggered, the dialogue is marked as "high priority", and the homeroom teacher is notified.

[0064] It should be noted that the response generation submodule 44 is implemented based on a large language model, specifically using GPT-3.5-turbo or an equivalent open-source model (such as ChatGLM3-6B). This submodule uses the personalized communication content generated by the content generation submodule 41 as the response framework, and the communication intent identified by the intent recognition submodule 42 and the sentiment tags output by the sentiment analysis submodule 43 as contextual conditions. It combines the dimensional features in the student's comprehensive profile data corresponding to the current intent, and calls the large language model to generate targeted response text. In actual deployment, the submodule maintains a locally deployed large language model, which is called via HTTP API. The following parameters are used during generation, for example: temperature=0.7 (balancing creativity and consistency), top_p=0.9, max_tokens=300. To prevent the generation of inappropriate content, output filtering rules are set: absolute evaluations (such as "your child is stupid") are prohibited, information about other students is prohibited, and medical diagnosis suggestions are prohibited. Before the generated response text is pushed to parents, it will be post-processed by a rule engine to replace sensitive words (such as replacing "poor student" with "has room for improvement").

[0065] It should be noted that the push timing optimization submodule 45 collects parents' interaction behavior data, including: information viewing time (accurate to the minute), information viewing duration (the time difference between the push notification and the user opening it), information viewing frequency (average number of views per day), and response time (the time difference between the push notification and the user entering a response). Based on this data, an LSTM-based time series prediction model is used to analyze parents' information viewing habits and active time windows.

[0066] Specifically, the submodule constructs a time series of parent activity, using hours as the time granularity, and statistically analyzes the average probability of parents viewing information at different time periods over the past 30 days, for example. and response probability For a certain time period t (e.g., 8:00~9:00), through The push score is calculated, where the weight coefficient α is initially set to 0.5, and is adaptively adjusted monthly based on the reward value of the feedback learning module. The push timing optimization submodule 45 selects the time period with the highest push score as the optimal push timing parameter. For example, if the optimal push timing for a parent is "19:00~20:00 on weekdays", then all non-urgent information will be pushed in batches during this time period. For urgent information (such as "student has not arrived at school"), the parameter is ignored and the information is pushed immediately. This submodule recalculates the optimal push timing for each parent every two weeks to adapt to changes in parents' schedules.

[0067] Example 4: Feedback Learning Module 6

[0068] This embodiment provides a detailed description of the feedback learning module 6.

[0069] The feedback learning module 6 collects feedback behavior data from parents on interactive content. The feedback behavior data includes: information reading status (read / unread), reading time (the time from when the information is displayed on the interface to when the user slides or closes it), whether the "helpful" button is clicked, and whether the parent actively initiates a follow-up conversation (whether a new question is entered within 10 minutes after receiving a system reply).

[0070] The aforementioned feedback behavior data is converted into monitoring signals, and then... The initial values ​​for each weighting coefficient are set as follows: w1=0.3, w2=0.2, w3=0.3, w4=0.2, and the standard reading time threshold is used to calculate the feedback reward value. For example, if the time is set to 15 seconds (based on A / B testing, users usually finish reading the core content within this time), the reward value R ranges from 0 to 1. For example, if the parent has read the content, the reading time is 20 seconds, the parent clicks "helpful," and initiates a follow-up conversation, then R = 0.3 × 1 + 0.2 × (20 / 15) + 0.3 × 1 + 0.2 × 1 = 0.3 + 0.267 + 0.3 + 0.2 = 1.067, which is truncated to 1.0. If the parent has not read the content, then R = 0.

[0071] The feedback reward value is used as the reward signal for reinforcement learning and input into the student data multidimensional representation learning network in the student data profile construction module 3. The network is then incrementally trained using the policy gradient method.

[0072] Specifically, the PPO algorithm is used to update network parameters. Training settings: A model update is triggered every 500 feedback samples collected, for example, with the hyperparameter pruned to ε=0.2, and the advantage function... The learning rate is calculated using the GAE method (λ=0.95), starting from 1e-5 and decaying exponentially with each update. To prevent catastrophic forgetting, each update randomly samples, for example, 200 old samples from the historical replay buffer for joint training. Through this incremental training, student profiles are dynamically updated (for example, if a student's classroom performance has significantly improved in the past week, the "classroom participation" dimension in their behavioral feature vector will be automatically increased), and communication strategies are continuously optimized (for example, if it is found that parents are positive about academic information but negative about behavioral information, the system will reduce the frequency of behavioral information pushes).

[0073] The specific process of incremental training:

[0074] First, at 2:00 AM every day, the feedback learning module 6 reads all feedback behavior data generated the previous day from the database.

[0075] Then, according to Calculate the reward value R for each piece of feedback.

[0076] Secondly, the reward value is used as a label and combined with the student profile network to form a training sample based on the input features (i.e., standardized structured academic data) before the interaction.

[0077] Then, the PPO algorithm is used to perform one round of parameter updates (one epoch, batch size=64).

[0078] Next, the updated network parameters are saved to the model server and will take effect at 8:00 AM the following day.

[0079] Finally, an offline evaluation is conducted monthly: the correlation between the prediction accuracy and reward value of the model before and after the update is calculated on the validation set. If the performance drops by more than 5%, the model is rolled back to the previous month's version and a manual review is triggered.

[0080] Through the incremental training described above, the student profiles are dynamically updated, communication strategies are continuously optimized, and the system has adaptive learning capabilities. With increased usage time, the information reading rate significantly improves under simulation testing, and the parent satisfaction score increases from 4.2 out of 5 to 4.7.

[0081] In the description of this specification, 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 at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0082] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0083] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A home-school communication and interaction system based on AI data empowerment, characterized in that, include: The student data collection module (1) is configured to access the school's academic affairs system, classroom management system and attendance system, and automatically collect multi-dimensional raw data of students according to a preset cycle. The multi-dimensional raw data includes academic performance data, classroom performance data, homework completion data, attendance data, extracurricular activity participation data and mental health assessment data. The multimodal data preprocessing module (2) is connected to the student data acquisition module (1) and is configured to perform data cleaning, missing value filling, format unification processing and outlier detection on the acquired multi-dimensional raw data to generate standardized structured academic data. The student data profile construction module (3) is connected to the multimodal data preprocessing module (2) and is configured to use a student data multidimensional representation learning network based on the Transformer architecture to extract features from the standardized structured academic data, generate personalized feature vectors of students in four dimensions: academic level, learning behavior, comprehensive quality and mental health, and fuse the personalized feature vectors of the four dimensions to construct comprehensive student profile data. The communication and interaction engine (4) is connected to the student data profile building module (3) and is configured to process the student comprehensive profile data based on the large language model to generate interactive content for communication with parents. The human-computer interaction interface module (5) is connected to the communication interaction engine (4) and is configured to provide a parent-side interaction interface and a teacher-side interaction interface. The feedback learning module (6) is connected to the human-computer interaction interface module (5) and the student data profile construction module (3) respectively. It is configured to collect feedback behavior data of parents on the interactive content and input the feedback behavior data as a supervision signal into the student data multidimensional representation learning network for incremental training.

2. The system according to claim 1, characterized in that, The communication and interaction engine (4) includes: The content generation submodule (41) is configured to receive interaction requests from parents, combine the student comprehensive profile data, and call the large language model to generate personalized communication content that is adapted to the individual characteristics of the current student. The intent recognition submodule (42) is configured to perform natural language processing on the text or voice information input by the parents and identify the communication intent type of the parents through the intent classification model. The intent types include academic consultation, behavior performance, homework status and school safety. The sentiment analysis submodule (43) is configured to perform sentiment analysis on the text or voice information input by parents and output sentiment tags; The response generation submodule (44) is configured to generate targeted response text based on the communication intent and the emotional tags, combined with the corresponding dimension features in the student comprehensive profile data; The push timing optimization submodule (45) is configured to collect parents' interactive behavior data, analyze parents' information viewing habits and active time windows based on the time series prediction model, and output the optimal push timing parameters.

3. The system according to claim 1, characterized in that, The student data profile construction module (3) uses a Transformer-based multidimensional representation learning network for student data, which includes: an input layer (31), a multi-head self-attention layer (32), a feedforward neural network layer (33), and an output layer (34). The multi-head self-attention layer (32) adopts an 8-head attention mechanism to calculate the correlation weights between elements in the input feature sequence; The output layer (34) is used to generate academic level feature vectors, learning behavior feature vectors, comprehensive quality feature vectors and mental health feature vectors.

4. The system according to claim 2, characterized in that, The intent recognition submodule (42) adopts a BERT-based intent classification model, which adds a fully connected classification layer on the basis of the pre-trained BERT model and fine-tunes the training through the cross-entropy loss function.

5. The system according to claim 2, characterized in that, The sentiment analysis submodule (43) adopts an LSTM-Attention-based sentiment analysis model, which includes a word embedding layer, a bidirectional LSTM layer, and an attention layer.

6. The system according to claim 2, characterized in that, The push timing optimization submodule (45) constructs a time series of parent activity, and calculates the average information viewing probability of parents in different time periods with hourly time granularity, and uses the push scoring function: Calculate the push score and select the time period with the highest push score as the optimal push timing parameter; in, Time period The probability of viewing information. Time period The probability of a response. These are the weighting coefficients.

7. The system according to claim 1, characterized in that, The feedback learning module (6) converts the collected feedback behavior data into feedback reward values: The feedback reward value is used as the reward signal for reinforcement learning, and the PPO algorithm is used to incrementally train the student data multidimensional representation learning network.

8. The system according to claim 1, characterized in that, The parent-side interactive interface provided by the human-computer interaction interface module (5) includes an information viewing unit, a question consultation unit, and a feedback evaluation unit. The teacher-side interactive interface provided by the human-computer interaction interface module (5) includes a data entry unit, a communication record viewing unit, and an intervention suggestion sending unit.

9. A home-school communication and interaction method based on AI data empowerment, applied to the system described in any one of claims 1 to 8, characterized in that, Includes the following steps: S1: Collect multi-dimensional raw data of students through the student data collection module (1); S2: The multi-dimensional raw data is preprocessed by the multi-modal data preprocessing module (2) to generate standardized structured industrial data; S3: Construct comprehensive student profile data based on the standardized structured academic data through the student data profile construction module (3); S4: In response to the interaction request from the parent, the communication interaction engine (4) generates communication interaction content for the parent based on the student comprehensive profile data; S5: Push the communication content to the parent's end through the human-computer interaction interface module (5); S6: Collect feedback behavior data of parents on interactive content through the feedback learning module (6), and perform incremental training on the student data profile construction module (3) based on the feedback behavior data.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of claim 9.

Images