Innovative sentence vector and hidden semi-Markov model combined personalized picture book intelligent recommendation system

Abstract

The invention discloses a personalized picture book intelligent recommendation system based on combination of innovative sentence vectors and a hidden semi-Markov model, which belongs to the technical field of education and comprises a data acquisition module, a sentence vector generation module, an interest tracking module, a recommendation generation module, the innovative sentence vectors, the hidden semi-Markov model and an expectation maximization (EM) algorithm. According to the method, an efficient and accurate personalized picture book recommendation system is successfully constructed through deep fusion of the innovative sentence vectors and the HSMM, the problem of interest drifting of students in the reading process is effectively solved, the innovative sentence vectors can comprehensively capture deep semantics and emotional tendencies of reading texts of the students, and the interestingness of the students is improved. A rich and multi-dimensional data basis is provided for interest analysis, the HSMM model allows the students to stay for different time durations in different potential interest states, the heterogeneity of the interest duration of the students is accurately described, and dynamic tracking of interest changes of the students is achieved.

Description

Technical Field

[0001] The present invention belongs to the field of educational technology, and in particular relates to a personalized picture book intelligent recommendation system that combines innovative sentence vectors with a hidden semi-Markov model. Background Art

[0002] Recommendation systems have been widely used in the fields of education and reading. Their core goal is to recommend books and learning materials that match students' interests and needs through in-depth analysis of user behavior data. This system has played an important role in improving learning efficiency and stimulating interest in reading. However, with the changes in the educational environment and individual student development, students' interests and needs show significant dynamic characteristics and often change significantly over time. For example, lower grade students are interested in fairy tale picture books, but as they get older, they gradually turn to popular science knowledge or literary classics picture books.

[0003] Existing recommendation systems have obvious limitations in processing time information. On the one hand, some systems fail to fully consider the impact of time factors on user behavior during the design and implementation process, resulting in a lack of timeliness and dynamic adaptability in recommendation results. This practice of ignoring the time dimension makes it difficult for recommendation systems to capture and respond to real-time changes in students' interests. On the other hand, although some systems try to introduce time information, they often convert the complexity of modeling into the difficulty of feature design, and compensate for the shortcomings of the model itself by adding a large number of time-related features. This approach not only increases the computational burden of the system, but also leads to feature redundancy and overfitting problems, affecting the stability and accuracy of the recommendation effect.

[0004] This limitation is particularly prominent in the field of picture book recommendations. As an important carrier of early childhood education, the recommendation effect of picture books is directly related to students' reading experience and learning outcomes. However, most picture book recommendation systems on the current market lack effective methods to track students' interest drift. These systems often make recommendations based on static user portraits or historical behavior data, and fail to fully consider the dynamic characteristics of students' interests changing over time. As a result, it is difficult for the recommendation system to capture subtle changes in students' interests in real time, resulting in the recommended picture books being out of touch with students' current interests and needs, and the recommendation effect is poor. This not only affects students' reading interest, but also hinders the improvement of their reading ability and knowledge level.

[0005] Based on this, the present invention designs a personalized picture book intelligent recommendation system that combines innovative sentence vectors with hidden semi-Markov models to solve the above problems. Summary of the Invention

[0006] The purpose of the present invention is to solve the problem that most picture book recommendation systems on the current market lack effective methods to track the drift of students' interests. These systems often make recommendations based on static user portraits or historical behavior data, and fail to fully consider the dynamic characteristics of students' interests changing over time. As a result, it is difficult for the recommendation system to capture subtle changes in students' interests in real time, resulting in a disconnect between the recommended picture books and students' current interests and needs, and poor recommendation effects. This not only affects students' reading interest, but also hinders the improvement of their reading ability and knowledge level. A personalized picture book intelligent recommendation system combining innovative sentence vectors with hidden semi-Markov models is proposed.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A personalized picture book intelligent recommendation system that combines innovative sentence vectors with hidden semi-Markov models, including a data acquisition module, a sentence vector generation module, an interest tracking module, a recommendation generation module, innovative sentence vectors, a hidden semi-Markov model, and an expectation-maximization (EM) algorithm.

[0009] The data collection module ensures data consistency, availability, and security through multi-dimensional data collection, data cleaning, data normalization, data verification, and quality control, providing a solid data foundation for subsequent interest analysis and recommendation generation.

[0010] The sentence vector generation module uses an efficient reasoning model to perform deep vectorization processing on students' reading content, where the efficient reasoning model uses BERT or GPT;

[0011] The interest tracking module is based on the Hidden Semi-Markov Model (HSMM) and models the student's interest changes as a transition process of latent states.

[0012] The recommendation generation module includes interest matching strategy, recommendation list generation and recommendation output mode;

[0013] The innovative sentence vectors use the efficient inference large-scale models BERT or GPT. These models have powerful natural language processing capabilities and can deeply understand the semantic structure and sentiment of text. The BERT model uses a bidirectional Transformer structure and generates sentence vectors containing rich semantic information through pre-training and fine-tuning. The pre-training tasks of the BERT model include masked language modeling and next sentence prediction.

[0014] The hidden semi-Markov model is used to model the changes in students' interests. By introducing state duration distribution, students are allowed to stay in different potential states for different lengths of time, so as to better simulate the heterogeneity of students' interests and focus.

[0015] The expectation-maximization (EM) algorithm gradually approaches the true value of the model parameters by alternately performing the expectation step (E) and the maximization step (M). In the E step, the posterior probability distribution of the potential state under the current parameters is calculated.

[0016] As a further description of the above technical solution:

[0017] The data collection module acquires multi-dimensional student data from the education platform database, including but not limited to reading content, historical reading records, academic performance, practice mastery, and oral practice performance. Specifically, reading content data includes the text of picture books read by students, reading time, and reading frequency; historical records cover students' past reading behaviors and preferences; performance data includes grades in various subjects and comprehensive evaluations; practice mastery reflects students' performance in various exercises; and oral practice data records students' oral expression ability and progress. In addition, the system also collects data from students' reading preferences, learning habits, and cognitive levels to more comprehensively understand students' individual needs and learning characteristics.

[0018] The collected data is cleaned, noise and invalid data are removed, and normalized to ensure data consistency and availability. Specifically, the data cleaning process includes identifying and removing outliers, missing values, and duplicate records in the data. The data normalization process includes standardizing numerical data to make it conform to the standard normal distribution, and normalizing non-numerical data, such as one-hot encoding or word embedding representation. In addition, the system also uses data integration and data conversion technologies to integrate and process data from different sources and formats to provide more comprehensive and accurate data support.

[0019] As a further description of the above technical solution:

[0020] The BERT model calculates the vector representation of text using the following formula:

[0021] BERT(x)=Transformer(Embedding(x))

[0022] Here, x represents the input text, Embedding(x) is the word vector representation of the text, and Transformer is the core component of the BERT model, responsible for extracting the deep features of the text. The generated sentence vector not only contains the surface vocabulary information of the text, but also captures multiple deep features such as sentence structure, semantic relationship, and emotional expression, providing a rich information foundation for subsequent interest analysis and recommendation. The BERT model is trained through the pre-training tasks of Masked Language Model (MLM) and Next Sentence Prediction (NSP), using the following formula:

[0023]

[0024] Among them, L represents the loss function of the BERT model, x i represents the i-th word in the input text, x mask To represent masked words, the GPT model is based on a unidirectional Transformer structure. Through large-scale pre-training, it generates sentence vectors with generative and semantic understanding capabilities. The pre-training task of the GPT model includes language modeling (LM), which is trained using the following formula:

[0025]

[0026] Among them, L represents the loss function of the GPT model, x i represents the i-th word in the input text, x 1:i-1 Represents the first i-1 words of the input text. The vector generation process includes text input, feature extraction, and vector output. The text input is the picture book text read by the students input into the model. The feature extraction model extracts the multi-dimensional features of the text including vocabulary, syntax, semantics, and emotion through a multi-level attention mechanism and transformation network. The calculation formula of the attention mechanism is as follows:

[0027]

[0028] Where Q, K, and V represent query, key, and value matrices respectively, and d k The vector output model outputs an innovative sentence vector representing the dimension of the key vector. This vector comprehensively reflects the surface information and deep semantic structure of the text. The sentence vector generated by the vector application is used for subsequent interest tracking and recommendation generation, providing a solid data foundation for accurately identifying students' interests and recommending suitable picture books.

[0029] As a further description of the above technical solution:

[0030] The HSMM model is defined by the following parameters:

[0031] Initial state distribution π=(π1,π2,...,π N ), where π i represents the probability that the student is initially in state i. The initial state distribution is estimated by analyzing the student's historical reading behavior and preferences. Based on the type and number of picture books the student has read in the past period of time, the initial probability of each type of picture book is calculated;

[0032] State transition probability matrix A=[a ij ] N×N , where a ijrepresents the probability of a student transferring from state i to state j. The state transition probability matrix reflects the dynamic changes of students' interests between different states. If a student switches from reading fairy tale picture books to reading science picture books, the value of the corresponding element in the state transition probability matrix will increase.

[0033] State duration distribution D = (D1, D2, ..., D N ), where D i represents the duration distribution of students in state i. The state duration distribution allows students to stay in different potential states for different lengths of time, which is used to better simulate the heterogeneity of students' interests and focus. Students may be interested in picture books on a specific theme for a period of time, and more interested in picture books on other themes at another time.

[0034] Observation probability matrix b = [b ik ] N×M , where b ik It represents the probability of observing data k when the student is in state i. The observation probability matrix reflects the correlation between the student's interest state and the observed data. When students observe data related to popular science knowledge when reading popular science picture books, the value of the corresponding element in the observation probability matrix will increase.

[0035] As a further description of the above technical solution:

[0036] Define multiple potential interest states, each representing a student's primary interest within a specific timeframe, including fairy tales, popular science, and literary classics. These states are dynamically adjusted based on the student's reading behavior and preferences to accommodate changes in student interests.

[0037] The model allows students to stay in different potential states for different lengths of time. By capturing the heterogeneity of state duration, it accurately reflects the dynamic changes in students' interests. For example, students may be very interested in picture books on a specific theme for a period of time, and more interested in picture books on other themes at another time.

[0038] In order to comprehensively evaluate the student's current interest status, the sentence vector is integrated with other multivariate data for analysis, and the student's comprehensive interest score is calculated using the following formula:

[0039]

[0040] Among them, S represents the student's comprehensive interest score, score(k) represents the score of the student's observed data k. By comprehensively considering the student's initial state distribution, state transition probability matrix, state duration distribution and observation probability matrix, it is used to more accurately evaluate the student's interest state and provide a more reliable data basis for personalized picture book recommendations.

[0041] As a further description of the above technical solution:

[0042] The interest matching strategy uses an intelligent matching algorithm based on the current interest status output by the interest tracking module to dynamically adjust the picture book recommendation strategy. The system first selects a set of picture books related to the student's current interest status from a preset picture book library. If the student's current interest status is popular science knowledge, the system will prioritize selecting picture books that match the student's interest from the popular science picture book library. In addition, the system also considers the student's historical reading data and multi-dimensional information of personalized needs to further screen and sort the selected picture books to ensure the accuracy and personalization of the recommendation results.

[0043] The recommendation list is generated by combining students' historical reading data, current interest status, and personalized needs, and adopting an intelligent recommendation algorithm to generate a recommendation list containing multiple picture books. The system first analyzes students' reading preferences and habits, reading time, reading frequency, and reading types based on their historical reading data, and builds a personalized reading profile of the students. Then, the system scores and sorts the picture books in the picture book library based on the students' current interest status and personalized reading profile. The scoring formula is as follows:

[0044]

[0045] Among them, R(p) represents the recommendation score of picture book p, a ij and b ik They represent the state transition probability and observation probability respectively, and score(k) represents the score of the student observing data k. Finally, the system sorts the picture books according to the recommendation scores and generates a recommendation list containing multiple picture books;

[0046] In the recommendation output method, the recommendation list is displayed to students through the education platform interface, and detailed picture book introductions and reading suggestions are provided to enhance the interactivity and practicality of the recommendation. The system displays the recommendation list to students in the form of pictures and texts, including picture book cover, title, introduction and author information. In addition, the system also provides corresponding reading suggestions, reading methods and reading skills based on the personalized needs of students. For science picture books, the system is used to provide relevant science knowledge explanations and experimental operation guides. The output method improves students' reading interest and promotes the improvement of students' reading ability and knowledge level.

[0047] As a further description of the above technical solution:

[0048] The pre-training tasks of the BERT model include the masked language model (MLM) and the next sentence prediction (NSP), which are trained using the following formula:

[0049]

[0050] Among them, L represents the loss function of the BERT model, x i represents the i-th word in the input text, x mask To represent masked words, the GPT model is based on a unidirectional Transformer structure. Through large-scale pre-training, it generates sentence vectors with generative and semantic understanding capabilities. The pre-training task of the GPT model includes language modeling (LM), which is trained using the following formula:

[0051]

[0052] Among them, L represents the loss function of the GPT model, x i represents the i-th word in the input text, x 1:i-1 Represents the first i-1 words of the input text;

[0053] The vector generation process includes text input, feature extraction, and vector output. Text input involves inputting the picture book text that students are reading into the model. The feature extraction model uses a multi-level attention mechanism and transformation network to extract the multi-dimensional features of the text, including vocabulary, syntax, semantics, and sentiment. The calculation formula of the attention mechanism is as follows:

[0054]

[0055] Where Q, K, and V represent query, key, and value matrices respectively, and d k The vector output model outputs an innovative sentence vector representing the dimension of the key vector. This vector comprehensively reflects the surface information and deep semantic structure of the text. The sentence vector generated by the vector application is used for subsequent interest tracking and recommendation generation, providing a solid data foundation for accurately identifying students' interests and recommending suitable picture books.

[0056] As a further description of the above technical solution:

[0057] The HSMM is used to model the changes in students’ interests by introducing state duration distribution, which allows students to stay in different latent states for different lengths of time, thereby better simulating the heterogeneity of students’ interests and focus;

[0058] In HSMM, multiple potential interest states are defined, each corresponding to a student's main interest tendency within a certain period of time. These states are dynamically adjusted based on students' reading behavior and preferences to adapt to changes in students' interests. Multiple states are defined for fairy tales, popular science knowledge, and literary classics, each representing a student's main interest tendency within a certain period of time.

[0059] The HSMM model introduces state duration distribution, allowing students to stay in different potential states for different lengths of time to better simulate the heterogeneity of students' interests and focus. This modeling approach enables the HSMM to more accurately capture the dynamic changes in students' interests and improve the accuracy and timeliness of recommendation results.

[0060] To construct the HSMM model, it is necessary to set the distribution probability of students' initial interest states, define the transition probability between states, and define the probability of observing different data in each state. The initial state distribution is estimated by analyzing students' historical reading behavior and preferences. The state transition probability matrix reflects the dynamic changes of students' interests between different states, and the observation probability matrix reflects the correlation between students' interest states and observed data.

[0061] Through iterative model updates, HSMM can dynamically track changes in students' interest states and capture interest drift in real time. It uses the Viterbi algorithm to decode the state sequence. The algorithm predicts students' interest states by calculating the most likely state sequence. The Viterbi algorithm calculates the most likely state sequence using the following formula:

[0062]

[0063] Among them, δ t (j) represents the maximum probability of state j at time t, a ij represents the probability of state i transitioning to state j, b j (o t ) indicates that data o is observed in state j t probability;

[0064] Through the Viterbi algorithm, HSMM can accurately predict students' interest status and provide a more reliable data basis for personalized picture book recommendations. The dynamic tracking mechanism enables the system to be implemented.

[0065] As a further description of the above technical solution:

[0066] The expectation maximization (EM) algorithm uses the forward-backward algorithm to calculate the posterior probability, which is calculated using the following formula:

[0067]

[0068] Among them, γ t (i) represents the posterior probability of state i at time t, α t (i) and β t (i) represents the forward probability and backward probability respectively, and P(O|λ) represents the probability of the observed data O under the model parameter λ.

[0069] As a further description of the above technical solution:

[0070] In the M step, based on the results calculated in the E step, the model parameters are updated to maximize the likelihood function of the observed data. The model parameters and initial state distribution are updated using the following formula:

[0071]

[0072] State transition probability matrix:

[0073]

[0074] Status duration distribution:

[0075]

[0076] Observation probability matrix:

[0077]

[0078] Among them, duration(t) represents the duration of the state, I(o t =k) ​​represents the observed data o t Is it equal to k? The parameter estimation process includes initialization parameters, iterative optimization and parameter output. The initialization parameters include setting the initial parameters of the HSMM model, the initial state distribution, the transition probability matrix and the observation probability matrix. The iterative optimization is to optimize the model parameters step by step by alternating E steps and M steps until the parameters converge. The parameter output is to output the optimized model parameters for subsequent interest state prediction and recommendation generation. Through the iterative optimization of the EM algorithm, the estimation accuracy of the HSMM model parameters is improved, thereby improving the accuracy of the prediction of students' interest states.

[0079] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0080] 1. In this invention, by deeply integrating innovative sentence vectors and hidden semi-Markov models (HSMMs), we successfully constructed an efficient and accurate personalized picture book recommendation system, effectively solving the problem of students' interest drifting during the reading process. The innovative sentence vectors can comprehensively capture the deep semantics and emotional tendencies of students' reading texts, providing a rich and multi-dimensional data foundation for interest analysis. The HSMM model allows students to stay in different potential interest states for different lengths of time, accurately depicting the heterogeneity of students' interest duration, and realizing dynamic tracking of students' interest changes.

[0081] 2. In this invention, recommendation accuracy is improved. In terms of picture book recommendation accuracy, the system's average accuracy has increased by 15%-20%, significantly reducing the situation where recommendation results do not match students' actual interests. In fairy tale picture book recommendations, the accuracy of traditional methods is 65%, while the accuracy of the system of this invention reaches over 80%. This improvement is due to the ability of innovative sentence vectors to capture the deep semantics and emotional tendencies of texts, and the HSMM model's accurate portrayal of the duration of students' interests. By combining the two, the system can more accurately identify students' interest status, thereby generating a recommendation list that better meets students' needs.

[0082] 3. In the present invention, reading interest is enhanced. Through tracking and analyzing students' reading behavior, it is found that after using the system, students' reading frequency and reading time increased by 20% and 25% respectively. This data shows that the picture books recommended by the system are more in line with students' interests and needs, and effectively stimulate students' enthusiasm for reading. The system tracks students' interest changes in real time and dynamically adjusts the recommendation strategy to ensure that the recommended picture books always match students' current interests. This personalized recommendation method not only improves students' reading interest, but also promotes the formation of students' reading habits.

[0083] 4. In the present invention, learning effects are improved. In terms of learning effects, the average scores of students using the system in reading comprehension tests increased by 10%-15%. In particular, in the recommendation of popular science picture books, students' mastery of relevant knowledge was significantly improved, and the test scores increased by an average of 12%. This improvement is due to the system's accurate identification of students' interests and personalized recommendations, which enables students to access more learning materials that meet their interests and needs, thereby improving learning effects.

[0084] 5. In the present invention, the system is adaptable in real time. It can capture changes in students' interests in real time and dynamically adjust the recommendation strategy to ensure that the recommended picture books always match the students' current interests. When students switch from fairy tale picture books to science picture books, the system can quickly identify this change and adjust the recommended content accordingly. This real-time adaptability enables the system to better meet students' personalized needs and improve the recommendation effect.

[0085] 6. In the present invention, the personalized recommendation system comprehensively considers the students' multi-dimensional data, including reading content, historical records, grades, practice mastery and oral practice performance, to achieve highly personalized recommendations. Each student's recommendation list is generated based on their unique interests and needs, which greatly improves the pertinence and effectiveness of the recommendation. This personalized recommendation method can not only improve students' reading interest, but also promote their all-round development.

[0086] 7. In the present invention, the user experience is optimized, the system interface is friendly, the recommendation results are clearly displayed, and detailed picture book introductions and reading suggestions are provided, which enhances the user's interactive experience. Student feedback shows that after using the system, their satisfaction with the recommendation results increased by 30%. This user experience optimization not only improves students' satisfaction with use, but also promotes students' continued use of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] Figure 1 This is a flow chart of a personalized picture book intelligent recommendation system that combines innovative sentence vectors with a hidden semi-Markov model proposed by the present invention. DETAILED DESCRIPTION

[0088] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0089] Please see the attached Figure 1 , the present invention provides a technical solution:

[0090] Step 1, data collection:

[0091] Identify relevant data tables in the education platform database that contain student reading content, history, grades, practice mastery, and oral practice data. These data tables include but are not limited to student information tables, reading record tables, grade tables, practice record tables, and oral practice tables. The student information table stores basic student information, such as student ID, name, and grade. The reading record table records students' reading behavior, including the picture book text read, reading time, and reading frequency. The grade table records students' scores in various subjects and comprehensive evaluations. The practice record table records students' performance in various exercises. The oral practice table records students' oral expression ability and progress. By identifying these data tables, ensure the comprehensiveness and accuracy of data collection;

[0092] Extract the required data from the database using SQL queries or other data extraction tools. Specific query statement examples:

[0093] SELECT

[0094] student_id, reading_content, reading_duration, reading_frequency, historical_records, grades, practice_performance, oral_practice;

[0095] FROM

[0096] student_data_table;

[0097] WHERE

[0098] student_id=[specific student ID];

[0099] In the above query, student_id represents the student's unique identifier, reading_content represents the picture book text the student is reading, reading_duration represents the student's reading duration, reading_frequency represents the student's reading frequency, historical_records represents the student's historical reading records, grades represents the student's grades, practice_performance represents the student's practice mastery, and oral_practice represents the student's oral practice performance. By executing this query, relevant data for a specific student ID is extracted from the database.

[0100] Integrate the extracted data into a unified data structure, such as a DataFrame, for subsequent processing. DataFrame is a commonly used data structure that facilitates data cleaning, data analysis, and data visualization. When integrating data, ensure that the data format and type are consistent for subsequent processing and analysis.

[0101] Clean the collected data, remove noise and invalid data, and perform normalization to ensure data consistency and availability. Specifically, the data cleaning process includes identifying and removing outliers, missing values, and duplicate records. The data normalization process includes standardizing numerical data to conform to the standard normal distribution and normalizing non-numerical data using methods such as one-hot encoding or word embedding. In addition, the system uses data integration and data conversion technologies to integrate and process data from different sources and formats to provide more comprehensive and accurate data support.

[0102] The system also ensures data integrity and consistency through data verification and quality control. Specifically, the system uses data validation and anomaly detection technologies to assess and monitor data quality, promptly identifying and addressing data quality issues to ensure data accuracy and reliability. In addition, the system uses data encryption and secure storage technologies to protect the privacy and security of student data and prevent data leakage and abuse.

[0103] In summary, through multi-dimensional data collection, data cleaning, data normalization, data verification, and quality control measures, the system ensures the consistency, availability, and security of the data, providing a solid data foundation for subsequent interest analysis and recommendation generation;

[0104] Step 2, sentence vector generation:

[0105] Text input:

[0106] Text preprocessing:

[0107] Word segmentation, segment the reading content of students. Use a tokenizer suitable for Chinese, such as Jieba. The Jieba tokenizer can effectively split Chinese text into a sequence of words, providing a basis for subsequent text processing;

[0108] Stop word removal, remove stop words in the text to reduce noise. Stop words refer to words with high frequency but weak meaning in the text, such as "de", "he", and "shi". Removing stop words reduces interference in understanding the semantics of the text;

[0109] Text cleaning, remove irrelevant information such as punctuation marks and special characters in the text. Through regular expressions or other text cleaning techniques, remove punctuation marks and special characters in the text to improve the neatness and readability of the text;

[0110] Input formatting, convert the preprocessed text into the input format required by the model, such as Tensor or PyTorch tensor. Convert the text into the input format required by the model, such as converting a text sequence into a sequence of word vectors and converting it into tensor form for the model to perform calculations and inferences;

[0111] Vector generation:

[0112] Model selection, select an efficient large model for inference, such as BERT or GPT, to ensure that the model has strong semantic understanding ability. Models such as BERT and GPT, based on large-scale pre-training, can deeply understand the semantics and emotions of the text, providing an accurate basis for subsequent interest analysis and recommendation;

[0113] Model loading, load the pre-trained model weights to ensure that the model is in the best state. The pre-trained model weights contain a large amount of language knowledge and semantic information, which can improve the performance and accuracy of the model;

[0114] Vector calculation:

[0115] Forward propagation, input the formatted text into the model for forward propagation calculation. For example, for the BERT model, use the following formula for forward propagation:

[0116] BERT(x)=Transformer(Embedding(x))

[0117] BERT(x) represents the output of the BERT model, Transformer represents the core component of the BERT model, and Embedding(x) represents the word vector representation of the text. Through forward propagation, the model can perform deep semantic analysis and feature extraction on the input text.

[0118] Vector extraction: Extract innovative sentence vectors from the model's output layer. These vectors contain rich semantic and sentiment information. For example, for the BERT model, the following formula is used to extract sentence vectors:

[0119] sentence_vector=pooling(BERT(x))

[0120] Here, sentence_vector represents a sentence vector, and pooling represents a pooling operation, which is used to extract sentence-level representations from the output of the BERT model. Through the pooling operation, the model output is converted into a fixed-length vector for subsequent interest analysis and recommendation.

[0121] Vector storage: Store the generated sentence vectors in memory or a database for subsequent use. Store the generated sentence vectors in memory or a database for subsequent interest tracking and recommendation generation.

[0122] Step 3, interest tracking:

[0123] Based on students' interest tendencies, we define multiple potential interest states, such as fairy tales, popular science knowledge, and literary classics. These states reflect students' interest preferences at different time periods and can more accurately capture the dynamic changes in students' interests. The fairy tale state corresponds to students' interest in fairy tale picture books, the popular science knowledge state corresponds to students' interest in popular science picture books, and the literary classic state corresponds to students' interest in literary classic picture books.

[0124] Set the initial probability distribution for each state, such as a uniform distribution or a distribution based on prior knowledge. Uniform distribution means that the initial probability of each state is equal, which is suitable for situations without prior knowledge. The distribution based on prior knowledge is adjusted according to the student's historical reading behavior and preferences to improve the accuracy of the model. For example, if the student has tended to read fairy tale picture books in the past, the initial probability of the fairy tale state is set higher;

[0125] The transition probabilities between states are set to ensure that the self-transition probability of each state is zero. A zero self-transition probability means that students will not stay in the same state but will move to other states. This setting better simulates the dynamic changes in students' interests. The non-self-transition probabilities are adjusted based on students' historical reading behavior and preferences to improve the accuracy of the model. If students frequently switch from fairy tale picture books to popular science picture books, the transition probability from the fairy tale state to the popular science state is set higher;

[0126] The probability of observing different data in each state is set to reflect the correlation between the state and the observed data. The observed data are the student's reading content, history, grades, practice mastery, and oral practice performance. The observation probability matrix is ​​adjusted according to the student's historical reading behavior and preferences to improve the accuracy of the model. If the student is more inclined to read popular science picture books in the popular science knowledge state, then the probability of observing popular science picture books in the popular science knowledge state is set higher;

[0127] Set the initial parameters of the EM algorithm, such as the number of iterations and the convergence threshold. The number of iterations determines the number of times the EM algorithm runs, and the convergence threshold determines when the algorithm stops iterating. These parameters are adjusted according to actual conditions to improve the accuracy of the model.

[0128] Based on the current model parameters, calculate the posterior probability of observing the current data in each state. The posterior probability reflects the probability of the student being in each state given the observed data. The posterior probability is calculated using the forward-backward algorithm. The forward-backward algorithm is a dynamic programming algorithm that recursively calculates the forward and backward probabilities to obtain the posterior probability.

[0129] Estimate the duration distribution of each state. The state duration distribution reflects the length of time students stay in each state. Use the exponential distribution to estimate the state duration. The exponential distribution is a continuous probability distribution that models the duration of the state well.

[0130] Based on the results of the E-step calculation, the model parameters are updated to maximize the likelihood function of the observed data. The likelihood function reflects the probability of the observed data appearing given the model parameters. By maximizing the likelihood function, more accurate model parameters can be obtained. For example, the transition probability matrix is ​​updated using the following formula:

[0131]

[0132] Among them, a ij represents the probability of transitioning from state i to state j, γ t (i) and β t (i) represents the posterior probability and forward probability respectively;

[0133] Repeat the E and M steps until the model parameters converge. Model parameter convergence means that the change of model parameters is less than the preset threshold. At this time, the model has reached a stable state.

[0134] Based on the current data and model parameters, the student's current interest state is predicted and the state sequence is decoded using the Viterbi algorithm. The Viterbi algorithm is a dynamic programming algorithm that recursively calculates the maximum probability path to obtain the most optimal state sequence.

[0135] Record the duration of each state for subsequent interest drift analysis. The state duration reflects the time students stay in each state, thereby better understanding the dynamic changes of students' interests.

[0136] Step 4, recommendation generation:

[0137] Interest Matching:

[0138] State mapping: Mapping the predicted interest state to the corresponding picture book recommendation strategy, such as fairy tale state corresponding to fairy tale picture book recommendation. Specifically, the system selects picture books related to the student's current interest state from the preset picture book library, such as popular science knowledge, fairy tales, and literary classics. For example, if the student's current interest state is popular science knowledge, the system will prioritize selecting picture books that match the student's interests from the popular science picture book library;

[0139] Strategy selection: Based on students' interests, the system selects the optimal recommendation strategy to ensure the accuracy of recommendations. The system considers students' historical reading data and multi-dimensional information of personalized needs, and further screens and sorts the selected picture books to ensure the accuracy and personalization of the recommendation results. The system adjusts the type and difficulty of recommended picture books based on students' historical reading preferences to meet students' personalized needs.

[0140] Generate a recommendation list:

[0141] Data fusion: The system integrates and analyzes students' historical reading data, current interests, and personalized needs. The system comprehensively considers students' historical reading behavior, current interests, and personalized needs to generate a recommendation list containing multiple picture books. The system adjusts the type and difficulty of recommended picture books based on students' historical reading preferences to meet students' personalized needs.

[0142] Recommendation algorithm application, applies collaborative filtering and content recommendation algorithms to generate a recommendation list containing multiple picture books, and uses the following formula for collaborative filtering recommendation:

[0143]

[0144] Among them, score(i, j) represents the similarity between picture book i and picture book j, r ui represents the rating of user u on picture book i, μ i represents the average rating of picture book i, and U represents the set of users;

[0145] List sorting: sorting the recommendation list based on recommendation scores or relevance to ensure the priority of recommendation results. The system sorts the recommendation list based on the recommendation scores output by the recommendation algorithm or the relevance between picture books to ensure the priority of recommendation results. The system puts picture books with higher recommendation scores at the front of the recommendation list to improve the accuracy and effectiveness of the recommendation.

[0146] Recommended output:

[0147] Interface display: The recommendation list is presented to students through the education platform interface, providing intuitive recommendation results. The system presents the recommendation list to students in the form of pictures and texts, including multiple information such as the picture book cover, title, introduction and author. The system displays the recommendation list in the form of cards on the interface, and each card contains information such as the picture book cover, title and introduction;

[0148] Reading suggestions: Provide detailed introductions and reading suggestions for each recommended picture book to enhance the interactivity and practicality of the recommendation. Based on the individual needs of students, the system provides corresponding reading suggestions for each recommended picture book, such as reading methods and reading skills. For popular science picture books, the system provides relevant popular science knowledge explanations and experimental operation guides.

[0149] Feedback collection: collect students' feedback on the recommendation results for subsequent model optimization and improvement. The system collects students' feedback on the recommendation results through the education platform interface, such as students' ratings and comments on the recommended picture books. The system sets a feedback button for students to rate and comment on the recommended picture books. By collecting students' feedback, the system continuously optimizes and improves the recommendation algorithm to improve the accuracy and effectiveness of the recommendation.

[0150] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A personalized picture book intelligent recommendation system based on innovative sentence vectors combined with hidden semi-Markov models, characterized by: It includes data acquisition module, sentence vector generation module, interest tracking module, recommendation generation module, innovative sentence vector, hidden semi-Markov model and expectation maximization algorithm; The data collection module ensures data consistency, availability, and security through multi-dimensional data collection, data cleaning, data normalization, data verification, and quality control, providing a solid data foundation for subsequent interest analysis and recommendation generation. The sentence vector generation module uses an efficient reasoning model to perform deep vectorization processing on students' reading content, where the efficient reasoning model uses BERT or GPT; The interest tracking module is based on the Hidden Semi-Markov Model (HSMM) and models the student's interest changes as a transition process of latent states. The recommendation generation module includes interest matching strategy, recommendation list generation and recommendation output mode; The innovative sentence vectors use the efficient inference large-scale models BERT or GPT. These models have powerful natural language processing capabilities and can deeply understand the semantic structure and sentiment of text. The BERT model uses a bidirectional Transformer structure and generates sentence vectors containing rich semantic information through pre-training and fine-tuning. The pre-training tasks of the BERT model include masked language modeling and next sentence prediction. The hidden semi-Markov model is used to model the changes in students' interests. By introducing state duration distribution, students are allowed to stay in different potential states for different lengths of time, so as to better simulate the heterogeneity of students' interests and focus. The expectation-maximization (EM) algorithm gradually approaches the true value of the model parameters by alternately performing the expectation step (E) and the maximization step (M). In the E step, the posterior probability distribution of the potential state under the current parameters is calculated.

2. The personalized picture book intelligent recommendation system based on innovative sentence vectors combined with hidden semi-Markov models according to claim 1 is characterized in that: The data collection module acquires multi-dimensional student data from the education platform database, including but not limited to reading content, historical reading records, academic performance, practice mastery, and oral practice performance. Specifically, reading content data includes the text of picture books read by students, reading time, and reading frequency; historical records cover students' past reading behaviors and preferences; performance data includes grades in various subjects and comprehensive evaluations; practice mastery reflects students' performance in various exercises; and oral practice data records students' oral expression ability and progress. In addition, the system also collects data from students' reading preferences, learning habits, and cognitive levels to more comprehensively understand students' individual needs and learning characteristics. The collected data is cleaned, noise and invalid data are removed, and normalized to ensure data consistency and availability. Specifically, the data cleaning process includes identifying and removing outliers, missing values, and duplicate records in the data. The data normalization process includes standardizing numerical data to make it conform to the standard normal distribution, and normalizing non-numerical data, such as one-hot encoding or word embedding representation. In addition, the system also uses data integration and data conversion technologies to integrate and process data from different sources and formats to provide more comprehensive and accurate data support.

3. The personalized picture book intelligent recommendation system based on innovative sentence vectors combined with hidden semi-Markov models according to claim 2 is characterized in that: The BERT model calculates the vector representation of text using the following formula: BERT(x)=Transformer(Embedding(x)) Here, x represents the input text, Embedding(x) is the word vector representation of the text, and Transformer is the core component of the BERT model, responsible for extracting the deep features of the text. The generated sentence vector not only contains the surface vocabulary information of the text, but also captures multiple deep features such as sentence structure, semantic relationship, and emotional expression, providing a rich information foundation for subsequent interest analysis and recommendation. The BERT model is trained through the pre-training tasks of Masked Language Model (MLM) and Next Sentence Prediction (NSP), using the following formula: Among them, L represents the loss function of the BERT model, x i represents the i-th word in the input text, x mask To represent masked words, the GPT model is based on a unidirectional Transformer structure. Through large-scale pre-training, it generates sentence vectors with generative and semantic understanding capabilities. The pre-training task of the GPT model includes language modeling (LM), which is trained using the following formula: Among them, L represents the loss function of the GPT model, x i represents the i-th word in the input text, x 1:i-1 Represents the first i-1 words of the input text. The vector generation process includes text input, feature extraction, and vector output. The text input is the picture book text read by the students input into the model. The feature extraction model extracts the multi-dimensional features of the text including vocabulary, syntax, semantics, and emotion through a multi-level attention mechanism and transformation network. The calculation formula of the attention mechanism is as follows: Where Q, K, and V represent query, key, and value matrices respectively, and d k The vector output model outputs an innovative sentence vector representing the dimension of the key vector. This vector comprehensively reflects the surface information and deep semantic structure of the text. The sentence vector generated by the vector application is used for subsequent interest tracking and recommendation generation, providing a solid data foundation for accurately identifying students' interests and recommending suitable picture books.

4. The personalized picture book intelligent recommendation system based on innovative sentence vectors combined with hidden semi-Markov models according to claim 3 is characterized in that: The HSMM model is defined by the following parameters: Initial state distribution π=(π1,π2,...,π N ), where π i represents the probability that the student is initially in state i. The initial state distribution is estimated by analyzing the student's historical reading behavior and preferences. Based on the type and number of picture books the student has read in the past period of time, the initial probability of each type of picture book is calculated; State transition probability matrix A=[a ij ] N×N , where a ij represents the probability of a student transferring from state i to state j. The state transition probability matrix reflects the dynamic changes of students' interests between different states. If a student switches from reading fairy tale picture books to reading science picture books, the value of the corresponding element in the state transition probability matrix will increase. State duration distribution D = (D1, D2, ..., D N ), where D i represents the duration distribution of students in state i. The state duration distribution allows students to stay in different potential states for different lengths of time, which is used to better simulate the heterogeneity of students' interests and focus. Students may be interested in picture books on a specific theme for a period of time, and more interested in picture books on other themes at another time. Observation probability matrix B = [b ik ] N×M , where b ik It represents the probability of observing data k when the student is in state i. The observation probability matrix reflects the correlation between the student's interest state and the observed data. When students observe data related to popular science knowledge when reading popular science picture books, the value of the corresponding element in the observation probability matrix will increase.

5. The personalized picture book intelligent recommendation system based on innovative sentence vectors combined with hidden semi-Markov models according to claim 4 is characterized in that: Define multiple potential interest states, each representing a student's primary interest within a specific timeframe, including fairy tales, popular science, and literary classics. These states are dynamically adjusted based on the student's reading behavior and preferences to accommodate changes in student interests. The model allows students to stay in different potential states for different lengths of time. By capturing the heterogeneity of state duration, it accurately reflects the dynamic changes in students' interests. For example, students may be very interested in picture books on a specific theme for a period of time, and more interested in picture books on other themes at another time. In order to comprehensively evaluate the student's current interest status, the sentence vector is integrated with other multivariate data for analysis, and the student's comprehensive interest score is calculated using the following formula: Among them, S represents the student's comprehensive interest score, score(k) represents the score of the student's observed data k. By comprehensively considering the student's initial state distribution, state transition probability matrix, state duration distribution and observation probability matrix, it is used to more accurately evaluate the student's interest state and provide a more reliable data basis for personalized picture book recommendations.

6. The personalized picture book intelligent recommendation system combining innovative sentence vectors and hidden semi-Markov models according to claim 5 is characterized in that: The interest matching strategy uses an intelligent matching algorithm based on the current interest status output by the interest tracking module to dynamically adjust the picture book recommendation strategy. The system first selects a set of picture books related to the student's current interest status from a preset picture book library. If the student's current interest status is popular science knowledge, the system will prioritize selecting picture books that match the student's interest from the popular science picture book library. In addition, the system also considers the student's historical reading data and multi-dimensional information of personalized needs to further screen and sort the selected picture books to ensure the accuracy and personalization of the recommendation results. The recommendation list is generated by combining students' historical reading data, current interest status, and personalized needs, and adopting an intelligent recommendation algorithm to generate a recommendation list containing multiple picture books. The system first analyzes students' reading preferences and habits, reading time, reading frequency, and reading types based on their historical reading data, and builds a personalized reading profile of the students. Then, the system scores and sorts the picture books in the picture book library based on the students' current interest status and personalized reading profile. The scoring formula is as follows: Among them, R(p) represents the recommendation score of picture book p, a ij and b ik They represent the state transition probability and observation probability respectively, and score(k) represents the score of the student observing data k. Finally, the system sorts the picture books according to the recommendation scores and generates a recommendation list containing multiple picture books; In the recommendation output method, the recommendation list is displayed to students through the education platform interface, and detailed picture book introductions and reading suggestions are provided to enhance the interactivity and practicality of the recommendation. The system displays the recommendation list to students in the form of pictures and texts, including picture book cover, title, introduction and author information. In addition, the system also provides corresponding reading suggestions, reading methods and reading skills based on the personalized needs of students. For science picture books, the system is used to provide relevant science knowledge explanations and experimental operation guides. The output method improves students' reading interest and promotes the improvement of students' reading ability and knowledge level.

7. The personalized picture book intelligent recommendation system combining innovative sentence vectors and hidden semi-Markov models according to claim 6 is characterized in that: The pre-training tasks of the BERT model include the masked language model (MLM) and the next sentence prediction (NSP), which are trained using the following formula: Among them, L represents the loss function of the BERT model, x i represents the i-th word in the input text, x mask To represent masked words, the GPT model is based on a unidirectional Transformer structure. Through large-scale pre-training, it generates sentence vectors with generative and semantic understanding capabilities. The pre-training task of the GPT model includes language modeling (LM), which is trained using the following formula: Among them, L represents the loss function of the GPT model, x i represents the i-th word in the input text, x 1:i-1 Represents the first i-1 words of the input text; The vector generation process includes text input, feature extraction, and vector output. Text input involves inputting the picture book text that students are reading into the model. The feature extraction model uses a multi-level attention mechanism and transformation network to extract the multi-dimensional features of the text, including vocabulary, syntax, semantics, and sentiment. The calculation formula of the attention mechanism is as follows: Where Q, K, and V represent query, key, and value matrices respectively, and d k The vector output model outputs an innovative sentence vector representing the dimension of the key vector. This vector comprehensively reflects the surface information and deep semantic structure of the text. The sentence vector generated by the vector application is used for subsequent interest tracking and recommendation generation, providing a solid data foundation for accurately identifying students' interests and recommending suitable picture books.

8. The personalized picture book intelligent recommendation system combining innovative sentence vectors and hidden semi-Markov models according to claim 7 is characterized in that: The HSMM is used to model the changes in students’ interests by introducing state duration distribution, which allows students to stay in different latent states for different lengths of time, thereby better simulating the heterogeneity of students’ interests and focus; In HSMM, multiple potential interest states are defined, each corresponding to a student's main interest tendency within a certain period of time. These states are dynamically adjusted based on students' reading behavior and preferences to adapt to changes in students' interests. Multiple states are defined for fairy tales, popular science knowledge, and literary classics, each representing a student's main interest tendency within a certain period of time. The HSMM model introduces state duration distribution, allowing students to stay in different potential states for different lengths of time to better simulate the heterogeneity of students' interests and focus. This modeling approach enables the HSMM to more accurately capture the dynamic changes in students' interests and improve the accuracy and timeliness of recommendation results. To construct the HSMM model, it is necessary to set the distribution probability of students' initial interest states, define the transition probability between states, and define the probability of observing different data in each state. The initial state distribution is estimated by analyzing students' historical reading behavior and preferences. The state transition probability matrix reflects the dynamic changes of students' interests between different states, and the observation probability matrix reflects the correlation between students' interest states and observed data. Through iterative model updates, HSMM can dynamically track changes in students' interest states and capture interest drift in real time. It uses the Viterbi algorithm to decode the state sequence. The algorithm predicts students' interest states by calculating the most likely state sequence. The Viterbi algorithm calculates the most likely state sequence using the following formula: Among them, δ t (j) represents the maximum probability of state j at time t, a ij represents the probability of state i transitioning to state j, b j (o t ) indicates that data o is observed in state j t probability; Through the Viterbi algorithm, HSMM can accurately predict students' interest status and provide a more reliable data basis for personalized picture book recommendations. The dynamic tracking mechanism enables the system to be implemented.

9. The personalized picture book intelligent recommendation system combining innovative sentence vectors and hidden semi-Markov models according to claim 8 is characterized in that: The expectation maximization (EM) algorithm uses the forward-backward algorithm to calculate the posterior probability, which is calculated using the following formula: Among them, γ t (i) represents the posterior probability of state i at time t, α t (i) and β t (i) represents the forward probability and backward probability respectively, and P(O|λ) represents the probability of the observed data O under the model parameter λ.

10. The personalized picture book intelligent recommendation system combining innovative sentence vectors and hidden semi-Markov models according to claim 9 is characterized in that: In the M step, based on the results calculated in the E step, the model parameters are updated to maximize the likelihood function of the observed data. The model parameters and initial state distribution are updated using the following formula: State transition probability matrix: Status duration distribution: Observation probability matrix: Among them, duration(t) represents the duration of the state, I(o t =k) ​​represents the observed data o t Is it equal to k? The parameter estimation process includes initialization parameters, iterative optimization and parameter output. The initialization parameters include setting the initial parameters of the HSMM model, the initial state distribution, the transition probability matrix and the observation probability matrix. The iterative optimization is to optimize the model parameters step by step by alternating E steps and M steps until the parameters converge. The parameter output is to output the optimized model parameters for subsequent interest state prediction and recommendation generation. Through the iterative optimization of the EM algorithm, the estimation accuracy of the HSMM model parameters is improved, thereby improving the accuracy of the prediction of students' interest states.

Images