A reading content reorganization method based on reading consistency

Abstract

The present invention relates to the technical field of reading content reorganization, and discloses a reading content reorganization method based on reading consistency, comprising: executing a server task allocation strategy to allocate a task amount to each server based on feature data; executing a user reading behavior modeling strategy to extract an embedding vector of a historical reading behavior sequence, and extracting the user's reading preference based on the embedding vector; if the article the user is currently reading belongs to the user's reading preference, executing an article paragraph semantic analysis strategy to extract the semantics of each paragraph in the article and calculate the article's coherence score; then executing a user preference calculation strategy to calculate an interest graph based on the user's social behavior and emotional state; executing a dynamic reorganization strategy to dynamically calculate the optimal next recommended paragraph; obtaining the user's click rate and dwell time, executing a feedback optimization strategy to optimize the selection of recommended paragraphs. This provides an intelligent and personalized reading experience and increases the user's reading interest.

Description

Technical Field

[0001] The present invention relates to the technical field of reading content reorganization, and in particular to a reading content reorganization method based on reading consistency. Background Art

[0002] In an era of information explosion, users often struggle to grasp the most important information when reading long articles, or their reading experience is compromised by unclear logic. Therefore, interest-based reading content reorganization technology has emerged. By analyzing users' reading behavior and preferences, it can intelligently adjust the order of article content to better align with their interests, improving information retrieval efficiency. Furthermore, combined with natural language processing technology, it can assess the coherence between paragraphs, optimize content logic, and ensure smooth reading. Furthermore, reading content reorganization can help users find key information faster, reduce wasted reading time, and enhance the user's reading experience.

[0003] Most recommendation systems recommend content based solely on users' historical clicks and reading preferences, but they still have shortcomings in the organization and logical optimization of reading content. On the one hand, existing technologies mainly focus on recommending articles as a whole, while ignoring the optimization of the internal structure of articles, resulting in recommended content that cannot adapt to the user's reading rhythm. On the other hand, current text reorganization methods are mostly based on keyword matching or simple semantic similarity calculations, which makes it difficult to accurately grasp the logical relationship between article paragraphs, resulting in possible breaks or logical confusion in the reorganized content. In addition, existing methods capture user interests in a relatively static manner and are unable to adjust recommended content in real time, resulting in poor personalized recommendation results. Therefore, there is still a lot of room for optimization in the content reorganization of existing reading recommendation systems, and more intelligent strategies are needed to improve the reading experience.

[0004] This solution proposes a reading content reorganization method based on reading consistency. Based on user interest graph and natural language processing technology, articles are intelligently reorganized to improve the reading experience. Summary of the Invention

[0005] The present invention provides a reading content reorganization method based on reading consistency, which facilitates solving the problems mentioned in the above background technology.

[0006] In a first aspect, the present application provides a reading content reorganization method based on reading consistency, which adopts the following technical solution: A reading content reorganization method based on reading consistency, comprising:

[0007] Set up m servers;

[0008] Acquire characteristic data of each server, including computing power, execution efficiency, load, and task transmission delay; execute a server task allocation strategy to allocate a task amount to each server based on the characteristic data;

[0009] For any reading user: obtain the user's historical reading behavior sequence before time t;

[0010] The historical reading behavior sequence includes each article that the user has read in the past

[0011] Execute the user reading behavior modeling strategy, extract the embedding vector of the historical reading behavior sequence, and extract the user reading preference based on the embedding vector;

[0012] Determine whether the article the user is currently reading belongs to the user's reading preferences;

[0013] If the article the user is currently reading is in the user's reading preference, the article paragraph semantic analysis strategy is executed to extract the semantics of each paragraph in the article and calculate the coherence score of the article;

[0014] Determine whether the paragraphs of the article can be reconstructed based on the article's coherence score;

[0015] If the article paragraph can be reconstructed, the user preference calculation strategy is executed to calculate the interest graph based on the user's social behavior and emotional state;

[0016] Based on the interest graph, a dynamic reorganization strategy is executed to dynamically calculate the optimal next recommended paragraph;

[0017] When the user finishes reading the current paragraph, recommend the next paragraph to the user;

[0018] Obtain user click-through rate and dwell time, implement feedback optimization strategies, and optimize the selection of recommended paragraphs.

[0019] Optimizing task allocation based on server feature data ensures efficient utilization of computing resources and improves the computational efficiency of the recommendation system. Server computing power, execution efficiency, load, and task transmission latency all affect the system's overall responsiveness and recommendation quality. By collecting this data and implementing a task allocation strategy, we can balance the load, avoid overloading some servers while leaving others idle, and improve system stability and responsiveness. For any user, capturing their historical reading behavior sequence and extracting embedding vectors allows for more accurate modeling of their interests. By calculating whether the currently read article matches the user's interests, we can accurately determine their reading tendencies and inform subsequent content optimization decisions. When article coherence is low, the system dynamically restructures paragraphs to better align with the user's interests, while also improving the readability of recommended content. Ultimately, based on the user's interest graph and dynamic reorganization strategy, the optimal next paragraph is recommended, enhancing the user's immersive reading experience. Furthermore, by capturing user click-through rates and dwell time and implementing feedback optimization strategies, we can further fine-tune the recommendation strategy, making recommendations more precise and improving overall user satisfaction.

[0020] Preferably, the execution server task allocation strategy allocates a task amount to each server based on the characteristic data, including:

[0021] Among them, a i is the amount of tasks assigned to server i, are the computing capabilities of server i and server j respectively, are the execution efficiencies of server i and server j respectively, are the current loads of server i and server j respectively, is the task transmission delay undertaken by server i, A is the total task amount, Combined computing power for all servers.

[0022] Improving server computing efficiency through precise task allocation can effectively reduce task execution time and enhance the stability of the recommendation system. In a multi-server environment, the computing power, execution efficiency, load, and task transmission delay of different servers can vary significantly. Indiscriminate task allocation can lead to under-tasks for servers with strong computing power and overload for servers with weaker computing power, thus affecting overall efficiency. Therefore, task allocation should consider the computing power, execution efficiency, current load, and task transmission delay of each server, and optimize task allocation based on the combined computing power of all servers. By rationally allocating tasks, computing resources are optimally utilized, reducing task wait time and improving the real-time performance of the recommendation system. Furthermore, by optimizing task transmission delay, tasks can be properly scheduled across servers, reducing the waste of computing resources due to communication overhead. Overall, a precise task allocation strategy not only reduces computing costs but also improves the efficiency of the recommendation algorithm, ensuring that users receive efficient and accurate recommendations.

[0023] Preferably, the method of executing a user reading behavior modeling strategy, extracting an embedding vector of a historical reading behavior sequence, and extracting a user reading preference based on the embedding vector includes:

[0024] The historical reading behavior sequence is represented as BOOK = [book1, book2,, book c ], where c is the total number of articles, book i represents the i-th article;

[0025] Use the embedding method to extract the embedding vector v of the i-th article i =Embed(book i ), where Embed(·) is the embedding method function, v i is the embedding vector of the i-th article;

[0026] Then the embedding vector of the historical reading behavior sequence is composed of V = [v1, v2,, v c ], V is the embedding vector matrix extracted from the historical reading behavior sequence;

[0027] Add positional encoding to each embedding vector: Where PosEnc(i) is the position encoding;

[0028] Update the embedding vector matrix Calculate user preferences using the self-attention mechanism:

[0029] Define the self-attention calculation:

[0030] Where W1, W2, and W3 are trainable parameter matrices, and Q1, Q2, and Q3 are query, key, and value matrices, respectively;

[0031] Calculate the self-attention weight: Among them, τ is the dimension of the embedding vector; after multi-layer Transformer calculation, the user preference vector is extracted: Γ=-∑ (i) logP(book i+1 |V), where Γ is the loss function, For users who use model predictions, see the next article book i+1 The probability of p T is the transpose of the user preference vector p, v i+1 For book i+1 The embedding vector, v j For book j Embedding vector of

[0032] Set the loss threshold ε Γ ;

[0033] Get the difference ΔΓ of the loss function update, if ΔΓ<ε Γ , then stop updating the loss function and get the optimal user preference vector p.

[0034] Optimizing personalized recommendations by modeling a user's historical reading behavior can accurately capture user interests and improve the relevance of recommended content. A user's historical reading behavior is one of the most important inputs to a recommendation system, reflecting their long-term interests and preferences. By constructing a sequence of a user's historical reading behavior and using embedding methods to extract vector representations of articles, users' reading habits can be converted into a mathematical model, making it suitable for deep learning model training. Furthermore, positional encoding enhances sequential information, enabling the model to learn the impact of reading order on user interests. Simultaneously, the self-attention mechanism (Transformer) captures the correlations between different articles and extracts deep-level user preference characteristics. Through multi-layer Transformer calculations, a personalized user interest vector is ultimately generated. This vector is then optimized based on a loss function to accurately predict the user's next reading content. Furthermore, setting a loss threshold effectively prevents model overfitting and improves the generalization ability of the recommendation system. Ultimately, this method not only improves the accuracy of the recommendation system but also adapts to the reading habits of different users, providing more personalized recommendation services.

[0035] Preferably, if the article currently being read by the user belongs to the user's reading preference, executing the article paragraph semantic analysis strategy, extracting the semantics of each paragraph in the article, and calculating the coherence score of the article, including:

[0036] The article that the user is currently reading is represented as G = [g1, g2, ..., g d ], where d is the total number of paragraphs in the article, g i Indicates the i-th paragraph;

[0037] Use natural language processing technology to extract the semantic meaning of each paragraph i =f θ (g i ) where f θ (g i ) is the natural language processing model, θ is the parameter of the natural language processing model, u i For paragraph g i semantics;

[0038] Compute the coherence score between any two paragraphs: Among them, u i For paragraph g i The semantic vector, u j For paragraph g j The semantic vector, u i ·u j is the semantic vector u i and semantic vector u j The dot product of ‖u i ‖ is the semantic vector u i The Euclidean norm of j ‖ is the semantic vector u j The Euclidean norm of ;

[0039] Calculate the global coherence score, Setting the coherence threshold

[0040] like Then the article has low coherence and the paragraphs can be reconstructed;

[0041] like Then the article has high coherence and the paragraphs cannot be reconstructed.

[0042] By calculating article coherence and optimizing paragraph reorganization, the user's reading fluency and reading experience can be improved. In recommendation systems, a user's reading experience is influenced not only by the content's relevance to their interests but also by the article's internal coherence. By using natural language processing (NLP) techniques to extract semantic information from each paragraph and calculate the coherence score between paragraphs, the overall logical coherence of the article can be determined. When the article's global coherence score falls below a set threshold, the system performs paragraph reorganization to optimize the article's structure and improve readability. This method effectively improves article content, making it more consistent with users' logical thinking habits, thereby enhancing the reading experience. Furthermore, by calculating the dot product of paragraph semantic vectors, the semantic similarity between paragraphs can be accurately measured, ensuring that the reorganized article remains consistent with the original theme without compromising the overall message conveyed. Ultimately, this method not only improves the quality of recommended content but also increases user retention time, enhancing user trust and satisfaction with the recommendation system.

[0043] Preferably, if the article paragraph can be reconstructed, a user preference calculation strategy is executed to calculate an interest graph based on the user's social behavior and emotional state, including:

[0044] Get the user's activity actions on social media [e1, e2, e3], where e1 is like, e2 is comment, and e3 is forward;

[0045] Building an interest graph Among them, R is the interest graph, They are user like weight, comment weight and forwarding weight respectively. and The number of likes, comments and reposts of users respectively;

[0046] Get the text T posted by the user;

[0047] Use the sentiment analyzer to calculate the user's sentiment score: in, is the sentiment score, f ∈ (·) is the sentiment analyzer, ∈ is the parameter of the sentiment analyzer;

[0048] when The user has positive emotions;

[0049] when The user has negative emotions;

[0050] when Then the user has neutral emotions;

[0051] Combined with sentiment scores, adjust the interest graph: Among them, δ is the adjustment coefficient, which is used to control the influence of sentiment analysis results on the interest graph. f(freq(T)) is a function of how frequently users mention topics in their texts. When the topic frequency is higher, it is given more weight.

[0052] By building an interest graph and combining it with sentiment analysis to optimize recommendation results, we can provide personalized content recommendations that better align with the user's emotional state. A user's activities on social media, such as likes, comments, and reposts, can effectively reflect their interest in different content. By building an interest graph and calculating interest preferences based on the weights of likes, comments, and reposts, we can accurately capture users' long-term interests. Furthermore, sentiment analysis combined with the text content posted by users can identify the user's current emotional state and adjust recommendation strategies accordingly. For example, when a user's sentiment score is positive, the system can recommend more positive content to enhance the user's emotional experience; when the user's sentiment score is negative, the system can avoid recommending content that may exacerbate negative emotions, thereby improving the user's psychological comfort. Ultimately, this method not only improves the accuracy of recommendations but also optimizes the user's emotional experience, making the recommendation system more intelligent and user-friendly.

[0053] Preferably, executing a dynamic reorganization strategy based on the interest graph and dynamically calculating the optimal next recommended paragraph includes:

[0054] The paragraph the user is currently reading is recorded as g * ;

[0055] For any unread paragraph g ′ , calculate paragraph g ′ Matching degree with interest graph:

[0056] Among them, u ′ For paragraph g ′ The semantic vector of R u ′ is the dot product of interest graph and semantic vector, ‖R‖, ‖u ′ ‖ are the moduli of interest graph and semantic vector respectively;

[0057] Calculate the optimal next recommended paragraph g - , Among them, u * For paragraph g * The semantic vector of is a weighting coefficient used to balance the impact of paragraph coherence and interest matching.

[0058] By combining interest graphs and semantic analysis to optimize recommended passages, the matching degree of recommended content can be improved, enhancing the user's reading experience. In recommendation systems, users' interests in different passages can vary significantly, necessitating a refined recommendation strategy. By calculating the matching degree between the currently read passage and the interest graph, we can ensure that the recommended content aligns with the user's long-term interests. Furthermore, when calculating the matching degree, we use the dot product method of the passage's semantic vectors to accurately measure the relevance of the recommended content. Furthermore, to ensure that the recommended passages not only meet interest preferences but also maintain reading fluency, a coherence weighting strategy is employed to achieve an optimal balance between coherence and interest matching in the recommended passages. Ultimately, this method significantly improves the quality of recommended content, ensuring that users maintain a high level of interest throughout the reading process, thereby increasing user stickiness and enhancing the overall effectiveness of the recommendation system.

[0059] Preferably, obtaining the user click rate and dwell time, executing the feedback optimization strategy, and optimizing the selection of recommended paragraphs include:

[0060] The user click-through rate is recorded as CTR, and the user's stay time in the recommended paragraph is recorded as Y;

[0061] Calculate user feedback scores in, Derivative the model parameters click rate and dwell time:

[0062] Among them, π is the model parameters click rate and dwell time, represents the impact of model parameter π on click-through rate, is the impact of model parameters on the length of stay, where p is the quality feature of the recommended paragraph;

[0063] Update parameters in, is the learning rate;

[0064] The selection of recommended paragraphs is optimized using the updated parameters.

[0065] Optimizing the recommendation model based on user feedback can continuously improve the intelligence of the recommendation system and make recommendations more precise. Users' click-through rate and dwell time are important indicators of recommendation quality. Therefore, by calculating user feedback scores and performing gradient updates on the model parameters, the performance of the recommendation system can be optimized. Specifically, by taking the derivatives of click-through rate and dwell time, we can determine the impact of model parameters on user feedback and adjust the model parameters accordingly. Furthermore, using a learning rate to control the update step size prevents model overfitting or slow convergence, thereby improving the stability of the recommendation system. Ultimately, this method can continuously optimize the selection of recommended paragraphs, allowing the recommendation system to continuously evolve as user behavior changes, improving user reading experience and satisfaction.

[0066] The present invention has the following beneficial effects:

[0067] 1. This reading content reorganization method based on reading consistency sets up m servers and obtains characteristic data for each server, including computing power, execution efficiency, load, and task transmission latency. This method effectively allocates tasks to the servers, improving computing resource utilization, reducing server overload, and thus optimizing overall task scheduling efficiency. By acquiring a user's historical reading behavior sequence before time t and integrating it with their reading preferences, it accurately captures their interests and ensures personalized and accurate content recommendations. By implementing a semantic analysis strategy for article paragraphs and extracting article coherence scores, it further optimizes the article presentation and ensures a smooth and logical reading experience. A dynamic reorganization strategy based on the interest graph accurately calculates the optimal next recommended paragraph, ensuring that the recommended content is more aligned with the user's current interests and context, enhancing the user's immersive reading experience. Furthermore, by obtaining user click-through rate and dwell time and implementing a feedback optimization strategy, it continuously adjusts and optimizes recommended content, ensuring that the system continuously adapts to the user's reading habits, improving the accuracy of the recommendation system and user satisfaction.

[0068] 2. This reading content reorganization method based on reading consistency can ensure that the server's computing resources are fully utilized and improve the efficiency of task processing by allocating a reasonable amount of tasks to each server and combining key indicators such as the server's computing power, execution efficiency, current load, and task transmission delay. By introducing the concept of comprehensive server computing power, the overall task scheduling is more scientific and reasonable, and can effectively avoid the waste or overload of server resources. By establishing a mathematical relationship between the task volume and the server characteristics, a quantitative calculation basis can be provided for task allocation, making task allocation more fair and efficient. By optimizing the task scheduling strategy, not only can the system's computing efficiency be improved, but also the task transmission delay can be reduced, and the response speed of the overall server cluster can be improved, thereby enhancing the user experience.

[0069] 3. This reading content reorganization method based on reading consistency can better capture users' long-term interests and reading habits by converting the user's historical reading behavior sequence into an embedding vector matrix, so that the recommendation system can more accurately match the user's needs. By adding position encoding to the embedding vector, the information of the article reading order can be retained, the model's ability in sequence modeling can be improved, and the recommendation results can be more consistent with the user's reading logic. By calculating user preferences through the self-attention mechanism, it is possible to more accurately capture the user's long-term interests, and extract more representative user preference vectors through multi-layer Transformer, thereby improving the accuracy of user interest modeling. By setting the loss function and adopting an optimization strategy, it can ensure that the final extracted user preference vector is optimal, thereby improving the recommendation system's matching degree with the user's reading interests and optimizing the recommendation effect.

[0070] 4. This reading content reorganization method based on reading consistency divides the article currently read by the user into multiple paragraphs and uses natural language processing technology to extract the semantic information of each paragraph. This can provide a deeper understanding of the structure and content of the article and improve the recommendation system's perception of the article content. By calculating the coherence score between article paragraphs, the logical structure of the article can be effectively evaluated to ensure the rationality and readability of the article content. By calculating the global coherence score, it is possible to determine whether the overall logic of the article conforms to the user's reading habits, and by setting a threshold, it is determined whether the article paragraphs need to be reconstructed to improve the readability and coherence of the article. By optimizing the article content structure, the user's reading experience can be improved, the reading content can be made more consistent with the user's logical thinking, and the efficiency of information acquisition can be improved.

[0071] 5. This reading content reorganization method based on reading consistency can more comprehensively analyze users' social interactions by obtaining users' likes, comments, and forwarding behaviors on social media, build an interest graph based on users' social behavior patterns, and improve the recommendation system's ability to capture user interests. By setting weights for different social behaviors and combining the user's text content on social platforms, the user's emotional state can be calculated more accurately, ensuring that the recommended content not only meets the user's interests but also matches the user's current emotions. By calculating the user's emotional score through the sentiment analyzer, the user's emotional state can be identified, and the interest graph can be adjusted accordingly to improve the contextual adaptability of the recommended content. By optimizing the interest graph based on the results of sentiment analysis, the recommended content can be made to meet the user's long-term interests and match the user's short-term emotional fluctuations, thereby improving the user's acceptance and satisfaction with the recommended content.

[0072] 6. This reading content reorganization method based on reading consistency can ensure that the next recommended paragraph meets the user's interests by calculating the matching degree between the paragraph the user is currently reading and the interest graph, thereby improving the relevance of the recommended content. By using the dot product method to calculate the matching degree, it is possible to maintain high computational efficiency while ensuring the accuracy of the recommendation results. By combining the weighted calculation method of the coherence score and the interest matching degree, the recommended content can be made more in line with the user's interests while ensuring the logical coherence of the article, thereby improving the user's reading coherence experience. By dynamically adjusting the weighting coefficient, the recommendation strategy can be adjusted in different reading situations to ensure the adaptability and intelligence of the recommendation results, thereby improving the user's reading stickiness and satisfaction.

[0073] 7. This reading consistency-based content reorganization method quantifies user interest in recommended content by obtaining click-through rates and dwell time, providing accurate feedback to the recommendation system. By calculating user feedback scores, the quality of recommended content can be more intuitively measured, ensuring that the optimization direction of recommended content meets user needs. By performing derivative analysis on the model parameters of click-through rate and dwell time, the parameters of the recommendation algorithm can be more finely tuned, making the optimization of recommended content more scientific and reasonable. By optimizing and updating the parameters using the learning rate, the performance of the recommendation system can be continuously improved, making the recommendation results more accurate and personalized, and enhancing the user experience and satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 Schematic diagram of the process of the present invention.

[0075] Figure 2 This is a flow chart of the dynamic reorganization strategy implemented in the present invention. DETAILED DESCRIPTION

[0076] 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 creative efforts are within the scope of protection of the present invention.

[0077] Example 1, refer to Figure 1 , a reading content reorganization method based on reading consistency, comprising:

[0078] Set up m servers;

[0079] Acquire characteristic data of each server, including computing power, execution efficiency, load, and task transmission delay; execute a server task allocation strategy to allocate a task amount to each server based on the characteristic data;

[0080] For any reading user: obtain the user's historical reading behavior sequence before time t;

[0081] The historical reading behavior sequence includes each article that the user has read in the past

[0082] Execute the user reading behavior modeling strategy, extract the embedding vector of the historical reading behavior sequence, and extract the user reading preference based on the embedding vector;

[0083] Determine whether the article the user is currently reading belongs to the user's reading preferences;

[0084] If the article the user is currently reading is in the user's reading preference, the article paragraph semantic analysis strategy is executed to extract the semantics of each paragraph in the article and calculate the coherence score of the article;

[0085] Determine whether the paragraphs of the article can be reconstructed based on the article's coherence score;

[0086] If the article paragraph can be reconstructed, the user preference calculation strategy is executed to calculate the interest graph based on the user's social behavior and emotional state;

[0087] Based on the interest graph, a dynamic reorganization strategy is executed to dynamically calculate the optimal next recommended paragraph;

[0088] When the user finishes reading the current paragraph, recommend the next paragraph to the user;

[0089] Obtain user click-through rate and dwell time, implement feedback optimization strategies, and optimize the selection of recommended paragraphs.

[0090] By setting up m servers and acquiring characteristic data for each server, including computing power, execution efficiency, load, and task transmission latency, tasks can be rationally allocated, thereby improving computing resource utilization, reducing the risk of server overload, and optimizing overall task scheduling. By capturing the user's historical reading behavior sequence before time t and integrating it with their reading preferences, the system accurately captures their interests, ensuring more personalized recommendations and enhancing the user experience. By performing a semantic analysis strategy on article paragraphs and extracting article coherence scores, the structure of the article can be optimized to ensure a smooth and logical reading experience. A dynamic reorganization strategy based on the interest graph accurately calculates the optimal next recommended paragraph, ensuring that the recommended content better aligns with the user's reading preferences and enhances user immersion. Furthermore, by analyzing user click-through rates and dwell time and employing a feedback optimization strategy, the recommendation system can be continuously adjusted to adapt to the user's reading habits, improving recommendation accuracy and user satisfaction.

[0091] The server task allocation strategy is executed to allocate task amounts to each server based on feature data, including: Among them, a i is the amount of tasks assigned to server i, are the computing capabilities of server i and server j respectively, are the execution efficiencies of server i and server j respectively, are the current loads of server i and server j respectively, is the task transmission delay undertaken by server i, A is the total task amount, Combined computing power for all servers.

[0092] By rationally allocating tasks to each server and incorporating key metrics such as computing power, execution efficiency, load, and task transmission latency, we can fully utilize computing resources and improve task processing efficiency. The concept of comprehensive server computing power makes task scheduling more scientific, avoiding resource waste and server overload. Establishing a mathematical relationship between task volume and server characteristics provides a quantitative basis for task allocation, thereby improving fairness and efficiency. Optimizing task scheduling strategies not only improves computing efficiency but also reduces task transmission latency, increasing the overall responsiveness of the server cluster and ultimately enhancing the user experience.

[0093] The method of executing a user reading behavior modeling strategy, extracting an embedding vector of a historical reading behavior sequence, and extracting a user reading preference based on the embedding vector includes:

[0094] The historical reading behavior sequence is represented as BOOK = [book1, book2,, book c ], where c is the total number of articles, book i represents the i-th article;

[0095] Use the embedding method to extract the embedding vector v of the i-th article i =Embed(book i ), where Embed(·) is the embedding method function, v i is the embedding vector of the i-th article;

[0096] Then the embedding vector of the historical reading behavior sequence is composed of V = [v1, v2,, v c ], V is the embedding vector matrix extracted from the historical reading behavior sequence;

[0097] Add positional encoding to each embedding vector: Where PosEnc(i) is the position encoding;

[0098] Update the embedding vector matrix Calculate user preferences using the self-attention mechanism:

[0099] Define the self-attention calculation:

[0100] Where W1, W2, and W3 are trainable parameter matrices, and Q1, Q2, and Q3 are query, key, and value matrices, respectively;

[0101] Calculate the self-attention weight: Among them, τ is the dimension of the embedding vector; after multi-layer Transformer calculation, the user preference vector is extracted:

[0102] Γ=-∑ (i) logP(book i+1 |V), where Γ is the loss function, For users who use model predictions, see the next article book i+1 The probability of p T is the transpose of the user preference vector p, v i+1 For book i+1 The embedding vector, v j For book j Embedding vector of

[0103] Set the loss threshold ε Γ ;

[0104] Get the difference ΔΓ of the loss function update, if ΔΓ<ε Γ , then stop updating the loss function and get the optimal user preference vector p.

[0105] By converting a user's historical reading behavior sequence into an embedding vector matrix, we can effectively capture their long-term reading habits, enabling the recommendation system to more accurately match user needs. Combined with the positional information encoding of the embedding vector, this preserves the order in which articles were read, enhancing the model's ability to model sequences and making recommendations more consistent with user logic. Using a self-attention mechanism to calculate user preferences captures long-term points of interest. A multi-layer Transformer extracts representative user preference vectors, improving the accuracy of interest modeling. By setting a loss function and employing an optimization strategy, we ensure that the user preference vectors are optimal, improving the recommendation system's matching of reading interests and ultimately optimizing recommendation effectiveness.

[0106] If the article the user is currently reading belongs to the user's reading preference, the article paragraph semantic analysis strategy is executed to extract the semantics of each paragraph in the article and calculate the coherence score of the article, including:

[0107] The article that the user is currently reading is represented as G = [g1, g2, ..., g d], where d is the total number of paragraphs in the article, g i Indicates the i-th paragraph;

[0108] Use natural language processing technology to extract the semantic meaning of each paragraph i =f θ (g i ) where f θ (g i ) is the natural language processing model, θ is the parameter of the natural language processing model, u i For paragraph g i semantics;

[0109] Compute the coherence score between any two paragraphs: Among them, u i For paragraph g i The semantic vector, u j For paragraph g j The semantic vector, u i ·u j is the semantic vector u i and semantic vector u j The dot product of ‖u i ‖ is the semantic vector u i The Euclidean norm of j ‖ is the semantic vector u j The Euclidean norm of ;

[0110] Calculate the global coherence score, Setting the coherence threshold

[0111] like Then the article has low coherence and the paragraphs can be reconstructed;

[0112] like Then the article has high coherence and the paragraphs cannot be reconstructed.

[0113] By dividing the article the user is currently reading into paragraphs and using natural language processing to extract semantic information from each paragraph, we can gain a deeper understanding of the article's structure and content, improving the recommendation system's ability to perceive the article's information. Calculating the coherence score between paragraphs effectively assesses the article's logical structure, ensuring the content's rationality and readability. By calculating the global coherence score, we can determine whether the article's overall logic aligns with the user's reading habits and set a threshold to determine whether paragraphs need to be restructured to optimize readability and coherence. Ultimately, by optimizing the article's content structure, we enhance the user's reading experience, make the reading more logical, and improve information retrieval efficiency.

[0114] If the article paragraph can be reconstructed, the user preference calculation strategy is executed to calculate the interest graph based on the user's social behavior and emotional state, including:

[0115] Get the user's activity actions on social media [e1, e2, e3], where e1 is like, e2 is comment, and e3 is forward;

[0116] Building an interest graph Among them, R is the interest graph, They are user like weight, comment weight and forwarding weight respectively. and The number of likes, comments and reposts of users respectively;

[0117] Get the text T posted by the user;

[0118] Use the sentiment analyzer to calculate the user's sentiment score: in, is the sentiment score, f ∈ (·) is the sentiment analyzer, ∈ is the parameter of the sentiment analyzer;

[0119] when The user has positive emotions;

[0120] when The user has negative emotions;

[0121] when Then the user has neutral emotions;

[0122] Combined with sentiment scores, adjust the interest graph: Among them, δ is the adjustment coefficient, which is used to control the influence of sentiment analysis results on the interest graph. f(freq(T)) is a function of how frequently users mention topics in their texts. When the topic frequency is higher, it is given more weight.

[0123] By analyzing users' likes, comments, and forwarding behaviors on social media, we can more comprehensively capture their social interactions. By combining social behavior patterns to construct an interest graph, we can improve the recommendation system's ability to capture user interests. By assigning weights to different social behaviors and integrating the textual content from the user's social platforms, we can accurately calculate the user's emotional state, ensuring that recommended content is both relevant to their interests and their current emotional state. By calculating the user's sentiment score through a sentiment analyzer, we can identify their emotional state and optimize the interest graph accordingly, making recommended content more contextually adaptable. Ultimately, by optimizing the interest graph based on sentiment analysis results, we ensure that recommended content aligns with users' long-term interests while also adapting to short-term mood swings, improving user acceptance and satisfaction with recommended content.

[0124] The method of executing a dynamic reorganization strategy based on the interest graph and dynamically calculating the optimal next recommended paragraph includes:

[0125] The paragraph the user is currently reading is recorded as g * ;

[0126] For any unread paragraph g ′ , calculate paragraph g ′ Matching degree with interest graph:

[0127] Among them, u ′ For paragraph g ′ The semantic vector of R u ′ is the dot product of interest graph and semantic vector, ‖R‖, ‖u ′ ‖ are the moduli of interest graph and semantic vector respectively;

[0128] Calculate the optimal next recommended paragraph g - , Among them, u * For paragraph g * The semantic vector of is a weighting coefficient used to balance the impact of paragraph coherence and interest matching.

[0129] In this embodiment, refer to Figure 2 .

[0130] By calculating the match between the user's current reading paragraph and their interest graph, we can ensure that the next recommended paragraph aligns with the user's interests, improving the relevance of recommended content. Using a dot product to calculate the match maintains computational efficiency while ensuring accurate recommendation results. A weighted calculation method combining coherence scores and interest matching ensures the logical coherence of the article while aligning recommended content with user interests, enhancing the reading experience. By dynamically adjusting the weighting coefficients, we can optimize recommendation strategies for different reading scenarios, making recommendation results more adaptable and intelligent, and improving user engagement and satisfaction.

[0131] The method of obtaining user click-through rate and dwell time, executing feedback optimization strategy, and optimizing the selection of recommended paragraphs includes:

[0132] The user click-through rate is recorded as CTR, and the user's stay time in the recommended paragraph is recorded as Y;

[0133] Calculate user feedback scores in, Derivative the model parameters click rate and dwell time:

[0134] Among them, π is the model parameters click rate and dwell time, represents the impact of model parameter π on click-through rate, is the impact of model parameters on the length of stay, where p is the quality feature of the recommended paragraph;

[0135] Update parameters in, is the learning rate;

[0136] The selection of recommended paragraphs is optimized using the updated parameters.

[0137] By analyzing user click-through rates and dwell time, we can quantify user interest in recommended content and provide accurate feedback to the recommendation system. Calculating user feedback scores provides a direct measure of the quality of recommended content, ensuring that optimization targets meet user needs. Derivative analysis of model parameters such as click-through rate and dwell time allows for precise adjustment of recommendation algorithm parameters, making the optimization process more scientific and rational. Using learning rate optimization for parameter updates can continuously improve the performance of the recommendation system, making recommendations more accurate and personalized, and enhancing user experience and satisfaction.

[0138] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0139] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A reading content reorganization method based on reading consistency, characterized in that: include: Set up m servers; Obtaining characteristic data of each server, the characteristic data including computing power, execution efficiency, load, and task transmission delay; Execute the server task allocation strategy and assign the task amount to each server based on the feature data; For any reading user: obtain the user's historical reading behavior sequence before time t; The historical reading behavior sequence includes each article that the user has read in the past; Execute the user reading behavior modeling strategy, extract the embedding vector of the historical reading behavior sequence, and extract the user reading preference based on the embedding vector; Determine whether the article the user is currently reading belongs to the user's reading preferences; If the article the user is currently reading is in the user's reading preference, the article paragraph semantic analysis strategy is executed to extract the semantics of each paragraph in the article and calculate the coherence score of the article; Determine whether the paragraphs of the article can be reconstructed based on the article's coherence score; If the article paragraph can be reconstructed, the user preference calculation strategy is executed to calculate the interest graph based on the user's social behavior and emotional state; Based on the interest graph, a dynamic reorganization strategy is executed to dynamically calculate the optimal next recommended paragraph; When the user finishes reading the current paragraph, recommend the next paragraph to the user; Obtain user click-through rate and dwell time, implement feedback optimization strategies, and optimize the selection of recommended paragraphs; The method of executing a user reading behavior modeling strategy, extracting an embedding vector of a historical reading behavior sequence, and extracting a user reading preference based on the embedding vector includes: Represent the historical reading behavior sequence as , where c is the total number of articles, represents the i-th article; Use the embedding method to extract the embedding vector of the i-th article ,in, is the embedding method function, is the embedding vector of the i-th article; The embedding vector of the historical reading behavior sequence is composed of , Embedding vector matrix extracted for historical reading behavior sequence; Add positional encoding to each embedding vector: ,in Encode for position; Update the embedding vector matrix ; Calculate user preferences using the self-attention mechanism: Define the self-attention calculation: ,in, is the trainable parameter matrix, where are query, key, and value matrices respectively; Calculate the self-attention weight: ,in, is the dimension of the embedding vector; After multi-layer Transformer calculation, the user preference vector is extracted: Layer( ; ,in, is the loss function, See the next article for users who use model predictions The probability of , where is the user preference vector The transpose of for The embedding vector of for Embedding vector of Setting loss thresholds ; Get the difference of the loss function update ,like , then stop updating the loss function and get the optimal user preference vector .

2. The reading content reorganization method based on reading consistency according to claim 1 is characterized in that: The server task allocation strategy is executed to allocate task amounts to each server based on feature data, including: ,in, is the amount of tasks assigned to server i, are the computing capabilities of server i and server j respectively, are the execution efficiencies of server i and server j respectively, are the current loads of server i and server j respectively, The transmission delay of the task undertaken by server i, is the total task volume, Combined computing power for all servers.

3. The reading content reorganization method based on reading consistency according to claim 1 is characterized in that: If the article the user is currently reading belongs to the user's reading preference, the article paragraph semantic analysis strategy is executed to extract the semantics of each paragraph in the article and calculate the coherence score of the article, including: Represent the article the user is currently reading as , where d is the total number of paragraphs in the article, Indicates the i-th paragraph; Use natural language processing technology to extract the semantic meaning of each paragraph in, is a natural language processing model, are the parameters of the natural language processing model, For paragraphs semantics; Compute the coherence score between any two paragraphs: ,in, For paragraphs The semantic vector of For paragraphs The semantic vector of is the semantic vector and semantic vectors The dot product of is the semantic vector The Euclidean norm of , is the semantic vector The Euclidean norm of ; Calculate the global coherence score, ; Setting the coherence threshold ; like , then the article has low coherence and the paragraphs can be reconstructed; like , then the article has high coherence and the paragraphs cannot be reconstructed.

4. The reading content reorganization method based on reading consistency according to claim 1 is characterized in that: If the article paragraph can be reconstructed, the user preference calculation strategy is executed to calculate the interest graph based on the user's social behavior and emotional state, including: Get user activity on social media ,in, For praise, For comments, For forwarding; Building an interest graph ,in, For interest graph, , , They are user like weight, comment weight and forwarding weight respectively. , and The number of likes, comments and reposts of users respectively; Get the text posted by the user ; Use the sentiment analyzer to calculate the user's sentiment score: ,in, Score for emotion, For sentiment analyzer, Parameters for the sentiment analyzer; when , then the user has positive emotions; when , the user has negative emotions; when , then the user has neutral emotions; Combined with sentiment scores, adjust the interest graph: ,in, is the adjustment coefficient used to control the influence of sentiment analysis results on the interest graph. It is a function of how frequently users mention topics in their texts. Topics with higher frequency are given more weight.

5. The reading content reorganization method based on reading consistency according to claim 3 is characterized in that: The method of executing a dynamic reorganization strategy based on the interest graph and dynamically calculating the optimal next recommended paragraph includes: Record the paragraph the user is currently reading as ; For any unread paragraph , calculate paragraph Matching degree with interest graph: ,in, For paragraphs The semantic vector of is the dot product of the interest graph and the semantic vector, are the modules of interest graph and semantic vector respectively; Calculate the optimal next recommended paragraph , ,in, For paragraphs The semantic vector of , is a weighting coefficient used to balance the impact of paragraph coherence and interest matching.

6. The reading content reorganization method based on reading consistency according to claim 5 is characterized in that: The method of obtaining user click-through rate and dwell time, executing feedback optimization strategy, and optimizing the selection of recommended paragraphs includes: The user click rate is recorded as , the user's stay time in the recommended paragraph is recorded as ; Calculate user feedback scores ,in, ; Derivative the model parameters click rate and dwell time: ,in, are model parameters, Represents model parameters Impact on click-through rate, is the effect of model parameters on the length of stay, where The quality characteristics of the recommended paragraphs; Update parameters ,in, is the learning rate; The selection of recommended paragraphs is optimized using the updated parameters.

Images