Reading management method and reading management device for electronic document

Abstract

The invention relates to the technical field of electronic document management, and discloses an electronic document reading management method which comprises the following steps: S1, a document information entropy calculation step: analyzing text contents of an electronic document based on an information entropy calculation method, and determining the complexity of different contents; s2, a user behavior data collection step: collecting behavior data of rolling, staying, returning, highlighting and noting when the user reads the electronic document, and also provides a reading management device of the electronic document, which comprises a document information entropy calculation module used for calculating the complexity of the electronic document text and setting a reading weight; and the user behavior monitoring module is used for collecting user reading behavior data. According to the method, recommendation sorting is optimized through reinforcement learning, interests and preferences are dynamically adjusted through temperature parameter regulation and control exploration and utilization balance and a time decay mechanism, the user control feeling is enhanced through interactive recommendation, and more intelligent, efficient and personalized reading management is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic document management, and specifically to a reading management method and a reading management device for electronic documents. Background Art

[0002] With the popularization of electronic documents, the way people obtain and read information has changed greatly. Whether for learning, working or daily browsing, users are exposed to a large number of electronic documents every day. However, in the face of a vast amount of information flow, how to quickly find content that meets personal needs has become a major pain point for users. An efficient reading management system should not only provide accurate content recommendations but also be able to dynamically adapt to changes in user interests to improve reading efficiency and experience.

[0003] Existing electronic document reading management systems can already provide a certain degree of content recommendation and personalized display. Some systems adopt recommendation algorithms based on historical click data, which can match user interests to a certain extent and improve the relevance of recommended content. At the same time, some reading platforms support simple interactive adjustments. For example, users can manually mark content they are not interested in or adjust the recommendation scope through filtering functions. These methods have improved the reading experience to a certain extent, enabling users to manage and obtain information more conveniently.

[0004] However, there are still some deficiencies in existing recommendation methods. Firstly, the recommendation mechanisms of many systems are relatively fixed and do not fully consider the dynamic changes in user interests, resulting in the recommended content being prone to solidification and lacking freshness. In addition, some systems tend to strengthen existing interests more during the recommendation process and lack reasonable exploration strategies, making it difficult for users to access new high-value information. On the other hand, the sorting method of recommended content is often based on static weights and fails to be adjusted in combination with the real-time feedback of users, resulting in insufficient recommendation accuracy. Moreover, existing recommendation systems mostly adopt fixed display modes and lack flexible interaction methods, making it impossible for users to actively adjust the recommended content and affecting the overall reading experience. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technology, the present invention provides a reading management method and a reading management device for electronic documents, which solve the problems of solidification of recommended content, lack of dynamic adaptability, difficulty in balancing exploration and accuracy, and single interaction method in the existing technology.

[0006] To achieve the above objectives, the present invention is realized through the following technical solutions: A reading management method for electronic documents, including the following steps: S1. Document information entropy calculation step: Analyze the text content of the electronic document based on the information entropy calculation method to determine the complexity of different contents; S2. User behavior data collection step: Collect the behavior data of the user's scrolling, pausing, returning, highlighting, and taking notes while reading the electronic document. S3. Information gain calculation step: Calculate the information gain of the user for different contents of the electronic document based on the information entropy calculation result and the user behavior data, which is used to evaluate the user's reading depth. S4. Bayesian inference progress calculation step: Update the user's reading state using the Bayesian inference method based on the user behavior data and the information gain calculation result, and dynamically calculate the user's reading progress. S5. Personalized optimization step: Adjust the personalized reading recommendation strategy based on the reading progress calculation result, and optimize the subsequent content recommendation order to match the user's reading habits and comprehension level.

[0007] Preferably, the document information entropy calculation step includes: Parse the text content of the electronic document, and calculate the information entropy value of each text paragraph based on the statistical model of the text content. Determine the complexity of each piece of text according to the information entropy value, and assign different reading weights to contents with different complexities; use an entropy-based algorithm to mark the important parts in the document for optimization in subsequent user behavior analysis.

[0008] Preferably, the information gain calculation step includes: Based on the complexity analysis result in the document information entropy calculation step, combined with the dwell time, scrolling behavior, and highlighting annotations in the user behavior data collection step, calculate the information gain of the user for each content paragraph in the document. Evaluate the user's understanding depth of different contents according to the information gain value, and dynamically adjust the user's reading progress according to this result; for contents with high information gain, recommend more detailed subsequent contents, while for contents with low information gain, relatively brief or overview contents can be recommended.

[0009] Preferably, the information gain calculation step includes: Calculate the initial information entropy of the document. Calculate the change value of the text information entropy after the user reads. Use the information gain calculation method to measure the user's reading depth. Adjust the contribution weight of this page to the reading progress according to the information gain result.

[0010] Preferably, the Bayesian inference progress calculation step includes: Set the user's reading prior probability. Calculate the relationship between the user's reading behavior pattern and the reading completion degree on a certain page. Use the Bayesian inference method to calculate the probability that the user actually reads this page; Combine the reading probabilities of different pages to dynamically adjust the reading progress, and the calculation result of the reading progress is used to dynamically adjust the subsequent content recommendation.

[0011] Preferably, the personalized optimization steps include: Adjust the reading time allocation strategy according to the user's reading behavior pattern; Predict the best learning path for the unread part of the user and adjust the recommendation order. The recommended reading path is dynamically adjusted based on the user's reading behavior; Generate a personalized reading report for the user, providing review and reinforcement learning suggestions.

[0012] The present invention also provides a reading management device for electronic documents, including: A document information entropy calculation module, which is used to calculate the complexity of the electronic document text and set the reading weight; A user behavior monitoring module, which is used to collect user reading behavior data, including the behavior data of scrolling, staying, returning, highlighting, and taking notes; An information gain calculation module, which is used to calculate the depth of understanding of the document content by the user's interaction behavior and adjust the reading progress weight; a Bayesian inference progress calculation module, which is used to dynamically update the reading progress according to the user behavior data; A personalized optimization module, which is used to adjust the reading recommendation strategy based on the user's historical reading behavior.

[0013] Preferably, the document information entropy calculation module includes: A text parsing unit, which is used to analyze the text structure of the electronic document; A word frequency statistics unit, which is used to count the occurrence probability of words in the document; An information entropy calculation unit, which is used to calculate the complexity of the text and set the reading time weight.

[0014] Preferably, the user behavior monitoring module includes: A time monitoring unit, which is used to record the staying time of the user on the page; A scrolling monitoring unit, which is used to monitor the scrolling behavior and page turning speed of the user; An interaction monitoring unit, which is used to record the behavior data of the user's highlighting, marking, and taking notes.

[0015] Preferably, the Bayesian inference progress calculation module includes: A behavior probability calculation unit, which is used to calculate the relationship between the user's reading behavior and the reading completion degree; A Bayesian inference unit, which is used to calculate the true reading probability of the user by using the Bayesian method; A progress dynamic adjustment unit for dynamically calculating the reading progress based on the reading probability and optimizing the learning path.

[0016] The present invention provides a reading management method and a reading management device for electronic documents. It has the following beneficial effects: 1. By using reinforcement learning to optimize the recommendation ranking, the present invention introduces softmax to regulate the probability distribution of the recommended content display. Compared with the traditional fixed ranking method, this strategy can adapt to the changes in user interests in real time, making the recommendation results more in line with personalized needs and avoiding the problem of single content.

[0017] 2. By introducing a temperature parameter, the present invention can intelligently adjust the balance between exploration and exploitation. The system will neither be limited to existing preferences nor blindly recommend low-correlation content. Compared with the traditional single recommendation method based on historical behavior, this method can expand the user's interest field while still maintaining the accuracy of the recommended content.

[0018] 3. By combining a time decay factor, the present invention makes the latest user behavior have a greater impact on the recommendation ranking. In this way, the system will not rely on outdated data for a long time, but can dynamically respond to the recent preference changes of users. Compared with the problem that the existing recommendation system is prone to "solidify" user interests, this solution ensures the timeliness and flexibility of the recommended content.

[0019] 4. By adopting an interactive recommendation strategy, such as sliding, folding, user-defined adjustment, etc., the present invention makes the presentation of the recommended content more flexible. Compared with the traditional fixed card display, the present invention can adapt to the operation habits of different users, improve the readability of the recommended information, reduce information overload, and improve the user's reading experience. Description of the Drawings

[0020] Figure 1 is a flowchart of the method of the present invention; Figure 2 is a schematic structural diagram of the present invention. Detailed Embodiments

[0021] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the specification of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0022] Please refer to the attached Figure 1 , the embodiments of the present invention provide a reading management method for electronic documents, including the following steps: S1. Steps for calculating document information entropy: Analyze the text content of an electronic document based on the information entropy calculation method to determine the complexity of different contents. The main purpose of the steps for calculating document information entropy is to quantify the complexity of each part in an electronic document, providing basic data support for subsequent reading behavior analysis, information gain calculation, and personalized optimization. By calculating the information entropy of a document, the difficulty of different contents can be measured, and the reading weights can be adjusted accordingly, enabling the reading management system to reasonably allocate the user's reading time and optimize subsequent recommendation strategies at the same time.

[0023] Generally, the text content of an electronic document contains multiple levels of information, and the complexity of different levels of information is different, resulting in differences in the cognitive load of users during reading. Therefore, as an option, based on the information entropy theory, the complexity of the text can be quantified and combined with the user's behavior data to more accurately evaluate the user's reading depth and understanding.

[0024] Specifically, this step completes the complexity analysis of the electronic document through four sub-processes: text parsing, word frequency statistics, information entropy calculation, and reading weight allocation. In some embodiments, natural language processing (NLP) technologies, such as topic modeling (LDA), word vector representation (Word2Vec), or semantic similarity analysis (BERT), can be combined to improve the accuracy of text complexity evaluation.

[0025] In this embodiment, first, the electronic document is structurally parsed to extract the text content and perform basic cleaning and preprocessing. Generally, text preprocessing includes: Removing stop words (such as words with no actual semantic contribution like "de", "shi", "zai", etc.); Word segmentation, splitting the text into independent words or phrases; Part-of-speech tagging, identifying different categories of words such as verbs, nouns, adjectives, etc.; Syntactic analysis, identifying the subject-predicate-object structure of sentences to better understand the hierarchy of the text content; In a possible implementation, the text parsing unit can use the TF-IDF (term frequency-inverse document frequency) method to measure the importance of words, or use deep learning semantic vector representations such as Word2Vec or BERT to extract richer semantic information.

[0026] In some embodiments, the complexity of the document content can be preliminarily analyzed through word frequency statistics. Generally, complex texts often contain more low-frequency words and long sentences, while simpler texts are usually composed of high-frequency words and short sentences. Therefore, the n-gram statistical method can be used to calculate the word frequency distribution in the text and perform modeling analysis on it.

[0027] As an option, the word frequency histogram of each paragraph can be calculated, and further the entropy value of the text can be calculated. In some embodiments, the system can also analyze the sentence length distribution of the text and the complexity of the subject-predicate-object structure to further evaluate the reading difficulty of the text.

[0028] In this embodiment, the calculation of information entropy is based on the following mathematical formula: Where: H(X) represents the information entropy value of the document or paragraph, reflecting the text complexity; P(x i ) represents the occurrence probability of the i-th word in the text; N is the total number of different words in the text; x i is a single word in the document.

[0029] The core idea of information entropy calculation is: when the word frequency distribution of a certain text is relatively uniform, the information entropy is higher, which means that the information contained in the text is more dispersed and the complexity is higher; on the contrary, if some words occupy a large proportion (such as a large number of repeated high-frequency words), the information entropy is lower, indicating that the complexity of the text is lower.

[0030] In a possible implementation, the information entropy can be calculated according to the following steps: Count the number of occurrences of each word in the document, calculate its relative frequency P(x i ), and calculate the total information entropy of the entire document using the above entropy formula.

[0031] Divide the document into multiple paragraphs, calculate the information entropy of each paragraph respectively, and obtain the complexity distribution of different contents.

[0032] In this embodiment, based on the calculation result of information entropy, the system will assign different reading weights to each text paragraph. Generally, paragraphs with higher information entropy contain more new information, and users may need more time to read these contents. Therefore, the system can use the following formula to calculate the reading time weight: Where: W(T) is the reading time weight of the current paragraph; H(X) is the information entropy of the current paragraph; M is the total number of paragraphs in the document; H(X j ) is the information entropy of the j-th paragraph; represents the sum of the information entropy values of all paragraphs in the entire document, which is the normalization denominator part and is used to ensure that the sum of the weights of each paragraph is equal to 1.

[0033] The function of this formula is to ensure that the content with higher information entropy occupies a larger proportion of the reading time, while the content with lower information entropy occupies a relatively smaller proportion. As an option, the weight can be dynamically adjusted during the reading process and further optimized in combination with the user's reading behavior.

[0034] In some embodiments, syntactic tree parsing can be further combined to analyze the structural complexity of the text. For example, the depth of the dependency tree of a sentence can be calculated to measure the syntactic complexity of the sentence. In addition, TF-IDF and cosine similarity can be combined to analyze the information redundancy degree between different paragraphs, so as to optimize the reading time allocation strategy.

[0035] As another possible implementation, a machine learning model (such as a random forest or a deep learning model) can be introduced. Through the feature training of a large amount of text, the system can more accurately predict the complexity of the text and optimize the accuracy of information entropy calculation.

[0036] The calculation steps of document information entropy quantify and evaluate the complexity of an electronic document through four key sub-processes: text parsing, word frequency statistics, information entropy calculation, and reading weight allocation. The calculation method of information entropy is based on the probability distribution formula and combined with the calculation method of reading time weight to ensure that the system can reasonably allocate the user's reading time and improve the intelligence of reading management. Further optimization solutions can combine natural language processing technologies such as syntactic structure analysis, topic modeling, and TF-IDF to improve the calculation accuracy and applicability.

[0037] S2. User behavior data collection step, collecting the behavior data of the user's scrolling, staying, returning, highlighting, and note-taking when reading an electronic document; The user behavior monitoring step is a key link for the system to obtain user reading interaction data. The main goal of this step is to collect the user's reading behavior information in real time and use this data to build a reading pattern to optimize the subsequent information gain calculation and personalized recommendation strategy.

[0038] Generally, the behavior characteristics of users when reading an electronic document are diverse, mainly including: page stay time, scrolling operation, interaction behaviors (clicking, marking, note-taking, etc.), look-back and skip-reading situations, etc. These behaviors can not only reflect the user's interest preferences for the text content but also provide indirect information about their depth of understanding of the content. Therefore, in this step, the system uses multi-dimensional monitoring means (such as mouse trajectory, touch interaction, keyboard input, etc.) to establish a user's reading behavior dataset to achieve accurate user behavior modeling.

[0039] Specifically, the core of this step includes four sub - processes: time monitoring, scrolling monitoring, interaction monitoring, and reading interest inference. Each sub - process is responsible for capturing user reading behavior data in different dimensions. In some embodiments, combined with machine learning methods (such as Long Short - Term Memory Network LSTM, Hidden Markov Model HMM), the system can predict the user's reading pattern and adjust the personalized recommendation strategy in real - time to optimize the user experience.

[0040] In this embodiment, the time monitoring unit is used to record the residence time of the user on each page or paragraph, and analyze the user's reading pattern based on this data. Generally, a longer residence time usually indicates that the user is interested in this part of the content or the content is more difficult to understand, while a brief browse may mean that the user is not concerned about this content or is already familiar with it.

[0041] As an option, the following formula can be used to calculate the reading time weight of a unit paragraph: Where: T w is the reading time weight of the current paragraph; T p represents the residence time of the user in the current paragraph; T t is the total reading time of the user in the entire document.

[0042] The function of this formula is to measure the time allocation of the user to different paragraphs, so as to infer the user's focus of attention. In some embodiments, the sliding window averaging method can be combined to smooth the time data to reduce the influence of outliers on the results.

[0043] In addition, time monitoring can also be combined with the analysis of reading rhythm changes. By calculating the time change rate between paragraphs, the user's interest fluctuations can be further identified. For example, if the residence time of a certain paragraph is much higher than that of the previous and subsequent paragraphs, this paragraph may contain key content, and the system can adjust the subsequent recommendation strategy accordingly.

[0044] In this embodiment, the scrolling monitoring unit is used to capture the scrolling behavior characteristics of the user, including scrolling speed, direction, pause position, etc. The scrolling behavior can reflect the user's browsing pattern of the document content: rapid scrolling usually indicates a skimming or browsing state, while multiple pauses or rollbacks may mean that a certain part of the content is more complex or the user has a higher degree of attention to it.

[0045] Generally, the scrolling speed can be calculated by the following formula: Where: S r is the scrolling speed (unit: pixels per second or lines per second); D s represents the scrolling distance (pixels or lines); T s is the time (seconds) used for scrolling.

[0046] In some embodiments, the system can calculate the scrolling acceleration, i.e., the rate of change of the scrolling speed over time, to analyze the user's reading rhythm. For example, areas with a large change in scrolling speed may be parts where the user quickly searches for information, while a low scrolling acceleration usually means the user is reading carefully.

[0047] In addition, to detect whether the user looks back at a certain piece of content, the scrolling direction can be recorded and the rollback frequency can be calculated. The more times of the rollback operation, the higher the complexity of the content segment or the more likely the user is interested in this part of the content.

[0048] In this embodiment, the interaction monitoring unit is used to record the user's active interaction behaviors during reading, including: marking, word highlighting, note-taking, clicking, etc. Generally, when the user performs an interaction operation on a certain piece of content, it usually means that the content has a high value to him / her.

[0049] In a possible implementation, an interaction behavior weighting model can be defined to calculate the weighted values of different types of interaction behaviors: I w = w1 × N h + w2 × N n + w3 × N c ; Where: I w is the weighted value of the interaction behavior; N h is the number of times the user highlights; N n is the number of times the user adds notes; N c is the number of times the user clicks on a certain part of the content; w1, w2, w3 are the weight coefficients of different interaction behaviors and can be adjusted according to experimental data.

[0050] Generally, the weights of highlighting and note-taking are higher, while the weight of ordinary clicking is lower, because the former usually means that the user is deeply thinking about a certain piece of content. In some embodiments, the TF-IDF method can be combined to analyze the keywords of the user's interaction and further infer their reading interest points.

[0051] In this embodiment, the system infers the user's reading interest and cognitive state based on the data of time monitoring, scrolling monitoring, and interaction monitoring. As an option, a Bayesian classifier or a Hidden Markov Model (HMM) can be used to predict the user's future reading needs according to the user's historical behaviors and dynamically adjust the content recommendation order.

[0052] Specifically, in a possible implementation, the system can calculate the reading interest distribution of the user on different categories of content: Where: P(Ci P(X|C) represents the probability of the user's interest in category C i ; P(X|C i ) represents the probability that the user performs certain actions under category C i ; P(C i ) is the prior probability of the category; P(X) is the probability distribution of the user's overall behavior.

[0053] Through this formula, the user's preference for different topic contents can be estimated, so as to optimize the recommendation algorithm and improve the accuracy of the personalized reading experience.

[0054] In some embodiments, eye-tracking technology can be combined to analyze the user's gaze movement trajectory to obtain more accurate reading attention data. In addition, sentiment analysis methods can be used to analyze the user's emotional fluctuations during reading, such as by facial expression recognition or voice emotion analysis, to judge the user's reaction to specific content, so as to adjust the recommendation strategy.

[0055] As another possible implementation, a reinforcement learning algorithm can be adopted to adjust the recommendation model in real time, enabling the system to be optimized according to the user's immediate feedback and gradually improving the recommendation effect.

[0056] The user behavior monitoring step comprehensively collects and analyzes the user's reading behavior through four key subprocesses: time monitoring, scrolling monitoring, interaction monitoring, and reading interest inference. Based on these data, the system can accurately infer the user's reading pattern, and combined with machine learning algorithms, optimize the personalized recommendation strategy to make the reading management system more intelligent and efficient.

[0057] S3. Information gain calculation step: Based on the information entropy calculation result and the user behavior data, calculate the information gain of the user for different contents of the electronic document, which is used to evaluate the user's reading depth; The user behavior analysis and recommendation strategy adjustment step (S3) is based on the user behavior monitoring in step S2, and further deeply mines the user behavior data to identify the user's reading interests, content preferences, and interaction patterns, and dynamically optimize the recommendation strategy accordingly. The core goal of this step is to improve the matching degree of the recommended content, enabling the user to obtain the document content that meets their own needs faster, thereby enhancing the reading experience.

[0058] In this step, the system conducts multi-dimensional analysis on various types of user behavior data in the document, including reading time distribution, scrolling behavior pattern, interaction frequency, and rollback operation, etc. In addition, the system combines machine learning and statistical inference technologies, such as Bayesian classification, Hidden Markov Model (HMM), and reinforcement learning, to dynamically adjust the recommendation strategy to ensure that the recommended content is highly consistent with the user's interests.

[0059] In this embodiment, the interest pattern analysis unit processes the user behavior data collected in step S2, and constructs a reading interest model of the user through feature extraction, classification, and pattern recognition. The system first extracts the behavior characteristics of the user in different content categories and represents them as behavior feature vectors: X=(x1,x2,...,x n ); where: x1,x2,...,x n represent behavior data in different dimensions, such as the residence time, click times, scrolling rate, etc. of a certain piece of content.

[0060] Generally, the user's interest in a certain type of content can be calculated by a Bayesian classifier to obtain its interest probability, using the "reading interest distribution" disclosed in S2.

[0061] In a possible implementation manner, the system continuously monitors the user's interest changes and continuously updates the user interest model based on new behavior data. For example, if the user frequently rolls back a certain type of content within a period of time, the system can adjust the interest weight of this category so that it occupies a higher priority in subsequent recommendations.

[0062] In addition, in some embodiments, the user's behavior patterns can be classified by a clustering algorithm (such as K-means) to identify different types of reading habits, such as fast browsing type, in-depth reading type, and jumping reading type, so as to optimize the personalized recommendation strategy.

[0063] Based on the interest pattern analysis, in this embodiment, the personalized recommendation strategy adjustment unit is responsible for adjusting the recommended content according to the dynamic changes of the user's interest. Generally, the recommendation strategy adjustment mainly includes operations such as content screening, order optimization, and dynamic update.

[0064] Specifically, the system first uses the interest model to screen the content in the document library to remove the content that the user may not be interested in; then, based on the user's interest weights for different content categories, the recommended content is sorted by priority to ensure that the most relevant content can be presented first; finally, the system dynamically updates the recommendation list in combination with real-time behavior data to match the user's immediate needs.

[0065] In some embodiments, the system can combine the hidden Markov model (HMM) to predict the user's next reading behavior, so as to make an early recommendation. The calculation method is as follows: where: P(X t+1 |X t ) represents the possible next reading behavior of the user after the current behavior X t ; P(X t+1 |Si ) represents the probability that the user performs action X i in state S; P(S t+1 |X i ) is the probability that the user is in state S t under the current action X; N1 is the number of all possible reading states. t i

[0066] This model can be used to predict the user's stay time, interaction methods, etc. on different contents, so as to adjust the recommended content in advance and improve the accuracy of recommendations.

[0067] In this embodiment, the recommendation algorithm optimization unit is responsible for continuously optimizing the recommendation strategy to ensure that the recommendation system can continuously adjust the content recommendation weight according to the user's feedback. In a possible implementation, the system uses a reinforcement learning algorithm to continuously update the recommendation strategy through the user's real-time feedback.

[0068] Reinforcement learning can be implemented through the Q-learning method: where: Q(s t , a t ) represents the Q value of performing action a t in state s, that is, the value of the recommended content; α is the learning rate, which determines the influence degree of new data on the model; r t is the user's immediate feedback on the recommended content (such as click, favorite, etc.); γ is the discount factor, which is used to measure the influence of future benefits; t is the maximum Q value of taking the optimal action a t+1 in the next state s. t+1

[0069] In some embodiments, reinforcement learning can be combined with the Multi-Armed Bandit algorithm to find the best balance between exploring new content and exploiting existing interest data. For example, the system can recommend content that the user has not read in some cases to test their interests, so as to optimize the long-term recommendation effect.

[0070] In another possible implementation, natural language processing (NLP) technology can be combined to analyze the notes and highlighted content marked by the user during reading to extract the user's focus points. For example, through the TF-IDF algorithm or topic modeling (LDA), the keywords that the user is most interested in in the document can be identified, and the recommended content can be adjusted accordingly.

[0071] In addition, in some embodiments, the system can combine sentiment analysis, utilize facial expression recognition or speech emotion analysis to identify the user's emotional reactions to different contents. For example, if the user shows a high degree of concentration or positive emotion towards certain contents, the system can further optimize the recommendation weights of the relevant contents.

[0072] The user behavior analysis and recommendation strategy adjustment step deeply analyzes the behavior data collected in S2, establishes a user interest model, and optimizes the personalized recommendation strategy based on this model. The system combines technologies such as Bayesian classification, Hidden Markov Model (HMM), and reinforcement learning, and is able to predict the user's reading interests and dynamically adjust the recommended contents. In addition, by combining NLP technology and sentiment analysis, the system can further improve the accuracy of the recommendation strategy to optimize the user's overall reading experience.

[0073] S4. Bayesian inference progress calculation step, based on the user behavior data and the information gain calculation results, uses the Bayesian inference method to update the user's reading status and dynamically calculate the user's reading progress; Step S4 user feedback processing and recommendation strategy adaptive optimization is to further optimize the recommendation strategy based on the user feedback information on the basis of step S3 user behavior analysis and recommendation strategy adjustment.

[0074] Generally, the user's feedback can be divided into two major categories: explicit feedback and implicit feedback. Explicit feedback includes interaction behaviors such as liking, collecting, and commenting that clearly express the user's preferences, while implicit feedback includes behavioral characteristics such as reading duration, scrolling speed, click frequency, and rollback times that do not require the user's active operation.

[0075] As an option, this step processes the user feedback data, calculates the quality score of the recommended content, and combines methods such as reinforcement learning and adaptive optimization to enable the recommendation strategy to be dynamically adjusted to improve the personalized accuracy of the recommended content.

[0076] Specifically, the core technologies of this step include: Collection and classification of user feedback data, summarizing and organizing explicit feedback and implicit feedback; Recommendation effect evaluation, calculating the comprehensive score of the recommended content using a weighted scoring function; Reinforcement learning optimization, adjusting the recommendation weights based on the user feedback data and optimizing the strategy; Recommendation diversity control, using the Multi-Armed Bandit (MAB) method to prevent the recommended contents from being overly concentrated; Adaptive parameter adjustment, combined with policy gradient optimization, to make the recommendation algorithm adapt to the changes in the user's interests.

[0077] In this embodiment, the collection unit of user feedback data is responsible for classifying and storing various interaction data of users during the process of reading electronic documents, so as to form a complete feedback data set.

[0078] Generally, user feedback can be represented as a feature vector: F = (f1, f2,..., f7); Where: f1 is the number of likes, that is, the number of times the user likes the content; f2 is the number of bookmarks, that is, the number of times the user adds the document to the bookmark; f3 is the number of comments, that is, the number of times the user comments; f4 is the dwell time, that is, the time the user reads the content (unit: second); f5 is the scrolling speed, that is, the speed at which the user scrolls the page (unit: pixel / second); f6 is the number of rollbacks, that is, the number of times the user returns to the previously read content; f7 is the bounce rate, that is, the probability that the user leaves without in-depth reading; F is the original recommendation ranking factor, representing the weight without time decay adjustment.

[0079] In a possible implementation manner, the system can use principal component analysis (PCA) to reduce the dimension of the above feedback data, so as to screen out the features that have the greatest impact on the optimization of the recommendation strategy. For example, if the dwell time and the number of rollbacks of the user show strong correlation in the historical data, the system can reduce some redundant features to improve the calculation efficiency.

[0080] In some embodiments, the timeliness weight of the feedback data can be calculated in combination with the time decay factor to reduce the impact of outdated data: F ′ = e -λt F; Where: F ′ is the feedback feature vector after time decay processing; λ is the time decay factor, representing the degree of decay of the data over time; t is the timestamp of the feedback data, the difference from the current time; e is the base of the natural logarithm; F is the original recommendation ranking factor, representing the weight without time decay adjustment.

[0081] In this embodiment, the recommendation effect evaluation unit quantifies the quality of the recommended content to determine whether the current recommendation strategy needs to be adjusted.

[0082] As an option, the recommendation effect can be calculated through a weighted scoring function as follows: Where: S is the comprehensive score of the recommended content; f i is the value of the i-th feedback feature; w i is the weight of the feedback feature, representing the degree of influence of this feature on the optimization of the recommendation strategy.

[0083] In some embodiments, the weight w iIt can be adaptively adjusted according to the user's historical interaction data. For example, if the system finds that the user has spent a long time on a certain type of content and has made many comments, it will automatically increase the weight values corresponding to the dwell time and the number of comments to enhance the influence of this feature.

[0084] In addition, in a possible implementation, the system can use regression analysis to calculate the contribution of different feedback features to the final score, ensuring that the optimized weight allocation is more reasonable.

[0085] In this embodiment, the recommendation strategy optimization unit combines the reinforcement learning method and continuously adjusts the parameters of the recommendation strategy according to the user feedback data.

[0086] As an option, the Q-learning method can be used to update the quality score of the recommended content (the specific formula has been given in step S3 and will not be repeated here).

[0087] In another implementation, the Policy Gradient method can be combined to make the recommendation strategy more intelligent by optimizing the policy function π(a|s). The goal of policy gradient optimization is to find the policy parameter θ that can maximize the long-term user satisfaction: Where: J(θ) is the objective optimization function, measuring the overall performance of the recommendation strategy; is the policy gradient; θ is the parameter vector of the recommendation strategy; π θ (a|s) represents the probability of selecting the recommended action a in state s; Q(s,a) is the Q value of this recommended content, that is, the predicted user satisfaction; is the mathematical expectation, used to calculate the weighted average gradient under all possible policies; is the logarithmic policy gradient, representing the derivative of the logarithm of the policy probability of taking action a with respect to the parameter θ in the given state s.

[0088] In this embodiment, to avoid the recommended content being overly concentrated in certain specific categories, the system combines the Multi-armed Bandit (MAB) method to balance between exploring new content and exploiting existing data.

[0089] As an option, the system can use the UCB (Upper Confidence Bound) strategy to calculate the priority of the recommended content (the formula has been given in step S3 and will not be repeated here).

[0090] In some embodiments, the value range of the exploration factor c can be dynamically adjusted to match the changing interests of different users. For example, if the system detects that the user's recent interests are relatively stable, it will reduce the exploration factor and decrease the recommendation frequency of new content; conversely, if the user's behavior pattern changes greatly, it will appropriately increase the exploration factor to increase the recommendation proportion of new content.

[0091] In addition, in another possible implementation, a temperature-softening (Softmax) strategy can be adopted to adjust the recommendation probability: where: P(a i ) is the probability of selecting the recommended content a i ; Q(a i ) is the quality score of this content; j is an index variable in the recommendation candidate set, used to traverse all recommendable content a j to obtain a normalized probability distribution; τ is the temperature parameter, which controls the degree of exploration. The larger the value, the more random the recommendation strategy is.

[0092] User feedback processing and adaptive optimization of the recommendation strategy (S4) achieve dynamic optimization of the recommendation strategy through the collection of user feedback data, evaluation of recommendation effects, reinforcement learning optimization, and diversity control.

[0093] The system combines algorithms such as Q-learning, policy gradient, UCB, and Softmax to continuously adjust the recommendation strategy, enabling it to adaptively optimize according to the changes in user interests, thereby enhancing the personalized experience of e-document reading.

[0094] S5. Personalized optimization step: Based on the calculation result of the reading progress, adjust the personalized reading recommendation strategy and optimize the subsequent content recommendation order to match the user's reading habits and comprehension level; The display of the recommendation result and the optimization of user interaction in step S5 are based on the user feedback processing and adaptive optimization of the recommendation strategy in step S4, further optimizing the display method of the recommended content and enhancing the user interaction experience.

[0095] Generally, the display method of the recommended content has a direct impact on the user's reading behavior and feedback. Therefore, in this step, the system not only considers the accuracy of the recommended content but also pays attention to the visual presentation of the recommended information, the content sorting method, the dynamic adjustment mechanism, and the personalized interaction optimization, thereby improving the user's reading interest and satisfaction.

[0096] As an option, this step combines methods such as interface layout optimization, interactive recommendation, reinforcement learning regulation, and personalized information annotation to make the display of the recommended content more intelligent and meet the reading preferences of different users.

[0097] Specifically, the core technologies of this step include: Optimization of the sorting of recommended content, adjusting the display order based on the user interest weight; Personalized recommendation display, dynamically adjusting the content presentation method in combination with the user's historical data; Interactive recommendation mechanism, enhancing the user experience through methods such as swiping, clicking, and folding; Dynamic interface adaptation, adaptively adjusting the recommendation module according to the device type and screen size; Information annotation and visualization enhancement, improving the information comprehension through labels, charts, etc.

[0098] In this embodiment, the recommendation result display unit optimizes the sorting of the content by calculating the personalized weight score of the recommended content, so as to ensure that the content most in line with the user's interests can be displayed preferentially.

[0099] Generally, the sorting of the recommended content is affected by multiple factors, including the user's historical interaction data, the characteristics of the current recommended content, the real-time changes of the user feedback, etc. Therefore, in this embodiment, the sorting score of the recommended content can be calculated through a weighted scoring function.

[0100] As an option, the aforementioned publicly disclosed weighted scoring function can be used to calculate the "weighted scoring function calculation": In addition, in another possible implementation, the collaborative filtering (CF) method can be combined to calculate the sorting score of the recommended content based on the preferences of similar users.

[0101] In some embodiments, a time decay factor can be combined to make the impact of the most recent user behavior on the sorting score greater. The calculation method of the time decay factor here is the same as that in step S4, and the formula will not be repeated, but it is still applied to this embodiment.

[0102] In this embodiment, in order to improve the user experience, the system can dynamically adjust the display mode of the recommended content in combination with the user's reading habits.

[0103] As an option, the recommended content can be displayed in a waterfall or pagination mode, and the scrolling method can be adjusted according to the user's behavior. For example: If the user is used to quickly browsing, the recommendation list adopts a continuous loading mode to improve the content fluency; If the user pays more attention to specific categories, partitioned recommendations can be adopted to display different categories of content in different modules.

[0104] In some embodiments, the recommended content can be classified according to the user's interest fields and different card styles can be used for differentiation. For example: News content can be displayed in the form of title + summary; Technical documents can be displayed in a table-of-contents style for easy quick positioning of key information; Multimedia content can be in the form of video preview + text introduction to improve readability.

[0105] In this embodiment, to enhance the user's active interaction ability, the recommendation system provides a series of interactive recommendation methods, including: Swipe selection recommendation: The user can adjust the recommended content by swiping left and right, similar to the "card-style recommendation" mode; Foldable recommendation: The user can expand or fold a certain type of content to reduce information overload; User-defined recommendation: The user can adjust the recommendation parameters, such as selecting the preferred content category, recommendation frequency, etc.

[0106] In a possible implementation, the system can combine reinforcement learning (RL) and continuously optimize the interaction method according to the user's interaction behavior. For example, if the system detects that the user is more inclined to swipe through the content, it will increase the proportion of "swipe-style recommendations" and reduce the use of the "pagination loading" method.

[0107] In this embodiment, to adapt to the reading needs of different devices, the recommendation system can automatically adjust the interface layout.

[0108] As an option, the system can dynamically adjust the display method of the recommended content according to the screen size and resolution. For example: On the PC side, the recommended content can adopt a grid layout to increase the information density; On the mobile side, the recommended content adopts a single-column streaming layout to improve the reading coherence.

[0109] In some embodiments, the system can combine gesture interaction to optimize the recommendation experience, such as: Two-finger zooming can be used to adjust the font size of the recommended content; Long pressing can be used to mark important content and trigger in-depth recommendations.

[0110] In this embodiment, to improve the readability of the recommended content, the system can add information tagging and visual elements to the recommended content.

[0111] As an option, the recommended content can include: Keyword highlighting, such as highlighting the terms that the user is interested in; Popularity indicator, using an icon to represent the popularity of the content; Recommendation reason, showing the "reason for recommendation" below the recommended content, such as "because you follow the XX topic".

[0112] In some embodiments, the system can combine data visualization technology to present the statistical information of the recommended content in the form of a chart. For example: A bar chart shows the reading popularity trend of the article; The word cloud graph visually presents the user's interest keywords.

[0113] In this embodiment, through technologies such as recommended content sorting optimization, personalized display, interactive recommendation, interface adaptive adjustment, and information visualization, the intelligent display of recommended content and the optimization of user interaction are realized.

[0114] The system combines reinforcement learning, collaborative filtering, and interface adaptation algorithms to continuously adjust the recommendation method, enabling it to be dynamically optimized according to the user's reading habits, thereby enhancing the user's personalized experience and the effectiveness of content recommendation.

[0115] The reading management device for electronic documents described below can be correspondingly referred to the reading management method for electronic documents described above.

[0116] Please refer to the appendix Figure 2 , the present invention also provides a reading management device for electronic documents, including: A document information entropy calculation module, configured to calculate the complexity of the electronic document text and set a reading weight; A user behavior monitoring module, configured to collect user reading behavior data, including behavior data of scrolling, staying, returning, highlighting, and taking notes; An information gain calculation module, configured to calculate the depth of understanding of the document content by the user's interaction behavior and adjust the reading progress weight; a Bayesian inference progress calculation module, configured to dynamically update the reading progress according to the user behavior data; A personalized optimization module, configured to adjust the reading recommendation strategy based on the user's historical reading behavior.

[0117] The device in this embodiment can be used to execute the above method embodiment, and its principle and technical effect are similar, which will not be elaborated here.

[0118] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for managing the reading of an electronic document, characterized in that: The following steps are involved: S1, a document information entropy calculation step, analyzing the text content of the electronic document based on the information entropy calculation method to determine the complexity of different contents; S2, user behavior data collection step, collecting the user's scrolling, staying, returning, highlighting and note-taking behavior data when reading the electronic document; S3, information gain calculation step, based on the information entropy calculation result and user behavior data, calculating the information gain of the user for different contents of the electronic document, which is used to evaluate the user's reading depth; S4, Bayesian reasoning progress calculation step, based on the user behavior data and the information gain calculation result, using the Bayesian reasoning method to update the user's reading status and dynamically calculate the user's reading progress; S5: Personalized optimization step: Based on the reading progress calculation results, adjust the personalized reading recommendation strategy and optimize the order of subsequent content recommendations to match the user's reading habits and understanding level.

2. The electronic document reading management method according to claim 1, characterized in that: The document information entropy calculation step includes: Parse the text content of electronic documents and calculate the information entropy value of each text paragraph based on the statistical model of the text content; According to the information entropy value, the complexity of each paragraph of text is determined, and different reading weights are assigned to content of different complexity; Use an entropy-based algorithm to mark important parts of the document for optimization in subsequent user behavior analysis.

3. The electronic document reading management method according to claim 1, characterized in that: The information gain calculation step comprises: Based on the complexity analysis results in the document information entropy calculation step, combined with the dwell time, scrolling behavior and highlighting in the user behavior data collection step, the information gain of the user for each content paragraph in the document is calculated; According to the information gain value, the user's understanding depth of different content is evaluated, and the user's reading progress is dynamically adjusted based on the result; For content with high information gain, more detailed follow-up content is recommended, while for content with low information gain, relatively brief or summary content can be recommended.

4. The electronic document reading management method according to claim 1, characterized in that: The information gain calculation step comprises: Calculate the initial information entropy of the document; Calculate the change in text information entropy after the user reads; The information gain calculation method is used to measure the user's reading depth; According to the information gain result, adjust the contribution weight of the page to the reading progress.

5. The electronic document reading management method according to claim 1, characterized in that: The Bayesian reasoning progress calculation step includes: Set the user's prior probability of reading; Calculate the relationship between the user's reading behavior pattern and reading completion on a certain page; Use Bayesian reasoning to calculate the probability that the user actually reads the page; The reading progress is adjusted dynamically based on the reading probability of different pages, and the reading progress calculation results are used to dynamically adjust subsequent content recommendations.

6. The electronic document reading management method according to claim 1, characterized in that: The personalized optimization step comprises: Adjust reading time allocation strategy based on user reading behavior patterns; Predict the best learning path for the user's unread parts and adjust the recommendation order. The recommended reading path is dynamically adjusted based on the user's reading behavior; Generate personalized reading reports for users, providing review and reinforcement learning suggestions.

7. An electronic document reading management device, characterized in that: The method for managing the reading of an electronic document according to any one of claims 1 to 6 comprises: Document information entropy calculation module, used to calculate the complexity of electronic document text and set reading weight; User behavior monitoring module, used to collect user reading behavior data, including scrolling, staying, returning, highlighting and note-taking behavior data; Information gain calculation module, used to calculate the depth of understanding of document content by user interaction behavior and adjust the reading progress weight; Bayesian reasoning progress calculation module, used to dynamically update reading progress based on user behavior data; The personalized optimization module is used to adjust the reading recommendation strategy based on the user's historical reading behavior.

8. The electronic document reading management device according to claim 7, characterized in that: The document information entropy calculation module includes: A text parsing unit for analyzing the text structure of an electronic document; The word frequency statistics unit is used to count the occurrence probability of words in the document; Information entropy calculation unit, used to calculate the complexity of the text and set the reading time weight.

9. The electronic document reading management device according to claim 7, characterized in that: The user behavior monitoring module includes: Time monitoring unit, used to record the time users stay on the page; A scroll monitoring unit, used to monitor the user's scrolling behavior and page turning speed; The interaction monitoring unit is used to record the user's highlight, annotation and note behavior data.

10. The electronic document reading management device according to claim 7, characterized in that: The Bayesian reasoning progress calculation module includes: A behavior probability calculation unit is used to calculate the relationship between the user's reading behavior and the reading completion degree; A Bayesian reasoning unit, used to calculate the user's true reading probability using a Bayesian method; The progress dynamic adjustment unit is used to dynamically calculate the reading progress based on the reading probability and optimize the learning path.

Images