An electronic book hierarchical compression pushing method based on behavior semantics

Abstract

The present application relates to the technical field of digital library and electronic book management, and particularly relates to an electronic book hierarchical compression pushing method based on behavior semantics; comprising the following steps: step one: collecting behavior data of reading electronic books, including page staying time, page turning direction and speed, highlighting and comment operation and chapter jumping track; step two: performing time sequence feature extraction and group aggregation on the behavior data to generate a behavior semantic feature vector; step three: generating hierarchical labels based on the behavior semantic feature vector, including core degree, difficulty index and interaction heat; step four: performing adaptive hierarchical compression according to the hierarchical labels to generate at least two versions of different compression quality levels; step five: responding to a reading request of a reader, dynamically selecting and pushing a compressed version of a corresponding chapter according to device information, network condition and historical reading mode of the current reader; the problem of how to perform adaptive hierarchical compression on electronic book content according to the behavior semantic features of different chapters is solved.

Description

Technical Field

[0001] This invention relates to the field of digital library and e-book management technology, specifically to a hierarchical compression and push method for e-books based on behavioral semantics. Background Technology

[0002] With the popularization of digital libraries and mobile reading, e-books have become an important carrier of knowledge dissemination. Currently, library e-book systems generally adopt a unified storage and delivery model for entire files, with common formats including PDF, EPUB, and MOBI. This traditional approach has the following technical drawbacks: 1. Waste of storage and transmission resources: In academic monographs and scanned documents, the information density and reader attention vary greatly between different chapters. However, existing systems use the same compression and transmission strategy for all chapters, resulting in a large number of low-attention chapters occupying valuable cloud storage and network bandwidth resources; 2. Poor mobile reading experience: On smartphones, tablets, and other mobile devices, complex content such as code snippets, mathematical formulas, and high-resolution images require refined formatting, while novel chapters or appendices have lower image quality requirements. The existing system cannot dynamically adjust the pushed version according to the characteristics of the chapter, resulting in slow loading for readers in weak network environments, or receiving low-quality versions on high-performance devices; third, the system lacks perception and feedback on reading quality: traditional e-book management systems only record coarse-grained indicators such as the number of borrowings and total reading time, and cannot perceive deeper semantic information such as the "difficulty level", "skipping ratio", and "rereading frequency" of specific chapters, thus failing to proactively optimize the delivery form of library resources; fourth, static compression lacks adaptability: existing compression technologies (such as PDF reduction and JPEG lossy compression) set parameters uniformly for the entire book, and once compression is completed, they remain fixed and cannot be dynamically adjusted according to changes in the reading behavior of the reader group.

[0003] Therefore, there is an urgent need for a method that can use the semantics of group reading behavior to guide the classification and dynamic compression and delivery of e-books in order to solve the above-mentioned technical problems. Summary of the Invention

[0004] To overcome the aforementioned technical problems in the prior art, this invention provides a method for hierarchical compression and push of e-books based on behavioral semantics. This method solves the problem of how to perform adaptive hierarchical compression of e-book content based on the behavioral semantic features of different chapters, generating multiple versions with different compression quality levels. It also solves the problem of how to dynamically select and push the most suitable compressed version based on the current reader's device capabilities, network conditions, and historical reading patterns when responding to a reader's reading request.

[0005] To achieve the above objectives, this invention provides a method for hierarchical compression and push of e-books based on behavioral semantics, comprising the following steps: Step 1: Collecting behavioral data from multiple readers during e-book reading, the behavioral data including at least page dwell time, page turning direction and speed, highlighting and annotation operations, and chapter jump trajectory; Step 2: Extracting temporal features and aggregating the behavioral data to generate a behavioral semantic feature vector for each chapter; Step 3: Based on the behavioral semantic feature vector, using a graph neural network combined with the e-book's table of contents structure, generating hierarchical tags for each chapter, the hierarchical tags including core content, difficulty index, and interaction popularity; Step 4: Performing adaptive hierarchical compression on the e-book source file according to the hierarchical tags to generate at least two versions with different compression quality levels; Step 5: Responding to readers' reading requests, dynamically selecting and pushing the compressed version of the corresponding chapter based on the current reader's device information, network status, and historical reading patterns.

[0006] Preferably, the behavioral data also includes font scaling frequency, scrolling speed, page reread count, and explicit evaluation marks by readers for specific paragraphs.

[0007] Preferably, the temporal feature extraction employs a temporal convolutional network or a long short-term memory network to encode each reader's continuous behavior on a single page into a page embedding vector, which is then aggregated by chapter into a mean vector and a variance vector.

[0008] Preferably, the graph neural network uses chapters as nodes and chapter order and parent-child relationships in the directory as edges. The initial features of the nodes are the behavioral semantic feature vectors generated in step two, and after graph convolution, normalized hierarchical labels are output.

[0009] Preferably, the adaptive hierarchical compression in step four includes: if the core score is higher than the first threshold, retaining the original image, vector graphics, and embedded fonts, and using lossless compression; if the core score is between the first threshold and the second threshold, reducing the image quality to a preset percentage and removing redundant metadata; if the core score is lower than the second threshold, extracting only the plain text layer, and calling a large language model to generate chapter summaries to replace the original image and complex layout.

[0010] Preferably, the dynamic selection and push in step five includes: prioritizing the push of the compressed version of the current chapter; predicting the next chapter the reader will read based on the reader's historical skipping frequency and pushing a lightweight version in advance; and automatically upgrading subsequent chapters to a higher compression quality version when continuous rapid page turning is detected.

[0011] Preferably, the collected behavioral data is processed with differential privacy noise on the client side, and only the aggregated group statistical features are stored on the server side, which cannot be inferred from individual readers.

[0012] Preferably, the hierarchical labels are updated periodically. When the amount of new behavioral data reaches a preset threshold, the hierarchical and compression process is re-executed, and the new version replaces the old version for storage.

[0013] The source file formats of the e-book include PDF, EPUB, MOBI, or PDF documents generated from scanned images. The compression engine module calls the corresponding parsing and compression toolchain according to the file format.

[0014] Preferably, the device information includes screen resolution, whether it supports high dynamic range rendering, and CPU computing power level, and the network status is obtained by measuring round-trip latency and download speed in real time.

[0015] The present invention has at least the following technical effects through the technical solution provided by the present invention: By retaining only the plain text skeleton or summary for low-interest chapters, the average storage footprint of large academic ebooks is reduced, mobile first-screen loading time is shortened, and storage and bandwidth costs are significantly reduced. High-interaction chapters automatically receive high-quality rendering, while low-value chapters do not waste bandwidth. Content can be quickly filtered through automatically generated chapter summaries, reducing unnecessary page-turning time and improving reading experience and information retrieval efficiency. Hierarchical tags are regularly updated with new behavioral data, and compression strategies are continuously optimized, making ebook resources "smarter" with use and possessing adaptive evolution capabilities. Privacy protection is built-in; differential privacy noise processing ensures that individual behavior is untraceable, meeting data compliance requirements. It is cross-format compatible, supporting multiple mainstream ebook formats such as PDF, EPUB, and MOBI, demonstrating strong adaptability. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the behavior semantics-based hierarchical compression and push method for e-books provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of temporal feature extraction and group aggregation in step two of this invention; Figure 3 This is a schematic diagram of the generation of hierarchical labels by the neural network in step three of this invention. Detailed Implementation

[0017] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0018] In this invention, the terms "system" and "network" are used interchangeably. "Multiple" refers to two or more; therefore, in this invention, "multiple" can also be understood as "at least two." "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / ", unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship. Furthermore, it should be understood that in the description of this invention, terms such as "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or order.

[0019] Please see Figure 1 This invention provides a method for hierarchical compression and push of e-books based on behavioral semantics, comprising: Step 1: collecting behavioral data from multiple readers during e-book reading, the behavioral data including at least page dwell time, page turning direction and speed, highlighting and annotation operations, and chapter jump trajectory; Step 2: extracting temporal features and aggregating the behavioral data to generate a behavioral semantic feature vector for each chapter; Step 3: generating hierarchical tags for each chapter based on the behavioral semantic feature vectors, using a graph neural network combined with the e-book's table of contents structure, the hierarchical tags including core content, difficulty index, and interaction popularity; Step 4: performing adaptive hierarchical compression on the e-book source file according to the hierarchical tags to generate at least two versions with different compression quality levels; Step 5: responding to readers' reading requests, dynamically selecting and pushing the compressed version of the corresponding chapter based on the current reader's device information, network status, and historical reading patterns.

[0020] In this embodiment of the invention, step one mainly involves collecting reading behavior data. With user authorization and anonymization, a lightweight collection module at the front end of the reader acquires reader reading behavior data, including at least: page dwell time, page turning direction and speed, highlighting, annotation, bookmarking, chapter jump trajectory, font scaling frequency, scrolling speed, number of page rereads, and explicit paragraph evaluation marks. The data is uploaded to the client after differential privacy noise addition, and the server only stores the group aggregation features and does not retain individual identifiable information; then step two is executed.

[0021] In embodiments of the present invention, such as Figure 2 and Figure 3As shown, step two mainly extracts behavioral semantic features. The behavioral data is encoded using a Temporal Convolutional Network (TCN) or a Long Short-Term Memory (LSTM) network, converting continuous behaviors on a single page into page embedding vectors. These vectors are then aggregated by chapter into mean and variance vectors, forming chapter-level behavioral semantic feature vectors. Next, step three generates chapter-level hierarchical labels. Specifically, a graph structure is constructed using chapters as nodes and chapter order and parent-child relationships in the table of contents as edges. The behavioral semantic feature vectors are used as initial node features, input to a Graph Neural Network (GNN) for graph convolution, and output normalized three-dimensional hierarchical labels: 1: Coreness, representing the importance of content and the necessity of retaining details; 2: Difficulty Index, representing the difficulty of understanding and the need for assistance; 3: Interaction Popularity, representing the actual frequency of reader use. The hierarchical labels are updated periodically, recalculated when the amount of new data reaches a threshold. Then, step four is executed.

[0022] In this embodiment of the invention, step four mainly performs adaptive hierarchical compression, performing three levels of hierarchical compression based on core density to generate at least two quality versions: When the core density exceeds the first threshold, retain the original image, vector graphics, and embedded fonts, and perform lossless compression. When the core count is between the first and second thresholds, reduce image quality and remove redundant metadata. When the coreness is less than the second threshold, only the plain text layer is retained, and a large language model is used to generate chapter summaries to replace images and complex layouts. It supports ebook source file formats such as PDF, EPUB, MOBI, and scanned PDF. The compression engine calls the corresponding parsing and compression toolchain according to the format.

[0023] In this embodiment of the invention, step five mainly performs dynamic selection and intelligent push. When responding to a reading request, the following information is considered to select the version: device information: resolution, HDR support, CPU computing power; network status: real-time RTT and download speed; personal reading mode: historical skipping frequency. The push strategy is as follows: prioritize pushing the best version of the current chapter; predict the next chapter and push a lightweight version in advance; when flipping pages rapidly, automatically upgrade subsequent chapters to a higher quality version; version switching is transparent to the user and does not interrupt reading.

[0024] In one implementation method, the method is illustrated in detail using an academic monograph titled "Introduction to Machine Learning" (PDF format, 12 chapters, approximately 300 pages) from a university's digital library as an example. The overall implementation process is detailed below: Step 1: Behavioral data collection. Embed the JavaScript SDK in the front end of the e-book reader (including web readers and mobile apps) and register the following event listeners: 1.1: Record the number of milliseconds the reader stays on each page and report it when page visibility changes or a page-turning event is triggered; 1.2: Record the page turning direction (next page / previous page) and page turning interval time to calculate the page turning rate; 1.3: Record the start position, end position, and length of the highlighted text; 1.4: Record the annotation content and the associated page ID; 1.5: Record the path of jumping directly from one chapter to another (e.g., jumping from page 10 of chapter 3 to page 1 of chapter 8).

[0025] The collected data undergoes preliminary filtering before reporting: false triggering events with a dwell time of less than 0.5 seconds (such as rapid scrolling) and abnormal dwell times exceeding 30 minutes (such as readers leaving the device without closing the page) are removed. The data is encapsulated in JSON format and reported to the server in batches every 10 seconds or after each chapter is completed. To protect reader privacy, this implementation employs differential privacy technology on the client side: random noise following a Laplace distribution is added to the statistical value of each behavioral event (such as dwell time), with the noise scale parameter set to 0.5, making it impossible for the server to accurately infer individual behavior from a single reported data point.

[0026] Step Two: After receiving the behavioral data, the server groups it by book ID and chapter ID. For a single reader's behavioral sequence on a single page, a Temporal Convolutional Network (TCN) is used for encoding. The TCN consists of 4 dilated convolutional layers with a kernel size of 3 and dilation rates of 1, 2, 4, and 8. The input is a continuous page behavior vector of length L (features for each step include: dwell time, whether page was turned back, whether there was a highlight, and whether there was a jump in / out), and the output is a 128-dimensional page embedding vector. For a chapter containing P pages, the embedding vectors of all pages are aggregated according to the reader dimension, and the mean and variance are calculated to obtain a 256-dimensional chapter semantic feature vector (128-dimensional mean + 128-dimensional variance). Then, the feature vectors of the chapter are averaged across all readers to obtain the initial chapter features at the group level. The above process can be expressed mathematically as follows: Let the embedding vector of reader r on page i of chapter c be... Then the reader's aggregation characteristics for this chapter are:

[0027] The characteristics of the group are:

[0028] Where R represents the total number of readers who have read this chapter.

[0029] Step 3: Generate hierarchical labels using a graph neural network, constructing a chapter relationship graph G=(V,E), where the node set V corresponds to the 12 chapters of the book, and the edge set E contains two types: Sequential edges: Chapter i points to Chapter i+1 (i=1..11) Directory hierarchy edge: If some chapters in the directory structure (such as 1.1, 1.2, 1.3) belong to the same parent chapter 1, then a bidirectional edge is established between the parent chapter and the child chapter.

[0030] A two-layer graph convolutional network (GCN) is used, with each layer having an output dimension of 64 and an activation function of ReLU. The initial features of the nodes are obtained in step two. (256-dimensional). After graph convolution, each node outputs a 3-dimensional vector, which is normalized to the [0,1] interval using the Sigmoid function, corresponding to core score, difficulty index, and interaction popularity, respectively. Assume that after training with behavioral data from multiple readers accumulated over 30 days, the core score, difficulty index, and interaction popularity values ​​for some chapters are obtained.

[0031] Step 4: Adaptive hierarchical compression. Based on the hierarchical tags, set the first threshold to 0.7 and the second threshold to 0.3; perform compression on the ebook source file (PDF): Assuming the core score of Chapter 3 is 0.85, which is greater than 0.7 (the first threshold), then a tool (such as mutool) is called to extract the original image, retain all vector graphics (for example, using pdfimages -j to retain JPEG2000 lossless format), embed fonts without subsetting, and finally generate a high-quality PDF fragment with a file size of about 95% of the original size.

[0032] Assuming the core score of Chapter 7 is 0.45, which falls between the first and second thresholds (i.e., 0.3 ≤ 0.45 ≤ 0.7), ImageMagick is used to reduce the quality of all images to 60%, remove embedded fonts (and replace them with system fonts), remove annotation metadata from the PDF, and compress the file size to approximately 50% of the original size.

[0033] Assuming the core score of Chapter 10 is 0.12, which is less than 0.3 (the second threshold), we use pdftotext to extract the plain text, discard all images, and then call a large language model (such as GPT-3.5-turbo) to generate a chapter summary of no more than 150 words. The compressed file size is about 2% of the original size.

[0034] The final product consists of three versions of the ebook: Full version: All chapters retain the highest quality; Standard version: High-core chapters retain high quality, medium-core chapters are moderately compressed, and low-core chapters consist of only text and summary; Lightweight version: All chapters use medium or lower compression strategies (suitable for mobile devices in weak network scenarios).

[0035] The three versions are stored in an object store (such as MinIO), and the version number is associated with the chapter ID.

[0036] Step 5: Dynamic Selection and Push Notification. When reader A requests to read "Introduction to Machine Learning" via mobile phone, the following push notification logic is executed: 1. Obtain reader A's historical reading pattern (query from the database): In the past 3 ebooks, this reader skipped pages 0.6 times (i.e., 60% of the chapters were skipped after only reading the first few pages), which is considered a high-frequency skipping reader.

[0037] 2. The current chapter is Chapter 3. Based on the device information and network conditions, the push scheduling module selects the standard version instead of the full version to save data traffic.

[0038] 3. Based on historical skipping frequency, it is predicted that reader A has a high probability of skipping later chapters, so only a lightweight version of Chapter 4 is preloaded as a placeholder.

[0039] 4. Real-time monitoring of page-turning events: When reader A rapidly turns pages in Chapter 3 (turning 8 pages within 5 seconds), the upgrade strategy is triggered, automatically upgrading Chapter 4 from the lightweight version to the standard version, and downloading the lightweight version of Chapter 5 in advance.

[0040] All version switching is completed in the background, without the reader's awareness. Ultimately, reader A's reading smoothness metrics show that the first screen loading time is shorter than the traditional method, and the data consumption throughout the book is also lower.

[0041] Furthermore, on the 1st of each month, the system analyzes newly added behavioral data from the past 30 days. If a book's new readership exceeds 100, a reclassification process is triggered: historical group characteristics are merged with new data characteristics (using an exponentially weighted moving average with weight α=0.7), the graph neural network is rerun, and the classification labels are updated. If the difference (cosine distance) between the new and old labels exceeds 0.1, recompression is triggered, a new version is generated, and the old version is stored, while the previous version is retained as a fallback. All behavioral data is uploaded after adding Laplacian noise on the client side. The server only stores the aggregated group statistical features, not the original individual behavioral sequences. For scenarios requiring model updates, a federated learning framework is used: each client calculates gradients locally, uploads the noisy gradients, and the server aggregates and updates the global model; the original data does not leave the client.

[0042] This invention reduces the average storage footprint of large academic ebooks by retaining only the plain text skeleton or summary for low-interest chapters, shortening the first-screen loading time on mobile devices and significantly reducing storage and bandwidth costs. High-interaction chapters automatically receive high-quality rendering, while low-value chapters do not waste bandwidth. Furthermore, content can be quickly filtered through automatically generated chapter summaries, reducing unnecessary page-turning time and improving the reading experience and information retrieval efficiency. Hierarchical tags are regularly updated with new behavioral data, and compression strategies are continuously optimized, making ebook resources increasingly "intelligent" and possessing adaptive evolution capabilities. Built-in privacy protection and differential privacy noise processing ensure that individual behavior is untraceable, complying with data compliance requirements. It is cross-format compatible, supporting multiple mainstream ebook formats such as PDF, EPUB, and MOBI, demonstrating strong adaptability.

[0043] Furthermore, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the present invention.

[0044] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above embodiments. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention.

[0045] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not describe the various possible combinations separately.

[0046] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0047] Furthermore, various different implementations of the present invention can be combined arbitrarily, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed in the present invention.

Claims

1. A hierarchical compression and push method for e-books based on behavioral semantics, characterized in that, Includes the following steps: Step 1: Collect behavioral data from multiple readers during the e-book reading process. The behavioral data includes at least page dwell time, page turning direction and speed, highlighting and annotation operations, and chapter jumping trajectory. Step 2: Extract temporal features and aggregate groups from the behavioral data to generate behavioral semantic feature vectors for each chapter; Step 3: Based on the behavioral semantic feature vector, use a graph neural network combined with the e-book's table of contents structure to generate hierarchical tags for each chapter. The hierarchical tags include core content, difficulty index, and interaction popularity. Step 4: Perform adaptive layered compression on the e-book source file according to the hierarchical tags to generate at least two versions with different compression quality levels; Step 5: Respond to the reader's reading request and dynamically select and push the compressed version of the corresponding chapter based on the reader's current device information, network status, and historical reading mode.

2. The method according to claim 1, characterized in that, The behavioral data also includes font scaling frequency, scrolling speed, page reread count, and explicit rating tags from readers for specific paragraphs.

3. The method according to claim 1, characterized in that, The temporal feature extraction employs a temporal convolutional network or a long short-term memory network to encode each reader's continuous behavior on a single page into a page embedding vector, which is then aggregated by chapter into a mean vector and a variance vector.

4. The method according to claim 1, characterized in that, The graph neural network uses chapters as nodes and chapter order and parent-child relationships in the directory as edges. The initial features of the nodes are the behavioral semantic feature vectors generated in step two. After graph convolution, normalized hierarchical labels are output.

5. The method according to claim 1, characterized in that, The adaptive hierarchical compression in step four includes: If the core density is higher than the first threshold, the original image, vector graphics, and embedded fonts are retained, and lossless compression is used. If the core count is between the first threshold and the second threshold, the image quality will be reduced to a preset percentage and redundant metadata will be removed. If the core content is lower than the second threshold, only the plain text layer is extracted, and a large language model is used to generate chapter summaries to replace the original images and complex layouts.

6. The method according to claim 1, characterized in that, The dynamic selection and push in step five includes: Prioritize pushing the compressed version of the current chapter; Based on the reader's historical skipping frequency, a lightweight version is pushed to them in advance; When rapid page turning is detected, subsequent chapters will be automatically upgraded to a higher compression quality version.

7. The method according to claim 1, characterized in that, Also includes: The collected behavioral data is processed with differential privacy noise on the client side, and only the aggregated group statistical features are stored on the server side, which cannot be inferred from individual readers.

8. The method according to claim 1, characterized in that, The hierarchical labels are updated periodically. When the amount of new behavioral data reaches a preset threshold, the hierarchical and compression process is re-executed, and the new version replaces the old version for storage.

9. The method according to claim 1, characterized in that, The source file formats of the e-book include PDF, EPUB, MOBI, or PDF documents generated from scanned images. The compression engine module calls the corresponding parsing and compression toolchain according to the file format.

10. The method according to claim 1, characterized in that, The device information includes screen resolution, whether it supports high dynamic range rendering, and CPU computing power level. The network status is obtained by measuring round-trip latency and download speed in real time.

Images