Digital book generation method based on semantic understanding

Abstract

The invention relates to the technical field of digitization, in particular to a digital book generation method based on semantic understanding. The method comprises the following steps: constructing an image-text association structure based on reference literatures, and extracting core knowledge points from the reference literatures and demand data; constructing a semantic unit network by analyzing a logical relationship between the core knowledge points, and binding a semantic template for each semantic unit based on the semantic unit network and a preset binding rule to obtain a unit template mapping table; performing content generation and verification on each semantic unit according to the logical relationship in the semantic unit network and the unit template mapping table, and matching and inserting a picture according to the image-text association structure to obtain a digital first draft of the book; updating the semantic unit network based on an editing instruction input by a user, and performing content generation and verification again to obtain a final digital book; according to the method, the generation quality of the digital book is improved, and the problems of poor logicality and low generation efficiency in a traditional book generation method are effectively solved.

Description

Technical Field

[0001] This application relates to the field of digital technology, specifically to a method for generating digital books based on semantic understanding. Background Technology

[0002] In today's era of rapid development in digital technology and the information industry, book digitization has become a core means of cultural inheritance, knowledge dissemination, and efficient resource utilization. Whether it is the rapid supplementation of professional knowledge in the academic research field, the demand for diversified teaching materials in the education field, or the immediate consumption of various popular science and lifestyle books in the mass reading market, all require book content to be generated and disseminated in a more efficient way.

[0003] Throughout the development of book creation and dissemination, traditional book creation models have long faced the predicament of inefficiency. These models rely heavily on manual labor, from conceptualizing the theme and building the framework to writing the content, organizing the logic, and then undergoing multiple rounds of revisions, optimizations, and format adjustments—each step requiring a significant investment of time and energy from the creators. Limited by factors such as labor costs, the boundaries of knowledge reserves, and the allocation of creative energy, the creation cycle of a traditional book often lasts for months or even years, making it difficult to quickly respond to the market's immediate demand for various types of knowledge content. This inefficiency, stemming from human limitations, is particularly pronounced when creating highly specialized books with complex content systems, severely restricting the speed of content production and the breadth of its dissemination.

[0004] Against this backdrop, book digitization technology has emerged and gradually developed. However, current mainstream book digitization technologies still have significant shortcomings and have failed to fundamentally solve the problem of low efficiency in traditional book creation. Existing technologies mostly focus on the digitization of published physical books, with scanning conversion based on optical character recognition (OCR) being the most widely used basic method. This method acquires images of book pages using high-resolution scanning equipment and then uses OCR technology to convert the text in the images into editable text. It has advantages such as fast processing speed and batch processing capabilities and has been applied in large-scale digitization projects such as the National Library of China. However, this method essentially only digitizes the book carrier and does not involve technological innovation in the book content generation process. Furthermore, existing digitization technologies cannot automatically complete the semantic analysis, logical organization, and structured generation of book content based on specific themes or knowledge needs, making it difficult to truly reduce the dependence of book creation on human intervention and failing to meet the actual need for rapidly generating high-quality book content.

[0005] The existing technologies mentioned above all suffer from the problems proposed in this background: they can only convert physical books into electronic text, without addressing the innovation of the book content generation process, and cannot solve the efficiency problem of traditional creation relying on manual labor and having a long cycle; they lack the ability to analyze and generate content, making it difficult to reduce the reliance on manual labor in creation and failing to meet the demand for efficient generation of high-quality books.

[0006] The information disclosed in this background section is intended only to enhance the understanding of the overall background of this application and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0007] The purpose of this application is to provide a semantic understanding-based method for generating digital books, which to some extent solves the problems raised in the background technology, improves the efficiency of digital book generation, and reduces the reliance on manual creation.

[0008] To achieve the above objectives, this application provides the following technical solution:

[0009] This application provides a method for generating digital books based on semantic understanding, including the following steps:

[0010] Receive references and determine the number of references; standardize the references and extract images and annotations to obtain the text-image association structure;

[0011] Receive user-input requirement data and extract core knowledge points from the requirement data and references; analyze the logical relationships between the core knowledge points and construct a semantic unit network composed of different semantic units based on the logical relationships;

[0012] Based on the semantic unit network analysis, the type and semantic density of each semantic unit are analyzed, and a semantic template is bound to each semantic unit based on the preset binding rules to obtain a unit template mapping table;

[0013] The processing order is determined based on the logical relationships in the semantic unit network, and content is generated and verified for each semantic unit based on the processing order and the unit template mapping table to obtain the initial text draft of the book; based on the image-text association structure, images are matched and inserted into the initial text draft of the book to obtain the initial digital draft of the book.

[0014] When an editing instruction is received from a user regarding the initial digital draft of the book, the semantic unit network is updated based on the editing instruction, and the content is regenerated and verified to obtain the final digital book.

[0015] As a preferred embodiment of the semantic understanding-based book digitization method described in this application, the method for determining the processing order specifically includes:

[0016] Traverse the semantic unit network and extract the logical relationship corresponding to each semantic unit; count the number of logical relationships in which the semantic unit is a successor unit in the preceding dependency and the relationship type is a pre-dependency; and use the number as the number of unsatisfied dependencies of the corresponding semantic unit.

[0017] All semantic units in the semantic unit network with zero unmet dependencies and a state of pending generation are designated as ready units, thus obtaining a ready unit queue.

[0018] Traverse the ready unit queue and generate and verify the content of any ready unit in the ready unit queue.

[0019] As a preferred embodiment of the semantic understanding-based book digitization method described in this application, the method for determining the processing order further includes:

[0020] After completing content generation and verification, obtain all semantic units that use the ready unit as a preceding unit to obtain the set of successor units corresponding to the ready unit;

[0021] Decrement the number of unmet dependencies of all successor units in the successor unit set by 1, and remove the ready unit from the ready unit queue;

[0022] Traverse the set of successor units corresponding to the ready unit. If the number of dependencies to be satisfied for any successor unit is 0 and the unit status of the successor unit is pending generation, then add the corresponding successor unit to the ready unit queue.

[0023] As a preferred embodiment of the semantic understanding-based book digitization method described in this application, the content generation and verification method specifically includes:

[0024] For any ready unit that performs content generation and verification, the text generation model is invoked to generate a detailed and coherent main text using the core content of the ready unit as the seed.

[0025] According to the unit template mapping table, obtain the semantic template bound to the ready unit; fill the text content into the predefined placeholders of the corresponding semantic template to form a complete, formatted content block;

[0026] The core content coverage and logical coherence of the content block are verified. If both the core content coverage and logical coherence of the content block pass the verification, the corresponding ready unit is determined to have completed content generation and verification, and the unit status of the corresponding ready unit is updated to verified.

[0027] As a preferred embodiment of the semantic understanding-based book digitization method described in this application, the content generation and verification method further includes:

[0028] If the core content coverage or logical coherence check of the content block fails, the text generation model will be called again for rewriting.

[0029] Set a maximum number of rewrites. If the number of rewrites of the ready unit reaches the maximum number of rewrites and still fails the core content coverage check or logical coherence check of the content block, then update the semantic status of the corresponding ready unit to require manual review.

[0030] When the ready queue is empty, all semantic units in the semantic unit network are traversed. If there are still semantic units in the state of "to be generated", the unit state of the corresponding semantic unit is updated to "requires manual review", and the semantic unit is generated and verified.

[0031] As a preferred embodiment of the semantic understanding-based book digitization method described in this application, the method for constructing the semantic unit network specifically includes:

[0032] Each core knowledge point is instantiated as an independent semantic unit, and each semantic unit includes at least core content, unit state, and semantic density; the semantic density is calculated as follows:

[0033] The text length features of the core content, the statistical features of keywords related to the core knowledge points in the reference materials, the complexity features of the logical structure presented by the core content, and the basic weight coefficients associated with the knowledge type are obtained respectively.

[0034] Based on the aforementioned basic weight coefficients, the text length feature, statistical feature, and complexity feature are summed to obtain the semantic density.

[0035] As a preferred embodiment of the semantic understanding-based book digitization method described in this application, the method for constructing the semantic unit network further includes:

[0036] The frequency of different core knowledge points appearing together in the same paragraph or adjacent sentences in the input text is statistically analyzed, and the sentence structure of the input text is analyzed by combining dependency parsing or semantic role labeling to identify the logical relationship between different core knowledge points that appear at the same frequency.

[0037] Each core knowledge point is transformed into a high-dimensional vector, and the cosine similarity between the high-dimensional vectors is calculated. The core knowledge points corresponding to the high-dimensional vectors with cosine similarity greater than the preset similarity threshold are selected, and the corresponding logical relationships are identified based on the type attributes and universality of the corresponding core knowledge points.

[0038] By treating all semantic units as nodes and the identified logical relationships as edges, a semantic unit network with a graph data structure is constructed.

[0039] As a preferred embodiment of the semantic understanding-based book digitization method described in this application, the binding rules are as follows:

[0040] Traverse each semantic unit in the semantic unit network, and use the knowledge type of the semantic unit as the first condition to filter out the semantic templates that contain the knowledge type of the corresponding semantic unit in all knowledge type ranges, and use them as candidate semantic templates for the corresponding semantic unit.

[0041] Determine the density level of all semantic units; using the density level of the semantic units as the second condition, select candidate semantic templates from the candidate semantic templates whose semantic density range matches the density level of the corresponding semantic units, and use them as the semantic templates of the corresponding semantic units.

[0042] If a semantic unit corresponds to multiple semantic templates, the final semantic template is selected based on predefined priority rules.

[0043] As a preferred embodiment of the semantic understanding-based book digitization method described in this application, the editing instructions include the location information of the editing occurrence, the operation action, and the operation content; the operation action includes modification, deletion, and addition.

[0044] Based on the location information of the editing, the corresponding content block is located, and the corresponding semantic unit is located according to the relationship between the content block and the semantic unit. The semantic unit is recorded as the target unit.

[0045] If the operation is deletion, the unit status of the target unit is updated to invalid; if the operation is addition, and an independent knowledge point is added to the target unit, a new semantic unit is generated based on the added knowledge point.

[0046] New semantic units are added to the semantic unit network, and the corresponding semantic unit's unit state is set to the pending generation state.

[0047] As a preferred embodiment of the semantic understanding-based book digitization method described in this application, if the operation is modification, the unit status of the target unit is updated to "user has modified", and it is determined whether the core content of the target unit has been modified; if the core content has been modified, the semantic density of the target unit is recalculated.

[0048] Traverse the semantic unit network to obtain all semantic units that use the target unit as a preceding unit, thus obtaining the affected units;

[0049] Update the semantic state of all target units and their corresponding affected units to the state to be generated;

[0050] Determine the processing order of all target units and their corresponding affected units, and regenerate and verify the content of all target units and their corresponding affected units based on the processing order and the unit template mapping table;

[0051] The regenerated target unit and the corresponding content blocks of the affected unit are reassembled to obtain the final digital book.

[0052] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0053] A text-image association structure is constructed based on references, and core knowledge points are extracted from references and requirement data. A semantic unit network is built by analyzing the logical relationships between these core knowledge points, ensuring the inherent logic and coherence of the generated content. The type and semantic density of each semantic unit are analyzed based on the semantic unit network, and semantic templates are bound to each semantic unit based on preset binding rules, achieving deep adaptation between content and presentation. The processing order is determined according to the logical relationships in the semantic unit network, and content generation and verification are performed for each semantic unit based on the processing order and the unit template mapping table. Images are matched and inserted based on the text-image association structure, improving the quality of the initial digital draft and reducing the user's editing burden. The semantic unit network is updated according to the user's editing instructions, and content generation and verification are performed again, improving the quality of the digital book and effectively solving the problems of poor logic, rigid format, and low iteration efficiency in traditional book generation methods. Attached Figure Description

[0054] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0055] Figure 1A flowchart illustrating a book digitization method based on semantic understanding provided in this application;

[0056] Figure 2 A flowchart illustrating the process order determination method for a semantic understanding-based book digitization generation method provided in this application. Detailed Implementation

[0057] The technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solution of this application, rather than limitations thereof. In the absence of conflict, the embodiments and technical features in the embodiments can be combined with each other.

[0058] like Figure 1 As shown in the figure, this embodiment introduces a method for generating digital books based on semantic understanding, including:

[0059] The process involves receiving and determining the number of references; standardizing the references and extracting images and annotations to obtain a text-image structure; the references are in PDF format; the standardization process involves using a PDF parsing tool to read the file header information and page layout metadata of the references, and using image processing algorithms to remove scanning noise, correct tilted pages, and filter annotations and watermark layers; the processed references are then converted into a machine-readable standardized format, such as Markdown or JSON.

[0060] The method for obtaining the image-text association structure is as follows: a PDF rendering tool is used to render the PDF page into an image; an edge detection algorithm is used to identify the image region in the image and extract the image; an image identifier is assigned to each image, and the page number and page coordinates of the image are recorded; the annotation text of the adjacent area of ​​the image is located by a text position association algorithm, and a mapping relationship is established between the image identifier and the annotation text and the corresponding text paragraph to form the image-text association structure.

[0061] The system receives user-inputted requirement data and extracts core knowledge points from the requirement data and references. The requirement data includes the book title, outline structure, and core content. The core content refers to the key information input by the user that can directly define the core direction, core knowledge scope, and core objectives of the book's digital generation.

[0062] The method for extracting the core knowledge points is as follows:

[0063] The user-inputted book title, outline structure, core content, and references are used as input text. A pre-trained semantic understanding model performs named entity recognition and keyword extraction on the input text, initially identifying all important professional terms, concept names, and entities in the input text to obtain candidate knowledge points.

[0064] Calculate the semantic relevance between the candidate knowledge points, and combine graph computing algorithms to calculate the importance score of each candidate knowledge point; based on the importance scores, filter out core knowledge points to eliminate redundant terms;

[0065] In this embodiment, the semantic relevance is calculated by calculating the cosine similarity of each candidate knowledge point in the vector space; optionally, the semantic relevance can also be obtained by statistically analyzing the frequency of co-occurrence of candidate knowledge points in the input text.

[0066] Analyze the logical relationships between the core knowledge points, and construct a semantic unit network composed of different semantic units based on the logical relationships;

[0067] Specifically, each core knowledge point is instantiated as an independent semantic unit. Each semantic unit includes at least a unit identifier, core content, knowledge type, unit state, and semantic density. The core content refers to a complete textual statement that summarizes the essence of the core knowledge point. It is obtained by using a key sentence extraction algorithm to extract the most generalized sentences from the reference materials associated with the core knowledge point.

[0068] Optionally, the core content can also be obtained by calling a text generation model, taking the core knowledge points as input, and automatically generating a complete text statement based on contextual semantics; the knowledge type is obtained by: based on a predefined relation type library, using a classifier or machine learning classification model to map the core knowledge points to the relation type library to obtain the knowledge type;

[0069] The unit status is used to identify the stage of the corresponding semantic unit in the book generation lifecycle, including: pending generation, verified, requiring manual review, modified by the user, and expired; pending generation is the initial state of all semantic units;

[0070] The semantic density is used to pre-evaluate the relative complexity and richness of the information carried by the core knowledge point before content generation. It is calculated as follows: the text length feature of the core content, the statistical features of keywords related to the core knowledge point in the reference materials, the complexity feature of the logical structure presented by the core content, and the basic weight coefficient associated with the knowledge type are obtained respectively; based on the basic weight coefficient, the text length feature, statistical feature, and complexity feature are summed to obtain the semantic density.

[0071] Optionally, the total number of characters in the core content and the total number of words after word segmentation are calculated to obtain the text length feature; the total number of times the keywords appear in the reference materials and the distribution range of the keywords in different chapters or paragraphs are statistically analyzed to obtain the statistical feature; the complexity feature is obtained by identifying and statistically analyzing the types and numbers of logical connectors in the input text; the basic weight coefficient is a predefined static weight associated with the knowledge type of the semantic unit, used to characterize the inherent information density baseline of different knowledge types.

[0072] The semantic unit network is constructed as follows:

[0073] The frequency of different core knowledge points appearing together in the same paragraph or adjacent sentences in the input text is statistically analyzed, and the sentence structure of the input text is analyzed in combination with dependency parsing or semantic role labeling to identify the logical relationship between different core knowledge points that appear at the same frequency.

[0074] Each core knowledge point is transformed into a high-dimensional vector, and the cosine similarity between the high-dimensional vectors is calculated. The core knowledge points corresponding to the high-dimensional vectors with cosine similarity greater than the preset similarity threshold are selected to filter core knowledge points that are not directly co-occurring in the text but are highly related in semantics. Based on the type attributes and universality of the corresponding core knowledge points, the logical relationship between the core knowledge points is inferred.

[0075] By treating all semantic units as nodes and the identified logical relationships as edges, a semantic unit network with a graph data structure is constructed.

[0076] Based on the semantic unit network analysis, the type and semantic density of each semantic unit are analyzed, and a semantic template is bound to each semantic unit based on a preset template library and binding rules to obtain a unit template mapping table; the template library stores different semantic templates, which are formatted fragments designed for different types of knowledge points in books and encapsulated with specific typesetting rules and style instructions; each semantic template corresponds to a template identifier, and the knowledge type range and semantic density range are set for each semantic template;

[0077] The specific binding rules are as follows:

[0078] Traverse each semantic unit in the semantic unit network, and use the knowledge type of the semantic unit as the first condition to filter out the semantic templates that contain the knowledge type of the corresponding semantic unit in all knowledge type ranges, and use them as candidate semantic templates for the corresponding semantic unit.

[0079] Determine the density level of all semantic units; classify semantic units with a semantic density greater than or equal to a preset first density threshold as high-density semantic units; classify semantic units with a semantic density less than the first density threshold but greater than or equal to a preset second density threshold as medium-density semantic units; classify semantic units with a semantic density less than the second density threshold as low-density semantic units.

[0080] Using the density level of semantic units as the second condition, candidate semantic templates that match the density level of the corresponding semantic units are selected from the candidate semantic templates. These templates are then used as the semantic templates for the corresponding semantic units. This ensures that high-density, complex concept-type semantic units are matched with templates that have the ability to be structured and decomposed, while low-density, simple knowledge-point-type semantic units are matched with templates that have a simple structure. This effectively avoids the problem of empty interface display or information overload caused by the mismatch between template structure and content complexity.

[0081] If a semantic unit corresponds to multiple semantic templates, the final semantic template is selected based on predefined priority rules. The priority rules include: counting the number of constraints of each semantic template and extracting the maximum value; taking the semantic template corresponding to the maximum value as the final semantic template corresponding to the semantic unit; if the maximum value corresponds to different semantic templates, comparing the user usage frequency and user modification number of the corresponding semantic templates, and selecting the semantic template with the highest user usage frequency or the fewest user modification number as the final semantic template corresponding to the semantic unit to improve the user experience.

[0082] The unit identifier of each semantic unit is associated with the corresponding semantic template to obtain the unit template mapping table.

[0083] The processing order is determined based on the logical relationships in the semantic unit network, and content generation and verification are performed on each semantic unit based on the processing order and the unit template mapping table to obtain the initial draft of the book text; for example Figure 2 As shown, the processing order is determined as follows:

[0084] The semantic unit network is traversed, and the logical relations corresponding to each semantic unit are extracted. The number of logical relations with a pre-dependency type, where the semantic unit acts as a successor unit in the pre-dependency, is counted, and this number is taken as the number of dependencies to be satisfied for the corresponding semantic unit. A pre-dependency refers to a strong order constraint between two semantic units, meaning that the correct understanding or generation of one semantic unit requires the correct generation and explanation of the other semantic unit. In a pair of semantic units with a pre-dependency logical relation, if the correct generation of semantic unit A requires the completion of semantic unit B, then semantic unit B is taken as the pre-unit, and semantic unit A as the successor unit. Here, A and B are two different semantic units, and the logical relation between A and B is a pre-dependency.

[0085] All semantic units in the semantic unit network with zero unmet dependencies and a state of pending generation are designated as ready units, thus obtaining a ready unit queue. These ready units do not need to wait for any prior knowledge and can begin processing immediately.

[0086] Traverse the ready unit queue and generate and validate content for any ready unit in the ready unit queue; after completing the content generation and validation, obtain all semantic units that use the ready unit as a predecessor unit to obtain the set of successor units corresponding to the ready unit; decrement the number of unmet dependencies of all successor units in the set of successor units by 1, and remove the ready unit from the ready unit queue; traverse the set of successor units corresponding to the ready unit, and if the number of unmet dependencies of any successor unit is 0 and the unit status of the successor unit is pending generation, add the corresponding successor unit to the ready unit queue;

[0087] When the ready queue is empty, all semantic units in the semantic unit network are traversed. If there are still semantic units in the state of "to be generated", it indicates that there may be circular dependencies in the semantic unit network. The state of the semantic units in the state of "to be generated" is updated to "requires manual review", and the semantic units are then generated and verified.

[0088] The specific methods for content generation and verification are as follows:

[0089] For any ready unit that performs content generation and verification, the text generation model is invoked to generate a detailed and coherent main text using the core content of the ready unit as the seed.

[0090] According to the unit template mapping table, obtain the semantic template bound to the ready unit; fill the text content into the predefined placeholders of the corresponding semantic template to form a complete, formatted content block;

[0091] The core content coverage and logical coherence of the content block are verified. If the core content coverage and logical coherence of the content block pass the verification, the corresponding ready unit is determined to have completed content generation and verification, and the unit status of the corresponding ready unit is updated to verified.

[0092] Specifically, it is determined whether the content block completely covers the key points and keywords described in the core content of the corresponding ready unit; if it completely covers them, the core content coverage of the content block is deemed to have passed the verification; it is also determined whether the content block is consistent with the logical relationship of the corresponding semantic unit in the ready unit network; if the logical relationship is consistent, the logical coherence of the content block is deemed to have passed the verification.

[0093] If the core content coverage or logical coherence check of the content block fails, the text generation model is called again for rewriting. A maximum number of rewrites is set. If the number of rewrites of the ready unit reaches the maximum number of rewrites and still fails the core content coverage or logical coherence check of the content block, the semantic status of the corresponding ready unit is updated to require manual review.

[0094] When there are no semantic units in the semantic unit network with a semantic state of "to be generated", the content blocks corresponding to all semantic units are assembled to obtain the first draft of the digital book; the specific assembly steps are as follows:

[0095] Using the logical relationships corresponding to all semantic units in the semantic unit network as the underlying framework, the core order and nesting relationships between each content block are determined.

[0096] Based on the core order and nesting relationship, the corresponding content blocks are filled into the container defined by the predefined book template; the book template provides title frames, page layouts, headers and footers, and other styles for chapters, sections, subsections, and other levels.

[0097] The table of contents of the book is automatically generated based on the chapter titles of the book template and the semantic units with title attributes in the semantic unit network.

[0098] It retrieves all references cited during the content generation process and automatically generates a unified list of references according to preset format specifications.

[0099] In this embodiment, for semantic units whose status requires manual review, a highlighted background or annotation box is inserted at the corresponding position in the initial draft of the book text, and the reason is noted, so as to accurately guide users to conduct subsequent manual review and revision.

[0100] Based on the image-text association structure, images are matched and inserted into the initial draft of the book text to obtain the digital draft of the book. Specifically, the image identifier, annotation text, corresponding text paragraph, and page number and page coordinate information of the image are extracted from the image-text association structure. All text paragraphs of the initial draft of the book text are traversed, and the corresponding text paragraphs in the image-text association structure are matched with the text paragraphs of the initial draft of the book text to locate the target paragraph where each image needs to be inserted.

[0101] For an image that matches the target paragraph, analyze the semantic relationship between the text content of the target paragraph and the annotation text corresponding to the image; calculate the semantic similarity between the text content of the target paragraph and the annotation text corresponding to the image using a pre-trained text semantic matching model; if the semantic similarity is greater than a preset similarity threshold, then determine that the corresponding image and the target paragraph have been successfully matched; according to the page coordinates of the successfully matched image, insert the image and the corresponding annotation text into the corresponding position of the target paragraph.

[0102] When inserting an image, the image format is adjusted according to the preset image layout rules, a uniform image number is assigned to the image, and the image number identifier in the annotation text is updated.

[0103] When an editing instruction is received from a user regarding the initial digital draft of the book, the semantic unit network is updated based on the editing instruction, and the content is regenerated and verified.

[0104] Specifically, the user's editing instructions are obtained; the editing instructions include the location information of the editing, the operation action, and the operation content, for example, a sentence is modified in the third paragraph of the first chapter; the operation action includes modification, deletion, and addition; the corresponding content block is located based on the location information of the editing, and the corresponding semantic unit is located according to the relationship between the content block and the semantic unit, and the semantic unit is recorded as the target unit;

[0105] If the operation is modification, the unit status of the target unit is updated to "user modified," and it is determined whether the core content of the target unit has been modified. If the core content has been modified, the semantic density of the target unit is recalculated. If the operation is deletion, the unit status of the target unit is updated to "invalid." If the operation is addition, and an independent knowledge point is added to the target unit, a new semantic unit is generated based on the added knowledge point. The new semantic unit is added to the semantic unit network, and the unit status of the corresponding semantic unit is set to "pending generation."

[0106] Traverse the semantic unit network to obtain all semantic units that use the target unit as a preceding unit, thus obtaining the affected units; update the semantic state of all target units and their corresponding affected units to the pending generation state; determine the processing order of all target units and their corresponding affected units, and regenerate and verify the content of all target units and their corresponding affected units based on the processing order and the unit template mapping table; reassemble the content blocks corresponding to the regenerated target units and their corresponding affected units to obtain the final digital book.

[0107] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0108] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of protection of this application, and these forms are all within the protection scope of this application.

Claims

1. A method for generating a book based on semantic understanding, comprising the following steps: receiving references and determining the number of references; normalizing the references, extracting pictures and annotations, and obtaining a picture-text association structure; receiving user input demand data, extracting core knowledge points from the demand data and the references, analyzing the logical relationships between the core knowledge points, and constructing a semantic unit network composed of different semantic units based on the logical relationships; analyzing the type and semantic density of each semantic unit based on the semantic unit network, binding a semantic template to each semantic unit based on a preset binding rule, and obtaining a unit template mapping table; determining a processing order based on the logical relationships in the semantic unit network, generating and checking the content of each semantic unit based on the processing order and the unit template mapping table, obtaining a book text draft, matching pictures into the book text draft based on the picture-text association structure, and obtaining a book digitalization draft; when receiving a user editing instruction for the book digitalization draft, updating the semantic unit network based on the editing instruction, and re-generating and checking the content to obtain a final digital book.

2. The method for generating a book in digital form based on semantic understanding according to claim 1, characterized in that, The method for determining the processing order specifically comprises: traversing the semantic unit network, extracting the logical relationship corresponding to each semantic unit, and counting the number of logical relationships with a relationship type of front dependency and the semantic unit as a successor unit, and taking the number as a to-be-satisfied dependency number of the corresponding semantic unit; taking all semantic units in the semantic unit network with a to-be-satisfied dependency number of 0 and a unit state of to-be-generated as ready units to obtain a ready unit queue; traversing the ready unit queue, and generating and checking the content of any ready unit in the ready unit queue.

3. The method of claim 2, wherein the semantic understanding-based book digitalization generation method is characterized by, The method for determining the processing order further comprises: after completing the content generation and checking, obtaining all semantic units taking the ready unit as a front unit to obtain a successor unit set corresponding to the ready unit; decreasing the to-be-satisfied dependency number of all successor units in the successor unit set by 1, and removing the ready unit from the ready unit queue; traversing the successor unit set corresponding to the ready unit, and if the to-be-satisfied dependency number of any successor unit is 0 and the unit state of the successor unit is to-be-generated, adding the corresponding successor unit to the ready unit queue.

4. The method of claim 3, wherein the semantic understanding-based book digitalization generation method is characterized by, The method for generating and checking the content specifically comprises: for any ready unit for which content generation and checking are performed, calling a text generation model to generate a detailed and coherent text content with the core content of the ready unit as a seed; obtaining a semantic template bound to the ready unit according to the unit template mapping table, and filling the text content into a placeholder predefined in the corresponding semantic template to form a complete and formatted content block. The core content coverage and logical coherence of the content block are checked, and if the core content coverage and logical coherence of the content block are both checked, it is determined that the corresponding ready unit completes content generation and checking, and the unit state of the corresponding ready unit is updated to checked.

5. The method of claim 4, wherein the semantic understanding-based book digitalization generation method is characterized by, The content generation and checking method further comprises: If the core content coverage or logical coherence of the content block is not passed, the text generation model is called again for rewriting; The maximum number of rewritings is set, and if the number of rewritings of the ready unit reaches the maximum number of rewritings and the core content coverage or logical coherence of the content block is still not passed, the semantic state of the corresponding ready unit is updated to human review required; When the ready queue is empty, all semantic units of the semantic unit network are traversed, and if there is still a semantic unit with a unit state of to be generated, the unit state of the corresponding semantic unit is updated to human review required, and the semantic unit is generated and checked.

6. The method of claim 5, wherein the semantic understanding-based book digitalization generation method is characterized by, The construction method of the semantic unit network specifically comprises: Each core knowledge point is instantiated as an independent semantic unit, and each semantic unit at least comprises core content, a unit state and semantic density; and the calculation method of the semantic density is as follows: Text length features of the core content, statistical features of keywords related to the core knowledge point in reference materials, complexity features of logical structures presented by the core content, and basic weight coefficients associated with the knowledge type are obtained respectively; Based on the basic weight coefficients, the text length features, statistical features and complexity features are summed to obtain the semantic density.

7. The method of claim 6, wherein the semantic understanding-based book digitalization generation method is characterized by, The construction method of the semantic unit network further comprises: The frequency of different core knowledge points appearing together in the same paragraph or adjacent sentences in the input text is counted, and the sentence structure of the input text is analyzed by combining dependency syntax analysis or semantic role labeling to identify the logical relationship of different core knowledge points appearing together at the same frequency; Each core knowledge point is converted into a high-dimensional vector, and the cosine similarity between high-dimensional vectors is calculated; Core knowledge points corresponding to high-dimensional vectors with a cosine similarity greater than a preset similarity threshold are selected, and the corresponding logical relationship is identified based on the type attribute and universality of the corresponding core knowledge point; All semantic units are taken as nodes, and the identified logical relationship is taken as an edge to construct a semantic unit network of a graph data structure.

8. The method of claim 7, wherein the semantic understanding-based book digitalization generation method is characterized by, The binding rule is specifically as follows: Each semantic unit in the semantic unit network is traversed, and all semantic templates whose knowledge type range contains the knowledge type of the corresponding semantic unit are selected as candidate semantic templates of the corresponding semantic unit, taking the knowledge type of the semantic unit as the first condition; The density levels of all semantic units are determined, and candidate semantic templates whose semantic density range matches the density level of the corresponding semantic unit are selected from the candidate semantic templates as the semantic templates of the corresponding semantic unit, taking the density level of the semantic unit as the second condition; If a semantic unit corresponds to multiple semantic templates, the final semantic template is selected based on a predefined priority rule.

9. The method of claim 8, wherein the semantic understanding-based book digitalization generation method is characterized by, The editing instruction comprises position information of where the editing occurs, operation action and operation content; the operation action comprises modification, deletion and addition; Based on the position information of where the editing occurs, the corresponding content block is located, and based on the relationship between the content block and the semantic unit, the corresponding semantic unit is located, and the semantic unit is recorded as a target unit; If the operation action is deletion, the unit state of the target unit is updated to invalid; if the operation action is addition, and an independent knowledge point is added in the target unit, a new semantic unit is generated based on the added knowledge point; The new semantic unit is added to the semantic unit network, and the unit state of the corresponding semantic unit is set to a to-be-generated state.

10. The method of claim 9, wherein the semantic understanding-based book digitalization generation method is characterized by, If the operation action is modification, the unit state of the target unit is updated to user modified, and it is judged whether the core content of the target unit is modified; if the core content is modified, the semantic density of the target unit is recalculated; All semantic units taking the target unit as a preceding unit are obtained by traversing the semantic unit network, and affected units are obtained; The semantic states of all target units and corresponding affected units are updated to a to-be-generated state; The processing order of all target units and corresponding affected units is determined, and all target units and corresponding affected units are re-generated and verified based on the processing order and the unit template mapping table; Based on the re-generated target units and corresponding affected units, the corresponding content blocks are re-assembled to obtain the final digital book.

Images