An electronic book recommendation method and system based on artificial intelligence

Abstract

The application relates to the technical field of data processing, and discloses an electronic book recommendation method and system based on artificial intelligence, which comprises the following steps: obtaining historical reading data of a target user, wherein the historical reading data comprises reading behavior data and text content of the read electronic books; extracting multi-dimensional features of the text, including information density, reasoning span, rhetoric intensity, theme distribution, knowledge field coverage and narrative mode features; training a user preference prediction model based on the text features and reading behavior labels; extracting the same features from the electronic books in a book library and inputting the prediction model to obtain interest scores; and screening target electronic books according to the interest scores and recommending the target electronic books to a user terminal. The application can significantly improve the accuracy and personalization level of the electronic book recommendation result by fusing multi-dimensional text features and user behavior data.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to an artificial intelligence-based e-book recommendation method and system. Background Technology

[0002] With the rapid development of the internet and the widespread use of e-reading devices, the variety and quantity of e-books are constantly increasing, covering various types such as novels, popular science, and academic works. Faced with a massive amount of e-book resources, how to provide content that meets the reading interests and needs of different users has become an urgent problem to be solved by e-reading platforms and content distribution systems.

[0003] Currently, common ebook recommendation methods mainly include collaborative filtering-based methods and content-based methods. Collaborative filtering-based methods rely on a large amount of historical user behavior data, making recommendations by analyzing the similarity of interests among different users. However, these methods perform poorly in cold-start scenarios and are easily affected by data sparsity issues. Content-based recommendation methods, on the other hand, analyze ebook metadata, tags, or summaries to match them with user interests and preferences, which can alleviate the cold-start problem to some extent. However, they often rely on shallow content features and struggle to accurately reflect differences in language expression, information presentation, and other aspects of ebooks.

[0004] In addition, some recommendation systems have attempted to introduce natural language processing technology to analyze the full text of e-books. However, due to limitations in the sophistication of the processing methods and the breadth of feature extraction, existing technologies often fail to fully reflect the content features of e-books at different levels, thus affecting the accuracy and personalization of recommendations.

[0005] Therefore, how to achieve more accurate and efficient recommendations by combining user reading behavior and the in-depth content characteristics of e-books from a vast amount of e-book resources remains an important technical problem that needs to be solved in the field of e-reading. Summary of the Invention

[0006] In view of this, the present invention proposes an e-book recommendation method and system based on artificial intelligence to solve the above problems.

[0007] On the one hand, the present invention proposes an artificial intelligence-based e-book recommendation method, comprising:

[0008] Obtain the target user's historical reading data, which includes the user's reading behavior data and the text content of the e-books already read;

[0009] Text features are extracted from the text content, including: information density, calculated as the ratio of the number of concept units after word segmentation and stop word filtering to the total number of words; reasoning span, calculated as the proportion of sentence pairs with implicit premises between adjacent sentences to the total number of sentence pairs, wherein the implicit premises are identified by a text implication judgment model; rhetorical intensity, calculated as the proportion of the frequency of rhetorical devices in the text content to the total number of sentences; topic distribution, obtained by inputting the text content into a topic model, outputting a topic probability vector, and taking the weighted average within a preset range; knowledge domain coverage, obtained by mapping entities and keywords in the text content to an external knowledge graph, counting the number of different domain tags and the proportion of the domain tags appearing in the whole book, and calculating based on the number and proportion; and narrative mode features, obtained by inputting the text into a plot structure classification model, identifying the plot structure type, and outputting the corresponding classification probability vector.

[0010] A user preference prediction model is trained based on the aforementioned text features and reading behavior labels;

[0011] Text features of ebooks in the library are extracted in the same way, input into the user preference prediction model, and interest scores are obtained.

[0012] Target ebooks are selected based on interest ratings and recommended to the target user's device.

[0013] Furthermore, when acquiring reading behavior data of target users, this includes:

[0014] Call the historical reading record interface from the reading application on the user's terminal to obtain a list of e-books that the user has read within a preset time range;

[0015] For each e-book that has been read, reading behavior data associated with the user is extracted. The reading behavior data includes reading time spent on each chapter, chapter completion rate, number of jumps, number of rereads, and records of bookmarking or annotation operations.

[0016] The corresponding full-text content of the e-book is obtained by accessing the e-book storage server or the user's local cache. After obtaining the full-text content, text decoding and format parsing are performed to remove the layout information and retain the plain text data for natural language processing.

[0017] The reading behavior data is associated with the corresponding text content and stored using eBook identifiers.

[0018] Furthermore, the acquisition of the information density includes:

[0019] The acquisition of information density includes: performing word segmentation on the text content and deleting stop words to obtain a sequence of word elements arranged in order; identifying word combinations that conform to preset structural rules as candidate units based on part-of-speech tagging results, wherein the structural rules stipulate that the first word element in the candidate unit is a noun or a noun derived from a verb, and the subsequent word elements are modifiers or nouns, and the length of the entire combination is not less than two word elements; merging and deduplicating the candidate units selected by the structural rules to form a set of concept units; retrieving all occurrence positions of each concept unit in the text content, recording the sentence position index of the first occurrence, and counting the total number of occurrences; dividing the number of occurrences of each concept unit by the total number of words in the text content to obtain the density value of the concept unit; summing the density values ​​of all concept units in each chapter to obtain the chapter information density value; and weighting the information density values ​​of all chapters according to the number of words in the chapter to obtain the information density of the text content.

[0020] Furthermore, obtaining the reasoning span includes:

[0021] The text content is segmented into sentences to obtain a sequentially arranged sentence sequence; a set of sentence pairs is constructed according to adjacency relationships, such that each sentence pair consists of the current sentence and its subsequent sentence; each sentence pair is input into the text entailment determination model, and the model outputs a determination result on whether the sentence pair contains an implicit premise; sentence pairs with implicit premises are marked as inference-related sentence pairs, and the number of inference-related sentence pairs is counted; the number of inference-related sentence pairs is divided by the total number of sentence pairs to obtain the inference span value, wherein the implicit premise is used to characterize the degree of semantic dependence between adjacent sentences and the density of cross-sentence inference.

[0022] Furthermore, obtaining the rhetorical intensity includes:

[0023] The text content is segmented into sentences to obtain a set of sentences, and each sentence is assigned its position index in the full text. Component parsing is performed on each sentence, taking sentences with word segmentation and part-of-speech tagging as input, and outputting tree data representing the hierarchical structure of phrase components. Based on the tree data, structural deviation is calculated; the structural deviation is the normalized value of the difference between the maximum depth of the phrase hierarchy in that sentence and the average depth of the phrase hierarchy in the entire text. Dependency parsing is performed on each sentence, taking sentences with word segmentation and part-of-speech tagging as input, and outputting a directed graph structure representing the dependency relationships between lexical units. Based on the dependency relationships, relation complexity is calculated; the relation complexity is the ratio of the number of dependency edges for non-core modifying, parallel, insertional, and comparative relations in that sentence to the number of lexical units in that sentence. Lexical unit merging is performed on each sentence. The process involves mapping lexical units with different morphological changes to a unified word form; calculating the repetition reinforcement ratio based on the merged lexical units, where the repetition reinforcement ratio is the ratio of the number of repeated lexical units within a preset interval to the total number of lexical units in the sentence; combining the structural deviation, relational complexity, and repetition reinforcement ratio with preset weights to obtain the rhetorical score of the sentence, and marking sentences with rhetorical scores greater than a preset threshold as high-rhetorical sentences; calculating the proportion of high-rhetorical sentences in the entire text as the rhetorical frequency component; calculating the distribution dispersion index of high-rhetorical sentences based on the position index and taking its normalized value as the position distribution component; mapping high-rhetorical sentences to the plot units determined by the plot structure classification model, calculating the proportion of high-rhetorical sentences in each plot unit to the total number of sentences in that unit, and taking its mean as the narrative coupling component; and summing the rhetorical frequency component, position distribution component, and narrative coupling component with preset weights to obtain the rhetorical intensity.

[0024] Furthermore, obtaining the topic distribution includes:

[0025] The text content is segmented into paragraphs to obtain a sequentially arranged set of paragraphs. Each paragraph is segmented and stop words are removed to obtain a paragraph word sequence. This paragraph word sequence is converted into a bag-of-words vector representation, which is a frequency statistics vector based on a preset vocabulary. The bag-of-words vector is input into a topic model, which is a probabilistic generation model based on latent Dirichlet assignment and obtained through parameter estimation using text training corpora containing labeled topic categories. The probability distribution vector of each paragraph under each preset topic is obtained from the output of the topic model. For each paragraph, the topic components with probability values ​​within a preset range are summed by multiplying their probability values ​​by the corresponding topic importance weights to obtain the topic concentration value of that paragraph. The topic concentration values ​​of all paragraphs are weighted and averaged according to the number of characters in each paragraph to obtain the topic distribution value of the text content. This topic distribution value represents the degree of concentration of the text content in the topic dimension and its correlation distribution characteristics with multiple topics.

[0026] Furthermore, the acquisition of the knowledge domain coverage includes:

[0027] The text content is segmented and stop words are removed to obtain a word sequence. Keyword extraction is performed on the word sequence, including sorting the words based on statistical results of word frequency and inverse document frequency combined with word position distribution information, and selecting the top-ranked words as keywords. Named entity recognition is performed on the word sequence, using a recognition model trained on training corpora containing entity category annotations to obtain entity words representing people, places, institutions, works, and proper nouns. The keywords and entity words are mapped to a node set of an external knowledge graph after string matching and synonym merging, obtaining the domain label corresponding to each node. The number of different domain labels is counted, and the proportion of the occurrence of nodes corresponding to each domain label in the full text to the total number of words in the full text is calculated. The number of domain labels and the proportion of domain labels are combined according to a preset calculation rule to obtain the knowledge domain coverage.

[0028] Furthermore, the acquisition of the narrative mode features includes:

[0029] The text content is segmented into plot units, which are achieved by identifying paragraph boundaries that indicate the progression of time, scene transitions, or changes in character relationships. Narrative structure elements are extracted for each plot unit, including the order of character appearances, event triggering conditions, conflict and turning point locations, and ending type. The extracted narrative structure elements are sequentially encoded into a structural feature sequence and input into a plot structure classification model. The plot structure classification model is trained on training corpus containing labeled plot structure categories and outputs the classification probability vector of the plot unit under each preset plot structure category. The classification probability vectors of all plot units are weighted and averaged according to the number of words in the plot unit to obtain the narrative pattern features of the text content.

[0030] Furthermore, the training of the user preference prediction model includes:

[0031] The target user's reading behavior tags are paired with the text features of the e-books they have read using e-book identifiers to form a training sample set of users, text features, and behavior tags. Feature normalization and missing value imputation are performed on the training sample set, which is then divided into a training set and a validation set. The training set is input into a prediction model with a multi-layer nonlinear mapping structure for iterative training. The prediction model takes text features and user behavior tags as input and outputs predicted user interest ratings for candidate e-books. A loss value is calculated based on the difference between the prediction results on the validation set and the actual user behavior, and this loss value is used to update the prediction model parameters until a preset convergence condition is met on the validation set, thus obtaining the user preference prediction model.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0033] This invention introduces multiple textual features into the e-book recommendation process, including information density, reasoning span, rhetorical intensity, topic distribution, knowledge domain coverage, and narrative pattern characteristics. This allows for a comprehensive characterization of e-book content across multiple dimensions, such as information capacity, logical reasoning span, language expression, topic concentration, knowledge domain scope, and narrative structure. Compared to existing methods that rely solely on keywords or topic tags, this approach better reflects the deeper semantic and stylistic features of e-books. By combining these multi-dimensional textual features with user reading behavior data and training a user preference prediction model, the method can simultaneously capture users' comprehensive preferences in content type, language expression, and knowledge domain, thereby significantly improving the accuracy of interest prediction. This method is applicable to e-book recommendations across different categories and genres, requires no manual rules, and possesses strong generalization and adaptability, facilitating its application in multi-platform and multilingual recommendation scenarios. Recommendations based on interest scores not only meet users' topical needs but also closely align with their reading preferences in terms of language style and narrative pattern, resulting in more accurate and personalized recommendations. Meanwhile, since the model's input features come from explicitly calculable text feature values ​​and user behavior data, the recommendation results are traceable to a certain extent, which helps improve the system's interpretability and user trust.

[0034] On the other hand, the present invention proposes an artificial intelligence-based e-book recommendation system, comprising:

[0035] The data acquisition module is configured to acquire the target user's historical reading data, which includes the user's reading behavior data and the text content of the e-books already read;

[0036] The feature extraction module is configured to extract text features from the text content. These text features include: information density, calculated as the ratio of the number of concept units (after word segmentation and stop word filtering) to the total number of words; reasoning span, calculated as the proportion of sentence pairs with implicit premises between adjacent sentences to the total number of sentence pairs, where implicit premises are identified using a text implication judgment model; rhetorical intensity, calculated as the proportion of the frequency of rhetorical devices in the text to the total number of sentences; topic distribution, obtained by inputting the text content into a topic model, outputting a topic probability vector, and taking a weighted average within a preset range; knowledge domain coverage, obtained by mapping entities and keywords in the text content to an external knowledge graph, statistically analyzing the number of different domain tags and the proportion of those tags appearing throughout the book, and calculating based on the number and proportion; and narrative pattern features, obtained by inputting the text into a plot structure classification model, identifying the plot structure type, and outputting the corresponding classification probability vector.

[0037] The rating module is configured to train a user preference prediction model based on the text features and reading behavior labels; and to extract the text features of e-books in the library in the same way, input them into the user preference prediction model, and obtain interest ratings.

[0038] The recommendation module is configured to filter target ebooks based on interest ratings and recommend them to the target user's terminal.

[0039] It should be noted that the AI-based e-book recommendation method of this invention has the same beneficial effects as its system, and will not be described in detail here. Attached Figure Description

[0040] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0041] Figure 1 A flowchart illustrating an AI-based e-book recommendation method provided in this embodiment of the invention.

[0042] Figure 2 This is a functional block diagram of an AI-based e-book recommendation system provided in an embodiment of the present invention. Detailed Implementation

[0043] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0044] See Figure 1 As shown, this embodiment of the invention provides an artificial intelligence-based e-book recommendation method, including:

[0045] S1: Obtain the target user's historical reading data, which includes the user's reading behavior data and the text content of the e-books already read;

[0046] S2: Extract text features from the text content. These features include: Information density, calculated as the ratio of the number of concept units after word segmentation and stop word filtering to the total number of words; Reasoning span, calculated as the proportion of sentence pairs with implicit premises between adjacent sentences to the total number of sentence pairs (implicit premises are identified using a text implication judgment model); Rhetorical intensity, calculated as the proportion of the frequency of rhetorical devices in the text to the total number of sentences; Topic distribution, obtained by inputting the text content into a topic model, outputting a topic probability vector, and taking the weighted average within a preset range; Knowledge domain coverage, obtained by mapping entities and keywords in the text content to an external knowledge graph, counting the number of different domain tags and the proportion of those tags appearing throughout the book, and calculating based on the number and proportion; and Narrative pattern features, obtained by inputting the text into a plot structure classification model, identifying the plot structure type, and outputting the corresponding classification probability vector.

[0047] S3: Train a user preference prediction model based on text features and reading behavior labels; extract text features from ebooks in the library in the same way, input them into the user preference prediction model, and obtain interest scores;

[0048] S4: Filter target e-books based on interest ratings and recommend them to target user terminals.

[0049] Understandably, this involves acquiring the target user's historical reading data, which includes the user's reading behavior data and the text content of read ebooks. User reading behavior data can be obtained through the user's terminal reading application by calling the history interface, and is limited to a preset time range. This reading behavior data includes reading time spent on each chapter, chapter completion rate, number of chapter jumps, number of rereads, and records of bookmarking or annotation operations. The text content of read ebooks is obtained by accessing the ebook storage server or the user's local cache files. After acquisition, text decoding and layout parsing are performed to remove layout information such as paragraph indentation, page breaks, and font tags, retaining only the plain text data used for natural language processing.

[0050] Text features are extracted from the text content, including the following items: Information density, obtained by segmenting the text and removing stop words to obtain a word sequence; identifying word combinations that meet the conditions based on preset structural rules to form a set of concept units, and calculating the ratio of the occurrence frequency of each concept unit in the set to the total number of words as the information density value; Inference span, formed by segmenting the text into sentence sequences and constructing a set of adjacent sentence pairs, inputting the sentence pairs into a pre-trained text entailment judgment model to determine whether there are implicit premises, and calculating the proportion of sentence pairs with implicit premises to the total number of sentence pairs; Rhetorical intensity, obtained by statistically analyzing the use of rhetoric in the text. The proportion of sentences containing rhetorical expressions to the total number of sentences can be identified through a combination of constituent syntax analysis and dependency syntax analysis. Topic distribution is calculated by inputting the text into a pre-trained topic model, outputting a topic probability vector, and then taking a weighted average within a preset probability range as the topic distribution value. Knowledge domain coverage involves mapping keywords and entities in the text to an external knowledge graph, counting the number of tags from different domains and the proportion of each tag appearing in the entire text, and combining the number and proportion to calculate the topic distribution. Narrative pattern features are calculated by segmenting the text into plot units and inputting them into a plot structure classification model, outputting a classification probability vector for the corresponding plot structure.

[0051] The user preference prediction model is trained based on text features and reading behavior labels. During the training process, the text features of the e-books that the user has read are paired with their corresponding reading behavior labels to form a training sample set, which is then input into the prediction model for iterative training. The prediction performance is evaluated on the validation set to update the model parameters.

[0052] Textual features of ebooks in the library are extracted in the same way, input into a user preference prediction model to obtain interest scores; target ebooks are selected based on interest scores, and the recommendation results are pushed to the target user's terminal device.

[0053] In some embodiments of this application, when acquiring reading behavior data of a target user, the following is included:

[0054] Call the historical reading record interface from the reading application on the user's terminal to obtain a list of e-books that the user has read within a preset time range;

[0055] For each ebook that has been read, reading behavior data associated with the user is extracted. This reading behavior data includes reading time spent on each chapter, chapter completion rate, number of jumps, number of rereads, and records of bookmarking or annotation operations.

[0056] The full text content of the corresponding e-book is obtained by accessing the e-book storage server or the user's local cache. After obtaining the full text content, text decoding and format parsing are performed to remove layout information and retain the plain text data for natural language processing.

[0057] Reading behavior data is associated with and stored with the corresponding text content using eBook identifiers.

[0058] It's important to note that when acquiring reading behavior data from a target user, the server first sends a data request to the user's terminal, invoking the historical reading record interface of the user's reading application. This interface can be a local API provided by the reading application or a cloud-based interface provided by the reading platform. This interface returns historical reading record data associated with the target user's account. To ensure data freshness and relevance, a time parameter can be passed to the interface call to limit the time range for data acquisition, such as the last 30 days, 90 days, or other time intervals configured by the platform. The historical reading records returned by the interface include a list of ebooks that the user has completed or is currently reading within that time range. Each ebook record includes at least an ebook identifier, reading start and end times, and reading progress information.

[0059] For each ebook that has been read, reading behavior data extraction and processing are performed. This data is obtained by calling the reading application's behavior log interface or parsing locally stored reading session log files. This data records the user's chapter entry and exit timestamps, page jump events, annotation and favorite operations during the reading process. The server categorizes each behavior record by chapter and calculates the reading dwell time based on the entry and exit timestamps; it calculates the chapter completion rate based on the proportion of read pages within a chapter to the total number of pages in the chapter; it counts the number of jumps and rereads within a chapter, where jumps are determined by detecting events of non-continuous changes in page number, and rereads are determined by identifying repeated entry events into already read chapters; simultaneously, it extracts records of the user's favorite and annotation operations within a chapter, and appends the occurrence time and location index to each record.

[0060] After extracting reading behavior data, it is also necessary to obtain the full-text content of the corresponding e-book. This can be achieved in two ways: first, by accessing the e-book storage server and requesting a plain text version or a standardized file version (such as EPUB, TXT, etc.) using the e-book identifier; second, by searching for the e-book cache file on the user's local terminal and uploading it to the server. Regardless of whether it comes from the server or the local cache, the obtained e-book file must undergo text decoding and format parsing before entering the analysis process. The decoding process converts the file according to its character encoding (such as UTF-8, GBK, etc.) to ensure consistency in subsequent processing; format parsing, for formats such as EPUB, HTML, and PDF, removes layout information, font tags, page breaks, image placeholders, etc., retaining only plain text data suitable for natural language processing, while preserving chapter titles, paragraph boundaries, and necessary structural markers to support subsequent feature calculations by chapter or paragraph.

[0061] Finally, the extracted reading behavior data is associated and stored with the corresponding plain text content of the ebook. The association is uniquely determined by the ebook identifier. A relational database or key-value database can be used for storage, with the ebook identifier as an index, recording the ebook's reading behavior data and text content in the same data structure. This storage structure facilitates rapid retrieval of the entire content of the target ebook and the corresponding user reading behavior during subsequent text feature extraction and model training, thereby improving processing efficiency.

[0062] In some embodiments of this application, obtaining information density includes:

[0063] Information density acquisition includes: performing word segmentation on the text content and deleting stop words to obtain a sequence of word elements arranged in order; identifying word combinations that conform to preset structural rules as candidate units based on part-of-speech tagging results, where the structural rules stipulate that the first word element in a candidate unit must be a noun or a noun derived from a verb, and subsequent word elements must be modifiers or nouns, with the entire combination length not less than two word elements; merging and deduplicating the candidate units selected by the structural rules to form a set of concept units; retrieving all occurrence positions of each concept unit in the text content, recording the sentence position index of the first occurrence, and counting the total number of occurrences; dividing the number of occurrences of each concept unit by the total number of words in the text content to obtain the density value of that concept unit; summing the density values ​​of all concept units in each chapter to obtain the chapter information density value; and weighting the information density values ​​of all chapters according to the number of words in each chapter to obtain the information density of the text content.

[0064] It should be noted that, firstly, word segmentation is performed on the text content. This segmentation can be achieved using a word segmentation tool that combines statistical probability with dictionary matching. During the initialization phase, this tool loads a dictionary corresponding to the text's language and determines the word segmentation positions based on a contextual probability model, thus segmenting the continuous character sequence into a meaningful sequence of word units. After word segmentation, stop word removal is performed on the word unit sequence. The stop word set can be determined by a pre-constructed high-frequency function word list, which contains functional words that contribute little to semantics, such as conjunctions, auxiliary words, and modal particles. After stop word filtering, a valid sequence of word units arranged in the original text order is obtained.

[0065] Next, part-of-speech tagging is performed on the word sequence. The tagging model can be trained based on a conditional random field or a bidirectional recurrent neural network. Based on the tagging results, word combinations that conform to preset structural rules are identified as candidate units. The structural rules stipulate that the first word in a candidate unit must be a noun or a noun derived from a verb, and subsequent words must be modifiers (such as adjectives or quantifiers) or nouns, with the entire combination having a length of at least two words. These rules aim to ensure that candidate units represent relatively independent and complete semantic concepts.

[0066] Subsequently, all candidate units that meet the rules are merged and deduplicated to obtain a set of concept units. During the merging process, candidate units with the same word sequence are treated as the same concept unit and merged to eliminate interference from duplicate counting.

[0067] After obtaining the set of concept units, each concept unit is searched throughout the entire text to determine all its occurrence positions, and the sentence number corresponding to the first occurrence position is recorded so that weighted analysis can be performed later based on the first occurrence position. At the same time, the total number of occurrences of the concept unit in the entire text is counted.

[0068] Then, the frequency of each concept unit is divided by the total number of words in the entire text to obtain the density value of that concept unit. The density values ​​of all concept units within a chapter are summed to obtain the chapter information density value. During this process, chapter boundaries can be determined based on the original structural markers of the ebook or the title positions extracted during the text parsing stage.

[0069] Finally, the information density values ​​of all chapters are weighted and averaged according to the word count of each chapter to obtain the overall information density value of the entire ebook. This information density value can characterize the richness of conceptual information carried by the ebook per unit text length and can be used in subsequent user preference modeling and ebook recommendation decisions.

[0070] In some embodiments of this application, obtaining the inference span includes:

[0071] The text content is segmented into sentences to obtain a sequentially arranged sentence sequence; a set of sentence pairs is constructed according to adjacency relationships, such that each sentence pair consists of the current sentence and its subsequent sentence; each sentence pair is input into the text entailment judgment model, and the output is the judgment result of whether the sentence pair has an implicit premise; sentence pairs with the judgment result of having an implicit premise are marked as inference-related sentence pairs, and the number of inference-related sentence pairs is counted; the number of inference-related sentence pairs is divided by the total number of sentence pairs to obtain the inference span value, where the implicit premise is used to characterize the degree of semantic dependence between adjacent sentences and the density of cross-sentence inference.

[0072] It should be noted that obtaining the reasoning span includes:

[0073] First, the text content is segmented into sentences, resulting in a sequentially arranged sentence sequence. Sentence segmentation can be achieved based on punctuation recognition, prioritizing the identification of sentence-ending punctuation such as periods, question marks, and exclamation marks, and combining this with word segmentation and part-of-speech tagging results to avoid misinterpreting punctuation marks in abbreviations and proper nouns as sentence boundaries. After segmentation, the original sentence order information is preserved for subsequent construction of adjacent sentence pairs.

[0074] Next, a set of sentence pairs is constructed based on adjacency relationships, such that each sentence pair consists of the current sentence and its immediate next sentence. The set of sentence pairs can be represented as an ordered sequence of binary tuples, where the first element is the word sequence of the current sentence and the second element is the word sequence of the next sentence.

[0075] Then, each sentence is input into the text entailment judgment model. The text entailment judgment model can be trained on training corpora containing entailment relation annotations. The model input consists of vectorized representations of two sentences, and the output is a binary classification result indicating the existence of implicit premises. The vectorized representations are obtained by mapping each sentence to a pre-trained word vector space after word segmentation and part-of-speech tagging, and then averaging or sequence encoding the word vectors. Implicit premises here represent background information that exists between the current sentence and subsequent sentences but is not explicitly stated in the text but is semantically necessary.

[0076] When the model outputs a judgment result indicating the existence of an implicit premise, the sentence pair is marked as a reasoning-related sentence pair, and its count is added to the statistical data structure. After processing all sentence pairs, the total number of reasoning-related sentence pairs is counted, and this number is divided by the total number of sentence pairs in the set to obtain the reasoning span value.

[0077] The reasoning span value reflects the degree of semantic dependence between adjacent sentences and the density of cross-sentence reasoning; a higher value indicates a tighter logical reasoning chain in the text. In subsequent user interest modeling, the reasoning span value can serve as an important feature input to the prediction model, reflecting the logical reasoning load, thereby improving the accuracy of recommendation results for reasoning and argumentative ebooks.

[0078] In some embodiments of this application, the textual entailment determination model is used to determine whether there is an implicit premise between adjacent sentence pairs. The textual entailment determination model is a type of binary classification model based on deep neural networks. It is obtained through supervised training on training corpora containing sentence pairs and their entailment relation labels. The input is two sentences arranged in sequence, and the output is a probability score indicating whether an implicit premise exists or not. Here, an "implicit premise" refers to a semantic condition that is not explicitly stated in the current sentence pair but is necessary for the validity of subsequent sentences.

[0079] The construction and annotation of training data can be based on the characteristics of long e-book texts: First, sentence sequences are extracted from a large-scale e-book corpus to generate a candidate sentence pair set. The candidate types include at least adjacent sentence pairs and non-adjacent sentence pairs within the same chapter. Then, a pre-screening rule is used to narrow down the annotation scope. The pre-screening rule includes sentence pairs based on referential and anaphoric trigger words, sentence pairs containing causal or conditional conjunctions, and sentence pairs with omissions. On this basis, manual annotation is organized. Annotators assign the label "exists an implicit premise" or "does not exist an implicit premise" to each sentence pair according to a unified guideline. Two independent annotators are used, and a consistency threshold is used as the pass condition. For inconsistent samples, a third party arbitrates to determine the label. To enhance the model's discriminative ability, hard-negative samples (sentence pairs with similar contexts but not dependent on implicit premises) and random sentence pairs across chapters can be constructed as easy-negative samples, so that positive and negative samples can form an effective comparison in terms of contextual similarity and semantic dependence.

[0080] Before model training, the training corpus undergoes the same preprocessing procedures as the inference stage, including sentence segmentation, word segmentation, and part-of-speech tagging, mapping sentences to fixed-length vector sequences. Sentence vector representations are generated by the embedding layer, which maps word units or sub-words to low-dimensional dense vectors using a lookup table. The model structure includes an encoding layer, an inter-sentence interaction layer, and a discriminative output layer: the encoding layer models the context of each sentence's vector sequence to obtain its respective sentence representation; the inter-sentence interaction layer aligns the representations of the two sentences to generate a joint representation that characterizes cross-sentence semantic dependencies; the discriminative output layer receives the joint representation and provides a binary classification probability score. During training, a binary classification loss function is used as the optimization objective, iteratively updating the model parameters until the preset convergence conditions and performance thresholds are reached on the validation set. To adapt to differences in e-book genres, training and validation sets can be constructed by stratified sampling according to genre, and genre stratification indicators are recorded during the validation phase for subsequent threshold setting and intra-domain fine-tuning.

[0081] In the inference phase, the target ebook is first segmented into sequential sentence columns, and sentence pairs are constructed based on adjacency relationships. Each sentence pair is input into the text entailment judgment model to obtain a probability score p indicating the existence of an implicit premise. Each sentence pair is then judged based on a decision threshold τ determined on the validation set. Pairs with p ≥ τ are considered inference-related. The threshold τ is selected by combining precision and recall performance on the validation set to obtain a stable decision boundary. The proportion of inference-related sentence pairs to the total number of sentence pairs is calculated to obtain the inference span value. To reduce bias caused by length and punctuation anomalies, extremely long sentences can be encoded in blocks and represented using max-pooling without destroying semantic units. Sentences containing abnormal punctuation are corrected according to rules during the segmentation phase. These processes are maintained consistently in both the training and inference phases to ensure uniformity.

[0082] To align with the overall workflow of this application, the textual implication determination model shares the same word segmentation, part-of-speech tagging, and sentence segmentation components as the extraction of information density, rhetorical intensity, topic distribution, knowledge domain coverage, and narrative pattern features, avoiding distribution drift caused by inconsistent preprocessing. The model output is only used as a basis for determining the inference span in the proportional statistics, without directly altering the calculation methods of other features. During the training of the user preference prediction model, the inference span is input as an independent numerical feature along with other text features and reading behavior labels to enhance the ability to characterize the "logical reasoning load" dimension of preferences. It is understandable that, without changing the input as sentence pairs, the output as binary classification probabilities, and the definition of inference span based on the sentence pair ratio, adjustments to the embedding dimension, encoding layer depth, interaction layer form, and threshold selection strategy are all alternative implementations of this application.

[0083] In some embodiments of this application, obtaining the rhetorical intensity includes:

[0084] The text content is segmented into sentences to obtain a set of sentences, and each sentence is assigned its position index in the full text. For each sentence, component parsing is performed, taking the segmented and part-of-speech-tagged sentences as input and outputting a tree-like structure representing the hierarchical structure of phrase components. Based on the tree-like structure, the structural deviation is calculated as the normalized value of the difference between the maximum depth of the phrase hierarchy in that sentence and the average depth of the phrase hierarchy in the entire text. For each sentence, dependency parsing is performed, taking the segmented and part-of-speech-tagged sentences as input and outputting a directed graph structure representing the dependency relationships between lexical units. Based on the dependency relationships, the relation complexity is calculated as the ratio of the number of dependency edges for non-core modifying, parallel, insertional, and comparative relations to the number of lexical units in that sentence. Finally, lexical merging is performed on each sentence. The process involves mapping lexical units with different morphological changes to a unified word form; calculating the repetition reinforcement ratio based on the merged lexical units, which is the ratio of the number of repeated lexical units within a preset interval to the total number of lexical units in the sentence; combining the structural deviation, relational complexity, and repetition reinforcement ratio with preset weights to obtain the rhetorical score of the sentence, and marking sentences with rhetorical scores greater than a preset threshold as high-rhetorical sentences; calculating the proportion of high-rhetorical sentences in the entire text as the rhetorical frequency component; calculating the distribution dispersion index of high-rhetorical sentences based on the position index and taking its normalized value as the positional distribution component; mapping high-rhetorical sentences to the plot units determined by the plot structure classification model, calculating the proportion of high-rhetorical sentences in each plot unit to the total number of sentences in that unit, and taking its mean as the narrative coupling component; and summing the rhetorical frequency component, positional distribution component, and narrative coupling component with preset weights to obtain the rhetorical intensity.

[0085] It should be noted that, firstly, the text content is segmented into sentences, resulting in a set of sentences arranged in sequence. Sentence segmentation can be achieved by combining punctuation rules with word segmentation results, prioritizing the identification of sentence-ending punctuation (such as periods, question marks, and exclamation marks) as segmentation boundaries, and correcting for cases that should not be segmented, such as abbreviations and time expressions. After sentence segmentation, each sentence is assigned its position index in the full text. The position index can be represented by the sentence sequence number in the full text or an index value based on character offset, to support subsequent distribution statistics.

[0086] Next, constituent syntactic analysis is performed on each sentence. Constituent syntactic analysis takes sentences that have already undergone word segmentation and part-of-speech tagging as input, and uses a pre-trained constituent syntactic analysis model to output the phrase constituent hierarchical structure of the sentence, represented as a tree data structure. Each node in the tree structure represents a phrase constituent, and its child nodes represent its direct components. Based on the tree structure, the structural deviation is calculated. Specifically, the maximum depth of the phrase hierarchical structure of the sentence is calculated, and the difference is calculated with the average phrase hierarchical structure depth of all sentences in the entire text. The difference is then normalized to a preset range to obtain the structural deviation value of the sentence.

[0087] Subsequently, dependency parsing is performed on each sentence. Dependency parsing also takes sentences with completed word segmentation and part-of-speech tagging as input, and uses a pre-trained dependency parsing model to output a dependency graph of the words in the sentence. The nodes of the dependency graph are words, and the edges represent the dependency type and direction between words. Based on this dependency graph, the relation complexity is calculated. Specifically, the number of dependency edges belonging to non-core modifying relations, parallel relations, insertion relations, and comparison relations in the sentence is counted, and this number is divided by the total number of words in the sentence to obtain the relation complexity value.

[0088] Next, lemmatization is performed on each sentence. Lemmatization is implemented through the lemmatization module, which maps the lexical units of verbs, nouns, adjectives, etc., under different morphological changes (tense, number, gender, etc.) to their basic lexical forms in order to reduce the statistical bias caused by lemmatization.

[0089] After word form merging is completed, the repetition reinforcement ratio is calculated. The repetition reinforcement ratio is the ratio of the number of repeated words within a preset interval (e.g., a window size of n words) to the total number of words in the sentence. It is used to reflect the degree to which word repetition in a sentence reinforces the meaning or emotion.

[0090] Next, the structural deviation, relational complexity, and repetition reinforcement ratio are weighted and summed according to preset weights to obtain the rhetorical score of the sentence. The preset weights can be obtained by performing feature importance analysis on a manually labeled high-rhetorical-sounding sample set. When the rhetorical score is greater than a preset threshold, the sentence is marked as a high-rhetorical-sounding sentence.

[0091] After tagging, the proportion of highly rhetorical sentences in the entire text is calculated to obtain the rhetorical frequency component. Based on the position index, the distribution dispersion index of highly rhetorical sentences in the entire text is calculated and normalized to the range of [0,1] to obtain the position distribution component.

[0092] Furthermore, the highly rhetorical sentences are mapped to the plot units determined by the plot structure classification model. The proportion of highly rhetorical sentences in each plot unit to the total number of sentences in that unit is counted, and the average of this proportion for all plot units is taken to obtain the narrative coupling component.

[0093] Finally, the rhetorical frequency component, positional distribution component, and narrative coupling component are weighted and summed according to preset weights to obtain the rhetorical intensity value of the text content. This rhetorical intensity value can quantify the complexity and rhetorical density of the text's linguistic expression, providing an important input reflecting the characteristics of language style for the user preference prediction model. The weighting can be based on regression fitting of historical user feedback data: collect the rhetorical feature values ​​of existing e-books and their corresponding user reading preference ratings (such as average reading time, collection rate, rating, etc.). Using this user feedback as the target variable, perform multiple regression analysis on structural deviation, relational complexity, and repetition reinforcement ratio, and the fitted regression coefficients are normalized and used as preset weights.

[0094] In some embodiments of this application, obtaining the topic distribution includes:

[0095] The text content is segmented into paragraphs to obtain a sequentially arranged set of paragraphs. Each paragraph is then segmented and stop words are removed to obtain a paragraph word sequence. This sequence is converted into a bag-of-words vector representation, which is a frequency vector based on a predefined vocabulary. The bag-of-words vector is input into a topic model, which is a probabilistic generation model based on latent Dirichlet assignment, and its parameters are estimated using training corpora containing labeled topic categories. The probability distribution vector of each paragraph under each predefined topic is obtained from the output of the topic model. For each paragraph, the topic components whose probability values ​​fall within a predefined range are summed by multiplying their probability values ​​by the corresponding topic importance weights to obtain the topic concentration value for that paragraph. The topic concentration values ​​of all paragraphs are weighted and averaged according to the number of characters in each paragraph to obtain the topic distribution value of the text content. This topic distribution value represents the degree of concentration of the text content in the topic dimension and its correlation distribution characteristics with multiple topics.

[0096] It should be noted that, firstly, the text content is divided into paragraphs, resulting in a set of paragraphs arranged in order. During the division, paragraph separators (such as line breaks and indentation marks) are used as boundaries, preserving the original paragraph order information and character position indices for subsequent weighted calculations.

[0097] Secondly, word segmentation is performed on each paragraph, and stop words are removed to obtain the word sequence of the paragraph. Word segmentation can be achieved by a word segmentation model trained on a domain e-book corpus, and the stop word list can be manually compiled based on high-frequency function words, auxiliary words, and functional words and validated on the training set.

[0098] Then, the paragraph word sequence is converted into a bag-of-words vector representation. The bag-of-words vector representation is constructed based on a pre-defined vocabulary, which is derived from a set of high-frequency and informative words in a large-scale e-book training corpus containing multiple topic categories. Each dimension of the vector corresponds to the frequency of a word's occurrence in the paragraph.

[0099] Next, the bag-of-words vectors are input into the topic model. The topic model is a probabilistic generative model based on Latent Dirichlet Allocation (LDA), obtained by parameter estimation on training corpora containing labeled topic categories. During training, the corpus is first segmented, stop words and low-frequency words are removed, and then the Gibbs sampling method is used to iteratively estimate the topic-word distribution and document-topic distribution until the convergence condition is met.

[0100] From the output of the topic model, obtain the probability distribution vector of each paragraph under each preset topic. For example, the probability values ​​of a paragraph under the topics of "history", "literature", "science", and "art" are 0.45, 0.25, 0.15, and 0.15, respectively.

[0101] For each paragraph's probability distribution vector, topic components with probability values ​​within a preset range (e.g., 0.2 to 0.6) are selected to avoid retaining only absolutely dominant topics or irrelevant noise topics. For each selected topic component, the sum of its probability value and its corresponding topic importance weight is calculated to obtain the topic concentration value for that paragraph. The topic importance weights can be determined based on statistical analysis of the relevance between topics and user preferences in the user's historical reading records, or based on feature importance analysis of the influence of topics on reading interest in manually labeled samples.

[0102] Finally, the topic concentration values ​​of all paragraphs are weighted and averaged according to the number of words in each paragraph to obtain the topic distribution value of the entire text. This topic distribution value is used to represent the degree of concentration of text content in the topic dimension, as well as its relevance distribution characteristics with multiple topics, providing quantitative feature input at the topic level for subsequent user interest prediction.

[0103] The topic model is obtained by estimating parameters on a training corpus containing manually labeled topic categories. The training corpus can be extracted from an e-book library and covers a variety of topic categories to ensure the model's topic coverage and generalization ability. Before training, the corpus is segmented, stop words and low-frequency words are removed, and a pre-defined vocabulary is constructed. During training, Gibbs sampling or variational inference methods are used to estimate the topic-word distribution and document-topic distribution parameters of the latent Dirichlet assignment model.

[0104] The topic components with probability values ​​within a preset range refer to those selected from the topic probability distribution vector of each paragraph, whose probability values ​​are greater than the lower limit of topic selection and less than or equal to the upper limit, to avoid selecting only absolutely dominant topics or noisy topics. Topic importance weights can be determined in either of the following ways: based on statistical calculations of the correlation between topic frequency and user preference in a manually annotated sample set; or based on the correspondence of different topics in the user's historical reading records.

[0105] In some embodiments of this application, the acquisition of knowledge domain coverage includes:

[0106] The text content is segmented and stop words are removed to obtain a word sequence. Keyword extraction is then performed on the word sequence, including sorting the words based on statistical results of word frequency and inverse document frequency combined with word position distribution information, and selecting the top-ranked words as keywords. Named entity recognition is then performed on the word sequence, using a recognition model trained on training corpora containing entity category annotations to obtain entity words representing people, places, institutions, works, and proper nouns. Keywords and entity words are then mapped to a node set in an external knowledge graph after string matching and synonym merging, obtaining the domain label corresponding to each node. The number of different domain labels is counted, and the proportion of the occurrence of nodes corresponding to each domain label in the entire text relative to the total number of words in the entire text is calculated. Finally, the number of domain labels and the proportion of domain labels are combined according to preset calculation rules to obtain the knowledge domain coverage.

[0107] It should be noted that, firstly, word segmentation is performed on the text content, and stop words are removed, resulting in a sequentially arranged sequence of word units. Word segmentation can employ a segmentation model based on a combination of statistics and rules, trained on a large-scale e-book corpus to ensure segmentation accuracy. The stop word list includes commonly used function words, auxiliary words, prepositions, and high-frequency meaningless words, and is manually reviewed and confirmed.

[0108] Next, keyword extraction is performed on the lexical sequence. Keyword extraction first calculates the term frequency (TF) of each lexical term in the full text, then calculates its inverse document frequency (IDF) in the external reference document set, and multiplies the two to obtain the TF-IDF value. Based on this, word position distribution information is introduced, that is, a position weighting coefficient is calculated based on the first occurrence position of the lexical term in the full text and its uniformity of distribution across different paragraphs. Finally, the TF-IDF value is multiplied by the position weighting coefficient to obtain a comprehensive score, and the lexical terms are sorted in descending order of comprehensive score. The top-ranking lexical terms are selected as keywords.

[0109] Next, named entity recognition is performed on the word sequence. This process is implemented using a recognition model trained on training corpora containing entity category labels (such as people, places, organizations, works, proper nouns, etc.). During training, word segmentation and part-of-speech tagging results are used as input features, combined with context window information, and entity words and their categories are identified using sequence labeling models such as Conditional Random Field (CRF) or Bidirectional Long Short-Term Memory Network (BiLSTM).

[0110] Subsequently, the keywords are matched with the entity tokens obtained from named entity recognition. Tokens that differ in form but have the same semantics are merged using synonyms. The merged token set is then mapped to the node set of an external knowledge graph. This mapping can be achieved through direct string matching or near-synonym matching based on vector similarity.

[0111] Extract the domain tag corresponding to each node from the knowledge graph, such as "physics," "literature," and "computer science." Count the number of different domain tags and calculate the proportion of the occurrence of the node corresponding to each domain tag in the entire text relative to the total number of words in the entire text.

[0112] Finally, according to preset calculation rules, the number of domain tags and the proportion of domain tags appearing are combined. For example, a weighted summation formula can be used, where the number of domain tags measures the breadth of knowledge coverage, and the proportion of domain tags appearing measures the depth of knowledge coverage. The weights can be determined based on the historical correlation statistics between user reading interests and domain tags, or obtained through feature importance analysis based on manually labeled sample sets. The calculated comprehensive value is the knowledge domain coverage of the text content, used to quantify the breadth and concentration of the text in the distribution of knowledge topics, providing knowledge structure feature input for the user preference prediction model.

[0113] In some embodiments of this application, the acquisition of narrative mode features includes:

[0114] The text content is segmented into plot units, which are achieved by identifying paragraph boundaries that indicate the progression of time, scene transitions, or changes in character relationships. Narrative structure elements are extracted for each plot unit, including the order of character appearances, event triggering conditions, conflict and turning point locations, and ending type. The extracted narrative structure elements are sequentially encoded into a structural feature sequence and input into a plot structure classification model. The plot structure classification model is trained on training corpus containing labeled plot structure categories and outputs the classification probability vector of the plot unit under each preset plot structure category. The classification probability vectors of all plot units are weighted and averaged according to the number of words in the plot unit to obtain the narrative pattern features of the text content.

[0115] It's important to note that, firstly, the text content is segmented into plot units. This segmentation is achieved by identifying paragraph boundaries that indicate time progression, scene transitions, or changes in character relationships. Specifically, a set of plot boundary recognition rules can be constructed, including: detecting the presence of explicit time markers (such as specific dates or time periods), location change indicators (such as "at the same time" or "on the other side"), or narrative statements involving changes in the relationships between major characters (such as "he met her for the first time"). When a paragraph that meets the boundary conditions is detected, it is used as the starting point of a new plot unit until the next boundary appears.

[0116] Then, narrative structure elements are extracted for each plot unit. These elements include: character appearance order: the main characters appearing in the plot unit and their order of appearance, obtained through named entity recognition and character referencing resolution methods; event triggering conditions: the key conditions or events that lead to the plot's progression within the plot unit, obtained through dependency parsing combined with trigger verb recognition; conflict and turning point locations: using emotion change detection and opposition relationship determination methods, the sentence indexes that indicate a shift from a balanced state to a conflict state or from a conflict to a resolution state in the narrative are identified; and ending type: the narrative direction is analyzed at the end of the plot unit to determine whether the ending of the unit is successful, unsuccessful, open, or suspenseful.

[0117] Next, the extracted narrative structural elements are encoded into structural feature sequences according to chronological order and plot development order. For example, the order of characters' appearances is encoded as a numerical sequence, event triggering conditions and conflict types are represented by classification codes, and positional indices are retained in the sequences so that the model can identify plot development patterns.

[0118] Subsequently, the structural feature sequences are input into the plot structure classification model. This model is trained on a training corpus containing manually labeled plot structure categories. The training corpus includes various types of e-book samples, and each plot unit in the sample is labeled with a structural category (such as "introduction-development-transition-conclusion", "multi-line interweaving", "flashback narrative", etc.). During training, the structural feature sequences are used as input to the model, and supervised learning methods (such as Bidirectional Long Short-Term Memory Network BiLSTM or Convolutional Neural Network CNN) are used for classification training, outputting the classification probability vector of each plot unit under each preset plot structure category.

[0119] Finally, the classification probability vectors of all plot units are weighted and averaged according to the number of words in each plot unit to obtain the overall narrative pattern feature of the text content. This feature is used to represent the type distribution of the text in the macro-narrative structure and serves as one of the inputs to the user preference prediction model to improve the matching degree of the recommendation results to the narrative style.

[0120] In some embodiments of this application, the training of the user preference prediction model includes:

[0121] The target user's reading behavior tags are paired with the text features of the e-books they have read using e-book identifiers to form a training sample set of users, text features, and behavior tags. Feature normalization and missing value imputation are performed on the training sample set, and it is divided into a training set and a validation set. The training set is input into a prediction model with a multi-layer nonlinear mapping structure for iterative training. The prediction model takes text features and user behavior tags as input and outputs the user's predicted interest rating for candidate e-books. The loss value is calculated by the difference between the prediction results on the validation set and the actual user behavior, and the prediction model parameters are updated using the loss value until the preset convergence condition is reached on the validation set, thus obtaining the user preference prediction model.

[0122] It's important to note that, firstly, the target user's reading behavior tags are paired with various textual features of the ebooks they have read, using ebook identifiers. Specifically, textual features (including information density, reasoning span, rhetorical intensity, thematic distribution, knowledge domain coverage, narrative pattern features, etc.) of each read ebook are extracted from the user's reading history and stored corresponding to the user's behavioral tags generated during the reading process. Behavioral tags may include chapter completion rate, average dwell time, number of times the user has added the bookmark or annotated it, etc., reflecting the user's actual level of interest in the ebook.

[0123] Then, data preprocessing is performed on the resulting set of user-text feature-behavior label triples. Preprocessing includes feature normalization, which maps each feature value to a uniform numerical range (e.g., 0 to 1) to avoid imbalanced effects on training caused by features with different dimensions; and missing value imputation, which fills in missing text features or behavior labels by using the average value of similar ebooks or similarity-based interpolation methods when some ebooks are missing.

[0124] Next, the preprocessed training sample set is divided into a training set and a validation set, for example, randomly divided in an 8:2 ratio, to ensure that user and book samples in the validation set do not appear repeatedly in the training set, thereby improving generalization ability.

[0125] Subsequently, the training set is input into the prediction model, which has a multi-layer nonlinear mapping structure, for iterative training. This prediction model may include an input layer, several hidden layers, and an output layer. The input layer receives a combined vector of text features and user behavior labels. The hidden layers realize high-order interactions between features through nonlinear activation functions. The output layer generates predicted user interest ratings for candidate e-books.

[0126] During training, the loss value is calculated by comparing the predicted results on the validation set with the actual user behavior. The loss value can be expressed in the form of mean squared error (MSE) or cross-entropy. The parameters of the prediction model are updated using this loss value. Gradient descent-based optimization methods (such as the Adam optimizer) are used to gradually adjust the model parameters until the loss value on the validation set converges to below a preset threshold, or training stops when the loss value decreases below a set value in several consecutive iterations.

[0127] See Figure 2 As shown, this embodiment of the invention provides an artificial intelligence-based e-book recommendation system, including:

[0128] The data collection module is configured to acquire the target user's historical reading data, which includes the user's reading behavior data and the text content of the e-books already read.

[0129] The feature extraction module is configured to extract text features from the text content. These features include: information density, calculated as the ratio of the number of concept units (after word segmentation and stop word filtering) to the total number of words; reasoning span, calculated as the proportion of sentence pairs with implicit premises between adjacent sentences to the total number of sentence pairs (implicit premises are identified using a text entailment judgment model); rhetorical intensity, calculated as the proportion of the frequency of rhetorical devices in the text to the total number of sentences; topic distribution, obtained by inputting the text content into a topic model, outputting a topic probability vector, and taking the weighted average within a preset range; knowledge domain coverage, obtained by mapping entities and keywords in the text content to an external knowledge graph, counting the number of different domain tags and the proportion of those tags appearing throughout the book, and calculating based on the number and proportion; and narrative pattern features, obtained by inputting the text into a plot structure classification model, identifying the plot structure type, and outputting the corresponding classification probability vector.

[0130] The rating module is configured to train a user preference prediction model based on text features and reading behavior labels; the text features of e-books in the library are extracted in the same way, input into the user preference prediction model, and interest ratings are obtained.

[0131] The recommendation module is configured to filter target ebooks based on interest ratings and recommend them to the target user's terminal.

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An AI-based e-book recommendation method, characterized in that, include: Obtain the target user's historical reading data, which includes the user's reading behavior data and the text content of the e-books already read; Text features are extracted from the text content, including: information density, calculated as the ratio of the number of concept units after word segmentation and stop word filtering to the total number of words; reasoning span, calculated as the proportion of sentence pairs with implicit premises between adjacent sentences to the total number of sentence pairs, wherein the implicit premises are identified by a text implication judgment model; rhetorical intensity, calculated as the proportion of the frequency of rhetorical devices in the text content to the total number of sentences; topic distribution, obtained by inputting the text content into a topic model, outputting probability distribution vectors under each preset topic, and performing a weighted average; knowledge domain coverage, obtained by mapping entities and keywords in the text content to an external knowledge graph, counting the number of different domain tags and the proportion of those domain tags in the entire book, and calculating based on the number and proportion; and narrative mode features, obtained by inputting the text content into a plot structure classification model, identifying plot structure types, and outputting the corresponding classification probability vectors. A user preference prediction model is trained based on the aforementioned text features and reading behavior labels; Text features of ebooks in the library are extracted in the same way, input into the user preference prediction model, and interest scores are obtained. Target ebooks are selected based on interest ratings and recommended to target users' devices; The acquisition of the reasoning span includes: The text content is segmented into sentences to obtain a sequentially arranged sentence sequence; a set of sentence pairs is constructed according to adjacency relationships, such that each sentence pair consists of the current sentence and its subsequent sentence; each sentence pair is input into the text entailment determination model, and the model outputs a determination result on whether the sentence pair contains an implicit premise; sentence pairs with implicit premises are marked as inference-related sentence pairs, and the number of inference-related sentence pairs is counted; the number of inference-related sentence pairs is divided by the total number of sentence pairs to obtain the inference span value, wherein the implicit premise is used to characterize the degree of semantic dependence between adjacent sentences and the density of cross-sentence inference; The acquisition of information density includes: performing word segmentation on the text content and deleting stop words to obtain a sequence of word elements arranged in order; identifying word combinations that conform to preset structural rules as candidate units based on part-of-speech tagging results, wherein the structural rules stipulate that the first word element in the candidate unit is a noun or a noun derived from a verb, and the subsequent word elements are modifiers or nouns, and the length of the entire combination is not less than two word elements; merging and deduplicating the candidate units selected by the structural rules to form a set of concept units; retrieving all occurrence positions of each concept unit in the text content, recording the sentence position index of the first occurrence, and counting the total number of occurrences; dividing the number of occurrences of each concept unit by the total number of words in the text content to obtain the density value of the concept unit; summing the density values ​​of all concept units in each chapter to obtain the chapter information density value; and weighting the information density values ​​of all chapters according to the number of words in the chapter to obtain the information density of the text content.

2. The AI-based e-book recommendation method according to claim 1, characterized in that, When acquiring reading behavior data of target users, this includes: Call the historical reading record interface from the reading application on the user's terminal to obtain a list of e-books that the user has read within a preset time range; For each e-book that has been read, reading behavior data associated with the user is extracted. The reading behavior data includes reading time spent on each chapter, chapter completion rate, number of jumps, number of rereads, and records of bookmarking or annotation operations. The corresponding full-text content of the e-book is obtained by accessing the e-book storage server or the user's local cache. After obtaining the full-text content, text decoding and format parsing are performed to remove the layout information and retain the plain text data for natural language processing. The reading behavior data is associated with the corresponding text content and stored using eBook identifiers.

3. The AI-based e-book recommendation method according to claim 1, characterized in that, The acquisition of the rhetorical intensity includes: The text content is segmented into sentences to obtain a sentence set, and each sentence is assigned its position index in the full text. Component parsing is performed on each sentence, taking sentences with word segmentation and part-of-speech tagging as input, and outputting tree-like data representing the hierarchical structure of phrase components. Based on the tree-like data, a structural deviation is calculated, which is the normalized value of the difference between the maximum depth of the phrase hierarchy in that sentence and the average depth of the phrase hierarchy in the entire text. Dependency parsing is performed on each sentence, taking sentences with word segmentation and part-of-speech tagging as input, and outputting a directed graph structure representing the dependency relationships between lexical units. Based on these dependency relationships, a relation complexity is calculated, which is the ratio of the number of dependency edges for non-core modifying, parallel, insertional, and comparative relationships in that sentence to the number of lexical units in that sentence. Finally, lexical unit merging is performed on each sentence. The merging process is achieved by mapping lexical units with different morphological changes to a unified word form. Based on the merged lexical units, a repetition reinforcement ratio is calculated, which is the ratio of the number of repeated lexical units within a preset interval to the total number of lexical units in the sentence. The structural deviation, relational complexity, and repetition reinforcement ratio are weighted and synthesized according to preset weights to obtain the rhetorical score of the sentence, and sentences with rhetorical scores greater than a preset threshold are marked as high-rhetorical sentences. The proportion of high-rhetorical sentences in the entire text is calculated as the rhetorical frequency component. Based on the position index, the distribution dispersion index of high-rhetorical sentences is calculated, and its normalized value is taken as the positional distribution component. High-rhetorical sentences are mapped to the plot units determined by the plot structure classification model, and the proportion of high-rhetorical sentences in each plot unit to the total number of sentences in that unit is calculated, and its mean is taken as the narrative coupling component. The rhetorical frequency component, positional distribution component, and narrative coupling component are weighted and summed according to preset weights to obtain the rhetorical intensity.

4. The AI-based e-book recommendation method according to claim 3, characterized in that, The acquisition of the topic distribution includes: The text content is segmented into paragraphs to obtain a sequentially arranged set of paragraphs. Each paragraph is segmented and stop words are removed to obtain a paragraph word sequence. This paragraph word sequence is converted into a bag-of-words vector representation, which is a frequency statistics vector based on a preset vocabulary. The bag-of-words vector is input into a topic model, which is a probabilistic generation model based on latent Dirichlet assignment and obtained through parameter estimation using text training corpora containing labeled topic categories. The probability distribution vector of each paragraph under each preset topic is obtained from the output of the topic model. For each paragraph, the topic components with probability values ​​within a preset range are summed by multiplying their probability values ​​by the corresponding topic importance weights to obtain the topic concentration value of that paragraph. The topic concentration values ​​of all paragraphs are weighted and averaged according to the number of characters in each paragraph to obtain the topic distribution value of the text content. This topic distribution value represents the degree of concentration of the text content in the topic dimension and its correlation distribution characteristics with multiple topics.

5. The AI-based e-book recommendation method according to claim 2, characterized in that, The acquisition of the knowledge domain coverage includes: The text content is segmented and stop words are removed to obtain a word sequence. Keyword extraction is performed on the word sequence, including sorting the words based on statistical results of word frequency and inverse document frequency combined with word position distribution information, and selecting the top-ranked words as keywords. Named entity recognition is performed on the word sequence, using a recognition model trained on training corpora containing entity category annotations to obtain entity words representing people, places, institutions, works, and proper nouns. The keywords and entity words are mapped to a node set of an external knowledge graph after string matching and synonym merging, obtaining the domain label corresponding to each node. The number of different domain labels is counted, and the proportion of the occurrence of nodes corresponding to each domain label in the full text to the total number of words in the full text is calculated. The number of domain labels and the proportion of domain labels are combined according to a preset calculation rule to obtain the knowledge domain coverage.

6. The AI-based e-book recommendation method according to claim 2, characterized in that, The acquisition of the narrative mode features includes: The text content is segmented into plot units, which are achieved by identifying paragraph boundaries that indicate the progression of time, scene transitions, or changes in character relationships. Narrative structure elements are extracted for each plot unit, including the order of character appearances, event triggering conditions, conflict and turning point locations, and ending type. The extracted narrative structure elements are sequentially encoded into a structural feature sequence and input into a plot structure classification model. The plot structure classification model is trained on training corpus containing labeled plot structure categories and outputs the classification probability vector of the plot unit under each preset plot structure category. The classification probability vectors of all plot units are weighted and averaged according to the number of words in the plot unit to obtain the narrative pattern features of the text content.

7. The AI-based e-book recommendation method according to claim 1, characterized in that, The training of the user preference prediction model includes: The target user's reading behavior tags are paired with the text features of the e-books they have read using e-book identifiers to form a training sample set of users, text features, and behavior tags. Feature normalization and missing value imputation are performed on the training sample set, which is then divided into a training set and a validation set. The training set is input into a prediction model with a multi-layer nonlinear mapping structure for iterative training. The prediction model takes text features and user behavior tags as input and outputs predicted user interest ratings for candidate e-books. A loss value is calculated based on the difference between the prediction results on the validation set and the actual user behavior, and this loss value is used to update the prediction model parameters until a preset convergence condition is met on the validation set, thus obtaining the user preference prediction model.

8. An AI-based e-book recommendation system, used to implement the AI-based e-book recommendation method according to any one of claims 1-7, characterized in that, The AI-based e-book recommendation system includes: The data acquisition module is configured to acquire the target user's historical reading data, which includes the user's reading behavior data and the text content of the e-books already read; The feature extraction module is configured to extract text features from the text content. These text features include: information density, calculated as the ratio of the number of concept units (after word segmentation and stop word filtering) to the total number of words; reasoning span, calculated as the proportion of sentence pairs with implicit premises between adjacent sentences to the total number of sentence pairs, where implicit premises are identified using a text implication judgment model; rhetorical intensity, calculated as the proportion of the frequency of rhetorical devices in the text to the total number of sentences; topic distribution, obtained by inputting the text content into a topic model, outputting probability distribution vectors under each preset topic, and performing a weighted average; knowledge domain coverage, obtained by mapping entities and keywords in the text content to an external knowledge graph, statistically analyzing the number of different domain tags and their proportion throughout the book, and calculating based on the number and proportion; and narrative pattern features, obtained by inputting the text content into a plot structure classification model, identifying plot structure types, and outputting corresponding classification probability vectors. The rating module is configured to train a user preference prediction model based on the text features and reading behavior labels; and to extract the text features of e-books in the library in the same way, input them into the user preference prediction model, and obtain interest ratings. The recommendation module is configured to filter target ebooks based on interest ratings and recommend them to the target user's terminal.

Images