Text content classification methods, devices, electronic devices and storage media

Abstract

This disclosure provides a text content classification method, apparatus, electronic device, and storage medium, relating to the field of artificial intelligence technology, and particularly to the field of deep learning technology. The specific implementation scheme is as follows: Obtain the name, description, and chapter title of the target text content; extract the name features of the name, the description features of the description, and the chapter title features of the chapter titles; based on the name features, description features, and chapter title features, as well as the correlations between these features, predict the labels of the target text content across multiple classification dimensions. Through the above processing, multi-label classification of text content can be achieved efficiently and accurately.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and more particularly to the field of deep learning technology, specifically to a text content classification method, apparatus, electronic device, and storage medium. Background Technology

[0002] In related technologies, text classification typically employs the following methods: collecting training samples; preprocessing the text; representing the processed text as vectors and determining appropriate feature extraction methods based on the text representation; training a model using traditional machine learning or deep learning methods; and classifying the text based on the trained model. However, these methods have drawbacks. When classifying each piece of text, each text belongs to only one category, meaning each text corresponds to only one label. This classification clearly fails to specifically reflect the characteristics of the text content. Furthermore, when the text to be classified is too long, the classification results of these methods are not accurate enough. Therefore, in related technical solutions, the text content classification method is relatively simplistic and fails to specifically reflect the characteristics of the text content. Summary of the Invention

[0003] This disclosure provides a method, apparatus, device, and storage medium for classifying text content.

[0004] According to one aspect of this disclosure, a text content classification method is provided, comprising: obtaining the name, description, and chapter title of the target text content; extracting name features of the name, description features of the description, and chapter title features of the chapter title; and predicting the labels of the target text content on multiple classification dimensions based on the name features, description features, and chapter title features, as well as the correlation between the name features, description features, and chapter title features.

[0005] Optionally, extracting the name features, the introduction features, and the chapter title features includes: extracting keywords from the name to obtain target words, extracting features from the target words, and obtaining the name features; filtering invalid words from the introduction to obtain target introduction content, extracting features from the target introduction content, and obtaining the introduction features; and performing invalid data cleaning on the chapter titles to obtain target chapter titles, extracting features from the target chapter titles, and obtaining the chapter title features.

[0006] Optionally, filtering invalid words in the introduction to obtain the target introduction content includes: using keyword extraction and whole-sentence filtering to filter invalid words in the introduction to obtain the target introduction content; and performing invalid data cleaning operations on chapter titles to obtain the target chapter titles includes: using methods to filter the format of chapter titles and extract the main body of the chapter titles to perform invalid data cleaning operations on the chapter titles to obtain the target chapter titles.

[0007] Optionally, by using a format that filters chapter titles and extracting the main body of the chapter titles, invalid data cleaning operations are performed on the chapter titles to obtain target chapter titles that include at least one of the following: deleting the chapter title number, deleting the chapter title suffix, deleting duplicate chapter titles, deleting chapter titles with a similarity greater than the similarity threshold, and deleting explanatory content from the chapter titles.

[0008] Optionally, based on name features, introduction features, and chapter title features, as well as the relationships between name features, introduction features, and chapter title features, the label of the target text content on multiple classification dimensions is predicted. This includes: using a prediction model, based on name features, introduction features, and chapter title features, as well as the relationships between name features, introduction features, and chapter title features, the label of the target text content on multiple classification dimensions is predicted. The prediction model is trained on multiple sets of sample data, which include: name features, introduction features, and chapter title features of the text content, as well as the label of the text content on multiple classification dimensions.

[0009] Optionally, the above method further includes: training the prediction model in the following ways: acquiring sentence pair sample data, using the sentence pair sample data for machine training to obtain an initial model, wherein the sentence pair sample data includes: sentence pairs and the semantics represented by the sentence pairs; training the initial model using domain corpus of the predetermined domain to which the text content belongs to obtain a language model for the predetermined domain, wherein the domain corpus includes: domain terms and the labels corresponding to the domain terms; and training the language model using multiple sets of sample data to obtain the prediction model.

[0010] Optionally, training the language model using multiple sets of sample data to obtain the prediction model includes: constructing a first loss function for individual labels in a multi-label dataset and a second loss function for the overall text content; and combining the first and second loss functions to train the language model to obtain the prediction model.

[0011] Optionally, the text content is a novel.

[0012] According to another aspect of this disclosure, an apparatus for classifying text content is provided, comprising: an acquisition module for acquiring the name, description, and chapter title of target text content; an extraction module for extracting name features of the name, description features of the description, and chapter title features of the chapter title; and a prediction module for predicting the labels of the target text content on multiple classification dimensions based on the name features, description features, and chapter title features, as well as the correlation between the name features, description features, and chapter title features.

[0013] Optionally, the extraction module includes: a first extraction unit, used to extract keywords from the name to obtain target words, extract features from the target words, and obtain name features; a second extraction unit, used to filter invalid words from the introduction to obtain target introduction content, extract features from the target introduction content, and obtain introduction features; and a third extraction unit, used to perform invalid data cleaning on chapter titles to obtain target chapter titles, extract features from the target chapter titles, and obtain chapter title features.

[0014] Optionally, the second extraction unit includes: a first processing subunit, used to filter invalid words in the introduction by using keyword extraction and whole sentence filtering to obtain the target introduction content; the third extraction unit includes: a second processing subunit, used to perform invalid data cleaning operation on the chapter titles by using the format of filtering chapter titles and extracting the main body of the chapter titles to obtain the target chapter titles.

[0015] Optionally, the second processing subunit is further configured to perform invalid data cleaning operations on the chapter titles by at least one of the following methods: using a filtered chapter title format to extract the main body of the chapter titles, to obtain the target chapter titles: deleting the chapter title sequence number, deleting the chapter title suffix, deleting duplicate chapter titles, deleting chapter titles with a similarity greater than a similarity threshold, and deleting explanatory content of the chapter titles.

[0016] Optionally, the prediction module includes: a prediction unit, used to use a prediction model to predict the labels of the target text content on multiple classification dimensions based on name features, introduction features, and chapter title features, as well as the association between name features, introduction features, and chapter title features. The prediction model is trained based on multiple sets of sample data, which include: name features, introduction features, and chapter title features of the text content, as well as the labels of the text content on multiple classification dimensions.

[0017] Optionally, the above apparatus further includes: a training module, comprising: a first training unit for acquiring sentence pair sample data and performing machine training using the sentence pair sample data to obtain an initial model, wherein the sentence pair sample data includes: sentence pairs and the semantics represented by the sentence pairs; a second training unit for training the initial model using domain corpus of a predetermined domain to which the text content belongs, to obtain a language model for the predetermined domain, wherein the domain corpus includes: domain terms and the labels corresponding to the domain terms; and a third training unit for training the language model using multiple sets of sample data to obtain a prediction model.

[0018] Optionally, the third training unit includes: a construction subunit for constructing a first loss function for individual labels in a multi-label language model and a second loss function for the overall text content; and a training subunit for training the language model by combining the first and second loss functions to obtain a prediction model.

[0019] Optionally, the text content in the aforementioned device is a novel.

[0020] According to another aspect of this disclosure, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform any of the methods described above.

[0021] According to another aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause a computer to perform any of the methods described above.

[0022] According to another aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements any of the methods described above.

[0023] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0024] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein:

[0025] Figure 1 This is a flowchart of a text content classification method provided according to embodiments of this disclosure;

[0026] Figure 2 This is a schematic diagram of a text content classification method provided according to an optional embodiment of this disclosure;

[0027] Figure 3 This is a structural block diagram of a text content classification device provided according to embodiments of the present disclosure;

[0028] Figure 4 This is a block diagram of an electronic device used to implement text content classification according to embodiments of the present disclosure. Detailed Implementation

[0029] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0030] Terminology Explanation

[0031] Natural Language Processing (NLP) is a discipline that studies language problems in human-computer interaction.

[0032] An Application Programming Interface (API) is a set of predefined interfaces (such as functions or HTTP interfaces) or conventions for connecting different components of a software system. It provides applications and developers with a set of routines accessible to them based on certain software or hardware, without requiring access to the source code or understanding of the details of the internal workings.

[0033] Naive Bayes method ( The Bayes model is a classification method based on Bayes' theorem and the assumption of conditional independence of features.

[0034] Logistic regression is a generalized linear regression analysis model that is commonly used in fields such as data mining, automated disease diagnosis, and economic forecasting.

[0035] Support Vector Machine (SVM) is a type of generalized linear classifier that performs binary classification of data using supervised learning. Its decision boundary is the maximum-margin hyperplane obtained by solving for the learning samples.

[0036] The K-Nearest Neighbor (KNN) algorithm is one of the simplest classification techniques in data mining. K-nearest neighbor means that each sample can be represented by its K nearest neighbors. The nearest neighbor algorithm is a method for classifying each record in a dataset.

[0037] XGBoost, an optimized distributed gradient boosting library, is designed for efficiency, flexibility, and portability. It implements machine learning algorithms within the Gradient Boosting framework. XGBoost provides parallel tree boosting (also known as GBDT, GBM), which can quickly and accurately solve many data science problems. The same code runs on major distributed environments (Hadoop, SGE, MPI) and can solve problems beyond billions of examples.

[0038] FastText is a fast text classification algorithm.

[0039] TextCNN is an algorithm that uses convolutional neural networks to classify text.

[0040] TextRNN is an algorithm that uses recurrent neural networks to classify text.

[0041] ERNIE 2.0 is a continuous learning semantic understanding framework that supports the incremental introduction of custom pre-trained tasks from different perspectives to capture lexical, syntactic, and semantic information from the corpus. These tasks train and update the model through multi-task learning, and whenever a new task is introduced, the framework can learn the new task without forgetting previously learned information.

[0042] BERT, short for Bidirectional Encoder Representation from Transformers, is a pre-trained language representation model. The BERT MASK method means that 15% of the original data is masked, and 85% is not. The masked 15% is processed in three ways: 80% is assigned a mask value, 10% is randomly assigned a value, and the remaining 10% retains its original value.

[0043] This disclosure provides a text content classification method. Figure 1 This is a flowchart of the text content classification method provided in the embodiments of this disclosure, such as... Figure 1 As shown, the method includes:

[0044] Step S102: Obtain the name, description, and chapter title of the target text content;

[0045] Step S104: Extract the name features of the name, the introduction features of the introduction, and the chapter title features of the chapter title;

[0046] Step S106: Based on name features, description features, and chapter title features, as well as the relationships between name features, description features, and chapter title features, predict the labels of the target text content on multiple classification dimensions.

[0047] Through the above steps, the name, description, and chapter titles of the target text content are obtained. Feature extraction is then performed on these names, descriptions, and chapter titles. Since the descriptions and chapter titles can comprehensively and accurately describe the main information of the text content, they can retain the most effective feature information for predicting the corresponding labels of the text content in each classification dimension. This makes the subsequent prediction and judgment process more efficient and accurate, improving the accuracy of the text content classification results. Finally, based on the obtained features of the name, description, and chapter titles, as well as the relationships between these features, the labels of the target text content in multiple classification dimensions are predicted. This enables multi-label classification of the text content, making the labels corresponding to the target text content more comprehensive and accurate. This solves the problem in related technical solutions where the text content classification method is relatively singular and cannot specifically reflect the characteristics of the text content, thus achieving efficient, accurate, and comprehensive classification results for the text content.

[0048] As an optional implementation, various methods can be adopted when extracting the name features, the introduction features, and the chapter title features. For example, the following methods can be used: extract keywords from the name to obtain the target words, extract the features of the target words to obtain the name features; filter out invalid words in the introduction to obtain the target introduction content, extract the features of the target introduction content to obtain the introduction features; perform invalid data cleaning on the chapter titles to obtain the target chapter titles, extract the features of the target chapter titles to obtain the chapter title features. By extracting keywords from the name, filtering invalid words from the introduction, and performing invalid data cleaning on the chapter titles, invalid information can be removed to the maximum extent, avoiding interference from noisy data, making the subsequent prediction results more accurate. After removing invalid information, feature extraction is performed on the target words, target introduction content, and target chapter titles. The obtained feature data will be used as a basis for judgment when predicting the corresponding tags of the text content in the future, which can further improve the accuracy of the text content classification results.

[0049] As an optional implementation, various methods can be used to filter invalid words in the introduction to obtain the target introduction content. For example, keyword extraction and whole-sentence filtering can be used to filter invalid words in the introduction to obtain the target introduction content. Similarly, various methods can be used to clean invalid data from chapter titles to obtain the target chapter titles. For example, filtering the chapter title format and extracting the main body of the chapter title can be used to clean invalid data from the chapter title to obtain the target chapter title. Two fine-grained filtering formats—keyword extraction and whole-sentence filtering—are available for the introduction content. These filtering methods are effective and highly adaptable, maximizing the retention of valid content in the current text introduction. By filtering the chapter title format and extracting the main body of the chapter title, invalid information in the chapter titles can be effectively filtered out, avoiding interference from noisy data and further improving the accuracy of the text content classification results.

[0050] As an optional implementation, a method of filtering chapter titles and extracting the main body of the chapter titles is used to perform invalid data cleaning on the chapter titles. When obtaining the target chapter titles, various methods can be used, such as at least one of the following: deleting the chapter title number, deleting the chapter title suffix, deleting duplicate chapter titles, deleting chapter titles with a similarity greater than a similarity threshold, and deleting explanatory content from the chapter titles. By deleting the above content, invalid data in the chapter titles can be filtered out to the maximum extent, avoiding interference from this invalid content with the predicted text classification labels and improving the accuracy of the text content classification label prediction results.

[0051] As an optional implementation, various methods can be used to predict the labels of target text content across multiple classification dimensions based on name features, description features, and chapter title features, as well as the relationships between these features. For example, a predictive model can be used to predict the labels of target text content across multiple classification dimensions based on name features, description features, and chapter title features, as well as the relationships between these features. The predictive model is trained on multiple sets of sample data, including the name features, description features, and chapter title features of the text content, as well as the labels of the text content across multiple classification dimensions. A predictive model trained on multiple sets of name features, description features, and chapter title features of the text content, as well as the labels of the text content across multiple classification dimensions, can achieve efficient and accurate multi-label classification prediction of text content, while improving the accuracy of the prediction results.

[0052] As an optional implementation, the prediction model can be trained in various ways, such as: acquiring sentence pair sample data; using the sentence pair sample data for machine training to obtain an initial model, wherein the sentence pair sample data includes: sentence pairs and the semantics represented by the sentence pairs; training the initial model using domain corpus of the predetermined domain to which the text content belongs to obtain a language model for the predetermined domain, wherein the domain corpus includes: domain terms and the corresponding labels of the domain terms; and training the language model using multiple sets of sample data to obtain the prediction model. By training the prediction model using sentence pairs and the semantics corresponding to the sentence pairs, the model's ability to understand and analyze word meanings and semantics is enhanced, enabling the model to more accurately judge the text content, thus facilitating subsequent classification and label prediction. Furthermore, by training the prediction model using domain terms of the predetermined domain to which the text content belongs and the corresponding labels of the domain terms, the prediction model's ability to analyze terms and corresponding labels in the predetermined domain can be enhanced, thereby improving the sensitivity and accuracy of classifying the target text content and optimizing the model's classification performance.

[0053] As an optional implementation, various methods can be used to train the language model using multiple sets of sample data to obtain the prediction model. For example, the following approach can be adopted: construct a first loss function for each individual label in a multi-label dataset and a second loss function for the overall text content; combine the first and second loss functions to train the language model and obtain the prediction model. By constructing the first and second loss functions, the prediction error of the model can be evaluated from both the perspective of each individual label and the overall text content, enabling the model to efficiently learn text content features during training, minimize test errors, further improve the accuracy of the prediction model, and also enhance the model's generalization ability.

[0054] As an optional implementation, the text content described above can be diverse, such as novels. By predicting the tags of novels, they can be categorized more comprehensively and meticulously, thereby better meeting users' reading needs and improving the accuracy of novel categorization retrieval and recommendation, ultimately enhancing the user experience. It should be noted that novels are merely an example of long-text content; other similar long-text content, such as online serialized content, can also be predicted using the above method for multi-tag classification.

[0055] Based on the above embodiments and optional embodiments, an optional implementation method is provided.

[0056] Long-text content (e.g., novels) constitutes a vast and widely read category. The book database stores key information such as novel titles, authors, synopses, all chapters, main text, and book classifications. Generally, novel classification tags are manually labeled by authors or publishers. While the novel database stores a wealth of key book information, the variety and coverage of classification tags are limited. This is manageable for popular books whose tags can be obtained from resource providers or existing transcoded resources, but for other books, tag classification presents challenges such as maintenance difficulties, a massive workload for manual labeling, extremely high costs, and uncertain returns. The incompleteness and lack of standardization in the book tag classification system make it difficult to manage data based on book classifications, hindering fine-grained novel tag classification within the database and impeding business scenarios such as general-purpose retrieval and recommendation.

[0057] The problems described above make it difficult to satisfy users in scenarios such as vertical category searches when using book tags as query terms for recall and ranking. This severely reduces the user's search and reading experience, thereby affecting user reading time and platform stickiness. It also hinders the creation of user profiles, the capture of user characteristics, and the identification of user behavioral intentions. It makes it impossible to identify user interests based on tags, making it difficult to implement personalized recommendations such as content push.

[0058] In related technologies, novel tag classification methods are generally based on manual annotation, with very few deep learning and machine learning methods applied to book tag classification. This is primarily because current Natural Language Processing (NLP) technology is not yet mature enough; novel texts are often too numerous and their content not refined enough, making it difficult to extract book tag features. Therefore, there are no universally applicable novel-related datasets, nor are there universal or industry-standard book tag annotation rules. Consequently, current techniques widely used for multi-classification of text struggle to classify extremely long texts with numerous tags.

[0059] This optional implementation aims to solve the above problems by using deep learning multi-label classification technology to mine multi-dimensional (currently seven dimensions, which may be added later as the classification becomes more refined): style, genre, background, plot, character identity, character personality, and fan fiction elements) tags of the novel using the novel's title, synopsis, and chapter titles.

[0060] The current requirement for novels is to tag them across seven dimensions: style, genre, setting, plot, character identity, character personality, and fan fiction elements. Because the tagging of novels, and indeed most books, is currently done manually, the available technologies and data are limited. For example, the mainstream 376 tags cover over 20,000 novels. Therefore, text classification in Natural Language Processing (NLP) is the preferred approach.

[0061] The text classification method and process used are as follows:

[0062] (1) Data Collection: Data sets are collected through web crawlers, Application Programming Interfaces (APIs), etc. It is crucial to ensure a balanced number of samples in each category, avoiding excessive imbalance. More training samples are better, as they improve the model's generalization ability. Simultaneously, ensure that the samples adequately cover the possible input space, not just common cases.

[0063] (2) Explore data and text preprocessing: Observe the characteristics of the data and perform corresponding preprocessing on the text according to the requirements.

[0064] (3) Feature engineering: Selecting text representation methods and feature extraction methods. The main representation methods are spatial vector models and word vector models. Different text representation methods have corresponding feature extraction methods.

[0065] (4) Training the classifier: Generally, there are traditional machine learning methods and deep learning methods. Traditional machine learning methods include Naive Bayes, Logistic Regression, Support Vector Machine, K-Nearest Neighbor (KNN) algorithm, and XGBoost. Deep learning methods are implemented by convolutional neural networks such as FastText, TextCNN, TextRNN, and Transformer.

[0066] The optional implementation of this disclosure employs a deep learning method. Deep learning, which is generally used for NLP text classification, is based on four neural network structures: NN, CNN, RNN, and Transformer.

[0067] NN refers to traditional neural network models, such as FastText. CNN refers to convolutional neural network models. Image network structures are used for text processing, such as TextCNN. Later, based on CNN, CharCNN, VDCNN, and others emerged. RNN refers to recurrent neural network models. RNNs have natural advantages for text processing, but ordinary RNNs suffer from gradient explosion / vanishing problems, leading to improved RNN models: LSTM and GRU. LSTM / GRU addresses the gradient problem in RNN networks based on the concept of gates. Based on the concept of context, there are also Bi-LSTM / Bi-GRU, etc. RNN, LSTM, and GRU are mentioned. Transformer is a new network architecture proposed in recent years, a more powerful feature extractor than CNN / RNN. The key point of Transformer is its self-attention mechanism. Currently, the most widely used and performing network models are based on Transformer. The optional implementation of this disclosure is based on the ERNIE model using a multi-head attention mechanism.

[0068] The aforementioned related technologies, which classify text content using tag-based multi-classification, are insufficient to meet current needs. "Multi-classification" refers to a system with multiple categories, where each sample's tag value can only belong to one category. However, for the multi-tag classification problem in this optional implementation, each sample's tag value may belong to more than one category. In the context of novels, this can be understood as a novel that is both a sweet romance and a domineering CEO romance, with a protagonist who is also cunning and manipulative, and a plot that is also set in a school setting. This novel would have four tags: sweet romance, cunning, domineering CEO, and school setting. These tags can be further categorized by style, character identity, character personality, plot, etc. Therefore, the number of class tags is uncertain; some samples may have only one class tag, while others may have as many as a dozen. Therefore, this optional implementation addresses "multi-tag text classification" rather than "multi-classification of text."

[0069] As described in the preceding requirements, a single data point in a classification task may have multiple dimensions, and each dimension may have more than one label. Furthermore, there may be interdependencies between category labels. For example, if a novel's character is a general, the genre is more likely to be "iron-blooded"; if the plot involves a son-in-law, the genre is more likely to be sweet romance or a lighthearted, fast-paced story. In contrast, in typical text multi-classification tasks, the categories are mutually exclusive. For instance, in sentiment analysis of a text, categories like happy, sad, and angry are independent of each other. Clearly, the current requirements better align with the definition of multi-label text classification. However, there are currently no mature solutions for multi-label classification problems in relevant technologies, primarily because the labels may have complex or hierarchical dependencies, and no mature model exists at present to uncover these relationships.

[0070] As can be seen from the requirements of the optional implementation methods disclosed herein, the requirements for novel materials are quite high. A book can contain tens of thousands of words, and it is necessary to find content that represents the book's style, character traits, plot features, and other characteristics. However, general deep learning models for text classification have limitations on the number of characters in the input text. Commonly used classification models are basically limited to 512 characters, with a maximum of 1024 characters. In fact, representing all the dimensional features of a book within 512 characters is quite challenging.

[0071] Annotating content in novels and other books is costly, and the datasets are limited. Deep learning models struggle to train a highly usable prediction model on extremely small datasets.

[0072] There are many multi-label learning algorithms in related technologies. Based on the problem-solving perspective, these algorithms can be divided into two main categories: one is problem-transformation-based methods, and the other is algorithm-applicability-based methods. Problem-transformation-based methods transform the problem data to make it suitable for existing algorithms; algorithm-applicability-based methods refer to extending a specific algorithm to handle multi-label data, thereby improving the algorithm and making it applicable to the data.

[0073] To address the issue that multi-classification models cannot resolve complex multi-label problems and hierarchical dependencies between labels, this optional implementation method directly performs multi-label classification by modifying the algorithm, rather than transforming the problem into different subsets. Compared to transforming the problem into multiple classifiers, this method has higher applicability, preserves the independence and interdependencies between labels, and offers greater practicality and universality.

[0074] The optional implementation of this disclosure achieves a significant data filtering and cleaning strategy for synopsis, chapter titles, etc., enhances the feature representation of the language model in the novel task, improves the commonly used multi-label text classification model classifier, and makes the novel perform well in the multi-dimensional classification task with 376 labels. Figure 2 This is a schematic diagram of a text content classification method provided according to an optional embodiment of this disclosure, such as... Figure 2 As shown, the specific process is as follows:

[0075] (1) Processing of novel materials

[0076] Novel content crawling: Obtain novel chapter content from the database, mainly by using the database that stores the chapter directory and the database that stores the chapter content. The novel title, synopsis and all chapter titles are searched according to the book identifier index. In this optional implementation method, the main content features of the book are mined from the chapter titles of the first 2000 chapters of each novel.

[0077] Invalid word filtering in the synopsis data: The synopses crawled from the book material database are raw and unprocessed, with mixed content. They include not only valid synopsis content but also numerous messages from the author to readers, such as recommendations for new or completed novels, announcements of publication or release on various platforms, update schedules, requests for monthly tickets, collections, subscriptions, comments, and donations, novel groups, various reader exchange groups, novel award information, emoticons in the form of (n*_*n), the novel platform's Uniform Resource Locator (URL), various non-semantic interjections such as emmmmmm, 886, 5555, and even default catch-all synopsis descriptions for novels without synopsis. To address this, a strategy was designed to segment the novel text into sentences based on specified delimiters and filter sentences containing invalid keywords. This primarily targets representative descriptive fields in the synopsis using regular expression matching and Chinese character detection, with two fine-grained filtering formats: keyword extraction and whole-sentence filtering. Not only is the filtering effect significant and the adaptability high, but it also retains the effective content of the current novel synopsis to the greatest extent.

[0078] Chapter title data cleaning and main content extraction: Novel chapter titles vary widely, but the main content of each chapter is generally summarized in the title, attracting readers' attention and being concise and easy to understand. Therefore, in order to use the main content of chapter titles for multi-label classification of novel text, it is necessary to perform very fine filtering of the prefixes and suffixes of chapter titles and some general data cleaning, as follows:

[0079] First, general novels are in serial form, and the explanatory information related to chapters often appears in the form of chapter titles, such as a one-day hiatus, a one-day leave, or the author taking a one-week vacation. These unpredictable explanations are mixed in the list of chapter titles but are not chapter titles. The main difference is that the explanatory content does not carry chapter numbers. Therefore, for novels with chapter numbers, the main body is obtained by identifying the chapter numbers and removing the numbers and suffixes. However, the chapter number formats of different novels are highly inconsistent, mainly including: Chapter 13 / Thirteenth Chapter / Thirteen Chapters / [Chapter 13] / Chapter Thirteen-Fourteen (two chapters combined) / Thirteen - Fourteen / Thirteenth Chapter_1 / Thirteen: / 13、 / The Strongest Son-in-law - 13 / Chapter 13 / Chapter 13 of the First Volume of the Heavenly Book / Thirteenth Episode / Thirteen Episodes / [Season I] Chapter Thirteen / Extra Chapter 13. There are even inverted title numbers. For these complex situations, a hierarchical title number filtering strategy is designed. It is divided into 10 levels to ensure that the complete chapter numbers are extracted to the greatest extent and to avoid misclassification.

[0080] Secondly, for novels without chapter numbers or those with a mix of chapter numbers, a strategy for handling titles without numbers is designed, and the content entry requirements are relaxed. Mainly, a targeted judgment is made on whether the chapter content is valid.

[0081] At the same time, for main body content with the same or similar prefixes, the prefix tree algorithm is used to filter out content with high repetition as much as possible, such as: "Jianghu 1_All over the World", "Jianghu 2_A pot of turbid wine to ask the sky". The prefix tree will match "Jianghu x_", and the cut content is more suitable for input into the network.

[0082] In addition, for cases where the main body of the two consecutive titles is the same or has very little difference, a content similarity determination algorithm (edit distance, tightness, bm25, etc.) is added, such as "Warcraft in Battle vs. The Unintentional Mage", "Warcraft in Another Battle vs. The Unintentional Mage". For many cases where the content similarity of consecutive titles exceeds the set threshold, content filtering is performed to make the sample data more high-quality.

[0083] Then, filter the chapter content suffixes, such as: Third Update, Additional Update, Supplementary Update, (Update Completed), [Please take a leave on the 20th], (Completed), (Request for votes and rewards), _Upper, —1. A hierarchical suffix filtering strategy is designed.

[0084] Finally, clean the invalid chapter data. If the chapter is similar to: End of the full text, End of this book, Final chapter, Grand finale, End, Prologue, Preface, Extra chapter. Clean up this content that has no practical meaning. Avoid interference in the model learning.

[0085] Training and evaluation sample selection: The samples are organized according to a certain pattern. After cleaning and filtering, some samples may have empty descriptions or insufficient chapter title text. In small datasets with multiple labels, it is essential to clean up weak and negative samples to reduce noise. Furthermore, it is necessary to compare the overlap of labels between the training and test sets to avoid model distortion caused by large differences in the similarity of label coverage and proportion between the training and test sets.

[0086] (2) Processing of the model

[0087] Using ERNIE (base 2.0 and above, including but not limited to ernie_2.3 and ernie_doc) pre-trained models: Compared to shallow lexical analysis models and BERT's Mask method, ERNIE incorporates more human knowledge into its data processing and model structure, such as phrases and proper nouns. The input for model pre-training is constructed from encyclopedic, news, and forum dialogue data, creating sentence pairs with contextual relationships. Internal lexical analysis tools are used to segment these sentence pairs at different granularities (characters, words, entities, etc.). During pre-training, word meaning, semantics, and inter-sentence representation features are learned, increasing the richness of semantic representation.

[0088] Post-pretraining of the language model using novel corpora: Pre-training the language model on general internet corpora is insufficient to achieve high availability for current technical tasks. To increase the sensitivity and accuracy of multi-label classification of novels, post-pretraining of the pre-trained model was performed using corpora such as novel label definition explanations, similar label introductions, and descriptions of various novel genres and styles. After fine-tuning the language model on the current novel corpora, the model's performance was significantly improved.

[0089] Feature engineering optimization: The calculation methods for output probabilities and cross-entropy were modified to enable the model to fully learn the dependencies between labels. The model was fine-tuned specifically for novel tasks. Optimizations were made across multiple aspects, including the number of convolutional layers, activation function, loss function, learning rate, forward and backpropagation of data after input into the network, and batch size. A unique evaluation formula for multi-label novel classification was developed. The model's prediction error was evaluated from two perspectives: single-dimensional (style, plot, character personality, character identity, fan fiction, background, genre, etc.) recall and overall recall at the book level. This enabled the model to learn the most effective novel features during training, minimized testing error, and improved the generalization ability of the multi-label novel classification model.

[0090] The deep learning model novel multi-label device: The prediction model obtained after feature engineering optimization training is the model that will be used in the actual multi-label device. In practical applications, single prediction and batch prediction are designed for multi-label prediction. By inputting the unique identifier of the book (such as book_id) into the built classification prediction device, the novel data in the material library can be organized according to the sample format organized during training and input into the network to obtain multiple labels corresponding to the novel in seven dimensions.

[0091] The optional implementation methods disclosed herein can improve the accuracy of multi-tag classification of novels, more accurately label novel categories, and help improve the user's search experience, reading experience, and recommendation accuracy in business operations. Through the optional implementation methods disclosed herein, the recommendation hit rate of novels and the user's reading time on the platform can be improved, and the supplementary tag information of novels in various dimensions can also be better applied to other strategies.

[0092] In this embodiment of the disclosure, a device for classifying text content is also provided. Figure 3 According to the structural block diagram of the text content classification device provided in the embodiments of this disclosure, such as Figure 3 As shown, the device includes: an acquisition module 31, an extraction module 32, and a prediction module 33. The device will be described below:

[0093] The acquisition module 31 is used to acquire the name, description, and chapter title of the target text content; the extraction module 32 is connected to the acquisition module 31 and is used to extract the name features of the name, the description features of the description, and the chapter title features of the chapter title; the prediction module 33 is connected to the extraction module 32 and is used to predict the labels of the target text content on multiple classification dimensions based on the name features, description features, and chapter title features, as well as the correlation between the name features, description features, and chapter title features.

[0094] As an optional embodiment, the extraction module 32 includes: a first extraction unit, used to extract keywords from the name to obtain target words for the name, extract features from the target words for the name, and obtain name features for the name; a second extraction unit, used to filter invalid words from the introduction to obtain target introduction content, extract features from the target introduction content, and obtain introduction features for the introduction; and a third extraction unit, used to perform invalid data cleaning operations on chapter titles to obtain target chapter titles, extract features from the target chapter titles, and obtain chapter title features for the chapter titles.

[0095] As an optional embodiment, the second extraction unit includes: a first processing subunit, used to filter invalid words in the introduction by using keyword extraction and whole sentence filtering to obtain the target introduction content; the third extraction unit includes: a second processing subunit, used to perform invalid data cleaning operation on the chapter titles by using the format of filtering chapter titles and extracting the main body of the chapter titles to obtain the target chapter titles.

[0096] As an optional embodiment, the second processing subunit is further configured to perform invalid data cleaning operations on the chapter titles by at least one of the following methods: using a filtered chapter title format to extract the main body of the chapter titles, to obtain the target chapter titles: deleting the chapter title sequence number, deleting the chapter title suffix, deleting duplicate chapter titles, deleting chapter titles with a similarity greater than a similarity threshold, and deleting explanatory content of the chapter titles.

[0097] As an optional embodiment, the prediction module 33 includes: a prediction unit, used to employ a prediction model to predict the labels of target text content on multiple classification dimensions based on name features, introduction features, and chapter title features, as well as the correlation between name features, introduction features, and chapter title features. The prediction model is trained based on multiple sets of sample data, which include: name features, introduction features, and chapter title features of the text content, as well as the labels of the text content on multiple classification dimensions.

[0098] As an optional embodiment, the above apparatus further includes: a training module, which includes: a first training unit for acquiring sentence pair sample data and performing machine training using the sentence pair sample data to obtain an initial model, wherein the sentence pair sample data includes: sentence pairs and the semantics represented by the sentence pairs; a second training unit for training the initial model using domain corpus of a predetermined domain to which the text content belongs to obtain a language model for the predetermined domain, wherein the domain corpus includes: domain terms and the labels corresponding to the domain terms; and a third training unit for training the language model using multiple sets of sample data to obtain a prediction model.

[0099] As an optional embodiment, the third training unit includes: a construction subunit for constructing a first loss function for individual labels in a multi-label language model and a second loss function for the overall text content; and a training subunit for training the language model by combining the first and second loss functions to obtain a prediction model.

[0100] As an optional embodiment, the text content in the above-described device is a novel.

[0101] The acquisition, storage, and application of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0102] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0103] Figure 4 A schematic block diagram of an example electronic device 400 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0104] like Figure 4 As shown, device 400 includes a computing unit 401, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 402 or a computer program loaded from storage unit 408 into random access memory (RAM) 403. RAM 403 may also store various programs and data required for the operation of device 400. The computing unit 401, ROM 402, and RAM 403 are interconnected via bus 404. Input / output (I / O) interface 405 is also connected to bus 404.

[0105] Multiple components in device 400 are connected to I / O interface 405, including: input unit 406, such as keyboard, mouse, etc.; output unit 407, such as various types of monitors, speakers, etc.; storage unit 408, such as disk, optical disk, etc.; and communication unit 409, such as network card, modem, wireless transceiver, etc. Communication unit 409 allows device 400 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0106] The computing unit 401 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 401 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 401 performs the various methods and processes described above, such as text content classification methods. For example, in some embodiments, the text content classification method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 408. In some embodiments, part or all of the computer program may be loaded and / or installed on device 400 via ROM 402 and / or communication unit 409. When the computer program is loaded into RAM 403 and executed by the computing unit 401, one or more steps of the text content classification method described above may be performed. Alternatively, in other embodiments, the computing unit 401 may be configured to perform the text content classification method by any other suitable means (e.g., by means of firmware).

[0107] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0108] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0109] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0110] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0111] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with embodiments of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0112] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0113] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0114] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for classifying text content, comprising: Obtain the name, synopsis, and chapter titles of the target text content, wherein the target text content is a novel; Extract the name features of the name, the introduction features of the introduction, and the chapter title features of the chapter titles. The name features are obtained by extracting features from the target words of the name, the target words are obtained by extracting the keywords of the name, the introduction features are obtained by extracting features from the target introduction content, and the chapter title features are obtained by extracting features from the target chapter titles. The target chapter titles are obtained by performing invalid data cleaning operations on the chapter titles using at least one of the following methods: deleting the chapter title number, deleting the chapter title suffix, deleting duplicate chapter titles, deleting chapter titles with a similarity greater than a similarity threshold, and deleting the explanatory content of the chapter titles. Based on the name feature, the introduction feature, and the chapter title feature, as well as the correlation between the name feature, the introduction feature, and the chapter title feature, the labels of the target text content on multiple classification dimensions are predicted. These multiple classification dimensions include: style, genre, background, plot, character identity, character personality, and fan fiction elements. The labels of the target text content on multiple classification dimensions are obtained by processing the name feature, the introduction feature, and the chapter title feature, as well as the correlation between these features, through a prediction model. The prediction model is trained using multiple sets of sample data through a first loss function for individual labels and a second loss function for the overall text content. The prediction model is obtained by training in the following manner: Obtain sentence pair sample data, and use the sentence pair sample data for machine training to obtain an initial model, wherein the sentence pair sample data includes: sentence pairs and the semantics represented by the sentence pairs; The initial model is trained using domain corpus of the predetermined domain to which the text content belongs, to obtain the language model for the predetermined domain, wherein the domain corpus includes: domain terms and the corresponding tags of the domain terms; The language model is trained using the multiple sets of sample data to obtain the prediction model.

2. The method according to claim 1, wherein, The extraction of name features, the description features, and the chapter title features includes: Keywords of the name are extracted to obtain target words, and features of the target words are extracted to obtain the name features of the name; Invalid words in the description are filtered out to obtain the target description content. Features of the target description content are extracted to obtain the description features of the description. The invalid data cleaning operation is performed on the chapter titles to obtain the target chapter titles. The features of the target chapter titles are extracted to obtain the chapter title features of the chapter titles.

3. The method according to claim 2, wherein, The step of filtering invalid words in the description to obtain the target description content includes: using keyword extraction and whole-sentence filtering to filter invalid words in the description to obtain the target description content; The step of performing invalid data cleaning on the chapter titles to obtain the target chapter titles includes: using a method that filters the format of the chapter titles and extracts the main body of the chapter titles, performing invalid data cleaning on the chapter titles to obtain the target chapter titles.

4. The method according to claim 3, wherein, The method of filtering the chapter titles and extracting the main body of the chapter titles involves performing invalid data cleaning on the chapter titles, resulting in target chapter titles that include at least one of the following: Delete the sequence number of the chapter title, delete the suffix of the chapter title, delete duplicate chapter titles, delete chapter titles with a similarity greater than the similarity threshold, and delete the descriptive content of the chapter title.

5. The method according to claim 1, wherein, The method includes: The prediction model is trained based on multiple sets of sample data, including: the name features, introduction features, and chapter title features of the text content, as well as the labels of the text content on multiple classification dimensions.

6. The method according to claim 5, wherein, The method further includes: Construct a first loss function for individual tags in a multi-label language model and a second loss function for the overall text content.

7. A device for classifying text content, comprising: The acquisition module is used to acquire the name, synopsis, and chapter title of the target text content, wherein the target text content is a novel; An extraction module is used to extract the name features of the name, the introduction features of the introduction, and the chapter title features of the chapter titles. The name features are obtained by extracting features from the target words of the name, which are obtained by extracting the keywords of the name. The introduction features are obtained by extracting features from the target introduction content. The chapter title features are obtained by extracting features from the target chapter titles. The target chapter titles are obtained by performing invalid data cleaning operations on the chapter titles using at least one of the following methods: deleting the chapter title number, deleting the chapter title suffix, deleting duplicate chapter titles, deleting chapter titles with a similarity greater than a similarity threshold, and deleting descriptive content from the chapter titles. The prediction module is used to predict the labels of the target text content on multiple classification dimensions based on the name feature, the introduction feature, and the chapter title feature, as well as the correlation between the name feature, the introduction feature, and the chapter title feature. The multiple classification dimensions include: style, genre, background, plot, character identity, character personality, and fan fiction elements. The labels of the target text content on multiple classification dimensions are obtained by processing the name feature, the introduction feature, and the chapter title feature, as well as the correlation between the name feature, the introduction feature, and the chapter title feature, using a prediction model. The prediction model is trained on a language model using multiple sets of sample data through a first loss function for individual labels and a second loss function for the text content as a whole. The device further includes: a training module, the training module comprising: The first training unit is used to acquire sentence pair sample data and perform machine training using the sentence pair sample data to obtain an initial model. The sentence pair sample data includes: sentence pairs and the semantics represented by the sentence pairs. The second training unit is used to train the initial model using domain corpus of the predetermined domain to which the text content belongs, so as to obtain a language model for the predetermined domain, wherein the domain corpus includes: domain terms and the corresponding tags of the domain terms; The third training unit is used to train the language model using the multiple sets of sample data to obtain the prediction model.

8. The apparatus according to claim 7, wherein, The extraction module includes: The first extraction unit is used to extract keywords from the name to obtain target words for the name, and extract features from the target words to obtain name features for the name. The second extraction unit is used to filter out invalid words in the introduction, obtain the target introduction content, extract the features of the target introduction content, and obtain the introduction features of the introduction. The third extraction unit is used to perform invalid data cleaning operations on the chapter titles to obtain target chapter titles, extract features of the target chapter titles, and obtain chapter title features of the chapter titles.

9. The apparatus according to claim 8, wherein, The second extraction unit includes: a first processing subunit, used to filter invalid words in the introduction by using keyword extraction and whole sentence filtering to obtain the target introduction content; The third extraction unit includes a second processing subunit, used to perform invalid data cleaning on the chapter titles by using a format that filters the chapter titles and extracting the main body of the chapter titles, to obtain the target chapter titles.

10. The apparatus according to claim 9, wherein, The second processing subunit is further configured to perform invalid data cleaning on the chapter titles by at least one of the following methods, using a format that filters the chapter titles and extracts the main body of the chapter titles, to obtain the target chapter titles: Delete the sequence number of the chapter title, delete the suffix of the chapter title, delete duplicate chapter titles, delete chapter titles with a similarity greater than the similarity threshold, and delete the descriptive content of the chapter title.

11. The apparatus according to claim 7, wherein, The prediction module includes: The prediction model is trained based on multiple sets of sample data, including: the name features, introduction features, and chapter title features of the text content, as well as the labels of the text content on multiple classification dimensions.

12. The apparatus according to claim 11, wherein, The third training unit includes: The construction subunit is used to construct a first loss function for individual tags in a multi-tag language model and a second loss function for the overall text content.

13. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 6.

14. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1 to 6.

15. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1 to 6.

Images