Content topic classification method and apparatus, electronic device, and storage medium

Abstract

The application discloses a content theme classification method and device, electronic equipment and a storage medium; the content to be classified can be acquired, and a historical publishing content corresponding to a theme distribution type of a publisher of the content to be classified is recognized; when the theme distribution type is a target type, attribute information corresponding to the publisher and theme information of historical publishing content of the publisher are acquired; based on content information of the content to be classified and the attribute information of the publisher and the theme information of the historical publishing content, a probability that the content to be classified belongs to each preset theme is predicted; and according to the probability, a target theme of the content to be classified is determined from the preset themes. When the theme distribution type of the historical publishing content of the publisher is the target type, the application can predict the theme of the content to be classified in combination with the attribute information of the publisher and the theme information of the historical publishing content, and the accuracy of content theme classification can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to a content theme classification method and device, electronic equipment and storage medium. BACKGROUND

[0002] With the continuous development of network information technology, the number of various content information on the Internet increases sharply. In order to analyze the content interested by users from the mass content, the content information needs to be processed. The content classification is a key technology for processing large-scale content information and plays a crucial role in information processing. The content classification is to classify the content data according to a certain classification system or standard to obtain one or more corresponding content theme labels.

[0003] At present, the content theme classification method of the traditional technology usually only classifies the theme of the content according to the content itself. However, the related information of the content is not fully utilized, and the accuracy of the theme classification result is low. SUMMARY

[0004] The embodiments of the present application provide a content theme classification method and device, electronic equipment and storage medium, which can improve the accuracy of content theme classification.

[0005] The embodiments of the present application provide a content theme classification method, which comprises:

[0006] obtaining the content to be classified, and identifying the theme distribution type of the historical published content of the publisher of the content to be classified;

[0007] when the theme distribution type is a target type, obtaining the attribute information corresponding to the publisher and the theme information of the historical published content of the publisher;

[0008] based on the content information of the content to be classified, and the attribute information of the publisher and the theme information of the historical published content, predicting the probability that the content to be classified belongs to each preset theme;

[0009] determining the target theme of the content to be classified from the preset themes according to the probability.

[0010] Correspondingly, the embodiments of the present application provide a content theme classification device, which comprises:

[0011] an identifying unit, configured to obtain the content to be classified, and identify the theme distribution type of the historical published content of the publisher of the content to be classified;

[0012] an obtaining unit, configured to, when the theme distribution type is a target type, obtain the attribute information corresponding to the publisher and the theme information of the historical published content of the publisher.

[0013] a prediction unit, configured to predict a probability that the content to be classified belongs to each preset theme based on content information of the content to be classified, and attribute information of a publisher and theme information of historical published content of the publisher;

[0014] a determination unit, configured to determine a target theme of the content to be classified from the preset themes according to the probability.

[0015] Optionally, in some embodiments of the present application, the prediction unit can also be configured to predict the probability that the content to be classified belongs to each preset theme based on content information of the content to be classified when the theme distribution type is a non-target type; and the determination unit can also be configured to determine the target theme of the content to be classified from the preset themes according to the probability.

[0016] Optionally, in some embodiments of the present application, the identification unit can be specifically configured to perform theme distribution analysis on the historical published content based on attribute information of the publisher corresponding to the content to be classified and theme information of the historical published content of the publisher, and determine the theme distribution type corresponding to the historical published content.

[0017] Optionally, in some embodiments of the present application, the identification unit can include a first fusion subunit, a first extraction subunit and a first prediction subunit, as follows:

[0018] The first fusion subunit is configured to fuse attribute information of the publisher corresponding to the content to be classified and theme information of the historical published content of the publisher, to obtain theme distribution information.

[0019] The first extraction subunit is configured to perform feature extraction on the theme distribution information, to obtain distribution feature information of the theme distribution information.

[0020] The first prediction subunit is configured to predict the theme distribution type corresponding to the historical published content according to the distribution feature information.

[0021] Optionally, in some embodiments of the present application, the theme distribution information includes at least one text unit.

[0022] The first extraction subunit can be specifically configured to extract word-level feature information of each text unit in the theme distribution information; process the word-level feature information of each text unit based on word-level feature information of a text unit corresponding to a context of the text unit; and fuse the processed word-level feature information of each text unit, to obtain the distribution feature information of the theme distribution information.

[0023] Optionally, in some embodiments of the present application, the first extraction subunit can be specifically configured to perform feature extraction on the topic distribution information by using the trained topic distribution identification model, to obtain distribution feature information of the topic distribution information.

[0024] Optionally, in some embodiments of the present application, the content topic classification apparatus can further include a training unit configured to train the topic distribution identification model. Specifically, the training unit is configured to obtain training data, the training data including attribute information of a sample publisher, topic information of sample historical publishing content of the sample publisher, and an expected topic distribution type of the sample historical publishing content; fuse the attribute information of the sample publisher and the topic information of the sample historical publishing content to obtain sample topic distribution information; perform feature extraction on the sample topic distribution information by using the topic distribution identification model, to obtain distribution feature information of the sample topic distribution information; predict an actual topic distribution type of the sample historical publishing content according to the distribution feature information; and adjust parameters of the topic distribution identification model according to the actual topic distribution type and the expected topic distribution type, to obtain a trained topic distribution identification model.

[0025] Optionally, in some embodiments of the present application, the prediction unit can include a second fusion subunit, a second extraction subunit, and a second prediction subunit, as follows:

[0026] The second fusion subunit is configured to fuse the content information of the content to be classified, and the attribute information of the publisher and the topic information of the historical publishing content, to obtain fused information.

[0027] The second extraction subunit is configured to perform feature extraction on the fused information, to obtain feature information of the fused information.

[0028] The second prediction subunit is configured to predict a probability that the content to be classified belongs to each preset topic according to the feature information.

[0029] Optionally, in some embodiments of the present application, the fused information includes at least one text unit. The second extraction subunit can be specifically configured to extract word-level feature information of each text unit in the fused information; process the word-level feature information of each text unit based on word-level feature information of the text unit corresponding to a context of the text unit; and fuse the processed word-level feature information of each text unit, to obtain the feature information of the fused information.

[0030] Optionally, in some embodiments of the present application, the step of “extracting word-level feature information of each text unit in the fused information” can include:

[0031] obtaining a content vector and a type vector corresponding to each text unit in the fused information, the type vector representing an information type to which the text unit belongs;

[0032] encoding position information of each text unit in the fused information to obtain a position vector corresponding to each text unit;

[0033] fusing the content vector, the type vector, and the position vector to obtain word-level feature information of each text unit in the fused information.

[0034] An electronic device provided in an embodiment of the present application includes a processor and a memory, the memory storing a plurality of instructions, and the processor loads the instructions to perform steps in a content topic classification method provided in an embodiment of the present application.

[0035] An embodiment of the present application further provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement steps in a content topic classification method provided in an embodiment of the present application.

[0036] In addition, an embodiment of the present application further provides a computer program product including a computer program or instructions, which, when executed by a processor, implement steps in a content topic classification method provided in an embodiment of the present application.

[0037] A content topic classification method, device, electronic device, and storage medium are provided in an embodiment of the present application. The method can obtain to-be-classified content and identify a topic distribution type corresponding to historical publishing content of a publisher of the to-be-classified content. When the topic distribution type is a target type, attribute information corresponding to the publisher and topic information of the historical publishing content of the publisher are obtained. Based on content information of the to-be-classified content and the attribute information of the publisher and the topic information of the historical publishing content, probabilities that the to-be-classified content belongs to each preset topic are predicted. According to the probabilities, a target topic of the to-be-classified content is determined from the preset topics. When the topic distribution type of the historical publishing content of the publisher is the target type, the topic of the to-be-classified content can be predicted in combination with the attribute information of the publisher and the topic information of the historical publishing content, and the accuracy of content topic classification can be improved. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort.

[0039] Figure 1a is a scenario schematic diagram of a content topic classification method provided by an embodiment of the present application;

[0040] Figure 1b is a flowchart of a content topic classification method provided by an embodiment of the present application;

[0041] Figure 1c is a structure schematic diagram of a content topic classification method provided by an embodiment of the present application;

[0042] Figure 1d is a structure schematic diagram of a content topic classification method provided by an embodiment of the present application;

[0043] Figure 2 is another flowchart of a content topic classification method provided by an embodiment of the present application;

[0044] Figure 3 is a structure schematic diagram of a content topic classification device provided by an embodiment of the present application;

[0045] Figure 4 is a structure schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0046] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0047] The embodiments of the present application provide a content topic classification method, device, electronic device and storage medium. The content topic classification device can be integrated in an electronic device, which can be a terminal or a server, etc.

[0048] It can be understood that the content topic classification method of the present embodiment can be executed on a terminal, on a server, or by the terminal and the server together. The above examples should not be understood as limiting the present application.

[0049] As shown in Figure 1a , taking the content topic classification method executed by the terminal and the server together as an example. The content topic classification system provided by the embodiments of the present application includes a terminal 10 and a server 11, etc. The terminal 10 and the server 11 are connected through a network, such as a wired or wireless network connection, etc., wherein the content topic classification device can be integrated in the server.

[0050] The server 11 can be configured to acquire the content to be classified, and identify a topic distribution type corresponding to historical publishing content of a publisher of the content to be classified; when the topic distribution type is a target type, acquire attribute information corresponding to the publisher and topic information of the historical publishing content of the publisher; based on content information of the content to be classified and the attribute information of the publisher and the topic information of the historical publishing content, predict a probability that the content to be classified belongs to each preset topic; and determine a target topic of the content to be classified from the preset topics according to the probability. The server 11 can be a single server, or a server cluster or a cloud server composed of multiple servers.

[0051] The terminal 10 can be configured to send the content to be classified to the server 11, and receive a topic classification result of the content to be classified sent by the server 11. The terminal 10 can include a mobile phone, a smart television, a tablet computer, a notebook computer, a smart home, a wearable electronic device, an AR (Augmented Reality) / VR (Virtual Reality) device, a vehicle-mounted terminal or a personal computer (PC, Personal Computer), and the like. The terminal 10 can further be provided with a client, which can be an application program client or a browser client, and the like.

[0052] The server 11 can be configured to perform the step of topic classification.

[0053] The content topic classification method provided by the embodiment of the present application relates to natural language processing and machine learning in the field of artificial intelligence. The present application can combine the attribute information of the publisher and the topic information of the historical publishing content to predict the topic of the content to be classified, thereby improving the accuracy of content topic classification.

[0054] The artificial intelligence (AI, Artificial Intelligence) is to use a digital computer or a machine controlled by a digital computer to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use the knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science, which tries to understand the essence of intelligence and produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machine has the functions of perception, reasoning and decision-making. Artificial intelligence technology is a comprehensive discipline, which involves a wide range of fields, including hardware and software technologies. The artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning, automatic driving, intelligent transportation and other fields.

[0055] Among them, natural language processing (Nature Language processing, NLP) is an important direction in the field of computer science and artificial intelligence. It studies the various theories and methods that can realize the effective communication between people and computers with natural language. Natural language processing is a science that integrates linguistics, computer science and mathematics. Therefore, the research in this field will involve natural language, that is, the language used in daily life, so it is closely related to the study of linguistics. Natural language processing technology usually includes text processing, semantic understanding, machine translation, robot question and answer, knowledge graph and other technologies.

[0056] Among them, machine learning (Machine Learning, ML) is a multi-disciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It is a branch of computer science that studies how computers simulate or implement human learning behavior to acquire new knowledge or skills, reorganize existing knowledge structure to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent. Its application covers various fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and example-based learning.

[0057] The following will be described in detail. It should be noted that the order of the following embodiments is not limited as the preferred order of the embodiments.

[0058] This embodiment will be described from the perspective of a content theme classification device, which can be integrated in an electronic device, which can be a server or a terminal device.

[0059] As shown in Figure 1b The specific process of the content theme classification method can be as follows:

[0060] 101, obtain the content to be classified, and identify the theme distribution type corresponding to the historical publishing content of the publisher of the content to be classified.

[0061] Among them, the content to be classified can include video, image, audio, text, etc., which is not limited by the embodiment. By classifying the content to be classified, one or more theme labels corresponding to the content to be classified can be obtained. The theme label can represent the theme content information contained in the content to be classified. Specifically, the theme label can be a content element contained in the content to be classified, for example, the content to be classified contains "cat" and "dog", and its theme label can be "cat" and "dog"; the theme label can also be the feeling of the user given by the scene corresponding to the content to be classified, such as "thrilling" and "funny", etc.; this embodiment is not limited.

[0062] The publisher specifically refers to a user publishing the content to be classified. The historical publishing content of the publisher can be content published by the publisher before publishing the content to be classified.

[0063] Optionally, in the embodiment, the step of identifying a topic distribution type corresponding to the historical publishing content of the publisher of the content to be classified can include:

[0064] Based on the attribute information corresponding to the publisher of the content to be classified and the topic information of the historical publishing content of the publisher, performing topic distribution analysis on the historical publishing content to determine the topic distribution type corresponding to the historical publishing content.

[0065] The attribute information corresponding to the publisher specifically refers to the user personal portrait of the publisher, which can include the identity document (Identity document, identity information), biography, level, nickname, and the like of the publisher. The topic information of the historical publishing content can specifically be one or more topic labels of the historical publishing content.

[0066] In the embodiment, the topic distribution type indicates topic distribution information, which can include a topic aggregation distribution type, a topic dispersion distribution type, and the like. The topic aggregation distribution type represents that the topics of the content published by the publisher have stability, that is, the content published by the publisher is generally concentrated in a few fixed topics, and the topic information distribution is relatively concentrated. The topic dispersion distribution type represents that the stability of the topic information corresponding to the content published by the publisher is weak, that is, the distribution of the topic information of the content published by the publisher is relatively dispersed, and the published content does not have a relatively fixed topic.

[0067] In some embodiments, the topic information of the historical publishing content can be pre-calculated and stored in a preset topic information storage area. In addition, the topic label of the historical publishing content can also be displayed on the content page corresponding to the historical publishing content. The embodiment can directly obtain the historical calculated topic information from the preset topic information storage area.

[0068] Optionally, in other embodiments, the topic analysis can also be performed in real time on the content information of at least one dimension of the historical publishing content to obtain the topic information of the historical publishing content. Specifically, for example, the historical publishing content is a video, the content information of the video in multiple dimensions such as image modal, voice modal, and text modal can be obtained, feature extraction is performed on the content information to obtain feature information in each dimension, and then the feature information in each dimension is fused to obtain fused feature information corresponding to the video, and the topic label of the video is predicted based on the fused feature information.

[0069] The historical published content is subjected to theme distribution analysis. There are various methods for theme distribution analysis, and the embodiments are not limited in this regard.

[0070] For example, training data such as theme information of sample published content and a corresponding theme distribution type can be obtained, and a theme distribution identification model can be trained based on the training data. The theme distribution identification model can be used to analyze the theme distribution of content.

[0071] For another example, the theme information of historical published content can also be subjected to clustering processing, such as clustering processing of theme labels of historical published content, to obtain at least one clustering cluster of theme labels. According to the number of clustering clusters and the number of historical published content, the theme distribution type corresponding to the historical published content can be determined. Specifically, the theme distribution type corresponding to the historical published content can be determined according to the ratio of the number of clustering clusters to the number of historical published content. If the ratio is greater than a preset value, it can be determined that the theme distribution type of the historical published content is a theme scattered distribution type. If the ratio is not greater than the preset value, it can be determined that the theme distribution type of the historical published content is a theme concentrated distribution type. The preset value can be set according to actual conditions, and the embodiments are not limited in this regard.

[0072] Optionally, in the embodiments, the step of “subjecting the historical published content to theme distribution analysis based on attribute information corresponding to a publisher of the content to be classified and theme information of historical published content of the publisher, and determining a theme distribution type corresponding to the historical published content” can include:

[0073] Fusing the attribute information corresponding to the publisher of the content to be classified and the theme information of the historical published content of the publisher to obtain theme distribution information;

[0074] Extracting features from the theme distribution information to obtain distribution feature information of the theme distribution information;

[0075] According to the distribution feature information, predicting the theme distribution type corresponding to the historical published content.

[0076] The fusion method of attribute information corresponding to a publisher and theme information of historical published content can be various, and the embodiments are not limited in this regard. For example, the fusion method can be splicing processing. Specifically, splicing processing of attribute information and theme information can be splicing attribute information after theme information or splicing theme information after attribute information.

[0077] The feature extraction of theme distribution information can specifically be convolution processing and pooling processing of theme distribution information. Through feature extraction, distribution feature information corresponding to the theme distribution information can be obtained.

[0078] Optionally, in the embodiment, the topic distribution information includes at least one text unit;

[0079] The step of "extracting features of the topic distribution information to obtain distribution feature information of the topic distribution information" can include:

[0080] extracting word-level feature information of each text unit in the topic distribution information;

[0081] processing the word-level feature information of each text unit based on the word-level feature information of the text unit corresponding to the context of the text unit;

[0082] fusing the processed word-level feature information of each text unit to obtain the distribution feature information of the topic distribution information.

[0083] In the embodiment, the word-level feature information of each text unit in the topic distribution information can be a word vector corresponding to each text unit in the topic distribution information, or can be feature information obtained by fusing a content vector, a type vector, and a position vector of each text unit in the topic distribution information.

[0084] In the embodiment, the text unit corresponding to the context of the text unit can be other text units in the topic distribution information except the text unit. The embodiment can fuse the word-level feature information of each context text unit corresponding to the text unit to obtain context feature information of the text unit, and then process the word-level feature information of the text unit based on the context feature information. The way of fusing the processed word-level feature information of each text unit can be various, such as weighted summation, and the embodiment does not limit this.

[0085] Specifically, the word-level feature information of each text unit can be processed based on the word-level feature information of the text unit corresponding to the context of the text unit through a time sequence model. The time sequence model can be a long short-term memory (LSTM), a bidirectional encoder representation from transformers (Bert), or the like.

[0086] Wherein, LSTM is a kind of time recurrent neural network, that is, a kind of recurrent neural network (RNN, recurrent neural network). LSTM is more suitable for extracting semantic features from time series data, and is often used to extract semantic features from context information in natural language processing tasks. LSTM can selectively forget part of the historical data, add part of the current input data, and finally integrate into the current state and produce an output state through three gate structures (input gate, forget gate and output gate).

[0087] Wherein, Bert is an open-source time series model based on the transformer (Transformer) structure. Bert is connected by multiple layers of bidirectional Transformers, which can be 12 layers or 24 layers, etc. Bert can be obtained by pre-training and fine-tuning. Bert mainly includes two tasks during training. The first task is to randomly remove words in the training corpus and replace them with masks (mask) to let the model predict the removed words. The second task is that each training data is a sentence pair, wherein part of the training data is a true correlation between the upper and lower sentences, and part of the training data is unrelated between the upper and lower sentences, and the model needs to judge the relationship between the upper and lower sentences in the training data. The model is optimized based on the loss value of the two tasks. The training process of Bert can fully utilize the context information, so that the model has stronger expression ability. After completing the pre-training of the model, the model can be fine-tuned for specific tasks. Fine-tuning is a common transfer learning technique in the field of deep learning, which can make the model better adapt to language knowledge in specific scenarios.

[0088] Wherein, the step of "extracting word-level feature information of each text unit in the topic distribution information" can include:

[0089] Obtain the content vector and type vector corresponding to each text unit in the topic distribution information, wherein the type vector represents the information type to which the text unit belongs;

[0090] Encode the position information of each text unit in the topic distribution information to obtain the position vector corresponding to each text unit;

[0091] Fuse the content vector, the type vector and the position vector to obtain the word-level feature information of each text unit in the topic distribution information.

[0092] Wherein, the content vector corresponding to the text unit can be a word vector of the text unit. It should be noted that the text unit can be a word or a character, and the present embodiment does not limit this.

[0093] The type vector of the text unit can represent the information type to which the text unit belongs, and specifically can represent whether the text unit belongs to the attribute information of the publisher or the topic information of the historical publishing content. If the text unit belongs to the attribute information of the publisher, the type vector can further represent which specific attribute the text unit belongs to in the attribute information, such as whether the text unit is the id of the publisher or the level, the introduction, etc.

[0094] The position vector of the text unit can represent the position of the text unit in the topic distribution information, which can be the beginning of a sentence, the end of a sentence, etc.

[0095] The fusion manner of the content vector, the type vector, and the position vector can be various, and the embodiment is not limited thereto. For example, the fusion manner can be splicing processing, and the splicing order is not limited, for example, the splicing can be performed in the order of the content vector, the type vector, and the position vector, or in the reverse order, i.e., in the order of the position vector, the type vector, and the content vector. The fusion manner can also be weighted fusion, etc., in which the weights corresponding to the content vector, the type vector, and the position vector are determined first, and then the fusion is performed according to the weights.

[0096] In an embodiment, the topic distribution information is obtained by fusing the attribute information of the publisher and the topic information of the historical publishing content, and specifically can be obtained by fusing the id, the level, the introduction, and the topic label of the historical publishing content, etc. The topic distribution information is segmented to obtain each text unit of the topic distribution information, and each text unit can be spliced by the ‘[SEP]’ separator, in which the symbol at the beginning is ‘[CLS]’, the byte length is controlled to be 512, if the byte length is greater than 512, the truncation is performed, and if the byte length is less than 512, the symbol ‘[PAD]’ is supplemented. The word sequence after splicing is converted into a vector sequence. The vector sequence includes a vector corresponding to each word, and specifically can include the word-level feature information corresponding to each text unit in the above embodiment. The obtained vector sequence can be used as the input of the encoder.

[0097] The symbol ‘[PAD]’ is specifically a placeholder, the symbol ‘[CLS]’ can be set at the beginning of the topic distribution information, and the symbols ‘[CLS]’ and ‘[SEP]’ can affect the weight learning of the model.

[0098] Specifically, for each word, the vector corresponding to the word (i.e., the word-level feature information of the text unit) can be obtained by fusing the word vector, the position vector and the type vector corresponding to the word; a position vector and a type vector can be generated for the word vector of each word, and the position vector and the type vector are added to the word vector. The type vector is used to represent information about what type the position vector belongs to (the types can include id, level, introduction, and subject tags related to the published content, etc.).

[0099] Optionally, in the embodiment, after obtaining the vector sequence, i.e., obtaining the word-level feature information of each text unit in the topic distribution information, the word-level feature information can be input into an encoder. The encoder can be a Bert model, which uses a Transformer structure to process the original vector through multiple layers of attention operations to convert the original vector into a vector representation with abstract semantics (which can be the word-level feature information of the processed text unit described in the above embodiment). Each vector representation in the sequence contains both semantic information of the word and current context information.

[0100] Optionally, in the embodiment, the step of "predicting the topic distribution type corresponding to the historical published content according to the distribution feature information" can include:

[0101] According to the distribution feature information, predicting the probability that the historical published content belongs to each preset topic distribution type;

[0102] According to the probability, determining the topic distribution type corresponding to the historical published content.

[0103] Among them, the preset topic distribution type with the maximum probability can be taken as the topic distribution type corresponding to the historical published content.

[0104] Among them, the probability that the historical published content belongs to each preset topic distribution type can be predicted by a classifier (specifically, a classification decoding layer). The classifier can be a support vector machine (SVM), a recurrent neural network, or a fully connected deep neural network (DNN), etc. The present embodiment does not limit this.

[0105] The classification decoding layer is a linear layer, which can take the distribution feature information as input, output a score of a numerical type, and then convert the score into a probability value in the range of 0-1 through a sigmoid function. The probability value can represent the probability that the topic of the published content is stable. If the probability that the output topic distribution type of the historical published content belongs to the topic aggregation distribution type is greater than a threshold value, it is indicated that the topic of the content published by the publisher is relatively stable.

[0106] The sigmoid function, that is, the S-shaped growth curve, can be used as an activation function in a neural network or in a logistic regression process to map a variable to a value range of 0-1.

[0107] Optionally, in the embodiment, the step of “extracting features from the topic distribution information to obtain distribution feature information of the topic distribution information” can include:

[0108] The distribution feature information of the topic distribution information is extracted through the trained topic distribution identification model.

[0109] The step of “predicting the topic distribution type corresponding to the historical published content according to the distribution feature information” can include:

[0110] The topic distribution type corresponding to the historical published content is predicted according to the distribution feature information through the trained topic distribution identification model.

[0111] The topic distribution identification model can be a neural network, such as a bidirectional encoder representation (Bert, Bidirectional Encoder Representations from Transformers), ALBERT (A Lite BERT), a long short-term memory network (LSTM, Long Short-Term Memory), a residual network (ResNet, Residual Network), or a densely connected convolutional network (DenseNet, Dense Convolutional Network). It can be understood that the topic distribution identification model of the embodiment is not limited to the above-mentioned several types.

[0112] Specifically, the topic distribution identification model can be a binary classification model, and the model structure can be as follows: Figure 1cAs shown, the topic distribution information obtained by fusing the attribute information of the publisher and the topic information of the historical publishing content is taken as input, then the word-level feature information of each text unit in the topic distribution information is extracted through the feature extraction layer of the topic distribution identification model, the word-level feature information of each text unit in the topic distribution information is processed by adopting the Transformer structure of the encoder to obtain the distribution feature information of the topic distribution information, and then according to the distribution feature information, the topic distribution type corresponding to the historical publishing content is predicted to determine whether the topic of the content published by the publisher has stability.

[0113] It should be noted that the topic distribution identification model of the embodiment is trained by a plurality of training data; the topic distribution identification model can be trained by other devices and then provided to the content topic classification device, or the content topic classification device can also train itself.

[0114] If the content topic classification device trains itself, before the step of “extracting the distribution feature information of the topic distribution information through the trained topic distribution identification model”, the content topic classification method can further include:

[0115] obtaining training data, wherein the training data includes attribute information of a sample publisher, topic information of sample historical publishing content of the sample publisher, and an expected topic distribution type of the sample historical publishing content;

[0116] fusing the attribute information of the sample publisher and the topic information of the sample historical publishing content to obtain sample topic distribution information;

[0117] extracting the distribution feature information of the sample topic distribution information through a topic distribution identification model;

[0118] predicting an actual topic distribution type of the sample historical publishing content according to the distribution feature information;

[0119] adjusting parameters of the topic distribution identification model according to the actual topic distribution type and the expected topic distribution type to obtain a trained topic distribution identification model.

[0120] Among them, the attribute information of the publisher can be pulled, including: publisher creator id, introduction, level, portrait and other information; and the information of the historical publishing content is pulled, including: the topic label of the content, the publisher of the content and other information. According to the publisher, the historical publishing content is grouped, the distribution of all topic labels corresponding to each publisher is counted, and these information is saved in the database. When constructing the training data, the corresponding data can be directly pulled from the database.

[0121] The step of "adjusting parameters of the topic distribution identification model according to the actual topic distribution type and the expected topic distribution type to obtain the trained topic distribution identification model" can include:

[0122] calculating a loss value corresponding to the topic distribution type prediction according to the predicted probability corresponding to the actual topic distribution type and the expected topic distribution type;

[0123] adjusting the parameters of the topic distribution identification model according to the loss value to obtain the trained topic distribution identification model.

[0124] The training process can adjust the parameters of the topic distribution identification model based on a back propagation algorithm, and optimize the parameters of the topic distribution identification model based on the predicted probability corresponding to the actual topic distribution type and the expected topic distribution type, so that the loss value corresponding to the topic distribution type prediction is less than a preset loss value. The preset loss value can be set according to actual conditions.

[0125] The loss value can be calculated using a cross-entropy loss function or a mean square error loss function, and the present embodiment does not limit this.

[0126] Optionally, in the present embodiment, the topic distribution type of the historical publishing content of the publisher can be pre-calculated and saved in a preset storage area. When the content to be classified needs to be classified, the topic distribution type of the historical publishing content of the publisher can be obtained from the preset storage area according to the identification information (such as id) corresponding to the publisher of the content to be classified. It can be understood that, since the content published by the publisher is constantly updated, the topic distribution type saved in the preset storage area needs to be updated regularly. Specifically, in some embodiments, the topic distribution identification model needs to be trained regularly using new training data.

[0127] 102, when the topic distribution type is the target type, obtaining attribute information corresponding to the publisher and topic information of the historical publishing content of the publisher.

[0128] The target type can be a topic aggregation distribution type. When the topic distribution type of the historical publishing content of the publisher corresponding to the content to be classified is a topic aggregation distribution type, the present embodiment can predict the topic of the content to be classified according to the content information of the content to be classified itself, and the topic information of the historical publishing content and the attribute information of the publisher. This is because the topic distribution of the content published by the publisher is a topic aggregation distribution type, which is stable, so the historical content published by the publisher and the attribute information of the publisher have a reference effect on the topic prediction of the content to be classified.

[0129] 103. predicting, based on the content information of the content to be classified and the attribute information of the publisher and the topic information of the historical published content, a probability that the content to be classified belongs to each preset topic.

[0130] The content information of the content to be classified can include content information of each dimension thereof. For example, if the content to be classified is a video, the content information can include content information of multiple dimensions such as an image modality, a voice modality, and a text modality. Specifically, for the content information of the image modality, the text information can be converted from the text in the image through OCR (Optical Character Recognition); for the content information of the voice modality, the voice information can be converted into text information through ASR (Automated Speech Recognition) technology. In some embodiments, the content information (such as a video picture) under the image modality can also be subjected to image recognition to obtain object description information of an object carried by the video picture.

[0131] Optionally, in the embodiment, the step of predicting, based on the content information of the content to be classified and the attribute information of the publisher and the topic information of the historical published content, a probability that the content to be classified belongs to each preset topic can include:

[0132] fusing the content information of the content to be classified, the attribute information of the publisher, and the topic information of the historical published content to obtain fused information;

[0133] extracting features of the fused information to obtain feature information of the fused information;

[0134] predicting, according to the feature information, a probability that the content to be classified belongs to each preset topic.

[0135] The content information of the content to be classified, the attribute information of the publisher, and the topic information of the historical published content can be fused in various ways, such as splicing processing, and the splicing order is not limited and can be set according to actual conditions. For example, the three can be spliced in the order of content information, attribute information, and topic information to obtain the fused information.

[0136] Optionally, in the embodiment, the fused information includes at least one text unit, and the step of extracting features of the fused information to obtain feature information of the fused information can include:

[0137] extracting word-level feature information of each text unit in the fused information;

[0138] The word-level feature information of each text unit is processed based on the word-level feature information of the context text unit corresponding to each text unit.

[0139] The word-level feature information of each processed text unit is fused to obtain feature information of the fused information.

[0140] The word-level feature information of each text unit in the fused information can be a word vector corresponding to each text unit in the fused information, or can be feature information obtained by fusing a content vector, a type vector, and a position vector of each text unit in the fused information, and the present embodiment does not limit this.

[0141] The context text unit corresponding to each text unit can be other text units in the fused information except the text unit. The present embodiment can fuse the word-level feature information of each context text unit corresponding to each text unit to obtain context feature information corresponding to the text unit, and then process the word-level feature information of the text unit based on the context feature information. There are various ways to fuse the word-level feature information of each processed text unit, such as weighted summation, and the present embodiment does not limit this.

[0142] Specifically, the word-level feature information of each text unit can be processed based on the word-level feature information of the context text unit corresponding to each text unit through a time sequence model. The time sequence model can be a long short-term memory (LSTM, Long Short-Term Memory), a bidirectional encoder representation (Bert, Bidirectional Encoder Representations from Transformers), or the like.

[0143] Optionally, in the present embodiment, the step of "extracting the word-level feature information of each text unit in the fused information" can include:

[0144] Obtaining a content vector and a type vector corresponding to each text unit in the fused information, the type vector representing an information type to which the text unit belongs;

[0145] Encoding the position information of each text unit in the fused information to obtain a position vector corresponding to each text unit;

[0146] Fusing the content vector, the type vector, and the position vector to obtain the word-level feature information of each text unit in the fused information.

[0147] The content vector corresponding to the text unit can be a word vector of the text unit. It should be noted that the text unit can be a word or a character, and the embodiment is not limited in this regard.

[0148] The type vector of the text unit can represent the information type to which the text unit belongs, that is, it can specifically represent whether the text unit belongs to content information of the content to be classified, or belongs to attribute information of the publisher, or belongs to topic information of the historical published content. If the text unit belongs to the attribute information of the publisher, the type vector can also represent which specific attribute the text unit belongs to in the attribute information, such as whether the text unit is the id of the publisher, or the level, the introduction, etc. If the text unit belongs to the content information of the content to be classified, the type vector can also represent which specific content the text unit belongs to in the content information of the content to be classified, such as the text unit can be a video title, a subtitle, etc.

[0149] The position vector of the text unit can represent the position of the text unit in the fused information, which can be specifically the beginning of a sentence, the end of a sentence, etc.

[0150] The fusion manner of the content vector, the type vector, and the position vector can be various, and the embodiment is not limited in this regard. For example, the fusion manner can be splicing processing, and the splicing order is not limited, for example, the splicing can be performed in the order of the content vector, the type vector, and the position vector, or in the reverse order, that is, in the order of the position vector, the type vector, and the content vector. The fusion manner can also be weighted fusion, and the weights corresponding to the content vector, the type vector, and the position vector are determined first, and then the fusion is performed according to the weights.

[0151] In a specific scenario, the content to be classified is video content, and the content information of the content to be classified can be the content information of each dimension of the video, which can specifically include a video title, and text information obtained by OCR or ASR, etc. The embodiment can perform word segmentation on the content information of the content to be classified, and perform word segmentation on the attribute information of the publisher and the topic information of the historical published content. Specifically, it can also be regarded as performing word segmentation processing on the fused information to obtain each text unit of the fused information. Then, the text units are spliced with a ‘[SEP]’ separator, the symbol at the beginning is ‘[CLS]’, the byte length is controlled to be 512, if the byte length is greater than 512, it is truncated, and if the byte length is less than 512, the symbol ‘[PAD]’ is supplemented, and the spliced word sequence is converted into a vector sequence. The vector sequence includes a vector corresponding to each word, which can specifically include the word-level feature information corresponding to each text unit described in the above embodiment.

[0152] Specifically, for each word, the vector corresponding to the word (i.e., the word-level feature information of the text unit) can be obtained by fusing the word vector, the position vector and the type vector corresponding to the word; a position vector and a type vector can be generated for the word vector of each word, and the position vector and the type vector are added to the word vector. The type vector is used to represent information about what type (types can include video titles, text information obtained through OCR or ASR, etc.) the position vector belongs to.

[0153] Optionally, in the embodiment, after obtaining the vector sequence, i.e., obtaining the word-level feature information of each text unit in the fused information, the word-level feature information can be input into an encoder. The encoder can be a BERT model. The transformer structure is used to convert the original vector into a vector representation with abstract semantics (which can be the word-level feature information of the processed text unit described in the above embodiment). The transformer multi-head self-attention (self-attention) mechanism enables each word vector to obtain global information from multiple angles, improving the expression ability of the vector. Through multi-layer self-attention operation, each word vector can more deeply understand the semantic information of the entire sentence.

[0154] Optionally, in the embodiment, the step of "predicting the probability that the content to be classified belongs to each preset theme according to the feature information" can include:

[0155] calculating the similarity between the feature information and the theme feature information corresponding to each preset theme;

[0156] determining the probability that the content to be classified belongs to each preset theme according to the similarity.

[0157] The similarity between the feature information and the theme feature information corresponding to each preset theme can be calculated, specifically, the vector distance between the feature vector corresponding to the feature information and the theme feature vector corresponding to each preset theme can be calculated, and the vector distance can represent the similarity between them. The greater the vector distance, the lower the similarity, and the smaller the probability that the content to be classified belongs to the preset theme; on the contrary, the smaller the vector distance, the higher the similarity, and the greater the probability that the content to be classified belongs to the preset theme. The calculation method of the vector distance can be various, and the embodiment does not limit it. For example, the vector distance can be a cosine distance, or an Euclidean distance, etc.

[0158] The classifier can predict the probability that the content to be classified belongs to each preset theme according to the feature information. The classifier can be a support vector machine (SVM), a recurrent neural network, a fully connected deep neural network (DNN), or the like. The present embodiment is not limited in this regard.

[0159] The classifier can take the feature information of the fused information as input, obtain the theme feature information of each preset theme, and perform similarity calculation on the input feature information and the theme feature information of all preset themes. Specifically, the input feature information can be point multiplied with the theme feature information of each preset theme. At this time, each preset theme can obtain a similarity score, and then a probability value in the interval of 0-1 can be obtained through a sigmoid function. The probability value represents the probability that the content to be classified belongs to the preset theme. After obtaining the probability values corresponding to all preset themes, the preset theme with a probability value greater than a preset threshold can be selected as the target theme of the content to be classified.

[0160] The step of "extracting features of the fused information to obtain feature information of the fused information" can include:

[0161] The feature information of the fused information is extracted through the trained theme recognition model.

[0162] The step of "predicting the probability that the content to be classified belongs to each preset theme according to the feature information" can include:

[0163] The probability that the content to be classified belongs to each preset theme is predicted according to the feature information through the trained theme recognition model.

[0164] The theme recognition model can be a neural network, such as a bidirectional encoder representation (Bert), ALBERT (A Lite BERT), a long short-term memory network (LSTM), a residual network (ResNet), or a dense connection convolutional network (DenseNet). It can be understood that the theme recognition model of the present embodiment is not limited to the above-mentioned types.

[0165] It should be noted that the subject recognition model of the embodiment is trained by a plurality of training data; the subject recognition model can be trained by other devices and provided to the content subject classification device, or the content subject classification device can also train itself.

[0166] If the content subject classification device trains itself, before the step of "extracting features of the fused information by the trained subject recognition model to obtain feature information of the fused information", the content subject classification method can further include:

[0167] Obtaining training data, the training data including content information of a sample content to be classified, attribute information of a sample publisher corresponding to the sample content to be classified, subject information of sample historical publishing content of the sample publisher, and expected probability of the sample content to be classified belonging to each preset subject;

[0168] Fusing the content information of the sample content to be classified, the attribute information of the sample publisher, and the subject information of the sample historical publishing content to obtain sample fused information;

[0169] Extracting features of the sample fused information by a subject recognition model to obtain feature information of the sample fused information;

[0170] Predicting actual probability of the sample content to be classified belonging to each preset subject according to the feature information;

[0171] Adjusting parameters of the subject recognition model based on the actual probability of the sample content to be classified belonging to each preset subject and the corresponding expected probability to obtain a trained subject recognition model.

[0172] The training process can adjust the parameters of the subject recognition model based on a back propagation algorithm, specifically, the parameters of the subject recognition model can be optimized based on the actual probability of the sample content to be classified belonging to each preset subject and the corresponding expected probability, so that the actual probability of the sample content to be classified belonging to each preset subject approaches the corresponding expected probability, specifically, the loss value between the actual probability of the sample content to be classified belonging to each preset subject and the corresponding expected probability can be less than a preset loss value, which can be set according to actual conditions.

[0173] The loss value can be calculated by using a cross-entropy loss function or a mean square error loss function, and the embodiment does not limit this.

[0174] In the training process of the subject identification model, the expected probability of the content of the sample to be classified belonging to each preset subject, i.e., the label of the training data, can be represented by a mulit-hot vector. Specifically, if the content of the sample to be classified contains a preset subject, the expected probability corresponding to the preset subject can be set to 1, otherwise, the expected probability can be set to 0.

[0175] 104. determining the target subject of the content to be classified from the preset subjects according to the probabilities.

[0176] In some embodiments, the preset subjects can be sorted according to the size of the probabilities, for example, from large to small, to obtain sorted preset subjects, and the first n preset subjects in the sorted preset subjects can be determined as the target subject of the content to be classified. In other embodiments, the preset subjects with probabilities greater than a preset threshold can be determined as the target subject of the content to be classified, and the preset threshold can be set according to actual conditions.

[0177] Optionally, in the embodiment, the content subject classification method can further include:

[0178] when the subject distribution type is a non-target type, predicting the probability of the content to be classified belonging to each preset subject based on the content information of the content to be classified;

[0179] determining the target subject of the content to be classified from the preset subjects according to the probabilities.

[0180] In the embodiment, the subject distribution type is a non-target type, specifically, the subject distribution type is not a subject clustering distribution type. When the subject distribution type of the historical published content of the publisher corresponding to the content to be classified is not a subject clustering distribution type, it indicates that the content published by the publisher does not have stability, that is, the historical published content and attribute information of the publisher do not have reference for the subject prediction of the content to be classified, and then the target subject of the content to be classified can be directly predicted based on the content information of the content to be classified itself.

[0181] In the embodiment, the step of predicting the probability of the content to be classified belonging to each preset subject based on the content information of the content to be classified can include:

[0182] extracting features from the content information of the content to be classified to obtain feature information of the content to be classified;

[0183] predicting the probability of the content to be classified belonging to each preset subject according to the feature information.

[0184] The specific steps of feature extraction of the content information of the content to be classified can refer to the steps of feature extraction of the fused information in the above-mentioned embodiments, and will not be described here.

[0185] The content information of the content to be classified can include content information of each dimension thereof. For example, if the content to be classified is a video, the content information thereof can include content information of multiple dimensions such as an image modality, a voice modality, and a text modality. Specifically, for the content information of the image modality, the text information can be obtained by converting the text in the image into text information through OCR (Optical Character Recognition). For the content information of the voice modality, the voice information can be converted into text information through ASR (Automated Speech Recognition) technology. In some embodiments, the content information (such as a video picture) in the image modality can also be subjected to image recognition to obtain object description information of an object carried by the video picture.

[0186] Optionally, in the embodiment, the content information of the content to be classified includes at least one text unit. The step of "extracting feature information of the content information of the content to be classified" can include:

[0187] extracting word-level feature information of each text unit in the content information;

[0188] processing the word-level feature information of each text unit based on the context of the text unit corresponding to each text unit;

[0189] fusing the processed word-level feature information of each text unit to obtain the feature information of the content information.

[0190] The word-level feature information of each text unit in the content information can be a word vector corresponding to each text unit in the content information, or can be feature information obtained by fusing a content vector, a type vector, and a position vector of each text unit in the content information, and the embodiment does not limit this.

[0191] In a specific scenario, the content to be classified is video content, and the content information of the content to be classified can be content information of each dimension of the video, which can specifically include a video title, text information obtained through OCR and ASR, and the like. The embodiment can perform word segmentation processing on the content information of the content to be classified to obtain each text unit of the content information of the content to be classified, and then splice each text unit with a '[SEP]' separator, where the symbol at the beginning is '[CLS]', the byte length is controlled to be 512, and if the byte length is greater than 512, it is truncated, and if the byte length is less than 512, it is supplemented with a '[PAD]' symbol. The word sequence after splicing is converted into a vector sequence. The vector sequence includes a vector corresponding to each word, which can specifically include the word-level feature information corresponding to each text unit in the content information described in the above embodiment.

[0192] Specifically, for each word, the vector (which can be word-level feature information) corresponding to the word (i.e., the text unit) can be obtained by fusing the word vector corresponding to the word, the position vector, and the type vector. A position vector and a type vector can be generated for the word vector of each word, and the position vector and the type vector are added to the word vector. The type vector is used to represent information about what type of vector (types can include video titles, text information obtained through OCR or ASR, and the like) the position vector is.

[0193] Optionally, in the embodiment, when the topic distribution type is a non-target type, the topic of the content to be classified can also be predicted through the topic recognition model in step 103.

[0194] In a specific scenario, the topic of the content to be classified can be predicted using a topic recognition model as shown in Figure 1d If the topic distribution type of the historical publishing content of the publisher corresponding to the content to be classified is a non-target type, the content information of the content to be classified is used as the input information of the topic recognition model. If the topic distribution type of the historical publishing content of the publisher corresponding to the content to be classified is a target type, the fused information of the content information of the content to be classified, the attribute information of the publisher, and the topic information of the historical publishing content is used as the input information of the topic recognition model.

[0195] Specifically, the input information is subjected to feature extraction by a feature extraction layer of the topic recognition model to obtain word-level feature information of each text unit in the input information, and then the word-level feature information of each text unit is processed by an encoder adopting a Transformer structure to obtain feature information corresponding to the input information. Then, according to the feature information and topic feature information of each label (i.e., a preset topic), the probability of the content to be classified belonging to each preset topic can be predicted. For example, the feature information corresponding to the input information and the topic feature information of each preset topic can be dot multiplied, at which time each preset topic can obtain a similarity score, and then a probability value in the interval of 0-1 is obtained through a sigmoid function, which represents the probability of the content to be classified belonging to the preset topic. After obtaining the probability values corresponding to all preset topics, a preset topic with a probability value greater than a preset threshold can be selected as the target topic of the content to be classified.

[0196] The application can train the topic distribution recognition model by using the attribute information of the publisher and the topic information of the historical publishing content, and divide the topic distribution type of the historical publishing content corresponding to the publisher into two types, i.e., a topic concentrated distribution type and a topic dispersed distribution type. In the case that the topic distribution type of the historical publishing content corresponding to the publisher of the content to be classified is the topic concentrated distribution type, the topic recognition model can interact the personal portrait information (i.e., the attribute information) of the publisher, the historical publishing content and the content information of the content to be classified, fully utilize the memory capacity of the neural network and the historical publishing content of the publisher whose topic of the publishing content is relatively stable, and significantly improve the accuracy of the topic classification of the content to be classified.

[0197] The content topic classification method provided by the application can be applied to multiple scenes, such as a scene of video topic classification. Specifically, users can select videos they want to watch from a large number of videos based on topic classification labels, and operators can manage videos according to topic classification labels. The topic classification label recognition system can generally be used in the video standardization process, i.e., after a video uploader uploads a video, the video platform describes the video with topic classification labels, which can reduce the labeling workload of labeling personnel on the one hand, and can also cluster videos through topic classification labels to improve the distribution quality of the video platform. When a user watches a video, the video platform can recommend related videos to the user according to the topic classification labels to improve the user's viewing experience. In addition, when a user searches for a video, high-quality topic classification labels can also help the user accurately obtain related videos.

[0198] From the above, the embodiment can acquire the content to be classified, and identify the historical publishing content corresponding to the topic distribution type of the publisher of the content to be classified; when the topic distribution type is a target type, acquire the attribute information corresponding to the publisher and the topic information of the historical publishing content of the publisher; based on the content information of the content to be classified, and the attribute information of the publisher and the topic information of the historical publishing content, predict the probability that the content to be classified belongs to each preset topic; and according to the probability, determine the target topic of the content to be classified from the preset topics. When the topic distribution type of the historical publishing content of the publisher is a target type, the application can combine the attribute information of the publisher and the topic information of the historical publishing content to predict the topic of the content to be classified, and can improve the accuracy of content topic classification.

[0199] According to the method described in the foregoing embodiment, the following will be further described in detail with the example that the content topic classification device is specifically integrated in a server.

[0200] The embodiment of the application provides a content topic classification method, as shown in the figure, and the specific process of the content topic classification method can be as follows: Figure 2

[0201] 201. The server acquires content to be classified.

[0202] The content to be classified can include video, image, audio, text, etc., and the embodiment is not limited thereto. Classification of the content to be classified can obtain one or more topic labels corresponding thereto. The topic label can represent the topic content information contained in the content to be classified.

[0203] 202. The server performs topic distribution analysis on the historical publishing content based on the attribute information corresponding to the publisher of the content to be classified and the topic information of the historical publishing content of the publisher, and determines the topic distribution type corresponding to the historical publishing content.

[0204] The attribute information corresponding to the publisher is specifically the user personal portrait of the publisher, which can include the identity document (Identity document, identity information), biography, level, nickname, etc. of the publisher. The topic information of the historical publishing content can be one or more topic labels of the historical publishing content.

[0205] ​In this embodiment, the subject distribution type indicates subject distribution information, which can include a subject concentrated distribution type, a subject dispersed distribution type, and the like. The subject concentrated distribution type indicates that the subject of the content published by the publisher is stable, that is, the content published by the publisher is generally concentrated in several fixed subjects, and the subject information distribution is relatively concentrated. The subject dispersed distribution type indicates that the stability of the subject information corresponding to the content published by the publisher is weak, that is, the distribution of the subject information of the content published by the publisher is relatively dispersed, and the published content does not have a relatively fixed subject.

[0206] Optionally, in this embodiment, the step of "performing subject distribution analysis on the historical publishing content of the publisher based on the attribute information corresponding to the publisher of the content to be classified and the subject information of the historical publishing content of the publisher, and determining the subject distribution type corresponding to the historical publishing content" can include:

[0207] fusing the attribute information corresponding to the publisher of the content to be classified and the subject information of the historical publishing content of the publisher to obtain subject distribution information;

[0208] performing feature extraction on the subject distribution information to obtain distribution feature information of the subject distribution information;

[0209] predicting the subject distribution type corresponding to the historical publishing content according to the distribution feature information.

[0210] The fusion manner of the attribute information corresponding to the publisher and the subject information of the historical publishing content can be various, and this embodiment does not limit the fusion manner. For example, the fusion manner can be splicing processing.

[0211] Optionally, in this embodiment, the subject distribution information includes at least one text unit.

[0212] The step of "performing feature extraction on the subject distribution information to obtain distribution feature information of the subject distribution information" can include:

[0213] extracting word-level feature information of each text unit in the subject distribution information;

[0214] processing the word-level feature information of each text unit based on the word-level feature information of the context text unit corresponding to each text unit;

[0215] fusing the processed word-level feature information of each text unit to obtain the distribution feature information of the subject distribution information.

[0216] 203. The server determines whether the subject distribution type is a target type. If yes, step 204 is performed; if no, step 206 is performed.

[0217] The target type can be specifically a topic aggregation distribution type.

[0218] 204. The server acquires attribute information of the publisher and topic information of historical publishing content of the publisher, and proceeds to step 205.

[0219] When the topic distribution type of the historical publishing content of the publisher corresponding to the content to be classified is the topic aggregation distribution type, the embodiment can predict the topic of the content to be classified according to the content information of the content to be classified, and the topic information of the historical publishing content and the attribute information of the publisher. This is because the topic distribution of the content published by the publisher is the topic aggregation distribution type, which has stability, and therefore the historical content published by the publisher and the attribute information of the publisher have a reference effect on the topic prediction of the content to be classified.

[0220] 205. The server predicts a probability that the content to be classified belongs to each preset topic based on the content information of the content to be classified, and the attribute information of the publisher and the topic information of the historical publishing content; determines a target topic of the content to be classified from the preset topics according to the probability, and ends the process.

[0221] The content information of the content to be classified can include content information of each dimension thereof. For example, if the content to be classified is a video, the content information can include content information of multiple dimensions such as image modalities, voice modalities, and text modalities.

[0222] Optionally, in the embodiment, the step of predicting the probability that the content to be classified belongs to each preset topic based on the content information of the content to be classified, and the attribute information of the publisher and the topic information of the historical publishing content can include:

[0223] fusing the content information of the content to be classified, and the attribute information of the publisher and the topic information of the historical publishing content to obtain fused information;

[0224] performing feature extraction on the fused information to obtain feature information of the fused information;

[0225] predicting the probability that the content to be classified belongs to each preset topic according to the feature information.

[0226] The content information of the content to be classified, the attribute information of the publisher, and the topic information of the historical publishing content are fused in various ways, such as splicing processing, and the splicing order is not limited and can be set according to actual conditions.

[0227] In some embodiments, the preset topics can be sorted according to the size of the probability, for example, from large to small, to obtain sorted preset topics, and the first n preset topics in the sorted preset topics are determined as the target topic of the content to be classified. In other embodiments, the preset topic with a probability greater than a preset threshold can also be determined as the target topic of the content to be classified, and the preset threshold can be set according to actual conditions.

[0228] 206. The server predicts a probability that the content to be classified belongs to each preset topic based on the content information of the content to be classified; determines a target topic of the content to be classified from the preset topics according to the probability, and ends the process.

[0229] In the case where the topic distribution type is a non-target type, that is, not belonging to the topic aggregation distribution type. When the topic distribution type of the historical published content of the publisher corresponding to the content to be classified does not belong to the topic aggregation distribution type, it indicates that the content published by the publisher does not have stability, that is, the historical published content and attribute information of the publisher have no reference for the topic prediction of the content to be classified. Therefore, the target topic of the content to be classified can be directly predicted based on the content information of the content to be classified itself.

[0230] The step of "predicting a probability that the content to be classified belongs to each preset topic based on the content information of the content to be classified" can include:

[0231] performing feature extraction on the content information of the content to be classified to obtain feature information of the content to be classified;

[0232] predicting a probability that the content to be classified belongs to each preset topic based on the feature information.

[0233] The specific steps of feature extraction on the content information of the content to be classified can refer to the steps of feature extraction on the fused information in the above embodiments, which will not be described here.

[0234] The content information of the content to be classified can include content information of each dimension thereof. For example, if the content to be classified is a video, the content information thereof can include content information of multiple dimensions such as image modalities, voice modalities, and text modalities.

[0235] Optionally, in the present embodiment, the content information of the content to be classified includes at least one text unit, and the step of "performing feature extraction on the content information of the content to be classified to obtain feature information of the content to be classified" can include:

[0236] extracting word-level feature information of each text unit in the content information;

[0237] The word-level feature information of each text unit is processed based on the word-level feature information of each text unit corresponding to the context of the text unit.

[0238] The word-level feature information of each processed text unit is fused to obtain the feature information of the content information.

[0239] As can be seen from the above, in the embodiment, the server can be used to obtain the content to be classified; the historical publishing content of the publisher of the content to be classified is analyzed based on the attribute information corresponding to the publisher of the content to be classified and the theme information of the historical publishing content, to determine the theme distribution type corresponding to the historical publishing content; it is determined whether the theme distribution type is a target type. If yes, the attribute information corresponding to the publisher and the theme information of the historical publishing content of the publisher are obtained; the probability that the content to be classified belongs to each preset theme is predicted based on the content information of the content to be classified, and the attribute information of the publisher and the theme information of the historical publishing content; according to the probability, the target theme of the content to be classified is determined from the preset theme, and the process is ended. If no, the probability that the content to be classified belongs to each preset theme is predicted based on the content information of the content to be classified; according to the probability, the target theme of the content to be classified is determined from the preset theme, and the process is ended. When the theme distribution type of the historical publishing content of the publisher is the target type, the attribute information of the publisher and the theme information of the historical publishing content can be combined to predict the theme of the content to be classified, and the accuracy of the content theme classification can be improved.

[0240] In order to better implement the above method, the embodiment of the application further provides a content theme classification device, as shown in Figure 3 The content theme classification device can include an identification unit 301, an acquisition unit 302, a prediction unit 303, and a determination unit 304, as follows:

[0241] (1) The identification unit 301;

[0242] The identification unit is configured to obtain the content to be classified, and identify the theme distribution type corresponding to the historical publishing content of the publisher of the content to be classified.

[0243] Optionally, in some embodiments of the application, the identification unit can be specifically configured to analyze the theme distribution of the historical publishing content based on the attribute information corresponding to the publisher of the content to be classified and the theme information of the historical publishing content of the publisher, to determine the theme distribution type corresponding to the historical publishing content.

[0244] Optionally, in some embodiments of the application, the identification unit can include a first fusion subunit, a first extraction subunit, and a first prediction subunit, as follows:

[0245] The first fusion sub-unit is configured to fuse attribute information corresponding to a publisher of the content to be classified and topic information of historical publishing content of the publisher, to obtain topic distribution information.

[0246] The first extraction sub-unit is configured to extract features of the topic distribution information, to obtain distribution feature information of the topic distribution information.

[0247] The first prediction sub-unit is configured to predict a topic distribution type corresponding to the historical publishing content according to the distribution feature information.

[0248] Optionally, in some embodiments of the present application, the topic distribution information includes at least one text unit.

[0249] The first extraction sub-unit can be specifically configured to extract word-level feature information of each text unit in the topic distribution information, process the word-level feature information of each text unit based on the word-level feature information of the context of the text unit corresponding to each text unit, and fuse the processed word-level feature information of each text unit, to obtain the distribution feature information of the topic distribution information.

[0250] Optionally, in some embodiments of the present application, the first extraction sub-unit can be specifically configured to extract features of the topic distribution information by using a trained topic distribution identification model, to obtain the distribution feature information of the topic distribution information.

[0251] Optionally, in some embodiments of the present application, the content topic classification apparatus can further include a training unit configured to train a topic distribution identification model. Specifically, the training unit is configured to obtain training data, the training data including attribute information of a sample publisher, topic information of sample historical publishing content of the sample publisher, and an expected topic distribution type of the sample historical publishing content, fuse the attribute information of the sample publisher and the topic information of the sample historical publishing content, to obtain sample topic distribution information, extract features of the sample topic distribution information by using the topic distribution identification model, to obtain distribution feature information of the sample topic distribution information, predict an actual topic distribution type of the sample historical publishing content according to the distribution feature information, and adjust parameters of the topic distribution identification model according to the actual topic distribution type and the expected topic distribution type, to obtain a trained topic distribution identification model.

[0252] (2) the obtaining unit 302;

[0253] The obtaining unit is configured to obtain attribute information corresponding to the publisher and topic information of historical publishing content of the publisher when the topic distribution type is a target type.

[0254] the prediction unit 303;

[0255] predict, based on the content information of the content to be classified, and the attribute information of the publisher and the topic information of the historical published content, a probability that the content to be classified belongs to each preset topic.

[0256] Optionally, in some embodiments of the present application, the prediction unit can include a second fusion subunit, a second extraction subunit and a second prediction subunit, as follows:

[0257] The second fusion subunit is configured to fuse the content information of the content to be classified, and the attribute information of the publisher and the topic information of the historical published content, to obtain fused information.

[0258] The second extraction subunit is configured to extract features of the fused information, to obtain feature information of the fused information.

[0259] The second prediction subunit is configured to predict, based on the feature information, a probability that the content to be classified belongs to each preset topic.

[0260] Optionally, in some embodiments of the present application, the fused information includes at least one text unit; the second extraction subunit can be specifically configured to extract word-level feature information of each text unit in the fused information; process the word-level feature information of each text unit based on the word-level feature information of the text unit corresponding to the context of each text unit; and fuse the processed word-level feature information of each text unit, to obtain the feature information of the fused information.

[0261] Optionally, in some embodiments of the present application, the step of “extracting word-level feature information of each text unit in the fused information” can include:

[0262] obtaining a content vector and a type vector corresponding to each text unit in the fused information, the type vector representing an information type to which the text unit belongs;

[0263] encoding position information of each text unit in the fused information, to obtain a position vector corresponding to each text unit;

[0264] fusing the content vector, the type vector and the position vector, to obtain the word-level feature information of each text unit in the fused information.

[0265] the determination unit 304;

[0266] A determining unit is used to determine the target topic of the content to be classified from the preset topics based on the probability.

[0267] Optionally, in some embodiments of this application, the prediction unit may also be used to predict the probability that the content to be classified belongs to each preset topic based on the content information of the content to be classified when the topic distribution type is a non-target type; the determination unit may also be used to determine the target topic of the content to be classified from the preset topics according to the probability.

[0268] As can be seen from the above, this embodiment can obtain the content to be classified through the identification unit 301 and identify the topic distribution type corresponding to the publisher's historical published content of the content to be classified; when the topic distribution type is the target type, the acquisition unit 302 obtains the attribute information corresponding to the publisher and the topic information of the publisher's historical published content; the prediction unit 303 predicts the probability that the content to be classified belongs to each preset topic based on the content information of the content to be classified, the attribute information of the publisher, and the topic information of the historical published content; the determination unit 304 determines the target topic of the content to be classified from the preset topics according to the probability. When the topic distribution type of the publisher's historical published content is the target type, this application can combine the publisher's attribute information and the topic information of the historical published content to predict the topic of the content to be classified, which can improve the accuracy of content topic classification.

[0269] This application also provides an electronic device, such as... Figure 4 The diagram shows a structural schematic of an electronic device involved in an embodiment of this application. This electronic device can be a terminal or a server, specifically:

[0270] The electronic device may include components such as a processor 401 with one or more processing cores, a memory 402 with one or more computer-readable storage media, a power supply 403, and an input unit 404. Those skilled in the art will understand that... Figure 4 The electronic device structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein:

[0271] The processor 401 is the control center of the electronic device, connects each part of the entire electronic device by various interfaces and lines, executes various functions of the electronic device and processes data by running or executing software programs and / or modules stored in the memory 402 and calling data stored in the memory 402, thereby overall detecting the electronic device. Optionally, the processor 401 can include one or more processing cores; preferably, the processor 401 can integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface and application program, etc., and the modem processor mainly processes wireless communication. It can be understood that the above-mentioned modem processor can also not be integrated into the processor 401.

[0272] The memory 402 can be used to store software programs and modules, and the processor 401 executes various functions and data processing by running the software programs and modules stored in the memory 402. The memory 402 can mainly include a storage program area and a storage data area, wherein the storage program area can store an operating system, at least one application program required by a function (such as a sound playing function, an image playing function, etc.) and the like; the storage data area can store data created according to the use of the electronic device and the like. In addition, the memory 402 can include a high-speed random access memory, and can also include a non-volatile memory, for example, at least one magnetic disk storage device, a flash memory device, or other volatile solid-state memory device. Accordingly, the memory 402 can also include a memory controller to provide the processor 401 with access to the memory 402.

[0273] The electronic device further includes a power supply 403 for supplying power to each component, and preferably, the power supply 403 can be logically connected to the processor 401 through a power management system, so as to realize functions such as management of charging, discharging and power consumption management through the power management system. The power supply 403 can also include one or more than one direct current or alternating current power supply, a recharging system, a power failure detection circuit, a power converter or inverter, a power state indicator and any other components.

[0274] The electronic device can further include an input unit 404, which can be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal input related to user settings and function control.

[0275] Although not shown, the electronic device can also include a display unit and the like, which will not be described here. Specifically, in the present embodiment, the processor 401 in the electronic device will load the executable file corresponding to the process of one or more than one application program into the memory 402 according to the following instructions, and run the application program stored in the memory 402 by the processor 401, thereby realizing various functions, as follows:

[0276] obtaining to-be-classified content, and identifying a topic distribution type corresponding to historical published content of a publisher of the to-be-classified content; when the topic distribution type is a target type, obtaining attribute information corresponding to the publisher and topic information of the historical published content of the publisher; predicting, based on content information of the to-be-classified content and the attribute information of the publisher and the topic information of the historical published content, a probability that the to-be-classified content belongs to each preset topic; and determining, according to the probability, a target topic of the to-be-classified content from the preset topics.

[0277] The specific implementation of each operation can be referred to the foregoing embodiments, which will not be described herein.

[0278] As can be seen, the embodiment can obtain to-be-classified content, and identify a topic distribution type corresponding to historical published content of a publisher of the to-be-classified content; when the topic distribution type is a target type, obtain attribute information corresponding to the publisher and topic information of the historical published content of the publisher; predict, based on content information of the to-be-classified content and the attribute information of the publisher and the topic information of the historical published content, a probability that the to-be-classified content belongs to each preset topic; and determine, according to the probability, a target topic of the to-be-classified content from the preset topics. When the topic distribution type of the historical published content of the publisher is the target type, the application can combine the attribute information of the publisher and the topic information of the historical published content to predict the topic of the to-be-classified content, thereby improving the accuracy of content topic classification.

[0279] Those skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructions, or by related hardware controlled by the instructions, which can be stored in a computer readable storage medium and loaded and executed by a processor.

[0280] To this end, the embodiment of the application provides a computer readable storage medium, which stores a plurality of instructions. The instructions can be loaded by a processor to execute the steps in any content topic classification method provided by the embodiment of the application. For example, the instructions can execute the following steps:

[0281] obtaining to-be-classified content, and identifying a topic distribution type corresponding to historical published content of a publisher of the to-be-classified content; when the topic distribution type is a target type, obtaining attribute information corresponding to the publisher and topic information of the historical published content of the publisher; predicting, based on content information of the to-be-classified content and the attribute information of the publisher and the topic information of the historical published content, a probability that the to-be-classified content belongs to each preset topic; and determining, according to the probability, a target topic of the to-be-classified content from the preset topics.

[0282] The specific implementation of each operation can refer to the foregoing embodiments, which will not be repeated here.

[0283] The computer readable storage medium can include a read only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0284] Due to the instructions stored in the computer readable storage medium, the steps of any of the content topic classification methods provided by the embodiments of the present application can be executed, and thus the beneficial effects of any of the content topic classification methods provided by the embodiments of the present application can be achieved. Details are described in the foregoing embodiments, which will not be repeated here.

[0285] According to an aspect of the present application, a computer program product or computer program is provided, which includes computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, so that the computer device executes the method provided in any of the various optional implementations of the content topic classification aspect.

[0286] The above describes in detail the content topic classification method, device, electronic device and storage medium provided by the embodiments of the present application. The principles and implementation manners of the present application are described by applying specific examples in this paper. The above description of the embodiments is only used to help understand the method and its core idea of the present application. Meanwhile, for those skilled in the art, the specific implementation manner and application range of the present application will be changed according to the idea of the present application. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A content topic classification method, characterized in that, include: Obtain the content to be categorized and identify the topic distribution type corresponding to the historical posts of the publisher of the content to be categorized; When the topic distribution type is a target type, obtain the attribute information corresponding to the publisher and the topic information of the publisher's historical published content; wherein, the target type is a topic cluster distribution type; Based on the content information of the content to be classified, the attribute information of the publisher, and the topic information of the historical published content, the probability of the content to be classified belonging to each preset topic is predicted; Based on the probability, determine the target topic of the content to be classified from the preset topics; When the topic distribution type is a non-target type, the probability that the content to be classified belongs to each preset topic is predicted based on the content information of the content to be classified; according to the probability, the target topic of the content to be classified is determined from the preset topics.

2. The method according to claim 1, characterized in that, The method of identifying the topic distribution types corresponding to the historical posts of the publisher of the content to be classified includes: Based on the attribute information of the publisher of the content to be classified and the topic information of the publisher's historical published content, a topic distribution analysis is performed on the historical published content to determine the topic distribution type corresponding to the historical published content.

3. The method according to claim 2, characterized in that, The step of performing a topic distribution analysis on the historical published content based on the attribute information of the publisher of the content to be classified and the topic information of the publisher's historical published content, and determining the topic distribution type corresponding to the historical published content, includes: The attribute information corresponding to the publisher of the content to be classified and the topic information of the publisher's historical published content are merged to obtain topic distribution information; Feature extraction is performed on the topic distribution information to obtain the distribution feature information of the topic distribution information; Based on the distribution feature information, predict the topic distribution type corresponding to the historical published content.

4. The method according to claim 3, characterized in that, The topic distribution information includes at least one text unit; The step of extracting features from the topic distribution information to obtain the distribution feature information of the topic distribution information includes: Extract word-level feature information from each text unit in the topic distribution information; Based on the word-level feature information of the text unit corresponding to the context of each text unit, the word-level feature information of each text unit is processed. The word-level feature information of each processed text unit is fused to obtain the distribution feature information of the topic distribution information.

5. The method according to claim 3, characterized in that, The step of extracting features from the topic distribution information to obtain the distribution feature information of the topic distribution information includes: The topic distribution recognition model, after training, is used to extract features from the topic distribution information to obtain the distribution feature information of the topic distribution information.

6. The method according to claim 5, characterized in that, Before extracting features from the topic distribution information using the trained topic distribution recognition model to obtain the distribution feature information of the topic distribution information, the method further includes: Acquire training data, which includes attribute information of the sample publisher, topic information of the sample publisher's historical published content, and the expected topic distribution type of the sample publisher's historical published content; The attribute information of the sample publisher and the topic information of the sample's historical published content are fused to obtain the sample topic distribution information; By using a topic distribution recognition model, feature extraction is performed on the topic distribution information of the samples to obtain the distribution feature information of the topic distribution information of the samples; Based on the distribution feature information, predict the actual topic distribution type of the historical published content of the sample; Based on the actual topic distribution type and the expected topic distribution type, the parameters of the topic distribution recognition model are adjusted to obtain the trained topic distribution recognition model.

7. The method according to claim 1, characterized in that, The method of predicting the probability that the content to be classified belongs to each preset topic based on the content information of the content to be classified, the attribute information of the publisher, and the topic information of the historical published content includes: The content information of the content to be classified, the attribute information of the publisher, and the topic information of the historical published content are merged to obtain the merged information; Feature extraction is performed on the fused information to obtain the feature information of the fused information; Based on the feature information, predict the probability that the content to be classified belongs to each preset topic.

8. The method according to claim 7, characterized in that, The fused information includes at least one text unit; the step of extracting features from the fused information to obtain feature information of the fused information includes: Extract word-level feature information from each text unit in the fused information; Based on the word-level feature information of the text units corresponding to the context of each text unit, the word-level feature information of each text unit is processed. The word-level feature information of each processed text unit is fused to obtain the feature information of the fused information.

9. The method according to claim 8, characterized in that, The extraction of word-level feature information from each text unit in the fused information includes: Obtain the content vector and type vector corresponding to each text unit in the fused information, wherein the type vector represents the information type to which the text unit belongs; The position information of each text unit in the fused information is encoded to obtain the position vector corresponding to each text unit; The content vector, the type vector, and the position vector are fused to obtain word-level feature information of each text unit in the fused information.

10. A content topic classification device, characterized in that, include: The identification unit is used to acquire the content to be classified and identify the topic distribution type corresponding to the historical published content of the publisher of the content to be classified; The acquisition unit is used to acquire the attribute information corresponding to the publisher and the topic information of the publisher's historical published content when the topic distribution type is the target type; wherein, the target type is a topic cluster distribution type; The prediction unit is used to predict the probability that the content to be classified belongs to each preset topic based on the content information of the content to be classified, the attribute information of the publisher, and the topic information of the historical published content. A determining unit is configured to determine the target topic of the content to be classified from the preset topics based on the probability. The prediction unit is further configured to predict the probability that the content to be classified belongs to each preset topic based on the content information of the content to be classified when the topic distribution type is a non-target type; the determination unit can also be configured to determine the target topic of the content to be classified from the preset topics according to the probability.

11. The apparatus according to claim 10, characterized in that, The identification unit is specifically used to perform topic distribution analysis on the historical published content based on the attribute information corresponding to the publisher of the content to be classified and the topic information of the publisher's historical published content, and to determine the topic distribution type corresponding to the historical published content.

12. The apparatus according to claim 11, characterized in that, The identification unit includes a first fusion subunit, a first extraction subunit, and a first prediction subunit; The first fusion subunit is used to fuse the attribute information corresponding to the publisher of the content to be classified with the topic information of the publisher's historical published content to obtain topic distribution information; The first extraction subunit is used to extract features from the topic distribution information to obtain the distribution feature information of the topic distribution information; The first prediction subunit is used to predict the topic distribution type corresponding to the historical published content based on the distribution feature information.

13. The apparatus according to claim 12, characterized in that, The topic distribution information includes at least one text unit; The first extraction subunit is specifically used to extract word-level feature information of each text unit in the topic distribution information; and to process the word-level feature information of each text unit based on the word-level feature information of the text unit in the context corresponding to each text unit. The word-level feature information of each processed text unit is fused to obtain the distribution feature information of the topic distribution information.

14. The apparatus according to claim 12, characterized in that, The first extraction subunit is specifically used to extract features from the topic distribution information using a trained topic distribution recognition model to obtain the distribution feature information of the topic distribution information.

15. The apparatus according to claim 14, characterized in that, The content topic classification device also includes a training unit; The training unit is used to train the topic distribution recognition model. Specifically, the training unit is used to acquire training data, which includes the attribute information of the sample publisher, the topic information of the sample publisher's historical published content, and the expected topic distribution type of the historical published content. The attribute information of the sample publisher and the topic information of the historical published content are fused to obtain the sample topic distribution information. The topic distribution recognition model is used to extract features from the sample topic distribution information to obtain the distribution feature information of the sample topic distribution information. Based on the distribution feature information, the actual topic distribution type of the historical published content is predicted. Based on the actual topic distribution type and the expected topic distribution type, the parameters of the topic distribution recognition model are adjusted to obtain the trained topic distribution recognition model.

16. The apparatus according to claim 10, characterized in that, The prediction unit includes a second fusion subunit, a second extraction subunit, and a second prediction subunit; The second fusion subunit is used to fuse the content information of the content to be classified, the attribute information of the publisher, and the topic information of the historical published content to obtain fused information; The second extraction subunit is used to extract features from the fused information to obtain the feature information of the fused information; The second prediction subunit is used to predict the probability that the content to be classified belongs to each preset topic based on the feature information.

17. The apparatus according to claim 16, characterized in that, The fused information includes at least one text unit; the second extraction subunit is specifically used to extract the word-level feature information of each text unit in the fused information; and to process the word-level feature information of each text unit based on the word-level feature information of the text units corresponding to the context of each text unit. The word-level feature information of each processed text unit is fused to obtain the feature information of the fused information.

18. The apparatus according to claim 17, characterized in that, The second extraction subunit is specifically used to obtain the content vector and type vector corresponding to each text unit in the fused information, wherein the type vector represents the information type to which the text unit belongs; to encode the position information of each text unit in the fused information to obtain the position vector corresponding to each text unit; and to fuse the content vector, the type vector and the position vector to obtain the word-level feature information of each text unit in the fused information.

19. An electronic device, characterized in that, It includes a memory and a processor; the memory stores an application program, and the processor runs the application program within the memory to perform the operations in the content topic classification method according to any one of claims 1 to 9.

20. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted for loading by a processor to perform the steps of the content topic classification method according to any one of claims 1 to 9.

21. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps in the content topic classification method according to any one of claims 1 to 9.

Images