A subject type mining method, device, equipment and storage medium

Abstract

The application relates to the technical field of artificial intelligence, in particular to a topic type mining method, device and equipment and a storage medium, the method comprises the following steps: obtaining target content to be mined, determining title information and label information of the target content; obtaining at least one piece of comment information related to the target content; based on a pre-established double-tower model, determining at least one first semantic vector corresponding to the target content by using the title information, the label information and the at least one piece of comment information; determining a second semantic vector matched with each first semantic vector in a pre-established semantic vector retrieval library to obtain a second semantic vector set; wherein the semantic vector retrieval library comprises second semantic vectors extracted from feature description texts of a plurality of candidate topic types based on the double-tower model; and determining a topic type of the target content according to the second semantic vector set. The application can improve the accuracy of topic type mining.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, and storage medium for mining topic types. Background Technology

[0002] In the era of rapid development of mobile internet, with the lowering of the threshold for content production, both text and image content and video content have experienced rapid development.

[0003] To enable users to accurately access high-quality content that interests them, recommendation systems have emerged. Current recommendation systems primarily employ content recommendation methods that involve reviewing and filtering text, image, or video content uploaded by content creators and placing it into a recommendation content pool. Then, based on user profile characteristics, recommendation algorithms such as collaborative recommendation, matrix factorization, supervised learning algorithms like Logistic Regression, deep learning-based models, and Gradient Boosting Decision Tree (GBDT) are combined with these profile characteristics to make recommendations.

[0004] Currently, content review and screening primarily involve manual review and machine processing. Manual review mainly checks for clickbait titles, lack of informative content, and misinformation. Machine processing mainly uses machine learning and natural language processing technologies to filter and process content, removing sensitive legal or pornographic content, and deduplicating clickbait and similar content. However, due to the complexity and diversity of content formats, as well as the limitations of the models' semantic understanding capabilities, many content quality issues requiring background knowledge processing are difficult to handle. Much content, such as in-depth, practical, and interesting content, is highly personalized and difficult to define. Different users have different experiences, making it difficult to accumulate a large number of samples, which hinders machine processing from identifying and judging, and also makes it difficult for human reviewers to maintain quality standards. Therefore, how to accurately extract thematic information from content to help users filter more attractive and high-quality content from massive amounts of data and improve user experience has become a research topic of great interest. Summary of the Invention

[0005] In view of the above-mentioned problems in the prior art, the purpose of the present invention is to provide a topic type mining method, apparatus, device and storage medium, which can improve the accuracy and efficiency of topic type mining.

[0006] To address the aforementioned problems, this invention provides a topic type mining method, comprising:

[0007] Obtain the target content to be mined, and determine the title information and tag information of the target content;

[0008] Obtain at least one comment related to the target content;

[0009] Based on a pre-established dual-tower model, at least one first semantic vector corresponding to the target content is determined using the title information, the tag information, and the at least one comment information;

[0010] For each first semantic vector, a matching second semantic vector is determined in a pre-established semantic vector retrieval library to obtain a second semantic vector set; wherein, the semantic vector retrieval library includes second semantic vectors extracted from feature description texts for multiple candidate topic types based on the dual-tower model;

[0011] The topic type of the target content is determined based on the second semantic vector set.

[0012] Another aspect of the present invention provides a topic type mining device, comprising:

[0013] The target content acquisition module is used to acquire the target content to be mined and determine the title information and tag information of the target content;

[0014] The comment information acquisition module is used to acquire at least one comment related to the target content;

[0015] The first semantic vector determination module is used to determine at least one first semantic vector corresponding to the target content based on a pre-established dual-tower model, using the title information, the tag information, and the at least one comment information.

[0016] The second semantic vector determination module is used to determine a matching second semantic vector for each first semantic vector in a pre-established semantic vector retrieval library to obtain a second semantic vector set; wherein, the semantic vector retrieval library includes second semantic vectors extracted from feature description texts for multiple candidate topic types based on the dual-tower model;

[0017] The topic type determination module is used to determine the topic type of the target content based on the second semantic vector set.

[0018] In another aspect, the present invention provides an electronic device including a processor and a memory, wherein the memory stores at least one instruction or at least one program, the at least one instruction or the at least one program being loaded and executed by the processor to implement the above-described topic type mining method.

[0019] In another aspect, the present invention provides a computer-readable storage medium storing at least one instruction or at least one program, wherein the at least one instruction or the at least one program is loaded and executed by a processor to implement the topic type mining method as described above.

[0020] Due to the above technical solution, the present invention has the following beneficial effects:

[0021] The topic type mining method of this invention, based on a dual-tower model, utilizes the title information, tag information, and at least one comment information of the target content to determine the topic type of the target content. It fully leverages posterior comment data to mine the topic type of the target content, improving the accuracy and efficiency of topic type mining, thereby accelerating the discovery of high-quality topic content. Furthermore, by utilizing user comment information to mine the topic type of target content, this invention can be applied to mining target content in different fields and of different types, exhibiting strong adaptability. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the present invention, the accompanying drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0023] Figure 1 This is a schematic diagram of the implementation environment provided in one embodiment of the present invention;

[0024] Figure 2 This is a flowchart of a topic type mining method provided in one embodiment of the present invention;

[0025] Figure 3 This is a schematic diagram of a comment example provided in one embodiment of the present invention;

[0026] Figure 4 This is a flowchart of a topic type mining method provided in another embodiment of the present invention;

[0027] Figure 5 This is a schematic diagram of the structure of a twin-tower model provided in one embodiment of the present invention;

[0028] Figure 6 This is a flowchart of a model training method provided in one embodiment of the present invention;

[0029] Figure 7 This is a schematic diagram of the structure of a subject-type mining device provided in one embodiment of the present invention;

[0030] Figure 8This is a schematic diagram of the structure of a topic type mining system provided in one embodiment of the present invention;

[0031] Figure 9 This is a schematic diagram of the structure of a server provided in one embodiment of the present invention. Detailed Implementation

[0032] Artificial Intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to have perception, reasoning, and decision-making capabilities. AI technology is a comprehensive discipline involving a wide range of fields, encompassing both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0033] The solutions provided in this invention relate to the fields of machine learning and natural language processing in artificial intelligence. Machine learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, and other disciplines. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Natural language processing (NLP) is a science that integrates linguistics, computer science, and mathematics. It studies various theories and methods that enable effective communication between humans and computers using natural language. Therefore, research in this field involves natural language, i.e., the language people use in daily life, and thus it is closely related to linguistic research. Natural language processing technologies typically include text processing, semantic understanding, machine translation, robot question answering, knowledge graphs, and other technologies.

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0036] First, the relevant terms involved in the embodiments of this invention are explained as follows:

[0037] PGC: Professional Generated Content, refers to content produced by professional content-producing institutions or organizations.

[0038] UGC: User Generated Content, refers to content produced and shared by end users, such as text, audio, images, and video on social media and online platforms. It emerged alongside the Web 2.0 concept, which emphasizes personalization, and represents a new way of using the internet, shifting from primarily downloading to a balance between downloading and uploading.

[0039] PUGC: Professional User Generated Content, which is content produced in UGC format that is relatively close to PGC.

[0040] MCN: Multi-Channel Network, is a product form of multi-channel network that unites PGC content and, with strong capital support, ensures the continuous output of content, thereby ultimately achieving stable commercial monetization.

[0041] Feeds, also known as news sources, are a data format used by websites to disseminate the latest information to users. They are typically arranged in a timeline format, which is the most original, direct, and basic form of displaying feeds. The aggregation of feeds in one place is called aggregation, and the software used for aggregation is called an aggregator. For users, an aggregator is software specifically designed for subscribing to websites and is generally also called a Really Simple Syndication (RSS) reader, feed reader, news reader, etc.

[0042] BERT (Bidirectional Encoder Representation from Transformers) is a pre-trained model proposed by Google in 2018. The core of the BERT model is a Transformer model used to extract text features, capable of fully describing character-level, word-level, sentence-level, and even inter-sentence relationship features.

[0043] Embedding, mathematically, represents a mapping relationship (F:X->Y). In natural language processing, embedding specifically refers to the mapping result from semantic space to vector space, that is, representing an entity with a low-dimensional vector.

[0044] Faiss: An open-source clustering and similarity search library from Facebook's AI team. It provides efficient similarity search and clustering for dense vectors, supports searching on the order of billions of vectors, and is currently the most mature approximate nearest neighbor search library.

[0045] Reference manual attached Figure 1 The illustration shows a schematic diagram of an implementation environment provided by an embodiment of the present invention, which may include a terminal 110 and a server 120. The terminal 110 and the server 120 may be directly or indirectly connected via wired or wireless communication, and this application does not impose any limitations on this connection.

[0046] The terminal 110 may be a smartphone, tablet computer, laptop computer, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. The terminal 110 may run one or more clients, which may be clients of the same type or clients of different types.

[0047] The server 120 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0048] In practical applications, the client can be a client for information-related applications, such as RSS readers, feed readers, news readers, etc. After opening, logging in, or clicking on the client, users can generate or browse text / video content. After browsing the content, users can also interact with it by commenting, liking, forwarding, or saving it. When a user generates content on the terminal 110, the terminal 110 can send the generated content to the server 120. The server 120 can review, filter, and process the content, and distribute it to other terminals 110. Other terminals 110 can display the content in the form of a feed stream for users to quickly refresh. When a user comments on the content on the terminal 110, the terminal 110 can send the comment information to the server 120. The server 120 can use the comment information to mine the topic type of the content, thereby accelerating the discovery of high-quality topic content and enabling better distribution of high-quality content.

[0049] It should be noted that, Figure 1 This is just one example.

[0050] Reference manual attached Figure 2 This illustrates the flow of a topic type mining method provided by an embodiment of the present invention, which can be applied to... Figure 1 In the server. Specifically, as shown in Figure 2 As shown, the method may include the following steps:

[0051] S201: Obtain the target content to be mined, and determine the title information and tag information of the target content.

[0052] In this embodiment of the invention, the target content can be content from various consulting applications or platforms, including image content, text content, and video content. The target content can be UGC content, PGC content, or PUGC content, and this embodiment of the invention does not impose any limitations.

[0053] In this embodiment of the invention, the tag information of the target content can be pre-annotated by machine algorithms and / or manually. For example, it can be annotated during the review, screening and processing process before the target content is distributed, or it can be input by the producer of the target content when producing the target content. The tags can represent the content information, sentiment classification, producer information and target user information of the target content.

[0054] S203: Obtain at least one comment related to the target content.

[0055] In this embodiment of the invention, a time window (e.g., 10 minutes) can be preset to acquire all comment information generated within this time window regarding the target content, thereby determining the topic type of the target content. The comment information may include at least one comment posted by at least one user who viewed the target content. In one example, such as... Figure 3 As shown, it displays an example of comments about beautiful female athletes, yielding comments such as "Beautiful. There are so many beautiful women in track and field, swimming, and volleyball!" and "Fairy!"

[0056] In one possible embodiment, after obtaining at least one comment related to the target content, the at least one comment can be preprocessed. Specifically, the preprocessing can include removing noise from the comment information, including removing the following types of noise: (1) repeated meaningless words, such as "hahahahahaha"; (2) repeated pinyin, such as "honghonghuohuohuanghuanghuhu"; (3) repeated personal names, such as "Zhang San Zhang San Zhang San Zhang San"; (4) incoherent comments, such as "the three are almost the pride of the street quickly established", such incoherent comments can be identified by a separate incoherence model; (5) comments with full facial expressions; (6) short comments, such as comments with 3 characters or less; (7) low-quality comments such as vulgarity, abuse, and spam, which can be identified by a separate recognition model. On the other hand, the comment information can be filtered from the perspective of the commenting user, combined with the user's profile. Generally, comments posted by users with the following characteristics tend to be of higher quality: (1) Title: Users with higher titles tend to be more active in their comments, and their comments should be given more attention; (2) Education level: People with higher education levels tend to have more complete thinking and use better vocabulary, making it less likely for them to post superficial comments like "good, good, good" or "thumbs up, thumbs up"; (3) Number of historical comment interactions: If a user has previously posted a comment that has received a lot of likes, it can help predict the quality of their latest comment to some extent; (4) User interest and preference information: By analyzing historical behavior, we can identify the characteristics of active users and increase their weight. For example, if a highly active user's core interest is technology and they consume mostly technology-related content, they may have some "deep" insights in this field.

[0057] In one possible embodiment, referencing the appendix to the specification... Figure 4 After obtaining at least one comment related to the target content, the process may further include filtering the at least one comment; specifically, the filtering process may include:

[0058] S204: Perform sentiment analysis on the at least one comment using a preset sentiment analysis model to determine the sentiment tendency of each comment; filter comment information with preset sentiment tendency from the at least one comment.

[0059] Specifically, to uncover high-quality topic content, the sentiment of comments can be analyzed, selecting at least one comment that is positive or neutral. These comments themselves have a positive sentiment, essentially using positive user feedback for topic type mining, and correlating the sentiment type of the comment with the content properties being mined. Specifically, a supervised learning model using a text classification task's convolutional neural network (TextCNN) and long short-term memory network (LSTM) can be employed to construct the preset sentiment analysis model, obtaining the sentiment of each comment. It should be noted that other existing sentiment analysis models can also be used to analyze the sentiment of comments; this embodiment of the invention is not limited to this.

[0060] S205: Based on the pre-established dual-tower model, determine at least one first semantic vector corresponding to the target content using the title information, the tag information, and the at least one comment information.

[0061] In this embodiment of the invention, a dual-tower model can be pre-established, and for each piece of comment information obtained (or each piece of filtered comment information), a first semantic vector is determined based on the dual-tower model using the title information, the tag information, and the comment information. Since there can be multiple pieces of comment information, there can also be multiple first semantic vectors, and the first semantic vector has a one-to-one correspondence with the comment information.

[0062] In one possible embodiment, the dual-tower model may include a first sub-model and a second sub-model, wherein the first sub-model may include a semantic representation layer and a fully connected layer;

[0063] The method of determining at least one first semantic vector corresponding to the target content based on the pre-established dual-tower model, using the title information, the tag information, and the at least one comment information, may include:

[0064] The semantic representation layer is used to process the title information and the tag information respectively to obtain the semantic vector of the title information and the semantic vector of the tag information;

[0065] The semantic representation layer is used to process the at least one comment message to obtain a semantic vector for each comment message;

[0066] For each comment, the semantic vectors of the comment, title, and tags are fused using the fully connected layer to obtain the first semantic vector corresponding to the target content.

[0067] In one possible embodiment, since the title information and tag information of the target content are fixed, the semantic vectors of the title information and the tag information can be extracted in advance using the BERT model. The semantic vectors of the title information and the tag information can be directly used during the topic type mining process.

[0068] In one example, refer to the appendix of the instruction manual. Figure 5 The dual-tower model can be divided into two sides (each side corresponding to a sub-model). The left sub-model is the first sub-model, also known as the User & Content Tower. This first sub-model may include a BERT model and a fully connected layer. The BERT model can extract semantic vectors from the target text's tag information, title information, and user comment information. The fully connected layer can fuse the obtained semantic vectors into a first semantic vector related to the target content. The right sub-model is the second sub-model, also known as the Target Tower. This second sub-model may also include a BERT model, which can extract second semantic vectors from the feature description text of multiple candidate topic types. Then, K-Nearest Neighbor (KNN) can be used for retrieval, or Faiss can be used for vector retrieval to determine the topic type of the target content.

[0069] In practical applications, considering both model performance and efficiency, a two-layer BERT model can be chosen as the actual model, significantly improving inference speed while sacrificing a small amount of accuracy. Of course, a more layered BERT model can also be chosen as the actual model; this embodiment of the invention does not impose any limitations.

[0070] S207: For each first semantic vector, determine the second semantic vector that matches it in the pre-established semantic vector retrieval library to obtain a second semantic vector set; wherein, the semantic vector retrieval library includes second semantic vectors extracted from feature description texts for multiple candidate topic types based on the dual-tower model.

[0071] In this embodiment of the invention, a second semantic vector can be extracted in advance from the feature description text of multiple candidate topic types based on the second sub-model of the dual-tower model, and the second semantic vector is stored in a semantic vector retrieval library. The second sub-model may include a BERT model. The multiple candidate topic types and their corresponding feature description texts can be predefined and have very rich semantic information. Generally, they are topics with great influence on the internet, of concern and interest to users, and capable of generating discussion. Specifically, the content can be classified according to different application scenarios and user needs, such as interesting content, in-depth content, and practical content. Then, combined with specific type categories and production and plot techniques, it can be further divided into many sub-topics. For example, game-related content can be divided into funny games, exciting operations, and creative plots; lifestyle-related content can be divided into street photography, selfies, imitation shows, contrast, curiosity, and suspense; internet celebrity content can be divided into jokes and pranks; food-related content can be divided into food recipes, food reviews, and mukbang; and pet-related content can be divided into animal daily life, animal stories, feeding and care, etc. In addition, there is more semantically rich and abstract user-perspective content. For example, through comments and interactions, users will provide a lot of related discussions on the topic, including practical content, interesting stories about online classes, beautiful female athletes, showing affection, and so on.

[0072] In one possible embodiment, determining a matching second semantic vector for each first semantic vector in a pre-established semantic vector retrieval library to obtain a second semantic vector set may include:

[0073] For each first semantic vector, determine the similarity between each second semantic vector in the semantic vector retrieval library and the first semantic vector;

[0074] Obtain a second semantic vector whose similarity to the first semantic vector meets a preset condition, and use it as the second semantic vector that matches the first semantic vector;

[0075] The second semantic vector set is obtained by summing the second semantic vectors determined by each of the first semantic vectors.

[0076] Specifically, obtaining a second semantic vector whose similarity to the first semantic vector meets a preset condition, as the second semantic vector matching the first semantic vector, may include:

[0077] Based on the magnitude of the similarity, each second semantic vector in the semantic vector retrieval library is sorted from largest to smallest;

[0078] The first, predetermined number of second semantic vectors are selected as the second semantic vectors that match the first semantic vector.

[0079] In practical applications, the similarity between vectors can be measured by the cosine of the angle between two vectors. Depending on the actual situation, one or more matching second semantic vectors can be selected for each first semantic vector. That is, the first semantic vector can match one second semantic vector or multiple second semantic vectors. All the second semantic vectors corresponding to at least one first semantic vector constitute the second semantic vector set.

[0080] S209: Determine the topic type of the target content based on the second semantic vector set.

[0081] In this embodiment of the invention, the candidate topic type that corresponds most frequently to the at least one comment information can be determined as the topic type of the target content. Since the first semantic vector corresponds one-to-one with the comment information, the candidate topic type corresponding to the first semantic vector can be used as the candidate topic type corresponding to the comment information. For each first semantic vector, the candidate topic type corresponding to the second semantic vector that matches it can be used to determine its corresponding candidate topic type.

[0082] In one possible embodiment, determining the topic type of the target content based on the second semantic vector set may include:

[0083] For each first semantic vector, the candidate topic type corresponding to the first semantic vector is determined based on the second semantic vector that matches it in the second semantic vector set;

[0084] Count the number of identical candidate topic types, and determine the topic type with the most identical candidate topic types as the topic type of the target content.

[0085] In practical applications, the K-Nearest Neighbors (KNN) classification algorithm can be used to determine the candidate topic type corresponding to each first semantic vector. KNN is one of the simplest classification methods in data mining. KNN means the K nearest neighbors; each sample can be represented by its K nearest neighbors. The core idea of ​​the KNN algorithm is that if most of a sample's K nearest neighbors in the feature space belong to a certain category, then the sample also belongs to that category and possesses the characteristics of samples in that category. This method determines the category of a sample based solely on the category of its one or a few nearest neighbors.

[0086] In practical applications, for each first semantic vector, Faiss can be used to perform vector retrieval, and one or more of the top-ranked second semantic vectors can be taken as the second semantic vectors that match the first semantic vector. The candidate topic type corresponding to the first semantic vector can be determined based on the obtained second semantic vectors.

[0087] Specifically, for each first semantic vector, if only one matching second semantic vector is selected, the candidate topic type corresponding to the second semantic vector can be directly used as the candidate topic type corresponding to the first semantic vector; if multiple matching second semantic vectors are selected, the candidate topic type corresponding to each second semantic vector can be determined separately, and the candidate topic type corresponding to the most multiple second semantic vectors can be determined as the candidate topic type corresponding to the first semantic vector.

[0088] Specifically, for the target content, if only one comment is obtained, there is only one first semantic vector, and the candidate topic type corresponding to the semantic vector can be directly used as the topic type of the target content; if multiple comment information is obtained, there are also multiple first semantic vectors, and the candidate topic type corresponding to the most multiple first semantic vectors can be determined as the topic type of the target content.

[0089] In one possible embodiment, referencing the appendix to the specification... Figure 6 The method may also include training a dual-tower model, such as Figure 6 As shown, the training dual-tower model may include:

[0090] S601: Obtain multiple preset candidate topic types and their corresponding feature description text;

[0091] S603: Obtain multiple preset sample contents and label each sample content with topic type information;

[0092] S605: For each sample content, obtain the corresponding comment information; associate the comment information with the topic type information of the sample content to generate a training sample pair corresponding to the sample content;

[0093] S607: Based on a preset machine learning algorithm, the dual-tower model is constructed using the feature description texts corresponding to the multiple candidate topic types and the training sample pairs corresponding to the multiple sample contents.

[0094] In this embodiment of the invention, multiple candidate topic types and corresponding feature description text can be predefined; sample content and corresponding comment information can be obtained from consulting applications or platforms, and then the corresponding topic type information can be manually labeled. Alternatively, multiple sample contents corresponding to each candidate topic type and corresponding comment information can be obtained from consulting applications or platforms.

[0095] In practical applications, a subset of corresponding content can be found through predefined high-quality seed comments, which can be predefined to belong to specific topic types. Then, by looking back at the content, the corresponding user comments can be selected, and comments that are associated with the corresponding topic type information can be generated. These association pairs are positive samples. Through these association pairs, a model for matching comment information text with topic type association can be constructed, namely the dual-tower model.

[0096] In one possible embodiment, after obtaining the comment information, the comment information can also be preprocessed, and the preprocessing process is similar to... Figure 2 The preprocessing procedure is the same in the illustrated method embodiment. Similarly, a preset sentiment analysis model can be used to perform sentiment analysis on the comment information, determine the sentiment tendency of each comment, and select positive or neutral comment information from the comment information for model training.

[0097] In one example, the sample construction is shown in Table 1. It can be obtained from multiple scenarios (such as short video content scenarios), corresponding to relevant user comments and topic types. By collecting and labeling a large amount of such corresponding data, it can be used as positive samples for mining topic types through comments to build a dual-tower model.

[0098] Table 1

[0099]

[0100] Specifically, a neural network model can be pre-constructed. The neural network model may include a first sub-model and a second sub-model. The first sub-model may include a semantic representation layer and a fully connected layer. The semantic representation layer may be a BERT model, and the second sub-model may also include a BERT model. The network can be trained using deep semantic matching. During the training process, the parameters of the neural network model can be adjusted until convergence, thus obtaining the dual-tower model.

[0101] In summary, the topic type mining method of this invention, based on a dual-tower model, utilizes the title information, tag information, and at least one comment information of the target content to determine the topic type of the target content. It fully leverages posterior comment data to mine the topic type of the target content, thereby improving the accuracy and efficiency of topic type mining and accelerating the discovery of high-quality topic content. Furthermore, by utilizing user comment information to mine the topic type of target content, this invention can be applied to mining target content in different fields and of different types, demonstrating strong adaptability.

[0102] The topic type mining method of this invention utilizes filtered, effective positive or neutral comment information to mine the topic types of high-quality content from the user's perspective. This accelerates the discovery of high-quality topic content, enables better distribution of such content, reduces the cost of manual screening, and improves operational efficiency. Simultaneously, it provides guidance and assistance to content creators, leading to the generation of more high-quality content that meets user needs on relevant content platforms, thus making the content ecosystem more complete and healthy.

[0103] Reference manual attached Figure 7 This illustrates the structure of a subject-type mining apparatus provided in one embodiment of the present invention. For example... Figure 7 As shown, the device may include:

[0104] The target content acquisition module 710 is used to acquire the target content to be mined and determine the title information and tag information of the target content;

[0105] The comment information acquisition module 720 is used to acquire at least one comment information related to the target content;

[0106] The first semantic vector determination module 730 is used to determine at least one first semantic vector corresponding to the target content based on a pre-established dual-tower model, using the title information, the tag information, and the at least one comment information.

[0107] The second semantic vector determination module 740 is used to determine a matching second semantic vector for each first semantic vector in a pre-established semantic vector retrieval library to obtain a second semantic vector set; wherein, the semantic vector retrieval library includes second semantic vectors extracted from feature description texts for multiple candidate topic types based on the dual-tower model;

[0108] The topic type determination module 750 is used to determine the topic type of the target content based on the second semantic vector set.

[0109] In one possible embodiment, the device may further include a filtering module for filtering the at least one comment, including: performing sentiment analysis on the at least one comment using a preset sentiment analysis model to determine the sentiment tendency of each comment; and filtering comment information with a preset sentiment tendency from the at least one comment.

[0110] In one possible embodiment, the apparatus may further include a dual-tower model training module, which may include:

[0111] The candidate topic type acquisition unit is used to acquire multiple preset candidate topic types and their corresponding feature description text;

[0112] The sample content acquisition unit is used to acquire multiple preset sample contents and to annotate the topic type information of each sample content.

[0113] The training sample pair association unit is used to obtain the corresponding comment information for each sample content; and associate the comment information with the topic type information of the sample content to generate a training sample pair corresponding to the sample content;

[0114] The dual-tower model construction unit is used to construct the dual-tower model based on a preset machine learning algorithm, using the feature description text corresponding to the multiple candidate topic types and the training sample pairs corresponding to the multiple sample contents.

[0115] It should be noted that the apparatus provided in the above embodiments is only illustrated by the division of the above functional modules when implementing its functions. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus provided in the above embodiments and the corresponding method embodiments belong to the same concept, and the specific implementation process can be found in the corresponding method embodiments, which will not be repeated here.

[0116] One embodiment of the present invention also provides a topic type mining system, as detailed in the appendix to the specification. Figure 8 The system may include a content production terminal 801, a content consumption terminal 802, an uplink and downlink content interface server 803, a content database 804, a dispatch center server 805, a manual review system 806, a machine processing system 807, an interactive and statistical reporting interface 808, an interactive content database 809, a topic content vector library 810, a topic mining service 811, and a topic mining model 812.

[0117] The content production end 801 is a content producer of PGC, UGC, MCN or PUGC, which can provide local or photographed text and image content, video or image collection content through mobile terminal or backend application programming interface (API) system.

[0118] The content production terminal 801 can also obtain the upload server interface address by communicating with the upstream and downstream content interface servers, and then upload local files. During the shooting process, the local image and text content can select matching music, filter templates, and image and text beautification functions, etc.

[0119] The content consumer terminal 802 can communicate with the upstream and downstream content interface server to obtain the index information of the accessed image or video file, and then download the corresponding streaming media file and play it through the local player.

[0120] The content consumer terminal 802 can also report user playback behavior data, buffering, loading time, playback clicks, etc. during the upload and download process to the server.

[0121] The content consumer terminal 802 can also consume interactive information of the content, including reporting interactive information such as comments, UGC short texts, likes, reposts, and collections of the content through the interaction and statistics reporting interface.

[0122] The uplink and downlink content interface server 803 can communicate directly with the content production end. Content submitted from the front end, including the title, publisher, summary, cover image, publication time, and photos, can be directly entered into the server and stored in the content database.

[0123] The uplink and downlink content interface server 803 can also write the metadata of the text and image content, such as the file size, cover image link, bitrate, file format, title, publication time, author, etc., into the content database.

[0124] The uplink and downlink content interface server 803 can also submit the uploaded files to the scheduling center server for subsequent content processing and transfer.

[0125] The content database 804 is the core database of content. All metadata of content published by producers is stored in this database, including file size of text and images, cover image link, bitrate, file format, title, publication time, author, etc. It also includes video file size, video format, whether it is original or first release, and the classification of content during the manual review process (including first, second, and third level classifications and tag information. For example, an article explaining Huawei mobile phones may be classified as technology at the first level, smartphones at the second level, domestic mobile phones at the third level, and tagged with Huawei and Mate 30).

[0126] When the uplink and downlink content interface server 803 receives a video file, it performs a standard transcoding operation on the content. After the transcoding is completed, it asynchronously returns metadata, which mainly includes information such as file size, bitrate, specifications, and cover image. This information is stored in the content database 804. During the manual review process, the information in the content database 804 can be read. At the same time, the results and status of the manual review are also sent back to the content database 804 to update the metadata of the content in the content database 804.

[0127] The scheduling center server 805 is responsible for the entire scheduling process of content flow. It can receive the content into the database through the uplink and downlink content interface server 803, and then obtain the content's metadata from the content database 804.

[0128] The scheduling center server 805 can schedule the manual review system 806 and the machine processing system 807 to process content, and control the scheduling order and priority. Content processing mainly includes machine processing and manual review. The core of machine processing is to call the deduplication service to check for completely duplicate and similar content. The deduplication results are written to the content database 804. Completely duplicate content will not be subject to secondary manual processing. The deduplication service can handle different bitrates, different resolutions, different sizes, partial black screens, the presence or absence of filters, the presence or absence of logos, and the insertion of advertising content and cropping of intros and outros in similar text and images, effectively reducing the content that needs to be processed in the manual review process. Specifically, the deduplication of text and images usually uses the similarity hash algorithm, while the deduplication of video content uses the construction of video embedding vectors.

[0129] The dispatch center server 805 can also schedule content to be manually reviewed by the manual review system 806. Finally, the content that passes the manual review system 806 can be provided to the content consumer 802 through a content distribution service, typically a recommendation engine, search engine, or a direct display page provided by the operator.

[0130] The scheduling center server 805 can also communicate with the topic mining service 811 to perform topic type mining and labeling based on comment information after scheduling is completed.

[0131] The manual review system 806 is a complex web database-based system that can read the original information of the text and image content in the content database 804 and perform a preliminary filtering by humans to determine whether the text and image content involves pornography, gambling, or other sensitive characteristics.

[0132] Based on the initial review, the manual review system 806 can also conduct a second review of the content, mainly by classifying and labeling or confirming the content. Since the text and image content itself is not yet fully mature enough to be processed entirely by machine learning such as deep learning, a second manual review is required on the machine processing machine. Through human-machine collaboration, the accuracy and efficiency of the text and image labeling itself can be improved.

[0133] The machine processing system 807 can provide deduplication services for text and video, mainly by vectorizing text and video (for text content, text is usually vectorized using BERT), then establishing vector indexes, and then determining the degree of similarity by comparing the distance between vectors; before performing the deduplication task, the title short text can be deduplicated first to reduce the amount of computation in subsequent processing.

[0134] The interactive and statistical reporting interface 808 communicates with the content consumer terminal 802 to receive reported interactive information such as UGC short texts of content comments, likes, reposts, and collections; the received comment interaction data can be written into the interactive content database 809 as a data source for subsequent comment mining and model training.

[0135] The interactive content database 809 can store user comment interaction data, providing raw sample data for training sentiment analysis models and topic mining models.

[0136] The topic content vector library 810 can uniformly store the embedding vectors of all candidate topic types, and can communicate with the topic mining service 811 to complete the specific topic mining process. The topic mining service 811 can search and query here.

[0137] The topic mining service 811 can service-encode the topic mining model, receive comment interaction information, accept scheduling services from the scheduling center server 805, complete the topic mining process, and communicate with the topic content vector library 810 to complete the specific topic type matching process and strategy.

[0138] The topic mining model 812 can read sample data from the interactive content database 809 and construct a corresponding topic mining model according to the model training method provided in the above method embodiment. The topic mining model is a dual-tower model that shares weight parameters. It can perform targeted model updates and upgrades based on the amount of collected and labeled samples and the required topic coverage, thereby improving the overall mining effect.

[0139] One embodiment of the present invention also provides an electronic device, which includes a processor and a memory, wherein the memory stores at least one instruction or at least one program, the at least one instruction or the at least one program being loaded and executed by the processor to implement the topic type mining method provided in the above method embodiments.

[0140] Memory can be used to store software programs and modules. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory. Memory can primarily include a program storage area and a data storage area. The program storage area can store the operating system, application programs required for the functions, etc.; the data storage area can store data created based on the use of the device, etc. Furthermore, memory can include high-speed random access memory, and can also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, memory can also include a memory controller to provide the processor with access to the memory.

[0141] The method embodiments provided in this invention can be executed in a terminal, server, or similar computing device; that is, the aforementioned electronic device may include a terminal, server, or similar computing device. Taking running on a server as an example, such as... Figure 9 The diagram illustrates the structural schematic of a server for running a topic type mining method according to an embodiment of the present invention. The server 900 can vary considerably depending on its configuration or performance, and may include one or more central processing units (CPUs) 910 (e.g., one or more processors) and memory 930, and one or more storage media 920 (e.g., one or more mass storage devices) for storing application programs 923 or data 922. The memory 930 and storage media 920 may be temporary or persistent storage. The program stored in the storage media 920 may include one or more modules, each module including a series of instruction operations on the server. Furthermore, the CPU 910 may be configured to communicate with the storage media 920 and execute the series of instruction operations in the storage media 920 on the server 900. Server 900 may also include one or more power supplies 960, one or more wired or wireless network interfaces 950, one or more input / output interfaces 940, and / or one or more operating systems 921, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, etc.

[0142] The input / output interface 940 can be used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of server 900. In one example, the input / output interface 940 includes a network interface controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In one example, the input / output interface 940 can be a radio frequency (RF) module for wireless communication with the Internet. This wireless communication can use any communication standard or protocol, including but not limited to Global System for Mobile Communications (GSM), General Packet Radio Service (GPRS), Code Division Multiple Access (CDMA), Wideband Code Division Multiple Access (WCDMA), Long Term Evolution (LTE), email, and Short Messaging Service (SMS).

[0143] Those skilled in the art will understand that Figure 9 The structure shown is for illustrative purposes only; the server 900 may also include more advanced components. Figure 9 The more or fewer components shown, or having the same Figure 9 The different configurations shown.

[0144] One embodiment of the present invention also provides a computer-readable storage medium, which can be disposed in an electronic device to store at least one instruction or at least one program related to implementing a topic type mining method, wherein the at least one instruction or the at least one program is loaded and executed by the processor to implement the topic type mining method provided in the above method embodiment.

[0145] Optionally, in embodiments of the present invention, the storage medium may include, but is not limited to, various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0146] One embodiment of the present invention also provides a computer program product or computer program including 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 executes the computer instructions, causing the computer device to perform the topic type mining methods provided in the various optional implementations described above.

[0147] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, specific embodiments have been described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims can be performed in a different order than that shown in the embodiments and still achieve the desired result. Additionally, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0148] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the apparatus embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0149] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0150] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A subject type mining method characterized by, The method comprises: obtaining target content to be excavated, and determining title information and label information of the target content; the target content is content in each information application or platform; the label information is used to represent content information, sentiment classification, producer information and target user information of the target content; and the target content comprises at least any one of picture content, video content or text content; obtaining at least one piece of comment information related to the target content; the comment information comprises at least one piece of comment published by at least one user who browses the target content; performing sentiment analysis on the at least one piece of comment information by using a preset sentiment analysis model, and determining sentiment tendency of each piece of comment information respectively; screening comment information with a preset sentiment tendency from the at least one piece of comment information; the preset sentiment tendency is positive or neutral; determining at least one first semantic vector corresponding to the target content based on a pre-established double-tower model, the title information, the label information and the at least one piece of screened comment information; the first semantic vector corresponds to the comment information one by one; determining a second semantic vector matched with each first semantic vector in a pre-established semantic vector retrieval library to obtain a second semantic vector set; wherein the semantic vector retrieval library comprises second semantic vectors extracted from feature description texts of a plurality of candidate theme types based on the double-tower model; for each first semantic vector, determining a candidate theme type corresponding to the first semantic vector according to the second semantic vector matched with the first semantic vector in the second semantic vector set; counting the number of the same candidate theme type, and determining the candidate theme type with the largest number as the theme type of the target content.

2. The method of claim 1, wherein, The double-tower model comprises a first sub-model and a second sub-model, and the first sub-model comprises a semantic representation layer and a fully connected layer; determining at least one first semantic vector corresponding to the target content based on the pre-established double-tower model, the title information, the label information and the at least one piece of comment information comprises: processing the title information and the label information by using the semantic representation layer respectively to obtain a semantic vector of the title information and a semantic vector of the label information; processing the at least one piece of comment information by using the semantic representation layer to obtain a semantic vector of each piece of comment information respectively; for each piece of comment information, fusing the semantic vector of the comment information, the semantic vector of the title information and the semantic vector of the label information by using the fully connected layer to obtain the first semantic vector corresponding to the target content.

3. The method according to claim 1 or 2, characterized in that, determining a second semantic vector matched with each first semantic vector in a pre-established semantic vector retrieval library to obtain a second semantic vector set comprises: for each first semantic vector, determining a similarity between each second semantic vector in the semantic vector retrieval library and the first semantic vector; obtaining a second semantic vector with a similarity to the first semantic vector satisfying a preset condition as the second semantic vector matched with the first semantic vector; The second semantic vectors determined by each of the first semantic vectors are aggregated to obtain a second semantic vector set.

4. The method of claim 3, wherein, The method further comprises training a double-tower model, and the training of the double-tower model comprises: obtaining a plurality of preset candidate topic types and feature description texts corresponding to the candidate topic types; obtaining a plurality of preset sample contents, and performing topic type information labeling on each of the sample contents; 5. The method according to claim 1 or 2, characterized in that, for each sample content, obtaining corresponding comment information, associating the comment information with the topic type information of the sample content, and generating a training sample pair corresponding to the sample content; based on a preset machine learning algorithm, constructing the double-tower model by using the feature description texts corresponding to the plurality of candidate topic types and the training sample pairs corresponding to the plurality of sample contents. comprise: a target content acquisition module configured to acquire target content to be mined, and determine title information and label information of the target content; the target content is content in each information application or platform; the label information is used to represent content information, sentiment classification, producer information, and target user information of the target content; and the target content comprises at least any one of picture content, video content, or text content; a comment information acquisition module configured to acquire at least one piece of comment information related to the target content; 6. A subject type mining apparatus characterized by comprising: the comment information comprises at least one piece of comment published by at least one user who browses the target content; a screening module configured to perform sentiment analysis on the at least one piece of comment information by using a preset sentiment analysis model, determine sentiment tendencies of each piece of comment information respectively, and screen comment information having a preset sentiment tendency from the at least one piece of comment information; a first semantic vector determination module configured to determine at least one first semantic vector corresponding to the target content by using the title information, the label information, and the screened at least one piece of comment information based on a double-tower model established in advance; the first semantic vector corresponds to the comment information one by one; a second semantic vector determination module configured to determine, for each of the first semantic vectors, a second semantic vector matched with the first semantic vector in a semantic vector retrieval library established in advance to obtain a second semantic vector set; the semantic vector retrieval library comprises second semantic vectors extracted for feature description texts of a plurality of candidate topic types based on the double-tower model; a topic type determination module configured to determine a topic type of the target content according to the second semantic vector set; the topic type determination module comprises: a candidate type acquisition unit configured to, for each first semantic vector, determine a candidate topic type corresponding to the first semantic vector according to a second semantic vector matched with the first semantic vector in the second semantic vector set. ​ ​ ​ The target type determination unit is configured to count the number of the same candidate theme types, and determine the candidate theme type with the largest number as the theme type of the target content.

7. The apparatus of claim 6, wherein, The double-tower model comprises a first sub-model and a second sub-model, the first sub-model comprises a semantic representation layer and a full connection layer; the first semantic vector determination module comprises: The first vector acquisition unit is configured to process the title information and the label information respectively by using the semantic representation layer, to obtain a semantic vector of the title information and a semantic vector of the label information; The second vector acquisition unit is configured to process the at least one piece of comment information by using the semantic representation layer, to obtain a semantic vector of each piece of comment information respectively; The vector fusion unit is configured to, for each piece of comment information, fuse the semantic vector of the comment information, the semantic vector of the title information and the semantic vector of the label information by using the full connection layer, to obtain a first semantic vector corresponding to the target content.

8. The apparatus of claim 6 or 7, wherein, The second semantic vector determination module comprises: The similarity acquisition unit is configured to, for each first semantic vector, determine the similarity between each second semantic vector in the semantic vector retrieval library and the first semantic vector; The third vector acquisition unit is configured to obtain the second semantic vector whose similarity with the first semantic vector satisfies a preset condition, as the second semantic vector matched with the first semantic vector; The vector set acquisition unit is configured to collect the second semantic vectors determined by each first semantic vector, to obtain a second semantic vector set.

9. The apparatus of claim 8, wherein, The third vector acquisition unit comprises: The sorting sub-unit is configured to sort each second semantic vector in the semantic vector retrieval library from large to small according to the size of the similarity; The vector acquisition sub-unit is configured to take the top preset number of second semantic vectors as the second semantic vectors matched with the first semantic vector.

10. The apparatus of claim 6 or 7, wherein, The device further comprises a double-tower model training module, the double-tower model training module comprises: The candidate theme type acquisition unit is configured to obtain a plurality of preset candidate theme types and feature description texts corresponding thereto; The sample content acquisition unit is configured to obtain a plurality of preset sample contents, and perform theme type information labeling on each sample content; The training sample pair association unit is configured to, for each sample content, obtain comment information corresponding thereto; associate the comment information with the theme type information of the sample content, to generate a training sample pair corresponding to the sample content; The double-tower model construction unit is configured to construct the double-tower model based on a preset machine learning algorithm, by using the feature description texts corresponding to the plurality of candidate theme types and the training sample pairs corresponding to the plurality of sample contents.

11. An electronic device, comprising: The electronic device comprises a processor and a memory, the memory stores at least one instruction or at least one program, the at least one instruction or the at least one program is loaded and executed by the processor to implement the theme type mining method according to any one of claims 1-5.

12. A computer-readable storage medium, characterized in that, The computer readable storage medium stores at least one instruction or at least one program, and the at least one instruction or at least one program is loaded and executed by the processor to implement the subject type mining method as claimed in any one of claims 1-5.

13. A computer program product comprising a computer program, characterized in that, The computer program is executed by the electronic device to implement the subject type mining method as claimed in any one of claims 1-5.

Images