Potential hot content identification method based on mainstream value vertical domain knowledge base

Abstract

The invention relates to the technical field of deep learning, and discloses a potential hot content identification method based on a mainstream value droop domain knowledge base, which comprises the following steps: constructing the mainstream value droop domain knowledge base, carrying out topic modeling through a hidden Dirichlet distribution model and extracting potential topic words, an event triple is extracted through a RoBERTa-CRF (Conditional Random Field) model, and an event atlas is constructed; nodes related to the hot spots are retrieved from the event atlas through an atlas retrieval engine, sub-graphs formed by the related nodes are converted into vector representation through a graph embedding technology, and a hot spot mode is formed; and by calculating Jaccard similarity, matching content features similar to the hotspot mode in the fusion graph representation network, and identifying potential hotspot content. According to the method, the forward enhanced information content is accurately identified and screened through thematic modeling, event extraction, atlas retrieval, similarity calculation and other technologies, and the real-time monitoring and prediction capability on the social hotspot trend can be enhanced.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a method for identifying potential hot content based on a knowledge base in a mainstream value vertical field. Background Art

[0002] The hot information contained in the knowledge base of mainstream value vertical categories is of great value for understanding the focus of public attention, guiding social public opinion, and formulating relevant policies. However, faced with massive amounts of data and information, how to effectively mine content with potential hot value from these data has become a huge challenge. The rise of social media and the popularization of the Internet have made information dissemination faster, but it has also brought about the problem of information overload, making it difficult to quickly identify the truly important hot spots from a large amount of information. In addition, with the increasing complexity and diversity of social events, traditional information processing methods have been unable to meet the needs of modern society for information screening and prediction. Therefore, developing a method that can accurately discover potential hot content from the knowledge base of mainstream value vertical categories is of great significance to improving information processing efficiency, optimizing decision support systems, strengthening public safety, and enhancing social governance capabilities.

[0003] Traditional methods for discovering potential hot content rely primarily on techniques such as keyword matching and frequency statistics. These methods are somewhat effective when handling simple, intuitive hot content discovery tasks. However, faced with complex text data and deep semantic information, these traditional methods are insufficient. They often cannot effectively handle the long-tail effect in text, that is, information that appears less frequently but is of great value. In addition, traditional methods lack sufficient flexibility and adaptability when dealing with emerging topics and rapidly changing hot topics. These methods often fail to deeply understand the context and topic structure of the text, resulting in large errors in the prediction and discovery of hot content. At the same time, they also have difficulty coping with the multimodal nature of information, such as text, images, and video, which limits their application capabilities in an all-media environment. Therefore, there is an urgent need for a new method that can overcome these shortcomings and achieve accurate discovery and prediction of potential hot content. Summary of the Invention

[0004] In order to solve the above problems, the present invention proposes a method for identifying potential hot content based on the mainstream value vertical category knowledge base. By using the mainstream value vertical category knowledge base, topic modeling is performed through LDA to extract keywords, and the RoBERTa-CRF model is used to extract event triplets to construct an event graph. Combined with graph embedding technology and Jaccard similarity calculation, it can accurately identify and screen positively reinforced information content, effectively improving the real-time monitoring and prediction capabilities of social hot trends.

[0005] The technical solution adopted in the present invention is as follows: Methods for identifying potential hot content based on knowledge bases in mainstream vertical value domains include: Build a knowledge base for mainstream vertical value categories, perform topic modeling and extract potential keywords using the latent Dirichlet distribution model, and extract event triplets and construct an event graph using the RoBERTa-CRF model; The graph retrieval engine retrieves nodes related to hotspots from the event graph, and converts the subgraph composed of related nodes into a vector representation using graph embedding technology to form a hotspot pattern; By calculating the Jaccard similarity, content features similar to hotspot patterns are matched in the fusion graph representation network to identify potential hotspot content.

[0006] Furthermore, the topic modeling is performed using the latent Dirichlet distribution model to extract potential topic words, and the event triples are extracted using the RoBERTa-CRF model to construct an event graph, including: Preprocess the text data in the mainstream value vertical domain knowledge base, input the preprocessed text data into the latent Dirichlet distribution model, train it to obtain the topic distribution and topic word set, and thus extract potential topic words; Based on the RoBERTa-CRF model, the pre-trained language representation capability of the RoBERTa layer is used to capture the deep semantic information in the text and the complex relationships between entities. The CRF layer is used to predict the relationships between entities and generate event triples, which include subject, predicate, and object. The event triples are mapped into a graph structure, with the subject and object in the event triples as nodes in the graph and the predicate as the edge connecting the nodes; the relationship between multiple event triples is explicitly represented in the graph, forming an overall network structure.

[0007] Furthermore, the graph retrieval engine retrieves nodes related to hot spots from the event graph, and converts the subgraph composed of related nodes into a vector representation through graph embedding technology to form a hot spot model, including: Input keywords related to hot topics through the graph search engine to locate nodes in the event graph that match the hot topic terms; the retrieved nodes include the main nodes of the hot topic events and their associated contextual information; The hotspot-related subgraph is converted into a vector representation through graph embedding technology. The graph embedding technology maps the nodes and edges in the event graph into a high-dimensional vector space by learning the neighborhood relationship and graph topology characteristics of the nodes in the graph structure.

[0008] Furthermore, the topic modeling and extraction of potential keywords using the latent Dirichlet distribution model include: For each document Select a topic distribution , for each word in the document , first select a topic from the topic distribution of the document , and then select words from the word distribution corresponding to the topic ;in, is the theme word distribution; is a dimensional vector, is the number of topics; is a dimensional vector, is the size of the vocabulary; In mathematical expression, given Topics and documents ,document The generation probability of the word for:

[0009] in, Indicates the topic Next word The probability of Represents a document Medium Theme The probability of is a topic-word distribution, is a document-topic distribution, Represents a document The number of words in .

[0010] Furthermore, the extraction of event triples by the RoBERTa-CRF model includes: Through the RoBERTa layer to input text Encode and get the vector representation of each word ; Learning the sequence probability of word labels based on vector representation through CRF layer ,in is a label sequence, and the CRF layer maximizes the conditional probability To predict the optimal label sequence for each word:

[0011] in, is the normalization factor, is the weight of the transferred feature, is the label of the i−1th word, is the label of the i-th word, Indicates the index of the tag; is an exponential function, is the length of the word sequence, is the number of label types; Event triples are constructed based on the output label sequence, thereby accurately extracting event information from the text.

[0012] Furthermore, the constructing of the event graph includes: The event triples are converted into an event graph, which represents the relationship between events using a graph structure. The subject and object in the event triples serve as nodes in the graph, and the predicate serves as the edge connecting the nodes:

[0013] in, Represents the event graph, is a collection of nodes, is the edge set; Each entity pair in the event triple By predicate Connected, expressions include:

[0014] Where R is the set of all possible predicate relations.

[0015] Furthermore, the subgraph composed of related nodes is converted into a vector representation by using graph embedding technology to form a hotspot pattern, including: For each node , generating a fixed-length node sequence by random walk method , as the contextual relationship of the event graph, where is the wandering length; The node sequence is analyzed by Skip-Gram model Perform embedding learning to maximize the co-occurrence probability between a node and its context:

[0016] in, is a node The context node set, is a node Generate context node probability.

[0017] Furthermore, the transition probability of the random walk method is:

[0018] in, is a node arrive The transfer weight, is the normalization constant; the Node2Vec model introduces two hyperparameters and Control the depth-first and breadth-first tendencies of the walk so that the transfer weight Defined as:

[0019] in, Representation node and Shortest path distance in a graph; hyperparameters Control the probability of returning a node, a hyperparameter Controlling the probability of exploring new nodes allows for balanced modeling of the local and global structures of the graph.

[0020] Furthermore, the node Generate context node Probability , expressed using the Softmax function:

[0021] in, Representation node and The dot product of the embedding vector is used to measure the similarity; if negative sampling is used for optimization, the objective function is approximated as:

[0022] in, is a Sigmoid function, and ; is the negative sampling distribution, is the number of negative samples; represents the mathematical expectation operation, is sampled from a negative distribution The elements sampled from ; After training, each node Embedding vector of That is its representation in d-dimensional vector space; for the entire subgraph , can generate the overall vector representation of the subgraph by aggregating the embedding vectors of all its nodes. The aggregation methods include averaging, weighted summation and graph pooling-based methods. Finally, the subgraph vector Expressed as:

[0023] in, Is an aggregate function.

[0024] Furthermore, the Jaccard similarity is calculated to match content features similar to the hotspot pattern in the fusion graph representation network to identify potential hotspot content, including: The Jaccard similarity is used to calculate the similarity between the subgraph vector and the fusion graph to represent other content features in the network. This is achieved by comparing the ratio of the intersection and union of the two sets. The calculation formula of the Jaccard similarity is:

[0025] in, and Represents two sets, is a collection and The number of elements in the intersection, and is a collection and The number of elements in the union; the Jaccard similarity range is between 0 and 1, and the larger the value, the higher the similarity between the two sets; Based on the event graph, each subgraph is represented as a feature set. These feature sets include node attributes, edge types, and related graph information. By calculating the Jaccard similarity of these feature sets, the similarity between subgraphs is quantified.

[0026] The beneficial effects of the present invention are: 1. This paper proposes a method that effectively improves the accuracy and expertise of discovering potential hot content by combining the LDA topic model with a domain knowledge base. This method can deeply explore the underlying themes of a specific domain and utilize the expertise and terminology in the domain knowledge base to provide the model with rich contextual information, thereby enhancing the professionalism of the text processing in a specific domain. This enhanced professionalism is particularly important for accurate information recognition in application scenarios that require in-depth expertise.

[0027] 2. This invention introduces novel considerations for real-time monitoring and prediction of social hotspot trends, directly addressing the dynamic and complex nature of information dissemination. By constructing an event graph and utilizing a graph search engine, this invention can capture the relationships and dynamic changes between events in real time, rapidly retrieving nodes related to hotspots, and thus enabling real-time monitoring and prediction of social hotspot trends, significantly improving the speed of response to social changes and trends.

[0028] 3. This invention is of great significance for optimizing information content screening and positive reinforcement. By extracting event triplets using the RoBERTa-CRF model and constructing an event graph, and through graph embedding technology and Jaccard similarity calculation, the invention can accurately identify and filter information content that is positively reinforced. This screening mechanism helps filter out irrelevant or negative information, ensuring the quality and value of information content, while also improving the model's ability to identify positive information content. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a flow chart of the method for identifying potential hot content based on the mainstream value vertical category knowledge base in Example 1 of the present invention.

[0030] Figure 2 This is an architecture diagram of a series of technologies for identifying abnormal and hot content features in Example 2 of the present invention. DETAILED DESCRIPTION

[0031] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, the specific embodiments of the present invention are now described. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. That is, the embodiments described are only part of the embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0032] The relevant terms involved in the present invention are explained as follows: The LDA model (Latent Dirichlet Allocation Model) is an unsupervised topic modeling method that can represent large-scale text data as a probability distribution over multiple topics. It identifies implicit topics within the text through topic decomposition and extracts keywords from each topic. For its theoretical foundation, please refer to [Blei, DM, Ng, AY, & Jordan, MI (2003). Latent Dirichlet Allocation. Journal of Machine Learning Research, 3, 993-1022.].

[0033] Graph embedding is a method for mapping graph data into low-dimensional vectors, which captures the graph's topological structure and the relationships between nodes. Specifically, the core idea of ​​graph embedding is to convert the graph structure into a representation in a vector space. The goal is to preserve as much node and edge information as possible from the original graph in this low-dimensional space, facilitating subsequent tasks such as node classification, link prediction, and clustering. For the theoretical basis, please refer to [Grover, A., & Leskovec, J. (2016). node2vec: Scalable Feature Learning for Networks. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and DataMining.].

[0034] Event Knowledge Graph: A graph structure built based on event triples, where nodes represent event entities and edges represent relationships between entities. Its theoretical basis can be found in the literature [Paulheim, H. (2017). Knowledge Graph Refinement: A Survey of Approaches and Evaluation Methods. Semantic Web, 8(3), 489-508.].

[0035] Fusion Graph Representation Network (FGN): A graph neural network technique that fuses features from multiple graph structures to extract global and local representations of content. For its theoretical foundation, see [Grover, A., & Leskovec, J. (2016). node2vec: Scalable Feature Learning for Networks. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.].

[0036] Example 1 Currently, the goal of discovering potential hot topics is to predict and identify information within a vast amount of information that is likely to attract widespread public attention or discussion, enabling timely responses to social hot topics, optimized content recommendations, and effective information monitoring and management. The challenges lie in accurately extracting features from massive amounts of complex data, effectively clustering them, and updating hot topics in real time. Furthermore, it's important to consider information authenticity verification and the polarization of public opinion.

[0037] This embodiment provides a method for identifying potential hot content based on the mainstream value vertical field knowledge base. It comprehensively utilizes technologies such as topic modeling, event extraction, graph retrieval and similarity calculation to discover potential hot content in the vertical field knowledge base, and can realize real-time monitoring of social hot trends and improve the prediction capabilities.

[0038] like Figure 1 As shown, this embodiment of the method for identifying potential hot content based on the mainstream value vertical category knowledge base includes: Build a knowledge base for mainstream vertical value categories, perform topic modeling and extract potential keywords using the latent Dirichlet distribution model, and extract event triplets and construct an event graph using the RoBERTa-CRF model; The graph retrieval engine retrieves nodes related to hotspots from the event graph, and converts the subgraph composed of related nodes into a vector representation using graph embedding technology to form a hotspot pattern; By calculating the Jaccard similarity, content features similar to hotspot patterns are matched in the fusion graph representation network to identify potential hotspot content.

[0039] Preferably, topic modeling is performed using a latent Dirichlet distribution model to extract potential topic words, and event triplets are extracted using a RoBERTa-CRF model to construct an event graph, including: Preprocess the text data in the mainstream value vertical domain knowledge base, input the preprocessed text data into the latent Dirichlet distribution model, train it to obtain the topic distribution and topic word set, and thus extract potential topic words; Based on the RoBERTa-CRF model, the pre-trained language representation capability of the RoBERTa layer is used to capture the deep semantic information in the text and the complex relationships between entities. The CRF layer is used to predict the relationships between entities and generate event triples. The event triples include subject, predicate, and object. The event triples are mapped into a graph structure, with the subject and object in the event triples as nodes in the graph and the predicate as the edge connecting the nodes; the relationship between multiple event triples is explicitly represented in the graph to form an overall network structure.

[0040] Preferably, nodes related to hot spots are retrieved from the event graph by a graph retrieval engine, and the subgraph composed of related nodes is converted into a vector representation by graph embedding technology to form a hot spot pattern, including: Input keywords related to hot topics through the graph search engine to locate nodes in the event graph that match the hot topic terms; the retrieved nodes include the main nodes of the hot topic events and their associated contextual information; The hotspot-related subgraphs are converted into vector representations through graph embedding technology. Graph embedding technology maps the nodes and edges in the event graph into a high-dimensional vector space by learning the neighborhood relationships and graph topology features of the nodes in the graph structure.

[0041] Specifically, the potential hot content identification method based on the mainstream value vertical category knowledge base in this embodiment can be implemented by the following steps: S1. Construction of a mainstream value vertical knowledge base: To mitigate the risks of generative AI in ideological and political security, mainstream value corpus support is required. The knowledge base is divided into a basic corpus and a question-and-answer corpus. The basic corpus includes authoritative datasets such as hundreds of millions of comments and comments, national laws and regulations, national public opinion data, and media case studies. The question-and-answer corpus includes 120,000 pairs of questions covering major sensitive issues and hot topics of public concern. Through human-machine collaboration, a diverse set of question-and-answer data is generated to meet the corpus requirements of generative AI in pre-training, fine-tuning, and reinforcement learning tasks.

[0042] S2, Topic Modeling: We use the LDA model to perform topic modeling on the text in the knowledge base. First, we preprocess the text data, including removing stop words, performing word segmentation, and performing word frequency analysis. The preprocessed text is then fed into the LDA model, which is trained to generate a topic distribution and a set of topic words. Each topic is represented by a set of high-frequency words, which serve as potential topic words.

[0043] S3, Event Triple Extraction: Based on the extracted keywords, the RoBERTa-CRF model is used to extract event triplets from the text. Leveraging RoBERTa's pre-trained language representation capabilities, the model captures deep semantic information and complex relationships between entities in the text. The CRF layer predicts relationships between entities (such as the subject, predicate, and object associated with the keyword) and generates event triplets (subject, predicate, object). Based on sequence annotation, the CRF layer further optimizes the order of entity and relationship labels to ensure syntactic and semantic consistency of the extracted event triplets.

[0044] S4, Event Graph Construction: The event graph is constructed by mapping event triples into a graph structure, with the "subject" and "object" in the triples as nodes in the graph, and the "predicate" as the edges connecting these nodes. In this way, the relationships between multiple triples are explicitly represented in the graph, forming an overall network structure.

[0045] S5, Hotspot-Related Node Retrieval: Using a graph search engine (such as the Neo4j engine based on the Cypher query language), enter keywords related to the hotspot and locate nodes in the event graph that match the hotspot keywords. The retrieved nodes include the main node of the hotspot event and its associated context (such as other related events, relationships, or child nodes).

[0046] S6, Subgraph Vectorization: Graph embedding techniques (such as Node2Vec or GraphSAGE) are used to convert hotspot-related subgraphs into vector representations. Graph embedding techniques map nodes and edges in a graph into a high-dimensional vector space by learning the neighborhood relationships and topological features of the graph structure. After subgraph vectorization, each hotspot subgraph can be represented by a unique vector.

[0047] S7, Potential Hotspot Content Matching: Subgraph vectors are matched with the content features of the fused graph representation network. The similarity between the two is measured by calculating Jaccard similarity. Jaccard similarity is a metric for measuring set similarity and is suitable for comparing subgraph features. This similarity calculation identifies potential hotspot content that is similar to the current hotspot pattern.

[0048] Among them, the purpose of step S1 is to prevent the risks of generative AI in ideological security and political security, and at the same time provide mainstream value corpus support for generative AI to meet its corpus needs in pre-training, fine-tuning, reinforcement learning and other tasks. Specific practices include collecting and integrating basic corpora from authoritative sources, such as People's Daily, speech and commentary, laws and regulations, public opinion data and media cases, to form a large-scale corpus. In addition, through the combination of man and machine, major sensitive issues and hot issues of concern to netizens are covered to form a diverse question and answer corpus. The collection and processing of these corpora help generative AI to follow the mainstream value orientation when processing texts, ensure the comprehensiveness, accuracy and correctness of the content, and thus achieve the safe, standardized and sustainable development of generative AI under the framework of full network mobilization, scientific professionalism and human-machine integration.

[0049] The purpose of step S2 is to structure the text content in the knowledge base, identify the implicit topics in the text through topic decomposition, and extract the keywords of each topic. These keywords will provide semantic guidance for the subsequent event triple extraction, ensuring that the model focuses on content related to hot topics. First, for each document Select a topic distribution , then for each word in the document , first select a topic from the topic distribution of the document , and then select words from the word distribution corresponding to the topic ,in is the theme The word distribution of . is a dimensional vector ( is the number of topics), the value of each dimension determines the prior probability of the corresponding topic appearing in the document. If If an element of is larger, then the prior probability of the topic corresponding to the element appearing in the document is relatively high. is a dimensional vector ( is the size of the vocabulary), controls the distribution of words under each topic, and similarly, if If an element of is larger, then the prior probability of the word corresponding to the element appearing in the topic is relatively high.

[0050] In mathematical expression, given Topics and documents ,document The generation probability of the word in is:

[0051] in, Indicates the topic Next word The probability of Represents a document Medium Theme The probability of is a topic-word distribution, is a document-topic distribution, Represents a document The number of words in .

[0052] The purpose of step S3 is to extract the subject-verb-object triples of events from the text based on the keywords, forming a structured event description and providing a data basis for event graph construction. Specifically, the model uses RoBERTa to analyze the input text. Encode and get the vector representation of each word The CRF layer then learns the sequence probabilities of word labels based on these vector representations. ,in is a label sequence. The CRF model maximizes the conditional probability To predict the optimal label sequence for each word, this can be achieved by the following formula:

[0053] in, is the normalization factor, is the weight of the transferred feature, and are the labels of the i−1th and i-th words respectively; is an exponential function, is the length of the word sequence, in the conditional probability In the calculation of , it is necessary to consider the information of each position in the sequence, from arrive Perform summation; is the number of label types for each position The word, its label Can be taken different values, among which Indicates the index of the tag.

[0054] Finally, the label sequence output by the model is used to construct event triples in the form of (entity 1, relation, entity 2), thereby accurately extracting event information from the text.

[0055] The purpose of step S4 is to convert the event triples into an event graph, representing the relationship between events using a graph structure, providing intuitive graph data support for the retrieval and analysis of hot information. The "subject" and "object" in the triples serve as nodes in the graph, and the "predicate" serves as the edge connecting these nodes. The graph construction can be expressed using the following formula to represent the connection between nodes and edges: ,in Representative graph, is a collection of nodes, is an edge set. For each entity pair in the event triple , which are expressed through the predicate connected, can be expressed as:

[0056] Here, R is the set of all possible predicate relationships. In practice, graph databases such as Neo4j can be used to store and query this structured data. Such databases provide efficient data retrieval performance, making querying and analyzing event graphs faster and more convenient.

[0057] The purpose of step S5 is to find the nodes related to the hot topic and their associated subgraphs through the graph retrieval engine, and provide target data for the construction of the hot spot model. The process of retrieving nodes related to the hot spot from the event graph through the graph retrieval engine involves using a graph database query language, such as Cypher, to perform efficient graph queries. The specific approach is to define queries to match nodes and relationships related to the hot spot. For example, if the hot spot is related to a specific event type, the query may look for nodes with specific labels or nodes with specific attribute values. In a graph database, such as Neo4j, this query can quickly find relevant nodes in the graph structure and retrieve other nodes and relationships connected to these nodes, thereby constructing an event graph related to the hot spot. This method can not only identify the direct participants of hot events, but also reveal the complex relationship network behind the event, providing support for further analysis and decision-making.

[0058] The purpose of step S6 is to map the structure, node features and edge relationships of the hotspot related subgraphs into a high-dimensional vector space. Through embedding, the complex relationships of the subgraphs can be compressed into a fixed-dimensional vector representation, which is convenient for similarity calculation with other content. The goal of graph embedding is to convert the subgraphs into Represented as a low-dimensional vector, where is a collection of nodes, is a set of edges. Given a subgraph , we hope to embed the function The node Mapping to d-dimensional vector space , while preserving the structural characteristics and semantic information of the graph. To achieve this goal, it is necessary to define the neighborhood relationship of the nodes, generate the context of the nodes through random walks, and then embed these contexts for learning.

[0059] In the Node2Vec model, the first step of the embedding process is to generate a neighborhood sequence of nodes through random walks. , generating a fixed-length node sequence through random walks ,in is the walk length. The transition probability of a random walk is defined as:

[0060] in, is a node arrive The transfer weight, Is the normalization constant. Node2Vec introduces two hyperparameters and Control the depth-first and breadth-first tendencies of the walk so that the transfer weight Defined as:

[0061] in, Representation node and The shortest path distance in the graph. Parameters Control the probability of returning a node, Controlling the probability of exploring new nodes allows for balanced modeling of the local and global structures of the graph.

[0062] Node sequence generated by random walk is regarded as the contextual relationship of the graph. Next, these sequences are embedded through the Skip-Gram model. The goal of Skip-Gram is to maximize the co-occurrence probability between a node and its context, which is defined as:

[0063] in, is a node The context node set, is a node Generate context node The probability of , using the Softmax function is expressed as:

[0064] in, Representation node and The dot product of the embedding vectors is used to measure their similarity. Since the cost of calculating the Softmax denominator is high, negative sampling is usually used for optimization, and the objective function is approximated as:

[0065] in, is a Sigmoid function, and there is ; is the negative sampling distribution, is the number of negative samples, represents the mathematical expectation operation, is sampled from a negative distribution After training, each node Embedding vector of That is its representation in d-dimensional vector space. , the overall vector representation of the subgraph can be generated by aggregating the embedding vectors of all its nodes. Common aggregation methods include averaging, weighted summation, or graph pooling-based methods. Finally, the subgraph vector representation is:

[0066] in, is an aggregation function, such as mean or maximum pooling. Through the above process, the hotspot-related subgraphs are converted into high-dimensional vector representations. These vectors not only retain the structural information of the graph, but also capture the semantic relationship between nodes, which facilitates subsequent similarity calculation and hotspot pattern matching.

[0067] The purpose of step S7 is to identify potential hot content similar to the current hot pattern by calculating the similarity between the hot pattern and other content features. The process of using Jaccard similarity to calculate the similarity between the subgraph vector and other content features in the fusion graph representation network is achieved by comparing the ratio of the intersection and union of the two sets. The calculation formula of Jaccard similarity is:

[0068] in, and Represents two sets, is a collection and The number of elements in the intersection, and is a collection and The number of elements in the union. The Jaccard similarity range is between 0 and 1, with larger values ​​indicating greater similarity between the two sets.

[0069] In the context of graphs, each subgraph can be represented as a feature set that contains node attributes, edge types, and other relevant graph information. By calculating the Jaccard similarity of these feature sets, we can quantify the similarity between subgraphs. Specifically, for the feature sets of two subgraphs, we first determine their shared features (the intersection), then determine the sum of all their unique features (the union), and finally apply the above formula to calculate the similarity.

[0070] It can be seen that the method of this embodiment can improve the efficiency and accuracy of identifying potential hot content from the mainstream value vertical category knowledge base by accurately identifying and screening positively reinforced information content through technologies such as topic modeling, event extraction, graph retrieval and similarity calculation, and enhance the real-time monitoring and prediction capabilities of social hot trends, thereby providing timely and valuable information support for decision makers. In addition, the method of this embodiment can also break through the limitations of existing technologies in discovering potential hot content based on domain knowledge bases, and achieve accurate perception of potential hot spots and abnormal content by constructing a large-scale mainstream value vertical category knowledge base, which is of great significance for quickly responding to social changes, guiding public opinion and optimizing information processing processes.

[0071] To sum up, the method of this embodiment builds a knowledge base of mainstream value vertical categories, comprehensively utilizes topic modeling, event extraction, graph retrieval and similarity calculation, accurately perceives potential hotspots and abnormal content, enhances the application of generative AI in ideology and political security, realizes the identification and screening of positive information content, significantly improves the positive screening ability of information content, and enhances the real-time monitoring and prediction of social hot trends.

[0072] Example 2 This embodiment is based on embodiment 1: This embodiment provides a method for identifying potential hot content based on a mainstream value vertical category knowledge base, including the following steps: Build a mainstream value vertical knowledge base, including basic corpus and question-and-answer corpus. The basic corpus includes authoritative data sets such as People's Daily, and the question-and-answer corpus covers major sensitive issues and hot topics of concern to netizens. The LDA model is used to perform topic modeling on the text in the knowledge base. The text data is preprocessed and then input into the model to train the topic distribution and topic word set. Based on the extracted keywords, the RoBERTa-CRF model is used to extract event triplets from the text, capturing the deep semantic information in the text and the complex relationships between entities. Construct an event graph, map event triples into a graph structure, and form an overall network structure; Retrieve nodes related to hot topics through the graph search engine and locate nodes in the event graph that match hot topic terms; Graph embedding technology is used to convert hotspot-related subgraphs into vector representations, and the neighborhood relationships and graph topology features of nodes in the graph structure are learned; The subgraph vectors are matched with the content features of the fusion graph representation network, and the potential hot content similar to the current hot pattern is identified by calculating the Jaccard similarity.

[0073] Specifically, the method of this embodiment was used to conduct experiments on two real data sets (a People's Daily corpus data set and a social media public opinion data set), respectively verifying the method's ability to discover potential hot content in different domain knowledge base environments.

[0074] Experimental Datasets: 1. People's Daily Corpus Dataset: This primarily covers mainstream values, policy interpretations, and hotly debated commentary. 2. Social Media Public Opinion Dataset: This dataset is scraped from the Sina website, labeled as rumor-like and quality-controlled. For privacy reasons, this dataset has been anonymized.

[0075] Baseline model: The comparison baseline methods include keyword frequency statistics (TF-IDF), traditional topic modeling (LDA), and deep learning event extraction model (BERT+CRF).

[0076] Evaluation indicators: Hotspot detection accuracy (Precision), recall rate (Recall) and F1 score are used as the main evaluation criteria.

[0077] Experimental scenario: For hot events in different data sets, compare the potential hotspot discovery performance of the method in this embodiment with that of the baseline model.

[0078] Table 1 - Experimental results of different methods on two datasets

[0079] As shown in Table 1, in the People's Daily corpus dataset, the method in this embodiment accurately identified hot policy interpretations and public opinion focal points by constructing an event graph and using graph embedding technology, achieving an F1 score of 0.88, a 10% improvement over the baseline BERT+CRF model. In a social media public opinion dataset, the method in this embodiment significantly improved its ability to monitor hot topic propagation paths and topic aggregation, achieving a hotspot detection accuracy of 0.85, 11.8% higher than the BERT+CRF baseline model.

[0080] like Figure 2 As shown in the figure, the recognition pattern of abnormal and hot content features and its series of technologies are presented. The left part describes how to perform feature matching through the representation of the all-media content fusion graph to identify abnormal and hot content; when the match is successful, the system can discover the anomaly or hot spot. The right part introduces the series of abnormal and hot content feature recognition technologies that support this process, including the use of mainstream value vertical category knowledge bases and event context knowledge graphs to identify the publishing entity, and the use of content precision understanding technology to process hot content, such as false information, sensitive images and dense videos, thereby achieving comprehensive identification and management of abnormal and hot content.

[0081] This demonstrates that the method of this embodiment significantly improves the accuracy and real-time performance of potential hotspot content discovery. The method also demonstrates strong generalization capabilities and can be applied to hotspot content discovery tasks across diverse fields. Through the above experiments and analysis, the method of this embodiment demonstrates its superiority in potential hotspot content discovery and provides an important reference for hotspot monitoring and prediction in related fields.

[0082] Example 3 This embodiment is based on embodiment 1: This embodiment provides a computer device including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the method for identifying potential hot content based on a mainstream value vertical domain knowledge base in Example 1. The computer program may be in source code form, object code form, executable file, or some intermediate form.

[0083] Example 4 This embodiment is based on embodiment 1: This embodiment provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the potential hotspot content identification method based on the mainstream value vertical category knowledge base of Example 1. The computer program may be in source code form, object code form, executable file or some intermediate form, etc. The storage medium includes: any entity or device capable of carrying computer program code, recording medium, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the storage medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the storage medium does not include electric carrier signals and telecommunication signals.

[0084] It should be noted that, for the sake of simplicity, the aforementioned method embodiments are described as a series of action combinations. However, those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily required by this application.

Claims

1. A method for identifying potential hot content based on a knowledge base of mainstream value vertical categories, characterized by: include: Build a knowledge base in the mainstream vertical domain, perform topic modeling and extract potential keywords through the latent Dirichlet distribution model, and extract event triplets and build an event graph through the RoBERTa-CRF model; The graph retrieval engine retrieves nodes related to hot spots from the event graph, and converts the subgraph composed of related nodes into vector representation through graph embedding technology to form a hot spot model; By calculating the Jaccard similarity, content features similar to the hotspot pattern are matched in the fusion graph representation network to identify potential hotspot content.

2. The potential hot content identification method based on the mainstream value vertical category knowledge base according to claim 1 is characterized in that: The topic modeling is performed by using a latent Dirichlet distribution model to extract potential topic words, and the event triples are extracted by using a RoBERTa-CRF model to construct an event graph, including: Preprocess the text data in the mainstream value vertical domain knowledge base, input the preprocessed text data into the latent Dirichlet distribution model, train to obtain the topic distribution and topic word set, and thus extract potential topic words; Based on the RoBERTa-CRF model, the deep semantic information in the text and the complex relationship between entities are captured through the pre-trained language representation capability of the RoBERTa layer, and the relationship between entities is predicted through the CRF layer to generate event triples, wherein the event triples include subject, predicate and object; The event triples are mapped into a graph structure, with the subject and object in the event triples as nodes in the graph and the predicate as the edge connecting the nodes; the relationship between multiple event triples is explicitly represented in the graph to form an overall network structure.

3. The potential hot content identification method based on the mainstream value vertical domain knowledge base according to claim 1 is characterized in that: The method of retrieving nodes related to hot spots from the event graph through the graph retrieval engine and converting the subgraph composed of related nodes into a vector representation through the graph embedding technology to form a hot spot model includes: Input keywords related to hot topics through the graph search engine to locate nodes in the event graph that match the hot topic words; the retrieved nodes include the main nodes of the hot events and their associated context information; The hotspot-related subgraphs are converted into vector representations through graph embedding technology. The graph embedding technology maps the nodes and edges in the event graph into a high-dimensional vector space by learning the neighborhood relationships and graph topology features of the nodes in the graph structure.

4. The potential hot content identification method based on the mainstream value vertical category knowledge base according to claim 1 is characterized in that: The topic modeling is performed through the latent Dirichlet distribution model and potential topic words are extracted, including: For each document Select a topic distribution , for each word in the document , first select a topic from the topic distribution of the document , and then select words from the word distribution corresponding to the topic ;in, Is the theme The word distribution of is a dimensional vector, is the number of topics; is a dimensional vector, is the size of the vocabulary; In mathematical expression, given Topics and documents ,document The generation probability of the word for: in, Indicated in the topic Next word The probability of Representation Document Medium Theme The probability of is a topic-word distribution, is a document-topic distribution, Representation Document The number of words in .

5. The potential hot content identification method based on the mainstream value vertical category knowledge base according to claim 1 is characterized in that: The event triples are extracted by the RoBERTa-CRF model, including: Through the RoBERTa layer to the input text Encode and get the vector representation of each word ; Learning the sequence probability of word labels based on vector representation through CRF layer ,in is a label sequence, and the CRF layer maximizes the conditional probability To predict the optimal label sequence for each word: in, is the normalization factor, is the weight of the transferred feature, is the label of the i−1th word, is the label of the i-th word, Indicates the index of the tag; is an exponential function, is the length of the word sequence, is the number of label types; Event triples are constructed based on the output label sequence, thereby accurately extracting event information from the text.

6. The potential hot content identification method based on the mainstream value vertical category knowledge base according to claim 1 is characterized in that: The constructing of the event graph includes: The event triples are converted into event graphs, and the association relationship of events is represented by a graph structure; the subject and object in the event triples are used as nodes in the graph, and the predicate is used as the edge connecting the nodes: in, Represents the event graph, is a collection of nodes, is the edge set; Each entity pair in the event triple By predicate Connected, expressions include: Where R is the set of all possible predicate relations.

7. The method for identifying potential hot content based on a mainstream value vertical domain knowledge base according to claim 6 is characterized in that: The subgraph consisting of related nodes is converted into a vector representation by using a graph embedding technology to form a hotspot pattern, including: For each node , generate a fixed-length node sequence through a random walk method , as the contextual relationship of the event graph, where is the walk length; The node sequence is analyzed by Skip-Gram model Perform embedding learning to maximize the co-occurrence probability between a node and its context: in, Is a node The context node set of Is a node Generate context node probability.

8. The potential hot content identification method based on the mainstream value vertical category knowledge base according to claim 7 is characterized in that: The transition probability of the random walk method is: in, Is a node arrive The transfer weight, is a normalization constant; the Node2Vec model introduces two hyperparameters and Control the depth-first and breadth-first tendencies of the walk so that the transfer weight Defined as: in, Representation Node and Shortest path distance in a graph; hyperparameters Control the probability of returning a node, a hyperparameter Control the probability of exploring new nodes to achieve a balanced modeling of the local and global structures of the graph.

9. The potential hot content identification method based on the mainstream value vertical category knowledge base according to claim 8 is characterized in that: The node Generate context node Probability , expressed using the Softmax function: in, Representation Node and The dot product of the embedding vector is used to measure the similarity; if negative sampling is used for optimization, the objective function is approximated as: in, is a Sigmoid function, and ; is the negative sampling distribution, is the number of negative samples; represents the mathematical expectation operation, is a negative sampling distribution The elements sampled from ; After training, each node The embedding vector That is its representation in d-dimensional vector space; for the entire subgraph , can generate the overall vector representation of the subgraph by aggregating the embedding vectors of all its nodes, and the aggregation methods include averaging, weighted summation and graph pooling-based methods; finally, the subgraph vector It is expressed as: in, Is an aggregate function.

10. The potential hot content identification method based on the mainstream value vertical domain knowledge base according to claim 1 is characterized in that: By calculating the Jaccard similarity, content features similar to the hotspot pattern are matched in the fusion graph representation network to identify potential hotspot content, including: The Jaccard similarity is used to calculate the similarity between the subgraph vector and the fusion graph to represent other content features in the network. This is achieved by comparing the ratio of the intersection and union of the two sets. The calculation formula of the Jaccard similarity is: in, and Represents two sets, Is a collection and The number of elements in the intersection, and Is a collection and The number of elements in the union; the value range of Jaccard similarity is between 0 and 1, and the larger the value, the higher the similarity between the two sets; Based on the event graph, each subgraph is represented as a feature set, which contains the attributes of the nodes, the types of the edges, and the related graph information. The similarity between the subgraphs is quantified by calculating the Jaccard similarity of these feature sets.

Images