Method and system for constructing false illegal advertisement knowledge base based on big data

Abstract

The application discloses a false illegal advertisement knowledge base construction method and system based on big data, which comprises the following steps: obtaining advertisement semantic representation vectors by performing multi-modal semantic alignment and joint representation learning on multi-source heterogeneous advertisement data, obtaining advertisement elements by performing hierarchical knowledge extraction, extracting legal elements in supervision and punishment labels, constructing advertisement illegal semantic hypergraphs by performing mapping and binding, constructing a false illegal advertisement knowledge base by performing entity alignment, relationship disambiguation and conflict resolution, and dynamically evolving and updating the false illegal advertisement knowledge base based on time decay and mapping and binding confidence. The method can not only improve the efficiency and accuracy of false illegal advertisement knowledge base construction, but also has good interpretability and can be directly applied to a false illegal advertisement knowledge base construction system.

Description

Technical Field

[0001] This invention relates to the field of advertising supervision technology, and in particular to a method and system for constructing a knowledge base of false and illegal advertisements based on big data. Background Technology

[0002] With the booming development of mobile internet and digital media, the channels for spreading false and illegal advertisements have expanded to complex online environments. As relevant laws and regulations continue to improve, higher requirements are being placed on the identification of false advertisements. Building a knowledge base capable of efficiently, accurately, and interpretably identifying false and illegal advertisements has become a key technological direction for improving the intelligence level of advertising supervision, and has significant practical significance and application value.

[0003] However, existing technologies for building knowledge bases for false and illegal advertisements still have significant shortcomings: First, traditional methods rely on keyword strategies or rule models, which are difficult to effectively process multi-source heterogeneous advertising data and ignore a large amount of cross-modal semantic information; second, knowledge extraction in existing technologies generally only stays at the "entity-relationship" level, lacking the ability to deeply map and bind advertising content features with specific legal provisions, resulting in poor explanation capabilities for violations and difficulty in directly serving regulatory enforcement; at the same time, most knowledge bases adopt a static or fixed-cycle full-update model, which cannot be rapidly and differentially evolved according to the dynamic evolution of advertising methods and feedback from manual review, resulting in poor timeliness and high maintenance costs for the knowledge base. Therefore, this invention proposes a method and system for constructing a knowledge base for false and illegal advertisements based on big data. Through multimodal semantic alignment and joint representation learning, multi-source heterogeneous data are uniformly mapped to the same semantic space. An innovative hierarchical knowledge extraction architecture is proposed, and an advertising illegality semantic hypergraph is constructed through a mapping and binding mechanism. Finally, the knowledge base is dynamically evolved and updated based on time decay and mapping and binding confidence. This effectively solves the problems of cross-modal semantic gap, knowledge-law disconnect, and rigid update mechanism in existing technologies, and provides a solid technical foundation for building an intelligent advertising supervision system with strong interpretability and high timeliness. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for constructing a knowledge base of false and illegal advertisements based on big data.

[0005] To achieve the above objectives, the present invention is implemented according to the following technical solution: This invention includes the following steps: Collect historical multi-source heterogeneous advertising data and corresponding regulatory penalty annotations, and perform multimodal semantic alignment and joint representation learning on the multi-source heterogeneous advertising data to obtain advertising semantic representation vectors; Hierarchical knowledge extraction is performed on the semantic representation vector of advertisements to obtain advertising elements, and legal elements are extracted from regulatory penalty annotations. By mapping and binding advertising elements with legal elements, a semantic hypergraph of advertising violations is constructed; Entity alignment, relation disambiguation, and conflict resolution are performed on the semantic hypergraph of illegal advertising to construct a knowledge base for false and illegal advertising. The knowledge base of false and illegal advertisements is dynamically evolved and updated based on time decay and mapping-bound confidence. The multi-source heterogeneous advertising data includes apparent internet advertising text, advertising images, keyframe sequences of advertising videos, and DOM structure data of advertising landing pages; The regulatory penalty label includes advertising laws and regulations and the penalty results; The advertising elements include the advertising entity, the advertising relationship, and the advertising event.

[0006] Furthermore, the method for obtaining the semantic representation vector of the advertisement includes: Collect historical multi-source heterogeneous advertising data and corresponding regulatory penalty annotations; Multi-source heterogeneous advertising data is preprocessed to obtain multimodal initial semantic vectors; the preprocessing operations include encoding advertising text through a pre-trained language model to obtain text semantic vectors, encoding advertising images through a hierarchical visual Transformer to obtain image semantic vectors, encoding the keyframe sequence of advertising video frame by frame and then aggregating it temporally to obtain video keyframe semantic vectors, and encoding the DOM tree from bottom to top through a tree-shaped long short-term memory network to obtain DOM structure semantic vectors. The initial multimodal semantic vectors are mapped to a unified cross-modal semantic space using a learnable projection matrix, and the semantic proximity between different modalities of the same advertisement sample is enhanced by a cross-modal alignment loss, as expressed in the following expression: ; in For cross-modal alignment loss function, , for Sample index of training batch, For modal index set, For image indexing, For video keyframe indexing, For DOM structure indexing, For text semantic vectors, For modal semantic vectors, For temperature coefficient, For positive sample pairs, For negative sample pairs; The semantic consistency score among the initial semantic vectors of the multimodal systems after modal alignment is calculated using a consistency metric function and expressed as follows: ; in The semantic consistency score between the text / image modality and other modalities. This is the consistency metric weight vector. , This is the initial semantic vector of the multimodal expressions after modal alignment. This is a consistency bias vector; After modal alignment, the initial multimodal semantic vector and semantic consistency score are gated and fused to obtain a multimodal fused semantic vector. Then, feature concatenation and projection are performed to obtain the advertising semantic representation vector. ; ; in For multimodal fusion semantic vectors, For advertising semantic representation vectors, For the semantic projection matrix of advertisements, This is the projection bias vector.

[0007] Furthermore, the method for hierarchical knowledge extraction to obtain advertising elements includes: The advertising semantic representation vector is injected into the modally aligned text semantic vector, and a conditional random field is used to jointly decode the entity boundary and type. The word-level tag score matrix is ​​calculated to obtain the advertising entity. The entity representation of the advertising entity includes entity type, text span, and semantic embedding. The entity interaction representation vector is obtained by concatenating the semantic embedding, interaction features and advertising semantic representation vectors of each advertising entity, calculating the relationship type probability and determining the advertising relationship. An event type query attention mechanism is used for event structure identification. Entity representations are aggregated to obtain event semantic vectors. The probability of existence for each event type is calculated to determine the advertising event. The expression is as follows: ; ; in For the first Event types for entities Attention weights For event query projection matrix, For the first Learnable query vectors for various event types As a feature dimension of the event semantic space, For the first The event semantic vector of a type of event, This represents the total number of entities identified in the advertising sample. For the first The entity representation vector of each entity.

[0008] Furthermore, the method for constructing the advertising violation semantic hypergraph includes: The semantic embedding of advertising entities is mapped to the legal semantic space. The cosine similarity between the semantic embedding of advertising entities and the semantic embedding of legal provisions in the same case of false and illegal advertising is calculated. The cosine similarity is normalized as the mapping binding confidence. The binding relationship between each advertising entity and each legal provision is established based on the mapping binding confidence to determine the binding pairs. The global mapping binding set is output. Extract the advertising entities contained in each advertising event, filter the binding pairs corresponding to the advertising entities in each advertising event, and form an event mapping binding subset; The advertising entity and advertising event are treated as advertising domain nodes, and the legal provisions and penalties are treated as legal domain nodes. Based on the event mapping binding subset, ad entity nodes and legal provision nodes within the same ad event are aggregated into a hyperedge. The attention weight of the ad semantic representation vector to each ad entity representation vector is calculated, and the hyperedge weight is calculated by combining the mapping binding confidence of each binding pair. The expression is as follows: ; ; in for The super-edge weight of advertising events for The event mapping of the advertising event is bound to a subset. For the first One advertising entity, For the first One legal provision, To establish confidence levels for mapping advertising entities to legal provisions. For the overall representation of advertising With the representation of advertising entities The attention weights are denoted by , where is the query projection matrix and is the key projection matrix. For the attention head dimension, for The set of advertising entity nodes for an advertising event; The advertising domain nodes, legal domain nodes, hyperedges, and hyperedge weights are assembled into an advertising violation semantic hypergraph.

[0009] Furthermore, the method for constructing a knowledge base for false and illegal advertisements includes: Obtain the semantic hypergraph of different false and illegal advertisements. Calculate the alignment score of the advertisement domain nodes based on the semantic embedding of the nodes and the hypergraph structure position. Perform alignment decisions and cluster merging based on the alignment scores. The expression is: ; in For advertising entities and Corresponding ad domain node and Alignment score, For balance coefficient, , For advertising domain nodes and semantic embedding, , For advertising domain nodes and The associated set of hyperedge indices; Calculate the competition priority of each legal provision node corresponding to the hyperedge in the semantic hypergraph of advertising violations, and perform relation disambiguation according to the competition priority. The expression is: ; in For legal provisions Corresponding legal provisions node Competition and cooperation priorities For advertising events Sino-French Term Mean confidence score for all mappings For the application of legal provisions in historical regulatory penalties The average penalty amount For advertising events The index of legal provisions in the text This is the time-related decay coefficient. For the current moment, For legal provisions The most recent revision time, , , These are the relational weight coefficients; The method for disambiguating relationships is as follows: select the legal node with the highest priority of competition and cooperation as the main legal node, set legal nodes with a priority greater than the competition and cooperation retention threshold as competition and cooperation related legal nodes and retain them in the super edge, and otherwise remove the remaining legal nodes from the super edge. The semantic hypergraph of advertising violations after entity alignment and relation disambiguation is transformed into a set of knowledge entries; the set of knowledge entries includes advertising entity nodes, main legal provision nodes, associated penalty result nodes, hyperedge weights, and timestamps; The conflict degree is calculated for any two sets of knowledge items using the following expression: ; in A collection of knowledge items and The degree of conflict For the corresponding set of advertising entity nodes and Jaccard similarity between them For mutually exclusive indicator functions of legal ontology, when the main legal provisions of two sets of knowledge entries... and In legal terms, it takes the value 1 when defined as a mutually exclusive relationship, and 0 otherwise. , For the corresponding advertising event and The weight of the superedge. , For the main legal provisions in the corresponding advertising incident and Node and The mean confidence score of all mappings bound together; When the conflict degree is greater than the conflict threshold, it is determined that there is a substantial conflict between two sets of knowledge items; otherwise, there is no substantial conflict. When there is a substantial conflict, if the difference in the superedge weights of the two knowledge item sets is greater than the confidence difference threshold, the knowledge item set with the larger superedge weight is retained; otherwise, the two knowledge item sets are subject to manual review. After resolving conflicts, the set of knowledge entries corresponding to the semantic hypergraph of illegal advertising is extracted, and knowledge relationships are constructed to form a knowledge base of false and illegal advertising. The knowledge relationships include a set of knowledge nodes, a set of knowledge relationships, a set of node confidence scores, and a source and time series set. The knowledge nodes include aligned advertising entity nodes, advertising event nodes, legal provision nodes, and penalty result nodes. The knowledge relationships are transformed from hyperedges after relation disambiguation and include the main legal provision node, the associated penalty result node, the hyperedge weight, the average confidence score of the mapping binding, and the timestamp. The node confidence score is based on a single false and illegal advertisement. The source and time series set includes the version number of the database entry, the knowledge relationship, and the source advertisement identifier.

[0010] Furthermore, the method for performing dynamic evolution updates includes: Obtain multi-source heterogeneous advertising data and corresponding regulatory penalty annotations for false and illegal advertisements to be processed, construct a semantic hypergraph of illegal advertisements to be processed and the knowledge relationship to be processed, and calculate the alignment score between the semantic hypergraph of illegal advertisements to be processed and the historical semantic hypergraph of illegal advertisements. If the highest alignment score is less than the first update threshold, the knowledge relationship to be processed will be directly added to the knowledge base of false and illegal advertisements. Conversely, the historical knowledge relationship with the highest alignment score is taken as the knowledge relationship to be updated, the dynamic weight of the knowledge relationship to be updated is calculated, and the update strategy is selected according to the dynamic weight. The dynamic weights are determined by factors and mapping-bound confidence scores, expressed as follows: ; ; in For the dynamic weights of the knowledge relationships to be updated, For time indexing, The overall confidence level of knowledge relationships. Time decay factor, The entry time for the knowledge relationships to be updated. For the half-life of knowledge, For hotspot reinforcement coefficient, As an indicator function, if the knowledge nodes in the knowledge relation This belongs to the set of high-frequency illegal patterns in the current period. Then the indicator function takes the value 1. , , The confidence weighting coefficient is... The reliability score for the source of the knowledge relationship to be updated. For all mappings of knowledge relationships to be updated, bind to a set. For the first One advertising entity, For the first One legal provision, To establish confidence levels for mapping advertising entities to legal provisions. The current citation frequency of the knowledge relationship to be updated. The highest reference frequency of knowledge relationships in the historical database; The update strategy is as follows: when the dynamic weight of the knowledge relationship to be updated is greater than the second update threshold, the knowledge relationship to be processed and the knowledge relationship to be updated are merged and expanded; otherwise, the knowledge relationship to be processed and the knowledge relationship to be updated are merged to remove duplicates.

[0011] Secondly, a knowledge base construction system for false and illegal advertisements based on big data includes: The semantic processing module is used to collect historical multi-source heterogeneous advertising data and corresponding regulatory penalty annotations, and to perform multimodal semantic alignment and joint representation learning on the multi-source heterogeneous advertising data to obtain advertising semantic representation vectors. Extraction module: used to perform hierarchical knowledge extraction from the semantic representation vector of advertisements to obtain advertising elements, and to extract legal elements from regulatory penalty annotations; Semantic Hypergraph Module: Used to map and bind advertising elements with legal elements to construct a semantic hypergraph of advertising violations; The knowledge base module is used to perform entity alignment, relation disambiguation, and conflict resolution on the semantic hypergraph of illegal advertising, construct a knowledge base of false and illegal advertising, and dynamically evolve and update the knowledge base of false and illegal advertising based on time decay and mapping binding confidence.

[0012] The beneficial effects of this invention are: This invention relates to a method and system for constructing a knowledge base for false and illegal advertisements based on big data. Compared with existing technologies, this invention has the following technical advantages: This invention uses multimodal semantic alignment and joint representation learning to model various heterogeneous data types in a unified manner. It can fully explore and utilize the complementarity and consistency between different modal information to generate more discriminative and robust advertising semantic representation vectors, laying a solid foundation for subsequent accurate knowledge extraction. This invention extracts factual and legal elements through a hierarchical knowledge extraction framework and constructs a semantic hypergraph of advertising violations through a mapping and binding mechanism. This enables advertising violation nodes in the knowledge base to be directly linked to specific legal provisions and penalties, improving the interpretability of the law and generating evidence chains to assist law enforcement judgments. This greatly enhances the usability of the knowledge base in advertising supervision practice. This invention uses a hypergraph structure to express complex, high-order relationships such as "multiple advertising elements are mapped to a single legal element" or "an advertisement violates multiple laws". Through entity alignment, relationship disambiguation and conflict resolution steps, it automatically identifies and unifies contradictory information from different data sources, effectively improving the factual accuracy and logical consistency of knowledge in the knowledge base. This invention employs a dynamic evolution update strategy based on time decay and mapping-bound confidence to perform differentiated updates to the knowledge base. This enables the knowledge base to respond quickly and adapt to new and mutated false and illegal advertising methods, significantly reducing manual maintenance costs and ensuring the long-term vitality of the knowledge base. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the steps of the method for constructing a knowledge base for false and illegal advertisements based on big data, as described in this invention. Detailed Implementation

[0014] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0015] The present invention relates to a method and system for constructing a knowledge base for false and illegal advertisements based on big data, comprising the following steps: like Figure 1 As shown, this embodiment includes the following steps: Collect historical multi-source heterogeneous advertising data and corresponding regulatory penalty annotations, and perform multimodal semantic alignment and joint representation learning on the multi-source heterogeneous advertising data to obtain advertising semantic representation vectors; Hierarchical knowledge extraction is performed on the semantic representation vector of advertisements to obtain advertising elements, and legal elements are extracted from regulatory penalty annotations. By mapping and binding advertising elements with legal elements, a semantic hypergraph of advertising violations is constructed; Entity alignment, relation disambiguation, and conflict resolution are performed on the semantic hypergraph of illegal advertising to construct a knowledge base for false and illegal advertising. The knowledge base of false and illegal advertisements is dynamically evolved and updated based on time decay and mapping-bound confidence. The multi-source heterogeneous advertising data includes apparent internet advertising text, advertising images, keyframe sequences of advertising videos, and DOM structure data of advertising landing pages; The regulatory penalty label includes advertising laws and regulations and the penalty results; The advertising elements include the advertising entity, the advertising relationship, and the advertising event.

[0016] In this embodiment, the method for obtaining the semantic representation vector of an advertisement includes: Collect historical multi-source heterogeneous advertising data and corresponding regulatory penalty annotations; Multi-source heterogeneous advertising data is preprocessed to obtain multimodal initial semantic vectors; the preprocessing operations include encoding advertising text through a pre-trained language model to obtain text semantic vectors, encoding advertising images through a hierarchical visual Transformer to obtain image semantic vectors, encoding the keyframe sequence of advertising video frame by frame and then aggregating it temporally to obtain video keyframe semantic vectors, and encoding the DOM tree from bottom to top through a tree-shaped long short-term memory network to obtain DOM structure semantic vectors. The initial multimodal semantic vectors are mapped to a unified cross-modal semantic space using a learnable projection matrix, and the semantic proximity between different modalities of the same advertisement sample is enhanced by a cross-modal alignment loss, as expressed in the following expression: ; in For cross-modal alignment loss function, , for Sample index of training batch, For modal index set, For image indexing, For video keyframe indexing, For DOM structure indexing, For text semantic vectors, For modal semantic vectors, For temperature coefficient, For positive sample pairs, For negative sample pairs; The semantic consistency score among the initial semantic vectors of the multimodal systems after modal alignment is calculated using a consistency metric function and expressed as follows: ; in The semantic consistency score between the text / image modality and other modalities. This is the consistency metric weight vector. , This is the initial semantic vector of the multimodal expressions after modal alignment. This is a consistency bias vector; After modal alignment, the initial multimodal semantic vector and semantic consistency score are gated and fused to obtain a multimodal fused semantic vector. Then, feature concatenation and projection are performed to obtain the advertising semantic representation vector. ; ; in For multimodal fusion semantic vectors, For advertising semantic representation vectors, For the semantic projection matrix of advertisements, This is the projection bias vector; In actual assessment, regulatory penalty records are extracted from the historical regulatory penalty database, and an index alignment is established between the advertising sample and the regulatory penalty record. The penalty record is bound to the corresponding advertising sample through the advertising publisher's identifier (advertiser's unified social credit code or platform store ID) and the publication timestamp. Based on the penalty records, the original page data streams of the advertising sample publishing platforms (e-commerce platform product promotion pages, social media information flow ads, search engine promotion results, and short video platform ad placement pages) are collected. The advertising content is crawled through distributed collection nodes, separated and stored according to modal type to form an original set of advertising samples. The original set of advertising samples includes original advertising text, original advertising images, original advertising videos, and original HTML documents. The original ad text is encoded using UTF-8 standardization, special characters are cleaned, duplicate text is removed, stop words are filtered, and word segmentation is performed to obtain the ad text. The original ad images are formatted, resolution is normalized, and pixel values ​​are standardized. Low-quality images are then screened out to obtain the ad images. The original ad video is temporally sampled to obtain a candidate frame sequence. The color histogram difference between adjacent frames is calculated to detect content abrupt changes. These abrupt changes are identified as content keyframes to obtain the ad video keyframe sequence. The original HTML document is input, the DOM tree structure is parsed, node tag types, node text content, CSS style properties, and node hierarchy are extracted, script nodes and style nodes are filtered, and semantic nodes (text blocks, image links, buttons, forms) are retained. The DOM structure data of the ad landing page is then output. Since each modal encoder is trained independently, its semantic space distribution is heterogeneous. It is necessary to map it to a unified cross-modal semantic space through a learnable projection matrix, and introduce cross-modal alignment loss to enhance the semantic proximity between different modalities of the same advertisement sample. The temperature parameter is set to 0.07. False and illegal advertisements often exhibit a semantic discrepancy between text and image, where the text claims medical efficacy but the image is merely that of ordinary food. To explicitly model this cross-modal semantic inconsistency, a semantic consistency score is calculated using a consistency metric function. A semantic consistency score closer to 0 indicates a higher degree of semantic discrepancy between the corresponding modalities. Furthermore, a text-image consistency constraint loss is introduced to enhance the model's ability to discriminate semantically discrepancies during training. The expression is as follows: ; in For the image-text consistency constraint loss, These are modal consistency labels, where 1 indicates that the text and image are semantically consistent, and 0 indicates that they are divergent. Text-image semantic consistency score; The multimodal fusion semantic vector is obtained by gating the fusion of other modalities with the text modality. When a modality deviates semantically from the text modality (low consistency score), the weight of that modality in the final representation is reduced, and the text semantics are used as a substitute basis to prevent false images from misleading the representation. By using gated fusion, the initial semantic vectors of other modalities are jointly represented to learn the initial semantic vector of the text. Finally, feature concatenation and projection are performed to obtain the semantic representation vector of the advertisement.

[0017] In this embodiment, the method for hierarchical knowledge extraction to obtain advertising elements includes: The advertising semantic representation vector is injected into the modally aligned text semantic vector, and a conditional random field is used to jointly decode the entity boundary and type. The word-level tag score matrix is ​​calculated to obtain the advertising entity. The entity representation of the advertising entity includes entity type, text span, and semantic embedding. The entity interaction representation vector is obtained by concatenating the semantic embedding, interaction features and advertising semantic representation vectors of each advertising entity, calculating the relationship type probability and determining the advertising relationship. An event type query attention mechanism is used for event structure identification. Entity representations are aggregated to obtain event semantic vectors. The probability of existence for each event type is calculated to determine the advertising event. The expression is as follows: ; ; in For the first Event types for entities Attention weights For event query projection matrix, For the first Learnable query vectors for various event types As a feature dimension of the event semantic space, For the first The event semantic vector of a type of event, This represents the total number of entities identified in the advertising sample. For the first Entity representation vectors of each entity; In practical evaluation, the semantic representation vector of the advertisement is injected into the modally aligned semantic vector of the text to enhance the entity recognition's ability to perceive cross-modal contexts. The expression is as follows: ; in This is a word-level context feature matrix. This is the sample index for the training batch. This is the word-level projection feature matrix of the advertising text (text semantic features after cross-modal projection). This is the global context projection matrix; Conditional random fields are used for joint decoding of entity boundaries and types. A word-level tag score matrix is ​​calculated to obtain the advertising entities (word-level tags exceeding a specified score are selected). The expression is: ; in This is a word-level tag score matrix. For the entity label projection matrix, For entity label bias vectors, The length of the word sequence in the advertising text. It is a collection of advertising entity types, including the advertising entity, goods / services, claimed efficacy, price, target audience, and publishing medium; The sequence labeling probability of a conditional random field is: ; in For the sequence labeling probability of a conditional random field, Label the entity with a sequence of tags. The partition function (normalization factor) is used to sum over all possible label sequences to normalize the probability. For the first Tags of each word Transfer to the Tags of each word Tag score, For from the first Tags of each word Transfer to the Tags of each word The transfer score, The tag transition matrix characterizes the transition compatibility between adjacent lexical tags. For the identified entities, the average pooling of their internal lexical context features is taken as the entity semantic embedding. The sequence labeling probability of the conditional random field is calculated, and finally the set of advertising entities is output. Each entity contains entity type, text span and semantic embedding. The semantic embedding, interaction features and advertising semantic representation vectors of each advertising entity are concatenated to obtain the entity interaction representation vector. The relationship type probability is calculated to determine the advertising relationship type, which includes claim-efficacy, product-belongs to category, subject-publishing advertisement, price-corresponding product, audience-applicable product, efficacy-comparison with competitors, and use-before and after effect. The interaction representation vector between each advertising entity is represented as follows: , representing entities and Interactive representation, , For entities and Semantic embedding, This represents the interaction features between two semantic embeddings. The contextual projection matrix represents the advertising relationships; these relationships include claims (efficacy), products (category), entities (advertising entities), prices (corresponding products), and target audiences (applicable products). Ad relationships are selected based on relationship type probability, with a probability threshold of 0.75. The relationship type probability is: ; in For relation type probabilities, Projection matrix for classifying relations, The relation classification bias vector; Advertising events are selected based on the probability of their occurrence, with a probability threshold of 0.7. The probability of each event type occurring is as follows: ; in There is a probability for each event type. For the first Types of events for The entity set of the advertising sample This is the weight vector for determining the existence of events. Bias for determining the existence of events; Output the product of word-level tag scores and sequence label probabilities of the conditional random field, the probability of relation type, the probability of event type existence, and the advertising entities, advertising relationships, and advertising events that are greater than the corresponding probability thresholds; The legal provisions in the regulatory penalty annotations are semantically encoded to obtain the semantic embedding of the legal provisions. The penalty results (penalty type, penalty amount, penalty authority) in the regulatory penalty annotations are structured into vectors and then projected through a multilayer perceptron to obtain the penalty result embedding. The legal elements are composed of the semantic embedding of the legal provisions and the penalty result embedding.

[0018] In this embodiment, the method for constructing an advertising violation semantic hypergraph includes: The semantic embedding of advertising entities is mapped to the legal semantic space. The cosine similarity between the semantic embedding of advertising entities and the semantic embedding of legal provisions in the same case of false and illegal advertising is calculated. The cosine similarity is normalized as the mapping binding confidence. The binding relationship between each advertising entity and each legal provision is established based on the mapping binding confidence to determine the binding pairs. The global mapping binding set is output. Extract the advertising entities contained in each advertising event, filter the binding pairs corresponding to the advertising entities in each advertising event, and form an event mapping binding subset; The advertising entity and advertising event are treated as advertising domain nodes, and the legal provisions and penalties are treated as legal domain nodes. Based on the event mapping binding subset, ad entity nodes and legal provision nodes within the same ad event are aggregated into a hyperedge. The attention weight of the ad semantic representation vector to each ad entity representation vector is calculated, and the hyperedge weight is calculated by combining the mapping binding confidence of each binding pair. The expression is as follows: ; ; in for The super-edge weight of advertising events for The event mapping of the advertising event is bound to a subset. For the first One advertising entity, For the first One legal provision, To establish confidence levels for mapping advertising entities to legal provisions. For the overall representation of advertising With the representation of advertising entities Attention weights To query the projection matrix, The key projection matrix, For the attention head dimension, for The set of advertising entity nodes for an advertising event; The advertising domain nodes, legal domain nodes, hyperedges, and hyperedge weights are assembled into an advertising violation semantic hypergraph. In actual evaluation, the advertising entity embedding and the legal provision embedding are located in different semantic spaces (ad content space and legal regulation space). It is necessary to use a learnable projection matrix to map the advertising entity embedding to the legal semantic space so that the cross-domain similarity calculation has semantic comparability. The mapping binding confidence of the semantic embedding of the advertising entity and the semantic embedding of the legal provision in the same case of false and illegal advertisement is calculated. The mapping binding confidence is obtained by normalizing the cosine similarity based on temperature scaling Softmax (temperature parameter is 0.1). The advertising entity with the mapping binding confidence greater than the confidence threshold (0.6) is bound to the legal provision, and the binding is rejected otherwise (to avoid noisy association). The binding pair includes the advertising entity, the legal provision and the mapping binding confidence. Taking the binding of two sets of advertising entities and legal provisions as an example, for the advertising entity "XX brand slimming capsules" (type of "goods and services") and the legal provision "Article 28 of the Advertising Law" (identification of false advertising - false efficacy), the normalized cosine similarity of the corresponding semantic embedding is calculated to be 0.87, which is greater than 0.6, so the corresponding advertising entity and legal provision are bound together; for the advertising entity "natural extracts" (type of "claimed ingredients") and the legal provision "Article 28 of the Advertising Law" (identification of false advertising - false efficacy), the normalized cosine similarity of the corresponding semantic embedding is calculated to be 0.36, which is less than 0.6, so the binding of the corresponding advertising entity and legal provision is rejected. The characterization of false and illegal advertisements is based on "events" as the basic unit (e.g., "a certain entity claims that a certain product has medical effects" constitutes an efficacy verification event). The global mapping binding set needs to be aggregated according to advertising events, so that each event only carries the legal provisions bound to its internal entities, ensuring the local consistency of the hyperedge semantics. Taking the advertising event as the core organizational unit, the advertising entity nodes within the event and their successfully bound legal provision nodes are aggregated into a hyperedge, breaking through the limitations of traditional binary edges and realizing the expression of complex illegal semantic structures of "multi-entity-multi-legal provisions". The semantic hypergraph of advertising violations only contains advertising entities and legal provisions that meet the binding relationship, and each legal provision is directly associated with the corresponding penalty result. For example, the description of an advertising event: "A certain health technology company (the advertising entity) promoted 'XX brand weight loss capsules' (product / service) on an e-commerce platform, claiming '20 pounds weight loss in 30 days' (claimed efficacy). The advertisement showed 'before and after comparison photos of consumer Xiao Li' (user testimonials) and marked 'suitable for postpartum obesity' (applicable population)." Hierarchical knowledge extraction identified the following advertising entities: A1 "a certain health technology company (advertising entity)", A2 "XX brand weight loss capsules (product / service)", A3 "claimed efficacy '20 pounds weight loss in 30 days' (claimed efficacy)", A4 "before and after comparison photos (user testimonials)", and A5 "postpartum obesity group (applicable group)". Mapping and binding were performed: A1 was bound to Article 56 of the Advertising Law (Advertiser's Responsibility) with a binding confidence level of 0.87; A2 and A3 were bound to Article 28 of the Advertising Law (False Advertising - False Efficacy) with a binding confidence level of 0.71; A2 was bound to Article 17 of the Advertising Law (Advertisements for Non-Medical Products Prohibit Disease Treatment Functions) with a binding confidence level of 0.81; A4 was bound to Article 28 of the Advertising Law (False Advertising - Fictitious User Testimonials) with a binding confidence level of 0.64; A5 was bound to Article 46 of the Advertising Law (Regulations on Advertising to Special Populations) with a binding confidence level of 0.63. All legal provisions were associated with penalty results (penalty type / fine, amount / 200,000 RMB, agency / market supervision bureau). There are a total of 11 nodes in the advertising event (5 advertising entity nodes, 1 advertising event node, and 4 legal provisions nodes), including 5 binding pairs. The same advertising entity A2 is simultaneously bound to two legal provisions B1 / B2. According to the event mapping binding subset, the advertising entity nodes and legal provision nodes in the same advertising event are aggregated into a hyperedge, and the corresponding hyperedge weight is calculated to be 0.1467. The advertising domain nodes, legal domain nodes, hyperedges, and hyperedge weights are assembled into an advertising violation semantic hypergraph.

[0019] In this embodiment, the method for constructing a knowledge base for false and illegal advertisements includes: Obtain the semantic hypergraph of different false and illegal advertisements. Calculate the alignment score of the advertisement domain nodes based on the semantic embedding of the nodes and the hypergraph structure position. Perform alignment decisions and cluster merging based on the alignment scores. The expression is: ; in For advertising entities and Corresponding ad domain node and Alignment score, For balance coefficient, , For advertising domain nodes and semantic embedding, , For advertising domain nodes and The associated set of hyperedge indices; Calculate the competition priority of each legal provision node corresponding to the hyperedge in the semantic hypergraph of advertising violations, and perform relation disambiguation according to the competition priority. The expression is: ; in For legal provisions Corresponding legal provisions node Competition and cooperation priorities For advertising events Sino-French Term Mean confidence score for all mappings For the application of legal provisions in historical regulatory penalties The average penalty amount For advertising events The index of legal provisions in the text This is the time-related decay coefficient. For the current moment, For legal provisions The most recent revision time, , , These are the relational weight coefficients; The method for disambiguating relationships is as follows: select the legal node with the highest priority of competition and cooperation as the main legal node, set legal nodes with a priority greater than the competition and cooperation retention threshold as competition and cooperation related legal nodes and retain them in the super edge, and otherwise remove the remaining legal nodes from the super edge. The semantic hypergraph of advertising violations after entity alignment and relation disambiguation is transformed into a set of knowledge entries; the set of knowledge entries includes advertising entity nodes, main legal provision nodes, associated penalty result nodes, hyperedge weights, and timestamps; The conflict degree is calculated for any two sets of knowledge items, expressed as follows: ; in A collection of knowledge items and The degree of conflict For the corresponding set of advertising entity nodes and Jaccard similarity between them For mutually exclusive indicator functions of legal ontology, when the main legal provisions of two sets of knowledge entries... and In legal terms, it takes the value 1 when defined as a mutually exclusive relationship, and 0 otherwise. , For the corresponding advertising event and The weight of the superedge. , For the main legal provisions in the corresponding advertising incident and Node and The mean confidence score of all mappings bound together; When the conflict degree is greater than the conflict threshold, it is determined that there is a substantial conflict between two sets of knowledge items; otherwise, there is no substantial conflict. When there is a substantial conflict, if the difference in the superedge weights of the two knowledge item sets is greater than the confidence difference threshold, the knowledge item set with the larger superedge weight is retained; otherwise, the two knowledge item sets are subject to manual review. After resolving conflicts, the set of knowledge entries corresponding to the semantic hypergraph of illegal advertising is extracted, and knowledge relationships are constructed to form a knowledge base of false and illegal advertising. The knowledge relationships include a set of knowledge nodes, a set of knowledge relationships, a set of node confidence scores, and a source and time series set. The knowledge nodes include aligned advertising entity nodes, advertising event nodes, legal provision nodes, and penalty result nodes. The knowledge relationships are transformed from hyperedges after relation disambiguation and include the main legal provision node, the associated penalty result node, the hyperedge weight, the average confidence score of the mapping binding, and the timestamp. The node confidence score is based on a single false and illegal advertisement. The source and time series set includes the version number of the database entry, the knowledge relationship, and the source advertisement identifier. In actual evaluation, heterogeneous expressions of the same advertiser, the same product, or the same efficacy claim may appear in different advertising samples (e.g., "XX weight loss capsules" and "XX slimming capsules" refer to the same product). It is necessary to determine that advertising entity nodes with similar semantics and structural associations in the advertising violation semantic hypergraph are the same real-world objects and merge them to eliminate knowledge redundancy. First, calculate the semantic embedding vector similarity and hypergraph structural similarity of the advertising domain nodes corresponding to the advertising entities in different advertising violation semantic hypergraphs. Perform weighted fusion to obtain the alignment score of the advertising domain nodes. If the alignment score is greater than the entity alignment threshold, it is determined to be the same advertising entity. Then, according to the weight of the associated hyperedge, the same advertising entities in different advertising violation semantic hypergraphs are weighted and aggregated into a unified advertising entity. Otherwise, the original advertising domain nodes are retained. Among them, the intersection of the legal provisions mapped and bound to the same advertising entity is taken as the legal domain node of the new hyperedge. The balance coefficient (semantic and structural balance) is set at 0.6, and the sample ad X semantic super... Figure 1 Semantic supersemantics of sample ad Y Figure 1 For example, the semantic cosine similarity is 0.993 and the hypergraph structure similarity is 0.75, corresponding to an alignment score of 0.8958 (greater than the threshold of 0.85). This means that the advertising semantic hypergraphs X1 and Y1 are the same advertising entity. The nodes are merged and the intersection of the two hyperedges is taken as the binding law set. A single hyperedge (the same advertising event) may contain multiple legal provisions at the same time (e.g., an advertising event violates both "Article 17 of the Advertising Law - Restrictions on Medical Advertising" and "Article 28 of the Advertising Law - False Advertising"). It is necessary to resolve the conflict and cooperation based on the priority of the application of the legal provisions to determine the main applicable legal provision and avoid multiple unapplicable legal provisions corresponding to the same illegal event in the knowledge base. Taking the relationship weight coefficients of 0.5 / 0.3 / 0.2 and the competition retention threshold of 0.65, and using the advertising event "an advertisement for 'XX brand probiotic powder' on an e-commerce platform, claiming to 'enhance immunity, treat constipation, and prevent colds'" as an example, we perform mapping and binding. The same superedge contains Article 17 of the Advertising Law (non-medical products are prohibited from involving therapeutic functions) / the confidence level of the legal binding mapping is 0.85; Article 28 of the Advertising Law (false efficacy) / the confidence level of the legal binding mapping is 0.92; Article 46 of the Advertising Law (regulation of special groups) / the confidence level of the legal binding mapping is 0.32; and Article 9 of the Advertising Law (advertisements shall not contain any disparaging information about others) / the confidence level of the legal binding mapping is 0.05. The corresponding competition priorities are calculated to be 0.8311 / 0.941 / 0.4648 / 0.3810 respectively. B2 is taken as the main rule, B1 is kept in the super edge, and B3 / B4 are taken out. After alignment and disambiguation, with a conflict threshold of 0.1, there may still be logically contradictory knowledge entries in the knowledge base of false and illegal advertisements (such as the same advertiser in samples from similar time periods, one super-edge is labeled as "legal advertising" and the other is labeled as "false advertising"). Conflict detection and resolution need to be carried out through legal ontology constraints and confidence differences. The following are three examples of conflict detection and resolution in the knowledge base of false and illegal advertisements. For the first pair of knowledge items Z1 / Z4, the hyperedge weights are 0.24 / 0.15 and the average mapping-binding confidence is 0.88 / 0.68. For the second pair of knowledge items Z1 / Z4, the hyperedge weights are 0.32 / 0.18 and the average mapping-binding confidence is 0.93 / 0.8. For the third pair of knowledge items Z1 / Z4, the hyperedge weights are 0.27 / 0.25 and the average mapping-binding confidence is 0.85 / 0.86. The corresponding conflict degree is calculated to be 0 / 0.1536 / 0.01305. Therefore, it is determined that there is no substantial conflict between the first pair of knowledge entries, and both knowledge entries can coexist in the knowledge base. When it is determined that there is a substantial conflict between the second pair of knowledge entries and the third pair of knowledge entries (the indicator function of the third pair of knowledge entries shows mutual exclusion). Calculate the difference in hyperedge weights for the two sets of knowledge items. In the second pair of knowledge items, the difference in hyperedge weights of 0.14 is greater than the confidence difference threshold of 0.1. The knowledge item with a hyperedge weight of 0.32 is retained, and the knowledge item with a hyperedge weight of 0.18 is removed. In the third pair of knowledge items, the difference in hyperedge weights of 0.024 is less than the confidence difference threshold of 0.1. Manual review is required. In the knowledge base of false and illegal advertisements, each false and illegal advertisement generates a knowledge relation (as a storage unit), and the knowledge relations form the knowledge base of false and illegal advertisements; The node confidence score is calculated from the semantic hypergraph of all illegal advertisements contained in a single fake and illegal advertisement, expressed as follows: ; in For false and illegal advertisements The set of node confidence scores, For false and illegal advertisements The knowledge base node set, For nodes The node confidence, The weight of the hyperedge to which the node is located. The mean confidence level is bound to the mappings involved in the node. When a node is used as the penalty result node, the average confidence score of its associated legal provisions is taken. In the same mapping and binding relationship, the sets of mapping and binding relationships involved in the advertising entity node and the legal provision node are different, so the node confidence scores are inconsistent.

[0020] In this embodiment, the method for performing dynamic evolution updates includes: Obtain multi-source heterogeneous advertising data and corresponding regulatory penalty annotations for false and illegal advertisements to be processed, construct a semantic hypergraph of illegal advertisements to be processed and the knowledge relationship to be processed, and calculate the alignment score between the semantic hypergraph of illegal advertisements to be processed and the historical semantic hypergraph of illegal advertisements. If the highest alignment score is less than the first update threshold, the knowledge relationship to be processed will be directly added to the knowledge base of false and illegal advertisements. Conversely, the historical knowledge relationship with the highest alignment score is taken as the knowledge relationship to be updated, the dynamic weight of the knowledge relationship to be updated is calculated, and the update strategy is selected according to the dynamic weight. The dynamic weights are determined by factors and mapping-bound confidence scores, expressed as follows: ; ; in For the dynamic weights of the knowledge relationships to be updated, For time indexing, The overall confidence level of knowledge relationships. Time decay factor, The entry time for the knowledge relationships to be updated. For the half-life of knowledge, For hotspot reinforcement coefficient, As an indicator function, if the knowledge nodes in the knowledge relation This belongs to the set of high-frequency illegal patterns in the current period. Then the indicator function takes the value 1. , , The confidence weighting coefficient is... The reliability score for the source of the knowledge relationship to be updated. For all mappings of knowledge relationships to be updated, bind to a set. For the first One advertising entity, For the first One legal provision, To establish confidence levels for mapping advertising entities to legal provisions. The current citation frequency of the knowledge relationship to be updated. The highest reference frequency of knowledge relationships in the historical database; The update strategy is as follows: when the dynamic weight of the knowledge relationship to be updated is greater than the second update threshold, the knowledge relationship to be processed and the knowledge relationship to be updated are merged and expanded; otherwise, the knowledge relationship to be processed and the knowledge relationship to be updated are merged to remove duplicates. In the actual evaluation, the time decay factor was set to 0.5, the knowledge half-life to 180 days, the hotspot reinforcement coefficient to 0.3, the confidence weight coefficient to 0.4 / 0.4 / 0.2, the first update threshold to 0.65, and the second update threshold to 1.2. Three sets of false and illegal advertisements to be processed were given, and the alignment scores with the illegal semantic hypergraphs of each historical advertisement in the knowledge base were calculated. The highest corresponding values ​​were 0.32 / 0.89 / 0.88. The highest alignment score of the semantic hypergraph between the first group of pending false and illegal advertisements and historical advertisements is 0.3, which is less than the first update threshold of 0.65. Therefore, the corresponding knowledge relationship to be processed is directly added to the knowledge base of false and illegal advertisements. The highest alignment score of the semantic hypergraph between the second group of pending false and illegal advertisements and historical advertisements is 0.89, which is greater than the first update threshold of 0.65. The corresponding historical knowledge relationship is taken as the knowledge relationship to be updated. The dynamic weight of the knowledge relationship to be updated is calculated to be 1.1, which is less than the second update threshold of 1. According to the rules, the knowledge relationship to be processed and the knowledge relationship to be updated are merged and expanded. The meaning of fusion and expansion is to merge the new information in the knowledge relationship to be processed (such as newly emerging illegal variant expressions) with the existing knowledge in the knowledge relationship to be updated to form a more complete knowledge expression. The highest alignment score of the semantic hypergraph between the third group of pending false and illegal advertisements and historical advertisements is 0.88, which is greater than the first update threshold of 0.65. The corresponding historical knowledge relationship is taken as the knowledge relationship to be updated. The dynamic weight of the knowledge relationship to be updated is calculated to be 0.91, which is less than the second update threshold of 1. According to the rules, the knowledge relationship to be processed and the knowledge relationship to be updated are merged and deduplicated. The meaning of merging and deduplication is that when the same or highly similar illegal knowledge entries already exist in the historical knowledge relationship, the knowledge relationship to be processed is only taken as a new instance of the entry without changing the knowledge expression of the knowledge entry itself. At the same time, entities, relationships or binding information that are duplicated with the historical knowledge relationship in the knowledge relationship to be processed are checked and removed to avoid knowledge redundancy.

[0021] Secondly, a knowledge base construction system for false and illegal advertisements based on big data includes: The semantic processing module is used to collect historical multi-source heterogeneous advertising data and corresponding regulatory penalty annotations, and to perform multimodal semantic alignment and joint representation learning on the multi-source heterogeneous advertising data to obtain advertising semantic representation vectors. Extraction module: used to perform hierarchical knowledge extraction from the semantic representation vector of advertisements to obtain advertising elements, and to extract legal elements from regulatory penalty annotations; Semantic Hypergraph Module: Used to map and bind advertising elements with legal elements to construct a semantic hypergraph of advertising violations; The knowledge base module is used to perform entity alignment, relation disambiguation, and conflict resolution on the semantic hypergraph of illegal advertising, construct a knowledge base of false and illegal advertising, and dynamically evolve and update the knowledge base of false and illegal advertising based on time decay and mapping binding confidence.

[0022] 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 method for constructing a knowledge base for false and illegal advertisements based on big data, characterized in that, Includes the following steps: S1. Collect historical multi-source heterogeneous advertising data and corresponding regulatory penalty annotations, and perform multimodal semantic alignment and joint representation learning on the multi-source heterogeneous advertising data to obtain advertising semantic representation vectors; S2. Hierarchical knowledge extraction is performed on the semantic representation vector of the advertisement to obtain the advertising elements, and legal elements in the regulatory penalty annotation are extracted. S3. Map and bind advertising elements with legal elements to construct a semantic hypergraph of advertising violations; S4. Perform entity alignment, relation disambiguation, and conflict resolution on the semantic hypergraph of illegal advertising to construct a knowledge base for false and illegal advertising. S5. Dynamically evolve and update the knowledge base of false and illegal advertisements based on time decay and mapping binding confidence. The multi-source heterogeneous advertising data includes apparent internet advertising text, advertising images, keyframe sequences of advertising videos, and DOM structure data of advertising landing pages; The regulatory penalty label includes advertising laws and regulations and the penalty results; The advertising elements include advertising entities, advertising relationships, and advertising events; The method for constructing the advertising violation semantic hypergraph includes: The semantic embedding of advertising entities is mapped to the legal semantic space. The cosine similarity between the semantic embedding of advertising entities and the semantic embedding of legal provisions in the same case of false and illegal advertising is calculated. The cosine similarity is normalized as the mapping binding confidence. The binding relationship between each advertising entity and each legal provision is established based on the mapping binding confidence to determine the binding pairs. The global mapping binding set is output. Extract the advertising entities contained in each advertising event, filter the binding pairs corresponding to the advertising entities in each advertising event, and form an event mapping binding subset; The advertising entity and advertising event are treated as advertising domain nodes, and the legal provisions and penalties are treated as legal domain nodes. Based on the event mapping binding subset, ad entity nodes and legal provision nodes within the same ad event are aggregated into a hyperedge. The attention weight of the ad semantic representation vector to each ad entity representation vector is calculated, and the hyperedge weight is calculated by combining the mapping binding confidence of each binding pair. The expression is as follows: ； ； in for The super-edge weight of advertising events for The event mapping of the advertising event is bound to a subset. For the first One advertising entity, For the first One legal provision, To establish confidence levels for mapping advertising entities to legal provisions. For the overall representation of advertising With the representation of advertising entities Attention weights To query the projection matrix, The key projection matrix, For the attention head dimension, for The set of advertising entity nodes for an advertising event; The advertising domain nodes, legal domain nodes, hyperedges, and hyperedge weights are assembled into an advertising violation semantic hypergraph.

2. The method for constructing a knowledge base for false and illegal advertisements based on big data according to claim 1, characterized in that, The method for obtaining the semantic representation vector of an advertisement includes: Collect historical multi-source heterogeneous advertising data and corresponding regulatory penalty annotations; Multi-source heterogeneous advertising data is preprocessed to obtain multimodal initial semantic vectors; the preprocessing operations include encoding advertising text through a pre-trained language model to obtain text semantic vectors, encoding advertising images through a hierarchical visual Transformer to obtain image semantic vectors, encoding the keyframe sequence of advertising video frame by frame and then aggregating it temporally to obtain video keyframe semantic vectors, and encoding the DOM tree from bottom to top through a tree-shaped long short-term memory network to obtain DOM structure semantic vectors. The initial multimodal semantic vectors are mapped to a unified cross-modal semantic space using a learnable projection matrix, and the semantic proximity between different modalities of the same advertisement sample is enhanced by a cross-modal alignment loss, as expressed in the following expression: ； in For cross-modal alignment loss function, , for Sample index of training batch, For modal index set, For image indexing, For video keyframe indexing, For DOM structure indexing, For text semantic vectors, For modal semantic vectors, For temperature coefficient, For positive sample pairs, For negative sample pairs; The semantic consistency score among the initial semantic vectors of the multimodal systems after modal alignment is calculated using a consistency metric function and expressed as follows: ； in The semantic consistency score between the text / image modality and other modalities. This is the consistency metric weight vector. , This is the initial semantic vector of the multimodal expressions after modal alignment. This is a consistency bias vector; After modal alignment, the initial multimodal semantic vector and semantic consistency score are gated and fused to obtain a multimodal fused semantic vector. Then, feature concatenation and projection are performed to obtain the advertising semantic representation vector. ； ； in For multimodal fusion semantic vectors, For advertising semantic representation vectors, For the semantic projection matrix of advertisements, This is the projection bias vector.

3. The method for constructing a knowledge base for false and illegal advertisements based on big data according to claim 1, characterized in that, The method for hierarchical knowledge extraction to obtain advertising elements includes: The advertising semantic representation vector is injected into the modally aligned text semantic vector, and a conditional random field is used to jointly decode the entity boundary and type. The word-level tag score matrix is ​​calculated to obtain the advertising entity. The entity representation of the advertising entity includes entity type, text span, and semantic embedding. The entity interaction representation vector is obtained by concatenating the semantic embedding, interaction features and advertising semantic representation vectors of each advertising entity, calculating the relationship type probability and determining the advertising relationship. An event type query attention mechanism is used for event structure identification. Entity representations are aggregated to obtain event semantic vectors. The probability of existence for each event type is calculated to determine the advertising event. The expression is as follows: ； ； in For the first Event types for entities Attention weights For event query projection matrix, For the first Learnable query vectors for various event types As a feature dimension of the event semantic space, For the first The event semantic vector of a type of event, This represents the total number of entities identified in the advertising sample. For the first The entity representation vector of each entity.

4. The method for constructing a knowledge base for false and illegal advertisements based on big data according to claim 1, characterized in that, The method for constructing a knowledge base for false and illegal advertisements includes: Obtain the semantic hypergraph of different false and illegal advertisements. Calculate the alignment score of the advertisement domain nodes based on the semantic embedding of the nodes and the hypergraph structure position. Perform alignment decisions and cluster merging based on the alignment scores. The expression is: ； in For advertising entities and Corresponding ad domain node and Alignment score, For balance coefficient, , For advertising domain nodes and semantic embedding, , For advertising domain nodes and The associated set of hyperedge indices; Calculate the competition priority of each legal provision node corresponding to the hyperedge in the semantic hypergraph of advertising violations, and perform relation disambiguation according to the competition priority. The expression is: ； in For legal provisions Corresponding legal provisions node Competition and cooperation priorities For advertising events Sino-French Term Mean confidence score for all mappings For the application of legal provisions in historical regulatory penalties The average penalty amount For advertising events The index of legal provisions in the text This is the time-related decay coefficient. For the current moment, For legal provisions The most recent revision time, , , These are the relational weight coefficients; The method for disambiguating relationships is as follows: select the legal node with the highest priority of competition and cooperation as the main legal node, set legal nodes with a priority greater than the competition and cooperation retention threshold as competition and cooperation related legal nodes and retain them in the super edge, and otherwise remove the remaining legal nodes from the super edge. The semantic hypergraph of advertising violations after entity alignment and relation disambiguation is transformed into a set of knowledge entries; the set of knowledge entries includes advertising entity nodes, main legal provision nodes, associated penalty result nodes, hyperedge weights, and timestamps; The conflict degree is calculated for any two sets of knowledge items using the following expression: ； in A collection of knowledge items and The degree of conflict For the corresponding set of advertising entity nodes and Jaccard similarity between them For mutually exclusive indicator functions of legal ontology, when the main legal provisions of two sets of knowledge entries... and In legal terms, it takes the value 1 when defined as a mutually exclusive relationship, and 0 otherwise. , For the corresponding advertising event and The weight of the superedge. , For the main legal provisions in the corresponding advertising incident and Node and The mean confidence score of all mappings bound together; When the conflict degree is greater than the conflict threshold, it is determined that there is a substantial conflict between two sets of knowledge items; otherwise, there is no substantial conflict. When there is a substantial conflict, if the difference in the superedge weights of the two knowledge item sets is greater than the confidence difference threshold, the knowledge item set with the larger superedge weight is retained; otherwise, the two knowledge item sets are subject to manual review. After resolving conflicts, the set of knowledge entries corresponding to the semantic hypergraph of illegal advertising is extracted, and knowledge relationships are constructed to form a knowledge base of false and illegal advertising. The knowledge relationships include a set of knowledge nodes, a set of knowledge relationships, a set of node confidence scores, and a source and time series set. The knowledge nodes include aligned advertising entity nodes, advertising event nodes, legal provision nodes, and penalty result nodes. The knowledge relationships are transformed from hyperedges after relation disambiguation and include the main legal provision node, the associated penalty result node, the hyperedge weight, the average confidence score of the mapping binding, and the timestamp. The node confidence score is based on a single false and illegal advertisement. The source and time series set includes the version number of the database entry, the knowledge relationship, and the source advertisement identifier.

5. The method for constructing a knowledge base for false and illegal advertisements based on big data according to claim 1, characterized in that, The method for dynamic evolution and updating includes: Obtain multi-source heterogeneous advertising data and corresponding regulatory penalty annotations for false and illegal advertisements to be processed, construct a semantic hypergraph of illegal advertisements to be processed and the knowledge relationship to be processed, and calculate the alignment score between the semantic hypergraph of illegal advertisements to be processed and the historical semantic hypergraph of illegal advertisements. If the highest alignment score is less than the first update threshold, the knowledge relationship to be processed will be directly added to the knowledge base of false and illegal advertisements. Conversely, the historical knowledge relationship with the highest alignment score is taken as the knowledge relationship to be updated, the dynamic weight of the knowledge relationship to be updated is calculated, and the update strategy is selected according to the dynamic weight. The dynamic weights are determined by factors and mapping-bound confidence scores, expressed as follows: ； ； in For the dynamic weights of the knowledge relationships to be updated, For time indexing, The overall confidence level of knowledge relationships. Time decay factor, The entry time for the knowledge relationships to be updated. For the half-life of knowledge, For hotspot reinforcement coefficient, As an indicator function, if the knowledge nodes in the knowledge relation This belongs to the set of high-frequency illegal patterns in the current period. Then the indicator function takes the value 1. , , The confidence weighting coefficient is... The reliability score for the source of the knowledge relationship to be updated. For all mappings of knowledge relationships to be updated, bind to a set. For the first One advertising entity, For the first One legal provision, To establish confidence levels for mapping advertising entities to legal provisions. The current citation frequency of the knowledge relationship to be updated. The highest reference frequency of knowledge relationships in the historical database; The update strategy is as follows: when the dynamic weight of the knowledge relationship to be updated is greater than the second update threshold, the knowledge relationship to be processed and the knowledge relationship to be updated are merged and expanded; otherwise, the knowledge relationship to be processed and the knowledge relationship to be updated are merged to remove duplicates.

6. A system for constructing a knowledge base for false and illegal advertisements based on big data, used to perform the method described in any one of claims 1-5, characterized in that, include: The semantic processing module is used to collect historical multi-source heterogeneous advertising data and corresponding regulatory penalty annotations, and to perform multimodal semantic alignment and joint representation learning on the multi-source heterogeneous advertising data to obtain advertising semantic representation vectors. Extraction module: used to perform hierarchical knowledge extraction from the semantic representation vector of advertisements to obtain advertising elements, and to extract legal elements from regulatory penalty annotations; Semantic Hypergraph Module: Used to map and bind advertising elements with legal elements to construct a semantic hypergraph of advertising violations; The knowledge base module is used to perform entity alignment, relation disambiguation, and conflict resolution on the semantic hypergraph of illegal advertising, construct a knowledge base of false and illegal advertising, and dynamically evolve and update the knowledge base of false and illegal advertising based on time decay and mapping binding confidence.

Images