Compliance decision-making support method and system for foreign-related enterprises based on knowledge graph reasoning

Abstract

The foreign-related enterprise compliance decision-making assistance method and system based on knowledge graph reasoning provided by the present invention relate to the field of data and knowledge graph technology. The present invention obtains compliance data related to global advertising, obtains a foreign-related advertising compliance knowledge graph after preprocessing; obtains information related to foreign-related enterprise product advertising, and constructs multimodal heterogeneous fusion data after data preprocessing, feature extraction and knowledge fusion processing; then based on Cypher query and knowledge graph reasoning, by calculating the violation risk confidence, obtains the interpretability result of knowledge graph reasoning; combines the pre-defined Cypher statement to perform sensitive word query, thereby obtaining the longest multi-hop query graph of sensitive words, and visualizes it; performs semantic similarity subgraph matching calculation on the sensitive word multi-hop reasoning graph and the foreign-related advertising compliance knowledge graph, and obtains compliant and accurately expressed alternative words; finally outputs the foreign-related advertising compliance plan. The present invention improves the efficiency and standardization of foreign-related enterprises in handling foreign advertising copywriting decisions.

Description

Technical Field

[0001] The present invention relates to the field of big data and knowledge graph technology, and in particular to a foreign-related enterprise compliance decision-making support method and system based on knowledge graph reasoning. Background Art

[0002] As domestic foreign-related enterprises frequently carry out cross-border business activities, their multilingual publicity materials such as product information and advertising copy have become an important means to shape brand image and expand overseas markets. However, due to differences in laws, regulations, industry standards, cultural customs, etc. in various countries, companies are prone to compliance issues in external publicity, such as using inappropriate sensitive words and violating local advertising regulations, which may not only damage the brand image, but also face the risk of legal proceedings.

[0003] Traditional compliance review mainly relies on manual review of laws, regulations, cases and professional judgment, which is labor-intensive, inefficient and easily affected by subjective factors, making it difficult to adapt to the real-time and accurate compliance needs in the context of globalization. The use of big data and knowledge graph technology can achieve real-time and accurate compliance needs, but the existing technology has the following main shortcomings:

[0004] 1) Insufficient richness of knowledge graph corpus: The existing system has failed to fully build a dedicated knowledge graph covering global advertising compliance knowledge, and lacks effective integration and structured expression of multi-source heterogeneous data such as international laws and regulations, industry standards, and company policies. 2) Difficulty in cross-language knowledge integration: There are more than 7,000 languages ​​in the world, and export-oriented companies trade with multiple language families, multiple ethnic groups, and multiple languages. Except for a few languages, most languages ​​have only a small amount of or even no annotated data. 3) Lack of explainability: When making compliance judgments, existing systems often lack a transparent display of the reasoning process, making it difficult for users to understand the basis for the system's judgment, affecting the credibility and acceptance of the decision. 4) Imperfect compliance plan recommendations: For identified sensitive words, existing systems often lack targeted compliance alternative word recommendation functions, which is not conducive to companies quickly correcting advertising texts that do not comply with regulations.

[0005] In view of this, the applicant filed this application after studying the existing technology. Summary of the invention

[0006] The present invention aims to provide a foreign-related enterprise compliance decision-making assistance method and system based on knowledge graph reasoning, so as to enable domestic foreign-related enterprises to make deeper reasoning decisions on possible illegal information in product information and advertising copy during external publicity, making the review of foreign-related enterprises' advertising compliance more intelligent, improving efficiency and standardization, and reducing manual labor intensity.

[0007] In order to solve the above technical problems, the present invention is implemented through the following technical solutions:

[0008] A foreign-related enterprise compliance decision-making assistance method based on knowledge graph reasoning, including:

[0009] S1, obtain global advertising-related compliance data, and perform document structuring and knowledge extraction preprocessing to obtain a foreign-related advertising compliance knowledge graph;

[0010] S2, obtains information related to product advertisements of foreign-related enterprises, outputs word fusion vectors after data preprocessing, feature extraction and knowledge fusion processing, and constructs multi-modal heterogeneous fusion data;

[0011] S3, according to the foreign-related advertising compliance knowledge graph and the word fusion vector, based on Cypher query and knowledge graph reasoning, a multi-hop reasoning graph of sensitive words is obtained, and the violation risk confidence of the multi-hop reasoning graph of sensitive words is calculated to obtain the interpretability result of the knowledge graph reasoning;

[0012] S4, based on the interpretability result of the knowledge graph reasoning, combined with the pre-defined Cyp statement, perform a sensitive word query on the foreign-related advertising compliance knowledge graph to obtain all multi-hop query graphs corresponding to the sensitive words, thereby obtaining the longest multi-hop query graph of sensitive words and visualizing it;

[0013] S5, performing semantic similarity subgraph matching calculation on the multi-hop reasoning graph of sensitive words and the foreign-related advertising compliance knowledge graph, and obtaining compliant and accurate alternative words in combination with a set threshold;

[0014] S6, outputting the foreign-related advertising compliance plan of the foreign-related enterprise based on the interpretable results of the knowledge graph reasoning and the compliant and accurately expressed alternative words.

[0015] Preferably, the global advertising-related compliance data includes advertising industry norms, advertising legal documents and advertising violation cases; the foreign-related enterprise product advertising-related information includes company information, product information and advertising copy information.

[0016] Preferably, the steps of document structuring and knowledge extraction preprocessing specifically include:

[0017] Clean and remove duplicates from the data and process missing values ​​to ensure data quality and consistency;

[0018] Classify and annotate data through text classification and keyword recognition and extraction technology;

[0019] Constructing a corpus, the corpus comprising: a sensitive word vocabulary, a legal document information table, a legal content information table, a company information table, a product information table, an advertising copy table, and a violation case table;

[0020] According to the corpus, the entities, relationships and attributes of the foreign-related compliance knowledge graph are constructed, and the information extraction technology is used to extract and cross-language align the foreign-related advertising compliance knowledge to obtain the foreign-related advertising compliance knowledge graph; wherein, the entities include sensitive words, alternative words, laws and regulations, legal documents, violation cases, advertising copy, product information and company information; the relationships include the legal provisions, violation cases and legal documents corresponding to the entities; the attributes include risk level, legal provision link, penalty scope and language type.

[0021] Preferably, the S2 specifically includes:

[0022] Performing data preprocessing on the acquired information related to the product advertisements of foreign-related enterprises, wherein the data preprocessing includes: converting traditional and simplified Chinese characters, converting full-width and half-width characters, removing stop words and special characters;

[0023] Extract key features from the pre-processed information related to foreign-related enterprise product advertisements, obtain corresponding key features and attributes, and construct structured data and unstructured data;

[0024] The word embedding method is used to vectorize structured data and unstructured data to obtain the word embedding vector of the corresponding information;

[0025] The word embedding vector is sent to the autoencoder to extract the latent semantic vector, and the word fusion vector is output, so as to obtain multimodal heterogeneous fusion data.

[0026] Preferably, constructing structured data specifically includes:

[0027] Use keyword extraction algorithms to extract thematic keywords and key descriptions of information related to foreign-related enterprise product advertisements as key features and attributes of structured data;

[0028] Perform text segmentation on the preprocessed information related to foreign-related enterprise product advertisements to obtain the corresponding word sequence:

[0029] W i =(w i1 ,w i2 ,...w ij ,...,w im );

[0030] Among them, W i represents the word sequence of the i-th text paragraph, m is the number of words in the i-th text paragraph; w ij represents the jth word in the i-th text paragraph;

[0031] For each word w ij , construct a node v ij, construct the edges of nodes according to the common relationship between words; if two words co-occur in the same window, an undirected edge is established between them;

[0032] Initialize each node v ij The weight Q ij , and then iteratively update and calculate each node weight until convergence. The update calculation method of the node weight is:

[0033]

[0034] Among them, N(v ij ) is the node v ij The set of adjacent nodes, d is the damping coefficient; Q ik Represents node v ik The corresponding weights;

[0035] The node weights are sorted and the K nodes with the highest weights are selected as keywords. For each selected keyword, contextual information is extracted from the preprocessed foreign-related enterprise product advertisement related information as the corresponding key description, thereby obtaining structured data.

[0036] Preferably, S3 is specifically:

[0037] Input the word fusion vector into the foreign-related advertising compliance knowledge graph for Cypher query to obtain a fusion vector query graph for each word;

[0038] Performing subgraph aggregation on the fusion vector query graph to obtain an aggregate graph, selecting sensitive word nodes in the aggregate graph to obtain a sensitive word multi-hop reasoning graph, and calculating the violation risk confidence of the sensitive word multi-hop reasoning graph;

[0039] According to the violation risk confidence, the multi-hop reasoning graph and sensitive words corresponding to the set violation risk confidence threshold are output to obtain the interpretable result of knowledge graph reasoning.

[0040] Preferably, S4 is specifically:

[0041] According to the interpretability result of the knowledge graph reasoning, the inferred sensitive words are embedded into the pre-defined Cy statements, and then the embedded Cy statements are sent to the foreign-related advertising compliance knowledge graph for query, so as to obtain all multi-hop query graphs corresponding to the sensitive words, thereby obtaining the longest multi-hop query graph of sensitive words;

[0042] The longest multi-hop query graph of the sensitive word is visualized in the form of an information box.

[0043] Preferably, S5 is specifically:

[0044] The multi-hop reasoning graph of sensitive words is used to perform semantic similarity sub-graph matching calculation with the foreign-related advertising compliance knowledge graph to obtain multiple multi-hop reasoning graphs of alternative words, and the alternative words corresponding to the multi-hop reasoning graphs of alternative words that exceed the set threshold are output, which are compliant and accurately expressed alternative words.

[0045] Preferably, the foreign-related advertising compliance plan consists of advertising copy, sensitive words, legal documents, violation cases and the compliant and accurately expressed alternative words.

[0046] The present invention also provides a foreign-related enterprise compliance decision-making support system based on knowledge graph reasoning, comprising:

[0047] The foreign-related advertising compliance knowledge graph construction module is used to obtain global advertising-related compliance data, and perform document structuring and knowledge extraction preprocessing to obtain the foreign-related advertising compliance knowledge graph;

[0048] The multi-source heterogeneous knowledge fusion module is used to obtain information related to foreign-related enterprise product advertisements, perform knowledge fusion processing, output word fusion vectors, and construct multi-modal heterogeneous fusion data;

[0049] A knowledge graph reasoning module is used to obtain a multi-hop reasoning graph of sensitive words based on the foreign-related advertising compliance knowledge graph and the word fusion vector, based on Cypher query and knowledge graph reasoning, and obtain an interpretable result of knowledge graph reasoning by calculating the violation risk confidence of the multi-hop reasoning graph of sensitive words;

[0050] A sensitive word visualization display module is used to perform sensitive word query on the foreign-related advertising compliance knowledge graph based on the interpretability results of the knowledge graph inference and in combination with pre-defined Cyp statements, to obtain all multi-hop query graphs corresponding to sensitive words, thereby obtaining the longest multi-hop query graph of sensitive words and visualizing it;

[0051] A sensitive word replacement word generation module is used to perform semantic similarity subgraph matching calculation on the sensitive word multi-hop reasoning graph and the foreign-related advertising compliance knowledge graph, and obtain compliant and accurate replacement words in combination with a set threshold;

[0052] A compliance solution generation module is used to output a foreign-related advertising compliance solution based on the interpretable results of the knowledge graph reasoning and the compliant and accurately expressed alternative words.

[0053] The present invention also provides a foreign-related enterprise compliance decision-making assistance device based on knowledge graph reasoning, including a processor and a memory, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement a foreign-related enterprise compliance decision-making assistance method based on knowledge graph reasoning as described above.

[0054] The present invention also provides a computer-readable storage medium, on which computer-readable instructions are stored. When the computer-readable instructions are executed by a processor of a device where the computer-readable storage medium is located, a foreign-related enterprise compliance decision-making support method based on knowledge graph reasoning is implemented as described above.

[0055] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0056] The foreign-related advertising compliance knowledge graph constructed by the present invention contains global advertising-related compliance data, has rich knowledge graph corpus, and effectively integrates and structuredly expresses multi-source heterogeneous data such as international laws and regulations, industry standards, and company policies.

[0057] The method of the present invention is based on Cypher query and knowledge graph reasoning to make compliance judgments on information related to foreign-related enterprise product advertisements, and visualize the interpretability results of knowledge graph reasoning to enhance users' credibility and acceptance of knowledge graph reasoning, so as to enable foreign-related enterprises to make deeper reasoning decisions on illegal information that may exist in product information and advertising copy during external publicity.

[0058] At the same time, the present invention combines the semantic similarity matching technology with the identified sensitive words to recommend compliant and accurate alternative words to users, which is beneficial for foreign-related enterprises to quickly correct advertising texts that do not comply with regulations, thereby improving the efficiency and standardization of foreign-related enterprises in handling external advertising copy and reducing manual labor intensity. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0060] Figure 1 A schematic diagram of a foreign-related enterprise compliance decision-making support method based on knowledge graph reasoning provided in Example 1.

[0061] Figure 2 This is a model diagram of the foreign-related advertising compliance knowledge graph provided for Example 2.

[0062] Figure 3 This is a sample diagram of the foreign-related advertising compliance knowledge graph interface provided in Example 2.

[0063] Figure 4 This is a display diagram of the longest multi-hop query graph interface for sensitive words provided in Example 2.

[0064] Figure 5 This is a front-end page display diagram of the longest multi-hop query graph of sensitive words provided in Example 2.

[0065] Figure 6 This is a diagram of the interface of the alternative word multi-hop reasoning graph provided in Example 2.

[0066] Figure 7 This is a diagram showing the results of downloading files for the compliance solution provided in Example 2.

[0067] Figure 8 A schematic diagram of a foreign-related enterprise compliance decision-making support device based on knowledge graph reasoning provided in Example 3.

[0068] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention claimed for protection, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0070] Embodiment 1

[0071] Embodiment 1 of the present invention provides a foreign-related enterprise compliance decision-making assistance method based on knowledge graph reasoning, which can be implemented by a foreign-related enterprise compliance decision-making assistance device based on knowledge graph reasoning (hereinafter referred to as the foreign-related enterprise compliance decision-making assistance device), and in particular, is executed by one or more processors in the foreign-related enterprise compliance decision-making assistance device.

[0072] In this embodiment, the foreign-related enterprise compliance decision-making assistance device may be an electronic device equipped with a processor, which carries a computer program of the foreign-related enterprise compliance decision-making assistance method based on knowledge graph reasoning and the computer program can be executed, such as a computer, a smart phone, a smart tablet, a workstation, etc., which is not limited here.

[0073] like Figure 1 As shown, a foreign-related enterprise compliance decision-making support method based on knowledge graph reasoning includes steps S1 to S6.

[0074] S1, obtain global advertising-related compliance data, and perform document structuring and knowledge extraction preprocessing to obtain a foreign-related advertising compliance knowledge graph.

[0075] In this embodiment, the Internet is used to collect global advertising-related compliance data, including regulatory documents, industry standards, company policies, etc., to structure the advertising compliance corpus data and extract knowledge, and to construct a foreign-related advertising compliance knowledge graph. The specific steps of construction include:

[0076] In this embodiment, it is assumed that the global advertising-related compliance data collected on the Internet is divided into six categories (other categories can also be set as needed, which are not limited here), namely advertising industry specifications, advertising legal documents, advertising illegal cases, advertising copy, product information and company information, defined as C = {C1, C2, C3, C4, C5, C6}, and the global data source S global After collection, we get the advertising compliance corpus dataset D;

[0077] The specific process can be expressed as follows:

[0078]

[0079] The Collect function represents the collection of data from the global data source S global According to data category C i The process of collecting data.

[0080] The obtained raw data D is then cleaned, classified and annotated to be organized into an organized and manageable form. First, the data is cleaned and preprocessed, including deduplication and processing of missing values, to ensure the quality and consistency of the data. Then, the data is classified into corresponding categories through text classification and keyword recognition and extraction technology. Next, a complete corpus is constructed, which contains structured information such as sensitive word vocabulary, legal regulations document information table, legal regulations content information table, company information table, product information table, advertising copy table, violation case table, etc., providing rich data resources for the subsequent knowledge graph construction.

[0081] Further, this step is expressed as:

[0082] D structured =Structure(D)

[0084] Among them, the Structure function represents the process of cleaning, classifying and labeling the advertising compliance corpus dataset D to ensure the quality and organization of the data.

[0085] Next, design the model layer. Define the entities, relationships, and attributes in the knowledge graph to provide specifications and guidance for the structure of the entire knowledge graph. In this embodiment, this step designs conceptual entities such as sensitive words, alternative words, laws and regulations, legal documents, illegal cases, advertising copy, product information, company information, and other relationships such as related laws, related illegal cases, and affiliated legal documents, as well as attributes such as risk level, legal link, penalty scope, and language type. This model layer can clearly reflect the organizational structure and semantic relationship of foreign-related advertising compliance knowledge, which is convenient for subsequent knowledge extraction, alignment, and application.

[0086] Furthermore, the expression is expressed as:

[0087] G schema =DesignSchema(D structured );

[0088] The DesignSchema function is responsible for defining entities, relationships, and attributes to build the schema layer G of the knowledge graph. schema .

[0089] Then, the information extraction technology is used to extract foreign-related advertising compliance knowledge from the constructed corpus and the foreign-related advertising compliance knowledge graph model layer constructed by the model layer design module.

[0090] Furthermore, the expression is:

[0091] G extracted =ExtractKnowledge(D structured ,G schema );

[0092] Among them, ExtractKnowledge means using D structured Structured Data and G schema The pattern layer is used as input to extract knowledge and obtain the knowledge triple set G of the knowledge graph. extracted .

[0093] Specifically, the ExtractKnowledge algorithm uses the Padd leNLP natural language processing tool and the UIE extraction tool. First, Padd leNLP is used to preprocess and analyze the corpus, including tasks such as entity recognition and relationship extraction. The (existing) named entity recognition model is used to extract entity information such as company names and product names from the text, and the (existing) relationship extraction model is used to extract the association between entities, such as the relationship between products and companies, and which laws and regulations sensitive words violate. Then, through the UIE extraction tool, an extraction template for specific entities and relationships is designed and executed, and the D structuredExtract specific instance information. For example, for company information, design a template to extract company name, location, enterprise size, etc.; for product information, design a template to extract product name, description, price, etc.

[0094] Next, cross-language alignment of foreign-related advertising compliance knowledge is carried out, mainly including multi-language sensitive word alignment, multi-language advertising text alignment, multi-language company information alignment, multi-language product information alignment, etc.

[0095] Furthermore, the expression is:

[0096] G aligned =AlignKnowledge(G extracted ,AlignTypes);

[0097] Among them, G aligned Represents the aligned knowledge triple set; AlignKnowledge algorithm is used to extract the knowledge triple set G extracted Perform cross-language alignment. AlignTypes is a set of alignment types, including all types that need to be aligned, such as multilingual sensitive words, advertising text, company information, and product information.

[0098] Specifically, the AlignKnowledge algorithm adopts a knowledge alignment method based on multi-task learning. It uses entity alignment to obtain multilingual knowledge semantic information through simultaneous learning of multiple tasks such as entity alignment and attribute alignment, and uses entity attribute alignment to enrich the multilingual semantic information of entities.

[0099] The processed foreign-related compliance knowledge data is stored in the graphic database for subsequent query, analysis and application.

[0100] Furthermore, the expression is: KG advertisement =Store(G aligned );

[0101] The Store function is responsible for storing the aligned knowledge triple set G aligned Store in a graph database to build a foreign advertising compliance knowledge graph KG advertisement .

[0102] Specifically, in this embodiment, the graph database uses Neo4j, which is a graph database management system that stores data in a graphical manner and provides powerful graph query and analysis functions. It is suitable for storing complex entity and relationship structures. Among them, Py2neo is the Python driver of Neo4j, which provides an interface for interacting with the Neo4j database to facilitate the operation and management of the graph database. The Py2neo library is used to establish a connection with the Neo4j database, and the construction and management of the knowledge graph are implemented through Python code. By creating nodes and relationships, the aligned knowledge triple set G aligned Stored in a Neo4j database.

[0103] S2, obtains information related to product advertisements of foreign-related enterprises, outputs word fusion vectors after data preprocessing, feature extraction and knowledge fusion processing, and constructs multi-modal heterogeneous fusion data;

[0104] In this embodiment, data resources from different sources and different types, including company information, product information, and advertising copy, are integrated and uniformly processed, and knowledge fusion is performed on these data to construct multi-modal heterogeneous fusion data for subsequent reasoning and decision-making.

[0105] Furthermore, first, given the original corpus T data ,The text preprocessing technology is used to preprocess it. The preprocessing steps include:

[0106] 1) Traditional-Simplified Conversion: For Chinese information containing traditional or simplified characters, use HanLP’s traditional-simplified conversion rules to ensure that the text information is unified into the same writing system.

[0107] 2) Full-width and half-width character conversion: Convert full-width characters (such as Chinese full-width numbers and English full-width symbols) to corresponding half-width characters to ensure the consistency of character encoding.

[0108] 3) Stop word removal: Use a predefined stop word list to remove irrelevant or low-information words in the text through string matching.

[0109] 4) Remove special characters: Remove special symbols, blank characters and other non-text elements in the text, and only retain meaningful text information.

[0110] Get the preprocessed corpus The expression of the above process is as follows:

[0111]

[0112] Next, we use the keyword extraction algorithm to extract the subject keywords and key descriptions from the advertising copy, company information, and product information as the key features and attributes of the structured data. Perform text segmentation to obtain word sequence W i =(w i1 ,w i2 ,...,w im ), where m is the number of words in the i-th text paragraph. For each word w ij , construct a node v ij , construct the edges of nodes according to the co-occurrence relationship between words. If two words co-occur in the same window, an undirected edge is established between them.

[0113] Initialize each node v ij The weight Q ij =1, iteratively calculate node weights until convergence:

[0114]

[0115] Among them, N(v ij ) is the node v ij The adjacent node set, d is the damping coefficient, generally taken as 0.85. ik Represents node v ik Corresponding weights. According to the node weight sorting, select the K nodes with the highest weight as keywords. For each keyword, extract its context information from the original text as the key description, thereby constructing structured data and unstructured data.

[0116] Then, the preprocessed structured and unstructured data are converted into vector form for unified knowledge fusion.

[0117] Structured data representation process: For text fields in company information and product information structured data, such as company profile and product description, the Word2Vec word embedding method is applied to map each word into a low-dimensional dense vector. For the entire text, the average pooling method is used to aggregate the word vectors to obtain the vector representation of the text. The formula is as follows:

[0118] S word2vec (input)=avgPooling({Word2Vec(w ij )|w ij ∈input});

[0119] S word2vec(input) represents the vector representation of the text field of structured data after being converted by the Word2Vec word embedding method. input represents the input text field, such as a company profile or product description. ij Represents each word in the text field.

[0120] Word2Vec(w ij ) represents word w ij The Word2Vec vector representation of . avgPooling() represents the operation of aggregating vocabulary vectors using average pooling.

[0121] Unstructured data representation: For unstructured data of advertising copy, first use the word segmentation tool jieba to perform text segmentation to obtain a word sequence. Then, the word embedding method is also used for the word segmentation results to convert each word into a vector. Finally, the same aggregation strategy as structured data is used to generate the overall vector representation of the advertising copy. The following is the public notice:

[0122] NS word2vec (input′)=avgPooling({Word2Vec(w ij )||w ij ∈seg(input′)});

[0123] NS word2vec (input′) represents the unstructured data vector representation. input′ represents the input advertisement copy. seg(input′) represents the word segmentation operation on the advertisement copy.

[0124] Next, knowledge fusion is performed. This step uses an autoencoder (AE) to fuse the vector representations of multi-source heterogeneous knowledge, extract latent semantic vectors, and form a word fusion vector of the advertising text. The expression is as follows:

[0125] Z=AE([S word2vec , NS word2vec ]);

[0126] S word2vec Represents the vector representation of structured data, NS word2vec Represents the vector representation of unstructured data. AE() represents the autoencoder model, which receives the input vector and generates a latent semantic vector Z by learning the representation of multi-source heterogeneous knowledge. Z is the fused vector representation extracted by the autoencoder, which has a higher level of semantic information and is used to represent the word fusion vector of the advertising text.

[0127] S3, according to the foreign-related advertising compliance knowledge graph and the word fusion vector, based on Cypher query and knowledge graph reasoning, a multi-hop reasoning graph of sensitive words is obtained, and by calculating the violation risk confidence of the multi-hop reasoning graph of sensitive words, the interpretability result of the knowledge graph reasoning is obtained.

[0128] In this embodiment, this step is responsible for conducting compliance review of advertising texts based on the foreign-related advertising compliance knowledge graph and multimodal heterogeneous fusion data, including advertising text knowledge graph query and sensitive word reasoning graph generation.

[0129] Advertisement text knowledge graph query steps: Take the word fusion vector of each word in the advertisement text as the query vector, and use the Cy query language to perform similarity query in the knowledge graph. The query goal is to find the graph entity that is closest to the query word vector (i.e., the most semantically similar). The query statement is as follows:

[0130] match(e.Entity)where similarly(e.embedding, $query_embedding)>$similarityThreshold return e;

[0131] Among them, $query_embedding represents the fusion vector of the query term, $similarityThreshold represents the set similarity threshold, and the similarly function represents the similarity between the calculated graph entity vector e.embedding and the query vector, where e represents the queried knowledge graph entity.

[0132] Collect the entity list returned by the Cypher query, each entity corresponds to a word in the ad text, and treat each word and its corresponding query result entity as a node pair. The first node in the node pair represents the word in the ad text, and the second node represents the matching entity in the knowledge graph. The edge between the node pairs represents the query relationship and can carry the similarity score between the query word and the matching entity. In this way, a word fusion vector query graph reflecting the relationship between each word in the ad text and the matching entity in the knowledge graph is constructed. The formula is as follows:

[0133] G = (V, E);

[0134] V=

[0135] {(w i , e i )|for each word-entity pair(w i , e i )returned by the Cypherquery};

[0136] E={(v i , v j , similarlyScore)|for each nodes(v i , v j )and theircorrespondingsimilarityscore)};

[0137] Among them, V represents the node set and E represents the edge set.

[0138] Sensitive word reasoning graph generation steps: First, subgraph aggregation is performed from the above word fusion vector query graph G to obtain the aggregate graph V agg , according to the node set V marked as sensitive words in the aggregation graph multi The edge set E between the nodes of sensitive words multi Construct a multi-hop reasoning graph G for sensitive words multi =(V multi , E multi ), the confidence is calculated based on the confidence of the pre-trained model, and the violation risk confidence of the multi-hop reasoning graph is evaluated. multi ). The sensitive word nodes V whose confidence exceeds the threshold threshold sensitive As the sensitive word output, the expression is:

[0139] V sensitive = {v i ∈V multi |Confidence(G multi )>threshold};

[0140] Output=(V sensitive , G multi );

[0141] At the same time, the multi-hop reasoning graph of sensitive words corresponding to the sensitive words is output as the interpretable result of knowledge graph reasoning.

[0142] S4. Based on the interpretability results of the knowledge graph reasoning and in combination with the pre-defined Cyp statements, a sensitive word query is performed on the foreign-related advertising compliance knowledge graph to obtain all multi-hop query graphs corresponding to the sensitive words, thereby obtaining the longest multi-hop query graph of sensitive words and visualizing it.

[0143] In this embodiment, a foreign-related advertising compliance knowledge graph query is performed based on sensitive words, and the query results such as legal documents, regulations, and violation cases associated with the sensitive words are visualized to enhance the user's understanding and trust in the system's judgment.

[0144] First, embed sensitive words into Cypher statements. The sensitive words here are the output of S3 step reasoning, and create a sensitive word list, which contains all sensitive words to be queried.

[0145] Create a general Cypher query template for knowledge graph query. The label of the sensitive word node in the knowledge graph is defined as SensitiveWord. The following is an example of a Cypher-based query template used to obtain a multi-hop query graph for each sensitive word. It can also be set according to actual conditions:

[0146] match(s:SensitiveWord)-[*1..depth]-(n)where s.word in V sensitive returns,n;

[0147] s represents the sensitive word node found; SensitiveWord represents the label of the sensitive word node; S.word represents the attribute of the sensitive word node, which is used to match the sensitive word to be queried; [*1..depth] means starting from the sensitive word node, at most depth relationships can be skipped, and depth is the maximum depth of the foreign-related advertising compliance knowledge graph to reach the final node n. n represents the final node connected to the sensitive word node through a multi-hop relationship.

[0148] The constructed Cypher query statement is sent to the query service port of the foreign-related advertising compliance knowledge graph through the py2neo interface in Python and the query results returned by the knowledge graph are received. The returned query results are usually data in JSON format, containing multi-hop query graph information associated with sensitive words, such as nodes, relationships, attributes, etc.

[0149] Compare the lengths of these multi-hop query graphs for sensitive words, and select the longest one as the longest multi-hop query graph for sensitive words. The length is determined by the number of nodes in the traversal path.

[0150] Next, the knowledge graph is visualized. The raw data of the longest multi-hop query graph of sensitive words is converted into the format required by the front-end visualization component. The displayed information includes legal documents, legal regulations, violation cases, etc.

[0151] S5, performing semantic similarity subgraph matching calculation on the multi-hop reasoning graph of sensitive words and the foreign-related advertising compliance knowledge graph, and combining the set threshold to obtain compliant and accurately expressed alternative words.

[0152] In this step, the foreign-related advertising compliance knowledge graph is inferred based on sensitive words, and a reasoning path is provided for the reasoning as an explainable result, providing users with compliant and accurately expressed alternative words.

[0153] This step receives the results inferred from S3, including the sensitive word V sensitive and its multi-hop reasoning graph information G multi . Using the depth-first search subgraph matching algorithm MatchSubgraph, in the foreign-related advertising compliance knowledge graph G knowledge Find the multi-hop reasoning graph G with sensitive words sensitive Semantically similar subgraphs, set the similarity threshold to screen out potential alternative words multi-hop reasoning graph G sub It is expressed as:

[0154] G sub =MatchSubgraph(G sensitive ,G knowledge ,threshold);

[0155] At the same time, extract the alternative word node V from the matched subgraph alt , the expression is:

[0156] V alt =ExtractAlternativeWords(G sub );

[0157] And output all alternative words and their corresponding multi-hop reasoning graph G in the form of front-end information boxes alt :

[0158] G alt =output(V alt ,G sub );

[0159] Among them, ExtractAlternativeWords represents the function of extracting alternative word nodes; output represents the output function.

[0160] S6, outputting a foreign-related advertising compliance plan based on the interpretable results of the knowledge graph reasoning and the compliant and accurately expressed alternative words.

[0161] In this embodiment, according to the results of the knowledge graph reasoning step, the sensitive word explainability display step, and the sensitive word alternative word generation step, a compliance plan for foreign-related advertisements is generated with one click for users to download and view, including the compliance plan data structure design step and the compliance plan data integration step:

[0162] Compliance plan data structure design steps: According to the designed compliance plan data structure template, fill the pre-processed advertising copy, inferred sensitive words, legal documents, violation cases, generated alternative words, etc. into the template to form the compliance result.

[0163] Compliance scheme data integration steps: Use the CSV data exchange format to serialize the instantiated compliance scheme data structure, use the openpyxl library to write the serialized data to a file, and set a suitable file name for users to download.

[0164] In summary, the present invention has the following beneficial effects:

[0165] The foreign-related advertising compliance knowledge graph constructed by the present invention contains global advertising-related compliance data, has rich knowledge graph corpus, and effectively integrates and structuredly expresses multi-source heterogeneous data such as international laws and regulations, industry standards, and company policies.

[0166] The method of the present invention is based on Cypher query and knowledge graph reasoning to make compliance judgments on information related to foreign-related enterprise product advertisements, and visualize the interpretability results of knowledge graph reasoning to enhance users' credibility and acceptance of knowledge graph reasoning, so as to enable foreign-related enterprises to make deeper reasoning decisions on illegal information that may exist in product information and advertising copy during external publicity.

[0167] At the same time, the present invention combines the semantic similarity matching technology with the identified sensitive words to recommend compliant and accurate alternative words to users, which is beneficial for foreign-related enterprises to quickly correct advertising texts that do not comply with regulations, thereby improving the efficiency and standardization of foreign-related enterprises in handling external advertising copy and reducing manual labor intensity.

[0168] Embodiment 2

[0169] The global advertising-related compliance data provided by the second embodiment of the present invention includes legal documents, Amazon advertising policies, Arabic laws, Japanese advertising-related policies, WeChat public accounts, and the People's Republic of China Illegal Case Library. At the same time, legal documents are derived from laws and regulations and documents of various countries, including Amazon advertising policies, Japan Consumer Affairs Agency's product representation law, domestic laws and regulations, and relevant legal documents of various countries on the WIPO Lex website. Illegal cases are derived from the West Kool Law Library and the People's Republic of China Illegal Case Library. In addition, the product information and advertising copy information in the information related to foreign-related enterprise product advertising are derived from multiple channels such as the People's Republic of China Illegal Case Library, Japan Amazon website, domestic court document website, and product information introductions on domestic and foreign shopping websites. Company information comes from national enterprise credit information data, etc., as shown in the following table.

[0170]

[0171]

[0172] The acquired raw data is cleaned, classified and annotated to be organized and manageable. First, the data is cleaned and preprocessed, including deduplication and missing value processing, to ensure the quality and consistency of the data. Then, the data is classified into different categories through text classification and keyword recognition and extraction technology. Next, a complete corpus is constructed, which contains structured information such as sensitive word vocabulary, legal regulations document information table, legal regulations content information table, company information table, product information table, advertising copy table, violation case table, etc., providing rich data resources for the subsequent knowledge graph construction.

[0173] The model layer designs the steps to define the entities, relationships and attributes in the knowledge graph, providing specifications and guidance for the structure of the entire knowledge graph. The model layer designs conceptual entities such as sensitive words, alternative words, legal regulations, legal documents, illegal cases, advertising copy, product information, company information, and other relationships such as related laws, related illegal cases, and legal documents, and attributes such as risk level, legal link, penalty scope, and language type. Figure 2 The knowledge graph model can clearly reflect the organizational structure and semantic relationship of foreign-related advertising compliance knowledge, which is convenient for subsequent knowledge extraction, alignment and application.

[0174] In the knowledge extraction step, information extraction technology is used to extract foreign-related advertising compliance knowledge from the constructed corpus and the foreign-related advertising compliance knowledge graph model layer.

[0175] In the knowledge alignment step, cross-language alignment of foreign-related advertising compliance knowledge is carried out, mainly including multi-language sensitive word alignment, multi-language advertising text alignment, multi-language company information alignment, multi-language product information alignment, etc.

[0176] In the data storage step, the processed foreign-related compliance knowledge data is stored in the Neo4j graph database, such as Figure 3 for subsequent query, analysis and application.

[0177] In this embodiment, the multimodal heterogeneous fusion data part includes steps such as data preprocessing, feature extraction, multi-source heterogeneous knowledge representation, and knowledge fusion.

[0178] Define the original corpus of advertising copy as T data="ruly,Step into"Eternal Auras,"Wherewitchcraft seamlessly blends and………with modern healthcare. Our remedies, born of ancient knowledge and scientific research, lead to holistic well-being. Amidst a lively crowd, find tailored potions and firearms forrejuvenation.?!!! Embark on your wellness journey with us today.";

[0179] After text preprocessing, it is converted into And extract the key description as “healthcare”.

[0180] For structured data of company information and product information, word embedding method is used for vector representation; for unstructured data of advertising copy, text segmentation is first performed and then the segmentation results are represented by vectors through word embedding method.

[0181] Given product information = "Product Weight: 950g, Colour Name: Yellow, Product Length: 27cm, Product Height: 9cm, Product Width / Depth: 16cm";

[0182] The vector representation after Word2Vec word embedding is: [0.123, -0.456, 0.789, ..., 0.098, -0.567, 0.890];

[0183] Ad copy = "Challenge the limit, refuse excess! With our competitive-grade bow and arrow set, every time you draw the bow for a full moon, it is a rejection of excessive complexity and a pursuit of perfection."

[0184] After Jieba word segmentation and Word2Vec word embedding, the vector representation is: [[0.23,-0.406,0.989,……-0.235,0.678,-0.901,0.456],[0.456,0.789,-0.123,……,0.321,-0.741,0.235],……[-0.789,0.123,0.456,……,0.678,-0.901,0.456],[0.321,-0.890,0.741,……,0.456,0.123,0.789]].

[0185] The word embedding vectors of company information, product information, and advertising copy are fed into the autoencoder, and the latent semantic vector is extracted from the hidden layer of the autoencoder and output as the word fusion vector of the advertising text.

[0186] The word embedding vectors from three different sources are output by the autoencoder as the word fusion vector of the advertising text: [[0.986, 0.872, …, -0.877], [-0.567, 0.321, …, 0.741], …, [0.875, -0.231, …, -0.235]].

[0187] In the advertisement text knowledge graph query part, the word fusion vector of the advertisement text is sent to the foreign-related advertisement compliance knowledge graph for Cy query to obtain the fusion vector query graph of each word. The query graphs of words such as firearms, witchcraft, healthcare, holistic, blends, and rehabilitation are obtained from the word fusion vector.

[0188] In the sensitive word reasoning graph generation part, subgraph aggregation is performed on the word fusion vector query graph, sensitive word nodes in the aggregation graph are selected, the sensitive word multi-hop reasoning graph is obtained, and the violation risk confidence of the multi-hop reasoning graph is calculated. Sensitive words that exceed the violation risk confidence threshold are output and the sensitive word multi-hop reasoning graph is output as the interpretable result of knowledge graph reasoning.

[0189] The query graph obtained from the advertisement text knowledge graph query is aggregated, and the nodes labeled as sensitive words are selected. The confidence is further calculated, and finally the sensitive words exceeding the threshold are output, which are firearms, witchcraft, and healthcare.

[0190] In the sensitive word query part, the inferred sensitive words are embedded in the pre-defined Cy statements, and the Cy statements are sent to the foreign-related advertising compliance knowledge graph for query, to obtain all multi-hop query graphs of sensitive words, and to filter out the longest multi-hop query graph of sensitive words that contains the inferred multi-hop reasoning graph of sensitive words. Figure 4 As shown in the figure, we can see that the longest multi-hop query graph is from the firearms node to the healthcare node. The graph contains 36 nodes and 103 relationships. The healthcare node is a sensitive word node, which contains the English alternative word health protection and the Chinese alternative word medical insurance. It also contains cross-language Chinese sensitive words such as medical care, Japanese sensitive words, and Arabic sensitive words. The relevant laws violated are the "Amazon Ads Guidelines and Acceptance Policies".

[0191] In the knowledge graph visualization part, the longest multi-hop query graph of sensitive words is visualized in the form of information boxes on the front-end page. Figure 5 As shown, the decision support includes a sensitive word display area and a specific description area. The sensitive word display area marks the three sensitive words "firearms", "witchcraft" and "healthcare" in red. The specific description area describes the three sensitive words in detail, including the legal documents they violate, the source of the documents, the status of the documents, the issuer of the documents, the link of the documents, the name of the regulations, the content of the regulations, the scope of punishment, violation cases and other information.

[0192] like Figure 6 As shown, in this embodiment, the sensitive word witchcraft has 9 alternative words, four of which are in Chinese, namely, deception, sorcery, magic, and evil arts, and five are in English, namely, necromantic, necromancy, enchantment, blackart, and black art.

[0193] In this embodiment, Figure 7 As shown, the compliance plan generation section provides detailed information such as the basis for punishment, the content of the regulations, alternative words, etc. for users to view.

[0194] The compliance plan is designed as a data structure template that includes advertising copy, sensitive words, legal documents, violation cases, alternative words, etc. The instance data of advertising copy, sensitive words, legal documents, violation cases, alternative words, etc. are used to generate a compliance plan file and provide it to the front end for display or download.

[0195] The present invention improves the credibility and acceptance of users for knowledge graph reasoning, and enables foreign-related enterprises to make deeper reasoning decisions on illegal information that may exist in product information and advertising copy during external publicity. At the same time, it recommends compliant and accurate alternative words to users, which helps foreign-related enterprises to quickly correct advertising texts that do not comply with regulations, thereby improving the efficiency and standardization of foreign-related enterprises in handling external advertising copy and reducing manual labor intensity.

[0196] Embodiment 3

[0197] like Figure 8 As shown, the third embodiment of the present invention further provides a foreign-related enterprise compliance decision-making auxiliary device based on knowledge graph reasoning, including:

[0198] The foreign-related advertising compliance knowledge graph construction module is used to obtain global advertising-related compliance data, and perform document structuring and knowledge extraction preprocessing to obtain the foreign-related advertising compliance knowledge graph;

[0199] The multi-source heterogeneous knowledge fusion module is used to obtain information related to foreign-related enterprise product advertisements. After data preprocessing, feature extraction and knowledge fusion processing, it outputs word fusion vectors to construct multi-modal heterogeneous fusion data.

[0200] A knowledge graph reasoning module is used to obtain a multi-hop reasoning graph of sensitive words based on the foreign-related advertising compliance knowledge graph and the word fusion vector, based on Cypher query and knowledge graph reasoning, and obtain an interpretable result of knowledge graph reasoning by calculating the violation risk confidence of the multi-hop reasoning graph of sensitive words;

[0201] A sensitive word visualization display module is used to perform sensitive word query on the foreign-related advertising compliance knowledge graph based on the interpretability results of the knowledge graph inference and in combination with pre-defined Cyp statements, to obtain all multi-hop query graphs corresponding to sensitive words, thereby obtaining the longest multi-hop query graph of sensitive words and visualizing it;

[0202] A sensitive word replacement word generation module is used to perform semantic similarity subgraph matching calculation on the sensitive word multi-hop reasoning graph and the foreign-related advertising compliance knowledge graph, and obtain compliant and accurate replacement words in combination with a set threshold;

[0203] A compliance solution generation module is used to output a foreign-related advertising compliance solution based on the interpretable results of the knowledge graph reasoning and the compliant and accurately expressed alternative words.

[0204] Embodiment 4

[0205] The fourth embodiment of the present invention also provides a foreign-related enterprise compliance decision-making assistance device based on knowledge graph reasoning, which includes a memory and a processor, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement the foreign-related enterprise compliance decision-making assistance method based on knowledge graph reasoning as described above.

[0206] Embodiment 5

[0207] The fifth embodiment of the present invention also provides a computer-readable storage medium, on which computer-readable instructions are stored. When the computer-readable instructions are executed by the processor of the device where the computer-readable storage medium is located, the foreign-related enterprise compliance decision-making support method based on knowledge graph reasoning as mentioned above is implemented.

[0208] In the several embodiments provided in the embodiments of the present invention, it should be understood that the disclosed devices and methods can also be implemented in other ways. The above-described device and method embodiments are merely illustrative. In addition, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0209] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A foreign-related enterprise compliance decision-making assistance method based on knowledge graph reasoning, characterized in that: include: S1, obtain global advertising-related compliance data, and perform document structuring and knowledge extraction preprocessing to obtain a foreign-related advertising compliance knowledge graph; S2, obtains information related to product advertisements of foreign-related enterprises, outputs word fusion vectors after data preprocessing, feature extraction and knowledge fusion processing, and constructs multi-modal heterogeneous fusion data; S3, according to the foreign-related advertising compliance knowledge graph and the word fusion vector, based on Cypher query and knowledge graph reasoning, a multi-hop reasoning graph of sensitive words is obtained, and the violation risk confidence of the multi-hop reasoning graph of sensitive words is calculated to obtain the interpretability result of the knowledge graph reasoning; S4, based on the interpretability result of the knowledge graph reasoning, combined with the pre-defined Cyp statement, perform a sensitive word query on the foreign-related advertising compliance knowledge graph to obtain all multi-hop query graphs corresponding to the sensitive words, thereby obtaining the longest multi-hop query graph of sensitive words and visualizing it; S5, performing semantic similarity subgraph matching calculation on the multi-hop reasoning graph of sensitive words and the foreign-related advertising compliance knowledge graph, and combining the set threshold to obtain compliant and accurate alternative words; S6, outputting the foreign-related advertising compliance plan of the foreign-related enterprise based on the interpretable results of the knowledge graph reasoning and the compliant and accurately expressed alternative words.

2. According to claim 1, a foreign-related enterprise compliance decision-making auxiliary method based on knowledge graph reasoning is characterized in that The global advertising-related compliance data include advertising industry norms, advertising legal documents and advertising violation cases; the foreign-related enterprise product advertising-related information includes company information, product information and advertising copy information.

3. According to claim 1, a foreign-related enterprise compliance decision-making assistance method based on knowledge graph reasoning is characterized in that ,The steps of document structuring and knowledge extraction preprocessing specifically include: Clean and remove duplicates from the data and process missing values ​​to ensure data quality and consistency; Classify and annotate data through text classification and keyword recognition and extraction technology; Constructing a corpus, the corpus comprising: a sensitive word vocabulary, a legal document information table, a legal content information table, a company information table, a product information table, an advertising copy table, and a violation case table; According to the corpus, the entities, relationships and attributes of the foreign-related compliance knowledge graph are constructed, and the information extraction technology is used to extract and cross-language align the foreign-related advertising compliance knowledge to obtain the foreign-related advertising compliance knowledge graph; wherein, the entities include sensitive words, alternative words, laws and regulations, legal documents, violation cases, advertising copy, product information and company information; the relationships include the legal provisions, violation cases and legal documents corresponding to the entities; the attributes include risk level, legal provision link, penalty scope and language type.

4. According to claim 1, a foreign-related enterprise compliance decision-making assistance method based on knowledge graph reasoning is characterized in that , the S2 specifically includes: Performing data preprocessing on the acquired information related to the product advertisements of foreign-related enterprises, wherein the data preprocessing includes: converting traditional and simplified Chinese characters, converting full-width and half-width characters, removing stop words and special characters; Extract key features from the pre-processed information related to foreign-related enterprise product advertisements, obtain corresponding key features and attributes, and construct structured data and unstructured data; The word embedding method is used to vectorize structured data and unstructured data to obtain the word embedding vector of the corresponding information; The word embedding vector is sent to the autoencoder to extract the latent semantic vector, and the word fusion vector is output, so as to obtain multimodal heterogeneous fusion data.

5. According to claim 4, a foreign-related enterprise compliance decision-making assistance method based on knowledge graph reasoning is characterized in that ,Constructing structured data specifically is: Use keyword extraction algorithms to extract thematic keywords and key descriptions of information related to foreign-related enterprise product advertisements as key features and attributes of structured data; Perform text segmentation on the preprocessed information related to foreign-related enterprise product advertisements to obtain the corresponding word sequence: IN i =(in i1 ,In i2 ,...In ij ,...,In im ); Among them, W i represents the word sequence of the i-th text paragraph, m is the number of words in the i-th text paragraph; w ij represents the jth word in the i-th text paragraph; For each word w ij , construct a node v ij , construct the edges of nodes according to the common relationship between words; if two words co-occur in the same window, an undirected edge is established between them; Initialize each node v ij The weight Q ij , and then iteratively update and calculate each node weight until convergence. The update calculation method of the node weight is: Among them, N(v ij ) is the node v ij The set of adjacent nodes, d is the damping coefficient; Q ik Represents node v ik The corresponding weights; The node weights are sorted and the K nodes with the highest weights are selected as keywords. For each selected keyword, contextual information is extracted from the preprocessed foreign-related enterprise product advertisement related information as the corresponding key description, thereby obtaining structured data.

6. According to claim 1, a foreign-related enterprise compliance decision-making assistance method based on knowledge graph reasoning is characterized in that , the S3 is specifically: Input the word fusion vector into the foreign-related advertising compliance knowledge graph for Cypher query to obtain a fusion vector query graph for each word; Performing subgraph aggregation on the fusion vector query graph to obtain an aggregate graph, selecting sensitive word nodes in the aggregate graph to obtain a sensitive word multi-hop reasoning graph, and calculating the violation risk confidence of the sensitive word multi-hop reasoning graph; According to the violation risk confidence, the multi-hop reasoning graph and sensitive words corresponding to the set violation risk confidence threshold are output to obtain the interpretable result of knowledge graph reasoning.

7. According to claim 6, a foreign-related enterprise compliance decision-making assistance method based on knowledge graph reasoning is characterized in that , the S4 is specifically: According to the interpretability result of the knowledge graph reasoning, the inferred sensitive words are embedded into the pre-defined Cy statements, and then the embedded Cy statements are sent to the foreign-related advertising compliance knowledge graph for query, so as to obtain all multi-hop query graphs corresponding to the sensitive words, thereby obtaining the longest multi-hop query graph of sensitive words; The longest multi-hop query graph of the sensitive word is visualized in the form of an information box.

8. According to claim 7, a foreign-related enterprise compliance decision-making assistance method based on knowledge graph reasoning is characterized in that , the S5 is specifically: The multi-hop reasoning graph of sensitive words is used to perform semantic similarity sub-graph matching calculation with the foreign-related advertising compliance knowledge graph to obtain multiple multi-hop reasoning graphs of alternative words, and the alternative words corresponding to the multi-hop reasoning graphs of alternative words that exceed the set threshold are output, which are compliant and accurately expressed alternative words.

9. According to claim 8, a foreign-related enterprise compliance decision-making assistance method based on knowledge graph reasoning is characterized in that The foreign-related advertising compliance plan consists of advertising copy, sensitive words, legal documents, violation cases and compliant and accurate alternative words.

10. A foreign-related enterprise compliance decision-making support system based on knowledge graph reasoning, characterized in that: include: The foreign-related advertising compliance knowledge graph construction module is used to obtain global advertising-related compliance data, and perform document structuring and knowledge extraction preprocessing to obtain the foreign-related advertising compliance knowledge graph; The multi-source heterogeneous knowledge fusion module is used to obtain information related to foreign-related enterprise product advertisements. After data preprocessing, feature extraction and knowledge fusion processing, it outputs word fusion vectors to construct multi-modal heterogeneous fusion data. A knowledge graph reasoning module is used to obtain a multi-hop reasoning graph of sensitive words based on the foreign-related advertising compliance knowledge graph and the word fusion vector, based on Cypher query and knowledge graph reasoning, and obtain an interpretable result of knowledge graph reasoning by calculating the violation risk confidence of the multi-hop reasoning graph of sensitive words; A sensitive word visualization display module is used to perform sensitive word query on the foreign-related advertising compliance knowledge graph based on the interpretability results of the knowledge graph inference and in combination with pre-defined Cyp statements, to obtain all multi-hop query graphs corresponding to sensitive words, thereby obtaining the longest multi-hop query graph of sensitive words and visualizing it; A sensitive word replacement word generation module is used to perform semantic similarity subgraph matching calculation on the sensitive word multi-hop reasoning graph and the foreign-related advertising compliance knowledge graph, and obtain compliant and accurate replacement words in combination with a set threshold; A compliance solution generation module is used to output a foreign-related advertising compliance solution based on the interpretable results of the knowledge graph reasoning and the compliant and accurately expressed alternative words.

Images