Multimodal sensitive word processing method, device, storage medium and computer equipment

Abstract

The present disclosure relates to the field of natural language processing technology and can be applied to the fields of financial public opinion monitoring, compliance review, and risk prevention and control. It provides a multimodal sensitive word processing method, device, storage medium, and computer equipment, including: obtaining information to be published, the information type of which includes at least one of images, text, audio, and video, and a pre-built knowledge base; identifying sensitive words in the information to be published based on a large language model and a sensitive word knowledge base to obtain an identification result; when there is sensitive content in the information to be published, determining corresponding replacement information based on the enterprise-level knowledge base, and the replacement information is used to replace the sensitive content in the information to be published to obtain the target published information; when there is no sensitive content in the information to be published, determining the information to be published as the target published information. This embodiment realizes automated compliance review of cross-modal content, effectively reduces the cost of manual review, and improves the accuracy and timeliness of sensitive information processing.

Description

Technical Field

[0001] The present disclosure relates to the technical field of natural language processing, and in particular to a multimodal sensitive word processing method, apparatus, storage medium, and computer equipment. Background Art

[0002] With the development of network technology and informatization, cyberspace has become a primary means of information dissemination, economic activities, social governance, and international communication, transforming people's lives and work patterns. However, the widespread use of the internet has also brought new challenges to content regulation. How to effectively filter out inappropriate and sensitive content while ensuring the free flow of information has become a key step in the information dissemination process.

[0003] In the financial sector, the rise of the "Internet + Finance" model has heightened requirements for the security and compliance of financial information, emphasizing the accuracy and efficiency of information processing and content distribution. While various sensitive word management tools exist, these technologies remain insufficient in the financial sector. For example, these technologies are often limited to sensitive word recognition in textual formats and struggle to cope with the complex environments of multimodal data such as images, audio, and video, limiting their effectiveness. Summary of the Invention

[0004] The embodiments of the present disclosure at least provide a multimodal sensitive word processing method, apparatus, storage medium, and computer equipment. By constructing a multimodal fusion sensitive word recognition system, automated compliance review of cross-modal content is achieved, effectively reducing manual review costs and improving the accuracy and timeliness of sensitive information processing.

[0005] The present disclosure provides a multimodal sensitive word processing method, including:

[0006] Acquire information to be published and a pre-built knowledge base; wherein the information type in the information to be published includes at least one of images, text, audio, and video; and the knowledge base includes a sensitive word knowledge base and an enterprise-level knowledge base;

[0007] performing sensitive word identification on the information to be published based on the large language model and the sensitive word knowledge base to obtain a sensitive word identification result; wherein the sensitive word identification result includes whether the information to be published contains sensitive content that matches the sensitive word knowledge base;

[0008] If the sensitive word recognition result indicates that the information to be published contains sensitive content, determining replacement information corresponding to the sensitive content based on the enterprise-level knowledge base; the replacement information is used to replace the sensitive content in the information to be published to obtain the target published information;

[0009] When the sensitive word recognition result indicates that there is no sensitive content in the information to be published, the information to be published is determined as the target published information.

[0010] In some possible embodiments, the enterprise-level knowledge base includes a vector database and a knowledge graph; the enterprise-level knowledge base is constructed by the following steps:

[0011] Acquire tracking data from various platforms and systems through preset tracking points; wherein the preset tracking points are deployed through a full tracking method or a customized tracking method; the tracking data includes at least one of text, images, audio, and video;

[0012] Performing vectorization processing on each of the buried point data, and constructing the vector database based on the vectorization processing results of each of the buried point data;

[0013] Based on the large language model, the semantic information of each of the buried data and the business scenarios corresponding to the buried data are extracted, and the association relationship between each of the buried data is determined based on the extraction results, and the knowledge graph is constructed based on the association relationship between each of the buried data.

[0014] In some possible embodiments, the sensitive word knowledge base includes a sensitive word vector database and a sensitive word knowledge graph; the sensitive word knowledge base is constructed by the following steps:

[0015] Based on preset sensitive word screening rules, semantic information of each embedded data, and the business scenario corresponding to the embedded data, each embedded data is screened for sensitive words to obtain multiple embedded data corresponding to different business scenarios; wherein the preset sensitive word screening rules include sensitive word screening rules corresponding to each business scenario;

[0016] For each business scenario, vectorize each of the sensitive word embedding data corresponding to the business scenario, and build a sensitive word vector data sub-library based on the vectorization results of each sensitive word embedding data;

[0017] Construct the sensitive word vector database based on the sensitive word vector data sub-library corresponding to each business scenario;

[0018] The semantic information of each of the sensitive word buried data is extracted based on the large language model, and the association relationship between each of the sensitive word buried data is determined based on the extraction results and the business scenarios corresponding to each of the sensitive word buried data, and the sensitive word knowledge graph is constructed based on the association relationship between each of the sensitive word buried data.

[0019] In some possible embodiments, identifying sensitive words in the information to be published based on the large language model and the sensitive word knowledge base includes:

[0020] Extracting features of each type of information in the information to be released, and performing multimodal feature alignment and feature fusion processing based on the feature extraction results to generate a joint feature vector and joint information corresponding to the joint feature vector;

[0021] Extracting semantic information and business scenarios of the joint information based on the large language model, and determining relevance information corresponding to the information to be released based on the extraction results and the sensitive word knowledge graph;

[0022] Determining similarity information corresponding to the information to be released based on the joint feature vector, the business scenario corresponding to the joint information, and the sensitive word vector database;

[0023] The sensitive word recognition result is determined based on the correlation information and the similarity information.

[0024] In some possible embodiments, determining replacement information corresponding to the sensitive content based on the enterprise-level knowledge base includes:

[0025] Extracting semantic information of the sensitive content and a business scenario corresponding to the sensitive content based on the large language model;

[0026] Determining multiple candidate replacement information corresponding to the sensitive content based on the semantic information of the sensitive content, the business scenario corresponding to the sensitive content, and the knowledge graph;

[0027] Vectorization processing is performed on the sensitive content and each candidate replacement information respectively, and the replacement information is determined based on the vectorization processing result and the vector database.

[0028] In some possible embodiments, after determining the replacement information corresponding to the sensitive content based on the enterprise-level knowledge base, the method further includes:

[0029] Pushing the sensitive content and replacement information corresponding to the sensitive content to a manual review terminal;

[0030] Receive the manual review result. If the review passes, update the sensitive content in the information to be published based on the replacement information to obtain the target published information; if the review fails, update the sensitive content in the information to be published based on the review replacement information indicated by the manual review result to obtain the target published information.

[0031] In some possible embodiments, after receiving the manual review result, the method further includes:

[0032] Based on the manual review results, the preset sensitive word screening rules, the sensitive word knowledge base, and the enterprise-level knowledge base are updated respectively; and

[0033] The embedded point data of each platform and system is periodically acquired through preset embedded points, and the preset sensitive word screening rules, the sensitive word knowledge base and the enterprise-level knowledge base are updated respectively based on the acquired embedded point data.

[0034] The present disclosure provides a multimodal sensitive word processing device, including:

[0035] An information acquisition module, configured to acquire information to be published and a pre-built knowledge base; wherein the information type in the information to be published includes at least one of images, text, audio, and video; and the knowledge base includes a sensitive word knowledge base and an enterprise-level knowledge base;

[0036] a sensitive word identification module, configured to identify sensitive words in the information to be published based on the large language model and the sensitive word knowledge base, and obtain a sensitive word identification result; wherein the sensitive word identification result includes whether the information to be published contains sensitive content that matches the sensitive word knowledge base;

[0037] an information replacement module, configured to, if the sensitive word recognition result indicates that the information to be published contains sensitive content, determine, based on the enterprise-level knowledge base, replacement information corresponding to the sensitive content; the replacement information is used to replace the sensitive content in the information to be published to obtain the target published information;

[0038] The information determination module is configured to determine the information to be published as the target published information when the sensitive word recognition result indicates that there is no sensitive content in the information to be published.

[0039] In some possible embodiments, the enterprise-level knowledge base includes a vector database and a knowledge graph; and the information acquisition module is further configured to:

[0040] Acquire tracking data from various platforms and systems through preset tracking points; wherein the preset tracking points are deployed through a full tracking method or a customized tracking method; the tracking data includes at least one of text, images, audio, and video;

[0041] Performing vectorization processing on each of the buried point data, and constructing the vector database based on the vectorization processing results of each of the buried point data;

[0042] Based on the large language model, the semantic information of each of the buried data and the business scenarios corresponding to the buried data are extracted, and the association relationship between each of the buried data is determined based on the extraction results, and the knowledge graph is constructed based on the association relationship between each of the buried data.

[0043] In some possible embodiments, the sensitive word knowledge base includes a sensitive word vector database and a sensitive word knowledge graph; the information acquisition module is further configured to:

[0044] Based on preset sensitive word screening rules, semantic information of each embedded data, and the business scenario corresponding to the embedded data, each embedded data is screened for sensitive words to obtain multiple embedded data corresponding to different business scenarios; wherein the preset sensitive word screening rules include sensitive word screening rules corresponding to each business scenario;

[0045] For each business scenario, vectorize each of the sensitive word embedding data corresponding to the business scenario, and build a sensitive word vector data sub-library based on the vectorization results of each sensitive word embedding data;

[0046] Construct the sensitive word vector database based on the sensitive word vector data sub-library corresponding to each business scenario;

[0047] The semantic information of each of the sensitive word buried data is extracted based on the large language model, and the association relationship between each of the sensitive word buried data is determined based on the extraction results and the business scenarios corresponding to each of the sensitive word buried data, and the sensitive word knowledge graph is constructed based on the association relationship between each of the sensitive word buried data.

[0048] In some possible embodiments, the sensitive word identification module is specifically used to:

[0049] Extracting features of each type of information in the information to be released, and performing multimodal feature alignment and feature fusion processing based on the feature extraction results to generate a joint feature vector and joint information corresponding to the joint feature vector;

[0050] Extracting semantic information and business scenarios of the joint information based on the large language model, and determining relevance information corresponding to the information to be released based on the extraction results and the sensitive word knowledge graph;

[0051] Determining similarity information corresponding to the information to be released based on the joint feature vector, the business scenario corresponding to the joint information, and the sensitive word vector database;

[0052] The sensitive word recognition result is determined based on the correlation information and the similarity information.

[0053] In some possible embodiments, the information replacement module is specifically configured to:

[0054] Extracting semantic information of the sensitive content and a business scenario corresponding to the sensitive content based on the large language model;

[0055] Determining multiple candidate replacement information corresponding to the sensitive content based on the semantic information of the sensitive content, the business scenario corresponding to the sensitive content, and the knowledge graph;

[0056] Vectorization processing is performed on the sensitive content and each candidate replacement information respectively, and the replacement information is determined based on the vectorization processing result and the vector database.

[0057] In some possible embodiments, the information replacement module is further configured to:

[0058] Pushing the sensitive content and replacement information corresponding to the sensitive content to a manual review terminal;

[0059] Receive the manual review result. If the review passes, update the sensitive content in the information to be published based on the replacement information to obtain the target published information; if the review fails, update the sensitive content in the information to be published based on the review replacement information indicated by the manual review result to obtain the target published information.

[0060] In some possible embodiments, the information replacement module is further configured to:

[0061] Based on the manual review results, the preset sensitive word screening rules, the sensitive word knowledge base, and the enterprise-level knowledge base are updated respectively; and

[0062] The embedded point data of each platform and system is periodically acquired through preset embedded points, and the preset sensitive word screening rules, the sensitive word knowledge base and the enterprise-level knowledge base are updated respectively based on the acquired embedded point data.

[0063] An embodiment of the present disclosure provides a computer device, comprising: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor and the memory communicate via the bus, and when the machine-readable instructions are executed by the processor, the multimodal sensitive word processing method described in any possible embodiment described above is performed.

[0064] An embodiment of the present disclosure provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the multimodal sensitive word processing method as described in any of the possible implementations described above is implemented.

[0065] The multimodal sensitive word processing method, apparatus, storage medium and computer equipment provided in the embodiments of the present disclosure specifically, first, obtain information to be published and a pre-built knowledge base; wherein the information type in the information to be published includes at least one of images, text, audio and video; the knowledge base includes a sensitive word knowledge base and an enterprise-level knowledge base; then, based on the large language model and the sensitive word knowledge base, sensitive word recognition is performed on the information to be published to obtain a sensitive word recognition result; wherein the sensitive word recognition result includes whether there is sensitive content in the information to be published that matches the sensitive word knowledge base; when the sensitive word recognition result indicates that there is sensitive content in the information to be published, replacement information corresponding to the sensitive content is determined based on the enterprise-level knowledge base; the replacement information is used to replace the sensitive content in the information to be published to obtain the target published information; when the sensitive word recognition result indicates that there is no sensitive content in the information to be published, the information to be published is determined as the target published information.

[0066] In this way, by building a multimodal, integrated sensitive word recognition system, we can comprehensively judge and process multimodal information, ensuring comprehensive identification of sensitive information. Furthermore, compliance review and replacement strategies can automatically trigger the replacement process based on the recognition results, automating the entire process from identification to replacement. This effectively reduces manual review costs and improves the accuracy and timeliness of sensitive information processing.

[0067] In order to make the above-mentioned objectives, features and advantages of the present disclosure more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following briefly introduces the drawings that need to be cited in the embodiments. The drawings herein are incorporated into and constitute a part of the specification. These drawings illustrate embodiments consistent with the present disclosure and, together with the specification, are used to illustrate the technical solutions of the present disclosure. It should be understood that the following drawings only illustrate certain embodiments of the present disclosure and should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without inventive effort.

[0069] Figure 1 A flowchart of a multimodal sensitive word processing method provided by an embodiment of the present disclosure is shown;

[0070] Figure 2 A flowchart showing a method for constructing an enterprise-level knowledge base provided by an embodiment of the present disclosure is shown;

[0071] Figure 3 A flowchart of a method for constructing a sensitive word knowledge base provided by an embodiment of the present disclosure is shown;

[0072] Figure 4 A flowchart of a sensitive word identification method provided by an embodiment of the present disclosure is shown;

[0073] Figure 5 A flowchart of a method for determining replacement information provided by an embodiment of the present disclosure is shown;

[0074] Figure 6 A schematic diagram of the structure of a multimodal sensitive word processing device provided by an embodiment of the present disclosure is shown;

[0075] Figure 7 A schematic structural diagram of a computer device provided by an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0076] In order to make the purpose, technical solutions and advantages of the embodiments of the present disclosure clearer, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. The components of the embodiments of the present disclosure generally described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present disclosure provided in the drawings is not intended to limit the scope of the disclosure for which protection is sought, but merely represents selected embodiments of the present disclosure. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present disclosure.

[0077] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings.

[0078] The term "and / or" herein simply describes an association relationship, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, and the existence of B alone. In addition, the term "at least one" herein represents any combination of at least one of any one or more of a plurality of items. For example, "including at least one of A, B, and C" can represent any one or more elements selected from the set consisting of A, B, and C.

[0079] To facilitate understanding of this embodiment, the execution subject of the multimodal sensitive word processing method provided by the embodiment of the present disclosure is first introduced in detail. The execution subject of the multimodal sensitive word processing method provided by the embodiment of the present disclosure is a computer device. The computer device can be a terminal device or a server. Among them, the terminal device can also be a mobile device, a user terminal, a terminal, a handheld device, a computing device, a vehicle-mounted device, a wearable device, etc. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud storage, big data and artificial intelligence platforms. Optionally, the method can also be applied to an implementation environment composed of a computer device and a server.

[0080] The following describes in detail the multimodal sensitive word processing method provided by the embodiment of the present application in conjunction with the accompanying drawings. Figure 1 FIG. 1 is a flowchart of a multimodal sensitive word processing method provided by an embodiment of the present disclosure, the method comprising the following steps S101 to S104:

[0081] S101, obtaining information to be published and a pre-built knowledge base.

[0082] It is understood that information to be released can be any content that an enterprise, organization, or individual intends to publish on a public platform. This can include at least one of the following types of content: images (e.g., promotional posters, product screenshots), text (e.g., press releases, social media copy), audio (e.g., podcasts, voice ads), and video (e.g., promotional videos, live broadcast recordings), as well as a combination of multiple types. For example, a company plans to release a piece of marketing content consisting of a promotional video (video type) and accompanying text description (text type).

[0083] In some possible embodiments, the information to be published may also include content such as web page links, which is not specifically limited here.

[0084] Here, knowledge bases include sensitive word knowledge bases and enterprise-level knowledge bases. A sensitive word knowledge base is a database containing a list of sensitive words. Specifically for the financial sector, this database may include sensitive terms or illegal expressions related to financial regulations and market operating standards; it may also include vocabulary from sensitive topics such as society and politics. An enterprise-level knowledge base is a knowledge base built based on the specific needs of an enterprise or field, and may contain information related to a company's business and policies.

[0085] Specifically, enterprise-level knowledge bases include vector databases and knowledge graphs. A vector database is a data repository that stores high-dimensional vector data. It maps unstructured data (such as text, images, audio, and video) or structured data (such as numerical values ​​and time series) into mathematical vectors through vectorization processing, thereby enabling rapid retrieval of similarities between information. For example, an enterprise can uniformly convert product manuals (text), promotional posters (images), and customer service voice recordings (audio) into vectors. Subsequently, by calculating the cosine similarity between vectors, the most relevant content can be quickly retrieved. A knowledge graph is a semantic network that uses a graph structure to represent entities (such as products, users, business scenarios, and sensitive words) and their relationships (such as "users purchase products" and "sensitive content in a certain business scenario"), revealing implicit relationships between data through semantic association analysis.

[0086] For example, referring to Figure 2 As shown, the enterprise-level knowledge base can be constructed through the following steps S201 to S203:

[0087] S201, obtaining tracking data of each platform and system through preset tracking points.

[0088] Specifically, when building an enterprise-level knowledge base, users can first capture user behavior, system logs, and other data through pre-deployed monitoring points within business systems or platforms. Enterprises can then choose to build a tracking software development kit (SDK) that covers multiple platforms (web, mobile, mini-programs, and desktop applications) to meet the collection needs of different technology stacks (Java, Go, Python, and C++ services). Tracking methods include full tracking (a single-step integration meets all collection requirements) and customized tracking (providing customized collection information). For example, a full tracking SDK can automatically capture user activity data across web, mobile, mini-programs, and desktop applications. A customized tracking SDK can also support customized tracking for specific business scenarios (such as user registration and payment).

[0089] Tracked data can include at least one of text (e.g., user reviews, product descriptions), images (e.g., product screenshots, promotional images), audio (e.g., customer service conversation recordings), and video (e.g., user operation recordings). For example, a social media platform collects user-posted text and image content and video interaction data through tracking.

[0090] S202, performing vectorization processing on each of the buried point data, and constructing the vector database based on the vectorization processing results of each of the buried point data.

[0091] Specifically, after acquiring tracking data from various platforms and systems, in order to build a vector database, for text data, word embedding models (such as Word2Vec and BERT) can be used to convert text into vectors. For image, audio, and video data, feature vectors can be extracted using models such as convolutional neural networks and autoencoders. For example, product images can be converted into 128-dimensional feature vectors, or customer service audio can be converted into text using speech recognition technology, which can then be further vectorized.

[0092] S203, extracting the semantic information of each of the buried data and the business scenarios corresponding to the buried data based on the large language model, and determining the association relationship between each of the buried data based on the extraction results, and constructing the knowledge graph based on the association relationship between each of the buried data.

[0093] Furthermore, in order to construct a knowledge graph, a large language model can be used to parse the buried data, perform semantic analysis on it, and extract entities and their attributes. Here, the large language model can use its own natural language understanding capabilities and multimodal data processing capabilities, and through context association to mine deep semantic features; then, based on the extraction results, the association relationship between each buried data is determined, and the knowledge graph is constructed. Among them, different large language models can be used for different modal data. For example, text data can use pre-trained language models such as BERT and GPT for semantic recognition; image data and video data can use large language models such as Qwen2-Vl for semantic analysis; audio data can be subjected to semantic analysis using large language models such as Qwen2-Audio.

[0094] For example, in the financial sector, the captured tracking data might include user activity records on a financial management app, including interactions such as clicking on the "fund details page," entering "purchase amount 10,000 yuan," and completing "payment confirmation," along with metadata such as page dwell time and operation timestamps. A large language model can be used to identify core entities such as users, fund products, and transaction orders. Attributes can include specific characteristics such as the user's age, risk preference, fund name, risk level, and order payment status and transaction time. Contextual information (such as user browsing path, purchase time, and historical transaction history) can be combined to label entities with business scenarios. For example, if a user continuously browses high-risk fund products and ultimately completes payment, the business scenario can be labeled "high-risk investment behavior." If a user purchases a certain fund type for the first time during a specific promotional event, the scenario can be labeled "new user participation in the event." This helps further refine the association logic between entities, for example, refining the association between "user A" and "fund B" to "user A purchased high-risk fund B during the promotional event." Finally, based on semantic information and business scenarios, we can identify the relationships between entities within each data point, thereby constructing a knowledge graph. For example, in a financial knowledge graph, we can establish triple relationships such as "user-holding-fund," "fund-belonging-fund type," and "user-participation-promotional activity," and enhance the semantic expression of these relationships through attributes such as timestamps and transaction amounts.

[0095] In some other embodiments, if the extracted buried data also includes link-type data, web page parsing can be performed on it first, and then the parsed data can be processed using a large language model to achieve the extraction of association relationships and the construction of a knowledge graph.

[0096] Similarly, the sensitive word knowledge base includes a sensitive word vector database and a sensitive word knowledge graph. The sensitive word vector database can be used to quickly retrieve and analyze similarity between sensitive words. The graph structure in the sensitive word knowledge graph can reveal the semantic associations and business scenario mappings between sensitive words. Figure 3 As shown, the sensitive word knowledge base can be constructed through the following steps S301 to S304:

[0097] S301, based on preset sensitive word screening rules and semantic information of each of the buried data and the business scenarios corresponding to the buried data, sensitive word screening is performed on each of the buried data to obtain multiple sensitive word buried data corresponding to different business scenarios.

[0098] Here, sensitive word screening rules refer to a set of rules used to identify words, phrases, or content that may be sensitive (e.g., involving legal or financial risks). Because different business scenarios have varying degrees of sensitivity to sensitive words, the preset sensitive word screening rules in this disclosure are tailored to each business scenario.

[0099] Specifically, by screening each embedded data for sensitive words based on the sensitive word screening rules corresponding to each business scenario and the semantic information of each embedded data and the business scenario corresponding to the embedded data, effective isolation and accurate identification of sensitive data in different business scenarios can be ensured.

[0100] S302: For each business scenario, vectorize each of the sensitive word buried data corresponding to the business scenario, and construct a sensitive word vector data sub-library based on the vectorization processing results of each sensitive word buried data.

[0101] Furthermore, after obtaining the sensitive word buried data corresponding to each business scenario, the sensitive word buried data corresponding to each business field can be vectorized and converted into vector form, and then a sensitive word vector data sub-library for each business scenario can be constructed based on the conversion results.

[0102] S303: Build the sensitive word vector database based on the sensitive word vector data sub-library corresponding to each business scenario.

[0103] It is understandable that after obtaining the sensitive word vector data sub-libraries corresponding to each business scenario, they can be integrated to obtain a sensitive word vector database. Then, by integrating the sub-libraries of each business scenario, a comprehensive sensitive word vector database can be formed.

[0104] S304, extracting semantic information of each of the sensitive word buried data based on the large language model, and determining the association relationship between each of the sensitive word buried data based on the extraction results and the business scenarios corresponding to each of the sensitive word buried data, and constructing the sensitive word knowledge graph based on the association relationship between each of the sensitive word buried data.

[0105] Specifically, after obtaining the sensitive word buried data corresponding to each business scenario, it is also necessary to use a large language model to extract the semantic information of each sensitive word buried data, analyze the contextual semantics of the sensitive words, and then determine the association relationship between each sensitive word buried data based on the extraction results and the business scenarios corresponding to each of the sensitive word buried data, and based on this, construct an association network between sensitive words to obtain a sensitive word knowledge graph.

[0106] In some possible embodiments, the sensitive word knowledge base may be constructed by obtaining sensitive word bases of various platforms as a basis, and then updating and further improving it using the above steps S301 to S304.

[0107] For example, with the development of emerging business scenarios such as digital currency and cross-border payments, enterprises or regulators can also adaptively expand existing data collection systems. This can be manifested in the expansion of tracking point deployment and the addition of deployment platforms. On the one hand, for new business scenarios such as digital currency trading platforms and cross-border payment systems, tracking points can be added at key business nodes (such as user registration, fund transfers, and transaction confirmations) to ensure that key data sources such as user behavior data, transaction flow information, and platform logs are fully captured. On the other hand, the deployment platform can also be extended from traditional web and mobile terminals to new technology carriers such as blockchain nodes and API interfaces to cover the data flow paths in distributed architectures such as digital currency wallets and cross-border payment gateways.

[0108] At the same time, by regularly executing data collection operations such as data embedding and adding new data embedding, a dynamically updated data stream can be formed, including real-time user behavior and system logs, as well as cross-platform, multi-modal complex data types. For each type of data, the above-mentioned knowledge base construction method can be re-executed to update the corresponding enterprise-level knowledge base and sensitive word knowledge base. Here, when updating the knowledge base, an incremental update method can be adopted, that is, only local adjustments are made to the sensitive words or business rules that have changed, so as to avoid the computational overhead brought by global reconstruction. For example, when a new sensitive word is added, only its vector representation and the local subgraph in the knowledge graph need to be updated.

[0109] Here, in the dynamic optimization process of the sensitive word knowledge base, the corresponding sensitive word screening rules are also updated, and the vector representation and the association relationship in the knowledge graph can also be updated synchronously.

[0110] In some other embodiments, newly collected embedded data can also be used to train or fine-tune the large language model to improve its semantic understanding ability for complex business scenarios. For example, in a fund sales scenario, by analyzing user purchase behavior and risk assessment records, the large language model can identify business scenarios such as "high-risk investment behavior" and "novice user participation in activities", and update the entity relationships in the knowledge graph accordingly (such as the triple relationship of "user-risk level-fund product"). In addition, for emerging business scenarios such as digital currency and cross-border payments, new entity types such as "blockchain address" and "smart contract" can be added to the knowledge base, and cross-domain association relationships such as "digital currency-trading platform-regulatory compliance" can be constructed at the same time.

[0111] S102: performing sensitive word recognition on the information to be published based on the large language model and the sensitive word knowledge base to obtain a sensitive word recognition result.

[0112] Here, when identifying sensitive words in information to be published, the information to be published can be compared with a sensitive word knowledge base based on a large language model to obtain a sensitive word identification result. The sensitive word identification result includes whether the information to be published contains sensitive content that matches the sensitive word knowledge base. The sensitive word identification result can also include the location of the sensitive content.

[0113] For example, in order to accurately identify sensitive words that may cause compliance risks or violate regulatory requirements, refer to Figure 4 As shown, when sensitive word recognition is performed on the information to be published based on the large language model and the sensitive word knowledge base, the following steps S401 to S404 may be included:

[0114] S401 , extracting features of each type of information in the information to be released, and performing multimodal feature alignment and feature fusion processing based on the feature extraction results to generate a joint feature vector and joint information corresponding to the joint feature vector.

[0115] Feature extraction refers to extracting representative and informative features from raw data. These features can be text features (such as keywords and syntactic structures) or image features (for example, objects or scenes in an image). Feature alignment involves matching data from different modalities (such as images, text, and audio) so that they can be uniformly represented in the same space. Feature fusion is then used to integrate the feature vectors corresponding to the aligned modal information into a joint feature vector. This joint feature vector combines data from all modalities to form a multidimensional feature representation.

[0116] Specifically, the information to be published may contain data from multiple modalities, such as text, images, and voice. Therefore, when extracting features, you can choose the appropriate feature extraction method for each type of information. For example, for text information, you can use natural language processing (NLP) techniques (such as TF-IDF and word embedding) to extract features and obtain text feature vectors. For image information, you can use computer vision techniques (such as convolutional neural networks (CNNs)) to extract features and obtain image feature vectors. For audio information, you can use speech recognition techniques (such as Mel-Frequency Cepstral Coefficients (MFCCs)) to extract features and obtain audio feature vectors.

[0117] It is understandable that the joint feature vector can also be used to obtain the joint information corresponding to the joint feature vector. The joint information represents a comprehensive expression of the semantic associations and complementary characteristics in cross-modal data. It not only contains the logical description of the text but also integrates the intuitive features of the image to form a more complete representation of the original multimodal data. For example, in financial information, a news article about the stock market may contain both charts and text content. Through feature extraction, financial terms (such as "stock gains" or "market fluctuations") in the news can be extracted, and the data in the chart (such as stock price trends) can be extracted. These features are then aligned and fused to generate a joint feature vector containing both text and chart information, as well as the joint information corresponding to this vector.

[0118] In the embodiment of the present disclosure, in the process of multimodal feature extraction, alignment and feature fusion processing of the information to be published, it is proposed to construct a hybrid deep learning model based on convolutional neural network (CNN) and recurrent neural network (RNN) that can process multiple data types, and enhance the model's sensitivity to subtle differences through multimodal training data training, so as to realize the ability to extract and integrate features from different modal data, and further realize the fusion and alignment processing of data.

[0119] S402: extracting semantic information and business scenarios of the joint information based on the large language model, and determining relevance information corresponding to the information to be released based on the extraction result and the sensitive word knowledge graph.

[0120] Here, deep semantic analysis of the joint information is performed using a large language model, extracting semantic information (e.g., the negative connotation of "plummet" in "a fund's net value plummeted") and business scenarios. Furthermore, combined with a knowledge graph of sensitive terms (e.g., including illegal terms like "promise of high returns" and their associated concepts), the correlation between the information to be released and each sensitive term in the knowledge graph is calculated.

[0121] For example, if the information to be released mentions "an analyst touts high returns in advance," the large language model will perform semantic analysis on it and, based on the business scenario and semantic analysis results, associate it with the sensitive word knowledge graph. It may determine that its correlation score with "high return promise" is 0.9 (high correlation). If the information to be released only mentions "a company's stock price fluctuations," the correlation score may be 0.3 (low correlation).

[0122] S403: Determine similarity information corresponding to the information to be released based on the joint feature vector, the business scenario corresponding to the joint information, and the sensitive word vector database.

[0123] Similarly, since the joint feature vector contains multiple attributes of the information to be released, it can also be used to calculate similarity with the vectors in the sensitive word vector database. The calculation results can be used to further determine whether the information to be released is similar to the sensitive words, thereby further evaluating its sensitivity.

[0124] In some possible embodiments, a dual search mechanism can be introduced to further improve the accuracy and response hit rate of sensitive word recognition. This dual search improves accuracy through a two-round search process. In the first round, a fast search algorithm (such as IVF-Flat) can be used to preliminarily screen the sensitive word vector database and return the top K candidate sensitive words. Simple filtering rules (such as length filtering and stop word removal) can be used to reduce subsequent computational effort. In the second round, the candidate sensitive words can be enhanced (e.g., by combining contextual semantics and sentiment) to construct enhanced vectors. A more complex similarity metric (such as deep learning models calculating semantic similarity) is then used to re-rank the candidate words. This re-ranking can incorporate business priority weights (e.g., "policy and regulatory sensitive words take precedence over advertising sensitive words") and can further incorporate historical user behavior data (such as whether the user has previously posted similar sensitive information) to dynamically adjust the ranking results. This two-round search and re-ranking process can improve the recall and precision of sensitive word recognition while maintaining efficiency, prioritizing the most relevant information and thus enhancing overall recognition effectiveness.

[0125] In some possible embodiments, in order to improve data processing efficiency and vector representation accuracy, data dimensionality reduction technology can be introduced as an optimization method to address the problem of excessively high dimensions of the generated joint feature vectors due to the large size of the data set. Specifically, when the dimensions of the joint feature vectors generated after feature extraction and fusion of the original data are high, similarity calculation or storage will generate large computing resource consumption and storage pressure, which may cause the semantic relationship between the vectors to be masked by noise, thereby reducing the accuracy of sensitive information identification. To this end, dimensionality reduction algorithms such as t-SNE (t-Distributed Stochastic Neighbor Embedding) can be used to map the high-dimensional joint feature vectors to a low-dimensional space (such as 2D or 3D), significantly reducing computational complexity while retaining the key semantic structure of the data.

[0126] S404: Determine the sensitive word recognition result based on the correlation information and the similarity information.

[0127] It is understood that based on the relevance information and similarity information, a comprehensive determination can be made as to whether the information to be released contains sensitive content that matches the sensitive word knowledge base. For example, if the relevance score of a piece of information in the information to be released is ≥ 0.8 and the similarity score is ≥ 0.75, then the information to be released is determined to contain sensitive content.

[0128] S103, when the sensitive word recognition result indicates that there is sensitive content in the information to be published, determining replacement information corresponding to the sensitive content based on the enterprise-level knowledge base; the replacement information is used to replace the sensitive content in the information to be published to obtain the target published information.

[0129] Specifically, if the sensitive word identification results indicate the presence of sensitive content in the information to be published, replacement information corresponding to the sensitive content can be located in the enterprise knowledge base. This replacement information not only complies with legal requirements but also maintains the core meaning and expressiveness of the original information. Through this replacement process, the sensitive content in the information to be published is converted into a compliant expression, resulting in a target release that meets the publication standards while retaining the original intent.

[0130] Specifically, in order to further improve the accuracy and applicability of replacement information, the present disclosure proposes a replacement information determination method, referring to Figure 5 As shown, the method includes the following steps S501 to S503:

[0131] S501 : Extracting semantic information of the sensitive content and a business scenario corresponding to the sensitive content based on the large language model.

[0132] It's understandable that analyzing sensitive content through a large language model can extract its semantic information and the business scenarios it relates to. For example, when processing text involving sensitive topics, a large language model can identify whether the content carries potential compliance risks, reputational risks, and other sensitive factors, and match it with relevant business scenarios (such as credit approval, investment consulting, and market transactions). Accurately identifying business scenarios provides context for subsequent selection of replacement information, ensuring that the replaced content better meets business needs and application scenarios.

[0133] For example, in a credit approval scenario, suppose a financial institution releases a credit policy interpretation and states, "For businesses in high-risk industries, such as real estate and entertainment, we will adopt stricter credit approval standards." The phrase "high-risk industries" might be considered sensitive due to its implications for industry-specific regulatory policies or market sensitivities. During analysis, the large language model identifies the potential compliance risk inherent in the phrase "high-risk industries," specifically avoiding direct mention of specific industries to mitigate potential legal or regulatory risks. It also identifies the business scenario involved in the content, namely, credit approval.

[0134] S502: Determine multiple candidate replacement information corresponding to the sensitive content based on the semantic information of the sensitive content, the business scenario corresponding to the sensitive content, and the knowledge graph.

[0135] Specifically, the semantic information of sensitive content extracted by the large language model, business scenarios, and pre-built knowledge graphs are matched. By traversing the knowledge graph, nodes and paths that match the sensitive content in terms of semantics, business scenarios, and compliance are found. The semantic similarity between the words or expressions represented by each node in the knowledge graph and the sensitive content is analyzed to ensure that the replacement information accurately conveys the original meaning. Specifically, in terms of business scenario matching, nodes that align with the business scenarios of the sensitive content can be screened to ensure the rationality of the replacement information in business applications. In terms of compliance matching, it is necessary to follow the laws, regulations, and regulatory requirements of the financial industry and exclude nodes that do not comply with the regulations. In this way, multiple alternative replacement information can be determined, which formally meet the strict requirements of laws and regulations, effectively avoid potential legal risks, fully consider the applicability of business scenarios, and can naturally integrate into the context of the original text, ensuring the accuracy and clarity of information conveyed.

[0136] For example, in an investment consulting context, if the original text states, "In the near future, we do not recommend investors participate in certain high-risk investment products," the term "high-risk investment products" may be considered sensitive due to concerns about investor protection and market stability. First, a large language model can be used to extract the semantics of "high-risk investment products," clarifying that they refer to investment products with a high risk level that could result in significant losses for investors. The system also identifies the business context of this content as investment consulting. Next, the system traverses the knowledge graph, searching for nodes and paths that are semantically similar to "high-risk investment products," are applicable in the investment consulting context, and meet regulatory compliance requirements. Ultimately, multiple alternative replacements may be identified, such as "investment products with higher risk levels," "investment options with greater uncertainty," and "investment projects that may face a higher risk of loss." These alternatives accurately convey the core message of the original text regarding high-risk investment products while also complying with investment consulting standards and requirements.

[0137] S503: Perform vectorization processing on the sensitive content and each candidate replacement information respectively, and determine the replacement information based on the vectorization processing result and the vector database.

[0138] To further optimize the selection of replacement information and improve its accuracy and applicability, the present disclosure utilizes vectorization technology to convert sensitive content and each candidate replacement information into vectors in a high-dimensional vector space. The similarity between these vectors is determined using pre-stored vector data in a vector database and a similarity calculation algorithm. The similarity calculation results quantify the degree of proximity between the candidate replacement information and the sensitive content. The higher the similarity, the closer the candidate replacement information is to the sensitive content, and the more likely it is to maintain the original intent and expression after replacement.

[0139] Furthermore, based on the similarity result determined by the vector database, the candidate replacement information may be sorted in descending order of similarity, and the candidate replacement information with the highest similarity may be selected as the final replacement information.

[0140] It is understandable that in order to further improve the quality and compliance of information release and ensure the semantic accuracy of the replacement information, this disclosure also proposes a manual review process for the replacement information, which may specifically include the following steps (a) to (b):

[0141] (a) Pushing the sensitive content and the replacement information corresponding to the sensitive content to the manual review terminal;

[0142] (b) receiving a manual review result; if the review passes, updating the sensitive content in the information to be published based on the replacement information to obtain the target published information; if the review fails, updating the sensitive content in the information to be published based on the review replacement information indicated by the manual review result to obtain the target published information.

[0143] Specifically, after obtaining the replacement information, the sensitive content and its corresponding replacement information are sent to the manual review end for review by the reviewer. After the manual reviewer completes the review of the replacement information, they receive their review results, which usually include two situations: review passed or review failed, as well as the review replacement information that may be attached when the review fails. If the review passes, the replacement information that passes the review is automatically adopted and updated to the corresponding sensitive content position in the information to be released, thereby generating the final target release information. This process ensures that the target release information is semantically consistent with the original intention and complies with the requirements of laws, regulations and business specifications. If the review fails, the sensitive content in the information to be released can be updated accordingly according to the review replacement information indicated in the manual review result, thereby generating the final target release information. Here, the review replacement information may be a new replacement plan proposed by the reviewer based on his or her own professional knowledge and experience according to the specific situation.

[0144] In some other embodiments, the pushed content may further include context information corresponding to the sensitive content, and auxiliary information such as similarity evaluation results between the replacement information and the sensitive content in the vector database.

[0145] After receiving the manual review results, the preset sensitive word screening rules, the sensitive word knowledge base, and the enterprise-level knowledge base can be updated based on the manual review results. Specifically, if the review results repeatedly show failures due to misidentification of sensitive words, the sensitive word screening rules can be optimized, and the screening threshold or screening strategy can be adjusted to reduce the misidentification rate. If reviewers frequently point out that certain replacement information is missing from the knowledge base, the enterprise-level knowledge base can be promptly updated to incorporate approved replacement information into the knowledge base, supplementing relevant domain knowledge, enriching the enterprise's knowledge base, and improving the completeness of the knowledge base. This inclusion process not only requires adding the replacement information itself, but also integrating its corresponding contextual information, business scenarios, and other aspects to construct a complete knowledge entry (see steps S201-S203 above for details). For the sensitive word database, misidentified sensitive words can be removed from the sensitive word knowledge base based on the manual review results, enabling dynamic updating and improvement of the sensitive word knowledge base.

[0146] In some possible embodiments, the content sent to the reviewer may include more than just sensitive content and replacement information. To assist the reviewer in making a more accurate and comprehensive judgment, a rich set of auxiliary information is also provided. For example, the complete contextual paragraph containing the sensitive content may be included, allowing the reviewer to understand the meaning and potential impact of the sensitive content from a broader context. Simultaneously, an assessment of the similarity between the sensitive content and the alternative replacement information in the vector database is provided, helping the reviewer quickly understand the semantic proximity between the replacement information and the sensitive content. During the review process, the reviewer may discover words that were not originally included in the sensitive word knowledge base but are sensitive in specific business scenarios. When these newly identified sensitive words are added to the sensitive word knowledge base, they also need to be vectorized and their associations with other sensitive words are determined based on their semantic information and business context, updating the sensitive word vector database and the sensitive word knowledge graph (see steps S301-S304 above for details).

[0147] In order to achieve effective management and optimization of the entire process of handling sensitive words, the present disclosure also proposes data collection, collation, statistics, and analysis of sensitive word handling results, builds data collection and processing capabilities based on technical components such as Kafka, Redis, and ElasticSearch, and builds data analysis and computing capabilities based on memory computing, distributed computing, and batch processing. On the basis of data support, it provides corresponding sensitive word management device operation results monitoring, operational analysis, statistical monitoring, indicator models, comprehensive management, and vocabulary update capabilities for the entire process. For example, through real-time monitoring and analysis of indicators such as the accuracy of sensitive word recognition, the approval rate of replacement information, and the frequency of sensitive word knowledge base updates, problems and deficiencies in the sensitive word management process can be discovered in a timely manner. At the same time, based on the results of data analysis, an indicator model can also be constructed to predict possible future trends in sensitive words, make preparations in advance, and enhance the initiative of sensitive word management.

[0148] S104: When the sensitive word recognition result indicates that the information to be published does not contain sensitive content, the information to be published is determined as the target published information.

[0149] Here, in the case that there is no sensitive content in the information to be published, the information to be published can be directly determined as the target publication information to realize the publication of the information.

[0150] The multimodal sensitive word processing method, apparatus, storage medium, and computer equipment provided in the embodiments of this disclosure, by constructing a multimodal fusion sensitive word recognition system, achieves comprehensive judgment and processing of multimodal information, ensuring comprehensive identification of sensitive information processing. Furthermore, compliance review and replacement strategies can automatically trigger the replacement process based on the identification results, achieving automated processing from identification to replacement, effectively reducing manual review costs and improving the accuracy and timeliness of sensitive information processing.

[0151] Those skilled in the art will understand that in the above method of the specific implementation method, the writing order of each step does not mean a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0152] Based on the same inventive concept, the embodiment of the present disclosure also provides a multimodal sensitive word processing device corresponding to the multimodal sensitive word processing method. Since the principle of solving the problem by the device in the embodiment of the present disclosure is similar to the above-mentioned multimodal sensitive word processing method in the embodiment of the present disclosure, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be repeated.

[0153] Reference Figure 6 FIG. 1 is a schematic diagram of a multimodal sensitive word processing device 600 provided in an embodiment of the present disclosure, wherein the device includes:

[0154] Information acquisition module 601 is used to acquire information to be published and a pre-built knowledge base; wherein the information type in the information to be published includes at least one of images, text, audio and video; the knowledge base includes a sensitive word knowledge base and an enterprise-level knowledge base;

[0155] Sensitive word identification module 602, configured to identify sensitive words in the information to be published based on the large language model and the sensitive word knowledge base, and obtain a sensitive word identification result; wherein the sensitive word identification result includes whether the information to be published contains sensitive content that matches the sensitive word knowledge base;

[0156] An information replacement module 603 is configured to, if the sensitive word recognition result indicates that the information to be published contains sensitive content, determine replacement information corresponding to the sensitive content based on the enterprise-level knowledge base; the replacement information is used to replace the sensitive content in the information to be published to obtain the target published information;

[0157] The information determination module 604 is configured to determine the information to be published as the target published information if the sensitive word recognition result indicates that the information to be published does not contain sensitive content.

[0158] In some possible embodiments, the enterprise-level knowledge base includes a vector database and a knowledge graph; the information acquisition module 601 is further configured to:

[0159] Acquire tracking data from various platforms and systems through preset tracking points; wherein the preset tracking points are deployed through a full tracking method or a customized tracking method; the tracking data includes at least one of text, images, audio, and video;

[0160] Performing vectorization processing on each of the buried point data, and constructing the vector database based on the vectorization processing results of each of the buried point data;

[0161] Based on the large language model, the semantic information of each of the buried data and the business scenarios corresponding to the buried data are extracted, and the association relationship between each of the buried data is determined based on the extraction results, and the knowledge graph is constructed based on the association relationship between each of the buried data.

[0162] In some possible embodiments, the sensitive word knowledge base includes a sensitive word vector database and a sensitive word knowledge graph; the information acquisition module 601 is further used to:

[0163] Based on preset sensitive word screening rules, semantic information of each embedded data, and the business scenario corresponding to the embedded data, each embedded data is screened for sensitive words to obtain multiple embedded data corresponding to different business scenarios; wherein the preset sensitive word screening rules include sensitive word screening rules corresponding to each business scenario;

[0164] For each business scenario, vectorize each of the sensitive word embedding data corresponding to the business scenario, and build a sensitive word vector data sub-library based on the vectorization results of each sensitive word embedding data;

[0165] Construct the sensitive word vector database based on the sensitive word vector data sub-library corresponding to each business scenario;

[0166] The semantic information of each of the sensitive word buried data is extracted based on the large language model, and the association relationship between each of the sensitive word buried data is determined based on the extraction results and the business scenarios corresponding to each of the sensitive word buried data, and the sensitive word knowledge graph is constructed based on the association relationship between each of the sensitive word buried data.

[0167] In some possible embodiments, the sensitive word identification module 602 is specifically configured to:

[0168] Extracting features of each type of information in the information to be released, and performing multimodal feature alignment and feature fusion processing based on the feature extraction results to generate a joint feature vector and joint information corresponding to the joint feature vector;

[0169] Extracting semantic information and business scenarios of the joint information based on the large language model, and determining relevance information corresponding to the information to be released based on the extraction results and the sensitive word knowledge graph;

[0170] Determining similarity information corresponding to the information to be released based on the joint feature vector, the business scenario corresponding to the joint information, and the sensitive word vector database;

[0171] The sensitive word recognition result is determined based on the correlation information and the similarity information.

[0172] In some possible embodiments, the information replacement module 603 is specifically configured to:

[0173] Extracting semantic information of the sensitive content and a business scenario corresponding to the sensitive content based on the large language model;

[0174] Determining multiple candidate replacement information corresponding to the sensitive content based on the semantic information of the sensitive content, the business scenario corresponding to the sensitive content, and the knowledge graph;

[0175] Vectorization processing is performed on the sensitive content and each candidate replacement information respectively, and the replacement information is determined based on the vectorization processing result and the vector database.

[0176] In some possible embodiments, the information replacement module 603 is further configured to:

[0177] Pushing the sensitive content and replacement information corresponding to the sensitive content to a manual review terminal;

[0178] Receive the manual review result. If the review passes, update the sensitive content in the information to be published based on the replacement information to obtain the target published information; if the review fails, update the sensitive content in the information to be published based on the review replacement information indicated by the manual review result to obtain the target published information.

[0179] In some possible embodiments, the information replacement module 603 is further configured to:

[0180] Based on the manual review results, the preset sensitive word screening rules, the sensitive word knowledge base, and the enterprise-level knowledge base are updated respectively; and

[0181] The embedded point data of each platform and system is periodically acquired through preset embedded points, and the preset sensitive word screening rules, the sensitive word knowledge base and the enterprise-level knowledge base are updated respectively based on the acquired embedded point data.

[0182] Based on the same technical concept, the embodiment of the present disclosure also provides a computer device. Figure 7 , which is a schematic diagram of the structure of a computer device 700 provided in an embodiment of the present disclosure, includes a processor 701, a memory 702, and a bus 703. The memory 702 is used to store execution instructions and includes a memory 7021 and an external memory 7022. The memory 7021, also referred to as internal memory, is used to temporarily store calculation data in the processor 701 and data exchanged with an external memory 7022, such as a hard disk. The processor 701 exchanges data with the external memory 7022 through the memory 7021.

[0183] In the embodiment of the present application, the memory 702 is specifically used to store application code for executing the solution of the present application, and the execution is controlled by the processor 701. That is, when the computer device 700 is running, the processor 701 communicates with the memory 702 via the bus 703, so that the processor 701 executes the application code stored in the memory 702, thereby performing the method described in any of the aforementioned embodiments.

[0184] The memory 702 may be, but is not limited to, a random access memory (RAM), a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), etc.

[0185] Processor 701 may be an integrated circuit chip with signal processing capabilities. Such processors may be general-purpose processors, including central processing units (CPUs) and network processors (NPs). They may also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. These processors may implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of the present invention. A general-purpose processor may be a microprocessor or any conventional processor.

[0186] It should be understood that the structure illustrated in the embodiments of the present application does not constitute a specific limitation on the computer device 700. In other embodiments of the present application, the computer device 700 may include more or fewer components than shown, or may combine or separate certain components, or arrange the components differently. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0187] The present disclosure also provides a computer-readable storage medium having a computer program stored thereon. When executed by a processor, the computer program executes the steps of the multimodal sensitive word processing method described in the above method embodiment. The storage medium may be a volatile or non-volatile computer-readable storage medium.

[0188] The present disclosure also provides a computer program product that carries program code. The program code includes instructions that can be used to execute the steps of the multimodal sensitive word processing method described in the above method embodiment. For details, please refer to the above method embodiment and will not be repeated here.

[0189] The computer program product may be implemented in hardware, software, or a combination thereof. In one embodiment, the computer program product is implemented as a computer storage medium. In another embodiment, the computer program product is implemented as a software product, such as a software development kit (SDK).

[0190] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system and device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In the several embodiments provided in the present disclosure, it should be understood that the disclosed system and method can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0191] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0192] In addition, each functional unit in each embodiment of the present disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0193] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present disclosure, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present disclosure. The aforementioned storage medium includes: various media that can store program code, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0194] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present disclosure, which are used to illustrate the technical solutions of the present disclosure, rather than to limit them. The scope of protection of the present disclosure is not limited thereto. Although the present disclosure has been described in detail with reference to the above-described embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above-described embodiments within the technical scope disclosed in the present disclosure, or replace some of the technical features therein with equivalents. Such modifications, changes, or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure, and should be included in the scope of protection of the present disclosure. Therefore, the scope of protection of the present disclosure shall be subject to the scope of protection of the claims.

Claims

1. A multimodal sensitive word processing method, characterized in that: include: Acquire information to be published and a pre-built knowledge base; wherein the information type in the information to be published includes at least one of images, text, audio, and video; and the knowledge base includes a sensitive word knowledge base and an enterprise-level knowledge base; performing sensitive word identification on the information to be published based on the large language model and the sensitive word knowledge base to obtain a sensitive word identification result; wherein the sensitive word identification result includes whether the information to be published contains sensitive content that matches the sensitive word knowledge base; If the sensitive word recognition result indicates that the information to be published contains sensitive content, determining replacement information corresponding to the sensitive content based on the enterprise-level knowledge base; the replacement information is used to replace the sensitive content in the information to be published to obtain the target published information; If the sensitive word recognition result indicates that the information to be published does not contain sensitive content, determining the information to be published as the target published information; The enterprise-level knowledge base includes a vector database and a knowledge graph; the enterprise-level knowledge base is constructed through the following steps: Acquire tracking data from various platforms and systems through preset tracking points; wherein the preset tracking points are deployed through a full tracking method or a customized tracking method; the tracking data includes at least one of text, images, audio, and video; Performing vectorization processing on each of the buried point data, and constructing the vector database based on the vectorization processing results of each of the buried point data; Extracting semantic information of each of the buried data and the business scenarios corresponding to the buried data based on the large language model, determining the association relationship between each of the buried data based on the extraction results, and constructing the knowledge graph based on the association relationship between each of the buried data; The sensitive word knowledge base includes a sensitive word vector database and a sensitive word knowledge graph; the sensitive word knowledge base is constructed by the following steps: Based on preset sensitive word screening rules, semantic information of each embedded data, and the business scenario corresponding to the embedded data, each embedded data is screened for sensitive words to obtain multiple embedded data corresponding to different business scenarios; wherein the preset sensitive word screening rules include sensitive word screening rules corresponding to each business scenario; For each business scenario, vectorize each of the sensitive word embedding data corresponding to the business scenario, and build a sensitive word vector data sub-library based on the vectorization results of each sensitive word embedding data; Construct the sensitive word vector database based on the sensitive word vector data sub-library corresponding to each business scenario; The semantic information of each of the sensitive word buried data is extracted based on the large language model, and the association relationship between each of the sensitive word buried data is determined based on the extraction results and the business scenarios corresponding to each of the sensitive word buried data, and the sensitive word knowledge graph is constructed based on the association relationship between each of the sensitive word buried data.

2. The method according to claim 1, characterized in that The identifying of sensitive words in the information to be published based on the large language model and the sensitive word knowledge base includes: Extracting features of each type of information in the information to be released, and performing multimodal feature alignment and feature fusion processing based on the feature extraction results to generate a joint feature vector and joint information corresponding to the joint feature vector; Extracting semantic information and business scenarios of the joint information based on the large language model, and determining relevance information corresponding to the information to be released based on the extraction results and the sensitive word knowledge graph; Determining similarity information corresponding to the information to be released based on the joint feature vector, the business scenario corresponding to the joint information, and the sensitive word vector database; The sensitive word recognition result is determined based on the correlation information and the similarity information.

3. The method according to claim 2, characterized in that The determining, based on the enterprise-level knowledge base, replacement information corresponding to the sensitive content includes: Extracting semantic information of the sensitive content and a business scenario corresponding to the sensitive content based on the large language model; Determining multiple candidate replacement information corresponding to the sensitive content based on the semantic information of the sensitive content, the business scenario corresponding to the sensitive content, and the knowledge graph; Vectorization processing is performed on the sensitive content and each candidate replacement information respectively, and the replacement information is determined based on the vectorization processing result and the vector database.

4. The method according to claim 2, characterized in that After determining the replacement information corresponding to the sensitive content based on the enterprise-level knowledge base, the method further includes: Pushing the sensitive content and replacement information corresponding to the sensitive content to a manual review terminal; Receive the manual review result. If the review passes, update the sensitive content in the information to be published based on the replacement information to obtain the target published information; if the review fails, update the sensitive content in the information to be published based on the review replacement information indicated by the manual review result to obtain the target published information.

5. The method according to claim 4, characterized in that After receiving the manual review result, the method further includes: Based on the manual review results, the preset sensitive word screening rules, the sensitive word knowledge base, and the enterprise-level knowledge base are updated respectively; and The embedded point data of each platform and system is periodically acquired through preset embedded points, and the preset sensitive word screening rules, the sensitive word knowledge base and the enterprise-level knowledge base are updated respectively based on the acquired embedded point data.

6. A multimodal sensitive word management device, characterized in that: include: An information acquisition module, configured to acquire information to be published and a pre-built knowledge base; wherein the information type in the information to be published includes at least one of images, text, audio, and video; and the knowledge base includes a sensitive word knowledge base and an enterprise-level knowledge base; a sensitive word identification module, configured to identify sensitive words in the information to be published based on the large language model and the sensitive word knowledge base, and obtain a sensitive word identification result; wherein the sensitive word identification result includes whether the information to be published contains sensitive content that matches the sensitive word knowledge base; an information replacement module, configured to, if the sensitive word recognition result indicates that the information to be published contains sensitive content, determine, based on the enterprise-level knowledge base, replacement information corresponding to the sensitive content; the replacement information is used to replace the sensitive content in the information to be published to obtain the target published information; An information determination module, configured to determine the information to be published as the target published information if the sensitive word recognition result indicates that the information to be published does not contain sensitive content; The enterprise-level knowledge base includes a vector database and a knowledge graph; the enterprise-level knowledge base is constructed through the following steps: Acquire tracking data from various platforms and systems through preset tracking points; wherein the preset tracking points are deployed through a full tracking method or a customized tracking method; the tracking data includes at least one of text, images, audio, and video; Performing vectorization processing on each of the buried point data, and constructing the vector database based on the vectorization processing results of each of the buried point data; Extracting semantic information of each of the buried data and the business scenarios corresponding to the buried data based on the large language model, determining the association relationship between each of the buried data based on the extraction results, and constructing the knowledge graph based on the association relationship between each of the buried data; The sensitive word knowledge base includes a sensitive word vector database and a sensitive word knowledge graph; the sensitive word knowledge base is constructed by the following steps: Based on preset sensitive word screening rules, semantic information of each embedded data, and the business scenario corresponding to the embedded data, each embedded data is screened for sensitive words to obtain multiple embedded data corresponding to different business scenarios; wherein the preset sensitive word screening rules include sensitive word screening rules corresponding to each business scenario; For each business scenario, vectorize each of the sensitive word embedding data corresponding to the business scenario, and build a sensitive word vector data sub-library based on the vectorization results of each sensitive word embedding data; Construct the sensitive word vector database based on the sensitive word vector data sub-library corresponding to each business scenario; The semantic information of each of the sensitive word buried data is extracted based on the large language model, and the association relationship between each of the sensitive word buried data is determined based on the extraction results and the business scenarios corresponding to each of the sensitive word buried data, and the sensitive word knowledge graph is constructed based on the association relationship between each of the sensitive word buried data.

7. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

8. A computer device comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 5 is implemented.

Images