A copywriting auditing method and system based on cross-modal natural language processing

Abstract

The application relates to the technical field of natural language processing, in particular to a script auditing method and system based on cross-modal natural language processing, which comprises the following steps: collecting multi-modal script data, constructing an industry-specific variant knowledge base after standardization preprocessing; processing the text data through a three-level analysis mechanism to generate standard text with a variant label; constructing a field compliance knowledge graph and a natural language processing model, inputting the standard text with the variant label, strengthening the compliance sensitive word features, realizing context semantic correlation analysis and compliance rule matching verification, and generating a text compliance verification result and standardized text features; extracting image key area features and video caption features, determining multi-modal semantic consistency through a cross-modal attention mechanism, and forming a multi-modal verification result; and executing differentiated auditing decisions according to risk levels and outputting a structured auditing result. The scheme improves the recognition accuracy and efficiency through a full-link collaborative architecture.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a text review method and system based on cross-modal natural language processing. Background Technology

[0002] Marketing copywriting, as a form of promotion relying on multiple media such as multimedia advertising, brochures, and emails, aims to convey product selling points and competitive advantages in a way that meets customer needs, thereby boosting sales and service promotion. Its compliance directly impacts the market competition landscape and the company's compliant operations. Therefore, compliance review of marketing copywriting is a crucial aspect of a company's market operations.

[0003] Currently, compliance review of marketing copy mainly relies on manual review or traditional computer technology to retrieve compliance clauses, determining compliance through manual comparison of each clause and traditional keyword matching. This approach requires significant investment of manpower and time, and is limited by subjective human judgment and insufficient semantic understanding of traditional technologies, resulting in low review efficiency and difficulty in guaranteeing accuracy. It fails to meet the actual needs of enterprises for rapid review of bulk marketing copy.

[0004] With the explosive growth in the number of marketing copy, the diversification of media, and the increasing sophistication of compliance requirements, the bottlenecks of existing review methods in terms of efficiency, cost, and accuracy are becoming increasingly apparent. There is an urgent need for a highly efficient, accurate, and low-cost marketing copy compliance review technology to address the core pain points of existing solutions and provide reliable support for enterprise marketing copy compliance review. Summary of the Invention

[0005] This invention achieves synergistic linkage of core capabilities through a full-link collaborative architecture, solving the technical problem of difficulty in identifying variant violations and improving identification accuracy and efficiency.

[0006] The technical solution proposed in this invention is: a text review method based on cross-modal natural language processing, the method comprising: Collect multimodal copywriting data, and after standardized preprocessing, build an industry-specific variant knowledge base; Based on an industry-specific variant knowledge base, text data is processed through a three-level parsing mechanism to generate standard text with variant tags. Construct a domain compliance knowledge graph and natural language processing model, input standard text with variant labels, strengthen compliance-sensitive word features, realize context semantic association analysis and compliance rule matching verification, and generate text compliance verification results and standardized text features; Based on standardized text features, key region features of images and subtitle features of videos are extracted. Multimodal semantic consistency is determined through a cross-modal attention mechanism to form a multimodal verification result. By combining text compliance verification results, multimodal verification results, and dynamic rule base, differentiated audit decisions are made according to risk level, and structured audit results are output.

[0007] Preferably, the specific construction process of the industry-specific variant knowledge base is as follows: Collect industry violation case databases, regulatory notices, and manually annotated variant samples, and integrate them to form the raw data of the knowledge base; Based on variant type, the original data is stored in a structured manner using a triplet format; Establish an automatic update mechanism to regularly expand the variant word entries based on newly generated violation cases; The expanded variant entries are manually reviewed and confirmed to ensure that the accuracy of the data entering the database meets the usage requirements, forming a dynamically iterative industry-specific variant knowledge base.

[0008] Preferably, the process for obtaining the standard text with variant tags is as follows: Based on the TF-IDF algorithm and variant word matching, the preprocessed copy data is traversed and detected to identify potential variant words; The restoration strategy is adapted to the variant type to obtain the restored text; The restored text is input into the BiLSTM model to calculate the semantic similarity between the restored text and the context. If the similarity reaches the preset similarity threshold, the standard text with variant labels is output; if the similarity does not reach the preset similarity threshold, a manual review process is triggered.

[0009] Preferably, the specific process for obtaining the text compliance verification result and the standardized text features is as follows: Organize industry compliance rules and construct a domain compliance knowledge graph that includes entities, relationships, and violation judgment standards; The TransE algorithm is used to transform entities and relationships in the domain compliance knowledge graph into vectors with preset dimensions, thereby enabling graph vector embedding. Based on the BERT-Base model, a global semantic attention layer is added, and feature weights are calculated by quantizing and fusing compliance-sensitive features. The embedded graph vectors and text word vectors are input into a multi-layer Transformer encoder for feature fusion, which enhances the feature representation of compliance-sensitive segments and enables contextual semantic association analysis and compliance rule matching verification.

[0010] Preferably, the specific process for obtaining the multimodal verification result is as follows: Based on standardized text features, ResNet-50 network is used to extract features from images. After fine-tuning with industry image datasets, key regions are located using RoIAlign. The CNN text detection and CTC subtitle recognition algorithms are used to detect and recognize text regions in video frames and convert them into standardized subtitle features. The semantic relevance between text and image, and between text and subtitle is calculated using cosine similarity. Semantic conflict features with semantic relevance below the semantic relevance threshold are given high attention weights. The features of each modality are weighted according to the attention weight and concatenated to generate a fusion vector. The similarity is calculated with the semantically consistent template vector to determine the semantic consistency of the multimodalities and form a multimodal verification result.

[0011] Preferably, the specific process of obtaining the dynamic rule base is as follows: Based on the violation judgment criteria in the domain compliance knowledge graph and combined with historical violation cases, the correspondence between violation characteristics and risk levels is extracted to generate an initial rule set; A rule validity verification algorithm is adopted to verify the initial rule set through historical audit samples, and rules that meet the accuracy requirements are selected and included in the dynamic rule base. Establish a rule update trigger mechanism to periodically identify new combinations of violation characteristics and generate candidate rules based on newly generated violation cases and changes in compliance rules; After validity verification, qualified candidate rules are added to the dynamic rule base, and the rule priorities are adjusted accordingly.

[0012] Preferably, the specific content of the differentiated review decision is as follows: By integrating the intensity of violation features in text compliance verification results, the degree of semantic conflict in multimodal verification results, and the rule matching results of dynamic rule base, a risk level assessment system is established. If the intensity of the violation feature is lower than the preset intensity threshold and there is no clear semantic conflict, and it meets the fault tolerance rules, it is judged as low risk and automatically marked as requiring a second sampling inspection; If the intensity of the violation reaches the preset intensity threshold or there is a clear semantic conflict, but the basis for the violation needs to be further confirmed, it is judged as medium risk, and an audit work order containing the violation fragment identifier and semantic analysis description is generated and pushed to the manual audit queue. If the intensity of the violation characteristics exceeds the preset intensity threshold or meets the mandatory blocking rules, it is judged as high risk, the publication of the copy is directly blocked, the relevant information of the violation is recorded and synchronized to the platform's risk control system.

[0013] Preferably, the update process of the domain compliance knowledge graph is as follows: Establish data sources for updating the compliance map, including newly added industry compliance policies, regulatory notifications, uncovered violation cases, and supplementary compliance rules information provided by human feedback; Configure map update trigger conditions, including periodic update triggers and event-driven updates; When the update trigger condition is met, extract the compliant entities, relationships and judgment criteria from the newly added data, and perform structured processing according to the preset graph data format; The graph incremental update algorithm is used to integrate newly added structured data into the existing domain compliance knowledge graph and update entity vectors and relationship mappings. The effectiveness of the updated graph is verified through compliance rule matching tests to ensure that the new content does not conflict with existing rules.

[0014] The present invention also provides a copywriting review system based on cross-modal natural language processing, the system being used to execute the copywriting review method based on cross-modal natural language processing.

[0015] The present invention also provides a computer-readable storage medium storing a computer program that is executed by a processor to implement the aforementioned copywriting review method based on cross-modal natural language processing.

[0016] The beneficial effects of this invention are: 1. Through a three-level parsing mechanism and a dynamically iterative industry-specific variant knowledge base, the system efficiently solves the challenges of identifying variant violations such as homophonic substitution and symbolic interpolation. Variant identification accuracy is improved to over 92%, and the false negative rate is reduced to below 8%. The knowledge base is automatically updated 24 hours a day and manually reviewed, adapting to new violation patterns across multiple industries, significantly reducing manual maintenance costs. The coverage rate of variant terms across multiple industries reaches over 90%, providing standardized text input for subsequent compliance verification and solidifying the foundation for audit accuracy.

[0017] 2. Integrating multimodal features from text, images, and video captions, key features are extracted using ResNet-50 and CTC networks. A cross-modal attention mechanism is employed to enhance the expression of semantic conflict features, achieving a multimodal semantic consistency judgment accuracy of ≥93%. The false negative rate for violations such as "text-image discrepancies" and "caption conflicts" has decreased from 75% to 12%. The processing latency for a single multimodal data entry is ≤300ms, enabling high-concurrency review. This ensures comprehensive review while meeting batch processing requirements, effectively reducing compliance risks and consumer complaints caused by multimodal information asymmetry.

[0018] 3. By leveraging the FP-Growth algorithm to discover novel combinations of violation features, a three-dimensional dynamic rule base is constructed. Combined with a domain compliance knowledge graph that is updated regularly every 72 hours and through event-driven updates, rapid adaptation and accurate matching of compliance rules are achieved. Differentiated risk decision-making reduces manual review workload by 67%, achieves a high-risk interception rate of ≥98%, and efficiently implements low-risk spot checks. This not only improves review efficiency (batch review efficiency increased by 60%), but also continuously adapts to changes in regulatory policies through collaborative optimization of the knowledge graph and rule base, ensuring compliance and flexibility across multiple industries. Attached Figure Description

[0019] Figure 1 A flowchart of a copywriting review method based on cross-modal natural language processing; Figure 2 This is a flowchart illustrating the review process of a copywriting review method based on cross-modal natural language processing. Detailed Implementation

[0020] The following description is intended to disclose the present invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art. The basic principles of the invention defined in the following description can be applied to other embodiments, modifications, improvements, equivalents, and other technical solutions that do not depart from the spirit and scope of the invention.

[0021] It is understood that the term "a" should be understood as "at least one" or "one or more," that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple, and the term "a" should not be understood as a limitation on the number.

[0022] like Figure 1 and Figure 2 As shown, a full-process intelligent and compliant copywriting review technology system is constructed through multimodal data fusion modeling, accurate parsing of variant texts, cross-modal semantic verification, dynamic rule iteration, and continuous optimization using knowledge graphs. The specific details are as follows: Multimodal text data is collected and preprocessed according to standardization to construct an industry-specific variant knowledge base. Based on this knowledge base, text data is processed through a three-level parsing mechanism to generate standard text with variant tags. A domain compliance knowledge graph and natural language processing model are constructed, and the standard text with variant tags is input to strengthen compliance-sensitive word features. Contextual semantic association analysis and compliance rule matching verification are then implemented to generate text compliance verification results and standardized text features. Based on the standardized text features, key image region features and video subtitle features are extracted. Multimodal semantic consistency is determined through a cross-modal attention mechanism to form multimodal verification results. Combining the text compliance verification results, multimodal verification results, and dynamic rule base, differentiated review decisions are executed according to risk levels, and structured review results are output.

[0023] Furthermore, the specific construction process of the industry-specific variant knowledge base is as follows: Collect industry violation case databases, regulatory notices, and manually annotated variant samples to form the raw data of the knowledge base; store the raw data in a structured format using triples according to variant type; establish an automatic update mechanism to regularly expand variant term entries based on newly generated violation cases; manually review and confirm the expanded variant term entries to ensure that the accuracy of the data entering the database meets the usage requirements, forming a dynamically iterative industry-specific variant knowledge base.

[0024] The text data acquisition utilizes multi-platform interface synchronization technology, covering core scenarios such as e-commerce product descriptions, financial advertising copy, and social media content. Image data acquisition employs visual content capture technology to extract visual elements such as product tags, advertising images, and combined text and image content. Video data acquisition utilizes frame extraction and subtitle synchronization technology to obtain related data such as video frames, embedded subtitles, and audio-to-text transcription. In the standardized preprocessing, text data uses the BERT word segmentation algorithm (to achieve Chinese word segmentation and term restoration), noise reduction (setting a noise word threshold ≤0.05 to remove low-frequency meaningless words), and format unification (unified to UTF-8 encoding). Image data uses deblurring (Gaussian filter kernel size). The system employs 3×3 methods for entity detection (YOLOv8 algorithm, confidence threshold ≥0.7) and key region localization (RoIAlign algorithm, confidence threshold ≥0.65). For video data, it uses frame filtering (1 frame per second), subtitle recognition (CTC network, accuracy ≥98%), and temporal normalization (sorted by timestamp). Modal feature alignment uses cross-modal feature mapping technology (based on cosine similarity mapping) to eliminate representational differences between text, image, and video data. The variant knowledge construction architecture is based on triple storage technology to build the association between variant words, standard words, and industry tags, generating an industry-specific variant knowledge base that includes a variant term set, industry classification index, and automatic update mechanism.

[0025] The industry-specific variant knowledge base has a three-layer architecture, consisting of a raw data layer, a structured storage layer, and an update layer. The raw data layer receives and integrates the industry violation case library, regulatory notice texts, and manually annotated variant samples; the structured storage layer is classified by variant type and stored in a triple format of (variant term, standard term, industry tag), and uses the Neo4j graph database to support dual-dimensional indexing by industry and variant type (index query latency ≤5ms); the update layer integrates an automatic update trigger module and a manual review module, which expands the entries based on new violation cases every 24 hours according to a preset cycle, and triggers an update when the accumulation of new violation cases is ≥100, and enters the database after manual review and confirmation (review confirmation rate ≥95%).

[0026] In detail, the specific implementation logic of this step is as follows: The multimodal data acquisition module collects data synchronously through multiple channels, including multi-platform interfaces, visual capture, and video parsing (relying on initially configured interface adaptation rules, visual capture parameters, and video parsing protocols). It achieves temporal alignment of multimodal data using timestamp synchronization technology. The module performs format verification and outlier removal on the collected data (adopting data rationality verification criteria based on industry standards and data characteristics). Text data is segmented and denoised using BERT to extract core text fragments. Image data is deblurred and subjected to YOLOv8 entity detection, then RoIAlign is used to locate key areas such as product labels and advertising slogans. Video data is extracted and CTC subtitles are recognized, separating visual frames from text subtitles. Finally, a variant knowledge construction architecture is used to model the core features of the text, generating an industry-specific variant knowledge base. The entire process from acquisition to modeling meets the real-time requirements of the review process (≤500ms).

[0027] Specifically, the equipment and process details for data acquisition and knowledge base construction are as follows: The acquisition device has a multimodal data synchronous input function, a built-in high-performance processing unit (main frequency ≥2GHz, to meet the real-time data processing requirements), and integrates a text acquisition unit, an image capture unit, and a video analysis unit. The process is divided into the following steps: First, the device is initialized, the multimodal data acquisition unit is started, and the parameters are configured (acquisition frequency 10Hz, resolution accuracy 0.01mm, interface protocol HTTP / 2, etc.). Then, data acquisition is performed, triggering the corresponding modal data acquisition according to the audit scenario, with the capability of single-modal or multimodal collaborative acquisition. Next, the acquired data is preprocessed, performing noise reduction, format normalization, and other operations on a modal basis. Text data is additionally processed by BERT word segmentation, image data is additionally processed by RoIAlign key region localization (confidence ≥0.65), and video data is additionally processed by subtitle extraction. Then, feature extraction and knowledge base construction are performed, extracting core features according to data type characteristics, inputting them into the variant knowledge construction architecture after modal alignment, outputting an industry-specific variant knowledge base and storing it (caching 10 knowledge base versions, storage latency ≤10ms). Finally, the processed data is quality checked, and the knowledge base is checked for term coverage and accuracy (variant term coverage ≥90%, accuracy ≥95% is considered qualified). If it is unqualified, the acquisition and construction are retried (retry ≤3 times).

[0028] Furthermore, the process of obtaining the standard text with variant tags is as follows: Based on the TF-IDF algorithm and variant word matching, the preprocessed text data is traversed and detected to identify potential variant words; a restoration strategy is adapted for each variant type to obtain the restored text; the restored text is input into the BiLSTM model to calculate the semantic similarity between the restored text and the context; if the similarity reaches the preset similarity threshold, the standard text with variant labels is output; if the similarity does not reach the preset similarity threshold, a manual review process is triggered.

[0029] The variant detection technology employs a strategy combining the TF-IDF algorithm with a variant lexicon for matching. TF-IDF is used to calculate the weights of text terms, and the variant lexicon is then used to accurately match potential variant words. Variant types include homophonic substitution, symbolic insertion, mixed pinyin writing, character form variation, word order adjustment, simplified / traditional character mixing, variant character replacement, numerical homophony, English abbreviation transcription, dialect homophony, radical replacement, and character splitting. The differentiated restoration strategies are as follows: homophonic substitution uses the pinyin edit distance algorithm (matching threshold ≥ 0.85); symbolic insertion uses a symbol-to-Chinese character mapping table; and mixed pinyin writing uses a pinyin-to-Chinese character bidirectional dictionary combined with a context-based semantic priority strategy. For other types, a semantic association restoration algorithm is used (based on BERT to calculate semantic similarity, with a threshold of ≥0.8); the semantic verification model adopts a BiLSTM network architecture (2 hidden layers, 256 neurons per layer, activation function Tanh, dropout=0.3), inputting the restored text and context fragments, and outputting a semantic similarity value (based on cosine similarity calculation); the preset similarity threshold is set to 0.85 based on industry review accuracy requirements. If the threshold is reached, the standard text with variant labels is output (the labels include variant type, original variant word, and standard word); if the threshold is not reached, manual review is triggered (review trigger rate ≤5%).

[0030] The semantic verification model is specifically a deep learning architecture that includes an input layer, a hidden layer, and an output layer. The input layer receives the restored text vector (768-dimensional, based on BERT encoding) and the context vector (768-dimensional); the hidden layer contains two layers of neurons, 256 neurons each, using the Tanh activation function, with dropout set to 0.3 to prevent overfitting; the output layer uses the Sigmoid activation function to output semantic similarity values ​​(mapped to the [0,1] interval); the model optimizer uses the Adam algorithm (initial learning rate of 0.001, iterative training until convergence with validation set loss ≤ 0.005), ultimately constructing the semantic association determination capability between the restored text and the context.

[0031] In detail, the specific implementation logic of this step is as follows: First, load the industry-specific variant knowledge base and initialize the parameters of the variant detection and restoration module; perform BERT tokenization and term weight calculation on the preprocessed text data, and combine with the variant vocabulary to match and detect potential variant words; for the detected variant words, call the corresponding restoration strategy according to their variant types to obtain the restored text; input the restored text and context fragments into the semantic verification model to calculate the semantic similarity; if the similarity ≥ 0.85, generate the corresponding variant label (indicating the variant type, original variant word, and standard word), and output it in association with the standard text; if the similarity < 0.85, mark it as a suspected incorrect restoration and trigger the manual review process (the manual review results are synchronously fed back to the knowledge base update layer); finally, output the standard text with variant labels or manual review work orders to ensure the accuracy of text parsing.

[0032] Specifically, the process details of variant parsing and standard text generation are as follows: The core processing unit integrates a variant detection module, a variant restoration module, and a semantic verification module, and is deployed on the audit server or cloud node; the process includes: first, load the industry-specific variant knowledge base and semantic verification model to complete parameter initialization; then perform BERT tokenization on the input text, calculate the term weights through the TF-IDF algorithm, and combine with the variant vocabulary to match and detect potential variant words; then adapt the restoration strategy according to the variant type and perform the variant restoration operation to generate the preliminary restored text; then input the preliminary restored text and context fragments into the semantic verification model to calculate the semantic similarity and compare it with the 0.85 threshold; next, output the standard text with variant labels or trigger manual review according to the comparison result; finally, perform quality verification on the parsing results. If the accuracy of the standard text ≥ 95% is qualified, otherwise, re-execute the parsing (the number of parsing times ≤ 2 times). The single-text processing delay during the parsing process meets the high-concurrency audit requirements (≤ 200ms), the variant recognition coverage rate ≥ 92%, and the restoration accuracy rate ≥ 93%.

[0033] Furthermore, the process of obtaining the text compliance verification results and the standardized text features is as follows: Sort out the industry compliance rules and build a domain compliance knowledge graph containing entities, relationships, and violation judgment criteria; use the TransE algorithm to transform the entities and relationships in the domain compliance knowledge graph into preset-dimensional vectors to achieve graph vector embedding; based on the BERT-Base model, add a global semantic attention layer, and calculate the feature weights through compliance-sensitive feature quantization and fusion; input the embedded graph vectors and text word vectors into a multi-layer Transformer encoder for feature fusion to strengthen the feature expression of compliance-sensitive segments, and achieve context semantic association analysis and compliance rule matching verification.

[0034] The industry compliance rules cover regulatory requirements for multiple industries such as e-commerce, finance, and social media, including core rules such as prohibition of absolute terms, mandatory risk warnings, and restrictions on prohibited content. The domain compliance knowledge graph is constructed in the form of "entity-relationship-violation judgment standard". Entities include compliance terms, violation terms, and regulatory clauses, while relationships include "prohibited association", "required association", and "association based on". The violation judgment standard clarifies the judgment conditions for various violation scenarios. The graph embedding algorithm adopts the TransE algorithm, which transforms entities and relationships into 256-dimensional vectors and constructs entity-relationship mappings through vector operations. The basic natural language processing model adopts the BERT-Base model, and the global semantic attention layer calculates feature weights based on compliance-sensitive features to enhance the feature expression of compliance-sensitive segments. The feature fusion architecture adopts a 6-layer Transformer encoder to achieve deep fusion of word vectors and graph vectors. The text compliance verification results include compliance / violation judgment, violation type, and violation basis. The standardized text features are high-dimensional vectors that integrate compliance semantics (768 dimensions, consistent with the BERT output dimension).

[0035] The feature weights of the global semantic attention layer are calculated using a weighted fusion algorithm: the weight calculation formula is as follows. ( For feature weights, =0.6 is the weighting coefficient. For the first The weight of compliant sensitive words for each term. The formula is used to strengthen the feature weights of compliance-sensitive words and highly correlated words of violation entities, thereby improving the accuracy of compliance verification. The Transformer encoder of the feature fusion architecture contains 6 layers, each of which adopts an 8-head attention mechanism and a feedforward neural network (hidden layer dimension 2048) to achieve deep fusion and dimension unification of vector features, and output 768-dimensional standardized text features.

[0036] In detail, the specific implementation logic of this step is as follows: First, sort out the compliance rules and regulatory provisions in various industries, extract compliance entities, violation entities and their associated relationships, clarify the violation judgment criteria, and construct a domain compliance knowledge graph; use the TransE algorithm to transform the entities and relationships in the knowledge graph into 256-dimensional vectors to complete the graph vector embedding; based on the BERT-Base model, add a global semantic attention layer, and calculate the feature weights through the quantization and fusion of compliance-sensitive features; convert the standard text with variant labels into 768-dimensional word vectors after BERT tokenization, and input them together with the embedded 256-dimensional graph vectors into a 6-layer Transformer encoder for feature fusion and enhancement; analyze the matching degree between the text semantics and the compliance rules through the compliance rule matching module (a similarity ≥ 0.7 is determined to match the violation rules) to generate the text compliance verification result; at the same time, extract the fused 768-dimensional high-dimensional vector as the standardized text feature; finally, output the text compliance verification result and the standardized text feature, and verify the reliability of the result through the compliance verification accuracy test (an accuracy ≥ 92% is considered qualified).

[0037] Specifically, the process details of knowledge graph construction, model training and feature extraction are as follows: The data processing and model training unit integrates a knowledge graph construction module, a model training module and a feature extraction module, and supports offline rule import and online model iteration; the specific process includes: First, collect industry compliance rules, regulatory provisions, and violation cases, extract entities, relationships and violation judgment criteria, construct a domain compliance knowledge graph and store it (using the Neo4j graph database for storage to achieve fast query of entities and relationships, with a query latency ≤ 3ms); Subsequently, use the TransE algorithm to perform vector embedding on the knowledge graph to generate 256-dimensional entity vectors and relationship vectors; Then, improve the BERT-Base model, add a global semantic attention layer, load the compliance-sensitive word library and graph vectors, and train the model until the loss of the validation set converges (the loss value ≤ 0.008); Next, input the standard text with variant labels into the trained model, and after word vector conversion, graph vector fusion, and feature enhancement, output the text compliance verification result and the standardized text feature; Finally, select 1000 typical compliance and violation texts to verify the result. A compliance verification accuracy ≥ 92% is considered qualified, and if it is unqualified, the model is re-optimized (the number of optimizations ≤ 2 times). The single-text processing latency after model training meets the high-concurrency review requirements (≤ 150ms), and the compliance verification accuracy ≥ 92%.

[0038] Furthermore, the specific process of obtaining the multi-modal verification result is as follows: Based on standardized text features, a ResNet-50 network is used to extract features from images. After fine-tuning with an industry image dataset, key regions are located using RoIAlign. CNN text detection and CTC subtitle recognition algorithms are used to detect and recognize text regions in video frames, converting them into standardized subtitle features. Cosine similarity is used to calculate the semantic correlation between text and images, and between text and subtitles. Semantic conflict feature pairs with semantic correlation below the semantic correlation threshold are given high attention weights. The features of each modality are weighted according to attention weights and concatenated to generate a fusion vector. The similarity is calculated with the semantically consistent template vector to determine multimodal semantic consistency and form a multimodal verification result.

[0039] The visual feature extraction network employs a ResNet-50 network architecture (containing 5 convolutional stages, with 3 / 4 / 6 / 3 residual blocks for Conv2-Conv5 respectively), transforming image features into 2048-dimensional vectors through GlobalAveragePooling. Industry image datasets, including 500,000 e-commerce product images and 200,000 financial advertising images, are used to fine-tune the network and improve the accuracy of key region feature extraction. The key region localization technology uses the RoIAlign algorithm (pooling resolution 7×7) to accurately locate core regions such as product labels and advertising slogans. The text detection algorithm uses a lightweight CNN architecture (MobileNetV2) to quickly identify text regions in video frames (IOU threshold 0.5). The text recognition algorithm uses a CTC network (encoder is BiLSTM, decoder is CTCBeamSearch) to achieve accurate recognition of subtitle text (accuracy ≥98%). The cross-modal attention mechanism calculates the semantic correlation of features across modalities using cosine similarity, with the formula: The high attention weight is calculated using a preset formula: weight value = 1 - semantic relevance (e.g., when the relevance is 0.4, the weight value = 0.6). This weight allocation strategy is validated based on 100,000 multimodal samples (40,000 e-commerce samples, 30,000 financial samples, and 30,000 social samples). Conflict feature pairs with semantic relevance < 0.6 account for 35% (38% / 31,200 e-commerce samples, 32% / 9,600 financial samples, and 36% / 10,800 social samples). Among these, the main conflict in e-commerce scenarios is "conflict between product description and image label" (62%), in finance scenarios it is "conflict between profit promotion and missing risk warning" (58%), and in social scenarios it is "semantic conflict between text content and subtitles" (55%). After adopting this weight allocation, the conflict recognition accuracy is improved by 12% compared to equal weights (e-commerce from 80% to 92%, finance from 82% to 94%, and social from 79% to 91%). The sample distribution follows a normal distribution (mean 0.45, variance 0.08), and is verified by paired t-tests. (p < 0.01) The difference was statistically significant. The semantically consistent template vector was generated based on 100,000 compliant multimodal samples (2048 + 768 + 768 = 3584 dimensions). The sample annotation standard was as follows: first, the annotators were trained on compliance rules (training time ≥ 8 hours). After unifying the annotation standard, a two-round annotation and sampling verification process was adopted. The manual annotation ratio was ≥ 80% (50% for senior compliance experts / 50,000 people, 30% for intermediate auditors / 30,000 people), and the machine pre-annotation ratio was 20% / 20,000 people. The annotation consistency was verified by Cohen's Kappa coefficient (Kappa ≥ 0.85). The sampling verification sampled 10,000 samples (3333 / 3333 / 3334 samples from various industries), and the annotation error rate was ≤ 2%. The industry distribution of the samples was consistent with the actual audit scenario (e-commerce 40%, finance 30%, social 30%), ensuring the scenario adaptability of the template vector. The multimodal verification results included semantic consistency / conflict determination, conflict modality identification, and conflict location information.

[0040] The cross-modal attention fusion architecture is specifically a four-level architecture including single-modal feature extraction, cross-modal association computation, feature fusion, and consistency determination. The single-modal feature extraction stage outputs 768-dimensional standardized text features, 2048-dimensional image key region features, and 768-dimensional standardized caption features, respectively. The cross-modal association calculation stage calculates the semantic association degree between text and image, and between text and caption, and assigns high weights to conflicting feature pairs with an association degree <0.6. The feature fusion stage concatenates the features of each modality according to the attention weight to generate a 3584-dimensional fusion vector. The consistency judgment stage calculates the similarity between the fusion vector and the semantically consistent template vector. A similarity of ≥0.7 is judged as semantically consistent, and <0.7 is judged as semantically conflicting.

[0041] In detail, the specific implementation logic of this step is as follows: First, load the 768-dimensional normalized text features, and initialize the parameters of the ResNet-50 image feature extraction network, video subtitle recognition module, and cross-modal attention module. After preprocessing the image data (using 3×3 Gaussian filtering for deblurring and resizing to 224×224), input it into the ResNet-50 network to extract 2048-dimensional visual features, and locate the key region features through the RoIAlign algorithm (7×7 pooling resolution). Extract frames from the video data (1 frame per second), detect the text region through MobileNetV2, input it into the CTC network to recognize the subtitle text and convert it into 768-dimensional normalized subtitle features. Calculate the semantic correlation between text and image, and text and subtitle through the cross-modal attention mechanism, and assign high attention weights (weight value = 1 - correlation) to the feature pairs with a correlation lower than 0.6. Concatenate the features of each modality weighted by the weights to generate a 3584-dimensional fusion vector, and calculate the similarity with the semantic-consistent template vector. If the similarity ≥ 0.7, it is determined to be semantically consistent; if the similarity < 0.7, it is determined to be semantically conflicting, and mark the conflicting modality and conflict position. Finally, output the multi-modal verification result to ensure the accurate determination of multi-modal semantic consistency.

[0042] Specifically, the process details of multi-modal feature extraction and consistency determination are as follows: The core processing unit integrates an image feature extraction module, a video subtitle recognition module, and a cross-modal fusion module, and is deployed on a high-performance computing node. The process includes: first, load the normalized text features, the pre-trained ResNet-50 network, and the CTC recognition model; then preprocess the input image, perform feature extraction and key region localization, and output 2048-dimensional image key region features; then extract frames from the input video, perform text region detection and subtitle recognition, and output 768-dimensional normalized subtitle features; then calculate the semantic correlation of each modality through the cross-modal attention mechanism, assign high weights to the conflicting feature pairs and perform feature fusion; next, compare the fusion vector with the semantic-consistent template vector to determine the multi-modal semantic consistency; finally, output the multi-modal verification result, and perform a conflict recognition accuracy test on the result (the conflict recognition accuracy ≥ 95% is qualified), and if it is unqualified, readjust the module parameters (the number of adjustments ≤ 2 times). The processing delay of a single multi-modal data ≤ 300ms, and the semantic consistency determination accuracy ≥ 93%.

[0043] Furthermore, the specific process of obtaining the dynamic rule library is as follows: Based on the violation judgment criteria in the domain compliance knowledge graph and combined with historical violation cases, the correspondence between violation characteristics and risk levels is extracted to generate an initial rule set. A rule validity verification algorithm is used to verify the initial rule set through historical audit samples, and rules that meet the accuracy requirements are selected and included in the dynamic rule base. A rule update trigger mechanism is established to periodically mine new combinations of violation characteristics based on newly generated violation cases and changes in compliance rules to generate candidate rules. After validity verification, qualified candidate rules are added to the dynamic rule base, and the rule priorities are adjusted accordingly.

[0044] The initial rule set is constructed in the format of violation feature - risk level - handling strategy. Violation features include text violation features, multimodal conflict features, etc., and risk levels are divided into low, medium, and high. The handling strategy clearly defines the review actions corresponding to the level. The rule validity verification algorithm is tested with 10,000 historical review samples to calculate the accuracy and recall of the rules, and rules with an accuracy of ≥90% and a recall of ≥85% are selected. The dynamic rule base adopts an architecture format of rule storage - priority management - update interface. Rule storage is organized in XML format by industry - rule type hierarchy. Priority management sets the priority according to the importance of the rule and the applicable scenario (levels 1-5, with level 5 being the highest). Update... The interface supports incremental addition and priority adjustment of rules; the risk level assessment system is constructed based on the intensity of violation characteristics (0-10 points, ≥8 points are significant, 3-7 points are clear, <3 points are weak), the degree of semantic conflict (0-10 points, ≥7 points are severe, 4-6 points are clear, <4 points are none), and the rule matching results. Low risk corresponds to weak violation characteristics (<3 points) and no clear semantic conflict (<4 points), medium risk corresponds to clear violation tendency (3-7 points) but requires manual confirmation, and high risk corresponds to significant violation characteristics (≥8 points) or meets the mandatory blocking rules; differentiated review decisions include actions such as automatic marking for sampling inspection, generating manual review work orders, and direct blocking of publication.

[0045] The rule update triggering mechanism is a dual-mode system combining periodic triggering and event-driven triggering. Regular triggers are executed on a 72-hour cycle, while event-driven triggers target scenarios such as the release of major compliance policies and the emergence of new violation patterns. New violation feature combination mining employs the FP-Growth algorithm, with minimum support dynamically adapted to the industry (e-commerce 0.05, finance 0.03, social media 0.06), based on the density of violation cases in the industry (e-commerce has the highest density, followed by social media, and then finance). Rule generation thresholds are set at "confidence ≥ 0.9, support ≥ 0.02, and lift ≥ 1.5," with lift used to filter rules with discriminative power (excluding commonly present non-violation-related features). Candidate rule validity is verified through a combination of historical sample testing and two rounds of manual review. The manual review criteria are: the first round of review is conducted independently by at least three experts with over three years of industry compliance experience (eliminating rules with questionable compliance or narrow scenario coverage). The first round of review involves narrow rules, with a second round of review by two senior experts (focusing on verifying the feasibility of rule implementation). Compliance experts must participate at least 70% of the team, and the review consensus must be at least 80% (number of votes in favor / total number of votes). Review dimensions include rule compliance (whether it complies with regulatory provisions), scenario coverage (whether it covers at least three sub-scenarios), and implementation feasibility (whether it can be implemented through existing systems). The rule priority adjustment strategy employs a three-dimensional approach combining dynamic weighting of risk level, call frequency, and false positive rate: high-risk rules receive a 2-point priority increase, medium-risk rules receive a 1-point increase, and low-risk rules receive no adjustment. Rules with ≥1000 monthly calls receive an additional 0.5 priority increase. Rules with a false positive rate ≤3% after being hit receive an additional 0.3 priority increase (false positive rate = number of false positive cases / total number of hit cases). The final priority is still limited to levels 1-5 to ensure accurate reach of high-priority rules. The overall pass rate is at least 90%.

[0046] In detail, the specific implementation logic of this step is as follows: First, load the domain compliance knowledge graph and historical violation cases (≥50,000 entries), extract the correspondence between violation characteristics and risk levels, and generate an initial rule set. Then, validate the initial rule set using 10,000 historical audit samples, selecting rules with an accuracy ≥90% and a recall ≥85% for inclusion in the dynamic rule base. Initialize the rule update trigger mechanism, setting a 72-hour regular update cycle and event-driven trigger conditions. Periodically mine new violation characteristic combinations according to the trigger mechanism, generating candidate rules (confidence ≥0.9, support ≥0.02, lift ≥1.5), and after validity validation (pass rate ≥90%), supplement them to the dynamic rule base. Dynamically adjust rule priorities according to risk level, call frequency, and false positive rate. Integrate text. The strength of violation features (0-10 points) in the compliance verification results, the degree of semantic conflict (0-10 points) in the multimodal verification results, and the rule matching results from the dynamic rule base are substituted into the risk level assessment system. Differentiated review decisions are executed based on the assessment results: low risk (violation feature < 3 points and semantic conflict < 4 points) is automatically marked as "requires secondary sampling inspection", medium risk (violation feature 3-7 points or semantic conflict 4-6 points) generates a review work order containing violation fragment identifiers and semantic analysis explanations and pushes it to the manual review queue, and high risk (violation feature ≥ 8 points or matching the mandatory interception rule) directly intercepts the publication of the copy and records the violation information. Finally, the review decision results and structured review results are output to ensure the compliance and efficiency of the review process.

[0047] Specifically, the process details for rule base construction, updating, and review decision execution are as follows: The rule management and review decision-making unit integrates a rule base construction module, a rule update module, and a decision execution module, supporting modular rule management and automated decision-making. The process includes: first, building an initial rule set based on a domain compliance knowledge graph and historical violation cases, and validating its effectiveness to generate a dynamic rule base; then, setting a rule update trigger mechanism to periodically trigger new rule mining and validity verification, adding qualified rule increments to the rule base and adjusting their priorities; next, loading text compliance verification results and multimodal verification results, and performing rule matching in conjunction with the dynamic rule base (matching latency ≤ 5ms); then, calculating the risk level through a risk level assessment system and executing differentiated review decisions; finally, outputting structured review results (including review conclusions, violation types, violation basis, and optimization suggestions), and feeding the review data back to the rule update module and model optimization module. The rule query latency of the dynamic rule base is ≤ 5ms, the accuracy rate of review decisions is ≥ 94%, and the high-risk interception rate is ≥ 98%.

[0048] Furthermore, the specific details of the differentiated review decision are as follows: A risk level assessment system is established by integrating the intensity of violation features in text compliance verification results, the degree of semantic conflict in multimodal verification results, and the rule matching results of the dynamic rule base. If the intensity of the violation features is lower than the preset intensity threshold and there is no clear semantic conflict, and it meets the fault tolerance rules, it is judged as low risk and automatically marked as requiring secondary sampling inspection. If the intensity of the violation features reaches the preset intensity threshold or there is a clear semantic conflict, but the basis for the violation needs further confirmation, it is judged as medium risk, and an audit work order containing violation fragment identifiers and semantic analysis explanations is generated and pushed to the manual review queue. If the intensity of the violation features is higher than the preset intensity threshold or meets the mandatory interception rules, it is judged as high risk, the publication of the text is directly blocked, the relevant violation information is recorded and synchronized to the platform's risk control system.

[0049] The preset intensity thresholds are as follows: low risk threshold < 3 points, medium risk threshold 3-7 points, and high risk threshold ≥ 8 points; semantic conflict level thresholds are: no explicit semantic conflict < 4 points, explicit semantic conflict 4-6 points, and severe semantic conflict ≥ 7 points; the fault tolerance rules include occasional "extreme" expressions in e-commerce copywriting without false advertising tendencies, and slightly inappropriate expressions in social copywriting without negative guidance; the mandatory blocking rules include explicit violations of regulatory provisions such as the promotion of prohibited drugs, violent text, and false promises of principal protection in finance; the review work order includes screenshots of the violation fragments, semantic analysis reports, and references to relevant compliance provisions, which facilitates rapid manual review; violation-related information includes the violation account ID, publication time, copy content, violation type, and blocking basis, which, after being synchronized to the platform's risk control system, supports subsequent operations such as accumulating the number of account violations and tiered penalties.

[0050] In detail, the specific implementation logic of this step is as follows: First, the text compliance verification results, multimodal verification results, and rule matching results from the dynamic rule base are loaded, and core indicators such as violation feature strength, semantic conflict degree, and rule matching level are extracted. Based on these indicators, a risk level assessment system is constructed, and a weighted summation is used to calculate the comprehensive risk score (violation feature strength weight 0.4, semantic conflict degree weight 0.35, rule matching level weight 0.25). The risk level is determined based on the comprehensive risk score: a score <3 is considered low risk, 3-7 is medium risk, and ≥8 is high risk. Low-risk text is automatically marked "requires secondary sampling" and stored in the sampling queue (sampling ratio...). 10%) For medium-risk copywriting, generate review work orders containing violation segment identifiers (highlighted text segments, marked conflicting areas in images / videos) and semantic analysis explanations (reasons of conflict, analysis of violation tendencies), and push them to the manual review queue according to priority (processing time ≤ 2 hours); For high-risk copywriting, directly intercept publication, return violation prompts to users (including violation type and applicable clauses), and record the violating account information and synchronize it to the platform's risk control system; The results of manual review are fed back to the dynamic rule base and model optimization module in real time. Copywriting that fails low-risk spot checks or is determined to be in violation in medium-risk reviews is added to the violation case library for rule and model optimization.

[0051] Specifically, the execution process for differentiated review decisions is detailed as follows: The review and decision-making unit integrates a risk assessment module, a work order generation module, an interception and execution module, and a result feedback module, supporting automated decision-making and manual collaboration. The process includes: first, loading various verification results and rule matching data; then, calculating the comprehensive risk score and level through the risk assessment module; next, executing corresponding decisions according to risk level: low risk triggers automatic marking and storage for random inspection; medium risk triggers work order generation and push; high risk triggers interception execution and information synchronization. The manual review module then receives medium-risk work orders, allowing reviewers to view violation details, semantic analysis reports, and related evidence, and provide conclusions such as "pass," "intercept," or "second review." Finally, all decision results (automatic and manual) are compiled into structured review results, fed back to users and the platform, and relevant data is synchronized to the optimization module. The average latency for decision execution is: low risk ≤100ms, medium risk ≤500ms, high risk ≤200ms; the accuracy rate for manual review work order processing is ≥98%, and the accuracy rate for high-risk interception is ≥99%.

[0052] Furthermore, the update process of the domain compliance knowledge graph is as follows: Establish data sources for graph updates, including newly added industry compliance policies, regulatory notifications, uncovered violation cases, and supplementary compliance rules information provided by manual feedback; set graph update trigger conditions, including periodic update triggers and event-driven updates; when the update trigger conditions are met, extract the compliance entities, relationships, and judgment criteria from the newly added data and perform structured processing according to the preset graph data format; adopt a graph incremental update algorithm to integrate the newly added structured data into the existing domain compliance knowledge graph, updating entity vectors and relationship mappings; verify the effectiveness of the updated graph through compliance rule matching tests to ensure that the newly added content does not conflict with existing rules and maintains the consistency and accuracy of the graph.

[0053] The updated graph data is obtained synchronously through multiple channels. Newly added industry compliance policies and regulatory notices are obtained through official interfaces or text crawling technology. Uncovered violation cases are obtained from missed detection feedback in the review process (≥50 cases trigger an update). Manual feedback information is submitted through the reviewer's interface. Regular updates are triggered on a 72-hour cycle, while event-driven updates are triggered for scenarios such as major policy releases and adjustments to industry compliance rules. Structured processing uses natural language processing technology (BERT word segmentation and entity recognition) to extract compliant entities, relationships, and judgment criteria from the new data and convert them into triples supported by the knowledge graph. The incremental update algorithm of the graph adopts a fusion strategy based on vector similarity (similarity ≥0.9 is judged as duplicate entities) to avoid duplicate entities and conflicting relationships and update entity vectors and relationship mappings. The compliance rule matching test uses 1,000 typical compliance and violation cases to verify the rule matching accuracy of the updated graph (accuracy decrease ≤5% is considered valid). The whole process optimization includes fine-tuning of model parameters and dynamic rule base supplementation to ensure that the review model and rule base are adapted to the new compliance requirements.

[0054] The graph incremental update architecture is specifically a five-level architecture: data acquisition, structured processing, conflict detection, incremental fusion, and validity verification. The data acquisition stage receives updated data from multiple sources; the structured processing stage transforms the data into standardized triples; the conflict detection stage identifies duplicate entities and conflicting relationships through vector similarity comparison (≥0.9) and rule verification; the incremental fusion stage integrates conflict-free data into the existing graph and updates entity vectors and relationship mappings; and the validity verification stage verifies the update effect through compliance rule matching tests (accuracy decrease ≤5%).

[0055] In detail, the specific implementation logic of this step is as follows: First, a multi-channel knowledge graph update data collection interface was established to collect newly added industry compliance policies, regulatory notifications, uncovered violation cases (≥50), and manual feedback information. Update trigger conditions were set, defining a 72-hour regular update cycle and event-driven trigger scenarios. When an update is triggered, the new data undergoes text parsing (BERT word segmentation and entity recognition) and structured processing to extract compliant entities, relationships, and judgment criteria, generating standardized triples. A conflict detection algorithm (vector similarity ≥0.9 for duplicates) identifies duplicate entities and conflicting relationships, merging duplicate entities and prioritizing conflicting relationships (official policies > regulatory notifications > manual feedback). An incremental knowledge graph update algorithm integrates the processed triples into the existing domain compliance knowledge graph, updating entity vectors and relationship mappings. Compliance rule matching tests were conducted using 1000 typical cases to ensure a rule matching accuracy decrease of ≤5%. Finally, based on the updated knowledge graph, the text compliance verification model parameters were fine-tuned (learning rate 0.0001, 50 iterations), and relevant rules were added to the dynamic rule base (≥10 rules / update) to optimize the entire review process.

[0056] Specifically, the process details for map updating and overall optimization are as follows: The knowledge graph management and optimization unit integrates a data acquisition module, an incremental update module, and a full-process optimization module, supporting automated updates of the graph and continuous optimization of the review system. The process includes: first, establishing multi-channel data acquisition interfaces, configuring a 72-hour update cycle and event-driven trigger conditions, and completing parameter initialization; then, collecting newly added update data according to the trigger conditions, performing structured processing and conflict detection; next, updating the domain compliance knowledge graph using an incremental update algorithm, and storing the updated graph version; then, verifying the graph's effectiveness through compliance rule matching tests (a decrease in rule matching accuracy ≤5% is considered acceptable); subsequently, based on the updated graph, fine-tuning the text compliance verification model parameters and supplementing the dynamic rule base; finally, selecting 2000 typical review scenarios to test the full-process review capability. A pass is considered acceptable if there is no significant decrease in review accuracy and violation interception rate (accuracy decrease ≤3%, interception rate decrease ≤2%), otherwise, rolling back to the previous version and updating again (update times ≤2). The graph update processing latency is ≤1000ms, and the review accuracy after full-process optimization is ≥93%.

[0057] In addition, the system also has the functions of model iteration and optimization, multi-industry batch review adaptation, and review scenario adaptiveness: Model iteration and optimization adjust model parameters and rules based on feedback from audit data (false positives, missed detections, manual review results, etc.). During the offline training phase, every ≥10,000 audit data entries (including both compliant and non-compliant data), the original training set is divided into training / validation / test sets in a 7:2:1 ratio. The text compliance verification model and semantic verification model are retrained, and model parameters are updated. During the online fine-tuning phase, the model parameters are fine-tuned every 1,000 new real-time audit data entries using an incremental learning algorithm (SGD optimizer, learning rate 0.0005) to avoid disasters. Difficulty in forgetting; multi-industry batch review adaptation supports simultaneous loading of compliance knowledge graphs, variant knowledge bases, and dynamic rule bases from multiple industries. Parallel scheduling is achieved using the Spark distributed computing framework, deployed across 8 computing nodes (4 cores, 8GB configuration). Task splitting strategy is based on a dual dimension of document type and industry tag (text documents are assigned to nodes 1-3, image documents to nodes 4-6, and video documents to nodes 7-8). Load balancing strategy combines round-robin with dynamic adjustment of resource utilization (when node CPU utilization ≥ 80% or memory utilization ≥ 85%). When the CPU / memory usage is less than 60%, new tasks are assigned to low-load nodes. Actual test data shows that: in the e-commerce industry, a single review of 10,000 text and image-mixed copy takes 50 seconds serially and 18.75 seconds in parallel (efficiency 320 pieces / second); in the financial industry, a single review of 10,000 text and video-mixed copy takes 53 seconds serially and 20.38 seconds in parallel (efficiency 310 pieces / second); in the social media industry, a single review of 10,000 plain text copy takes 48 seconds serially and 17.14 seconds in parallel (efficiency 340 pieces / second). The mixed review of three industries (3333 items each, totaling 9999 items) took 51 seconds serially and 19.61 seconds in parallel (efficiency of 320 items / second), which is 60% ≥ 40% more efficient than the single-industry serial review. The node load balancing error is ≤ 10% (CPU utilization of each node fluctuates within a range of 55%-65%). The review scenario is adaptive and supports automatic recognition of e-commerce product copywriting, financial advertisements, social content and other scenarios. The corresponding review strategy can be switched through configuration files (such as strengthening risk warning verification for financial advertisements and strengthening variant recognition and image-text consistency verification for e-commerce copywriting).

[0058] The system has data security protection functions. Sensitive information such as audit data, model parameters, and knowledge base data are stored in encrypted form (the encryption algorithm conforms to the AES-256 security standard), and the transmission process uses the TLS 1.3 encryption protocol. The data anonymization strategy is detailed: user identity information is partially masked (the middle 4 digits of the mobile phone number are replaced with "***", the first 6 and last 4 digits of the ID card number are retained, and the middle 8 digits are replaced with "********", the surname is retained, and the given name is replaced with ""); payment information is processed using SHA-256 salted hashing (the salt value is a 32-bit random string, stored independently in an encrypted database, and physically isolated from the hash result); privacy fields in the illegal documents (such as address, contact information, and bank card number) are also protected. The system automatically identifies and replaces "[Privacy Information]" placeholders using regular expression matching, achieving an accuracy rate of ≥99% in de-identification. It features system self-checking and fault alerting functions, performing regular automatic self-checks (≤5 minutes per cycle). Detection items include memory leak detection (memory growth ≤5% after 24 hours of continuous operation), model inference accuracy verification (accuracy ≥92% when randomly selecting 100 labeled data points for testing), interface connectivity detection (ping test response time ≤10ms), and storage media health status detection. When a fault is detected, it provides dual alerts via the system alarm interface and email, along with fault codes and suggested troubleshooting steps. Low-power mode implementation technology: automatically triggered after 30 seconds of no-review tasks, using a PMIC (Power Management Chip) dynamic... The power supply voltage was adjusted (from 1.2V to 0.8V), the CPU frequency was reduced from 2GHz to 500MHz, and the power supply to non-core hardware units such as video parsing and image capture was shut down (their power supply circuits were cut off). Unnecessary peripheral interfaces such as USB and HDMI were disabled, and computing resources were dynamically scheduled to keep only the data acquisition interface and the core communication module running. At the same time, redundant disk read and write was disabled, temporary data was cached in memory (cache capacity ≤ 512MB), and background log writing was disabled (only critical alarm logs were retained). Ultimately, the overall power consumption was ≤ 10mA (the measured standby power consumption was 8.2mA, a reduction of 81.8% compared to the normal operating power consumption of 45mA). After detecting the audit trigger signal, the system woke up within ≤ 50ms and resumed full-power operation (power supply was cut off). The voltage rises back to 1.2V, the CPU frequency recovers to 2GHz, and the core modules and peripheral interfaces are activated sequentially. The audit log mechanism is strengthened: log recording requirements include operator ID, operation timestamp (accurate to milliseconds), operation type (automatic review / manual review / rule update / model optimization / data export), audit object ID (unique document identifier), processing result, data source (platform name and interface address), and operation terminal information (IP address and device model). Logs adopt an immutable chained storage structure with a retention period of ≥180 days. They support multi-dimensional filtering and querying by operator, time range, operation type, audit object ID, etc., and support exporting in Excel and CSV formats to meet regulatory audit and fault tracing needs.

[0059] The processes described above with reference to the flowcharts in the embodiments disclosed in this invention can be implemented as computer software programs. The embodiments disclosed in this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), it performs the functions defined in the methods of this application. It should be noted that the computer-readable medium described above in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wire segments, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to: wireless segments, wire segments, optical fibers, RF, etc., or any suitable combination thereof.

[0060] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0061] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The purpose of the present invention has been fully and effectively achieved. The functions and structural principles of the present invention have been shown and explained in the embodiments. Without departing from the stated principles, the implementation of the present invention may have any variations or modifications.

Claims

1. A copywriting review method based on cross-modal natural language processing, characterized in that, The method includes: Collect multimodal copywriting data, and after standardized preprocessing, build an industry-specific variant knowledge base; Based on an industry-specific variant knowledge base, text data is processed through a three-level parsing mechanism to generate standard text with variant tags. Construct a domain compliance knowledge graph and natural language processing model, input standard text with variant labels, strengthen compliance-sensitive word features, realize context semantic association analysis and compliance rule matching verification, and generate text compliance verification results and standardized text features; Based on standardized text features, key region features of images and subtitle features of videos are extracted. Multimodal semantic consistency is determined through a cross-modal attention mechanism to form a multimodal verification result. By combining text compliance verification results, multimodal verification results, and dynamic rule base, differentiated audit decisions are made according to risk level, and structured audit results are output.

2. The document review method based on cross-modal natural language processing according to claim 1, characterized in that, The specific construction process of the industry-specific variant knowledge base is as follows: Collect industry violation case databases, regulatory notices, and manually annotated variant samples, and integrate them to form the raw data of the knowledge base; Based on variant type, the original data is stored in a structured manner using a triplet format; Establish an automatic update mechanism to regularly expand the variant word entries based on newly generated violation cases; The expanded variant entries are manually reviewed and confirmed to ensure that the accuracy of the data entering the database meets the usage requirements, forming a dynamically iterative industry-specific variant knowledge base.

3. The document review method based on cross-modal natural language processing according to claim 2, characterized in that, The process of obtaining the standard text with variant tags is as follows: Based on the TF-IDF algorithm and variant word matching, the preprocessed copy data is traversed and detected to identify potential variant words; The restoration strategy is adapted to the variant type to obtain the restored text; The restored text is input into the BiLSTM model to calculate the semantic similarity between the restored text and the context. If the similarity reaches the preset similarity threshold, the standard text with variant labels is output; if the similarity does not reach the preset similarity threshold, a manual review process is triggered.

4. The document review method based on cross-modal natural language processing according to claim 3, characterized in that, The specific process for obtaining the text compliance verification result and the standardized text features is as follows: Organize industry compliance rules and construct a domain compliance knowledge graph that includes entities, relationships, and violation judgment standards; The TransE algorithm is used to transform entities and relationships in the domain compliance knowledge graph into vectors with preset dimensions, thereby enabling graph vector embedding. Based on the BERT-Base model, a global semantic attention layer is added, and feature weights are calculated by quantizing and fusing compliance-sensitive features. The embedded graph vectors and text word vectors are input into a multi-layer Transformer encoder for feature fusion, which enhances the feature representation of compliance-sensitive segments and enables contextual semantic association analysis and compliance rule matching verification.

5. A text review method based on cross-modal natural language processing according to claim 4, characterized in that, The specific process for obtaining the multimodal verification results is as follows: Based on standardized text features, ResNet-50 network is used to extract features from images. After fine-tuning with industry image datasets, key regions are located using RoIAlign. The CNN text detection and CTC subtitle recognition algorithms are used to detect and recognize text regions in video frames and convert them into standardized subtitle features. The semantic relevance between text and image, and between text and subtitle is calculated using cosine similarity. Semantic conflict features with semantic relevance below the semantic relevance threshold are given high attention weights. The features of each modality are weighted according to the attention weight and concatenated to generate a fusion vector. The similarity is calculated with the semantically consistent template vector to determine the semantic consistency of the multimodalities and form a multimodal verification result.

6. The document review method based on cross-modal natural language processing according to claim 5, characterized in that, The specific process of obtaining the dynamic rule base is as follows: Based on the violation judgment criteria in the domain compliance knowledge graph and combined with historical violation cases, the correspondence between violation characteristics and risk levels is extracted to generate an initial rule set; A rule validity verification algorithm is adopted to verify the initial rule set through historical audit samples, and rules that meet the accuracy requirements are selected and included in the dynamic rule base. Establish a rule update trigger mechanism to periodically identify new combinations of violation characteristics and generate candidate rules based on newly generated violation cases and changes in compliance rules; After validity verification, qualified candidate rules are added to the dynamic rule base, and the rule priorities are adjusted accordingly.

7. A text review method based on cross-modal natural language processing according to claim 6, characterized in that, The specific details of the differentiated review decision are as follows: By integrating the intensity of violation features in text compliance verification results, the degree of semantic conflict in multimodal verification results, and the rule matching results of dynamic rule base, a risk level assessment system is established. If the intensity of the violation feature is lower than the preset intensity threshold and there is no clear semantic conflict, and it meets the fault tolerance rules, it is judged as low risk and automatically marked as requiring a second sampling inspection; If the intensity of the violation reaches the preset intensity threshold or there is a clear semantic conflict, but the basis for the violation needs to be further confirmed, it is judged as medium risk, and an audit work order containing the violation fragment identifier and semantic analysis description is generated and pushed to the manual audit queue. If the intensity of the violation characteristics exceeds the preset intensity threshold or meets the mandatory blocking rules, it is judged as high risk, the publication of the copy is directly blocked, the relevant information of the violation is recorded and synchronized to the platform's risk control system.

8. A text review method based on cross-modal natural language processing according to claim 7, characterized in that, The update process of the domain compliance knowledge graph is as follows: Establish data sources for updating the compliance map, including newly added industry compliance policies, regulatory notifications, uncovered violation cases, and supplementary compliance rules information provided by human feedback; Configure map update trigger conditions, including periodic update triggers and event-driven updates; When the update trigger condition is met, extract the compliant entities, relationships and judgment criteria from the newly added data, and perform structured processing according to the preset graph data format; The graph incremental update algorithm is used to integrate newly added structured data into the existing domain compliance knowledge graph and update entity vectors and relationship mappings. The effectiveness of the updated graph is verified through compliance rule matching tests to ensure that the new content does not conflict with existing rules.

9. A document review system based on cross-modal natural language processing, characterized in that, The system is used to perform a text review method based on cross-modal natural language processing as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that is executed by a processor to implement a copy review method based on cross-modal natural language processing as described in any one of claims 1-8.

Images