Large model sensitive large content filtering method in learning scene

Abstract

The invention discloses a large-model sensitive large-content filtering method in a learning scene, and the method comprises the steps: extracting an image-side visual coding feature and a text-side semantic coding feature of a learning problem from the learning problem inputted by a user through employing a neural network model based on deep learning; and multi-modal feature fusion is further carried out on the characters and the pictures to form more comprehensive and accurate question joint representation, and the joint representation can fully mine the deep relation between the characters and the pictures, so that sensitive or improper contents in complex input are effectively identified, and automatic and intelligent filtering of the compound learning questions is realized. The problems that a traditional single-mode filtering scheme is narrow in coverage, low in detection accuracy, low in manual auditing efficiency and the like are solved, the recognition capacity of a platform for sensitive content is improved, and then a firm and reliable safety guarantee is provided for an intelligent education platform.

Description

Technical Field

[0001] The present application relates to the field of large models, and more specifically, to a large model-sensitive large content filtering method in a learning scenario. Background Art

[0002] With the widespread application of large models in learning scenarios such as education, it is increasingly common for users to raise questions in various forms, including text and images. To maintain a healthy and safe educational environment and prevent inappropriate or sensitive content (such as off-topic content and illegal information) from entering the platform, it is imperative to build efficient large-model sensitive content filtering solutions. On the one hand, traditional manual review methods are inefficient and highly subjective, making it difficult to meet the requirements of large-scale automated management. On the other hand, filtering methods that rely solely on a single modality (such as plain text or pure images) cannot fully cover the complex and diverse user input. For example, text filtering technology cannot identify inappropriate content in images, while image filtering methods tend to overlook potential problems in text descriptions, resulting in significant blind spots in sensitive content detection.

[0003] Existing large-scale sensitive content filtering solutions primarily rely on feature extraction and analysis of a single modality (such as pure text or images). While these approaches have achieved some success in their respective fields, they often struggle with complex learning problems that integrate text and images. Furthermore, while some systems attempt to combine manual review with automated detection, manual intervention not only increases operational costs but also makes it difficult to ensure consistent and timely review standards. Consequently, current solutions generally suffer from significant shortcomings such as limited detection accuracy, insufficient coverage, and low automation.

[0004] Therefore, we look forward to an optimized large-model-sensitive and large-content filtering method in learning scenarios. Summary of the Invention

[0005] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a large-model sensitive large-content filtering method in a learning scenario, which extracts the image-side visual coding features and text-side semantic coding features of the learning problems from the learning problems input by the user by adopting a neural network model based on deep learning, and further performs multimodal feature fusion on the two to form a more comprehensive and accurate joint representation of the problem, which can fully explore the deep connection between text and pictures, thereby effectively identifying sensitive or inappropriate content in complex inputs, and realizing automated and intelligent filtering of complex learning problems. In this way, the problems of narrow coverage, low detection accuracy and low efficiency of manual review of traditional single-modal filtering solutions are solved, and the platform's ability to identify sensitive content is improved, thereby providing a solid and reliable security guarantee for the intelligent education platform.

[0006] According to one aspect of the present application, a large model-sensitive large content filtering method in a learning scenario is provided, which includes:

[0007] Obtaining a learning question input by a user, where the learning question includes a text description and image data;

[0008] Extract visual features from image data to obtain visual encoding features of the learning problem image side;

[0009] Extract semantic features from text description to obtain semantic encoding features of the learning problem text;

[0010] Perform multimodal feature fusion on the learning problem image visual encoding features and the learning problem text semantic encoding features to obtain a joint multimodal feature representation of the learning problem;

[0011] Based on the joint representation of multimodal features of the learning question, it is determined whether the large model answers the learning question input by the user.

[0012] Beneficial effects: Compared with the existing technology, the present application provides a large-model sensitive large-content filtering method in a learning scenario, which extracts the image-side visual coding features and text-side semantic coding features of the learning problems input by the user by adopting a neural network model based on deep learning, and further fuses the two for multimodal features to form a more comprehensive and accurate joint representation of the problem. This joint representation can fully explore the deep connection between text and pictures, thereby effectively identifying sensitive or inappropriate content in complex inputs, and realizing automated and intelligent filtering of complex learning problems. This application solves the problems of narrow coverage, low detection accuracy and low efficiency of manual review of traditional single-modality filtering solutions, improves the platform's ability to identify sensitive content, and thus provides a solid and reliable security guarantee for the intelligent education platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0014] Figure 1 Flowchart of a large model-sensitive large content filtering method in a learning scenario according to an embodiment of the present application;

[0015] Figure 2 Schematic diagram of data flow of a large model-sensitive large content filtering method in a learning scenario according to an embodiment of the present application;

[0016] Figure 3 This is a flowchart of sub-step S4 of the large model-sensitive large content filtering method in a learning scenario according to an embodiment of the present application. DETAILED DESCRIPTION

[0017] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described herein.

[0018] As used in this application and the claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not intended to refer to the singular but may include the plural. Generally speaking, the terms "comprises" and "include" only indicate the inclusion of the steps and elements specifically identified, and these steps and elements do not constitute an exclusive list. A method or apparatus may also include other steps or elements.

[0019] Although the present application makes various references to certain modules in the system according to embodiments of the present application, any number of different modules can be used and run on the user terminal and / or server. The modules are illustrative only, and different aspects of the system and method can use different modules.

[0020] Flowcharts are used in this application to illustrate the operations performed by the systems according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, the various steps may be processed in reverse order or simultaneously, as needed. Furthermore, other operations may be added to these processes, or one or more operations may be removed from these processes.

[0021] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described herein.

[0022] It should be noted in advance that the acquisition and processing of all information or data in the present invention are carried out in compliance with relevant data protection laws and policies and with the authorization given by the owner of the corresponding device.

[0023] In the technical solution of this application, a large-model-sensitive large-content filtering method in a learning scenario is proposed. Figure 1 This is a flowchart of a large model-sensitive large content filtering method in a learning scenario according to an embodiment of the present application. Figure 2 Schematic diagram of data flow of the large model sensitive large content filtering method in the learning scenario according to the embodiment of the present application. Figure 1 and Figure 2 As shown, the large model-sensitive large content filtering method in the learning scenario according to the embodiment of the present application includes the following steps: S1, obtaining a learning question input by a user, where the learning question includes a text description and image data; S2, extracting visual features from the image data to obtain visual coding features of the learning question image side; S3, extracting semantic features from the text description to obtain semantic coding features of the learning question text side; S4, performing multimodal feature fusion of the learning question image side visual coding features and the learning question text side semantic coding features to obtain a joint representation of multimodal features of the learning question; S5, determining whether the large model answers the learning question input by the user based on the joint representation of multimodal features of the learning question.

[0024] In particular, the S1 obtains the learning questions input by the user, and the learning questions include text descriptions and image data. It should be understood that the learning questions raised by users are increasingly showing multimodal characteristics, including both text descriptions and image data. It is difficult to accurately identify and judge complex or hidden sensitive content by relying solely on single-modal information (such as only text or only pictures). For example, some inappropriate information may only appear when the text and pictures are combined, and analyzing the text or pictures alone is prone to omission risks. In the technical solution of the present application, by obtaining the learning questions input by the user, it can help the system adapt to the diversity of question expressions in actual teaching interactions, ensure that the system can fully receive and understand the user's real needs, and provide richer and more comprehensive information input for the sensitive content detection model, thereby improving the system's understanding and recognition capabilities of learning problems in complex and diverse forms of expression, realizing automated, efficient and highly accurate content security management, and providing a strong guarantee for the education platform to create a healthy and safe interactive environment.

[0025] In particular, S2 extracts visual features from the image data to obtain visual encoding features for the learning problem image. Specifically, in the technical solution of this application, a visual feature extractor based on a dilated convolutional neural network model is used to extract visual features from the image data to obtain a visual feature encoding vector for the learning problem image as the visual encoding features for the learning problem image. It should be understood that user-submitted image data often contains dense formula derivations, experimental setup diagrams, or annotated schematics. These contents may require accurate recognition due to the presence of sensitive symbols (such as inappropriate signs or graphic elements that imply negative information) in local areas. Traditional convolutional neural networks, due to their limited receptive field, tend to overlook discretely distributed sensitive feature points in an image. The dilated convolutional neural network introduces "holes" between standard convolution kernels, i.e., skip sampling, which allows the network to significantly expand its receptive field without increasing the number of parameters or computational burden. This mechanism enables the model to effectively balance local details and global structure, resulting in visual feature encoding vectors that not only have high resolution but also reflect complex spatial relationships, thereby significantly improving the detection capability of hidden sensitive content. Furthermore, this approach reduces spatial information loss, providing a more complete and discriminative visual feature foundation for subsequent multimodal fusion. By extracting visual features from image data based on a dilated convolutional neural network model, it not only meets the requirements of large multimodal models for high-quality, rich visual representations, but also significantly enhances the system's ability to automatically and intelligently identify complex and sensitive content. This initiative effectively ensures the learning platform's safe filtering of problematic images, providing solid technical support for creating a healthy and trustworthy educational environment.

[0026] In particular, the S3 extracts semantic features from the text description to obtain semantic coding features of the learning question text. Specifically, in the technical solution of the present application, a semantic feature extractor based on the Bert model is used to extract semantic features from the text description to obtain a semantic feature coding vector of the learning question text as the semantic coding feature of the learning question text. It should be understood that the questions raised by users usually contain rich and complex text information. These text descriptions not only carry the core content of the question, but may also contain sensitive or inappropriate information. In order to accurately understand and extract deep semantic features in the text, the system introduces a semantic feature extractor based on the Bert model. The Bert model is an advanced pre-trained language representation model. Through the bidirectional Transformer structure, it can simultaneously consider the left and right contexts of the context to achieve comprehensive capture of the text semantics. Different from the traditional unidirectional language model, Bert performs self-supervised learning through the Masked Language Model and Next Sentence Prediction tasks, which gives it powerful semantic understanding and context association capabilities. When processing text descriptions of learning questions, the Bert model can effectively parse multi-level linguistic information, including word meaning, syntactic structure, and context, thereby generating a highly expressive semantic feature encoding vector for the learning question text. This vector not only captures the literal meaning of the text but also reflects underlying intent and subtle semantic differences, facilitating the accurate identification of sensitive content. Using a semantic feature extractor based on the Bert model not only meets the need for a deep understanding of complex text semantics, but also enhances the accuracy and robustness of sensitive content detection, effectively supporting the system's intelligent processing of multimodal learning problems and ensuring the content security and service quality of the education platform.

[0027] In particular, the S4 performs multimodal feature fusion of the learning problem on the visual encoding features of the learning problem image side and the semantic encoding features of the learning problem text side to obtain a joint representation of the learning problem multimodal features. It should be understood that there is often redundancy and noise interference between visual features and text features in the original space, and traditional simple feature splicing or weighted fusion methods are difficult to capture this abnormal pattern of semantic separation of images and texts but logical association. Therefore, in order to solve the core contradiction between collinearity and nonlinear interaction of multimodal data, a joint representation that can characterize the fine-grained association of image and text features is constructed. In the technical solution of the present application, multimodal feature fusion of the learning problem visual encoding features of the learning problem image side and the semantic encoding features of the learning problem text side is performed to obtain a joint representation of the learning problem multimodal features. That is, by constructing a performance-oriented graph structure driven by the attributes of the features themselves, the complex and detailed interactive relationship between visual and semantic features is deeply revealed and utilized. In this process, principal component analysis (PCA) is first used to deeply decouple visual and textual feature vectors, removing redundant information from the original features (such as general descriptors irrelevant to the image in BERT text features) while retaining the core components that best represent sensitive content. This decoupling effectively eliminates high-frequency noise common in educational contexts (such as repeated formula symbols in math problems or standard process terminology in lab reports), allowing subsequent analysis to focus on truly discriminative feature combinations. A PCKR network then maps the low-dimensional PCs into a high-dimensional nonlinear space, where anomalous image features (such as intentionally added micro-watermarks) are strongly correlated with specific keywords in the text (such as suggestive phrases). Furthermore, the degree of synergistic anomaly between cross-modal features is quantified using a synergistic factor correlation topology matrix. This correlation strength calculation not only considers local feature similarity but also captures risk patterns unique to educational contexts through the task-specific nature of the synergistic factor (for example, defining "image-text semantic conflict" as a performance metric), enabling the system to identify seemingly unrelated feature combinations that are potentially harmful. This design enables the system to identify combined sensitive content that is difficult to detect with traditional methods, such as splitting illegal information into text in local areas of images and scattered text prompts; subsequently, the introduction of graph convolutional networks solves the problem of long-range dependency of cross-modal features. In the complex learning problems submitted by students, sensitive signals may be scattered in multiple fragments of different modalities. Through the graph structure defined by the synergistic effectiveness factor correlation topology matrix, the model can establish multi-hop connections between image region features and text fragment features. In this way, the bottleneck of traditional fusion methods that are difficult to capture nonlinear and high-order relationships between multimodal features is effectively solved, so that the final joint representation of multimodal features not only reflects the comprehensive semantics and visual content of text and image information, but also enhances the sensitive content detection model's understanding and discrimination capabilities for complex inputs. In a specific example of the present application, such as Figure 3As shown, the S4 includes: S41, extracting the characteristic principal components of the visual coding features of the learning problem image end and the semantic coding features of the learning problem text end, and performing cross-modal kernel function correlation on them to obtain a set of kernel correlation coding vectors between the principal components of the multimodal features of the learning problem; S42, performing principal component cross-domain deep fusion and global information integration based on graph convolutional coding on the set of kernel correlation coding vectors between the principal components of the multimodal features of the learning problem to obtain a joint representation of the multimodal features of the learning problem.

[0028] Specifically, S41 extracts the characteristic principal components of the visual encoding features of the learning problem image and the semantic encoding features of the learning problem text, and performs cross-modal kernel function correlation on them to obtain a set of kernel correlation encoding vectors between the principal components of the multimodal features of the learning problem. That is, in an embodiment of the present application, first, a characteristic principal component analysis is performed on the visual feature encoding vectors of the learning problem image and the semantic feature encoding vectors of the learning problem text to obtain a set of visual feature principal component encoding vectors of the learning problem image and a set of semantic feature principal component encoding vectors of the learning problem text. It should be understood that since visual features and text features may carry high-dimensional noise or implicit linear correlation during the extraction process, direct cross-modal correlation analysis is easily interfered with by irrelevant features, making it difficult for the subsequent multimodal fusion module to effectively capture the essential feature correlations that truly affect the judgment of sensitive content. This redundancy comes not only from the dimensional expansion within a single modality feature (such as the high number of channels of image convolution features), but also from the potential overlap of cross-modal features in semantic expression (such as the repeated expression of text descriptions and image content under a specific theme). Therefore, in the technical solution of the present application, feature principal component analysis is performed on the visual feature coding vector of the learning problem image side and the semantic feature coding vector of the learning problem text side to obtain a set of visual feature principal component coding vectors of the learning problem image side and a set of semantic feature principal component coding vectors of the learning problem text side.

[0029] That is, through orthogonal feature decoupling and reconstruction, the original high-dimensional features are projected into the low-dimensional principal component space, thereby stripping away redundant information and highlighting the deep semantic association structure between cross-modal features. In this process, principal component analysis extracts orthogonal basis vectors through covariance matrix decomposition, suppressing secondary variation factors while retaining the direction of maximum variance, so that the principal component sets of visual and text features represent the most discriminative independent semantic dimensions in their respective modalities. The dimensionality compression and feature decoupling achieved through principal component analysis enable the graph convolutional neural network to focus on the substantial associations between cross-modal features rather than surface noise in subsequent processing, thereby more efficiently identifying sensitive content feature patterns implicit in text-image combinations.

[0030] In a specific example of the present application, the following principal component analysis formula is used to perform feature principal component analysis on the visual feature encoding vector of the learning problem image side and the semantic feature encoding vector of the learning problem text side to obtain a set of visual feature principal component encoding vectors of the learning problem image side and a set of semantic feature principal component encoding vectors of the learning problem text side; wherein the principal component analysis formula is:

[0031]

[0032] in, is the first covariance matrix, is the visual feature encoding vector of the image side of the learning problem, is the length of the visual feature encoding vector at the image end of the learning problem, is the principal component analysis of the features, is the set of principal component encoding vectors of the visual features of the image side of the learning problem, are the first, second, and third principal component encoding vectors of the learning problem image side visual feature. and A learning problem image-side visual feature principal component encoding vector, is the multimodal feature diagonal matrix of the learning problem, and They are and The corresponding eigenvalues;

[0033]

[0034] in, is the second covariance matrix, Encode the semantic feature vector of the learning problem text side, is the principal component analysis of the features, is the multimodal feature diagonal matrix of the learning problem, and They are and The corresponding eigenvalues, is a set of principal component encoding vectors of semantic features of the learning problem text side, are the first, second, and third principal component encoding vectors of the semantic features of the learning problem text end, respectively. and A principal component encoding vector of semantic features of the learning problem text side.

[0035] Furthermore, each corresponding set of principal component encoding vectors for the learning problem image-side visual features and the learning problem text-side semantic features is fed into a cross-modal component kernelization association architecture to obtain a set of kernel association encoding vectors between principal components of the learning problem's multimodal features. It should be understood that while principal component analysis strips away redundant information and constructs a low-dimensional orthogonal feature space, the semantic associations between cross-modal features often have deep nonlinear dependencies. For example, when a user uploads an image of a math problem that contains irrelevant images and the accompanying text uses metaphorical vocabulary to circumvent conventional filtering, simple principal component dimensionality reduction cannot directly reveal the deep synergistic anomalies between the image principal components (such as unusual edge textures) and the text principal components (such as ambiguous word vectors). Traditional linear association models find it difficult to capture the complex interactions of such cross-modal features in low-dimensional space. Therefore, in order to construct a nonlinear association bridge between cross-modal principal components and enable the image and text principal components to form a computable association code in the high-dimensional feature space, in the technical solution of the present application, each corresponding set of the learning problem image-side visual feature principal component coding vectors and the learning problem text-side semantic feature principal component coding vectors in the set of learning problem image-side semantic feature principal component coding vectors are respectively input into the cross-modal component kernelized association architecture to obtain a set of kernel association coding vectors between the multimodal feature principal components of the learning problem.

[0036] Here, the principal component kernel association coding network leverages the nonlinear mapping capabilities of deep neural networks to project orthogonalized principal component pairs into a high-dimensional abstract space, breaking through the limitations of linear separability and thereby exploring potential sensitive semantic association patterns between cross-modal principal components. Through the forward propagation of the neural network, the kernel association coding network not only learns explicit correlations between principal component pairs (such as the correspondence between color distributions in image principal components and color descriptive vocabulary in text principal components), but also captures implicit coupling relationships (such as the unusual linkage between the semantic deviations of geometric shape principal components and homophonic words in the text). This encoding process transforms the previously isolated image and text principal components into vector representations that carry cross-modal association information, providing cross-modal semantic units for the subsequent graph convolutional network to construct a global interaction topology. This enables sensitive content detection based on fine-grained collaborative patterns between multimodal features rather than single-modal features. This mechanism effectively addresses the problem of traditional methods underdetecting sensitive content that involves collaborative image and text. Furthermore, the vectorized representation of association coding avoids the subjectivity and lag of manually defined cross-modal rules, improving the system's adaptive detection capabilities for new types of sensitive content.

[0037] In a specific example of the present application, each corresponding set of principal component coding vectors of visual features on the image side of the learning problem and the principal component coding vectors of semantic features on the text side of the learning problem are input into the cross-modal component kernelized association architecture to obtain a set of kernel association coding vectors between principal components of multimodal features of the learning problem using the following association coding formula; wherein, the association coding formula is:

[0038]

[0039] in, express activation function, and are the learnable parameter weight matrix and the learnable parameter bias vector, The first kernel correlation encoding vector in the set of multimodal feature principal components of the learning problem is A learning problem multimodal feature principal component kernel correlation encoding vector, and For each group of corresponding learning problems, the principal component encoding vector of the visual features on the image side and the principal component encoding vector of the semantic features on the text side are included.

[0040] Specifically, S42 performs graph convolutional coding-based principal component cross-domain deep fusion and global information integration on the set of kernel correlation coding vectors between principal components of multimodal features of learning problems to obtain a joint representation of multimodal features of learning problems. That is, in an embodiment of the present application, first, the learning problem multimodal synergy efficiency factor is calculated between any two kernel correlation coding vectors between principal components of multimodal features of learning problems in the set of kernel correlation coding vectors between principal components of multimodal features of learning problems to obtain a learning problem multimodal synergy efficiency factor correlation topology matrix. It should be understood that although the feature principal component vectors after kernel correlation coding already carry the nonlinear correlation information of the image and text principal components, there are still potential synergistic or antagonistic relationships between these correlation units that require further structured expression. For example, when a user uploads a physics problem image containing an illegal experimental device, and the accompanying text uses professional terminology to describe dangerous operating steps, kernel correlation coding may generate multiple cross-modal correlation vectors (such as the correlation between the principal component of the device shape and the principal component of the verb phrase, and the correlation between the principal component of the color and the principal component of the degree adverb). However, the contribution of different correlation vectors to the judgment of sensitive content varies significantly. Traditional methods that homogenize all correlation vectors can easily dilute key sensitive signals with non-core correlation noise. Therefore, to establish a dynamic importance assessment system for cross-modal correlation vectors, the technical solution of this application calculates the multimodal synergistic effectiveness factor of the learning problem between any two kernel correlation coding vectors between the principal components of the multimodal features of the learning problem in the set of kernel correlation coding vectors to obtain the multimodal synergistic effectiveness factor correlation topology matrix of the learning problem.

[0041] Here, the synergy factor, a customizable correlation strength metric, can flexibly define the synergistic interaction mechanism between correlation vectors based on the specific patterns of sensitive content in educational scenarios (e.g., illegal information often accompanies the unusual coupling of text and image semantics, and off-topic content often exhibits loose associations between text and image). By converting the synergy factor value into edge weights in a topological matrix, the system can quantify the contribution of different cross-modal associations to the sensitivity of the current learning problem, providing a structured prior for subsequent graph convolutional networks. This dynamic weight allocation mechanism enables the graph convolutional network to focus on high-threat correlation subgraphs and extract the collaborative propagation patterns of cross-modal sensitive content through neighborhood aggregation, significantly improving the ability to capture hidden illegal content (e.g., suggestive information scattered across multiple correlation vectors) while reducing the risk of misjudgment caused by excessive focus on weak associations.

[0042] In a specific example of the present application, the following formula is used to calculate the learning problem multimodal synergistic effectiveness factor between any two learning problem multimodal feature principal component kernel correlation coding vectors in the set of learning problem multimodal feature principal component kernel correlation coding vectors to obtain the learning problem multimodal synergistic effectiveness factor correlation topology matrix; wherein, the formula is:

[0043]

[0044]

[0045] in, and Respectively and A learning problem multimodal feature principal component kernel correlation encoding vector, for and The multimodal collaborative effectiveness factor of the learning problem between is the multimodal synergistic effectiveness factor correlation topology matrix of the learning problem, is the scale parameter.

[0046] Subsequently, the set of kernel correlation encoding vectors between principal components of the multimodal features of the learning problem and the correlation topology matrix of the multimodal performance operator of the learning problem are input into the graph convolutional neural network model to obtain the multimodal significant joint feature matrix of the learning problem as the joint representation of the multimodal features of the learning problem. Considering that traditional methods often have difficulty in effectively capturing the deep correlation of cross-modal features when processing the fusion features of images and text, especially the potential semantic and spatial connections between feature components of different modalities. Since user input in learning scenarios may contain mixed information of both images and text, independent analysis of a single modality tends to ignore the complementarity between the two. For example, sensitive elements in an image may be legitimized by text descriptions, while implicit issues in the text may be masked by image content. Therefore, in order to achieve global correlation modeling of multimodal features, thereby improving the comprehensiveness and accuracy of sensitive content detection, in the technical solution of this application, the set of kernel correlation encoding vectors between principal components of the multimodal features of the learning problem and the correlation topology matrix of the multimodal performance operator of the learning problem are input into the graph convolutional neural network model to obtain the multimodal significant joint feature matrix of the learning problem.

[0047] Through graph convolution operations, the model can perform multi-level information transmission and aggregation in the graph structure, so that the representation of each principal component feature node not only contains its own information, but also integrates the associated features of adjacent nodes. This mechanism can break through the limitations of traditional linear or shallow fusion methods. For example, local sensitive areas in image features may have nonlinear correlations with keywords in text descriptions, and graph convolution can capture such cross-modal fine-grained matching patterns through multi-hop propagation. At the same time, the introduction of the performance operator correlation topology matrix further strengthens the quantitative expression of the interaction strength between features, allowing the model to distinguish the influence weights of different modal feature components on the final detection results, avoiding misjudgments due to feature redundancy or noise interference. This not only improves the automation and coverage of the filtering system, but also reduces operating costs by reducing manual intervention, while ensuring the consistency and real-time nature of audit standards, providing technical support for building a safe and reliable large-scale model learning environment.

[0048] In a specific example of the present application, the set of kernel correlation encoding vectors between principal components of the multimodal features of the learning problem and the multimodal performance operator correlation topology matrix of the learning problem are input into the graph convolutional neural network model using the following convolution formula to obtain the multimodal significant joint feature matrix of the learning problem; wherein, the graph convolution formula is:

[0049]

[0050] in, represents the graph convolutional neural network model, The multimodal salient joint feature matrix for the learning problem.

[0051] In particular, S5 determines whether the large model answers the user-entered learning question based on the joint representation of the multimodal features of the learning question. Specifically, in an embodiment of the present application, the multimodal salient joint feature matrix of the learning question is first passed through a classifier-based sensitive content detection model to obtain a detection result. The detection result is used to indicate whether the learning question is sensitive content. It should be understood that the multimodal salient joint feature matrix of the learning question not only contains independent features of the text and image, but also encodes the interactive relationship between the two, such as the correlation between sensitive symbols in the image and suggestive words in the text description. Traditional rule-based filtering methods have difficulty handling such complex patterns of cross-modal coupling, while classifiers, through supervised learning, can adaptively extract key discriminative features, such as identifying the synergy between local texture anomalies in an image and semantic deviations from the educational theme of the text. The classifier learns the mapping relationship between features and sensitive content categories through training data, constructing a decision boundary from the high-dimensional feature space to the binary classification result (sensitive / non-sensitive). This approach not only strengthens the discriminative ability of the fused features, but also improves the intelligence and reliability of the entire content filtering process.

[0052] Finally, in response to the learning question being sensitive content, the big model refuses to answer the learning question, and in response to the learning question not being sensitive content, the big model answers the learning question. Considering that sensitive content often has a negative impact on the teaching quality and platform atmosphere, and may even lead to legal or ethical risks, it must be intercepted in a timely manner through automated means. That is, in an embodiment of the present application, when the detection results confirm that the question is sensitive content, the big model actively refuses to answer, effectively preventing the dissemination of inappropriate information; and for non-sensitive content, it ensures that users can smoothly obtain the required knowledge support. This targeted response based on sensitive content identification results realizes automated and intelligent content management, helps maintain the security and credibility of the education platform, improves user experience, and promotes the healthy and sustainable development of the platform.

[0053] In summary, the large-model sensitive large-content filtering method in the learning scenario according to the embodiment of the present application is explained, which extracts the image-side visual coding features and text-side semantic coding features of the learning problems from the learning problems input by the user by adopting a neural network model based on deep learning, and further fuses the two for multimodal features to form a more comprehensive and accurate joint representation of the problem. This joint representation can fully explore the deep connection between text and pictures, thereby effectively identifying sensitive or inappropriate content in complex inputs, and realizing automated and intelligent filtering of complex learning problems. In this way, the problems of narrow coverage, low detection accuracy and low efficiency of manual review of traditional single-modality filtering solutions are solved, and the platform's ability to identify sensitive content is improved, thereby providing a solid and reliable security guarantee for the intelligent education platform.

[0054] While various embodiments of the present disclosure have been described above, the above descriptions are illustrative, non-exhaustive, and not intended to be limiting of the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is selected to best explain the principles of the embodiments, their practical applications, or improvements to existing technologies, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A large model-sensitive large content filtering method in a learning scenario, characterized by: include: Obtaining a learning question input by a user, where the learning question includes a text description and image data; Extract visual features from image data to obtain visual encoding features of the learning problem image side; Extract semantic features from text description to obtain semantic encoding features of the learning problem text; Performing multimodal feature fusion of the learning problem on the visual coding features of the learning problem image and the semantic coding features of the learning problem text to obtain a joint representation of the multimodal features of the learning problem, including: extracting the feature principal components of the visual coding features of the learning problem image and the semantic coding features of the learning problem text, and performing cross-modal kernel function correlation on them to obtain a set of kernel correlation coding vectors between principal components of the multimodal features of the learning problem; performing principal component cross-domain deep fusion and global information integration based on graph convolutional coding on the set of kernel correlation coding vectors between principal components of the multimodal features of the learning problem to obtain a joint representation of the multimodal features of the learning problem; Based on the joint representation of multimodal features of the learning question, it is determined whether the large model answers the learning question input by the user.

2. The large model-sensitive large content filtering method in a learning scenario according to claim 1 is characterized in that: Extract visual features from image data to obtain the visual encoding features of the learning problem image, including: A visual feature extractor based on a deep neural network model is used to extract visual features from image data to obtain a visual feature encoding vector of the learning problem image side as the visual encoding feature of the learning problem image side.

3. The large model-sensitive large content filtering method in a learning scenario according to claim 2 is characterized in that: The deep neural network model is a dilated convolutional neural network model.

4. The large model-sensitive large content filtering method in a learning scenario according to claim 3 is characterized in that: Extract semantic features from text descriptions to obtain semantic encoding features for learning problem texts, including: A semantic feature extractor based on the Bert model is used to extract semantic features from the text description to obtain a semantic feature encoding vector of the learning question text as the semantic encoding feature of the learning question text.

5. The large model-sensitive large content filtering method in a learning scenario according to claim 4 is characterized in that: Extracting the principal components of the visual encoding features of the learning problem image and the semantic encoding features of the learning problem text, and performing cross-modal kernel function correlation on them to obtain a set of kernel correlation coding vectors between principal components of the multimodal features of the learning problem, including: Performing feature principal component analysis on the visual feature encoding vector of the learning question image and the semantic feature encoding vector of the learning question text to obtain a set of visual feature principal component encoding vectors of the learning question image and a set of semantic feature principal component encoding vectors of the learning question text; Each corresponding set of principal component coding vectors of visual features on the image side of the learning problem and the principal component coding vectors of semantic features on the text side of the learning problem are input into the cross-modal component kernelized association architecture to obtain a set of kernel association coding vectors between principal components of multimodal features of the learning problem.

6. The large model-sensitive large content filtering method in a learning scenario according to claim 5 is characterized in that: The kernel correlation coding vectors between principal components of the multimodal features of the learning problem are subjected to deep cross-domain principal component fusion and global information integration based on graph convolutional coding to obtain a joint representation of the multimodal features of the learning problem, including: Calculating the learning problem multimodal synergistic effectiveness factor between any two learning problem multimodal feature principal component kernel correlation coding vectors in the set of learning problem multimodal feature principal component kernel correlation coding vectors to obtain the learning problem multimodal synergistic effectiveness factor correlation topology matrix; performing topological regularization of the multimodal kernelized feature space layout of the learning problem on each of the kernel correlation coding vectors between principal components of the learning problem multimodal features in the set of kernel correlation coding vectors between principal components of the learning problem multimodal features to obtain a set of optimized kernel correlation coding vectors between principal components of the learning problem multimodal features; The set of kernel correlation encoding vectors between principal components of the optimized learning problem multimodal features and the correlation topology matrix of the learning problem multimodal synergistic effectiveness factors are input into the graph convolutional neural network model to obtain the learning problem multimodal significant joint feature matrix as the joint representation of the learning problem multimodal features.

7. The large model-sensitive large content filtering method in a learning scenario according to claim 6 is characterized in that: Based on the joint representation of multimodal features of the learning problem, determine whether the large model answers the input learning problem, including: Determine whether a learning question is sensitive content based on the joint representation of its multimodal features; In response to the learning question belonging to sensitive content, the large model refuses to answer the learning question, and in response to the learning question not belonging to sensitive content, the large model answers the learning question.

8. The large model-sensitive large content filtering method in a learning scenario according to claim 7 is characterized in that: Based on the joint representation of multimodal features of the learning question, determine whether the learning question is sensitive content, including: The multimodal significant joint feature matrix of the learning problem is passed through a classifier-based sensitive content detection model to obtain a detection result. The detection result is used to indicate whether the learning problem belongs to sensitive content.

Images