Content auditing method and device based on AI knowledge base, electronic equipment and computer readable storage medium

Abstract

The method comprises the steps of obtaining a to-be-uploaded file and a corresponding file type, preprocessing the to-be-uploaded file, generating a to-be-audited file, calling a corresponding auditing model in an AI knowledge base for auditing according to the type of the to-be-audited file, and generating a corresponding auditing result; the multi-hop query request of the user is obtained according to the auditing result, a modification suggestion is generated, and finally the modification suggestion is returned to the user and is encrypted and stored, and more intelligent and accurate auditing is achieved by means of the powerful knowledge representation and reasoning ability of the knowledge base. The modification suggestions provided when the auditing is not passed can help the user to quickly understand the problem and perform targeted modification, so that the efficiency of the content creator is improved, the times of repeated submission and auditing are reduced, the auditing cost is reduced, and the quality and efficiency of the whole auditing process are improved.

Description

Technical Field

[0001] The present invention relates to the field of AI intelligent review, and particularly to a content review method, device, electronic device, and computer-readable storage medium based on an AI knowledge base. Background Art

[0002] In the era of information explosion, a large amount of content such as various texts, images, and videos is generated. Reviewing this content to ensure its compliance with regulations and quality standards is an important task for many platforms and institutions. Traditional review methods often rely on manual labor, with low efficiency and strong subjectivity. With the development of artificial intelligence technology, review methods based on rules and machine learning have gradually been applied. However, when faced with complex knowledge associations and diverse content, these methods are difficult to conduct comprehensive and accurate reviews, and when the review fails, they are unable to provide targeted modification suggestions for users. As a semantic network, the knowledge base can integrate and represent a large amount of knowledge, providing a new idea for solving the above problems. Summary of the Invention

[0003] To improve the above problems, the present invention proposes a content review method based on an AI knowledge base, specifically including the following steps:

[0004] Step 1: Obtain the file to be uploaded and the corresponding file type, preprocess the file to be uploaded, and generate a file to be reviewed;

[0005] Step 2: According to the type of the file to be reviewed, retrieve the corresponding review model in the AI knowledge base for review, and generate a corresponding review result;

[0006] Step 3: Obtain the multi-hop query request of the user according to the review result, and generate a modification suggestion;

[0007] Obtain the multi-hop query request, and the intelligent interaction model of the AI knowledge base converts the multi-hop query request into the optimal association path starting from the summary content entity node in the AI knowledge base, and generates a modification suggestion;

[0008] Specifically, the intelligent interaction model includes a collaborative representation learning module, a modification suggestion set sorting module, and an association path learning module; the collaborative representation learning module is divided into an AI knowledge base learning representation module and a multi-hop query request learning representation module;

[0009] Among them, the AI knowledge base learning representation module represents the entities and relationships of the industry specification information in the AI knowledge base as vectors, that is Among them, h represents the head entity vector in the industry specification semantic triple, t represents the tail entity vector in the industry specification semantic triple, and r represents the relationship between the head entity vector and the tail entity vector in the industry specification semantic triple. The relationship between the head entity vector h and the tail entity vector t in the industry specification semantic triple is interacted through a tensor by the NTN scoring function, and the semantic relationship between the head entity vector and the tail entity vector is obtained. represents a weight matrix related to the relationship r. represents a relationship matrix related to the relationship r. This relationship matrix is used to concatenate the head entity vector h and the tail entity vector t and then perform a linear transformation. [h;t] represents a new vector formed by concatenating the head entity vector h and the tail entity vector t in a preset order. b r represents the bias vector related to the relationship r, and g(*) represents a non-linear activation function.

[0010] The multi-hop query request learning representation module represents the multi-hop query request. The representation vector of the multi-hop query request is obtained through the average semantic vector of the last hidden layer in the bidirectional long short-term memory network model. Then, the representation vector of the multi-hop query request is converted through the fully connected layer in the bidirectional long short-term memory network to obtain a new representation vector of the multi-hop query request, and the new representation vector of the multi-hop query request is mapped to a vector space with the same dimension as the AI knowledge base learning representation module to obtain a set of modification suggestions.

[0011] Step 4: Return the modification suggestions represented by the modification suggestion entity nodes to the user, and store the modification suggestions in encrypted form.

[0012] Specifically, in Step 2, when the file to be reviewed is a document, the document review model based on natural language processing in the AI knowledge base is used for review, and the corresponding document review result is generated.

[0013] When the file to be reviewed is an image, feature extraction is performed on the preprocessed image, and it is input into the image review model in the AI knowledge base for review. The target detection algorithm is used to detect the key targets in the image, and by comparing with the illegal image feature library, the image containing illegal content is identified, and the corresponding image review result is generated.

[0014] When the file to be reviewed is a video, the preprocessed video file is split into image frames, key frames are extracted, and the extraction is performed through the image review model in the AI knowledge base. In addition, the audio features of the audio part of the video are extracted, the audio is converted into text using speech recognition technology, and then text review is performed through the document review model based on natural language processing in the AI knowledge base to detect whether there is illegal audio content in the video, and finally the corresponding video review result is generated.

[0015] Specifically, in step 3, during the calculation of the scoring function NTN model, it is necessary to construct a loss function L by constructing the relationship between the positive triple information and the negative triple information, that is

[0016]

[0017] where f + (h, r, t) represents training the positive triple through the NTN scoring function, f + (h, r, t) represents training the positive triple through the NTN scoring function, f - (h, r, t) represents training the negative triple through the NTN scoring function, and exp(*) represents the exponential function.

[0018] Specifically, in step 3, the average semantic vector of the last hidden layer in the bidirectional long short-term memory network model is used as the representation vector e of the multi-hop query request q q , that is

[0019] e q = Mean(Bi lstm([CLS], q1, q2,... q n , [SEP]))

[0020] where the function Mean(*) represents calculating the average semantic vector of the multi-hop query request q; then, the full connection layer in the bidirectional long short-term memory network is used to convert e q to obtain the representation vector e of the new multi-hop query request q′ , that is

[0021]

[0022] where represents a weight matrix related to the relationship r in the triple (h, r, t) corresponding to the semantic information of the industry norms in the AI knowledge base, represents a relationship matrix related to the relationship r in the triple (h, r, t) corresponding to the semantic information of the industry norms in the AI knowledge base, b rDenote the bias vector related to the relationship r, and g(*) represents a non-linear activation function.

[0023] It can be understood that before performing the above steps 1-4, the AI knowledge base needs to be pre-trained.

[0024] In addition, the present invention also proposes a content review device based on an AI knowledge base. The device includes a document to be reviewed acquisition module, a document review module, a review document modification suggestion module, and a review document storage module, where:

[0025] The document to be reviewed acquisition module is used to acquire the file to be uploaded and the corresponding file type, preprocess the file to be uploaded, and generate a file to be reviewed;

[0026] The document review module is used to retrieve the corresponding review model in the AI knowledge base according to the type of the file to be reviewed, and generate a corresponding review result;

[0027] The review document modification suggestion module is used to obtain the multi-hop query request of the user according to the review result, and generate a modification suggestion;

[0028] Obtain the multi-hop query request, and the intelligent interaction model of the AI knowledge base converts the multi-hop query request into the optimal association path starting from the abstract content entity node in the AI knowledge base, and generates a modification suggestion;

[0029] Specifically, the intelligent interaction model includes a collaborative representation learning module, a modification suggestion set sorting module, and an association path learning module; the collaborative representation learning module is divided into an AI knowledge base learning representation module and a multi-hop query request learning representation module;

[0030] Among them, the AI knowledge base learning representation module represents the entities and relationships of the industry specification information in the AI knowledge base, that is Among them, h represents the head entity vector in the industry specification semantic triple, t represents the tail entity vector in the industry specification semantic triple, r represents the relationship between the head entity vector and the tail entity vector in the industry specification semantic triple, and the relationship between the head entity vector h and the tail entity vector t in the industry specification semantic triple is interacted through a tensor by the NTN scoring function, and the semantic relationship between the head entity vector and the tail entity vector is obtained, Denote a weight matrix related to the relationship r, Denote a relationship matrix related to the relationship r, which is used to splice the head entity vector h and the tail entity vector t and then perform a linear transformation. [h;t] represents a new vector formed by splicing the head entity vector h and the tail entity vector t in a preset order, b rDenote the bias vector related to the relationship r, and g(*) represents a non-linear activation function;

[0031] The multi-hop query request learning representation module represents the multi-hop query request. The representation vector of the multi-hop query request is obtained through the average semantic vector of the last hidden layer in the bidirectional long short-term memory network model. Then, the representation vector of the multi-hop query request is transformed through the fully connected layer in the bidirectional long short-term memory network to obtain a new representation vector of the multi-hop query request, and the new representation vector of the multi-hop query request is mapped to a vector space with the same dimension as the AI knowledge base learning representation module to obtain a set of modification suggestions;

[0032] The audit document storage module is used to return the modification suggestions represented by the modification suggestion entity nodes to the user and store the modification suggestions in encrypted form.

[0033] Specifically, in the document audit module, when the file to be audited is a document, the document audit model based on natural language processing in the AI knowledge base is used for auditing, and the corresponding document audit result is generated.

[0034] When the file to be audited is an image, feature extraction is performed on the preprocessed image, and it is input into the image audit model in the AI knowledge base for auditing. The target detection algorithm is used to detect the key targets in the image, and by comparing with the illegal image feature library, the image containing illegal content is identified, and the corresponding image audit result is generated.

[0035] When the file to be audited is a video, the preprocessed video file is split into image frames, key frames are extracted, and they are audited through the image audit model in the AI knowledge base. In addition, the audio features of the audio part of the video are extracted, the audio is converted into text using speech recognition technology, and then text audit is performed through the document audit model based on natural language processing in the AI knowledge base to detect whether there is illegal audio content in the video, and finally the corresponding video audit result is generated.

[0036] Specifically, in the audit document modification suggestion module, in the process of calculating the scoring function NTN model, it is necessary to construct a loss function L by constructing the relationship between the positive triple information and the negative triple information, that is

[0037]

[0038] where f + (h,r,t) represents training the positive triple through the NTN scoring function, f + (h,r,t) represents training the positive triple through the NTN scoring function, f -(h, r, t) represents the training of negative triples through the NTN scoring function, and exp(*) represents the exponential function.

[0039] Specifically, in the review document modification suggestion module, the average semantic vector of the last hidden layer in the bidirectional long short-term memory network model is used as the representation vector e of the multi-hop query request q q , that is

[0040] e q = Mean(Bi lstm([CLS], q1, q2,... q n , [SEP]))

[0041] Among them, the function Mean(*) represents calculating the average semantic vector of the multi-hop query request q; then, e is transformed through the fully connected layer in the bidirectional long short-term memory network q to obtain the representation vector e of the new multi-hop query request q′ , that is

[0042]

[0043] Among them represents a weight matrix related to the relationship r in the triple (h, r, t) corresponding to the semantic information of the industry norms in the AI knowledge base, represents a relationship matrix related to the relationship r in the triple (h, r, t) corresponding to the semantic information of the industry norms in the AI knowledge base, b r represents the bias vector related to the relationship r, and g(*) represents a non-linear activation function.

[0044] It can be understood that the above-mentioned AI knowledge base-based content review device further includes a corresponding training module for pre-training the AI knowledge base.

[0045] Among them, the present invention also provides an electronic device, which includes a processor and a memory: the memory is used to store program code and transmit the program code to the processor, and the processor is used to execute the above-mentioned AI knowledge base-based content review method according to the instructions in the program code.

[0046] Moreover, the present invention also provides a computer-readable storage medium, which is used to store program code and transmit the program code to the processor, and the processor is used to execute the above-mentioned AI knowledge base-based content review method according to the instructions in the program code.

[0047] The present invention obtains the file to be uploaded and the corresponding file type, preprocesses the file to be uploaded to generate a file to be reviewed, retrieves the corresponding review model in the AI knowledge base according to the type of the file to be reviewed, and generates the corresponding review result. When the review is abnormal, a multi-hop query request of the user is obtained according to the review result, and a modification suggestion is generated. Finally, the modification suggestion is returned to the user and encrypted for storage.

[0048] The present invention utilizes the powerful knowledge representation and reasoning capabilities of the knowledge base to achieve more intelligent and accurate review. The modification suggestions provided when the review fails can help users quickly understand the problems and make targeted modifications, improving the efficiency of content creators, reducing the number of repeated submissions for review, lowering the review cost, and enhancing the quality and efficiency of the entire review process. Brief Description of the Drawings

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0050] Figure 1 is a schematic flowchart of a content review method based on an AI knowledge base;

[0051] Figure 2 is a schematic diagram of the modules of a content review device based on an AI knowledge base. Detailed Embodiments

[0052] As Figure 1 shown, a content review method based on an AI knowledge base has the following specific steps:

[0053] Step 1: Obtain the file to be uploaded and the corresponding file type, and preprocess the file to be uploaded to generate a file to be reviewed;

[0054] It can be understood that when the system administrator faces a large number of files to be uploaded, it is necessary to determine whether the content of the file to be uploaded does not conform to the corresponding industry rules. Therefore, the system administrator needs to review the file to be uploaded according to relevant industry regulations, policy documents, industry technical standards, and operation specifications. And different review methods need to be retrieved according to different types of files to be uploaded. Further, the file type of the file to be uploaded can be obtained. It can be understood that the files to be uploaded include types such as documents, tables, images, videos, and audios.

[0055] It is understandable that when the file to be uploaded is a document, data cleaning of the document to be uploaded is required, such as HTML tags, special characters, garbled codes, etc., and stop words are deleted. When the file to be uploaded is an image, the size ratio and resolution of the image to be uploaded need to be adjusted, followed by grayscale conversion, image smoothing operation through mean filtering, and finally edge detection through the Canny operator. When the file to be uploaded is a video, color correction is required, the video resolution and frame rate are adjusted, and the video frames are smoothed through Gaussian denoising. In addition, those skilled in the art also know how to perform preprocessing operations on other files such as tables and audio.

[0056] Step 2: According to the type of the file to be reviewed, retrieve the corresponding review model in the AI knowledge base for review and generate the corresponding review result.

[0057] Exemplarily, when the file to be reviewed is a document, the document review model based on natural language processing in the AI knowledge base is used for review. The document review model uses the BERT model in the deep learning model to perform semantic understanding and feature extraction on the document. By matching with the preset sensitive word library and illegal semantic pattern library, it is judged whether there is illegal content in the text, such as sensitive words, illegal remarks, false information, etc., and the corresponding document review result is generated.

[0058] Exemplarily, when the file to be reviewed is an image, feature extraction is performed on the preprocessed image and input into the image review model in the AI knowledge base for review. The image review model has pre-trained various image features, such as people, scenes, objects, etc. At the same time, object detection algorithms, such as the YOLO algorithm, are used to detect key objects in the image. By comparing with the illegal image feature library, images containing illegal content are identified and the corresponding image review result is generated.

[0059] Exemplarily, when the file to be reviewed is a video, the preprocessed video file is split into a series of image frames, and key frames are extracted. The key frame extraction is based on shot change detection and motion analysis techniques to ensure that important content is not missed. Then, the key frames are reviewed using the same process as the image review, that is, through the image review model in the AI knowledge base. In addition, the audio part of the video is also processed, audio features are extracted, the audio is converted into text using speech recognition technology, and then text review is performed through the document review model based on natural language processing in the AI knowledge base to detect whether there is illegal audio content in the video, and finally the corresponding video review result is generated.

[0060] It is understandable that the AI knowledge base needs to be pre-trained before the review. Specifically, corresponding data is selected from the system for training. Exemplarily, promotional material samples are collected from the promotional pages of e-commerce platforms, advertising channels, etc. The promotional material samples include various types of data. For pictures, image recognition tools are used to identify and label elements such as products, logos, and text in the pictures, and OCR technology is used to extract the text information in the pictures. For videos, key frames are extracted through video processing software, image recognition and text recognition are performed on the key frames, and at the same time, the audio information of the video is extracted and converted into text.

[0061] Exemplarily, when training the AI knowledge base, customer service voice conversations and text conversation data can also be exported from the system's customer service system. For voice conversations, speech recognition technology is used to convert them into text, and then the text is cleaned to remove invalid information such as duplicate content, typos, and garbled codes, and lexical and syntactic analysis is performed.

[0062] Moreover, the AI knowledge base can also collect materials such as company-internal documents, meeting records, and emails regarding business, policies and regulations, R & D, production, and rules and regulations, convert them into text format, and perform preprocessing to extract key information and question-and-answer pairs. The review system of this embodiment effectively controls the compliance of the output content and avoids common problems that may occur in many advertisements and violate the Advertising Law and the Food Safety Law.

[0063] In the feature extraction process of the AI knowledge base training, when performing feature extraction on text data, for the preprocessed text data, word vector models such as Word2Vec and GloVe are used to convert each word in the text into a vector representation, and then models such as convolutional neural networks and recurrent neural networks are used to process the word vectors to extract the semantic features of the text. When performing feature extraction on pictures, convolutional neural networks such as VGG and ResNet are used to perform feature extraction on the pictures to obtain the feature vectors of the pictures. When performing feature extraction on videos, combining the key frame image features and video time series information, 3D convolutional neural network models are used to perform feature extraction on the videos, and at the same time, the fusion of audio features and video image features is considered. The extracted text, picture, and video features are fused, for example, by means of splicing, weighted summation, etc., to form a unified feature representation as the input for subsequent model training.

[0064] During the process of training and optimizing the AI knowledge base, the model of the AI knowledge base is usually selected. Exemplarily, a model with the Transformer architecture is selected, such as BERT, GPT, etc. These models show good performance in natural language processing tasks and can effectively process the fused feature data. During the training process, the fused feature data is divided into a training set, a validation set, and a test set. The training set is used to train the model. During the training process, appropriate hyperparameters are set, such as the learning rate, the number of iterations, the batch size, etc. The cross-entropy loss function is used to calculate the difference between the model prediction result and the true label, and optimization algorithms such as stochastic gradient descent, Adagrad, Adadelta, etc. are used to update the model parameters, continuously reducing the value of the loss function and improving the accuracy of the model. During training, it is also necessary to evaluate and optimize the training effect of the AI knowledge base: the validation set is used to evaluate the model during the training process, and metrics such as the accuracy, recall rate, F1 value, etc. of the model are monitored. According to the evaluation results, the hyperparameters of the model are adjusted or the model structure is improved, such as increasing or decreasing the number of model layers, adjusting the attention mechanism, etc., to improve the performance of the model. Finally, the test set is used to test the optimized model to verify the generalization ability of the model.

[0065] Finally, the trained model is applied to build the AI knowledge base, and various types of business data and their corresponding feature representations and answers are stored in the database in a certain format, such as using a relational database or a non-relational database. At the same time, new system update data is regularly collected. Exemplarily, such as e-commerce business data, and the new data is processed according to the above data collection, preprocessing, feature extraction, and model training processes, and then the new data and the updated model parameters are applied to the update of the knowledge base to ensure that the knowledge base can timely reflect the latest business information and customer needs.

[0066] Step 3: Obtain the multi-hop query request of the user according to the review result and generate a modification suggestion.

[0067] Specifically, when the review result is not passed, the intelligent interaction model of the AI knowledge base is triggered. After the system administrator inputs the corresponding multi-hop query request through the interaction interface of the intelligent interaction model, the intelligent interaction model converts the multi-hop query request into a modification suggestion generated by the optimal association path starting from the abstract content entity node in the AI knowledge base.

[0068] Further, the intelligent interaction model includes a collaborative representation learning module, a modification suggestion set ranking module, and an associated path learning module. The collaborative representation learning module is divided into an AI knowledge base learning representation module and a multi-hop query request learning representation module. The AI knowledge base learning representation module represents the entities and relationships of the industry standard information in the AI knowledge base in vector form. It can be understood that the industry standard information includes relevant industry regulations, policy documents, industry technical standards, and operation specifications as described above. The multi-hop query request learning representation module represents the user's multi-hop query request in vector form and maps it to a vector space with the same dimension as the AI knowledge base learning representation module to obtain a modification suggestion set. Then, the modification suggestion set ranking module extracts all entity nodes under all associated paths starting from the abstract content entity node of the file to be reviewed from the AI knowledge base and filters out the entity nodes related to the multi-hop query request as modification suggestion entity nodes. The associated path learning module iteratively trains the ranked associated paths, and at the same time embeds the associated paths into a low-dimensional vector space for training and updating to obtain an updated intelligent interaction model.

[0069] Further, in the AI knowledge base learning representation module, it is necessary to embed the semantic information of the industry standards in the AI knowledge base into a low-dimensional space vector to enhance the accuracy of the modification suggestions generated by the intelligent interaction model. Specifically, the AI knowledge base learning representation module generates corresponding triples (h, r, t) according to the semantic information of the industry standards in the AI knowledge base and represents them through the scoring function in the Neural Tensor Network (NTN) model. Specifically, the relationship between the semantic triples of the industry standards in the AI knowledge base is represented by the NTN scoring function, that is

[0070]

[0071] Specifically, h represents the head entity vector in the semantic triple of the industry standard, t represents the tail entity vector in the semantic triple of the industry standard, and r represents the relationship between the head entity vector and the tail entity vector in the semantic triple of the industry standard. The relationship between the head entity vector h and the tail entity vector t in the semantic triple of the industry standard is interacted through a tensor by the NTN scoring function, and the semantic relationship between the head entity vector and the tail entity vector is obtained. represents a weight matrix related to the relationship r. represents a relationship matrix related to the relationship r. This relationship matrix is used to splice the head entity vector h and the tail entity vector t and then perform a linear transformation. [h;t] represents a new vector formed by splicing the head entity vector h and the tail entity vector t in a preset order, b rDenote the bias vector related to the relation r, and g(*) represents a non-linear activation function, specifically the hyperbolic tangent function.

[0072] Furthermore, during the calculation of the scoring function in the NTN model, it is necessary to construct a loss function L by constructing the relationship between the positive triple information and the negative triple information, that is

[0073]

[0074] where f + (h, r, t) represents training the positive triple through the NTN scoring function, and f + (h, r, t) represents training the positive triple through the NTN scoring function, and f - (h, r, t) represents training the negative triple through the NTN scoring function. exp(*) represents the exponential function. It can be understood that the positive triple refers to the modification suggestions having an associated path with the entity nodes in the abstract content of the document to be reviewed, and the negative triple refers to the modification suggestions of other associated paths, that is, irrelevant modification suggestions.

[0075] Furthermore, the multi-hop query request learning representation module also needs to perform vector representation on the multi-hop query request of the user and map it to a vector space with the same dimension as the corresponding triple (h, r, t) generated from the semantic information of the industry specification. Specifically, the multi-hop query request learning representation module obtains the semantic information hidden in the multi-hop query request and performs deep bidirectional semantic representation by combining the context information, that is, given the multi-hop query request q=(q1, q2,...q n ), where q1, q2,...q n represent the segmented query keywords, n represents a natural number. Set the entity node A in the abstract content of the document to be reviewed ∈ E, and the set of entity nodes of the modification suggestion S q ∈ E, where E is the set of entity nodes of the semantic information of the industry specification in the AI knowledge base. Set the head position marker [CLS] and the tail position marker [SEP] at the head and tail of the multi-hop query request q respectively. Input the multi-hop query request q with the inserted position markers into the bidirectional long short-term memory network model in the multi-hop query request learning representation module, and use the average semantic vector of the last hidden layer in the bidirectional long short-term memory network model as the representation vector e q of the multi-hop query request q, that is

[0076] e q =Mean(Bi lstm([CLS], q1, q2,...q n , [SEP]))

[0077] Among them, the function Mean(*) represents calculating the average semantic vector of the multi-hop query request q. Then, through the fully connected layer in the bidirectional long short-term memory network, e q is transformed to obtain the representation vector e q′ of the new multi-hop query request, that is

[0078]

[0079] where represents a weight matrix related to the relationship r in the triple (h, r, t) corresponding to the semantic information generation of the industry norms in the AI knowledge base, represents a relationship matrix related to the relationship r in the triple (h, r, t) corresponding to the semantic information generation of the industry norms in the AI knowledge base, b r represents the bias vector related to the relationship r, and g(*) represents a non-linear activation function, specifically the hyperbolic tangent function. By transforming the multi-hop query request, the representation vector of the new multi-hop query request is obtained, and the representation vector e q′ of the new multi-hop query request is mapped to a vector space with the same dimension as the AI knowledge base learning representation module to obtain a set of modification suggestions.

[0080] Furthermore, when iteratively training the multi-hop query request learning representation module, when the final obtained entity node belongs to the corresponding modification suggestion, f(e a , e q′ , e Sq ) > 0, and when the final obtained entity node does not belong to the corresponding modification suggestion, f(e a , e q′ , e Sq ) < 0, where e a represents the vector of the abstract content entity node of the file to be reviewed, e q′ represents the representation vector of the new multi-hop query request, and e Sq represents the representation vector of the modification suggestion entity node.

[0081] Furthermore, in the modification suggestion set sorting module, all entity nodes under all associated paths starting from the abstract content entity node of the file to be reviewed are extracted from the AI knowledge base, and during the process of screening entity nodes related to the multi-hop query request as modification suggestion entity nodes, the modification suggestion set sorting module also needs to calculate the output scores of all entity nodes under all associated paths and the multi-hop query request q.

[0082] Specifically, in the process of screening and modifying the proposed entity nodes, since there are multiple associated paths in the path from the abstract content entity node to the proposed modification entity node, it is necessary to calculate the optimal associated path from all the associated paths. In the process of calculating the optimal associated path, the proposed modification set sorting module uses the global attention mechanism and the encoding and decoding functions in the Transformer network to represent all the associated paths in vectors. Therefore, let P j (k) represent the associated path from the abstract content entity node to the proposed modification entity node. It can be understood that the associated path includes multiple jump nodes encountered in the multi-hop query request. Therefore, the associated path P j (k) is represented as:

[0083] P j (k) = [(a, p1, c a1 ), (c1, p2, c2),..., (c m-1 , p m , c m )]

[0084] Where P j (k) represents the k-th associated path from the abstract content entity node of the file to be reviewed to the j-th proposed modification entity node. a represents the abstract content entity node of the file to be reviewed, and p1, p2,..., p m represent the jump paths existing in the k-th associated path. c1, c2,..., c m-1 , c m represent m jump nodes existing in the k-th associated path.

[0085] Furthermore, through the encoding function in the Transformer network, the vectors of all the associated paths from the abstract content entity node to the j-th proposed modification entity node after encoding are obtained and the vector of the k-th associated path after encoding At the same time, through the encoding function in the Transformer network, the vectors of the jump nodes in all the associated paths from the abstract content entity node to the j-th proposed modification entity node after encoding are also obtained and the vectors of the m jump nodes existing in the k-th associated path after encoding And the vectors of all the associated paths after encoding are obtained and the vector of the k-th associated path after encoding and the vectors of the jump nodes in all the associated paths and the vectors of the m jump nodes existing in the k-th associated path after encoding And the first association information R1 between them is obtained, that is

[0086]

[0087] Among them, P j (k) represents the k-th associated path from the abstract content entity node to the j-th modified suggestion entity node, represents the vector of the jump nodes in all associated paths from the abstract content entity node to the j-th modified suggestion entity node obtained by encoding through the encoding function in the Transformer network, represents the encoded vector of the k-th associated path obtained by encoding through the encoding function in the Transformer network represents the vector of m jump nodes existing in the k-th associated path obtained by encoding through the encoding function in the Transformer network represents the vector of m jump nodes existing in the k-th associated path obtained by encoding through the encoding function in the Transformer network

[0088] Furthermore, obtain e in the collaborative representation learning module a e q′ and where e a represents the vector of the abstract content entity node of the file to be reviewed, and e q′ represents the representation vector of the new multi-hop query request, the representation vector of the j-th modified suggestion entity node, and splice the vectors of e a e q′ and and obtain the second associated information R2 through the multi-layer perceptron in the Transformer network, that is Then, obtain the output score higher than the preset threshold by passing the first associated information R1 and the second associated information R2 through the activation function sigmoid, that is Finally, obtain the sorted associated paths, and obtain the optimal associated path and the modified suggestion entity node corresponding to the optimal associated path. Furthermore, perform iterative training on the sorted associated paths through the associated path learning module.

[0089] Step 4: Return the modified suggestions represented by the modified suggestion entity nodes to the user, and encrypt and store the modified suggestions.

[0090] Specifically, the AI knowledge base uses the consensus mechanism of the blockchain to authenticate and encrypt the obtained modified suggestions, and records the authentication and storage time at the same time, so as to ensure the data quality of the modified suggestions in the intelligent interaction process, and can also avoid duplicate storage of the queried modified suggestions.

[0091] Exemplarily, the above-mentioned AI knowledge base can be applied to promotional materials generated by businesses such as e-commerce. In particular, legal and quality control reviews are carried out through the AI knowledge base. Specifically, the AI knowledge base also needs to collect promotional material samples from multiple channels such as the historical promotional materials of e-commerce platforms, legal case databases, and industry standard documents, mark the legal and quality control issues therein, such as false propaganda, trademark infringement, incorrect product parameters, etc., and construct a corresponding training data set. When the system administrator reviews the promotional text through the AI knowledge base, the AI knowledge base processes the pictures and text in the promotional material samples respectively, and retrieves the corresponding review models in the AI knowledge base for review respectively. Exemplarily, when the system administrator reviews the promotional material sample of skin care products, the AI model detects that the statement "completely remove freckles after using for one week" in the text does not conform to medical common sense and advertising law regulations, and determines it as false propaganda. At this time, the system administrator can further obtain the modification suggestions of the AI knowledge base through multi-hop queries. Exemplarily, the obtained modification suggestion is "After using for a period of time, it helps to improve the freckle problem". This way of placing the AI knowledge base in a quasi-objective and neutral position can reduce the fear of manual review of the system administrator of the e-commerce platform, thereby greatly reducing the review pressure of administrators such as quality control and legal affairs.

[0092] Moreover, the platform customer service staff can also conduct multi-round conversations through the AI knowledge base to understand the specific information parameters of the corresponding products in the e-commerce platform. Exemplarily, the customer service staff asks "Does the [model number] smartphone support wireless charging and what is the maximum charging power?" The customer service enters the question into the customer service Q&A system. The AI knowledge base will extract the key information "[model number] smartphone", "wireless charging", and "charging power", and retrieve them in the AI knowledge base. If there is corresponding information in the knowledge base, directly return the answer "[model number] smartphone supports wireless charging, and the maximum charging power is [X]W" to; if there is no direct answer, the AI knowledge base generates an answer through multi-hop queries from the product technical parameters to the charging-related information and returns it to the customer service staff.

[0093] Moreover, the AI knowledge base can also provide corresponding Q&A services for administration, human resources, finance, etc. within the company. Specifically, the AI knowledge base collects materials such as rules and regulations, process documents, and frequently asked questions (FAQs) from the administrative, human resources, and finance departments of the company to form an original data set. Natural language processing technology is used to clean, annotate, and classify the original data. For example, the process of requisitioning office supplies in the administrative category, the performance appraisal standards in the human resources category, and the invoice issuing norms in the finance category are respectively annotated and classified. At the same time, the AI knowledge base constructs structured information by analyzing the correlation relationships between different data, such as the connection between employee training in the human resources department and training expense reimbursement in the finance department, and regularly updates the data source. According to company policy changes, business adjustments, etc., the knowledge base is updated in real time or regularly to ensure the timeliness and accuracy of knowledge. Further, company employees can conduct corresponding intelligent Q&A based on the AI knowledge base. Employees input questions through the company's internal communication platform or a dedicated Q&A client. The system uses natural language understanding technology to parse the questions, extract key information, and determine the domain to which the questions belong. Specifically, it can determine the information belonging to specific departments in administration, human resources, or finance. According to the question domain and key information, retrieval and reasoning are performed in the AI knowledge base. If there is a directly matching answer in the knowledge base, it is directly returned to the employee; if there is no direct answer, relevant information is integrated through multi-hop queries to generate an answer. For example, when an employee asks "How to reimburse travel expenses after participating in external training", the system associates the training regulations in human resources and the travel expense reimbursement process in finance through multi-hop queries to generate a detailed answer. Feedback on the Q&A results is collected. Employees can evaluate the answers, and the system records the feedback information for optimizing the Q&A rules and algorithms of each model in the AI knowledge base.

[0094] It can be understood that the AI knowledge base can also be applied to more fields, such as intelligent assistants, which can assist enterprise leaders in real-time grasping data information on various business lines such as the company's R & D, products, sales, etc. It can also supervise the production progress from multiple dimensions based on the production scenarios of specific factories and enterprises, and can also determine whether the language skills of sales and customer service personnel are professional and accurate according to real-time conversations between sales and customer service.

[0095] Moreover, the AI knowledge base can also be applied in the financial field, where it can analyze a large amount of customer data and financial transaction data, more accurately evaluate the credit risk and market risk of customers, and provide a scientific basis for credit approval. When applied in the media field, the AI knowledge base can assist journalists in news writing, creating creative content, generating video scripts, etc., improving the efficiency and quality of content creation. It can provide personalized recommendations for news, videos, music, etc. based on data such as users' browsing history and interests, enhancing the user experience. It can also quickly review news, social media content, videos, etc., identify and filter out bad information, false information, etc., and maintain the health and safety of the network environment.

[0096] Embodiment 2

[0097] As Figure 2 shown, a content review device based on the AI knowledge base includes a to-be-reviewed document acquisition module, a document review module, a review document modification suggestion module, and a review document storage module, where:

[0098] The to-be-reviewed document acquisition module is used to obtain the file to be uploaded and its corresponding file type, preprocess the file to be uploaded, and generate a to-be-reviewed file;

[0099] It can be understood that when the system administrator faces a large number of files to be uploaded, it is necessary to determine whether the content of the file to be uploaded violates the corresponding industry rules. Therefore, the system administrator needs to review the file to be uploaded according to relevant industry regulations, policy documents, industry technical standards, and operation specifications. And different review methods need to be invoked according to different types of files to be uploaded. Further, obtaining the file type to be uploaded, it can be understood that the files to be uploaded include types such as documents, tables, images, videos, and audios.

[0100] It can be understood that when the file to be uploaded is a document, data cleaning needs to be performed on the document to be uploaded, such as HTML tags, special characters, garbled characters, etc., and stop words need to be removed. When the file to be uploaded is an image, the size ratio and resolution of the image to be uploaded need to be adjusted, then grayscale conversion is performed, image smoothing operation is carried out through mean filtering, and finally edge detection is performed through the Canny operator. When the file to be uploaded is a video, color correction is required, the video resolution and frame rate are adjusted, and the video frames are smoothed through Gaussian denoising. In addition, those skilled in the art also know how to perform preprocessing operations on other files such as tables and audios.

[0101] The document review module is used to retrieve the corresponding review model in the AI knowledge base for review according to the type of the to-be-reviewed file, and generate the corresponding review result.

[0102] Exemplarily, when the document to be reviewed is a document, the document review model based on natural language processing in the AI knowledge base is used for review. The document review model uses the BERT model in the deep learning model to perform semantic understanding and feature extraction on the document. By matching with the preset sensitive word library and illegal semantic pattern library, it is judged whether there is illegal content in the text, such as sensitive words, illegal remarks, false information, etc., and the corresponding document review result is generated.

[0103] Exemplarily, when the document to be reviewed is an image, feature extraction is performed on the preprocessed image and input into the image review model in the AI knowledge base for review. The image review model has pre-trained various image features, such as people, scenes, objects, etc. At the same time, object detection algorithms, such as the YOLO algorithm, are used to detect key objects in the image. By comparing with the illegal image feature library, images containing illegal content are identified, and the corresponding image review result is generated.

[0104] Exemplarily, when the document to be reviewed is a video, the preprocessed video file is split into a series of image frames, and key frames are extracted. The key frame extraction is based on shot change detection and motion analysis techniques to ensure that important content is not missed. Then, the key frames are reviewed using the same process as image review, that is, reviewed through the image review model in the AI knowledge base. In addition, the audio part of the video is also processed, audio features are extracted, the audio is converted into text using speech recognition technology, and then text review is performed through the document review model based on natural language processing in the AI knowledge base to detect whether there is illegal audio content in the video, and finally the corresponding video review result is generated.

[0105] It can be understood that the AI knowledge base needs to be pre-trained before review. Specifically, corresponding data is selected from the system for training. Exemplarily, promotional material samples are collected from the promotional pages of e-commerce platforms, advertising channels, etc. The promotional material samples include various types of data. For pictures, image recognition tools are used to identify and label elements such as products, logos, and text in the pictures, and OCR technology is used to extract the text information in the pictures. For videos, key frames are extracted through video processing software, image recognition and text recognition are performed on the key frames, and at the same time, the audio information of the video is extracted and converted into text.

[0106] Exemplarily, when training the AI knowledge base, customer service voice conversations and text conversation data can also be exported from the system's customer service system. For voice conversations, speech recognition technology is used to convert them into text, and then the text is cleaned to remove invalid information such as duplicate content, typos, and garbled characters, and lexical and syntactic analysis is performed.

[0107] Moreover, the AI knowledge base can also collect materials such as company-internal documents, meeting records, emails, etc. regarding business, policies, R & D, production, and rules and regulations, convert them into text format, and perform preprocessing to extract key information and Q&A pairs.

[0108] During the feature extraction process of training the AI knowledge base, when performing feature extraction on text data, for the preprocessed text data, word vector models such as Word2Vec and GloVe are used to convert each word in the text into a vector representation, and then models such as convolutional neural networks and recurrent neural networks are used to process the word vectors to extract the semantic features of the text. When performing feature extraction on images, convolutional neural networks such as VGG and ResNet are used to perform feature extraction on the images to obtain the feature vectors of the images. When performing feature extraction on videos, combining the key frame image features and video time series information, models such as 3D convolutional neural networks are used to perform feature extraction on the videos, and at the same time, the fusion of audio features and video image features is considered. The extracted text, image, and video features are fused, for example, by means of splicing, weighted summation, etc., to form a unified feature representation as the input for subsequent model training.

[0109] During the process of training and optimizing the AI knowledge base, model selection is usually performed on the AI knowledge base. Exemplarily, models with the Transformer architecture such as BERT and GPT are selected. These models show good performance in natural language processing tasks and can effectively process the fused feature data. During the training process, the fused feature data is divided into a training set, a validation set, and a test set. The training set is used to train the model. During the training process, appropriate hyperparameters such as the learning rate, the number of iterations, and the batch size are set. The cross-entropy loss function is used to calculate the difference between the model prediction result and the true label, and optimization algorithms such as stochastic gradient descent, Adagrad, and Adadelta are used to update the model parameters, continuously reducing the value of the loss function and improving the accuracy of the model. During training, it is also necessary to evaluate and optimize the training effect of the AI knowledge base: the validation set is used to evaluate the model during the training process, and indicators such as the accuracy rate, recall rate, and F1 value of the model are monitored. According to the evaluation results, the hyperparameters of the model are adjusted or the model structure is improved, such as increasing or decreasing the number of model layers, adjusting the attention mechanism, etc., to improve the performance of the model. Finally, the optimized model is tested using the test set to verify the generalization ability of the model.

[0110] Finally, apply the trained model to build an AI knowledge base, store various business data, their corresponding feature representations, and answers in a database in a certain format, such as using a relational database or a non-relational database. At the same time, regularly collect new system update data, such as e-commerce business data for example, process the new data according to the above processes of data collection, preprocessing, feature extraction, and model training, and then apply the new data and updated model parameters to the update of the knowledge base to ensure that the knowledge base can timely reflect the latest business information and customer needs.

[0111] The audit document modification suggestion module is used to obtain the multi-hop query request of the user according to the audit result and generate modification suggestions.

[0112] Specifically, when the audit result is not passed, the intelligent interaction model of the AI knowledge base is triggered. After the system administrator inputs the corresponding multi-hop query request through the interaction interface of the intelligent interaction model, the intelligent interaction model converts the multi-hop query request into a modification suggestion generated by the optimal association path starting from the summary content entity node in the AI knowledge base.

[0113] Furthermore, the intelligent interaction model includes a collaborative representation learning module, a modification suggestion set sorting module, and an association path learning module. The collaborative representation learning module is divided into an AI knowledge base learning representation module and a multi-hop query request learning representation module. Among them, the AI knowledge base learning representation module represents the entities and relationships of the industry specification information in the AI knowledge base in vectors. It can be understood that the industry specification information is such as the relevant industry regulations, policy documents, industry technical standards, and operation specifications mentioned above. The multi-hop query request learning representation module represents the multi-hop query request of the user in vectors and maps it to a vector space with the same dimension as the AI knowledge base learning representation module to obtain a modification suggestion set. Then, the modification suggestion set sorting module extracts the entity nodes under all association paths starting from the summary content entity node of the document to be audited in the AI knowledge base and screens the entity nodes related to the multi-hop query request as modification suggestion entity nodes; the association path learning module extracts the association paths related to the summary content entity node of the document to be audited from the AI knowledge base, embeds the association paths into a low-dimensional vector space, and then trains and updates them to obtain an updated intelligent interaction model.

[0114] Further, in the AI knowledge base learning and representation module, it is necessary to embed the semantic information of industry norms in the AI knowledge base into a low-dimensional space vector to enhance the accuracy of the modification suggestions generated by the intelligent interaction model. Specifically, the AI knowledge base learning and representation module generates corresponding triples (h, r, t) according to the semantic information of industry norms in the AI knowledge base, and represents them through the scoring function in the Neural Tensor Network (NTN) model. Specifically, the relationship between the semantic triples of industry norms in the AI knowledge base is represented by the NTN scoring function, that is

[0115]

[0116] Specifically, h represents the head entity vector in the semantic triple of industry norms, t represents the tail entity vector in the semantic triple of industry norms, and r represents the relationship between the head entity vector and the tail entity vector in the semantic triple of industry norms. The relationship between the head entity vector h and the tail entity vector t in the semantic triple of industry norms is interacted through a tensor by the NTN scoring function, and the semantic relationship between the head entity vector and the tail entity vector is obtained. represents a weight matrix related to the relationship r, represents a relationship matrix related to the relationship r. This relationship matrix is used to splice the head entity vector h and the tail entity vector t and then perform a linear transformation. [h;t] represents a new vector formed by splicing the head entity vector h and the tail entity vector t in a preset order, b r represents the bias vector related to the relationship r, and g(*) represents a non-linear activation function, specifically the hyperbolic tangent function.

[0117] Further, in the process of calculating the NTN model of the scoring function, it is necessary to construct a loss function L by constructing the relationship between the positive triple information and the negative triple information, that is

[0118]

[0119] where f + (h, r, t) represents training the positive triple through the NTN scoring function, f + (h, r, t) represents training the positive triple through the NTN scoring function, f - (h, r, t) represents training the negative triple through the NTN scoring function, and exp(*) represents the exponential function. It can be understood that the positive triple refers to the modification suggestions having an associated path with the abstract content entity nodes of the document to be reviewed, and the negative triple refers to the modification suggestions of other associated paths, that is, irrelevant modification suggestions.

[0120] Furthermore, the multi-hop query request learning and representation module also needs to represent the multi-hop query request of the user in vector form and map it to a vector space with the same dimension as the triple (h, r, t) generated corresponding to the semantic information of the industry standard. Specifically, the multi-hop query request learning and representation module obtains the semantic information hidden in the multi-hop query request and performs deep bidirectional semantic representation in combination with the context information, that is, given the multi-hop query request q = (q1, q2,... q n ), where q1, q2,... q n represent the segmented query keywords, n represents a natural number, and it is assumed that the abstract content entity node A ∈ E of the file to be reviewed and the set of modified suggestion entity nodes S q ∈ E, where E is the set of entity nodes of the semantic information of the industry standard in the AI knowledge base. At the head and tail of the multi-hop query request q, the head position marker [CLS] and the tail position marker [SEP] are respectively set, and the multi-hop query request q with the inserted position markers is input into the bidirectional long short-term memory network model in the multi-hop query request learning and representation module, and the average semantic vector of the last hidden layer in the bidirectional long short-term memory network model is used as the representation vector e q of the multi-hop query request q, that is

[0121] e q = Mean(Bi lstm([CLS], q1, q2,... q n , [SEP]))

[0122] where the function Mean(*) represents calculating the average semantic vector of the multi-hop query request q. Then, through the fully connected layer in the bidirectional long short-term memory network, e q is transformed to obtain the representation vector e q′ of the new multi-hop query request, that is

[0123]

[0124] where represents a weight matrix related to the relationship r in the triple (h, r, t) generated corresponding to the semantic information of the industry standard in the AI knowledge base, represents a relationship matrix related to the relationship r in the triple (h, r, t) generated corresponding to the semantic information of the industry standard in the AI knowledge base, b r represents the bias vector related to the relationship r, and g(*) represents a non-linear activation function, specifically the hyperbolic tangent function. By transforming the multi-hop query request, the representation vector of the new multi-hop query request is obtained, and the representation vector e q′Map it to a vector space with the same dimension as the AI knowledge base learning representation module to obtain a set of modification suggestions.

[0125] Further, when iteratively training the multi-hop query request learning representation module, when the finally obtained entity node belongs to the corresponding modification suggestion, f(e a ,e q′ ,e Sq )>0, when the finally obtained entity node does not belong to the corresponding modification suggestion, where, e a represents the vector of the abstract content entity node of the file to be reviewed, e q′ represents the representation vector of the new multi-hop query request, represents the representation vector of the modification suggestion entity node.

[0126] Further, in the modification suggestion set sorting module, when extracting entity nodes under all associated paths starting from the abstract content entity node of the file to be reviewed from the AI knowledge base and screening entity nodes related to the multi-hop query request as modification suggestion entity nodes, the modification suggestion set sorting module also needs to calculate the output scores of entity nodes under all associated paths and the multi-hop query request q.

[0127] Specifically, in the process of screening modification suggestion entity nodes, since there are multiple associated paths in the path from the abstract content entity node to the modification suggestion entity node, it is necessary to calculate the optimal associated path from all associated paths. In the process of calculating the optimal associated path, the modification suggestion set sorting module uses the global attention mechanism and the encoding and decoding functions in the Transformer network to represent all associated paths with vectors. Therefore, use P j (k) to represent the associated path from the abstract content entity node of the file to be reviewed to the modification suggestion entity node. It can be understood that the associated path includes multiple jump nodes encountered in the multi-hop query request. Therefore, the associated path P j (k) is expressed as:

[0128] P j (k) = [(a, p1, c a1 ), (c1, p2, c2),..., (c m-1 , p m , c m )]

[0129] where, P j (k) represents the k-th associated path from the abstract content entity node of the file to be reviewed to the j-th modification suggestion entity node. a represents the abstract content entity node of the file to be reviewed, p1, p2,..., p mDenote the jump path existing in the k-th associated path, c1, c2, ..., c m-1 , c m Denote the m jump nodes existing in the k-th associated path.

[0130] Furthermore, obtain the vectors of all associated paths from the summary content entity node to the j-th modified suggestion entity node after encoding through the encoding function in the Transformer network and the vector of the k-th associated path after encoding At the same time, also obtain the vectors of the jump nodes in all associated paths from the summary content entity node to the j-th modified suggestion entity node after encoding through the encoding function in the Transformer network and the vectors of the m jump nodes existing in the k-th associated path after encoding And obtain the vectors of all the associated paths after encoding and the vector of the k-th associated path after encoding and the vectors of the jump nodes in all associated paths and the vectors of the m jump nodes existing in the k-th associated path after encoding And obtain the first association information R1 between them, that is

[0131]

[0132] where, P j (k) represents the k-th associated path from the summary content entity node to the j-th modified suggestion entity node, represents the vector of the jump nodes in all associated paths from the summary content entity node to the j-th modified suggestion entity node obtained through the encoding function in the Transformer network, represents the vector of the k-th associated path obtained through the encoding function in the Transformer network represents the vectors of the m jump nodes existing in the k-th associated path obtained through the encoding function in the Transformer network represents the vectors of the m jump nodes existing in the k-th associated path obtained through the encoding function in the Transformer network

[0133] Furthermore, obtain e in the collaborative representation learning module a , e q′ , where e a represents the vector of the summary content entity node of the file to be reviewed, e q′ represents the representation vector of the new multi-hop query request, The representation vector of the j-th modified suggestion entity node, which combines e a and e q′ and performs vector splicing, and obtains the second correlation information R2 through the multi-layer perceptron in the Transformer network, that is Then, the first correlation information R1 and the second correlation information R2 are used to obtain the output score higher than the preset threshold through the sigmoid activation function, that is Finally, the sorted correlation paths are obtained, and the optimal correlation path and the modified suggestion entity node corresponding to the optimal correlation path are obtained. Further, the sorted correlation paths are iteratively trained through the correlation path learning module.

[0134] The audit document storage module is used to return the modified suggestions represented by the modified suggestion entity nodes to the user, and encrypt and store the modified suggestions.

[0135] Specifically, the AI knowledge base uses the consensus mechanism of the blockchain to authenticate and encrypt the obtained modified suggestions, and records the authentication and storage time at the same time, so as to ensure the data quality of the modified suggestions in the intelligent interaction process, and can also avoid duplicate storage of the obtained modified suggestions.

[0136] Exemplarily, the above AI knowledge base can be applied to promotional materials generated by businesses such as e-commerce. In particular, legal and quality control audits are carried out through the AI knowledge base. Specifically, the AI knowledge base also needs to collect promotional material samples from multiple channels such as e-commerce platform historical promotional materials, legal case libraries, and industry standard documents, mark the legal and quality control issues in them, such as false promotion, trademark infringement, product parameter errors, etc., and construct a corresponding training data set. When the system administrator audits the promotional text through the AI knowledge base, the AI knowledge base processes the pictures and text in the promotional material samples respectively, and invokes the corresponding audit models in the AI knowledge base for auditing respectively. Exemplarily, when the system administrator audits the promotional material sample of skin care products, the AI model detects that the statement "completely remove freckles after using for one week" in the text does not conform to medical common sense and advertising law regulations, and determines it as false promotion. At this time, the system administrator can further obtain the modified suggestions of the AI knowledge base through multi-hop query. Exemplarily, the obtained modified suggestion is "it helps to improve the freckle problem after using for a period of time". This way of placing the AI knowledge base in a quasi-objective and neutral position can reduce the fear of manual auditing of system administrators on e-commerce platforms, thereby greatly reducing the auditing pressure on administrators such as quality control and legal affairs.

[0137] Moreover, the platform customer service staff can also conduct multi-round conversations through the AI knowledge base to understand the specific information parameters of the corresponding products on the e-commerce platform. Exemplarily, the customer service staff asks, "Does the [model] smartphone support wireless charging and what is the maximum charging power?" The customer service enters the question into the customer service Q&A system. The AI knowledge base extracts the key information "[model] smartphone", "wireless charging", and "charging power" and retrieves them in the AI knowledge base. If the corresponding information exists in the knowledge base, it directly returns the answer "[model] smartphone supports wireless charging and the maximum charging power is [X]W" to the customer service; if there is no direct answer, the AI knowledge base generates an answer by performing multi-hop queries from the product technical parameters to the charging-related information and returns it to the customer service staff.

[0138] Moreover, the AI knowledge base can also provide corresponding Q&A for administration, human resources, finance, etc. within the company. Specifically, the AI knowledge base collects materials such as the rules and regulations, process documents, and frequently asked questions (FAQ) of the administrative, human resources, and finance departments of the company to form an original data set. Natural language processing technology is used to clean, annotate, and classify the original data. For example, the office supplies requisition process for the administrative category, the performance appraisal standards for the human resources category, and the invoice issuing specifications for the finance category are respectively annotated and classified. At the same time, the AI knowledge base constructs structured information by analyzing the correlation relationships between different data, such as the connection between employee training in the human resources department and training expense reimbursement in the finance department, and regularly updates the data source. According to company policy changes, business adjustments, etc., the knowledge base is updated in real-time or regularly to ensure the timeliness and accuracy of the knowledge. Further, company employees can conduct corresponding intelligent Q&A based on the AI knowledge base. Employees enter questions through the company's internal communication platform or a dedicated Q&A client. The system uses natural language understanding technology to parse the questions, extract key information, and determine the domain to which the questions belong. Specifically, it can determine the information belonging to a specific department of administration, human resources, or finance. According to the question domain and key information, it conducts retrieval and reasoning in the AI knowledge base. If there is a directly matching answer in the knowledge base, it is directly returned to the employee; if there is no direct answer, relevant information is integrated through multi-hop queries to generate an answer. For example, when an employee asks, "How to reimburse travel expenses after participating in external training", the system generates a detailed answer by performing multi-hop queries to associate the training regulations of human resources and the travel expense reimbursement process of finance. Feedback collection is carried out on the Q&A results. Employees can evaluate the answers, and the system records the feedback information for optimizing the Q&A rules and algorithms of each model in the AI knowledge base.

[0139] It can be understood that the AI knowledge base can also be applied to more fields, such as intelligent assistants, which can assist enterprise leaders in real-time grasping of data information on various business lines such as the company's R & D, products, sales, etc. It can also supervise the production progress from multiple dimensions based on the production scenarios of specific factories and enterprises, and can also determine whether the sales and customer service personnel's speech is professional and accurate according to the real-time conversations of sales and customer service.

[0140] Moreover, the AI knowledge base can also be applied in the financial field, where it can analyze a large amount of customer data and financial transaction data to more accurately assess the credit risk and market risk of customers, providing a scientific basis for credit approval. The AI knowledge base is applied in the media field to assist journalists in news writing, creative content creation, video script generation, etc., improving the efficiency and quality of content creation. It can provide personalized news, video, music and other content recommendations for users based on data such as the user's browsing history and interests, enhancing the user experience, and quickly reviewing news, social media content, videos, etc., identifying and filtering out bad information, false information, etc., to maintain the health and safety of the network environment.

[0141] The present invention also relates to an electronic device, characterized in that the electronic device includes a processor and a memory: the memory is used to store program codes and transmit the program codes to the processor, and the processor is used to execute a method for content review based on the AI knowledge base described in the above method embodiment according to the instructions in the program codes.

[0142] The present invention also relates to a computer-readable storage medium, characterized in that the computer-readable storage medium is used to store program codes and transmit the program codes to a processor, and the processor is used to execute a method for content review based on the AI knowledge base described in the above method embodiment according to the instructions in the program codes.

[0143] The above description is only a preferred embodiment of the present disclosure and an explanation of the applied technical principles. Those skilled in the art should understand that the scope of disclosure involved in the present disclosure is not limited to the technical solutions formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above disclosure concept. For example, the technical solutions formed by mutually replacing the above features with the (but not limited to) technical features with similar functions disclosed in the present disclosure.

Claims

1. A content review method based on AI knowledge base, characterized in that The following steps are involved: Step 1: Obtain the file to be uploaded and the corresponding file type, pre-process the file to be uploaded, and generate a file to be reviewed; Step 2: According to the type of the document to be reviewed, the corresponding review model in the AI ​​knowledge base is retrieved for review, and the corresponding review result is generated; Step 3: Obtain the user's multi-hop query request according to the review result, and generate modification suggestions; A multi-hop query request is obtained, and the intelligent interaction model of the AI ​​knowledge base converts the multi-hop query request into an optimal association path through the AI ​​knowledge base with the summary content entity node as the starting node, and generates a modification suggestion; Specifically, the intelligent interaction model includes a collaborative representation learning module, a modification suggestion set sorting module and an associated path learning module; the collaborative representation learning module is divided into an AI knowledge base learning representation module and a multi-hop query request learning representation module; The AI ​​knowledge base learning representation module vectorizes the entities and relationships of the industry specification information in the AI ​​knowledge base, that is, Among them, h represents the head entity vector in the industry specification semantic triple, t represents the tail entity vector in the industry specification semantic triple, r represents the relationship between the head entity vector and the tail entity vector in the industry specification semantic triple, and the relationship between the head entity vector h and the tail entity vector t in the industry specification semantic triple is interacted through a tensor through the NTN scoring function, and the semantic relationship between the head entity vector and the tail entity vector is obtained. represents a weight matrix associated with relation r, represents a relationship matrix related to the relationship r, which is used to concatenate the head entity vector h and the tail entity vector t and then perform linear transformation, [h; t represents a new vector formed by concatenating the head entity vector h and the tail entity vector t in a preset order, b r Represents the bias vector related to the relationship r, g(*) represents a nonlinear activation function; The multi-hop query request learning representation module performs vector representation on the multi-hop query request, obtains the representation vector of the multi-hop query request through the average semantic vector of the last hidden layer in the bidirectional long short-term memory network model, and then converts the representation vector of the multi-hop query request through the fully connected layer in the bidirectional long short-term memory network to obtain a new representation vector of the multi-hop query request, and maps the representation vector of the new multi-hop query request to the vector space of the same dimension as the AI ​​knowledge base learning representation module to obtain a set of modification suggestions; Step 4: Return the modification suggestion represented by the modification suggestion entity node to the user, and store the modification suggestion in encrypted form.

2. According to claim 1, a method for content review based on an AI knowledge base is characterized in that: In step 2, when the file to be reviewed is a document, the document review model based on natural language processing in the AI ​​knowledge base is used to review it and generate a corresponding document review result; When the file to be reviewed is an image, feature extraction is performed on the preprocessed image and the image is input into the image review model in the AI ​​knowledge base for review. The target detection algorithm is used to detect key targets in the image, and the image containing illegal content is identified by comparing it with the illegal image feature library, and the corresponding image review result is generated; When the file to be reviewed is a video, the preprocessed video file is split into image frames, key frames are extracted, and the image review model in the AI ​​knowledge base is used for review. In addition, the audio features of the audio part of the video are extracted, and the audio is converted into text using speech recognition technology. The text is then reviewed through a document review model based on natural language processing in the AI ​​knowledge base to detect whether there is any illegal audio content in the video, and finally generate a corresponding video review result.

3. According to claim 1, a method for content review based on an AI knowledge base is characterized in that: In step 3, in the process of calculating the score function NTN model, it is necessary to construct the loss function L by constructing the relationship between the positive triple information and the negative triple information, that is, Among them, f + (h, r, t) represents the positive triplet trained by the NTN score function, f + (h, r, t) represents the positive triplet trained by the NTN score function, f - (h, r, t) represents the training of negative triples through the NTN score function, and exp(*) represents the exponential function.

4. According to claim 1, a method for content review based on an AI knowledge base is characterized in that: In step 3, the average semantic vector of the last hidden layer in the bidirectional long short-term memory network model is used as the representation vector e of the multi-hop query request q q ,Right now e q =Mean(Bilstm([CLS],q1,q2,...q n ,[SEP])) The function Mean(*) represents the calculation of the average semantic vector of the multi-hop query request q; then the fully connected layer in the bidirectional long short-term memory network is used to convert e q Transform to get the new representation vector e of the multi-hop query request q′ ,Right now in represents a weight matrix associated with the relation r in the triple (h, r, t) corresponding to the semantic information generation of the industry specification in the AI ​​knowledge base, represents a relation matrix related to relation r in the triple (h, r, t) corresponding to the semantic information generation of industry specifications in the AI ​​knowledge base, b r Represents the bias vector associated with the relation r, and g(*) represents a nonlinear activation function.

5. A content review device based on AI knowledge base, characterized in that The device includes a module for obtaining documents to be reviewed, a module for reviewing documents, a module for suggesting revisions to the reviewed documents, and a module for storing the reviewed documents, wherein: A module for acquiring documents to be reviewed, used to obtain files to be uploaded and corresponding file types, pre-process the files to be uploaded, and generate files to be reviewed; The document review module is used to retrieve the corresponding review model in the AI ​​knowledge base for review according to the type of the document to be reviewed, and generate the corresponding review result; An audit document modification suggestion module is used to obtain a user's multi-hop query request based on the audit result and generate modification suggestions; A multi-hop query request is obtained, and the intelligent interaction model of the AI ​​knowledge base converts the multi-hop query request into an optimal association path through the AI ​​knowledge base with the summary content entity node as the starting node, and generates a modification suggestion; Specifically, the intelligent interaction model includes a collaborative representation learning module, a modification suggestion set sorting module and an associated path learning module; the collaborative representation learning module is divided into an AI knowledge base learning representation module and a multi-hop query request learning representation module; The AI ​​knowledge base learning representation module vectorizes the entities and relationships of the industry specification information in the AI ​​knowledge base, that is, Among them, h represents the head entity vector in the industry specification semantic triple, t represents the tail entity vector in the industry specification semantic triple, r represents the relationship between the head entity vector and the tail entity vector in the industry specification semantic triple, and the relationship between the head entity vector h and the tail entity vector t in the industry specification semantic triple is interacted through a tensor through the NTN scoring function, and the semantic relationship between the head entity vector and the tail entity vector is obtained. represents a weight matrix associated with relation r, represents a relationship matrix related to the relationship r, which is used to concatenate the head entity vector h and the tail entity vector t and then perform linear transformation, [h; t represents a new vector formed by concatenating the head entity vector h and the tail entity vector t in a preset order, b r Represents the bias vector related to the relationship r, g(*) represents a nonlinear activation function; The multi-hop query request learning representation module performs vector representation on the multi-hop query request, obtains the representation vector of the multi-hop query request through the average semantic vector of the last hidden layer in the bidirectional long short-term memory network model, and then converts the representation vector of the multi-hop query request through the fully connected layer in the bidirectional long short-term memory network to obtain a new representation vector of the multi-hop query request, and maps the representation vector of the new multi-hop query request to the vector space of the same dimension as the AI ​​knowledge base learning representation module to obtain a set of modification suggestions; The review document storage module is used to return the modification suggestions represented by the modification suggestion entity node to the user and store the modification suggestions in encrypted form.

6. The AI ​​knowledge base content review device according to claim 5, characterized in that: In the document review module, when the file to be reviewed is a document, the document review model based on natural language processing in the AI ​​knowledge base is used to review it and generate a corresponding document review result; When the file to be reviewed is an image, feature extraction is performed on the preprocessed image and the image is input into the image review model in the AI ​​knowledge base for review. The target detection algorithm is used to detect key targets in the image, and the image containing illegal content is identified by comparing it with the illegal image feature library, and the corresponding image review result is generated; When the file to be reviewed is a video, the preprocessed video file is split into image frames, key frames are extracted, and the image review model in the AI ​​knowledge base is used for review. In addition, the audio features of the audio part of the video are extracted, and the audio is converted into text using speech recognition technology. The text is then reviewed through a document review model based on natural language processing in the AI ​​knowledge base to detect whether there is any illegal audio content in the video, and finally generate a corresponding video review result.

7. The AI ​​knowledge base content review device according to claim 5, characterized in that: In the document review modification suggestion module, in the process of calculating the score function NTN model, it is necessary to construct the loss function L by constructing the relationship between the positive triple information and the negative triple information, that is, Among them, f + (h, r, t) represents the positive triplet trained by the NTN score function, f + (h, r, t) represents the positive triplet trained by the NTN score function, f - (h, r, t) represents the training of negative triples through the NTN score function, and exp(*) represents the exponential function.

8. The AI ​​knowledge base content review device according to claim 5, characterized in that: In the document review modification suggestion module, the average semantic vector of the last hidden layer in the bidirectional long short-term memory network model is used as the representation vector e of the multi-hop query request q q ,Right now e q =Mean(Bilstm([CLS],q1,q2,...q n ,[SEP])) The function Mean(*) represents the calculation of the average semantic vector of the multi-hop query request q; then the fully connected layer in the bidirectional long short-term memory network is used to convert e q Transform to get the new representation vector e of the multi-hop query request q′ ,Right now in represents a weight matrix associated with the relation r in the triple (h, r, t) corresponding to the semantic information generation of the industry specification in the AI ​​knowledge base, represents a relation matrix related to relation r in the triple (h, r, t) corresponding to the semantic information generation of industry specifications in the AI ​​knowledge base, b r Represents the bias vector associated with the relation r, and g(*) represents a nonlinear activation function.

9. An electronic device, characterized in that: The electronic device includes a processor and a memory: the memory is used to store program code and transmit the program code to the processor, and the processor is used to execute the AI ​​knowledge base content review method described in any one of claims 1 to 4 according to the instructions in the program code.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store program code and transmit the program code to a processor, and the processor is used to execute the AI ​​knowledge base content review method described in any one of claims 1 to 4 according to the instructions in the program code.

Images