An AI knowledge base construction method and system for cross-domain collaboration

Abstract

The application belongs to the technical field of knowledge engineering, and discloses an AI knowledge base construction method and system for cross-domain collaboration, which comprises the following steps: based on a symbolic dynamic knowledge graph, an adversarial neural network is used to decouple the domain-independent intention and the domain feature vector in a user query, a generalized intention vector and a personalized weight parameter are generated; the semantic similarity between the generalized intention vector and the dynamic knowledge graph is calculated, a multi-hop retrieval is used to recall multi-domain high-relevance knowledge fragments by using HNSW, and a cross-domain knowledge set is obtained; a variational Bayesian uncertainty estimation method is used to scan the knowledge fragments in the cross-domain knowledge set and shield illegal content, and a personalized cross-domain answer is generated in combination with the personalized weight parameter. Through the generation of the generalized intention vector and the personalized weight parameter, the application realizes user intention decoupling and personalized weighting, adapts the query to multiple domains, and improves the relevance of knowledge retrieval.

Description

Technical Field

[0001] This invention relates to the field of knowledge engineering technology, and in particular to a method and system for constructing an AI knowledge base for cross-domain collaboration. Background Technology

[0002] Currently, the construction of enterprise-level AI knowledge bases mainly relies on multimodal data integration and structured knowledge representation technologies. By integrating heterogeneous data sources such as text, images, and audio, semantically related knowledge graphs are constructed to support intelligent retrieval and reasoning. Typical processes include knowledge extraction, entity alignment, relation modeling, and graph storage. The application of deep learning-based cross-modal embeddings (such as CLIP and ALBEF) and graph neural networks (GNNs) significantly improves the completeness of knowledge representation. In cross-domain scenarios, traditional methods often adopt modular architecture designs, such as building sub-knowledge bases for different domains and achieving initial collaboration through ontology mapping, or using pre-trained language models (such as BERT and GPT) to generate universal semantic vectors to reduce the impact of domain differences. These methods rely on static knowledge modeling, supplemented by a regular manual review and update mechanism, and have achieved certain effectiveness in vertical fields such as healthcare and finance.

[0003] However, existing methods still face two major limitations: First, insufficient semantic consistency in cross-domain knowledge fusion. Heterogeneous modal data suffer from vector alignment bias due to differences in feature spaces, and conventional multimodal fusion (such as attention weighting) fails to reveal the potential relationship structure between modeling domains, leading to semantic ambiguity or information loss when mapping cross-domain entity relationships. Second, the real-time performance and generalization of dynamic knowledge evolution are limited. Static graphs rely on full-scale retraining to adapt to new domains, and incremental update mechanisms often focus on local optimization within a single domain, lacking constraints on the collaborative evolution of global knowledge and making it difficult to balance domain specificity with the expression of generalization intent. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for constructing an AI knowledge base for cross-domain collaboration to solve the problem of unified representation of cross-modal knowledge and continuous collaborative evolution of dynamic knowledge graphs.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for constructing an AI knowledge base for cross-domain collaboration, comprising the following steps: S1. Collect cross-domain raw data for preprocessing, extract heterogeneous modal vectors at the same dimension, and obtain a unified cross-modal vector set through attention-weighted fusion; S2. An entity embedding matrix is ​​aggregated based on a cross-modal unified vector set, and a relation type matrix is ​​generated by combining an inductive network. The entity embedding matrix and the relation type matrix are used as a symbolic dynamic knowledge graph. S3. Based on symbolic dynamic knowledge graph, adversarial neural network is used to decouple domain-independent intent and domain feature vector in user query to generate generalized intent vector and personalized weight parameters. S4. Calculate the semantic similarity between the generalized intent vector and the dynamic knowledge graph, and retrieve highly relevant knowledge fragments from multiple domains to obtain a cross-domain knowledge set. S5. Scan out knowledge fragments in the cross-domain knowledge set and block illegal content, and generate personalized cross-domain answers by combining personalized weight parameters; S6. Based on personalized cross-domain answers and user feedback logs, the dynamic knowledge graph is continuously evolved to form a cross-domain collaborative AI knowledge base.

[0008] As a preferred embodiment of the AI ​​knowledge base construction method for cross-domain collaboration described in this invention, in S1, the cross-domain raw data includes text data, image data, and time-series data corresponding to enterprise internal knowledge documents, business logs, and monitoring data; The preprocessing includes filtering, noise reduction, and normalization of the cross-domain raw data.

[0009] As a preferred embodiment of the AI ​​knowledge base construction method for cross-domain collaboration described in this invention, the step of obtaining a cross-modal unified vector set includes: Semantic vectors are extracted from the preprocessed cross-domain raw data, and then the semantic vectors are projected and aligned in real time to obtain heterogeneous modal vectors of the same dimension. By utilizing the hybrid head attention dynamic routing mechanism, the similarity scores between heterogeneous modal vectors of the same dimension are calculated and used as weights. The results are then weighted and fused to output a unified cross-modal vector set.

[0010] As a preferred embodiment of the AI ​​knowledge base construction method for cross-domain collaboration described in this invention, S2 specifically comprises: Aggregate the cross-modal unified vector set to obtain the entity embedding matrix; By analyzing common neighbor nodes in the entity embedding matrix through inductive network analysis, cross-domain relationship edges are obtained and sparse branch pruning is performed to generate a relationship matrix, which is then linked with the entity embedding matrix to construct a symbolic dynamic knowledge graph.

[0011] As a preferred embodiment of the AI ​​knowledge base construction method for cross-domain collaboration described in this invention, S3 specifically includes: Extract the query semantic vector from the user's original query text, and use the generator of the adversarial neural network to separate the domain-independent intent and domain features; The semantic rationality of the symbolic dynamic knowledge graph is verified by using an adversarial neural network discriminator, and finally the generalized intent vector is obtained through the training mechanism of the adversarial neural network. Personalized weight parameters are generated based on query semantic vectors and user feedback logs.

[0012] As a preferred embodiment of the AI ​​knowledge base construction method for cross-domain collaboration described in this invention, S4 specifically comprises: The cosine similarity between the generalization intent vector and the knowledge fragment vectors in the dynamic knowledge graph is calculated and combined into a set of cosine similarity values. Based on the set of cosine similarity values, the knowledge fragments in the dynamic knowledge graph are clustered to obtain a cross-domain candidate knowledge fragment set. The HNSW algorithm is used to build a multi-layer graph for the cross-domain candidate knowledge fragment set. Starting from the high-level graph, the search is carried out, and the node with the highest cosine similarity to the generalization intention vector is visited as the starting node. Then, the next-hop node with the most similarity to the cross-domain candidate knowledge fragment is found among the neighbors of the starting node. Based on the next-hop node, HNSW is used to search for neighbor nodes until a multi-domain knowledge fragment sequence is built. Sort the sequence of knowledge fragments from multiple domains in descending order of cosine similarity to generate a cross-domain knowledge set.

[0013] As a preferred embodiment of the AI ​​knowledge base construction method for cross-domain collaboration described in this invention, S5 specifically includes: The system scans knowledge fragments in cross-domain knowledge sets, removes illegal knowledge fragments, generates compliant knowledge fragments, and combines personalized weight parameters to generate weighted knowledge vectors. Based on weighted knowledge vectors, a diverse candidate answer framework is constructed by generating semantic variants through GAN, and a bundle search decoding and source tagging are used to finally output personalized cross-domain answers.

[0014] As a preferred embodiment of the AI ​​knowledge base construction method for cross-domain collaboration described in this invention, in S6, the continuous evolution of the dynamic knowledge graph based on personalized cross-domain answers and user feedback logs includes: Extract semantic feature vectors from users' personalized cross-domain answers; Extract user behavior feature vectors from user feedback logs; The semantic feature vector and the user behavior feature vector are concatenated into a local feature vector, and forward propagation is performed through a fully connected neural network to obtain the locally generated logical prediction value. Weights are assigned to local feature vectors, and logical prediction values ​​are aggregated to obtain a global knowledge vector.

[0015] As a preferred embodiment of the AI ​​knowledge base construction method for cross-domain collaboration described in this invention, the AI ​​knowledge base constituting cross-domain collaboration includes: The global knowledge vector is fused with the historical entity embedding matrix to obtain the updated entity embedding matrix, and the cosine similarity of each entity vector in the updated entity embedding matrix is ​​calculated and combined into a cosine similarity vector. The updated relation matrix is ​​obtained by fusing the historical relation matrix and the cosine similarity vector. Based on the updated entity embedding matrix and the updated relation matrix, a new symbolic dynamic knowledge graph is constructed through graph links, thereby forming a cross-domain collaborative AI knowledge base.

[0016] Secondly, this invention provides an AI knowledge base construction system for cross-domain collaboration, comprising: The data acquisition module is used to collect raw data from across domains for preprocessing, extract heterogeneous modal vectors of the same dimension from the model, and obtain a unified cross-modal vector set through attention-weighted fusion. The knowledge graph generation module is used to aggregate entity embedding matrices from cross-modal unified vector sets, combine them with inductive networks to generate relation type matrices, and use the entity embedding matrix and relation type matrix as a symbolic dynamic knowledge graph. The intent analysis module is used to symbolize dynamic knowledge graphs. It uses adversarial neural networks to decouple domain-independent intents and domain feature vectors in user queries, and generates generalized intent vectors and personalized weight parameters. The semantic graph transition module is used to calculate the semantic similarity between the generalized intent vector and the dynamic knowledge graph, and to retrieve and recall highly relevant knowledge fragments from multiple domains to obtain a cross-domain knowledge set. The qualified knowledge generation module is used to scan out knowledge fragments in the cross-domain knowledge set and block illegal content, and generate personalized cross-domain answers by combining personalized weight parameters; The knowledge graph evolution module is used to continuously evolve the dynamic knowledge graph with personalized cross-domain answers and user feedback logs, forming a cross-domain collaborative AI knowledge base.

[0017] The beneficial effects of this invention are as follows: By generating generalized intent vectors and personalized weight parameters, this invention achieves user intent decoupling and personalized weighting, adapts queries to multiple domains, improves knowledge retrieval relevance, and achieves the beneficial effects of enhancing response accuracy and personalization; by continuously evolving the dynamic knowledge graph, it achieves real-time updates of the knowledge base, integrates user feedback data, maintains synergy, and improves cross-domain adaptability and evolution capabilities. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Fig. 1 A flowchart illustrating a method for constructing an AI knowledge base for cross-domain collaboration; Fig. 2 A schematic diagram of a system for building an AI knowledge base for cross-domain collaboration; Fig. 3 Flowchart for generating a cross-modal unified vector set; Fig. 4 A flowchart for generating symbolic dynamic knowledge graphs. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0023] Reference Figs. 1-4 This is one embodiment of the present invention, which provides a method for constructing an AI knowledge base for cross-domain collaboration, including the following steps:

[0024] S1. Collect cross-domain raw data for preprocessing, extract heterogeneous modal vectors of the same dimension from the model, and obtain a unified cross-modal vector set through attention-weighted fusion.

[0025] Cross-domain raw data includes text data, image data, and time-series data corresponding to enterprise internal knowledge documents, business logs, and monitoring data; Furthermore, this includes knowledge texts stored in document form such as corporate rules and regulations, operating procedures, product manuals, frequently asked questions, training materials, meeting minutes, and service manuals; text data generated from various business scenarios, such as work order records, customer service dialogue texts, and email or IM conversation summaries; image data used to present business status and structure, such as on-site equipment photos, process diagrams, monitoring screen screenshots, topology diagrams, and report screenshots; and time-series data collected in chronological order, such as transaction indicator curves, number of interface calls, response time, CPU / memory utilization, disk I / O, network traffic, and number of alarm triggers generated from various business activities and monitoring links. By connecting with existing enterprise document management platforms, collaborative office platforms, knowledge base platforms, business databases, and log middleware, text and time-series data are acquired through API calls, offline database extraction, message queue subscriptions, and batch file imports using data acquisition adapters. For monitoring screens, report pages, and industrial control interfaces, image data and its associated timestamps and metadata are periodically captured through screenshot interfaces and export functions provided by the monitoring terminal or by deploying lightweight acquisition agents. During the acquisition process, data source identifiers, business domain tags, and time stamps are uniformly added to provide a complete raw data foundation for subsequent normalization processing and knowledge base construction.

[0026] Preprocessing includes filtering, noise reduction, and normalization of the cross-domain raw data; Using a dedicated encoder, semantic vectors are extracted from the preprocessed cross-domain raw data, and then the semantic vectors are projected and aligned in real time to obtain heterogeneous modal vectors of the same dimension. Furthermore, for text data, the BERT (Bidirectional Encoder Representations from Transformers) tokenizer is used to segment the text and add [CLS] and [SEP] special markers to form a segmented sequence. The segmented sequence is then input into the BERT model to obtain the keyword feature vector output from the last hidden layer. Image data is preprocessed (adjusted to 224×224 pixels and normalized) and then input into the ResNet-50 model (Residual Network) for processing, sequentially through an initial convolutional layer, a max pooling layer, and four sets of residual blocks (conv2_x to conv5_x). Finally, a global average pooling layer outputs a visual feature vector. For audio data, a Wav2Vec (Wave-to-Vector) network is used, generating acoustic feature vectors through five convolutional layers. The keyword feature vector, visual feature vector, and acoustic feature vector are combined into a heterogeneous feature vector set, which is then projected and aligned to a heterogeneous modality vector of the same dimension through a linear transformation.

[0027] Furthermore, the BERT model uses text data and is trained through a self-supervised pre-training method. Its core tasks include masked language modeling and next-sentence prediction. The optimization objective is to maximize the accuracy of context word representations. The ResNet-50 model uses image data and is trained through supervised learning. It uses the cross-entropy loss function to optimize parameters and combines a stochastic gradient descent (SGD) optimizer and weight decay regularization to suppress overfitting. The final output is an image classification probability distribution. The Wav2Vec network uses large-scale unlabeled raw audio waveform data and is trained through a self-supervised contrastive learning method. Its core mechanism is that the quantization module generates potential feature targets and uses a noise contrastive estimation loss function to optimize the encoder network, minimizing the distance between the representation vectors of similar audio segments and maximizing the distance between dissimilar segments.

[0028] It should be noted that the core technical concept of cross-modal semantic alignment is to map different information types (such as text, images, and audio) into vector representations of the same dimension through encoding and projection alignment methods.

[0029] Using a hybrid head attention dynamic routing mechanism, the similarity scores between heterogeneous modal vectors of the same dimension are calculated and used as weights. Weighted fusion is then performed to output a fused cross-modal unified vector set. Furthermore, the heterogeneous modal vectors of the same latitude are input into the hybrid head attention dynamic routing mechanism. Each heterogeneous modal vector of the same latitude undergoes a bilinear mapping operation to generate a projection representation. The projection representation is then subjected to a transpose matrix multiplication operation to calculate a similarity score. The similarity score is normalized by the Softmax function to generate weights. The hybrid head attention dynamic routing mechanism performs a weighted summation operation on the heterogeneous modal vectors of the same latitude based on the weights. Each heterogeneous modal vector of the same latitude is first multiplied with the fused projection matrix and then weighted. The weighted summation result is output as a fused cross-modal unified vector set.

[0030] It should be noted that the hybrid head attention dynamic routing mechanism is an attention optimization architecture that selects the most relevant subset of attention heads to process multimodal inputs through dynamic routing strategies. Its core is to calculate similarity scores, perform Softmax normalization of weights, and finally achieve cross-modal feature fusion through weighted fusion projection, ultimately outputting a cross-modal unified vector set in a unified semantic space.

[0031] S2. Based on the cross-modal unified vector set, it is mapped to the entity embedding matrix through a graph neural network, combined with the inductive network to generate the relation matrix, and the incremental graph convolutional network is used to obtain the symbolic dynamic knowledge graph.

[0032] A graph neural network is used to map a cross-modal unified vector into an entity embedding matrix. Furthermore, the cross-modal unified vector set input graph convolutional network performs message passing operations, updating node representations by aggregating the cross-modal features of neighboring nodes. Next, layer normalization is performed on the aggregated node representations to scale them to a standard distribution. Then, all normalized node representations are concatenated along the feature dimensions to form a single high-dimensional vector. This single high-dimensional vector is the multimodal fusion vector that integrates multimodal semantic associations. The entity nodes corresponding to the multimodal fusion vector are then directly arranged row-wise according to their index order to form an entity embedding matrix.

[0033] It should be noted that the graph convolutional network uses a cross-modal unified vector set, is trained through supervised learning, optimizes the cross-entropy loss function using the gradient descent algorithm, and updates node representations using a neighborhood aggregation mechanism, ultimately outputting classification or link prediction results.

[0034] By analyzing the common neighbor nodes in the entity embedding matrix through inductive network analysis, cross-domain relationship edges are obtained and sparse branch pruning is performed to generate a relationship matrix. Furthermore, the entity embedding matrix is ​​processed by inductive network to calculate the distribution of common neighbor nodes among the entities in the entity embedding matrix to identify shared neighbor patterns; cross-domain relationship edges are obtained based on the shared neighbor pattern recognition results to capture cross-domain associations; sparse branch pruning is performed on the obtained cross-domain relationship edges to filter low-weight connections and achieve simplification; and the sparsed cross-domain relationship edges are integrated to generate a relationship matrix. Furthermore, the inductive network uses an entity embedding matrix, is trained through self-supervised learning, and employs a contrastive loss function to optimize the model, minimizing the distance between entities with similar shared neighbor patterns in the vector space, and dynamically pruning low-weight cross-domain relationship edges based on KL divergence.

[0035] Based on entity embedding matrices and relation matrices, a symbolic dynamic knowledge graph is constructed by combining graph convolutional networks with residual links. Furthermore, the entity embedding matrix and relation matrix are used as inputs for multi-layer feature extraction through a graph convolutional network. Each layer of the graph convolutional network applies an aggregation function in conjunction with the representations of neighboring nodes. Specifically, the aggregation function operation involves collecting the embeddings of all neighboring nodes of the target node, weighting the neighbor representations using the relation matrix, and performing a weighted average aggregation to obtain the aggregation result. Next, a linear transformation operation is performed on the aggregation result, applying a transformation matrix to generate a linear output. Then, a ReLU activation function operation is applied to achieve a non-linear mapping to generate an activation output. Finally, residual links are used to add the activation output to the input entity embeddings to generate the output representation of the current network layer, connecting the input (the input features of the current graph convolutional network layer) and the output to alleviate the gradient vanishing problem. Residual links allow information (the input entity embedding matrix of the current graph convolutional network layer) to skip some computational layers. The output representation generated after multi-layer feature extraction by the graph convolutional network is used to update entities and relations, ultimately constructing a symbolic dynamic knowledge graph.

[0036] S3. Based on symbolic dynamic knowledge graph, adversarial neural networks are used to decouple domain-independent intent and domain feature vector in user queries, generating generalized intent vectors and personalized weight parameters.

[0037] The query semantic vector is extracted from the user's original query text using the BERT model; Furthermore, the user's original query text is converted into a TokenID sequence containing [CLS] and [SEP] tags by the BERT tokenizer; the TokenID sequence is input into the BERT model, and the BERT model calculates the hidden state through a multi-layer Transformer encoder; the vector of the [CLS] tag position is extracted from the last hidden state output by the BERT model as the query semantic vector; It should be noted that the user's original query text refers to the initial natural language question or request submitted by the user to the system, which usually contains click-through rate, command verbs, and sentiment expression behavior markers.

[0038] Based on query semantic vectors, a generator using an adversarial neural network separates domain-independent intents and domain features; Furthermore, the query semantic vector is input into the generator of the adversarial neural network. The query semantic vector is first input into a shared fully connected layer, which undergoes a linear transformation and ReLU activation to generate a shared feature representation. Then, the shared feature representation is input into two independent branches in parallel. Branch one performs a linear transformation through an independent fully connected layer and is mapped by a Sigmoid activation function to generate a domain-independent intent feature vector. This branch ensures that the vector retains the core intent semantics by minimizing the intent classification cross-entropy loss, generating a domain-independent intent feature vector. Branch two is forward-propagated through a gradient reversal layer, performs a linear transformation through an independent fully connected layer, and is mapped by a Sigmoid activation function to generate a domain feature vector. This branch minimizes the semantic rationality loss, flips the gradient sign during backward propagation in the gradient reversal layer, performs gradient ascent operation to increase the loss value, and forces the vector to strengthen the domain discriminativeness. The outputs of the two are orthogonal, realizing the decoupling of intent and domain semantics, and generating a domain feature vector. The optimization objectives of the generator of the adversarial neural network include minimizing the intent classification loss to retain the core semantics and maximizing the domain discriminative gain to deceive the discriminator of the adversarial neural network, forcing the generator to strip away domain-related information. Furthermore, the adversarial neural network uses user query semantic vectors, true intent labels, and logical rule constraints from a symbolic dynamic knowledge graph to train through an adversarial training framework: the generator employs joint optimization of intent classification loss and semantic rationality loss, combined with a gradient reversal layer to achieve adversarial game between the generator and the discriminator; the discriminator distinguishes between legal and illegal feature combinations through cross-entropy loss, synchronously updates the graph's logical rules to achieve dynamic adaptation, and finally outputs a generalized intent vector that conforms to the graph's semantics.

[0039] It should be noted that domain-independent intent refers to the core semantic objective in a user query that is not restricted by a specific domain.

[0040] Domain features are a structured encapsulation of domain knowledge, containing three core semantic categories: vocabulary, rules, and data patterns, and are aligned with the knowledge graph in real time through dynamic weights.

[0041] Based on domain-independent intent and domain features, an adversarial neural network discriminator is used to verify the semantic rationality of the symbolic dynamic knowledge graph; Furthermore, the domain-independent intent vector and the domain feature vector are sequentially concatenated along the feature dimensions to generate a unified high-dimensional combined vector. This high-dimensional combined vector is then input into the discriminator of the adversarial neural network. The discriminator of the adversarial neural network invokes the explicit logical rules of the symbolic dynamic knowledge graph for semantic matching verification. If the high-dimensional combined vector conforms to the logical constraints of the symbolic dynamic knowledge graph, the discriminator of the adversarial neural network outputs a high confidence level. If the combination violates the rules of the symbolic dynamic knowledge graph, the discriminator of the adversarial neural network outputs a low confidence level and generates a symbolic contradiction marker.

[0042] Obtain the generalized intent vector through the training mechanism of the adversarial neural network; Furthermore, the generator of the adversarial neural network (ANN) receives feedback from the discriminator of the AAN, forcing it to learn the constraints of the symbolic dynamic knowledge graph by minimizing the intent classification loss and maximizing the semantic plausibility loss. The operation of minimizing the intent classification loss involves calculating the cross-entropy loss between the generalized intent vector output by the AAN generator and the true intent label using an independent intent classifier, and backpropagating this loss gradient to reduce intent prediction bias. The operation of maximizing the semantic plausibility loss involves converting the confidence output of the AAN discriminator for the combination of domain-independent intent features and domain features into a loss function (formula: semantic plausibility loss = 1 - discriminator confidence). During backpropagation through the gradient inversion layer, the gradient value is multiplied by -1 to achieve sign inversion. The AAN generator performs a gradient ascent operation to increase the loss value, forcing the combination of domain-independent intent features and domain features to improve the discriminator confidence and satisfy the logical constraints of the symbolic dynamic knowledge graph. Meanwhile, the adversarial neural network (ANN) discriminator, optimized through cross-entropy loss, distinguishes between legitimate and illegitimate feature combinations. Combined with a gradient reversal layer, it inversely drives the ANN generator to align with the semantics of the knowledge graph. During training, the symbolic dynamic knowledge graph updates its logical rules in real time, and the ANN discriminator synchronously adjusts its verification logic to ensure the dynamic adaptability of domain-independent intent features and domain features. Finally, the optimized ANN generator outputs a generalized intent vector.

[0043] Personalized weight parameters are generated using the Sigmoid activation function based on query semantic vectors and user feedback logs. Furthermore, the query semantic vector and user feedback logs are input into the feature fusion layer; the feature fusion layer performs a dot product operation on the query semantic vector and user feedback logs to generate fused feature values; the fused feature values ​​are input into the Sigmoid activation function, which non-linearly maps the fused feature values ​​to the (0,1) interval, and the mapping result is used as personalized weight parameters.

[0044] S4. Calculate the semantic similarity between the generalized intent vector and the dynamic knowledge graph, and use HNSW multi-hop retrieval to recall highly relevant knowledge fragments from multiple domains to obtain a cross-domain knowledge set.

[0045] The cosine similarity is calculated using the cosine similarity formula and combined with the cosine similarity between the generalized intention vector and the vectors of each knowledge segment in the dynamic knowledge graph to form a set of cosine similarity values. The formula for calculating cosine similarity is: ; in, For cosine similarity, To generalize the intent vector, This refers to the vector of knowledge fragments in a dynamic knowledge graph.

[0046] Clustering is performed using the K-means algorithm based on the cosine similarity value set. The cluster boundary similarity threshold is obtained, and knowledge fragments in the dynamic knowledge graph with a cosine similarity greater than the cluster boundary similarity threshold (usually ranging from 0.5 to 0.8) are identified as relevant items, forming a cross-domain candidate knowledge fragment set. Furthermore, the cosine similarity values ​​of all fragment pairs in the dynamic knowledge graph constitute a cosine similarity value set; the cosine similarity value set is input into the K-means algorithm, which randomly initializes K cluster centroids; each value in the cosine similarity value set is assigned to the nearest cluster based on its cosine distance to the K cluster centroids; after assignment, the K-means algorithm recalculates the mean of all cosine similarity values ​​within each cluster as the new cluster centroid; the assignment and updating of cluster centroids is repeated until the change in cluster centroids is less than the convergence threshold or the maximum number of iterations is reached; after clustering, the cluster boundaries separating related and unrelated items are identified, and the cluster boundary similarity threshold is defined as the cosine similarity value corresponding to the maximum distance between different clusters; if the cosine similarity value of any fragment pair in the dynamic knowledge graph is greater than the cluster boundary similarity threshold, it is determined to be a related item; all related items are combined to form a cross-domain candidate knowledge fragment set.

[0047] The HNSW (HierarchicalNavigableSmallWorld) algorithm is used to build a multi-layer graph for cross-domain candidate knowledge fragment sets. A greedy search strategy is adopted, starting from the high-level graph and approaching it layer by layer downwards, prioritizing the node with the highest cosine similarity to the generalization intent vector as the starting node. Furthermore, the HNSW (HierarchicalNavigableSmallWorld) algorithm is used to construct a multi-layer graph for the cross-domain candidate knowledge fragment set. First, in the base layer, the vector nodes of all cross-domain candidate knowledge fragment sets are fully connected to form a topology. By calculating the Euclidean distance, a fixed number of nearest neighbors are strictly selected for each node to establish a dense connection at the bottom layer. Then, the nodes are assigned a level according to the exponential decay probability rule, and a high-level sparse subgraph is recursively generated. The scale of the neighbor connections of the nodes is reduced proportionally with each level increase to achieve level sparsity. The high-level graph contains sparse nodes and long-distance connections, while the bottom-level graph contains all nodes of the cross-domain candidate knowledge fragment set and short-distance connections. The entry point of the high-level graph (the highest-level node) is used as the starting node for the search. The greedy search strategy starts from the entry point of the high-level graph, calculates the cosine similarity between the query vector (generalization intent vector) and the neighbors of the current node in the high-level graph, and selects the node with the highest cosine similarity to update the current node. If there is no closer neighbor node in the high-level graph, it descends to the next level of the graph and repeats the greedy search with the termination node of the high-level graph as the new entry point. After descending to the bottom level graph, it performs an exact search in the bottom level graph, maintains the k nodes with the highest cosine similarity through a dynamic candidate queue, and finally returns the node with the highest cosine similarity as the starting node.

[0048] HNSW is used to find the next-hop node that is most similar to the cross-domain candidate knowledge fragment among the neighbors of the starting node. Based on the next-hop node, HNSW is used cyclically to find neighbor nodes until a multi-domain knowledge fragment sequence is established. Furthermore, the greedy search strategy starts from the highest-level entry point, calculates the cosine similarity between the generalization intention vector and all neighboring nodes of the starting node at the highest level, and selects the neighboring node with the highest cosine similarity as the next-hop node. Based on the next-hop node, it iterates through its neighboring nodes, recalculates the cosine similarity between the generalization intention vector and its neighboring nodes, and continuously updates the current node as the next-hop node with the highest cosine similarity. If the cosine similarity of all neighboring nodes in the current layer is lower than that of the current node, it descends to the next layer of the graph, using the current node as the new entry point and repeating the greedy search process. After descending layer by layer to the bottom layer of the graph, an exact search is performed in the bottom layer graph: a candidate queue is dynamically maintained to store the k nodes with the highest current cosine similarity, and the neighboring nodes are iterated through based on the next-hop node, and the candidate queue is updated. The search terminates at the local minimum node in the underlying graph. The final output node set must meet the following conditions: the cosine similarity between all nodes and the generalization intent vector is strictly greater than the preset cluster boundary similarity threshold (usually ranging from 0.5 to 0.8); the node set is sorted from high to low according to the cosine similarity value; the node set meets the relevance judgment criteria of the cross-domain candidate knowledge fragment set, forming a multi-domain knowledge fragment sequence.

[0049] Sort the sequence of knowledge fragments from multiple domains in descending order of cosine similarity to generate a cross-domain knowledge set.

[0050] S5. Use variational Bayesian uncertainty estimation to scan knowledge fragments in cross-domain knowledge sets and block illegal content, and combine personalized weight parameters to generate personalized cross-domain answers.

[0051] Based on cross-domain knowledge sets, Bayesian uncertainty estimation is used to filter knowledge fragments with confidence levels below the Bayesian compliance confidence boundary, thereby generating compliance knowledge fragments. Furthermore, the Bayesian compliance confidence score is calculated based on the features of each knowledge fragment in the cross-domain knowledge set; the variance of the posterior probability distribution of each knowledge fragment is evaluated using the Bayesian uncertainty estimation method; the Bayesian compliance confidence score is compared with the preset Bayesian compliance confidence boundary (usually ranging from 0.85 to 0.95); all knowledge fragments with a Bayesian compliance confidence score lower than the Bayesian compliance confidence boundary are filtered out; the remaining knowledge fragments constitute compliant knowledge fragments.

[0052] It should be noted that the Bayesian compliance confidence boundary is dynamically set based on industry regulatory requirements, corporate risk tolerance, and historical posterior distribution characteristics.

[0053] A weighted knowledge vector is generated based on compliance knowledge fragments combined with personalized weight parameters; Furthermore, personalized weight parameters are assigned weight coefficients to each compliance knowledge fragment. First, the identifier of the compliance knowledge fragment is parsed, and a matching item is retrieved in the personalized weight parameter matrix based on the identifier to obtain the scalar weight coefficient value corresponding to the identifier. The scalar weight coefficient value is then uniquely assigned to the specified compliance knowledge fragment. The compliance knowledge fragment is then input into the XLM-RoBERTa model. The XLM-RoBERTa model encodes the input content through multi-layer neural network transformation, generating a dense vector representation with fixed dimensions. The dense vector captures the semantic features and logical relationships of the compliance knowledge fragment, forming a numerical representation. Finally, L2 normalization is performed on the numerical dense vector to output a standardized basic knowledge vector. The basic knowledge vector is multiplied element-wise with the corresponding weight coefficient in the personalized weight parameters to obtain a weighted knowledge vector. All weighted knowledge vectors are summed element-wise to generate the final weighted knowledge vector.

[0054] Furthermore, the LM-RoBERTa model is trained using a large-scale cross-lingual parallel corpus covering 100 languages. It achieves cross-lingual semantic alignment through a self-supervised learning framework: First, dynamic masking language modeling is used to randomly mask 15% of the words in the input text. The model predicts the masked words based on the context and optimizes the cross-entropy loss. At the same time, unsupervised cross-lingual contrastive learning is combined to force the similar semantic texts in different languages ​​to minimize the distance in the vector space. The full-word masking strategy is further applied to improve the ability to capture complete semantics. Finally, the sentence order of the document is shuffled through a sentence rearrangement task and the model is trained to restore the original order to enhance long-distance dependency modeling, and finally generate knowledge vectors.

[0055] Based on weighted knowledge vectors, a diverse candidate answer framework is constructed by generating semantic variants through GAN, and a bundle search decoding and source tagging are used to finally output personalized cross-domain answers. Furthermore, the weighted knowledge vector serves as input to the generative adversarial network (GAN), driving the generator to produce semantic variant vectors. The discriminator evaluates the rationality of the semantic variant vectors and optimizes the generation quality through adversarial training. After diversity screening, the generated semantic variants constitute diverse candidate answer frameworks. A beam search decoding process is employed to handle these diverse candidate answer frameworks. A candidate sequence pool with a beam search width of k is initialized. Based on multiple semantic path vectors within the diverse candidate answer frameworks, the Transformer decoder generates lexical probability distributions step-by-step. The k partial sequences with the highest sum of lexical probabilities are retained, and low-probability branches are pruned. Source tagging records the weighted knowledge vector indices corresponding to the generated lexical terms and the generation path decision points are superimposed. The beam search decoding outputs the complete sequence with the highest probability as the personalized cross-domain answer.

[0056] Furthermore, the generative adversarial network (GAN) uses weighted knowledge vectors as input data and is trained using an adversarial training framework. The weighted knowledge vectors are then input to the generator. The generator processes the input data through three stacked fully connected layers: the input layer performs a linear transformation and then the Leaky ReLU activation function to generate an activation output; the activation output is input to the first hidden layer, performs a linear transformation, and then the Leaky ReLU activation function to generate a new activation output; the new activation output is input to the second hidden layer, performs a linear transformation, and then the Leaky ReLU activation function to generate the hidden layer output; the hidden layer output is input to the output layer, performs a linear transformation, and then the Tanh activation function to generate a semantic variant vector. The weight matrix of each generator layer is constrained using spectral normalization to maintain the Lipschitz constant. The discriminator uses a one-dimensional convolutional structure to process the input data. The process involves: input semantic variant vectors or real samples undergoing convolution in the first convolutional layer, followed by spectral normalization and LeakyReLU activation to generate the first convolutional output; the first convolutional output undergoes convolution in the second convolutional layer, followed by spectral normalization and LeakyReLU activation to generate the second convolutional output; the second convolutional output is then processed by a fully connected layer to generate a confidence score; during adversarial training, the generator minimizes the Wasserstein distance loss function to optimize parameters; the discriminator maximizes the Wasserstein distance loss function and adds a gradient penalty term to optimize parameters; an alternating update strategy is used to train the network: the discriminator parameters are updated five times, followed by the generator parameters once; the Adam optimizer is used to perform parameter update operations; finally, the output semantic variant vectors form a diverse candidate answer framework.

[0057] S6. Based on personalized cross-domain answers and user feedback logs, knowledge feature gradients are integrated through federated distillation, and graph neural networks are combined to continuously evolve the dynamic knowledge graph, forming a cross-domain collaborative AI knowledge base.

[0058] Based on personalized cross-domain answers, semantic feature vectors are extracted through a bidirectional long short-term memory network; Furthermore, personalized cross-domain answers are used as input sequences and fed into a bidirectional long short-term memory (LSTM) network. The LSM network generates forward latent states by processing the sequence word-by-word in the forward layer and backward latent states by processing the sequence word-by-word in the backward layer. The LSM network then concatenates and fuses the forward and backward latent states at each word position to form a complete latent state vector. This complete latent state vector, along with the global semantic information of the sequence, constitutes a semantic feature vector.

[0059] It should be noted that the bidirectional long short-term memory network uses labeled text sequence data and is trained through supervised learning. The labeled text sequence data is input into the bidirectional long short-term memory network. The forward propagation layer processes the input data word by word in sequence, calculating input weights through the input gate, calculating historical memory retention weights through the forget gate, generating new memory states through the candidate memory layer, updating memory content in memory units, and generating forward hidden states through the output gate. The backward propagation layer processes the input data word by word in reverse sequence, generating backward hidden states through the same gating mechanism. The forward and backward hidden states at each word position are concatenated and fused to form a complete hidden state vector. The complete hidden state vector is then input into the network. The connection layer performs a linear transformation to generate predicted scores, which are then activated by the Softmax function to output the probability distribution of word labels. The error between the predicted probability distribution and the true labels is calculated using the cross-entropy loss function. The error gradient is propagated along the time step using the backpropagation algorithm to calculate the gradient of the network parameters. Gradient pruning is applied to limit the gradient norm threshold; if the gradient norm exceeds the set threshold, the gradient value is scaled. The Adam optimizer is used to update the network weight parameters based on the pruned gradient. The optimization process is repeated iteratively until the loss converges. Finally, the complete hidden state vector is output as a semantic feature vector that integrates bidirectional semantics. This vector aggregates the global context information of the sequence to form the semantic feature vector.

[0060] User behavior feature vectors are extracted using a multilayer perceptron based on user feedback logs. Furthermore, after data cleaning, noise removal, and unified structure, the user feedback logs are input into a multilayer perceptron. The number of nodes in the input layer of the multilayer perceptron is strictly matched with the feature dimension. The input layer receives the user feedback logs and then outputs them to the hidden layer. The hidden layer performs a linear transformation through a fully connected layer and applies the ReLU activation function to introduce non-linear feature interactions. The output layer generates user behavior feature vectors through a linear transformation. Further, the multilayer perceptron uses the preprocessed user feedback log data, is trained using supervised learning, optimizes parameters using the backpropagation algorithm, calculates the prediction error using the mean squared error loss function, and applies Dropout regularization to suppress overfitting, finally outputting the user behavior feature vectors.

[0061] Concatenate the semantic feature vector and the user behavior feature vector into a local feature vector; Furthermore, the semantic feature vector and the user behavior feature vector need to be aligned in dimensions to ensure that their feature structures are compatible. Then, through a concatenation operation, all elements of the semantic feature vector are placed in the front and all elements of the user behavior feature vector are placed in the back, generating a local feature vector with dimensions equal to the sum of the two.

[0062] Based on local feature vectors, forward propagation is performed through a fully connected neural network to obtain locally generated logical prediction values. Furthermore, the fully connected neural network uses local feature vector data, is trained through supervised learning, optimizes parameters using the cross-entropy loss function, and combines the Adam optimizer and L2 regularization to suppress overfitting. The local feature vector first enters the input layer, where the number of nodes strictly matches the feature dimension. The local feature vector is then passed to the hidden layer, where the input is weighted and summed using the weight matrix and bias vector, and then processed by the ReLU activation function. The ReLU activation function performs a non-linear mapping on the input values, retaining positive values ​​and setting negative values ​​to zero, thus blocking the transmission of negative signals and introducing non-linear transformation capability. The activation value of the output of the previous layer is passed to the input of the next layer layer by layer. Finally, it reaches the output layer, where the local feature vector is generated.

[0063] The local feature vector is weighted using the KL divergence dynamic weight allocation mechanism, and the logical prediction values ​​are aggregated using an adaptive aggregation operator to obtain the global knowledge vector. Furthermore, based on the KL divergence dynamic weight allocation mechanism, the KL divergence value between each local feature vector and the target distribution is calculated. Based on the KL divergence value, weight coefficients are calculated through exponential normalization, making the weight coefficients negatively correlated with the KL divergence value, thus achieving dynamic weight allocation. Each local feature vector is multiplied by its corresponding weight coefficient to generate a weighted local feature vector. The weighted local feature vector is input into an adaptive aggregation operator, which dynamically adjusts the learning rate by combining first-order moment estimation and second-order moment estimation. The adaptive aggregation operator performs an aggregation operation on the weighted logical prediction values ​​to generate a global knowledge vector.

[0064] The updated entity embedding matrix is ​​obtained by fusing the global knowledge vector with the historical entity embedding matrix using gradient-enhanced graph convolution. Furthermore, the historical entity embedding matrix and the global knowledge vector are input into the gradient-boosted graph convolution. Gradient-Boosted graph convolution is a hybrid architecture combining gradient boosting and graph convolutional networks (GCNs) to improve node representation accuracy through iterative optimization of residual features. The gradient-boosted graph convolution initializes the historical entity embedding matrix as the initial node features. Subsequently, it iteratively optimizes the residual features through multiple layers: in each round, the residual between the current node feature and the target feature is calculated; the residual is input into a single-layer graph convolutional network (GCN) to aggregate neighbor features and output a residual correction; the residual correction is added to the historical features through residual concatenation to generate updated node features; simultaneously, the global knowledge vector is concatenated and fused with the node features in each iteration to ensure global semantic injection; finally, the updated entity embedding matrix is ​​output. The updated entity embedding matrix integrates global knowledge vector information and historical entity embedding matrix information.

[0065] Furthermore, the training of the gradient-enhanced graph convolutional network relies on graph structure data and global knowledge vectors, and adopts a phased iterative optimization strategy for training: first, the initial node embeddings are generated through standard graph convolution, and then in the multi-layer residual optimization stage, the residual between the current node embedding and the true label is calculated in each round. The residual correction amount is output by aggregating the neighbor features using a single-layer graph convolutional network, and the node features are updated through residual connections. At the same time, the global knowledge vector and the node features are concatenated and fused. Finally, the graph convolution weights and residual correction parameters are jointly optimized through gradient backpropagation.

[0066] Calculate the cosine similarity of each entity vector in the updated entity embedding matrix and combine them into a cosine similarity vector; Furthermore, extract all entity vectors from the updated entity embedding matrix; traverse each pair of entity vectors in the updated entity embedding matrix and calculate the cosine similarity of each pair of entity vectors; combine the cosine similarity calculation results of all entity vector pairs in order to form a cosine similarity vector.

[0067] A relation matrix is ​​obtained by dynamically weighting and fusing the historical relation matrix and cosine similarity vector. The relation matrix is ​​then normalized to obtain an updated relation matrix. Furthermore, the historical relation matrix and the cosine similarity vector are fused through dynamic weighted fusion. The dynamic weighted fusion process allocates weight coefficients according to a preset weight ratio (usually ranging from 0 to 1). The weighted historical relation matrix and the weighted cosine similarity vector are then added element-wise to generate a relation matrix. The relation matrix is ​​then normalized. The normalization process uses a maximum-minimum scaling method to linearly transform all elements of the relation matrix to the interval between 0 and 1, generating an updated relation matrix.

[0068] Based on the updated entity embedding matrix and the updated relation matrix, a new symbolic dynamic knowledge graph is constructed by combining graph convolutional networks with residual links, thereby forming a cross-domain collaborative AI knowledge base; Furthermore, the updated entity embedding matrix and the updated relation matrix are input into the graph convolutional network; the graph convolutional network generates a new entity embedding by aggregating neighbor entity features and fusing the updated relation matrix information; the new entity embedding and the updated entity embedding matrix are added together through residual links to generate an enhanced entity embedding; the enhanced entity embedding and the updated relation matrix are combined to form a new symbolic dynamic knowledge graph.

[0069] This embodiment also provides an AI knowledge base construction system for cross-domain collaboration, including: The data acquisition module is used to collect raw data from across domains for preprocessing, extract heterogeneous modal vectors of the same dimension from the model, and obtain a unified cross-modal vector set through attention-weighted fusion. The knowledge graph generation module is used to aggregate entity embedding matrices from cross-modal unified vector sets, combine them with inductive networks to generate relation type matrices, and use the entity embedding matrix and relation type matrix as a symbolic dynamic knowledge graph. The intent analysis module is used to symbolize dynamic knowledge graphs. It uses adversarial neural networks to decouple domain-independent intents and domain feature vectors in user queries, and generates generalized intent vectors and personalized weight parameters. The semantic graph transition module is used to calculate the semantic similarity between the generalized intent vector and the dynamic knowledge graph, and to retrieve and recall highly relevant knowledge fragments from multiple domains to obtain a cross-domain knowledge set. The qualified knowledge generation module is used to scan out knowledge fragments in the cross-domain knowledge set and block illegal content, and generate personalized cross-domain answers by combining personalized weight parameters; The knowledge graph evolution module is used to continuously evolve the dynamic knowledge graph with personalized cross-domain answers and user feedback logs, forming a cross-domain collaborative AI knowledge base.

[0070] This embodiment also provides a computer device applicable to the construction method of AI knowledge base for cross-domain collaboration, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the AI ​​knowledge base construction method for cross-domain collaboration proposed in the above embodiment.

[0071] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0072] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the method for constructing an AI knowledge base for cross-domain collaboration as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0073] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for constructing an AI knowledge base for cross-domain collaboration, characterized in that, Includes the following steps: S1. Collect cross-domain raw data for preprocessing, extract heterogeneous modal vectors at the same dimension, and obtain a unified cross-modal vector set through attention-weighted fusion; S2. An entity embedding matrix is ​​aggregated based on a cross-modal unified vector set, and a relation type matrix is ​​generated by combining an inductive network. The entity embedding matrix and the relation type matrix are used as a symbolic dynamic knowledge graph. S3. Based on symbolic dynamic knowledge graph, adversarial neural network is used to decouple domain-independent intent and domain feature vector in user query to generate generalized intent vector and personalized weight parameters. S4. Calculate the semantic similarity between the generalized intent vector and the dynamic knowledge graph, and retrieve highly relevant knowledge fragments from multiple domains to obtain a cross-domain knowledge set. S5. Scan out knowledge fragments in the cross-domain knowledge set and block illegal content, and generate personalized cross-domain answers by combining personalized weight parameters; S6. Based on personalized cross-domain answers and user feedback logs, the dynamic knowledge graph is continuously evolved to form a cross-domain collaborative AI knowledge base.

2. The method for constructing an AI knowledge base for cross-domain collaboration as described in claim 1, characterized in that, In S1, the cross-domain raw data includes text data, image data, and time-series data corresponding to enterprise internal knowledge documents, business logs, and monitoring data; The preprocessing includes filtering, noise reduction, and normalization of the cross-domain raw data.

3. The method for constructing an AI knowledge base for cross-domain collaboration as described in claim 2, characterized in that, The acquisition of the cross-modal unified vector set includes: Semantic vectors are extracted from the preprocessed cross-domain raw data, and then the semantic vectors are projected and aligned in real time to obtain heterogeneous modal vectors of the same dimension. By utilizing the hybrid head attention dynamic routing mechanism, the similarity scores between heterogeneous modal vectors of the same dimension are calculated and used as weights. The results are then weighted and fused to output a unified cross-modal vector set.

4. The method for constructing an AI knowledge base for cross-domain collaboration as described in claim 1, characterized in that, S2 specifically refers to: Aggregate the cross-modal unified vector set to obtain the entity embedding matrix; By analyzing common neighbor nodes in the entity embedding matrix through inductive network analysis, cross-domain relationship edges are obtained and sparse branch pruning is performed to generate a relationship matrix, which is then linked with the entity embedding matrix to construct a symbolic dynamic knowledge graph.

5. The method for constructing an AI knowledge base for cross-domain collaboration as described in claim 1, characterized in that, S3 specifically refers to: Extract the query semantic vector from the user's original query text, and use the generator of the adversarial neural network to separate the domain-independent intent and domain features; The semantic rationality of the symbolic dynamic knowledge graph is verified by using an adversarial neural network discriminator, and finally the generalized intent vector is obtained through the training mechanism of the adversarial neural network. Personalized weight parameters are generated based on query semantic vectors and user feedback logs.

6. The method for constructing an AI knowledge base for cross-domain collaboration as described in claim 1, characterized in that, S4 specifically refers to: The cosine similarity between the generalization intent vector and the knowledge fragment vectors in the dynamic knowledge graph is calculated and combined into a set of cosine similarity values. Based on the set of cosine similarity values, the knowledge fragments in the dynamic knowledge graph are clustered to obtain a cross-domain candidate knowledge fragment set. The HNSW algorithm is used to build a multi-layer graph for the cross-domain candidate knowledge fragment set. Starting from the high-level graph, the node with the highest cosine similarity to the generalization intention vector is visited as the starting node. Then, the next-hop node with the most similarity to the cross-domain candidate knowledge fragment is found among the neighbors of the starting node. Based on the next-hop node, the HNSW algorithm is used repeatedly to find neighbor nodes until a multi-domain knowledge fragment sequence is established. Sort the sequence of knowledge fragments from multiple domains in descending order of cosine similarity to generate a cross-domain knowledge set.

7. The method for constructing an AI knowledge base for cross-domain collaboration as described in claim 1, characterized in that, S5 specifically refers to: The system scans knowledge fragments in cross-domain knowledge sets, removes illegal knowledge fragments, generates compliant knowledge fragments, and combines personalized weight parameters to generate weighted knowledge vectors. Based on weighted knowledge vectors, a diverse candidate answer framework is constructed by generating semantic variants through GAN. Then, a beam search decoding method is used to superimpose source annotation, and finally, personalized cross-domain answers are output.

8. The method for constructing an AI knowledge base for cross-domain collaboration as described in claim 1, characterized in that, In S6, the dynamic knowledge graph is continuously evolved based on personalized cross-domain answers and user feedback logs, including: Extract semantic feature vectors from users' personalized cross-domain answers; Extract user behavior feature vectors from user feedback logs; The semantic feature vector and the user behavior feature vector are concatenated into a local feature vector, and the logical prediction value is obtained by performing forward propagation through a fully connected neural network. Weights are assigned to local feature vectors, and logical prediction values ​​are aggregated to obtain a global knowledge vector.

9. The method for constructing an AI knowledge base for cross-domain collaboration as described in claim 8, characterized in that, The AI ​​knowledge base constituting cross-domain collaboration includes: The updated entity embedding matrix is ​​obtained by fusing the global knowledge vector with the historical entity embedding matrix. Calculate the cosine similarity of each entity vector in the updated entity embedding matrix and combine them into a cosine similarity vector; The updated relation matrix is ​​obtained by fusing the historical relation matrix and the cosine similarity vector. Based on the updated entity embedding matrix and the updated relation matrix, a new symbolic dynamic knowledge graph is constructed through graph links, thereby forming a cross-domain collaborative AI knowledge base.

10. A cross-domain collaborative AI knowledge base construction system, based on the cross-domain collaborative AI knowledge base construction method according to any one of claims 1 to 9, characterized in that, include, The data acquisition module is used to collect raw data from across domains for preprocessing, extract heterogeneous modal vectors of the same dimension from the model, and obtain a unified cross-modal vector set through attention-weighted fusion. The knowledge graph generation module is used to aggregate entity embedding matrices from cross-modal unified vector sets, combine them with inductive networks to generate relation type matrices, and use the entity embedding matrix and relation type matrix as a symbolic dynamic knowledge graph. The intent analysis module is used to symbolize dynamic knowledge graphs. It uses adversarial neural networks to decouple domain-independent intents and domain feature vectors in user queries, and generates generalized intent vectors and personalized weight parameters. The semantic graph transition module is used to calculate the semantic similarity between the generalized intent vector and the dynamic knowledge graph, and to retrieve and recall highly relevant knowledge fragments from multiple domains to obtain a cross-domain knowledge set. The qualified knowledge generation module is used to scan out knowledge fragments in the cross-domain knowledge set and block illegal content, and generate personalized cross-domain answers by combining personalized weight parameters; The knowledge graph evolution module is used to continuously evolve the dynamic knowledge graph with personalized cross-domain answers and user feedback logs, forming cross-domain collaborative AI knowledge.

Images