Intention-driven network configuration template generation method

Abstract

The invention discloses an intention-driven network configuration template generation method, which is characterized in that automatic translation from a user intention to a network configuration instruction is realized by constructing a comprehensive model fusing a graph neural network and a natural language processing model, and specifically comprises the following steps: S1, constructing a network configuration instruction graph; s2, preparing a data set; s3, finely adjusting the BERT model; s4, instruction embedding generation; s5, intention embedding generation; s6, calculating matching similarity; according to the method, knowledge contained in a network configuration document is mined, an instruction graph capable of representing the relation between network configuration instructions is constructed, an incidence relation data set of a private network service deployment intention and the network configuration instructions is prepared, a graph neural network and a natural language processing pre-training model are fused, and through contrast learning training tuning, the network configuration instructions are optimized. The translation from the private network deployment intention to the network configuration instruction sequence is realized, that is, the network configuration template is automatically generated, and the pressure of a service expert on reasoning a configuration scheme in a new scene is relieved.

Description

Technical Field

[0001] The present invention relates to the technical field of network management, and particularly to an intention-driven network configuration template generation method. Background Art

[0002] In recent years, with the rapid growth of the number of mobile devices, the continuous enrichment of media content, and the increasing demand for high-quality data, wireless data traffic has surged. The explosive development of telecommunications services has made the current network infrastructure denser and more complex. Achieving efficient network management and control has become a key problem in the future network development. In this process, operators need to handle a large number of tasks such as parameter configuration, resource allocation, and security policy setting to ensure the stable operation of services and network security. However, due to the complexity of 5G networks and the diversity of business requirements, service provisioning often faces challenges of tight time and heavy tasks. Therefore, efficient management and control strategies and technical means are needed to support.

[0003] However, in terms of translating user intentions into network configurations, it is necessary to accurately analyze the business requirements of customers and manually write network configuration scripts that meet these requirements. This not only requires extremely high professional skills, but also human errors may affect the service quality. At the same time, intention translation templates and state machines are mostly limited to specific business scenarios. With the diversification of customer needs and the rapid development of network technologies, network devices, business documents, and business types have become increasingly complex. Manually handling these complex and changeable configuration requirements has become increasingly difficult, which not only raises the professional threshold requirements for business experts, but also limits the flexibility and scalability of the system. To solve the above problems, we propose an intention-driven network configuration template generation method. Summary of the Invention

[0004] In view of the deficiencies of the prior art, the present invention provides an intention-driven network configuration template generation method to solve the problems raised in the background art.

[0005] The above technical objectives of the present invention are achieved through the following technical solutions:

[0006] An intention-driven network configuration template generation method, which realizes the automatic translation of user intentions into network configuration instructions by constructing a comprehensive model that integrates a graph neural network and a natural language processing model, specifically including the following steps:

[0007] S1. Construction of a network configuration instruction graph: Model the instruction reference part of the network configuration manual into an instruction graph, including instruction nodes, directory nodes, and parameter nodes, and capture the shared parameters, inclusion relationships, and precedence relationships between the nodes;

[0008] S2. Dataset Preparation: Based on the scenario deployment section in the network configuration manual, construct a "private network deployment intention - network configuration instruction" dataset for training the model to achieve the translation from intention to instruction;

[0009] S3. BERT Model Fine - Tuning: Fine - tune the BERT Chinese pre - trained model and enhance the model's intention understanding ability using the professional terms in the network configuration document;

[0010] S4. Instruction Embedding Generation: Use the GraphSAGE graph neural network model to generate the embedding features of instruction nodes, and capture the complex dependencies between nodes by aggregating the neighbor node information in the instruction graph;

[0011] S5. Intention Embedding Generation: Through the fine - tuned BERT model, convert the natural - language intention description input by the user into an intention embedding vector;

[0012] S6. Matching Similarity Calculation: According to the similarity evaluation model of intention embedding and instruction embedding, calculate the matching degree between the two, and generate a matching configuration instruction template.

[0013] Preferably, the instruction graph construction step includes the following modeling of nodes and edges:

[0014] a) Nodes: Include directory nodes, instruction nodes, and parameter nodes. Among them, the directory node corresponds to the directory structure to which the instruction belongs, the instruction node corresponds to the network configuration instruction, and the parameter node represents the input and output parameters of the instruction;

[0015] b) Edges: Include the associated edges between instruction nodes and parameter nodes, the inclusion edges between directory nodes, the inclusion edges between directory nodes and instruction nodes, and the directed edges representing the precedence relationship of instructions.

[0016] Preferably, in the dataset preparation step, the network configuration intention of each scenario is associated with the corresponding instruction sequence for the training and verification of the intention translation model.

[0017] Preferably, the GraphSAGE graph neural network model uses the local neighbor features of instruction nodes and generates high - quality instruction embedding representations containing context dependencies by iteratively aggregating neighbor node features.

[0018] Preferably, the intention embedding generation step optimizes the intention representation through the masked language model task of the BERT model to enhance the model's understanding ability of the language and professional terms in the network configuration document, thereby generating semantically rich intention embeddings.

[0019] Preferably, in the step of calculating the matching similarity, the cosine similarity or multi-layer perceptron method based on the intent embedding and the instruction embedding is used to quantify the fit between the user intent and the specific network configuration instruction.

[0020] Preferably, by selecting InfoNCE (InfoNCE) or binary cross entropy (BCEWithLogit) as the loss function, the matching accuracy of the model for the user intent is optimized during the training process.

[0021] Preferably, after the model training is completed, based on the intent description input by the user, a set of network configuration instructions matching the intent can be automatically generated to assist network administrators in efficiently deploying network configurations.

[0022] In summary, the present invention mainly has the following beneficial effects:

[0023] 1. By mining the knowledge contained in the network configuration documents, the present invention constructs an instruction graph that can represent the relationships between network configuration instructions, prepares a correlation dataset between the private network service deployment intent and the network configuration instructions. Through automation technology, this intelligent model can quickly generate a configuration script template, reducing the time for business experts to manually write and modify scripts. This high efficiency not only reduces the difficulty of service activation, but also reduces labor costs and improves the speed of service activation.

[0024] 2. The present invention integrates a graph neural network and a natural language processing pre-training model, and through contrastive learning training and optimization, realizes the translation from the private network deployment intent to the network configuration instruction sequence, that is, automatically generates a network configuration template, and can input a new scenario description and output the instruction sequence corresponding to the new scenario, reducing the pressure on business experts to infer configuration solutions in new scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a schematic overview of the training stage of the technical solution of the present invention;

[0026] Figure 2 is a schematic diagram of the composition method of the instruction graph of the present invention;

[0027] Figure 3 is a schematic diagram of the calculation process of the BERT model of the present invention;

[0028] Figure 4 is a schematic diagram of the calculation process of the GraphSAGE model of the present invention;

[0029] Figure 5 is a schematic diagram of the model inference application of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions of the embodiments of the present invention in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

[0031] The following embodiments are used to illustrate the present invention, but cannot be used to limit the scope of protection of the present invention. The conditions in the embodiments can be further adjusted according to specific conditions. Any simple improvement to the method of the present invention under the premise of the concept of the present invention shall fall within the scope of protection required by the present invention.

[0032] Reference Figures 1 - 5 , an intention-driven network configuration template generation method, which realizes the automatic translation from user intention to network configuration instructions by constructing a comprehensive model that integrates a graph neural network and a natural language processing model, specifically including the following steps:

[0033] S1. Construction of the network configuration instruction graph: Model the instruction reference part of the network configuration manual into an instruction graph, including instruction nodes, directory nodes, and parameter nodes, and capture the shared parameters, inclusion relationships, and precedence relationships between the nodes;

[0034] S2. Preparation of the dataset: Based on the scenario deployment part in the network configuration manual, construct a "private network deployment intention - network configuration instruction" dataset for training the model to achieve the translation from intention to instruction;

[0035] S3. Fine-tuning of the BERT model: Fine-tune the pre-trained BERT Chinese model, and use the professional terms in the network configuration document to enhance the intention understanding ability of the model;

[0036] S4. Generation of instruction embeddings: Use the GraphSAGE graph neural network model to generate the embedding features of the instruction nodes, and capture the complex dependencies between the nodes by aggregating the neighbor node information in the instruction graph;

[0037] S5. Generation of intention embeddings: Through the fine-tuned BERT model, convert the natural language intention description input by the user into an intention embedding vector;

[0038] S6. Calculation of the matching similarity: According to the similarity evaluation model of the intention embedding and the instruction embedding, calculate the matching degree between the two to generate a matching configuration instruction template.

[0039] Reference Figures 1 - 5 , the modeling of the following nodes and edges is included in the instruction graph construction step:

[0040] a) Nodes: including directory nodes, instruction nodes, and parameter nodes, where the directory nodes correspond to the directory structure to which the instruction belongs, the instruction nodes correspond to network configuration instructions, and the parameter nodes represent the input and output parameters of the instructions;

[0041] b) Edges: including the associated edges between instruction nodes and parameter nodes, the inclusion edges between directory nodes, the inclusion edges between directory nodes and instruction nodes, and the directed edges representing the precedence relationship of instructions;

[0042] Among them, the instruction reference part can be modeled into a graph for the subsequent training of the graph neural network model. The instruction graph can be mathematically modeled as follows: Let the instruction graph G be composed of a node set V and an edge set Composition

[0043]

[0044] The node set V contains three types of nodes, directory nodes V Directory , instruction nodes V Instruction and parameter nodes V Paramter ;

[0045] V = V Directory ∪V Instruction ∪V Paramter

[0046] Each node has a name attribute. The name attribute of the directory node V Directory is the name of the directory, the name attribute of the instruction node V Instruction is the name of the instruction, and the name attribute of the parameter node V Paramter is the name of the parameter;

[0047] The node set contains four types of edges, namely and and and Four types of edges;

[0048]

[0049] is the edge from the instruction node V Instruction to the parameter node V Paramter , representing that the instruction has this parameter; for example, in the following formula, instruction i has parameter p;

[0050]

[0051] is the edge from the directory node V Directory to the directory node V DirectoryThe edge represents the inclusion relationship of the directory structure; for example, in the following formula, directory i is the parent directory of directory p;

[0052]

[0053] is the directory node V Directory The edge pointing to the instruction node V Instruction represents the instructions included in a directory. For example, in the following formula, directory i is the parent directory of directory p. To reduce the number of edges, only the leaf directory points to the instruction node. For example, in the following formula, directory i includes instruction p;

[0054]

[0055] is the directory node V Instruction The edge pointing to the instruction node V Instruction is used to represent the pre-order relationship of instructions. In the following formula, instruction i appears before instruction p, and there is a parameter sharing relationship between instruction i and instruction p;

[0056]

[0057] Reference Figures 1 - 5 , in the data set preparation step, the network configuration intention of each scenario is associated with the corresponding instruction sequence for the training and verification of the intention translation model. The GraphSAGE graph neural network model uses the local neighbor features of the instruction nodes and generates high-quality instruction embedding representations containing context dependencies by iteratively aggregating the neighbor node features;

[0058] Among them, the BERT model is a natural language processing model based on the Transformer architecture, which captures long-distance dependencies through the self-attention mechanism;

[0059] In the present invention, assume there is an input sequence x = [x 1 , x 2 , …, x n , where n is the length of the sequence. In the MLM task, we randomly select some words in the sequence for masking. The specific process is as follows: First, we select each word x i in the sequence for masking with a probability of 15%. For the selected word x i , there are the following several

[0060] With an 80% probability, replace x i with the special '[MASK]' token, that is

[0061] With a 10% probability, replace x iRandomly replace with other word x rand , that is

[0062] There is a 10% probability of keeping x i unchanged, that is

[0063] Let x masked represent the processed sequence, which includes the masked words;

[0064] Next, the model M will predict the masked words based on the context. Specifically, what the model outputs is the probability distribution P(x i |x masked ) of all words at each position. To train the model, we minimize the following loss function:

[0065]

[0066] To improve the graph model's understanding and representation ability of instruction nodes, the instruction reference part and scenario deployment part of the network configuration manual are adopted, and the BERT Chinese model is fine-tuned through the masked language model (MLM) task. The resulting model is denoted as M fine-tuned , this fine-tuning not only makes the BERT model more adaptable to the language style and professional terms of the network configuration manual, but also provides high-quality initial features of node embeddings for the subsequent graph model. These features can more accurately capture the semantic relationships and context information between instructions, thereby enhancing the graph model's representation ability of instruction nodes. In addition, the fine-tuned BERT model can also output accurate intent embeddings for user intents, which helps to better understand user needs and achieve an accurate match between user intents and instruction nodes. In this way, richer information is provided for the graph model, further improving the performance and practicality of the model.

[0067] Reference Figures 1 - 5 , the intent embedding generation step optimizes the intent representation through the masked language model task of the BERT model to enhance the model's understanding ability of the language and professional terms of the network configuration document, thereby generating semantically rich intent embeddings;

[0068] Among them, it is necessary to use the graph model to capture the complex graph structure relationships of instructions in the network configuration manual through the input instruction graph, and in this way, the model's understanding and utilization of the complex relationships between instructions can be enhanced, and semantic embeddings are output for each instruction. The graph model is suitable for processing this complex graph structure. By iteratively aggregating the information of neighbor nodes, the graph model can gradually capture and learn the multi-level relationships between nodes from local to global. This method ensures that the model can understand the direct and indirect relationships between different instructions, thereby better mapping the structural information in the manual.

[0069] In the instruction graph, the fine-tuned BERT model is used to generate initial language embedding features for each instruction on the graph. Specifically, a string of text, which is a concatenation of the Chinese name, English name and functional description of each instruction, parameter and directory, is input into the fine-tuned BERT model to obtain its corresponding language embedding. These language embeddings are used as the initial features of each node in the graph. Set (I) is denoted as the complete set of instructions. The initial features are defined as , u∈V, then:

[0070]

[0071] GraphSAGE is an effective graph neural network model, which is mainly used for node representation learning of large-scale graph data. In this study, the GraphSAGE model is adapted to the instruction graph to generate instruction embedding from the instruction graph. Some key technical details of GraphSAGE are:

[0072] 1. Sampling neighbors: GraphSAGE uses a fixed-size neighbor sampling strategy, which means that for each node in the graph, the model randomly selects a fixed number of neighbor nodes from its neighbors. This method reduces the amount of computation and can handle the degree differences of nodes that are common in real-world graphs;

[0073] 2. Neighbor aggregation: GraphSAGE can introduce a variety of aggregation functions (such as mean, pooling, LSTM, etc.) to aggregate the features of a node's neighbor nodes. The purpose of these aggregation functions is to effectively integrate the information of neighbor nodes into the features of the current node, thereby learning the comprehensive expression of the node;

[0074]

[0075] in, is the neighbor set of node v, W k is the weight matrix of the kth layer, is the feature representation of node u at the k-1 layer, σ is a nonlinear activation function, and the mean aggregation function is used here to aggregate the neighbor node features of a node;

[0076] 3. Node update: In each iteration, the new features of the node are updated by combining its own features with the aggregated neighbor features. This update process usually involves nonlinear activation functions.

[0077]

[0078] in is the feature representation of node v at the k-1 layer, || represents the connection of vectors, W'k Another weight is used for updating the nodes of the current layer. After outputting the node embeddings of the last k-layer network, the node embeddings of the type of instruction are extracted as the final embeddings of the instructions. Let each instruction embedding vector be p i , i ∈ set(I), then we have:

[0079]

[0080] Reference Figures 1 - 5 , in the step of calculating the matching similarity, based on the cosine similarity between the intent embedding and the instruction embedding or the multi-layer perceptron method, the degree of fit between the user intent and the specific network configuration instruction is quantified;

[0081] Among them, after the fine-tuning of the BERT model is completed, using its enhanced representation ability for business professional terms after fine-tuning, deep embedding vectors are generated for each user intent. These intent embeddings are obtained by inputting the intent description text into the fine-tuned BERT Chinese model. The model will capture the semantic information in the text and encode it into vectors of a fixed dimension. These embedding vectors not only accurately reflect the meaning of the intent in the semantic space, but also provide a basis for subsequent intent recognition and instruction matching;

[0082] There is already a fine-tuned BERT Chinese model M fine-tund and a set of string texts of user intent descriptions, denoted as intent. For each intent description text intent, the model will generate a corresponding intent embedding vector, denoted as e. The specific formula is as follows:

[0083] e = M fine-tuned (intent),

[0084] These embedding vectors have rich semantic information. Because the fine-tuned BERT model has been optimized and trained on a corpus in a specific domain, it can better understand and represent the language and terms in the professional field. The intent embedding vectors generated in this way can be used in a variety of downstream tasks, such as intent recognition, text classification, and instruction matching. In the intent recognition task, these embedding vectors can be used as input features to train a classification model to accurately distinguish different user intents. In the instruction matching task, the intent embedding vectors can help find the instruction that best matches the user intent, so as to achieve efficient instruction execution and meet user needs. Therefore, the intent embedding vectors generated by the fine-tuned BERT Chinese model not only accurately reflect the meaning of the intent in the semantic space, but also provide a solid foundation for subsequent intent recognition and instruction matching, further improving the performance and practicality of the system;

[0085] After generating the instruction embedding and the intent embedding, it is necessary to quantify the matching degree between the two. To calculate this matching similarity, various methods can be used, such as the scaled cosine similarity or the method of multi-layer perceptron. By calculating the low matching score between the instruction embedding and the intent embedding, a similarity between 0 and 1 can be obtained. This similarity intuitively reflects the fit between the user's intent and the specific instruction: the closer the similarity is to 1, the higher the match between the user's intent and the instruction; on the contrary, the lower the similarity, the worse the match between the two.

[0086] In the following formula, let intent represent a text of intent description, e be the intent description embedding of intent, and let the set of all instructions be set(I), p i can be used as the embedding of each instruction. The calculation formula of the scaled cosine similarity ConsineSimilarity between the intent embedding and the instruction embedding is as follows:

[0087]

[0088] In the above formula, f(x)·g(x) represents the dot product of the vectors of f(x) and g(y), and ||f(x)|| and ||g(y)|| represent taking the modulus of the vectors of f(x) and g(y);

[0089] If the method of multi-layer perceptron is used, the intent embedding and the instruction embedding can be concatenated and input into the multi-layer perceptron, and finally a similarity between 0 and 1 is output, denoted as MLPScore. Its calculation formula is as follows

[0090] MLPScore(x,y) = σ(W n φ(…φ(W 2 φ(W 1 [e;p i +b 1 )+b 2 )…)+b n ), i ∈ set(I)

[0091] In the above formula, [f(x);g(y)] is the vector concatenation of the intent embedding f(x) and the instruction embedding g(y), W 1 , W 2 ,…, W n are the weight matrices of each layer, b 1 , b 2 ,…, b n are the bias vectors of each layer, φ is a non-linear activation function, here the ReLU function is taken, and σ is the activation function of the output layer, here the Sigmoid function is taken to ensure that the output value is between 0 and 1, representing the matching similarity.

[0092] ReferenceFigures 1 - 5 , by selecting Information Noise Contrastive Estimation (InfoNCE) or Binary Cross Entropy (BCEWithLogit) as the loss function, the matching accuracy of the model to the user's intention is optimized during the training process;

[0093] Among them, for the selection of the loss function and model training, a loss function needs to be selected to calculate the loss for training the model, so that the model can more accurately match the instructions corresponding to the intention. Two methods, namely the InfoNCE loss function or the BCE loss function, can be used;

[0094] After the model training is completed, based on the intention description input by the user, a set of network configuration instructions matching the intention can be automatically generated to assist network administrators in efficiently deploying network configurations;

[0095] After the model training is completed, the model inference application can be entered. In the inference stage, when the user inputs the network configuration intention, the model can output a set of configuration instructions related to the intention. Further, by using the already constructed instruction graph, subgraphs can be extracted from the output instruction set and then converted into the UML diagram of the expert model, which is convenient for network management experts to further implement network configurations.

[0096] Working principle: Please refer to Figures 1 - 5 As shown, the present invention constructs an instruction graph that can represent the relationship between network configuration instructions by mining the knowledge contained in network configuration documents, prepares a dataset of the association relationship between private network service deployment intentions and network configuration instructions. On this basis, the present invention integrates a graph neural network and a natural language processing pre-trained model, and through contrastive learning training and optimization, realizes the translation of private network deployment intentions into network configuration instruction sequences, that is, automatically generates a network configuration template, reducing the pressure on business experts in reasoning about configuration solutions in new scenarios.

[0097] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intent-driven network configuration template generation method, characterized in that: By building a comprehensive model that integrates graph neural networks and natural language processing models, the automatic translation of user intent to network configuration instructions is achieved, which includes the following steps: S1. Network configuration instruction graph construction: Model the instruction reference part of the network configuration manual into an instruction graph, including instruction nodes, directory nodes, and parameter nodes, and capture the shared parameters, inclusion relationships, and precedence relationships between nodes; S2. Dataset preparation: Based on the scenario deployment part in the network configuration manual, a "private network deployment intention-network configuration instruction" dataset is constructed to train the model to achieve the translation from intention to instruction; S3. BERT model fine-tuning: Fine-tune the BERT Chinese pre-trained model and use the professional terms in the network configuration document to enhance the model's intent understanding ability; S4. Instruction embedding generation: Using the GraphSAGE graph neural network model, we generate embedding features for instruction nodes and capture the complex dependencies between nodes by aggregating neighbor node information in the instruction graph. S5. Intent embedding generation: The natural language intent description entered by the user is converted into an intent embedding vector through the fine-tuned BERT model; S6. Matching similarity calculation: Based on the similarity evaluation model of intent embedding and instruction embedding, the matching degree between the two is calculated to generate a matching configuration instruction template.

2. The method for generating an intent-driven network configuration template according to claim 1, characterized in that: The instruction graph construction step includes the modeling of the following nodes and edges: a) Node: including directory nodes, instruction nodes and parameter nodes. The directory node corresponds to the directory structure to which the instruction belongs, the instruction node corresponds to the network configuration instruction, and the parameter node represents the input and output parameters of the instruction; b) Edges: including association edges between instruction nodes and parameter nodes, inclusion edges between directory nodes, inclusion edges between directory nodes and instruction nodes, and directed edges representing instruction precedence relationships.

3. The method for generating an intent-driven network configuration template according to claim 1, characterized in that: In the data set preparation step, the network configuration intent of each scene is associated with the corresponding instruction sequence for training and verification of the intent translation model.

4. The method for generating an intent-driven network configuration template according to claim 1, characterized in that: The GraphSAGE graph neural network model utilizes the local neighbor features of instruction nodes and generates high-quality instruction embedding representations containing contextual dependencies by iteratively aggregating neighbor node features.

5. The method for generating an intent-driven network configuration template according to claim 1, characterized in that: The intent embedding generation step optimizes the intent representation through the masked language model task of the BERT model to enhance the model's ability to understand the language and professional terms of the network configuration document, thereby generating semantically rich intent embeddings.

6. The method for generating an intent-driven network configuration template according to claim 1, characterized in that: In the matching similarity calculation step, the degree of fit between the user intention and the specific network configuration instruction is quantified based on the cosine similarity or multi-layer perceptron method of intent embedding and instruction embedding.

7. The method for generating an intent-driven network configuration template according to claim 1, characterized in that: By selecting Information Noise Contrastive Estimation (InfoNCE) or Binary Cross Entropy (BCEWithLogit) as the loss function, the model's matching accuracy to user intent is optimized during the training process.

8. The method for generating an intent-driven network configuration template according to claim 1, characterized in that: After the model training is completed, a set of network configuration instructions matching the intent can be automatically generated based on the intent description input by the user, thereby assisting network administrators in efficiently deploying network configurations.

Images