Expert model construction method and device based on network configuration manual

Abstract

The invention discloses an expert model construction method and device based on a network configuration manual, and the key points of the technical scheme are that the method comprises the following steps: a data processing step; a pre-training step; an augmented text generation step; an instruction clustering step; entity connection: on the basis of a clustering result, performing unified management on repeated parameters among the instructions by adopting rule matching and parameter reference technologies; through the steps of automatic data processing, instruction clustering, entity connection and the like, the instructions are extracted from the network configuration manual and the expert model is generated, so that the steps of manually looking up the manual and configuring one by one are reduced, information and relationships can be automatically extracted from newly added network configuration data, a more flexible solution is provided, and the user experience is improved. In the face of a new service scene or a configuration requirement, the configuration knowledge base can be automatically updated, manual configuration does not need to be started from the beginning, the flexibility and the strain capacity of network configuration are improved, and a network administrator can more easily cope with a complex service scene.

Description

Technical Field

[0001] The present invention relates to the technical field of network configuration management, and in particular to a method and device for constructing an expert model based on a network configuration manual. Background Art

[0002] In recent years, with the rapid increase in the number of mobile devices and the surge in demand for high-quality data, telecommunication services have developed rapidly, resulting in a surge in wireless data traffic and increasingly complex network infrastructure. In order to adapt to this trend, efficient network management and configuration have become particularly important. However, in the current network environment, even with the help of scripting tools, service activation and network management still require a lot of manual operations, especially in complex business scenarios. There is a close working relationship between network devices and network elements, and the configuration process usually requires the coordination of various devices. Therefore, when configuring the network, it is necessary to manually consult a large number of technical manuals to complete parameter configuration, making the service activation process cumbersome and time-consuming.

[0003] To address this problem, academia and industry have proposed some exploratory solutions for intelligent network management in recent years, such as intent-driven networking (IBN). This method allows users to describe the required network goals by providing an intent interface without having to go into detailed implementation details. Intent-driven networking achieves intelligent management by automating the conversion, verification, deployment, configuration, and optimization of network status. However, in order to ensure the accuracy of intent conversion and configuration operations, intent-driven networking requires a lot of domain knowledge support.

[0004] However, in the existing network configuration management, configuration tasks mainly rely on manual review and execution, and technicians need to manually extract knowledge from documents and configure instructions. This method is easier to implement when the network is simple, but as the network environment becomes increasingly complex, manual configuration methods face obvious limitations. The manual review and configuration process is time-consuming and labor-intensive, and it is easy to cause network failures due to human errors. Existing methods rely too much on the experience and skills of individual technicians, and knowledge sharing and team collaboration are limited, which is not conducive to knowledge accumulation and inheritance. To solve the above problems, we propose an expert model construction method and device based on the network configuration manual. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a method and device for constructing an expert model based on a network configuration manual, which improves the efficiency and accuracy of network management and reduces the errors in manual configuration by automatically extracting and organizing configuration instructions. At the same time, through a structured and visual knowledge management method, configuration management in a complex network environment is made more intuitive, more efficient, and more adaptable, thereby solving the problems raised in the background technology.

[0006] The above technical objectives of the present invention are achieved through the following technical solutions:

[0007] The expert model construction method based on the network configuration manual includes the following steps:

[0008] Data processing steps: Parse the configuration instructions in the network configuration manual and use regular expressions to extract instruction names, parameter names, and parameter values ​​to generate preliminary semi-structured data;

[0009] Pre-training step: using the pre-trained BERT model to process the semi-structured data, generating semantic embeddings for each configuration instruction using a masked language model task to capture the semantic relationship between instructions;

[0010] Augmented text generation step: Generate augmented text based on semantic replacement to obtain several positive and negative example texts for each instruction for subsequent comparative learning;

[0011] Instruction clustering step: Using the instruction-cluster model, the augmented text is clustered through contrastive learning to summarize the association between instructions and obtain the reference relationship and logical order of configuration parameters;

[0012] Entity connection step: Based on the clustering results, rule matching and parameter reference technology are used to uniformly manage the repeated parameters between instructions, convert related instruction parameters into shared parameters, and generate an expert model diagram of the network configuration.

[0013] Preferably, the regular expression used in the data processing step is used to identify the specific format of the configuration instruction and generate structured data including instruction name, parameter name and parameter value.

[0014] Preferably, in the pre-training step, the BERT model generates instruction embedding representations based on a network management term set and an instruction set as pre-training data using a masked language model task.

[0015] Preferably, the augmented text generation step includes the following process:

[0016] Generate two augmented texts as positive example texts of the same configuration instruction by semantic replacement;

[0017] Other augmented texts are used as negative example texts for subsequent instruction contrastive learning tasks.

[0018] Preferably, the instruction clustering step adopts a Kmeans-based contrastive learning clustering algorithm, performs cluster analysis on the generated embedded representation based on the augmented text of the instruction, and identifies the configuration order between mutually related instructions.

[0019] Preferably, the entity connection step sets instruction parameters with the same parameters as shared parameters based on a rule matching method to achieve automatic filling of configuration parameters and reference between instructions.

[0020] The present invention also provides an expert model construction device based on a network configuration manual, comprising:

[0021] A data processing module is used to parse the configuration instructions in the network configuration manual, extract the instruction name, parameter name and parameter value using regular expressions, and generate preliminary configuration information containing structured data;

[0022] A pre-training module, connected to the data processing module, performs a masked language model task based on a BERT model to generate an embedded representation of configuration instructions to capture the semantic relationship between instructions;

[0023] An augmented text generation module, connected to the pre-training module, for generating positive and negative example texts for each configuration instruction to support contrastive learning clustering tasks;

[0024] An instruction clustering module, connected to the augmented text generation module, performs cluster analysis on instruction embedding based on the instruction-cluster model, and summarizes the logical association relationship and execution order between instructions;

[0025] The entity connection module is connected to the instruction clustering module, performs sharing processing on instruction parameters based on rule matching and parameter reference, and generates an expert model diagram of configuration parameters.

[0026] Preferably, the data processing module uses regular expressions to parse the network configuration manual, extract structured data of the instructions and remove redundant information.

[0027] Preferably, the augmented text generation module generates two augmented texts as positive example texts and other augmented texts as negative example texts for comparative learning and clustering.

[0028] Preferably, the entity connection module realizes unified management of multiple instruction parameters by referencing shared parameters, and finally generates an expert model diagram of the configuration parameters.

[0029] In summary, the present invention mainly has the following beneficial effects:

[0030] 1. Through automated data processing, instruction clustering, and entity connection, instructions are extracted from the network configuration manual and an expert model is generated. This automated process reduces the steps of manually consulting the manual and configuring one by one, greatly improving configuration efficiency. The automated process also avoids omissions and errors in manual operations, and reduces the impact of human errors on network configuration stability through clustering and parameter references.

[0031] 2. The present invention structures the instructions in the originally lengthy and difficult-to-interpret configuration manual into a visual model diagram. This structuring process not only improves the readability of the information, but also makes it easier for network administrators to understand and master the complex network configuration process. The structuring and visualization of configuration knowledge makes the network configuration process more intuitive, supports the accumulation and sharing of knowledge, and thus facilitates teamwork and knowledge inheritance.

[0032] 3. The BERT model and instruction-cluster clustering technology can automatically extract information and relationships from the newly added network configuration data. This dynamic adaptability can respond to changing business needs and network conditions in real time, providing more flexible solutions. When facing new business scenarios or configuration requirements, the configuration knowledge base can be automatically updated without manual configuration from scratch, which improves the flexibility and adaptability of network configuration. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flow chart of the model building method in the embodiment of the present invention;

[0034] Figure 2 It is a flow chart of the relationship extraction of the instruction-cluster clustering model of the model building method in the embodiment of the present invention;

[0035] Figure 3 It is a structural frame view of the model building device in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings of the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the scope of protection of the present invention.

[0037] The following examples are used to illustrate the present invention, but they cannot be used to limit the scope of protection of the present invention. The conditions in the examples can be further adjusted according to specific conditions. Simple improvements to the method of the present invention under the premise of the concept of the present invention belong to the scope of protection claimed in the present invention.

[0038] refer to Figure 1-Figure 3 ,The expert model building method based on the network configuration manual includes the following steps:

[0039] Data processing steps: Parse the configuration instructions in the network configuration manual and use regular expressions to extract instruction names, parameter names, and parameter values ​​to generate preliminary semi-structured data;

[0040] Pre-training step: using the pre-trained BERT model to process the semi-structured data, generating semantic embeddings for each configuration instruction using a masked language model task to capture the semantic relationship between instructions;

[0041] Augmented text generation step: Generate augmented text based on semantic replacement to obtain several positive and negative example texts for each instruction for subsequent comparative learning;

[0042] Instruction clustering step: Using the instruction-cluster model, the augmented text is clustered through contrastive learning to summarize the association between instructions and obtain the reference relationship and logical order of configuration parameters;

[0043] Entity connection step: Based on the clustering results, rule matching and parameter reference technology are used to uniformly manage the repeated parameters between instructions, convert related instruction parameters into shared parameters, and generate an expert model diagram of the network configuration.

[0044] refer to Figure 1-Figure 3 , the regular expression used in the data processing step is used to identify the specific format of the configuration instruction and generate structured data including instruction name, parameter name and parameter value;

[0045] Among them, the business documents in the field of network control are messy and difficult to understand, usually containing a large number of professional technical terms and complex configuration details. Its specific form and content can be abstracted into the following table:

[0046] Network management and control service documentation

[0047]

[0048] refer to Figure 1-Figure 3 In the pre-training step, the BERT model generates instruction embedding representation based on the network management term set and instruction set as pre-training data using the masked language model task.

[0049] refer to Figure 1-Figure 3 , the augmented text generation step includes the following process:

[0050] Generate two augmented texts as positive example texts of the same configuration instruction by semantic replacement;

[0051] Other augmented texts are used as negative example texts for subsequent instruction contrastive learning tasks.

[0052] refer to Figure 1-Figure 3,The instruction clustering step adopts a Kmeans-based contrastive learning clustering algorithm, performs cluster analysis on the generated embedding representation based on the augmented text of the instruction, and identifies the configuration order between interrelated instructions.

[0053] refer to Figure 1-Figure 3 , the entity connection step sets the instruction parameters with the same parameters as shared parameters based on the rule matching method to achieve automatic filling of configuration parameters and reference between instructions;

[0054] Among them, the instruction configuration information in the network configuration manual usually has a regular structure, such as "instruction name: parameter name 1 = parameter value 1, parameter name 2 = parameter value 2..." In the data processing step, the configuration manual is traversed in full text using a regular expression: (\S+): ((\S+) = ([\S]+))(, (\S+) = ([\S]+))* full text traversal, where \S+ represents any non-blank character sequence, and ((\S+) = ([\S]+)) represents a parameter key-value pair, thereby roughly extracting the configuration instructions with the above regular structure. For specific instructions that still contain redundant comments, regular matching is repeated multiple times to gradually remove redundant data and obtain the specific instruction information of the required scenario, including instruction name, parameter name and parameter value. For example, when configuring the VLAN (virtual local area network) of a network switch, the configuration instruction may appear in the following form in the configuration manual: "Configure VLAN: VLAN_ID = 10, IP = 192.168.1.1, MASK = 255.255.255.0", where "Configure VLAN" is the command name, followed by a series of parameter names and parameter values, which are VLAN ID, IP address and subnet mask. Regular expressions are used to identify command parameter names and parameter values. The appropriate regular expression is: Configure VLAN: VLAN_ID = (\d+), IP = (\S+), MASK = (\S+), where \d+ matches one or more digits, representing the value of VLAN_ID; \S+ matches a string of characters that do not contain blank characters, which is suitable for matching IP addresses and subnet masks; each pair of parameter names and values ​​is separated by a comma. By using this method to obtain each command name and its parameter name and parameter value information, redundant information in the network configuration manual can be removed, thereby realizing the automatic extraction of network configuration information and providing data support for subsequent entity linking technology;

[0055] The relationship extraction step includes pre-training and clustering. The pre-training module uses BERT-base-cased as the base language model, with the parameter volume set to 110M. The network management terminology set and instruction set are used as pre-training data. The mask language model is used as the task to pre-train BERT to obtain the embedded representation of each instruction. It is assumed that the pre-training data contains multiple network firewall configuration instructions, covering specific configurations such as security policy settings, port rule definitions, and IP address filtering. The typical instruction is: "Set security policy: policy ID = 1234, source IP = 192.168.0.1, destination IP =192.168.0.2, action = allow", the pre-trained model can identify "setting security policy" as the core operation by semantically analyzing the instruction text, and accurately capture the relationship between the four parameters of "policy ID", "source IP", "destination IP" and "action". The mask object in the mask language model is the entity in the control field, that is, the term masked by the mask is predicted based on the context. The model uses the mask language model task to infer that the action of "allow" refers to the release policy in network communication through the context, so that the model can understand the meaning of the instruction and the role of its parameters;

[0056] The clustering module uses the instruction-cluster model to process the embedded representation, discover and summarize the relationship between instructions. By analyzing the appearance patterns of the same parameters such as "source IP" and "destination IP" in different instructions and the manual context, the clustering model can identify interrelated network policies. In addition, by identifying the connection between "action = allow" and specific IP address and port rules, the model can extract the logical relationship of network policies and obtain the specific order of network configuration execution. The specific algorithm flow of the instruction-cluster model is as follows;

[0057] 1. Construct augmented text: obtain two augmented texts for each original text based on semantic replacement. is an instruction set containing M instructions. Each instruction corresponds to a natural language description text. For each natural language description text, two different semantic similarity calculation models are used to replace synonyms of the text respectively, and two different augmented texts are obtained as the positive example text set of this instruction. The original instruction x i For example, the two augmented texts are used as positive examples. The two augmented texts of all other instructions in the same batch are used as x i Negative example i≠j, that is, the augmented text from the same original text is called the positive example corresponding to the original text, and the remaining augmented texts are called the negative examples corresponding to the original text;

[0058] 2. N rounds of clustering based on contrastive learning: The input of each round is a batch of original text X and its corresponding two augmented texts X (1) , X (2) , after the encoder ψ(·), the respective vectors are expressed as Where N is the number of original samples in the current batch, D 1 For each sentence, the encoder encodes the vector dimension, and the encoding E will enter the Kmeans-based clusterer and be encoded with the cluster center Calculate the distance-based clustering loss, where k is the number of clusters. The clustering loss is calculated as follows:

[0059]

[0060] Encoding of augmented text (1) 、E (2) It will enter the instance comparison learning module. The strength comparison learning module separates overlapping classes by separating different instance vectors and encodes the text E (1) 、E (2) These vector representations are mapped to the space where the instance contrast loss is applied through a fully connected mapping layer. Where D 2 To map the dimensions of the vectors, the vector representations of all augmented texts are arranged in order. The subscripts of two augmented texts (i.e., positive pairs) from the same original text in Z are 2 times. The contrast loss of the entire batch is calculated on all positive pairs. τ is the temperature coefficient. The smaller the value, the more obvious the contrast learning effect will be.

[0061]

[0062] Calculate the joint loss to jointly train the encoders,

[0063]

[0064] 3. Sample selection mechanism based on entropy: In each round of clustering, the model calculates the entropy of samples in each batch and selects p samples x with the largest entropy value. i …x i+p-1 , the entropy value is calculated as follows:

[0065]

[0066] where z i For sample x i The encoding, μ k is the code of the kth cluster center, α is the scaling factor, and then, for sample x i …x i+p-1The model calculates the two cluster centers closest to each sample by using the cosine example and obtains the text data corresponding to the cluster center p ,data q Finally, the query pair (x i ,data p ),(x i ,data q );

[0067] 4. Fine-tuning module based on large language model: After N rounds of clustering, the model will accumulate N×p×2 query pairs. At this time, each query pair is combined with the pre-prepared task description and task example in turn to form a complete prompt text input to the large language model. After that, the large language model will output its inference results based on the prompt text, and then calculate the fine-tuning loss of the large language model by comparing the inference results with the sample examples:

[0068]

[0069] where z i For sample x i The encoding, μ k is the encoding of the cluster center inferred by the large language model, α is the scaling factor, and the loss is back-propagated to the encoder to train the encoder;

[0070] After clustering, we have obtained the instruction clusters with the same parameters and the rough logical order of their configuration. Finally, we need to physically connect the clustering results to simplify the repeated configuration of different instructions, that is, convert the same parameters into shared parameters through parameter reference. This means that the instructions configured subsequently can directly reference the parameters of the previously configured instructions, thereby realizing the automatic filling of configuration parameters. The physical connection is performed in an instruction subset through the rule matching method, and the association between instructions is established based on whether the same parameters are referenced. For example, a network firewall needs to configure security policies and access control in turn. By converting IP parameters into shared parameters for reference, the configuration of subsequent instructions can be simplified and unified adjustment and modification can be facilitated. Based on the results of physical connection, the network can intelligently associate related instructions, and finally build a network configuration expert model to realize automatic service activation based on the network configuration manual.

[0071] refer to Figure 1-Figure 3 The present invention also provides an expert model construction device based on a network configuration manual, comprising:

[0072] A data processing module is used to parse the configuration instructions in the network configuration manual, extract the instruction name, parameter name and parameter value using regular expressions, and generate preliminary configuration information containing structured data;

[0073] A pre-training module, connected to the data processing module, performs a masked language model task based on a BERT model to generate an embedded representation of configuration instructions to capture the semantic relationship between instructions;

[0074] An augmented text generation module, connected to the pre-training module, for generating positive and negative example texts for each configuration instruction to support contrastive learning clustering tasks;

[0075] An instruction clustering module, connected to the augmented text generation module, performs cluster analysis on instruction embedding based on the instruction-cluster model, and summarizes the logical association relationship and execution order between instructions;

[0076] The entity connection module is connected to the instruction clustering module, performs sharing processing on instruction parameters based on rule matching and parameter reference, and generates an expert model diagram of configuration parameters.

[0077] refer to Figure 1-Figure 3 The data processing module uses regular expressions to parse the network configuration manual, extracts the structured data of the instructions and removes redundant information. The augmented text generation module generates two augmented texts as positive example texts and other augmented texts as negative example texts for comparative learning and clustering. The entity connection module achieves unified management of multiple instruction parameters by referencing shared parameters, and finally generates an expert model diagram of the configuration parameters.

[0078] Working principle: Please refer to Figure 1-Figure 3 As shown, the present invention obtains structured information by performing regular extraction on the detailed content of each command (or instruction) in the existing business documents, and further clusters the semi-structured information based on semantic similarity using a clustering algorithm to obtain clustering results, thereby extracting the relationship between the information, and finally modeling the reference relationship based on the associated parameters between the commands to realize the visualization of business document knowledge, so that network administrators can more easily cope with complex business scenarios.

[0079] Although the embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that, unless otherwise defined, the technical or scientific terms used in the present invention should have the common meanings understood by those having ordinary skills in the field to which the present invention belongs.

[0080] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An expert model construction method based on a network configuration manual, characterized in that: The following steps are involved: Data processing steps: Parse the configuration instructions in the network configuration manual and use regular expressions to extract instruction names, parameter names, and parameter values ​​to generate preliminary semi-structured data; Pre-training step: using the pre-trained BERT model to process the semi-structured data, generating semantic embeddings for each configuration instruction using a masked language model task to capture the semantic relationship between instructions; Augmented text generation step: Generate augmented text based on semantic replacement to obtain several positive and negative example texts for each instruction for subsequent comparative learning; Instruction clustering step: Using the instruction-cluster model, the augmented text is clustered through contrastive learning to summarize the association between instructions and obtain the reference relationship and logical order of configuration parameters; Entity connection step: Based on the clustering results, rule matching and parameter reference technology are used to uniformly manage the repeated parameters between instructions, convert related instruction parameters into shared parameters, and generate an expert model diagram of the network configuration.

2. The method for constructing an expert model based on a network configuration manual according to claim 1, characterized in that: The regular expression used in the data processing step is used to identify the specific format of the configuration instruction and generate structured data including instruction name, parameter name and parameter value.

3. The method for constructing an expert model based on a network configuration manual according to claim 1, characterized in that: In the pre-training step, the BERT model generates instruction embedding representations using a masked language model task based on a network management term set and an instruction set as pre-training data.

4. The method for constructing an expert model based on a network configuration manual according to claim 1, characterized in that: The augmented text generation step includes the following process: Generate two augmented texts as positive example texts of the same configuration instruction by semantic replacement; Other augmented texts are used as negative example texts for subsequent instruction contrastive learning tasks.

5. The method for constructing an expert model based on a network configuration manual according to claim 1, characterized in that: The instruction clustering step adopts a Kmeans-based contrastive learning clustering algorithm to perform cluster analysis on the generated embedding representation based on the augmented text of the instruction and identify the configuration order between mutually related instructions.

6. The method for constructing an expert model based on a network configuration manual according to claim 1, characterized in that: The entity connection step sets the instruction parameters with the same parameters as shared parameters based on a rule matching method, so as to realize automatic filling of configuration parameters and reference between instructions.

7. An expert model construction device based on a network configuration manual, characterized in that: include: A data processing module is used to parse the configuration instructions in the network configuration manual, extract the instruction name, parameter name and parameter value using regular expressions, and generate preliminary configuration information containing structured data; A pre-training module, connected to the data processing module, performs a masked language model task based on a BERT model to generate an embedded representation of configuration instructions to capture the semantic relationship between instructions; An augmented text generation module, connected to the pre-training module, for generating positive and negative example texts for each configuration instruction to support contrastive learning clustering tasks; An instruction clustering module, connected to the augmented text generation module, performs cluster analysis on instruction embedding based on the instruction-cluster model, and summarizes the logical association relationship and execution order between instructions; The entity connection module is connected to the instruction clustering module, performs sharing processing on instruction parameters based on rule matching and parameter reference, and generates an expert model diagram of configuration parameters.

8. The expert model building device based on the network configuration manual according to claim 7, characterized in that: The data processing module uses regular expressions to parse the network configuration manual, extract structured data of instructions and remove redundant information.

9. The expert model building device based on the network configuration manual according to claim 7, characterized in that: The augmented text generation module generates two augmented texts as positive example texts and other augmented texts as negative example texts for comparative learning and clustering.

10. The expert model building device based on network configuration manual according to claim 7, characterized in that: The entity connection module realizes unified management of multiple instruction parameters by referencing shared parameters, and finally generates an expert model diagram of the configuration parameters.

Images