Network policy intention translation system and method based on retrieval enhancement

Abstract

The invention discloses a network strategy intention translation system and method based on retrieval enhancement. The system comprises an intention acquisition module, a retrieval enhancement module, an intention translation module, a network strategy generation module, a network strategy verification module, a network strategy issuing module and the like. The method comprises the following steps: (1) an intention acquisition module acquires a network strategy intention input by a user in a natural language; (2) a retrieval enhancement module retrieves a retrieval document highly related to the business scene; (3) an intention translation module generates an expected intermediate intention template; (4) a network strategy generation module converts the expected intermediate intention template into a specific network configuration strategy; (5) a network strategy verification module verifies the network configuration strategy; and (6) the network strategy issuing module converts the network configuration strategy into a specific network configuration command and pushes the specific network configuration command to corresponding network equipment.

Description

Technical Field

[0001] The present invention relates to a network policy intention translation system and method based on retrieval enhancement, belonging to the field of information technology, in particular to the field of network policy intelligent configuration technology. Background Art

[0002] With the expansion of network scale and the diversification of network service types, the deployment and management of network policies are facing increasing complexity. Traditional network policy deployment methods have significant limitations. Such methods mainly rely on manual operations of network administrators, which require high-level policy goals to be converted into specific configuration commands and deployed to different network devices. This is not only inefficient, but also significantly increases management costs. The professional skills of network administrators have become the key to the successful implementation of network policies. In order to ensure accuracy, administrators need to be proficient in multiple protocols and configuration details, which places high demands on their professional level. In addition, in large-scale network environments, especially when network policies need to be adjusted frequently, the traditional manual configuration method is obviously no longer applicable. The rigidity and inefficiency of this method make it difficult to adapt to the rapid development of business and cannot meet the needs of modern network management.

[0003] Intent-Based Networking (IBN) is a new network management paradigm that automatically converts user intent into specific network configurations and policies by understanding and analyzing user intent. In IBN, users only need to declare their desired network behavior or target state without detailing how to achieve these goals. The IBN system is responsible for interpreting these high-level intents and converting them into low-level commands and configurations that can be executed on network devices. The core advantage of IBN is that it simplifies network management tasks, enabling network administrators to manage the network by defining the expected results of the network without having to delve into complex configuration details. This network management paradigm reduces manual operations through automation, improves network reliability and security, and accelerates the deployment of new services, allowing the network to respond more flexibly and agilely to changes in business needs.

[0004] Although most existing IBN systems have introduced natural language processing (NLP) technologies such as named entity recognition (NER) to achieve intent parsing, the ambiguity and diversity in natural language make it difficult for NLP technology to fully and accurately understand and process user intent, which limits the accuracy and robustness of the parsing results. Secondly, the entity recognition and intent parsing processes of many IBN systems mainly rely on pre-trained models and limited training data. However, due to the limited data related to network intent, the large model is not exposed to enough scenarios and cases during learning, which affects its ability to understand and process complex or novel intents. Therefore, this design cannot capture the data of new needs and emerging scenarios in the network operation process in a timely manner in actual applications. This limits the applicability and flexibility of the system to a certain extent, and easily leads to data distortion and insufficient generalization ability. In addition, the existing technical solutions often formulate a set of fixed matching rules based on the manuals provided by specific manufacturers to adapt to a certain brand of network equipment. Although this method simplifies the management and configuration of single-brand equipment to a certain extent, it greatly limits the flexibility and scalability of the system. In actual applications, enterprise networks usually contain multiple devices from different vendors, each of which may follow different standards and protocols. Therefore, the traditional method that relies on fixed rules is difficult to effectively support the unified management and policy deployment of devices across manufacturers, which not only increases the complexity of network management, but also may lead to configuration errors and compatibility issues, affecting the stable operation and service quality of the entire network.

[0005] For example, Chinese patents CN112507730A and CN116933098A both use earlier natural language processing technologies, such as NER and finite state machines (FSM). Although these technologies perform well in specific environments, their accuracy and robustness are significantly insufficient when processing complex natural language inputs rich in professional terms and technical details. In particular, when processing intent descriptions that contain a large number of professional terms and technical details, these technologies may lead to misjudgments or ignore important information due to lack of contextual understanding capabilities. Although Chinese patent CN117527620A introduces a knowledge base to assist intent-driven 5G slice service orchestration, the knowledge retrieval technology it uses is still in its infancy and fails to effectively utilize the advantages of the knowledge base. Due to the low accuracy of knowledge retrieval, the system may deviate when matching user needs and fail to accurately extract the information needed to support decision-making.

[0006] In summary, how to use natural language processing technology and knowledge base to optimize network management processes, enhance knowledge retrieval effects, and achieve efficient and accurate intent translation has become a technical problem that urgently needs to be solved in the field of network strategy intent translation. Summary of the invention

[0007] In view of this, the purpose of the present invention is to utilize the powerful language processing capabilities of large language models, combined with the retrieval enhancement technology of two-stage hybrid retrieval, to invent a network policy intent translation system and method to achieve efficient and accurate translation and deployment of network policy intent.

[0008] In order to achieve the above object, the present invention proposes a network policy intent translation system based on retrieval enhancement, and the system includes the following modules:

[0009] Intent acquisition module: The function of this module is to obtain the network policy intent input by the user in natural language, and forward the network policy intent to the intent translation module and the retrieval enhancement module;

[0010] Retrieval enhancement module: This module receives the network policy intent from the intent acquisition module, uses the keyword retrieval algorithm and vector retrieval algorithm based on weighted factors to efficiently retrieve retrieval documents that are highly relevant to the business scenario from the knowledge base, and sends them to the intent translation module.

[0011] Intent translation module: The functions of this module are: (a) to perform deep semantic analysis on the network policy intent received from the intent acquisition module to identify key network configuration requirements and convert them into an initial intermediate intent template in a specific format; (b) to receive the retrieval document sent from the retrieval enhancement module, and the intent translation module optimizes the initial intermediate intent template based on the initial intermediate intent template and the retrieval document to obtain an optimized intermediate intent template; (c) the intent translation module interacts with the user to modify the optimized intermediate intent template to meet expectations, and finally obtains the expected intermediate intent template and sends it to the network policy generation module;

[0012] Network policy generation module: The function of this module is to convert the expected intermediate intent template received from the intent translation module into a specific network configuration policy, i.e., a domain-specific language DSL, using a network policy generation algorithm, and send it to the network policy verification module;

[0013] Network policy verification module: The function of this module is to conduct comprehensive validity and consistency verification of the network configuration policy received from the network policy generation module; the verification is based on the preset rule base and network model to ensure that each network configuration policy complies with the network's security policy, performance requirements and business logic; after successful verification, the network configuration policy is sent to the network policy delivery module;

[0014] Network policy delivery module: The function of this module is to receive the successfully verified configuration policy set from the network policy verification module, and convert the network configuration policy into specific network configuration commands according to the specific type and protocol version of the network device. These network configuration commands will then be pushed to the corresponding network devices to implement the required network changes. During the entire process, the network policy delivery module is also responsible for monitoring the status of the configuration process to ensure that all network devices correctly receive the network configuration commands and can execute these changes, thereby realizing the user's initial intention.

[0015] The specific contents of the search enhancement module that uses a keyword search algorithm based on weighted factors to efficiently search for search documents that are highly relevant to the business scenario from the knowledge base are as follows:

[0016] First, the keyword weight model is used to extract the keywords in the user's network strategy intention and assign corresponding weights to them to form a list of keywords and their weights. in Indicates keywords, Indicates keywords At the same time, the user's network strategy intention is processed through traditional word segmentation methods to generate a keyword list with default weights. in Indicates keywords, with a default weight of w c is 0.1; then, merge lists M and C to get a new list L;

[0017] If the keywords in M If they appear in C at the same time, their final weight is the sum of the two weights:

[0018]

[0019] If the keyword only appears in one of the lists, the weight in the original list is used; next, the list L is used as input, and the keyword retrieval algorithm based on the weighted factor is executed to retrieve documents related to the specific business scenario from the knowledge base, and a score is calculated for each document, and then the topk1 most relevant documents and their scores are output;

[0020] The document score is calculated according to the following formula:

[0021]

[0022] In the above formula, Q is the network policy intention input by the user in natural language, D is all documents, |Q| is the number of terms in the query statement, Q i is the i-th term in the query sentence Q, wl (Q i ) is the i-th term Q in the query statement Q i The weight score, N is the total number of documents, DF(Q i ) contains the term Q i The number of documents, f(Q i ,D) is the term Q i The number of occurrences in document D, w l (Q i ) is the term Q i The weight of k3 is a positive coefficient used to control the saturation of word frequency. k1 and b are parameters used to control the impact of document length on the score. |D| is the document length, and ave|D| is the average length of all documents.

[0023] The retrieval enhancement module uses a vector retrieval algorithm to efficiently retrieve retrieval documents that are highly relevant to the business scenario from the knowledge base. The specific content is:

[0024] In the system initialization phase, the retrieval enhancement module vectorizes all documents, that is, converts the content of each document into a vector representation in a high-dimensional space to better capture the semantic information of the document. Each document obtains a unique vector index, which not only captures the content information of the document, but also reflects the semantic similarity between documents.

[0025] For a given network policy intent input by a user in natural language, the retrieval enhancement module vectorizes it to form a query vector, and before traversing, uses a keyword retrieval algorithm based on weighted factors to efficiently retrieve retrieval documents that are highly relevant to the business scenario from the knowledge base, calculates the similarity score between the query vector and the retrieval document vector, and finally outputs the topk2 documents with the highest scores as retrieval results; the similarity calculation method includes but is not limited to cosine similarity and Euclidean distance.

[0026] The intent translation module is a large network model based on the Transformer architecture, which can capture key information of the user's network strategy intention by using the self-attention mechanism; the intent translation module is first pre-trained on a large amount of general text data to ensure that it has a wide range of language understanding and generation capabilities; then, it is fine-tuned using network data from different manufacturers to improve its performance on network-related tasks; the fine-tuning dataset not only includes existing network data, but also generates a large amount of synthetic data through the capabilities of the intent translation module large model itself. These synthetic data are generated based on the existing network dataset through the generation capability of the model, aiming to increase the diversity and complexity of the dataset to cover more network application scenarios; in order to avoid the catastrophic forgetting problem of the model during the fine-tuning process, the LoRA (Low-Rank Adaptation) fine-tuning method is adopted to update only part of the model's weights, so as to quickly adapt to new tasks while maintaining the original model capabilities, without significantly increasing the number of model parameters.

[0027] The specific content of the interaction between the intent translation module and the user is: the user puts forward his opinions and suggestions on the current optimized intermediate intent template in the form of a free-style dialogue through natural language, and specifically points out the parts that need to be adjusted; the intent translation module automatically adjusts and optimizes the intermediate intent template based on the user's feedback, and generates a new version template that is more in line with the user's actual needs, so as to better reflect the user's actual needs. This process can be iterated repeatedly until the user is satisfied.

[0028] In order to achieve the adaptation and support of equipment from multiple manufacturers, the knowledge base in the search enhancement module always preserves the latest network management knowledge and strategies by introducing the latest network configuration guides, best practices and technical information of equipment from different manufacturers, making the search results more robust.

[0029] The present invention also proposes a network policy intent translation method based on retrieval enhancement, the method comprising the following steps:

[0030] (1) The intent acquisition module acquires the network policy intent input by the user in natural language, and forwards the network policy intent to the intent translation module and the retrieval enhancement module;

[0031] (2) The retrieval enhancement module receives the network policy intent sent by the intent acquisition module, and uses the keyword retrieval algorithm and vector retrieval algorithm based on weighted factors to efficiently retrieve retrieval documents that are highly relevant to the business scenario from the knowledge base and send them to the intent translation module;

[0032] (3) The intent translation module performs deep semantic analysis on the network policy intent received from the intent acquisition module to identify key network configuration requirements and converts them into an initial intermediate intent template in a specific format; the intent translation module receives the retrieval document sent from the retrieval enhancement module, and optimizes the initial intermediate intent template based on the initial intermediate intent template and the retrieval document to obtain an optimized intermediate intent template; the intent translation module interacts with the user to modify the optimized intermediate intent template to meet expectations, and finally obtains the expected intermediate intent template and sends it to the network policy generation module;

[0033] (4) The network policy generation module converts the expected intermediate intent template received from the intent translation module into a specific network configuration policy, i.e., a domain-specific language (DSL), using a network policy generation algorithm, and sends the template to the network policy verification module;

[0034] (5) The network policy verification module performs comprehensive validity and consistency verification on the network configuration policy received from the network policy generation module; the verification is based on the preset rule base and network model to ensure that each network configuration policy complies with the network's security policy, performance requirements and business logic; after successful verification, the network configuration policy is sent to the network policy delivery module;

[0035] (6) The network policy delivery module receives the successfully verified configuration policy set from the network policy verification module, and converts the network configuration policy into specific network configuration commands based on the specific type and protocol version of the network device. These network configuration commands are then pushed to the corresponding network devices to implement the required network changes. During the entire process, the network policy delivery module is also responsible for monitoring the status of the configuration process to ensure that all network devices correctly receive the network configuration commands and can execute these changes, thereby realizing the user's initial intention.

[0036] The beneficial effects of the present invention are as follows: the system and method of the present invention not only adopt a large language model based on the Transformer architecture, but also fine-tune it through data unique to the network field, and combine it with retrieval enhancement technology, which greatly improves the system's adaptability to multi-manufacturer equipment. During the fine-tuning process, the system has the ability to widely absorb network data from different manufacturers, which not only enhances the model's ability to understand various device configuration instructions, but also enables it to generate intermediate intent templates that are more in line with the equipment of each manufacturer. Through the application of retrieval enhancement technology, the system can quickly and accurately retrieve relevant information from a multi-manufacturer data knowledge base, and further optimize the intermediate intent template to ensure that it not only meets the user's original intention, but also can adapt to network equipment from different manufacturers. The system and method of the present invention not only improves the ability to handle complex network management tasks, but also significantly improves the accuracy and efficiency of service responses, effectively meeting the needs of supporting device diversity in network management. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a structural diagram of the network strategy intent translation system based on retrieval enhancement proposed by the present invention.

[0038] Figure 2 It is a flow chart of the network policy intent translation method based on retrieval enhancement proposed by the present invention. DETAILED DESCRIPTION

[0039] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings.

[0040] See also Figure 1 , introduces the network strategy intention translation system based on retrieval enhancement proposed by the present invention, and the system includes the following modules:

[0041] Intent acquisition module: The function of this module is to obtain the network policy intent input by the user in natural language, and forward the network policy intent to the intent translation module and the retrieval enhancement module;

[0042] For example, the intent acquisition module obtains the network policy intent input by the user in natural language: I want to watch a 2k variety show on iQiyi at 9 o'clock tonight.

[0043] Retrieval enhancement module: This module receives the network policy intent from the intent acquisition module, uses the keyword retrieval algorithm and vector retrieval algorithm based on weighted factors to efficiently retrieve retrieval documents that are highly relevant to the business scenario from the knowledge base, and sends them to the intent translation module.

[0044] Intent translation module: The functions of this module are:

[0045] (a) Perform deep semantic analysis on the network policy intent received from the intent acquisition module to identify key network configuration requirements and convert them into an initial intermediate intent template in a specific format;

[0046] For the above example, the intent translation module converts it into an initial intermediate intent template in a specific format as follows:

[0047] endpoint=local, service=video streaming, traffic=iQiyi, time=9:00, resolution=2K, bandwidth=5Mbps-10Mbps

[0048] (b) receiving the retrieval document sent from the retrieval enhancement module, the intent translation module optimizes the initial intermediate intent template according to the initial intermediate intent template and the retrieval document to obtain an optimized intermediate intent template; (c) the intent translation module interacts with the user to modify the optimized intermediate intent template to meet expectations, and finally obtains the expected intermediate intent template and sends it to the network strategy generation module;

[0049] Network policy generation module: The function of this module is to convert the expected intermediate intent template received from the intent translation module into a specific network configuration policy, i.e., a domain-specific language DSL, using a network policy generation algorithm, and send it to the network policy verification module;

[0050] In the embodiment, the specific working content of the network policy generation algorithm is as follows: by parsing the expected intermediate intent template, automatically configure network parameters, including protocol type, source IP address, destination IP address, service time, bandwidth and resolution, etc., to optimize user experience and ensure service quality, and finally convert it into a network configuration policy in JSON format.

[0051] Network policy verification module: The function of this module is to conduct comprehensive validity and consistency verification of the network configuration policy received from the network policy generation module; the verification is based on the preset rule base and network model to ensure that each network configuration policy complies with the network's security policy, performance requirements and business logic; after successful verification, the network configuration policy is sent to the network policy delivery module;

[0052] Network policy delivery module: The function of this module is to receive the successfully verified configuration policy set from the network policy verification module, and convert the network configuration policy into specific network configuration commands according to the specific type and protocol version of the network device. These network configuration commands will then be pushed to the corresponding network devices to implement the required network changes. During the entire process, the network policy delivery module is also responsible for monitoring the status of the configuration process to ensure that all network devices correctly receive the network configuration commands and can execute these changes, thereby realizing the user's initial intention.

[0053] The specific contents of the search enhancement module that uses a keyword search algorithm based on weighted factors to efficiently search for search documents that are highly relevant to the business scenario from the knowledge base are as follows:

[0054] First, the keyword weight model is used to extract the keywords in the user's network strategy intention and assign corresponding weights to them to form a list of keywords and their weights. in Indicates keywords, Indicates keywords At the same time, the user's network strategy intention is processed through traditional word segmentation methods to generate a keyword list with default weights. in Indicates keywords, with a default weight of w c is 0.1; then, merge lists M and C to get a new list L;

[0055] If the keywords in M If they appear in C at the same time, their final weight is the sum of the two weights:

[0056]

[0057] If a keyword appears in only one of the lists, the weight in the original list is used;

[0058] For the above example, the new list L obtained by the search enhancement module is as follows:

[0059] ('iqiyi': 0.9, '2k': 0.9, 'variety show': 0.4, 'I want': 0.1, '9 o'clock': 0.2, 'tonight': 0.1, 'watch': 0.1, 'one': 0.1)

[0060] Next, take the list L as input and execute the keyword retrieval algorithm based on weighted factors to retrieve documents related to the specific business scenario from the knowledge base, calculate a score for each document, and then output the topk1 most relevant documents and their scores;

[0061] The document score is calculated according to the following formula:

[0062]

[0063] In the above formula, Q is the network policy intention input by the user in natural language, D is all documents, |Q| is the number of terms in the query statement, Q i is the i-th term in the query sentence Q, w l (Q i ) is the i-th term Q in the query statement Q i The weight score, N is the total number of documents, DF(Q i ) contains the term Q i The number of documents, f(Q i ,D) is the term Q i The number of occurrences in document D, w l (Q i ) is the term Q i The weight of k3 is a positive coefficient used to control the saturation of word frequency. k1 and b are parameters used to control the impact of document length on the score. |D| is the document length, and ave|D| is the average length of all documents.

[0064] For the above example, the most relevant document scores obtained by the retrieval enhancement module are [0.5361, 0.3369, 0.3360......, 0, 0], and the corresponding top three documents are: (1) From Huawei: The recommended minimum bandwidth for watching 2K videos on iQiyi is 10Mbps, and the actual bandwidth may be 16Mbps or higher; if multiple devices share the same network connection, the total bandwidth required will increase accordingly for each additional device watching the video; (2) From Cisco: The server bandwidth required for watching 2K videos on Sohu Video is 3Mbps to 10Mbps, and 6Mbps and above is recommended; (3) From Huawei: The recommended server bandwidth for watching 2K videos on LeTV is 6Mbps and above.

[0065] The retrieval enhancement module uses a vector retrieval algorithm to efficiently retrieve retrieval documents that are highly relevant to the business scenario from the knowledge base. The specific content is:

[0066] In the system initialization phase, the retrieval enhancement module vectorizes all documents, that is, converts the content of each document into a vector representation in a high-dimensional space to better capture the semantic information of the document. Each document obtains a unique vector index, which not only captures the content information of the document, but also reflects the semantic similarity between documents.

[0067] For a given network policy intent input by a user in natural language, the retrieval enhancement module vectorizes it to form a query vector, and before traversing, uses a keyword retrieval algorithm based on weighted factors to efficiently retrieve retrieval documents that are highly relevant to the business scenario from the knowledge base, calculates the similarity score between the query vector and the retrieval document vector, and finally outputs the topk2 documents with the highest scores as retrieval results; the similarity calculation method includes but is not limited to cosine similarity and Euclidean distance.

[0068] For the above example, the retrieval enhancement module finally outputs the topk2 documents with the highest scores, i.e. the retrieval results are as follows:

[0069] (1) The recommended minimum bandwidth for watching 2K videos on iQiyi is 10Mbps, and the actual bandwidth may be 16Mbps or higher. If multiple devices share the same network connection, the total bandwidth required will increase with each additional device watching the video. (2) The local IP address is 192.168.0.1, and the user device type is Huawei router.

[0070] For the above example, the intent translation module optimizes the initial intermediate intent template based on the initial intermediate intent template and the search results. The optimized intermediate intent template is as follows:

[0071] endpoint=192.168.0.1, service=video streaming, traffic=iQiyi, time=9:00, resolution=2K, bandwidth=10Mbps-*, device_type=Huawei Router.

[0072] The intent translation module is a large network model based on the Transformer (Vaswani A. Attention is all you need [J]. Advances in Neural Information Processing Systems, 2017.) architecture, which can use the self-attention mechanism to capture the key information of the user's network strategy intention; the intent translation module is first pre-trained on a large amount of general text data to ensure that it has a wide range of language understanding and generation capabilities; then, it is fine-tuned using network data from different manufacturers to improve its performance on network-related tasks; the fine-tuning dataset not only includes existing network data, but also generates a large amount of synthetic data through the ability of the intent translation module large model itself. These synthetic data are generated based on the existing network dataset through the generation ability of the model, aiming to increase the diversity and complexity of the dataset to cover more network application scenarios; in order to avoid the catastrophic forgetting problem of the model during the fine-tuning process, the LoRA (Low-Rank Adaptation) fine-tuning method is adopted to update only part of the model's weights, so as to quickly adapt to new tasks while maintaining the original model capabilities, without significantly increasing the number of model parameters.

[0073] The specific content of the interaction between the intent translation module and the user is: the user puts forward his opinions and suggestions on the current optimized intermediate intent template in the form of a free-style dialogue through natural language, and specifically points out the parts that need to be adjusted; the intent translation module automatically adjusts and optimizes the intermediate intent template based on the user's feedback, and generates a new version template that is more in line with the user's actual needs, so as to better reflect the user's actual needs. This process can be iterated repeatedly until the user is satisfied.

[0074] For the above example, suppose the user is not satisfied with the optimized intermediate intent template and provides more specific modification suggestions through the dialogue: the time is 9 pm, that is, 21:00. The intent translation module automatically adjusts and optimizes the intermediate intent template based on the user's feedback and generates a new version template that better meets the user's actual needs, namely the expected intermediate intent template as follows:

[0075] endpoint=192.168.0.1, service=video streaming, traffic=iQiyi, time=21:00, resolution=2K, bandwidth=10Mbps-*, device_type=Huawei Router.

[0076] For the above example, the network policy generation module uses the network policy generation algorithm to convert the expected intermediate intent template received from the intent translation module into a specific network configuration policy as follows:

[0077]

[0078]

[0079] In order to achieve the adaptation and support of equipment from multiple manufacturers, the knowledge base in the search enhancement module always preserves the latest network management knowledge and strategies by introducing the latest network configuration guides, best practices and technical information of equipment from different manufacturers, making the search results more robust.

[0080] See also Figure 2 , introduces a network policy intent translation method based on retrieval enhancement proposed by the present invention, the method comprising the following steps:

[0081] (1) The intent acquisition module acquires the network policy intent input by the user in natural language, and forwards the network policy intent to the intent translation module and the retrieval enhancement module;

[0082] (2) The retrieval enhancement module receives the network policy intent sent by the intent acquisition module, and uses the keyword retrieval algorithm and vector retrieval algorithm based on weighted factors to efficiently retrieve retrieval documents that are highly relevant to the business scenario from the knowledge base and send them to the intent translation module;

[0083] (3) The intent translation module performs deep semantic analysis on the network policy intent received from the intent acquisition module to identify key network configuration requirements and converts them into an initial intermediate intent template in a specific format; the intent translation module receives the retrieval document sent from the retrieval enhancement module, and optimizes the initial intermediate intent template based on the initial intermediate intent template and the retrieval document to obtain an optimized intermediate intent template; the intent translation module interacts with the user to modify the optimized intermediate intent template to meet expectations, and finally obtains the expected intermediate intent template and sends it to the network policy generation module;

[0084] (4) The network policy generation module converts the expected intermediate intent template received from the intent translation module into a specific network configuration policy, i.e., a domain-specific language (DSL), using a network policy generation algorithm, and sends the template to the network policy verification module;

[0085] (5) The network policy verification module performs comprehensive validity and consistency verification on the network configuration policy received from the network policy generation module; the verification is based on the preset rule base and network model to ensure that each network configuration policy complies with the network's security policy, performance requirements and business logic; after successful verification, the network configuration policy is sent to the network policy delivery module;

[0086] (6) The network policy delivery module receives the successfully verified configuration policy set from the network policy verification module, and converts the network configuration policy into specific network configuration commands based on the specific type and protocol version of the network device. These network configuration commands are then pushed to the corresponding network devices to implement the required network changes. During the entire process, the network policy delivery module is also responsible for monitoring the status of the configuration process to ensure that all network devices correctly receive the network configuration commands and can execute these changes, thereby realizing the user's initial intention.

[0087] The inventor conducted a large number of experiments on the method proposed in the present invention. The experiment was simulated on a Linux server using a Python environment, aiming to evaluate the effectiveness of the network policy intent translation technology solution based on retrieval enhancement proposed in the present invention. During the experiment, the ability of the large model of the intention translation module to parse user input and generate an initial intermediate intent template was first evaluated through a series of network policy intents input by users in natural language. The results show that the model can accurately capture the user's network policy intent and generate an initial intermediate intent template that meets expectations. Then, the effectiveness of the keyword weight allocation mechanism was verified by combining the keyword weight model with the traditional word segmentation library. The experimental data show that the mechanism can effectively highlight the importance of keywords, thereby improving the efficiency and accuracy of subsequent processing steps. In addition, the experiment also evaluated the performance of the retrieval enhancement module. By comparing knowledge items from different sources, it was proved that the module can quickly locate the most relevant knowledge in a large amount of information, significantly improving the response speed and accuracy of the system. The experiment also examined the system's ability to adjust and optimize the intermediate intent template based on user feedback. The results show that when users provide more specific modification suggestions, the system can respond quickly and generate an expected intermediate intent template that better meets user needs. Finally, the experiment verified the system's ability to generate corresponding network configuration policies based on the expected intermediate intent template. The generated network configuration policies not only accurately reflect the user's network policy intentions, but also take into account the requirements of actual network conditions. By comparing the consistency between the generated network configuration policies and the expected goals, it is confirmed that the network policy generation algorithm can effectively transform complex user network policy intentions into specific network configuration policies.

[0088] The above experimental results prove that the system and method proposed in the present invention are feasible and effective.

Claims

1. A network policy intent translation system based on retrieval enhancement, characterized by: The system includes the following modules: Intent acquisition module: The function of this module is to obtain the network policy intent input by the user in natural language, and forward the network policy intent to the intent translation module and the retrieval enhancement module; Retrieval enhancement module: This module receives the network policy intent from the intent acquisition module, uses the keyword retrieval algorithm and vector retrieval algorithm based on weighted factors to efficiently retrieve retrieval documents that are highly relevant to the business scenario from the knowledge base, and sends them to the intent translation module. Intent translation module: The functions of this module are: (a) to perform deep semantic analysis on the network policy intent received from the intent acquisition module to identify key network configuration requirements and convert them into an initial intermediate intent template in a specific format; (b) to receive the retrieval document sent from the retrieval enhancement module, and the intent translation module optimizes the initial intermediate intent template based on the initial intermediate intent template and the retrieval document to obtain an optimized intermediate intent template; (c) the intent translation module interacts with the user to modify the optimized intermediate intent template to meet expectations, and finally obtains the expected intermediate intent template and sends it to the network policy generation module; Network policy generation module: The function of this module is to convert the expected intermediate intent template received from the intent translation module into a specific network configuration policy, i.e., a domain-specific language DSL, using a network policy generation algorithm, and send it to the network policy verification module; Network policy verification module: The function of this module is to conduct comprehensive validity and consistency verification of the network configuration policy received from the network policy generation module; the verification is based on the preset rule base and network model to ensure that each network configuration policy complies with the network's security policy, performance requirements and business logic; after successful verification, the network configuration policy is sent to the network policy delivery module; Network policy delivery module: The function of this module is to receive the successfully verified configuration policy set from the network policy verification module, and convert the network configuration policy into specific network configuration commands according to the specific type and protocol version of the network device. These network configuration commands will then be pushed to the corresponding network devices to implement the required network changes. During the entire process, the network policy delivery module is also responsible for monitoring the status of the configuration process to ensure that all network devices correctly receive the network configuration commands and can execute these changes, thereby realizing the user's initial intention.

2. The network policy intent translation system based on retrieval enhancement according to claim 1 is characterized in that: The specific contents of the search enhancement module that uses a keyword search algorithm based on weighted factors to efficiently search for search documents that are highly relevant to the business scenario from the knowledge base are as follows: First, the keyword weight model is used to extract the keywords in the user's network strategy intention and assign corresponding weights to them to form a list of keywords and their weights. in Indicates keywords, Indicates keywords At the same time, the user's network strategy intention is processed through traditional word segmentation methods to generate a keyword list with default weights. in Indicates keywords, with a default weight of w c is 0.1; Then, merge the lists M and C to get a new list L; If the keywords in M If they both appear in C at the same time, their final weight is the sum of the two weights: If the keyword only appears in one of the lists, the weight in the original list is used; next, the list L is used as input, and the keyword retrieval algorithm based on the weighted factor is executed to retrieve documents related to the specific business scenario from the knowledge base, and a score is calculated for each document, and then the topk1 most relevant documents and their scores are output; The document score is calculated according to the following formula: In the above formula, Q is the network policy intention input by the user in natural language, D is all documents, |Q| is the number of terms in the query statement, Q i is the i-th term in the query sentence Q, w l (Q i ) is the i-th term Q in the query statement Q i The weight score, N is the total number of documents, DF(Q i ) contains the term Q i The number of documents, f(Q i ,D) is the term Q i The number of occurrences in document D, w l (Q i ) is the term Q i The weight of k3 is a positive coefficient used to control the saturation of word frequency. k1 and b are parameters used to control the impact of document length on the score. |D| is the document length, and ave|D| is the average length of all documents.

3. The network policy intention translation system based on retrieval enhancement according to claim 1 is characterized in that: The retrieval enhancement module uses a vector retrieval algorithm to efficiently retrieve retrieval documents that are highly relevant to the business scenario from the knowledge base. The specific content is: In the system initialization phase, the retrieval enhancement module vectorizes all documents, that is, converts the content of each document into a vector representation in a high-dimensional space to better capture the semantic information of the document. Each document obtains a unique vector index, which not only captures the content information of the document, but also reflects the semantic similarity between documents. For a given network policy intent input by a user in natural language, the retrieval enhancement module vectorizes it to form a query vector, and before traversing, uses a keyword retrieval algorithm based on weighted factors to efficiently retrieve retrieval documents that are highly relevant to the business scenario from the knowledge base, calculates the similarity score between the query vector and the retrieval document vector, and finally outputs the topk2 documents with the highest scores as retrieval results; the similarity calculation method includes but is not limited to cosine similarity and Euclidean distance.

4. The network policy intent translation system based on retrieval enhancement according to claim 1 is characterized in that: The intent translation module is a large network model based on the Transformer architecture, which can capture key information about the user's network strategy intent using the self-attention mechanism. The intent translation module is first pre-trained on a large amount of general text data to ensure that it has a wide range of language understanding and generation capabilities. It is then fine-tuned using network data from different vendors to improve its performance on network-related tasks. The fine-tuning dataset not only includes the existing network data, but also generates a large amount of synthetic data through the capabilities of the large model of the intent translation module. These synthetic data are generated based on the existing network dataset through the generation capability of the model, aiming to increase the diversity and complexity of the dataset to cover more network application scenarios. In order to avoid the catastrophic forgetting problem of the model during the fine-tuning process, the LoRA (Low-Rank Adaptation) fine-tuning method is adopted to update only part of the model's weights, so as to quickly adapt to new tasks while maintaining the original model capabilities, without significantly increasing the number of model parameters.

5. The network policy intention translation system based on retrieval enhancement according to claim 1 is characterized in that: The specific content of the interaction between the intent translation module and the user is: the user puts forward his opinions and suggestions on the current optimized intermediate intent template in the form of a free-style dialogue through natural language, and specifically points out the parts that need to be adjusted; the intent translation module automatically adjusts and optimizes the intermediate intent template based on the user's feedback, and generates a new version template that is more in line with the user's actual needs, so as to better reflect the user's actual needs. This process can be iterated repeatedly until the user is satisfied.

6. The network policy intention translation system based on retrieval enhancement according to claim 1 is characterized in that: In order to achieve the adaptation and support of equipment from multiple manufacturers, the knowledge base in the search enhancement module always preserves the latest network management knowledge and strategies by introducing the latest network configuration guides, best practices and technical information of equipment from different manufacturers, making the search results more robust.

7. A network policy intent translation method based on retrieval enhancement, characterized in that: The method comprises the following steps: (1) The intent acquisition module acquires the network policy intent input by the user in natural language, and forwards the network policy intent to the intent translation module and the retrieval enhancement module; (2) The retrieval enhancement module receives the network policy intent sent by the intent acquisition module, and uses the keyword retrieval algorithm and vector retrieval algorithm based on weighted factors to efficiently retrieve retrieval documents that are highly relevant to the business scenario from the knowledge base and send them to the intent translation module; (3) The intent translation module performs deep semantic analysis on the network policy intent received from the intent acquisition module to identify key network configuration requirements and converts them into an initial intermediate intent template in a specific format; the intent translation module receives the retrieval document sent from the retrieval enhancement module, and optimizes the initial intermediate intent template based on the initial intermediate intent template and the retrieval document to obtain an optimized intermediate intent template; the intent translation module interacts with the user to modify the optimized intermediate intent template to meet expectations, and finally obtains the expected intermediate intent template and sends it to the network policy generation module; (4) The network policy generation module converts the expected intermediate intent template received from the intent translation module into a specific network configuration policy, i.e., a domain-specific language (DSL), using a network policy generation algorithm, and sends the template to the network policy verification module; (5) The network policy verification module performs comprehensive validity and consistency verification on the network configuration policy received from the network policy generation module; the verification is based on the preset rule base and network model to ensure that each network configuration policy complies with the network's security policy, performance requirements and business logic; after successful verification, the network configuration policy is sent to the network policy delivery module; (6) The network policy delivery module receives the successfully verified configuration policy set from the network policy verification module, and converts the network configuration policy into specific network configuration commands based on the specific type and protocol version of the network device. These network configuration commands are then pushed to the corresponding network devices to implement the required network changes. During the entire process, the network policy delivery module is also responsible for monitoring the status of the configuration process to ensure that all network devices correctly receive the network configuration commands and can execute these changes, thereby realizing the user's initial intention.

Images