Network management racing item question-answering system based on knowledge graph

Abstract

The invention discloses a network management racing item question answering system based on a knowledge graph, and belongs to the technical field of knowledge graph teaching. In the system, the association relationship among massive network equipment, protocol configuration and fault cases is stored in a knowledge graph structuralization mode, and natural language processing and graph database retrieval technologies are combined; core knowledge points of complex problems can be quickly positioned, and strong scene adaptability and expansion potential are shown through modular design and a dynamic cooperation mechanism. The knowledge graph management module supports incremental updating and cleaning fusion of multi-source heterogeneous data, can integrate newest competition cases and technical specifications in time, and guarantees the frontier and authority of a knowledge base; and the interaction optimization module automatically identifies a knowledge blind area according to the score of the user on the answer and triggers graph completion. The self-feedback and self-iteration mechanism not only reduces the manual maintenance cost, but also provides a technical basis for cross-domain knowledge migration.

Description

Technical Field

[0001] The present invention belongs to the field of knowledge graph teaching technology, and specifically is a network management competition question and answer system based on knowledge graph. Background Art

[0002] In recent years, with the increasing complexity and diversity of network system management competitions, the need for knowledge management and application has become increasingly urgent. As a key competition format in the information technology field, the depth and breadth of network system management knowledge plays a crucial role in improving participants' skills and achieving their performance. With the rapid development of data science and network technology, traditional knowledge management methods are no longer sufficient to meet the current demands of network system management competitions. Therefore, building a comprehensive, efficient, and intelligent network system management knowledge graph is crucial. Global search engine giants such as Google, Baidu, and Sogou have leveraged linked open data to build their own knowledge graphs, significantly improving search quality and deepening the application of semantic search. Google's knowledge graph integrates vast amounts of information to provide users with more intuitive and rich search results. Baidu's "Zhixin" and Sogou's "Zhilifang" leverage Chinese context to provide users with a more precise and personalized search experience. These successful cases not only demonstrate the enormous potential of knowledge graphs in the search field but also provide valuable experience and inspiration for building knowledge graphs for network system management competitions. In the field of education, the continuous emergence and integration of linked datasets has made it easier for researchers in the field of educational informatics to access and utilize these resources. Several well-known linked data platforms, such as the Linked Open Drug Dataset, LikedLifeData, and Bio2RDF, provide strong data support for teaching and research in the biomedical field. Similarly, network system management competitions also require a robust knowledge graph to support their teaching and research.

[0003] However, traditional network management knowledge systems mostly rely on static databases and rule engines, which make it difficult to effectively associate scattered heterogeneous knowledge, resulting in fragmented query results and a lack of contextual reasoning capabilities. The user interaction mechanism is rigid and cannot dynamically optimize answer generation strategies based on real-time feedback. Knowledge updates rely on manual maintenance and lag behind technological evolution. The degree of coupling between modules is high, making it difficult to support flexible expansion and adaptive learning across scenarios. Summary of the Invention

[0004] The purpose of the present invention is to provide a network management competition question and answer system based on knowledge graph in order to solve the above problems.

[0005] The technical solution adopted by the present invention is as follows: a network management competition question-answering system based on knowledge graph, the system comprising: a user interaction layer module, a question analysis module, a knowledge retrieval module, an answer generation module, an interaction optimization module, a system monitoring and evaluation module, and a knowledge graph management module;

[0006] The interactive optimization module is internally configured with: a feedback collection and annotation submodule, a model optimization submodule, and a strategy adjustment and deployment submodule;

[0007] The front-end interface input port of the user interaction layer module is directly connected to the natural language processing interface of the question analysis module;

[0008] The structured output port of the problem analysis module is connected to the query engine input port of the knowledge retrieval module through a standardized data protocol.

[0009] The result output terminal of the knowledge retrieval module is connected to the data input interface of the answer generation module.

[0010] The multimodal rendering port of the answer generation module is connected back to the front-end display interface of the user interaction layer, and the text and chart answers are returned to the user interface.

[0011] The feedback collection port of the user interaction layer is connected to the data input port of the interaction optimization module;

[0012] The strategy adjustment port of the interactive optimization module is respectively embedded with the model parameter interface of the question analysis module, the similarity threshold regulator of the knowledge retrieval module and the priority controller of the answer generation module.

[0013] The indicator collection end of the system monitoring and evaluation module is connected to the performance log output end of all modules through distributed probes

[0014] The data update interface of the knowledge graph management module is bidirectionally connected to the storage engine of the knowledge retrieval module and the external data source input pipeline.

[0015] In a preferred embodiment, the user interaction layer module is provided with a visual front-end interface based on Web technology and a standardized API interface. The front-end interface adopts a responsive design and supports multi-terminal adaptation. The core components include a question input box, an answer display panel and a user feedback control, wherein the feedback control embeds a five-star rating and tag selection function to capture explicit evaluations. The API interface follows the RESTful specification and provides a request response mechanism in JSON format to facilitate third-party system integration and automated calls. Interface interaction data is collected in real time through point-of-sale technology to collect user behavior logs, page dwell time, answer expansion times and question redirection paths. These logs are initially formatted and then transmitted to downstream modules for in-depth analysis to ensure comprehensive analysis of user intentions and efficient use of feedback.

[0016] In a preferred embodiment, the problem parsing module is provided with a natural language processing pipeline, including four core links: word segmentation, part-of-speech tagging, syntactic analysis and intent recognition. The word segmentation link adopts a word segmentation tool that combines a hybrid dictionary and a statistical model, taking into account both professional terminology recognition and new word discovery capabilities; the part-of-speech tagging link uses a hidden Markov model and a conditional random field to jointly label entity types, with network device names as nouns and configuration actions as verbs. The syntactic analysis link parses the subject-verb-object structure based on the dependency tree algorithm to extract the core entities and relationships in the problem; the intent recognition link integrates rule templates and deep learning classifiers to map problems to preset categories, fault diagnosis, configuration queries or performance optimization. The parsed structured data is encapsulated as a unified query instruction to provide accurate input for knowledge retrieval.

[0017] In a preferred embodiment, the knowledge retrieval module is internally provided with a dual mechanism of a graph database query engine and a semantic expansion algorithm. The graph database query engine constructs dynamic search statements based on the Cypher language, quickly locates associated nodes in the knowledge graph according to the parsed entities and relationships, and matches the relationship chain between router models and compatible protocols. The semantic expansion algorithm uses a pre-trained language model to calculate the cosine similarity between user questions and knowledge base content, generalizes and completes fuzzy queries, and expands "routing setting problems" to associated scenarios of "static routing configuration errors" and "dynamic routing protocol conflicts." After the retrieval results are sorted by relevance and filtered for redundancy, high-confidence entity attributes, relationship paths, and context descriptions are output, providing multi-dimensional data support for answer generation.

[0018] In a preferred embodiment, the answer generation module is internally provided with a dual-track strategy of a template matching engine and a natural language generation model. The template matching engine predefines the answer framework for multiple scenarios. The fault type template includes cause location, solution steps and prevention suggestions. The retrieval results are injected into the template through slot filling technology to generate standardized answers. The natural language generation model trains end-to-end text generation capabilities based on the Transformer architecture, dynamically combines knowledge fragments for complex reasoning scenarios, and integrates multiple device configuration conflicts into a coherent solution description. The answer output link introduces multimodal rendering technology, which supports the mixed display of text, flowcharts and code blocks. When explaining the network topology, a visual connection diagram is generated simultaneously to improve the comprehensibility and operational guidance of the answer.

[0019] In a preferred embodiment, the feedback collection and annotation submodule directly captures the user's explicit evaluation through the front-end interactive control, and records the user behavior to infer implicit feedback. The raw data is filtered out of noise by matching high-frequency keywords with regular expressions, and two types of high-value samples are screened based on the active learning algorithm: low-certainty samples whose model prediction confidence is close to the threshold, and high-diversity samples whose semantic difference with the annotated data exceeds a preset ratio. The screened samples are pushed to the annotation platform, manually annotated according to priority, and finally integrated into a triple structured data set containing annotation results, original logs, and behavioral indicators for downstream model iterative optimization.

[0020] The marking priority formula is:

[0021]

[0022] Where: α represents the weight coefficient of high-frequency words, which is dynamically calculated based on the frequency of occurrence of regular matching keywords in historical data. The higher the frequency, the larger the α value.

[0023] f k The number of times high-frequency keywords are matched in the current sample is logarithmically scaled to reduce the long-tail distribution bias.

[0024] β represents the confidence sensitivity coefficient, which is preset with a fixed value according to the model type (such as classification or retrieval model) and is used to amplify the impact of low-confidence samples.

[0025] c represents the model's confidence in predicting the sample, τ is the preset threshold (such as 0.5), and the smaller the difference, the more uncertain the model is.

[0026] γ represents the diversity gain coefficient, which is adaptively adjusted based on the semantic coverage of the current annotated dataset. The lower the coverage, the higher the \gammaγ value.

[0027] sim(s,Slabeled) represents the average semantic similarity between the current sample s and the labeled dataset Slabeled, and the cosine similarity is calculated using the pre-trained word vector.

[0028] In a preferred embodiment, the model optimization submodule dynamically adjusts the core model parameters of the question-answering system through user feedback data to improve the accuracy of intent recognition and answer generation. The specific process includes: based on the reinforcement learning framework, converting the user's score for the answer into a reward signal, using the policy gradient algorithm to optimize the decision weight of the answer generation model, reducing the priority of the corresponding template for low-scoring answers or adjusting the temperature parameter of the natural language generation model to control the output diversity; at the same time, introducing an active learning mechanism, by calculating the KL divergence of the model prediction confidence and the distribution of the labeled data, screening out the unlabeled samples with the largest amount of information, and preferentially pushing them to the manual labeling platform for iterative training; combining incremental learning technology to regularly use the newly labeled data to fine-tune the pre-trained model to avoid resource consumption of full training, and balancing the model stability and adaptability by dynamically adjusting the learning rate and batch size, and ultimately achieving the self-evolution of the system in continuous interaction;

[0029] The policy gradient dynamic weight formula is:

[0030]

[0031] in:

[0032] Δw represents the amount of model weight update, which directly determines the direction and magnitude of strategy adjustment.

[0033] η represents the reinforcement learning rate, which is adaptively adjusted according to the volatility of recent user feedback. The greater the feedback volatility, the smaller the η value is to suppress oscillation.

[0034] Reward represents the normalized reward value mapped to the user rating. The higher the rating, the closer the reward is to 1, and vice versa.

[0035] represents the policy gradient, which indicates the sensitivity of the probability of action a in state s to the weight w under the current policy.

[0036] λ represents the historical loss attenuation coefficient, which is dynamically calculated based on the average loss of the past N rounds of training. The higher the loss, the larger λ is to accelerate weight convergence.

[0037] history_loss represents the sliding average of historical training losses, which is used to suppress overfitting of the model to noise feedback.

[0038] This formula combines immediate user feedback with long-term training stability through a dual mechanism of reward-driven and historical constraints. The first term, guided by user ratings, drives the model towards high-reward strategies. The second term dynamically suppresses weight mutations through historical losses, preventing policy oscillations caused by a single low rating. Compared to traditional policy gradient algorithms, this formula introduces adaptive λ and η, enabling the model to achieve both agility and robustness in complex interaction scenarios.

[0039] In a preferred embodiment, the policy adjustment and deployment submodule leverages a multi-armed bandit algorithm to dynamically optimize and smoothly roll out policies. First, the optimized new policy model is encapsulated as an independent service node and deployed to a grayscale environment via containerization, running in parallel with the baseline policy. Based on a real-time traffic distribution mechanism, the system directs some user requests to the new policy node. It also collects metrics such as response latency, answer accuracy, and user ratings. Using a Bayesian optimization algorithm, it calculates the expected return for each policy version and dynamically adjusts traffic weights to maximize overall performance. During deployment, a monitoring component continuously tracks service health. If it detects abnormal metrics, such as a sudden increase in error rate or latency exceeding a threshold, it automatically triggers a version rollback and notifies operations personnel to intervene. For proven new policies, hot reloading is used to update online model parameters and rule bases, dynamically adjusting semantic similarity thresholds or answer generation template priorities, enabling iterations without service interruption. An adaptive learning rate mechanism driven by historical data further ensures that the magnitude of model weight updates matches the current system stability, avoiding service fluctuations caused by aggressive adjustments. Ultimately, this ensures both agile policy evolution and high system availability.

[0040] In a preferred embodiment, the system monitoring and evaluation module is internally provided with a real-time performance dashboard and automated auditing tools. The performance dashboard tracks key indicators, including average response latency, query hit rate, and user satisfaction score, stores historical data in a time series database, and generates trend reports. The automated auditing tool embeds an anomaly detection algorithm, an isolation forest model identifies sudden increases in invalid queries, and a rule engine intercepts potential malicious requests such as SQL injection attacks. The evaluation phase uses an A / B testing framework to compare the effects of different algorithm versions and the accuracy of traditional rule engines and deep learning models. The test results are analyzed through statistical significance to generate optimization suggestions, driving continuous iteration of the system.

[0041] In a preferred embodiment, the knowledge graph management module is internally provided with a data cleaning pipeline, an incremental update mechanism, and a distributed storage architecture. The data cleaning pipeline removes noise data through regular matching, filters non-technical forum comments, merges duplicate entries using an entity alignment algorithm, and unifies "Cisco router" and "Cisco router" into standardized terms. The incremental update mechanism regularly crawls the latest competition cases and technical documents, identifies new entities and relationships through difference comparison, and automatically adds configuration parameters for new firewalls. The distributed storage architecture is based on graph database sharding technology, partitioning and storing large-scale knowledge nodes according to dimensions such as device type and protocol level, and combining with a parallel computing engine to improve the response efficiency of complex queries, ensuring the high availability and scalability of the knowledge graph.

[0042] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0043] 1. In the present invention, the efficiency and accuracy of the knowledge service of the network management competition are significantly improved through the deep integration of knowledge management and intelligent interaction technology. The system uses knowledge graphs to structuredly store the relationship between massive network equipment, protocol configurations and fault cases. Combined with natural language processing and graph database retrieval technology, it can quickly locate the core knowledge points of complex problems, and automatically complete the technical term dependencies in fuzzy queries through semantic expansion, providing contestants with multi-dimensional and multi-strength knowledge support. The interactive optimization module dynamically captures user feedback and behavioral data, continuously optimizes the answer generation strategy and knowledge retrieval threshold, ensures that the system is adaptively adjusted in real-time interaction, and prioritizes the recommendation of verified solutions for high-frequency fault scenarios. At the same time, it actively explores potential problems to improve knowledge coverage, and ultimately realizes a full-link closed loop from knowledge storage to intelligent decision-making.

[0044] 2. The present invention demonstrates strong scenario adaptability and expansion potential through modular design and dynamic collaboration mechanism. The knowledge graph management module supports incremental updates and cleaning integration of multi-source heterogeneous data, and can timely integrate the latest competition cases and technical specifications to ensure the cutting-edge and authoritative nature of the knowledge base; the reinforcement learning and active learning framework embedded in the interactive optimization module enables the system to autonomously evolve its reasoning capabilities during continuous use, automatically identify knowledge blind spots and trigger graph completion through user ratings of answers. This self-feedback and self-iteration mechanism not only reduces manual maintenance costs, but also provides a technical foundation for cross-domain knowledge transfer, extending the network configuration rule reasoning logic to security policy optimization scenarios, and providing an integrated intelligent support platform for the teaching, training and actual combat of network management competitions. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a block diagram of the overall system of the present invention;

[0046] Figure 2 This is the internal system block diagram of the interactive optimization module in the present invention. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0048] Reference Figure 1-2 ,

[0049] A network management competition question-answering system based on a knowledge graph, comprising: a user interaction layer module, a question parsing module, a knowledge retrieval module, an answer generation module, an interaction optimization module, a system monitoring and evaluation module, and a knowledge graph management module;

[0050] The internal settings of the interactive optimization module include: feedback collection and annotation submodule, model optimization submodule and strategy adjustment and deployment submodule;

[0051] The front-end interface input port of the user interaction layer module is directly connected to the natural language processing interface of the question analysis module;

[0052] The structured output port of the question analysis module is connected to the query engine input port of the knowledge retrieval module through a standardized data protocol.

[0053] The result output of the knowledge retrieval module is connected to the data input interface of the answer generation module.

[0054] The multimodal rendering port of the answer generation module is connected back to the front-end display interface of the user interaction layer, returning the text and graphic answers to the user interface.

[0055] The feedback collection port of the user interaction layer is connected to the data input port of the interaction optimization module;

[0056] The strategy adjustment port of the interactive optimization module is embedded in the model parameter interface of the question parsing module, the similarity threshold regulator of the knowledge retrieval module, and the priority controller of the answer generation module.

[0057] The indicator collection end of the system monitoring and evaluation module is connected to the performance log output end of all modules through distributed probes

[0058] The data update interface of the knowledge graph management module bidirectionally connects the storage engine of the knowledge retrieval module and the external data source input pipeline.

[0059] The user interaction layer module is equipped with a visual front-end interface based on Web technology and a standardized API interface. The front-end interface adopts a responsive design and supports multi-terminal adaptation. The core components include a question input box, an answer display panel, and a user feedback control. The feedback control embeds a five-star rating and tag selection function to capture explicit evaluations. The API interface follows the RESTful specification and provides a request and response mechanism in JSON format to facilitate third-party system integration and automated calls. Interface interaction data is collected in real time through tracking technology to collect user behavior logs, page dwell time, answer expansion times, and question redirection paths. These logs are initially formatted and then transmitted to downstream modules for in-depth analysis to ensure comprehensive analysis of user intentions and efficient use of feedback.

[0060] The question parsing module is equipped with a natural language processing pipeline, including four core links: word segmentation, part-of-speech tagging, syntactic analysis, and intent recognition. The word segmentation link uses a hybrid dictionary and statistical model word segmentation tool, taking into account both professional terminology recognition and new word discovery capabilities; the part-of-speech tagging link uses a hidden Markov model and conditional random fields to jointly annotate entity types, with network device names as nouns and configuration actions as verbs. The syntactic analysis link uses a dependency tree algorithm to parse the subject-verb-object structure and extract the core entities and relationships in the question; the intent recognition link integrates rule templates and deep learning classifiers to map questions to preset categories for fault diagnosis, configuration queries, or performance optimization. The parsed structured data is encapsulated into a unified query instruction to provide accurate input for knowledge retrieval.

[0061] The knowledge retrieval module is internally equipped with a dual mechanism: a graph database query engine and a semantic expansion algorithm. The graph database query engine constructs dynamic search statements based on the Cypher language, quickly locating associated nodes in the knowledge graph based on parsed entities and relationships, and matching the relationship chain between router models and compatible protocols. The semantic expansion algorithm uses a pre-trained language model to calculate the cosine similarity between user questions and knowledge base content, generalizing and completing fuzzy queries, expanding "routing setup issues" to related scenarios such as "static routing configuration errors" and "dynamic routing protocol conflicts." After relevance sorting and redundancy filtering, the search results output high-confidence entity attributes, relationship paths, and contextual descriptions, providing multi-dimensional data support for answer generation.

[0062] The answer generation module is internally configured with a dual-track strategy consisting of a template matching engine and a natural language generation model. The template matching engine predefines the answer framework for multiple scenarios. The fault template includes cause location, solution steps, and prevention suggestions. The search results are injected into the template through slot filling technology to generate standardized answers. The natural language generation model trains end-to-end text generation capabilities based on the Transformer architecture, dynamically combines knowledge fragments for complex reasoning scenarios, and integrates multiple device configuration conflicts into a coherent solution description. The answer output stage introduces multimodal rendering technology, which supports the mixed display of text, flowcharts, and code blocks. When explaining the network topology, a visual connection diagram is generated simultaneously to improve the comprehensibility and operational guidance of the answer.

[0063] The feedback collection and annotation submodule directly captures explicit user feedback through front-end interactive controls, while also tracking user behavior to infer implicit feedback. Raw data is filtered for noise using regular expression matching of high-frequency keywords. An active learning algorithm then selects two types of high-value samples: low-certainty samples whose model prediction confidence approaches a threshold, and high-diversity samples whose semantic differences from the annotated data exceed a preset ratio. These selected samples are then pushed to the annotation platform for manual annotation based on priority. Ultimately, they are integrated into a structured dataset consisting of three tuples: annotation results, raw logs, and behavioral metrics, for use in downstream model iteration and optimization.

[0064] The marking priority formula is:

[0065]

[0066] Where: α represents the weight coefficient of high-frequency words, which is dynamically calculated based on the frequency of occurrence of regular matching keywords in historical data. The higher the frequency, the larger the α value.

[0067] f k The number of times high-frequency keywords are matched in the current sample is logarithmically scaled to reduce the long-tail distribution bias.

[0068] β represents the confidence sensitivity coefficient, which is preset with a fixed value according to the model type (such as classification or retrieval model) and is used to amplify the impact of low-confidence samples.

[0069] c represents the model's confidence in predicting the sample, τ is the preset threshold (such as 0.5), and the smaller the difference, the more uncertain the model is.

[0070] γ represents the diversity gain coefficient, which is adaptively adjusted based on the semantic coverage of the current annotated dataset. The lower the coverage, the higher the \gammaγ value.

[0071] sim(s,Slabeled) represents the average semantic similarity between the current sample s and the labeled dataset Slabeled, and the cosine similarity is calculated using the pre-trained word vector.

[0072] The model optimization submodule dynamically adjusts the core model parameters of the question-answering system through user feedback data to improve the accuracy of intent recognition and answer generation. The specific process includes: based on the reinforcement learning framework, converting the user's answer rating into a reward signal, using the policy gradient algorithm to optimize the decision weight of the answer generation model, reducing the priority of the corresponding template for low-scoring answers or adjusting the temperature parameter of the natural language generation model to control the output diversity; at the same time, an active learning mechanism is introduced to screen out the unlabeled samples with the largest amount of information by calculating the KL divergence of the model prediction confidence and the distribution of the labeled data, and preferentially push them to the manual labeling platform for iterative training; combined with incremental learning technology, the pre-trained model is regularly fine-tuned using the newly labeled data to avoid the resource consumption of full training, and the stability and adaptability of the model are balanced by dynamically adjusting the learning rate and batch size, ultimately achieving the self-evolution of the system in continuous interaction;

[0073] The policy gradient dynamic weight formula is:

[0074]

[0075] in:

[0076] Δw represents the amount of model weight update, which directly determines the direction and magnitude of strategy adjustment.

[0077] η represents the reinforcement learning rate, which is adaptively adjusted according to the volatility of recent user feedback. The greater the feedback volatility, the smaller the η value is to suppress oscillation.

[0078] Reward represents the normalized reward value mapped to the user rating. The higher the rating, the closer the reward is to 1, and vice versa.

[0079] represents the policy gradient, which indicates the sensitivity of the probability of action a in state s to the weight w under the current policy.

[0080] λ represents the historical loss attenuation coefficient, which is dynamically calculated based on the average loss of the past N rounds of training. The higher the loss, the larger λ is to accelerate weight convergence.

[0081] history_loss represents the sliding average of historical training losses, which is used to suppress overfitting of the model to noise feedback.

[0082] This formula combines immediate user feedback with long-term training stability through a dual mechanism of reward-driven and historical constraints. The first term, guided by user ratings, drives the model towards high-reward strategies. The second term dynamically suppresses weight mutations through historical losses, preventing policy oscillations caused by a single low rating. Compared to traditional policy gradient algorithms, this formula introduces adaptive λ and η, enabling the model to achieve both agility and robustness in complex interaction scenarios.

[0083] The policy adjustment and deployment submodule leverages a multi-armed bandit algorithm to dynamically tune and smoothly roll out policies. The optimized new policy model is first packaged as an independent service node and deployed to a grayscale environment using containerization, running in parallel with the baseline policy. Based on a real-time traffic distribution mechanism, the system directs some user requests to the new policy node. Metrics such as response latency, answer accuracy, and user ratings are collected. A Bayesian optimization algorithm is used to calculate the expected return for each policy version, dynamically adjusting traffic weights to maximize overall performance. During deployment, a monitoring component continuously tracks service health. If anomalies are detected, such as a sudden increase in error rate or latency exceeding a threshold, a version rollback is automatically triggered and operations personnel are notified for investigation. For proven new policies, hot reloading is used to update online model parameters and rule bases, dynamically adjusting semantic similarity thresholds and answer generation template priorities, enabling iterations without service interruption. An adaptive learning rate mechanism driven by historical data further ensures that model weight updates align with current system stability, avoiding service fluctuations caused by aggressive adjustments. Ultimately, this ensures both agile policy evolution and high system availability.

[0084] The system monitoring and evaluation module includes a real-time performance dashboard and automated auditing tools. The performance dashboard tracks key metrics, including average response latency, query hit rate, and user satisfaction scores. It stores historical data in a time-series database and generates trend reports. The automated auditing tools incorporate anomaly detection algorithms, using an isolation forest model to identify sudden increases in invalid queries, and a rules engine to intercept potentially malicious requests such as SQL injection attacks. The evaluation phase utilizes an A / B testing framework to compare the performance of different algorithm versions and the accuracy of traditional rule engines and deep learning models. Statistical significance analysis of test results generates optimization recommendations, driving continuous system iteration.

[0085] The knowledge graph management module incorporates a data cleaning pipeline, an incremental update mechanism, and a distributed storage architecture. The data cleaning pipeline uses regular matching to remove noisy data, filter non-technical forum comments, and merge duplicate entries using an entity alignment algorithm, standardizing the terms "Cisco router" and "Cisco router." The incremental update mechanism regularly crawls the latest competition cases and technical documentation, identifies new entities and relationships through difference comparison, and automatically adds configuration parameters for new firewalls. The distributed storage architecture, based on graph database sharding technology, partitions and stores large-scale knowledge nodes by device type, protocol level, and other dimensions. Combined with a parallel computing engine, it improves the efficiency of complex query responses, ensuring the high availability and scalability of the knowledge graph.

[0086] From the above we can know:

[0087] In the present invention, the efficiency and accuracy of the knowledge service of the network management competition are significantly improved through the deep integration of knowledge management and intelligent interaction technology. The system uses knowledge graphs to structuredly store the relationship between massive network equipment, protocol configurations and fault cases. Combined with natural language processing and graph database retrieval technology, it can quickly locate the core knowledge points of complex problems, and automatically complete the technical term dependencies in fuzzy queries through semantic expansion, providing contestants with multi-dimensional and multi-strength knowledge support. The interactive optimization module dynamically captures user feedback and behavioral data, continuously optimizes the answer generation strategy and knowledge retrieval threshold, ensures that the system is adaptively adjusted in real-time interaction, and prioritizes the recommendation of verified solutions for high-frequency fault scenarios. At the same time, it actively explores potential problems to improve knowledge coverage, and ultimately realizes a full-link closed loop from knowledge storage to intelligent decision-making.

[0088] In the present invention, through modular design and dynamic collaboration mechanism, it demonstrates strong scenario adaptability and expansion potential. The knowledge graph management module supports incremental updates and cleaning integration of multi-source heterogeneous data, and can timely integrate the latest competition cases and technical specifications to ensure the cutting-edge and authoritative nature of the knowledge base; the reinforcement learning and active learning framework embedded in the interactive optimization module enables the system to autonomously evolve its reasoning capabilities during continuous use, and automatically identify knowledge blind spots and trigger graph completion through user ratings of answers. This self-feedback and self-iteration mechanism not only reduces manual maintenance costs, but also provides a technical foundation for cross-domain knowledge transfer, extending the network configuration rule reasoning logic to security policy optimization scenarios, and providing an integrated intelligent support platform for the teaching, training and actual combat of network management competitions.

[0089] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprises" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements that are not explicitly listed, or also includes elements that are inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device that includes the element.

[0090] The above description is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A network management competition question-answering system based on knowledge graph, characterized by: The system includes: a user interaction layer module, a question analysis module, a knowledge retrieval module, an answer generation module, an interaction optimization module, a system monitoring and evaluation module, and a knowledge graph management module; The interactive optimization module is internally configured with: a feedback collection and annotation submodule, a model optimization submodule, and a strategy adjustment and deployment submodule; The front-end interface input port of the user interaction layer module is directly connected to the natural language processing interface of the question analysis module; The structured output port of the problem analysis module is connected to the query engine input port of the knowledge retrieval module through a standardized data protocol; The result output terminal of the knowledge retrieval module is connected to the data input interface of the answer generation module. The multimodal rendering port of the answer generation module is connected back to the front-end display interface of the user interaction layer to return the text and graphic answers to the user interface; The feedback collection port of the user interaction layer is connected to the data input port of the interaction optimization module; The strategy adjustment port of the interactive optimization module is respectively embedded with the model parameter interface of the question analysis module, the similarity threshold regulator of the knowledge retrieval module and the priority controller of the answer generation module. The indicator collection end of the system monitoring and evaluation module is connected to the performance log output end of all modules through distributed probes The data update interface of the knowledge graph management module is bidirectionally connected to the storage engine of the knowledge retrieval module and the external data source input pipeline.

2. The network management competition question-answering system based on knowledge graph according to claim 1, characterized in that: The user interaction layer module is equipped with a visual front-end interface based on Web technology and a standardized API interface. The front-end interface adopts a responsive design and supports multi-terminal adaptation. The core components include a question input box, an answer display panel, and a user feedback control. The feedback control is embedded with a five-star rating and tag selection function to capture explicit evaluations. The API interface follows the RESTful specification and provides a request-response mechanism in JSON format, facilitating third-party system integration and automated calls.

3. The network management competition question-answering system based on knowledge graph according to claim 1, characterized in that: The question parsing module is equipped with a natural language processing pipeline, including four core links: word segmentation, part-of-speech tagging, syntactic analysis and intent recognition. The word segmentation link adopts a word segmentation tool that combines a hybrid dictionary and a statistical model, taking into account both professional terminology recognition and new word discovery capabilities. The part-of-speech tagging link jointly labels entity types through a hidden Markov model and a conditional random field. The syntactic analysis link parses the subject-verb-object structure based on the dependency tree algorithm to extract the core entities and relationships in the question. The intent recognition link integrates rule templates and deep learning classifiers to map questions to preset categories. The parsed structured data is encapsulated as a unified query instruction to provide accurate input for knowledge retrieval.

4. The network management competition question-answering system based on knowledge graph according to claim 1, characterized in that: The knowledge retrieval module is internally equipped with a dual mechanism of a graph database query engine and a semantic expansion algorithm. The graph database query engine constructs dynamic search statements based on the Cypher language, and quickly locates related nodes in the knowledge graph according to the parsed entities and relationships. The semantic expansion algorithm uses a pre-trained language model to calculate the cosine similarity between user questions and knowledge base content, and generalizes and completes fuzzy queries. After the search results are sorted by relevance and filtered for redundancy, they output high-confidence entity attributes, relationship paths and context descriptions, providing multi-dimensional data support for answer generation.

5. The network management competition question-answering system based on knowledge graph according to claim 1, characterized in that: The answer generation module is internally configured with a dual-track strategy of a template matching engine and a natural language generation model; the template matching engine predefines answer frameworks for multiple scenarios, and the fault templates include cause location, solution steps, and prevention suggestions. The retrieval results are injected into the template through slot filling technology to generate standardized answers; the natural language generation model trains end-to-end text generation capabilities based on the Transformer architecture, dynamically combines knowledge fragments for complex reasoning scenarios, and integrates multiple device configuration conflicts into a coherent solution description; the answer output link introduces multimodal rendering technology, supports the mixed display of text, flowcharts, and code blocks, and simultaneously generates a visual connection diagram when explaining the network topology, thereby improving the comprehensibility and operational guidance of the answers.

6. The network management competition question-answering system based on knowledge graph according to claim 1, characterized in that: The feedback collection and annotation submodule directly captures users' explicit comments through front-end interactive controls, while simultaneously recording user behavior to infer implicit feedback. The raw data is filtered for noise using regular expression matching of high-frequency keywords, and two types of high-value samples are screened based on an active learning algorithm: low-certainty samples whose model prediction confidence is close to a threshold, and high-diversity samples whose semantic differences from the annotated data exceed a preset ratio. The screened samples are pushed to the annotation platform, where they are manually annotated according to priority and ultimately integrated into a triplet structured dataset containing annotation results, raw logs, and behavioral indicators for iterative optimization of downstream models. The marking priority formula is: Where: α represents the weight coefficient of high-frequency words, which is dynamically calculated based on the frequency of occurrence of regular matching keywords in historical data. The higher the frequency, the larger the α value. f k The number of times a high-frequency keyword matches in the current sample is logarithmically scaled to reduce the long-tail distribution bias; β represents the confidence sensitivity coefficient, which is preset with a fixed value according to the model type (such as classification or retrieval model) and is used to amplify the impact of low-confidence samples; c represents the model's confidence in predicting the sample, τ is a preset threshold (such as 0.5), and the smaller the difference, the more uncertain the model is; γ represents the diversity gain coefficient, which is adaptively adjusted based on the semantic coverage of the current annotated dataset. The lower the coverage, the higher the \gammaγ value; sim(s,Slabeled) represents the average semantic similarity between the current sample s and the labeled dataset Slabeled, and the cosine similarity is calculated using the pre-trained word vector.

7. The network management competition question-answering system based on knowledge graph according to claim 1, characterized in that: The model optimization submodule dynamically adjusts the core model parameters of the question-answering system based on user feedback data to improve the accuracy of intent recognition and answer generation. The specific process includes: based on the reinforcement learning framework, converting user ratings of answers into reward signals, using the policy gradient algorithm to optimize the decision weights of the answer generation model, reducing the priority of corresponding templates for low-rated answers or adjusting the temperature parameters of the natural language generation model to control output diversity; at the same time, an active learning mechanism is introduced to screen out the unlabeled samples with the most information by calculating the KL divergence of the model prediction confidence and the distribution of labeled data, and preferentially pushing them to the manual labeling platform for iterative training; The policy gradient dynamic weight formula is: in: Δw represents the amount of model weight update, which directly determines the direction and magnitude of strategy adjustment; η represents the reinforcement learning rate, which is adaptively adjusted according to the volatility of recent user feedback. The greater the feedback volatility, the smaller the η value is to suppress oscillation; Reward represents the normalized reward value mapped to the user rating. The higher the rating, the closer the reward is to 1, and vice versa. represents the policy gradient, which indicates the sensitivity of the probability of action a in state s to the weight w under the current policy; λ represents the historical loss attenuation coefficient, which is dynamically calculated based on the average loss of the past N rounds of training. The higher the loss, the larger λ is to accelerate weight convergence; history_loss represents the sliding average of historical training losses, which is used to suppress overfitting of the model to noise feedback; This formula combines immediate user feedback with long-term training stability through a dual mechanism of reward-driven and historical constraints. The first term is guided by user ratings, driving the model to iterate towards high-reward strategies. The second term dynamically suppresses weight mutations through historical losses to avoid policy fluctuations caused by a single low rating. Compared with traditional policy gradient algorithms, this formula introduces adaptive λ and η, allowing the model to achieve both agility and robustness in complex interaction scenarios.

8. The network management competition question-answering system based on knowledge graph according to claim 1, characterized in that: The strategy adjustment and deployment submodule uses the multi-armed bandit algorithm as the core to achieve dynamic tuning and smooth launch of the strategy. First, the optimized new strategy model is encapsulated as an independent service node, deployed to the grayscale environment through containerization technology, and runs in parallel with the baseline strategy; based on the real-time traffic distribution mechanism, the system directs some user requests to the new strategy node, and at the same time collects response delay, answer accuracy and user rating indicators, uses the Bayesian optimization algorithm to calculate the expected benefits of each strategy version, and dynamically adjusts the traffic weight to maximize overall performance; during the deployment process, the monitoring component continuously tracks the health status of the service. If abnormal indicators are detected, such as a sudden increase in error rate or delay exceeding the threshold, the version rollback is automatically triggered, and the operation and maintenance personnel are notified to intervene and investigate.

9. The network management competition question-answering system based on knowledge graph according to claim 1, characterized in that: The system monitoring and evaluation module is internally equipped with a real-time performance dashboard and automated audit tools. The performance dashboard tracks key indicators, including average response latency, query hit rate, and user satisfaction score, and stores historical data in a time series database to generate trend reports. Automated audit tools embed anomaly detection algorithms, use the isolation forest model to identify sudden increases in invalid queries, and use the rule engine to intercept potential malicious requests such as SQL injection attacks. The evaluation phase uses an A / B testing framework to compare the effectiveness of different algorithm versions and the accuracy of traditional rule engines and deep learning models. The test results are analyzed through statistical significance to generate optimization suggestions, driving continuous iteration of the system.

10. The network management competition question-answering system based on knowledge graph according to claim 1, characterized in that: The knowledge graph management module is internally equipped with a data cleaning pipeline, an incremental update mechanism, and a distributed storage architecture; The data cleaning pipeline removes noisy data through regular matching, filters non-technical forum comments, merges duplicate entries using an entity alignment algorithm, and unifies "Cisco router" and "Cisco router" into standardized terms; the incremental update mechanism regularly crawls the latest competition cases and technical documents, identifies new entities and relationships through difference comparison, and automatically adds configuration parameters for new firewalls; the distributed storage architecture is based on graph database sharding technology, partitioning and storing large-scale knowledge nodes according to dimensions such as device type and protocol level, and combining it with a parallel computing engine to improve the response efficiency of complex queries and ensure the high availability and scalability of the knowledge graph.

Images