Large model rapid question-answering method and system based on traffic engineering image-text database

Abstract

The invention belongs to the technical field of model questioning and answering, and discloses a large-model rapid questioning and answering method and system based on a traffic engineering image-text database. Comprising the following steps: establishing a database, and clustering the database to construct a traffic engineering problem type set in each business field; obtaining a question language input by a user, and matching a user question type; analyzing the data in the question type of the user and a question language input by the user to generate a plurality of question answers, calculating a confidence score of each question answer, selecting the question answers based on the confidence scores, and recommending the question answers to the user; obtaining user feedback data according to the push result, analyzing the user feedback data, generating an update judgment coefficient, and executing an update instruction based on the update judgment coefficient; according to the method, whether updating is needed or not is judged according to the quantitative score of the feedback item type, the fluctuation coefficient and the updating judgment coefficient generated by the accumulation coefficient, so that the updating of the business domain module is more targeted, and resource waste caused by blind updating can be effectively avoided.

Description

Technical Field

[0001] This invention relates to the field of model question answering technology, and more specifically, to a method and system for rapid question answering of large models based on a traffic engineering graphic database. Background Technology

[0002] Traffic engineering, as a core area supporting urban operation and infrastructure construction, encompasses multiple business scenarios including urban roads, highways, bridges and tunnels, and rail transit, involving a vast amount of professional specifications, engineering cases, design drawings, and technical standards. With the deepening of digital transformation in the industry, the demand for efficient access to professional knowledge and rapid problem-solving among engineering technicians, managers, and researchers is becoming increasingly urgent. Intelligent question-and-answer systems have become a crucial technological support for improving industry efficiency. Traditional question-answering systems often employ keyword matching or shallow semantic analysis, which fails to accurately identify specialized entities and complex logical relationships within the transportation engineering field. This often results in irrelevant answers or matching to unrelated question types, demonstrating insufficient professional adaptability and failing to meet the high accuracy requirements of engineering scenarios. Furthermore, traditional question-answering systems generate answers solely based on text retrieval, lacking the logical support of structured knowledge graphs. This leads to disorganized answer structures and imprecise professional logic, making it difficult for users to verify the compliance and credibility of the answers. Additionally, existing systems rely on manual maintenance for updates, lacking an automated iteration mechanism based on user feedback. This prevents timely responses to knowledge updates and changes in user needs within the transportation engineering field, resulting in poor system adaptability.

[0003] In view of this, the present invention proposes a rapid question-answering method and system based on a large model of traffic engineering graphic database to solve the above problems. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art and achieve the above objectives, the present invention provides the following technical solution: A fast question-answering method based on a large-scale model of a traffic engineering image and text database, including the following methods: The structured and unstructured data of traffic engineering in various business areas are integrated to form a database. The database is clustered to obtain multiple traffic engineering problem types. Based on the traffic engineering problem types, a set of traffic engineering problem types for each business area is constructed. Obtain the question language input by the user, preprocess the question language and input it into the traffic engineering question type set, match the corresponding traffic engineering question type, and record it as the user question type; The system analyzes the data within the user's question type and the language of the question input by the user to obtain the first type of evidence and the second type of evidence. Based on the first type of evidence and the second type of evidence, it generates multiple question answers, calculates the confidence score of each question answer, and selects question answers to recommend to the user based on the confidence scores. Based on the push notification results, user feedback data is obtained, the user feedback data is analyzed, update judgment coefficients are generated, and update instructions are executed based on the update judgment coefficients.

[0005] Furthermore, the method for constructing the traffic engineering problem type set under each business domain is as follows: Traffic image and text data from various business domains are retrieved from the database. Each piece of traffic image and text data is treated as a sample, and a traffic engineering image and text sample set for the corresponding business domain is constructed based on all samples. After preprocessing the traffic engineering image and text sample sets for each business domain, feature extraction is performed to obtain multiple problem feature parameters and image-text matching feature parameters. A problem-image-text matching coupling neighborhood radius and a minimum core problem sample number threshold are set. Within the problem-image-text matching coupling neighborhood radius, thresholds for each problem feature parameter and each image-text matching feature parameter are also set. All samples in the traffic engineering image and text sample set are traversed, and based on the set problem-image-text matching coupling neighborhood radius... The domain radius and minimum core problem sample number threshold are used to determine the core problem points, boundary problem points, and noise problem points in the traffic engineering image and text sample set. After removing the noise problem points, the core problem points and boundary problem points are merged to form a sample set to be clustered. Density clustering algorithm is used to perform clustering operations on the sample set to be clustered, resulting in multiple independent clusters. Each independent cluster represents a type of traffic engineering problem. Based on the problem feature parameters and image and text matching feature parameters contained in each independent cluster, each independent cluster is named to obtain the corresponding traffic engineering problem type. Based on all traffic engineering problem types, a traffic engineering problem type set under each business domain is constructed.

[0006] Furthermore, the method for matching user question types is as follows: A pre-constructed traffic engineering terminology knowledge graph is used, where nodes represent professional terms and edges represent logical relationships between them. Each professional term node in the knowledge graph is assigned a pre-defined domain weight, and the logical relationships between professional terms are pre-defined with a pre-defined association strength. Based on the user-selected business domain, professional terms are extracted from each traffic engineering problem type within the corresponding business domain's traffic engineering problem type set. Based on the extracted professional terms, a set of type-specific professional terms is constructed for each traffic engineering problem type. The user's input question language is identified, yielding multiple professional terms, which are denoted as user question terms. A user-specific professional term set is constructed based on these user question terms. Based on the type-specific professional term set and the user-specific professional term set, a score is calculated for each traffic engineering problem type, and the traffic engineering problem type with the highest score is selected as the user question type.

[0007] Furthermore, the method for calculating the score for each type of traffic engineering problem is as follows: The process involves obtaining the frequency of each term in the user's terminology set within the corresponding terminology sets for each traffic engineering problem type. The frequencies of all terms are summed to obtain the total frequency for each traffic engineering problem type. This total frequency is then normalized to obtain the first score for each traffic engineering problem type. The user's terminology set and the terms in each terminology set are mapped to a traffic engineering terminology knowledge graph. For each term in the user's terminology set, it is paired with each term in each terminology set to form multiple term pairs. Based on each term pair... For each term pair, the shortest path between the two professional terms in the traffic engineering terminology knowledge graph is found. The product of the association strength of all edges on the shortest path is calculated, and then the product is multiplied by the domain weights of the two professional terms in the term pair and the path association coefficient to obtain the association score of each term pair. The association scores of all term pairs corresponding to the same traffic engineering problem type are averaged to obtain the mean association score, and the mean association score is normalized to obtain the second score for each traffic engineering problem type. The first score and the second score are weighted and summed to obtain the score value for each traffic engineering problem type.

[0008] Furthermore, the method for generating answers to multiple questions is as follows: Based on traffic image and text data within the user question type, each traffic image and text data is encoded to obtain a modal vector. A multimodal vector database is built based on all modal vectors. Simultaneously, core entities, their attributes, and logical relationships between them are extracted from all traffic image and text data within the user question type to construct a structured traffic engineering knowledge graph. A bidirectional association index is established between the multimodal vector database and the structured traffic engineering knowledge graph. Based on all traffic image and text data within the user question type, multiple sub-question types are defined, and a corresponding structured answer template is configured for each sub-question type. The structured answer template includes a keyword framework for guiding knowledge retrieval, a logical framework for organizing answer output, and a standard index for associating authoritative evidence. The system receives user-input question language and matches it to the corresponding sub-question type using a pre-defined classification model. Core entities are extracted from the user-input question language and denoted as user-extracted entities. The user-extracted entities are populated into the structured answer template corresponding to the sub-question type, generating a retrieval query instruction containing the retrieval target, constraints, and output requirements. Based on the retrieval query instruction, a bidirectional retrieval operation is performed, using the user-extracted entities in the retrieval query instruction and the keywords extracted from the keyword framework as retrieval terms. Semantic retrieval is performed in the multimodal vector database, calculating the cosine similarity of each modality vector in the multimodal vector database. Modality vectors with a cosine similarity greater than a preset cosine similarity threshold are taken as the first type of evidence. Starting from the user-extracted entities in the retrieval query instruction, a multi-hop association traversal query is performed in the structured traffic engineering knowledge graph to obtain the core entities, attributes, and logical relationships associated with the user-extracted entities, which are taken as the second type of evidence. The first and second types of evidence are input into a preset large language model, and based on the sub-question type structured answer template, the first and second types of evidence are merged to finally generate multiple question answers.

[0009] Furthermore, the method for calculating the confidence score of each question's answer is as follows: Each question answer is converted into a question semantic vector, and the user-input question language is converted into a user semantic vector. The cosine similarity between the user semantic vector and each question semantic vector is calculated, and the normalized cosine similarity is used as the semantic similarity score for each question answer. Multiple core entities are extracted from the user-input question language to form a first core entity set. Multiple core entities are extracted from each question answer to form a second core entity set for each question answer. All core entities in the structured traffic engineering knowledge graph are merged to form a third core entity set. The intersection of the second and third core entity sets is used as the key entity set. Each core entity in the first core entity set is denoted as... The first core entity is defined as each core entity in the key entity set. Any first core entity is combined with any key core entity to form an entity pair. A shortest path query algorithm is executed in the structured traffic engineering knowledge graph to obtain the shortest relation hop count for each entity pair. The average shortest relation hop count for all entity pairs corresponding to the answer to the same question is calculated, and the average shortest relation hop count is normalized to obtain the coverage score for each question answer. The number of references to the normative standard index in each question answer is obtained, and the number of references is normalized to obtain the evidence citation score for each question answer. The semantic similarity score, coverage score, and evidence citation score are weighted and summed to obtain the confidence score for each question answer.

[0010] Furthermore, the method for recommending question answers to users based on confidence scores is as follows: Based on the confidence scores of each question's answer, the questions and answers are sorted in descending order of confidence scores to obtain a question-answer recommendation table. The question-answer with the highest ranking in the recommendation table is recommended to the user. If the recommended question-answer does not meet the user's needs, the recommendations are iteratively updated according to the order of the question-answer recommendation table until the user's needs are met.

[0011] Furthermore, the update instructions include business domain module updates and large model updates; among them, business domain module updates refer to updating the question-answering business domain modules in the large language model, and large model updates refer to updating the entire large language model; obtain the update judgment coefficient of each business domain, and compare it with the update judgment coefficient threshold corresponding to each business domain. If the update judgment coefficient is less than the update judgment coefficient threshold, then the corresponding business domain module is updated. Obtain the total number of updates in the question-answering business domain module of the large language model, denoted as the update count. Compare the update count with the set update count threshold. If the update count is greater than the update count threshold, then perform a large model update.

[0012] Furthermore, the method for generating the updated judgment coefficients is as follows: Pre-build feedback data tables for each business area. Each feedback data table contains multiple feedback item types, and each feedback item type has a corresponding quantitative score range. Obtain the quantitative score of each feedback item type after each user feedback in each business area. Then, weightedly sum the quantitative scores of all feedback item types in the same business area to obtain the quantitative comprehensive score for each business area. The process involves obtaining the quantitative comprehensive score for each business area under a time series. Based on the quantitative comprehensive score under the time series, the standard deviation of the quantitative comprehensive score is calculated, and the standard deviation is normalized to obtain the volatility coefficient for each business area. Based on the quantitative comprehensive score under the time series, a quantitative comprehensive score function is constructed, and the integral value of the quantitative comprehensive score function under the time series is calculated. The integral value is normalized to obtain the cumulative coefficient for each business area. The volatility coefficient is incremented by one to obtain the benchmark volatility coefficient, and the cumulative coefficient is divided by the benchmark volatility coefficient to obtain the update judgment coefficient for each business area.

[0013] A large-scale rapid question-answering system based on a traffic engineering image and text database, implementing the aforementioned large-scale rapid question-answering method based on a traffic engineering image and text database, includes: The problem type set construction module is used to integrate structured and unstructured data of traffic engineering in various business areas to form a database. The database is clustered to obtain multiple traffic engineering problem types. Based on the traffic engineering problem types, a traffic engineering problem type set for each business area is constructed. The question type matching module is used to obtain the question language input by the user, preprocess the question language and input it into the traffic engineering question type set, match the corresponding traffic engineering question type, and record it as the user question type. The question-answer generation module analyzes the data within the user's question type and the language of the question input by the user to obtain the first type of evidence and the second type of evidence. Based on the first type of evidence and the second type of evidence, it generates multiple question answers, calculates the confidence score of each question answer, and selects question answers to recommend to the user based on the confidence scores. The update judgment module is used to obtain user feedback data based on the push results, analyze the user feedback data, generate update judgment coefficients, and execute update instructions based on the update judgment coefficients.

[0014] The technical effects and advantages of this invention, based on a large-scale rapid question-answering method and system using a traffic engineering graphic database, are as follows: This invention employs a dual-dimensional scoring mechanism—combining literal frequency matching of user-specific terminology with question-type terminology and knowledge graph-driven professional logical association calculation—to perform question type matching. This significantly improves the accuracy and professional suitability of traffic engineering question type matching. Simultaneously, when generating question answers, it utilizes a bidirectional retrieval operation involving dense vector semantic retrieval and multi-hop traversal of the knowledge graph to obtain dual evidence from both the factual and professional logical perspectives. This ensures both the semantic fit between the evidence and the user's question and strengthens the professional logical support of the answer, making the generated question answers more professional and rigorous. This invention uses a weighted scoring system that integrates three dimensions—semantic similarity, professional knowledge coverage, and authoritative evidence citation—to push generated answers to questions. This not only effectively avoids irrelevant answers caused by single semantic matching, but also accurately measures the logical relevance and authority of the answers within the professional knowledge system, making the confidence score more scientific and interpretable. This invention uses an update judgment coefficient generated from the quantitative score, fluctuation coefficient, and cumulative coefficient of the feedback project type to determine whether an update is needed. This makes updates to business domain modules more targeted and effectively avoids the waste of resources caused by blind updates. At the same time, by triggering large-scale model updates through a threshold for the number of module updates, it not only ensures the efficiency of local business optimization but also promotes the iteration of the overall system capabilities when needs in multiple domains are upgraded. This continuously improves the accuracy, professionalism, and user experience of the answers, ensuring that the system always adapts to knowledge updates and changes in user expectations in the field of transportation engineering. Attached Figure Description

[0015] Figure 1 This is a flowchart of the question-and-answer method of the present invention; Figure 2 This is a system module block diagram of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] In one embodiment, a fast question-answering method for large models based on a traffic engineering graphic database is disclosed, such as... Figure 1 As shown, the method includes: The structured and unstructured data of traffic engineering in various business areas are integrated to form a database. The database is clustered to obtain multiple traffic engineering problem types. Based on the traffic engineering problem types, a set of traffic engineering problem types for each business area is constructed. Obtain the question language input by the user, preprocess the question language and input it into the traffic engineering question type set, match the corresponding traffic engineering question type, and record it as the user question type; The system analyzes the data within the user's question type and the language of the question input by the user to obtain the first type of evidence and the second type of evidence. Based on the first type of evidence and the second type of evidence, it generates multiple question answers, calculates the confidence score of each question answer, and selects question answers to recommend to the user based on the confidence scores. Based on the push notification results, user feedback data is obtained, the user feedback data is analyzed, update judgment coefficients are generated, and update instructions are executed based on the update judgment coefficients.

[0018] Through the above technical solution, this application first integrates structured data (such as industry standard clauses, design parameter tables, engineering acceptance standards, and equipment model parameters) and unstructured data (such as engineering drawings, construction logs, case reports, technical papers, on-site video frames, and expert interview minutes) from various business areas in the field of traffic engineering, including urban road traffic management, highway maintenance, bridge and tunnel management, and intelligent control of rail transit, to construct a database. It then uses a density-based DBSCAN clustering algorithm to cluster the data, constructing a set of traffic engineering problem types for each business area, thereby clustering the data. Next, it obtains the user-input question language, preprocesses it, and inputs it into the traffic engineering problem set, matching it with the corresponding traffic engineering problem type, which is then recorded as the user question type. The preprocessing process includes word segmentation, noise reduction, and stop word filtering. During matching, a dual-dimensional scoring mechanism is used, combining literal frequency matching of user professional terminology and question type terminology with knowledge graph-driven professional logical association calculation, to perform question type matching, making the matching results more accurate. Based on the data under the corresponding question type and the user-input question language analysis, multiple question answers are generated, and based on each question... The system uses a confidence score to select the most reliable answer to recommend to the user. The confidence score is calculated by weighting three dimensions: semantic similarity score (cosine similarity between the user's question and answer), coverage score (average shortest relation hop count for entity pairs in the knowledge graph), and evidence citation score (number of explicit citations in standards and norms). This ensures the highest confidence score is pushed to the user, and the source of evidence and confidence score are displayed, improving the traceability and credibility of the answer. Finally, user feedback data is obtained based on the push results. This data is analyzed to generate update judgment coefficients, and update instructions are executed based on these coefficients. Update judgment coefficients generated from the quantitative scores, fluctuation coefficients, and cumulative coefficients of feedback project types in various business domains determine whether an update is needed. This makes updates to business domain modules more targeted and effectively avoids resource waste from blind updates. Simultaneously, a threshold for the number of module updates triggers large-scale model updates, ensuring the efficiency of local business optimizations while driving overall system capability iteration when multiple domain requirements upgrade. This continuously improves the accuracy, professionalism, and user experience of the answers, ensuring the system remains adaptable to knowledge updates and changing user expectations in the transportation engineering field.

[0019] The method for constructing traffic engineering problem type sets under each business domain is as follows: Traffic image and text data for each business domain are retrieved from the database. Each piece of traffic image and text data is treated as a sample, and a traffic engineering image and text sample set for the corresponding business domain is constructed based on all samples. After preprocessing the traffic engineering image and text sample sets for each business domain, feature extraction is performed to obtain multiple problem feature parameters and image-text matching feature parameters. A problem-image-text matching coupling neighborhood radius and a minimum core problem sample quantity threshold are set, and thresholds for each problem feature parameter and each image-text matching feature parameter are set within the problem-image-text matching coupling neighborhood radius. All samples in the traffic engineering image and text sample set are traversed, and based on the set parameters… By defining the problem-image-text matching coupling neighborhood radius and the minimum core problem sample number threshold, the problem core points, problem boundary points, and problem noise points in the traffic engineering image-text sample set are determined. After removing the problem noise points, the problem core points and problem boundary points are merged to form a sample set to be clustered. Density clustering algorithm is used to perform clustering operations on the sample set to be clustered, resulting in multiple independent clusters. Each independent cluster represents a type of traffic engineering problem. Based on the problem feature parameters and image-text matching feature parameters contained in each independent cluster, each independent cluster is named to obtain the corresponding traffic engineering problem type. Based on all traffic engineering problem types, a traffic engineering problem type set under each business domain is constructed.

[0020] The above scheme provides a specific method for constructing traffic engineering problem type sets under various business domains. This embodiment uses the "urban road traffic management" business domain as a specific example. All traffic graphic data belonging to the "urban road traffic management" business domain are obtained from the traffic engineering graphic database. Each piece of traffic graphic data is treated as a sample. Based on all samples, a traffic engineering graphic sample set for the corresponding business domain is constructed. These samples include graphic information related to traffic light settings, traffic signs and markings, lane management, etc. Then, the obtained traffic engineering graphic sample set is preprocessed, including image sharpening, optical character recognition (OCR), and feature extraction, ultimately yielding multiple problem feature parameters (such as traffic flow, accidents, etc.). The system uses traffic flow rate, congestion duration, pedestrian density, and other parameters to determine the matching characteristics of text and images (such as the degree of matching between text descriptions in images and actual traffic conditions, the degree of matching between signal phase diagrams and timing tables, and the degree of matching between monitoring screenshots and congestion text descriptions). Based on the experience and expertise of those in the field, a pre-set threshold for the problem-text-image matching coupling neighborhood and the minimum core problem sample size is established. Within this neighborhood, thresholds for each problem feature and each text-image matching feature are set (such as traffic flow thresholds, congestion duration thresholds, and thresholds for various matching degrees). All samples in the traffic engineering text-image sample set are iterated through, and based on the set problem-text-image matching coupling neighborhood radius and the minimum core problem sample size threshold, the system determines the matching characteristics of the traffic flow rate, congestion duration, pedestrian density, etc. This paper identifies the core problem points, boundary points, and noise points in the traffic engineering image and text sample set. The determination of these points is based on the core point, boundary point, and noise point partitioning in the DBSCAN clustering algorithm, combined with a pre-set problem-image / text matching coupling neighborhood radius and a minimum core problem sample number threshold. This partitioning method is existing technology and will not be elaborated upon here. After removing noise points, the core problem points and boundary points are merged to form a sample set to be clustered. A density-based clustering algorithm is then used to perform clustering operations on this sample set, resulting in multiple independent clusters. Each independent cluster represents a class of traffic engineering problems. The density-based clustering algorithm is also an existing technology in clustering methods. Without going into too much detail; based on the problem feature parameters and image-text matching feature parameters contained in each independent cluster, each independent cluster is named to obtain the corresponding traffic engineering problem type. For example, if the problem feature parameters of a certain independent cluster are high traffic flow, congestion duration, and traffic light timing during the morning rush hour, and its image-text matching feature parameters are high matching degree of signal phase diagram and high matching degree of peak congestion monitoring screenshot, then this independent cluster can be named the "Peak Hour Urban Road Intersection Traffic Light Timing Optimization Problem". Based on all traffic engineering problem types, a traffic engineering problem type set under the "Urban Road Traffic Management" domain is constructed; in this way, each traffic engineering problem type is associated with corresponding image-text data and feature parameters, which facilitates rapid question answering in the future.This approach not only integrates fragmented data into clusters of traffic engineering questions with clear professional attributes, addressing the pain point of traditional data classification that relies solely on keywords and fails to connect to specific professional scenarios, but also, by eliminating noisy data and merging core and boundary issues, it refines a precise set of business domain question types. This provides an efficient retrieval entry point for subsequent rapid question answering, significantly improving response speed and answer accuracy.

[0021] The method for matching user question types is as follows: A traffic engineering terminology knowledge graph is pre-constructed, where nodes represent professional terms and edges represent logical relationships between them. A domain weight is preset for each professional term node in the knowledge graph, and a correlation strength is preset for the logical relationships between each professional term. Based on the user's selected business domain, professional terms are extracted from each traffic engineering question type in the corresponding business domain's traffic engineering question type set. Based on the extracted professional terms, a set of type-specific professional terms is constructed for each traffic engineering question type. The user's input question language is identified, yielding multiple professional terms, which are denoted as user question terms. A set of user-specific professional terms is constructed based on these user question terms. Based on the set of type-specific professional terms and the user-specific professional term set, a score is calculated for each traffic engineering question type, and the traffic engineering question type with the highest score is selected as the user question type. The method for calculating the score for each traffic engineering problem type is as follows: Obtain the frequency of each term in the user's terminology set within the corresponding type's terminology set for each traffic engineering problem type. Add the frequencies of all terms to obtain the total frequency for each traffic engineering problem type. Normalize the total frequency to obtain the first score for each traffic engineering problem type. Map the terms in the user's terminology set and each type's terminology set to a traffic engineering terminology knowledge graph. For each term in the user's terminology set, pair it with each term in the type's terminology set to form multiple... For each term pair, the shortest path between the two terms in the traffic engineering terminology knowledge graph is found. The product of the association strengths of all edges on the shortest path is calculated, and then the product is multiplied by the domain weights of the two terms in the term pair and the path association coefficient to obtain the association score of each term pair. The average association scores of all term pairs corresponding to the same traffic engineering problem type are calculated to obtain the mean association score, and the mean association score is normalized to obtain the second score for each traffic engineering problem type. The first score and the second score are weighted and summed to obtain the score value for each traffic engineering problem type.

[0022] The above solution provides a specific method for matching the corresponding traffic engineering problem type based on the user's input question language. First, based on the experience and professional knowledge of those in the field, combined with common sense, a traffic engineering terminology knowledge graph is pre-constructed, with professional terms as nodes and logical relationships between terms as edges. Domain weights are preset for each professional term node, and association strengths are preset for each logical relationship. For example, in the business domain of "urban road traffic management," there are professional term nodes such as peak hours, intersections, signal timing, and congestion. Based on experience, the domain weight of peak hours is set to 0.8, the domain weight of intersections is set to 0.9, the domain weight of signal timing is set to 0.9, and the domain weight of congestion is set to 0.8. The association strength of intersection → signal timing is 0.9, the association strength of peak hours → congestion is 0.8, and the association strength of signal timing → congestion is 0.9. Based on the user's selected business domain, extract professional terms for each traffic engineering problem type in the corresponding business domain's traffic engineering problem type set. Based on the extracted professional terms, construct a set of type professional terms for each traffic engineering problem type. Recognize the user's input question language to obtain multiple professional terms, which are recorded as user question terms. Construct a set of user professional terms based on user question terms. The extraction of professional terms for each traffic engineering problem type in the above-mentioned traffic engineering problem type set can be obtained by word segmentation, entity recognition, or domain dictionary matching of its description text, specification documents, and case data. The recognition of user's input question language is obtained by preprocessing the question language text and then using domain dictionary matching and language models (such as BERT) to identify and map it into standard professional terms. These are all existing technologies and will not be described in detail. Then, obtain the frequency of each professional term in the user professional term set in the type professional term set corresponding to each traffic engineering problem type. Add up the frequencies of all professional terms to obtain the total frequency of each traffic engineering problem type. Rank the total frequency. The first score is obtained for each traffic engineering problem type through a single-processing method. This first score measures the degree of literal overlap between the terminology in the user's terminology set and the various traffic engineering problem types. A higher first score indicates that the terminology in the user's terminology set appears more frequently in that traffic engineering problem type, indicating a stronger relevance. Then, the terminology in the user's terminology set and the terminology sets of each type are mapped to a traffic engineering terminology knowledge graph. For each terminology in the user's terminology set, it is paired with each terminology in the type's terminology set to form multiple term pairs. Based on each term pair, the shortest path connecting the two terminology nodes in each term pair is found in the traffic engineering terminology knowledge graph. The shortest path is the path with the fewest edges connecting the two terminology nodes in the terminology pair. The product of the association strengths of all edges on the shortest path is calculated, and then this product is multiplied sequentially by the domain weights and path association coefficients of the two terminology nodes in the terminology pair to obtain the association score for each term pair. The specific association score can be expressed by the following formula:

[0023] in, This refers to the collection of user-specific terms. The collection of professional terms and types, Volume 1 A term pair consisting of professional terms This indicates the association score of the corresponding term pair. express Domain weights express Domain weights express arrive The shortest path of association, This represents the product of the association strengths of all edges on the corresponding shortest path. express and The path correlation coefficient; where the path correlation coefficient = 1 / (1 + path length); for example, if a certain term pair is intersection and congestion relief, the shortest path found in the traffic engineering terminology knowledge graph is: intersection Signal timing Congestion relief corresponds to Intersection The correlation strength of the signal timing multiplied by the signal timing The correlation strength of congestion relief; path length is the number of shortest path edges between two technical terms in the traffic engineering terminology knowledge graph (e.g., intersection). Signal timing Congestion mitigation is determined by the number of edges in the shortest path. If there is a node connecting the intersection and the congestion mitigation point, the path length is 1, and the path correlation coefficient is 1 / (1+1). Finally, the correlation scores of all term pairs are averaged to obtain the mean correlation score. This mean correlation score is then normalized to obtain a second score for each traffic engineering problem type. The second score measures the strength of the professional logical correlation between the professional terms in the user's terminology set and the professional terms of each traffic engineering problem type. A higher second score indicates a strong correlation within the professional knowledge system. Finally, the first and second scores are weighted and summed. The weights can be determined manually based on experience to obtain the score value for each traffic engineering problem type. A higher score indicates a greater correlation with the user's input question language; therefore, the traffic engineering problem type with the highest score is taken as the user's question type. This method, by integrating a two-dimensional scoring mechanism that combines literal frequency matching of user professional terms and question type terms with knowledge graph-driven professional logical correlation calculation, significantly improves the accuracy and professional adaptability of traffic engineering problem type matching, making the matching results more interpretable and domain-specific.

[0024] Based on traffic image and text data within the user question type, each traffic image and text data is encoded to obtain a modal vector. A multimodal vector database is built based on all modal vectors. Simultaneously, core entities, their attributes, and logical relationships between them are extracted from all traffic image and text data within the user question type to construct a structured traffic engineering knowledge graph. A bidirectional association index is established between the multimodal vector database and the structured traffic engineering knowledge graph. Based on all traffic image and text data within the user question type, multiple sub-question types are defined, and a corresponding structured answer template is configured for each sub-question type. The structured answer template includes a keyword framework for guiding knowledge retrieval, a logical framework for organizing answer output, and a standard index for associating authoritative evidence. The system receives user-input question language and matches it to the corresponding sub-question type using a pre-defined classification model. Core entities are extracted from the user-input question language and denoted as user-extracted entities. The system populates user-extracted entities into the structured answer template corresponding to the sub-question type, generating a retrieval query instruction containing the retrieval target, constraints, and output requirements. Based on the retrieval query instruction, a bidirectional retrieval operation is performed, using the user-extracted entities in the retrieval query instruction and the keywords extracted from the keyword framework as retrieval terms. Semantic retrieval is performed in the multimodal vector database, calculating the cosine similarity of each modality vector in the multimodal vector database. Modality vectors with a cosine similarity greater than a preset cosine similarity threshold are selected as the first type of evidence. Starting from the user-extracted entities in the retrieval query instruction, a multi-hop association traversal query is performed in the structured traffic engineering knowledge graph to obtain the core entities, attributes, and logical relationships associated with the user-extracted entities, which are then used as the second type of evidence. The first and second types of evidence are input into a preset large language model, and based on the sub-question type structured answer template, the first and second types of evidence are merged to finally generate multiple question answers. The method for calculating the confidence score of each question answer is as follows: Each question answer is converted into a question semantic vector; the user-input question language is converted into a user semantic vector; the cosine similarity between the user semantic vector and each question semantic vector is calculated; the normalized cosine similarity is used as the semantic similarity score of each question answer; multiple core entities are extracted from the user-input question language to form a first core entity set; multiple core entities are extracted from each question answer to form a second core entity set; all core entities in the structured traffic engineering knowledge graph are merged to form a third core entity set; the intersection of the second and third core entity sets is used as the key entity set; and the first core entity set... Each core entity is designated as the first core entity, and each core entity in the key entity set is designated as a key core entity. Any first core entity and any key core entity are combined into an entity pair. The shortest path query algorithm is executed in the structured traffic engineering knowledge graph to obtain the shortest relation hop count for each entity pair. The average shortest relation hop count for all entity pairs corresponding to the answer to the same question is calculated, and the average shortest relation hop count is normalized to obtain the coverage score for each question answer. The number of references to the normative standard index in each question answer is obtained, and the number of references is normalized to obtain the evidence citation score for each question answer. The semantic similarity score, coverage score, and evidence citation score are weighted and summed to obtain the confidence score for each question answer. The method for selecting and recommending question answers to users based on confidence scores is as follows: Based on the confidence scores of each question answer, the question answers are sorted in descending order of confidence scores to obtain a question answer recommendation table; the question answer ranked first in the question answer recommendation table is recommended to the user; if the recommended question answer does not meet the user's needs, the question answer is iteratively recommended according to the order of the question answer recommendation table until it meets the user's needs.

[0025] The above scheme provides a specific method for generating question answers. First, based on the text and image data within the user question type, each text and image data is encoded to obtain a modality vector. A multimodal vector database is built based on all modality vectors. Specifically, all relevant text and image data within the matched user question type can be extracted, including industry standard texts, engineering cases, design drawings, standard drawing sets, video materials, etc. Then, these different forms of data are converted into a dense vector using a multimodal encoder (such as CLIP, ViT+BERT), denoted as a modality vector. A multimodal vector database is built based on these modality vectors for subsequent semantic similarity retrieval. Simultaneously, the text and image data within the user question type... This paper extracts core entities (such as "intersection", "signal timing", and "Webster model"), their attributes (such as the location and number of lanes of an "intersection", the duration of "signal timing", and the use case of "Webster model"), and the logical relationships between core entities (such as "intersection → adopt → Webster model", "signal timing → impact → congestion relief"). A structured traffic engineering knowledge graph is constructed with core entities as nodes, their attributes as node features, and the logical relationships between core entities as edges. A bidirectional association index is established between the multimodal vector database and the structured traffic engineering knowledge graph. The bidirectional association index is used to establish the Chinese multimodal vector database. The mapping relationship between files / fragments and entities / relationships in the structured traffic engineering knowledge graph enables bidirectional communication, allowing vector retrieval results to jump to knowledge graph entities and knowledge graph entities to quickly locate relevant multimodal data. Based on data within user question types, multiple sets of sub-question types are set, and a corresponding structured answer template is configured for each sub-question type. The structured answer template includes a keyword framework for guiding knowledge retrieval, an answer logic framework for organizing answer output, and a standard index for associating authoritative evidence. Specifically, user question types can be further subdivided into more specific sub-question types (e.g., under "Urban Road Traffic Management," it can be divided into "Intersection Signal Timing Optimization," "Illegal Parking Management," and "Non-standard Parking Management"). (e.g., "motor vehicle lane planning"), and configure a structured answer template for each sub-question type; the keyword framework in the structured answer template defines the core professional terminology set under the sub-question type (e.g., "peak hours", "intersection", "signal timing"), which is used to guide the search to prioritize these terms; the answer logic framework specifies the organizational structure of the answer (e.g., "first explain the cause → then give the method → ​​finally cite the standard"), ensuring that the output answer conforms to professional expression habits; and the standard index associates the industry standards and national standards corresponding to the sub-question type, ensuring the authority and compliance of the answer; it receives the question language input by the user and matches it to the corresponding sub-question type through a preset classification model;The core entities are extracted from the user-input question language and denoted as user-extracted entities. These user-extracted entities are then populated into the structured solution template corresponding to the sub-question type, generating a retrieval query instruction containing the retrieval target, constraints, and output requirements. Specifically, a pre-trained traffic engineering classification model (such as BERT or TextCNN) is used to accurately match the user-input question language to the corresponding sub-question type. Then, NER technology is used to extract core entities from the user-input question language (such as "peak hours," "urban arterial road intersections," "signal timing," and "congestion relief"), denoted as user-extracted entities, which are then populated into the structured solution template corresponding to the sub-question type. The system generates a search query instruction containing the search objective, constraints, and output requirements. For example, "Search for methods to optimize signal timing at intersections on urban main roads during peak hours, requiring the use of the Webste model and outputting the implementation steps for time-segmented adjustments." This instruction includes the search objective, constraints, and output format requirements. Based on the search query instruction, a bidirectional search operation is performed. Using the user-extracted entities in the search query instruction and the keywords extracted from the template keyword framework as search terms, semantic search is performed in the multimodal vector database. The cosine similarity of each modality vector in the multimodal vector database is calculated. Modality vectors with a cosine similarity greater than a preset cosine similarity threshold are taken as the first type of evidence. The preset cosine similarity threshold can be adjusted based on... Based on experience and the expertise of those in the field, this approach allows for the acquisition of multimodal vectors highly semantically related to the user's question, serving as the first type of evidence. Starting with the user-extracted entity from the search query, a multi-hop associative traversal query is performed within the structured traffic engineering knowledge graph (e.g., starting from "intersection," traversing associative queries such as "signal timing → Webster model → time-segment adjustment → congestion mitigation") to obtain the core entities, attributes, and logical relationships associated with the user-extracted entity. These related entities, attributes, and logical relationships serve as the second type of evidence. Finally, the first and second types of evidence are input into a large language model adapted to the traffic engineering field (e.g., LLaMA, Qwen model), based on the sub-questions. The system uses structured answer templates configured within each question type, integrating primary and secondary evidence to organize answers and ultimately generate multiple question-answer pairs. Finally, it calculates the confidence score of each answer and sorts them in descending order based on that score, creating a question-answer recommendation table. The top-ranked answer in the table is then recommended to the user. If the recommended answer doesn't meet the user's needs, the system iteratively recommends answers according to the table until the user's needs are met. This solution, through sub-question type classification and structured answer template configuration, provides precise and professional guidance for knowledge retrieval and answer generation, avoiding the waste of resources from indiscriminate retrieval.By employing a bidirectional retrieval operation combining dense vector semantic retrieval and multi-hop traversal of the knowledge graph, dual evidence is obtained from both the factual and professional logical perspectives. This ensures both the semantic relevance of the evidence to the user's question and strengthens the professional logical support of the answer.

[0026] The method for calculating the confidence score of each question's answer is as follows: Each question's answer is converted into a question semantic vector; the user's input question language is converted into a user semantic vector; the cosine similarity between the user semantic vector and each question's semantic vector is calculated; and the normalized cosine similarity is used as the semantic similarity score for each question's answer. The language vector conversion method uses a language model from the traffic engineering domain (such as Sentence-BERT or BERT-Whitening). The natural language text of the user's question and each question's answer is encoded into a fixed-dimensional semantic vector, and the cosine similarity between the user semantic vector and each question's semantic vector is calculated. A higher cosine similarity indicates a greater semantic fit between the answer and the user's question. The higher the value, the better. Extract multiple core entities from the user-input question language to form a first core entity set. Extract multiple core entities from each question answer to form a second core entity set for each question answer. Merge all core entities in the structured traffic engineering knowledge graph to form a third core entity set. The intersection of the second and third core entity sets is taken as the key entity set. The key entity set can filter out invalid entities in each question answer that do not belong to the domain knowledge graph, ensuring that subsequent calculations are based only on domain-recognized professional entities, thus improving the professionalism of the scoring. Each core entity in the first core entity set is designated as the first core entity, and each core entity in the key entity set is designated as the key core entity. Any first... Core entities are paired with any key core entities to form entity pairs. A shortest path query algorithm is then executed within the knowledge graph to obtain the shortest relation hop count for each entity pair. Specifically, Breadth-First Search (BFS) or Dijkstra's algorithm can be used to find the shortest path between each entity pair. The relation hop count refers to the number of relation edges required to travel from one core entity to another. For example, the hop count for "intersection → signal timing → Webster model" is 2. The average shortest relation hop count for all entity pairs is calculated. A lower average shortest relation hop count indicates a stronger connection between the answer entity and the user's question entity within the professional knowledge system. Therefore, the average shortest relation hop count is normalized to obtain a coverage score. A higher coverage score indicates a stronger connection between the answer entity and the user's question entity within the professional knowledge system. The fewer the average shortest relation hops, the better. Next, the number of references to the standard index in each question's answer is obtained, and the number of references is normalized to obtain an evidence citation score. A higher evidence citation score indicates more references and greater authority. Finally, the semantic similarity score, coverage score, and evidence citation score are weighted and summed, with each weight determined based on the experience and expertise of those in the field, thus yielding the confidence score for each question's answer. By integrating the weighted scores of semantic similarity, professional knowledge coverage, and authoritative evidence citation, not only is the issue of irrelevant answers caused by single semantic matching effectively avoided, but the logical relevance and authority of the answer within the professional knowledge system can also be accurately measured, making the confidence score more scientific and interpretable.

[0027] The update instructions include business domain module updates and large model updates. Business domain module updates refer to updating the question-answering business domain modules within the large language model, while large model updates refer to updating the entire large language model. The update judgment coefficients for each business domain are obtained and compared with their respective update judgment coefficient thresholds. If an update judgment coefficient is less than the threshold, the corresponding business domain module is updated. The total number of updated question-answering business domain modules in the large language model is obtained and recorded as the update count. This update count is compared with a set update count threshold. If the update count is greater than the threshold, the large model is updated. The method for generating the update judgment coefficient is as follows: First, a feedback data table is pre-constructed for each business domain. This table contains multiple feedback item types, each with a corresponding quantitative score range. The quantitative scores for each feedback item type after each user feedback are obtained for each business domain. The quantitative scores for all feedback item types within the same business domain are weighted and summed to obtain the quantitative comprehensive score for each business domain. Then, the quantitative comprehensive score for each business domain is obtained in time series. Based on the quantitative comprehensive score in time series, the standard deviation of the quantitative comprehensive score is calculated, and the standard deviation is normalized to obtain the fluctuation coefficient for each business domain. Based on the quantitative comprehensive score in time series, a quantitative comprehensive score function is constructed. The integral value of the quantitative comprehensive score function in time series is calculated, and the integral value is normalized to obtain the cumulative coefficient for each business domain. Finally, the fluctuation coefficient is incremented by one to obtain the baseline fluctuation coefficient. The cumulative coefficient is divided by the baseline fluctuation coefficient to obtain the update judgment coefficient for each business domain.

[0028] The above scheme provides specific update judgment methods. Firstly, update instructions include updates to business domain modules and updates to the overall model. Updates to business domain modules refer to updating the question-answering business domain modules within the overall language model. For example, if an anomaly occurs in the urban road traffic management business domain, then that business domain is updated separately, including updating its multimodal vector database, structured traffic engineering knowledge graph, question type set, and answer templates. Updates to the overall language model refer to updating the entire model, such as updating the interaction interface and collaborative logic, synchronizing the interaction interfaces between the overall model and each question-answering business domain module, ensuring the synergy of evidence retrieval, knowledge graph invocation, and template filling, and performing performance calibration and evaluation, recalibrating the answer confidence assessment mechanism, etc. Specifically, the update judgment coefficients for each business domain are obtained. The higher the update judgment coefficient, the better the module performance and the higher the feedback score. Based on the experience and professional knowledge of personnel in this field, corresponding update judgment coefficient thresholds are set for each business domain. The update judgment coefficients are compared with the update judgment coefficient thresholds. If the update judgment coefficient is less than the update judgment coefficient threshold, it indicates that the module has an anomaly, and the corresponding business domain module is updated. At the same time, the number of business domain module updates in the large model is obtained and recorded as the update count. Similarly, an update count threshold is set based on experience and professional knowledge. The update count is compared with the set update count threshold. When the update count is greater than the update count threshold, it indicates that the large module has more module anomalies and the overall score is not high, so the large model is updated.

[0029] The specific method for generating update judgment coefficients is as follows: Feedback data tables for each business domain are pre-constructed. These tables contain multiple feedback item types, each with a corresponding quantitative score range. Feedback item types can be manually defined, including aspects such as answer accuracy, usability, interactive experience, and satisfaction. Within each business domain, users score each feedback item type against the defined quantitative score range, resulting in a quantitative score for each item type. The quantitative scores of all feedback item types within the same business domain are weighted and summed to obtain a comprehensive quantitative score for that business domain. The weight of each feedback item type can be manually set; for example, prioritizing answer accuracy and interactive experience increases their weight ratio. The comprehensive quantitative score indicates the user's satisfaction level in this Q&A session; a higher value indicates higher satisfaction, suggesting excellent module performance in that business domain. The comprehensive quantitative score for each business domain is then obtained over time. Based on this score, the standard deviation is calculated, and the standard deviation is normalized to obtain the volatility coefficient for each business domain. The volatility coefficient indicates the performance of the corresponding business domain module within that time series. The stability of the quantitative comprehensive score listed is assessed; a smaller value indicates better performance of the business domain module. Based on the quantitative comprehensive score under time series, a quantitative comprehensive score function is constructed, and the integral value of the quantitative comprehensive score function under time series is calculated. The integral value is normalized to obtain the cumulative coefficient for each business domain. The cumulative coefficient indicates the overall performance of the business domain module under the corresponding time series; a larger value indicates a more stable and high-quality overall user feedback performance of the business domain module within the period, and a lower necessity for module updates. Therefore, the fluctuation coefficient is incremented by one to obtain the baseline fluctuation coefficient, and the cumulative coefficient is set to the baseline fluctuation coefficient to obtain the update judgment coefficient for each business domain. In this way, the update judgment coefficient generated by the quantitative score, fluctuation coefficient, and cumulative coefficient of the feedback item type is used to determine whether an update is needed, making the updates of business domain modules more targeted and effectively avoiding the waste of resources caused by blind updates. At the same time, triggering large-scale model updates by using a threshold for the number of module updates ensures the efficiency of local business optimization and promotes the iteration of the overall system capabilities when multiple domain requirements are upgraded, thereby continuously improving the accuracy, professionalism, and user experience of the answers, and ensuring that the system always adapts to the knowledge updates and changes in user expectations in the field of transportation engineering.

[0030] In one embodiment, a large-scale rapid question-answering system based on a traffic engineering graphic database is disclosed, such as... Figure 2 As shown, the system includes: The problem type set construction module is used to integrate structured and unstructured data of traffic engineering in various business areas to form a database. The database is clustered to obtain multiple traffic engineering problem types. Based on the traffic engineering problem types, a traffic engineering problem type set for each business area is constructed. The question type matching module is used to obtain the question language input by the user, preprocess the question language and input it into the traffic engineering question type set, match the corresponding traffic engineering question type, and record it as the user question type. The question-answer generation module analyzes the data within the user's question type and the language of the question input by the user to obtain the first type of evidence and the second type of evidence. Based on the first type of evidence and the second type of evidence, it generates multiple question answers, calculates the confidence score of each question answer, and selects question answers to recommend to the user based on the confidence scores. The update judgment module is used to obtain user feedback data based on the push results, analyze the user feedback data, generate update judgment coefficients, and execute update instructions based on the update judgment coefficients.

[0031] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0032] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0033] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A fast question-answering method for large models based on traffic engineering graphic databases, characterized in that, The methods include: The structured and unstructured data of traffic engineering in various business areas are integrated to form a database. The database is clustered to obtain multiple traffic engineering problem types. Based on the traffic engineering problem types, a set of traffic engineering problem types for each business area is constructed. Obtain the question language input by the user, preprocess the question language and input it into the traffic engineering question type set, match the corresponding traffic engineering question type, and record it as the user question type; The system analyzes the data within the user's question type and the language of the question input by the user to obtain the first type of evidence and the second type of evidence. Based on the first type of evidence and the second type of evidence, it generates multiple question answers, calculates the confidence score of each question answer, and selects question answers to recommend to the user based on the confidence scores. Based on the push notification results, user feedback data is obtained, the user feedback data is analyzed, update judgment coefficients are generated, and update instructions are executed based on the update judgment coefficients.

2. The fast question-answering method for large models based on traffic engineering graphic databases according to claim 1, characterized in that, The method for constructing the traffic engineering problem type set under each business domain is as follows: Traffic image and text data from various business areas are retrieved from the database. Each piece of traffic image and text data is taken as a sample, and a traffic engineering image and text sample set for the corresponding business area is constructed based on all samples. After preprocessing the traffic engineering image and text sample sets for each business area, feature extraction is performed to obtain multiple problem feature parameters and image and text matching feature parameters. Set the problem-image-text matching coupling neighborhood radius and the minimum core problem sample number threshold, and set the thresholds for each problem feature parameter and each image-text matching feature parameter within the problem-image-text matching coupling neighborhood radius; traverse all samples in the traffic engineering image-text sample set, and determine the problem core points, problem boundary points, and problem noise points in the traffic engineering image-text sample set according to the set problem-image-text matching coupling neighborhood radius and the minimum core problem sample number threshold; After removing noisy points, the core points and boundary points of the problem are merged to form a sample set to be clustered. The density clustering algorithm is then used to perform clustering operations on the sample set to be clustered, resulting in multiple independent clusters. Each independent cluster represents a type of traffic engineering problem. Based on the problem feature parameters and image-text matching feature parameters contained in each independent cluster, each independent cluster is named to obtain the corresponding traffic engineering problem type. Based on all traffic engineering problem types, a traffic engineering problem type set under each business domain is constructed.

3. The rapid question-answering method for large models based on traffic engineering graphic databases according to claim 1, characterized in that, The method for matching user question types is as follows: A pre-constructed knowledge graph of traffic engineering terminology is used, where nodes represent professional terms and edges represent logical relationships between professional terms. Each professional term node in the knowledge graph is assigned a pre-defined domain weight, and the logical relationships between professional terms are pre-defined with a pre-defined association strength. Based on the user-selected business domain, professional terms are extracted from each traffic engineering problem type within the corresponding business domain's traffic engineering problem type set. Based on the extracted professional terms, a set of type professional terms for each traffic engineering problem type is constructed. The language of the questions input by users is identified, and multiple professional terms are obtained and recorded as user question terms. A set of user professional terms is constructed based on the user question terms. Based on the set of professional terms for different types and the set of professional terms for users, the score for each type of traffic engineering problem is calculated, and the traffic engineering problem type with the highest score is taken as the user problem type.

4. The rapid question-answering method for large models based on traffic engineering graphic databases according to claim 3, characterized in that, The method for calculating the score for each type of traffic engineering problem is as follows: Obtain the frequency of each professional term in the user's professional term set in the corresponding type professional term set for each traffic engineering problem type. Add up the frequencies of all professional terms to get the total frequency for each traffic engineering problem type. Normalize the total frequency to get the first score for each traffic engineering problem type. The user's professional terminology collection and the professional terms in various types of professional terminology collections are mapped to the traffic engineering terminology knowledge graph. For each professional term in the user's professional terminology collection, it is paired with each professional term in the type of professional terminology collection to form multiple sets of term pairs. Based on each set of term pairs, the shortest association path between the two professional terms in each set of term pairs is found in the traffic engineering terminology knowledge graph. Calculate the product of the association strengths of all edges on the shortest association path, and then multiply the product sequentially by the domain weights of the two terms in the term pair and the path association coefficient to obtain the association score for each term pair. Calculate the mean of the association scores for all term pairs corresponding to the same traffic engineering problem type, and then normalize the mean association score to obtain the second score for each traffic engineering problem type. Finally, weighted sum the first score and the second score to obtain the final score for each traffic engineering problem type.

5. The rapid question-answering method for large models based on traffic engineering graphic databases according to claim 1, characterized in that, The method for generating answers to multiple questions is as follows: Based on traffic image and text data within user question types, each traffic image and text data is encoded to obtain a modal vector. A multimodal vector database is built based on all modal vectors. Simultaneously, core entities, attributes of core entities, and logical relationships between core entities are extracted from all traffic image and text data within user question types to construct a structured traffic engineering knowledge graph. A bidirectional association index is established between the multimodal vector database and the structured traffic engineering knowledge graph. Based on all traffic image and text data within user question types, multiple sets of sub-question types are set, and a corresponding structured answer template is configured for each sub-question type. The structured answer template includes a keyword framework for guiding knowledge retrieval, an answer logic framework for organizing answer output, and a standard index for associating authoritative evidence. The system receives user-input question language and matches it to the corresponding sub-question type through a preset classification model. Core entities are extracted from the user-input question language and denoted as user-extracted entities. The user-extracted entities are then filled into the structured answer template corresponding to the sub-question type to generate a retrieval query instruction containing the retrieval target, constraints, and output requirements. Based on the retrieval query command, a bidirectional retrieval operation is performed. Using the user-extracted entity from the retrieval query command and the keywords extracted from the keyword framework as retrieval terms, semantic retrieval is performed in the multimodal vector database. The cosine similarity of each modality vector in the multimodal vector database is calculated, and modality vectors with a cosine similarity greater than a preset cosine similarity threshold are selected as the first type of evidence. Starting from the user-extracted entity in the retrieval query command, a multi-hop association traversal query is performed in the structured traffic engineering knowledge graph to obtain the core entities, attributes, and logical relationships associated with the user-extracted entity, which serve as the second type of evidence. The first and second types of evidence are then input into a preset large language model. Based on the structured answer template for sub-question types, the first and second types of evidence are merged to finally generate multiple question answers.

6. The fast question-answering method for large models based on traffic engineering graphic databases according to claim 5, characterized in that, The method for calculating the confidence score of the answers to each question is as follows: Each question answer is converted into a question semantic vector, and the user-input question language is converted into a user semantic vector. The cosine similarity between the user semantic vector and each question semantic vector is calculated, and the normalized cosine similarity is used as the semantic similarity score of each question answer. Multiple core entities are extracted from the user-input question language to form a first core entity set, and multiple core entities are extracted from each question answer to form a second core entity set for each question answer. All core entities in the structured traffic engineering knowledge graph are merged to form a third core entity set. The intersection of the second core entity set and the third core entity set is taken as the key entity set; each core entity in the first core entity set is denoted as the first core entity, and each core entity in the key entity set is denoted as the key core entity. Any first core entity and any key core entity are combined into an entity pair. The shortest path query algorithm is executed in the structured traffic engineering knowledge graph to obtain the shortest relation hop count for each entity pair. Calculate the average shortest relation hop count for all entity pairs corresponding to the answer to the same question, and normalize the average shortest relation hop count to obtain the coverage score for each question answer; obtain the number of references to the canonical standard index in each question answer, and normalize the number of references to obtain the evidence citation score for each question answer; weight and sum the semantic similarity score, coverage score, and evidence citation score to obtain the confidence score for each question answer.

7. The fast question-answering method for large models based on traffic engineering graphic databases according to claim 6, characterized in that, The method for recommending question answers to users based on confidence scores is as follows: Based on the confidence scores of each question's answer, the questions and answers are sorted in descending order of confidence scores to obtain a question and answer recommendation table; Recommend the highest-ranking answers to questions in the question-answer recommendation list to users; If the recommended answers do not meet the user's needs, the system will iteratively recommend answers to each question according to the order of the question-answer recommendation list until the answers meet the user's needs.

8. The fast question-answering method for large models based on traffic engineering graphic databases according to claim 1, characterized in that, The update instructions include business domain module updates and large model updates; among them, business domain module updates refer to updating the question-answering business domain module in the large language model, and large model updates refer to updating the entire large language model; Obtain the update judgment coefficient for each business domain, and compare it with the update judgment coefficient threshold corresponding to each business domain. If the update judgment coefficient is less than the update judgment coefficient threshold, then update the business domain module for the corresponding business domain. Obtain the total number of updates in the question-answering business domain module of the large language model, denoted as the update count. Compare the update count with the set update count threshold. If the update count is greater than the update count threshold, then perform a large model update.

9. The fast question-answering method for large models based on traffic engineering graphic databases according to claim 8, characterized in that, The method for generating the updated judgment coefficients is as follows: Pre-build feedback data tables for each business area. Each feedback data table contains multiple feedback item types, and each feedback item type has a corresponding quantitative score range. Obtain the quantitative score of each feedback item type after each user feedback in each business area, and weight and sum the quantitative scores of all feedback item types in the same business area to obtain the quantitative comprehensive score of each business area. Obtain the quantitative comprehensive score for each business area under the time series; calculate the standard deviation of the quantitative comprehensive score based on the time series; normalize the standard deviation to obtain the volatility coefficient for each business area; construct the quantitative comprehensive score function based on the quantitative comprehensive score under the time series; calculate the integral value of the quantitative comprehensive score function under the time series; normalize the integral value to obtain the cumulative coefficient for each business area. Add one to the volatility coefficient to obtain the benchmark volatility coefficient, and divide the cumulative coefficient by the benchmark volatility coefficient to obtain the update judgment coefficient for each business area.

10. A large-scale rapid question-answering system based on a traffic engineering graphic database, implementing the large-scale rapid question-answering method based on a traffic engineering graphic database as described in any one of claims 1-9, characterized in that, The system includes: The problem type set construction module is used to integrate structured and unstructured data of traffic engineering in various business areas to form a database. The database is clustered to obtain multiple traffic engineering problem types. Based on the traffic engineering problem types, a traffic engineering problem type set for each business area is constructed. The question type matching module is used to obtain the question language input by the user, preprocess the question language and input it into the traffic engineering question type set, match the corresponding traffic engineering question type, and record it as the user question type. The question-answer generation module analyzes the data within the user's question type and the language of the question input by the user to obtain the first type of evidence and the second type of evidence. Based on the first type of evidence and the second type of evidence, it generates multiple question answers, calculates the confidence score of each question answer, and selects question answers to recommend to the user based on the confidence scores. The update judgment module is used to obtain user feedback data based on the push results, analyze the user feedback data, generate update judgment coefficients, and execute update instructions based on the update judgment coefficients.

Images