Intelligent question answering system method for air traffic control communication business knowledge

Abstract

The invention belongs to the technical field of air traffic control and conduction, and particularly relates to an intelligent question answering system method for air traffic control and conduction business knowledge. According to the method and the system, the refined and context-aware information retrieval capability of the air traffic control communication service knowledge and the expert-level air traffic control communication service knowledge question and answer capability are provided, the knowledge management and application efficiency is effectively improved, and powerful support is provided for digital transformation of related services. Comprising the following steps: S1, a document content extraction and storage process; s2, a content retrieval and generation process; s3, content slice optimization is carried out through extraction and similarity calculation of adjacent paragraphs or contents, complete contents are in the same slice, and the content retrieval efficiency is improved; according to the method, the label library and the business rule library are continuously iteratively optimized, so that the provided service is always in the optimal state.

Description

Technical Field

[0001] The present invention belongs to the technical field of air traffic control communication and navigation, and particularly relates to an intelligent question-answering system method for air traffic control communication and navigation business knowledge. Background Art

[0002] At present, there are a wide variety of knowledge materials for civil aviation air traffic control communication and navigation, including standard operation specifications, fault case materials, emergency disposal procedures, job operation guides, safety learning materials, equipment technical materials, etc. Most of the knowledge materials are in paper form, and the degree of digitization is low. In the daily operation and maintenance of air traffic control communication and navigation, it depends on manual searching of paper materials to query equipment-related operation and maintenance operations, fault disposal materials, etc., and it is impossible to conduct dynamic correlation analysis between knowledge materials and equipment operation and maintenance. There is a lack of digital communication and navigation knowledge management tools, which affects the operation and maintenance efficiency of air traffic control communication and navigation.

[0003] How to quickly retrieve and accurately locate the similar fault disposal process when a device fails, so as to provide data support for emergency troubleshooting decision-making; how to effectively manage the continuously accumulated and updated knowledge materials, improve the management efficiency of knowledge materials, and provide refined, content-level, and full-life-cycle management; how to dynamically associate knowledge materials with equipment operation and maintenance to deeply explore the application value of knowledge materials are all the difficulties and pain points of current air traffic control communication and navigation knowledge management.

[0004] Combined with the actual requirements of communication and navigation operation and maintenance, based on technologies such as large language models and RAG, this project conducts targeted training to learn the knowledge data of communication and navigation experts, constructs an air traffic control communication and navigation knowledge algorithm model, provides full-life-cycle and refined management of knowledge materials, realizes content-level extraction of various formats of materials, provides accurate full-text retrieval and semantic retrieval, knowledge positioning, and intelligent question-answering dialogue, improves the utilization rate and digitization degree of knowledge materials, and provides data support and decision-making assistance for air traffic control operation efficiency.

[0005] When the existing RAG architecture processes document content, although it has a certain degree of generality, when facing professional documents in specific fields, such as air traffic control communication and navigation business documents, the performance of key links such as content extraction, classification, slicing, and knowledge retrieval is often unsatisfactory. These documents usually have unique layout formats and complex knowledge structures, and the traditional RAG architecture is difficult to process such documents efficiently and accurately, resulting in low knowledge organization and management efficiency, and insufficient accuracy and relevance in knowledge query and retrieval and business knowledge question-answering. Specifically, in the content extraction link, the model often has difficulty accurately identifying the specific professional document structure and domain professional information, resulting in information loss or errors; in the document classification stage, the complex business associations make the existing classification standards difficult to meet the requirements. This limitation in processing ability directly affects the subsequent knowledge organization and application, and ultimately leads to difficulties in meeting the high requirements of professional users for accuracy and relevance in knowledge retrieval and business question-answering, and also unable to achieve efficient knowledge management and utilization.

[0006] The existence of these problems highlights the limitations of the existing RAG architecture in processing professional field documents. Therefore, there is an urgent need for a new framework that can provide customized processing solutions for specific business needs to better adapt to the complex structure and special requirements of professional documents, thereby improving the efficiency of knowledge management and application. Summary of the Invention

[0007] In view of the deficiencies in the existing technology, the present invention provides a method for an intelligent question-answering system for air traffic control communication and navigation service knowledge, which provides refined and context-aware information retrieval capabilities for air traffic control communication and navigation service knowledge, as well as expert-level question-answering capabilities for air traffic control communication and navigation service knowledge, effectively improving the efficiency of knowledge management and application, and providing strong support for the digital transformation of related businesses.

[0008] To solve the above technical problems, the technical solution adopted by the present invention is: a method for an intelligent question-answering system for air traffic control communication and navigation service knowledge, including the following steps:

[0009] S1. Document content extraction and warehousing process;

[0010] S11. Content extraction and in-depth analysis:

[0011] S111. Text structuring process: After obtaining the document materials, the system will first perform structured analysis and layout recognition on the document content, converting the originally chaotic information into text data with a clear hierarchical structure;

[0012] S112. Key information extraction: Through natural language processing technology and document parsing algorithms, the text content is deeply analyzed to accurately extract the entity and relationship knowledge elements therein, converting unstructured data into a structured data form;

[0013] S12. Classification and slicing process:

[0014] S121. Text slicing: The system will perform fine-grained division of the document content according to logical units, paragraphs, sentences or semantic units to facilitate subsequent retrieval and invocation. This slicing processing method improves the content management efficiency and makes information retrieval or knowledge question-answering more accurate and efficient;

[0015] S122. Content classification: Based on the air traffic control communication and navigation service knowledge framework and the document-associated equipment information, the sliced content is intelligently classified, and different types of information are marked with business labels through association analysis algorithms to achieve accurate content management and retrieval;

[0016] S13. Document vectorization and persistence:

[0017] S131, the document content processed by S11 and S12 will be converted into a vectorized expression and stored in a vector database for subsequent retrieval and call;

[0018] S2, content retrieval and generation process;

[0019] S21. Problem analysis rewrite:

[0020] S211. When a user asks a question or query, the system analyzes and understands the question, extracts the key words, semantics, and business terms, and rewrites the original question using a large language model to better meet search requirements.

[0021] S22, Multi-way retrieval recall:

[0022] S221, Vector Similarity Matching: Accurate matching based on semantic vectors generated by deep learning;

[0023] S222, Keyword matching: ensure comprehensive retrieval of basic information;

[0024] S223,Multiple retrieval strategies complement each other to maximize the recall of content fragments and knowledge items related to the question;

[0025] S23. Reordering of search results:

[0026] S231. The initial search results usually contain a large amount of information, including both highly relevant content and less relevant or even irrelevant data. The system optimizes the results through the following methods: BM25 keyword score calculation, vector similarity calculation, and label similarity calculation;

[0027] S232: Through the optimization process in step S231, the system can improve the accuracy and relevance of the search results, laying the foundation for subsequent generation of higher-quality answers;

[0028] S24. Large language model generates answers:

[0029] S241. Based on the optimized search results, the system calls the large language model for in-depth processing.

[0030] S242, context integration: integrating multi-dimensional information of user questions and search results;

[0031] S243, Logical Reasoning: Intelligent reasoning based on knowledge base content and business rules;

[0032] S244. Structured output: Generate the final answer with a clear and rigorous logical framework.

[0033] Furthermore, it also includes step S3: optimizing content slicing by extracting and calculating the similarity of adjacent paragraphs or content, so that the complete content is within the same slice, improving the content retrieval efficiency, including the following steps:

[0034] S31. Content preprocessing: First, standardize the air traffic control communication and navigation business materials, including adjusting the document resolution, removing noise and interference elements, and conducting a preliminary analysis of the layout to identify key area titles, text, and charts;

[0035] S32. Feature extraction: Obtain the visual features and context semantic features of the text content through a multi-dimensional feature extraction method. At the same time, perform unique feature encoding on the chart area for subsequent classification;

[0036] S33. Similarity calculation: Based on the extracted multi-dimensional features, use the cosine similarity algorithm to calculate the similarity values between adjacent paragraphs or content blocks, as shown in Formula 1, and construct a content association matrix;

[0037]

[0038] where, v i and v j are the feature vectors of two content blocks i and j. Then, filter out highly relevant content segments through a dynamic threshold (T), as shown in Formula 2:

[0039] T = μ + kσ Formula 2;

[0040] where, μ is the average similarity of the current document, σ is the standard deviation, and k is an empirical coefficient;

[0041] S34. Content slicing optimization: Based on the content association matrix, perform intelligent slicing on the document content, group content with high similarity and logical coherence into the same block. At the same time, for chart content, achieve accurate recognition and separate division through shape feature and surrounding text matching. The shape matching of the chart area uses the improved Hausdorff distance, and the similarity between two shapes A and B is defined as, as shown in Formula 3:

[0042]

[0043] Combined with the surrounding text semantic similarity S text , comprehensively determine whether it belongs to the same slice, as shown in Formula 4:

[0044] ChartMatchScore = α·d H (A, B) + β·S text (α + β = 1) Formula 4;

[0045] When ChartMatchScore > θ, it is grouped into the same slice;

[0046] S35. Label tree construction: By analyzing the sliced content blocks, a hierarchical classification algorithm is used to construct a label tree step by step from the base to the subdivision. Each node represents a classification label, and its weight is determined by the importance and frequency of the content block. The node weight ω is jointly calculated by the TF-IDF frequency and the semantic significance score. See Equation 5:

[0047] ω = λ·TF-IDF(t) + (1 - λ)·Softmax(f DL (t)) Equation 5;

[0048] where f DL is the semantic feature vector output by the deep learning model, and λ ∈ [0, 1] is the balance coefficient. In this process, a deep learning model is used to perform topic model analysis on the text to automatically extract high-level semantic features;

[0049] S36. Model training and optimization: The labeled sample data is input into the neural network model for training. A multi-task learning framework is used to optimize both the content slicing and label classification tasks simultaneously. During the training process, the model parameters are continuously adjusted through cross-validation and performance evaluation to ensure robustness on different types of documents. The total loss function is the weighted sum of the losses of the two tasks. See Equation 6:

[0050] L total = γL slice + (1 - λ)L classify Equation 6;

[0051] where L slice is the contrastive loss of the slicing task, and L classify is the cross-entropy loss of the classification task. γ is dynamically adjusted through gradient normalization;

[0052] S37. Precise identification and application: The final label tree structure is used in the subsequent content retrieval system to achieve fast positioning and efficient invocation of relevant knowledge. Through the automated processing of air traffic control business materials, the accuracy of data extraction and the precision of classification are significantly improved, providing strong support for intelligent management and decision-making.

[0053] Furthermore, it also includes step S4. In the case where there are different names for the same content entity, to improve the accuracy of keyword search, a synonym dictionary is used to replace the keywords of the content entity in the business materials, covering the core concepts and common expressions of air traffic control communication and navigation services to ensure the consistency of business names. It includes the following steps:

[0054] S41. Synonym dictionary construction:

[0055] S411. Term extraction: Extract different names of the content entity from the air traffic control communication and navigation business materials;

[0056] S412. Establish mapping: Classify different expressions of the same subject into synonym groups, define the main word, i.e., the standardized name, and the subordinate words, i.e., the aliases.

[0057] S413. Manual verification: Have domain experts review the synonym mapping relationship to ensure semantic consistency.

[0058] S42. Document preprocessing and keyword matching:

[0059] S421. Word segmentation and entity recognition: Use NLP tools to segment the document and recognize business-related entities.

[0060] S422. Synonym matching: Traverse the keywords in the document and query the synonym dictionary.

[0061] S43. Search index construction:

[0062] S431. Inverted index generation: Build an inverted index for each main word, record the document positions containing the word, and attach a synonym extension field to the index. The definition of the inverted index is shown in Formula 7:

[0063] Index(m) = {(d, pos d ) | d ∈ D, m ∈ MainTerms(d) ∪ Synonyms(m)} Formula 7;

[0064] Among them, MainTerms(d) is the set of main words in document d, and Synonyms(m) is the set of subordinate words of main word m, which are mapped through the synonym dictionary.

[0065] S432. Weight assignment: Set different weights for the main words and subordinate words, with higher weights for the main words to improve the search ranking.

[0066] S44. Query processing:

[0067] S441. Query expansion: According to the synonym dictionary, replace the search term with the main word and associate all synonyms.

[0068] S442. Index retrieval: Use the expanded keyword set, main word + synonyms to retrieve the inverted index and return all matching documents.

[0069] Furthermore, it also includes step S5. Carry out customized label tree recognition and filling work according to the applicable site and business entity described in the business materials, and train a content classification and label extraction model using the existing materials and manual labels, including the following steps:

[0070] S51. Label system construction and data preparation:

[0071] S511, label tree structuring: sort out the equipment hierarchical relationship based on historical air traffic control navigation data and build a tree structure;

[0072] S512, document-label alignment: associate manually annotated document segments with corresponding label paths to form a training set;

[0073] S513, domain dictionary enhancement: integrated business synonym dictionary to solve the device alias problem;

[0074] S52. Hierarchical label prediction:

[0075] S521, hierarchical constraint: The prediction result of the upper-level label is used as the candidate range of the lower-level label. The hierarchical label prediction is subject to the conditional probability constraint, see formula 8:

[0076]

[0077] Among them, l i is the i-th layer label, Children(l i ) is l i-1 The sub-tag collection of

[0078] S522, Multi-label processing: support associating multiple label paths with the same document;

[0079] S53, Processing and Verification:

[0080] S531, logical check: device and classification level consistency check;

[0081] S532, manual intervention interface: provides a visual label tree editing interface to support experts to correct prediction results and provide feedback to the model;

[0082] S54, Model Iterative Optimization:

[0083] S541, Active Learning: Automatically screen low-confidence samples and give them priority to manual labeling. Sample selection is based on confidence scores, see Formula 9:

[0084]

[0085] Where P(l|d) is the probability of the model predicting label l, and L is the total number of candidate labels;

[0086] S542. Dynamic update: When a new device model is added or the classification standard is revised, incremental training is triggered.

[0087] Furthermore, the method further includes step S6, which uses a multi-way search and recall method combining the BM25 keyword matching algorithm and vector similarity matching, including the following steps:

[0088] S61. Input preprocessing: Improve and expand the user input question through a large language model, divide the user input into chunks, perform keyword analysis, and expand the same business terms;

[0089] S62. BM25 retrieval: Set weight enhancements for fields such as device type, protocol type, and device name, support wildcard matching, see Formula 10:

[0090]

[0091] Among them, ω f is the field weight, k1 = 1.2, b = 0.75 are empirical parameters, |d| is the document length, and avgdl is the average document length;

[0092] S63. Vector similarity query: Calculate the cosine similarity between the query vector and the document vector. Let the query vector be q, the document vector be d, and the similarity be, see Formula 11:

[0093]

[0094] Only retain the documents that satisfy vector_sim(d, q) ≥ τ, where τ = 0.6 is the preset threshold;

[0095] S64. Union fusion: Combine the results of BM25 and vector retrieval to ensure coverage of more documents;

[0096] S65. Duplicate removal: Remove duplicates by document ID and retain the version with a higher score;

[0097] S66. Quantity control: If the combined result exceeds the preset upper limit, truncate it according to the original score.

[0098] Furthermore, it also includes step S7. Re-rank through an algorithm that multiplies the BM25 keyword matching by a label similarity and then adds the vector similarity, including the following steps:

[0099] S71. Input and preprocessing:

[0100] S711. Input: The document list in the recall stage, including the BM25 score, vector similarity score, document content, label tree path, and user query statement;

[0101] S712. Preprocessing:

[0102] S7121. Query label path extraction: Use a pre-trained model to match and extract possible label paths from the query;

[0103] S7122. Document label path parsing: Extract the label tree path from the metadata of the recalled documents;

[0104] S72. Label Similarity Calculation: Measure the matching degree between the query label path and the document label path;

[0105] S721. Calculation Method:

[0106] S7211. Path Alignment: Align the query label path and the document label path hierarchically;

[0107] S7212. Hierarchical Similarity Calculation: Use Jaccard similarity to calculate the matching degree of each layer, assign a weight to each layer, and use exponential weights that increase as the hierarchy deepens. The similarity of the i-th layer is as shown in Formula 12:

[0108]

[0109] S7213. Weighted Similarity Calculation: Multiply the Jaccard similarity of each layer by the corresponding weight, then sum and divide by the total weight, as shown in Formula 13:

[0110]

[0111] where q i and d i represent the label sets of the query and the document at the i-th layer respectively, n is the depth of the label tree, ω i is the weight of the i-th layer, and ω i = 2 i The weights increase exponentially with depth;

[0112] S73. Comprehensive Sorting Score Calculation:

[0113] S731. Combine the BM25 score, label similarity, and vector similarity to calculate the final sorting score of the document, as shown in Formula 14:

[0114] final score(d) = α·label_sim(q,d) + β·BM25(d,q) + γ·vector_sim(d,q) Formula 14;

[0115] where BM25(d,q) is the BM25 score of the document, label_sim(q,d) is the label similarity between the query and the document, vector_sim(d,q) is the vector similarity score of the document, and α, β, γ are weight parameters;

[0116] S74. Sort according to the final score to obtain the result.

[0117] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0118] By optimizing the RAG technology, intelligent embedding vector conversion is realized for air traffic control communication and navigation service documents, enabling target retrieval to accurately capture the core elements of business knowledge and their internal relationships. Secondly, through large language models and integrating advanced artificial intelligence technologies, an expert-level business knowledge Q&A system is constructed to provide more accurate and reliable Q&A services. Finally, to further improve the intelligent level of the system, the present invention continuously iteratively optimizes the tag library and business rule library to ensure that the provided services are always in the best state. BRIEF DESCRIPTION OF THE DRAWINGS

[0119] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0120] Figure 1 It is a schematic flowchart of a method for an intelligent Q&A system of air traffic control communication and navigation service knowledge of the present invention;

[0121] Figure 2 It is a schematic flowchart of step S3 in the method of the intelligent Q&A system of the present invention;

[0122] Figure 3 It is a schematic flowchart of step S4 in the method of the intelligent Q&A system of the present invention;

[0123] Figure 4 It is a schematic flowchart of step S5 in the method of the intelligent Q&A system of the present invention;

[0124] Figure 5 It is a schematic flowchart of step S6 in the method of the intelligent Q&A system of the present invention;

[0125] Figure 6 It is a schematic flowchart of step S7 in the method of the intelligent Q&A system of the present invention;

[0126] Figure 7 It is a schematic diagram of the hierarchical tag tree of air traffic control communication and navigation equipment in step S5 of the method of the intelligent Q&A system of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0127] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0128] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention.

[0129] All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0130] It should be noted that like reference numerals and letters denote like items in the following drawings. Therefore, once an item is defined in one drawing, it does not require further definition and explanation in subsequent drawings.

[0131] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", etc. indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as a limitation of the present invention.

[0132] RAG (Retrieval-Augmented Generation): A method that combines information retrieval technology and generative AI to enhance the capabilities of large language models, enabling them to obtain factual information from external data sources and generate answers based on it.

[0133] LLM (Large Language Model): Large language models refer to deep learning models trained on a large amount of text data, which can understand and generate human language.

[0134] Layout recognition: Refers to a technology in the field of computer vision used to analyze layout elements in images or scanned documents, such as titles, paragraphs, pictures, etc., and identify and locate them.

[0135] Such asFigures 1-6 As shown, it is a schematic diagram of the specific implementation process in each step of the intelligent Q&A system method for the knowledge of air traffic control communication and navigation services of the present invention.

[0136] Based on the basic RAG architecture, the present invention aims at the unique document layout format and knowledge structure in the field of air traffic control communication and navigation. Through innovative designs in key links such as document content extraction, classification slicing, knowledge retrieval and recall, and answer generation, it realizes high-precision content extraction of air traffic control communication and navigation service documents, and improves the accuracy of content classification and slicing. The present invention can effectively optimize the knowledge organization and management process. When querying and retrieving knowledge, it can efficiently and accurately recall relevant content; in the business knowledge Q&A session, it can match question tags, accurately identify the question subject, generate more targeted answers, and improve the knowledge response quality and efficiency of the system in air traffic control communication and navigation applications.

[0137] An intelligent Q&A system method for air traffic control communication and navigation service knowledge includes the following steps:

[0138] S1. Document content extraction and warehousing process;

[0139] S11. Content extraction and in-depth analysis:

[0140] S111. Text structuring: After obtaining the document materials, the system will first perform structured analysis and layout recognition on the document content, converting the originally disordered information into text data with a clear hierarchical structure;

[0141] S112. Key information extraction: Through natural language processing technology and document parsing algorithms, perform in-depth analysis of the text content, accurately extract the entity and relationship knowledge elements therein, and convert unstructured data into a structured data form;

[0142] S12. Classification slicing processing:

[0143] S121. Text slicing: The system will perform fine-grained division of the document content according to logical units, such as paragraphs, sentences or semantic units, for subsequent retrieval and invocation. This slicing processing method improves the content management efficiency and makes information retrieval or knowledge Q&A more accurate and efficient;

[0144] S122. Content classification: Based on the air traffic control communication and navigation service knowledge framework and the document-related equipment information, perform intelligent classification on the sliced content, and through the association analysis algorithm, mark different types of information with business labels respectively for accurate content management and retrieval;

[0145] S13. Document vectorization and persistence:

[0146] S131. The document content processed through S11 and S12 will be converted into a vectorized representation and stored in a vector database for subsequent retrieval and invocation;

[0147] S2. Content retrieval and generation process;

[0148] S21. Problem analysis and rewriting:

[0149] S211. When a user poses a question or query request, the system analyzes and understands the question, extracts the keywords, semantics, and business term nouns of the question, and rewrites the original question through a large language model to make it more in line with the retrieval requirements;

[0150] S22. Multi-channel retrieval and recall:

[0151] S221. Vector similarity matching: Precise matching based on semantic vectors generated by deep learning;

[0152] S222. Keyword matching: Ensure comprehensive retrieval of basic information;

[0153] S223. The multi-channel retrieval strategies complement each other to recall content fragments and knowledge entries related to the question to the greatest extent;

[0154] S23. Retrieval result re-ranking:

[0155] S231. The preliminary results obtained from the retrieval usually contain a large amount of information, including both highly relevant content and possibly data with low or no relevance. The system optimizes the results through the following methods: BM25 keyword score calculation, vector similarity calculation, and label similarity calculation;

[0156] S232. Through the optimization process in step S231, the system can improve the accuracy and relevance of the retrieval results, laying a foundation for generating higher-quality answers in the subsequent steps;

[0157] S24. Generating answers using a large language model:

[0158] S241. Based on the optimized retrieval results, the system invokes a large language model for in-depth processing;

[0159] S242. Context integration: Integrate multi-dimensional information of the user's question and retrieval results;

[0160] S243. Logical reasoning: Perform intelligent reasoning based on the content of the knowledge base and business rules;

[0161] S244. Structured output: Generate the final answer in a clear and rigorous logical framework.

[0162] Through the orderly connection of the above steps S21, S22, S23, and S24, the system realizes a complete closed-loop from user query to accurate answer, significantly improving the quality and efficiency of knowledge response in air traffic control communication and navigation services, and providing reliable support for relevant decision-making.

[0163] Content extraction and slicing of air traffic control communication and navigation knowledge documents: For the special layout format and document content classification of air traffic control communication and navigation service materials, the present invention proposes a series of methods for automatic content recognition and extraction, document content segmentation, and content tag tree extraction, aiming to improve the extraction accuracy of air traffic control communication and navigation knowledge content and the accuracy of slice classification.

[0164] For the layout of power flow diagrams, configuration diagrams, and standard documents in the air traffic control industry, train and optimize a layout recognition model to increase the content extraction accuracy.

[0165] Further implement steps S11 and S12 through step S3;

[0166] Including step S3, optimize content slicing by extracting and calculating the similarity of adjacent paragraphs or content, so that the complete content is within the same slice, improving the content retrieval efficiency, including the following steps:

[0167] S31. Content preprocessing: First, standardize the air traffic control communication and navigation service materials, including adjusting the document resolution, removing noise and interference elements, and preliminarily analyzing the layout to identify key area titles, text, and charts;

[0168] S32. Feature extraction: Through a multi-dimensional feature extraction method, obtain the visual features and context semantic features of the text content. At the same time, perform unique feature encoding on the chart area for subsequent classification;

[0169] S33. Similarity calculation: Based on the extracted multi-dimensional features, use the cosine similarity algorithm to calculate the similarity value between adjacent paragraphs or content blocks, as shown in Formula 1, and construct a content association matrix;

[0170]

[0171] Among them, v i and v j are the feature vectors of two content blocks i and j. Then, filter out highly relevant content segments through a dynamic threshold (T), as shown in Formula 2:

[0172] T = μ + kσ Formula 2;

[0173] Among them, μ is the average similarity of the current document, σ is the standard deviation, and k is an empirical coefficient;

[0174] S34. Content Slicing Optimization: Based on the content correlation matrix, the document content is intelligently sliced, and the content with high similarity and logical coherence is grouped into the same block. At the same time, for chart content, through shape feature and surrounding text matching, accurate recognition and separate division are achieved. The shape matching of the chart area uses the improved Hausdorff distance. The similarity between two shapes A and B is defined as shown in Formula 3:

[0175]

[0176] Combined with the semantic similarity S of the surrounding text text , comprehensively determine whether it is the same slice, as shown in Formula 4:

[0177] ChartMatchScore = α·d H (A, B)+β·S text (α + β = 1) Formula 4;

[0178] ChartMatchScore > θ

[0179] When it is classified as the same slice;

[0180] S35. Label Tree Construction: By analyzing the sliced content blocks, a label tree is constructed step by step from the basic level to the detailed level using a hierarchical classification algorithm. Each node represents a classification label, and its weight is determined by the importance and frequency of the content block. The node weight ω is jointly calculated by the TF-IDF frequency and the semantic significance score, as shown in Formula 5:

[0181] ω = λ·TF-IDF(t)+(1 - λ)·Softmax(f DL (t)) Formula 5;

[0182] Among them, f DL is the semantic feature vector output by the deep learning model, and λ ∈ [0, 1] is the balance coefficient. In this process, a deep learning model is used to perform topic model analysis on the text and automatically extract high-level semantic features;

[0183] S36. Model Training and Optimization: The labeled sample data is input into the neural network model for training. A multi-task learning framework is used to optimize the content slicing and label classification tasks simultaneously. During the training process, the model parameters are continuously adjusted through cross-validation and performance evaluation to ensure robustness on different types of documents. The total loss function is the weighted sum of the losses of the two tasks, as shown in Formula 6:

[0184] L total = γL slice +(1 - λ)L classify Formula 6;

[0185] Among them, Lslice is the contrast loss for the slicing task, L classify is the cross-entropy loss for the classification task, and γ is dynamically adjusted by gradient normalization;

[0186] S37. Precise Identification and Application: The final label tree structure is used in the subsequent content retrieval system to achieve rapid positioning and efficient invocation of relevant knowledge. Through the automated processing of air traffic control business materials, the accuracy of data extraction and the precision of classification have been significantly improved, providing strong support for intelligent management and decision-making.

[0187] The above-mentioned method of step S3 makes full use of multi-dimensional features and deep learning technologies, significantly improving the processing ability of complex layout documents, and is an important innovation in solving the problem of knowledge extraction in the field of air traffic control communication and navigation.

[0188] Content Body Identification and Label Tree Extraction: In terms of content body identification, the present invention adopts advanced OCR technology combined with a professional layout recognition algorithm to comprehensively analyze the input document. Through multi-dimensional feature extraction, the system can accurately identify and extract core elements such as text, tables, and pictures in the document. In terms of document content classification, based on the professional knowledge system in the field of air traffic control business, a hierarchical document label tree structure is constructed, and labels are divided according to the air traffic control equipment information described in the document: it can be divided into primary equipment types, secondary equipment types, tertiary equipment types, equipment, equipment boards, and sub-modules, etc., to achieve label classification of document content and ensure that key information can be quickly identified and processed.

[0189] In steps S4 and S5, during the processes of steps S12 and S13, the main body of the existing slice is identified and a label tree is generated. At the same time, steps S4 and S5 will also identify the main body of the user's question after step S21 and before step S23, and then generate a label tree.

[0190] Step S4. Unification of Business Names. In the case where there are different names for the same content body, to improve the accuracy of keyword search, a synonym dictionary is used to replace the keywords of the content body in the business materials with synonyms, covering the core concepts and common expressions of air traffic control communication and navigation business, ensuring that the business names are consistent, including the following steps:

[0191] S41. Construction of Synonym Dictionary:

[0192] S411. Term Extraction: Extract different names of the content body from the air traffic control communication and navigation business materials; such as "radar system" and "radio detection and ranging equipment";

[0193] S412. Establish a mapping: Classify different expressions of the same subject into synonym groups, define a subject, i.e., a standardized name, and a subordinate term, i.e., an alias; for example, subject: "ADS-B", subordinate terms: ["Automatic Dependent Surveillance-Broadcast"];

[0194] S413, Manual Verification: Domain experts review synonym mapping relationships to ensure semantic consistency;

[0195] S42. Document preprocessing and keyword matching:

[0196] S421, Word Segmentation and Entity Recognition: Use NLP tools to segment documents and identify business-related entities, such as device names and protocol types.

[0197] S422, synonym matching: traverse the keywords in the document and query the synonym dictionary;

[0198] S43. Search index construction:

[0199] S431. Inverted index generation: Build an inverted index for each main word, record the location of documents containing the word, and append synonym extension fields to the index, such as a list of associated subordinate words for the main word "ADS-B". The definition of the inverted index is as shown in Formula 7:

[0200] Index(m)={(d,pos d )|d∈D,m∈MainTerms(d)∪Synonyms(m)} Formula 7;

[0201] Among them, MainTerms(d) is the set of main terms in document d, and Synonyms(m) is the set of subordinate terms of main term m, mapped through the synonym dictionary;

[0202] S432. Weight Assignment: Differentiated weights are assigned to the main and subordinate terms, with the main term receiving a higher weight, improving search ranking. For example, the main term "ADS-B" has a weight of 1.0, and the subordinate term "Automatic Dependent Surveillance-Broadcast" has a weight of 0.5. When the user enters "Automatic Dependent Surveillance-Broadcast System," the system identifies the subordinate term as matching the main term "ADS-B" and returns documents containing the main term. Weight sorting ensures that documents with the most main term occurrences are displayed first.

[0203] S44. Query processing:

[0204] S441, query expansion: according to the synonym dictionary, the search term is replaced with the main word (such as "VHF → "very high frequency") and all synonyms are associated;

[0205] S442. Index Retrieval: Use the expanded keyword set and the main word + synonym to retrieve the inverted index, and return all matching documents.

[0206] Example of Step S4: Original document: In a certain equipment manual, it is written that "secondary radar needs to be used in conjunction with an S-mode transponder";

[0207] After replacement: Standardized to "secondary surveillance radar needs to be used in conjunction with an S-mode transponder" (assuming that "secondary radar" is a subordinate word and "secondary surveillance radar" is the main word);

[0208] User search: When the user searches for "secondary radar", the system automatically associates the main word and returns all documents containing "secondary surveillance radar";

[0209] This step significantly improves the search recall rate and accuracy through business name standardization and query expansion, and is especially suitable for the air traffic control communication and navigation business field with strong professionalism and many term variants.

[0210] Customized Tag Tree Recognition:

[0211] Step S5. According to the applicable site and business entity described in the business materials, carry out customized tag tree recognition and filling work, and train a content classification and tag extraction model using existing materials and manual tags, including the following steps:

[0212] S51. Tag System Construction and Data Preparation:

[0213] S511. Tag Tree Structuring: Sort out the equipment hierarchical relationship from historical air traffic control communication and navigation materials, construct a tree structure, and the organizational form is as Figure 7 shown:

[0214] S512. Document-Tag Alignment: Associate the manually annotated document fragments with the corresponding tag paths to form a training set; for example: Document description "Antenna frequency calibration of THALES radar" → Tag path: Information detection system → Secondary surveillance radar → THALES single secondary radar → Outdoor antenna unit;

[0215] S513. Domain Dictionary Enhancement: Integrate the business synonym dictionary (reusing the previous solution) to solve the problem of equipment aliases;

[0216] S52. Hierarchical Tag Prediction:

[0217] S521. Hierarchical Constraint: The prediction result of the upper-level tag is used as the candidate range of the lower-level tag (for example, when predicting the first level as "surveillance equipment", the second level is only selected from "radar system" / "ADS-B", etc.), and the hierarchical tag prediction obeys the conditional probability constraint, see Formula 8:

[0218]

[0219] Among them, l i is the label of the i-th layer, and Children(l i ) is the set of sub-labels of l i-1 ;

[0220] S522, Multi-label processing: Support associating multiple label paths with the same document, such as involving radar and navigation equipment at the same time;

[0221] S53, Processing and verification:

[0222] S531, Logical verification: Check the consistency between the equipment and the classification hierarchy. For example, "RS VHF transmitting system" must belong to "Communication / Air-ground communication / VHF radio station";

[0223] S532, Manual intervention interface: Provide a visual label tree editing interface to support experts in correcting the prediction results and feedback them to the model;

[0224] S54, Model iterative optimization:

[0225] S541, Active learning: Automatically screen low-confidence samples and give them to manual annotation first. The sample selection is based on the confidence score. See formula 9:

[0226]

[0227] Among them, P(l|d) is the probability that the model predicts label l, and L is all candidate labels;

[0228] S542, Dynamic update: When a new equipment model is added or the classification standard is revised, trigger incremental training.

[0229] Example of step S5, input document: "RS4200 Radio Parameter Viewing Operation Guide Manual" with keywords: such as "RS4200 radio";

[0230] Output label tree: Communication / Air-ground communication equipment / VHF radio station / RS4200 radio;

[0231] This step realizes the accurate hierarchical classification of the document content through the dual drive of semantic understanding and business rules.

[0232] Through air traffic control knowledge retrieval, recall and re-ranking:

[0233] The knowledge retrieval, recall, and re-ranking processes are two crucial core stages in the entire retrieval process, jointly constituting an efficient and accurate knowledge retrieval mechanism. First, in the knowledge retrieval and recall stage, we adopt a multi-path retrieval and recall strategy. By simultaneously initiating multiple different retrieval paths, this method can effectively improve the comprehensiveness and accuracy of the retrieval, thus providing a richer and more reliable data basis for subsequent processing. Second, in the retrieved documents, we introduce an advanced re-ranking algorithm. This process aims to intelligently re-arrange the initially retrieved documents according to a pre-set weight and scoring mechanism. Specifically, the system automatically extracts the key features of each document and comprehensively evaluates its relevance and importance through a complex scoring mechanism, finally generating the optimal order that meets the user's needs to help the user find the required information more quickly.

[0234] Steps S22 and S23 are further implemented through steps S6 and S7;

[0235] The retrieval uses multi-path retrieval and recall:

[0236] Step S6: Use a combination of the BM25 keyword matching algorithm and vector similarity matching. This stage aims for a large number and comprehensiveness, rather than pursuing the degree of matching relevance, and includes the following steps:

[0237] S61. Input preprocessing: Improve and expand the user input question through a large language model, block the user input, perform keyword analysis, and expand the same business nouns;

[0238] S62. BM25 retrieval: Set weights for fields such as device type, protocol type, and device name to enhance support for wildcard matching (such as "ILS-*"), as shown in Formula 10:

[0239]

[0240] where ω f is the field weight (such as for device name ω f = 2.0, for protocol type ω f = 1.5), k1 = 1.2, b = 0.75 are empirical parameters, |d| is the document length, and avgdl is the average document length;

[0241] S63. Vector similarity query: Calculate the cosine similarity between the query vector and the document vector. Let the query vector be q and the document vector be d, and the similarity is as shown in Formula 11:

[0242]

[0243] Only retain the documents that satisfy vector_sim(d, q) ≥ τ, where τ = 0.6 is a preset threshold;

[0244] S64, Union Fusion: Combine the results of BM25 and vector retrieval to ensure coverage of more documents;

[0245] S65, Deduplication: Deduplicate by document ID and retain the version with a higher score (such as BM25 score or vector similarity score);

[0246] S66, Quantity Control: If the merged result exceeds the preset upper limit (such as 2000 articles), truncate according to the original score.

[0247] Re-ranking method based on air traffic control communication business classification and label tree:

[0248] Step S7, Re-rank through an algorithm that multiplies BM25 keyword matching by a label similarity and then adds vector similarity, including the following steps:

[0249] S71, Input and Preprocessing:

[0250] S711, Input: The document list in the recall stage, including BM25 score, vector similarity score, document content, label tree path, and user query statement;

[0251] S712, Preprocessing:

[0252] S7121, Query Label Path Extraction: Use a pre-trained model to match and extract possible label paths from the query; for example: query "RS radio indicator status" → extract label path ["Communication", "Ground-air communication equipment", "VHF radio", "RS VHF radio"];

[0253] S7122, Document Label Path Parsing: Extract the label tree path from the metadata of the recalled document; such as ["Communication", "Ground-air communication equipment", "VHF radio", "RS VHF radio"];

[0254] S72, Label Similarity Calculation: Measure the matching degree between the query label path and the document label path;

[0255] S721, Calculation Method:

[0256] S7211, Path Alignment: Align the query label path and the document label path hierarchically; if the path lengths are inconsistent, empty sets can be appended to the shorter path; for example: query path: ["Communication", "Ground-air communication equipment", "VHF radio", ""], document path: ["Communication", "Ground-air communication equipment", "VHF radio", "RS VHF radio"];

[0257] S7212. Hierarchical Similarity Calculation: Calculate the matching degree of each layer using Jaccard similarity, assign a weight to each layer, and use an exponential weight that increases as the hierarchy deepens. The similarity of the i-th layer is shown in Equation 12:

[0258]

[0259] S7213. Weighted Similarity Calculation: Multiply the Jaccard similarity of each layer by the corresponding weight, then sum and divide by the total weight. See Equation 13:

[0260]

[0261] where q i and d i represent the label sets of the query and the document at the i-th layer respectively, n is the depth of the label tree, ω i is the weight of the i-th layer, and ω i = 2 i The weight increases exponentially with depth;

[0262] S73. Comprehensive Ranking Score Calculation:

[0263] S731. Combine the BM25 score, label similarity, and vector similarity to calculate the final ranking score of the document. See Equation 14:

[0264] final score(d) = α·label_sim(q, d)+β·BM25(d, q)+γ·vector_sim(d, q) Equation 14;

[0265] where BM25(d, q) is the BM25 score of the document, label_sim(q, d) is the label similarity between the query and the document, vector_sim(d, q) is the vector similarity score of the document, and α, β, γ are weight parameters. Usually, α + β + γ = 1 (e.g., α = 0.6, β = 0.3, γ = 0.1);

[0266] S74. Sort according to the final score to obtain the result.

[0267] Example of Step S7. Input query: "What are the statuses of the status lights on the RS4200 radio panel?";

[0268] Retrieved documents:

[0269] Document A: BM25 score: 0.9, vector similarity score: 0.8, label path: ["Communication", "Ground-air communication equipment", "VHF radio", "OTE radio"];

[0270] Document B: BM25 score: 0.7, vector similarity score: 0.85, label path: ["Communication", "Ground-air communication equipment", "VHF radio", "RS VHF radio"];

[0271] Calculation process:

[0272] Extract the query label path: Query path: ["Communication", "Ground-air communication equipment", "VHF radio", "RS radio"];

[0273] Calculate the label similarity: Document A: label_sim = 0.47, Document B: label_sim = 1.0;

[0274] Calculate the comprehensive score:

[0275] Document A:

[0276] final_score = 0.6 * 0.47 + 0.3 * 0.9 + 0.1 * 0.8 = 0.282 + 0.27 + 0.08 = 0.632;

[0277] Document B:

[0278] final_score = 0.6 * 1.0 + 0.3 * 0.7 + 0.1 * 0.85 = 0.6 + 0.21 + 0.085 = 0.895;

[0279] Sorting result: Document B > Document A

[0280] This re-ranking method realizes accurate ranking of the recall results by combining the BM25 score, label similarity, and vector similarity. Through dynamic weight adjustment and fuzzy matching, the flexibility and accuracy of the ranking are further improved.

[0281] The present invention has been described exemplarily above in conjunction with the accompanying drawings. Obviously, the specific implementation of the present invention is not limited by the above embodiments. Those skilled in the art can make various modifications or variations to the present invention without departing from the technical concept of the present invention, and these modifications or variations will of course fall within the protection scope of the present invention.

Claims

1. An intelligent question-answering system method for air traffic control communication business knowledge, characterized in that: It includes the following steps: S1. Document content extraction and warehousing process; S11. Content extraction and in-depth analysis: S111. Text structuring process: After obtaining the document materials, the system will first perform structural analysis and layout recognition on the document content, converting the originally chaotic information into text data with a clear hierarchical structure; S112. Key information extraction: Through natural language processing technology and document parsing algorithms, the text content is deeply analyzed to accurately extract the entity and relationship knowledge elements therein, converting unstructured data into a structured data form; S12. Classification and slicing process: S121. Text slicing: The system will perform fine-grained division of the document content according to logical units, paragraphs, sentences or semantic units for subsequent retrieval and invocation. This slicing method improves the content management efficiency and makes information retrieval or knowledge Q&A more accurate and efficient; S122. Content classification: Based on the air traffic control communication, navigation and surveillance business knowledge framework and the document-related equipment information, the sliced content is intelligently classified, and through the association analysis algorithm, different types of information are marked with business labels respectively for accurate content management and retrieval; S13. Document vectorization and persistence: S131. The document content processed by S11 and S12 will be converted into a vectorized expression form and stored in the vector database for subsequent retrieval and invocation; S2. Content retrieval and generation process; S21. Question analysis and rewriting: S211. When the user raises a question or query request, the system will analyze and understand the question, extract the keywords, semantics and business term nouns of the question, and rewrite the original question through a large language model to make it more in line with the retrieval requirements; S22. Multi-channel retrieval and recall: S221. Vector similarity matching: Accurate matching based on the semantic vectors generated by deep learning; S222. Keyword matching: Ensure comprehensive retrieval of basic information; S223. The multi-channel retrieval strategies complement each other to recall content fragments and knowledge entries related to the question to the greatest extent; S23. Retrieval result re-ranking: S231. The preliminary results obtained by retrieval usually contain a large amount of information, including both highly relevant content and possibly lowly or even unrelated data mixed in. The system optimizes the results through the following methods: BM25 keyword score calculation, vector similarity calculation, label similarity calculation; S232. Through the optimization process in step S231, the system can improve the accuracy and relevance of the retrieval results, laying a foundation for generating higher-quality answers subsequently; S24. Large language model generates answers: S241. Based on the optimized retrieval results, the system calls a large language model for in-depth processing; S242. Context integration: Integrate multi-dimensional information of the user's question and retrieval results; S243. Logical reasoning: Intelligent reasoning based on the content of the knowledge base and business rules; S244. Structured output: Generate the final answer with a clear and rigorous logical framework.

2. The intelligent question-answering system method for air traffic communication and navigation service knowledge according to claim 1, wherein: It also includes step S3. Through the extraction and similarity calculation of adjacent paragraphs or content, content slicing optimization is carried out to ensure that the complete content is within the same slice, improving the content retrieval efficiency, including the following steps: S31. Content preprocessing: First, standardize the air traffic control communication and navigation business materials, including adjusting the document resolution, removing noise and interference elements, and conducting a preliminary analysis of the layout to identify key area titles, text, and charts; S32. Feature extraction: Through a multi-dimensional feature extraction method, obtain the visual features and context semantic features of the text content. At the same time, perform unique feature encoding on the chart area for subsequent classification; S33. Similarity calculation: Based on the extracted multi-dimensional features, use the cosine similarity algorithm to calculate the similarity values between adjacent paragraphs or content blocks, as shown in Formula 1, and construct a content association matrix; where v i and v j are the feature vectors of two content blocks i and j, and then high-correlation content segments are selected through a dynamic threshold (T), as shown in Equation 2: T = μ + kσ Formula 2; where μ is the average similarity of the current document, σ is the standard deviation, and k is an empirical coefficient; S34. Content slicing optimization: Based on the content association matrix, perform intelligent slicing on the document content, grouping content with high similarity and logical coherence into the same block. At the same time, for chart content, through shape feature and surrounding text matching, achieve accurate recognition and separate classification. The shape matching of the chart area uses the improved Hausdorff distance. The similarity between two shapes A and B is defined as, as shown in Formula 3: Combined with the semantic similarity S of the surrounding text text , comprehensively determine whether it is the same slice, see Equation 4: ChartMatchScore = α·d H (A,B) + β·S text (α + β = 1) Formula 4; When ChartMatchScore > θ, it is classified into the same slice; S35. Label tree construction: By analyzing the sliced content blocks, use a hierarchical classification algorithm to construct a label tree step by step from the basic level to the detailed level. Each node represents a classification label, and its weight is determined by the importance and occurrence frequency of the content block. The node weight ω is jointly calculated by the TF-IDF frequency and semantic significance score, as shown in Formula 5: ω = λ·TF-IDF(t)+(1 - λ)·Softmax(f DL (t)) Formula 5; Among them, f DL is the semantic feature vector output by the deep learning model, λ ∈ [0, 1] is the balance coefficient. In this process, the deep learning model is used to perform topic model analysis on the text to automatically extract high-level semantic features; S36. Model training and optimization: Input the labeled sample data into a neural network model for training. Adopt a multi-task learning framework to optimize the content slicing and label classification tasks simultaneously. During the training process, continuously adjust the model parameters through cross-validation and performance evaluation to ensure robustness on different types of documents. The total loss function is the weighted sum of the two-task losses, as shown in Formula 6: L total = γL slice + (1 - λ)L classify Formula 6; Among them, L slice is the contrastive loss of the slicing task, and L classify is the cross-entropy loss of the classification task, and γ is dynamically adjusted by gradient normalization; S37. Accurate recognition and application: The final label tree structure is used in the subsequent content retrieval system to achieve rapid positioning and efficient invocation of relevant knowledge. Through the automated processing of air traffic control business materials, the accuracy of data extraction and the precision of classification are greatly improved, providing strong support for intelligent management and decision-making.

3. The intelligent question-answering system method for air traffic control and communication business knowledge according to claim 1, wherein: It also includes step S4. In the case where the same content subject has different names, to improve the accuracy of keyword search, a synonym dictionary is used to replace the keywords of the content subject in the business materials, covering the core concepts and common expressions of air traffic control communication and navigation business to ensure the consistency of business names, including the following steps: S41. Synonym dictionary construction: S411. Term extraction: Extract different names of the content subject from the air traffic control communication and navigation business materials; S412. Establish mapping: Classify different expressions of the same subject into synonym groups, defining the main word, i.e., the standardized name, and the secondary word, i.e., the alias; S413. Manual verification: The domain experts review the synonym mapping relationships to ensure semantic consistency; S42. Document preprocessing and keyword matching: S421. Word segmentation and entity recognition: Use NLP tools to segment the documents and recognize business-related entities; S422. Synonym matching: Traverse the keywords in the documents and query the synonym dictionary; S43. Search index construction: S431. Inverted index generation: Build an inverted index for each main term, record the document positions containing the term, and append a synonym extension field to the index. The definition of the inverted index is shown in Formula 7: Index(m) = {(d, pos d ) | d ∈ D, m ∈ MainTerms(d) ∪ Synonyms(m)} Formula 7; where MainTerms(d) is the set of main terms in document d, and Synonyms(m) is the set of subordinate terms of main term m, which are mapped through the synonym dictionary; S432. Weight assignment: Set different weights for main terms and subordinate terms, with higher weights for main terms to improve search ranking; S44. Query processing: S441. Query expansion: According to the synonym dictionary, replace the search terms with main terms and associate all synonyms; S442. Index retrieval: Use the expanded keyword set, main term + synonyms to retrieve the inverted index and return all matching documents.

4. The intelligent question-answering system method for air traffic control and navigation service knowledge according to claim 3, wherein: It also includes step S5. According to the applicable sites and business entities described in the business materials, carry out customized label tree recognition and filling work, and train a content classification and label extraction model using existing materials and manual labels, including the following steps: S51. Label system construction and data preparation: S511. Label tree structuring: Sort out the equipment hierarchical relationship from historical air traffic control communication and navigation materials to construct a tree structure; S512. Document-label alignment: Associate the manually annotated document fragments with the corresponding label paths to form a training set; S513. Domain dictionary enhancement: Integrate the business synonym dictionary to solve the problem of equipment aliases; S52. Hierarchical label prediction: S521. Hierarchical constraint: The prediction results of the upper-level labels are used as the candidate range for the lower-level labels. The hierarchical label prediction is subject to the conditional probability constraint, as shown in Formula 8: Among them, l i is the label of the i-th layer, and Children(l i ) is the set of sub-labels of l i-1 ; S522. Multi-label processing: Support associating multiple label paths with the same document; S53. Processing and verification: S531. Logical verification: Check the consistency between equipment and classification levels; S532. Manual intervention interface: Provide a visual label tree editing interface to support experts in correcting the prediction results and feedback them to the model; S54. Model iterative optimization: S541. Active learning: Automatically screen the low-confidence samples and give them to manual annotation first. The sample selection is based on the confidence score, as shown in Formula 9: where P(l|d) is the probability of the model predicting label l, and L is all candidate labels; S542. Dynamic update: When a new equipment model is added or the classification standard is revised, trigger incremental training.

5. The intelligent question-answering system method for air traffic control and communication service knowledge according to claim 1, wherein: It also includes step S6. Use a multi-way retrieval and recall method that combines the BM25 keyword matching algorithm and vector similarity matching, including the following steps: S61. Input preprocessing: Improve and expand the user input questions through large language models, block the user input, analyze keywords, and expand the same business nouns; S62, BM25 Retrieval: Increase the weights for fields such as device type, protocol type, and device name, support wildcard matching. See Equation 10: where ω f is the field weight, k1 = 1.2, b = 0.75 are empirical parameters, |d| is the document length, and avgdl is the average document length; S63, Vector Similarity Query: Calculate the cosine similarity between the query vector and the document vector. Let the query vector be q, the document vector be d, and the similarity be. See Equation 11: Only retain the documents that satisfy vector_sim(d, q) ≥ τ, where τ = 0.6 is the preset threshold; S64, Union Fusion: Combine the results of BM25 and vector retrieval to ensure coverage of more documents; S65, Duplicate Removal: Remove duplicates by document ID and retain the version with a higher score; S66, Quantity Control: If the combined result exceeds the preset upper limit, truncate it according to the original score.

6. The intelligent question-answering system method for air traffic control and communication service knowledge according to claim 5, characterized in that: It also includes Step S7, re-rank using the algorithm of multiplying the BM25 keyword match by a label similarity plus the vector similarity, including the following steps: S71, Input and Preprocessing: S711, Input: The document list in the recall stage, including the BM25 score, vector similarity score, document content, label tree path, and user query statement; S712, Preprocessing: S7121, Query Label Path Extraction: Use a pre-trained model to match and extract possible label paths from the query; S7122, Document Label Path Parsing: Extract the label tree path from the metadata of the recalled documents; S72, Label Similarity Calculation: Measure the matching degree between the query label path and the document label path; S721, Calculation Method: S7211, Path Alignment: Align the query label path and the document label path hierarchically; S7212, Hierarchical Similarity Calculation: Use the Jaccard similarity to calculate the matching degree of each layer, assign a weight to each layer, and use an exponential weight that increases as the layer deepens. The similarity of the i-th layer is. See Equation 12: S7213, Weighted Similarity Calculation: Multiply the Jaccard similarity of each layer by the corresponding weight, then sum and divide by the total weight. See Equation 13: where q i and d i represent the label sets of the query and the document at the i-th layer respectively, n is the depth of the label tree, ω i is the weight of the i-th layer, ω i = 2 i The weight grows exponentially with depth; S73, Comprehensive Ranking Score Calculation: S731, Combine the BM25 score, label similarity, and vector similarity to calculate the final ranking score of the document. See Equation 14: final score(d) = α · label_sim(q, d) + β · BM25(d, q) + γ · vector_sim(d, q); Formula 14 Among them, BM25(d, q) is the BM25 score of the document, label_sim(q, d) is the label similarity between the query and the document, vector_sim(d, q) is the vector similarity score of the document, and α, β, γ are weight parameters; S74, Sort according to the final score to obtain the result.

Images