Airport operation and maintenance intelligent question and answer system retrieval method and device, and electronic equipment

Abstract

This invention discloses a retrieval method, device, and electronic device for an intelligent question-and-answer system for airport operations and maintenance, relating to the field of airport operations and maintenance technology. The method includes: semantically rewriting and expanding user question dialogues to obtain a hybrid index containing question keywords and question vectors; performing a hybrid retrieval in a vector library using the hybrid index to obtain retrieval results containing document metadata. The vector library includes document metadata and index information corresponding to multiple document fragments. The document fragments are generated after segmentation, vectorization, and knowledge processing of airport software operations and maintenance documents; inputting the retrieval results into a large language model; using the large language model for question association reasoning; and outputting the question reasoning results. This invention solves the technical problem in related technologies where, in airport operations and maintenance scenarios, the use of retrieval enhancement generation technology results in low retrieval recall, leading to missing, incorrect, or even misleading answers, thus reducing user trust.

Description

Technical Field

[0001] This invention relates to the field of airport operation and maintenance technology or other related fields. Specifically, it relates to a retrieval method and device, and electronic equipment for an intelligent question-and-answer system for airport operation and maintenance. Background Technology

[0002] With the deepening of smart airport construction, airport operation and support systems are becoming increasingly complex, involving multiple key subsystems such as check-in, departure, flight information display, baggage sorting, and energy monitoring. Their operation and maintenance work places extremely high demands on real-time performance, accuracy, and professionalism. Current airport operation and maintenance methods rely on manual consultation of paper manuals, documents, or internal knowledge bases, resulting in long response delays, low retrieval efficiency, and difficulty in handling sudden system failures. In recent years, intelligent question-answering systems based on retrieval-enhanced generation technology have been introduced into airport operation and maintenance scenarios, aiming to improve the efficiency of operation and maintenance decision-making by semantically retrieving external knowledge bases and combining them with large language models to generate accurate answers. However, retrieval-enhanced generation systems suffer from severe insufficient retrieval recall in the context of professional airport operation and maintenance, leading to missing, incorrect, or even misleading answers, seriously restricting the system's usability and user trust.

[0003] There is currently no effective solution to the above problems. Summary of the Invention

[0004] This invention provides a retrieval method, device, and electronic device for an intelligent question-and-answer system for airport operations and maintenance, in order to at least solve the technical problem in related technologies where, in airport operations and maintenance scenarios, the use of retrieval enhancement generation technology results in low retrieval recall, leading to missing, incorrect, or even misleading answers, and reducing user trust.

[0005] According to one aspect of the present invention, a retrieval method for an intelligent question-and-answer system for airport operation and maintenance is provided, comprising: receiving a user question dialogue, and performing question semantic rewriting and question expansion on the user question dialogue to obtain a hybrid index containing question keywords and question vectors; using the hybrid index to perform a hybrid retrieval in a pre-built vector library to obtain retrieval results containing document metadata, wherein the vector library includes document metadata and index information corresponding to multiple document fragments, the document fragments being document fragments generated after fragment segmentation, vectorization, and knowledge processing of airport software operation and maintenance documents, the document fragments being segmented according to a multi-level heading structure, each document fragment retaining its parent heading; inputting the retrieval results into a large language model, using the large language model to perform question association reasoning, and outputting question reasoning results.

[0006] Optionally, constructing the vector library includes: acquiring all airport software operation and maintenance documents from the historical process and determining the document type of each document; performing document preprocessing and document type conversion on all the airport software operation and maintenance documents based on the document type; segmenting the converted airport software operation and maintenance documents using a preset segmentation tool to obtain multiple initial document fragments, wherein each document fragment retains its parent title and appends the original document name as a source identifier; performing vectorization and Gaussian mixture clustering on the initial document fragments, and sending the clustered document fragment set to a large language model, whereby the large language model generates fragment summaries to complete the knowledge processing operation; constructing a vector index associated with each document fragment in the document fragment set based on the result of vectorization of the initial document fragments; constructing a keyword index based on the fragment summary corresponding to each document fragment; and constructing the vector library based on the fragment summary, vector index, and keyword index of each document fragment in the document fragment set.

[0007] Optionally, based on the document type, the steps of preprocessing and converting all airport software operation and maintenance documents include: if the document type is a text document, standardizing the expression and writing of the document title and subheadings at all levels, and adding corresponding heading styles; if the document type is a table document, standardizing the names of each table on each page, and extracting the header row from the first row of each table to obtain the header field; if the document type is an image document, a scanned document, or a document with font insertion, converting the airport software operation and maintenance document into a text document or a table document.

[0008] Optionally, the steps of preprocessing and converting all airport software operation and maintenance documents based on the document type include: when the document type is a text document, using a title segmentation tool to convert the preprocessed airport software operation and maintenance documents into one-to-one correspondences of document titles; when the document type is a table document, using a title segmentation tool to convert the table names of each page of the preprocessed airport software operation and maintenance documents into first-level headings, and converting the header row of the first table in each page into a second-level heading.

[0009] Optionally, the step of segmenting the converted airport software operation and maintenance document using a preset segmentation tool to obtain multiple initial document fragments includes: segmenting the converted airport software operation and maintenance document according to the heading styles of each level using a title segmentation tool to obtain initial document fragments corresponding to multiple levels of headings; if the size of any initial document fragment is detected to be greater than a preset threshold, performing a secondary segmentation of the initial document fragment using a text segmentation tool; after the segmentation process, retaining the headings of each level for the initial document fragments, and appending the standardized document name of the airport software operation and maintenance document to which each initial document fragment belongs as source information to the text content of the initial document fragment.

[0010] Optionally, the step of using a text segmentation tool to perform secondary segmentation on the initial document fragment includes: generating a semantic boundary-based text segmentation tool using a lightweight natural language model for the operation and maintenance domain; obtaining a continuous sentence sequence of the context window of the initial document fragment; inputting the initial document fragment and the continuous sentence sequence of the context window into the text segmentation tool, and using the text segmentation tool to perform secondary segmentation on the initial document fragment.

[0011] Optionally, the steps of vectorizing and performing Gaussian mixture clustering on the initial document fragments include: vectorizing each initial document fragment to obtain an initial vector; performing vector dimensionality reduction on the initial vectors to obtain a target document vector; performing associative semantic clustering on the target document vectors using a Gaussian mixture clustering algorithm to obtain a document fragment set; and if the number of document fragments in the document fragment set is greater than a preset threshold, re-vectorizing, performing vector dimensionality reduction, and performing associative semantic clustering on the document fragments in the document fragment set to obtain a new document fragment set.

[0012] According to another aspect of the present invention, a retrieval device for an intelligent question-and-answer system for airport operation and maintenance is also provided, comprising: a semantic rewriting unit, configured to receive user question dialogues and perform question semantic rewriting and question expansion on the user question dialogues to obtain a hybrid index containing question keywords and question vectors; a hybrid retrieval unit, configured to use the hybrid index to perform hybrid retrieval in a pre-built vector library to obtain retrieval results containing document metadata, wherein the vector library includes document metadata and index information corresponding to multiple document fragments, the document fragments being document fragments generated after fragment segmentation, vectorization, and knowledge processing of airport software operation and maintenance documents, the document fragments being segmented according to a multi-level heading structure, and each document fragment retaining its parent heading; and a question reasoning unit, configured to input the retrieval results into a large language model, use the large language model to perform question association reasoning, and output question reasoning results.

[0013] Optionally, when constructing the vector library, the retrieval device of the airport operation and maintenance intelligent question-and-answer system includes: a software operation and maintenance document acquisition unit, used to acquire all airport software operation and maintenance documents in the historical process and determine the document data type of each airport software operation and maintenance document; a preprocessing unit, used to perform document preprocessing and document type conversion on all airport software operation and maintenance documents based on the document data type; and a document segmentation unit, used to segment the converted airport software operation and maintenance documents using a preset segmentation tool to obtain multiple initial document fragments, wherein each document fragment retains its parent title and appends the original document name as a source identifier; and directs... The quantization processing unit is used to perform vectorization processing and Gaussian mixture clustering processing on the initial document fragments, and send the clustered document fragment set to the large language model, which generates fragment summaries to complete the knowledge processing operation; the vector index construction unit is used to construct a vector index associated with each document fragment in the document fragment set based on the result of vectorization processing of the initial document fragments; the keyword index construction unit is used to construct a keyword index based on the fragment summary corresponding to each document fragment; and the vector library construction unit is used to construct the vector library based on the fragment summary, vector index, and keyword index of each document fragment in the document fragment set.

[0014] Optionally, the preprocessing unit includes: a text standardization module, used to standardize the expression and writing of the file title and subheadings of the airport software operation and maintenance document when the document type is a text document, and to add the corresponding heading styles of the airport software operation and maintenance document; a table standardization module, used to standardize the names of the tables on each page of the airport software operation and maintenance document when the document type is a table document, and to extract the header row of the first row of the table on each page to obtain the table header field; and a document conversion module, used to convert the airport software operation and maintenance document into a text document or a table document when the document type is an image document, a scanned document, or a font insertion document.

[0015] Optionally, the preprocessing unit includes: a document title conversion module, used to convert the preprocessed airport software operation and maintenance document into a one-to-one correspondence using a title segmentation tool when the document type is a text document; and a table name conversion module, used to convert the table names of each page of the preprocessed airport software operation and maintenance document into first-level headings using a title segmentation tool when the document type is a table document, and to convert the header row of the first table in each page into a second-level heading.

[0016] Optionally, the document segmentation unit includes: a first segmentation module, used to segment the converted airport software operation and maintenance document according to the title style of each level using a title segmentation tool to obtain initial document fragments corresponding to multi-level titles; a second segmentation module, used to perform secondary segmentation of the initial document fragment using a text segmentation tool when the fragment size of any initial document fragment is detected to be greater than a preset threshold; and a source information attachment module, used to retain the titles of each level of the initial document fragment after the segmentation process, and attach the standard document name of the airport software operation and maintenance document to which each initial document fragment belongs as source information to the text content of the initial document fragment.

[0017] Optionally, the second segmentation module includes: a segmentation tool generation submodule, used to generate a text segmentation tool based on semantic boundaries using a lightweight natural language model for the operation and maintenance domain; a context acquisition submodule, used to acquire the continuous sentence sequence of the context window of the initial document fragment; and a segmentation submodule, used to input the initial document fragment and the continuous sentence sequence of the context window into the text segmentation tool, and use the text segmentation tool to perform secondary segmentation on the initial document fragment.

[0018] Optionally, the vectorization processing unit includes: a vectorization processing module for vectorizing each initial document fragment to obtain an initial vector; a dimensionality reduction processing module for performing vector dimensionality reduction processing on the initial vector to obtain a target document vector; a clustering processing module for performing associative semantic clustering processing on the target document vector using a Gaussian mixture clustering algorithm to obtain a document fragment set; and a reprocessing module for re-vectorizing, vector dimensionality reduction, and associative semantic clustering processing on the document fragments in the document fragment set if the number of document fragments in the document fragment set exceeds a preset threshold, to obtain a new document fragment set.

[0019] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the retrieval method of the airport operation and maintenance intelligent question-and-answer system described above.

[0020] According to another aspect of the present invention, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the retrieval method of the airport operation and maintenance intelligent question-and-answer system described above.

[0021] According to another aspect of the present invention, a computer program product is also provided, including a computer program that, when executed by a processor, implements the steps of the retrieval method of the airport operation and maintenance intelligent question-and-answer system described in any one of the above embodiments.

[0022] In this disclosure, user question dialogues are received, and the question dialogues are semantically rewritten and expanded to obtain a hybrid index containing question keywords and question vectors. Using the hybrid index, a hybrid search is performed in a pre-built vector library to obtain search results containing document metadata. The vector library includes document metadata and index information corresponding to multiple document fragments. The document fragments are generated after segmentation, vectorization, and knowledge processing of airport software operation and maintenance documents. The document fragments are segmented according to a multi-level heading structure, and each document fragment retains its parent heading. The search results are input into a large language model, and the large language model is used to perform question association reasoning to output the question reasoning results.

[0023] In this disclosure, user question dialogues can be semantically rewritten and expanded (converting fragmented queries from operations personnel, such as "system slow") into multiple semantically equivalent expressions, resulting in a hybrid index containing question keywords and question vectors. Then, a pre-built vector library is used for hybrid retrieval. This library includes document metadata and index information corresponding to multiple document fragments. Each document fragment retains its parent title path, completely avoiding contextual breaks caused by fixed-length segmentation. The segmented document fragments are then vectorized, and multiple fragments are clustered. Finally, the retrieval results are input into a large language model, which performs question association reasoning, outputting question reasoning results that cover complete context, are traceable in origin, and are free from misinterpretation. This significantly improves the recall coverage of relevant documents and the completeness and accuracy of answers, greatly enhancing the retrieval recall rate in airport operations scenarios, ensuring answer accuracy, and increasing user trust. This addresses the technical problem in related technologies where retrieval enhancement generation techniques in airport operations scenarios suffer from low retrieval recall, leading to missing, incorrect, or even misleading answers, thus reducing user trust. Attached Figure Description

[0024] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0025] Figure 1 This is a flowchart of an optional airport operation and maintenance intelligent question-and-answer system according to an embodiment of the present invention;

[0026] Figure 2This is a flowchart of an intelligent question-and-answer method for airport operation and maintenance according to an embodiment of the present invention;

[0027] Figure 3 This is a schematic diagram of an optional knowledge processing procedure according to an embodiment of the present invention;

[0028] Figure 4 This is a schematic diagram of an optional index construction according to an embodiment of the present invention;

[0029] Figure 5 This is a schematic diagram of a retrieval device for an optional intelligent question-and-answer system for airport operation and maintenance according to an embodiment of the present invention;

[0030] Figure 6 This is a hardware structure block diagram of an electronic device (or mobile device) for a retrieval method of an airport operation and maintenance intelligent question-and-answer system according to an embodiment of the present invention. Detailed Implementation

[0031] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0032] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0033] To facilitate understanding of the present invention by those skilled in the art, some terms or nouns involved in the various embodiments of the present invention are explained below:

[0034] Retrieval-Augmented Generation (RAG) is a technique that combines information retrieval with generative language models. It retrieves relevant document fragments from external knowledge bases as context to assist large language models in generating accurate and interpretable answers, thereby improving the reliability and professionalism of question-answering systems.

[0035] Recall ratio is the ratio of relevant information retrieved from a (vector) database to the total number of relevant documents. It is an indicator that measures the success of a retrieval system in retrieving relevant documents from a document collection, i.e., the percentage of relevant documents retrieved out of all relevant documents.

[0036] Hybrid retrieval refers to the simultaneous use of sparse representations (such as keyword matching) and dense representations (such as vector semantic similarity) in the information retrieval process, and the results are weighted and fused to balance precise matching and semantic understanding, thereby improving the overall recall and precision.

[0037] Hierarchical retrieval is an information retrieval method used to explore and retrieve large amounts of data organized in a hierarchical structure. This method improves the efficiency and accuracy of retrieval by organizing data into multiple levels. In hierarchical retrieval, each level of the hierarchy contains specific information, allowing users to progressively refine their search and find the data they need more accurately.

[0038] Knowledge enhancement refers to the semantic clustering, induction, and abstraction of original text fragments. Through steps such as vectorization, dimensionality reduction, clustering, and large-scale model generation of summaries, discrete document fragments are transformed into higher-order knowledge units (such as summary descriptions), which makes up for the defects of incomplete original document structure or scattered information, and enhances the completeness of retrieval content and reasoning ability.

[0039] GMM clustering, or Gaussian Mixture Models, is a clustering algorithm based on Gaussian Mixture Models. It is an unsupervised learning method based on probability density functions. It groups data by linear combination of multiple Gaussian distributions, supports soft clustering (probability assignment), and is suitable for scenarios with complex data distributions, such as semantic recognition, speech recognition, and image processing.

[0040] Uniform Manifold Approximation and Projection (UMAP) is a nonlinear dimensionality reduction technique that constructs an adjacency graph of high-dimensional data and optimizes its topology in a low-dimensional space to achieve efficient vector dimensionality reduction that preserves both local and global structure. It is often used for visualization or to provide a more suitable input space for clustering algorithms (such as GMM).

[0041] Sparse representation refers to a vector form of text that uses statistical features such as word frequency and inverse document frequency (e.g., BM25). It has high vector dimension but sparse elements, and is good at capturing precise keyword matching. It performs well in semantically ambiguous or term matching scenarios.

[0042] Dense representation refers to mapping text into low-dimensional continuous vectors (such as 256-dimensional or 768-dimensional vectors) through embedding models (such as Sentence-BERT). It can capture semantic similarity and contextual relationships and is suitable for semantic retrieval and synonym matching.

[0043] The BM25 algorithm, or Best Matching 25, is a text relevance scoring algorithm based on a probabilistic model. It sparsely scores documents by calculating the term frequency of the query term in the document, the normalized document length, and the inverse document frequency weight of the term. It is the most mainstream and stable ranking method in keyword retrieval.

[0044] Hybrid Index is a dual index structure that simultaneously builds keyword indexes (sparse) and vector indexes (dense), supporting multi-way recall and result fusion, and significantly improving retrieval coverage under different expression methods.

[0045] Question rewriting transforms a user's original short query into multiple semantically equivalent expressions through synonym replacement, semantic expansion, or sentence transformation, thereby expanding the search scope and bridging the semantic gap between user expression and document expression.

[0046] Question expansion refers to automatically adding relevant keywords, term variations, or contextual keywords to the original query based on semantic models or knowledge graphs, thereby enhancing the expressive richness of the query and improving the hit rate in sparse retrieval.

[0047] Document chunking refers to the process of dividing a raw document into multiple semantically complete text segments based on semantic boundaries (such as heading levels). It is a fundamental preprocessing step in building a search index and directly affects the accuracy and completeness of search results.

[0048] Large Language Model (LLM) refers to a deep neural network model with powerful language understanding and generation capabilities. It is used to generate natural language responses based on the retrieved context and is the core of the generation mechanism in the RAG system.

[0049] It should be noted that the retrieval method and apparatus of the airport operation and maintenance intelligent question and answer system in this disclosure can be used in the field of airport operation and maintenance technology to improve the retrieval recall rate of the airport operation and maintenance intelligent question and answer system, and can also be used in any field other than the field of airport operation and maintenance technology to improve the retrieval recall rate of the airport operation and maintenance intelligent question and answer system. This disclosure does not limit the application field of the retrieval method and apparatus of the airport operation and maintenance intelligent question and answer system.

[0050] It should be noted that the information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, and displayed data, etc.) collected in this public disclosure are information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data are handled with corresponding access points, allowing users to choose to authorize or refuse. For example, this system has interfaces with relevant users or organizations. Before obtaining relevant information, a request to obtain the information needs to be sent to the aforementioned user or organization through the interface, and the relevant information is obtained only after receiving consent from the aforementioned user or organization.

[0051] It should be noted that in this disclosure, customer information is collected and analyzed, and users are provided with corresponding operation entry points to choose whether to agree to or reject the automated decision results; if the user chooses to reject, the process will proceed to the expert decision-making process.

[0052] The following embodiments of the present invention can be applied to various airport operation and maintenance intelligent question-and-answer systems / applications / devices. The present invention is applicable to implementation scenarios such as smart airport operation and maintenance, and civil aviation information system support, for example, in scenarios such as intelligent question-and-answer systems for professional technicians, fault diagnosis assistants, and operation and maintenance knowledge base retrieval platforms.

[0053] This invention significantly improves the retrieval recall rate of professional domain texts through document segmentation that preserves multi-level heading structures, GMM-based clustering, knowledge processing, and hierarchical hybrid index construction. In scenarios where the airport operation and maintenance knowledge base has more than 100,000 document fragments, the recall rate of relevant documents can be increased to over 85%, which is more than 30% higher than the traditional fixed-block RAG method. It also effectively reduces the problems of out-of-context answers and semantic mismatch, achieving more accurate, complete, and traceable intelligent question-answering output.

[0054] The present invention will now be described in detail with reference to various embodiments.

[0055] Example 1

[0056] According to an embodiment of the present invention, an embodiment of a retrieval method for an intelligent question-and-answer system for airport operation and maintenance is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0057] Figure 1 This is a flowchart of an optional airport operation and maintenance intelligent question-and-answer system retrieval method according to an embodiment of the present invention, such as... Figure 1 As shown, the method includes the following steps:

[0058] Step S101: Receive user question dialogue, and perform question semantic rewriting and question expansion on the user question dialogue to obtain a hybrid index containing question keywords and question vectors.

[0059] It should be noted that in step S101, the system receives a natural language question dialogue input by the user, such as "the check-in system is slow." The system then semantically rewrites the original question using a pre-trained language model, generating multiple semantically equivalent but differently expressed variants, such as "how to handle passenger check-in service delays" and "what to do about check-in terminal lag," forming a set of semantically expanded queries. Further, the system performs keyword extraction and vector embedding on the original question and its expanded forms. Keyword extraction is based on TF-IDF (Term Frequency-Inverse Document Frequency) or the TextRank algorithm (an unsupervised text analysis algorithm based on graph ranking) to extract discriminative terms, such as "check-in system," "slow response," "delay," and "lag," forming a sparse keyword set. Simultaneously, a lightweight embedding model is used to map each query text into a dense vector of fixed dimensions, obtaining a question vector containing semantic information.

[0060] To improve search coverage, the extracted keyword set is combined with the corresponding question vector to construct a hybrid index for each query. This hybrid index contains both sparse keyword fields and dense vector fields, enabling subsequent search stages to trigger a dual-channel search mechanism based on term matching and semantic similarity in parallel.

[0061] Step S102: Using a hybrid index, perform a hybrid search in a pre-built vector library to obtain search results containing document metadata. The vector library includes document metadata and index information corresponding to multiple document fragments. The document fragments are generated after segmenting, vectorizing, and processing knowledge of airport software operation and maintenance documents. The document fragments are segmented according to a multi-level heading structure, and each document fragment retains its parent heading.

[0062] It should be noted that in step S102, the hybrid index is used to perform parallel hybrid retrieval in a pre-built vector library. This vector library contains document metadata and index information corresponding to multiple document fragments. Each document fragment comes from the structured processing of airport software operation and maintenance documents. The document fragments are segmented according to a multi-level heading structure in a predetermined format (e.g., Markdown format). For example, the segmentation boundaries are based on #, ##, ###, and ####, and the smallest segmentation unit is a fourth-level heading. Each fragment can completely retain all its parent heading paths during segmentation to form a hierarchical semantic context. At the same time, the source document name can be appended as metadata to the end of the fragment.

[0063] It should be noted that, in this embodiment, after vectorization, each document fragment not only retains the dense vector representation of its text content, but also simultaneously constructs independent vectors and keyword indexes for its respective headings at all levels. This allows the semantics of the headings to participate in matching as an independent retrieval dimension, avoiding semantic dilution caused by relying solely on the body text of the fragment.

[0064] Furthermore, after the document fragments are vectorized, they undergo a knowledge processing process. The summaries or inductive texts generated by GMM clustering and the large model are injected back into the vector library as enhanced fragments, sharing the same metadata identifier with the original fragments. This forms a dual-channel expression of "original content - extracted summary", enabling hybrid retrieval to simultaneously capture matching signals from the original expression and the semantic abstraction layer.

[0065] Step S103: Input the search results into the large language model, use the large language model to perform question association reasoning, and output the question reasoning results.

[0066] This embodiment can input multiple document fragments and their metadata returned by the mixed retrieval into the large language model. The model performs cross-fragment semantic association reasoning based on the input context fragments, title hierarchy structure, source identifiers and question rewrite versions, identifies the logical relationships, complementary relationships or conflict relationships between fragments, and performs semantic integration of fragment content in combination with title context to generate coherent and well-structured question reasoning results.

[0067] What helps enhance the structural integrity of the answer is that the large language model can identify multiple fragments that point to the same solution during the reasoning process, or infer the system module and processing flow level to which the question belongs through the title level. This naturally presents a four-dimensional relational response structure of "question-module-method-source" in the output, avoiding logical breaks caused by splicing isolated fragments.

[0068] Through the above steps, user question dialogues can be received, and the question dialogues can be semantically rewritten and expanded to obtain a hybrid index containing question keywords and question vectors. Using the hybrid index, a hybrid search is performed in a pre-built vector library to obtain search results containing document metadata. The vector library includes document metadata and index information corresponding to multiple document fragments. The document fragments are generated after fragmentation, vectorization, and knowledge processing of airport software operation and maintenance documents. The document fragments are segmented according to a multi-level heading structure, and each document fragment retains its parent heading. The search results are input into a large language model, which is used to perform question association reasoning and output the question reasoning results. In this embodiment, user question dialogues can be semantically rewritten and expanded (converting fragmented queries from maintenance personnel, such as "system slow") into multiple semantically equivalent expressions, resulting in a hybrid index containing question keywords and question vectors. Then, a pre-built vector library is used for hybrid retrieval. This library includes document metadata and index information corresponding to multiple document fragments. Each document fragment retains its parent title path, completely avoiding contextual breaks caused by fixed-length segmentation. The segmented document fragments are then vectorized, and multiple fragments are clustered. Finally, the retrieval results are input into a large language model, which performs question association reasoning, outputting question reasoning results that cover complete context, are traceable in origin, and are free from misinterpretation. This significantly improves the recall coverage of relevant documents and the completeness and accuracy of answers, greatly enhancing the retrieval recall rate in airport maintenance scenarios, ensuring answer accuracy, and increasing user trust. This addresses the technical problem in related technologies where retrieval enhancement generation techniques in airport maintenance scenarios suffer from low retrieval recall, leading to missing, incorrect, or even misleading answers, thus reducing user trust.

[0069] Optionally, the construction of the vector library includes: acquiring all airport software operation and maintenance documents from the historical process and determining the document type of each document; performing document preprocessing and document type conversion on all airport software operation and maintenance documents based on the document type; segmenting the converted airport software operation and maintenance documents using a preset segmentation tool to obtain multiple initial document fragments, wherein each document fragment retains its parent title and appends the original document name as a source identifier; performing vectorization and Gaussian mixture clustering on the initial document fragments, and sending the clustered document fragment set to a large language model, which generates fragment summaries to complete the knowledge processing operation; constructing a vector index for each document fragment in the associated document fragment set based on the result of vectorization of the initial document fragments; constructing a keyword index based on the fragment summary corresponding to each document fragment; and constructing the vector library based on the fragment summary, vector index, and keyword index of each document fragment in the document fragment set.

[0070] It should be noted that in the process of building the vector library, all historically accumulated airport software operation and maintenance documents are first obtained, and the document type of each document is identified, including text documents (e.g., Word documents), table documents (e.g., Excel documents), image documents (e.g., PDF), font insertion documents (e.g., TXT documents), etc., to provide a type basis for subsequent processing.

[0071] Furthermore, based on document type, preprocessing is performed on all documents, including standardizing naming formats, correcting heading levels, unifying table header structures, and completing missing headings. Optionally, based on document type, preprocessing and document type conversion steps are performed on all airport software operation and maintenance documents, including: when the document type is text-based, standardizing the expression and writing of the document title and subheadings at all levels, and adding corresponding heading styles for each level; when the document type is table-based, standardizing the names of each table on each page, and extracting the header field from the first row of each table; when the document type is image-based, scanned, or font-insertion-based, converting the airport software operation and maintenance documents into text-based or table-based documents.

[0072] It should be noted that during the preprocessing and type conversion of airport software operation and maintenance documents based on document type, when the document type is text-based, the document title and subheadings at all levels are standardized to ensure that the title content accurately reflects its technical content, eliminates ambiguous expressions and non-standard abbreviations, and at the same time, the writing is standardized to ensure that the title is semantically complete, grammatically correct, and free of typos. In addition, the corresponding style level is matched to each level of the title according to the text style system, such as "Heading 1" corresponding to a first-level title, "Heading 2" corresponding to a second-level title, and so on, up to "Heading 4" corresponding to a fourth-level title, so that the title structure is strictly aligned with the semantic level of the document.

[0073] Optionally, based on the document type, the steps of preprocessing and converting all airport software operation and maintenance documents include: when the document type is text, using a title segmentation tool to convert the document titles of the preprocessed airport software operation and maintenance documents one-to-one; when the document type is table, using a title segmentation tool to convert the table names of each page of the preprocessed airport software operation and maintenance documents into first-level headings, and converting the header row of the first table in each page into a second-level heading.

[0074] It should be noted that during the preprocessing and type conversion of airport software operation and maintenance documents based on document type, when the document type is text document, a heading segmentation tool is used to perform hierarchical heading mapping on the standardized document. For example, the style content marked as "Heading 1", "Heading 2", "Heading 3", and "Heading 4" in the text document Word is converted one by one into the "#", "##", "###", and "####" level identifiers in Markdown format. This ensures that the semantic level of the original document is precisely aligned with the predetermined heading structure (such as Markdown heading structure), without omissions, skipping levels, or misjudgments. In Markdown, the heading structure uses the hash symbol (#) as the heading identifier. The hierarchy is determined by the number of hash symbols, forming a clear nested relationship from level one to four: Level one headings begin with a single hash symbol (#heading content), representing the highest-level main heading in the document; Level two headings begin with two hash symbols (##heading content), used to divide main chapters or modules; Level three headings begin with three hash symbols (###heading content), used to describe sub-functions or subsystems; and Level four headings begin with four hash symbols (####heading content), serving as the finest-grained operational unit or specific scenario description. Each heading level is on its own line, with spaces allowed before and after, but cannot be nested within paragraph text. This structure can form a tree-like hierarchy, facilitating tool parsing and semantic segmentation. This structure does not rely on styles, fonts, or formatting information; it clearly expresses the logical hierarchy of text solely through character sequences, providing machine-recognizable semantic boundaries for document segmentation.

[0075] Furthermore, when the document type is a table, each worksheet undergoes name standardization. For example, for Excel spreadsheets, the original sheet names are uniformly adjusted to reflect the content theme. Simultaneously, the first row of each table is extracted as the header row. Regardless of whether the row contains merged cells, its actual semantic boundaries are parsed to form a clearly structured set of header fields. Additionally, when the document type is an image, scanned, or text-embedded document, optical character recognition and layout analysis tools can be used to extract text from the image content. Combined with table recognition algorithms, the table structure in the image is restored, and the extracted results are reconstructed into a text-based or table-based document format. This ensures that unstructured content is converted into a standardized text carrier that can be recognized by the segmentation tool, avoiding semantic loss due to image embedding. This ensures that the original maintenance documents have a clear structure, standardized titles, and a unified format before entering the segmentation process.

[0076] Furthermore, when the document type is a table, this embodiment uses a heading segmentation tool to perform a structured transformation on each page of the table. For example, in the case of an Excel spreadsheet, the name of the sheet containing the table is used as the top-level semantic unit of the page content and converted into a first-level heading (#) in Markdown format, so that the module or functional area to which the table belongs is explicitly identified in the document structure. At the same time, the content of the header field of the first row of the table, regardless of whether it contains merged cells, is parsed into a semantically complete set of phrases and embedded as a second-level heading (##) in the converted Markdown document, forming a two-layer structure of "module-sub-function".

[0077] It should be noted that the heading segmentation tool in this embodiment does not rely on visual layout or font style during conversion. Instead, it identifies heading levels based on the document object model or structured metadata. This ensures that even in scenarios with many table pages and inconsistent heading names, it can still output structurally consistent Markdown documents, avoiding hierarchical errors caused by manual intervention. This ensures that maintenance documents from different sources are uniformly converted into a Markdown structure with a clear semantic hierarchy after type conversion, providing a structurally parsable and semantically inheritable input foundation for subsequent segmentation based on heading levels.

[0078] Optionally, the step of using a preset segmentation tool to segment the converted airport software operation and maintenance document to obtain multiple initial document fragments includes: using a title segmentation tool to segment the converted airport software operation and maintenance document according to the title styles of each level to obtain initial document fragments corresponding to multiple levels of titles; if the size of any initial document fragment is detected to be greater than a preset threshold, using a text segmentation tool to perform secondary segmentation on the initial document fragment; after the segmentation process, retaining the titles of each level of the initial document fragment, and attaching the standard document name of the airport software operation and maintenance document to which each initial document fragment belongs as source information to the text content of the initial document fragment.

[0079] It should be noted that when this embodiment uses a preset segmentation tool to segment the converted airport software operation and maintenance document, it uses the defined title level symbols "#", "##", "###", and "####" as semantic boundaries to segment the document content level by level. Each segmentation unit corresponds to all subsequent content under a title level, forming an initial document fragment with a fourth-level title as the smallest semantic unit. This ensures that each fragment retains its complete title path from the first level to the current level in terms of structure, and constructs a four-level semantic nesting relationship of "system-module-function-operation".

[0080] Furthermore, when the text length of any initial document fragment exceeds a preset character threshold, the text segmentation tool is activated to perform secondary segmentation on that fragment. Secondary segmentation is based on a fixed character count window or semantic breakpoints, but the segmentation operation does not disrupt the title hierarchy. All generated sub-fragments must fully inherit all parent titles of their parent fragments, ensuring that each sub-fragment can still trace back to the complete semantic path of the original document, avoiding contextual breaks due to excessive content length. After completing all segmentation operations, the standardized name of its original airport software maintenance document (e.g., "XXX Software Maintenance Manual-V1.0-20250302.md") is appended to the end of each initial document fragment, clearly marked with the fixed prefix identifier "[Source]", creating a one-to-one mapping between the fragment content and its original document source, facilitating source verification and answer tracing for subsequent search results. This ensures that each document fragment possesses complete hierarchical inheritance in semantic structure, meets processing constraints in content size, and maintains clear identification in source attribution, providing the smallest semantic unit with structural integrity and semantic traceability for hybrid retrieval.

[0081] Optionally, the step of using a text segmentation tool to perform secondary segmentation on the initial document fragment includes: generating a semantic boundary-based text segmentation tool using a lightweight natural language model for the operation and maintenance domain; obtaining a continuous sentence sequence of the context window of the initial document fragment; inputting the initial document fragment and the continuous sentence sequence of the context window into the text segmentation tool, and using the text segmentation tool to perform secondary segmentation on the initial document fragment.

[0082] It should be noted that in the step of using a text segmentation tool to perform secondary segmentation on the initial document fragments, a lightweight natural language model for the airport operation and maintenance field can be pre-built. This model learns common semantic breakpoint patterns in operation and maintenance scenarios by fine-tuning professional corpora such as airport operation and maintenance manuals, fault reports, and operation logs. These breakpoints include boundary features such as the end of operation steps, switching of abnormal conditions, parameter configuration changes, and termination of system response descriptions. This results in a fine-grained segmentation capability suitable for operation and maintenance text structures.

[0083] Furthermore, the context window of the initial document fragment is obtained. This window contains a continuous sentence sequence consisting of several sentences preceding and following the current fragment. The window length is set to a fixed number of sentences based on the language habits of the operation and maintenance documents, such as three sentences before and after, to maintain local semantic coherence and avoid disrupting the action chain or causal logic due to isolated segmentation. In addition, the initial document fragment and the extracted context window can be input together into the text segmentation tool, so that the model can make segmentation decisions not only based on the semantics within the current fragment, but also combined with the surrounding context to determine whether there is a natural semantic termination point. For example, "If the above method is ineffective, then execute the following emergency process" is a reasonable segmentation boundary, and "Observe the system status after restarting the service" is a complete semantic unit that can be used as a segmentation position.

[0084] By using secondary segmentation, instead of relying on fixed word counts or punctuation marks, the segmentation points are dynamically identified by combining semantic patterns in the operations and maintenance domain with local context. This preserves the integrity of the operation process, the logical closed loop of the fault description, and the coherence of the processing steps, thereby improving the semantic usability of the segmented fragments.

[0085] Optionally, the steps of vectorizing and performing Gaussian mixture clustering on the initial document fragments include: vectorizing each initial document fragment to obtain an initial vector; performing vector dimensionality reduction on the initial vectors to obtain a target document vector; performing association semantic clustering on the target document vectors using a Gaussian mixture clustering algorithm to obtain a set of document fragments; and if the number of document fragments in the set of document fragments exceeds a preset threshold, re-vectorizing, performing vector dimensionality reduction, and performing association semantic clustering on the document fragments in the set of document fragments to obtain a new set of document fragments.

[0086] It should be noted that when vectorizing the initial document fragments, a pre-trained domain-adaptive embedding model is used to map the text content of each initial document fragment into a dense vector representation of a fixed dimension. This vector captures the semantic associations between terms, operations, fault phenomena and system components in the operation and maintenance scenario, forming the initial vector and preserving the semantic distribution characteristics of the original text.

[0087] Furthermore, vector dimensionality reduction is performed on the initial vectors using the Uniform Manifold Approximation and Projection (UMAP) algorithm. This algorithm projects vectors from the high-dimensional embedding space to the low-dimensional space, reducing computational complexity while preserving local neighborhood structure. This generates target document vectors, allowing semantically similar fragments to exhibit a closer clustering trend in the low-dimensional space. Further, this embodiment can employ Gaussian Mixture Clustering (GMM) to perform unsupervised clustering of the target document vectors. This algorithm models data distribution based on a linear combination of multiple Gaussian distributions, allowing individual document fragments to be probabilistically assigned to multiple clusters. It supports soft clustering assignment and is suitable for scenarios with semantic overlap and ambiguous boundaries in operational knowledge, such as overlapping areas of "slow response" and "high latency" in the semantic space, which can be encompassed by the same cluster.

[0088] If the number of document fragments in the document fragment set exceeds a preset threshold, vectorization, dimensionality reduction, and Gaussian mixture clustering must be performed on all document fragments in the set to form a new document fragment set. This process can be iterated until the clustering granularity reaches the inductive range of semantic summarization, avoiding semantic confusion caused by excessively large clustering in a single operation. This automatically aggregates semantically similar but dispersed operational knowledge fragments into semantic clusters with inherent logical connections, providing an input basis for generating structured summaries for subsequent large models and improving the coverage completeness and expression consistency of knowledge abstraction.

[0089] The following describes in detail another optional implementation method.

[0090] In existing technologies, airport software operation and maintenance issues have complex contextual backgrounds, including system environment, configuration parameters, and dependencies. Existing RAG systems typically employ a fixed-size text chunking strategy, which often disrupts the contextual integrity of the problem during the segmentation process. For example, a description of a "departure system database deadlock" problem might be split into multiple discontinuous text chunks, preventing the system from retrieving a complete solution; or a search for "emergency procedures for the XXX system" might only retrieve partial fragments, or even fragments of emergency procedures from another software, leading to completely incorrect answers. Furthermore, operations personnel often use concise, fragmented language to describe problems (e.g., "slow check-in system"), while knowledge base documents might describe similar problems in a formal and complete manner (e.g., "handling methods for passenger check-in system response times exceeding thresholds"). This mismatch in expression results in a large number of relevant documents being unretrievable.

[0091] In this embodiment of the invention, a method for improving retrieval recall is proposed. This method fully preserves the contextual semantics when segmenting documents, and at the same time adopts knowledge processing and hierarchical retrieval methods to avoid one-sided answers.

[0092] Figure 2 This is a flowchart of an intelligent question-and-answer method for airport operation and maintenance according to an embodiment of the present invention, such as... Figure 2 As shown, the main components include a data preprocessing flow and a question-and-answer interaction flow. The data preprocessing flow includes steps such as document preprocessing, document type conversion, image storage, document segmentation, image content recognition and extraction, knowledge processing, and index construction. The question-and-answer interaction (input / output) flow, i.e., the interaction flow between the user and the intelligent question-and-answer system, includes steps such as user input, security detection, question rewriting, question expansion, hybrid retrieval, reordering, reflection, and response synthesis. Among these, document preprocessing, document type conversion, document segmentation, knowledge processing, index construction, and hybrid retrieval are the main improvement schemes for improving retrieval recall involved in this invention, and the following content revolves around these improvement schemes.

[0093] Part 1, Document Preprocessing. The quality of document data largely determines the accuracy of the answer. This step mainly aims to standardize the content format of the input documents, so that different types of documents can be uniformly converted into document types that can accurately preserve the contextual semantics during the document type conversion stage. This is mainly done through manual review.

[0094] Airport software maintenance involves document types such as Word and Excel. Strictly standardize file naming (e.g., "name-version number-timestamp" format). For example, for Word files, ensure the file title and subheadings are correctly expressed and written, and that corresponding heading styles have been added, ensuring all content has a parent heading. For Excel files, ensure sheet page names are accurate and standardized, the first row of the table (including merged cells) is the header row, and header fields are accurately defined. For other types of documents, tools or manual conversion to Word or Excel files can be used for further processing.

[0095] Part 2, Document Type Conversion. The document segmentation stage requires selecting a specific segmentation strategy. If fixed-length segmentation or segmentation based on certain special delimiters is chosen, even though many segmentation tools include parameters to control the size of the overlap between adjacent text fragments, the relationships between multiple document fragments will still be lost, making the fragments relatively isolated and hindering overall semantic retrieval. To preserve contextual semantics as much as possible, document fragments should retain headings at all levels. Markdown format documents naturally support heading parsing at all levels. Combined with dedicated Markdown segmentation tools, documents can be quickly segmented, ensuring that each document fragment retains its semantics relatively completely.

[0096] Convert the standardized files from step 1 to Markdown format. Various open-source tools can be used to achieve this. For Word documents, ensure that all heading levels strictly correspond to Markdown document headings; for Excel documents, ensure that sheet page names are converted to first-level headings and heading rows are converted to second-level headings.

[0097] Part 3, Document Segmentation. Use an open-source or self-developed Markdown segmentation tool in conjunction with a regular text segmentation tool to segment the document. During segmentation, you can specify segmentation according to Markdown heading styles such as "#", "##", "###", "####", etc. The smallest segmentation unit is determined based on the actual document structure; most documents use third- or fourth-level headings as the smallest segmentation unit. After segmentation, if the document fragments are still large, perform a second segmentation using a regular text segmentation tool, ensuring that all headings are retained in the second segmentation. Simultaneously, append the document name specified in Step 1 as source information to each segmented text content with a special identifier to facilitate understanding by the larger model.

[0098] Part 4, Knowledge Processing. In some cases, the answer to a user's question is located in a wide range of document fragments, or the answer is not in a specific document fragment. It requires summarizing and generalizing to arrive at the answer. Especially when the document itself does not provide an overview or summary, it can easily lead to incomplete or incorrect document fragments being retrieved.

[0099] Knowledge processing techniques can effectively compensate for this deficiency. The method involves first vectorizing the original text fragments, then reducing the dimensionality of the high-dimensional vectors before using the Generative Model (GMM) clustering algorithm for semantic clustering. The clustered fragment set is then sent to a larger model to generate an overview or summary. If the number of summarized fragments after clustering is still large, a second clustering can be performed, repeating the above process to generate the final overview or summary. Generally, for clustering summaries exceeding three iterations, the second-to-last and last summaries are taken, marked with processing identifiers, and added as part of the document fragment set to continue the subsequent process.

[0100] Figure 3 This is a schematic diagram of an optional knowledge processing procedure according to an embodiment of the present invention, such as... Figure 3 As shown, after obtaining the segmented text, vectorization, vector dimensionality reduction, GMM clustering, and large model summarization can be performed to obtain overview / summary text fragments. Then, it is determined whether the fragment is greater than 1. If not, a set of text fragments is generated.

[0101] Part 5, Index Construction. Generally, constructing a hybrid index of "keywords (sparse representation) + vectors (dense representation)" improves recall. The calculation method for sparse representation (using the BM25 algorithm as an example) is as follows:

[0102]

[0103] in, Q Indicates Query, express Q A morpheme after parsing; d This represents a search result document; Indicates morpheme The weights; R ( , d ) indicates a morpheme With Documents d The relevance score.

[0104] The calculation method for dense representations (usually vector cosine similarity) is as follows:

[0105]

[0106] in A , B Represent two n-dimensional vectors:

[0107]

[0108]

[0109] Based on the documents segmented in step 3, recall can be further improved in this hybrid retrieval mode. However, performance remains poor when fragments are too large and user questions are short, or when the query expression and document expression are inconsistent. The following three methods can be used to avoid this situation: 1) Construct separate "keyword + vector" indexes for each level of heading extracted in step 3. Merge these indexes after separate retrieval of the hybrid retrieval stage and document fragment content. This can greatly increase the recall rate of query questions based on summaries of headings or segmented content. 2) Use a large model to extract keywords and phrases from each document fragment, construct separate "keyword + vector" indexes, and merge them after separate retrieval of the hybrid retrieval stage and document fragment content. 3) Construct multi-level indexes to achieve hierarchical retrieval. Borrowing from the knowledge processing clustering method in step 4, extract a set of keywords and their synonyms, summaries, or more colloquial aliases from a group of semantically clustered document fragments. Establish a mapping relationship between keywords, summaries, and this group of fragments. Simultaneously, construct separate keyword indexes for keywords and vector indexes for summaries. Generally, a two-level index is sufficient. During the retrieval phase, priority is given to searching for "keywords + abstract" content. Based on the top N matched abstracts, the corresponding document fragments are found and then a secondary mixed search is performed.

[0110] Figure 4 This is a schematic diagram of an optional index construction according to an embodiment of the present invention, such as... Figure 4As shown, during the index building process, keyword indexes and vector indexes can be constructed for related document fragments, headings at all levels, and keyword phrases, enabling hybrid retrieval.

[0111] Part 6, Hybrid Retrieval. Combine the question expansion strategy (i.e., question synonym rewriting) and use the index built in step 5 for retrieval. Generally, methods 1) and 2) do not need to be used simultaneously. Choose one of them according to the actual effect. 3) can be combined with methods 1) or 2) to retrieve separately and then merged. Subsequently, a re-sorting model is used to uniformly sort and filter.

[0112] The calculation formula for hybrid retrieval can adopt a weighted fusion method, combining the scores of keyword retrieval and vector retrieval. The specific calculation formula is as follows: S = α × S_sparse +(1- α )× S_dense ;

[0113] in: α For weight parameters (0≤ α ≤1), used to balance the contributions of the two retrieval techniques; S_sparse Scoring for sparse models; S_dense Scoring for dense models.

[0114] In actual use in airport operation and maintenance scenarios, α It can be dynamically adjusted. Experimental testing has shown that when the question contains a large number of technical terms related to airport software (factual question), it will... α Setting the parameter to around 0.7 yields the best search results; for general questions (semantic questions), [the parameter should be set to 0.7]. α Setting the parameter to around 0.2 yields the best search results.

[0115] The following is an explanation of another implementation method (the explanation of steps that do not involve optimization and improvement in the process is omitted). It is assumed that operation and maintenance documents for multiple software deployed at the airport have been collected, and the operation and maintenance personnel can get the answers to their queries based on the contents of these operation and maintenance documents. The RAG system needs to use these documents as the search objects.

[0116] Step 1. Document Preprocessing. All documents are manually reviewed and modified to standardize their format. The following example uses the Word document "XXX Operation and Maintenance Manual V1-2025-3-2.docx".

[0117] Step 1.1 Standardize the naming format. Rename the document to "XXX Software Operation and Maintenance Manual-V1.0-20250302.docx".

[0118] Step 1.2 Standardize title content. Review the main title and all levels of headings in the document to ensure that the titles accurately summarize the content, without ambiguity or typos. For example, the subheading "6.1 Special Process Handling" includes not only instructions for handling special processes but also instructions for handling other abnormal situations, so it should be changed to "6.1 Special Processes and Abnormal Handling"; the appendix section does not have any headings, so an "Appendix" heading should be added.

[0119] Step 1.3 Standardize heading format. Review the heading format and style of each level in the document to ensure they correspond to the Word style headings. The correspondence is shown in Table 1 below.

[0120] Table 1. Standard Title Format Instructions

[0121]

[0122] Step 2. Document Type Conversion. Convert the Word document to Markdown format. Use the open-source Word SDK library to read the file's heading styles, text paragraphs, tables, and images (images are stored separately and will not be detailed here). Then, organize the content according to the Markdown format and convert it. Note that the heading formats at each level must strictly correspond to the Markdown document heading formats, as shown in Table 2 below.

[0123] Table 2. Document Type Conversion Diagram

[0124]

[0125] In addition, ordered and unordered lists (including nested lists) in Word documents need to correspond to the list format in Markdown to ensure that the large model understands its structural semantics correctly.

[0126] Step 3. Document Segmentation. Use open-source or self-developed Markdown segmentation software to segment the document, specifying segmentation according to Markdown levels one to four, with the smallest segmentation unit being level four headings. After segmentation, evaluate each segment. If the document segment is still large, perform a second segmentation using a regular text segmentation tool, ensuring that all heading levels are retained in the second segmentation. Simultaneously, append the standardized document name from Step 1 as source information to each segmented text content with a special identifier to facilitate understanding by the larger model.

[0127] Step 4. Knowledge Processing. The segmented document fragments are vectorized using an embedding model. Then, dimensionality reduction techniques are used to process the high-dimensional vectors. Next, the GMM clustering algorithm is used to cluster multiple document fragments with high semantic relevance. Finally, the clustered fragment set is sent to a larger model to generate an overview or summary. This process is repeated until the final overview or summary is generated. The penultimate and final overviews or summaries are taken and added to the document fragment set as part of the fragments, while simultaneously assigning a "knowledge processing" label to these fragments.

[0128] Step 5. Index building.

[0129] Step 5.1 Vectorize all document fragments and construct a "keyword + vector" index, where the keyword index uses a sparse model and the vector index uses a dense (embedding) model.

[0130] Step 5.2 Vectorize the headings at each level extracted in Step 3 as text content separately, and construct a "keyword + vector" index, which is then merged and stored with the index constructed in Step 5.1.

[0131] Step 5.3 uses the method in Step 4 (this step can be performed simultaneously with Step 4 to improve construction efficiency) to extract a set of keywords and their synonyms, summaries and their aliases from a set of semantically similar document fragments. At the same time, a "keyword + summary" hybrid index is constructed separately for this set of extracted content, so that the index metadata of the original document fragments contains the identifier of this secondary index (e.g., the unique key of the secondary index), thus realizing the mapping between the primary and secondary indexes.

[0132] Step 6. Hybrid Search. When maintenance personnel inquire about "how to handle slow XXX software," the question is rewritten into two semantically similar questions using a problem expansion strategy, such as "how to handle slow XXX interface response" and "how to resolve XXX software lag." The index built in Step 5 is then used to search for both the original and rewritten questions. First, a search is performed using the document fragment index and the various levels of heading indexes. Simultaneously, a secondary index search is conducted to obtain the matched secondary index identifiers. These identifiers are then filtered using metadata to obtain the mapped primary index. Based on this, a second "document fragment + various levels of heading" index search is performed. During the search process, a finely tuned lightweight large language model is used to determine that the query type is semantic, and the weight parameter α of the hybrid search is dynamically adjusted to 0.2. The fragments retrieved using the above methods are merged, and a re-ranking model is used for unified sorting and filtering.

[0133] Based on the format and type of airport software operation and maintenance documents, this invention provides an optimized solution for intelligent question-answering retrieval and recall suitable for airport operation and maintenance scenarios. It can effectively improve the retrieval and recall rate of relevant document fragments, thereby improving the accuracy and quality of large model answers. It is suitable for semantic and scope-based question scenarios, and performs particularly well in knowledge base retrieval with large document size and complex content.

[0134] This embodiment employs standardized heading structures and formats during document preprocessing, enabling subsequent segmentation to accurately divide the document based on Markdown hierarchical semantics. This avoids contextual breaks caused by content truncation and improves the semantic integrity of the fragments. Furthermore, by binding document segmentation to heading levels and preserving complete heading paths in secondary segmentation, each document fragment possesses a traceable semantic context. This helps the retrieval system identify structured semantic clues when matching user queries, enhancing semantic alignment capabilities.

[0135] This embodiment employs a lightweight natural language model geared towards the operations and maintenance (O&M) domain to generate semantic boundary segmentation points. This allows the secondary segmentation process to adapt to the natural semantic breakpoints of operational procedures, fault descriptions, and system responses within O&M texts, improving the semantic coherence within fragments. In the knowledge processing stage, Gaussian mixture clustering is introduced after vectorization and UMAP dimensionality reduction to automatically aggregate semantically similar but dispersed O&M knowledge fragments into semantic clusters. This provides structured input for the large model to generate summaries, improving the breadth of knowledge abstraction and the consistency of expression. Optionally, by constructing a multi-level hybrid index, independent index layers are established for the original document fragments, titles at each level, keywords, and cluster summaries. This allows the retrieval process to simultaneously initiate queries from three dimensions: semantic granularity, structural identifiers, and summary guidance, enhancing the responsiveness to short, fragmented queries.

[0136] This embodiment introduces a two-level index mapping mechanism to achieve a hierarchical retrieval path of "keywords / summary → document cluster → original fragments", enabling the retrieval system to prioritize the semantic summary layer and then selectively recall relevant fragments, thereby reducing noise interference in the retrieval of massive documents.

[0137] The following is a detailed description with reference to another embodiment.

[0138] Example 2

[0139] The retrieval device of the airport operation and maintenance intelligent question and answer system provided in this embodiment includes multiple implementation units, each of which corresponds to the implementation steps in the above embodiment one.

[0140] Figure 5 This is a schematic diagram of a retrieval device for an optional intelligent question-and-answer system for airport operation and maintenance according to an embodiment of the present invention, such as... Figure 5As shown, the retrieval device of the airport operation and maintenance intelligent question and answer system may include: a semantic rewriting unit 51, a hybrid retrieval unit 52, and a question reasoning unit 53.

[0141] The semantic rewriting unit 51 is used to receive user question dialogues and perform question semantic rewriting and question expansion on the user question dialogues to obtain a hybrid index containing question keywords and question vectors.

[0142] The hybrid retrieval unit 52 is used to perform hybrid retrieval in a pre-built vector library using a hybrid index to obtain retrieval results containing document metadata. The vector library includes document metadata and index information corresponding to multiple document fragments. The document fragments are generated after segmenting, vectorizing, and processing knowledge of airport software operation and maintenance documents. The document fragments are segmented according to a multi-level heading structure, and each document fragment retains its parent heading.

[0143] The question reasoning unit 53 is used to input the retrieval results into the large language model, use the large language model to perform question association reasoning, and output the question reasoning results.

[0144] The retrieval device of the aforementioned airport operation and maintenance intelligent question-and-answer system can receive user question dialogues through the semantic rewriting unit 51, and perform question semantic rewriting and question expansion on the user question dialogues to obtain a hybrid index containing question keywords and question vectors. Through the hybrid retrieval unit 52, the hybrid index is used to perform hybrid retrieval in a pre-built vector library to obtain retrieval results containing document metadata. The vector library includes document metadata and index information corresponding to multiple document fragments. The document fragments are generated after segmenting, vectorizing, and knowledge processing of airport software operation and maintenance documents. The document fragments are segmented according to a multi-level heading structure, and each document fragment retains its parent heading. The retrieval results are input into the large language model through the question reasoning unit 53, and the large language model is used to perform question association reasoning to output the question reasoning results. In this embodiment, user question dialogues can be semantically rewritten and expanded (converting fragmented queries from operations personnel, such as "system slow") into multiple semantically equivalent expressions, resulting in a hybrid index containing question keywords and question vectors. Then, a pre-built vector library is used for hybrid retrieval. This library includes document metadata and index information corresponding to multiple document fragments. Each document fragment retains its parent title path, completely avoiding contextual breaks caused by fixed-length segmentation. The segmented document fragments are then vectorized, and multiple fragments are clustered. Finally, the retrieval results are input into a large language model, which performs question association reasoning, outputting question reasoning results that cover complete context, are traceable in origin, and are free from misinterpretation. This significantly improves the recall coverage of relevant documents and the completeness and accuracy of answers. This greatly enhances the retrieval recall rate in airport operations scenarios, ensures answer accuracy, and increases user trust. This addresses the technical problem in related technologies where retrieval enhancement generation techniques in airport operations scenarios suffer from low retrieval recall, leading to missing, incorrect, or even misleading answers, thus reducing user trust.

[0145] Optionally, when constructing the vector library, the retrieval device of the airport operation and maintenance intelligent question-and-answer system includes: a software operation and maintenance document acquisition unit, used to acquire all airport software operation and maintenance documents in the historical process and determine the document data type of each airport software operation and maintenance document; a preprocessing unit, used to perform document preprocessing and document type conversion on all airport software operation and maintenance documents based on the document data type; and a document segmentation unit, used to segment the converted airport software operation and maintenance documents using a preset segmentation tool to obtain multiple initial document fragments, wherein each document fragment retains its parent title and appends the original document name as a source identifier; and directs... The quantization processing unit performs vectorization and Gaussian mixture clustering on the initial document fragments, and sends the clustered document fragment set to the large language model, which generates fragment summaries to complete the knowledge processing operation. The vector index construction unit constructs a vector index for each document fragment in the associated document fragment set based on the vectorization results of the initial document fragments. The keyword index construction unit constructs a keyword index based on the fragment summary corresponding to each document fragment. The vector library construction unit constructs a vector library based on the fragment summary, vector index, and keyword index of each document fragment in the document fragment set.

[0146] Optionally, the preprocessing unit includes: a text standardization module, used to standardize the expression and writing of the file title and subheadings of the airport software operation and maintenance document when the document type is text, and to add the corresponding heading styles for each level of the airport software operation and maintenance document; a table standardization module, used to standardize the names of the tables on each page of the airport software operation and maintenance document when the document type is table, and to extract the header row from the first row of the table on each page to obtain the table header field; and a document conversion module, used to convert the airport software operation and maintenance document into a text document or a table document when the document type is image document, scanned document, or font insertion document.

[0147] Optionally, the preprocessing unit includes: a document title conversion module, used to convert the document titles of the preprocessed airport software operation and maintenance document to correspond one-to-one using a title segmentation tool when the document type is a text document; and a table name conversion module, used to convert the table names of each page of the preprocessed airport software operation and maintenance document into first-level headings and convert the header row of the first table in each page into second-level headings when the document type is a table document.

[0148] Optionally, the document segmentation unit includes: a first segmentation module, used to segment the converted airport software operation and maintenance document according to the title style of each level using a title segmentation tool to obtain initial document fragments corresponding to multi-level titles; a second segmentation module, used to perform secondary segmentation of the initial document fragment using a text segmentation tool when the fragment size of any initial document fragment is detected to be greater than a preset threshold; and a source information attachment module, used to retain the titles of each level of the initial document fragment after the segmentation process, and attach the standard document name of the airport software operation and maintenance document to which each initial document fragment belongs as source information to the text content of the initial document fragment.

[0149] Optionally, the second segmentation module includes: a segmentation tool generation submodule, used to generate a text segmentation tool based on semantic boundaries using a lightweight natural language model for the operation and maintenance domain; a context acquisition submodule, used to acquire the continuous sentence sequence of the context window of the initial document fragment; and a segmentation submodule, used to input the initial document fragment and the continuous sentence sequence of the context window into the text segmentation tool, and use the text segmentation tool to perform secondary segmentation on the initial document fragment.

[0150] Optionally, the vectorization processing unit includes: a vectorization processing module for vectorizing each initial document fragment to obtain an initial vector; a dimensionality reduction processing module for performing vector dimensionality reduction processing on the initial vector to obtain a target document vector; a clustering processing module for performing association semantic clustering processing on the target document vector using a Gaussian mixture clustering algorithm to obtain a document fragment set; and a reprocessing module for re-vectorizing, vector dimensionality reduction, and association semantic clustering processing on the document fragments in the document fragment set if the number of document fragments in the document fragment set exceeds a preset threshold, to obtain a new document fragment set.

[0151] The retrieval device of the aforementioned airport operation and maintenance intelligent question-and-answer system may also include a processor and a memory. The aforementioned semantic rewriting unit 51, hybrid retrieval unit 52, question reasoning unit 53, etc., are all stored in the memory as program units, and the processor executes the aforementioned program units stored in the memory to realize the corresponding functions.

[0152] The aforementioned processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured, and the comparison results can be sent to the target terminal by adjusting the kernel parameters.

[0153] The aforementioned memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0154] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored computer program, wherein, when the computer program is running, it controls the device where the computer-readable storage medium is located to execute the retrieval method of the airport operation and maintenance intelligent question-and-answer system of any one of the above embodiments.

[0155] According to another aspect of the present invention, an electronic device is also provided, including one or more processors and a memory, wherein the memory is used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the retrieval method of the airport operation and maintenance intelligent question-and-answer system of any one of the above embodiments.

[0156] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the retrieval method of the airport operation and maintenance intelligent question-and-answer system described in various embodiments of this application.

[0157] This application also provides a computer program product, including a non-volatile computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the retrieval method of the airport operation and maintenance intelligent question-and-answer system described in various embodiments of this application.

[0158] Figure 6 This is a hardware structure block diagram of an electronic device (or mobile device) for a retrieval method in an intelligent question-and-answer system for airport operation and maintenance, according to an embodiment of the present invention. Figure 6 As shown, an electronic device may include one or more ( Figure 6 The processor 602 (which may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.) and a memory 604 for storing data may also be included. In addition, it may include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the I / O interface), a network interface, a keyboard, a power supply, and / or a camera. Those skilled in the art will understand that... Figure 6 The structure shown is for illustrative purposes only and does not limit the structure of the electronic device described above. For example, the electronic device may also include components that are more... Figure 6 The more or fewer components shown, or having the same Figure 6 The different configurations shown.

[0159] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0160] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0161] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0162] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0163] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0164] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0165] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A retrieval method for an intelligent question-and-answer system for airport operation and maintenance, characterized in that, include: Receive user question dialogues, and perform question semantic rewriting and question expansion on the user question dialogues to obtain a hybrid index containing question keywords and question vectors; Using the hybrid index, a hybrid search is performed in a pre-built vector library to obtain search results containing document metadata. The vector library includes document metadata and index information corresponding to multiple document fragments. The document fragments are generated after segmenting, vectorizing, and processing knowledge of airport software operation and maintenance documents. The document fragments are segmented according to a multi-level heading structure, and each document fragment retains its parent heading. The search results are input into a large language model, which is then used to perform question association reasoning and output the question reasoning results.

2. The retrieval method according to claim 1, characterized in that, The construction of the vector library includes: Obtain all airport software operation and maintenance documents from the historical process, and determine the document type for each of the airport software operation and maintenance documents; Based on the document data type, all the airport software operation and maintenance documents are preprocessed and converted in document type. The converted airport software operation and maintenance document is segmented using a preset segmentation tool to obtain multiple initial document fragments. Each document fragment retains its parent title and is appended with the original document name as a source identifier. The initial document fragments are vectorized and subjected to Gaussian mixture clustering. The clustered document fragment set is then sent to a large language model, which generates fragment summaries to complete the knowledge processing operation. Based on the result of vectorization processing of the initial document fragments, a vector index is constructed that associates each document fragment in the document fragment set; A keyword index is built based on the segment summary corresponding to each document segment; The vector library is constructed based on the fragment summary, vector index, and keyword index of each document fragment in the document fragment set.

3. The retrieval method according to claim 1, characterized in that, Based on the document data type, the steps for preprocessing and converting all the airport software operation and maintenance documents include: When the document type is a text document, the file title and subheadings of the airport software operation and maintenance document are standardized in expression and writing, and the corresponding heading styles of the airport software operation and maintenance document are added. When the document type is a table, the names of each table on each page of the airport software operation and maintenance document are standardized, and the header row of the first row of each table is extracted to obtain the table header field. If the document type is an image document, a scanned document, or a document with font insertion, the airport software maintenance document will be converted into a text document or a table document.

4. The retrieval method according to claim 3, characterized in that, Based on the document data type, the steps for preprocessing and converting all the airport software operation and maintenance documents include: When the document type is a text document, a title segmentation tool is used to convert the preprocessed airport software operation and maintenance document title to correspondence. When the document type is a table, a title splitting tool is used to convert the table names of each page of the airport software operation and maintenance document into first-level titles, and the title row of the first table on each page into a second-level title.

5. The retrieval method according to claim 2, characterized in that, The steps of segmenting the converted airport software operation and maintenance document using a preset segmentation tool to obtain multiple initial document fragments include: The converted airport software operation and maintenance document was segmented according to the heading styles of each level using a heading segmentation tool to obtain initial document fragments corresponding to multi-level headings; If the size of any of the initial document fragments is detected to be greater than a preset threshold, a text segmentation tool is used to perform secondary segmentation on the initial document fragment; After the segmentation process, the headings at each level are retained for the initial document fragments, and the standard document name of the airport software operation and maintenance document to which each initial document fragment belongs is appended as source information to the text content of the initial document fragment.

6. The retrieval method according to claim 5, characterized in that, The steps for performing secondary segmentation on the initial document fragment using a text segmentation tool include: A text segmentation tool based on semantic boundaries is generated using a lightweight natural language model designed for the operations and maintenance field. Obtain the sequence of consecutive sentences in the context window of the initial document fragment; The initial document fragment and the continuous sentence sequence of the context window are input into the text segmentation tool, and the text segmentation tool is used to perform secondary segmentation on the initial document fragment.

7. The retrieval method according to claim 2, characterized in that, The steps of vectorizing and performing Gaussian mixture clustering on the initial document fragment include: Each of the initial document fragments is vectorized to obtain an initial vector; The initial vector is subjected to vector dimensionality reduction processing to obtain the target document vector; The target document vector is subjected to semantic clustering using a Gaussian mixture clustering algorithm to obtain a set of document fragments. If the number of document fragments in the document fragment set exceeds a preset threshold, the document fragments in the document fragment set are re-vectorized, vectorized, and subjected to semantic clustering to obtain a new document fragment set.

8. A retrieval device for an intelligent question-and-answer system for airport operation and maintenance, characterized in that, include: The semantic rewriting unit is used to receive user question dialogues and perform question semantic rewriting and question expansion on the user question dialogues to obtain a hybrid index containing question keywords and question vectors; The hybrid retrieval unit is used to perform hybrid retrieval in a pre-built vector library using the hybrid index to obtain retrieval results containing document metadata. The vector library includes document metadata and index information corresponding to multiple document fragments. The document fragments are generated after segmenting, vectorizing, and processing knowledge of airport software operation and maintenance documents. The document fragments are segmented according to a multi-level heading structure, and each document fragment retains its parent heading. The question reasoning unit is used to input the retrieval results into the large language model, use the large language model to perform question association reasoning, and output the question reasoning results.

9. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the retrieval method of the airport operation and maintenance intelligent question-and-answer system according to any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the retrieval method of the airport operation and maintenance intelligent question-and-answer system according to any one of claims 1 to 7.

Images