Intelligent decision-making and automated declaration method and system for transportation technology project

Abstract

This application discloses a method for intelligent decision-making and automated application for transportation technology projects, belonging to the field of data processing technology. It includes: performing multi-source data mining to obtain heterogeneous web page data; calculating the similarity correlation matrix of the heterogeneous web page data to generate a quantitative decision feature map; using a large-capacity file retrieval module to recursively segment private unstructured regulatory documents, mapping them to high-dimensional dense feature vectors and constructing a local inverted index database; performing approximate nearest neighbor retrieval based on similarity distance in the local inverted index database to generate a set of enhanced retrieval suggestions; acquiring long professional logic texts through server-pushed event streaming; traversing the document's underlying abstract syntax tree of the target formatted template, and injecting and formatting the basic short text and long professional logic texts onto specified table and paragraph structure nodes based on a coordinate offset matching algorithm to generate assembled application text. This method improves the accuracy of application text generation.

Description

Technical Field

[0001] This application belongs to the field of data processing technology, and in particular relates to a method and system for intelligent decision-making and automated application for transportation technology projects. Background Technology

[0002] In the wave of the information age, transportation engineering management faces profound demands for transformation. Currently, the transportation industry still faces numerous pain points in areas such as the construction of information infrastructure, the integration and sharing of data resources, and professional management concepts, hindering the industry's intelligent transformation and development. With the deepening of the national institutional reform, the functions of transportation management in many places have been further integrated and optimized, and the institutional and mechanism reforms of transportation management have driven urban governance back towards the core goal of "reducing costs and increasing efficiency." At the same time, intelligent vehicles, intelligent transportation, and smart cities are interdependent, and the need for deep integration and mining of multi-source traffic information, such as vehicle sensors, image acquisition, and infrastructure status perception, is becoming increasingly urgent, jointly contributing to the intelligentization of modern traffic management.

[0003] However, key information such as policy documents, statistical data, standards and norms, and academic literature in the transportation industry are scattered across dozens of different government platforms and academic databases. Manual retrieval is not only time-consuming and labor-intensive, but also makes it difficult to conduct in-depth cross-analysis and correlation mining of massive amounts of multi-source and heterogeneous data. This results in a lack of macro-strategic perspective and accurate data support for project selection, making it impossible to form an effective decision-making loop. At the same time, writing documents such as "Science and Technology Project Proposals," "Feasibility Study Reports," and "Project Approval Forms" requires a lot of effort in typesetting, tabulation, and standardized text filling, making it difficult for researchers to focus their core energy on the core technology innovation in transportation itself. In addition, due to the lack of real-time access to industry internal norms and private data, general-purpose large language models are prone to "illusions" when generating technical routes, assessment targets, and techno-economic indicators. The generated content may seem reasonable, but it lacks industry basis, is detached from the unit's management system and historical technological accumulation, and cannot be directly used in the formal approval process.

[0004] To overcome the aforementioned technical challenges, the industry urgently needs to introduce RAG (Retrieval-Augmented Generation) technology. By combining language generation with real-time information retrieval, RAG technology can proactively search vast, up-to-date information databases, utilizing external data sources as the basis for artificial intelligence responses. This significantly reduces errors and illusions, enhancing the performance of generative AI models in specific domains. Summary of the Invention

[0005] This application aims to address at least one of the technical problems existing in the prior art. To this end, this application proposes a method and system for intelligent decision-making and automated application for transportation science and technology projects. By comprehensively utilizing RAG technology and large-scale models, it reduces model illusions, ensures compliance, and significantly reduces the repetitive document writing workload for researchers in the science and technology project application process. It provides an efficient, accurate, and compliant method for intelligent decision-making and automated application for transportation industry science and technology projects. To address the aforementioned problems, according to the first aspect of this application, a method for intelligent decision-making and automated application for transportation technology projects is provided, the method comprising: It acquires multimodal input instructions including those for the target project, performs multi-source data mining through a concurrent scheduling engine, and obtains heterogeneous web page data from pre-configured external government platforms and academic databases. Based on cascaded lexical analysis and dimensionality reduction clustering modules, the similarity correlation matrix of the heterogeneous web page data between different sources is calculated, and a quantitative decision feature map including trend direction is generated. The system retrieves private unstructured regulatory documents and uses a recursive segmentation strategy to segment these documents based on a large-capacity file retrieval module. The segments are then mapped to high-dimensional dense feature vectors and a local inverted index database is constructed. Using the multimodal input command as the query center, an approximate nearest neighbor retrieval based on similarity distance is performed in the local inverted index database. Redundant similar paragraphs are filtered by calculating the hash value of the returned fragments from the initially retrieved text blocks, and a set of enhanced retrieval suggestions is generated. A project outline is generated based on a quantitative decision feature map. The enhanced retrieval prompt set and the project outline are input into a pre-deployed large language model. Professional logic long text is obtained by pushing event streams through the server. Traverse the underlying abstract syntax tree of the target formatted template document, and inject and format the basic short text and professional logic long text respectively into the specified table and paragraph structure nodes based on the coordinate offset matching algorithm to generate the assembled application text.

[0006] According to one embodiment of this application, obtaining heterogeneous web page data from a pre-configured external government affairs platform and academic database includes: By launching a headless browser instance to automatically load dynamic web pages, injecting preset anti-detection configuration parameters, forcibly suspending the execution flow until the cell nodes associated with the specified cascading class selector are loaded, intercepting the underlying asynchronous request data, and extracting the required academic literature titles, authors, and journal source content nodes to obtain heterogeneous web page data.

[0007] According to one embodiment of this application, calculating the similarity association matrix of the heterogeneous web page data between different sources includes: By merging the descriptive text from heterogeneous web page data, and performing lexical segmentation on the descriptive text based on a pre-built professional dictionary in the field of transportation technology, multiple heterogeneous source texts are obtained. For any two heterogeneous source texts from multiple heterogeneous source texts, extract the high-weighted keyword set based on term frequency-inverse document frequency respectively; Calculate the ratio of the number of elements in the intersection of two keyword sets to the number of elements in their union, and quantify this ratio as the data source correlation coefficient between the technical research direction and the policy support direction; After traversing all heterogeneous source texts, a similarity correlation matrix is ​​obtained between heterogeneous web page data from different sources.

[0008] According to one embodiment of this application, the method of using a recursive segmentation strategy to segment private unstructured regulatory documents based on a large-capacity file retrieval module, mapping them to high-dimensional dense feature vectors, and constructing a local inverted index database includes: Based on the large-capacity file retrieval module, a parser is dynamically assigned to extract plain text according to the suffix type of private unstructured regulatory documents, and a recursive character segmenter based on the preset character length limit and the number of overlapping characters is applied to generate a text block matrix. The text block matrix is ​​converted into floating-point feature vectors using a pre-trained Chinese deep embedding model to obtain high-dimensional dense feature vectors. Then, a similarity retrieval index is constructed in the local file system using a clustering quantization algorithm to obtain a local inverted index database.

[0009] According to one embodiment of this application, the step of filtering redundant and similar paragraphs by calculating the hash value of the returned fragments from the initially retrieved text blocks to generate a set of enhanced search suggestions includes: The initially retrieved text blocks are stripped of whitespace characters, punctuation marks, and format control characters, and then standardized and reduced in dimensionality. Calculate the hash value of the normalized string using a specified secure hash algorithm, and query the visited hash set in the current retrieval period to block the text mapped by duplicate feature vectors; The retained fragments are concatenated into a set of search result fragments in the format of search content block number-source file name-text content, thus generating a set of search enhancement suggestions.

[0010] According to one embodiment of this application, the step of traversing the document's underlying abstract syntax tree of the target formatting template, and injecting and formatting the basic short text and professional logical long text respectively onto the specified table and paragraph structure nodes based on a coordinate offset matching algorithm, generates the assembled declaration text, including: Nested traversal is performed based on the matching rules of the preloaded dictionary mapping; If a specific template cell text is identified in the document table abstract syntax tree as containing a mapped key value and meets the preset double cell offset rewrite condition, then the text is moved across the next adjacent physical cell, positioned in the second successor cell, its paragraph object is cleared and the target value is written. At the same time, the built-in style rewriter is used to force the font of the target value to be configured as SimSun font and the specified font size attribute system, generating the assembled declaration text.

[0011] According to one embodiment of this application, the method further includes: when the target formatting template is a preset pure table implementation node plan class template, triggering the bypass branch execution logic, interrupting the communication call request to the large-capacity file retrieval module and the large language model, directly receiving the two-dimensional item array containing multiple field attributes pushed by the front-end component, cyclically locating the base table row in the target document object and performing deep copy propagation node, writing the discrete fields in the two-dimensional item array into the corresponding columns of the clone node in sequence, and applying global zero indentation and center alignment constraint instructions to achieve pure dynamic assembly of a specific structure form.

[0012] According to a second aspect of this application, a smart decision-making and automated application system for transportation technology projects is provided, the system comprising: The acquisition module is used to acquire multimodal input instructions including the target project, and to perform multi-source data mining through a concurrent scheduling engine to acquire heterogeneous web page data from pre-configured external government platforms and academic databases; The first processing module is used to calculate the similarity correlation matrix of the heterogeneous web page data from different sources based on the cascaded lexical analysis and dimensionality reduction clustering module, and generate a quantitative decision feature map including trend direction. The second processing module is used to obtain private unstructured regulatory documents. Based on the large-capacity file retrieval module, the private unstructured regulatory documents are segmented using a recursive segmentation strategy, mapped into high-dimensional dense feature vectors, and a local inverted index database is constructed. The third processing module is used to perform approximate nearest neighbor retrieval based on similarity distance in the local inverted index database with the multimodal input command as the query center, and to filter redundant similar paragraphs by calculating the hash value of the returned fragments from the initially retrieved text blocks, and to generate a set of enhanced retrieval prompts. The fourth processing module is used to generate a project outline based on the quantitative decision feature map, input the retrieval enhancement prompt set and the project outline into the pre-deployed large language model, and obtain the professional logic long text through server push event streaming; The fifth processing module is used to traverse the document's underlying abstract syntax tree of the target formatted template, and based on the coordinate offset matching algorithm, inject and format the basic short text and the professional logic long text into the specified table and paragraph structure nodes respectively, generating the assembled application text.

[0013] According to a third aspect of this application, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the intelligent decision-making and automated application method for transportation technology projects as described in the first aspect above.

[0014] According to a fourth aspect of this application, a non-transitory computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the intelligent decision-making and automated application method for transportation technology projects as described in the first aspect above.

[0015] According to a fifth aspect of this application, a chip is provided, the chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run programs or instructions to implement the intelligent decision-making and automated application method for transportation technology projects as described in the first aspect.

[0016] According to a sixth aspect of this application, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the intelligent decision-making and automated application method for transportation technology projects as described in the first aspect above.

[0017] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application.

[0018] This application provides a method for intelligent decision-making and automated application for transportation technology projects, which has the following advantages over existing technologies: (1) This application integrates concurrent multi-source heterogeneous data acquisition with natural language processing deep analysis technology, compressing the traditional manual process of collecting and comparing policy, academic, and standard information, which would take several days, to the second level. Through automated topic clustering and cross-data source correlation analysis, it provides research management personnel with quantitative project planning suggestions such as "policy hot topics," "technological focus areas," and "potential market opportunities," breaking through information barriers and providing support for scientific decision-making.

[0019] (2) This application innovatively introduces RAG technology and a local vector database, enabling the large language model to strictly anchor to the proprietary industry regulations and historical typical project documents uploaded by the unit when generating highly professional content such as "technical and economic indicators" and "innovation points and intellectual property rights". Combined with a redundant text filtering mechanism based on hash algorithm, it provides the model with clean and high information density context constraints, improves the consistency between the generated application materials and industry standards, reduces model illusion, and improves the compliance of business texts.

[0020] (3) This application, by constructing a dynamic form mapping dictionary and an injection algorithm based on the traversal of the underlying abstract syntax tree, is able to handle nested table templates containing complex merged cells and coordinate offsets across rows. Furthermore, by implementing a bypass execution mechanism in pure table mode, the system balances the needs of freely generating large models with filling strictly structured forms. By simply inputting a core idea summary of less than a hundred words into the interface, a document such as a "Feasibility Study Report" of several thousand words, fully conforming to national official document formatting standards, can be generated with a single click. This achieves an automated closed loop from divergent thinking to standardized document implementation, reshaping the formatting process and greatly releasing scientific research productivity. Attached Figure Description

[0021] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart illustrating the intelligent decision-making and automated application method for transportation technology projects provided in this application embodiment; Figure 2 A schematic diagram of the intelligent decision-making and automated application architecture for science and technology projects in the transportation industry provided for the embodiments of this application; Figure 3 This is a flowchart illustrating the intelligent capture and NLP analysis and evaluation process for multi-source heterogeneous data in an embodiment of this application. Figure 4 This is a flowchart illustrating the knowledge base vectorization construction and intelligent retrieval enhancement processing based on the RAG architecture in an embodiment of this application. Figure 5 This is a flowchart illustrating the abstract syntax tree-level Word template dynamic parsing, large model streaming content generation, and coordinate offset injection in an embodiment of this application. Figure 6 This is a schematic diagram of the intelligent decision-making and automated application system for transportation technology projects provided in this application embodiment; Figure 7 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0023] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate to allow embodiments of this application to be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0024] The following description, in conjunction with the accompanying drawings, details the intelligent decision-making and automated application method for transportation technology projects, the intelligent decision-making and automated application system for transportation technology projects, the electronic equipment, and the readable storage medium provided in this application, through specific embodiments and application scenarios.

[0025] Among them, the intelligent decision-making and automated application method for transportation technology projects can be applied to the terminal, specifically executed by the hardware or software in the terminal.

[0026] The terminal includes, but is not limited to, portable communication devices such as mobile phones or tablets with touch-sensitive surfaces (e.g., touchscreen displays and / or touchpads). It should also be understood that, in some embodiments, the terminal may not be a portable communication device, but rather a desktop computer with touch-sensitive surfaces (e.g., touchscreen displays and / or touchpads).

[0027] The following embodiments describe a terminal including a display and a touch-sensitive surface. However, it should be understood that the terminal may include one or more other physical user interface devices such as a physical keyboard, mouse, and joystick.

[0028] The intelligent decision-making and automated application method for transportation technology projects provided in this application embodiment can be executed by an electronic device or a functional module or entity within an electronic device that can realize the intelligent decision-making and automated application method for transportation technology projects. The electronic devices mentioned in this application embodiment include, but are not limited to, mobile phones, tablets, computers, cameras, and wearable devices. The following uses an electronic device as the execution subject to illustrate the intelligent decision-making and automated application method for transportation technology projects provided in this application embodiment.

[0029] In the management and application process of science and technology projects in the transportation industry, researchers and management departments still face extremely heavy workloads in preliminary information research, decision-making argumentation, and application material preparation. Traditional transportation science and technology project initiation and application rely heavily on manual operations, which has the following significant limitations: First, there is a serious data silo problem in the early stages of scientific research intelligence gathering. While the connectivity of transportation systems is constantly improving, the stages of transportation infrastructure construction vary across different regions. Key information such as relevant macro-level policy documents, industry news, transportation statistics, national standards and regulations, and cutting-edge academic literature are scattered across various government websites and academic databases. Manual retrieval is not only time-consuming and labor-intensive, but also makes it difficult to conduct in-depth cross-analysis and correlation mining of massive amounts of multi-source, heterogeneous data. This results in the selection of scientific and technological projects and the formulation of requirements often lacking a macro-strategic perspective and accurate data support, failing to form an effective project decision-making loop.

[0030] Second, the process of writing project application materials is highly repetitive and formatted. Writing documents such as the "Science and Technology Project Proposal," "Feasibility Study Report," and "Project Approval Form" requires a significant amount of effort in typesetting, tabulation, and filling in standardized text. This inefficient textual labor makes it difficult for researchers to focus their core energy on the innovation of core transportation technologies (such as large-scale transportation model applications, vehicle-road cooperation, digital twins, and low-altitude transportation).

[0031] Third, while generative AI technologies such as large language models have made groundbreaking progress in the field of natural language processing, their direct application to technology management in the transportation sector faces severe technical bottlenecks and legal risks. Transportation industry documents are diverse in content, inconsistent in format, and densely packed with technical jargon. Due to a lack of up-to-date industry standards and proprietary data, general-purpose large language models are prone to generating "illusions" when generating technical roadmaps, assessment targets, and techno-economic indicators—that is, the generated content appears reasonable but is inconsistent with reality or lacks any industry basis. Furthermore, due to the lack of a real-time updated internal knowledge base, application materials generated by large models often deviate from the organization's actual management system and historical technological accumulation, making them unsuitable for direct use in formal approval processes.

[0032] Figure 1 This is a flowchart illustrating the intelligent decision-making and automated application method for transportation technology projects provided in this application embodiment, such as... Figure 1 As shown, the intelligent decision-making and automated application method for the transportation technology project includes steps 110, 120, 130, 140, 150 and 160.

[0033] Step 110: Obtain multimodal input instructions including the target project, and perform multi-source data mining through a concurrent scheduling engine to obtain heterogeneous web page data from pre-configured external government platforms and academic databases; In some embodiments, obtaining heterogeneous web page data from pre-configured external government platforms and academic databases includes: By launching a headless browser instance to automatically load dynamic web pages, injecting preset anti-detection configuration parameters, forcibly suspending the execution flow until the cell nodes associated with the specified cascading class selector are loaded, intercepting the underlying asynchronous request data, and extracting the required academic literature titles, authors, and journal source content nodes to obtain heterogeneous web page data.

[0034] Step 120: Based on the cascaded lexical analysis and dimensionality reduction clustering module, calculate the similarity correlation matrix of the heterogeneous web page data between different sources, and generate a quantitative decision feature map including trend direction; In some embodiments, calculating the similarity association matrix of the heterogeneous web page data between different sources includes: By merging the descriptive text from heterogeneous web page data, and performing lexical segmentation on the descriptive text based on a pre-built professional dictionary in the field of transportation technology, multiple heterogeneous source texts are obtained. For any two heterogeneous source texts from multiple heterogeneous source texts, extract the high-weighted keyword set based on term frequency-inverse document frequency respectively; Calculate the ratio of the number of elements in the intersection of two keyword sets to the number of elements in their union, and quantify this ratio as the data source correlation coefficient between the technical research direction and the policy support direction; After traversing all heterogeneous source texts, a similarity correlation matrix is ​​obtained between heterogeneous web page data from different sources.

[0035] It is easy to understand that this application targets technology management in the field of transportation planning and construction. Figure 2 This application provides a schematic diagram of an intelligent decision-making and automated application architecture for technology projects in the transportation industry. Figure 2 As shown, the overall architecture is built on the Python Flask microservice framework, following the design principles of front-end and back-end separation and high modular decoupling. The topology is divided into a data persistence layer, a core service layer, an interface routing layer, and a front-end presentation layer from bottom to top.

[0036] The data persistence layer primarily includes the local file system's result storage directory (RESULT_DIR) and knowledge base storage directory (KNOWLEDGE_BASE_DIR), as well as a local high-speed vector database built using FAISS (Facebook AI Similarity Search) to support text slice persistence and near-line retrieval. The core service layer comprises four independent execution engines, providing global call services through a singleton pattern. (1) Mining Service (web crawler data mining engine): responsible for the automated collection of multi-source heterogeneous data on the Internet; (2) Analysis Service (Big Data Analysis Engine): Responsible for in-depth mining and quantitative modeling of massive amounts of text; (3) Rag Service (Search Enhancement Engine): Manages document slicing, embedding model calls, and similarity matching; (4) Gen Service (Large Model Generation and Assembly Engine): It interfaces with third-party large language models and is responsible for the dynamic compilation and assembly of target documents.

[0037] The interface routing layer uses Flask's Blueprint mechanism to divide business logic into main_routes (main view and page routes), mining_routes (mining and analysis interfaces), and gen_routes (generation and knowledge base upload interfaces), effectively ensuring resource allocation and stability for large-scale model streaming requests and large-scale vector retrieval under high concurrency.

[0038] Figure 3 This is a flowchart illustrating the intelligent capture and NLP analysis and evaluation process for multi-source heterogeneous data according to an embodiment of this application. Figure 3 As shown, intelligent mining and deep analysis of multi-source heterogeneous data constitute the core of the "Decision Assistant," aiming to solve the problems of isolated and inefficient early-stage scientific research intelligence collection. To provide objective and comprehensive data support for the initiation and demand analysis of science and technology projects, this module integrates a powerful customized web crawler framework and natural language processing tools in the backend. Table 1 shows the different data mining strategies and underlying algorithm implementation mechanisms provided in the embodiments of this application.

[0039] After acquiring raw data from multiple sources, it needs to be transformed into a structured project decision report. The Analysis Service module integrates statistical methods and lexical analysis algorithms. To avoid differences in operating environments, a smooth degradation compatibility mode is designed: if the Jieba word segmentation dependency library exists in the environment, the accurate word segmentation algorithm based on Hidden Markov Model and TF-IDF (Term Frequency-Inverse Document Frequency) is enabled; if the dependency is missing, it degrades to the basic mode of extracting Chinese and English word elements using regular expressions (such as r'[\u4e00-\u9fa5]{2,}|[a-zA-Z]{3,}'). Rules and policies Ministry of Transport official website (static / pseudo-dynamic) The system utilizes an HTTP request library to construct GET requests with a spoofed User-Agent header and sends them to an endpoint with a specific channel ID and search scope parameter. A parsing library identifies sets of nodes in the DOM tree that bear the data-label or search-result attribute class. A built-in relative path absolute completion algorithm is used to reconstruct hyperlinks pointing to the details page. News Updates China Transportation News Network (POST and JSON interfaces) By bypassing front-end page rendering, we directly simulate the underlying POST request chain of the API, constructing a JSON payload containing site ID and pagination parameters. We directly capture the JSON structure in the response and implement deserialization based on the date string's timestamp and a descending sorting algorithm to obtain the latest industry trends. Statistical data Ministry of Transport Statistical Data Platform (Complex JSON) Construct an x-www-form-urlencoded request payload for a specific site code. Parse a deeply nested result Docs JSON array object. Clean the pipeline using regular expressions to remove unwanted characters from the raw returned data. <em>Use HTML highlighted tags to extract clean statistical summary text. < / em> Standards and Specifications National Standards Full-Text Public Disclosure System (Complex Nested Tables) Traversing the HTML root node and Hierarchical Approach. For dynamically encrypted document encoding, the system deploys a multi-level extraction and degradation strategy: it first attempts to capture internal function parameters from the hyperlink's on-click event using regular expressions; if this fails, it degrades to parsing the hcno query parameter in the href attribute; finally, it assembles the encrypted access link using a URL Encode algorithm. academic literature CNKI (China National Knowledge Infrastructure) (Strong anti-scraping measures / dynamic rendering engine) An automated testing framework environment is built using Selenium and Web Driver Manager. Options parameters that bypass crawler detection are injected into the Chrome driver. An explicit wait mechanism (ExplicitWait) is used to forcibly block the execution thread until the characteristic CSS class name `td.name` is fully mounted in the Document Object Model, ensuring that higher-order DOM element extraction is performed after the Ajax asynchronous request is completely completed, thereby obtaining the paper title, author sequence, and journal source.

[0040] Table 1 Taking keyword extraction and TF-IDF weight calculation as an example, the title and description fields of all crawled data sources are merged, and the merged long text is substituted into the domain dictionary. For any feature word $w_i$, the system calculates its TF-IDF weight using the following formula: After filtering the stop word set, the system selects the words of the ranking class (n, vn, nr) with the highest word frequency and feature weight to form the core word cloud matrix of "policy hot words" and "technological research focus".

[0041] Furthermore, the topic relevance analysis aims to reveal the degree of overlap in core concerns between different data sources (such as government regulations and policies and academic papers published by the research community). For any two heterogeneous data types A and B, the system extracts their respective sets of preceding keywords. and Calculate the Jaccard similarity coefficient between two sets to quantify their association: Based on the correlation coefficient quantification threshold, it can automatically generate insightful conclusions with multiple dimensions, such as generating high-value strategic recommendation texts like: "Around a certain technology field, which has received key policy attention, and in combination with market demand and industry development trends, it is recommended to carry out relevant science and technology research and development projects," directly providing decision support for science and technology management personnel.

[0042] Analysis Service extracts keyword sets for each data source, such as the policy keyword set K_policy and the paper keyword set K_paper. For any two sources A and B, their correlation is the intersection of their keywords divided by the union of their keywords, i.e., |K_A∩K_B| / |K_A∪K_B|. The similarity value for each pair of data sources is stored with keys such as policy_paper. If represented as a matrix in the patent text, the similarity values ​​between each pair of data sources are filled into an n×n symmetric matrix, with the diagonal set to 1, thus obtaining the similarity correlation matrix.

[0043] Step 130: Obtain private unstructured regulatory documents. Based on the large-capacity file retrieval module, the private unstructured regulatory documents are segmented using a recursive segmentation strategy, mapped into high-dimensional dense feature vectors, and a local inverted index database is constructed. In some embodiments, the method of using a recursive segmentation strategy to segment private unstructured regulatory documents based on a large-capacity file retrieval module, mapping them to high-dimensional dense feature vectors, and constructing a local inverted index database includes: Based on the large-capacity file retrieval module, a parser is dynamically assigned to extract plain text according to the suffix type of private unstructured regulatory documents, and a recursive character segmenter based on the preset character length limit and the number of overlapping characters is applied to generate a text block matrix. The text block matrix is ​​converted into floating-point feature vectors using a pre-trained Chinese deep embedding model to obtain high-dimensional dense feature vectors. Then, a similarity retrieval index is constructed in the local file system using a clustering quantization algorithm to obtain a local inverted index database.

[0044] Step 140: Using the multimodal input command as the query center, perform approximate nearest neighbor retrieval based on similarity distance in the local inverted index database. Filter redundant similar paragraphs by calculating the hash value of the returned fragments from the initially retrieved text blocks, and generate a set of enhanced retrieval suggestions. In some embodiments, the step of filtering redundant and similar paragraphs by calculating the hash value of the returned fragments from the initially retrieved text blocks to generate a set of enhanced search suggestions includes: The initially retrieved text blocks are stripped of whitespace characters, punctuation marks, and format control characters, and then standardized and reduced in dimensionality. Calculate the hash value of the normalized string using a specified secure hash algorithm, and query the visited hash set in the current retrieval period to block the text mapped by duplicate feature vectors; The retained fragments are concatenated into a set of search result fragments in the format of search content block number-source file name-text content, thus generating a set of search enhancement suggestions.

[0045] Step 150: Generate a project outline based on the quantitative decision feature map, input the retrieval enhancement prompt set and the project outline into the pre-deployed large language model, and obtain professional logic long text through server push event streaming; Figure 4 This is a flowchart illustrating the knowledge base vectorization construction and intelligent retrieval enhancement processing based on the RAG architecture, as described in this application embodiment. Figure 4 As shown, to completely solve the "illusion" problem of large language models in vertical domains, the key innovation of this application lies in the deep integration of the RAG architecture. By transforming the black-box state of the generation process into an "open-book exam" based on enterprise proprietary knowledge, the compliance of technical indicators and standards in the text generation process is effectively guaranteed.

[0046] When users upload supporting files (such as .pdf, .txt, .docx) containing internal technical requirements and national industry standards in batches through the "Industry Knowledge Base Management and Intelligent Retrieval" interface, Rag Service triggers a vectorization construction process. First, the system relies on the extended library to perform classification routing based on the uploaded file extensions, calling PyPDFLoader, Text Loader, or Docx2txtLoader to strip unstructured documents into a uniform plain text stream. Then, a context-aware recursive character segmentation strategy, RecursiveCharacterTextSplitter, is implemented. In rigorous documents such as traffic engineering standards, technical clauses often span long clauses. To balance the granularity accuracy of the vector database during retrieval with preserving complete technical context, the system sets the baseline text chunk size (chunk_size) to 800 characters and forces the overlap window length (chunk_overlap) to 100 characters. This 100-character redundant coverage, achieved through a mathematically defined sliding window algorithm, ensures that even if a key "technical and economic indicator" phrase is located at a segmentation boundary, its core semantics and contextual modifiers are completely encapsulated within at least one sub-block, preventing errors caused by model comprehension gaps. Next, by introducing a deep pre-trained embedding model (such as BAAI / bge-small-zh-v1.5), the system performs forward propagation inference on each generated discrete text block, mapping it to a high-dimensional dense floating-point feature vector (Dense Vector Embedding). Finally, these vector sets and their corresponding metadata (including document source information) are fed into the FAISS clustering engine, establishing and persisting a multi-dimensional index file at the disk level, constructing a high-throughput local vector database.

[0047] When a user initiates the intelligent document generation process, if the "Knowledge Base Enhanced Generation" switch is enabled, the system will automatically concatenate the "Project Topic" and "General Project Idea" submitted by the user in the dynamic form as a search probe feature statement (Query). The system also transforms this probe statement into a target vector through an embedding model and executes an L2 Euclidean distance (or cosine similarity) retrieval algorithm in the FAISS inverted index space, typically extracting the Top-K (e.g., k=5) closest candidate document blocks. However, in actual normative documents, there are many highly homogenized disclaimers or repetitive technical requirements sections. Directly submitting all candidate blocks to a large model may result in the context window being filled with meaningless repetitive text, leading to information density collapse during the input phase. Therefore, this system innovatively implements document block deduplication logic based on the SHA-256 hash algorithm. The specific algorithm flow is as follows: (1) For each candidate document block, execute the character cleaning pipeline to remove all spaces, carriage returns, and punctuation marks, and convert the character encoding to all lowercase to generate a normalized verification string. .

[0048] (2) Calculate the secure hash value of the normalized string using a hash algorithm. .

[0049] (3) The system maintains a runtime hash verification set seen_content_hashes. If the hash value H of the current candidate block already exists in the set, the candidate block is marked as a duplicate block and discarded immediately in the loop; otherwise, it is retained and its hash value is added to the verification set. Multiple independent, high-density contextual fragments purified through this process will be combined into a retrieval prompt body with special tags (such as [Source: XXX standard.pdf]), and pushed into the input queue of the large language model along with the user's original command.

[0050] Step 160: Traverse the document's underlying abstract syntax tree of the target formatted template, and inject and format the basic short text and professional logic long text respectively into the specified table and paragraph structure nodes based on the coordinate offset matching algorithm to generate the assembled application text.

[0051] In some embodiments, the step of traversing the document's underlying abstract syntax tree of the target formatted template, injecting and formatting the basic short text and the professional logical long text respectively onto the specified table and paragraph structure nodes based on a coordinate offset matching algorithm, generates the assembled declaration text, including: Nested traversal is performed based on the matching rules of the preloaded dictionary mapping; If a specific template cell text is identified in the document table abstract syntax tree as containing a mapped key value and meets the preset double cell offset rewrite condition, then the text is moved across the next adjacent physical cell, positioned in the second successor cell, its paragraph object is cleared and the target value is written. At the same time, the built-in style rewriter is used to force the font of the target value to be configured as SimSun font and the specified font size attribute system, generating the assembled declaration text.

[0052] Figure 5 The flowchart for the abstract syntax tree-level Word template dynamic parsing, large model streaming content generation, and coordinate offset injection in this application embodiment is as follows: Figure 5 As shown, this section serves as the central assembly workshop for the entire application architecture. Its task is to perfectly structure and fit together fragmented user input, long text outputs from large models, and complex Word template layouts.

[0053] Generating a feasibility study report of several thousand words typically faces network latency of tens of seconds to several minutes on the inference cluster. To prevent request timeouts due to prolonged blocking at the front-end gateway and to alleviate user anxiety, the system backend employs a generator and streaming architecture in the ` / generate_stream` route. The system wraps business logic functions with Flask's `stream_with_context` and enforces the HTTP message content type as `text / event-stream`, thus establishing SSE protocol communication. During generation, the core scheduler traverses the configuration files of predefined templates. Based on the type of triggering request (e.g., generating only the `table` section or generating the `report` section), it extracts the target field set to be processed (e.g., project name, main research content, technical and economic indicators, innovation points). The system attaches highly customized prompt word boundaries to different fields and combines the retrieved RAG context to initiate network calls to large language model interfaces (e.g., Deep Seek-chat nodes compatible with the Open AI underlying protocol). During this cycle, the generator continuously pushes the execution status identifiers serialized into JSON format (such as {"status":"Generating [key technical and economic indicators]..."} or {"status":"Successfully retrieved knowledge base data"}) to the front end in real time via the yield instruction, achieving transparent progress awareness.

[0054] Implanting text generated from large models into Word documents containing numerous complex cell merging and line-crossing layouts is a technically challenging operation. The system abandons traditional regular expression-based string replacement schemes and instead operates on the underlying XML and abstract syntax tree interface library python-docx. In the handler logic for specific templates: (1) Offset writing of table nodes: The program recursively traverses all Table->Row->Cell instances under the root of the target Document tree. After converting the traversed target cells to plain text and removing internal spaces, it performs an exact equality comparison with the keys of the mapping dictionary (such as [undertaking unit]). Once the preset anchor point is hit, the system does not write the data to the anchor point itself, but uses the coordinate offset algorithm for addressing. Usually, the addressing target is set to the cell one position after the current iteration index (i.e., row.cells[i+1]). However, in order to solve the specific complex merged cells that are common in Chinese official documents (such as the contact information field that spans multiple columns), the system introduces a specific set of exception keys (two_cell_offset_keys). For data key values ​​that fall into this set, the system modifies the pointer span and locates the target area two columns away (i.e., row.cells[i+2]). Subsequently, the system destroys all remaining paragraph nodes and character entities within the addressed target and implements a forced style overwrite based on the SimSun_GB2312 Chinese font library, with the font size limited to 12pt (Pt(12)). This satisfies the standardized typesetting audit requirements of government agencies and transportation authorities.

[0055] (2) Dynamic growth of paragraphs in long report sections: To address the large-scale unstructured report requirements in template_2, the system abandons operation within a small table. The algorithm detects specific anchor paragraphs in the text (such as the second occurrence of the specific keyword "Appendix 1") and calls the insert_paragraph_before node insertion function. The system iteratively processes each generated professional chapter (such as "Assessment Objectives and Technical and Economic Indicators"), dynamically constructing first-level headings based on the built-in Chinese numerical enumeration dictionary (I, II, III...). Subsequently, it forcibly assigns a boldface style to the outline headings and configures the hundreds of words of report text output by the large language model to 1.5 line spacing and 14-point Song font, achieving high-quality layout and continuous growth of the text.

[0056] The final submission text is assembled by the Gen Service and various template handlers. The Gen Service first merges the user input and the content generated by the large model, and selects the corresponding template handler; handler.process_document opens the Word template, iterates through the table, row, cell or body paragraph, and matches field anchors according to the placeholder_map; ordinary fields are written to adjacent cells, special fields are written to target cells offset two cells according to two_cell_offset_keys, and long text report areas are inserted with chapter paragraphs via insert_paragraph_before; after writing, the font and size are uniformly set, and it is saved as a docx file with a UUID suffix, which the front end then downloads according to the returned path.

[0057] In some embodiments, the method further includes: when the target formatting template is a preset pure table implementation node plan class template, triggering bypass branch execution logic, interrupting the communication call request to the large-capacity file retrieval module and the large language model, directly receiving the two-dimensional item array containing multiple field attributes pushed by the front-end component, cyclically locating the base table row in the target document object and performing deep copy propagation node, writing the discrete fields in the two-dimensional item array into the corresponding columns of the clone node in sequence, and applying global zero indentation and center alignment constraint instructions to achieve pure dynamic assembly of a specific structure form.

[0058] Specifically, when a request to generate the "Science and Technology Project Implementation Node Plan Table" (template_4) is encountered, the system's execution state machine undergoes a fundamental fork. This type of template business is essentially a summary of the phase schedules for various project management nodes (such as finalization of procurement documents, winning bids, contract signing, etc.), with no room for divergent writing driven by a large model. Therefore, when the target template belongs to the context of template_4, the generation scheduler suspends asynchronous task network request calls to the RAG module and LLM, and switches to a bypass branch: the system directly unpacks the multi-dimensional array object of projects transmitted from the front end (a list of multiple project dictionaries, all_content['projects']) from the user_inputs payload. The algorithm locates and obtains the first table instance, doc.tables, and calls the underlying handle to forcibly remove the existing test blank rows used as placeholders. Then, table.add_row() is frequently activated within the array loop to multiply the required number of new rows. For each cloned row, the specific attribute text is directly assigned using the iterative counting index and key-value mapping relationship (e.g., filling bidding_result_date into cell 5 and final_acceptance_time into cell 9). Furthermore, the accompanying auxiliary function set_cell_font performs extreme normalization on each of the cloned paragraphs—completely resetting the paragraph spacing, left and right indentation, and first-line indentation to zero (i.e., Pt(0)), and supplementing this with a fully centered alignment attribute setting, thus achieving automated construction of a purely data-driven form.

[0059] Finally, the entire document abstract syntax tree, which has undergone complex reconstruction, is serialized into a disk file via a file system stream. The system uses `uuid.uuid4()` to generate a prefix code with an extremely low collision rate, and concatenates it with the secure delivery name built into the template configuration file to ensure that pointer overwrite disasters will never occur when multiple users trigger saving simultaneously under asynchronous concurrent threads. The backend transmits the successful disk write command along with the absolute path of the dynamically mapped UUID to the browser host, activating the terminal's implicit link to spontaneously download, thus achieving a closed loop for the entire lifecycle service.

[0060] The intelligent decision-making and automated application method for transportation technology projects provided in this application can be implemented by a transportation technology project intelligent decision-making and automated application system. This application uses the transportation technology project intelligent decision-making and automated application system executing the method as an example to illustrate the transportation technology project intelligent decision-making and automated application system provided in this application.

[0061] This application also provides an intelligent decision-making and automated application system for transportation technology projects, such as... Figure 6As shown, the intelligent decision-making and automated application system for transportation technology projects includes: an acquisition module 610, a first processing module 620, a second processing module 630, a third processing module 640, a fourth processing module 650, and a fifth processing module 660.

[0062] The acquisition module 610 is used to acquire multimodal input instructions including the target project, and to perform multi-source data mining through a concurrent scheduling engine to acquire heterogeneous web page data from pre-configured external government platforms and academic databases; The first processing module 620 is used to calculate the similarity correlation matrix of the heterogeneous web page data from different sources based on the cascaded lexical analysis and dimensionality reduction clustering module, and generate a quantitative decision feature map including trend direction. The second processing module 630 is used to obtain private unstructured regulatory documents. Based on the large-capacity file retrieval module, the private unstructured regulatory documents are segmented using a recursive segmentation strategy, mapped into high-dimensional dense feature vectors, and a local inverted index database is constructed. The third processing module 640 is used to perform an approximate nearest neighbor retrieval based on similarity distance in the local inverted index database with the multimodal input command as the query center, and to filter redundant similar paragraphs by calculating the hash value of the returned fragments from the initially retrieved text blocks, and to generate a set of enhanced retrieval prompts. The fourth processing module 650 is used to generate a project outline based on the quantitative decision feature map, input the retrieval enhancement prompt set and the project outline into a pre-deployed large language model, and obtain professional logic long text through server push event streaming; The fifth processing module 660 is used to traverse the document's underlying abstract syntax tree of the target formatted template, and inject the basic short text and the professional logic long text into the specified table and paragraph structure nodes respectively based on the coordinate offset matching algorithm, thereby generating the assembled application text.

[0063] In another implementation, the intelligent decision-making and automated application system for science and technology projects in the transportation industry includes: a multi-source mining and analysis unit, configured to respond to front-end commands to perform concurrent data capture operations, covering policies, news, statistical data, standards and specifications, and academic literature, and to perform lexical feature analysis and multi-dimensional cross-correlation calculations on the acquired data to output quantitative insights into strategic and project needs; a retrieval enhancement generation and allocation unit, configured to integrate multimodal text segmentation operators, high-dimensional deep embedding models, and local vector database indexes, responding to retrieval requests to return structured industry standard background fragments with anti-duplicate filtering and contextual association characteristics; and a large language streaming request form. The `yuan` module is configured to send a composite query load, which aggregates system prompts, user outline input, and retrieval background fragments, to the underlying natural language dialogue model. Simultaneously, it constructs a communication pipeline to listen for real-time generated logic fragments of the model and distributes status feedback sequences block by block via a server-pushed event streaming channel. The dynamic document assembly engine unit is configured to parse pre-made physical document templates of underlying file types, provide adaptable abstract syntax tree deep access interfaces and injection overwrite operations for placeholder configurations or pure table dynamic arrays for different business attributes, and complete and persist the final state business application files conforming to industry standards under a unique naming mechanism that ensures concurrent write security.

[0064] The dynamic document assembly engine unit is further configured with multiple independent processor modules. Each processor module encapsulates an independent placeholder mapping dictionary and a model trigger field list to enable differentiated large model calls and heterogeneous document rendering for templates in different fields such as project proposals, feasibility study reports, and science and technology project approval forms.

[0065] The intelligent decision-making and automated application system for transportation technology projects provided in this application embodiment can achieve... Figures 1 to 5 The various processes implemented in the intelligent decision-making and automated application method for transportation technology projects will not be described in detail here to avoid repetition.

[0066] In some embodiments, such as Figure 7 As shown, this application embodiment also provides an electronic device 700, including a processor 701, a memory 702, and a computer program stored in the memory 702 and executable on the processor 701. When the program is executed by the processor 701, it implements the various processes of the above-described intelligent decision-making and automated application method embodiment for transportation technology projects and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0067] It should be noted that the electronic devices in the embodiments of this application include the mobile electronic devices and non-mobile electronic devices described above.

[0068] This application also provides a non-transitory computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the various processes of the above-described intelligent decision-making and automated application method for transportation technology projects and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0069] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0070] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-mentioned intelligent decision-making and automated application method for transportation technology projects.

[0071] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0072] This application also provides a chip, which includes a processor and a communication interface. The communication interface and the processor are coupled. The processor is used to run programs or instructions to implement the various processes of the above-described intelligent decision-making and automated application method for transportation technology projects, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0073] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0074] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element. Furthermore, it should be noted that the scope of the methods and systems in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0075] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the intelligent decision-making and automated application method for transportation technology projects of various embodiments of this application.

[0076] In the description of this application, "first feature" and "second feature" may include one or more of the features.

[0077] In the description of this application, "multiple" means two or more.

[0078] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

[0079] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0080] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for intelligent decision-making and automated application for transportation technology projects, characterized in that, The method includes: It acquires multimodal input instructions including those for the target project, performs multi-source data mining through a concurrent scheduling engine, and obtains heterogeneous web page data from pre-configured external government platforms and academic databases. Based on cascaded lexical analysis and dimensionality reduction clustering modules, the similarity correlation matrix of the heterogeneous web page data between different sources is calculated, and a quantitative decision feature map including trend direction is generated. The system retrieves private unstructured regulatory documents and uses a recursive segmentation strategy to segment these documents based on a large-capacity file retrieval module. The segments are then mapped to high-dimensional dense feature vectors and a local inverted index database is constructed. Using the multimodal input command as the query center, an approximate nearest neighbor retrieval based on similarity distance is performed in the local inverted index database. Redundant similar paragraphs are filtered by calculating the hash value of the returned fragments from the initially retrieved text blocks, and a set of enhanced retrieval suggestions is generated. A project outline is generated based on a quantitative decision feature map. The enhanced retrieval prompt set and the project outline are input into a pre-deployed large language model. Professional logic long text is obtained by pushing event streams through the server. Traverse the underlying abstract syntax tree of the target formatted template document, and inject and format the basic short text and professional logic long text respectively into the specified table and paragraph structure nodes based on the coordinate offset matching algorithm to generate the assembled application text.

2. The intelligent decision-making and automated application method for transportation technology projects according to claim 1, characterized in that, The process of obtaining heterogeneous web page data from pre-configured external government platforms and academic databases includes: By launching a headless browser instance to automatically load dynamic web pages, injecting preset anti-detection configuration parameters, forcibly suspending the execution flow until the cell nodes associated with the specified cascading class selector are loaded, intercepting the underlying asynchronous request data, and extracting the required academic literature titles, authors, and journal source content nodes to obtain heterogeneous web page data.

3. The intelligent decision-making and automated application method for transportation technology projects according to claim 1, characterized in that, The calculation of the similarity correlation matrix between the heterogeneous web page data from different sources includes: By merging the descriptive text from heterogeneous web page data, and performing lexical segmentation on the descriptive text based on a pre-built professional dictionary in the field of transportation technology, multiple heterogeneous source texts are obtained. For any two heterogeneous source texts from multiple heterogeneous source texts, extract the high-weighted keyword set based on term frequency-inverse document frequency respectively; Calculate the ratio of the number of elements in the intersection of two keyword sets to the number of elements in their union, and quantify this ratio as the data source correlation coefficient between the technical research direction and the policy support direction; After traversing all heterogeneous source texts, a similarity correlation matrix is ​​obtained between heterogeneous web page data from different sources.

4. The intelligent decision-making and automated application method for transportation technology projects according to claim 1, characterized in that, The large-capacity file retrieval module uses a recursive segmentation strategy to segment private unstructured regulatory documents, mapping them to high-dimensional dense feature vectors and constructing a local inverted index database, including: Based on the large-capacity file retrieval module, a parser is dynamically assigned to extract plain text according to the suffix type of private unstructured regulatory documents, and a recursive character segmenter based on the preset character length limit and the number of overlapping characters is applied to generate a text block matrix. The text block matrix is ​​converted into floating-point feature vectors using a pre-trained Chinese deep embedding model to obtain high-dimensional dense feature vectors. Then, a similarity retrieval index is constructed in the local file system using a clustering quantization algorithm to obtain a local inverted index database.

5. The intelligent decision-making and automated application method for transportation technology projects according to claim 1, characterized in that, The step of filtering redundant and similar paragraphs by calculating the hash value of the returned fragments from the initially retrieved text blocks and generating a set of enhanced search suggestions includes: The initially retrieved text blocks are stripped of whitespace characters, punctuation marks, and format control characters, and then standardized and reduced in dimensionality. Calculate the hash value of the normalized string using a specified secure hash algorithm, and query the visited hash set in the current retrieval period to block the text mapped by duplicate feature vectors; The retained fragments are concatenated into a set of search result fragments in the format of search content block number-source file name-text content, thus generating a set of search enhancement suggestions.

6. The intelligent decision-making and automated application method for transportation technology projects according to claim 1, characterized in that, The document's underlying abstract syntax tree, which traverses the target formatted template, is used to inject and format basic short text and professional logical long text respectively, based on a coordinate offset matching algorithm, and overlay them onto the specified table and paragraph structure nodes to generate the assembled application text, including: Nested traversal is performed based on the matching rules of the preloaded dictionary mapping; If a specific template cell text is identified in the document table abstract syntax tree as containing a mapped key value and meets the preset double cell offset rewrite condition, then the text is moved across the next adjacent physical cell, positioned in the second successor cell, its paragraph object is cleared and the target value is written. At the same time, the built-in style rewriter is used to force the font of the target value to be configured as SimSun font and the specified font size attribute system, generating the assembled declaration text.

7. The intelligent decision-making and automated application method for transportation technology projects according to claim 1, characterized in that, The method further includes: when the target formatting template is a preset pure table implementation node plan template, triggering the bypass branch execution logic, interrupting the communication call request to the large-capacity file retrieval module and the large language model, directly receiving the two-dimensional project array containing multiple field attributes pushed by the front-end component, cyclically locating the base table row in the target document object and performing deep copy propagation node, writing the discrete fields in the two-dimensional project array into the corresponding columns of the clone node in sequence, and applying global zero indentation and center alignment constraint instructions to achieve pure dynamic assembly of a specific structure form.

8. A smart decision-making and automated application system for transportation technology projects, implemented using the smart decision-making and automated application method for transportation technology projects as described in any one of claims 1 to 7, characterized in that, The system includes: The acquisition module is used to acquire multimodal input instructions including the target project, and to perform multi-source data mining through a concurrent scheduling engine to acquire heterogeneous web page data from pre-configured external government platforms and academic databases; The first processing module is used to calculate the similarity correlation matrix of the heterogeneous web page data from different sources based on the cascaded lexical analysis and dimensionality reduction clustering module, and generate a quantitative decision feature map including trend direction. The second processing module is used to obtain private unstructured regulatory documents. Based on the large-capacity file retrieval module, the private unstructured regulatory documents are segmented using a recursive segmentation strategy, mapped into high-dimensional dense feature vectors, and a local inverted index database is constructed. The third processing module is used to perform approximate nearest neighbor retrieval based on similarity distance in the local inverted index database with the multimodal input command as the query center, and to filter redundant similar paragraphs by calculating the hash value of the returned fragments of the initially retrieved text blocks, and to generate a set of enhanced retrieval prompts. The fourth processing module is used to generate a project outline based on the quantitative decision feature map, input the retrieval enhancement prompt set and the project outline into the pre-deployed large language model, and obtain the professional logic long text through server push event streaming; The fifth processing module is used to traverse the document's underlying abstract syntax tree of the target formatted template, and based on the coordinate offset matching algorithm, inject and format the basic short text and the professional logic long text respectively to the specified table and paragraph structure nodes to generate the assembled application text.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the intelligent decision-making and automated application method for transportation technology projects as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the intelligent decision-making and automated application method for transportation technology projects as described in any one of claims 1 to 7.

Images