Large model-based intelligent matching method and system for export customs declaration commodity names

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to an export customs declaration commodity name intelligent matching system and method based on a large model. Through real-time acquisition and standardization of multi-source heterogeneous data, extraction and vectorization of multi-modal features of commodity information, generation of classification candidate items based on a hybrid expert big language model, and Pareto optimal decision based on multi-target optimization, intelligent and efficient matching of commodity names and customs codes is realized, and the matching efficiency of commodity names and customs codes is improved. And the compliance, the timeliness and the economic benefit of the customs declaration strategy are improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and specifically relates to an intelligent matching system and method for export customs declaration commodity names based on a large model. Background Technology

[0002] In current technological practices, the classification of customs codes for exported goods primarily relies on a knowledge-driven paradigm centered on human experience. Specifically, this model is typically led by customs brokers or customs experts with extensive industry knowledge. The workflow mainly includes: first, receiving and interpreting descriptive documents for exported goods, such as product name, specifications, materials, and uses; second, leveraging their accumulated professional knowledge and understanding of the Harmonized Commodity Description and Coding System (HMS), constructing an initial classification direction in their minds; third, verifying and revising the initial judgment by consulting static customs tariff books, past customs declaration case databases, and regulatory announcements scattered across various government websites; and finally, selecting the most suitable customs code from multiple potential codes based on experience for declaration. In the past, when international trade agreements were relatively stable, the pace of commodity iteration was slow, and the frequency of regulatory policy changes was low, this model, relying on the individual abilities of experts, largely ensured the normal operation of customs declaration business through its comprehensive judgment of complex and ambiguous information.

[0003] However, with the deepening of global economic integration, the emergence and frequent iteration of regional trade agreements (such as the Regional Comprehensive Economic Partnership, RCEP), and the refinement and dynamic development of port regulatory policies in various countries, the aforementioned human expert paradigm based on static knowledge and linear processes has gradually revealed its inherent limitations at the principle level. These limitations are not simply manifested as inefficiency or occasional human errors, but rather a deeper structural contradiction arising from a fundamental mismatch between the technological paradigm and the application scenario. The reason for this lies in the fact that modern international trade customs declaration and classification has evolved from a single-objective classification search problem into a complex decision-making problem involving global optimization under dynamic, high-dimensional, and multi-constraint conditions. First, the explosive growth of data dimensions makes traditional manual processing methods unsustainable. The physical properties, chemical composition, and processing technology of commodities, along with the massive amounts of information such as dynamically changing agreement tariffs, rules of origin, specific port regulatory conditions, and even sudden trade control measures, together constitute a complex and coupled decision space. This data is not only heterogeneous in origin and format, distributed across hundreds of fragmented information platforms, but more importantly, it exhibits non-linear interactions. For example, a slight difference in material composition can lead to significant tariff fluctuations after applying a specific trade agreement. Secondly, the diversification and conflict of decision-making objectives exacerbate the complexity of the problem. Enterprises no longer simply seek the "correct" code, but rather strive for a Pareto optimal solution that minimizes tariff costs, speeds up customs clearance, and reduces supply chain risk while ensuring 100% compliance. These objectives often inherently conflict; for instance, a code enjoying an extremely low agreement tariff rate may correspond to stricter regulatory conditions and longer inspection times, increasing logistics costs and time uncertainty. Traditional human expert models, with their inherent cognitive bandwidth and computing power, are fundamentally incapable of processing such high-dimensional information flows in real time, let alone quantitatively weighing and dynamically optimizing multiple conflicting objectives. The inherent limitations of this model lead to a trade-off in performance: in pursuit of ultimate accuracy and compliance, companies have to invest a lot of manpower and time in exhaustive information verification, which seriously sacrifices operational efficiency; conversely, if they rely on experience to make quick decisions in pursuit of efficiency, they are very likely to fail to capture the latest policy changes or ignore the deep relationships between data, which can lead to classification errors, economic losses or compliance risks.

[0004] Therefore, designing a novel technical framework that can effectively integrate and understand multi-source heterogeneous trade data, and on this basis, establish a decision-making model that can dynamically perceive changes in the policy environment and intelligently weigh multi-dimensional optimization objectives, thereby breaking through the fundamental bottlenecks faced by existing technical paradigms in dealing with modern complex trade scenarios, has become a key challenge and an urgent technical problem for those skilled in the art. Summary of the Invention

[0005] The technical problem to be solved by this invention is that the existing method of classifying export customs declaration commodities based on the experience of human experts has fundamental limitations in terms of information processing bandwidth, dynamic adaptability and quantitative decision-making ability when faced with the explosive growth of data dimensions, fragmentation of multi-source heterogeneous information and complex decision-making scenarios in modern international trade, as well as the conflict between multiple optimization objectives such as compliance, cost and timeliness. It cannot achieve the globally optimal customs declaration strategy while ensuring compliance.

[0006] To achieve the above-mentioned objectives, this invention provides an intelligent matching system and method for export customs declaration commodity names based on a large model. The aim is to build an automated technical framework that can integrate multi-source heterogeneous data, dynamically perceive the policy environment, and make multi-objective optimization decisions, thereby systematically solving the above-mentioned technical problems.

[0007] To achieve this objective, the present invention provides an intelligent matching system for export customs declaration commodity names based on a large model. The system is configured within a server cluster with data storage and computing capabilities. The system includes: A real-time data acquisition system is configured to acquire structured, semi-structured, and unstructured data related to international trade from multiple preset, heterogeneous external data sources in real time, and process it into a unified, standardized data format that can be used by downstream models. Further, the real-time data acquisition system includes: a static data retrieval module, a real-time data stream subscription and processing module, and a knowledge graph construction module.

[0008] The static data retrieval module is internally configured with a distributed web crawler agent cluster targeting the official websites of customs authorities of major global economies. This agent cluster performs retrieval tasks at least once daily, retrieving data from national customs tariff databases, compilations of commodity classification decisions, rules of origin documentation, and trade control policy announcements. For structured tariff databases, the module directly parses the data tables, extracting field information including HS Code, legal units of measurement, most-favored-nation (MFN) tariff rates, agreement tariff rates, VAT rates, and consumption tax rates. For unstructured policy announcements published in document form, the module first invokes a document layout analysis model based on a convolutional neural network to identify and separate text regions, tables, and images. Subsequently, an optical character recognition (OCR) engine converts the text regions into a machine-readable text stream. Finally, a named entity recognition model based on a bidirectional long short-term memory network and conditional random field (Bi-LSTM-CRF) architecture automatically extracts and labels key entity information from the text stream, including policy effective date, expiration date, applicable commodity scope, involved countries or regions, and specific regulatory conditions.

[0009] The real-time data stream subscription and processing module establishes persistent connections via an application programming interface (API) to the data stream services of customs data publishing platforms and international business information providers at no fewer than fifty major global ports. This module utilizes a consumer cluster deployed on a distributed message queue system (such as Apache Kafka) to subscribe to and receive real-time data messages regarding temporary adjustments to tariff rates, sudden trade control measures, congestion at specific ports, and fluctuations in shipping rates. Each received message is encapsulated into a data packet containing a data payload, a data source identifier, and a timestamp accurate to milliseconds.

[0010] The knowledge graph construction module, acting as a terminal of a data processing pipeline, receives data from the static data retrieval module and the real-time data stream subscription and processing module. This module first performs data cleaning and normalization operations, including removing duplicate data, unifying entity naming (e.g., mapping "United States of America," "US," and "America" ​​to unique country entity IDs), and standardizing the format of numerical values ​​such as tax rates and dates. After processing, the module injects all data into a dynamically updated trade knowledge graph. The knowledge graph is stored using an attribute graph database (such as Neo4j), where nodes represent core entities such as commodities, customs codes, countries, ports, trade agreements, and legal terms, while edges represent specific, directed relationships between these entities. For example, "Commodity A" <--[Applicable Code] --> "HS Code X", "HS Code X" --[Applicable to 'Vietnam' under the RCEP Agreement] --> "5% Agreement Tax Rate", "HS Code X" --[Must be met at 'Shanghai Yangshan Port'] --> "Regulatory Condition B". The structure of the knowledge graph enables the explicit expression and efficient querying of complex policy relationships.

[0011] The system provided by this invention further includes: A feature extraction system is configured to receive product description information input by a user and convert it into a unified, high-dimensional feature vector that comprehensively represents the product's intrinsic attributes. Further, the feature extraction system includes: a deep semantic parsing module, a visual feature extraction module, and a feature fusion and alignment module.

[0012] The deep semantic parsing module receives natural language descriptive text about the product input by the user. This module first preprocesses the input text, including word segmentation, part-of-speech tagging, and dependency parsing. Then, a named entity recognition model, finely tuned from massive amounts of customs declaration data, is invoked to accurately identify and extract predefined product attribute entities from the text, specifically including: core product name, material composition and its content, processing technology, specifications, functions, and brand. The extracted entities form a structured list of product attributes in key-value pairs. Next, the module serializes this structured attribute list into a descriptive text and inputs it into a sentence embedding model based on the Transformer architecture. This model encodes the text into a 1024-dimensional text semantic vector.

[0013] The visual feature extraction module receives product photos or design drawings uploaded by the user. This module employs a deep convolutional neural network based on ResNet, specifically an architecture with 152 convolutional layers. This network has undergone transfer learning and fine-tuning on an internal product image dataset containing over ten million images covering all customs product category sections. During feature extraction, the module resizes the input product image to 224x224 pixels and feeds it into the network. It then extracts the output of the network's last global average pooling layer to generate a 2048-dimensional visual feature vector.

[0014] The feature fusion and alignment module receives a 1024-dimensional text semantic vector and a 2048-dimensional visual feature vector generated by the aforementioned module. This module first maps the dimension of the text semantic vector from 1024 to 2048 dimensions using a linear projection layer to align it with the dimension of the visual feature vector. Then, it performs element-wise addition on the dimension-aligned text vector and the visual feature vector, and normalizes the result using the L2 norm. Finally, it passes the fused vector through a multilayer perceptron (MLP) containing two fully connected layers and a GeLU activation function to generate a unified product feature vector with a dimension of 4096 that simultaneously represents the product's textual description and visual appearance.

[0015] The system provided by this invention further includes: A classification candidate generation system is disclosed, the core of which is a generative large language model with deep domain adaptation and dynamic policy awareness capabilities. This subsystem receives the unified commodity feature vector and, in conjunction with real-time policy information retrieved from the knowledge graph, generates a candidate list containing multiple potential customs codes and their confidence levels. Further, the classification candidate generation system includes: an information retrieval module, a hybrid expert (MoE) model inference core, and a candidate output module.

[0016] The information retrieval module, upon receiving a unified product feature vector, parses the core product name, material, and other key information contained within it, and uses this as query keywords to initiate a query request to the trade knowledge graph database. This query aims to retrieve all currently effective policies, regulations, tax rates, and regulatory conditions directly related to the product. For example, if the product description includes "country of origin: Vietnam," the module will focus on retrieving RCEP agreement clauses related to Vietnam. All retrieved relevant information is integrated into a context prompt text.

[0017] The core of the Hybrid Expert (MoE) model inference is based on a Transformer model with 7 billion parameters, containing only a decoder. This model has 32 Transformer layers, a hidden layer dimension of 4096, and employs a Grouped-Query Attention mechanism to optimize inference performance. Its core feature is that in every alternate Transformer layer, the standard feed-forward network is replaced by a Hybrid Expert module. Each Hybrid Expert module consists of a gating network and multiple independent expert sub-networks. The gating network is a lightweight linear classifier responsible for calculating a probability distribution pointing to multiple experts based on the hidden state of the currently processed token, and selecting the two experts with the highest activation scores to process the token. These expert sub-networks are specifically trained during the model fine-tuning phase to handle knowledge from different domains and are divided into three expert groups: The first expert group, comprising multiple sub-networks of experts, is known as the "International Trade Agreements Expert Group." This group of experts fine-tunes a corpus containing the full texts, explanatory notes, and relevant case law of major global trade agreements (such as RCEP, CPTPP, and USMCA), specializing in understanding and applying rules of origin, tariff schedules, and agreement-specific terminology.

[0018] The second expert group, comprising multiple sub-networks of experts, is known as the "General Rules of Classification Expert Group." This group of experts fine-tunes a corpus containing the World Customs Organization's Harmonized System Commodity Descriptions and Coding Notes, classification guidelines issued by national customs authorities, and historical classification decisions. They specialize in understanding and applying the Harmonized System's General Rules of Classification, category notes, chapter notes, and subheading notes.

[0019] The third expert group, comprising multiple expert sub-networks, is known as the "Port Practices and Special Supervision Expert Group." The experts in this group are trained using data collected by the real-time data stream subscription and processing module, which gathers specific operating procedures, inspection and quarantine requirements, and temporary control notices for each port. They specialize in handling special regulatory requirements and non-tariff barrier information for specific ports.

[0020] During inference, as the input sequence containing product feature vectors and contextual cues flows through the model, the gating network implements token-level dynamic routing. For example, tokens describing the material of a product in the sequence are routed to the "General Rules for the Classification of Goods and Services" for processing, while tokens describing trade terms are routed to the "International Trade Agreements Expert Group" and the "Port Practices and Special Regulation Expert Group" for collaborative processing.

[0021] The structured candidate output module processes the output of the hybrid expert model's inference core. The final layer of the model is designed to generate an output header in a specific JSON format. This output header converts the natural language classification reasons generated by the model into a structured list containing no more than five of the most likely customs code candidates. For each candidate code, a confidence score calculated from the model's internal attention score is attached, along with a brief explanation of the classification basis citing specific regulations or annotations.

[0022] The system provided by this invention further includes: A Pareto optimal decision-making system receives a list of customs code candidates output by the classification candidate generation subsystem, calculates the performance of each candidate code on multiple preset business objectives, and finally determines a non-dominated solution set (i.e., the Pareto optimal frontier). Further, the Pareto optimal decision-making system includes: a cost and risk quantification module, and a ranking algorithm execution module.

[0023] The cost and risk quantification module calculates a target vector V=[C, R, T] consisting of three core indicators for each customs code in the candidate list.

[0024] Indicator C, or "Estimated Total Cost," is calculated as follows: C = (FOB price + International freight + Insurance) × (1 + (Base Most Favored Nation (MFN) tariff rate × (1 - Agreement preferential exemption rate) + Importing country VAT rate + Consumption tax rate)) + Port fixed handling fees. All variables in this formula, such as the various tariff rates corresponding to different codes, freight rates for specific routes, and handling fees at the target port, are obtained in real-time by querying the aforementioned trade knowledge graph.

[0025] The indicator R, or "Compliance Risk Index," is calculated using a weighted scoring model: R = w1 × H + w2 × P + w3 × U. Here, H represents the historical controversy level of the code, derived by counting the number of classification disputes associated with that code in the knowledge graph; P represents the regulatory complexity involved in the code, quantified by the number of valid regulatory edges connected to the code node; and U represents recent policy uncertainty, a binary variable: U = 1 if the relevant regulations have changed within the past 30 days, otherwise U = 0. The weighting coefficients w1, w2, and w3 are pre-defined empirical values, and their sum is 1.

[0026] The metric T, or "estimated customs clearance time," is not a fixed value but is predicted using a pre-trained Gradient Boosting Decision Tree model. The model's input features include candidate customs codes, import / export countries, destination port, carrier, and declaration month. Its output is the 90th percentile estimated customs clearance time (in hours) under those conditions. The model is incrementally trained daily using the latest customs clearance records to maintain its predictive accuracy.

[0027] The sorting algorithm execution module receives the set of target vectors corresponding to all candidate codes. This module implements an optimization process based on the Non-Dominated Sorting Genetic Algorithm (NSGA-II). First, all candidate solutions are considered as the initial population. Then, the algorithm performs a fast non-dominated sort on the population, stratifying it into different Pareto levels. Next, solutions within the same level are selected by calculating the crowding degree of each solution. Offspring populations are generated through simulated binary crossover and polynomial mutation operations. After merging the parent and offspring generations, non-dominated sorting and crowding degree calculation are performed again to select a new parent population. This process terminates after iterating through a pre-set number of generations (e.g., 100 generations). Finally, the algorithm outputs all solutions in the first Pareto level (Pareto Front 1). These solutions collectively constitute an "optimal customs declaration strategy set," where improvement of any strategy necessarily comes at the expense of at least one other objective. The system ultimately presents this set of strategies to the user in the form of a visual chart, and by default recommends the solution that is closest to the Utopia point (i.e., the combination point of the theoretical optimal values ​​of each objective) in the normalized objective space as the "equilibrium optimal suggestion".

[0028] This invention also provides a method for intelligent matching of commodity names for export customs declarations based on a large model. The method is executed based on the aforementioned system and includes the following steps: Step S100: The system receives export commodity information submitted by the user through its user interface. The information includes at least a natural language description of the commodity and optionally a physical image of the commodity.

[0029] Step S200: The real-time data acquisition system runs continuously in the background, constantly collecting trade-related policies, regulations, tax rates, and port data from external data sources, and updating the trade knowledge graph database in real time to ensure that the information environment on which the system makes decisions is up-to-date.

[0030] Step S300: The feature extraction system receives the product information input in step S100, calls its internal text deep semantic parsing module and image visual feature extraction module to process the text and image respectively, and generates the unified product feature vector through the feature fusion and alignment module.

[0031] Step S400: The classification candidate generation system receives the unified commodity feature vector generated in step S300. Its internal information retrieval module first queries the trade knowledge graph based on the commodity features to obtain relevant real-time policy context. Subsequently, the commodity feature vector and policy context are jointly input into the hybrid expert model inference core; the model uses its dynamic expert routing mechanism for deep reasoning, and finally, the structured candidate output module generates a structured list containing multiple customs code candidates and their confidence levels.

[0032] Step S500: The Pareto optimal decision-making subsystem based on multi-objective optimization receives the candidate list generated in step S400. Its internal cost and risk quantification module encodes each candidate in the list and, by querying the knowledge graph and calling the prediction model, calculates the corresponding target vector consisting of the estimated overall cost (C), compliance risk index (R), and estimated clearance time (T).

[0033] Step S600: The sorting algorithm execution module receives the target vector set of all candidate codes and executes the non-dominated sorting algorithm to calculate the non-dominated solution set that constitutes the Pareto optimal front.

[0034] Step S700: The system presents the set of non-dominated solutions calculated in step S600 to the user. This presentation involves visually displaying the distribution of each solution across cost, risk, and time in a three-dimensional coordinate system, allowing the user to intuitively select based on their business preferences (e.g., cost priority, time priority, or risk aversion) on the Pareto front. The system also provides a default recommended solution based on equilibrium considerations. This concludes a complete intelligent matching and optimization decision-making process.

[0035] Beneficial effects: The intelligent matching system and method for export customs declaration commodity names based on a large model can, on the basis of a comprehensive perception of the dynamic trade environment, use advanced artificial intelligence models for deep semantic understanding and professional reasoning. It can not only generate highly accurate classification candidates, but also perform quantitative evaluation and optimization from multiple business dimensions such as cost, risk, and timeliness. It provides users with data-driven, globally optimal customs declaration decision support, thereby significantly improving the accuracy, economy and efficiency of customs declaration. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention.

[0037] Figure 2 for Figure 1 Block diagram of the real-time data acquisition system.

[0038] Figure 3 for Figure 1 Block diagram of the feature extraction system.

[0039] Figure 4 for Figure 1 Block diagram of the candidate generation system for classifying data.

[0040] Figure 5 for Figure 1 Block diagram of the Pareto optimal decision system.

[0041] Figure 6 This is a flowchart illustrating the method of the present invention.

[0042] The attached diagram is labeled as follows: 10. Real-time acquisition system; 11. Static data retrieval module; 12. Real-time data stream subscription and processing module; 13. Knowledge graph construction module; 20. Feature extraction system; 21. Deep semantic parsing module; 22. Visual feature extraction module; 23. Feature fusion and alignment module; 30. Classification candidate generation system; 31. Information retrieval module; 32. Hybrid expert (MoE) model inference core; 33. Structured candidate output module; 40. Pareto optimal decision system; 41. Cost and risk quantification module; 42. Ranking algorithm execution module. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0044] Please see Figures 1 to 6This invention provides an intelligent matching system and method for export customs declaration commodity names based on a large model. The core design idea of ​​this system is to construct a closed-loop automated framework that integrates data perception, multimodal feature understanding, domain expert knowledge reasoning, and multi-objective optimization decision-making to address the complexity and dynamism of commodity classification in modern international trade.

[0045] Reference Figure 1 The system is deployed on a server cluster with distributed data storage and massively parallel computing capabilities. This cluster can physically consist of high-performance computing instances provided by a cloud service provider or hosted in a self-built enterprise data center. Logically, the system is divided into four collaborative and highly cohesive subsystems: a real-time acquisition system 10, a feature extraction system 20, a classification candidate generation system 30, and a Pareto optimal decision-making system 40. These four subsystems exchange data through well-defined application programming interfaces (APIs), forming a complete data processing and analysis pipeline from environmental perception to final decision-making.

[0046] The internal structure and working mechanism of each subsystem will be explained in detail below.

[0047] First, refer to Figure 2 The real-time data acquisition system 10 is the perception foundation of the entire system. Its function is to build and dynamically maintain a comprehensive, accurate, and real-time global trade knowledge environment. This subsystem 10 consists of three parts: a static data retrieval module 11, a real-time data stream subscription and processing module 12, and a knowledge graph construction module 13.

[0048] Specifically, the static data retrieval module 11 periodically retrieves relatively stable but regularly updated official data. Internally, it is implemented as a distributed crawler cluster based on the Scrapy framework, deployed on multiple virtual private servers in different geographical locations to avoid blocking due to excessive access frequency from a single IP address. The module's task configuration file pre-sets over two hundred Uniform Resource Locators (URLs) for target data sources, covering the official websites of major trading partners such as the World Customs Organization (WCO), the General Administration of Customs of China, the US International Trade Commission (USITC), and the Directorate-General for Taxation and Customs Union of the European Commission. The task scheduling system uses the APScheduler library, with a default execution cycle of once every 12 hours. For structured data published in the target data source as HTML tables or directly downloadable CSV / XLSX files, such as Harmonized Customs System (HS) tariff tables from various countries, the module directly uses the Pandas library for parsing, accurately extracting key fields such as customs codes to 10 or 12 digits, Chinese and English product name descriptions, legal first and second units of measurement, Most Favored Nation (MFN) tariff rates, tariff rates under various Free Trade Agreements (FTAs), export tax rebate rates, import VAT rates, consumption tax rates, and regulatory condition codes. For structured data, the module verifies the completeness and consistency of fields using a preset rule base, removing invalid or redundant fields before generating a standardized data format.

[0049] Furthermore, for unstructured or semi-structured documents published as PDFs or scanned images, such as compilations of customs classification decisions, detailed rules of origin, or temporary policy announcements, the module initiates a more complex processing flow. This flow first invokes a document layout analysis model based on the Mask R-CNN architecture. This model, pre-trained on a dataset containing hundreds of thousands of labeled trade document images, can identify and segment text paragraphs, tables, seals, and figures in the document with an accuracy of over 98.5%. Subsequently, for the identified text regions, the system invokes a Tesseract OCR engine instance optimized for financial and legal documents to convert the imaged text into a UTF-8 encoded text stream. Finally, this text stream is fed into a Named Entity Recognition (NER) model based on the BERT-Bi-LSTM-CRF architecture. This NER model predefines over thirty entity labels, such as POLICY_ID, EFFECTIVE_DATE, EXPIRATION_DATE, APPLICABLE_PRODUCT_SCOPE, INVOLVED_COUNTRY, and REGULATORY_CONDITION_CODE. Through fine-tuning on a large amount of manually annotated policy and regulatory text, it achieves high-precision automated extraction of key information. For unstructured data, optical character recognition technology is used to extract key constraints from regulatory documents, and natural language processing technology is combined to generate semi-structured data.

[0050] Furthermore, the standardized structured and semi-structured data are correlated with historical data to construct a unified commodity feature vector index table, while the real-time updated dynamic policy information is synchronized to the dynamic knowledge graph construction module 13.

[0051] In parallel, the real-time data stream subscription and processing module 12 is responsible for capturing highly timely dynamic information. This module establishes persistent subscription connections with data service terminals of over eighty major global commercial information providers and key ports (such as Shanghai Port, Ningbo-Zhoushan Port, and Los Angeles Port) via standard WebSocket or RESTful API protocols. To ensure high throughput and reliability of data reception, the core of this module is a consumer group deployed on an Apache Kafka cluster. Each external data source corresponds to an independent Kafka topic. For example, the topic `fx.rates.realtime` receives real-time exchange rate data released by the International Monetary Fund, the topic `shipping.freight.baltic_dry_index` receives fluctuation data of the Baltic Dry Index, and the topic `port.status.shanghai.yangshan` receives real-time updates on congestion, berth availability, and customs inspection pressure at specific terminals in Shanghai Yangshan Port. Whenever new data is published, the data provider's push service sends a message to the corresponding Kafka topic. The module's consumer instances then immediately pull these messages. Each message is encapsulated in a predefined JSON format, which mandates the inclusion of metadata fields such as payload, source_identifier, data_type, and capture_timestamp_ms (capture timestamp accurate to milliseconds), ensuring data traceability and timeliness.

[0052] All data acquired from the static data retrieval module 11 and the real-time data stream subscription and processing module 12 is ultimately fed into the knowledge graph construction module 13. This module is the endpoint of the data processing pipeline and the starting point for knowledge accumulation. Its workflow begins with rigorous cleaning and normalization of the input data. The data cleaning steps include: using content-based hashing algorithms (such as MD5 or SHA-256) to detect and remove completely duplicate data records; applying fuzzy matching algorithms based on edit distance (Levenshtein distance) and Jaro-Winkler similarity to align and unify entity names, for example, mapping different expressions such as "People's Republic of China", "China", and "PRC" to a unique internal entity ID country:CN; and using regular expressions to perform format validation and standardization transformation on numerical data (such as tax rate percentages and currency amounts) and time-based data (such as dates and timestamps). After cleaning and normalization, the module injects this clean and standardized data into a dynamic trade knowledge graph implemented using the Neo4j graph database. The schema of this knowledge graph is meticulously designed, with core node types including Commodity, HS Code, Country, Port, Trade Agreement, Policy Document, and Regulation. Nodes are connected by edges with direction and attributes to represent rich semantic relationships. For example, a Commodity node can be connected to an HS Code node via an edge named CLASSIFIED_AS; an HS Code node can be connected to a Country node via an HAS_TARIFF_IN edge with rate and condition attributes; a Trade Agreement node (such as RCEP) can be connected to multiple Country nodes via INCLUDES_MEMBER edges and to specific Regulation nodes via DEFINES_RULE edges. This graph structure makes complex relationships that were originally implicit in multiple isolated documents and databases explicit, providing highly efficient query support for subsequent context-aware retrieval and complex reasoning. Updates to the knowledge graph are transactional, ensuring data consistency and integrity.

[0053] Furthermore, next, refer to Figure 3The feature extraction system 20 will be described below. The mission of this subsystem 20 is to receive raw, unstructured product information input by the user and transform it into a machine-understandable, high-information-density, standardized mathematical representation—a unified product feature vector. This subsystem 20 consists of a deep semantic parsing module 21, a visual feature extraction module 22, and a feature fusion and alignment module 23.

[0054] Specifically, when a user enters a natural language description of a product on the system's front-end interface, such as "Women's long-sleeved T-shirt, 100% pure cotton knit fabric, round neck, white, brand ABC, for everyday wear," the description text is first fed into the deep semantic parsing module 21. This module executes a sophisticated natural language processing (NLP) process. The first step is text preprocessing, including Chinese word segmentation using the Jieba word segmentation library, part-of-speech tagging using Peking University's part-of-speech tagging set, and dependency parsing using graph-based algorithms to reveal the grammatical relationships between words. The second step is key attribute entity extraction. The module calls a named entity recognition (NER) model that has been domain-adaptively fine-tuned on massive amounts of real customs declaration data in the "product name and specifications" section. This model is based on the RoBERTa-wwm-ext architecture. It can accurately identify and extract a series of predefined product attribute entities from the text and organize them in a structured form as key-value pairs. For the example above, the extraction result might be: {Core Product Name: "T-shirt", Gender / Purpose: "Women's", Sleeve Length: "Long Sleeve", Material Composition: {"Cotton": 1.0}, Processing Technology: "Knitted", Collar Type: "Crew Neck", Color: "White", Brand: "ABC", Declaration Element_Purpose: "Daily Wear"}. The third step, to preserve the semantic relationships of these attributes and generate a global text representation, involves reserializing this structured list of key-value pairs into a standardized descriptive text, such as: "The core product name is T-shirt, the gender / purpose is women's, the sleeve length is long sleeve, the material is 100% cotton, and the processing technology is knitted." This serialized text is then input into a Sentence-BERT-based sentence embedding model. This model is specially trained to map sentences with similar meanings but different expressions to neighboring positions in the vector space. The model outputs a dense vector of dimension 1024, i.e., a text semantic vector, which encapsulates all the core semantic information in the product description text.

[0055] Meanwhile, if users upload actual photos or design drawings of products, this image data is passed to the visual feature extraction module 22. The core of this module is a deep convolutional neural network (CNN), specifically a ResNet-152 model pre-trained on the ImageNet dataset. To better adapt it to the specific task of product classification, we performed transfer learning on an internally constructed dataset of over 20 million classified product images covering all 97 product chapters of the Harmonized System. The fine-tuning process employed a two-stage strategy: first, all layers of the ResNet-1e52 except the last fully connected layer were frozen, and only the classification head was trained to adapt to the new product categories; then, all layers were unfrozen with a small learning rate (e.g., 1e-5) for end-to-end fine-tuning. During actual feature extraction, input product images of arbitrary sizes were first preprocessed, including cropping the central region and scaling to a standard size of 224x224 pixels, followed by normalization. The processed image is fed into a fine-tuned ResNet-152 network, but instead of performing the final classification, the output of the penultimate layer, the Global Average Pooling layer, is extracted. This output is a 2048-dimensional vector, the visual feature vector, which captures visual features such as the shape, texture, color, and overall appearance of the product at the pixel level.

[0056] Finally, the 1024-dimensional text semantic vector and 2048-dimensional visual feature vector generated by the aforementioned two modules are fed into the feature fusion and alignment module 23. To fuse these two features from different sources and with different dimensions, this module first uses a learnable linear projection layer to linearly transform the 1024-dimensional text semantic vector to 2048 dimensions, ensuring consistency with the dimension of the visual feature vector. The weight matrix of this projection layer (1024x2048 dimensions) is trained on a dataset containing paired product text and images using contrastive learning, aiming to make matching text-image pairs closer in the projected space and mismatched pairs farther apart. After dimension alignment, the module performs element-wise addition and fusion of the projected text vector and the original visual feature vector to obtain a preliminary fused vector. To enhance the model's stability and generalization ability, this fused vector undergoes L2 norm normalization. Finally, the normalized vector is fed into a multilayer perceptron (MLP) consisting of two fully connected layers for deep fusion and nonlinear transformation. The hidden layer of this MLP has a dimension of 4096, and the activation function is a Gaussian error linear unit (GeLU). The final output 4096-dimensional vector is a unified product feature vector that can comprehensively and uniformly represent both the textual and visual attributes of the product. If the user does not provide an image, the visual feature vector is replaced with a zero vector in this step.

[0057] Furthermore, referring to Figure 4 The classification candidate generation system 30 will be described below. This subsystem 30 is the reasoning core of this invention. It receives standardized, unified commodity feature vectors and simulates the thought process of an experienced customs brokerage team to generate a candidate list containing multiple high-quality classification suggestions. This subsystem 30 consists of an information retrieval module 31, a hybrid expert (MoE) model reasoning core 32, and a structured candidate output module 33.

[0058] In practice, the information retrieval module 31 first extracts core keywords, such as product name, material, country of origin, and country of destination, from the source information of the received unified commodity feature vector (i.e., the structured attributes extracted by the deep semantic parsing module 21). Then, based on these keywords, it dynamically constructs a Cypher query statement and initiates a query request to the Neo4j knowledge graph database maintained by the knowledge graph construction module 13. For example, for a "cotton T-shirt" with "Vietnam" as its country of origin, the query statement will focus on retrieving nodes and their relationships related to the nodes Commodity (name='T-shirt'), Material (name='cotton'), and Country (name='Vietnam'), such as HSCode, TradeAgreement (name='RCEP'), Regulation, and TariffRate. The query results, namely all currently effective policies, regulations, tax rate information, regulatory conditions, and historical classification precedents directly related to the current commodity, are summarized and formatted into a well-structured natural language text, which is called the context prompt.

[0059] Subsequently, the unified product feature vector (converted into a text sequence in a specific format) is concatenated with the contextual cue text to form a complete input sequence, which is then fed into the Hybrid Expert (MoE) model inference core 32. The underlying model of this core is a deeply optimized autoregressive Transformer large language model with 7 billion parameters. Its architecture comprises 32 Transformer decoder layers, with a hidden layer dimension of 4096. The attention mechanism employs Grouped-Query Attention (GQA) to significantly reduce computational load and memory usage during inference while maintaining high-quality output. Its most crucial innovation lies in replacing the standard feedforward neural network (FFN) with a Hybrid Expert (MoE) module after every other Transformer layer, totaling 16 MoE layers.

[0060] In this embodiment, each MoE module consists of a lightweight gating network and 48 specialized expert sub-networks. The gating network is essentially a simple linear classifier; its input is the hidden state of the currently processed token, and its output is a probability distribution vector of length 48, normalized by the Softmax function. This vector represents the "weight" or "fitness" of assigning the current token to each expert. Based on this distribution, the system selects the two experts with the highest probabilities (Top-2 Gatings) to process the hidden state of the token in parallel, and the final output is the weighted sum of the outputs of these two experts. This mechanism allows the model to dynamically and intelligently activate a small subset of the most relevant expert networks based on the token content when processing a sequence, thereby achieving a knowledge capacity and reasoning ability far exceeding the parameter size with a computational cost far less than activating the entire model's parameters.

[0061] These 48 expert subnetworks are not homogeneous; rather, they were purposefully trained during the model fine-tuning phase to become "experts" in different professional fields and organized into three expert groups. The first expert group, the "International Trade Agreements Expert Group," consists of 16 expert subnetworks. Its training corpus includes the legal texts, official interpretations, rules of origin certification guidelines, and relevant international court precedents of all major trade agreements (such as RCEP, CPTPP, and USMCA). This group specializes in interpreting complex tariff reduction commitment schedules and determining whether goods meet the rules of origin criteria of specific agreements. The second expert group, the "General Rules for the Classification of Goods Expert Group," consists of 24 expert subnetworks and is the largest expert group. It was fine-tuned on a massive corpus containing the entire WCO Harmonized System Commentary, customs classification guidelines from over twenty major countries, and millions of historical classification decisions. This group's experts have achieved a level of understanding and application of the six General Rules for the Harmonized System, as well as various commentaries, chapter notes, and subheading notes, reaching the level of top human experts. The third expert group, the "Port Practices and Special Supervision Expert Group," consists of eight expert sub-networks. Its training data primarily comes from the real-time data stream subscription and processing module 12, which collects specific operating procedures, phytosanitary requirements, technical barriers to trade (TBT) notifications, and various temporary trade control measures for each port. This group of experts specializes in handling non-tariff barriers and compliance requirements related to specific import and export port operations. During inference, this architecture enables refined dynamic knowledge retrieval. For example, a token describing "100% cotton" in the input sequence will likely be routed by the gating network to an expert in the "General Rules for Commodity Classification Expert Group" for processing; while a token describing "originating in Vietnam" may simultaneously activate experts in both the "International Trade Agreements Expert Group" and the "Port Practices and Special Supervision Expert Group" to collaboratively assess its eligibility under the RCEP agreement and potential special quarantine requirements at the port of destination.

[0062] Furthermore, the output of the hybrid expert model inference core 32 is a piece of natural language text containing classification logic. This text is then passed to the structured candidate output module 33. This module parses the model's free text output and converts it into a machine-readable, standardized data structure. To this end, the final layer of the large language model is specially fine-tuned to generate output according to a predefined JSON schema. This module ensures that the final output is a JSON array, where each object represents a Customs Code (HS Code) candidate scheme. Each object contains the following fields: hs_code (10-digit Customs Code), description (corresponding Chinese product name), confidence_score (a floating-point number between 0 and 1, calculated from the model's internal attention score and output probability, representing the model's confidence in the classification suggestion), and justification (a concise explanation of the classification basis, usually citing specific clauses from the Harmonized System Notes or specific regulations). This module is typically configured to output the top 5 candidates with the highest confidence scores to form a comprehensive and high-quality candidate list.

[0063] Finally, refer to Figure 5 The Pareto optimal decision-making system 40 is described below. This subsystem 40 receives a structured candidate list output by the classification candidate generation subsystem 30. Its goal is to move from a purely "compliance probability" level to a "business optimality" level that comprehensively considers cost, risk, and timeliness, providing users with a set of non-dominated decisions. This subsystem 40 consists of a cost and risk quantification module 41 and a ranking algorithm execution module 42.

[0064] Specifically, the cost and risk quantification module 41 calculates a target vector V=[C, R, T] consisting of three key business indicators for each customs code in the candidate list.

[0065] The first indicator, C, represents the "Estimated Overall Cost." This module calculates this value by making precise queries to the knowledge graph. The formula is: C = (FOB price of goods × (1 + importing country's VAT rate + consumption tax rate)) + Customs duty + International freight and insurance + Port charges. Where, Customs duty = CIF price of goods × Comprehensive customs duty rate. The comprehensive customs duty rate is intelligently selected based on the current goods and trading partners (e.g., the lower of the Most Favored Nation (MFN) rate and the specific agreement rate). All variables in the formula, such as the tax rates corresponding to different codes, real-time shipping and insurance rates from the port of origin to the port of destination, and standard operating charges at the port of destination, are obtained through real-time queries to the trade knowledge graph.

[0066] The second indicator, R, represents the "Compliance Risk Index." This is a composite assessment model, calculated as: R = 0.5 × H + 0.3 × P + 0.2 × U. H (Historical Dispute Level) is quantified and normalized by querying the number of "classification disputes" or "customs inquiries" event nodes associated with the HS code in the knowledge graph. P (Regulatory Complexity) is measured by the number of valid Regulation nodes directly or indirectly connected to the HS code node; a higher number indicates more complex regulatory constraints. U (Recent Policy Uncertainty) is a Boolean variable. If any regulations related to the code have been added, deleted, or revised within the past 30 days (by querying the last_modified_date attribute of relevant nodes in the knowledge graph), then U = 1; otherwise, U = 0. The weighting coefficients of 0.5, 0.3, and 0.2 are empirical values ​​derived from extensive historical case analysis and can be adjusted by the system administrator.

[0067] The third metric, T, represents the "estimated customs clearance time." Since customs clearance time is influenced by numerous complex factors and is difficult to calculate using a deterministic formula, this module uses a pre-trained Gradient Boosting Decision Tree (GBDT) model for prediction. This model is implemented using the LightGBM framework and has over 50 input features, including: customs code, import / export countries, port of origin and port of destination, carrier credit rating, customs broker qualifications, declared amount, declared month, and whether special regulatory conditions are involved. The model is incrementally trained daily using the latest anonymized customs clearance records obtained from the customs system interface to continuously optimize its prediction accuracy. The model's output is the 90th percentile predicted time (in hours) for the shipment to complete the customs release process, representing a relatively conservative and reliable timeliness estimate.

[0068] Once the target vectors [C, R, T] of all candidate codes are computed, they are passed as a set to the sorting algorithm execution module 42. The goal of this module is to find a set of Pareto optimal solutions among these three conflicting objectives (typically, low cost may imply high risk, and low risk may require longer preparation time). This module implements a classic, efficient multi-objective optimization algorithm—the Non-Dominated Sorting Genetic Algorithm with Elite Strategy (NSGA-II). The algorithm proceeds as follows: all candidate solutions (i.e., candidate HS codes) are treated as the initial population. The algorithm then performs a fast non-dominated sort on the population, distributing the solutions into different Pareto fronts. The solutions in Front 1 are the best and are not dominated by any other solutions. Next, the crowding distance is calculated for each solution, a value that measures the density of solutions around it, used to maintain solution diversity within the same front. Subsequently, superior individuals are selected from the current population to generate the offspring population through simulated binary crossover and polynomial mutation operations. The parent and offspring populations are merged, and non-dominated sorting and crowding calculations are performed again to select a new population for the next generation. This iterative process continues for a preset number of generations (e.g., 100 generations) or until the population converges. Finally, the algorithm outputs all solutions in the Pareto First Front 1. This set of solutions is the "optimal customs clearance strategy set," where any improvement in any solution (e.g., reducing costs) necessarily leads to the deterioration of at least one other objective (e.g., increasing risk or extending customs clearance time).

[0069] This invention also provides a corresponding intelligent matching method for export customs declaration commodity names based on a large model, which is executed based on the aforementioned system. (Refer to...) Figure 6 The specific process of this method includes the following steps: Step S100: The user submits the export commodity information to be declared through the system's front-end web interface or API interface. Required information is a natural language description of the commodity, while optional information includes photographs of the commodity, design drawings, chemical composition reports, etc.

[0070] Step S200: This step is a background, continuous task. The real-time data acquisition system 10 operates 24 / 7 without user intervention. Its internal static data retrieval module 11 and real-time data stream subscription and processing module 12 continuously acquire the latest trade data from various global data sources, and through the knowledge graph construction module 13, update and expand the trade knowledge graph database in real time to ensure that the external environmental information used by the system for any decision is the latest and most accurate.

[0071] Step S300: The feature extraction system 20 receives the product information submitted by the user in step S100. If the information contains both text and images, the deep semantic parsing module 21 and the visual feature extraction module 22 are called respectively for processing to generate a text semantic vector and a visual feature vector. Subsequently, the feature fusion and alignment module 23 fuses these two vectors to generate a unified product feature vector with a dimension of 4096.

[0072] Step S400: The classification candidate generation system 30 receives the unified product feature vector generated in step S300. First, the information retrieval module 31 queries the knowledge graph based on product attributes to retrieve relevant contextual information. Then, the hybrid expert (MoE) model inference core 32 takes the product features and contextual information as input and uses its dynamic expert routing mechanism for deep semantic understanding and professional inference. Finally, the structured candidate output module 33 formats the model's inference results to generate a JSON list containing multiple (usually no more than 5) high-confidence customs code candidates.

[0073] Step S500: The Pareto optimal decision system 40 receives the candidate list generated in step S400. Its internal cost and risk quantification module 41 encodes each candidate in the list and, by querying the knowledge graph and calling the GBDT prediction model, accurately calculates the target values ​​for the three dimensions of estimated overall cost (C), compliance risk index (R), and estimated customs clearance time (T).

[0074] Step S600: The sorting algorithm execution module 42 receives all candidate codes and their corresponding [C, R, T] target vector sets, and executes the NSGA-II algorithm. After a series of iterative operations such as non-dominated sorting, crowding calculation, crossover, and mutation, the non-dominated solution set constituting the Pareto optimal front is finally calculated.

[0075] Step S700: The system presents the set of non-dominated solutions (i.e., the set of optimal customs declaration strategies) calculated in step S600 to the user through the front-end interface. A typical presentation method is to visualize each optimal solution as a point in an interactive 3D scatter plot, with the three coordinate axes representing cost, risk, and time, respectively. Users can rotate and zoom the chart to intuitively observe the trade-offs between various options and make their own choices on the Pareto front based on their business priorities (e.g., for urgent orders, the option with the smallest coordinate value on the time axis can be selected). In addition, the system also calculates the Euclidean distance of each solution to the utopian point (i.e., the combination point of the theoretically achievable optimal values ​​of the three objectives) in the normalized objective space, and marks the solution with the closest distance as the "equilibrium optimal suggestion" as the default recommended solution. At this point, a complete intelligent matching process from product description to optimal customs declaration strategy generation is concluded.

[0076] Example: To verify the technical effect of the present invention, a specific application example is provided.

[0077] A garment export company located in Zhejiang, China, needs to declare a batch of garments destined for Japan to Chinese customs. The user enters the following information into the system interface: Product description text: "100% cotton women's knit T-shirt, short sleeve, crew neck, white, no brand, regular order." Product image: The user uploaded a clear photo of a white short-sleeved T-shirt.

[0078] Trade information: Origin: China (CN), Destination: Japan (JP), Port of Destination: Tokyo (TYO).

[0079] The system execution flow is as follows: The knowledge graph of the real-time data acquisition system 10 already includes the latest tariff rates under the China-Japan RCEP agreement, the inspection and quarantine requirements of Japanese customs for clothing products, and the recent average customs clearance efficiency data of the Port of Tokyo.

[0080] The feature extraction system 20 receives the above information. The deep semantic parsing module 21 extracts the structured attributes: {core product name: "T-shirt", gender / use: "women's", material composition: {"cotton": 1.0}, processing technology: "knitted", sleeve length: "short sleeve", ...}, and generates a 1024-dimensional text vector. The visual feature extraction module 22 processes the T-shirt photo and generates a 2048-dimensional visual vector. The feature fusion and alignment module 23 fuses the two to generate a final 4096-dimensional unified product feature vector.

[0081] The classification candidate generation system 30 receives the vector. The information retrieval module 31 queries the knowledge graph to retrieve the rules of origin for textiles and tariff reduction clauses for China in the RCEP agreement. When processing the input, the hybrid expert model inference core 32's gating network routes tokens related to "cotton" and "knitwear" to the "General Rules of the Trade Act" panel, and tokens related to "export to Japan" and "RCEP" to the "International Trade Agreements Panel." After model inference, the structured candidate output module 33 generates the following candidate list: Candidate 1: HS code 6109.10.00 (cotton T-shirts, undershirts, etc.), confidence level 0.95, classification basis: directly corresponding to product name and material.

[0082] Candidate 2: HS code 6109.90.90 (knitted or crocheted T-shirts made of other textile materials), confidence level 0.65, classification basis: as an alternative if there is a dispute about the cotton content or processing method.

[0083] The Pareto optimal decision system 40 receives this list. The cost and risk quantification module 41 calculates the objective vector for the two candidate options: For 6109.10.00: C (Cost): According to the knowledge graph, this code is subject to the RCEP agreement tariff rate, and the tariff is 0. Assuming an FOB price of 10,000 yuan, freight and insurance of 500 yuan, and a Japanese consumption tax of 10%, the total cost is approximately 11,550 yuan.

[0084] R (Risk): According to the knowledge graph, this code has clear classification, few historical disputes, and low regulatory complexity. The calculated risk index R = 0.12.

[0085] T (duration): Using the GBDT model, the predicted customs clearance time is T=24 hours.

[0086] For 6109.90.90: C (Cost): The Most Favored Nation (MFN) tariff rate in Japan for this code may be higher than the RCEP agreement rate, or the agreement may not apply. Assuming an increase in tariffs of ¥800, the total cost is approximately ¥12,350.

[0087] R (Risk): This code falls under the "Other" category, which allows for greater interpretation and carries a slightly higher likelihood of customs questioning. The calculated risk index R = 0.35.

[0088] T (duration): Due to the high risk, inspection may be triggered. The GBDT model predicts that the customs clearance time will be T=48 hours.

[0089] The sorting algorithm execution module 42 receives these two solutions. Since 6109.10.00 is superior to 6109.90.90 in all three dimensions of cost, risk, and time, 6109.10.00 is the only Pareto optimal solution.

[0090] The system ultimately recommends HS code 6109.10.00 to the user and displays its estimated costs, risks, and duration.

[0091] As can be seen from the above embodiments, the intelligent matching system and method for export customs declaration commodity names based on a large model provided by the present invention can, on the basis of a comprehensive perception of the dynamic trade environment, use advanced artificial intelligence models for deep semantic understanding and professional reasoning. It can not only generate highly accurate classification candidates, but also perform quantitative evaluation and optimization from multiple business dimensions such as cost, risk, and timeliness, providing users with data-driven, globally optimal customs declaration decision support, thereby significantly improving the accuracy, economy and efficiency of customs declaration.

[0092] The above are merely specific embodiments of the present invention, but the technical features of the present invention are not limited thereto. Any simple changes, equivalent substitutions, or modifications made based on the present invention to solve essentially the same technical problems and achieve essentially the same technical effects are all covered within the protection scope of the present invention.

Claims

1. A smart matching system for export customs declaration commodity names based on a large model, which is set up in a server cluster, characterized in that, The system includes: The real-time acquisition system (10) is configured to acquire various types of data related to international trade from multiple heterogeneous external data sources, and inject the processed data into a dynamically updated trade knowledge graph (13) to build a global trade knowledge environment. The feature extraction system (20) is configured to receive product information input by the user, which includes product description text and optional product images, and process the product information into a unified product feature vector that can comprehensively represent the inherent attributes of the product; The classification candidate generation system (30) is configured to receive the unified commodity feature vector and, in conjunction with real-time strategy information retrieved from the trade knowledge graph (13), reason and generate a candidate list containing multiple potential customs codes and their confidence levels. And, the Pareto optimal decision system (40) is configured to receive the candidate list and, for each candidate customs code, quantify its performance on multiple preset business objectives and determine an optimal customs declaration strategy set consisting of non-dominated solutions by performing a multi-objective optimization algorithm.

2. The intelligent matching system for export customs declaration commodity names based on a large model as described in claim 1, characterized in that, The real-time acquisition system (10) includes: The static data retrieval module (11) is configured with a distributed web crawler agent cluster. It is configured to perform data retrieval tasks on the tariff database, commodity classification decisions, rules of origin and trade control policy announcements of the official websites of major global economies at a preset frequency. It also uses document layout analysis model, optical character recognition engine and named entity recognition model to process unstructured documents in order to automatically extract key entity information. The real-time data stream subscription and processing module (12), which establishes persistent connections with the data stream services of customs data publishing platforms at major global ports and international business information providers through an application programming interface, is configured to utilize a consumer cluster deployed on a distributed message queue system to subscribe to and receive real-time data messages regarding temporary adjustments to tariff rates, trade control measures, port congestion status, and shipping rate fluctuations; and The knowledge graph construction module (13) is configured to receive data from the static data retrieval module (11) and the real-time data stream subscription and processing module (12), perform data cleaning, entity naming unification and numerical format standardization operations, and inject the processed data into the trade knowledge graph (13).

3. The intelligent matching system for export customs declaration commodity names based on a large model as described in claim 2, characterized in that, The trade knowledge graph constructed and maintained by the knowledge graph construction module (13) is stored in an attribute graph database; The nodes of the knowledge graph represent core entities, which include at least commodities, customs codes, countries, ports, trade agreements, and legal provisions. The edges of the knowledge graph represent specific and directed relationships between the core entities. These relationships include, at a minimum, the applicable coding relationship between a commodity and a customs code, the applicable agreement tariff rate relationship between a customs code and a specific trade agreement and a country, and the regulatory conditions that must be met between a customs code and a specific port.

4. The intelligent matching system for export customs declaration commodity names based on a large model as described in claim 1, characterized in that, The feature extraction system (20) includes: The deep semantic parsing module (21) is configured to receive the natural language description text of the product input by the user, and accurately extract and organize a structured product attribute list from the text by calling the named entity recognition model fine-tuned by the customs declaration data. Then, the attribute list is serialized into descriptive text and input into a sentence embedding model based on the Transformer architecture to encode and generate a text semantic vector. The visual feature extraction module (22) is configured to receive product images uploaded by the user, and employ a deep convolutional neural network based on a residual network to extract the output of the global average pooling layer of the network to generate a visual feature vector; and The feature fusion and alignment module (23) is configured to receive the text semantic vector and the visual feature vector, align their dimensions through a linear projection layer, perform element-level addition and normalization, and finally generate the unified product feature vector through a multilayer perceptron.

5. The system according to claim 4, characterized in that, In the feature fusion and alignment module (23), the linear projection layer maps the dimension of the text semantic vector to the same dimension as the visual feature vector; after the element-level addition operation, the fused vector is normalized by the L2 norm; the multilayer perceptron contains two fully connected layers and the GeLU activation function, and the dimension of the unified commodity feature vector output by it is 4096.

6. The system according to claim 1, characterized in that, The classification candidate generation system (30) includes: The information retrieval module (31) is configured to, upon receiving the unified commodity feature vector, parse the core product name and material information contained therein as query keywords, initiate a query request to the trade knowledge graph (13) to retrieve the currently effective policies, regulations, tax rates and regulatory conditions related to the commodity, and integrate the search results into a contextual prompt text. The hybrid expert model inference core (32) is based on a Transformer model containing only a decoder, configured to receive the concatenated unified product feature vector and the contextual prompt text as input, and to perform inference using its internal dynamic expert routing mechanism; and The candidate output module (33) is configured to post-process the output of the hybrid expert model inference core (32) by converting the classification reasons generated by the model into a structured list containing no more than 5 customs code candidates, each accompanied by a confidence score and a description of the classification basis through an output header that can generate a specific JSON format.

7. The system according to claim 6, characterized in that, In the hybrid expert model inference core (32), a hybrid expert module is set in every Transformer layer. Each hybrid expert module consists of a gating network and multiple independent expert sub-networks. The gating network is configured to calculate a probability distribution pointing to multiple experts based on the hidden state of the token being processed, and select the two experts with the highest activation scores to process the token based on this distribution. The multiple expert subnetworks were divided into three expert groups, namely: A panel of experts specializing in understanding and applying rules of origin and tariff reductions, comprising multiple sub-networks of experts in international trade agreements; The Expert Group on General Rules for the Classification of Goods, which comprises multiple expert sub-networks, specializes in understanding and applying the Harmonized System's General Rules for Classification, category notes, chapter notes, and subheading notes; as well as The Port Practice and Special Regulation Expert Group is a network of multiple expert sub-networks specializing in handling specific port regulatory requirements and non-tariff barrier information.

8. The system according to claim 1, characterized in that, The Pareto optimal decision system (40) includes: The cost and risk quantification module (41) is configured to, for each customs code in the candidate list, calculate a target vector consisting of three core indicators: estimated overall cost (C), compliance risk index (R), and estimated clearance time (T), by querying the trade knowledge graph and calling a pre-trained gradient boosting decision tree model; and The sorting algorithm execution module (42) is configured to receive the target vector set corresponding to all candidate codes, and to perform non-dominated sorting, crowding calculation, crossover and mutation operations on the initial population by implementing an optimization process based on a non-dominated sorting genetic algorithm. After a preset number of iterations, it finally outputs all solutions that constitute the first Pareto level to form the optimal customs declaration strategy set.

9. The system according to claim 8, characterized in that, The cost and risk quantification module (41) quantifies the three core indicators in the following way: The estimated total cost (C) is calculated by taking into account the FOB price of the goods, international freight, insurance premiums, and various tax rates and port fixed operating fees corresponding to the customs code, which are obtained in real time by querying the trade knowledge graph. The compliance risk index (R) is calculated using a weighted scoring model, and its calculation factors include: historical controversy degree (H) obtained by statistically analyzing the number of classification dispute cases associated with the knowledge graph, regulatory complexity (P) quantified by the number of valid regulatory edges connected to the coded node, and recent policy uncertainty (U) determined based on whether relevant regulations have recently changed. The estimated customs clearance time (T) is predicted using the gradient boosting decision tree model, whose input features include the candidate customs code, import / export country, destination port, carrier, and declaration month.

10. A method for intelligent matching of export customs declaration commodity names based on a large model, the method being executed based on the system described in any one of claims 1-9, comprising the following steps: Step S100: Receive export commodity information submitted by the user through the user interface. The information includes at least a natural language description of the commodity and optionally a physical image of the commodity. Step S200: The real-time acquisition system runs continuously in the background, continuously collecting trade-related policies, regulations, tax rates and port data from external data sources, and updating the trade knowledge graph database in real time to ensure that the information environment on which the system makes decisions is up-to-date. Step S300: The feature extraction system receives the product information input in step S100, calls the text deep semantic parsing module and the image visual feature extraction module inside the feature extraction system to process the text and the image respectively, and generates the unified product feature vector through the feature fusion and alignment module; Step S400: The classification candidate generation system receives the unified commodity feature vector generated in step S300; firstly, it queries the trade knowledge graph based on the commodity features to obtain the relevant real-time policy context; Subsequently, Commodity feature vectors and policy context are jointly input into the hybrid expert model inference core; deep inference is performed using a dynamic expert routing mechanism, and finally a structured list containing multiple customs code candidates and their confidence levels is generated by the structured candidate output module; Step S500: The Pareto optimal decision system based on multi-objective optimization receives the candidate list generated in step S400; the cost and risk quantification module inside the Pareto optimal decision system encodes each candidate in the list, and calculates the target vector consisting of the estimated overall cost, compliance risk index and estimated clearance time by querying the knowledge graph and calling the prediction model. Step S600: The sorting algorithm execution module receives the target vector set of all candidate codes and executes the non-dominated sorting algorithm to calculate the non-dominated solution set that constitutes the Pareto optimal front. Step S700: Present the set of non-dominated solutions calculated in step S600 to the user; the presentation method is to visualize the distribution of each solution in three dimensions of cost, risk and time in a three-dimensional coordinate system, allowing users to make intuitive choices on the Pareto front according to their own business preferences; at the same time, a default recommendation scheme based on equilibrium considerations is provided.

Images