A customs document structured identification method and system based on multi-source fusion and large model self-checking

Abstract

The application discloses a customs document structured identification method and system based on multi-source fusion and large model self-checking, aiming at the technical defects of existing customs document identification technology, such as page evidence collapse, field semantic drift and output result credibility unable to be self-checked, a multi-source evidence fusion and double-constraint fusion generation mechanism is constructed, multi-dimensional evidence such as OCR text, document image, paging information, business prompt word and standard field template is integrated, OCR text misplacement and misalignment problems are corrected through cross-modal space alignment, field semantic drift is avoided relying on double constraints of business prompt word and standard template, through field-level, structure-level and table row-level difference comparison of two rounds of differentiated reasoning results, risk scores are quantitatively calculated, automatic evaluation of identification result credibility is realized, and the problems of traditional technology, such as customs document field identification misplacement, semantic identification deviation and uncontrollable result credibility, are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of automated customs document processing technology, and in particular to a method and system for structured identification of customs documents based on multi-source fusion and large-scale model self-checking. Background Technology

[0002] In the fields of international trade and customs clearance automation, accurate and efficient structured text extraction from various customs documents is a core foundation for improving customs clearance efficiency, reducing manual document review costs, and achieving intelligent customs supervision. Current mainstream customs document structured recognition solutions generally adopt a technical architecture of OCR layout parsing combined with general large-scale model text extraction. First, tools such as MinerU and PaddleOCR are used to parse the document layout and restore the tables. The processed plain text content is then input into a third-party large-scale model API, which, along with fixed business prompts, extracts core customs clearance fields such as invoice number, contract number, material number, quantity, amount, and country of origin.

[0003] Existing technical solutions suffer from three major underlying technical defects in practical applications, which severely restrict the accuracy and reliability of document recognition: First, there is a problem of page layout collapse. Traditional solutions compress two-dimensional document page information into a one-dimensional plain text input model without loss, ignoring the characteristic that the core fields of customs documents are highly dependent on the table rows and columns and page spatial positions. When OCR has recognition deviations such as splitting rows, misaligning columns, or truncating across pages, the large model cannot correct the field attribution by relying solely on text semantics, which can easily cause mismatches in key fields such as material number, quantity, amount, and currency.

[0004] Second, there is a problem of semantic drift in fields. Customs document templates are highly heterogeneous, and the semantics of the same field identifier differ in different document scenarios. Fixed prompt words cannot adapt to semantic changes in multiple scenarios. For example, partNo can correspond to customer part number, manufacturer part number, or product model. Fixed prompt words are prone to model migration recognition errors, causing semantic drift in fields and reducing the accuracy of structured recognition.

[0005] Third, there is a vacuum in the credibility of delivery. In real customs delivery scenarios, there is no manually labeled gold standard data. Traditional models only execute a single structured output and cannot independently determine the accuracy of the output results. There is no automatic detection mechanism when the model has problems such as missing fields, fabricated values, or field mismatches. It must rely on full manual review, which greatly increases the workload of manual review and restricts the improvement of customs clearance automation efficiency. Summary of the Invention

[0006] The main technical problem addressed by this invention is to provide a method and system for structured identification of customs documents based on multi-source fusion and large-scale model self-checking. By reconstructing the multimodal feature input space and utilizing hierarchical dual-constraint guidance, it achieves highly reliable structured document extraction in a fully localized, real-world delivery environment without gold standard certification. It avoids semantic drift of fields by relying on dual constraints of business prompts and standard templates. Through comparison of field-level, structure-level, and table row-level differences in the results of two rounds of differentiated inference, it quantifies and calculates risk scores, and achieves automatic evaluation of the credibility of the identification results. This effectively solves the problems of misidentification of customs document fields, semantic recognition deviation, and uncontrollable credibility of results in traditional technologies.

[0007] To solve the above-mentioned technical problems, one technical solution adopted by the present invention is: to provide a method for structured identification of customs documents based on multi-source fusion and large-scale model self-examination, comprising the following steps: S1. Access the customs document file to be identified, load the preset multi-source configuration dataset, which includes business prompt words, standard field templates and task type mapping data, parse the customs document file, extract the OCR text sequence, and retain the original image feature tensor of the document and the spatial geometric bounding box coordinate set of each text block. S2. Construct a multimodal dual-channel input matrix, perform language embedding processing on the OCR text and visual feature embedding processing on the original document image, and achieve the association matching of text features and visual space features through cross-modal space alignment; parse the constraint relationship between business prompt words and standard field templates, and perform hierarchical adjudication of field output permissions through conflict hierarchical resolution operators to form dual-constraint fusion recognition rules. S3. Based on a locally deployed isomorphic large model, differential prompt words are used to perform two rounds of independent reasoning, outputting a first version of the rough structured result and a second version of the verified structured result respectively; S4. The pluggable customs business verification engine is used to perform business normalization, global optimal row matching and vertical attribute completion correction on the two rounds of inference results to obtain two rounds of standardized structured results. S5. Construct a multi-dimensional risk scoring model to quantify the structural differences, semantic differences of key fields, and differences in table row alignment between the two rounds of standardized structured results, and obtain the risk score. S6. Compare the risk score with the preset global risk control threshold. If the risk score is lower than the threshold, the final structured result will be automatically output. If the risk score is higher than the threshold, the risk score will be marked as high and automatically routed to the manual review stage.

[0008] In a preferred embodiment of the present invention, in step S1, the specific method of parsing customs documents is as follows: firstly, the original text stream of the PDF format document is read; if the quality of the original text stream is not up to standard, a local lightweight multimodal visual language OCR model is called for parsing to fully capture the textual semantic features, image visual features, and spatial topological features of the document. The multi-source configuration dataset is a localized pre-configured dataset with no external API calls throughout the process. The business prompt words are used for soft constraint guidance for field positioning, and the standard field template is used to define the rigid schema boundaries and default values ​​of the structured output.

[0009] In a preferred embodiment of the present invention, in step S2, the multimodal dual-channel input matrix is ​​a concatenation matrix of text embedding features and image visual embedding features. By constructing a text query matrix and an image key matrix to calculate the attention distribution, the model associates and matches the image visual grid information when parsing text fields, correcting field attribution errors caused by OCR text splitting, misalignment, and page truncation.

[0010] In a preferred embodiment of the present invention, the cross-modal space alignment is implemented based on the model attention layer, relying on the visual continuity of the horizontal baseline of the document page to correct the misalignment of table fields and retain the topological features of the two-dimensional layout of the document.

[0011] In a preferred embodiment of the present invention, step S2 includes the following decision logic of the conflict hierarchical resolution operator: If a field exists in both the business prompt word and the standard field template, it is marked as an output field, and the field result is generated with the business prompt word as the positioning constraint and the standard template as the format boundary. If a field only exists in business prompts, it is downgraded to a semantic reference context and does not participate in the final result output; If a field exists only in the standard field template, the field content will be automatically completed based on the template default value, field type, and contextual evidence.

[0012] In a preferred embodiment of the present invention, in step S3, the differentiation mechanism of the dual-round independent reasoning is as follows: the first round of reasoning adopts a low-temperature sampling and greedy search strategy, combined with dual-constraint prompts to generate a first version of coarse structured results. The second round of inference injects adversarial reverse verification prompts, performs game-theoretic verification inference based on the same local large model weights, and generates a second version of the verification structured result. .

[0013] In a preferred embodiment of the present invention, step S4 includes the following correction operations of the pluggable customs business verification engine: completing the standardization and normalization of currency, country of origin, and packaging type through a business knowledge graph; using a global optimal row matching algorithm to achieve optimal alignment and merging of table array row items; and automatically inheriting the unified attributes of document headers to complete missing fields in table rows.

[0014] In a preferred embodiment of the present invention, the global optimal row matching algorithm is based on the bipartite graph optimal matching strategy, using material number, item number, and quantity as core anchor fields to achieve accurate alignment and redundancy merging of nested table array row items.

[0015] In a preferred embodiment of the present invention, in step S5, the calculation formula of the risk scoring model is: ; in, For nested structured data with missing differential values, For key fields, semantic distance and numerical tolerance, The row item alignment difference entropy is obtained by performing a globally optimal row matching on the two sets of table row items. The risk score is defined as follows: the higher the risk score, the greater the inconsistency between the two rounds of structured recognition results and the higher the recognition risk. These are the weighting coefficients for each dimension. Differentiated weights for different key fields for customs clearance. The first round of standardized structured results is obtained after performing business normalization and attribute completion correction on the first rough structured results. The second round of standardized structured results is obtained after business normalization and attribute completion correction of the second version of the review structured results, where k is the field index in the set of key field indexes for customs clearance; and They are respectively and The value of the kth key clearance field after business normalization; and They are respectively and The table is a collection of row items, where the subscript "line" indicates a substructure of the row item.

[0016] A structured identification system for customs documents based on multi-source fusion and large-scale model self-checking includes: The document parsing and multi-source feature acquisition module is used to access customs documents, complete OCR parsing, image feature and spatial coordinate acquisition, and load multi-source configuration datasets. The multi-source evidence fusion and dual-constraint generation module is used to construct a multimodal input matrix, perform cross-modal space alignment, and complete field output constraint adjudication through conflict hierarchical resolution. The isomorphic large model dual-round inference module is used for local deployment of large models, completing two rounds of differentiated independent inference, and outputting initial and final versions of structured data; The business rule correction module is used to normalize document fields, match table rows, and correct attribute completion. The risk assessment and risk control routing module is used to quantitatively calculate the risk score of the identification results and complete the automatic output or manual review of the routing based on the threshold.

[0017] The beneficial effects of this invention are: 1. Construct a multi-source evidence fusion cross-modal recognition mechanism. This invention abandons the one-dimensional text linear processing mode and integrates multi-dimensional evidence such as text semantics, image layout, spatial coordinates, and business templates. By correcting OCR recognition deviation through cross-modal spatial alignment, it completely solves the industry pain point of misaligned rows and columns in customs document forms and greatly improves the stability and accuracy of structured recognition.

[0018] 2. It pioneered a dual-constraint fusion generation mechanism to completely avoid the problem of semantic drift in fields. Through the layered adjudication of soft positioning constraints of business prompt words and rigid format constraints of standard templates, it adapts to the semantic differences of fields in heterogeneous documents. It retains the experience of business scenario recognition while strictly standardizing the output structure, eliminating problems such as custom fields in the model, semantic mismatch, and format drift.

[0019] 3. Achieve autonomous quality control in scenarios without gold standard certification, significantly reducing the cost of manual review. Relying on the self-checking mechanism of the two-round game of the isomorphic large model, the risk of the result can be quantitatively identified without manual annotation of standard answers. It automatically distinguishes between reliable results and high-risk work orders, and only performs manual review on abnormal work orders, greatly reducing the workload of full manual review and significantly improving the efficiency of customs document clearance automation.

[0020] 4. The fully localized deployment offers high security and is adapted to customs business scenarios. The entire model and inference process are deployed on the local server, eliminating the need to call external large model APIs and avoiding the risk of customs business data leakage. At the same time, the pluggable business rule engine can adapt to different customs business scenarios, making it highly versatile and scalable. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] The present invention provides a method for structured identification of customs documents based on multi-source fusion and large-scale model self-checking. This method reconstructs a multimodal feature input space and utilizes hierarchical dual-constraint guidance to achieve highly reliable structured document extraction in a fully localized, real-world delivery environment without gold standard certification. The method includes the following specific steps: S1. Document Sample Access, Multi-Source Template Binding and OCR Parsing: Loading Multi-Source Configuration Datasets Based on File Channel Attributes , in: These are business-side prompts used to provide location guidance (soft constraints). Set rigid SCHema boundaries and default values ​​for the system's standard field templates; This involves mapping task types. Subsequently, the system prioritizes reading the raw text stream. When the raw text stream quality is poor (e.g., scanned documents), the system automatically calls a locally deployed lightweight multimodal visual language OCR model for parsing. During the extraction process, in addition to extracting markdown or... Beyond the text sequence, the more crucial aspect is preserving the image feature tensor. and each text block The spatial geometric bounding box coordinate set enables the capture of complete multimodal topological features of customs documents; S2. Construct a multimodal dual-channel input matrix, perform language embedding processing on the text and visual feature embedding processing on the original document image, and achieve the association matching of text features and visual space features through cross-modal space alignment; parse the constraint relationship between business prompt words and standard field templates, and perform hierarchical adjudication of field output permissions through conflict hierarchical resolution operators to form dual-constraint fusion recognition rules; Specifically, to solve the problem of misaligned serialization, the system does not transmit text... Instead of using it as the sole input, it reconstructs the multimodal input moments. .

[0023] OCR Text Channel: Transmits text sequences Page separators and identified primary table cues are converted into token embeddings. .

[0024] Image Evidence Channel: This channel displays the rendered image of the original document page. After processing by the local visual encoder, the output contains spatial visual feature embeddings that include local image details. Cross-modal alignment is introduced in the attention layer of the model.

[0025] Let query matrix Key matrix By calculating the attention distribution, the model can directly "look back" at the visual grid around the field in the image when parsing a specific text field. When the general OCR divides the "quantity" and "amount" that were originally in the same row into two rows, the multimodal large model can correct its row classification perception based on the visual continuity of the horizontal baseline in the image.

[0026] Based on this, the system further dynamically parses multi-source templates to construct a set of business prompt words. With standard field template The hierarchical constraint relationship between them, specifically, the system first... The field names, field aliases, field location descriptions, example values, and extraction rules appearing in the data are semantically parsed to form a set of business location fields. At the same time, The field names, field types, array levels, default values, and required constraints defined in the schema are parsed in a structured manner to form a standard schema field set. Subsequently, the system calculates the intersection fields, business extension fields, and template missing fields of the two, and resolves them through a conflict-layering operator. Implement tiered adjudication for field output permissions.

[0027] when The business location field falls into When within the defined schema boundaries, this field is marked as an "outputtable field," and its business prompt content will be promoted to a field location constraint to guide the model in locating the corresponding value from OCR text and page images. Fields in When defining schema boundaries, this field is not directly written to the final JSON. Instead, it is downgraded to a "semantic understanding context reference," used only to assist the model in understanding the document context, table meaning, and field origin. The system uses a conflict layering resolution operator. During the output stage, this type of field is filtered to prevent it from entering the final structured result. The conflict resolution operator can be represented as:

[0028] In other words, for fields that exist in both business prompts and standard templates, the system allows output and uses business prompts to enhance positioning; for fields that exist only in business prompts, the system only retains their semantic reference function and does not allow output; for fields that exist only in standard templates, the system completes them based on the default values, field types, and contextual evidence in the standard templates. Through the above mechanism, the system retains business-side sample-level experience while strictly limiting the structural boundaries of the final JSON, eliminating the problems of non-schema field diffusion, model illusion field generation, and output format drift from the physical source.

[0029] S3. Isomorphic Game-Based Two-Round Inference and Result Collection: On a fully localized server, a two-round adversarial inference closed loop based on the ILMSC mechanism is enabled. In the first inference, the locally deployed large language model utilizes extremely low-temperature sampling, combined with the multimodal input matrix and hierarchical dual-constraint prompts constructed in S2, to perform a greedy search to obtain a deterministic inference path, producing a rough initial version of the JSON prediction result. .

[0030] In the asynchronous process following the completion of the first round, the system hides the prediction, injects adversarial reverse-picking perturbation prompts into the large model (e.g., "Recheck the table for cross-row errors in key numerical items, correct any potential illusions"), and initiates a second round of game-theoretic review inference on the same local weight instance, producing a second version of the review data. .

[0031] S4. The system loads a pluggable business correction plugin pool engine tailored to specific customs scenarios, and asynchronously performs rule correction and business semantic normalization on the two versions of results output by S3. The core operations include: Multi-level business normalization: Based on the internal knowledge graph, non-standard aliases or symbols are mapped to standard customs currency codes (such as USD), standard country codes, and customs packaging coding systems.

[0032] Global Optimal Row Matching Algorithm (GORM): For complex array row misalignment, extract the core identity anchor fields such as material number, item number, quantity, and unit price of each row, and use the bipartite graph optimal matching algorithm to calculate the globally optimal alignment and merging of the two-stage outputs on the LINES nested array.

[0033] Vertical attribute inheritance completer: When a uniform attribute (such as country of origin) is missing within lines, it automatically fills in the header data, corrects it, and outputs the final prediction result for the first stage. .

[0034] S5. The system loads the comprehensive risk scoring model and calculates the multidimensional uncertainty entropy value of the semantic consistency between the first prediction and the second review. The scoring model introduces a conflict function.

[0035] ; in, Calculate the missing value of nested JSON structures; We introduce weighted coefficients for core clearance key fields (e.g., the weight of total amount conflict is much higher than that of packaging description conflict), calculate the semantic distance or numerical tolerance of aliases after semantic normalization, quantify the entropy of alignment differences between table row items, and use this multivariate uncertainty function to quantitatively derive the final risk score. , representing the trust entropy value of the model's output result.

[0036] S6. Compare the risk score with the preset global risk control threshold. If the risk score is lower than the threshold, the final structured result will be automatically output. If the risk score is higher than the threshold, the risk score will be marked as high and automatically routed to the manual review stage.

[0037] The system sets global security risk control thresholds. : when The system issues an automated release instruction (needs_review=false) and automatically outputs the final structured JSON result to the customs clearance system; when This indicates that there is a serious semantic conflict in the prediction path generated independently by the two games of the large model (implying key item illusion or loss). The system marks the document as a high-risk interception marker, generates a highlighted conflict diagnosis report containing details of the conflict between the two rounds of games, and automatically routes the work order to the manual customs declaration review sequence (needs_review=true), requiring manual execution of targeted targeting. This mechanism realizes a trustworthy self-detection of the system's "unawareness of whether its output is correct".

[0038] A structured identification system for customs documents based on multi-source fusion and large-scale model self-checking includes: The document parsing and multi-source feature acquisition module is used to access customs documents, complete OCR parsing, image feature and spatial coordinate acquisition, and load multi-source configuration datasets. The multi-source evidence fusion and dual-constraint generation module is used to construct a multimodal input matrix, perform cross-modal space alignment, and complete field output constraint adjudication through conflict hierarchical resolution. The isomorphic large model dual-round inference module is used for local deployment of large models, completing two rounds of differentiated independent inference, and outputting initial and final versions of structured data; The business rule correction module is used to normalize document fields, match table rows, and correct attribute completion. The risk assessment and risk control routing module is used to quantitatively calculate the risk score of the identification results and complete the automatic output or manual review of the routing based on the threshold.

[0039] The beneficial effects of this invention, a method and system for structured identification of customs documents based on multi-source fusion and large-scale model self-checking, are as follows: By reconstructing the multimodal feature input space and utilizing hierarchical dual-constraint guidance, it achieves highly reliable structured document extraction in a fully localized, real-world delivery environment without gold standard certification. It avoids semantic drift of fields by relying on the dual constraints of business prompt words and standard templates. Through comparison of field-level, structure-level, and table row-level differences in the results of two rounds of differentiated reasoning, it quantifies and calculates risk scores, and achieves automatic evaluation of the credibility of the identification results. This effectively solves the problems of misidentification of customs document fields, semantic recognition deviation, and uncontrollable credibility of results in traditional technologies.

[0040] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for structured identification of customs documents based on multi-source fusion and large-scale model self-checking, characterized in that, Includes the following steps: S1. Access the customs document file to be identified, load the preset multi-source configuration dataset, which includes business prompt words, standard field templates and task type mapping data, parse the customs document file, extract the OCR text sequence, and retain the original image feature tensor of the document and the spatial geometric bounding box coordinate set of each text block. S2. Construct a multimodal dual-channel input matrix, perform language embedding processing on the OCR text and visual feature embedding processing on the original document image, and achieve the association matching of text features and visual space features through cross-modal space alignment; parse the constraint relationship between business prompt words and standard field templates, and perform hierarchical adjudication of field output permissions through conflict hierarchical resolution operators to form dual-constraint fusion recognition rules. S3. Based on a locally deployed isomorphic large model, differential prompt words are used to perform two rounds of independent reasoning, outputting a first version of the rough structured result and a second version of the verified structured result respectively; S4. The pluggable customs business verification engine is used to perform business normalization, global optimal row matching and vertical attribute completion correction on the two rounds of inference results to obtain two rounds of standardized structured results. S5. Construct a multi-dimensional risk scoring model to quantify the structural differences, semantic differences of key fields, and differences in table row alignment between the two rounds of standardized structured results, and obtain the risk score. S6. Compare the risk score with the preset global risk control threshold. If the risk score is lower than the threshold, the final structured result will be automatically output. If the risk score is higher than the threshold, the risk score will be marked as high and automatically routed to the manual review stage.

2. The method for structured identification of customs documents based on multi-source fusion and large-scale model self-checking as described in claim 1, characterized in that, In step S1, the specific method for parsing customs documents is as follows: firstly, the original text stream of the PDF format document is read; if the quality of the original text stream is not up to standard, a local lightweight multimodal visual language OCR model is called for parsing to fully capture the textual semantic features, image visual features, and spatial topological features of the document. The multi-source configuration dataset is a localized pre-configured dataset with no external API calls throughout the process. The business prompt words are used for soft constraint guidance in field positioning, and the standard field template is used to define the rigid schema boundaries and default values ​​of the structured output.

3. The method for structured identification of customs documents based on multi-source fusion and large-scale model self-checking as described in claim 1, characterized in that, In step S2, the multimodal dual-channel input matrix is ​​a concatenation matrix of text embedding features and image visual embedding features. By constructing a text query matrix and an image key matrix, the attention distribution is calculated so that the model can associate and match the image visual grid information when parsing text fields, and correct field attribution errors caused by OCR text splitting, misalignment, and page truncation.

4. The method for structured identification of customs documents based on multi-source fusion and large-scale model self-checking as described in claim 3, characterized in that, The cross-modal alignment is achieved based on the model attention layer. It relies on the visual continuity of the horizontal baseline of the document page to correct the misalignment of table fields and preserve the topological features of the two-dimensional layout of the document.

5. The method for structured identification of customs documents based on multi-source fusion and large-scale model self-checking as described in claim 1, characterized in that, In step S2, the adjudication logic of the conflict hierarchical resolution operator includes: If a field exists in both the business prompt word and the standard field template, it is marked as an output field, and the field result is generated with the business prompt word as the positioning constraint and the standard template as the format boundary. If a field only exists in business prompts, it is downgraded to a semantic reference context and does not participate in the final result output; If a field exists only in the standard field template, the field content will be automatically completed based on the template default value, field type, and contextual evidence.

6. The method for structured identification of customs documents based on multi-source fusion and large-scale model self-checking as described in claim 1, characterized in that, In step S3, the differentiation mechanism of the dual-round independent reasoning is as follows: the first round of reasoning adopts a low-temperature sampling and greedy search strategy, combined with dual-constraint prompts to generate a first version of coarse structured results. The second round of inference injects adversarial reverse verification prompts, performs game-theoretic verification inference based on the same local large model weights, and generates a second version of the structured verification result. .

7. The method for structured identification of customs documents based on multi-source fusion and large-scale model self-checking as described in claim 1, characterized in that, In step S4, the correction operation of the pluggable customs business verification engine includes: completing the standardization and normalization of currency, country, and packaging type through the business knowledge graph; using the global optimal row matching algorithm to achieve optimal alignment and merging of table array row items; and automatically inheriting the unified attributes of the document header to complete missing fields in the table rows.

8. The method for structured identification of customs documents based on multi-source fusion and large-scale model self-checking as described in claim 7, characterized in that, The global optimal row matching algorithm is based on the bipartite graph optimal matching strategy, using material number, item number, and quantity as core anchor fields to achieve accurate alignment and redundancy merging of nested table array row items.

9. The method for structured identification of customs documents based on multi-source fusion and large-scale model self-checking as described in claim 1, characterized in that, In step S5, the calculation formula for the risk scoring model is as follows: ; in, For nested structured data with missing differential values, For key fields, semantic distance and numerical tolerance, The row item alignment difference entropy is obtained by performing a globally optimal row matching on the two sets of table row items. The risk score is defined as follows: the higher the risk score, the greater the inconsistency between the two rounds of structured recognition results and the higher the recognition risk. These are the weighting coefficients for each dimension. Differentiated weights for different key fields for customs clearance. The first round of standardized structured results is obtained after performing business normalization and attribute completion correction on the first rough structured results. The second round of standardized structured results is obtained after business normalization and attribute completion correction of the second version of the review structured results, where k is the field index in the set of key field indexes for customs clearance; and They are respectively and The value of the kth key clearance field after business normalization; and They are respectively and The table is a collection of row items, where the subscript "line" indicates a substructure of the row item.

10. A customs document structured recognition system based on multi-source fusion and large-scale model self-checking as described in any one of claims 1-9, characterized in that, include: The document parsing and multi-source feature acquisition module is used to access customs documents, complete OCR parsing, image feature and spatial coordinate acquisition, and load multi-source configuration datasets. The multi-source evidence fusion and dual-constraint generation module is used to construct a multimodal input matrix, perform cross-modal space alignment, and complete field output constraint adjudication through conflict hierarchical resolution. The isomorphic large model dual-round inference module is used for local deployment of large models, completing two rounds of differentiated independent inference, and outputting initial and final versions of structured data; The business rule correction module is used to normalize document fields, match table rows, and correct attribute completion. The risk assessment and risk control routing module is used to quantitatively calculate the risk score of the identification results and complete the automatic output or manual review of the routing based on the threshold.

Images