Tender requirement automatic extraction method and system based on table structure and semantic understanding

Abstract

The present application relates to the technical field of table structure recognition, and specifically discloses a tender demand automatic extraction method and system based on table structure and semantic understanding, which acquires and receives a to-be-processed tender document, extracts a target table region set to be structured, performs OCR structured processing, constructs a table logical structure, and generates a structured HTML document; converts the structured HTML document into a DAG, labels a hierarchical semantic label as a first-level label for each data cell, and performs cross-cell semantic alignment and completion; constructs a semantic completeness evaluation model during the execution of the semantic propagation process, and evaluates the completeness of the first-level label corresponding to each data cell; performs five types of entity extraction for the first-level label that passes the completeness evaluation, and outputs the final entity result after correction. The present application can improve the accuracy, completeness and consistency of the tender demand extraction result, and reduce the workload of manual sorting and auditing.

Description

Technical Field

[0001] This invention relates to the field of table structure recognition technology, and more specifically, to a method and system for automatically extracting bidding requirements based on table structure and semantic understanding. Background Technology

[0002] In the field of bidding and tendering, as the size and complexity of bidding documents continue to increase, how to automatically extract structured information from tables has become a crucial technical issue for improving bidding and tendering efficiency and reducing manual review costs. Traditional methods mainly rely on manual reading and data entry, which is not only inefficient but also prone to omissions and errors. Therefore, the industry has gradually introduced automated technologies such as OCR, table recognition, and information extraction.

[0003] Existing technologies mainly fall into three categories. The first category is based on OCR and rule matching: OCR is first used to recognize text and coordinates, then fields are extracted through row and column partitioning and regular expression rules. This method is effective for tables with fixed formats, but it is prone to row and column misalignment in scenarios such as borderless tables, merged cells, multi-level headers, and multi-page tables. Furthermore, the rules struggle to understand logical relationships in natural language, resulting in poor adaptability to different bidding templates and requiring frequent maintenance of the rule base. The second category is based on end-to-end multimodal large-scale models: PDF pages or images are directly input into a visual language model, which generates structured results in one go. While reducing manual rule generation, this approach suffers from high inference costs, high token consumption, insufficient output stability, and susceptibility to illusions. Additionally, when scan quality is poor or table lines are missing, the model's understanding of the table structure is prone to errors. The third category is based on deep learning-based Table Structure Recognition (TSR) methods, such as LGPMA, TableFormer, and TATR. These methods can detect cell boundaries and generate HTML structures relatively well, but they mainly focus on the physical structure of the table and lack effective handling of multi-level headers, semantic inheritance relationships, and cross-cell references that are common in bidding scenarios. For example, when "Technical Requirements" is used as a merged header, its semantics cannot be automatically propagated to the lower-level data cells, resulting in a lack of complete hierarchical semantics in subsequent information extraction. At the same time, most TSR methods do not provide a coordination mechanism between structural confidence and subsequent extraction processes, making it difficult to achieve closed-loop error correction.

[0004] Therefore, it is necessary to provide a method and system for automatically extracting bidding requirements based on table structure and semantic understanding to solve the above technical problems. In order to solve the above problems, a technical solution is provided. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, this invention provides a method and system for automatically extracting bidding requirements based on table structure and semantic understanding. This addresses the problems of existing technologies in effectively parsing complex structures such as multi-level headers and merged cells and conveying hierarchical semantics when processing bidding forms, as well as the lack of a collaborative error correction mechanism between structural confidence and information extraction processes.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The method for automatically extracting bidding requirements based on table structure and semantic understanding includes the following steps: The system receives and processes tender documents, selects the parsing method based on the tender document type to locate the target table area, and generates a set of target table areas to be structured and processed. Based on the target table area set, perform OCR structured processing to construct the table's logical structure and generate a structured HTML document according to the positional relationship between cells; By converting structured HTML documents into DAGs, performing topological sorting and recursive semantic propagation of header semantics to leaf nodes, labeling each data cell with hierarchical semantic tags as first-level tags, and performing cross-cell semantic alignment and completion; By constructing a semantic integrity evaluation model during the semantic propagation process, the integrity of the first-level labels corresponding to each data cell is evaluated. For first-level labels that pass the integrity assessment, a domain-pre-trained encoder is used, combined with a hierarchical softmax classifier and contrastive learning, to extract five types of entities for each first-level label. By sequentially performing schema structure validation, low-confidence region secondary validation, and rule standardization correction on the entity extraction results, the final entity results are output and encapsulated into a standardized JSON output file.

[0007] As a further aspect of the present invention, the five types of entities specifically include: technical parameter entity, quantity and specification entity, delivery requirement entity, commercial terms entity, and qualification requirement entity.

[0008] As a further aspect of the present invention, the structured HTML document is converted into a DAG, and topological sorting is performed to recursively propagate the semantics of the table header to the leaf nodes. Each data cell is labeled with a hierarchical semantic tag as a first-level tag. The specific steps are as follows: A Directed Acyclic Graph (DAG) is constructed by treating each table cell as a graph node and the containment relationships between table headers and sub-headers formed across rows or columns as directed edges. Perform a topological sort on the DAG, propagate semantics in the order from the top-level header to the leaf node data cells, inherit the semantic labels of the parent nodes to the child nodes layer by layer, and combine them with the semantics of the current node so that each data cell obtains a hierarchical semantic label and is marked as a first-level label.

[0009] As a further aspect of the present invention, during the semantic propagation process, a semantic integrity evaluation model is constructed to assess the integrity of the first-level labels corresponding to each data cell. The specific steps are as follows: Get the set of parent nodes of the data cell, extract the semantic tags of the parent nodes, perform standardization processing, and construct the set of candidate semantic tags for the current node; Calculate semantic coverage metrics, depth coverage factor, and semantic matching degree based on the candidate semantic tag set; A semantic completeness evaluation model is constructed based on semantic coverage, deep coverage factor, and semantic matching degree, and a semantic completeness index is output. The completeness of the first-level label corresponding to the data cell is evaluated based on the semantic completeness index.

[0010] As a further aspect of the present invention, the semantic coverage index, depth coverage factor, and semantic matching degree are calculated based on the candidate semantic tag set, specifically as follows: The number of valid semantic tags in the first-level tags of the current node is counted. Based on the maximum number of semantic tags that appear in the current business scenario in the historical bidding knowledge base, the semantic coverage index is calculated based on the ratio of valid semantic tags to the maximum number of semantic tags.

[0011] As a further aspect of the present invention, the parent node in the DAG is defined as the first-level header node, the longest propagation path length of the current node to the parent node is recorded, the maximum propagation depth is calculated according to the DAG structure, and the depth coverage factor is calculated based on the ratio of the longest propagation path length to the maximum propagation depth.

[0012] As a further aspect of the present invention, a semantic matching evaluation mechanism is constructed to calculate the semantic matching degree by combining the effective semantic tags of the current node into a first-level semantic vector and extracting the standard semantic template vector corresponding to the current table type from the historical bidding knowledge base.

[0013] As a further aspect of the present invention, the integrity of the first-level label corresponding to the data cell is evaluated based on the semantic integrity index; specifically, the semantic integrity index is compared with a preset integrity threshold. If the semantic integrity index is greater than or equal to the preset integrity threshold, the integrity of the first-level label corresponding to the data cell is qualified; if the semantic integrity index is less than the preset integrity threshold, the integrity of the first-level label corresponding to the data cell is unqualified.

[0014] The automatic bidding requirement extraction system based on table structure and semantic understanding is applied to the above-mentioned automatic bidding requirement extraction method based on table structure and semantic understanding. It includes a document input interface, a table structure processing unit, a DAG hierarchical table semantic propagation unit, an integrity assessment unit, a five-type entity hierarchical extraction unit, and a post-processing and verification unit. The document input interface is used to receive the tender documents to be processed, select the parsing method according to the tender document type to locate the target table area, and generate a set of target table areas to be structured and processed. The table structuring unit is used to perform OCR structuring based on the target table area set, construct the table logical structure, and generate a structured HTML document based on the positional relationship between cells; The DAG hierarchical table semantic propagation unit is used to convert a structured HTML document into a DAG, perform topological sorting and recursively propagate the semantics of the table header to the leaf nodes, label each data cell with a hierarchical semantic tag as a first-level tag, and perform cross-cell semantic alignment and completion. The integrity assessment unit is used to assess the integrity of the first-level labels corresponding to each data cell by constructing a semantic integrity evaluation model during the semantic propagation process. The five-class hierarchical entity extraction unit is used to extract five types of entities for each first-level label that has passed the integrity assessment. It employs a domain-pre-trained encoder, combined with a hierarchical softmax classifier and contrastive learning. The post-processing and verification unit is used to sequentially perform schema structure verification, low-confidence region secondary verification, and rule standardization correction on the entity extraction results, and output the final entity results, which are then encapsulated into a standardized JSON output file.

[0015] A readable storage medium storing a computer program, which, when executed by a processor, is used to implement the steps of the automatic extraction method for bidding requirements based on table structure and semantic understanding as described above.

[0016] The technical effects and advantages of this invention's automatic extraction method and system for bidding requirements based on table structure and semantic understanding are as follows: This invention accurately locates target table areas and completes structured reconstruction through adaptive parsing for different file types; utilizing the DAG semantic propagation mechanism and cross-cell semantic completion mechanism, it transforms multi-level table headers, merged cells, and complex semantic relationships such as "same as above" and "see above table for details" into complete hierarchical semantic tags, effectively solving the problem that traditional OCR only recognizes text and struggles to understand business semantics; and dynamically assesses semantic inheritance quality through a semantic completeness evaluation model, avoiding invalid propagation and semantic omissions, thus improving the accuracy of table semantic modeling. By integrating domain-specific pre-trained encoders, hierarchical Softmax classifiers, and contrastive learning techniques, the system achieves accurate identification of core requirements such as technical parameters, quantity specifications, delivery requirements, business terms, and qualification requirements. Simultaneously, a closed-loop quality control mechanism is formed through schema structure verification, secondary verification of low-confidence regions, and rule standardization correction. This effectively reduces the impact of OCR recognition errors, structural parsing errors, and model inference biases, thereby significantly improving the accuracy, completeness, and consistency of bidding requirement extraction results. It also reduces the workload of manual processing and review, providing a reliable data foundation for intelligent bidding evaluation, regulatory review, procurement analysis, and digital management of bidding.

[0017] This invention not only enables the automated and structured extraction of requirements information from complex bidding documents, but also ensures the accuracy, consistency, and traceability of the extraction results. It significantly reduces the workload of manual review and data entry, improves the efficiency of digital processing of bidding and tendering business, and provides reliable data support for applications such as intelligent bid evaluation, regulatory review, procurement analysis, and the construction of bidding and tendering knowledge bases. Attached Figure Description

[0018] Figure 1 A flowchart of an automatic bidding requirement extraction method based on table structure and semantic understanding provided in an embodiment of the present invention; Figure 2 This is a system block diagram of an automatic bidding requirement extraction system based on table structure and semantic understanding, provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of this invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described technical solutions are only a part of this invention, and not all of it. All other technical solutions obtained by those skilled in the art based on the technical solutions of this invention without inventive effort are within the scope of protection of this invention.

[0020] Example 1: This is a flowchart of the automatic extraction method for bidding requirements based on table structure and semantic understanding provided in this embodiment of the invention. Figure 1The execution entity of the method shown can be a software and / or hardware device. The execution entity of this application can include, but is not limited to, at least one of the following: user equipment, network equipment, etc. User equipment can include, but is not limited to, computers, smartphones, personal digital assistants (PDAs), and the aforementioned electronic devices. Network equipment can include, but is not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of computers or network servers. Cloud computing is a type of distributed computing, consisting of a super virtual computer composed of a group of loosely coupled computers. This embodiment does not limit this. Steps S1 to S6 are detailed as follows: Step S1: Obtain the tender documents to be processed, select the parsing method according to the tender document type to locate the target table area, and generate a set of target table areas to be structured; file types include PDF files, image files and Word documents; Step S2 involves performing Optical Character Recognition (OCR) structured processing based on the target table region set to construct the table's logical structure and generate a structured Hypertext Markup Language (HTML) document based on the positional relationships between cells. The OCR structured processing flow includes layout analysis, text detection, character recognition, row and column clustering, and merged cell reasoning. The table's logical structure is constructed based on the cell positional relationships, and a structured HTML document is output. Each tag in the HTML document's cell nodes carries text content, row index, column index, cross-row and cross-column information, and structural confidence parameters. Low-confidence regions are filtered based on a preset confidence threshold to generate a set of low-confidence regions. Step S3 involves converting the structured HTML document into a directed acyclic graph (DAG), performing topological sorting and recursive semantic propagation of the header semantics to the leaf nodes, labeling each data cell with hierarchical semantic tags as first-level tags, and performing cross-cell semantic alignment and completion. Step S4: During the semantic propagation process, a semantic integrity evaluation model is constructed to assess the integrity of the first-level labels corresponding to each data cell. Step S5: For the first-level labels that have passed the integrity assessment, a domain-pre-trained encoder is used, combined with a hierarchical softmax classifier and contrastive learning, to extract five types of entities for each first-level label; the five types of entities specifically include: technical parameter entity, quantity and specification entity, delivery requirement entity, business condition entity, and qualification requirement entity. Step S6 involves sequentially performing schema structure verification, secondary verification of low-confidence regions, and rule standardization correction on the entity extraction results, outputting the final entity results, and encapsulating them into a standardized JSON output file.

[0021] It should be noted that all parameter data used in model calculations must be dimensionless beforehand.

[0022] Preferably, cross-cell semantic alignment and completion are performed, and the specific steps are as follows: By performing cross-cell semantic association analysis on data cells, referential relationships such as "same as above", "same as left", and "see table above" are identified, and the true value is obtained by tracing back to the nearest non-empty cell. Inferring the execution unit and completing the dimensionality based on the keywords in the table header; Further identify special structural information such as notes rows and total rows, and inject them as global constraints into the current table semantic space to achieve contextual semantic completion.

[0023] Preferably, the target table area is located by selecting the parsing method according to the type of tender document. The specific steps are as follows: File types include PDF files, image files, and Word documents; For PDF and image files, the target table area in the page is identified by the layout analysis algorithm, and the table boundary coordinates are extracted. For Word files, the Document Object Model (DOM) is directly parsed to obtain the target table area and its structural information.

[0024] Preferably, by converting the structured HTML document into a directed acyclic graph (DAG), performing topological sorting and recursive semantic propagation of the header semantics to the leaf nodes, and labeling each data cell with hierarchical semantic tags as first-level tags, the specific steps are as follows: By treating each table cell as a graph node and the containment relationships between the table header and sub-headers formed across rows or columns as directed edges, a directed acyclic graph (DAG) is constructed. Perform a topological sort on the DAG, propagate semantics in the order from the top-level header to the leaf node data cells, inherit the semantic labels of the parent nodes to the child nodes layer by layer, and combine them with the semantics of the current node so that each data cell obtains a hierarchical semantic label and is marked as a first-level label.

[0025] In this embodiment of the invention, the technical parameter table in a server procurement project tender document is used as an example to illustrate the specific implementation process of the invention. The tender document is stored in PDF format and includes multiple tables such as a list of procured equipment, a technical parameter table, and a commercial terms table.

[0026] The system receives the tender documents to be processed and identifies the file type as PDF. It then uses a layout analysis algorithm to parse the PDF pages, locating target table areas by detecting table borders, text arrangement patterns, and regional layout features, and extracting table boundary coordinate information. For example, if a technical parameter table area is detected on the first page, its corresponding coordinate range is obtained, and this area is added to the set of tables to be processed. If the input file is a Word document, the system directly parses the Document Object Model (DOM) to extract table objects and their row and column structure information.

[0027] Next, OCR structured processing is performed on the target table area. The system generates a structured HTML document after completing text recognition, cell detection, row and column clustering, and merged cell reasoning. The structured HTML document is then converted into a directed acyclic graph (DAG). Cells such as "Technical Requirements," "Server Configuration," "Parameter Name," and "Parameter Indicator" serve as graph nodes, and the hierarchical relationships between multi-level table headers are represented as directed edges. For example, the "Technical Requirements" node points to the "Server Configuration" node, which in turn points to the "CPU Core Count" node. A topological sort is performed on the DAG, and semantic propagation is performed from the top-level header to the leaf nodes. During propagation, the semantics of parent nodes are inherited layer by layer to child nodes, ensuring that the data cell corresponding to "CPU Core Count" receives a complete semantic label: "Technical Requirements → Server Configuration → CPU Core Count".

[0028] Similarly, the data cell corresponding to "memory capacity" is assigned the semantic tag: "Technical Requirements → Server Configuration → Memory Capacity". The complete semantic path described above serves as the first-level tag for the corresponding data cell.

[0029] After constructing the first-level labels, cross-cell semantic alignment and completion are performed. First, the referential relationships within the table are identified. For example, if the content of the cell in the second row and first column is "same as above," the system traces back to the nearest non-empty cell based on the row and column position to determine its actual value as "server configuration," and then adds this content to the current cell. Next, unit inference processing is performed. For instance, if the table header contains keywords such as "capacity" or "memory," but the data content is only "32," the system automatically completes the unit "GB" based on the historical bidding knowledge base and the table header semantics, forming the standard expression "32GB."

[0030] Further identify special structural information in the table. For example, a note row is found at the bottom of the table: "Note: All devices must support remote management functions." This note content is parsed as a global constraint and injected into the current table's semantic space, so that subsequent extracted technical parameters such as the number of CPU cores and memory capacity inherit this note constraint. At the same time, if a "Total" row is detected, the total quantity information is extracted and a correlation constraint relationship is established with the number of devices to achieve cross-row semantic association completion.

[0031] After the above processing, the issues of missing referents, missing units, and missing contextual semantics in the original table were effectively corrected. Ultimately, each data cell obtained complete hierarchical semantic labels and contextual business semantic information, providing an accurate data foundation for subsequent semantic completeness assessment, extraction of five types of entities, and knowledge base correction, thereby improving the accuracy and reliability of automatically extracting requirement information from complex bidding documents.

[0032] Preferably, during the semantic propagation process, a semantic integrity evaluation model is constructed to assess the integrity of the primary labels corresponding to each data cell. The specific steps are as follows: Get the set of parent nodes of the data cell, extract the semantic tags of the parent nodes, perform standardization processing, and construct the set of candidate semantic tags for the current node; The system calculates the semantic coverage metric based on the ratio of valid semantic tags to the maximum number of valid semantic tags in the current node's first-level tags, according to the maximum number of semantic tags appearing in the current business scenario in the historical bidding knowledge base. It also calculates the number of valid semantic tags inherited by the current node. The number of valid semantic tags refers to the set of tags that can participate in business semantic expression after standardization. For example, for the "CPU core count" cell, "technical requirements," "server configuration," and "parameter name" are all valid semantic tags, while non-business information such as table border numbers and sorting numbers are not included in the statistics. Simultaneously, it calculates the maximum number of semantic tags that may appear in the current business scenario based on the historical bidding knowledge base and predefined bidding templates. For example, server procurement bidding documents typically contain four layers of semantics: "equipment category," "technical requirements," "parameter name," and "parameter index," corresponding to a maximum of 4 semantic tags. The semantic coverage metric reflects the proportion of semantic information inherited by the current node relative to the complete semantic information. When the semantic coverage metric is close to 1, it indicates that the current node has acquired relatively complete semantic expression capabilities.

[0033] By defining the parent node in the DAG as the first-level header node, recording the longest propagation path length from the current node to the parent node, calculating the maximum propagation depth based on the DAG structure, and calculating the depth coverage factor based on the ratio of the longest propagation path length to the maximum propagation depth; By combining the valid semantic tags of the current node into a first-level semantic vector, and extracting the standard semantic template vector corresponding to the current table type from the historical bidding knowledge base, a semantic matching evaluation mechanism is constructed to calculate the semantic matching degree. The calculation formula is as follows: ; In the formula: For semantic matching degree, It is a first-level semantic vector. It is a standard semantic template vector.

[0034] A semantic completeness evaluation model is constructed based on semantic coverage, deep coverage factor, and semantic matching degree, and the semantic completeness index is output. The calculation formula is as follows: ; In the formula: As a semantic completeness metric, The weighting coefficient for semantic coverage. For semantic coverage, These are the weighting coefficients of the depth coverage factor. For depth coverage factor, The weight coefficients for semantic matching degree This represents the semantic matching degree.

[0035] The completeness of the first-level label corresponding to the data cell is evaluated based on the semantic completeness index. Specifically, the semantic completeness index is compared with a preset completeness threshold. If the semantic completeness index is greater than or equal to the preset completeness threshold, the completeness of the first-level label corresponding to the data cell is qualified; if the semantic completeness index is less than the preset completeness threshold, the completeness of the first-level label corresponding to the data cell is unqualified.

[0036] The semantic integrity evaluation model described above can evaluate the semantic inheritance status of the current node in real time and dynamically decide whether to continue propagating based on the degree of semantic integrity. This avoids the semantic redundancy diffusion problem that exists in the traditional DAG recursive propagation process, reduces the amount of invalid computation, improves the semantic modeling accuracy and processing efficiency of complex bidding forms, and enables the final data cells to carry complete, accurate and business-meaning hierarchical semantic labels, providing a high-quality semantic foundation for subsequent cross-cell semantic alignment, knowledge base correction and extraction of five types of entities.

[0037] In one embodiment of the present invention, taking the technical parameter table in a server procurement project tender document as an example, firstly, OCR structured recognition and table structure parsing are completed, and the table is converted into a directed acyclic graph (DAG) structure. The header nodes, such as "Equipment Category," "Technical Requirements," "Parameter Name," and "Parameter Indicator," constitute upper-level semantic nodes, while specific parameter data, such as "CPU Core Count" and "Memory Capacity," constitute leaf nodes. Subsequently, semantic propagation is performed according to the topological sorting result, passing the business semantics carried by the parent nodes layer by layer to the data cells. For example, for the data cell corresponding to "CPU Core Count," semantic tags such as "Equipment Category," "Technical Requirements," and "Parameter Name" from its parent node set are obtained, and these tags are standardized to uniformly map synonymous expressions. Invalid tags that do not participate in the business semantic expression, such as table serial numbers and border numbers, are removed, thereby constructing a candidate semantic tag set corresponding to the current node.

[0038] After constructing the candidate semantic tag set, the number of valid semantic tags inherited by the current node is counted, and the standard semantic template corresponding to the current business scenario is retrieved from the historical bidding knowledge base. For example, in the server procurement scenario, the knowledge base statistics show that complete business semantics typically includes four layers of semantics: "equipment category," "technical requirements," "parameter name," and "parameter index." Therefore, the maximum number of semantic tags is set to 4. Subsequently, the semantic coverage index is calculated by the ratio between the number of valid semantic tags of the current node and the maximum number of semantic tags. If the current node has obtained three layers of semantics: "equipment category," "technical requirements," and "parameter name," the corresponding semantic coverage is 0.75, indicating that the current node has inherited 75% of the complete business semantic information.

[0039] Meanwhile, the propagation depth of the current node is analyzed based on the DAG structure. Specifically, the top-level header node is used as the propagation starting point. The longest propagation path length from the current node to the top-level header node is calculated, and the depth coverage factor is calculated by combining this with the maximum propagation depth in the entire DAG structure. For example, when the "CPU core count" node is located in the fourth-level semantic structure, and three levels of semantic inheritance have already been completed, the depth coverage factor is 0.75. This metric reflects the degree of completion of the current node in the entire hierarchical semantic propagation process, evaluating the sufficiency of semantic inheritance from a structural perspective.

[0040] To further evaluate the matching degree between the semantic information obtained by the current node and the standard business semantics, a semantic matching evaluation mechanism is constructed. First, the first-level label corresponding to the current node is converted into a first-level semantic vector. Simultaneously, a standard semantic template vector corresponding to the current table type is extracted from the historical bidding knowledge base. Then, the cosine similarity algorithm is used to calculate the semantic matching degree between the first-level semantic vector and the standard semantic template vector. When the semantic matching degree is close to 1, it indicates that the semantic structure inherited by the current node is highly consistent with the standard semantic structure in the historical bidding scenario, and can accurately express the business meaning of the current data cell.

[0041] After obtaining semantic coverage, depth coverage factor, and semantic matching degree, a semantic completeness evaluation model is constructed using these three indicators to comprehensively evaluate the completeness of the first-level label corresponding to the current node. If the calculated semantic completeness index is greater than or equal to the preset completeness threshold, the first-level label corresponding to the current data cell is deemed complete, semantic propagation is stopped, and the current first-level label is output as the final level semantic label. For example, "Equipment Category → Technical Requirements → Parameter Name → Number of CPU Cores" is used as the final first-level label for this data cell. Conversely, if the semantic completeness index is lower than the preset threshold, it indicates that the current node still has semantic missing issues, semantic propagation continues, or a cross-cell semantic completion mechanism is triggered to supplement the missing semantics.

[0042] For example, in actual bidding documents, the first two columns of the row containing "Memory Capacity" might contain the phrase "Same as above." In this case, a cross-cell semantic alignment mechanism is used to trace back to the corresponding content in the previous row, automatically supplementing the current node with semantic information such as "Server" and "Computing Configuration," and then recalculating the semantic completeness index. When the completed semantic completeness meets the preset requirements, the first-level label is deemed to be complete. For cases where incomplete semantic paths such as "Parameter Name → Number of CPU Cores" are obtained due to OCR recognition errors or missing table structures, the semantic completeness index calculated in the end will not meet the threshold requirements because both the semantic coverage and depth coverage factor are low. In such cases, semantic completion and propagation operations will continue to be performed until a complete business semantic chain is formed.

[0043] Through the above methods, this invention can comprehensively evaluate the first-level tags from three dimensions: semantic coverage, hierarchical propagation, and semantic matching. It enables real-time monitoring and dynamic control of the semantic inheritance state of data cells, avoids the semantic redundancy and ineffective propagation problems existing in the traditional DAG recursive propagation process, improves the semantic modeling accuracy and processing efficiency of complex bidding forms, and enables the final data cells to carry complete, accurate, and business-logical hierarchical semantic tags, providing a reliable semantic foundation for subsequent knowledge base correction, extraction of five types of entities, and post-processing verification.

[0044] Preferably, for the first-level labels that pass the integrity assessment, a domain-pre-trained encoder is used, combined with a hierarchical softmax classifier and contrastive learning, to extract five types of entities for each first-level label. The specific steps are as follows: A complete set of qualified first-level labels is obtained. These hierarchical semantic labels are concatenated with the cell text content to form a semantically enhanced text sequence. This sequence is then input into a domain-specific continuously pre-trained encoder for feature encoding, outputting a feature vector representation containing contextual semantic information. For example, "Technical Requirements·Parameter Name·CPU Core Count" is fused with the corresponding cell content "≥8 cores" to construct the input sequence: "Technical Requirements·Parameter Name·CPU Core Count: ≥8 cores". This semantically enhanced text sequence is then input into the domain-specific continuously pre-trained encoder for feature encoding. The domain-specific continuously pre-trained encoder is obtained by continuously pre-training on a general BERT or RoBERTa model using historical bidding documents, procurement requirement documents, tender documents, and technical specification documents. This allows the model to learn domain-specific terminology, parameter expressions, and business logic relationships, thereby outputting a feature vector representation containing contextual semantic information.

[0045] Entity classification is performed using a hierarchical Softmax classification structure based on feature vector representation. First, a primary classification layer is established to coarsely classify the business domain to which the input labels belong, outputting three primary categories: technical requirements, business requirements, and qualification requirements. Then, based on the primary classification results, the corresponding secondary classification layers are established. The technical requirements category is further subdivided into technical parameter entities and quantity specification entities; the business requirements category is further subdivided into delivery requirement entities and business condition entities; and the qualification requirements category corresponds to qualification requirement entities. This hierarchical classification structure reduces mutual interference between categories and improves the ability to distinguish similar categories.

[0046] A contrastive learning mechanism is introduced to construct an entity feature optimization model, which narrows the distance between entities of the same category in the feature space while increasing the distance between entities of different categories. Specifically, entities of the same category are constructed as positive sample pairs; for example, "CPU core count ≥ 8 cores" and "memory capacity ≥ 32GB" both belong to technical parameter entities. Entities of different categories are constructed as negative sample pairs; for example, "CPU core count ≥ 8 cores" and "delivery within 30 days" belong to technical parameter entity and delivery requirement entity, respectively. By continuously narrowing the distance between entities of the same category in the feature space through a contrastive learning loss function, while increasing the distance between entities of different categories, the model can learn more obvious category boundary features, thereby improving classification accuracy in complex scenarios and solving the problem of ambiguous semantic boundaries between entities of different categories.

[0047] The classification probability corresponding to each entity category is calculated, and the category with the highest probability is selected as the final entity type. When the classification result belongs to the technical parameter entity, the parameter name, parameter index, and parameter constraint information are extracted; when the classification result belongs to the quantity specification entity, the quantity value, unit of measurement, and specification model information are extracted; when the classification result belongs to the delivery requirement entity, the delivery time, delivery location, and acceptance requirements information are extracted; when the classification result belongs to the commercial terms entity, the payment method, performance bond, and after-sales service conditions information are extracted; when the classification result belongs to the qualification requirement entity, the enterprise qualification, personnel qualification, and certification certificate requirements information are extracted.

[0048] Preferably, the entity extraction results are sequentially subjected to schema structure validation, secondary validation of low-confidence regions, and rule standardization correction to output the final entity results, which are then encapsulated into a standardized JSON output file. The specific steps are as follows: Obtain the initial JSON structured results from the extraction of five types of entities, and read the first-level tag path, entity category, entity content, and entity confidence information of the corresponding entities. At the same time, obtain the structure confidence parameters recorded in the OCR structuring stage and the semantic integrity evaluation results output in the semantic propagation stage, and construct the set of entities to be verified.

[0049] Perform schema structure validation. A standard schema template for bidding requirements is pre-established. The extracted results are then validated for structural legality, checking whether the initial JSON structured result contains predefined fields corresponding to technical parameters, quantity specifications, delivery requirements, business conditions, and qualification requirements. Simultaneously, the field types are verified to meet requirements, such as whether quantity fields are numeric, date fields conform to date formats, and business amount fields conform to amount formats. When missing fields, abnormal field names, or hierarchical structure errors are found, exception information is automatically recorded, and the validation prompt is reconstructed based on the exception content. This triggers a targeted retry in the large language model, generating a structured result that conforms to the schema specification.

[0050] After schema validation, low-confidence regions are filtered based on the overall credibility evaluation results. For target cells with an overall credibility below a preset threshold, the corresponding HTML structure fragment, original OCR text content, and first-level tag semantic paths are extracted to construct a local validation context. This local validation context is then separately input into a large language model for secondary validation to obtain the validation results. The validation results are compared with the initial extraction results. If the two results match, the entity's credibility is increased; if discrepancies exist, the final entity content is redefined by combining historical bidding knowledge base matching results and validation confidence, thereby eliminating error propagation problems caused by OCR misidentification, table parsing errors, and model inference bias.

[0051] After completing the low-confidence region verification, rule correction processing is performed. The rule base is invoked to standardize the extracted results, including unit unification, numerical format unification, and business logic consistency verification. Unit unification converts units of measurement such as "piece," "set," and "unit" into a unified standard format; numerical format unification converts data such as amounts, dates, and percentages into standard representations; and business logic consistency verification verifies whether entities meet business relationship requirements, such as ensuring that equipment quantity corresponds to equipment name, delivery time corresponds to delivery requirement entity, and qualification requirements match procurement category. When logical conflicts are detected, the rule correction strategy is automatically invoked to adjust or mark abnormal results.

[0052] Example 2: In this embodiment of the invention, the method for automatically extracting bidding requirements based on table structure and semantic understanding is applied to the server procurement bidding documents of a large data center construction project to illustrate the overall implementation process of the invention. The bidding documents for this project were issued by a bidding agency and have more than 300 pages in total, including a large number of complex tables such as equipment list, technical parameter table, commercial terms table, qualification requirements table, and delivery plan table. Due to the different layout methods used in different chapters, some tables have multi-level headers, merged cells, cross-page tables, and semantic references such as "same as above" and "see previous table for details." Traditional methods based on rule matching or simple OCR recognition are difficult to accurately extract bidding requirement information. Therefore, the method of the present invention is used to automatically parse the bidding documents.

[0053] The system receives the tender documents to be processed and automatically identifies the file type. Since the tender documents are in PDF format, the layout analysis module is invoked to examine the content of each page and identify table areas. The layout analysis module combines text distribution features, border line features, and visual layout features to locate table boundaries and extract corresponding coordinate information. For example, a server technical parameter table is identified in the technical specifications section, a payment terms table in the commercial terms section, and a qualification requirements table in the qualification review section. All identified table areas are constructed into a target table area set and used as input for subsequent structured processing.

[0054] OCR structured processing is performed on the target table area. First, text recognition and cell detection are completed. Then, the table's logical structure is constructed through row and column clustering analysis and cell merging reasoning. For example, in a server technical parameter table, a multi-level header structure such as "Equipment Category," "Technical Requirements," "Parameter Name," and "Parameter Indicators" is recognized, and the recognition results are converted into a structured HTML document. The HTML document not only stores the cell text content but also records row and column coordinates, cross-row and cross-column relationships, and hierarchical structure information between cells, thus forming a structured expression that reflects the true logical relationships of the table.

[0055] After obtaining the structured HTML document, the HTML structure is further converted into a Directed Acyclic Graph (DAG). Each table cell serves as a node in the graph, and the containment relationships between multi-level table headers are represented by directed edges. For example, the "Device Category" node points to the "Technical Requirements" node, which in turn points to the "Parameter Name" node, which ultimately relates to specific data nodes such as "CPU Core Count" and "Memory Capacity." The DAG is then topologically sorted, and semantic information is recursively propagated from the top-level header to the leaf nodes. After propagation, the data cell corresponding to "CPU Core Count" receives a complete hierarchical semantic label of "Device Category → Technical Requirements → Parameter Name → CPU Core Count," and the data cell corresponding to "Memory Capacity" receives a hierarchical semantic label of "Device Category → Technical Requirements → Parameter Name → Memory Capacity." These hierarchical semantic labels serve as the first-level labels for the corresponding data cells.

[0056] Because actual bidding documents frequently contain cross-cell references such as "same as above," "same as left," and "see table above," a further cross-cell semantic alignment and completion process is performed. For example, in the row containing "Memory Capacity," the first two columns display "same as above." Based on the row and column position, the system automatically backtracks to the nearest non-empty cell, identifies the actual value as "Server" and "Computing Configuration," and adds it to the current node. Simultaneously, for parameter values ​​where only numbers are entered without specifying the unit, such as "128," the system automatically infers the unit as "GB" based on the "Memory Capacity" keyword in the table header and the historical bidding knowledge base, completing it as "128GB." Furthermore, the note "All servers should support remote management functions" appearing at the bottom of the table is identified as a global constraint and injected into the current table's semantic space to achieve contextual semantic completion.

[0057] After semantic propagation and semantic completion are completed, a semantic completeness evaluation model is constructed to assess the completeness of the first-level labels corresponding to each data cell. First, the number of valid semantic labels in the current first-level label is counted, and a standard semantic template for the current business scenario is obtained from the historical bidding knowledge base. For example, in a server procurement scenario, the complete semantics recorded in the knowledge base typically include a four-layer structure: "equipment category," "technical requirements," "parameter name," and "parameter index." Then, semantic coverage, depth coverage factor, and semantic matching degree are calculated, and a semantic completeness index is calculated based on these three indicators. When the semantic completeness reaches a preset threshold, the current first-level label is deemed complete; if it is below the threshold, semantic propagation continues or a semantic completion mechanism is triggered. For example, the semantic completeness of the "CPU core count" node reaches 0.86 after calculation, which is higher than the preset threshold of 0.80, therefore it is deemed semantically complete. Data cells that only obtain partial semantic labels due to missing OCR information will have their semantic information supplemented until the requirements are met.

[0058] For first-level labels that pass the integrity assessment, the entity extraction stage begins. First, the first-level labels are fused with the corresponding data content, for example, to form semantically enhanced text such as "Equipment Category → Technical Requirements → Parameter Name → CPU Core Count: ≥32 cores". Then, an encoder continuously pre-trained in the bidding domain is used to encode the text's features, and a hierarchical Softmax classifier is used to perform hierarchical classification. During classification, the current entity is first determined to belong to the technical requirement, business requirement, or qualification requirement category, and then further subdivided into five entity categories: technical parameters, quantity specifications, delivery requirements, business conditions, and qualification requirements. Simultaneously, a contrastive learning mechanism is introduced to improve entity classification accuracy by narrowing the feature distance between entities of the same category and widening the feature differences between entities of different categories. For example, "CPU core count ≥32 cores" is identified as a technical parameter entity, "delivery within 30 days" is identified as a delivery requirement entity, and "ISO9001 certification" is identified as a qualification requirement entity.

[0059] After extracting the five types of entities, the post-processing and verification stage begins. First, a schema structure verification is performed, checking the structural legality of the extracted JSON results to verify compliance with the predefined bidding requirements schema specifications. If missing fields or hierarchical structure anomalies are found, a correction prompt is automatically constructed, and the large language model is invoked to regenerate the results. Subsequently, a comprehensive credibility model is constructed based on OCR confidence, structural recognition confidence, semantic completeness, and entity classification confidence to filter low-confidence regions. For entity results corresponding to low-confidence regions, their HTML fragments, original OCR text, and semantic tag paths are extracted and separately input into the large language model for secondary verification, combined with historical knowledge base for conflict arbitration and result correction. Finally, rule standardization processing is performed, uniformly converting amounts, dates, quantities, and units, and checking for consistency between equipment quantity and name, delivery time and contract terms, and other business logic.

[0060] After the above processing, a standardized JSON output file is finally generated. For example, it automatically outputs structured results such as technical parameter entities "CPU core count ≥ 32 cores" and "Memory capacity ≥ 128GB", commercial terms entities "Payment within 30 days of contract signing", and qualification requirements entities "ISO9001 quality management system certification". Through the processing flow of this invention, the information extraction work that originally required manually reading hundreds of pages of bidding documents can be completed automatically in a shorter time, effectively improving the accuracy, completeness, and consistency of extracting requirements from complex bidding documents.

[0061] Example 3: An automatic bidding requirement extraction system based on table structure and semantic understanding, including a document input interface, a table structuring processing unit, a DAG-level table semantic propagation unit, an integrity assessment unit, a five-type entity hierarchical extraction unit, and a post-processing and verification unit; the document input interface is connected to the table structuring processing unit, the table structuring processing unit is connected to the DAG-level table semantic propagation unit, the DAG-level table semantic propagation unit is connected to the integrity assessment unit, the integrity assessment unit is connected to the five-type entity hierarchical extraction unit, and the five-type entity hierarchical extraction unit is connected to the post-processing and verification unit; The document input interface is used to receive the tender documents to be processed, select the parsing method according to the tender document type to locate the target table area, and generate a set of target table areas to be structured and processed. The table structuring unit is used to perform OCR structuring based on the target table area set, construct the table logical structure, and generate a structured HTML document based on the positional relationship between cells; The DAG hierarchical table semantic propagation unit is used to convert structured HTML documents into directed acyclic graphs (DAGs), perform topological sorting and recursively propagate the semantics of the table header to the leaf nodes, label each data cell with hierarchical semantic tags as first-level tags, and perform cross-cell semantic alignment and completion. The integrity assessment unit is used to assess the integrity of the first-level labels corresponding to each data cell by constructing a semantic integrity evaluation model during the semantic propagation process. The five-class hierarchical entity extraction unit is used to extract five types of entities for each first-level label that has passed the integrity assessment. It employs a domain-pre-trained encoder, combined with a hierarchical softmax classifier and contrastive learning. The post-processing and verification unit is used to sequentially perform schema structure verification, low-confidence region secondary verification, and rule standardization correction on the entity extraction results, and output the final entity results, which are then encapsulated into a standardized JSON output file.

[0062] like Figure 2 The diagram shown is a system block diagram of an automatic bidding requirement extraction system based on table structure and semantic understanding according to an embodiment of the present invention, which can be used to execute... Figure 1 The steps in the method embodiments shown are implemented in a similar manner and have similar technical effects, and will not be repeated here.

[0063] A readable storage medium storing a computer program, which, when executed by a processor, is used to implement the steps of the automatic extraction method for bidding requirements based on table structure and semantic understanding as described above.

[0064] The readable storage medium can be a computer storage medium or a communication medium. A communication medium includes any medium that facilitates the transfer of computer programs from one location to another. A computer storage medium can be any available medium accessible to a general-purpose or special-purpose computer. For example, a readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application-Specific Integrated Circuit (ASIC). Alternatively, the ASIC can be located in a user equipment. Of course, the processor and the readable storage medium can also exist as discrete components in a communication device. The readable storage medium can be a read-only memory (ROM), random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0065] The present invention also provides a program product including executable instructions stored in a readable storage medium. At least one processor of the device can read the executable instructions from the readable storage medium, and the at least one processor executes the executable instructions to cause the device to implement the methods provided in the various embodiments described above.

[0066] In the embodiments of the above-described device, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0067] Through the above embodiments, this invention, by adaptively parsing different types of files, can accurately locate target table areas and complete structured reconstruction; by utilizing the DAG semantic propagation mechanism and cross-cell semantic completion mechanism, it transforms multi-level table headers, merged cells, and complex semantic relationships such as "same as above" and "see above table for details" into complete hierarchical semantic tags, effectively solving the problem that traditional OCR only recognizes text but cannot understand business semantics; by dynamically evaluating the quality of semantic inheritance through a semantic integrity evaluation model, it avoids invalid propagation and semantic loss, improving the accuracy of table semantic modeling; by combining domain-specific continuous pre-trained encoders, hierarchical Softmax classifiers, and contrastive learning techniques, it achieves accurate identification of core requirement information such as technical parameters, quantity specifications, delivery requirements, business conditions, and qualification requirements; at the same time, by forming a closed-loop quality control mechanism through schema structure verification, secondary verification of low-confidence regions, and rule standardization correction, it effectively reduces the impact of OCR recognition errors, structural parsing errors, and model inference biases, thereby significantly improving the accuracy, completeness, and consistency of bidding requirement extraction results, reducing the workload of manual processing and review, and providing a reliable data foundation for intelligent bidding, regulatory review, procurement analysis, and digital management of bidding.

[0068] This invention not only enables the automated and structured extraction of requirements information from complex bidding documents, but also ensures the accuracy, consistency, and traceability of the extraction results. It significantly reduces the workload of manual review and data entry, improves the efficiency of digital processing of bidding and tendering business, and provides reliable data support for applications such as intelligent bid evaluation, regulatory review, procurement analysis, and the construction of bidding and tendering knowledge bases.

[0069] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

[0070] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for automatically extracting bidding requirements based on table structure and semantic understanding, characterized in that, Includes the following steps: The system receives and processes tender documents, selects the parsing method based on the tender document type to locate the target table area, and generates a set of target table areas to be structured and processed. Based on the target table area set, perform OCR structured processing to construct the table's logical structure and generate a structured HTML document according to the positional relationship between cells; By converting structured HTML documents into DAGs, performing topological sorting and recursive semantic propagation of header semantics to leaf nodes, labeling each data cell with hierarchical semantic tags as first-level tags, and performing cross-cell semantic alignment and completion; By constructing a semantic integrity evaluation model during the semantic propagation process, the integrity of the first-level labels corresponding to each data cell is evaluated. For first-level labels that pass the integrity assessment, a domain-pre-trained encoder is used, combined with a hierarchical softmax classifier and contrastive learning, to extract five types of entities for each first-level label. By sequentially performing schema structure validation, low-confidence region secondary validation, and rule standardization correction on the entity extraction results, the final entity results are output and encapsulated into a standardized JSON output file.

2. The method for automatically extracting bidding requirements based on table structure and semantic understanding according to claim 1, characterized in that, The five types of entities include: technical parameter entities, quantity and specification entities, delivery requirement entities, commercial terms entities, and qualification requirement entities.

3. The method for automatically extracting bidding requirements based on table structure and semantic understanding according to claim 1, characterized in that, By converting the structured HTML document into a DAG, performing topological sorting, recursively propagating the semantics of the table header to the leaf nodes, and labeling each data cell with hierarchical semantic tags as first-level tags, the specific steps are as follows: A Directed Acyclic Graph (DAG) is constructed by treating each table cell as a graph node and the containment relationships between table headers and sub-headers formed across rows or columns as directed edges. Perform a topological sort on the DAG, propagate semantics in the order from the top-level header to the leaf node data cells, inherit the semantic labels of the parent nodes to the child nodes layer by layer, and combine them with the semantics of the current node so that each data cell obtains a hierarchical semantic label and is marked as a first-level label.

4. The method for automatically extracting bidding requirements based on table structure and semantic understanding according to claim 3, characterized in that, During semantic propagation, a semantic integrity evaluation model is constructed to assess the integrity of the primary labels corresponding to each data cell. The specific steps are as follows: Get the set of parent nodes of the data cell, extract the semantic tags of the parent nodes, perform standardization processing, and construct the set of candidate semantic tags for the current node; Calculate semantic coverage metrics, depth coverage factor, and semantic matching degree based on the candidate semantic tag set; A semantic completeness evaluation model is constructed based on semantic coverage, deep coverage factor, and semantic matching degree, and a semantic completeness index is output. The completeness of the first-level label corresponding to the data cell is evaluated based on the semantic completeness index.

5. The method for automatically extracting bidding requirements based on table structure and semantic understanding according to claim 4, characterized in that, The semantic coverage metric, depth coverage factor, and semantic matching degree are calculated based on the candidate semantic tag set, specifically as follows: The number of valid semantic tags in the first-level tags of the current node is counted. Based on the maximum number of semantic tags that appear in the current business scenario in the historical bidding knowledge base, the semantic coverage index is calculated based on the ratio of valid semantic tags to the maximum number of semantic tags.

6. The method for automatically extracting bidding requirements based on table structure and semantic understanding according to claim 5, characterized in that, By defining the parent node in the DAG as the first-level header node, recording the longest propagation path length from the current node to the parent node, calculating the maximum propagation depth based on the DAG structure, and calculating the depth coverage factor based on the ratio of the longest propagation path length to the maximum propagation depth.

7. The method for automatically extracting bidding requirements based on table structure and semantic understanding according to claim 6, characterized in that, By combining the valid semantic tags of the current node into a first-level semantic vector, and extracting the standard semantic template vector corresponding to the current table type from the historical bidding knowledge base, a semantic matching evaluation mechanism is constructed to calculate the semantic matching degree.

8. The method for automatically extracting bidding requirements based on table structure and semantic understanding according to claim 4, characterized in that, The completeness of the first-level label corresponding to the data cell is evaluated based on the semantic completeness index. Specifically, the semantic completeness index is compared with a preset completeness threshold. If the semantic completeness index is greater than or equal to the preset completeness threshold, the completeness of the first-level label corresponding to the data cell is qualified; if the semantic completeness index is less than the preset completeness threshold, the completeness of the first-level label corresponding to the data cell is unqualified.

9. A system for automatically extracting bidding requirements based on table structure and semantic understanding, applied to the method for automatically extracting bidding requirements based on table structure and semantic understanding as described in any one of claims 1-8, characterized in that, It includes a document input interface, a table structured processing unit, a DAG-level table semantic propagation unit, an integrity assessment unit, a hierarchical extraction unit for five types of entities, and a post-processing and verification unit. The document input interface is used to receive the tender documents to be processed, select the parsing method according to the tender document type to locate the target table area, and generate a set of target table areas to be structured and processed. The table structuring unit is used to perform OCR structuring based on the target table area set, construct the table logical structure, and generate a structured HTML document based on the positional relationship between cells; The DAG hierarchical table semantic propagation unit is used to convert a structured HTML document into a DAG, perform topological sorting and recursively propagate the semantics of the table header to the leaf nodes, label each data cell with a hierarchical semantic tag as a first-level tag, and perform cross-cell semantic alignment and completion. The integrity assessment unit is used to assess the integrity of the first-level labels corresponding to each data cell by constructing a semantic integrity evaluation model during the semantic propagation process. The five-class hierarchical entity extraction unit is used to extract five types of entities for each first-level label that has passed the integrity assessment. It employs a domain-pre-trained encoder, combined with a hierarchical softmax classifier and contrastive learning. The post-processing and verification unit is used to sequentially perform schema structure verification, low-confidence region secondary verification, and rule standardization correction on the entity extraction results, and output the final entity results, which are then encapsulated into a standardized JSON output file.

10. A readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it is used to implement the steps of the automatic extraction method for bidding requirements based on table structure and semantic understanding as described in any one of claims 1-8.

Images