Quality inspection and labeling system for bid invitation file generation based on deep learning

Abstract

The invention discloses a quality inspection and labeling system for bid invitation file generation based on deep learning, and the system comprises a data standardization module which is used for collecting items and historical texts and executing data standardization, and generating a standardized text sequence; the semantic coding and candidate generation module is used for tensorizing the standardized text sequence and generating a candidate text of the bid invitation file; the text decoding module is used for executing limited CRF decoding on the candidate text and outputting an aligned labeling result file; the hypergraph labeling module is used for constructing a hypergraph and globally decoding the hypergraph to generate a multi-level labeling result; the selective re-checking module is used for calculating uncertainty, triggering manual re-checking and outputting a confirmed labeling result and an uncertainty record; the consistency verification module is used for verifying hierarchical consistency to form a final labeling result; and the structured export module is used for exporting the bid invitation file finished product text and the structured annotation data. According to the invention, quality inspection and labeling of bid invitation file generation are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of text labeling, and particularly relates to a quality inspection and labeling system for tender document generation based on deep learning. BACKGROUND

[0002] In the existing tender document generation and management process, most systems still rely on manual editing, template filling or rule-based text splicing methods. Such methods generally have problems of rigid structure and insufficient semantic understanding when dealing with different project types, industry standards and regulatory provisions. Manual editing, although controllable, is inefficient and prone to text omissions, clause reference errors, and format inconsistencies. While template systems improve generation speed, they struggle to meet the differentiated needs of projects, especially in scenarios involving complex provisions, evaluation criteria and legal constraints. In such cases, the generated content often does not meet industry standards or the chapter structure is chaotic. At the same time, existing labeling systems based on natural language processing mostly use single-task models or rule-driven recognition, which have limited ability to recognize multi-level and nested entity structures in tender documents. The lack of contextual dependency constraints between labels leads to inaccurate key field labeling and ambiguous semantic boundaries, affecting automatic file quality inspection and compliance review.

[0003] In addition, traditional Conditional Random Fields (CRF) or Named Entity Recognition (NER) fail to effectively combine document structure and regulatory order in the label prediction process, often relying solely on word-level context for prediction while ignoring the chapter logic relationship of the text, which can result in label boundary crossing or misplacement. For cross and nested entities in complex documents, conventional sequence labeling models struggle to correctly identify hierarchical relationships, making it difficult to accurately locate provisions, sub-provisions and attached conditions. At the same time, existing systems rely heavily on manual sampling for label quality evaluation, lack quantitative judgment of model uncertainty, and cannot effectively identify high-risk label areas, leading to high manual review costs, low labeling reliability, and the inability to achieve dynamic quality control and continuous optimization.

[0004] Therefore, how to provide a quality inspection and labeling system for tender document generation based on deep learning is a problem that needs to be solved by those skilled in the art. SUMMARY

[0005] One object of the present application is to provide a quality inspection and labeling system for tender document generation based on deep learning. The present application uses restricted CRF, hierarchical hypergraph CRF and evidence-driven review to ensure legal label order, accurate nested entities and controllable risks. By triggering a small amount of manual intervention based on uncertainty, the present application significantly reduces labeling costs and improves efficiency. At the same time, the present application outputs finished texts and structured data, facilitating review and traceability, system integration and batch landing applications.

[0006] A quality inspection and annotation system for tender document generation based on deep learning, according to an embodiment of the present invention, includes: The data standardization module is used to collect project and historical texts and perform data standardization to generate a standardized text sequence. The semantic encoding and candidate generation module is used to tensor standardized text sequences and generate candidate texts for tender documents through deep encoding and decoding. The text decoding module is used to load order constraints and perform restricted CRF decoding on candidate text, outputting an aligned annotation result file; The hypergraph annotation module is used to construct a hypergraph based on hierarchical relationships and globally decode it to generate multi-level annotation results. The selective review module is used to calculate uncertainty and trigger manual review, and output the confirmed annotation results and uncertainty records. The consistency verification module is used to perform label normalization, boundary calibration, and hierarchical consistency verification to generate the final annotation results. The structured export module is used to backfill key elements and generate chapter indexes, exporting the finished text of the tender document and structured annotation data.

[0007] Optionally, modules can be integrated using the following methods: Data preparation was carried out using raw corpora including basic project information, historical bidding documents and industry rule texts. The raw corpora were then standardized to form a standardized text sequence. Based on standardized text sequences, a deep neural network encoder is used to vectorize the text and generate candidate texts for tender documents. A multi-task NER semantic annotation model is constructed based on candidate texts, consisting of three structural layers: text decoding layer, hyperlink annotation layer, and evidence evaluation layer. For candidate texts, finite state automata-constrained CRF decoding is used in the text decoding layer for sequence labeling. Label transitions that do not conform to the order are masked to obtain the labeling result file. Under the joint constraints of the annotation result file and the candidate text, the hypergraph annotation layer is enabled and the hierarchical hypergraph CRF method is used for global decoding. Hierarchical annotation is performed on nested and overlapping entities to generate multi-level annotation results. Quality measurement is carried out around the multi-level annotation results. At the evidence evaluation level, the Dirichlet CRF method of evidence theory is used to evaluate uncertainty. Based on the preset threshold, manual review and selective sample collection are triggered and the review conclusions are combined to form the confirmed annotation results. Based on the confirmed annotation results and uncertainty records, label normalization and boundary calibration, hierarchical consistency verification and conflict resolution are performed to obtain the final annotation results of the batch documents; Based on the final annotation results and candidate texts, generate the final tender document text and structured annotation data, which include annotations of key elements and chapter structure indexes.

[0008] Optionally, the formation of the standardized text sequence specifically includes: Text is collected in batches from project platforms and enterprise document libraries. Basic project information, historical bidding documents and industry rule texts are registered as raw corpus. Character encoding and line terminators are unified, control characters and abnormal whitespace are removed, original paragraphs and hierarchical marks are retained, empty documents and damaged documents are removed, and document-level deduplication is performed to form an initial screening set. The initial screening set is formatted uniformly, including uniform Chinese and English punctuation and character width, and standardized expression of numbers, dates and measurements. Domain-specific word segmentation is used to complete basic annotation and retain necessary structural labels. Irrelevant items are cleaned up according to the rules of disuse and low frequency. Common recognition errors are corrected in batches according to the error correction table. Annotated sentence sequence and word list are output. The system performs duplicate and near-duplicate detection on the annotated sentence sequence and merges similar content. It also cleans up layout noise and header / footer interference, constructs a structured thesaurus and transcribs the row and column relationships, standardizes chapter numbering and list styles, regulates indentation and paragraph spacing, and generates standardized text sequences, index views, and quality records.

[0009] Optionally, the process of generating the candidate text of the tender document specifically includes: Receive a standardized text sequence, divide it into fixed-size text blocks according to paragraph and sentence boundaries, establish a batch processing list, complete vocabulary mapping and index encoding, generate the input text tensor for the deep neural network, and record the correspondence between the text blocks and the original text positions; The deep neural network encoder is invoked to perform vectorization processing on the input text tensor. The deep neural network encoder consists of an embedding layer, a position encoding layer, a multi-head attention layer and a feedforward transformation layer. It sequentially completes word embedding mapping, position label addition, context dependency capture and semantic feature extraction. Through nonlinear activation and inter-layer normalization, a multi-dimensional semantic representation output is obtained to form a document semantic vector for text generation, and the batch index is retained to maintain the sample order. The document semantic vector is input into the text generation decoder, which generates candidate text for the tender document sentence by sentence. Punctuation and whitespace normalization, proper noun and numerical format restoration, paragraph merging and title placeholder replacement are performed. The structure is proofread according to length and integrity rules, and the candidate text for the tender document is output.

[0010] Optionally, the construction of the multi-task NER semantic annotation model specifically includes: Receive candidate text and corresponding semantic vectors, build multi-level indexes of documents, paragraphs, sentences and words, generate batch processing list and task number, load entity tag list and chapter level list, load domain thesaurus, chapter order list and clause mapping table, complete bidirectional mapping from text to index, configure maximum sequence length and truncation strategy, generate effective bit mask and padding rules, initialize cache, log directory and checkpoint path, verify input integrity and output ready data packet; A three-layer structure for constructing a multi-task NER semantic annotation model based on ready data packets consists of a text decoding layer, a hyperlink annotation layer, and an evidence evaluation layer. The input and output ports of each layer are bound according to a unified tensor specification. Parameter initialization strategies and random seeds are set, and three types of control data are generated at the same time: a list of label transfer constraints, a list of hierarchical relationships, and uncertainty evaluation rules. Candidate text and semantic vectors are sent to the text decoding layer via a data pipeline as initial input.

[0011] Optionally, obtaining the annotation result file specifically includes: The text decoding layer in the multi-task NER semantic annotation model is invoked to read candidate texts and their corresponding semantic vectors, load the label transfer constraint list and chapter order list, segment the candidate texts into observation sequences at the word granularity, and establish document, sentence and word-level index mappings to form a constraint resource list; In the text decoding layer, a finite state automaton constraint is established based on the constraint resource list. The chapter order and label constraint relationship of the candidate text are formally defined, and the logical order of the text is mapped to a finite set of states. Allowed label transition relationships are registered as legal state transitions, and disallowed label transition relationships are registered as illegal state transitions. A restricted state transition table is constructed and embedded with the transition matrix of the conditional random field. When performing label prediction, the text decoding layer calls the restricted state transition table at each position while calculating the label score to verify the legality of the path in real time. Paths that do not meet the state transition conditions are marked with a mask, and paths that meet the conditions are retained and their scores are accumulated. Thus, the path search process of the conditional random field is constrained by the finite state automaton, forming a restricted decoding context. The system performs restricted Viterbi decoding with the observation sequence and restricted decoding context as input. It unfolds step by step according to the legal path recorded in the restricted decoding context. The label selection at each step is dynamically planned only within the set of valid states until the end of the sequence. The legal path with the highest score is selected as the optimal output to form the annotation result. At the same time, state convergence and sequence consistency verification are performed to ensure that the decoding path enters the set of final states from the starting state. The output is an annotation result file aligned with the document, paragraph, sentence and word index.

[0012] Optionally, the process of generating the multi-level annotation results specifically includes: Read the annotation result file and candidate text, load the hierarchical relationship list, align the text fragments corresponding to each label in the annotation result with the candidate text, extract text segments with nested relationships, generate an initial set of entity nodes, and at the same time, according to the hierarchical relationship list, combine entities with adjacent, contained and parallel relationships into parent node and child node pairs to form a node-relationship mapping table, and output the hypergraph construction data package; In the hypergraph annotation layer, a hierarchical hypergraph structure model is established by taking the hypergraph construction data package as input. Each entity node is regarded as a semantic representation of a text fragment. Node pairs with inclusion or parallel relationships are connected by hyperedges to form a hypergraph network with parent-child parallel multi-level links. Each node carries a context feature vector, and each hyperedge corresponds to hierarchical dependency constraints and mutual influence weights. In the computation stage, the potential function of the conditional random field is used to jointly model the annotation score within the node and the association score between hyperedges to form a hierarchical hypergraph-CRF representation framework. Global decoding is performed in the hierarchical hypergraph-CRF structure. Joint reasoning is performed on the annotation status of all nodes and the node-relationship mapping table. The label consistency and boundary dependency relationship between parent and child nodes are optimized layer by layer. Hierarchical annotation is performed on nested or overlapping entity regions. Annotation results that conform to multi-level logic are output and aligned with candidate text indexes to generate multi-level annotation results.

[0013] Optionally, the process of forming the confirmed annotation results specifically includes: Read the multi-level annotation result file and candidate text, load the uncertainty assessment rules, and collect the current label, label candidate list, predicted probability, context window text, parent-child and parallel relationship identifiers of each level node into an evidence input data package. Establish a Dirichlet parameter set for each node, record the evidence strength and total evidence amount of each label in the label order, map negative values ​​to non-negative values ​​and perform normalization processing, and retain the node identifier, text start and end position, level, timestamp and source channel to form a node-level evidence ledger. Based on the node-level evidence ledger, the upper and lower bounds of the confidence interval are calculated for each node according to the Dirichlet parameters. The interval width, the median of the interval, and the current label position are recorded. The evidence difference degree is calculated for adjacent nodes and parent-child nodes. The structural consistency score is generated by combining the maximum difference and the average difference within the window. The confidence interval width and the structural consistency score are weighted and combined into an uncertainty score. The threshold system in the uncertainty assessment rules is called for comparison. The threshold system includes a basic threshold set according to the label category, a correction coefficient set according to the level, and a minimum trigger line set according to the text type. Nodes with scores higher than the threshold are marked as high uncertainty, and nodes with scores not higher than the threshold are marked as stable. At the same time, uncertainty records are generated. For nodes marked as highly uncertain, a selective manual review queue is established. Review tasks are issued in order of score from high to low, key information first, and cross-layer impact first. Corresponding text fragments, context windows, and candidate labels are extracted for manual confirmation. After collecting the review conclusions, confirmation labels and new confidence levels are generated by combining them with the original predictions according to the evidence merging strategy. The evidence merging strategy includes increasing the evidence strength of the corresponding label when the review conclusion is consistent with the original prediction, prioritizing the review conclusion and lowering the evidence strength of the original prediction when the review conclusion is inconsistent with the original prediction, and deciding based on the priority of the parent node and the importance of the label when there are parallel conclusions. The confirmed labeling results are then output.

[0014] Optionally, obtaining the final annotation result specifically includes: Read the confirmed annotation results and uncertainty records, load the candidate text, unify the label name, category code, symbols and whitespace, merge and replace synonymous labels, perform batch correction of format anomalies, and generate a standardized label list. Based on the standardized tag list, boundary calibration is performed on the start and end positions of each tag. Word segmentation boundary snapping is used to ensure that the start and end are aligned with legal segmentation points. The continuity of quantity phrases is restored by grouping numbers and units of measurement. The calibrated boundary mapping table is output. Based on the calibrated boundary mapping table, hierarchical consistency verification and conflict resolution are carried out to detect parent-child overlap, parallel mutual exclusion and cross-level affiliation conflicts. The primary labels are retained according to hierarchical priority and confidence, the range of subordinate labels is adjusted and duplicates and overlaps are removed, and after sorting and renumbering, the sub-results are merged to form the final annotation result of the batch documents.

[0015] Optionally, the output of the finished text of the tender document and the structured annotation data specifically includes: Based on the final annotation results and candidate text, the labels are aligned with the original text positions one by one, and the manuscript skeleton is generated according to chapters, sections, articles, and clauses. The key elements are filled in according to the correspondence between the labels and fields to form the finished draft. Based on the finished draft, generate chapter structure index and element index, establish multi-level anchor points for documents, chapters, clauses, sentences and words, output searchable bookmarks and jump links, export structured labeled dataset, including field name, field value, text start and end position, level and tag type, and synchronously write version number, timestamp and source record; Perform consistency and integrity checks on the final draft and structured annotation data, detect missing, duplicate, and out-of-bounds items, finalize the layout and export the file, and generate an archiveable final tender document text and structured annotation data.

[0016] The beneficial effects of this invention are: This invention utilizes a conditional random field decoding mechanism constrained by finite state automata to incorporate chapter order and label transfer logic into the constraint scope, ensuring that the generated results conform to industry standards and text logic at the structural level. This significantly reduces label out-of-bounds, chapter misalignment, and clause confusion, fundamentally improving the accuracy and consistency of annotation.

[0017] This invention achieves global modeling and joint reasoning of nested entities and multi-level clause relationships in bidding documents through a hierarchical hypergraph CRF model, enabling the accurate identification and expression of complex structures such as parent-child clauses, parallel items, and cross-references, thus overcoming the limitations of traditional sequence labeling models in hierarchical text processing.

[0018] This invention introduces an uncertainty assessment mechanism based on Dirichlet evidence theory, which dynamically quantifies the confidence interval of the model output. When a high uncertainty region is detected, selective manual review is automatically triggered, making human-machine collaboration the core link of quality control. This ensures the reliability of the system output and reduces the burden of manual review.

[0019] After annotation, this invention ensures that the final output document content fully complies with industry standards in terms of semantics, structure, and layout through a process of standardization, boundary calibration, and consistency verification. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0021] Fig. 1 This is a flowchart of a quality inspection and annotation system for generating tender documents based on deep learning, as proposed in this invention. Fig. 2 This is a schematic diagram of a quality inspection and annotation system for generating tender documents based on deep learning, as proposed in this invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0023] refer to Figs. 1-2 A quality inspection and annotation system for tender document generation based on deep learning, comprising: The data standardization module is used to collect project and historical texts and perform data standardization to generate a standardized text sequence. The semantic encoding and candidate generation module is used to tensor standardized text sequences and generate candidate texts for tender documents through deep encoding and decoding. The text decoding module is used to load order constraints and perform restricted CRF decoding on candidate text, outputting an aligned annotation result file; The hypergraph annotation module is used to construct a hypergraph based on hierarchical relationships and globally decode it to generate multi-level annotation results. The selective review module is used to calculate uncertainty and trigger manual review, and output the confirmed annotation results and uncertainty records. The consistency verification module is used to perform label normalization, boundary calibration, and hierarchical consistency verification to generate the final annotation results. The structured export module is used to backfill key elements and generate chapter indexes, exporting the finished text of the tender document and structured annotation data.

[0024] In this embodiment, the modules are interconnected using the following method: Data preparation was carried out using raw corpora including basic project information, historical bidding documents and industry rule texts. The raw corpora were then standardized to form a standardized text sequence. Based on standardized text sequences, a deep neural network encoder is used to vectorize the text and generate candidate texts for tender documents. A multi-task NER semantic annotation model is constructed based on candidate texts, consisting of three structural layers: text decoding layer, hyperlink annotation layer, and evidence evaluation layer. For candidate texts, finite state automata-constrained CRF decoding is used in the text decoding layer for sequence labeling. Label transitions that do not conform to the order are masked to obtain the labeling result file. Under the joint constraints of the annotation result file and the candidate text, the hypergraph annotation layer is enabled and the hierarchical hypergraph CRF method is used for global decoding. Hierarchical annotation is performed on nested and overlapping entities to generate multi-level annotation results. Quality measurement is carried out around the multi-level annotation results. At the evidence evaluation level, the Dirichlet CRF method of evidence theory is used to evaluate uncertainty. Based on the preset threshold, manual review and selective sample collection are triggered and the review conclusions are combined to form the confirmed annotation results. Based on the confirmed annotation results and uncertainty records, label normalization and boundary calibration, hierarchical consistency verification and conflict resolution are performed to obtain the final annotation results of the batch documents; Based on the final annotation results and candidate texts, generate the final tender document text and structured annotation data, which include annotations of key elements and chapter structure indexes.

[0025] In this embodiment, the formation of the standardized text sequence specifically includes: Text is collected in batches from project platforms and enterprise document libraries. Basic project information, historical bidding documents and industry rule texts are registered as raw corpus. Character encoding and line terminators are unified, control characters and abnormal whitespace are removed, original paragraphs and hierarchical marks are retained, empty documents and damaged documents are removed, and document-level deduplication is performed to form an initial screening set. The initial screening set is formatted uniformly, including uniform Chinese and English punctuation and character width, and standardized expression of numbers, dates and measurements. Domain-specific word segmentation is used to complete basic annotation and retain necessary structural labels. Irrelevant items are cleaned up according to the rules of disuse and low frequency. Common recognition errors are corrected in batches according to the error correction table. Annotated sentence sequence and word list are output. The system performs duplicate and near-duplicate detection on the annotated sentence sequence and merges similar content. It also cleans up layout noise and header / footer interference, constructs a structured vocabulary and transcribs the row and column relationships, standardizes chapter numbering and list styles, regulates indentation and paragraph spacing, and generates standardized text sequences, index views, and quality records.

[0026] In this embodiment, the process of generating the candidate text of the tender document specifically includes: Receive a standardized text sequence, divide it into fixed-size text blocks according to paragraph and sentence boundaries, establish a batch processing list, complete vocabulary mapping and index encoding, generate the input text tensor for the deep neural network, and record the correspondence between the text blocks and the original text positions; The deep neural network encoder is invoked to perform vectorization processing on the input text tensor. The deep neural network encoder consists of an embedding layer, a position encoding layer, a multi-head attention layer and a feedforward transformation layer. It sequentially completes word embedding mapping, position label addition, context dependency capture and semantic feature extraction. Through nonlinear activation and inter-layer normalization, a multi-dimensional semantic representation output is obtained to form a document semantic vector for text generation, and the batch index is retained to maintain the sample order. The document semantic vector is input into the text generation decoder, which generates candidate text for the tender document sentence by sentence. Punctuation and whitespace normalization, proper noun and numerical format restoration, paragraph merging and title placeholder replacement are performed. The structure is proofread according to length and integrity rules, and the candidate text for the tender document is output.

[0027] In this embodiment, the construction of the multi-task NER semantic annotation model specifically includes: Receive candidate text and corresponding semantic vectors, build multi-level indexes of documents, paragraphs, sentences and words, generate batch processing list and task number, load entity tag list and chapter level list, load domain thesaurus, chapter order list and clause mapping table, complete bidirectional mapping from text to index, configure maximum sequence length and truncation strategy, generate effective bit mask and padding rules, initialize cache, log directory and checkpoint path, verify input integrity and output ready data packet; A three-layer structure for constructing a multi-task NER semantic annotation model based on ready data packets consists of a text decoding layer, a hyperlink annotation layer, and an evidence evaluation layer. The input and output ports of each layer are bound according to a unified tensor specification. Parameter initialization strategies and random seeds are set, and three types of control data are generated at the same time: a list of label transfer constraints, a list of hierarchical relationships, and uncertainty evaluation rules. Candidate text and semantic vectors are sent to the text decoding layer via a data pipeline as initial input.

[0028] In this embodiment, obtaining the annotation result file specifically includes: The text decoding layer in the multi-task NER semantic annotation model is invoked to read candidate texts and their corresponding semantic vectors, load the label transfer constraint list and chapter order list, segment the candidate texts into observation sequences at the word granularity, and establish document, sentence and word-level index mappings to form a constraint resource list; In the text decoding layer, a finite state automaton constraint is established based on the constraint resource list. The chapter order and label constraint relationship of the candidate text are formally defined, and the logical order of the text is mapped to a finite set of states. Allowed label transition relationships are registered as legal state transitions, and disallowed label transition relationships are registered as illegal state transitions. A restricted state transition table is constructed and embedded with the transition matrix of the conditional random field. When performing label prediction, the text decoding layer calls the restricted state transition table at each position while calculating the label score to verify the legality of the path in real time. Paths that do not meet the state transition conditions are marked with a mask, and paths that meet the conditions are retained and their scores are accumulated. Thus, the path search process of the conditional random field is constrained by the finite state automaton, forming a restricted decoding context. The system performs restricted Viterbi decoding with the observation sequence and restricted decoding context as input. It unfolds step by step according to the legal path recorded in the restricted decoding context. The label selection at each step is dynamically planned only within the set of valid states until the end of the sequence. The legal path with the highest score is selected as the optimal output to form the annotation result. At the same time, state convergence and sequence consistency verification are performed to ensure that the decoding path enters the set of final states from the starting state. The output is an annotation result file aligned with the document, paragraph, sentence and word index.

[0029] This invention introduces CRF decoding with finite state automata constraints into the text decoding layer, explicitly embedding chapter order and tag transfer rules into path search, filtering out illegal transfers in real time and retaining only the compliant optimal path, significantly reducing the risk of out-of-order and out-of-bounds labeling, improving the accuracy and consistency of key element recognition, reducing post-processing workload, ensuring that the results are strictly aligned with the document index and have traceability.

[0030] In this embodiment, the process of generating the multi-level annotation results specifically includes: Read the annotation result file and candidate text, load the hierarchical relationship list, align the text fragments corresponding to each label in the annotation result with the candidate text, extract text segments with nested relationships, generate an initial set of entity nodes, and at the same time, according to the hierarchical relationship list, combine entities with adjacent, contained and parallel relationships into parent node and child node pairs to form a node-relationship mapping table, and output the hypergraph construction data package; In the hypergraph annotation layer, a hierarchical hypergraph structure model is established by taking the hypergraph construction data package as input. Each entity node is regarded as a semantic representation of a text fragment. Node pairs with inclusion or parallel relationships are connected by hyperedges to form a hypergraph network with parent-child parallel multi-level links. Each node carries a context feature vector, and each hyperedge corresponds to hierarchical dependency constraints and mutual influence weights. In the computation stage, the potential function of the conditional random field is used to jointly model the annotation score within the node and the association score between hyperedges to form a hierarchical hypergraph-CRF representation framework. Global decoding is performed in the hierarchical hypergraph-CRF structure. Joint reasoning is performed on the annotation status of all nodes and the node-relationship mapping table. The label consistency and boundary dependency relationship between parent and child nodes are optimized layer by layer. Hierarchical annotation is performed on nested or overlapping entity regions. Annotation results that conform to multi-level logic are output and aligned with candidate text indexes to generate multi-level annotation results.

[0031] This invention, based on hierarchical hypergraph CRF, unifies entities and parent-child parallel relationships into a structured expression that can be jointly reasoned, so that nesting and overlapping boundaries are consistently determined under global constraints, significantly improving the recall and accuracy of clause element extraction, reducing cross-level mislabeling and duplicate labeling, and outputting multi-level results aligned with text positions.

[0032] In this embodiment, the process of forming the confirmed annotation result specifically includes: Read the multi-level annotation result file and candidate text, load the uncertainty assessment rules, and collect the current label, label candidate list, predicted probability, context window text, parent-child and parallel relationship identifiers of each level node into an evidence input data package. Establish a Dirichlet parameter set for each node, record the evidence strength and total evidence amount of each label in the label order, map negative values ​​to non-negative values ​​and perform normalization processing, and retain the node identifier, text start and end position, level, timestamp and source channel to form a node-level evidence ledger. Based on the node-level evidence ledger, the upper and lower bounds of the confidence interval are calculated for each node according to the Dirichlet parameters. The interval width, the median of the interval, and the current label position are recorded. The evidence difference degree is calculated for adjacent nodes and parent-child nodes. The structural consistency score is generated by combining the maximum difference and the average difference within the window. The confidence interval width and the structural consistency score are weighted and combined into an uncertainty score. The threshold system in the uncertainty assessment rules is called for comparison. The threshold system includes a basic threshold set according to the label category, a correction coefficient set according to the level, and a minimum trigger line set according to the text type. Nodes with scores higher than the threshold are marked as high uncertainty, and nodes with scores not higher than the threshold are marked as stable. At the same time, uncertainty records are generated. For nodes marked as highly uncertain, a selective manual review queue is established. Review tasks are issued in order of score from high to low, key information first, and cross-layer impact first. Corresponding text fragments, context windows, and candidate labels are extracted for manual confirmation. After collecting the review conclusions, confirmation labels and new confidence levels are generated by combining them with the original predictions according to the evidence merging strategy. The evidence merging strategy includes increasing the evidence strength of the corresponding label when the review conclusion is consistent with the original prediction, prioritizing the review conclusion and lowering the evidence strength of the original prediction when the review conclusion is inconsistent with the original prediction, and deciding based on the priority of the parent node and the importance of the label when there are parallel conclusions. The confirmed labeling results are then output.

[0033] This invention introduces Dirichlet-based evidence modeling to quantify and label uncertainty and generate uncertainty records. Combined with threshold-driven selective manual review and evidence merging strategies, it can significantly improve the accuracy of high-risk segments, reduce overall review costs, and form a traceable quality loop while maintaining structural consistency.

[0034] In this embodiment, obtaining the final annotation result specifically includes: Read the confirmed annotation results and uncertainty records, load the candidate text, unify the label name, category code, symbols and whitespace, merge and replace synonymous labels, perform batch correction of format anomalies, and generate a standardized label list. Based on the standardized tag list, boundary calibration is performed on the start and end positions of each tag. Word segmentation boundary snapping is used to ensure that the start and end are aligned with legal segmentation points. The continuity of quantity phrases is restored by grouping numbers and units of measurement. The calibrated boundary mapping table is output. Based on the calibrated boundary mapping table, hierarchical consistency verification and conflict resolution are carried out to detect parent-child overlap, parallel mutual exclusion and cross-level affiliation conflicts. The primary labels are retained according to hierarchical priority and confidence, the range of subordinate labels is adjusted and duplicates and overlaps are removed, and after sorting and renumbering, the sub-results are merged to form the final annotation result of the batch documents.

[0035] In this embodiment, the output of the finished text of the tender document and the structured annotation data specifically includes: Based on the final annotation results and candidate text, the labels are aligned with the original text positions one by one, and the manuscript skeleton is generated according to chapters, sections, articles, and clauses. The key elements are filled in according to the correspondence between the labels and fields to form the finished draft. Based on the finished draft, generate chapter structure index and element index, establish multi-level anchor points for documents, chapters, clauses, sentences and words, output searchable bookmarks and jump links, export structured labeled dataset, including field name, field value, text start and end position, level and tag type, and synchronously write version number, timestamp and source record; Perform consistency and integrity checks on the final draft and structured annotation data, detect missing, duplicate, and out-of-bounds items, finalize the layout and export the file, and generate an archiveable final tender document text and structured annotation data.

[0036] Example 1: To verify the feasibility of this invention in practice, it was applied to an automated generation and intelligent quality inspection scenario for a bidding document. The experiment selected historical data covering three types of projects: construction engineering, equipment procurement, and service outsourcing, totaling 1860 documents with approximately 4.6 million words of text. The original data suffered from inconsistent formats, overlapping clauses, and missing numbering. Traditional processing methods primarily relied on manual editing and template splicing, supplemented by regular expressions and rule bases for field filling and format revision. These methods have limited ability to express cross-chapter references and hierarchical relationships between clauses, making it difficult to reliably handle the diverse language and complex structures of projects, often resulting in misalignment, omissions, and duplicate annotations.

[0037] Deploying the system of this invention on the same dataset, the data standardization module first completes unified encoding, sentence segmentation, noise reduction, error correction, and near-duplication merging. Then, the semantic encoding and candidate generation module outputs candidate texts that conform to the style guidelines. In the annotation chain, the text decoding module introduces CRF decoding constrained by finite state automata to explicitly shield non-compliant label transfers, ensuring sequence consistency and compliance. The hypergraph annotation module performs joint reasoning on parent-child, parallel, and inclusion relationships using hierarchical hypergraph CRF, accurately handling nested and overlapping entities. The evidence verification module calculates uncertainty based on Dirichlet-CRF and triggers selective manual verification, enabling targeted verification and evidence merging of high-risk segments. Finally, the consistency verification and structured export module outputs the finished text and a compatible dataset.

[0038] To differentiate itself from traditional methods, this invention synchronously replicates the common process of combining rules and templates on the same batch of data. First, a framework is generated based on a general template; then, the rule base matches fields and performs format replacement; finally, manual review and revision are performed. This approach exhibits significant rule conflicts and coverage blind spots when handling multi-layered clauses, cross-chapter citations, and custom terms, requiring substantial manual backfilling and multiple rounds of proofreading. The manual review is intensive, and consistency depends heavily on experience. Table 1 provides a comparison of the two methods across key metrics: Table 1 Performance Comparison of the Invention System and Traditional Methods

[0039]

[0040] As shown in Table 1, this invention significantly outperforms traditional methods in key end-to-end metrics. The average processing time per document is reduced from 4.6 seconds to 0.72 seconds, an efficiency improvement of 84.3%; the text generation completeness rate increases from 80.8% to 93.5%, an improvement of 12.7 percentage points; the tag order consistency rate increases from 88.2% to 96.9%, an improvement of 8.7 percentage points; the nested entity recognition accuracy rate increases from 85.9% to 94.1%, an improvement of 9.5 percentage points; the hierarchical consistency error decreases from 7.1% to 2.1%, a reduction of 70.4%; the manual review time decreases from 24.0 minutes to 6.3 minutes, a reduction of 73.5%; and the overall annotation accuracy rate increases from 87.2% to 97.8%, improving overall processing efficiency from 1.0 times to 6.4 times. The comparative results show that this invention achieves stable improvements in key metrics such as generation completeness rate, tag order consistency, nested entity recognition, and manual review time, achieving a multiple-fold increase in processing efficiency while retaining traceable uncertainty records to meet compliance review requirements.

[0041] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A quality inspection and annotation system for tender document generation based on deep learning, characterized in that, include: The data standardization module is used to collect project and historical texts and perform data standardization to generate a standardized text sequence. The semantic encoding and candidate generation module is used to tensor standardized text sequences and generate candidate texts for tender documents through deep encoding and decoding. The text decoding module is used to load order constraints and perform restricted CRF decoding on candidate text, outputting an aligned annotation result file; The hypergraph annotation module is used to construct a hypergraph based on hierarchical relationships and globally decode it to generate multi-level annotation results. The selective review module is used to calculate uncertainty and trigger manual review, and output the confirmed annotation results and uncertainty records. The consistency verification module is used to perform label normalization, boundary calibration, and hierarchical consistency verification to generate the final annotation results. The structured export module is used to backfill key elements and generate chapter indexes, exporting the finished text of the tender document and structured annotation data.

2. The quality inspection and annotation system for tender document generation based on deep learning according to claim 1, characterized in that, The modules are connected in the following way: Data preparation is carried out based on the original corpus, and data standardization is performed to form a standardized text sequence; Based on standardized text sequences, a deep neural network encoder is used to vectorize the text and generate candidate texts for tender documents. A multi-task NER semantic annotation model is constructed based on candidate texts, consisting of three structural layers: text decoding layer, hyperlink annotation layer, and evidence evaluation layer. For candidate texts, finite state automata-constrained CRF decoding is used in the text decoding layer for sequence labeling. Label transitions that do not conform to the order are masked to obtain the labeling result file. Under the joint constraints of the annotation result file and the candidate text, the hypergraph annotation layer is enabled and the hierarchical hypergraph CRF method is used for global decoding. Hierarchical annotation is performed on nested and overlapping entities to generate multi-level annotation results. Quality measurement is carried out around the multi-level annotation results. At the evidence evaluation level, the Dirichlet CRF method of evidence theory is used to evaluate uncertainty and form confirmed annotation results. Based on the confirmed annotation results and uncertainty records, label normalization and boundary calibration, hierarchical consistency verification and conflict resolution are performed to obtain the final annotation results; Based on the final annotation results and candidate texts, generate the final tender document text and structured annotation data.

3. The quality inspection and annotation system for tender document generation based on deep learning according to claim 2, characterized in that, The formation of the standardized text sequence specifically includes: Text is collected in batches from project platforms and enterprise document libraries. Basic project information, historical bidding documents and industry rule texts are registered as raw corpus. Character encoding and line terminators are unified, control characters and abnormal whitespace are removed, original paragraphs and hierarchical marks are retained, empty documents and damaged documents are removed, and document-level deduplication is performed to form an initial screening set. The initial screening set is formatted uniformly, and the domain-specific word segmentation is called to complete the basic annotation while retaining the necessary structural labels. The annotated sentence sequence and vocabulary are then output. The system performs duplicate and near-duplicate detection on the annotated sentence sequence and merges similar content. It also constructs a structured vocabulary and transcribs the row and column relationships, standardizes chapter numbering and list styles, regulates indentation and paragraph spacing, and generates standardized text sequences, index views, and quality records.

4. The quality inspection and annotation system for tender document generation based on deep learning according to claim 2, characterized in that, The process of generating the candidate text of the tender document specifically includes: Receive a standardized text sequence, divide it into fixed-size text blocks according to paragraph and sentence boundaries, establish a batch processing list, complete vocabulary mapping and index encoding, and generate the input text tensor for the deep neural network; A deep neural network encoder is invoked to perform vectorization processing on the input text tensor. The deep neural network encoder consists of an embedding layer, a position encoding layer, a multi-head attention layer, and a feedforward transformation layer. It sequentially completes word embedding mapping, position tag addition, context dependency capture, and semantic feature extraction. Through nonlinear activation and inter-layer normalization, a multi-dimensional semantic representation output is obtained to form a document semantic vector. The document semantic vector is input into the text generation decoder, which generates candidate text for the tender document sentence by sentence. The structure is proofread according to the length and integrity rules, and the candidate text for the tender document is output.

5. The quality inspection and annotation system for tender document generation based on deep learning according to claim 2, characterized in that, The construction of the multi-task NER semantic annotation model specifically includes: Receive candidate text and corresponding semantic vector, build multi-level index, generate batch processing list and task number, load chapter order table, complete bidirectional mapping from text to index, generate valid bit mask and padding rules, verify input integrity and output ready data packet; A three-layer structure for constructing a multi-task NER semantic annotation model based on ready data packets consists of a text decoding layer, a hyperlink annotation layer, and an evidence evaluation layer. The input and output ports of each layer are bound according to a unified tensor specification. Parameter initialization strategies and random seeds are set, and three types of control data are generated at the same time: a list of label transfer constraints, a list of hierarchical relationships, and uncertainty evaluation rules. Candidate text and semantic vectors are sent to the text decoding layer via a data pipeline as initial input.

6. The quality inspection and annotation system for tender document generation based on deep learning according to claim 2, characterized in that, The specific steps involved in obtaining the annotation result file are as follows: The text decoding layer in the multi-task NER semantic annotation model is invoked to read candidate texts and their corresponding semantic vectors, load the label transfer constraint list and chapter order list, segment the candidate texts into observation sequences at the word granularity, and establish document, sentence and word-level index mappings to form a constraint resource list; In the text decoding layer, a finite state automaton constraint is established based on the constraint resource list. The chapter order and label constraint relationship of the candidate text are formally defined, and the logical order of the text is mapped to a finite set of states. A restricted state transition table is constructed and the transition matrix of the conditional random field is embedded. When performing label prediction, the text decoding layer calls the restricted state transition table at each position while calculating the label score. The legality of the path is verified in real time. Paths that do not meet the state transition conditions are marked with a mask, and paths that meet the conditions are retained and their scores are accumulated to form a restricted decoding context. The system performs restricted Viterbi decoding with the observation sequence and restricted decoding context as input. It unfolds step by step according to the legal path recorded in the restricted decoding context. The label selection at each step is dynamically planned only within the set of valid states until the end of the sequence. The legal path with the highest score is selected as the optimal output to form the labeling result. At the same time, state convergence and sequence consistency verification are performed, and the labeling result file is output.

7. The quality inspection and annotation system for tender document generation based on deep learning according to claim 2, characterized in that, The process of generating the multi-level annotation results specifically includes: Read the annotation result file and candidate text, load the hierarchical relationship list, align the text fragments corresponding to each label in the annotation result with the candidate text, extract text segments with nested relationships, generate an initial set of entity nodes, and at the same time, according to the hierarchical relationship list, combine entities with adjacent, contained and parallel relationships into parent node and child node pairs to form a node-relationship mapping table, and output the hypergraph construction data package; In the hypergraph annotation layer, a hierarchical hypergraph structure model is established by taking the hypergraph construction data package as input. Each entity node is regarded as a semantic representation of a text fragment. Node pairs with inclusion or parallel relationships are connected by hyperedges to form a hypergraph network with parent-child parallel multi-level links. Each node carries a context feature vector, and each hyperedge corresponds to hierarchical dependency constraints and mutual influence weights. In the computation stage, the potential function of the conditional random field is used to jointly model the annotation score within the node and the association score between hyperedges to form a hierarchical hypergraph-CRF representation framework. Global decoding is performed in the hierarchical hypergraph-CRF structure. Joint reasoning is performed on the annotation status of all nodes and the node-relationship mapping table. The label consistency and boundary dependency relationship between parent and child nodes are optimized layer by layer. Hierarchical annotation is performed on nested or overlapping entity regions. Annotation results that conform to multi-level logic are output and aligned with candidate text indexes to generate multi-level annotation results.

8. The quality inspection and annotation system for tender document generation based on deep learning according to claim 2, characterized in that, The process of forming the confirmed annotation results specifically includes: Read the multi-level annotation result file and candidate text, load the uncertainty assessment rules, and compile the current label, label candidate list, prediction probability, context window text, parent-child and parallel relationship identifiers of each level node into an evidence input data package. Establish a Dirichlet parameter set for each node, record the evidence strength and total evidence amount of each label in label order, and form a node-level evidence ledger. Based on the node-level evidence ledger, the upper and lower bounds of the confidence interval are calculated for each node according to the Dirichlet parameters. The interval width, the median of the interval, and the current label position are recorded. The evidence difference degree is calculated for adjacent nodes and parent-child nodes, and a structural consistency score is generated. The confidence interval width and the structural consistency score are weighted and combined into an uncertainty score. The threshold system in the uncertainty assessment rules is called for comparison. Nodes with scores higher than the threshold are marked as high uncertainty, and nodes with scores not higher than the threshold are marked as stable. At the same time, uncertainty records are generated. For nodes marked as high uncertainty, a selective manual review queue is established. Review tasks are issued in order of score from high to low, key information first, and cross-layer impact first. Corresponding text fragments, context windows, and candidate labels are extracted for manual confirmation. After collecting the review conclusions, confirmation labels and new confidence levels are generated by combining them with the original predictions according to the evidence merging strategy. The confirmed labeling results are then output.

9. A quality inspection and annotation system for tender document generation based on deep learning according to claim 2, characterized in that, The final annotation result is obtained specifically by: Read the confirmed annotation results and uncertainty records, unify the label names, category codes, symbols and spaces, merge and replace synonymous labels, perform batch correction of format anomalies, and generate a standardized label list; Based on the standardized label list, perform boundary calibration on the start and end positions of each label and output the calibrated boundary mapping table; Based on the calibrated boundary mapping table, hierarchical consistency verification and conflict resolution are carried out. After sorting and renumbering, the sub-results are merged to form the final annotation result of the batch documents.

10. A quality inspection and annotation system for tender document generation based on deep learning according to claim 2, characterized in that, The output of the finished text of the tender document and the structured annotation data specifically includes: Based on the final annotation results and candidate text, the labels are aligned with the original text positions one by one, and the manuscript skeleton is generated according to chapters, sections, articles, and clauses. The key elements are filled in according to the correspondence between the labels and fields to form the finished draft. Generate chapter structure index and feature index based on the finished draft, output searchable bookmarks and jump links, and export structured labeled dataset; Perform consistency and integrity checks on the final draft and structured annotation data to generate archiveable final tender document text and structured annotation data.

Images