A building procurement information intelligent auditing, credit scoring and fidelity method based on BERT+NER

Abstract

This invention discloses an intelligent review, credit scoring, and authenticity verification system and method for construction procurement information based on BERT+NER, belonging to the interdisciplinary field of digitalization in the construction industry and natural language processing. The system includes: a text preprocessing unit, a construction-domain fine-tuned BERT semantic classification unit, a procurement-specific NER entity extraction unit, a regularization and logic verification unit, a manual review and incremental iteration unit, a dynamic credit scoring unit for entities, and a server-side weighting and traffic limiting control unit. This invention employs a closed-loop collaborative solution with vertical domain-specific model fine-tuning, a three-stage threshold decision-making process, seven types of structured entity extraction, dual format and logic verification, lifelong credit binding, and automatic traffic scheduling. It addresses industry pain points such as low efficiency of manual review, inability to identify fraudulent procurement, unstructured fields, chaotic formats, lack of long-term penalties, and unreasonable traffic distribution. It enables 24 / 7 fully automated review, accurate identification of authenticity, data standardization, and long-term ecosystem governance, and is suitable for the full lifecycle management of procurement information on construction B2B platforms, digital cloud exhibition platforms, and engineering supply and demand matching platforms.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing, semantic understanding in the vertical field of the construction industry, BERT domain pre-training and incremental fine-tuning, NER named entity structured extraction, regular expression format and business logic compliance verification, dynamic quantification of enterprise credit, automatic server-side traffic scheduling and exposure weight control. In particular, it relates to a full-link automated intelligent review and ecosystem authenticity protection system and method for engineering bidding, building material procurement, equipment leasing, and labor subcontracting scenarios, which realizes semantic-level authenticity identification, automatic structuring of key fields, automatic format logic verification, closed-loop iteration of manual review, lifelong binding of enterprise credit, automatic traffic restriction and demotion of false information, and automatic weighting and top placement of high-quality demand. Background Technology

[0002] Construction industry B2B trading platforms, 3D digital cloud exhibition platforms, and engineering supply and demand matching platforms generate massive amounts of procurement demand information daily, covering various scenarios such as material procurement, equipment leasing, labor subcontracting, project bidding, and cooperation inquiries. Currently, the industry generally adopts a crude management approach of manual review and keyword blocking, which has systemic technical flaws that cannot be addressed through conventional methods. 1. Manual review is costly, inefficient, and cannot cover the entire platform 24 / 7. The massive amount of procurement information relies on manual reading, judgment, verification, and entry. The average daily processing volume of a single platform is usually less than 200 items. The labor cost is high, the review cycle is long, and a large amount of information backlog occurs at night, on holidays, and during sudden peak periods. Spam and illegal information cannot be intercepted in real time, and the platform's content ecosystem continues to deteriorate.

[0003] 2. Keyword filtering alone, lacking deep semantic understanding capabilities, makes it unable to identify disguised fraudulent procurement. Traditional solutions can only block explicit violations such as "WeChat," "advertisement," and "promotion," and are completely ineffective against price gouging, fraudulent schemes, fake inquiries, and lead generation marketing texts disguised as legitimate procurement. For example, texts like "Welcome to call for inquiries, price negotiable, discounts for bulk orders" may not contain explicit violations, but semantically they lack genuine procurement intent. Current technology cannot accurately identify these texts based on contextual logic and the completeness of procurement intent, resulting in a persistently high penetration rate of fraudulent information.

[0004] 3. Lack of dedicated structured extraction capabilities for construction procurement; data cannot be standardized for storage and intelligent matching. Procurement texts contain core business fields such as category, specifications, quantity, budget, delivery date, project location, and quotation deadline. Existing technology cannot achieve automatic identification, boundary positioning, normalized extraction, and structured storage. It can only rely on manual processing and entry, resulting in supply and demand data that cannot be used for intelligent retrieval, accurate push, business opportunity matching, and big data analysis, leading to extremely low platform matching efficiency.

[0005] 4. Lack of automated format and business logic validation mechanisms leads to rampant invalid information. Procurement documents in the industry commonly suffer from problems such as inconsistent amount writing, inconsistent date formats, non-standard quantity units, conflicting date logic, inverted budget ranges, and significant mismatches between project quantity and budget. Existing technologies lack dedicated regular expression rule libraries and business logic validation engines, making it impossible to achieve automatic interception, automatic error correction, and standardized mandatory constraints, resulting in an extremely low information compliance rate on the platform.

[0006] 5. Without a credit binding mechanism and a closed-loop punishment mechanism, violators repeatedly engage in malicious activities such as false posting, spamming, expired requests, and harassing traffic diversion without quantitative punishment, posting restrictions, or traffic limits. Violators can continue to post illegal information by changing accounts, making platform governance extremely costly and failing to form a long-term constraint mechanism.

[0007] 6. The lack of differentiated traffic scheduling mechanism drowns out high-quality and genuine demand, leading to a negative cycle in the platform ecosystem. The uniform exposure rules for genuine and effective procurement and spam and illegal information prevent high-quality demand from receiving weighted traffic, while false information occupies a large amount of distribution resources. This results in a continuous decrease in the willingness of clients to publish, a continuous decrease in the effective response rate of service providers, and a continuous decrease in the platform's supply and demand activity, making it impossible to form a healthy self-circulating ecosystem.

[0008] Existing technologies only achieve fragmented keyword filtering and simple format verification, failing to form a closed-loop technical system encompassing semantic identification, structured extraction, compliance verification, credit quantification, and traffic scheduling. In particular, they lack BERT domain fine-tuning models for construction procurement scenarios, dedicated NER entity annotation systems, and server automatic scheduling algorithms that link credit and weights. This presents significant technological gaps and industry pain points, necessitating a complete technical solution that is scalable, fully automated, highly accurate, and provides long-term governance. Summary of the Invention

[0009] (a) The technical problem to be solved by the present invention 1. Solve the technical problems of relying on manual review of construction procurement information, which is costly, slow, and unable to be monitored 24 / 7; 2. Address the technical shortcomings of traditional keyword filtering, which cannot determine semantic intent and cannot identify disguised fraudulent procurement and price-rigging texts; 3. Resolve the technical deficiencies of core business fields in construction procurement that cannot be automatically extracted, cannot be structured and stored in the database, and lack support for intelligent matching; 4. Address industry pain points such as inconsistent procurement information formats, inconsistent unit names, non-standard dates, and the inability to automatically verify and block contradictory business logic; 5. Address the issues of lacking a quantitative credit system for corporate entities, lack of penalties for violations, lack of incentives for high-quality performance, and inability to achieve long-term ecological governance; 6. Solve the technical problems of differentiated platform traffic distribution, drowning out high-quality demand, resource consumption by false information, and continuous decline in platform supply and demand activity.

[0010] (II) Core Technical Solution of the Invention This invention discloses an intelligent review, credit scoring, and authenticity verification system for building procurement information based on BERT+NER. The system adopts a seven-layer closed-loop linkage architecture, including: a procurement text access and standardized preprocessing unit, a building domain fine-tuning BERT procurement intent semantic classification unit, a building procurement-specific NER named entity extraction unit, a regular expression format and business logic consistency verification unit, a manual review and model incremental iteration unit, a dynamic credit scoring unit for enterprise entities, and a server-side exposure weight and automatic traffic limiting control unit.

[0011] Each unit adopts a strongly coupled collaborative mechanism that enables data exchange, result sharing, and closed-loop iteration, rather than a simple superposition of independent functions, to jointly achieve fully automated review, authenticity verification, structured processing, compliance verification, credit quantification, traffic scheduling, and ecosystem authenticity assurance.

[0012] 1. Procurement Text Input and Standardized Preprocessing Unit Unified access to engineering procurement, building material procurement, equipment leasing, labor subcontracting, and bidding inquiry documents submitted by enterprises, and execution of standardized cleaning specifically for construction procurement scenarios: - Remove HTML tags, special control characters, and invalid emojis; - Full-width characters are uniformly converted to half-width characters; - Normalize consecutive whitespace characters, repeated punctuation marks, and redundant newline characters; - Automatic detection and filtering of garbled text. Outputs standardized text in a uniform format, free of interference, that can be directly fed into the model for processing, providing a stable and consistent input basis for subsequent BERT classification and NER extraction, and avoiding a decrease in model inference accuracy due to text noise.

[0013] 2. Fine-tuning BERT's semantic classification unit for procurement intent in the construction sector This unit does not use direct inference with the open-source BERT, but instead builds a binary classification model for the construction procurement vertical domain based on the BERT-base-chinese framework. This model is specifically designed to achieve semantic-level discrimination between "genuine and valid procurement" and "fake / advertising / price gouging / lead generation intent." Model training and inference mechanism: - Training dataset: It uses 15,000 manually labeled data, including 7,500 real and valid procurement data (covering building materials, equipment, labor, bidding, and prefabricated components) and 7,500 fake / price-rigging / advertising / invalid inquiries (covering fake procurement, lead generation marketing, fraudulent solutions, and inquiries without substantial needs). All of them come from real scenarios in the construction industry and do not contain general miscellaneous data.

[0014] - Model Structure: The input text is encoded into a token sequence by a Tokenizer, then encoded by BERT, and the resulting 768-dimensional feature vector is obtained. This vector is then connected to a fully connected layer (768→2) and a Softmax activation function. The output is the true purchase probability and the false intent probability. The model is trained using the cross-entropy loss function.

[0015] - Training parameters: learning rate 2e−5, batch size 32, epoch=3, validation set accuracy stably reached 92.5%.

[0016] - Three-stage threshold decision mechanism: Probability ≥ 0.7 → judged as REAL genuine procurement, proceeding to subsequent sampling and verification; Probability ≤ 0.3 → judged as FAKE false intent, the system automatically rejects; Probability 0.3~0.7 → judged as UNCERTAIN questionable, automatically entering manual review of the work order.

[0017] This invention addresses the technical problems of inaccurate semantics and high misjudgment rate of intent in general BERT models in the construction procurement scenario by fine-tuning domain-specific corpora. It achieves a recall rate of 88% and an accuracy rate of 92% for identifying fake procurement, which is far higher than general keyword filtering solutions.

[0018] 3. Named Entity Extraction Unit (NER) for Building Procurement A NER extraction model specifically designed for building procurement operations is constructed using a BERT+BiLSTM+CRF concatenated architecture. This model enables automatic boundary location, sequence labeling, and structured normalization output of core business fields. Model design and processing logic: - Encoding layer: Employs BERT-base-chinese to output a 768-dimensional context vector for each token, capturing long-distance semantic dependencies; - Decoding layer: Bidirectional BiLSTM (128-dimensional hidden layer) further enhances contextual features and is connected to the CRF layer to achieve globally optimal sequence labeling, avoiding entity boundary fragmentation and label conflicts; - Dedicated labeling system: Adopting the BIOES labeling standard, the entity type is strictly limited to the 7 core categories of building procurement: Product (PROD), Specification (SPEC), Quantity (QUANT), Unit (UNIT), Budget (BUDGET), Project Location (LOC), and Deadline (DATE); - Training dataset: 10,000 real texts of building procurement, with full manual annotation of entity boundaries and types; - Post-processing normalization: Performs special regular expression corrections on the CRF output, including converting Chinese numerals to Arabic numerals, relative dates to absolute dates, extracting upper and lower limits of budget ranges, and standardizing quantity units.

[0019] The final output is structured JSON data, which is directly stored in the business database, providing a standard data source for subsequent logical validation, credit scoring, retrieval ranking, and intelligent matching, realizing fully automatic conversion from text data to structured business data.

[0020] 4. Regular expression format and business logic consistency verification unit It features a built-in regular expression rule library and business logic constraint engine specifically designed for building procurement scenarios, performing fully automatic four-fold validation on text format and business rationality. If any validation fails, the application is automatically rejected and the precise error location and modification suggestions are returned. 1. Amount format validation: Matches standardized numerical values, ranges, and unit formats, and blocks ambiguous, non-standard, and abnormal amount expressions; 2. Date format validation: Enforces the YYYY-MM-DD standard format, blocking invalid dates, abbreviated dates, and illegal dates; 3. Quantity Unit Verification: Matches against the legally mandated unit database for the construction industry, blocking non-standard abbreviations, non-standard units, and quantities without units; 4. Business logic consistency verification: - Deadline ≥ Current system date; - Lower limit of budget range ≤ Upper limit of budget range - Quantity ≥ 0, no negative work volume - The budget level, project scope, and product categories are basically reasonable.

[0021] This unit implements dual mandatory constraints on format compliance and business rationality, intercepting invalid, erroneous, and contradictory procurement information at the source, thereby improving the overall data quality of the platform.

[0022] 5. Manual review and model incremental iteration unit The automated review results employ a three-tiered triage mechanism: - Clear violation: BERT determines that the code is FAKE or fails regular expression validation → automatic rejection, and publication will not be allowed; - Clear compliance: If BERT determines that the value is REAL and the confidence level is >0.9 and all regular expressions pass, it will automatically enter the release queue. - Model questionable: Confidence level 0.3~0.7 → Automatically generate manual review work order and push it to the review backend.

[0023] Reviewers perform one-click verification, authenticity labeling, entity completion and correction. Information that passes verification enters the release process. All verified and labeled data is periodically fed back to the training set. Model incremental fine-tuning is performed weekly, with the learning rate set to 1 / 10 of the base learning rate and epoch=1. This achieves continuous iteration of model accuracy and continuous reduction of misclassification rate, forming a complete technical closed loop of "data → model → verification → data optimization → better model".

[0024] 6. Dynamic Credit Scoring Unit for Enterprises This unit is not a business operation rule set by humans, but a credit quantification calculation mechanism that is automatically executed by the server based on the full-process review results. It uses the enterprise's unified social credit code as the unique entity identifier to achieve lifelong binding of credit score, which cannot be canceled or reset, and cannot be cleared or circumvented.

[0025] Credit scoring rules: - Initial credit score: 80 points, passing threshold: 60 points, maximum score: 100 points; - Bonus points: +2 points for each genuine and valid purchase that has been verified throughout the entire process; +3 points for each completed closed-loop transaction on the platform; - Deduction mechanism: -5 points for each automatic rejection; -10 points for each user complaint that is verified; -20 points for malicious false posting, price manipulation, or intentional violation. - Automatic access control: Credit score < 60: The system automatically limits the frequency of API calls, with a maximum of ≤ 3 posts per week, and all requests are subject to mandatory manual pre-review. - Credit score < 40: The system will permanently freeze the right to publish procurement information and include the user in the platform's blacklist.

[0026] This unit uses technical means to make the behavior of entities traceable, violations punishable, and high-quality content incentivized, thereby curbing industry chaos such as repeated violations and malicious postings from the source.

[0027] 7. Server-side exposure weight and automatic traffic limiting adjustment unit This unit is a server-side automatic traffic scheduling and ranking algorithm that directly binds enterprise credit scores with search ranking weights, homepage information flow recommendation probabilities, and global distribution quotas. This enables automated scheduling of high-quality demand traffic with weighted priority and false information traffic with automatically reduced priority, solving the technical problems of wasted traffic resources and low supply-demand matching efficiency.

[0028] Regulation algorithm and weight mapping rules: The initial relevance score (Score_raw) for procurement information is calculated using BM25 / TF-IDF. The final ranking score is calculated using the following formula: Score_final = Score_raw × w The weighting coefficient w is mapped from the credit score segments: - Credit score ≥ 80: w=2.0, weighted top placement, homepage information feed recommendation ratio increased to 60%; - Credit score 60~79: w=1.2, normal exposure, standard recommendation probability; - Credit score 40~59: w=0.5, ranking downgraded, search results significantly lower, traffic limited distribution; - Credit score < 40: w=0, information is only visible to the publisher, and will not be exposed, recommended, or searched globally.

[0029] If a credit score remains stable at ≥90 for three consecutive months, the system will automatically award a high-quality procurement entity label, granting it exclusive recommendation slots and additional traffic quotas to further improve the matching efficiency of high-quality entities.

[0030] (III) Beneficial Effects of the Invention 1. Achieve 24 / 7 fully automated initial review, significantly reducing labor costs and review delays. The entire process is automated, with only about 15% of questionable texts requiring manual review. The daily processing capacity has increased from 200 to over 5,000 texts, reducing the workload of manual review by more than 70%. This enables instant submission, instant review, and instant distribution of information, with no backlog or delays.

[0031] 2. Semantic-level accurate identification of fake and fraudulent procurement, breaking through the technical bottleneck of keyword filtering. Utilizing a finely tuned BERT model specifically for the construction procurement field, it judges authenticity from the perspective of context and the integrity of procurement intent. It has an accuracy rate of 92% and a recall rate of 88% in identifying fake pricing, traffic-driving advertisements, and fake inquiries, which is far higher than the effective interception rate of less than 40% of traditional keyword filtering. The amount of fake information published has decreased by more than 80%.

[0032] 3. Enables fully automated structured extraction of core business fields, supporting the platform's dedicated NER model for intelligent matching to automatically extract and normalize 7 major categories of core procurement entities. The entity recognition F1 value reaches over 90%, and text data is automatically converted into structured business data without manual input. It can be directly used for intelligent retrieval, precise push, business opportunity matching, and big data statistical analysis, significantly improving matching efficiency.

[0033] 4. Dual automatic format and logic verification significantly improves platform information compliance and data quality. Through a dedicated regular expression rule library and business logic engine, four mandatory constraints are implemented to ensure format standardization, date validity, unit standardization, and logical rationality. The compliance rate of procurement information format has increased from the industry average of 58% to over 95%, and invalid and contradictory information is fully intercepted.

[0034] 5. Lifetime binding of entity credit and automated punishment to achieve long-term governance of the platform ecosystem. The unified social credit code is used as the unique identifier. Credit scores cannot be reset or circumvented. Violations are automatically quantified and deducted, and posting permissions are automatically restricted and traffic is limited and rights are reduced. Good behaviors are automatically awarded points and weighted incentives. From a technical perspective, this achieves "violations have consequences and compliance has benefits", and completely solves the industry problem of repeated violations and repeated prohibitions.

[0035] 6. Server-side differentiated traffic scheduling activates a positive cycle of supply and demand on the platform. High-credit genuine demand is automatically weighted and prioritized, and its traffic is amplified. Low-credit false information is automatically downgraded, limited, and hidden. The platform's traffic resources are concentrated on high-quality and effective demand. The effective response rate of high-credit procurement information is increased by more than 3 times. The platform's monthly active users and supply and demand matching volume are significantly improved, forming a healthy and sustainable digital ecosystem.

[0036] 7. The closed loop of manual review and incremental model iteration enables continuous self-optimization of model accuracy. Review data is fed back into the training set on a regular basis, and incremental fine-tuning is automatically performed every week. The model's misjudgment rate continues to decrease, and the recognition accuracy improves by 2% to 3% per month. No manual retraining or redeployment is required, and the long-term performance continues to improve. Attached Figure Description

[0037] Figure 1. Overall flowchart of the intelligent review process for procurement information in this invention Figure 2. Flowchart of the fine-tuning BERT semantic classification model in the construction field according to the present invention. Figure 3. Schematic diagram of the entity extraction model architecture for the NER specific to building procurement in this invention. Figure 4. Flowchart of the consistency verification process between the regular expression format and business logic of this invention. Figure 5. Schematic diagram of the linkage control between enterprise credit score and exposure weight in this invention. Detailed Implementation

[0038] The present invention will be further described in detail below with reference to specific embodiments. These embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0039] Example 1: Automatic review and release of compliant procurement information throughout the entire process. A building materials company submitted a procurement document: "Seeking to purchase 500 cubic meters of C30 concrete, with a budget of 350,000 to 400,000 yuan, delivery location in Xihu District, Hangzhou City, and deadline for quotation is May 20, 2026."

[0040] The system performs the entire process: 1. The text input unit performs text cleaning, removing redundant spaces and line breaks, and outputs standard text; 2. The BERT semantic classification model inference showed a true purchase probability of 0.96, which was determined to be REAL, and the process proceeded directly to the next step. 3. NER model automatically extracts structured data: Category = Concrete, Specification = C30, Quantity = 500, Unit = Cubic meters, Budget = 350,000-400,000 RMB, Location = Xihu District, Hangzhou City, Deadline = 2026-05-20; 4. Regular expression and logical validation: Date format is valid, later than the current date; amount format is valid; unit is correct; all logical validations pass. 5. Confidence level > 0.9 and all validations passed, no manual review is required, and it will automatically enter the release queue; 6. The system automatically adds 2 to the company's credit score, bringing the current credit score to 78. The corresponding weight coefficient is 1.2, and the search ranking and exposure are normal.

[0041] Example 2: Semantic Recognition and Automatic Rejection of Fake Procurement The company submitted the following text: "Welcome to inquire via WeChat. We offer quotes for all kinds of building materials; prices are negotiable, discounts for bulk orders, and long-term cooperation is welcome." The text contains no explicit prohibited keywords and cannot be blocked by traditional keyword filtering methods.

[0042] The system of this invention processes: The BERT domain fine-tuning model inference has a false intent probability of 0.85, which is determined to be FAKE. The system automatically rejects the application, does not publish it, and reduces the company's credit score by 5 points, thus completing a fully automated violation interception.

[0043] Example 3: Automatic format and logic error checks reject the enterprise's submitted procurement text: "Request to purchase 100 tons of steel bars, budget 1,000,000 to 800,000 yuan, deadline 2023-01-01".

[0044] The system performs verification automatically: 1. The lower limit of the budget range (1 million) is greater than the upper limit (800,000), which is logically contradictory; 2. The deadline of January 1, 2023 is earlier than the current system date, which is a logical violation. The system will automatically reject the initial review and return a precise modification prompt: "Budget range is inverted, please correct; deadline cannot be earlier than the current date". The company can resubmit after making the correction.

[0045] Example 4: Manual review and incremental iteration of text with questionable model information The company submitted the following text: "We need a batch of engineering materials. We will purchase them if the price is right. We will discuss the details upon inspection."

[0046] The BERT inference confidence level is 0.55, which is in the range of 0.3 to 0.7. It is determined to be UNCERTAIN, and the system automatically generates a manual review work order.

[0047] After review, the auditors determined it to be a genuine purchase but with a brief description. They then added additional information about the product category and entity, approved the purchase, and published it. This text and the annotation results were automatically added to the backflow dataset and participated in weekly incremental model tuning to continuously optimize the model's accuracy in recognizing ambiguous purchase texts.

[0048] Example 5: Linked scheduling of credit score and exposure weight Company A consistently publishes genuine and valid procurement information with no violations. It has a credit score of 88 and a weighting coefficient of 2.0. Its procurement requests are given a weighted top position in keyword search results, have a higher probability of being recommended in the homepage information feed, and have a significantly higher effective response rate than ordinary users.

[0049] Company B has repeatedly posted false information, which has been automatically rejected and reported. As a result, its credit score has dropped to 55 points, with a weighting coefficient of 0.5. Its posting of requests has been demoted and its search ranking has dropped significantly. In addition, the system automatically limits it to a maximum of 3 posts per week, and all posts are subject to mandatory manual pre-review.

[0050] Company C has repeatedly posted malicious and false information, resulting in a credit score below 40. The system has permanently frozen its posting privileges, and its historical posts will not be exposed, retrieved, or recommended globally.

[0051] Example 6: NER Entity Extraction and Normalization Example Enter text: "Purchase approximately 200 tons of HRB400E rebar Φ25, budget 700,000-750,000 RMB, Tianjin Binhai New Area, quote required before May 30th."

[0052] NER model extraction + post-processing normalization output JSON: { "Category":"Rebar", Specifications: "HRB400E / Φ25", Quantity: 200, Unit: "tons" Budget: 700,000 - 750,000 Location: Tianjin Binhai New Area Deadline: "2026-05-30" The relative date "May 30th" is automatically converted to the standard date of the current year, completing fully automatic structured processing.

Claims

1. A BERT+NER-based building procurement information intelligent auditing, credit scoring and fidelity system, characterized in that, It includes seven units that are linked in a closed loop: the procurement text access and standardization preprocessing unit, which receives construction procurement, bidding, leasing and subcontracting texts, performs noise reduction, full-width and half-width conversion, whitespace normalization and garbled character filtering, and outputs standard noise-free text for review; A fine-tuned BERT procurement intent semantic classification unit for the construction domain was developed using a BERT-base-chinese model fine-tuned with a construction procurement-specific annotated corpus. Feature vectors were fed into a fully connected layer and Softmax to achieve binary classification. A three-tiered threshold was set: ≥0.7 for genuine procurement, ≤0.3 for false intent, and 0.3~0.7 for suspicious intent, outputting corresponding intent labels. The model was trained using 15,000 annotated construction scene data points with a learning rate of 2e-5, a batch size of 32, 3 epochs, and a validation set accuracy of 92.5%. The dedicated Named Entity Extraction (NER) unit for construction procurement employs a BERT+BiLSTM+CRF concatenated architecture. Based on BIOES rules, it annotates seven core procurement entities, including category, specifications, quantity, unit, budget, location, and deadline. It performs post-processing including number normalization, date standardization, and unit unification, outputting structured business data. The regular expression format and business logic consistency verification unit has a built-in library of regular expressions specifically for construction procurement. It performs four-fold verification: amount format, date format (YYYY-MM-DD), standardized units, and business logic. Logical constraints include deadline ≥ current date, lower budget limit ≤ upper budget limit, and quantity ≥ 0. Any failure to pass this verification automatically results in rejection and an error message. The manual review and model incremental iteration unit automatically generates review work orders for questionable texts with confidence levels of 0.3 to 0.7, reviews and re-feeds the labeled data back to the training set on a regular basis, performs incremental fine-tuning once a week, with a learning rate of 1 / 10 of the base learning rate and epoch=1, to achieve continuous iteration of model accuracy; The enterprise credit dynamic scoring unit uses the unified social credit code as its unique identifier, with an initial credit score of 80 points, a passing score of 60 points, and a maximum score of 100 points; genuine publication adds 2 points, completed transaction adds 3 points, automatic rejection deducts 5 points, successful complaint deducts 10 points, and malicious falsehood deducts 20 points. Credit scores < 60 automatically limit weekly postings to ≤ 3 and require mandatory pre-review; credit scores < 40 permanently freeze posting privileges. The server-side exposure weight and traffic limiting automatic adjustment unit calculates the final ranking score Score_final = Score_raw × w based on the credit score segmented mapping weight coefficient w. Credit scores ≥ 80 → w = 2.0 for weighted top placement; 60~79 → w = 1.2 for normal exposure; 40~59 → w = 0.5 for reduced traffic; < 40 → w = 0 for no global exposure. Simultaneously, the homepage information feed recommendation ratio for high-credit entities is increased to 60%.

2. The system of claim 1, wherein, The training dataset for the fine-tuned BERT procurement intent semantic classification unit in the construction field is exclusive data for the vertical construction industry. Real procurement samples cover building materials, equipment, labor, prefabricated components, and engineering bidding scenarios, while fake samples cover fake procurement, price manipulation, fraudulent schemes, lead generation marketing, and price inquiries without substantial needs. It does not contain general domain miscellaneous data.

3. The system of claim 1, wherein, In the NER named entity extraction unit specifically for building procurement, the BiLSTM hidden layer has a dimension of 128, and the CRF layer is used for global sequence labeling optimization to avoid entity boundary fragmentation and label conflicts. The post-processing process includes converting Chinese numerals to Arabic numerals, converting relative dates to absolute dates, and extracting the upper and lower limits of the budget range.

4. The system of claim 1, wherein, The standardized unit library of the regular expression format and business logic consistency verification unit includes tons, sets, cubic meters, meters, square meters, units, sets, pieces, groups, and batches. It intercepts non-standard abbreviations and illegal units, and prohibits logical violations such as invalid dates, budget inversions, and expired deadlines.

5. The system of claim 1, wherein, In the manual review and model incremental iteration unit, the labeled data that has passed the review is fed back into the training set in batches every week. When performing incremental fine-tuning, the bottom coding layer of BERT is frozen, and only the top fully connected layer and CRF layer are fine-tuned, so as to reduce training overhead and deployment complexity while ensuring accuracy improvement.

6. The system of claim 1, wherein, The credit score of the enterprise entity's dynamic credit scoring unit is permanently bound to the unified social credit code. Account cancellation and re-registration will not reset, clear the score, or avoid deduction records. All accounts under the same entity share the same credit score system.

7. A BERT+NER-based building procurement information intelligent auditing, credit scoring and fidelity method, characterized in that, The processor performs the following steps: Step S1: Receive the original text of building procurement, bidding, leasing, and subcontracting, and perform preprocessing such as noise reduction, full-width / half-width conversion, whitespace normalization, and garbled character filtering to generate standard noise-free text to be reviewed; Step S2: Input the standard text into the finely tuned BERT model for the construction procurement domain, perform binary classification of procurement intent, and determine it as genuine procurement, false intent, or questionable text according to a three-segment threshold; Step S3: For genuine intent text, extract 7 types of core business entities through a dedicated NER model with BERT+BiLSTM+CRF architecture, perform normalization post-processing, and output structured data; Step S4: Call the regular expression rule library and logic engine to perform four-fold verification of amount, date, unit, and business logic. If the format or logic violates the rules, the initial review will be automatically rejected and a modification prompt will be returned. Step S5: For questionable text with a threshold of 0.3~0.7, a manual review work order will be automatically generated. The reviewed and labeled data will be fed back to the training set weekly, and incremental fine-tuning will be performed to continuously optimize the model accuracy. Step S6: Using the unified social credit code as the main identifier, the credit score will be automatically calculated based on the compliance of the publication, transaction behavior, and complaint records. The initial score is 80 points. For genuine publication, the score increases by 2 points; for rejection, it decreases by 5 points; for complaints, it decreases by 10 points; for malicious falsehood, it decreases by 20 points. Publication is restricted if the credit score is less than 60 points, and permanently frozen if it is less than 40 points. Step S7: Based on the segmented mapping weight coefficient of the credit score, the final score for retrieval and ranking will be calculated. High credit scores will be weighted and placed at the top, low credit scores will be given reduced access, and blacklisted individuals will not be exposed globally, completing the intelligent review process and ensuring the authenticity of the ecosystem.

Images