Intelligent question and answer system for mining field bidding based on semantic understanding

Abstract

The present application relates to the field of mining industry bidding intelligent question and answer system based on semantic understanding, belongs to the field of bidding question and answer system, including: data layer, configured to build mining bidding dynamic knowledge base; the data layer includes: data cleaning and structuring module, vector database and knowledge graph construction module, standard question and answer pair library; model layer, connected with the data layer, configured to bidding field exclusive semantic understanding and question and answer model; the model layer includes: field fine-tuning model, multi-round dialogue and question disassembly engine; application layer, connected with the model layer, configured to provide personalized intelligent interaction service; the application layer includes: multi-dimensional user portrait module, risk iteration recommendation module. The present application, relying on three-layer architecture, completes data governance, model optimization and scene landing, realizes the intelligentization of the whole process of mining bidding, risk pre-control, long-term knowledge sedimentation, personalized service for thousands of people, and comprehensively improves the work efficiency and compliance level of bidding and procurement.

Description

Technical Field

[0001] This invention relates to the field of bidding and tendering question-and-answer systems, specifically a semantic understanding-based intelligent question-and-answer system for bidding and tendering in the mining industry. Background Technology

[0002] With the comprehensive advancement of national policies on "Artificial Intelligence+", digital economy development, and digital transformation of state-owned enterprises, intelligent upgrading in the bidding and tendering field has become an inevitable trend in industry development. Many regions in China have already implemented general-purpose AI-assisted bidding and tendering systems, achieving basic functions such as bid evaluation and document screening. However, the mining bidding and tendering scenario has highly specific industry characteristics, and existing technical solutions have many shortcomings: Insufficient adaptability to vertical scenarios: Mining bidding involves specialized businesses such as underground engineering construction, procurement of fully mechanized mining equipment, and verification of mining safety qualifications, which involve a large number of industry-specific terms, technical parameters, safety standards, and local regulations. General-purpose models have not been trained on mining-related corpora, resulting in low accuracy in understanding professional expressions and an inability to accurately complete core tasks such as qualification review and technical clause verification. Furthermore, similar foreign systems cannot be directly reused due to significant differences in regulatory frameworks and business logic compared to domestic systems.

[0003] Data fragmentation hinders effective knowledge reuse: Mining companies accumulate massive amounts of unstructured data such as tender documents, regulatory documents, equipment parameters, and engineering specifications from previous years. In the traditional model, this data is stored in isolation, and PDFs, drawings, and paper documents are difficult to retrieve and utilize. The industry generally relies on the experience of senior employees to conduct business, and the training cycle for new employees is as long as 6 months, resulting in a significant knowledge transfer gap and preventing companies from transforming their data assets into business capabilities.

[0004] Therefore, those skilled in the art have provided a semantic understanding-based intelligent question-answering system for bidding in the mining industry to address the problems mentioned in the background section. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing general AI models, such as poor adaptability to mining scenarios, low data utilization, low manual efficiency, lagging compliance risks, service homogenization, and difficulty in system integration. It provides a semantic understanding-based intelligent question-and-answer system for bidding and tendering in the mining field. This system relies on a three-layer architecture to complete data governance, model optimization, and scenario implementation, achieving intelligent assistance throughout the entire bidding and tendering process, proactive risk prevention, long-term knowledge accumulation, and personalized services tailored to each individual, comprehensively improving the efficiency and compliance level of bidding and procurement.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A semantic understanding-based intelligent question-answering system for bidding and tendering in the mining industry, including: The data layer is configured to build a dynamic knowledge base for mining bidding; the data layer includes: The data cleaning and structuring module is used to clean the collected multi-source heterogeneous bidding data, and to perform entity identification and relationship extraction on the process nodes, risk types and legal basis in the cleaned data, transforming unstructured data into structured data. The vector database and knowledge graph construction module is used to segment unstructured text data into semantically complete segments, generate text embedding vectors and store them in the vector database, and construct a four-layer association knowledge graph based on the graph database, which includes nodes of the bidding process, risk types, response solutions and legal basis. The standard question-and-answer pair library stores standard question-and-answer pairs covering four scenarios: document preparation, qualification review, bid evaluation decision-making, and post-bid management. The model layer, connected to the data layer, is configured as a semantic understanding and question-answering model specific to the bidding and tendering field; the model layer includes: The domain fine-tuning model, based on a general pre-trained large model, is fine-tuned using mining-specific corpus in the data layer to improve the semantic understanding of mining professional terms. A multi-turn dialogue and question decomposition engine is used to receive users' natural language queries, perform semantic parsing and automatic decomposition of complex questions including multiple logical levels, and generate context-coherent multi-turn responses; The application layer, connected to the model layer, is configured to provide personalized intelligent interaction services; the application layer includes: The multi-dimensional user profile module is used to build user profiles based on job type, operation history, and learning progress, and to provide differentiated content push to users with different profiles; The risk iteration recommendation module is used to calculate the risk weight of each bidding stage based on historical risk data, and prioritize push verification tasks and guidance content for high-risk stages.

[0007] As a further aspect of the present invention: the data cleaning and structuring module further includes: The OCR unit is used to extract text information from image and text data; Text slicing units are used to segment unstructured documents into semantically complete segments using text segmentation tools; A vectorization unit is used to convert text fragments into high-dimensional vectors through an embedding model and store them in the vector database; The vector database and knowledge graph construction module uses the Neo4j graph database to construct the four-layer association knowledge graph. The four-layer association knowledge graph is used to perform association queries on the risks of excessively high qualifications in the construction qualification setting process and their corresponding legal basis, solutions and cases.

[0008] As a further aspect of the present invention: the domain fine-tuning model is based on a general pre-trained model with 7B parameters and fine-tuned using LoRA low-rank adaptation technology; the model layer is configured to perform semantic understanding of mining professional terms such as technical parameters of coal mine fully mechanized mining equipment, underground engineering construction specifications, and safety mark certificate requirements for mining products, and to output guidance information for qualification review and document compliance verification.

[0009] As a further aspect of the present invention: the multi-turn dialogue and problem decomposition engine is specifically used for: When receiving an inquiry about how to set qualification conditions for bidding on coal mining equipment, the inquiry is automatically broken down into three sub-questions: requirements for safety mark certificates for mining products, performance thresholds for similar projects in the past three years, and prohibited situations. The answers to the three sub-questions are obtained respectively, and the answers are integrated into a coherent response for output.

[0010] As a further aspect of the present invention: the job types in the multi-dimensional user profile module include: bidding business personnel, bid evaluation experts, and risk control audit personnel; The application layer is configured to: prioritize pushing bidding document preparation templates and compliance clause libraries to the bidding business personnel; prioritize pushing risk verification lists to the risk control audit personnel; and match bidding knowledge learning content to users whose job type is new employee.

[0011] As a further aspect of the present invention: the risk iteration recommendation module is specifically used for: Based on logistic regression or gradient boosting tree models, weights are assigned to risk types such as expired qualification documents and ambiguous technical terms; priority is given to pushing verification tasks and modification suggestions to users for stages with risk weights higher than preset thresholds.

[0012] As a further embodiment of the present invention: the system further includes a dynamic update and integration module; the dynamic update and integration module is used for: The knowledge base of the data layer is dynamically updated by periodically capturing and updating laws and policies through an automatic law monitoring script and combining it with expert review. By integrating with the enterprise's existing OA and ERP systems through API interfaces, data transmission can be achieved through single-interaction responses.

[0013] As a further aspect of the present invention: the data layer is further configured to process multimodal content, including text, images, and drawings; The application layer also includes a supervisory risk control brain, which is used to conduct compliance inspections and risk screenings on the multimodal content before or after the release of the tender documents, including at least the automatic detection and early warning of discriminatory and exclusive clauses.

[0014] As a further aspect of the present invention: the system is applied to the automated processing of the core bidding process, which includes review of the completeness, legality and compliance, fairness, consistency and semantics of the bidding project.

[0015] As a further embodiment of the present invention: the system further includes an evaluation and standardization output module; the evaluation and standardization output module is used for: The system is iteratively optimized through three methods: automated testing, manual evaluation, and pilot verification. Output technical specifications for a semantic understanding-based intelligent question-answering system for bidding in the mining industry.

[0016] Compared with the prior art, the beneficial effects of the present invention are: Specifically addressing pain points in the mining industry with strong vertical adaptability: This invention is custom-developed for the unique business, regulations, and technical specifications of bidding in the coal and mining sectors. By fine-tuning the large model with mining-specific corpus and building an industry knowledge graph, it completely solves the problem of general AI models' insufficient understanding of mining professional terminology, underground engineering, and mining equipment parameters, filling the technological gap in the subdivided field of intelligent bidding in the mining industry.

[0017] Activate enterprise data assets and achieve long-term knowledge accumulation: Transform unstructured tender documents, regulations and technical documents accumulated by the enterprise over the years into structured knowledge bases and question and answer bases, establish a monthly dynamic update mechanism, break the traditional "mentoring" experience transfer model, compress the business proficiency cycle of new employees from 6 months to 3 months, and realize the reusability and inheritance of the enterprise's core knowledge assets.

[0018] Significantly improves operational efficiency and reduces labor costs: Through intelligent preparation, automatic review, one-click Q&A and other functions, the time for preparing bidding documents and the time for qualification review are shortened, reducing a large amount of repetitive manual labor. Every year, it can save enterprises a number of expenses such as labor costs, objection handling costs, and losses from failed bids, resulting in significant economic benefits.

[0019] Upgraded risk control model enables proactive risk prevention: Abandoning the traditional post-event verification risk control model, a weighted algorithm is built based on historical risk data to automatically identify high-risk links and give priority warnings. It accurately investigates potential risks such as discriminatory clauses, qualification fraud, and parameter violations, effectively reducing the review error rate and bidding objection rate, and strengthening the enterprise's compliance defense.

[0020] Personalized services adapt to all positions, with excellent human-computer interaction experience: Based on multi-dimensional user profiles, "personalized service" is achieved. Differentiated content is pushed to different roles such as bidding personnel, risk control personnel, evaluation experts, and new employees, which fits the actual work needs of each position and solves the problem of homogenization of traditional system services.

[0021] With strong integration and scalability, and high promotional value, the system can seamlessly connect with existing OA and ERP systems of enterprises and adapt to existing office systems. The overall design adopts a "basic framework + industry plug-in" approach, and can be quickly migrated to bidding scenarios of other state-owned enterprises such as construction, transportation, and power simply by replacing the industry corpus. It can not only be fully promoted within subsidiaries of large mining groups, but also form a standardized solution for external output, with broad market prospects.

[0022] Multimodal support covering all business scenarios: compatible with the review of multiple types of documents such as text, images, and drawings; adaptable to the business characteristics of mining bidding with a large number of engineering drawings and equipment drawings; functions cover the entire process of bidding, evaluation, and post-bid management, with a high degree of integration. Attached Figure Description

[0023] Figure 1 This is a structural diagram of an intelligent question-answering system for bidding in the mining industry based on semantic understanding. Detailed Implementation

[0024] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0025] As mentioned in the background section of this application, research has revealed that existing mining bidding processes involve specialized business areas such as underground engineering construction, procurement of fully mechanized mining equipment, and verification of mining safety qualifications. These areas contain a large number of industry-specific terms, technical parameters, safety standards, and local regulations. General-purpose models, lacking training with mining corpora, exhibit low accuracy in understanding professional expressions and are unable to accurately complete core tasks such as qualification review and technical clause verification. Furthermore, similar foreign systems, due to significant differences in regulatory frameworks and business logic compared to domestic systems, cannot be directly reused and thus possess certain shortcomings.

[0026] To address the aforementioned shortcomings, this application discloses a semantic understanding-based intelligent question-answering system for bidding and tendering in the mining industry. Relying on a three-layer architecture, it completes data governance, model optimization, and scenario implementation, realizing intelligent assistance for the entire bidding and tendering process in the mining industry, proactive risk prevention and control, long-term knowledge accumulation, and personalized services tailored to each individual, thereby comprehensively improving the efficiency and compliance level of bidding and procurement work.

[0027] The following will describe in detail, with reference to the accompanying drawings, how the solution of this application solves the above-mentioned technical problems.

[0028] Please see Figure 1In this embodiment of the invention, a semantic understanding-based intelligent question-and-answer system for bidding in the mining industry includes: a data layer configured to construct a dynamic knowledge base for mining bidding; the data layer includes: a data cleaning and structuring module, used to clean the collected multi-source heterogeneous bidding data, and to perform entity recognition and relation extraction on process nodes, risk types, and legal basis in the cleaned data, transforming unstructured data into structured data; a vector database and knowledge graph construction module, used to segment unstructured text data into semantically complete fragments, generate text embedding vectors and store them in the vector database, and construct a four-layer association knowledge graph based on the graph database, consisting of bidding process nodes, risk types, response plans, and legal basis; a standard question-and-answer pair library, storing standard question-and-answer pairs covering four scenarios: document preparation, qualification review, bid evaluation decision-making, and post-bid management; and a model layer connected to the data layer. The system is configured as a semantic understanding and question-answering model specifically for the bidding and tendering field. The model layer includes: a domain-fine-tuned model, based on a general pre-trained large model, fine-tuned using mining-specific corpora from the data layer to improve semantic understanding of mining terminology; a multi-turn dialogue and question decomposition engine, used to receive user natural language queries, perform semantic parsing and automatic decomposition of complex questions involving multiple logical levels, and generate context-coherent multi-turn responses; and an application layer, connected to the model layer, configured to provide personalized intelligent interactive services. The application layer includes: a multi-dimensional user profile module, used to build user profiles based on job type, operation history, and learning progress, and provide differentiated content recommendations for users with different profiles; and a risk iteration recommendation module, used to calculate the risk weights of each bidding and tendering stage based on historical risk data, and prioritize pushing verification tasks and guidance content for high-risk stages.This application employs a three-layer architecture—data layer, model layer, and application layer—working in synergy. On one hand, the data layer handles the cleaning, structuring, vector storage, and construction of a four-layer relational knowledge graph of multi-source heterogeneous bidding data. It also includes a standard question-and-answer database covering all core business scenarios, effectively addressing the problems of fragmented and unstructured mining bidding data, low utilization of unstructured data, and scattered business knowledge that is difficult to retrieve and reuse. This enables the accumulation of enterprise bidding data assets and the systematic management of knowledge, providing comprehensive, accurate, and reliable underlying knowledge support for upper-layer intelligent applications. On the other hand, the model layer utilizes domain-specific fine-tuning models to adapt general large-scale models to mining scenarios, significantly improving the system's semantic understanding of mining-specific professional terminology and business specifications. Coupled with a multi-turn dialogue and problem decomposition engine, it can efficiently parse complex and composite business problems, ensuring the coherence of multi-turn interaction contexts, and breaking through the limitations of traditional methods. The system addresses the bottlenecks of insufficient adaptability of AI models in the vertical field of mining bidding and tendering, and weak ability to handle complex business problems. Simultaneously, it leverages a multi-dimensional user profiling module at the application layer to deliver differentiated services based on job position, operational behavior, and learning status, overcoming the homogenization problem of traditional systems. Combined with a risk iteration recommendation module, it achieves early warning and precise control of high-risk aspects based on risk weights, shifting bidding risk control from post-event remediation to pre-event prevention. The overall architecture is logically clear and highly coupled, providing a one-stop solution for the entire process of mining bidding document preparation, qualification review, bid evaluation decision-making, and post-bid management, including intelligent Q&A, business assistance, compliance risk control, and personalized services. This significantly reduces the intensity of manual operations, improves overall work efficiency and compliance control, and the overall framework has good scalability, facilitating subsequent functional iterations and cross-industry application migration.

[0029] In this embodiment, the data cleaning and structuring module further includes: an OCR unit for extracting text information from image and text data; a text slicing unit for segmenting unstructured documents into semantically complete fragments using a text segmentation tool; a vectorization unit for converting text fragments into high-dimensional vectors through an embedding model and storing them in a vector database; and a vector database and knowledge graph construction module for constructing a four-layer relational knowledge graph using the Neo4j graph database. The four-layer relational knowledge graph is used to perform relational queries on the risks of excessively high qualifications in the construction qualification setting process and their corresponding legal basis, solutions, and cases. By accurately extracting text information from graphic bidding documents such as drawings and scanned PDFs using the OCR unit, and segmenting lengthy unstructured documents into semantically complete independent fragments using the text slicing unit, and then converting the text fragments into high-dimensional vectors and storing them in a vector database using the vectorization unit, the entire process of structured processing of multi-source heterogeneous data is fully realized. This completely solves the problem of difficulty in parsing and retrieving large amounts of paper scans, engineering drawings, and long text PDFs in the mining bidding field. At the same time, a four-layer relational knowledge graph is constructed using the Neo4j graph database, which can deeply bind bidding processes, risks, plans, and regulations, allowing for rapid linkage and querying of various business-related information, further strengthening the underlying data support capabilities, and improving the overall retrieval efficiency and knowledge association accuracy of the system.

[0030] Specific examples: OCR Unit Application Example: For scanned documents of mining underground engineering construction qualifications and technical description images attached to equipment drawings, the OCR unit automatically recognizes the text content in the image, such as extracting text information such as "Grade A Qualification for General Contracting of Mining Engineering Construction" and "MA Certification for Explosion-proof Standards for Underground Equipment", and completes the conversion of non-image and text data.

[0031] Text slicing unit application example: A 200-page PDF document on mining bidding regulations is split into segments according to semantic boundaries using a text segmentation tool, with each segment's length controlled to 512 characters. The resulting set of segments is then obtained. Each All text segments are semantically independent to avoid fragmented sentences affecting subsequent parsing.

[0032] Vectorization unit application example: split text fragments Embedded computation is performed, and the corresponding high-dimensional storage vector is obtained through data layer vectorization unit operations. , all vectors All queries are stored in a vector database; when a user enters a search query, the query statement is also sent to the vectorization unit to generate a query vector. Subsequent searches will be performed using vector similarity calculations. Quickly match relevant content, among which, The query vector is obtained by encoding the retrieval question using a vectorization unit. A high-dimensional vector of historical text fragments stored in the pointer database; The pointer similarity calculation function is used to compare the semantic similarity between the query vector and the vectors stored in the database. The system matches and outputs the bidding information with the strongest relevance based on the similarity value from high to low.

[0033] Example of a four-layer knowledge graph query: Taking the construction qualification setting process node as the query entry point, the knowledge graph automatically associates the risk type "qualification setting too high" and simultaneously retrieves the corresponding Ministry of Housing and Urban-Rural Development qualification standard clauses, reference cases of qualification setting for similar projects in the past 3 years, list of compliant qualification names, and other response plans and legal basis. Staff can obtain a complete set of reference content in one stop.

[0034] In this embodiment, the domain fine-tuning model is based on a general pre-trained model with 7B parameters and fine-tuned using LoRA low-rank adaptation technology. The model layer is configured to perform semantic understanding of mining professional terms such as technical parameters of coal mine fully mechanized mining equipment, underground engineering construction specifications, and safety mark certificate requirements for mining products, and to output guidance information for qualification review and document compliance verification. The model utilizes a general-purpose pre-trained model with 7B parameters and employs LoRA low-rank adaptation technology for industry-specific fine-tuning. This approach leverages a medium-parameter model to balance semantic understanding capabilities with ease of deployment, adapting to existing hardware environments such as enterprise local servers and GPU clusters. Furthermore, LoRA technology updates only a small number of low-rank parameters, effectively reducing training computation, time, and manpower costs, and avoiding resource waste associated with full-scale fine-tuning. Simultaneously, by embedding mining terminology and importing industry-specific corpora for targeted training, the model thoroughly addresses the issues of inadequate understanding and interpretation of professional content such as technical parameters of coal mine fully mechanized mining equipment, underground engineering construction specifications, and safety certification requirements for mining products. This enhances the model's professional adaptability in mining bidding scenarios, focusing the optimized model's capabilities on two core aspects: qualification review and document compliance verification. This ensures that the intelligent output aligns with industry rules and regulations, guaranteeing accurate and reliable review results and improving compliance control throughout the entire bidding process.

[0035] Specific examples: LoRA low-rank adaptation fine-tuning example: Based on a 7B parameter pre-trained model, the entire process is divided into a model training phase and a business inference phase. Define the original model weight matrix. As the core weight matrix, LoRA introduces two low-rank matrices. , Set low-rank dimension ,in, Refers to the dimension of the model input features. Refers to the dimension of the model's output features. This refers to low-rank dimensions, which are much smaller than the original dimensions of the model, thereby significantly reducing the number of trainable parameters.

[0036] During the model training phase, the backbone weight matrix W is permanently frozen. All parameters of matrix W do not participate in gradient updates and remain unchanged from their original pre-training values. Only the parameters of the low-rank matrices A and B are iteratively optimized. The optimization process is as follows: ① Sample Input: The mining bidding-specific annotation corpus (including parameters of coal mine fully mechanized mining equipment, underground engineering construction specifications, safety qualification clauses of mining MA, bidding compliance Q&A samples, etc.) is sent into the model in batches. The corpus text is first encoded by the data layer vectorization unit to obtain the batch input feature vector X. ② Forward propagation calculation: according to After completing the forward operation, the predicted vector Y is output. The difference between this predicted vector and the manually labeled standard business label vector is calculated to construct the classification loss function. The loss function is used to quantify the deviation between the model's current output and the standard answer in the mining industry. ③ Backpropagation: Solving the loss function only for low-rank matrices A and B. gradient , The main weight matrix W does not generate gradients or update because it is frozen; ④ Parameter Iterative Update: The adaptive gradient optimizer Adam is used with a fixed learning rate. Based on the gradient values, all elements in A and B are cumulatively corrected. The update formula is: ; ; in, To preset the training learning rate, , The matrix parameters before iteration, , These are the matrix parameters after iterative updates; ⑤ Iterative convergence: Repeat the complete process of "sample input - forward propagation calculation - backpropagation gradient - parameter iterative update", with each round of the loop counted as one iteration, and continue for multiple rounds of iteration until the loss function is achieved. Once the convergence stabilizes and the accuracy of identifying mining test samples reaches the preset threshold of 98%, the iteration stops, and the learning of knowledge in domains A and B is completed.

[0037] After the iteration is completed, only the low-rank matrices A and B after training convergence are saved, and together with the frozen and unchanged backbone weight matrix W, they are deployed to the model layer domain fine-tuning model.

[0038] During the business reasoning phase (the actual system operation phase), the raw data entered by staff, such as bidding inquiry texts, tender documents, and qualification certificates, are first processed by the vectorization unit of the data layer to extract features and generate the model input feature vector X. After X is input into the domain fine-tuning model of the model layer, the calculation is performed according to the following formula: ; Wherein, X is the text feature vector output by the data layer, which is the input data of the model layer; W is the frozen 7B pre-trained model backbone weight matrix, which retains the model's general language capabilities; A and B are the trained LoRA low-rank matrices, which carry knowledge specific to the mining bidding field; Y is the final output feature vector of the model layer, which is passed to downstream functional modules to realize core businesses such as mining professional terminology parsing, qualification review, and document compliance verification.

[0039] Example of professional terminology recognition and understanding: When the system reads mining-specific expressions such as "technical parameters of scraper conveyor in fully mechanized mining face", "construction specifications for underground roadway excavation", and "KA safety certification for mining products", the unadjusted general model is prone to conceptual confusion and parameter misinterpretation. However, the model after fine-tuning in this solution can accurately identify the meaning of terms, technical indicators and compliance requirements, and the accuracy of professional terminology understanding is greatly improved.

[0040] Application example in the qualification review process: In the qualification review of bidding for fully mechanized mining equipment, the model automatically analyzes the materials submitted by the bidders, checks the validity of the safety mark certificate for mining products, the enterprise's construction qualification level, and the supply performance of similar equipment in the past three years, and quickly identifies problems such as expired certificates, inconsistent qualification levels, and unsatisfactory performance, and generates standardized review opinions.

[0041] Application example of document compliance verification: When verifying the compliance of bidding documents for underground engineering projects, the model combines mining management regulations and bidding laws and regulations to automatically screen the content of clauses. It can accurately identify violations such as the failure to specify underground construction safety requirements and the setting of exclusive conditions in technical clauses. It also lists the risk points, corresponding legal provisions, and directly applicable modification solutions.

[0042] In this embodiment, the multi-turn dialogue and question decomposition engine is specifically used to: when receiving an inquiry about how to set qualification conditions for bidding on coal mining equipment, automatically decompose the inquiry into three sub-questions: requirements for safety mark certificates for mining products, performance thresholds for similar projects in the past three years, and prohibited situations; obtain the answers to each of the three sub-questions, and integrate the answers into a coherent response. This setting clarifies the decomposition logic and execution process for typical complex inquiries, enabling automated decomposition and processing of complex business issues in the mining bidding field. It solves the shortcomings of traditional intelligent systems that cannot parse multi-layered logical problems and can only answer simple questions. By decomposing sub-questions, retrieving answers separately, and then integrating and outputting them, it significantly improves the processing capability of complex business inquiries, ensuring logical integrity and comprehensiveness of the question-and-answer content. At the same time, it continues the contextual coherence of multi-turn dialogue, aligning with the progressive consultation habits in bidding work, improving the human-computer interaction experience and the practicality of the question-and-answer service, and further strengthening the system's implementation effect in actual business scenarios.

[0043] Specific examples: Example of complex problem decomposition logic: A user inputs a complete query: "How to set qualification conditions for bidding on coal mine fully mechanized mining equipment?" The engine decomposes this into three independent sub-problems according to business rules, represented as a set: ; in, Requirements for safety mark certificates for mining products The performance threshold for similar projects in the past 3 years This is an agreement for prohibited situations.

[0044] Example of sub-question answer retrieval: The system retrieves the knowledge base and model reasoning for each of the three sub-questions, obtaining the corresponding answer sets: ; Requirements for the validity period and certification scope of MA safety mark certificates for mining applications; Output performance threshold standards such as years of performance, number of projects, and equipment specifications; List situations in which dishonest enterprises and non-compliant partners are prohibited from participating in bidding.

[0045] Example of integrated answer output: The engine will The responses are reorganized and linked together according to business logic to form a coherent and complete text, instead of displaying the answers to individual sub-questions in a scattered manner. For example, the integrated text would be: "Companies participating in this tender for fully mechanized coal mining machines must provide a valid safety certificate for mining products and must have a track record of supplying similar coal mining machines in the past three years. Companies with a history of dishonesty in bidding or major safety accidents are not allowed to participate in this tender." This ensures that the responses conform to business reading habits.

[0046] Example of extended multi-round continuous questioning: If a user continues to ask "Can the expired security certificate be remedied?" based on the current question and answer, the system will continue to analyze the new question by relying on the context memory function and linking it to the previous dialogue scenario, without losing the previous business context and always maintaining the logical continuity of the dialogue.

[0047] In this embodiment, the job types in the multi-dimensional user profile module include: bidding business personnel, bid evaluation experts, and risk control review personnel. The application layer is configured as follows: for bidding business personnel, priority is given to pushing bid document preparation templates and compliance clause libraries; for risk control review personnel, priority is given to pushing risk verification lists; and for users whose job type is new employee, bidding knowledge learning content is matched. This setting breaks the limitations of the unified service model of traditional bidding intelligent systems. Based on the different work scenarios and core needs of bidding business personnel, bid evaluation experts, risk control review personnel, and new employees, it accurately matches corresponding functions and materials, effectively improving the office efficiency of personnel in different positions, reducing interference from invalid information, and at the same time, it pushes customized learning content for new employees, speeding up their business onboarding, giving full play to the value of system knowledge accumulation and auxiliary office work, making the system service more in line with the actual work needs of all positions in mining bidding, and significantly enhancing the overall practicality and job adaptability.

[0048] Specific examples: Example of job category and tag definition: Define the system user job tag set as: ;in Representing the bidding personnel, Representatives of the bid evaluation experts, Representing risk control auditors, For new employees, the system automatically matches the corresponding job title tags based on the login account.

[0049] Example of service push notification from bidding personnel: When the user tag is When (bidding personnel) are presenting tender documents, the system homepage prioritizes displaying various mining tender document preparation templates, industry compliance clause libraries, sample cases, etc. Staff can directly call the templates to carry out document preparation work, saving the steps of manually organizing materials and checking compliance clauses one by one.

[0050] Example of service push notification for risk control review personnel: When the user tag is When (risk control auditors) are present, the system automatically loads a standardized risk checklist, the basis for judging non-compliance clauses, historical risk cases, etc. Staff can then conduct compliance screening item by item by referring to the checklist and quickly identify various risk points in the bidding documents.

[0051] Example of new employee service push notification: When the user tag is When training new employees, the system blocks complex and professional review tools and pushes targeted learning content such as basic bidding and tendering processes, legal interpretations, introductory Q&A, and practical training to help new employees master business knowledge step by step and shorten the training cycle.

[0052] Example of expert matching for evaluation: When the user tag is When evaluating bids, the system prioritizes pushing specific content such as evaluation scoring standards, technical parameter comparison rules, and performance verification points to provide standardized reference for the evaluation work and ensure that the evaluation process is standardized and fair.

[0053] In this embodiment, the risk iteration recommendation module is specifically used to: assign weights to risk types such as expired qualification documents and ambiguous technical terms based on logistic regression or gradient boosting tree models; and prioritize pushing verification tasks and modification suggestions corresponding to stages with risk weights higher than preset thresholds to users. This setting relies on logistic regression and gradient boosting tree models to quantify and assign values ​​to various risks in bidding, changing the traditional problem of manual risk level judgment based on experience and chaotic investigation order. It realizes the digital and standardized determination of risk levels, and the system can automatically prioritize pushing high-weight risk items for verification, transforming risk control work from passive post-event rectification to proactive pre-event warning, greatly improving the efficiency of risk investigation and the ability to identify key hidden dangers, reducing compliance issues, bidding objections, and bid failures caused by human negligence, while unifying risk judgment standards across the entire unit, ensuring consistent risk control work, and further strengthening the application value of the system in bidding compliance management scenarios.

[0054] Specific examples: Example of risk weight calculation model and threshold setting: A logistic regression model is used to fit and calculate risk characteristics. The formula for calculating the comprehensive risk score is as follows: ; Where S is the risk weight score. For the intercept term, Preset feature weighting coefficients for each risk. Values ​​are assigned to corresponding risk characteristics. The system has preset risk notification thresholds. Only when When a risk is identified, it is prioritized for push notification.

[0055] Example of weighting different risk types: Weighting calculations for typical risks in mining bidding, including the score for the risk "Qualification documents not yet valid". Risk score calculation for "ambiguous wording of technical terms" The system has a preset push threshold. By comparison, we can obtain .

[0056] Example of prioritizing high-risk processes: After scanning the tender documents, the system identifies two issues simultaneously: expired qualification documents and ambiguous technical terms. Based on the weighted scoring rules, the system prioritizes pushing the verification tasks, risk descriptions, and rectification suggestions corresponding to the expired qualification documents to risk control personnel, ensuring that major compliance risks are addressed as soon as possible.

[0057] Example of delayed processing for low-risk items: For risks of "ambiguous technical terms" with scores below the threshold, the system will classify them into the regular verification queue. Reminders and rectification guidance will be provided after the high-risk items have been processed, so as to ensure that the risk control work is prioritized and carried out in an efficient and orderly manner.

[0058] In this embodiment, the system also includes a dynamic update and integration module. This module is used to: periodically capture updated regulations and policies through an automated regulatory monitoring script, and dynamically update the knowledge base at the data layer in conjunction with expert review; and integrate with the enterprise's existing OA and ERP systems via API interfaces to achieve data transmission for single-interaction responses. This setup, on the one hand, achieves routine updates of regulations and policies through automated scripts combined with expert review, solving the problem of outdated content and inability to match the latest regulations and industry requirements in traditional knowledge bases, ensuring the timeliness and compliance of the system's knowledge content. On the other hand, it achieves seamless integration with the enterprise's OA and ERP systems through API interfaces, breaking down data silos between multiple systems, eliminating the need for manual cross-platform data transfer, simplifying operation processes, improving data transmission efficiency, and strictly controlling the single-interaction response speed to ensure smooth and stable system operation under multi-system linkage. This lowers the learning and usage threshold for employees, allowing the intelligent question-and-answer system to deeply integrate into the enterprise's existing office system, further improving overall work collaboration efficiency and system adaptability.

[0059] Specific examples: Example of regulatory update response time: The system deploys a Python script for automatic regulatory monitoring, sets the response time limit for regulatory updates to 3 working days, and defines a formula for determining update completion: ; in, Total time consumed The official release date of the new regulations. The system requires a certain amount of time for expert review and completion of knowledge base iteration. When bidding-related regulations and internal procurement systems are updated, the script immediately captures the content, which is then reviewed and corrected by business experts and quickly synchronized to the knowledge base, knowledge graph, and question-and-answer database to ensure that the content queried and used by business personnel is always compliant and valid.

[0060] Cross-system API integration example: The system communicates with the enterprise's OA and ERP systems via a RESTful API interface. The data transmission time is denoted as t. The system requires the total response time for a single interaction to meet the following requirements: ; When employees initiate a bidding process approval in the OA system, they can directly access the bidding document template and compliance risk list in this system. Data such as project budget and supplier files in the ERP system can also be retrieved synchronously, without the need to manually export or import files throughout the process.

[0061] Example of dynamic knowledge base update: The system executes an automatic update process on a monthly basis. The monitoring script conducts a daily full-network inspection of bidding regulations, mining industry standards, and group management regulations. After summarizing incremental data, it is pushed to the business expert team for review, correcting errors, and adding new scenario Q&A examples. A full knowledge base iteration is completed every month to continuously adapt to the ever-changing external policies and internal business needs.

[0062] Example of multi-system collaborative work: After risk control personnel retrieve basic project information from the ERP system, they can synchronize it to this system with one click to conduct document compliance review. After the review is completed, the risk report is automatically sent back to the OA system to generate approval attachments. The entire process of data flow is smooth, which greatly reduces cross-system operation steps and improves the efficiency of office collaboration.

[0063] In this embodiment, the data layer is also configured to process multimodal content, including text, images, and drawings. The application layer also includes a supervisory risk control brain, which is used to conduct compliance checks and risk screenings on multimodal content before or after the release of bidding documents, including at least the automatic detection and warning of discriminatory and exclusive clauses. This setup breaks through the limitation of traditional systems that only support plain text processing, and is compatible with various types of materials such as text, images, and engineering drawings that frequently appear in mining bidding scenarios. Relying on the supervisory risk control brain, it realizes a dual risk control mode of pre-release screening and post-release batch inspection, which can automatically identify and promptly warn against discriminatory and exclusive clauses and other violations. It comprehensively covers the compliance review needs of the entire bidding process, makes up for the functional shortcomings of single text review, improves the comprehensiveness and accuracy of risk identification, effectively prevents various hidden compliance risks, and further strengthens the system's comprehensive risk control capabilities and scenario adaptability in complex bidding scenarios such as mining engineering and equipment procurement.

[0064] Specific examples: Example of defining multimodal data types and processing sets: Define the set of multimodal data types that the system can process as follows: ;in Represents text files, Represents image files, Representing engineering drawing files, the system can parse and perform compliance screening on all types of files within the collection.

[0065] Example of text content risk control screening: For plain text materials such as bidding announcements and technical tender documents, the supervisory risk control brain analyzes the clauses sentence by sentence to quickly identify exclusive clauses such as "restricting bidding to designated brand equipment" and discriminatory clauses such as "setting qualification thresholds that exceed industry standards." At the same time, the risk location and risk level are marked and rectification suggestions are given.

[0066] Example of reviewing image-based documents: For image-based documents such as scanned qualification certificates, on-site photos of equipment, and paper documents with official seals, the system uses OCR technology to extract text and image information, verify the certificate validity period, qualification level, and seal information, identify issues such as image tampering and qualification fraud, and trigger risk warnings.

[0067] Example of engineering drawing review: Connect with mining-specific engineering drawings such as underground roadway construction drawings and fully mechanized mining equipment assembly drawings. The system analyzes the technical parameters and construction standards marked on the drawings, compares them with industry standards and bidding requirements, and identifies potential problems such as parameter marking errors and construction standards that do not comply with safety regulations.

[0068] Example of full-process risk control application: Before the official release of the bidding documents, staff upload a complete set of materials for preliminary review to eliminate non-compliant clauses and problematic materials in advance; after the documents are released, the system starts the batch inspection function to conduct a second review of multiple published bids and attachments, forming a dual risk control guarantee and reducing the probability of bidding objections and compliance incidents.

[0069] In this embodiment, the system is applied to the automated processing of the core bidding process, which includes integrity review, legality and compliance review, fairness review, consistency review, and text semantic review of the bidding project. By fully integrating intelligent capabilities into the entire review process of the bidding project, the system overcomes the shortcomings of traditional tools that are limited in function and can only cover a few review types. It relies on intelligent means to unify review standards, improve review efficiency and accuracy, reduce subjective biases and omissions caused by manual review, and achieve standardized and automated operation of the bidding review process. This comprehensively reduces manual workload and compliance risks, further broadens the system's business coverage, and strengthens the system's practical value and implementation capability in the entire bidding process of the mining industry.

[0070] Specific examples: Example of defining a set of review types: Define the core review types for bidding and tendering covered by the system as a set: ;in For integrity review, For legal and compliance review, For fairness review, For consistency review, For text semantic review.

[0071] Completeness review example: The system checks the components of the tender documents, attachments, signed documents, etc., item by item, and counts missing data according to preset standards. If it detects that the equipment technical parameter table is not attached or the bid evaluation method attachment is missing, it will immediately prompt that the data is missing and list the missing items.

[0072] Example of legality and compliance review: The system compares the terms with national and local bidding regulations and mining industry management regulations to check whether they comply with legal provisions. For example, it identifies problems such as bidding processes not meeting the requirements for bid opening and evaluation management, and qualification requirements violating industry regulations, and marks the corresponding legal basis.

[0073] Fairness review example: The system identifies biased and restrictive content in documents, accurately identifies clauses that undermine the fairness of bidding, such as exclusive brand designation, regional barriers, and unreasonable performance thresholds, determines the risk level, and provides directions for modification.

[0074] Example of consistency review: The system cross-compares the content of the preceding and following chapters, technical parameters, construction period requirements, pricing rules and other information in the tender documents. If it finds inconsistencies such as the equipment parameter being agreed upon as 1140V in the preceding text and marked as 660V in the following text, it will automatically locate the conflict and issue an alert.

[0075] Text semantic review example: The system analyzes the logic of the statements and identifies problems such as semantic ambiguity, ambiguous expression, and incoherent sentences. For example, it identifies vague descriptions without clear definition, such as "the relevant equipment meets the requirements," and prompts staff to optimize the wording to ensure that the document's semantics are rigorous and accurate.

[0076] In this embodiment, the system also includes an evaluation and standardization output module. This module is used to iteratively optimize the system through automated testing, manual evaluation, and pilot verification; and to output technical specifications for a semantically understanding-based intelligent question-and-answer system for bidding in the mining industry. By combining automated testing, manual evaluation, and pilot verification in a multi-dimensional evaluation approach, the system's operational stability, functional completeness, and business adaptability can be comprehensively tested. Operational vulnerabilities and functional defects can be identified promptly and iterative optimizations can be completed, continuously refining system performance and user experience. Simultaneously, standardized technical specification documents can be compiled and output, solidifying project technical solutions, implementation standards, and application processes into formal ONNA enterprise standards. This not only facilitates unified operational standards for internal personnel, enabling maintenance and promotion, but also provides a complete basis for external technology output and industry standard applications, significantly improving the standardization, replicability, and industry promotion value of technological achievements.

[0077] Specific examples: Example of a three-dimensional evaluation system: Defining system evaluation methods as a set: ;in Represents automated testing, Represents human evaluation, The three methods, representing pilot verification, are combined to form a complete three-dimensional evaluation system.

[0078] Example of automated testing application: The system runs preset test cases in batches to detect indicators such as response speed, recognition accuracy, and concurrency capability, and calculates the overall pass rate R, setting a passing threshold. ,when The system automatically identifies abnormal modules, and technicians use the test reports to locate code or algorithm problems and complete the repair and optimization.

[0079] Example of manual evaluation application: An evaluation team composed of bidding and tendering business experts and technical personnel simulates the entire process of a daily office operating system to check the accuracy of Q&A, the rationality of risk identification, the ease of use of the interface, etc., record logical flaws and interaction problems that occur during use, generate a manual evaluation report, and promote targeted optimization.

[0080] Pilot verification application example: Select several subsidiaries under the group as pilot units, put them into real bidding business scenarios for a long period of trial operation, collect feedback from front-line staff on usage and business adaptation issues, and adjust the functional logic and parameter configuration according to the actual working conditions on site to ensure that the system is adapted to the business models of different branches.

[0081] Standardized document output example: After the system completes all testing and iterative optimization, based on the module organization technical architecture, module functions, operation process, technical indicators, operation and maintenance specifications, etc., a technical specification for a semantic understanding-based intelligent question and answer system for bidding in the mining field is compiled. This document can serve as a unified standard for internal use, training, and operation and maintenance within the enterprise. It can also serve as a core material for applying for coal and chemical industry standards, helping to promote the industrialization of technological achievements.

[0082] This invention adopts a three-layer architecture: data layer, model layer, and application layer. It utilizes OCR, text slicing, vectorization, and Neo4j knowledge graph technology to achieve full-process structured processing of multi-source heterogeneous data. Leveraging LoRA low-rank adaptation technology, it performs targeted fine-tuning of a large-scale, 7B-parameter general model for mining scenarios, effectively improving the model's understanding of mining terminology and technical specifications. Coupled with a complex problem decomposition and multi-turn dialogue engine, it can intelligently parse complex business queries while maintaining consistent interaction logic. The system constructs multi-dimensional user profiles based on job type, operation history, and learning progress, providing differentiated services for bidding personnel, risk control personnel, bid evaluation experts, and new employees. Simultaneously, it uses a risk weighting algorithm to provide early warnings for high-risk processes, shifting risk control from post-event rectification to pre-event prevention. A dynamic update mechanism ensures that the knowledge base content keeps pace with changes in regulations and policies, and it seamlessly integrates with enterprise OA and ERP systems through API interfaces. Breaking down data silos and ensuring efficient and stable interactive responses, the system also possesses multimodal content review capabilities for text, images, and drawings, comprehensively covering five core bidding review scenarios: completeness, compliance, fairness, etc. Combined with a three-dimensional evaluation system of automated testing, manual evaluation, and pilot verification for continuous iteration and optimization, and outputting standardized technical specification documents, the entire solution specifically addresses industry pain points such as low data utilization, poor adaptability of general models, low manual efficiency, prominent compliance risks, single service models, and difficulties in system integration in traditional bidding processes. It not only significantly reduces labor costs, shortens the training cycle for new employees, and strengthens compliance control, but also possesses strong replicability and scalability. It meets the intelligent office needs of mining companies throughout the entire process and can also provide a feasible standardized solution for the intelligent transformation of bidding in multiple fields such as construction and energy, demonstrating outstanding practical value, economic value, and industry promotion value.

[0083] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A semantic understanding-based intelligent question-answering system for bidding and tendering in the mining industry, characterized in that: include: The data layer is configured to build a dynamic knowledge base for mining bidding. The data layer includes: The data cleaning and structuring module is used to clean the collected multi-source heterogeneous bidding data, and to perform entity identification and relationship extraction on the process nodes, risk types and legal basis in the cleaned data, transforming unstructured data into structured data. The vector database and knowledge graph construction module is used to segment unstructured text data into semantically complete segments, generate text embedding vectors and store them in the vector database, and construct a four-layer association knowledge graph based on the graph database, which includes nodes of the bidding process, risk types, response solutions and legal basis. The standard question-and-answer pair library stores standard question-and-answer pairs covering four scenarios: document preparation, qualification review, bid evaluation decision-making, and post-bid management. The model layer, connected to the data layer, is configured as a semantic understanding and question-answering model specific to the bidding and tendering field; the model layer includes: The domain fine-tuning model, based on a general pre-trained large model, is fine-tuned using mining-specific corpus in the data layer to improve the semantic understanding of mining professional terms. A multi-turn dialogue and question decomposition engine is used to receive users' natural language queries, perform semantic parsing and automatic decomposition of complex questions including multiple logical levels, and generate context-coherent multi-turn responses; The application layer, connected to the model layer, is configured to provide personalized intelligent interaction services; the application layer includes: The multi-dimensional user profile module is used to build user profiles based on job type, operation history, and learning progress, and to provide differentiated content push to users with different profiles; The risk iteration recommendation module is used to calculate the risk weight of each bidding stage based on historical risk data, and prioritize push verification tasks and guidance content for high-risk stages.

2. The intelligent question-answering system for bidding in the mining field based on semantic understanding as described in claim 1, characterized in that, The data cleaning and structuring module further includes: The OCR unit is used to extract text information from image and text data; Text slicing units are used to segment unstructured documents into semantically complete segments using text segmentation tools; A vectorization unit is used to convert text fragments into high-dimensional vectors through an embedding model and store them in the vector database; The vector database and knowledge graph construction module uses the Neo4j graph database to construct the four-layer association knowledge graph. The four-layer association knowledge graph is used to perform association queries on the risks of excessively high qualifications in the construction qualification setting process and their corresponding legal basis, solutions and cases.

3. The intelligent question-answering system for bidding in the mining field based on semantic understanding according to claim 2, characterized in that, The domain fine-tuning model is based on a general pre-trained model with 7B parameters and is fine-tuned using LoRA low-rank adaptation technology. The model layer is configured to perform semantic understanding of the technical parameters of coal mine fully mechanized mining equipment, underground engineering construction specifications, and safety mark certificate requirements for mining products, and to output guidance information for the qualification review and document compliance verification processes.

4. The intelligent question-and-answer system for bidding in the mining field based on semantic understanding according to claim 3, characterized in that, The multi-turn dialogue and problem decomposition engine is specifically used for: When receiving an inquiry about how to set qualification conditions for bidding on coal mining equipment, the inquiry is automatically broken down into three sub-questions: requirements for safety mark certificates for mining products, performance thresholds for similar projects in the past three years, and prohibited situations. The answers to the three sub-questions are obtained respectively, and the answers are integrated into a coherent response for output.

5. The intelligent question-and-answer system for bidding in the mining field based on semantic understanding according to claim 4, characterized in that, The job types in the multi-dimensional user profile module include: bidding business personnel, bid evaluation experts, and risk control audit personnel; The application layer is configured to: prioritize pushing bidding document preparation templates and compliance clause libraries to the bidding business personnel; prioritize pushing risk verification lists to the risk control audit personnel; and match bidding knowledge learning content to users whose job type is new employee.

6. The intelligent question-and-answer system for bidding in the mining field based on semantic understanding according to claim 5, characterized in that, The risk iterative recommendation module is specifically used for: Based on logistic regression or gradient boosting tree models, weights are assigned to expired qualification documents and ambiguous technical terms; priority is given to pushing verification tasks and modification suggestions to users for links with risk weights higher than preset thresholds.

7. The intelligent question-and-answer system for bidding in the mining field based on semantic understanding according to claim 6, characterized in that, The system also includes a dynamic update and integration module; the dynamic update and integration module is used for: The knowledge base of the data layer is dynamically updated by periodically capturing and updating laws and policies through an automatic law monitoring script and combining it with expert review. By integrating with the enterprise's existing OA and ERP systems through API interfaces, data transmission can be achieved through single-interaction responses.

8. The intelligent question-answering system for bidding in the mining field based on semantic understanding according to claim 7, characterized in that, The data layer is also configured to process multimodal content, including text, images, and drawings; The application layer also includes a supervisory risk control brain, which is used to conduct compliance inspections and risk screenings on the multimodal content before or after the release of the tender documents, including at least the automatic detection and early warning of discriminatory and exclusive clauses.

9. The intelligent question-and-answer system for bidding in the mining field based on semantic understanding according to claim 8, characterized in that, The system also includes an evaluation and standardization output module; the evaluation and standardization output module is used for: The system is iteratively optimized through three methods: automated testing, manual evaluation, and pilot verification. Output technical specifications for a semantic understanding-based intelligent question-answering system for bidding in the mining industry.

Images