A tendering and purchasing bid evaluation method and system based on large language model technology
By processing bidding documents using a large language model, extracting features and entities, and performing hard index and multi-domain analysis, the shortcomings of manual review in electronic bidding systems are solved, and an intelligent and efficient bidding evaluation process is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUANGZHOU HIGH-TECH ENG CONSULTING CO LTD
- Filing Date
- 2025-07-16
- Publication Date
- 2026-05-19
AI Technical Summary
In existing electronic bidding systems, the evaluation process relies on manual review, which results in insufficient objectivity and intelligence, consumes a lot of manpower, and affects the efficiency of the evaluation.
A bidding and procurement evaluation method based on large language model is adopted. The bidding and tender documents are processed by LLM large language model, features and entities are extracted, hard index analysis and multi-domain analysis are performed, and a list of candidate suppliers is generated.
It improved the efficiency and intelligence of standardized processing of bidding data, enabled objective and accurate evaluation of suppliers, reduced labor costs, and improved bid evaluation efficiency.
Smart Images
Figure CN120805927B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of large language model technology, and in particular to a bidding and procurement evaluation method and system based on large language model technology. Background Technology
[0002] The scientific and rational nature of a company's bidding and procurement strategies has a significant impact on its economic benefits and high-quality development. Currently, bidding and procurement evaluation refers to the process by which multiple evaluation experts review, score, and rank the bid documents to select the best bidder. It is one of the important steps in the bidding and procurement process.
[0003] Electronic bidding is a method of managing, publishing, and processing bidding and procurement processes using the internet and digital technologies. With the increasing maturity of next-generation information technologies, emerging technologies such as big data, artificial intelligence, and cloud computing have been widely applied across various industries. Large language models, as one of these emerging technologies, can not only provide strong technical support for electronic bidding but also offer new ideas and methods for each stage of the electronic bidding process.
[0004] However, current electronic bidding technology solutions are usually in the stage of release and information collection, but for the specific review process, the traditional evaluation method still adopts manual review, which leads to insufficient objectivity and intelligence, and also requires a lot of manpower, affecting the efficiency of evaluation. Summary of the Invention
[0005] To address the aforementioned technical problems, this invention aims to provide a bidding and procurement evaluation method and system based on large language model technology.
[0006] The objective of this invention is achieved through the following technical solution:
[0007] Firstly, this invention proposes a bidding and procurement evaluation method based on large language model technology, comprising the following steps:
[0008] S1 obtains the tender documents and bid documents;
[0009] S2 processes tender documents based on the LLM large language model and extracts the corresponding tender clause feature set, which includes one or more tender clauses.
[0010] S3 processes tender documents based on the LLM large language model, extracts tender content entities from the tender documents, and aligns the tender content entities to the corresponding tender terms.
[0011] S4 performs hard indicator analysis based on the bidding terms and the corresponding bid content entities to obtain the hard indicator analysis results.
[0012] When the hard indicator analysis results pass, S5 further performs matching analysis on the bidding terms and corresponding bid content entities based on different domain parsing engines to obtain the domain analysis results corresponding to the bidding terms; further, it performs consistency analysis based on the domain analysis results to obtain the comprehensive analysis results of supplier bid matching degree.
[0013] Preferably, the method further includes:
[0014] Based on the bid matching results of each supplier, S6 generates a list of candidate suppliers from the N suppliers with the highest bid matching degree.
[0015] Preferably, step S1 includes:
[0016] Obtain the tender documents provided by the tendering party, which include technical specifications and procurement terms and conditions documents, etc.
[0017] Obtain the tender documents provided by the supplier, which include the tender letter, qualification certificate, financial statements, construction plan, quotation, etc.
[0018] The document formats for tender documents and bid documents include PDF, DOCX, XLSX, JPEG, etc.
[0019] Preferably, step S2 includes:
[0020] Based on the LLM large language model, the unstructured data of the tender document is first extracted by the parsing tool, including the original text, paragraph marks and tables, and then the unstructured data is transformed into a processable text sequence.
[0021] The semantic understanding module performs keyword matching or contextual analysis on the text sequence to obtain the tender terms.
[0022] By combining the clause and rule templates and the fine-tuning model, structured bidding clause information is output, which includes clause items and clause content.
[0023] Preferably, step S3 includes:
[0024] Based on the LLM large language model, the unstructured data of the tender documents, including raw text, paragraph marks and tables, are first extracted by parsing tools and then transformed into a processable text sequence.
[0025] The semantic parsing model based on domain fine-tuning extracts key entities from the obtained text sequence to obtain the bidding content entity, which includes entity category and entity content.
[0026] Semantic similarity is calculated based on the obtained bid content entities and bidding terms. The bid content entities and bidding terms are then associated based on semantic similarity to obtain an alignment set.
[0027] Preferably, step S5 includes:
[0028] Based on the defined domains, the corresponding domain parsing engines are invoked respectively. The defined domains include the technical domain, the financial domain, and the legal domain, and the domain parsing engines include the technical domain parsing engine, the financial domain parsing engine, and the legal domain parsing engine.
[0029] The domain parsing engine has a scoring model. The scoring model is used to match and analyze the bidding terms and the corresponding bidding content entities to obtain the corresponding domain scores. The domain scores include technical domain scores, financial domain scores and legal domain scores.
[0030] Consistency analysis based on domain scores yields a comprehensive analysis result of supplier bid matching.
[0031] Secondly, the present invention also proposes a bidding and procurement evaluation system based on large language model technology, the system comprising a processor, wherein the processor is used to execute the bidding and procurement evaluation method based on large language model technology as described in any embodiment of the first aspect.
[0032] The beneficial effects of this invention are as follows: The bidding and procurement evaluation method and system based on large language model technology proposed in this invention can use the bidding documents of the bidding party and the bid documents of the suppliers as a basis. It extracts the corresponding bidding clauses from the bidding documents and uses them as a basis to extract entities from the bid documents, aligning the bid content entities contained in the bid documents to the corresponding bidding clauses. This improves the efficiency and intelligence level of standardized processing of bidding data. Furthermore, based on the aligned bidding clauses and bid content entities, hard index analysis is performed. This allows for analysis of suppliers' compliance with index-related clauses, thereby selecting suppliers that meet the standard requirements. For suppliers that meet the basic requirements, a multi-domain parsing analysis scheme is further used to analyze non-index-related clauses or bid content exceeding the standard. The suitability of supplier conditions to clauses is analyzed from different domain perspectives, yielding a comprehensive analysis result. This helps improve the objectivity and accuracy of quantitative evaluation of non-standard bid content entities based on large language model technology, thereby improving the intelligence level and efficiency of supplier selection. Attached Figure Description
[0033] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.
[0034] Figure 1 This is a flowchart illustrating a bidding and procurement evaluation method based on large language model technology, as shown in the embodiment.
[0035] Figure 2 This is a schematic diagram illustrating the logical framework of a bidding and procurement evaluation method based on large language model technology, as shown in the embodiment. Detailed Implementation
[0036] The present invention will be further described in conjunction with the following application scenarios.
[0037] See Figure 1 , Figure 2 It demonstrates a bidding and procurement evaluation method based on large language model technology, including the following steps:
[0038] S1 obtains the tender documents and bid documents;
[0039] S2 processes tender documents based on the LLM large language model and extracts the corresponding tender clause feature set, which includes one or more tender clauses.
[0040] S3 processes tender documents based on the LLM large language model, extracts tender content entities from the tender documents, and aligns the tender content entities to the corresponding tender terms.
[0041] S4 performs hard indicator analysis based on the bidding terms and the corresponding bid content entities to obtain the hard indicator analysis results.
[0042] When the hard indicator analysis results pass, S5 further performs matching analysis on the bidding terms and corresponding bid content entities based on different domain parsing engines to obtain the domain analysis results corresponding to the bidding terms; further, it performs consistency analysis based on the domain analysis results to obtain the comprehensive analysis results of supplier bid matching degree.
[0043] Preferably, the method further includes:
[0044] Based on the bid matching results of each supplier, S6 generates a list of candidate suppliers from the N suppliers with the highest bid matching degree.
[0045] The above-described embodiments of the present invention use the tender documents from the tendering party and the bid documents from the suppliers as a basis. Relevant tender clauses are extracted from the tender documents, and then entity extraction is performed on the bid documents accordingly. The bid content entities contained in the bid documents are aligned with the corresponding tender clauses, thereby improving the efficiency and intelligence level of standardized processing of tender data. Further, hard index analysis is performed based on the aligned tender clauses and bid content entities. This allows for analysis of suppliers' compliance with index-related clauses, thereby selecting suppliers that meet the standard requirements. For suppliers that meet the basic requirements, a multi-domain analytical scheme is further used to analyze non-index-related clauses or bid content exceeding the standard. The suitability of supplier conditions to clauses is analyzed from different domain perspectives, yielding a comprehensive analysis result. This helps improve the objectivity and accuracy of quantitative evaluation of non-standard bid content entities based on a large language model, thereby enhancing the intelligence level of supplier selection.
[0046] Among them, large language models have excellent natural language processing capabilities. These models are pre-trained on a large scale and learn the syntax, structure and semantics of language from massive amounts of bidding and procurement data. With the help of large language models, bidding and procurement companies can more easily input supplier data. By combining this large amount of data information, a more scientific and effective bidding and tendering scoring index system can be established, thereby selecting the best supplier. This is of great value to improving the electronic bidding and tendering management level of bidding and procurement companies.
[0047] Preferably, step S1 includes:
[0048] Obtain the tender documents provided by the tendering party, which include technical specifications and procurement terms and conditions documents, etc.
[0049] Obtain the tender documents provided by the supplier, which include the tender letter, qualification certificate, financial statements, construction plan, quotation, etc.
[0050] The document formats for tender documents and bid documents include PDF, DOCX, XLSX, JPEG, etc.
[0051] During the bidding process, both the bidding party and suppliers submit documents in various formats (such as PDF contracts, scanned copies, Excel spreadsheets, etc.). Traditionally, these documents are read manually, and key information is recorded, which is time-consuming and prone to errors. This invention automatically extracts structured text from bidding documents and tender documents based on a large language model, and further extracts key information, thus improving the efficiency and intelligence of document data processing.
[0052] Preferably, step S2 includes:
[0053] Based on the LLM large language model, the unstructured data of the tender document is first extracted by the parsing tool, including the original text, paragraph marks and tables, and then the unstructured data is transformed into a processable text sequence.
[0054] The semantic understanding module performs keyword matching or contextual analysis on the text sequence to obtain the tender terms.
[0055] By combining the clause and rule templates and the fine-tuning model, structured bidding clause information is output, which includes clause items and clause content.
[0056] In one scenario, the LLM large language model uses built-in OCR or PDF parsing tools to extract the raw text of bidding documents, identify paragraphs, tables, and other structures, and transform unstructured data into a processable text sequence. Through self-attention mechanisms and positional encoding techniques, it achieves keyword matching (e.g., "bid security") and contextual analysis (e.g., determining clause relevance). Combining bidding clause sentence structure rules and domain-specific fine-tuning models, it outputs structured JSON results.
[0057] Based on a large language model that incorporates domain-specific fine-tuning technology, it is able to intelligently extract bidding clauses from tender documents, thereby improving the level of intelligence in obtaining key clause information.
[0058] Preferably, step S3 includes:
[0059] Based on the LLM large language model, the unstructured data of the tender documents, including raw text, paragraph marks and tables, are first extracted by parsing tools and then transformed into a processable text sequence.
[0060] The semantic parsing model based on domain fine-tuning extracts key entities from the obtained text sequence to obtain the bidding content entity, which includes entity category and entity content.
[0061] Semantic similarity is calculated based on the obtained bid content entities and bidding terms. The bid content entities and bidding terms are then associated based on semantic similarity to obtain an alignment set.
[0062] In one scenario, a domain-fine-tuned semantic understanding model specifically includes:
[0063] Based on the Llama base model, a domain-fine-tuned semantic parsing engine is trained using a large amount of bidding document data and related policy and regulatory documents as training data. It extracts target text from text sequences based on methods such as paragraph tagging or logical paragraph segmentation. During entity recognition of the target text, key entity extraction is performed through a key entity recognition function set by the parsing engine. The key entity recognition function is as follows:
[0064] in, This represents the probability that the target text belongs to entity i. This represents the trainable weight matrix for entity i. This represents the 768-dimensional text feature vector obtained after processing the target text using an LLM text encoder. This represents the bias term corresponding to entity i; This indicates that variable j belongs to the entity set. ; This represents the trainable weight matrix of entity j. This represents the bias term for the corresponding entity j;
[0065] Based on the probability that the target text belongs to each entity, the entity with the highest probability is selected as the key entity of the target text.
[0066] The key entities pre-defined include registered capital, ISO certification, historical litigation, technical standards, service plans, and product parameters.
[0067] In one scenario, the LLM large language model uses built-in OCR or PDF parsing tools to extract the original text of the tender document, identify paragraphs, tables, and other structures, and transform unstructured data into a processable text sequence. Based on a domain-fine-tuned LLM semantic parsing engine, entity extraction is performed on the text sequence to obtain the entity categories and corresponding entity content. Based on the obtained entity categories, they are associated with the corresponding tender clause items based on semantic similarity. The corresponding tender text content and tender clause text content are then aligned to obtain an alignment set.
[0068] Based on the large language model, the tender documents are further processed. The entity content of the tender documents is first extracted, and then the test content is aligned with the corresponding tender terms. This achieves cross-document content matching between the tender documents and the tender documents, and realizes the automatic extraction of the supplier's response content to the tender terms. This lays the foundation for subsequent targeted matching analysis of the tender terms and the corresponding tender content entities.
[0069] Preferably, step S4 includes:
[0070] Hard indicator analysis is performed on the bidding terms and corresponding bid content entities. The hard indicator analysis function used is as follows:
[0071]
[0072] in, This represents the deviation factor of hard indicators. This represents the total number of tender terms, with variable k representing the k-th tender term. This represents the weight of the k-th tender clause. Let represent the clause attribute factor of the k-th tender clause, where the k-th tender clause is a conditional clause. Otherwise, when the k-th tender clause is a non-conditional clause, ; This represents the content of the k-th tender clause. BERT vector; This represents the BERT vector representing the content of the bid entity corresponding to the k-th tender clause. This represents a conditional function, when... and cosine similarity Greater than the set threshold hour, Otherwise when hour, ;
[0073] Based on the obtained hard index deviation factor ,when Less than the set deviation threshold When the hard indicator analysis result is obtained, it is considered passed; otherwise, when Greater than or equal to the set deviation threshold At that time, the hard indicator analysis result was "not passed".
[0074] In one scenario, a threshold is set. , or settings .
[0075] In one scenario, regarding the weighting of bidding terms The settings are configured so that when the corresponding clause is a breach of contract clause or a standard regulatory clause, The larger the value, for example When the corresponding terms are ordinary terms, The smaller the value, for example .
[0076] In one scenario, based on the hard indicator deviation factor The range of values is , among which when This indicates that the hard criteria are fully met. When a deviation threshold is set... .
[0077] In one scenario, tender documents for procurement or technical services may include conditional clauses (generally mandatory minimum standards) such as: independent legal person status (requiring bidders to be legal persons or other organizations with independent civil liability capacity; branch offices must provide authorization letters and business licenses from their parent companies); technical team personnel configuration (specifying the number or qualifications of experts participating in the technical services); specific performance requirements (requiring bidders to provide evidence of completing two similar projects within the past three years (e.g., hospital project experience for medical equipment procurement), but not limited to specific administrative regions or industries); proof of financial capacity (requiring bidders to provide audited financial statements or bank credit certificates showing sound financial condition and no significant losses within the past three years); and technical qualification certification (specifying that the tendered products must pass national mandatory certification (e.g., 3C certification) or industry-specific standards (e.g., ISO). The requirements include: 13485 Medical Device Quality Management System; no related restrictions (prohibiting entities that have provided design or consulting services for the preliminary stages of the bidding project from participating in the bidding to prevent conflicts of interest); hardware equipment requirements (e.g., minimum requirements for server parameters); material parameter requirements (e.g., supplied materials must meet relevant standards and parameter requirements), etc. Non-conditional clauses include optional accessory brands (the scoring clauses require specifying the brands of hardware, locks, etc., but bidders are allowed to choose other brands (proof of equivalent quality is required)); recommended technical parameters (bidders are allowed to provide technical solutions "equivalent" to the parameters recommended in the bidding documents), etc.
[0078] The conditional clauses can be obtained through pre-calibration or by adaptive calibration based on historical standard data.
[0079] In one scenario, if the supplier's hard indicator analysis result is "not passed", the supplier's bid evaluation result is directly "not passed"; if the supplier's hard indicator analysis result is "passed", then the analysis proceeds to the next step.
[0080] In the above-described embodiments of the present invention, when performing hard indicator analysis, the content of the bidding terms and the corresponding bidding entity content are transformed into mathematical vectors, and the similarity of the vectors is used to intelligently analyze and review whether the supplier meets the standards, thereby reviewing the supplier's qualifications and other hard indicators, which helps to improve the efficiency of supplier review.
[0081] Considering that reviews of unconditional or standard-level content often lack unified standards, they typically rely on subjective human evaluation. Traditional expert reviews require the participation of experts from different fields, leading to significant coordination difficulties and high human resource costs. While some existing technologies utilize deep learning models to simulate expert reviews of unconditional metrics, in practical applications, these trained models cannot cover the diverse considerations arising from different real-world situations, resulting in insufficient objectivity and accuracy in the review results.
[0082] In response to the above situation, this invention also proposes a technical solution based on a multi-domain parsing engine to objectively and professionally analyze the content of supplier tender responses from different domain perspectives, so as to improve the objectivity and accuracy of reviewing non-indicator tender terms responses based on a large language model.
[0083] Preferably, step S5 includes:
[0084] Based on the defined domains, the corresponding domain parsing engines are invoked respectively. The defined domains include the technical domain, the financial domain, and the legal domain, and the domain parsing engines include the technical domain parsing engine, the financial domain parsing engine, and the legal domain parsing engine.
[0085] The domain parsing engine has a scoring model. The scoring model is used to match and analyze the bidding terms and the corresponding bidding content entities to obtain the corresponding domain scores. The domain scores include technical domain scores, financial domain scores and legal domain scores.
[0086] Consistency analysis based on domain scores yields a comprehensive analysis result of supplier bid matching.
[0087] Among them, the technical field analysis engine focuses on analyzing the technical feasibility of the bid content entities under different bidding terms. The technical field analysis engine is trained based on product performance evaluation data, historical technical solution review report data and corresponding review results as training datasets, so as to score the technical field of the bid content entities under the bidding terms.
[0088] The financial domain analysis engine focuses on analyzing the economic / cost effects of bid content entities under different bidding terms. The technical domain analysis engine is trained using product cost-benefit data, corporate financial analysis data, and corresponding review results as training datasets, thereby enabling it to score bid content entities under bidding terms in the financial domain.
[0089] The legal domain analysis engine focuses on analyzing the risk of violations of the bidding content entities under different bidding terms. The technical domain analysis engine is trained based on standard regulations, historical judgment documents, penalty case data, and corresponding review results as training datasets, thereby enabling it to score the legal domain of the bidding content entities under the bidding terms.
[0090] The domain score ranges from 0 to 10;
[0091] In one scenario, the domain parsing engine is an LLM model with a built-in analysis module. Different domain parsing engines are equipped with LoRA adapters corresponding to the domain. The LoRA adapter components are used to fine-tune the LLM large language model for the domain, so that the LLM large language model can make professional responses for the corresponding domain, thereby obtaining the analysis and evaluation results of the corresponding domain.
[0092] Preferably, consistency analysis is performed based on domain scores, specifically including:
[0093] Obtain domain score sets from different domain parsing engines for the tender terms and corresponding bid content entities. Among them, the domain score set At least including ; These represent the technical field score, financial field score, and legal field score for the current k-th tender clause, respectively.
[0094] The consistency quantification factor is calculated based on domain scoring, and the consistency quantification factor calculation function used is as follows:
[0095]
[0096] in, Represents the consistency quantification factor, variable Represented as the domain score set The elements in Represents the domain score set The total number of scores in the fields, Represents the domain score set The average score across all areas, This represents the historical domain score standard deviation for the k-th tender clause, obtained from the standard deviation of scores for each domain in the historical analysis data for the k-th tender clause; This indicates the maximum set score difference value;
[0097] Based on the obtained consistency quantification factor ,when Less than the set threshold At that time, the matching score corresponding to the k-th bidding clause is recorded as the current domain score set. Average scores across all areas Otherwise when Greater than or equal to the set threshold If a conflict occurs, a conflict flag is set for the current k-th bidding clause, and the matching score of the current record corresponding to the k-th bidding clause is recorded as the current domain score set. Minimum score for each area ;
[0098] The comprehensive analysis results of supplier bid matching degree are obtained based on the matching degree scores of each bidding clause.
[0099] In one scenario, the maximum score difference is set. .
[0100] In one scenario, a comprehensive analysis report of the supplier's bids is generated based on the matching scores of each bidding clause. This report includes the supplier's matching scores for different bidding clauses, as well as corresponding conflict markers. The conflict markers allow professionals to review the conflict situations and re-determine the corresponding matching scores. The comprehensive analysis report also includes the supplier's overall matching score, which is obtained through methods such as weighted averaging, average, summarizing, or weighted summarizing the matching scores for different bidding clauses.
[0101] The above-described embodiments of the present invention evaluate the suitability of suppliers to bidding terms from the perspectives of experts in different fields. By training analytical models for different fields, a matching analysis of the bid content is performed based on a professional field perspective, thus completing the field evaluation. Further, consistency analysis and quantification are conducted based on the evaluation results from different fields to extract the consistency and conflict characteristics of the evaluation results. When the consistency of evaluation results based on knowledge from different fields meets the standards, the matching analysis results based on different fields are considered to have high credibility, and a final comprehensive evaluation result is obtained based on field scores. However, when the consistency analysis results do not meet the standards, the matching analysis results from different fields are considered to conflict (in this case, a single result cannot reflect the objectivity of the analysis results), and the analysis results are further marked with conflict so that the final matching evaluation result is determined through expert analysis. By simulating a joint review method by experts from different fields, the objectivity and accuracy of the review can be improved, as well as the efficiency of the review.
[0102] By setting rules to obtain the comprehensive matching score of suppliers for different bidding terms, it can adapt to the needs of different bid evaluation requests and be used in different bid evaluation scenarios.
[0103] In one scenario, based on the comprehensive analysis results of supplier bid matching, only the parts marked as conflicting need to undergo further multi-domain expert review, effectively reducing the manpower cost and efficiency of the review.
[0104] This invention also proposes a bidding and procurement evaluation system based on large language model technology, including a processor, wherein the processor is used to execute the above-described... Figure 1 The bidding and procurement evaluation method based on large language model technology is shown, along with the specific implementation methods for each step. This invention will not be described again here.
[0105] It should be noted that the functional units / modules in the various embodiments of the present invention can be integrated into one processing unit / module, or each unit / module can exist physically separately, or two or more units / modules can be integrated into one unit / module. The integrated unit / module described above can be implemented in hardware or in the form of software functional units / modules.
[0106] From the above description of the embodiments, those skilled in the art will clearly understand that the embodiments described herein can be implemented in hardware, software, firmware, middleware, code, or any suitable combination thereof. For hardware implementation, the processor can be implemented in one or more of the following units: Application-Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field-Programmable Gate Array (FPGA), processor, controller, microcontroller, microprocessor, other electronic units designed to implement the functions described herein, or combinations thereof. For software implementation, some or all of the processes of the embodiments can be implemented by a computer program instructing the associated hardware. During implementation, the program can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of a computer program from one place to another. Storage media can be any available medium accessible to a computer. Computer-readable media can include, but is not limited to, RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code having the form of instructions or data structures and accessible to a computer.
[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should be able to analyze that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the essence and scope of the technical solutions of the present invention.
Claims
1. A bidding and procurement evaluation method based on large language model technology, characterized in that, Includes the following steps: S1 obtains the tender documents and bid documents; S2 processes tender documents based on the LLM large language model and extracts the corresponding tender clause feature set, which includes one or more tender clauses. S3 processes tender documents based on the LLM large language model, extracts tender content entities from the tender documents, and aligns the tender content entities to the corresponding tender terms. S4 performs hard indicator analysis based on the bidding terms and corresponding bid content entities, and obtains the hard indicator analysis results, specifically... include: This involves a hard indicator analysis of the bidding terms and corresponding bid content entities, using the following hard indicator analysis function: in, This represents the deviation factor of hard indicators. This represents the total number of tender terms, with variable k representing the k-th tender term. This represents the weight of the k-th tender clause. Let represent the clause attribute factor of the k-th tender clause, where the k-th tender clause is a conditional clause. Otherwise, when the k-th tender clause is a non-conditional clause, ; This represents the content of the k-th tender clause. BERT vector; This represents the BERT vector representing the content of the bid entity corresponding to the k-th tender clause. This represents a conditional function, when... and cosine similarity Greater than the set threshold hour, Otherwise when hour, ; Based on the obtained hard index deviation factor ,when Less than the set deviation threshold When the hard indicator analysis result is obtained, it is considered passed; otherwise, when Greater than or equal to the set deviation threshold When the hard indicator analysis result is "not passed", When the hard indicator analysis results pass, S5 further performs matching analysis on the bidding terms and corresponding bid content entities based on different domain parsing engines to obtain the domain analysis results corresponding to the bidding terms; further, it performs consistency analysis based on the domain analysis results to obtain the comprehensive analysis results of supplier bid matching degree.
2. The bidding and procurement evaluation method based on large language model technology according to claim 1, characterized in that, The method also includes: Based on the bid matching results of each supplier, S6 generates a list of candidate suppliers from the N suppliers with the highest bid matching degree.
3. The bidding and procurement evaluation method based on large language model technology according to claim 1, characterized in that, Step S1 includes: Obtain the tender documents provided by the tendering party, which include technical specifications and procurement terms and conditions. Obtain the tender documents provided by the supplier, which include a letter of tender, qualification certificate, financial statements, construction plan, and quotation. The document formats for the tender documents and bid documents include PDF, DOCX, XLSX, and JPEG.
4. The bidding and procurement evaluation method based on large language model technology according to claim 1, characterized in that, Step S2 includes: Based on the LLM large language model, the unstructured data of the tender document, including the original text, paragraph marks and tables, is first extracted by the parsing tool and then transformed into a processable text sequence. The semantic understanding module performs keyword matching or contextual analysis on the text sequence to obtain the tender terms. By combining the clause and rule templates and the fine-tuning model, structured bidding clause information is output, which includes clause items and clause content.
5. The bidding and procurement evaluation method based on large language model technology according to claim 4, characterized in that, Step S3 includes: Based on the LLM large language model, the unstructured data of the tender documents, including raw text, paragraph marks and tables, are first extracted by parsing tools and then transformed into a processable text sequence. The semantic parsing model based on domain fine-tuning extracts key entities from the obtained text sequence to obtain the bidding content entity, which includes entity category and entity content. Semantic similarity is calculated based on the obtained bid content entities and bidding terms. The bid content entities and bidding terms are then associated based on semantic similarity to obtain an alignment set.
6. The bidding and procurement evaluation method based on large language model technology according to claim 5, characterized in that, Domain-based fine-tuning semantic understanding models specifically include: Based on the Llama base model, a domain-fine-tuned semantic parsing engine is trained using a large amount of bidding document data and related policy and regulatory documents as training data. It extracts target text from text sequences based on paragraph tagging or logical paragraph segmentation. During entity recognition of the target text, key entity extraction is performed using a key entity recognition function set by the parsing engine. The key entity recognition function is as follows: in, This represents the probability that the target text belongs to entity i. This represents the trainable weight matrix for entity i. This represents the 768-dimensional text feature vector obtained after processing the target text using an LLM text encoder. This represents the bias term corresponding to entity i; This indicates that variable j belongs to the entity set. ; This represents the trainable weight matrix of entity j. This represents the bias term for the corresponding entity j; Based on the probability that the target text belongs to each entity, the entity with the highest probability is selected as the key entity of the target text.
7. The bidding and procurement evaluation method based on large language model technology according to claim 1, characterized in that, Step S5 includes: Based on the defined domains, the corresponding domain parsing engines are invoked respectively. The defined domains include the technical domain, the financial domain, and the legal domain, and the domain parsing engines include the technical domain parsing engine, the financial domain parsing engine, and the legal domain parsing engine. The domain parsing engine has a scoring model. The scoring model is used to match and analyze the bidding terms and the corresponding bidding content entities to obtain the corresponding domain scores. The domain scores include technical domain scores, financial domain scores and legal domain scores. Consistency analysis based on domain scores yields a comprehensive analysis result of supplier bid matching.
8. The bidding and procurement evaluation method based on large language model technology according to claim 7, characterized in that, Consistency analysis is performed based on domain scores, specifically including: Obtain domain score sets from different domain parsing engines for the tender terms and corresponding bid content entities. Among them, the domain score set At least including ; These represent the technical field score, financial field score, and legal field score for the current k-th tender clause, respectively. The consistency quantification factor is calculated based on domain scoring, and the consistency quantification factor calculation function used is as follows: in, Represents the consistency quantification factor, variable Represented as the domain score set The elements in Represents the domain score set The total number of scores in the fields, Represents the domain score set The average score across all areas, This represents the historical domain score standard deviation for the k-th tender clause, obtained from the standard deviation of scores for each domain in the historical analysis data for the k-th tender clause; This indicates the maximum set score difference value; Based on the obtained consistency quantification factor ,when Less than the set threshold At that time, the matching score corresponding to the k-th bidding clause is recorded as the current domain score set. Average scores across all areas Otherwise when Greater than or equal to the set threshold If a conflict occurs, a conflict flag is set for the current k-th bidding clause, and the matching score of the current record corresponding to the k-th bidding clause is recorded as the current domain score set. Minimum score for each area ; The comprehensive analysis results of supplier bid matching degree are obtained based on the matching degree scores of each bidding clause.
9. A bidding and procurement evaluation system based on large language model technology, characterized in that, Includes a processor, wherein the processor is used to execute the bidding and procurement evaluation method based on large language model technology as described in any one of claims 1-8 above.