Method for intelligently matching bidding and tendering documents based on large model

Abstract

The invention provides a method for intelligently matching bidding and tendering documents based on a large model. The method comprises the following steps: step 1, preprocessing the bidding and tendering documents; step 2, carrying out data cleaning on the pre-processed document so as to establish a pre-training data set; step 3, processing the txt text after data cleaning, and splitting the txt text into a plurality of index items; 4, questioning each index in the index text by using an instruction following function in the large model so as to generate a fine tuning data set; 5, pre-training and fine tuning are carried out based on the base large model; and step 6, constructing the cue words, the index items and the corresponding bidding file fragments into a template, inputting the template into the trained large language model, and guiding the trained large language model to perform reasoning and answering. The answering accuracy of the large model in the professional field is improved, so that the model is more excellent in the answering process.

Description

Technical Field

[0001] The present application relates to the technical field of target recognition and matching, and in particular to a method for intelligently matching bidding documents based on a large model. Background Art

[0002] In the bidding process, reviewing the bid documents to ensure that they meet the requirements of the bidding documents is a key step. Traditionally, this process mainly relies on manual review, and the reviewers need to check each indicator in each bid document one by one to ensure that it meets the requirements of the bidding documents. However, the manual review process is cumbersome and time-consuming, especially when the number of bid documents is large, it is difficult for the reviewers to complete the review of all documents in a short time. In addition, different reviewers may have different understandings and judgment criteria of the indicators, resulting in inconsistency and subjectivity in the review results. Due to the intensity and complexity of manual review, reviewers may become fatigued after working for a long time, which leads to errors and omissions in the review. Especially in large-scale bidding projects, the number of bid documents is huge, and manual review is difficult to cope with, and it is impossible to complete the processing and review of all documents in a short time.

[0003] In order to solve these problems, some companies have tried to introduce computer-aided tools to improve the efficiency of auditing. Existing technologies have alleviated some of the problems in the audit of bid documents to a certain extent, but there are still many challenges to be solved. Most of these tools rely on simple keyword matching and rule judgment, and it is difficult to handle complex indicator analysis and judgment requirements. For example, some systems screen bid documents through preset rules, but these rules often cannot cover all possible situations, resulting in the accuracy and comprehensiveness of the audit results being limited. Large language model technology provides a new solution for the audit of bid documents. Through training on a large amount of data, large models can understand and process complex language and indicator relationships, and provide more accurate and efficient audit capabilities. These models can not only quickly process and analyze a large number of bid documents and improve audit efficiency, but also enhance audit consistency and reduce deviations caused by human subjectivity through standardized algorithms and model training. In addition, large models can identify subtle differences and errors, reduce omission rates, adapt to complex indicator analysis requirements, and significantly improve the accuracy and comprehensiveness of audits. Although large language model technology can be used to match bids, there are still some challenges in the audit of bid documents using large language model technology.

[0004] Specifically, the following issues are included

[0005] 1. Traditional bidding documents are usually a large number of electronic documents, which contain rich content, including text, tables and other information. These contents need to be converted into a format and structure that can be understood by the big model. However, this conversion process is not simple. How to effectively extract and structure this information so that the big model can accurately understand and process it is a problem that needs to be solved. Ensuring that the model can comprehensively and accurately parse different types of data is the key to improving audit efficiency.

[0006] 2. When processing bidding documents, large models often face the challenge of understanding the differences between professional terms and different indicators. Bidding documents in different fields may contain a large number of professional terms, and the meanings and usage scenarios of these terms may be very complex. In addition, the relationship between different indicators and the similarities in specific situations are also key issues that the model needs to understand. How to enable large models to accurately understand and distinguish these professional terms and indicator relationships is another important challenge to improve audit accuracy. Ensuring that the model can not only recognize professional terms, but also correctly understand their meaning in different contexts is crucial to achieving high-accuracy audits. Summary of the invention

[0007] This application provides a method for intelligently matching bidding documents based on a large model, which can be used to solve the technical problem that the large model cannot accurately understand the characteristics of a specific field.

[0008] The present application provides a method for intelligently matching bidding documents based on a large model, the method comprising:

[0009] Step 1: Pre-process the bidding documents;

[0010] Step 2: Clean the preprocessed documents to establish a pre-training data set; data cleaning includes deduplication and filtering of useless information

[0011] Step 3: Process the txt text after data cleaning and split the txt text into multiple indicator items;

[0012] Step 4: Use the instruction following function in the large model to ask questions about each indicator in the indicator text to generate a fine-tuning data set;

[0013] Step 5: Pre-train and fine-tune based on the base large model;

[0014] Step 6: Construct the prompts, indicator items and corresponding bidding document fragments into a template, and input the template into the trained large language model to guide the trained large language model to infer the answer.

[0015] Furthermore, the bidding documents are pre-processed, including:

[0016] Step 11: Obtain bidding and tender documents from various public platforms;

[0017] Step 12: Use the hash value of the bidding documents to remove duplicates by calculating the SHA-256 hash value of each document and deleting the documents corresponding to the duplicate hash values;

[0018] Step 13: Convert the word document into a string format. The conversion method is as follows:

[0019] (S, T) = R (D)

[0020] Where D represents the input word document, S represents the output document content in string form, T represents the extracted table data list, and R represents the function that converts the Word document into a string.

[0021] Step 14, using the table mark | as a separator to extract T from the table; converting the content of the table T into an HTML format string;

[0022] The | separator is used to identify table rows in a document;

[0023] Step 15, get all the table contents by identifying the table tag;

[0024] Step 16, for the converted text string, i.e., the txt text, find the chapter title by identifying the field such as the chapter title, and segment the txt text according to the chapter title to form different paragraphs; wherein the chapter title includes "Chapter XX";

[0025] The bold text in the word document is wrapped with ** when it is converted into a text string.

[0026] Furthermore, the pre-training dataset in step 2 is established by the following method:

[0027] Step 21, removing duplicate information in the pre-processed document, i.e., the txt text, where the duplicate information includes conventional descriptions in the bidding documents, including bidder instructions, government procurement policies, qualification review and bid evaluation;

[0028] Step 22, removing useless information in the data including personal information and corporate information;

[0029] Step 23, use the txt text after data cleaning as a pre-training data set for subsequent processing.

[0030] Furthermore, in step three, the txt text after data cleaning is processed and the txt text is split into multiple indicator items, including:

[0031] Step 31, set a json format template; use the chapter title **Chapter** as the key value, and the content of each chapter as the value;

[0032] Step 32, determine the highest level sequence number, traverse each chapter content, find the highest level sequence number, including "one," "two," and "three"; the highest level sequence number is used as the basis for content division; according to the highest level sequence number found, divide each chapter content into multiple parts;

[0033] Step 33, processing the content of each part: for each part of the content, continue to search for the highest level sequence number within each part; if a part has a secondary sequence number, further divide it according to the secondary sequence number, and for each secondary sequence number, divide the content again;

[0034] Step 34, content separation, in the content of each serial number, use ":" and "." as the basis for separating index items; further subdivide the content into multiple index items.

[0035] Step 35, start parsing from the outermost chapter title, extract only the bottom-level key and value pairs and concatenate them as each index item; save the extracted index items into the index txt text;

[0036] Step 36, split the table content into indicator items, and combine each row of the table content with the title to form a clear indicator item.

[0037] Furthermore, in step 4, fine-tuning the dataset is achieved by the following method:

[0038] Step 41, set prompts, use the instruction following capability of the large model to generate response indicators in the bidding document for each indicator in the indicator txt text; each indicator includes data size comparison, data range, special symbol recognition and relevant industry standards; wherein the prompt is: please generate response indicators that meet or do not meet the requirements according to the following indicator requirements; in each indicator, the response indicator meets the requirements of the indicator item;

[0039] Data size comparison, including comparison of different units and application of “minimum requirements” and “maximum limits”;

[0040] The scope of the data, including a clear definition of the “scope”;

[0041] Special symbol recognition, emphasizing the impact of “≥”, “+”, and “-” on the meaning of indicator items;

[0042] Relevant industry standards, including different quality management system requirements;

[0043] Step 42, set the prompt, use the command following ability of the large model, answer according to the indicator item and the corresponding response indicator, and judge whether the response indicator meets the requirements of the indicator item; the prompt is as follows;

[0044] Prompt: Please judge whether the response indicator meets the requirements of the indicator item. If it does, answer "yes", otherwise answer "no".

[0045] Step 43: Integrate the prompts, indicator items, and corresponding response indicators in step 42 and save them as a fine-tuning dataset for model fine-tuning:

[0046] {P, I1, I2, A}

[0047] Among them, P is the prompt, I2 is the indicator item of the bidding document, I1 is the response indicator generated by I2 through the instruction following ability of the large language model, and A is the answer generated by the large model.

[0048] Furthermore, in step five, pre-training and fine-tuning are performed based on the base large model, including:

[0049] Step 51, using an open source large base model as a base model;

[0050] Step 52, incremental pre-training of the base large model; the warm-up phase is used to use a lower learning rate of 1e-4 at the beginning of training, and gradually increase it to a preset learning rate to stabilize the training process;

[0051]

[0052] where η t The learning rate at step t; η max is the preset maximum learning rate, which is the learning rate that the model should reach after the warmup phase ends; t Warmup is the total number of steps in the warmup phase;

[0053] Step 53, in the incremental pre-training process, the pre-training data set is mixed with a small amount of general knowledge domain data for training to keep the model's memory of old knowledge; in the mixed data set, the general knowledge domain data accounts for 70%; the training is carried out in stages, and new data is gradually added;

[0054] Step 54, fine-tune LoRA on the pre-trained model by introducing low-rank matrix decomposition into the fully connected layer weight matrix of the model;

[0055] W∈R d×k , A∈R d×r and B∈R r×k , where r<<min(d,k)

[0056] Where W ≈ AB, A and B are low-rank matrices obtained through training, and r is the rank of matrices A and B.

[0057] Furthermore, in step six, the prompts, indicator items and corresponding bidding document fragments are constructed into a template, and the template is input into the trained large language model to guide the trained large language model to reason and answer, including:

[0058] Step 61, using the txt text in the bidding document after data cleaning in step 2; then segmenting the txt file into multiple segments in a manner that each paragraph has 300 characters and each segmentation repeats 100 characters;

[0059] P i =T[(i-1)·(NR)+1:(i-1)·(NR)+N]

[0060] Where P i represents the i-th segment, T represents the txt text after data cleaning, N represents the fixed character length of the paragraph, which is 300, and R represents the repeated characters each time the segmentation is performed, which is 100;

[0061] Step 62, for each index {I1, I2, I3...I n} and multiple segments of the bidding document {P1, P2, P3…P n} Perform similarity matching;

[0062] Step 63: Set the index {I1, I2, I3…I n}Text and fragments {P1,P2,P3…P n} is converted into a vector. The vectorized indices are {V(I1), V(I2), V(I3)…V(I n )} and {V(P1),V(P2),V(P3)…V(P n )};

[0063] Step 64, for each indicator vector V(I i ) and each bidding segment vector V(P j ), calculate the cosine similarity between them:

[0064]

[0065] Among them, V(I i ).V(P j ) represents the dot product of vectors, ∥V(I i )∥ and∥V(P j )∥respectively represent the modulus length of the vector;

[0066] Step 65, for each indicator I i , find the bidding segment P that maximizes the cosine similarity j ;

[0067] Step 66, each indicator I i , corresponding to the bidding document fragment P j And the prompt are spliced ​​into a template; the specific template is as follows:

[0068] Prompt: Determine whether the following indicators are consistent with the content in the clip. Please answer yes or no.

[0069] Index I i :Specific indicator content;

[0070] Fragment P j : Specific fragment content;

[0071] Step 67: Input the template content into the trained model and generate the answer using the reasoning and generation capabilities of the large language model.

[0072] This application innovatively proposes to convert the bidding documents into text strings, and proposes a systematic method to structure the complex content of the bidding documents through hierarchical serial numbers and specific separators. The core is to use the highest and secondary serial numbers to divide the layers layer by layer, and to subdivide the content into clear indicators through specific symbols, thereby improving the processability and readability of the data and laying the foundation for subsequent model training.

[0073] This application innovatively proposes a method for training large models. During the model training process, gradual pre-training is performed by adopting an initial low learning rate to avoid gradient explosion caused by excessive learning rate. At the same time, the pre-training data set is mixed with a small amount of general knowledge data, and a phased training strategy is adopted. In addition, LoRA technology is introduced during the training process to reduce computing power consumption through low-rank matrix decomposition, thereby improving training efficiency. For the construction of the fine-tuning set, the particularity of the indicators in the bidding documents is focused on, covering aspects such as data comparison, range, special symbols, etc., in order to improve the accuracy of the large model's answers in these professional fields, so that the model can perform better when answering.

[0074] This application innovatively proposes a result analysis method, which vectorizes the indicators and bid texts, performs similarity matching to find the most relevant text fragments, and uses specific prompt words to splice the indicators with the matching text fragments so that the model can correctly judge the results. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1A flow chart of the complete technical solution for large model training provided in the embodiment of the present application;

[0076] Figure 2 A flowchart of the technical solution for processing bidding documents provided in the embodiment of the present application;

[0077] Figure 3 Flowchart of the indicator processing technical solution provided for the embodiment of this application DETAILED DESCRIPTION

[0078] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below in conjunction with the accompanying drawings.

[0079] Before specifically introducing the method provided by the present application, the definitions involved in the present application are first explained so that those skilled in the art can better understand the present application. The definitions of the present application include:

[0080] Definition 1: A large model is a language representation model that has a large amount of training data, contains many parameters, and has strong understanding ability;

[0081] Definition 2: Bidding documents refer to the documents prepared by the purchaser for bidding activities, which describe in detail the background, technical requirements, bid evaluation criteria and contract terms of the bidding project;

[0082] Definition 3: Bidding documents refer to the documents that the bidder needs to prepare and submit according to the requirements of the bidding documents, including the bidder’s information, technical solutions and other contents that respond to the requirements of the bidding documents;

[0083] Definition 4: A pre-training dataset contains a large amount of text and is used for pre-training large models. It relies on a large amount of data to learn language rules, knowledge, and context understanding capabilities.

[0084] Definition 5: Fine-tuning datasets are used to further train pre-trained large models to adapt to specific tasks or application scenarios.

[0085] Definition 6: GPT-4 large model, i.e. a commercial large language model, has powerful task processing capabilities and can perform complex natural language processing tasks;

[0086] Definition 7: Command following function refers to the ability of the large language model to understand and execute specific tasks given by the user. This enables the large model to generate corresponding outputs based on the natural language instructions provided by the user.

[0087] Definition 8: Prompts are important text inputs used to guide the model to generate the required content. Prompts are designed differently for different tasks.

[0088] Definition 9: The performance of a large language model with public code and training data is generally difficult to compare with that of a commercial large language model.

[0089] Definition 10: Incremental pre-training means continuing training based on the existing pre-trained model (Qwen2-7B) to achieve better performance on a specific task;

[0090] Definition 11: The warmup phase is the phase in which the learning rate gradually increases from a smaller initial value to the preset learning rate to avoid gradient explosion caused by an excessively large learning rate at the beginning of training.

[0091] The method provided in this application includes:

[0092] Step 1: Pre-process the bidding documents;

[0093] Step 2: Clean the pre-processed documents, including deduplication and filtering useless information, to establish a pre-training data set;

[0094] Step 3: Process the txt text after data cleaning and split the txt text into multiple indicator items;

[0095] Step 4: Use the instruction following function in the GPT-4 large model to ask questions about each indicator in the indicator text to generate a fine-tuning data set;

[0096] Step 5: Pre-train and fine-tune based on the base large model;

[0097] Step 6: Construct the prompts, indicator items and corresponding bidding document fragments into a template, and input the template into the trained large language model to guide the trained large language model to infer the answer.

[0098] The detailed steps are described as follows:

[0099] For step 1, pre-processing of the bidding documents includes the following steps:

[0100] Step 11: Obtain bidding and tender documents from various public government and enterprise bidding platforms; the corresponding platforms provide a large amount of public bidding information as the main data source;

[0101] Step 12: Use the hash value of the bidding documents to remove duplicates by calculating the SHA-256 hash value of each document and deleting the documents corresponding to the duplicate hash values;

[0102] Step 13: Convert the word document into a string format. The conversion method is as follows:

[0103] (S, T) = R (D)

[0104] Where D represents the input word document, S represents the output document content in string form, T represents the extracted table data list, and R represents the function that converts the Word document into a string.

[0105] Step 14, using the table mark | as a separator to extract T from the table; converting the content of the table T into an HTML format string;

[0106] The | separator is used to identify table rows in a document;

[0107] Step 15, get all the table contents by identifying the table tag;

[0108] Step 16, for the converted text string, i.e., the txt text, find the chapter title by identifying the field such as the chapter title, and segment the txt text according to the chapter title to form different paragraphs; wherein the chapter title includes "Chapter XX";

[0109] The bold text in the word document is wrapped with ** when it is converted into a text string;

[0110] Step 2: Create a pre-training data set, including the following steps:

[0111] Step 21, removing duplicate information in the pre-processed document, i.e., the txt text, where the duplicate information includes conventional descriptions in the bidding documents, including bidder instructions, government procurement policies, qualification review and bid evaluation;

[0112] Step 22, removing useless information in the data including personal information and corporate information;

[0113] Step 23, use the cleaned txt text as a pre-training data set for subsequent processing;

[0114] Step 3: Split the cleaned txt text into multiple indicator items, including the following steps:

[0115] Step 31, set a json format template; use the chapter title **Chapter** as the key value, and the content of each chapter as the value;

[0116] Step 32, determine the highest level sequence number, traverse each chapter content, find the highest level sequence number, including "one," "two," and "three"; the highest level sequence number is used as the basis for content division; according to the highest level sequence number found, each chapter content is divided into multiple parts; for example, "one," appearing in the text will be marked as one part, followed by "two," and "three."

[0117] Step 33, processing the content of each part: for each part of the content, continue to search for the highest level sequence number (such as "1.", "(1)") within each part; if a part has a secondary sequence number, further divide it according to the secondary sequence number, and for each secondary sequence number, divide the content again;

[0118] Step 3.4: Content separation: In the content of each serial number, use ":" and "." as the basis for separating indicator items; further subdivide the content into multiple indicator items.

[0119] The details are as follows:

[0120] {"Chapter 4 Bidder Qualification Requirements":{

[0121] "I. Qualification requirements":"The bidder must have the corresponding medical device business license.",

[0122] "2. Performance requirements":{

[0123] "1. In the past three years": "There should be at least three successful cases of similar projects.",

[0124] "2.Customer Feedback":"Written feedback from customers is required."}},

[0125] "Chapter 5 Bidding Contents and Requirements":{

[0126] "I. Project Overview": "The medical equipment for this tender is 4 sets of surgical microscopes and other equipment required for the surgical care center and ophthalmology department.",

[0127] "2. Technical and service requirements":{

[0128] "1. Applicable Departments": "For surgical microsurgery.",

[0129] "2. Configuration requirements": "Three people and six eyes, main surgeon mirror, side assistant mirror, opposite assistant mirror, full HD video recording system, vascular fluorescence angiography function, yellow fluorescence tumor angiography function.",

[0130] "3.":"The latest apochromatic technology, high resolution, large depth of field.",

[0131] "4. Fusion Optical Technology": "Using the fusion characteristics of the human eye, the left and right optical paths are designed for depth of field and resolution respectively. One optical path maximizes the depth of field, and the other optical path maximizes the resolution.",

[0132] "5.":"The optical system magnification ratio is 1:6.",

[0133] "6. Magnification": "12.5x eyepiece, magnification between 1.0-16.0x, electric adjustment with manual emergency knob."}}}

[0134] Step 35, start parsing from the outermost chapter title, extract only the bottom-level key and value pairs and concatenate them as each index item; save the extracted index items into the index txt text;

[0135] Step 36, split the table content into indicator items, combine each row of the table content with the title to form a clear indicator item, so that the model can clearly understand the specific requirements of each indicator; the details are as follows:

[0136] Table 1: Some business requirements

[0137] Serial number type Require 1 Delivery time 90 days from the date of contract signing 2 Delivery location First Affiliated Hospital Binhai Campus

[0138] Indicators:

[0139] Serial number: 1; Type: Delivery time; Requirement: 90 days from the date of contract signing;

[0140] Serial number: 2; Type: Delivery location; Requirement: Binhai Campus of the First Affiliated Hospital;

[0141] Step 4: Generate a fine-tuning dataset, including the following steps:

[0142] The large model can effectively distinguish indicator items and response indicators in most cases, but there are still some special cases that need further study, mainly including: data size comparison, data range, special symbol recognition and related industry standards.

[0143] Step 41, set prompts, use the instruction following capability of the large model to generate response indicators in the bidding document for each indicator in the indicator txt text; each indicator includes data size comparison, data range, special symbol recognition and relevant industry standards; wherein the prompt is: please generate response indicators that meet or do not meet the requirements according to the following indicator requirements; in each indicator, the response indicator meets the requirements of the indicator item;

[0144] Data size comparison, including comparison of different units and application of “minimum requirements” and “maximum limits”;

[0145] For example: Index item: The storage capacity of the hard disk shall not be less than 500GB and shall not exceed 1TB.

[0146] Response indicators: The hard disk storage capacity is 800GB.

[0147] The scope of the data, including a clear definition of the “scope”;

[0148] Index items: Zoom range: Electric zoom adjustment, working distance range 225mm--600mm. Meet different surgical requirements for deep and shallow parts.

[0149] Response indicators: Zoom range: Electric zoom adjustment, working distance range 400mm-600mm. Meet different surgical requirements for deep and shallow parts.

[0150] Special symbol recognition, emphasizing the impact of “≥”, “+”, and “-” on the meaning of indicator items;

[0151] Index items: Main lighting power 300W (±5%), xenon lamp independent power supply electrical system.

[0152] Response indicators: Main lighting power 285W, xenon lamp independent power supply electrical system.

[0153] Relevant industry standards, including different quality management system requirements;

[0154] Index items: Equipment should comply with GB 9706 and ISO 9001 standards to ensure quality and safety.

[0155] Response indicators: Medical equipment meets GB9706 standards.

[0156] Step 42, set the prompt, use the command following ability of the large model, answer according to the indicator item and the corresponding response indicator, and judge whether the response indicator meets the requirements of the indicator item; the prompt is as follows;

[0157] Prompt: Please judge whether the response indicator meets the requirements of the indicator item. If it does, answer "yes", otherwise answer "no".

[0158] Step 43: Integrate the prompts, indicator items, and corresponding response indicators in step 42 and save them as a fine-tuning dataset for model fine-tuning:

[0159] {P, I1, I2, A}

[0160] Among them, P is the prompt, I2 is the indicator item of the bidding document, I1 is the response indicator generated by I2 through the instruction following ability of the large language model, and A is the answer generated by the large model;

[0161] Step 5: Pre-train and fine-tune the large model, including the following steps:

[0162] Step 51, using the open source large base model Qwen2-7B (with 7 billion parameters) as the base model;

[0163] Step 52, incremental pre-training of the base large model Qwen2-7B; the warmup stage is used to use a lower learning rate of 1e-4 at the beginning of training, and gradually increase it to a preset learning rate to stabilize the training process;

[0164]

[0165] where η t The learning rate at step t; η max is the preset maximum learning rate, which is the learning rate that the model should reach after the warmup phase ends; t Warmup is the total number of steps in the warmup phase;

[0166] Step 53, to prevent catastrophic forgetting during large model training, during the incremental pre-training process, the pre-training data set is mixed with a small amount of general knowledge data to maintain the model's memory of old knowledge; the general knowledge data mainly comes from Wikipedia; in the mixed data set, the proportion of general knowledge data should be properly controlled, accounting for 70%, to ensure that the model can absorb new knowledge while maintaining its understanding of existing knowledge; adopt a phased training method to gradually add new data so that the model can transition to new knowledge more smoothly and reduce the risk of sudden forgetting;

[0167] Step 54, perform LoRA fine-tuning on the pre-trained Qwen2-7B model; apply LoRA technology to specific fields (bidding text processing), introduce low-rank matrix decomposition in the fully connected layer weight matrix of the model, and only adjust a small number of parameters (A and B), which greatly reduces the consumption of computing resources, thereby performing fine-tuning efficiently;

[0168] W∈R d×k , A∈R d×r and B∈R r×k , where r<<min(d,k)

[0169] Where W≈AB, A and B are low-rank matrices obtained through training, and r is the rank of matrices A and B;

[0170] The large model that is pre-trained and fine-tuned can understand the professional terms and differences between different indicators in the bidding documents;

[0171] Step 6: Guide the trained large model to answer questions, including the following steps:

[0172] Step 61, using the txt text in the bidding document after data cleaning in step 2; then segmenting the txt file into multiple segments in a manner that each paragraph has 300 characters and each segmentation repeats 100 characters;

[0173] P i =T[(i-1)·(NR)+1:(i-1)·(NR)+N]

[0174] Where P i represents the i-th segment, T represents the txt text after data cleaning, N represents the fixed character length of the paragraph, which is 300, and R represents the repeated characters each time the segmentation is performed, which is 100;

[0175] Step 62, for each index {I1, I2, I3...I n} and multiple segments of the bidding document {P1, P2, P3…P n} Perform similarity matching;

[0176] Step 63: Set the index {I1, I2, I3…I n}Text and fragments {P1,P2,P3…P n} is converted into a vector. The vectorized indices are {V(I1), V(I2), V(I3)…V(I n )} and {V(P1),V(P2),V(P3)…V(P n )};

[0177] Step 64, for each indicator vector V(I i ) and each bidding segment vector V(P j ), calculate the cosine similarity between them:

[0178]

[0179] Among them, V(I i ).V(P j ) represents the dot product of vectors, ∥V(I i )∥ and∥V(P j )∥respectively represent the modulus length of the vector;

[0180] Step 65, for each indicator I i , find the bidding segment P that maximizes the cosine similarity j ;

[0181] Step 66, each indicator I i , corresponding to the bidding document fragment P j And the prompt are spliced ​​into a template; the specific template is as follows:

[0182] Prompt: Determine whether the following indicators are consistent with the content in the clip. Please answer yes or no.

[0183] Index I i: Specific indicators; for example, zoom range: electric zoom adjustment, working distance range 225mm--600mm. Meet different deep and shallow surgical requirements.

[0184] Fragment P j : Specific content of the clip; for example: 1. Configuration requirements: three people with six eyes, main surgeon mirror, side assistant mirror, opposite assistant mirror, full HD video recording system, vascular fluorescence angiography function, yellow fluorescence tumor angiography function. 2. Zoom range: electric adjustment zoom, working distance range 300mm-500mm. Meet the requirements of different deep and shallow surgeries. 3. Maximum working distance 800mm. 4. The total content of benzene, toluene, ethylbenzene and xylene is 30mg / kg.

[0185] Step 67, input the template content to the trained model, and generate answers using the reasoning and generation capabilities of the large language model.

[0186] The adoption of this technical solution has brought significant technical effects and advantages. First, compared with the traditional manual processing method, the system significantly improves the efficiency of bidding document processing. By converting complex bidding documents into multiple discernible indicators and using the reasoning ability of large models, human errors and subjective biases are reduced, ensuring the fairness of the bidding process. This not only greatly reduces the need for manual intervention, but also significantly improves the accuracy of document processing and review, thereby ensuring the reliability of the results.

[0187] The present application is further described below through a specific embodiment.

[0188] A hospital surgical microscope procurement bidding case:

[0189] A hospital planned to purchase a batch of surgical microscopes to improve the accuracy and success rate of surgical operations. The hospital published a bidding notice through an open bidding platform, attracting many suppliers to bid.

[0190] The bidding documents contain a large amount of information, including basic information such as project name, filing number, total procurement budget, etc. The procurement package also contains key information such as information on different equipment and technical requirements (specific indicators of the purchased goods, such as model, configuration, product description).

[0191] For the convenience of comparison, the bidding documents are first converted into text strings, and then data preprocessing is performed. Then they are segmented according to the chapter numbers, and the content is further subdivided according to the hierarchical numbers of the content of different chapters. In the content of each number, ":" and "." are used as the basis for separating the index items, so that the information is refined into multiple index items. According to the json format, the chapter number is used as the key value, and the content in each number is extracted as the value value. The json file is parsed starting from the outermost chapter title, and only the bottom-level key and value pairs are extracted and spliced ​​as each index item, finally forming a structured data.

[0192] Secondly, we convert the text of different bidding documents and extract the corresponding txt text. After dividing the txt text of the bidding document into multiple paragraphs, we match each indicator item of the bidding document to find the most similar segment. We construct a template with the given prompt, indicator item and corresponding bidding document segment, and input the template into the trained large language model to guide the trained large language model to reason and answer.

[0193] A report will be generated for each bid document, summarizing the answers to all retrieved indicators, making it easier for experts to review the key points that do not comply with the bidding documents.

[0194] Through the above steps, the hospital successfully matched the bidding documents that met the requirements of the bidding documents, ensuring the transparency and efficiency of the procurement process. The large language model performed well in matching various indicators of the bidding documents, helping the hospital select the most suitable surgical microscope supplier, greatly improving procurement efficiency and accuracy.

[0195] The above-described embodiments of the present application do not constitute a limitation on the protection scope of the present application.

Claims

1. A method for intelligently matching bidding documents based on a large model, characterized in that: The method comprises: Step 1: Pre-process the bidding documents; Step 2: Clean the preprocessed documents to establish a pre-training data set; data cleaning includes deduplication and filtering of useless information Step 3: Process the txt text after data cleaning and split the txt text into multiple indicator items; Step 4: Use the instruction following function in the large model to ask questions about each indicator in the indicator text to generate a fine-tuning data set; Step 5: Pre-train and fine-tune based on the base large model; Step 6: Construct the prompts, indicator items and corresponding bidding document fragments into a template, and input the template into the trained large language model to guide the trained large language model to infer the answer.

2. The method according to claim 1, characterized in that Pre-processing of bidding documents, including: Step 11: Obtain bidding and tender documents from various public platforms; Step 12: Use the hash value of the bidding documents to remove duplicates by calculating the SHA-256 hash value of each document and deleting the documents corresponding to the duplicate hash values; Step 13: Convert the word document into a string format. The conversion method is as follows: (S,T)=R(D) Where D represents the input word document, S represents the output document content in string form, T represents the extracted table data list, and R represents the function that converts the Word document into a string. Step 14, using the table mark | as a separator to extract T from the table; converting the content of the table T into an HTML format string; The | separator is used to identify table rows in a document; Step 15, get all the table contents by identifying the table tag; Step 16, for the converted text string, i.e., the txt text, find the chapter title by identifying the field such as the chapter title, and segment the txt text according to the chapter title to form different paragraphs; wherein the chapter title includes "Chapter XX"; The bold text in the word document is wrapped with ** when it is converted into a text string.

3. The method according to claim 1, characterized in that In step 2, the pre-training dataset is established by the following method: Step 21, removing duplicate information in the pre-processed document, i.e., the txt text, where the duplicate information includes conventional descriptions in the bidding documents, including bidder instructions, government procurement policies, qualification review and bid evaluation; Step 22, removing useless information in the data including personal information and corporate information; Step 23, use the txt text after data cleaning as a pre-training data set for subsequent processing.

4. The method according to claim 1, characterized in that Step 3: Process the txt text after data cleaning and split the txt text into multiple indicator items, including: Step 31, set a json format template; use the chapter title **Chapter** as the key value, and the content of each chapter as the value; Step 32, determine the highest level sequence number, traverse each chapter content, find the highest level sequence number, including "one", "two", "three"; the highest level sequence number is used as the basis for content division; according to the highest level sequence number found, divide each chapter content into multiple parts; Step 33, processing the content of each part: for each part of the content, continue to search for the highest level sequence number within each part; if a part has a secondary sequence number, further divide it according to the secondary sequence number, and for each secondary sequence number, divide the content again; Step 34, content separation, in the content of each sequence number, use ":" and "." as the basis for separating indicator items; further subdivide the content into multiple indicator items. Step 35, start parsing from the outermost chapter title, extract only the bottom-level key and value pairs and concatenate them as each index item; save the extracted index items into the index txt text; Step 36, split the table content into indicator items, and combine each row of the table content with the title to form a clear indicator item.

5. The method according to claim 1, characterized in that Step 4: Fine-tune the dataset by the following method: Step 41, set prompts, use the instruction following capability of the large model to generate response indicators in the bidding document for each indicator in the indicator txt text; each indicator includes data size comparison, data range, special symbol recognition and relevant industry standards; wherein the prompt is: please generate response indicators that meet or do not meet the requirements according to the following indicator requirements; in each indicator, the response indicator meets the requirements of the indicator item; Data size comparison, including comparison of different units and application of "minimum requirements" and "maximum limits"; The scope of the data, including a clear definition of the "scope"; Special symbol recognition, emphasizing the impact of "≥", "+", and "-" on the meaning of indicator items; Relevant industry standards, including different quality management system requirements; Step 42, set the prompt, use the command following ability of the large model, answer according to the indicator item and the corresponding response indicator, and judge whether the response indicator meets the requirements of the indicator item; the prompt is as follows; Prompt: Please judge whether the response indicator meets the requirements of the indicator item. If it does, answer "yes", otherwise answer "no". Step 43: Integrate the prompts, indicator items and corresponding response indicators in step 4.2 and save them as a fine-tuning dataset for model fine-tuning: {P, I1, I2, A} Among them, P is the prompt, I2 is the indicator item of the bidding document, I1 is the response indicator generated by I2 through the instruction following ability of the large language model, and A is the answer generated by the large model.

6. The method according to claim 1, characterized in that Step 5: Pre-train and fine-tune based on the base model, including: Step 51, using an open source large base model as a base model; Step 52, incremental pre-training of the base large model; the warmup phase is used to use a lower learning rate of 1e-4 at the beginning of training, and gradually increase it to a preset learning rate to stabilize the training process; where η t The learning rate at step t; η max is the preset maximum learning rate, which is the learning rate that the model should reach after the warmup phase ends; t Warmup is the total number of steps in the warmup phase; Step 53, in the incremental pre-training process, the pre-training data set is mixed with a small amount of general knowledge domain data for training to keep the model's memory of old knowledge; in the mixed data set, the general knowledge domain data accounts for 70%; the training is carried out in stages, and new data is gradually added; Step 54, fine-tune LoRA on the pre-trained model by introducing low-rank matrix decomposition into the fully connected layer weight matrix of the model; W∈R d×k , A∈R d×r and B∈R r×k , where r<<min(d,k) Where W ≈ AB, A and B are low-rank matrices obtained through training, and r is the rank of matrices A and B.

7. The method according to claim 1, characterized in that Step 6: Construct the prompts, indicators and corresponding bidding document fragments into a template, and input the template into the trained large language model to guide the trained large language model to reason and answer, including: Step 61, using the txt text in the bidding document after data cleaning in step 2; then segmenting the txt file into multiple segments in a manner that each paragraph has 300 characters and each segmentation repeats 100 characters; P i =T[(i-1)·(N-R)+1:(i-1)·(N-R)+N] Where P i represents the i-th segment, T represents the txt text after data cleaning, N represents the fixed character length of the paragraph, which is 300, and R represents the repeated characters each time the segmentation is performed, which is 100; Step 62, for each index {I1, I2, I3...I n } and multiple segments of the bidding document {P1, P2, P3…P n } Perform similarity matching; Step 63: Set the index {I1, I2, I3…I n }Text and fragments {P1,P2,P3…P n } is converted into a vector. The vectorized indices are {V(I1), V(I2), V(I3)…V(I n )} and {V(P1),V(P2),V(P3)…V(P n )}; Step 64, for each indicator vector V(I i ) and each bidding segment vector V(P j ), calculate the cosine similarity between them: Among them, V(I i ).V(P j ) represents the dot product of vectors, ∥V(I i )∥ and∥V(P j )∥respectively represent the modulus length of the vector; Step 65, for each indicator I i , find the bidding segment P that maximizes the cosine similarity j ; Step 66, each indicator I i , corresponding to the bidding document fragment P j And the prompt are spliced ​​into a template; the specific template is as follows: Prompt: Determine whether the following indicators are consistent with the content in the clip. Please answer yes or no. Index I i :Specific indicator content; Fragment P j : Specific fragment content; Step 67, input the template content to the trained model, and generate answers using the reasoning and generation capabilities of the large language model.

Images