A method and system for positioning the contents of review points in a bid document

Abstract

The application discloses a kind of bidding file in review key content positioning method and system, including to the tender document is parsed, identify the review element information in text, extract review element and review standard;The bidding document is parsed, and hierarchical structure information is extracted and based on serial number matching, font attribute and independence rule are corrected to title, obtain target title and corresponding text content;First similarity of review element and target title is calculated respectively, second similarity of review standard and the text content under title is calculated, vectorization coding and cosine similarity calculation are carried out;According to the review type dynamic adjustment title weight and content weight, weighted fusion is carried out to obtain comprehensive matching score;Score is normalized, and when business and technical content are mixed, conflict resolution is realized by local context analysis and integrity principle, and finally determine the positioning information of review element in bidding document, generate the positioning report containing page number, paragraph position and matching score.

Description

Technical Field

[0001] This invention belongs to the technical field of bid document content location, specifically relating to a method and system for locating key review points in bid documents. Background Technology

[0002] During the bidding process, evaluation experts or bidders need to compare and analyze a large number of bidding documents and bids. The bidding documents contain detailed evaluation elements, including commercial, technical, pricing, qualification, and corresponding evaluation criteria; when preparing their bids, bidders need to ensure that their responses accurately cover these evaluation points.

[0003] In existing technologies, identifying key points for reviewing tender documents mainly relies on manual reading or simple keyword matching. However, the following problems exist in practical applications: 1. Complex document structure: Tender documents are mostly in PDF format, containing complex layouts, tables, images and multi-level headings. Traditional text extraction tools have difficulty accurately identifying the hierarchical structure of the document, resulting in content positioning errors.

[0004] 2. Insufficient semantic understanding: Simple keyword matching cannot understand the semantic relationship between the review criteria and the content of the bid response, which is prone to missed detection or false detection.

[0005] 3. Content overlap and confusion: In actual bidding documents, business content and technical content are sometimes mixed in the same chapter or paragraph, and existing methods are difficult to classify and locate the mixed content in a refined manner.

[0006] Therefore, it is necessary to design a method and system for locating key review points in tender documents to solve the above problems. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for locating key review points in tender documents, which aims to overcome the problems of complex document structure, insufficient semantic understanding, and overlapping and confusing content in existing technologies. It has the characteristics of accurately parsing document structure, deeply understanding semantic content, and effectively handling review points with overlapping and confusing content.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for locating key evaluation points in a tender document includes the following steps: The tender documents are parsed, and a pre-trained language model is used to identify the review elements in the text. The review elements and review criteria are extracted according to the preset review element naming rules. The tender documents are parsed, the hierarchical structure information of the tender documents is extracted using a document structure recognition model, and the extracted titles are corrected based on the title rules to obtain the target titles and the corresponding text content of the target titles. Calculate the first similarity between the review elements and the target title; including: The Sentence-BERT model was used to convert the review elements and the target title into 768-dimensional semantic vectors, respectively, to obtain the review element vector. and target title vector ; The first similarity is calculated using the cosine similarity formula. : ; The first similarity score reflects whether the chapter titles cover the themes of the review elements. Calculate the second similarity between the review criteria and the main text content under the target title; The first and second similarities are weighted according to the review type to obtain the comprehensive matching score; The comprehensive matching score is normalized and integrated to determine the positioning information of the evaluation elements in the bid documents.

[0009] Preferably, the bidding documents are parsed and the evaluation elements and evaluation criteria are extracted, including: The system receives uploaded tender documents and uses a document parsing library to parse them. For scanned PDF files, it uses an OCR engine to convert them into text; for original electronic PDF files, it directly extracts the text stream. Remove header, footer, and page number noise data, divide the text into blocks according to chapters and paragraphs, and complete data preprocessing; The segmented text is input into a pre-trained BERT model, which is fine-tuned using a historical bidding dataset and outputs a category label for each text segment. The label set is {business, technology, quotation, qualification, form, response}. Extract the review elements from the model output as the review dimension names, and extract the review criteria as specific scoring requirements or qualification requirements.

[0010] Preferably, the tender document is parsed to obtain the target title and corresponding body content, including: The PP-Structure module in PaddleOCR is called to perform layout recognition on the tender document. The PP-Structure module outputs the region types in the document, which include titles, body text, tables and images. Based on the vertical coordinates and indentation of the text blocks on the page, a preliminary document tree is constructed; A rule engine is introduced to perform secondary correction on the initially identified titles. The rule engine includes the following rules: Rule 1, Sequence matching, checks whether the text matches a preset Chinese or numeric regular expression. If it matches, the confidence level increases by 0.3. Rule 2, font attribute: extract the font size of a line of text. If the font size of that line is greater than 1.2 times the average font size of the three adjacent lines of text below it, the confidence level is increased by 0.3. Rule 3, Independence: If the line spacing between lines in the text is significantly greater than the line spacing, the confidence level increases by 0.2. If the overall confidence level is lower than the preset threshold, the corresponding recognition result will be downgraded from title to body text; otherwise, it will be retained as the target title, and all text and tables under the target title up to the next title will be determined as the corresponding body text content.

[0011] Preferably, calculating the second similarity includes: The BERT model was used to extract key information from the entire description of the review criteria to obtain key information about the review criteria. The T5-PEGASUS-PGN model is used to generate a summary of the text content under the target title to obtain the text summary information. The Sentence-BERT model was used to convert the key information of the review criteria and the information of the main text summary into 768-dimensional semantic vectors, respectively, to obtain the key information vector of the review criteria. and text summary information vector If the length of the main text exceeds the preset value, a segmented averaging or max pooling strategy will be adopted. The second similarity is calculated using the cosine similarity formula. : ; The second similarity score reflects whether the content of the chapter substantially responds to the stated review criteria.

[0012] Preferably, the first similarity and the second similarity are weighted according to the review type, including: The title weight is dynamically adjusted based on the types of evaluation elements identified from the tender documents. and content weight Among them, when the review type is technical or formal, Take 0.3 and Take 0.7 when the review type is business or qualification. Take 0.7 and Take 0.3 when the review type is quotation type. Take 0.2 and Take 0.8 when the review type is unknown. Take 0.5 and Set 0.5 as the default balancing strategy; Calculate the overall matching score: ; in The first similarity score, This represents the second similarity.

[0013] Preferably, the comprehensive matching score is subjected to normalization analysis and fusion processing, including: The overall matching scores of all candidate chapters are normalized using Min-Max, mapped to a score range of 0 to 100. The normalized scores are: ,in The highest overall matching score among all candidate chapters. The lowest overall matching score among all candidate chapters; To resolve conflicts and integrate business and technical content in tender documents.

[0014] Preferably, the fusion and conflict resolution process includes: If the overall matching score of the same paragraph for technical elements and the overall matching score for business elements both exceed the preset threshold, then the paragraph will be associated with both technical review points and business review points to achieve multi-tag association. If the system requires a single attribution, then the second similarity is based on the technical elements. And the second similarity for business elements If all conditions exceed a preset threshold, a local contextual analysis is performed: if the paragraph is closely surrounded by other clearly defined business element paragraphs and is located in the core area of ​​a business chapter, then the paragraph is determined to be a business element based on the principle of integrity, in order to maintain the consistency of the chapter's logic; otherwise, if Greater than If so, it should be classified as a technical response. Greater than This is classified as a business response; Generate a location report, which includes: review point name, matched bid document page number, paragraph start position, match score, and keywords used for identification.

[0015] Preferably, a system for locating key review points in a bid document is provided for executing the method for locating key review points in a bid document. The system includes: The tender document parsing module is used to parse tender documents, use a pre-trained language model to identify review element information in the text, and extract review elements and review criteria according to preset review element naming rules. The tender document parsing module is used to parse tender documents, extract hierarchical structure information of tender documents using a document structure recognition model, and correct the extracted titles based on title rules to obtain the target title and the corresponding text content of the target title; The similarity calculation module is used to calculate the first similarity between the review elements and the target title, and the second similarity between the review criteria and the main text content under the target title. The weighted fusion module is used to calculate the first similarity and the second similarity based on the review type to obtain a comprehensive matching score; The positioning output module is used to perform normalization analysis and fusion processing on the comprehensive matching score to determine the positioning information of the review elements in the tender documents.

[0016] Preferably, it also includes a computer device, including a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to run the method for locating the key points of the tender document.

[0017] Preferably, it further includes a computer-readable storage medium storing computer instructions for causing a computer to execute the method for locating key review points in the tender document.

[0018] The beneficial effects of the system and method for locating key review points in tender documents provided by this invention are as follows: 1. Accurate structure recognition: This invention combines PP-Structure with title rule correction, effectively solving the problem of misidentification of titles in PDF documents and establishing an accurate document hierarchy index.

[0019] 2. In-depth semantic matching: This invention adopts a dual-path similarity calculation of element-title and standard-text, which considers both the macro-matching of chapter affiliation and the micro-matching of content response, thereby improving the positioning accuracy.

[0020] 3. Dynamic weight adjustment: This invention adjusts the weight of the title and content for different review types, whether technical or commercial, to adapt to the different writing habits of different document categories.

[0021] 4. Integration of Mixed Content: This invention effectively solves the positioning conflict problem when business and technical content are mixed through normalization and fusion analysis, thereby improving the robustness of the system.

[0022] 5. Good scalability: As technology develops and new demands emerge, this method can be easily expanded and upgraded in terms of functionality, such as integrating a natural language understanding module to improve semantic understanding. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a flowchart illustrating the extraction process of review elements and criteria for tender documents in an embodiment of the present invention. Figure 3 This is a flowchart illustrating the process of recognizing the title and body of a tender document in an embodiment of the present invention. Figure 4 This is a flowchart of a method and system for locating key review points in a tender document, as described in an embodiment of the present invention. Figure 5 This is a schematic diagram of the structure of a computer device in an embodiment of the present invention. Detailed Implementation

[0024] Example 1: like Figure 1 As shown, a method for locating key evaluation points in a tender document includes the following steps: The tender documents are parsed, and a pre-trained language model is used to identify the review elements in the text. The review elements and review criteria are extracted according to the preset review element naming rules. The tender documents are parsed, the hierarchical structure information of the tender documents is extracted using a document structure recognition model, and the extracted titles are corrected based on the title rules to obtain the target titles and the corresponding text content of the target titles. Calculate the first similarity between the review elements and the target title; including: The Sentence-BERT model was used to convert the review elements and the target title into 768-dimensional semantic vectors, respectively, to obtain the review element vector. and target title vector ; The first similarity is calculated using the cosine similarity formula. : ; The first similarity score reflects whether the chapter titles cover the themes of the review elements. Calculate the second similarity between the review criteria and the main text content under the target title; The first and second similarities are weighted according to the review type to obtain the comprehensive matching score; The comprehensive matching score is normalized and integrated to determine the positioning information of the evaluation elements in the bid documents.

[0025] like Figure 2 As shown, preferably, the bidding documents are parsed and the evaluation elements and evaluation criteria are extracted, including: The system receives uploaded tender documents and uses a document parsing library to parse them. For scanned PDF files, it uses an OCR engine to convert them into text; for original electronic PDF files, it directly extracts the text stream. Remove header, footer, and page number noise data, divide the text into blocks according to chapters and paragraphs, and complete data preprocessing; The segmented text is input into a pre-trained BERT model, which is fine-tuned using a historical bidding dataset and outputs a category label for each text segment. The label set is {business, technology, quotation, qualification, form, response}. Extract the review elements from the model output as the review dimension names, and extract the review criteria as specific scoring requirements or qualification requirements.

[0026] like Figure 3 As shown, preferably, the tender document is parsed to obtain the target title and corresponding body content, including: The PP-Structure module in PaddleOCR is called to perform layout recognition on the tender document. The PP-Structure module outputs the region types in the document, which include titles, body text, tables and images. Based on the vertical coordinates and indentation of the text blocks on the page, a preliminary document tree is constructed; A rule engine is introduced to perform secondary correction on the initially identified titles. The rule engine includes the following rules: Rule 1, Sequence matching, checks whether the text matches a preset Chinese or numeric regular expression. If it matches, the confidence level increases by 0.3. Rule 2, font attribute: extract the font size of a line of text. If the font size of that line is greater than 1.2 times the average font size of the three adjacent lines of text below it, the confidence level is increased by 0.3. Rule 3, Independence: If the line spacing between lines in the text is significantly greater than the line spacing, the confidence level increases by 0.2. If the overall confidence level is lower than the preset threshold, the corresponding recognition result will be downgraded from title to body text; otherwise, it will be retained as the target title, and all text and tables under the target title up to the next title will be determined as the corresponding body text content.

[0027] Preferably, calculating the second similarity includes: The BERT model was used to extract key information from the entire description of the review criteria to obtain key information about the review criteria. The T5-PEGASUS-PGN model is used to generate a summary of the text content under the target title to obtain the text summary information. The Sentence-BERT model was used to convert the key information of the review criteria and the information of the main text summary into 768-dimensional semantic vectors, respectively, to obtain the key information vector of the review criteria. and text summary information vector If the length of the main text exceeds the preset value, a segmented averaging or max pooling strategy will be adopted. The second similarity is calculated using the cosine similarity formula. : ; The second similarity score reflects whether the content of the chapter substantially responds to the stated review criteria.

[0028] Preferably, the first similarity and the second similarity are weighted according to the review type, including: The title weight is dynamically adjusted based on the types of evaluation elements identified from the tender documents. and content weight Among them, when the review type is technical or formal, Take 0.3 and Take 0.7 when the review type is business or qualification. Take 0.7 and Take 0.3 when the review type is quotation type. Take 0.2 and Take 0.8 when the review type is unknown. Take 0.5 and Set 0.5 as the default balancing strategy; Calculate the overall matching score: ; in The first similarity score, This represents the second similarity.

[0029] Preferably, the comprehensive matching score is subjected to normalization analysis and fusion processing, including: The overall matching scores of all candidate chapters are normalized using Min-Max, mapped to a score range of 0 to 100. The normalized scores are: ,in The highest overall matching score among all candidate chapters. The lowest overall matching score among all candidate chapters; To resolve conflicts and integrate business and technical content in tender documents.

[0030] Preferably, the fusion and conflict resolution process includes: If the overall matching score of the same paragraph for technical elements and the overall matching score for business elements both exceed the preset threshold, then the paragraph will be associated with both technical review points and business review points to achieve multi-tag association. If the system requires a single attribution, then the second similarity is based on the technical elements. And the second similarity for business elements If all conditions exceed a preset threshold, a local contextual analysis is performed: if the paragraph is closely surrounded by other clearly defined business element paragraphs and is located in the core area of ​​a business chapter, then the paragraph is determined to be a business element based on the principle of integrity, in order to maintain the consistency of the chapter's logic; otherwise, if Greater than If so, it should be classified as a technical response. Greater than This is classified as a business response; Generate a location report, which includes: review point name, matched bid document page number, paragraph start position, match score, and keywords used for identification.

[0031] Preferably, a system for locating key review points in a bid document is provided for executing the method for locating key review points in a bid document. The system includes: The tender document parsing module is used to parse tender documents, use a pre-trained language model to identify review element information in the text, and extract review elements and review criteria according to preset review element naming rules. The tender document parsing module is used to parse tender documents, extract hierarchical structure information of tender documents using a document structure recognition model, and correct the extracted titles based on title rules to obtain the target title and the corresponding text content of the target title; The similarity calculation module is used to calculate the first similarity between the review elements and the target title, and the second similarity between the review criteria and the main text content under the target title. The weighted fusion module is used to calculate the first similarity and the second similarity based on the review type to obtain a comprehensive matching score; The positioning output module is used to perform normalization analysis and fusion processing on the comprehensive matching score to determine the positioning information of the review elements in the tender documents.

[0032] Preferably, it also includes a computer device, including a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to run the method for locating the key points of the tender document.

[0033] Preferably, it further includes a computer-readable storage medium storing computer instructions for causing a computer to execute the method for locating key review points in the tender document.

[0034] Example 2: like Figure 4 As shown, this embodiment provides an example of applying the method in a specific way, including the following steps: Step S1: Tender document analysis and evaluation element extraction: For document recognition and text acquisition, the system receives tender documents uploaded by users, typically in PDF or DOCX format. It then uses a common document parsing library to parse the files. For scanned PDFs, it first uses an OCR engine to convert them to text; for original electronic PDFs, it directly extracts the text stream. Next, it preprocesses the text by removing noise such as headers, footers, and page numbers, and dividing the text into blocks according to chapters and paragraphs.

[0035] Next, based on BERT's feature recognition, the segmented text is input into a pre-trained BERT model. This model has been fine-tuned using a historical bidding dataset and is capable of recognizing proper nouns in the bidding field. Naming rules are then applied for classification, and the model outputs a category label for each text segment. The label set is defined as C = {Business, Technical, Quotation, Qualification, Form, Response}.

[0036] Extraction of elements and standards: Element: refers to the name of the dimension to be reviewed, such as "Project Manager Qualification", "After-sales Service Commitment", and "Completeness of Technical Solution".

[0037] The assessment standard refers to specific scoring or qualification requirements, such as: having a PMP certificate earns 2 points, providing 7... 24-hour service and solutions must include a topology diagram.

[0038] Input text snippet: Project Manager. The project manager must have a Level 1 Construction Engineer Certificate, which earns 5 points; and a Safety Production Assessment Certificate (B Certificate), which earns 2 points.

[0039] BERT feature analysis model output: Category: Qualification / Business; Review Elements: Project Manager Qualification; Review Criteria: Possess a Level 1 Construction Engineer Certificate; Possess a Safety Production Assessment Certificate (B Certificate); Structured Storage: The system stores this information in the database with ID Req_001 and type Type=Qualification.

[0040] Step S2: Tender document parsing and title / text recognition: Based on PP-Structure structure recognition, since tender documents are mostly complex PDFs, the system calls the PP-Structure module in PaddleOCR. This module can not only recognize text and tables, but also page layout (PP-DocLayout), including 20 common categories: document title, paragraph title, text, page number, abstract, table of contents, references, footnotes, headers, footers, algorithms, formulas, formula numbers, images, tables, figure and table titles (figure titles, table titles, and figure / table titles), stamps, figures, sidebar text, and reference content.

[0041] Next, the initial layout of the tender document is extracted: the model outputs the region types in the document, including Title, Text, Table, and Figure; then, the hierarchy is constructed according to the position of the text blocks and the layout information: based on the vertical coordinates (Y-axis) of the regions on the page and the indentation relationship, the document tree is initially constructed.

[0042] Next, title rule correction is performed. PP-Structure might misinterpret bold, large font in the body text as a title. This step introduces a rule engine for secondary validation: Rule 1: Number Matching: Check if the text matches the regular expression. If it matches, increase the confidence level by 0.3.

[0043] Rule 2: Font attribute: Extract the font size of the text. If the font size of a line of text is greater than 1.2 times the average font size of the three adjacent lines below it, increase the confidence level by 0.3.

[0044] Rule 3 Independence: If the line spacing of a certain line is significantly greater than the line spacing, increase the confidence level by 0.2.

[0045] Judgment: If the overall confidence level is lower than the threshold of 0.6, the title will be downgraded to the main text.

[0046] Example: PP-Structure initial recognition result: A bolded text "Introduction of Key Personnel" was identified as the title.

[0047] Rule correction: Serial number matching: No serial number (confidence score incremented by 0).

[0048] Font attributes: 14pt font size, 12pt body text size below, scale 1.16 (<1.2, confidence level +0).

[0049] Independence: The line is located in the middle of the paragraph, with normal vertical spacing (confidence +0).

[0050] Correction result: The overall confidence level is low, and it is determined that "Introduction of Key Personnel" is a subheading or emphasis sentence in the main text, rather than a chapter title. The true chapter title should be "3.2 Project Team Configuration" above.

[0051] Final output: The target title is "3.2 Project Team Configuration", and the corresponding body content is all text and tables under this title up to the next title.

[0052] Step S3: Calculate the similarity between the review elements and the title: Vectorization encoding uses the Sentence-BERT model to convert text into 768-dimensional semantic vectors.

[0053] Velem: A vector of review elements; Vtitle: A vector of the target title of the tender document.

[0054] Next, similarity is calculated using the cosine similarity formula to determine the first similarity score, which reflects whether the chapter title covers the theme of the review elements.

[0055] Example: Evaluation element: "After-sales service system"; Tender document title: Chapter 5 After-sales and Technical Support; Calculation result: Stitle("After-sales service system" vs "Chapter 5 After-sales and Technical Support") = 0.88.

[0056] Step S4: Calculate the similarity between the review criteria and the key information in the body text under the title: Key information extraction for review criteria: The BERT model is used to extract key information from all the descriptive content in the review criteria, thus obtaining the key information of the review criteria.

[0057] To obtain the abstract information of the text under the title, the T5-PEGASUS-PGN model is used to generate the abstract information of the text content under the target title.

[0058] Vectorization encoding uses the Sentence-BERT model (paraphrase-multilingual-MiniLM-L12-v2) to convert the text into a 768-dimensional semantic vector.

[0059] Vstd: The vector of key information for the review criteria; Vcontent: The vector of the summary information of the main text under the target title. If the content is too long, a segmented averaging or max pooling strategy is adopted.

[0060] Next, similarity is calculated using the cosine similarity formula to determine the second similarity, which reflects whether the chapter content substantially responds to the review criteria.

[0061] Example: Evaluation criteria: "Provide proof of localized service outlets" -> Key information extracted: "Localized service outlets"; Tender document body: "...Our company has an office in this area, located at..., which can provide rapid response..." -> Summary information of the body content: "Has an office in this area"; Calculation result: Scontent(“Local service outlets” vs “Local offices”) = 0.92.

[0062] Step S5: Weighted calculation based on review type: The system dynamically adjusts the title weight based on the review element types identified in step S1. and content weight .

[0063] The configuration strategy table is shown in Table 1 below: Table 1: Configuration Strategy Table;

[0064] Overall score calculation: Score = α × Stitle + β × Scontent; Example: Scenario A, Technical Category: Element "Technical Solution"; Stitle=0.6, Scontent=0.9; Score=0.3×0.6+0.7×0.9=0.18+0.63=0.81.

[0065] Scenario B Qualification Category: Element "Business License", Stitle=0.9, Scontent=0.4, Score=0.7×0.9+0.3×0.4=0.63+0.12=0.75.

[0066] Without dynamic weights, the average score for scenario A might be low, leading to missed detections; scenario B might be misjudged due to low content similarity; weighting can more accurately reflect the actual response situation.

[0067] Step S6: Result Normalization Analysis and Fusion: Normalization processing: Min-Max normalization is performed on the scores of all candidate chapters, mapping them to the [0, 100] interval for easier display. NormScore = (Score...) Scoremin) / (Scoremax Scoremin)×100.

[0068] Integration and conflict resolution are crucial for addressing situations where business and technical content are mixed in tender documents, such as the "Project Overview" section which discusses both the technical approach and the company's business capabilities.

[0069] Multi-tag association: If the same paragraph's score exceeds the threshold (e.g., 75 points) for both "technical elements" and "business elements," then the paragraph is associated with both review points simultaneously. Optimal attribution: If the system requires single attribution, then the range of Scontent for the same type of element is compared.

[0070] Assuming both Scontent_tech and Scontent_biz score high (>75) and above, a local contextual analysis is introduced. If the paragraph is closely surrounded by other clearly defined business element paragraphs, i.e., located in the core area of ​​the business section, then based on the "holistic principle"—where the content of the business and technical sections in the tender document is fixed and the information is holistic—the paragraph is classified as a business element to maintain the consistency of the section's logic. Otherwise, if Scontent_tech > Scontent_biz, it is preferentially classified as a technical response. Conversely, it is classified as a business response, referring to the evaluation point location example shown in Table 2.

[0071] Table 2: Examples of Key Review Points;

[0072] Next, the location output is generated into a location report, which includes: the name of the review point, the page number of the matched tender document, the paragraph start position, the matching score, and highlighted keywords.

[0073] Example: Conflict Scenario: The 10th heading and its body text in the tender document are as follows: "Project Comprehensive Description," "Scaffolding Erection Organization and Process (see table below)." The table outlines the scaffolding erection organization and process. The steps and requirements are as follows: 1. Organization of Scaffolding Construction Personnel: Scaffolding is constructed by professional scaffolders from the main structure construction team. During construction, the workforce is adjusted and supplemented as needed based on the on-site progress to ensure construction progress, quality, and safety. 2. Construction Process: Site leveling and compaction → Foundation bearing capacity test and material preparation → Positioning and setting of continuous pads and bases → Uprights → Longitudinal and transverse ground bracing → Longitudinal and transverse horizontal bracing → Setting of unloading wire ropes → Uprights → Longitudinal and transverse horizontal bracing → External diagonal bracing / scissor bracing → Wall ties → Laying scaffolding. The main construction methods for erecting scaffolding using insert-type disc-lock steel pipe scaffolding are as follows: → Install guardrails → Install safety nets. See the table below for the key points of each step. Key points for erecting insert-type disc-lock steel pipe scaffolding: Foundation spacing and positioning. According to structural requirements, measure the distance between the inner and outer uprights and the wall at the four corners of the building and mark them; use a steel tape measure to straighten and mark the upright positions, and use small bamboo sticks to mark the upright marks; the base plate and base should be accurately placed on the positioning line, and the base plate must be laid flat and not suspended. The base plate size for the scaffolding foundation should be a wooden base plate or channel steel with a length of not less than 2 spans, a thickness of not less than 50mm, and a width of not less than 200mm. The upright foundation should be leveled and compacted according to the plan requirements, and drainage measures should be taken; on soil foundations, the bottom of the uprights must be equipped with base plates and adjustable bases. The uprights should be installed in a way that meets the specifications. Adjustable bases should be installed at the bottom of the uprights, or pads can be placed underneath. When there is a large difference in foundation height, adjustable bases can be used to adjust the uprights so that the connecting plates for the same horizontal bar on adjacent uprights are on the same horizontal plane. For spacing and positioning, measure the distance between the inner and outer uprights and the wall at the four corners of the building according to structural requirements and mark them. Use a steel tape measure to mark the upright positions and mark them with small bamboo strips. Pads and bases should be accurately placed on the positioning lines, and pads must be laid flat and not suspended. The pads for the scaffolding foundation should be made of wood or channel steel with a length of not less than two spans, a thickness of not less than 50mm, and a width of not less than 200mm. The upright foundation should be leveled and compacted according to the plan requirements, and drainage measures should be taken. Water-related measures: The bottom of the uprights on the soil foundation must be equipped with pads and adjustable bases, meeting the specifications. Adjustable bases should be installed at the bottom of the uprights, or pads can be placed underneath. When there is a large difference in foundation elevation, adjustable bases can be used to adjust the uprights so that the connecting plates for the same horizontal bar on adjacent uprights are on the same horizontal plane. Spacing and positioning: According to structural requirements, measure the distance between the inner and outer uprights and the wall at the four corners of the building and mark them; use a steel tape measure to straighten and mark the upright positions, and use small bamboo sticks to mark the upright positions; pads and bases should be accurately placed on the positioning lines, and pads must be laid flat and not suspended. The pads for the scaffolding foundation should be wooden pads or channel steel with a length of not less than two spans, a thickness of not less than 50mm, and a width of not less than 200mm.The foundation for the poles should be leveled and compacted according to the plan requirements, and drainage measures should be taken. A base plate and adjustable base must be installed at the bottom of the poles on the soil foundation, and these must meet the specifications. An adjustable base should be installed at the bottom of the poles, or a base plate can be placed underneath. When there is a large difference in foundation elevation, the adjustable base can be used to adjust the poles so that the connecting plates for installing the same horizontal bar on adjacent poles are on the same horizontal plane. Matching Results: Matching technical element "Construction Process": Score=75; Matching business element "Similar Performance": Score=72; Fusion Processing: Both are above the threshold of 70. In the location report, the system marks this paragraph as a response to both "Construction Process" and "Similar Performance," and highlights it with different colors on the front end, such as blue for technical and red for business, to alert review experts that this paragraph contains mixed information. By comparing the full text, it was found that the preceding paragraph belongs to the business element "Financial Status," and the preceding paragraph of this paragraph belongs to the business element "Price Reasonableness." Based on the "holistic principle," this paragraph belongs to the business element "Similar Performance."

[0074] Example 3: like Figure 5 As shown, embodiments of the present invention also provide a computer device. Figure 5 Taking a single processor 10 as an example, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other using different buses and can be mounted on a common motherboard or otherwise installed as needed. The processor can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices, such as display devices coupled to the interface. In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations, for example, as a server array, a group of blade servers, or a multiprocessor system.

[0075] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0076] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0077] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0078] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0079] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0080] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

Claims

1. A method for locating key evaluation points in a tender document, characterized in that, Includes the following steps: The tender documents are parsed, and a pre-trained language model is used to identify the review elements in the text. The review elements and review criteria are extracted according to the preset review element naming rules. The tender documents are parsed, the hierarchical structure information of the tender documents is extracted using a document structure recognition model, and the extracted titles are corrected based on the title rules to obtain the target titles and the corresponding text content of the target titles. Calculate the first similarity between the review elements and the target title; including: The Sentence-BERT model was used to convert the review elements and the target title into 768-dimensional semantic vectors, respectively, to obtain the review element vector. and target title vector ; The first similarity is calculated using the cosine similarity formula. : ； The first similarity score reflects whether the chapter titles cover the themes of the review elements. Calculate the second similarity between the review criteria and the main text content under the target title; The first and second similarities are weighted according to the review type to obtain the comprehensive matching score; The comprehensive matching score is normalized and integrated to determine the positioning information of the evaluation elements in the bid documents.

2. The method for locating key evaluation points in a tender document according to claim 1, characterized in that, The tender documents were analyzed to extract the evaluation elements and criteria, including: The system receives uploaded tender documents and uses a document parsing library to parse them. For scanned PDF files, it uses an OCR engine to convert them into text; for original electronic PDF files, it directly extracts the text stream. Remove header, footer, and page number noise data, divide the text into blocks according to chapters and paragraphs, and complete data preprocessing; The segmented text is input into a pre-trained BERT model, which is fine-tuned using a historical bidding dataset and outputs a category label for each text segment. The label set is {business, technology, quotation, qualification, form, response}. Extract the review elements from the model output as the review dimension names, and extract the review criteria as specific scoring requirements or qualification requirements.

3. The method for locating key review points in a tender document according to claim 1, characterized in that, The tender documents are parsed to obtain the target titles and corresponding body content, including: The PP-Structure module in PaddleOCR is called to perform layout recognition on the tender document. The PP-Structure module outputs the region types in the document, which include titles, body text, tables and images. Based on the vertical coordinates and indentation of the text blocks on the page, a preliminary document tree is constructed; A rule engine is introduced to perform secondary correction on the initially identified titles. The rule engine includes the following rules: Rule 1, Sequence matching, checks whether the text matches a preset Chinese or numeric regular expression. If it matches, the confidence level increases by 0.

3. Rule 2, font attribute: extract the font size of a line of text. If the font size of that line is greater than 1.2 times the average font size of the three adjacent lines of text below it, the confidence level is increased by 0.

3. Rule 3, Independence: If the line spacing between lines in the text is significantly greater than the line spacing, the confidence level increases by 0.

2. If the overall confidence level is lower than the preset threshold, the corresponding recognition result will be downgraded from title to body text; otherwise, it will be retained as the target title, and all text and tables under the target title up to the next title will be determined as the corresponding body text content.

4. The method for locating key review points in a tender document according to claim 1, characterized in that, Calculating the second similarity includes: The BERT model was used to extract key information from the entire description of the review criteria to obtain key information about the review criteria. The T5-PEGASUS-PGN model is used to generate a summary of the text content under the target title to obtain the text summary information. The Sentence-BERT model was used to convert the key information of the review criteria and the information of the main text summary into 768-dimensional semantic vectors, respectively, to obtain the key information vector of the review criteria. and text summary information vector If the length of the main text exceeds the preset value, a segmented averaging or max pooling strategy will be adopted. The second similarity is calculated using the cosine similarity formula. : ； The second similarity score reflects whether the content of the chapter substantially responds to the stated review criteria.

5. The method for locating key review points in a tender document according to claim 1, characterized in that, The first similarity and the second similarity are weighted according to the review type, including: The title weight is dynamically adjusted based on the types of evaluation elements identified from the tender documents. and content weight Among them, when the review type is technical or formal, Take 0.3 and Take 0.7 when the review type is business or qualification. Take 0.7 and Take 0.3 when the review type is quotation type. Take 0.2 and Take 0.8 when the review type is unknown. Take 0.5 and Set 0.5 as the default balancing strategy; Calculate the overall matching score: ； in The first similarity score, This represents the second similarity.

6. The method for locating key review points in a tender document according to claim 1, characterized in that, The comprehensive matching score is subjected to normalization analysis and fusion processing, including: The overall matching scores of all candidate chapters are normalized using Min-Max, mapped to a score range of 0 to 100. The normalized scores are: ,in The highest overall matching score among all candidate chapters. The lowest overall matching score among all candidate chapters; To resolve conflicts and integrate business and technical content in tender documents.

7. The method for locating key review points in a tender document according to claim 6, characterized in that, The fusion and conflict resolution process includes: If the overall matching score of the same paragraph for technical elements and the overall matching score for business elements both exceed the preset threshold, then the paragraph will be associated with both technical review points and business review points to achieve multi-tag association. If the system requires a single attribution, then the second similarity is based on the technical elements. And the second similarity for business elements If all conditions exceed a preset threshold, a local contextual analysis is performed: if the paragraph is closely surrounded by other clearly defined business element paragraphs and is located in the core area of ​​a business chapter, then the paragraph is determined to be a business element based on the principle of integrity, in order to maintain the consistency of the chapter's logic; otherwise, if Greater than If so, it should be classified as a technical response. Greater than This is classified as a business response; Generate a location report, which includes: review point name, matched bid document page number, paragraph start position, match score, and keywords used for identification.

8. A system for locating key review points in a tender document, used to execute the method for locating key review points in a tender document as described in any one of claims 1-8, characterized in that, The system includes: The tender document parsing module is used to parse tender documents, use a pre-trained language model to identify review element information in the text, and extract review elements and review criteria according to preset review element naming rules. The tender document parsing module is used to parse tender documents, extract hierarchical structure information of tender documents using a document structure recognition model, and correct the extracted titles based on title rules to obtain the target title and the corresponding text content of the target title; The similarity calculation module is used to calculate the first similarity between the review elements and the target title, and the second similarity between the review criteria and the main text content under the target title. The weighted fusion module is used to calculate the first similarity and the second similarity based on the review type to obtain a comprehensive matching score; The positioning output module is used to perform normalization analysis and fusion processing on the comprehensive matching score to determine the positioning information of the review elements in the tender documents.

9. A computer device, characterized in that, It includes a memory and a processor, which are interconnected and communicate with each other. The memory stores computer instructions, and the processor executes the computer instructions to run the method for locating the key points of the tender document as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the method for locating key review points in a tender document as described in any one of claims 1-8.

Images