A bidding scheme intelligent review system

Abstract

The application provides a bidding scheme intelligent review system, characterized by comprising a data management module for extracting file content with different document forms in externally input bidding-related files; a knowledge base construction module connected to the data management module, used for processing the file content with different document forms to generate semantic slices corresponding to the different file content, converting all the semantic slices into vector data by a text embedding model and mapping them to the same semantic space, and constructing a knowledge base of the file content based on the vector data; and an intelligent review module connected to the knowledge base construction module, used for reviewing bidding files of a bidding project by using the knowledge base. The knowledge base formed by the vector data embedded by the text embedding model has strong semantic association and logicality, and the bidding-related files provide professional knowledge support for the knowledge base.

Description

Technical Field

[0001] This invention relates to the field of bidding and tendering, and in particular to an intelligent review system for bidding proposals. Background Technology

[0002] Tender documents are a collection of legally binding documents used by the tendering party and the tendering parties in procurement activities. The tender document is the sole guiding document for a procurement activity, outlining the tendering party's requirements, while the tender document is a legally binding document submitted by the tendering party to demonstrate its procurement capabilities in response to the tendering party's requirements. The tender document must meet drafting specifications and the tendering party's requirements, and must provide the tendering party's proposed solution and price quote. In a procurement activity, multiple tendering parties submit their own tender documents to compete. The evaluation committee reviews the legality and competitiveness of the tender documents and selects one tender document as the winning bid.

[0003] Because bid documents have clearly defined writing guidelines, and for large-scale bidding and procurement activities, the content of bid proposals is often quite complex, bidders need to review their own bid documents before participating in the evaluation process. This review can be done by either entrusting a third party or the bidding party to review them, ensuring the legality and compliance of the bid documents and minimizing any non-compliance with the bidding requirements. Traditionally, this review is conducted manually by experienced professionals in the relevant technical fields. However, manual review is inefficient with a large volume of repetitive bid documents and may result in omissions. Furthermore, due to the wide range of industries involved in bidding projects, it is impossible to establish a unified expert database for the professional review of a large number of bid documents. Summary of the Invention

[0004] To address the related problems existing in the prior art, the present invention provides an intelligent bidding scheme review system, comprising: The data management module is used to extract file content with different document formats from externally input bidding-related documents; The knowledge base construction module, connected to the data management module, is used to process the file content in different document formats to generate semantic slices corresponding to different file content. It converts all semantic slices into vector data and maps them to the same semantic space through a text embedding model, and constructs a knowledge base for the file content based on the vector data. The intelligent review module, connected to the knowledge base construction module, is used to review the bid documents of the project to be bid using the knowledge base.

[0005] Preferably, the bidding-related documents include the tender documents corresponding to the project to be bid on, historical bidding reference templates, project information, and industry standards.

[0006] Preferably, when generating the semantic slice, the data management module obtains the attribute features of the file content to form corresponding metadata; The attribute features include the source of the file content, chapter level, content type, and timestamp.

[0007] Preferably, the document content is in the form of text, images, or tables; The knowledge base construction module includes: The first semantic processing unit is used to directly extract text-based data from the file content and generate the semantic slice associated with the text-based data. The second semantic processing unit is used to analyze the image-format data in the file content and generate the semantic slice associated with the image-format data; The third semantic processing unit is used to analyze the tabular data in the file content and generate the semantic slice associated with the tabular data.

[0008] Preferably, the constructed knowledge base includes: The vector layer is used to store the vector data and the metadata associated with the file content to form a retrieval index for the knowledge base; The document layer is used to store the structured content in the semantic slices; The relationship layer establishes and saves the connection relationships between the various contents of the file based on the chapter hierarchy of the file content; The semantic slice sequence number is used to associate the vector data in the vector layer, the structured content in the document layer, and the connection relationship in the relation layer.

[0009] Preferably, the intelligent review module includes: The document processing unit is used to unify the chapter directory of the tender documents of the project to be tendered into the text format of the corresponding chapter level of the tender documents in the knowledge base; The chapter review unit, connected to the document processing unit, is used to match and review the tender documents by chapter according to the document content of the tender documents in the knowledge base, and output the chapter review results.

[0010] Preferably, the relationship layer also has a pre-set scoring rule library for the project to be tendered. The scoring rule library includes the scoring type of each tender item in the tender document of the project to be tendered, and the scoring standard corresponding to the tender item whose scoring type is objective scoring item. The intelligent review module includes: The scoring type determination unit is used to determine the scoring type of each bid item in the bid document of the project to be bid, and output the type determination result; The first scoring unit, connected to the scoring type judgment unit, is used to provide the bidding item with the scoring type of subjective scoring item to the user for subjective scoring based on the type judgment result, and to obtain the corresponding subjective scoring result input by the user. The second scoring unit is connected to the scoring type judgment unit and is used to automatically score the bid item whose scoring type is objective scoring item according to the type judgment result and the corresponding scoring standard, and output the corresponding objective scoring result. The intelligent review module then outputs a review result corresponding to the tender document based on the subjective and objective scoring results.

[0011] Preferably, it also includes a tender document intelligent generation module, which is connected to the knowledge base construction module and is used to automatically generate the tender document for the project to be tendered based on the knowledge base; The intelligent tender document generation module includes: The requirement elicitation unit is used to obtain the tender document generation requirements input by the user; The chapter generation unit, connected to the requirement acquisition unit, is used to construct a chapter directory of the tender document corresponding to the tender project to be tendered in the knowledge base based on the tender document generation requirements and the knowledge base. The content retrieval unit, connected to the chapter generation unit, is used to retrieve the structured content of the corresponding semantic slice from the knowledge base for each chapter according to the chapter directory. The draft generation unit is connected to the content retrieval unit and the chapter generation unit respectively, and generates a context-coherent draft document based on all the recalled structured content and the chapter directory. The adjustment unit, connected to the draft generation unit, is used to provide the draft document to the user for modification and adjustment, and to obtain the modified draft document as the final completed tender document.

[0012] Preferably, the content recall unit includes: The retrieval subunit is used to retrieve relevant vector data from the vector layer in the knowledge base by using the chapter topic of each chapter as the retrieval condition, and output the retrieval results; The recall subunit, connected to the retrieval subunit, is used to recall the structured content of the semantic slice corresponding to the vector data based on the relevance of the retrieval results.

[0013] The present invention can produce the following beneficial effects: By building a knowledge base that relates to cross-industry and cross-modal document content related to bidding projects, it is possible to conduct batch and cross-industry intelligent reviews of bid documents. Compared with traditional manual reviews, this breaks through the technical and professional limitations of the review process, improves review efficiency, and ensures the accuracy of the review while avoiding omissions. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the intelligent bidding scheme review system of the present invention; Figure 2 This is a schematic diagram of the knowledge base construction module of the present invention; Figure 3 This is a schematic diagram of the intelligent review module of the present invention; Figure 4 This is a schematic diagram of the intelligent tender document generation module of the present invention; Figure 5 This is a schematic diagram of the content recall unit of the present invention; In the attached image: 1. Data Management Module; 2. Knowledge Base Construction Module; 21. First Semantic Processing Unit; 22. Second Semantic Processing Unit; 23. Third Semantic Processing Unit; 3. Intelligent Review Module; 31. Document Processing Unit; 32. Chapter Review Unit; 33. Scoring Type Judgment Unit; 34. First Scoring Unit; 35. Second Scoring Unit; 4. Intelligent Tender Generation Module; 41. Requirements Acquisition Unit; 42. Chapter Generation Unit; 43. Content Retrieval Unit; 431. Retrieval Subunit; 432. Retrieval Subunit; 44. Draft Generation Unit; 45. Adjustment Unit Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0017] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0018] This invention addresses the problems mentioned in the background section by providing an intelligent bidding scheme review system, such as... Figure 1 As shown, it includes: Data management module 1 is used to extract file content with different document formats from externally input bidding-related documents; Knowledge base construction module 2, connected to data management module 1, is used to process file content in different document formats to generate semantic slices corresponding to different file content. It converts all semantic slices into vector data through a text embedding model and maps them to the same semantic space. Based on the vector data, it constructs a knowledge base for the file content. The intelligent review module 3, connected to the knowledge base construction module 2, is used to review the bid documents of the project to be bid using the knowledge base.

[0019] Specifically, data management module 1 supports batch file uploads.

[0020] Data management module 1 receives bidding-related documents in batches, serving as a professional knowledge reserve to support the knowledge base. Furthermore, it extracts the content of documents in different document formats to meet the multimodal review requirements of formal bidding documents.

[0021] In this solution, different document formats include text, images, and tables.

[0022] Furthermore, the knowledge base construction module 2 obtains the file content extracted by the data management module 1, and obtains semantic slices of all file content.

[0023] Specifically, for text content, semantic slices are obtained by treating each text paragraph as a unit.

[0024] Similarly, for image content, each image is used as a unit to obtain the corresponding semantic slice.

[0025] Similarly, for table content, the structured content of each table is used as a unit to obtain the corresponding semantic slice.

[0026] Furthermore, the knowledge base construction module 2 performs semantic understanding on the above three types of semantic slices and performs unified processing on the semantic slices.

[0027] Specifically, the knowledge base construction module 2 uses a text embedding model to convert all semantic slices into unified high-dimensional vector data, thereby achieving mapping within the same semantic space. Furthermore, since the text embedding model associates similar semantics and the vector data is associated with the corresponding semantic slices, the vector data also has the same logical relationship as the semantic slices.

[0028] In a preferred embodiment of the present invention, the bidding-related documents include the tender documents corresponding to the project to be bid on, historical bidding reference templates, project information, and industry standards.

[0029] Specifically, the tender documents for the project to be tendered, which serve as the tender requirements for the corresponding tender documents, also include procurement requirements, time requirements, etc.

[0030] Similarly, the reference template is the historical winning bid document, which is the bid document of a successful bidder.

[0031] Similarly, project materials include technical solutions, design documents, architecture diagrams, and flowcharts. These project materials form the foundation for this tender; therefore, a new set of project materials corresponding to each new tender project must be uploaded.

[0032] Similarly, industry standards include national standards, industry norms, and related laws and regulations.

[0033] Specifically, bidding-related documents are used to form a knowledge base, serving as the criteria for reviewing bidding documents.

[0034] in, One tender document corresponds to one project to be tendered. As a guiding document in the bidding process, the tender document provides a reference for the chapter structure of the tender documents, which is used for subsequent review of the chapter structure of the tender documents or for generating the chapter structure of the tender documents for the project to be tendered.

[0035] The historical bid reference template is the format framework of historical winning bid documents. Generally speaking, the reference template is used in the generation process of the intelligent bid document generation module 4 to provide a benchmark framework for the large model to formulate bid documents.

[0036] Project data forms the core of the knowledge base, including multimodal document content such as implementation overviews, construction drawings, and material lists. Different bidding-related materials will span multiple professional fields. When generating bid documents, the intelligent bid document generation module 4 will also call relevant materials based on semantic associations and use them as content prompts (the intelligent bid document review and intelligent generation will be mentioned later).

[0037] Industry standards are generally used to assist in checking the standardization of tender documents and the legality of their content.

[0038] In a preferred embodiment of the present invention, when generating semantic slices, the data management module 1 obtains the attribute features of the file content to form corresponding metadata; The attribute features include the source of the file content, chapter level, content type, and timestamp.

[0039] Specifically, the bidding documents and historical bid documents have a clear chapter directory forming a multi-chapter structure. When the data management module 1 obtains the original bidding / bid documents, it parses the names of the bidding / bid parties in the bidding / bid documents to obtain the source of the document content; it parses the chapter hierarchy of the bidding / bid document directory, treating each chapter as part of metadata; it determines the content type of each semantic slice as one of text, image, and table; and the data management module 1 extracts the timestamp of the corresponding moment in the system when the above parsing begins.

[0040] For project materials and industry standards, these two types of documents may not have a clear chapter hierarchy. Therefore, it is only necessary to extract the relevant acquisition source, determine the content type of each semantic slice based on semantic analysis, and record the corresponding timestamp during the determination.

[0041] Furthermore, the data management module 1 forms the aforementioned attribute features into textual metadata, which is then attached as tags to the corresponding semantic slices.

[0042] The following describes several specific embodiments to facilitate a better understanding of the implementation of this technical solution by those skilled in the art: Example 1: Data management module 1 receives the externally uploaded tender document A, then extracts its content, slices it into semantic slices, and assigns text-type metadata to each slice. The metadata includes the source file and chapter level. The source file is always tender document A; the chapter level is the chapter number of the semantic slice's directory. This metadata is then assigned to the semantic slice corresponding to tender document A. Knowledge base construction module 2 converts the semantic slices into vector data using a text embedding model and stores the vector data. During subsequent review of tender documents, the module checks the source file portion of the metadata to confirm whether the tender document corresponds to the project being tendered; and it aligns the tender document's directory with the chapter level.

[0043] Example 2: Data management module 1 receives an externally uploaded historical bid reference template B, then extracts the content of this template, slices the content into segments, and assigns text-type metadata to each semantic segment. The metadata includes the source file, chapter level, and content type. Specifically, the source file is always historical bid reference template B; the chapter level is the chapter directory number where the semantic segment is located; and the content type is categorized according to the semantic segment's modality. This metadata, formed by the above attributes, is assigned to the corresponding semantic segment of historical bid reference template B. Knowledge base construction module 2 converts the semantic segments into vector data using a text embedding model and stores the vector data. When generating bid documents later, it queries the source file in the metadata to confirm the target content for generating the bid document, determines the chapter hierarchy structure of the generated bid document through the chapter level, and determines the type of content to be inserted under each chapter through the content type.

[0044] Example 3: Data management module 1 receives externally uploaded project data C, extracts the content of project data C, slices the content of project data C, and assigns text-type metadata to each semantic slice. The metadata includes the source file and content type. The source file is always project data C; the content type is classified according to the modality of the semantic slice. The metadata formed by the above attribute features is assigned to the corresponding semantic slice of project data C. Knowledge base construction module 2 converts semantic slices into vector data using a text embedding model and stores the vector data. During subsequent bid document generation, it retrieves and recalls several strongly correlated vector data by querying multiple identical contents in the metadata. This serves as contextual hints for inputting into the main model and as an index for recalling original materials in the review report.

[0045] Example 4: Data management module 1 receives industry standard D uploaded from an external source, extracts the content of this industry standard D file, slices the content of industry standard D, and assigns text-type metadata to each semantic slice. The metadata includes the source file and content type. The source file is always industry standard D; the content type is classified according to the modality of the semantic slice. The metadata formed by the above attribute features is assigned to the semantic slice corresponding to industry standard D. Knowledge base construction module 2 converts the semantic slices into vector data using a text embedding model and stores the vector data. During subsequent bid document generation, several strongly correlated vector data are retrieved and recalled by querying multiple identical contents in the metadata, serving as contextual hints input to the overall model.

[0046] In a preferred embodiment of the present invention, such as Figure 2 As shown, knowledge base construction module 2 includes: The first semantic processing unit 21 is used to directly extract text-based data from the file content and generate semantic slices associated with the text-based data. The second semantic processing unit 22 is used to analyze the image-type data in the file content and generate semantic slices associated with the image-type data. The third semantic processing unit 23 is used to analyze the tabular data in the file content and generate semantic slices associated with the tabular data.

[0047] When generating semantic slices, the knowledge base construction module 2 uses corresponding semantic understanding methods to perform semantic understanding based on different modalities.

[0048] Specifically, semantic understanding methods include using deep learning, relation extraction (RE), and event extraction (EE) for text, using OCR to parse the content of images to generate structured text, and using table-transformer to generate natural language summaries corresponding to the table content.

[0049] On the one hand, for text data, the first semantic processing unit 21 divides the extracted text into several paragraphs, and takes each paragraph as a semantic slice of the smallest unit, as a semantic slice related to the text data.

[0050] Specifically, the first semantic processing unit 21 is based on deep learning to capture the semantic information of each text segment, extracts the semantic context features through LSTM, and identifies named entities in the segments. Named entities are proper nouns that appear in the tender documents.

[0051] Subsequently, two or more named entities are extracted and identified through RE, and the contextual relationships between the named entities are parsed to obtain the semantic relationships between the named entities.

[0052] Subsequently, event information in the text paragraph is extracted using EE, trigger words of events in the text are identified based on preset prompt words, the identified trigger words are classified into the corresponding event types, the named entities related to the trigger words are identified, the event elements played by the named entities in the corresponding triggered event types are determined, and the entire event information is represented in a structured manner.

[0053] The first semantic processing unit 21 combines the source, chapter structure, extracted named entities, semantic relationships, and event information from the metadata to construct structured semantic tags, which are then attached to the corresponding text paragraphs.

[0054] On the other hand, for image data, the second semantic processing unit 22 parses the image slices through OCR, then calls the large model to analyze the image content, generates a structured text description containing image type, visual elements, and technical information as a semantic slice of the image data, and obtains the URL of the image, and associates the image URL with the vector converted from the semantic slice of the image.

[0055] The visual elements include the image style, the content of the text within the image, and the top-to-bottom relationships of each component in the image. The technical information includes information such as the recall, precision, and processing speed for each slice.

[0056] On the other hand, for tabular data, the third semantic processing unit 23 parses the row and column structure of the table, converts its content into corresponding structured data, and calls the large language model to understand the structured data, generating a natural language summary that summarizes its theme and core content as a semantic slice of the tabular data.

[0057] Specifically, the table-transformer identifies the semantics of rows, columns, and cells through table structure recognition, extracts the structured information of rows, columns, and tables, generates natural language summaries based on the structured information, and uses them as semantic slices of table modal data.

[0058] Each semantic slice of text, image, and table obtained through the three semantic processing units is used as a basic unit, which facilitates the processing and retrieval of the content corresponding to a single semantic slice.

[0059] In a preferred embodiment of the present invention, the constructed knowledge base includes: The vector layer is used to store vector data and metadata associated with the file content to form a retrieval index for the knowledge base; The document layer is used to store the structured content of semantic slices; The relationship layer, based on the chapter hierarchy of file content, establishes and saves the connection relationships between the contents of various files; By using the slice number of the semantic slice, the vector data in the vector layer, the structured content in the document layer, and the connection relationships in the relation layer are associated.

[0060] The slice ID consists of multiple segments of numbers, in the form of XX-XXX-XX, where each segment is a sub-sequence number, and each sub-sequence number corresponds to an index. The slice ID is also assigned to the corresponding vector data as part of the metadata.

[0061] Specifically, the vector layer, as a vector database, stores various vector data in sub-vector databases of the vector database, including bidding requirements database, historical template database, project information database, and industry standard database.

[0062] Each sub-vector database has a corresponding sub-index number.

[0063] The tender requirements database stores vectors corresponding to semantic slices of the tender documents, and the portion of the slice number corresponding to these vectors that represents the type of tender-related documents is the sub-number corresponding to the tender requirements database.

[0064] The historical template database stores vectors corresponding to semantic slices of historical bid documents, and the portion of the slice number corresponding to these vectors that represents the bid-related document type is the sub-number corresponding to the historical template database.

[0065] The project data database stores vectors corresponding to semantic slices containing technical solutions, design documents, architecture diagrams, flowcharts, etc., and the part of the slice number corresponding to these vectors that represents the type of bidding-related documents is the sub-number corresponding to the project data database.

[0066] The industry standard database stores vectors corresponding to semantic slices containing industry standards and related regulations. Among the slice numbers corresponding to these vectors, the part used to represent the sub-sequence number of the tender-related document type is the sub-sequence number corresponding to the industry standard database.

[0067] Vectors in the same sub-vector database can be retrieved by using the sub-index of the same part, which facilitates the strong correlation of the retrieved content.

[0068] Furthermore, the uploaded different bidding-related documents may correspond to different technical fields. When the knowledge base construction module 2 performs unified processing on the bidding-related documents, it understands the semantics of the text, images and tables and forms semantic slices. Based on the understood semantics, it determines the technical field to which the corresponding semantic slice belongs, and each technical field corresponds to a sub-serial number.

[0069] Furthermore, corresponding sub-serial numbers are set according to the order in which the bidding-related documents are uploaded.

[0070] Furthermore, when the knowledge base construction module 2 slices the bidding-related documents, it assigns a sub-sequence number to the semantic slices in each file according to the slicing order and keeps a count. When all slices in a file have been assigned the sub-sequence numbers corresponding to the order, the count is cleared, and the count is restarted when the semantic slices of the next file are assigned new sub-sequence numbers corresponding to the order.

[0071] The following describes a specific embodiment to facilitate understanding of the implementation of this technical solution by those skilled in the art: The naming convention for slice numbers is as follows: Sub-slice 1 - Sub-slice 2 - Sub-slice 3 - Sub-slice 4. Sub-slice 1 corresponds to the type of tender-related document, sub-slice 2 corresponds to the technical field, sub-slice 3 corresponds to the document number, and sub-slice 4 corresponds to the position of the semantic slice in the uploaded file. In the vector database, the sub-slice number for the tender requirements database is 1, for the historical template database it is 2, for the project data database it is 3, and for the industry standards database it is 4; the sub-slice number for the construction industry is 01, and for marine engineering it is 02. We are now uploading a construction industry project document, which is the 52nd uploaded tender-related document, containing multiple technical paragraphs and images inserted within the paragraphs. Then, each paragraph and each image is sliced. The slice ID of paragraph 1 is 3-01-52-001, the slice ID of paragraph 2 is 3-01-52-002, and the slice ID of image 1 is 3-01-52-003, which is between paragraph 2 and paragraph 3. The slice ID of image 1 is 3-01-52-003, and so on, until all content is sliced.

[0072] In practical applications, the naming rules for slice numbers and the content corresponding to sub-slices are determined by technical personnel based on the uploaded bidding-related documents and the construction of the knowledge base.

[0073] Through contrastive learning or multi-task learning, semantically related cross-modal vectors are aggregated in the corresponding sub-vector database.

[0074] Specifically, the knowledge base construction module 2 transforms the unified semantic slices of each modality into high-dimensional vectors through a pre-set and fine-tuned text embedding model, realizing isomorphic mapping of multimodal content in the same semantic space, so that data from different modalities can be processed or accessed in the same way.

[0075] In the contrastive learning task, training sample pairs are first constructed, including positive sample pairs with semantic matching and negative sample pairs with semantic mismatch. Then, the contrastive loss is calculated to maximize the cosine similarity of positive sample pairs and minimize the cosine similarity of negative sample pairs, thereby enhancing the semantic discrimination ability of the text embedding model and making subsequent semantic associations more accurate.

[0076] At the same time, multiple learning tasks can be constructed to assist in loss calculation.

[0077] After training, all vectors in the same sub-vector database will be aggregated. For vectors with different semantics, separate aggregation will be implemented. That is, multiple vectors with high semantic relevance will be aggregated into a set. Furthermore, the semantic similarity between any two cross-modal vectors can be directly calculated using cosine similarity.

[0078] Since the length of a vector has no physical meaning in cross-modal vector computation, while cosine similarity only focuses on direction, the level of cosine similarity can accurately reflect whether the semantics are similar.

[0079] Similarly, the document layer serves as a retrieval library, storing the structured content of semantic slices, which is obtained through the semantics of the document content.

[0080] Specifically, each semantic slice stores structured content in JSON format, including structured semantic tags, URLs, and natural language summaries. The JSON also stores the slice ID of the corresponding semantic slice and establishes a mapping with the corresponding vector data.

[0081] The structured content stored in the document layer can be directly accessed when reviewing inspection results or generating relevant materials.

[0082] Specifically, the document layer directly stores the original text of the text data, and the text paragraphs processed by the first semantic processing unit 21 are attached with structured semantic tags. When the tender document intelligent generation module 4 generates the tender document, it is necessary to recall relevant materials. For the text, the tender document intelligent generation module 4 retrieves relevant vector data in the vector library through similarity, associates the slice ID of the metadata on the vector data with the content of the structured semantic tag through JSON, and recalls the original text paragraphs related to the structured semantic tag.

[0083] Furthermore, the document layer stores the URL of the image data processed by the second semantic processing unit 22 and the structural description of the image. When the tender document intelligent generation module 4 generates the tender document, it is necessary to recall the relevant materials. For the image, the tender document intelligent generation module 4 retrieves the relevant vector data in the vector library through similarity, associates the slice ID of the metadata on the vector data with the URL through JSON, and retrieves the corresponding image by indexing the URL.

[0084] Furthermore, the document layer stores the structured data of the table and the natural language summary obtained by the third semantic processing unit 23. When the tender document intelligent generation module 4 generates the tender document, it is necessary to recall the relevant materials. For the table, the tender document intelligent generation module 4 retrieves the relevant vector data in the vector library through similarity, associates the slice ID of the metadata on the vector data with the natural language summary through JSON, and recalls the structured data of the table corresponding to the natural language summary.

[0085] The content of similarity retrieval will be mentioned later.

[0086] Similarly, the relational layer, as a relational database, stores the chapter structure composed of the chapter directory of each file's content and the connection relationships between the contents of each file.

[0087] Specifically, multiple related tables are created at the relationship layer.

[0088] In the associated table, slice IDs related to the chapter level are stored in the same table based on the slice ID and chapter level in the metadata, and the connection relationship between file contents is stored based on the JSON containing the slice ID.

[0089] in, The structure of a chapter table of contents is generally divided into first-level headings and subheadings. Each first-level heading includes several levels of subheadings. A chapter table of contents consists of multiple first-level headings, and the chapter structure is composed of multiple first-level headings and multiple subheadings.

[0090] The content of a document under a first-level heading is related to the content of multiple documents under the corresponding first-level heading, forming a structured primary subordinate relationship.

[0091] The content of a document under a subheading is related to the content of the document under the next subheading, forming a secondary subordinate relationship.

[0092] The content of multiple documents with the same level of subheadings is interconnected, forming a structured relationship of similar content.

[0093] The connection relationships include first subordinate relationships, second subordinate relationships, and content similarity relationships, which are stored in different association tables of the relationship layer.

[0094] Vector data, structured content, and relationships are linked based on slice IDs and corresponding JSON. The JSON is then linked to the corresponding structured semantic tags, URLs, and natural language summaries in the document layer. The relevant structured content is retrieved as the original material and provided to the tender document intelligent generation module 4. The order of chapters in the generated tender document is determined by the connection relationship.

[0095] In a preferred embodiment of the present invention, such as Figure 3 As shown, the intelligent review module 3 includes: The document processing unit 31 is used to unify the chapter directory of the tender documents of the project to be tendered into the text format of the corresponding chapter level of the tender documents in the knowledge base; The chapter review unit 32 and the file processing unit 31 are used to match and review the tender documents by chapter according to the content of the tender documents in the knowledge base, and output the chapter review results.

[0096] After receiving the tender documents to be reviewed, the intelligent review module 3 needs to call the semantically related vectors of the content of the tender documents to be reviewed during the intelligent review. However, it is not possible to directly compare the original documents through vectors. Therefore, the multimodal content of the tender documents to be reviewed is first semantically understood, sliced ​​and vectorized to form semantic slices that can be processed uniformly.

[0097] Specifically, the intelligent review module 3 receives and parses the tender documents to be reviewed, slicing the text, images, and tables into segments. For text content, it obtains the corresponding semantic segments for each text paragraph; for image content, it obtains the corresponding semantic segments for each image; and for table content, it obtains the corresponding semantic segments for each table's structured content. Each semantic segment is then assigned temporary metadata, including chapter level and content type.

[0098] Subsequently, the chapter structure of the tender documents is reviewed. The document processing unit 31 unifies the chapter indexes of the tender documents and the tender documents into a structured format, and then aligns the chapter indexes of the tender documents with those of the tender documents. Then, semantic alignment matching is performed using a large model, and the review results are output based on the matching results.

[0099] Specifically, the unified structured format is: chapter ID - chapter name. The format of each section in the chapter directory is standardized until all chapter directories are processed and correspond one-to-one, avoiding matching problems caused by format confusion.

[0100] Furthermore, the chapter review unit 32 begins to review the documents after the document processing unit 31 has finished processing them. It pre-sets prompt words to determine the purpose of the review, which includes reviewing the completeness of the content, consistency of expression, and responsiveness of instructions in the tender documents. It also sets review rules to determine whether the content completeness, consistency of expression, and responsiveness of instructions are met. The chapter review unit 32 calls the large language model, which combines the prompt words and review rules to conduct a multi-channel parallel review of the chapter structure of the tender documents.

[0101] in, When reviewing the completeness of the content, if the tender document does not have a corresponding chapter table of contents or the core content is missing, the chapter table of contents will be judged as missing. When reviewing the consistency of expression, if the content of the bidding chapter's table of contents has semantic conflicts with the bidding requirements, omits key indicators, or has expression deviations, then that chapter is judged to be inconsistent. The review of instruction responsiveness is used to comprehensively determine whether the chapter list of the tender document to be reviewed is acceptable. If the chapter list does not fully meet the tender requirements or is semantically inconsistent, it is judged as unacceptable. Only when the chapter list of the tender document to be reviewed is consistent with the chapter list of the tender document and the wording is consistent, will the chapter pass the chapter review.

[0102] To address issues such as missing chapter headings and inconsistencies between chapter content and bidding requirements, the large language model generates corresponding solutions or modification suggestions based on the judgment results and the content of the tender documents.

[0103] Furthermore, once the large language model has completed its review of all chapters of the tender documents to be reviewed, it extracts the review results for each chapter and integrates them into a chapter review report for users to query.

[0104] In addition, the chapter review unit 32 also saves the review log of each chapter index, which is used to trace the reasoning process of the current chapter index.

[0105] In a preferred embodiment of the present invention, the relationship layer also has a pre-set scoring rule library for the bidding project. The scoring rule library includes the scoring type of each bidding item in the bidding document of the bidding project, and the scoring standard corresponding to the bidding item whose scoring type is objective scoring item. The intelligent review module 3 also includes: The scoring type judgment unit 33 is used to judge the scoring type of each bid item in the bid document of the project to be bid, and output the type judgment result; The first scoring unit 34 is connected to the scoring type judgment unit 33. It is used to provide the bidding items with the scoring type of subjective scoring items to the user for subjective scoring based on the type judgment result, and to obtain the corresponding subjective scoring result input by the user. The second scoring unit 35 is connected to the scoring type judgment unit 33. It is used to automatically score the bid items with the scoring type of objective scoring items according to the type judgment result and the corresponding scoring standard, and output the corresponding objective scoring result. The intelligent review module 3 then outputs the review results corresponding to the tender documents based on the subjective and objective scoring results.

[0106] Specifically, the scoring type determination unit 33 receives and parses the bid documents of the project to be bid on, obtains multiple bid items in the document content, and extracts each bid item as a segment to be scored. Then, it determines whether each segment to be scored is a subjective or objective scoring item. The rule engine drives the segment to be scored to be routed to the corresponding scoring unit, routing the segment corresponding to the subjective scoring item to the first scoring unit 34, and the segment corresponding to the objective scoring item to the second scoring unit 35.

[0107] Furthermore, on the one hand, the first scoring unit 34 has a built-in subjective scoring strategy, including subjective scoring prompts and subjective scoring rules. First, the fine-tuned large language model scores the segments based on the subjective scoring strategy. Then, professionals review and modify the scoring results of each segment to obtain subjective scoring results that meet the bidding requirements.

[0108] On the other hand, the second scoring unit 35 has a built-in objective scoring strategy, including objective scoring prompts and objective scoring rules. The fine-tuned large language model directly scores the objective scoring items based on the objective scoring strategy. Generally speaking, the content of the objective scoring items can be intuitively quantified with data. The large language model scores directly based on the quantification standards and data comparison, and outputs the objective scoring results.

[0109] Specifically, the first scoring unit 34 and the second scoring unit 35 each have built-in scoring prompts that differ from each other, but they generally include scoring details, score threshold requirements, and traceability standards.

[0110] The scoring details are pre-defined within two scoring units, including multiple scoring methods for each subjective and objective scoring item. The scoring rules determine the scoring principles for objective scoring items, such as scoring technology or evaluating business value. The scoring details are subordinate to the scoring rules, and scoring is conducted through a combination of judgments. Score thresholds are set based on the scoring rules, defining whether the final scoring result is acceptable. The traceability specifications are also implemented within the two scoring units, used to associate the scoring results with the corresponding scoring items after scoring.

[0111] Both subjective and objective rating items undergo a single scoring process by the large model, which automatically generates modification suggestions for rating items whose scores fall below a certain threshold.

[0112] After the large model is scored, the deductions for each subjective and objective scoring item are marked. The scoring items with deductions are recorded as deduction items, and a scoring explanation is listed next to each scoring item, including the scoring rules and the basis for subjective and objective scoring.

[0113] Furthermore, the intelligent review module 3 counts the number of segments corresponding to each rating item, takes the rated segments as valid hit segments, and generates a correlation mapping dataset of rating results and hit segments. This dataset is then sent to the client for users to query and trace.

[0114] in, For segments that meet the scoring criteria and are scored, the intelligent review module 3 classifies them as valid hit segments and saves them to the dataset.

[0115] For unmatched or duplicate segments, which do not contribute to the scoring process or may interfere with it, the intelligent review module 3 classifies them as invalid segments and does not include them in the scoring report.

[0116] In a preferred embodiment of the present invention, the relational database corresponding to the relational layer includes a keyword library, and the intelligent review module 3 further includes keyword verification after performing subjective and objective scoring.

[0117] Specifically, the keyword database includes keywords related to project names, system names, commissioning units, and administrative divisions.

[0118] The project name is used to determine the project plan in the tender documents; the system name is used to confirm the technical platform and equipment for implementing the project plan; the commissioning unit is used to confirm the name and qualifications of the tendering party / bidder; and the administrative division is used to confirm the location for implementing the project plan.

[0119] Keyword verification is used to determine whether there are any errors in the submission format of the tender documents. Tender documents that do not meet the requirements of the tendering party will not be adopted.

[0120] Specifically, the intelligent review module 3 performs a full-text search on the tender documents to be reviewed, locates keywords in the document content based on the keyword database, and performs a matching check to ensure that these keywords are expressed correctly.

[0121] Furthermore, the intelligent review module 3 segments and vectorizes the keywords in the document content to obtain vector data corresponding to multiple keywords. High-dimensional semantic feature matching is performed on the vector data corresponding to the keywords to calculate the most similar vector data. These vector data are then retrieved from the vector database. The content of these vector data corresponds to synonyms, near-synonyms, or similar-looking characters of the keywords in the document content. Then, a large language model is used to perform deep semantic understanding on the retrieved vector data and match it with the document content. Content in the document content that does not match the keywords is intelligently removed and transformed into content corresponding to the keywords. After matching and modifying all document content, all keyword parts in the document content are highlighted, helping users avoid hidden compliance risks such as typos, misuse of old names and aliases, and ensuring the accuracy of key information in the bidding documents.

[0122] The document content may be vague in some aspects, or use simple descriptions or abbreviations, potentially leading to multiple uses for a single word. In such cases, vector data of possible synonyms, near-synonyms, or similar-looking characters in the vector library should be retrieved for judgment. The large language model obtains a unique understanding result through deep semantic understanding and removes vectors that do not match the result.

[0123] In a preferred embodiment of the present invention, cross-modal consistency review is further included after keyword detection.

[0124] Specifically, the intelligent review module 3 integrates the review results from each review channel and aligns the document content of different modalities in the tender documents based on these review results, ensuring that the text, charts, and tables correspond one-to-one.

[0125] Specifically, the intelligent review module 3 performs deep semantic understanding on two slices related during the review process, whether they are of the same modality or cross-modality. It combines the association relationships stored in the relational database through slice IDs to determine whether the two slices can correspond semantically. If there is a situation where the text description does not match the semantics of the adjacent illustration or there are too few similar semantics, the problem is saved and output to the problem list in the review report.

[0126] Specifically, the intelligent review module 3 has a built-in semantic matching threshold. Based on the semantic repetition of named entities in two semantic segments, the named entities of the text are obtained through the first semantic processing unit 21, the named entities of the image are parsed by OCR through the second semantic processing unit 22, and the named entities of the table are extracted by NER from the natural language summary obtained by the third semantic processing unit 23.

[0127] During the semantic matching process, if the semantic repetition of two semantic segments is lower than the matching threshold, the two segments are judged to be semantically inconsistent, and the issue is saved to the issue list; if the semantic repetition of two semantic segments is higher than the matching threshold, the two segments are judged to be semantically related, and the semantic relationship review is passed.

[0128] Furthermore, after all review processes are completed, the intelligent review module 3 integrates all review results and generates a structured review report. The review report includes a list of issues, deductions, scoring explanations, and modification suggestions. Moreover, each issue in the report can be traced back to the knowledge base via links provided in the scoring result-hit segment association mapping dataset. The original text, image, and tabular data are linked through the segment IDs in the links, enabling an evidence-based and traceable review process.

[0129] Specifically, the intelligent review module 3 returns the integrated review report and tender documents to the user.

[0130] in, The review report includes summaries described in text paragraphs, such as summaries of chapter structure review and cross-modal consistency checks, and highlights deducted points in red with scoring explanations next to them; In addition, the materials used when rating the scoring items are retrieved from the relationship layer by the intelligent review module 3, and users can use this link to access the original materials in the document layer for viewing; In addition, the problem list is presented in tabular form, with automatically generated modification suggestions for each problem; In the tender documents, the keywords to be reviewed are highlighted and accompanied by the names from the keyword database for users to check a second time, to prevent inconsistencies with the official names or the use of old names with changed addresses.

[0131] In a preferred embodiment of the present invention, similarly as Figure 1 As shown, it also includes a tender document intelligent generation module 4, which is connected to the knowledge base construction module 2 and is used to automatically generate tender documents for the project to be tendered based on the knowledge base; like Figure 4 As shown, the intelligent tender document generation module 4 includes: The requirement acquisition unit 41 is used to acquire the tender document generation requirements input by the user; Chapter generation unit 42, connected to requirement acquisition unit 41, is used to generate chapter directory of bidding documents corresponding to the bidding project in the knowledge base based on the requirements of the tender document and the knowledge base. The content retrieval unit 43, connected to the chapter generation unit 42, is used to retrieve the structured content of the corresponding semantic slice from the knowledge base for each chapter based on the chapter directory. The draft generation unit 44 is connected to the content retrieval unit 43 and the chapter generation unit 42 respectively. Based on all the recalled structured content, it generates a contextually coherent draft document according to the chapter directory. The adjustment unit 45 is connected to the draft generation unit 44 and is used to provide the draft document to the user for modification and adjustment, and to obtain the modified draft document as the final completed tender document.

[0132] The intelligent tender document generation module 4 is used to generate tender documents for projects to be tendered when the user has not uploaded tender documents.

[0133] Specifically, the chapter generation unit 42 parses the user requirements obtained from the requirement acquisition unit 41, retrieves the scoring details from the relational database through RAG retrieval enhancement based on the parsing results, and retrieves the vectors in the historical template vector library through the slice ID, automatically generating a chapter directory that is strongly associated with the scoring items and has a clear hierarchy to form the chapter structure.

[0134] Furthermore, the draft generation unit 44 obtains several structured contents recalled by the content recall unit 43, integrates them with user needs as contextual prompts containing multimodal information, and submits the contextual prompts to the large language model. The model generates chapter drafts that are highly relevant to user needs and logically coherent based on the contextual prompts.

[0135] Furthermore, users can modify document drafts using the built-in rich text editor.

[0136] Specifically, users can easily view the recalled structured content through a rich text editor, and can also copy the required structured content and embed it into the main text to complete the final version of the tender document.

[0137] The tender documents generated using this invention will be saved, and their content and structured data will be synchronously updated in the knowledge base, serving as a reference template for future projects. Furthermore, the final tender document version will be seamlessly submitted to the intelligent review module 3, initiating the subsequent automated quality review process.

[0138] Specifically, the generated tender documents will be used as historical templates. The knowledge base construction module 2 will perform unified semantic slicing and semantic association on the content, and then convert the semantic slices into vectors and store them in the historical template database of the vector layer, along with relevant metadata.

[0139] In a preferred embodiment of the present invention, such as Figure 5 As shown, the content recall unit 43 includes: The retrieval subunit 431 is used to retrieve relevant vector data from the vector layer of the knowledge base by taking the chapter topic of each chapter as the retrieval condition and output the retrieval results. The recall subunit 432 is connected to the retrieval subunit 431, and is used to recall the structured content of the semantic slice corresponding to the vector data based on the relevance of the retrieval results.

[0140] Specifically, the retrieval subunit 431 performs Top-K similarity retrieval in parallel across four sub-vector databases in the knowledge base for each chapter in the chapter catalog, using the chapter's topic as the retrieval condition. The recall subunit 432, based on the retrieval results, recalls K multimodal vector data highly related to the chapter topic in each sub-vector database. It matches the metadata content of the vector data with the content of structured semantic tags, URLs, and natural language summaries, and recalls the corresponding structured data, as well as the structured semantic tags, URLs, and natural language summaries, which are used as contextual prompts for subsequent draft generation and as original materials for easy viewing by users.

[0141] Specifically, the retrieval subunit 431 searches for the K vectors with the highest similarity in each vector database. These vectors are strongly related to the topic of the corresponding chapter. After the retrieval is completed, the recall subunit 432 recalls all the retrieved vectors according to the information sent by the retrieval subunit 431, for subsequent use.

[0142] In this invention, the review process for tender documents is closed-loop. Two embodiments are provided below to help those skilled in the art more easily understand the implementation of this technical solution: Example 1: Users possess a bid document for a project to be tendered and upload it to the intelligent bidding scheme review system of this invention. The intelligent review module 3 retrieves the tender document for the project from the knowledge base, aligns the chapters of the bid document with the tender document in a structured manner, and then reviews the completeness of the content, consistency of expression, and compliance with the requirements of the tender document's chapter directory. It then performs subjective and objective scoring on each segment of the bid document, executing the corresponding subjective or objective scoring. Finally, it checks the keywords and cross-modal consistency in the bid document. Each review channel outputs a review result. After all review items are completed, the intelligent review module 3 integrates all the above review results to form a complete review report, including a list of issues, deductions, scoring explanations, and modification suggestions, as well as bid documents with highlighted keywords. The review report also includes links to source materials. The review report and bid document are returned to the client. The client decides whether to adopt the bid document based on the review results. If not, modifications can be made based on the review report. Users can retrieve relevant materials from the knowledge base for modification, and the modified bid document can be saved to the knowledge base to form new relevant materials.

[0143] Example 2: If a user does not possess the tender documents for a project but needs to generate one, they input their requirements into the intelligent tender document review system of this invention. The intelligent tender document generation module 4 first retrieves the corresponding tender documents, related tender templates, project materials, etc., from the knowledge base based on the project to be tendered. This information is then combined with the user's input requirements to form contextual prompts, which are input into a large language model to generate a well-structured and highly contextualized draft tender document. The user can then modify the draft using a rich text editor and the accompanying source materials until it meets the client's needs and the tender requirements, thus forming a complete tender document. The generated tender document can also be sent to the intelligent review module 3 for review, generating a review report. The user decides whether to adopt the tender document based on the review results. After confirming the final version, the generated tender document is automatically submitted to the knowledge base construction module 2 and saved as new material in a hierarchical manner within the knowledge base, forming a logical closed loop.

[0144] The above description is merely a preferred embodiment of the present invention and does not limit the implementation and protection scope of the present invention. Those skilled in the art should realize that any equivalent substitutions and obvious changes made based on the description and illustrations of the present invention should be included within the protection scope of the present invention.

Claims

1. A smart bidding proposal review system, characterized in that, include: The data management module is used to extract file content with different document formats from externally input bidding-related documents; The knowledge base construction module, connected to the data management module, is used to process the file content in different document formats to generate semantic slices corresponding to different file content. It converts all semantic slices into vector data and maps them to the same semantic space through a text embedding model, and constructs a knowledge base for the file content based on the vector data. The intelligent review module, connected to the knowledge base construction module, is used to review the bid documents of the project to be bid using the knowledge base.

2. The intelligent review system for bidding proposals according to claim 1, characterized in that, The bidding-related documents include the tender documents, historical bidding reference templates, project information, and industry standards corresponding to the project to be bid on.

3. The intelligent review system for bidding proposals according to claim 1, characterized in that, When generating the semantic slice, the data management module obtains the attribute features of the file content to form corresponding metadata; The attribute features include the source of the file content, chapter level, content type, and timestamp.

4. The intelligent review system for bidding proposals according to claim 1, characterized in that, The document content can be in text, image, or table format. The knowledge base construction module includes: The first semantic processing unit is used to directly extract text-based data from the file content and generate the semantic slice associated with the text-based data. The second semantic processing unit is used to analyze the image-format data in the file content and generate the semantic slice associated with the image-format data; The third semantic processing unit is used to analyze the tabular data in the file content and generate the semantic slice associated with the tabular data.

5. The intelligent review system for bidding proposals according to claim 3, characterized in that, The constructed knowledge base includes: The vector layer is used to store the vector data and the metadata associated with the file content to form a retrieval index for the knowledge base; The document layer is used to store the structured content in the semantic slices; The relationship layer establishes and saves the connection relationships between the various contents of the file based on the chapter hierarchy of the file content; The semantic slice sequence number is used to associate the vector data in the vector layer, the structured content in the document layer, and the connection relationship in the relation layer.

6. The intelligent review system for bidding proposals according to claim 3, characterized in that, The intelligent review module includes: The document processing unit is used to unify the chapter directory of the tender documents of the project to be tendered into the text format of the corresponding chapter level of the tender documents in the knowledge base; The chapter review unit, connected to the document processing unit, is used to match and review the tender documents by chapter according to the document content of the tender documents in the knowledge base, and output the chapter review results.

7. The intelligent review system for bidding proposals according to claim 5, characterized in that, The relationship layer also has a pre-set scoring rule library for the project to be tendered. The scoring rule library includes the scoring type of each tender item in the tender document of the project to be tendered, and the scoring standard corresponding to the tender item whose scoring type is objective scoring item. The intelligent review module includes: The scoring type determination unit is used to determine the scoring type of each bid item in the bid document of the project to be bid, and output the type determination result; The first scoring unit, connected to the scoring type judgment unit, is used to provide the bidding item with the scoring type of subjective scoring item to the user for subjective scoring based on the type judgment result, and to obtain the corresponding subjective scoring result input by the user. The second scoring unit is connected to the scoring type judgment unit and is used to automatically score the bid item whose scoring type is objective scoring item according to the type judgment result and the corresponding scoring standard, and output the corresponding objective scoring result. The intelligent review module then outputs a review result corresponding to the tender document based on the subjective and objective scoring results.

8. The intelligent review system for bidding proposals according to claim 5, characterized in that, It also includes a tender document intelligent generation module, which is connected to the knowledge base construction module and is used to automatically generate the tender document for the project to be tendered based on the knowledge base; The intelligent tender document generation module includes: The requirement elicitation unit is used to obtain the tender document generation requirements input by the user; The chapter generation unit, connected to the requirement acquisition unit, is used to construct a chapter directory of the tender document corresponding to the tender project to be tendered in the knowledge base based on the tender document generation requirements and the knowledge base. The content retrieval unit, connected to the chapter generation unit, is used to retrieve the structured content of the corresponding semantic slice from the knowledge base for each chapter according to the chapter directory. The draft generation unit is connected to the content retrieval unit and the chapter generation unit respectively, and generates a context-coherent draft document based on all the recalled structured content and the chapter directory. The adjustment unit, connected to the draft generation unit, is used to provide the draft document to the user for modification and adjustment, and to obtain the modified draft document as the final completed tender document.

9. The intelligent review system for bidding proposals according to claim 8, characterized in that, The content recall unit includes: The retrieval subunit is used to retrieve relevant vector data from the vector layer in the knowledge base by using the chapter topic of each chapter as the retrieval condition, and output the retrieval results; The recall subunit, connected to the retrieval subunit, is used to recall the structured content of the semantic slice corresponding to the vector data based on the relevance of the retrieval results.

Images