Bidding file generation method and device
By breaking down the tender documents, identifying intent, and retrieving qualifications from the database, and combining this with a large language model to generate tender documents, the problems of low efficiency and insufficient accuracy in tender document generation are solved, and a close correspondence between the tender documents and the tender documents is achieved.
Patent Information
- Application Number
- CN202511243375.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-02
- Publication Date
- 2026-01-02
AI Technical Summary
Existing technologies for generating tender documents are inefficient, inaccurate, and lack specificity. Automated tools struggle to understand the semantic intent of clauses, leading to matching errors and a lack of flexibility in templated responses.
The tender documents are converted into text data using optical character recognition technology, and then segmented into sentences and classified into clauses. The intent is identified and key points for response are extracted. The tender documents are generated using a qualification database and a large language model. Prompt words are constructed by combining intent and key points for response to improve accuracy and relevance.
It improves the completeness and accuracy of tender documents, ensuring that the generated tender documents closely match the requirements of the tender documents, and solves the problems of lack of flexibility and specificity of traditional template-based methods.
Smart Images
Figure CN121257480A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of natural language generation, and in particular to a method and apparatus for generating tender documents. Background Technology
[0002] The preparation of tender documents is a crucial step for enterprises to participate in competition. It requires analyzing each clause of the tender documents, summarizing the requirements, matching the enterprise's qualifications, and writing corresponding responses. Because tender documents often contain numerous clauses, including both explicit objective qualification requirements and subjective technical solutions and service commitments, traditional manual processing methods require extensive data collection, are inefficient, and prone to omissions, resulting in incomplete responses in the generated tender documents.
[0003] To address these issues, existing solutions utilize automated tools to match keywords or fill in templates in tender documents, thereby improving the efficiency of bid generation. However, these methods struggle to truly understand the semantic intent of the clauses, leading to incorrect matching data and reducing the accuracy of the bid responses. Furthermore, when generating bids, subjective responses regarding solutions and service concepts are required. The automated tools available are limited in their templates and cannot flexibly adapt to the specific tender documents, resulting in bids with poor relevance and logical coherence.
[0004] Therefore, how to improve the relevance of the responses in the tender documents to the tender documents while ensuring their accuracy is a technical problem that needs to be solved. Summary of the Invention
[0005] This application provides a method and apparatus for generating tender documents, which can solve the problems of low accuracy of tender documents and poor relevance of the response content in tender documents to the tender documents in the prior art.
[0006] Some embodiments of this application provide a method for generating tender documents, including:
[0007] Receive the tender documents and break them down into several clauses to be responded to;
[0008] For each of the pending response clauses, determine the intent of the pending response clause and extract the response points corresponding to the intent from the pending response clauses;
[0009] For each of the aforementioned clauses to be responded to, the corresponding objective response data is obtained by retrieving from a preset qualification database based on the corresponding intent and the key points of the response.
[0010] Based on all the objective response data, all the intents, and all the response points, prompt words are constructed, and the prompt words are input into a preset large language model to obtain subjective response data;
[0011] The tender document is generated based on all the objective response data and the subjective response data.
[0012] Compared to existing technologies, the above embodiments have the following beneficial effects: First, by splitting and identifying the intent of the tender documents, complex text content is transformed into structured clauses to be responded to, and key response points corresponding to the intent are extracted, ensuring the accuracy and relevance of clause analysis. Second, based on the identified intent and key response points, an automatic search is performed in a pre-set qualification database to obtain objective response data that meets the hard requirements of the tender documents, thereby avoiding omissions and inefficiencies in the manual search process and improving the completeness and data accuracy of the tender documents. Furthermore, when generating subjective content such as scheme descriptions, technical measures, or service commitments, by integrating the retrieved objective response data and key response points, prompt words are constructed and input into a large language model. This allows the large language model to generate more targeted subjective response data based on the objective response data that meets the hard requirements, ensuring that the final tender documents closely adhere to the specific requirements of the tender documents, and solving the problems of lack of flexibility and relevance in traditional template-based methods.
[0013] Furthermore, the receipt of the tender documents involves splitting the tender documents into several pending response clauses, including:
[0014] The terms to be answered include: objective terms to be answered and subjective terms to be answered;
[0015] The tender document was converted from a scanned copy into text data using optical character recognition technology.
[0016] The text data of the tender document is segmented into sentences to obtain several statements;
[0017] For each statement, it is classified into an objective clause to be responded to or a subjective clause to be responded to using a preset classification model.
[0018] Compared to existing technologies, the above embodiments have the following advantages: By automatically converting scanned tender documents into text using optical character recognition technology, and further processing them by sentence segmentation and clause classification, the inefficiency and inaccuracy of manual tender document segmentation can be effectively solved. Furthermore, by using a classification model to distinguish between objective and subjective clauses, subsequent intent recognition and responses become more targeted, avoiding semantic interference caused by the mixing of clauses of different natures and improving the accuracy of response generation.
[0019] Furthermore, for each of the pending response clauses, determining the intent of the pending response clause and extracting response points corresponding to the intent from the pending response clauses includes:
[0020] When the clause to be responded to is the objective clause to be responded to, the first intent is identified from the corresponding statement, and the corresponding slot information is extracted from the statement according to the first intent, and the slot information is used as the response key point;
[0021] When the clause to be responded to is the subjective clause to be responded to, the second intent corresponding to the statement is parsed and the second intent is divided into several sub-intents. Based on the several sub-intents, the sub-intent sequence is labeled for each word in the corresponding statement, and the response points corresponding to each sub-intent are determined.
[0022] Compared to existing technologies, the above embodiments offer the following advantages: In objective clause scenarios, the slot information extraction method accurately captures explicit requirements such as time, quantity, and qualifications, ensuring the accuracy and compliance of objective responses. In subjective clause scenarios, the intent is further decomposed into sub-intents, and key response points are identified based on sequence labeling methods. This not only improves the depth of semantic understanding but also ensures comprehensive coverage of complex clauses. Through differentiated clause processing methods, the "generalization failure" problem of a single model for different clause types is effectively avoided. This ensures that the generated responses possess both accurate matching of objective clauses and flexible expression and logical integrity of subjective clauses, thereby improving the overall accuracy and relevance of bid document responses.
[0023] Further, the step of retrieving corresponding objective response data from a preset qualification database based on the corresponding intent and response points includes:
[0024] The qualification database includes: multiple qualification data and a first feature vector corresponding to each qualification data;
[0025] The intent and the key points of the response are vectorized to obtain the second feature vector corresponding to the intent;
[0026] Calculate the cosine similarity between each of the first feature vectors and the second feature vectors;
[0027] The objective response data is generated based on the top-ranking qualification data according to the cosine similarity.
[0028] Compared to existing technologies, the above embodiments offer the following advantages: Converting intent and response key points into feature vectors and performing similarity matching with the vectors of qualification data in the database effectively captures semantic-level associations, rather than solely relying on keyword matching, thus avoiding retrieval omissions or errors caused by expression differences in traditional methods. Generating objective response results by filtering data with high similarity not only improves retrieval efficiency but also ensures the reliability and consistency of the results. Furthermore, by continuously updating the qualification database, the database maintains high adaptability of the qualification data, thereby continuously improving the comprehensiveness of subsequently generated tender documents.
[0029] Further, the step of constructing prompt words based on all the objective response data, all the intents, and all the response points, and inputting the prompt words into a preset large language model to obtain subjective response data, includes:
[0030] For each objective response data, a corresponding prompt word is constructed based on the corresponding intent and the key points of the response;
[0031] For each of the aforementioned prompt words, it is input into the large language model, and the large language model generates corresponding advantage analysis content;
[0032] The subjective response data is obtained based on all the advantages analyzed.
[0033] Compared to existing technologies, the above embodiments offer the following advantages: Based on objective response data, prompt words are constructed by combining intent and key response points. These prompt words are then input into a large language model to generate subjective response data, thereby solving the problems of rigidity and inflexibility inherent in traditional template-based responses. Furthermore, guided by prompt words, the large language model can generate more targeted solution descriptions, advantage analyses, and service commitments within contextual constraints, thus enhancing the professionalism and persuasiveness of subjective responses. Simultaneously, generating corresponding targeted prompt words for each response point and summarizing the advantage analyses of each response point ensures both targeted and comprehensive response coverage.
[0034] Another embodiment of this application also provides a tender document generation device, including: a tender document splitting module, an intent extraction module, an objective response data retrieval module, a subjective response data generation module, and a tender document generation module;
[0035] The tender document splitting module is used to receive the tender document and split it into several clauses to be responded to.
[0036] The intent extraction module is used to determine the intent of each pending response clause and extract the response points corresponding to the intent from the pending response clause.
[0037] The objective response data retrieval module is used to retrieve corresponding objective response data from a preset qualification database for each of the terms to be responded to, based on the corresponding intent and the key points of the response.
[0038] The subjective response data generation module is used to construct prompt words based on all the objective response data, all the intentions, and all the response points, and input the prompt words into a preset large language model to obtain subjective response data;
[0039] The tender document generation module is used to generate a tender document based on all the objective response data and the subjective response data.
[0040] Furthermore, the tender document splitting module includes: an OCR recognition unit, a text segmentation unit, and a clause classification unit; the tender document splitting module is used to receive the tender document and split the tender document into several clauses to be responded to, including:
[0041] The terms to be answered include: objective terms to be answered and subjective terms to be answered;
[0042] The OCR recognition unit is used to convert the tender document from a scanned document into text data using optical character recognition technology.
[0043] The text segmentation unit is used to segment the text data of the tender document into sentences to obtain several statements.
[0044] The clause classification unit is used to classify each statement into an objective clause to be responded to or a subjective clause to be responded to, using a preset classification model.
[0045] Further, the intent extraction module includes: an objective intent extraction unit and a subjective intent extraction unit; the intent extraction module is used to determine the intent of each pending response clause and extract response points corresponding to the intent from the pending response clauses, including:
[0046] The objective intent extraction unit is used to identify a first intent from the corresponding statement when the clause to be responded to is the objective clause to be responded to, and extract the corresponding slot information from the statement according to the first intent, and use the slot information as the response key point;
[0047] The subjective intent extraction unit is used to parse the second intent corresponding to the statement when the clause to be responded to is the subjective clause to be responded to, and to divide the second intent into several sub-intents. Based on the several sub-intents, the sub-intent sequence is labeled for each word segment in the corresponding statement, and the response points corresponding to each sub-intent are determined.
[0048] Further, the objective response data retrieval module includes: a vectorization unit, a cosine similarity calculation unit, and an objective response data generation unit; the objective response data retrieval module is used to retrieve corresponding objective response data from a preset qualification database based on the corresponding intent and response key points, including:
[0049] The qualification database includes: multiple qualification data and a first feature vector corresponding to each qualification data;
[0050] The vectorization unit is used to vectorize the intent and the response points to obtain a second feature vector corresponding to the intent;
[0051] The cosine similarity calculation unit is used to calculate the cosine similarity between each of the first feature vectors and the second feature vectors.
[0052] The objective response data generation unit is used to generate the objective response data based on the qualification data with the highest cosine similarity.
[0053] Furthermore, the subjective response data generation module includes: a prompt word construction unit, a large language model analysis unit, and an analysis result unification unit; the subjective response data generation module is used to construct prompt words based on all the objective response data, all the intentions, and all the response points, and input the prompt words into a preset large language model to obtain subjective response data, including:
[0054] The prompt word construction unit is used to construct corresponding prompt words for each objective response data according to its corresponding intent and response key points;
[0055] The large language model analysis unit is used to input each of the prompt words into the large language model and generate corresponding advantage analysis content through the large language model.
[0056] The unified analysis result unit is used to obtain the subjective response data based on all the advantages analysis content. Attached Figure Description
[0057] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0058] Figure 1 This is a flowchart illustrating a method for generating tender documents provided in some embodiments of this application;
[0059] Figure 2 This is a schematic diagram of the structure of a tender document generation device provided in some embodiments of this application. Detailed Implementation
[0060] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0061] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.
[0062] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.
[0063] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.
[0064] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.
[0065] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).
[0066] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.
[0067] Existing solutions improve bid generation efficiency by using automated tools to match keywords or fill in templates in tender documents. However, these methods struggle to truly understand the semantic intent of the clauses, leading to incorrect matching data and reducing the accuracy of the bid responses. Furthermore, when generating bids, subjective responses regarding solutions and service concepts are required. The automated tools available for these responses have limited templates and cannot flexibly adapt to the specific tender documents, resulting in bids with poor relevance and logical coherence.
[0068] Please refer to Figure 1 To address the issues of low accuracy in existing tender documents and poor relevance of responses to tender documents, this application provides a tender document generation method, comprising steps S101 to S105, specifically:
[0069] S101: Receive the tender documents and break them down into several clauses to be responded to.
[0070] Furthermore, in some embodiments of this application, the receiving of the tender documents, which involves splitting the tender documents into several pending response clauses, includes:
[0071] The terms to be answered include: objective terms to be answered and subjective terms to be answered;
[0072] The tender document was converted from a scanned copy into text data using optical character recognition technology.
[0073] The text data of the tender document is segmented into sentences to obtain several statements;
[0074] For each statement, it is classified into an objective clause to be responded to or a subjective clause to be responded to using a preset classification model.
[0075] Preferably, in some embodiments of this application, the step of converting the tender document from a scanned document into text data using optical character recognition technology includes: the optical character recognition technology is a deep learning-based OCR (Optical Character Recognition) engine used for text extraction from the scanned document; firstly, the title, paragraph, and table areas in the tender document are identified based on a U-Net segmentation network; further, the text data in each area is extracted using OCR technology; finally, a hybrid model combining regular expressions and BiLSTM-CRF (Bidirectional Long Short-Term Memory with Conditional Random Field) is used to extract the numbering hierarchy and list format from the text data in each area, thereby forming a hierarchical text list.
[0076] Preferably, in some embodiments of this application, the step of segmenting the text data of the tender document into several sentences includes: for each text data in the formed hierarchical text list, segmenting the text data into several sentences by recognizing punctuation marks in the text data.
[0077] Preferably, in some embodiments of this application, classifying each statement into an objective clause to be responded to or a subjective clause to be responded to using a preset classification model includes: performing word segmentation on each statement using the BERT (Bidirectional Encoder Representations from Transformers) model to obtain several words; determining whether there is a corresponding preset objective clause expression or preset subjective clause expression among the several words corresponding to each local area; wherein, the preset objective clause expression includes keywords such as "must provide proof of…", "not less than… quantity", "meets… standard", "not lower than… indicator", or "submit… document / certificate"; the preset subjective clause expression includes descriptions such as "please explain…", "need to describe…", "should elaborate…", "provide implementation plan / technical route", or "detailed service commitment"; statements with preset objective clause expressions are classified as objective clauses to be responded to, and statements with preset subjective clause expressions are classified as subjective clauses to be responded to.
[0078] This application utilizes optical character recognition (OCR) technology to automatically convert scanned tender documents into text, and further performs sentence segmentation and clause classification, effectively solving the problems of low efficiency and insufficient accuracy associated with manual tender document splitting. Furthermore, by employing a classification model, statements are distinguished into objective and subjective clauses, making subsequent intent recognition and responses more targeted, avoiding semantic interference caused by the mixing of clauses of different natures, and improving the accuracy of response generation.
[0079] S102: For each of the pending response clauses, determine the intent of the pending response clause and extract the response points corresponding to the intent from the pending response clauses.
[0080] Furthermore, in some embodiments of this application, the step of determining the intent of each pending response clause and extracting response points corresponding to the intent from the pending response clause includes:
[0081] When the clause to be responded to is the objective clause to be responded to, the first intent is identified from the corresponding statement, and the corresponding slot information is extracted from the statement according to the first intent, and the slot information is used as the response key point;
[0082] When the clause to be responded to is the subjective clause to be responded to, the second intent corresponding to the statement is parsed and the second intent is divided into several sub-intents. Based on the several sub-intents, the sub-intent sequence is labeled for each word in the corresponding statement, and the response points corresponding to each sub-intent are determined.
[0083] Preferably, in some embodiments of this application, when the clause to be responded to is the objective clause to be responded to, identifying the first intent from the corresponding statement and extracting the corresponding slot information from the statement according to the first intent includes: when each statement is segmented by the BERT model and it is determined that the statement contains a preset objective clause expression, the statement is further semantically parsed by the BERT model, and its intent classification is determined according to the semantic parsing result, thereby determining the first intent. For example, the statement "The bidder shall be equipped with no less than 5 senior engineers" contains the preset objective clause expression "no less than... number", and its semantics are parsed as a requirement for qualifications, and the first intent is determined to be "qualification requirements"; further, each word obtained after the statement is segmented by the CRF model is slot-labeled, such as "no less than 5", and finally the first intent and slot information pair {intent: "qualification requirements", slot: {experience years: "≥3 years"}} is obtained.
[0084] Preferably, in some embodiments of this application, when the clause to be responded to is the objective clause to be responded to, identifying the first intent from the corresponding statement and extracting the corresponding slot information from the statement according to the first intent includes: when each statement is segmented using a BERT model and it is determined that the statement contains a preset subjective clause expression, the statement is further semantically parsed using a BERT model, and its intent classification is determined based on the semantic parsing results to obtain a second intent. At this time, the second intent is the primary intent, and the second intent is subsequently decomposed to obtain several sub-intents. For example, the statement "The bidder needs to provide an implementation plan and explain the progress control and risk management measures" corresponds to the second intent "project management", and the decomposed sub-intents are {implementation plan, progress control, risk management}. For each sub-intent, the sub-intent is labeled using a CRF model to obtain the response points for each sub-intent, such as "progress control" corresponding to "explain progress control".
[0085] In objective clause scenarios, this application employs a slot information extraction method to accurately capture explicit requirements such as time, quantity, and qualifications, ensuring the accuracy and compliance of objective responses. In subjective clause scenarios, the intent is further decomposed into sub-intents, and key response points are identified based on sequence labeling methods. This not only enhances the depth of semantic understanding but also ensures comprehensive coverage of complex clauses. Through differentiated clause processing, the "generalization failure" problem of a single model for different clause types is effectively avoided. This ensures that the generated responses possess both precise matching of objective clauses and flexible expression and logical integrity of subjective clauses, thereby improving the overall accuracy and relevance of the tender document responses.
[0086] S103: For each of the aforementioned clauses to be responded to, retrieve the corresponding objective response data from a preset qualification database based on the corresponding intent and the key points of the response.
[0087] Furthermore, in some embodiments of this application, the step of retrieving corresponding objective response data from a preset qualification database based on the corresponding intent and response key points includes:
[0088] The qualification database includes: multiple qualification data and a first feature vector corresponding to each qualification data;
[0089] The intent and the key points of the response are vectorized to obtain the second feature vector corresponding to the intent;
[0090] Calculate the cosine similarity between each of the first feature vectors and the second feature vectors;
[0091] The objective response data is generated based on the top-ranking qualification data according to the cosine similarity.
[0092] Preferably, in some embodiments of this application, the step of vectorizing the intent and the response points to obtain a second feature vector corresponding to the intent includes: for the first intent of the objective clause to be responded to, using the BERT model for word segmentation, converting the first intent and the corresponding slot information into an embedded vector representation to obtain a second feature vector corresponding to the first intent; for the second intent of the subjective clause to be responded to, for each sub-intent corresponding to the second intent, using the BERT model, converting each sub-intent and the response points corresponding to the sub-intent into a corresponding embedded vector representation to obtain a second feature vector corresponding to each sub-intent.
[0093] Preferably, in some embodiments of this application, the qualification database consists of two databases: a bidding case database and an enterprise database. The bidding case database stores qualification data such as historical successful and failed bidding cases within the enterprise, while the enterprise database stores objective and quantifiable qualification data such as qualification certificates, test reports, contract performance records, and technical specifications. For each sub-intention corresponding to a second feature vector, the corresponding qualification data is retrieved from the bidding case database; for each first intention corresponding to a first feature vector, the corresponding qualification data is retrieved from the enterprise database.
[0094] Preferably, in some embodiments of this application, since the qualification data retrieved by the first intent is objective quantitative data, it needs to be preprocessed to obtain objective response data that conforms to the specifications. The specific steps are as follows: according to a preset template, the qualification data is filled into the template to obtain the objective response data. The qualification data retrieved from the bidding case database by each sub-intent does not require further processing.
[0095] This application converts intent and response key points into feature vectors and performs similarity matching with the vectors of qualification data in the database. This effectively captures semantic-level connections, rather than relying solely on keyword matching, avoiding retrieval omissions or errors caused by expression differences in traditional methods. By selecting data with high similarity to generate objective response results, not only is retrieval efficiency improved, but the reliability and consistency of the results are also guaranteed. Furthermore, by continuously updating the qualification database, the database maintains high adaptability of the qualification data, thereby continuously improving the comprehensiveness of subsequently generated tender documents.
[0096] S104: Based on all the objective response data, all the intents, and all the response points, construct prompt words and input the prompt words into a preset large language model to obtain subjective response data.
[0097] Further, the step of constructing prompt words based on all the objective response data, all the intents, and all the response points, and inputting the prompt words into a preset large language model to obtain subjective response data, includes:
[0098] For each objective response data, a corresponding prompt word is constructed based on the corresponding intent and the key points of the response;
[0099] For each of the aforementioned prompt words, it is input into the large language model, and the large language model generates corresponding advantage analysis content;
[0100] The subjective response data is obtained based on all the advantages analyzed.
[0101] Preferably, in some embodiments of this application, the step of constructing a corresponding prompt word for each objective response data according to its corresponding intent and response key points includes: for a first intent, concatenating the first intent, the slot information (and response key points) corresponding to the first intent, and the objective response data obtained through template filling to form a corresponding prompt word; for a second intent, concatenating each sub-intent, the response key points corresponding to the sub-intent, and the case data (and qualification data in the case database) retrieved according to the sub-intent to form a corresponding prompt word.
[0102] Preferably, in some embodiments of this application, each prompt word is input into a large language model to obtain the advantage analysis content corresponding to each prompt word, and all advantage analysis content is spliced together to obtain subjective response data.
[0103] This application, based on objective response data, constructs prompt words by combining intent and key response points, and inputs these prompt words into a large language model to generate subjective response data, thereby solving the problems of rigidity and lack of flexibility in traditional template-based responses. Furthermore, guided by prompt words, the large language model can generate more targeted solution descriptions, advantage analyses, and service commitments under contextual constraints, thus enhancing the professionalism and persuasiveness of subjective responses. Simultaneously, it generates corresponding targeted prompt words for each response point and summarizes the advantage analysis content of each response point, ensuring both targeted and comprehensive response coverage.
[0104] S105: Generate a tender document based on all the objective response data and the subjective response data.
[0105] Preferably, in some embodiments of this application, during the process of generating a tender document, according to a preset tender document template, the objective response data corresponding to the first intent is first filled into the corresponding area, then the advantage analysis content generated by inputting the prompt words corresponding to each first intent into the large language model is filled into the area after the corresponding objective response data, and finally the advantage analysis content generated according to each sub-intent is filled into the corresponding area of the template, thereby obtaining the final tender document.
[0106] Preferably, in some embodiments of this application, after the tender document is generated, the user proofreads the tender document, generates feedback results, and further uses the feedback results to optimize the preceding BERT model.
[0107] In summary, the bid document generation method provided in this application has the following advantages over existing technologies: First, by splitting and identifying the intent of the bid document, complex text content is transformed into structured clauses to be responded to, and key response points corresponding to the intent are extracted, ensuring the accuracy and relevance of clause parsing. Second, based on the identified intent and key response points, an automatic search is performed in a preset qualification database to obtain objective response data that meets the hard requirements of the bid document, thereby avoiding omissions and inefficiencies in the manual search process and improving the completeness and data accuracy of the bid document. Furthermore, when generating subjective content such as scheme descriptions, technical measures, or service commitments, by integrating the retrieved objective response data and key response points, prompt words are constructed and input into a large language model. This allows the large language model to generate more targeted subjective response data based on the objective response data that meets the hard requirements, ensuring that the final bid document closely conforms to the specific requirements of the bid document, and solving the problem of lack of flexibility and relevance in traditional template-based methods.
[0108] like Figure 2 As shown, based on the above-described method embodiments, an embodiment of this application provides a tender document generation device, including: a tender document splitting module 201, an intent extraction module 202, an objective response data retrieval module 203, a subjective response data generation module 204, and a tender document generation module 205.
[0109] Further, in some embodiments of this application, the tender document splitting module 201 is used to receive the tender document and split it into several clauses to be responded to; the intent extraction module 202 is used to determine the intent of each clause to be responded to and extract the response points corresponding to the intent from the clauses to be responded to; the objective response data retrieval module 203 is used to retrieve the corresponding objective response data from a preset qualification database for each clause to be responded to based on its corresponding intent and response points; the subjective response data generation module 204 is used to construct prompt words based on all the objective response data, all the intents, and all the response points, and input the prompt words into a preset large language model to obtain subjective response data; the tender document generation module 205 is used to generate a tender document based on all the objective response data and the subjective response data.
[0110] Further, in some embodiments of this application, the tender document splitting module 201 includes: an OCR recognition unit, a text segmentation unit, and a clause classification unit; the tender document splitting module 201 is used to receive a tender document and split the tender document into several clauses to be responded to, including: wherein the clauses to be responded to include: objective clauses to be responded to and subjective clauses to be responded to; the OCR recognition unit is used to convert the tender document from a scanned document into text data through optical character recognition technology; the text segmentation unit is used to segment the text data of the tender document into several sentences; the clause classification unit is used to classify each sentence into an objective clause to be responded to or a subjective clause to be responded to using a preset classification model.
[0111] Further, in some embodiments of this application, the intent extraction module 202 includes: an objective intent extraction unit and a subjective intent extraction unit; the intent extraction module 202 is used to determine the intent of each pending response clause and extract the response points corresponding to the intent from the pending response clause, including: the objective intent extraction unit is used to identify a first intent from the corresponding statement when the pending response clause is the pending response objective clause, and extract the corresponding slot information from the statement according to the first intent, and use the slot information as the response points; the subjective intent extraction unit is used to parse the second intent of the corresponding statement when the pending response clause is the pending response subjective clause, and divide the second intent into several sub-intents, and perform sub-intent sequence annotation on each word segment in the corresponding statement according to the several sub-intents, and determine the response points corresponding to each sub-intent.
[0112] Further, in some embodiments of this application, the objective response data retrieval module 203 includes: a vectorization unit, a cosine similarity calculation unit, and an objective response data generation unit; the objective response data retrieval module 203 is used to retrieve corresponding objective response data from a preset qualification database according to the corresponding intent and response points, including: wherein the qualification database includes: multiple qualification data and a first feature vector corresponding to each qualification data; the vectorization unit is used to vectorize the intent and the response points to obtain a second feature vector corresponding to the intent; the cosine similarity calculation unit is used to calculate the cosine similarity between each first feature vector and the second feature vector; the objective response data generation unit is used to generate the objective response data based on a plurality of qualification data with the highest cosine similarity.
[0113] Further, in some embodiments of this application, the subjective response data generation module 204 includes: a prompt word construction unit, a large language model analysis unit, and an analysis result unification unit; the subjective response data generation module 204 is used to construct prompt words based on all the objective response data, all the intentions, and all the response points, and input the prompt words into a preset large language model to obtain subjective response data, including: the prompt word construction unit is used to construct corresponding prompt words for each objective response data based on its corresponding intention and response points; the large language model analysis unit is used to input each prompt word into the large language model and generate corresponding advantage analysis content through the large language model; the analysis result unification unit is used to obtain the subjective response data based on all the advantage analysis content.
[0114] In summary, the tender document generation device provided in this application has the following advantages over the prior art: First, by splitting the tender document and identifying intent, the complex text content is transformed into structured clauses to be responded to, and the response points corresponding to the intent are extracted, ensuring the accuracy and relevance of the clause parsing. Second, based on the identified intent and response points, an automatic search is performed in a preset qualification database to obtain objective response data that meets the hard requirements of the tender document, thereby avoiding omissions and inefficiencies in the manual search process and improving the completeness and data accuracy of the tender document. Furthermore, when generating subjective content such as scheme descriptions, technical measures, or service commitments, by integrating the retrieved objective response data and response points, prompt words are constructed and input into a large language model, enabling the large language model to generate more targeted subjective response data based on the objective response data that meets the hard requirements, ensuring that the final generated tender document closely conforms to the specific requirements of the tender document, and solving the problem of lack of flexibility and relevance in traditional template-based methods.
[0115] It is understood that the above-described device embodiments correspond to the method embodiments of this application, and can implement the tender document generation method provided by any of the above-described method embodiments of this application.
[0116] It should be noted that the device embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided in this application, the connection relationships between modules indicate that they have communication connections, which can specifically be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.
[0117] Based on the above embodiments of the bid document generation method, another embodiment of this application provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the bid document generation method of any embodiment of this application.
[0118] For example, in this embodiment, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete this application. The one or more module units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the terminal device.
[0119] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.
[0120] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.
[0121] Based on the above-described method embodiments, another embodiment of this application provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the tender document generation method described in any of the above-described method embodiments of this application.
[0122] The modules / units integrated in the device / terminal equipment, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.
Claims
1. A method of generating a bid document, characterized by, The method comprises the following steps: receiving a bidding document, and splitting the bidding document into a plurality of to-be-answered clauses; for each to-be-answered clause, determining an intention of the to-be-answered clause, and extracting a response key corresponding to the intention from the to-be-answered clause; for each to-be-answered clause, searching a pre-set qualification database according to the intention and the response key corresponding to the to-be-answered clause to obtain corresponding objective response data; constructing a prompt word according to all the objective response data, all the intentions and all the response keys, and inputting the prompt word into a pre-set large language model to obtain subjective response data; generating a bidding document according to all the objective response data and the subjective response data.
2. The method of claim 1, wherein, The method comprises the following steps: receiving a bidding document, and splitting the bidding document into a plurality of to-be-answered clauses, including: wherein the to-be-answered clauses comprise to-be-answered objective clauses and to-be-answered subjective clauses; converting the bidding document from a scanned copy into text data through optical character recognition technology; performing sentence processing on the text data of the bidding document to obtain a plurality of sentences; 3. The method of claim 2, wherein the bid document is generated by: for each sentence, classifying it into the to-be-answered objective clauses or the to-be-answered subjective clauses through a pre-set classification model. The method comprises the following steps: when the to-be-answered clause is the to-be-answered objective clause, identifying a first intention from the corresponding sentence, and extracting corresponding slot information from the sentence according to the first intention, wherein the slot information is taken as the response key; 4. The method of claim 2, wherein the bid file is generated by: when the to-be-answered clause is the to-be-answered subjective clause, analyzing a second intention of the corresponding sentence, and dividing the second intention into a plurality of sub-intentions, wherein each sub-intention is used to perform sub-intention sequence labeling on each word in the corresponding sentence to determine the response key corresponding to each sub-intention. The method comprises the following steps: wherein the qualification database comprises a plurality of qualification data and a first feature vector corresponding to each qualification data; vectorizing the intention and the response key to obtain a second feature vector corresponding to the intention; calculating the cosine similarity of each first feature vector and the second feature vector; generating the objective response data according to a plurality of qualification data with higher cosine similarity.
5. The method of claim 1, wherein, The method comprises the following steps: for each objective response data, constructing a corresponding prompt word according to the intention and the response key corresponding to the objective response data; for each prompt word, inputting it into the large language model to generate corresponding advantage analysis content through the large language model; obtaining the subjective response data according to all the advantage analysis content.
6. A bid document generating apparatus characterized by comprising: The method comprises the following steps: The tender document splitting module, the intention extraction module, the objective answer data retrieval module, the subjective answer data generation module, and the bid document generation module; The tender document splitting module is configured to receive a tender document and split the tender document into a plurality of to-be-answered clauses. The intention extraction module is configured to determine an intention of each to-be-answered clause and extract an answer point corresponding to the intention from the to-be-answered clause. The objective answer data retrieval module is configured to retrieve corresponding objective answer data from a preset qualification database according to the intention and the answer point of each to-be-answered clause. The subjective answer data generation module is configured to construct a prompt word according to all the objective answer data, all the intentions, and all the answer points, input the prompt word into a preset large language model, and obtain subjective answer data. The bid document generation module is configured to generate a bid document according to all the objective answer data and the subjective answer data.
7. A bid document generating apparatus as claimed in claim 6, wherein The tender document splitting module includes an OCR recognition unit, a text sentence segmentation unit, and a clause classification unit. The to-be-answered clauses include to-be-answered objective clauses and to-be-answered subjective clauses. The OCR recognition unit is configured to convert the tender document from a scanned copy into text data through optical character recognition technology. The text sentence segmentation unit is configured to perform sentence segmentation processing on the text data of the tender document to obtain a plurality of sentences. The clause classification unit is configured to classify each sentence as a to-be-answered objective clause or a to-be-answered subjective clause through a preset classification model.
8. A bid document generating apparatus as claimed in claim 7, wherein, The intention extraction module includes an objective intention extraction unit and a subjective intention extraction unit. The objective intention extraction unit is configured to identify a first intention from the corresponding sentence and extract corresponding slot information from the sentence as an answer point according to the first intention when the to-be-answered clause is a to-be-answered objective clause. The subjective intention extraction unit is configured to parse a second intention of the corresponding sentence, divide the second intention into a plurality of sub-intentions, perform sub-intention sequence labeling on each word in the corresponding sentence according to the plurality of sub-intentions, and determine an answer point corresponding to each sub-intention when the to-be-answered clause is a to-be-answered subjective clause.
9. The bid document generating apparatus of claim 7 wherein, The objective answer data retrieval module includes a vectorization unit, a cosine similarity calculation unit, and an objective answer data generation unit. The vectorization unit is configured to convert the intention and the answer point into vectors. The cosine similarity calculation unit is configured to calculate the cosine similarity between the vectors. The objective answer data generation unit is configured to retrieve corresponding objective answer data from a preset qualification database according to the intention and the answer point. The qualification database comprises multiple qualification data and a first feature vector corresponding to each qualification data. The vectorization unit is configured to vectorize the intent and the response gist to obtain a second feature vector corresponding to the intent. The cosine similarity calculation unit is configured to calculate a cosine similarity between each first feature vector and the second feature vector. The objective response data generation unit is configured to generate the objective response data according to a plurality of qualification data with higher cosine similarity.
10. The bid document generating apparatus of claim 6 wherein, The subjective response data generation module comprises a prompt word construction unit, a large language model analysis unit, and an analysis result unification unit. The subjective response data generation module is configured to construct a prompt word according to all the objective response data, all the intents, and all the response gists, and input the prompt word into a preset large language model to obtain subjective response data, which comprises: The prompt word construction unit is configured to construct a corresponding prompt word for each objective response data according to the corresponding intent and the response gist thereof. The large language model analysis unit is configured to input each prompt word into the large language model to generate corresponding advantage analysis content through the large language model. The analysis result unification unit is configured to obtain the subjective response data according to all the advantage analysis content.