Bid text generation method and system based on large language model and storage medium
By building a historical material library and using a large language model to automatically filter and generate bid texts, the problem of insufficient personalization and intelligence of bid text generation in the existing technology is solved, and efficient and professional bid text generation is achieved.
Patent Information
- Application Number
- CN202411070765.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-06
- Publication Date
- 2024-11-19
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
When the existing technology automatically generates bid text, it is difficult to reflect the bidder's personalized needs, and the degree of intelligence and automation is limited, resulting in insufficient controllability of text generation and low efficiency.
By building a historical material library, including a first-level outline material library, a second-level outline material library and a paragraph material library, combined with bid evaluation methods and business bid technical evaluation requirements, a large language model is used to automatically filter and generate bid texts, and gradually generate first-level data Outline, secondary outline and paragraph text, and combine.
The automated, intelligent and controllable generation of bid texts has been realized, the quality and generation efficiency of texts have been improved, the professionalism and pertinence of the text have been ensured, and the intentions and project needs of the bidders are accurately reflected.
Smart Images

Figure CN118982007A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of electronic bidding and tendering, and in particular to a bidding text generation method, system and storage medium based on a large language model. Background Art
[0002] With the rapid development of artificial intelligence technology, large language models have shown powerful capabilities in various text generation tasks. In recent years, large language models based on deep learning, such as GPT-3 and BERT, have been widely used in many fields such as text translation, dialogue systems, and content creation. By learning language patterns from large corpora, large language models can generate natural, fluent, and semantically coherent text content, providing new ideas and tools for automated text generation in various industries.
[0003] In the bidding process, the writing of bid texts is a crucial step. The traditional way of writing bid texts mainly relies on manual work, which is not only inefficient but also difficult to ensure the consistency and accuracy of the texts. In order to improve efficiency and text quality, some companies have begun to try to use artificial intelligence technology, especially large language models, to assist in the generation of bid texts. Specifically, through preset templates and rules, combined with historical bidding data, try to automatically generate some text content.
[0004] However, when the current technology automatically generates bid texts, the generated texts are difficult to reflect the bidders' personalized needs and intentions, lacking pertinence, resulting in insufficient controllability of text generation. In addition, although preset templates and rules can be provided for large language models, their intelligence and automation are still limited and cannot meet the needs of writing bid texts as much as possible. Therefore, how to automatically, intelligently, and controllably generate bid texts to improve the quality and generation efficiency of bid texts is a difficult problem currently faced. Summary of the invention
[0005] This application provides a bidding text generation method, system and storage medium based on a large language model, which can automatically, intelligently and controllably generate bidding texts, thereby improving the quality and generation efficiency of bidding texts. This application provides the following technical solutions:
[0006] In a first aspect, the present application provides a bidding text generation method based on a large language model, the method comprising:
[0007] Collect historical bidding documents, and build a historical material library based on the historical bidding documents, wherein the historical material library includes a primary outline material library, a secondary outline material library, and a paragraph material library;
[0008] Obtaining a bid evaluation method, selecting the most matching first-level outline material from the first-level outline material library based on the bid evaluation method, inputting the bid evaluation method and the first-level outline material into a preset large language model, and outputting a first-level outline of the bidding text;
[0009] Obtaining the commercial bid and technical bid review requirements, selecting the most matching secondary outline material from the secondary outline material library based on the commercial bid and technical bid review requirements, inputting the primary outline of the bidding text, the commercial bid and technical bid review requirements, and the secondary outline material into the large language model, and outputting the secondary outline of the bidding text;
[0010] Acquire keywords, select the most matching paragraph materials from the paragraph material library based on the keywords, input the secondary outline of the bidding text, the keywords and the paragraph materials into the large language model, and output the paragraph text of the bidding document;
[0011] The first-level outline, the second-level outline and the paragraph text of the bidding document are combined to generate the final bidding text of the bidding document.
[0012] In a specific implementation scheme, the historical bidding documents are collected, and a historical material library is constructed based on the historical bidding documents. The historical material library includes a primary outline material library, a secondary outline material library, and a paragraph material library, including:
[0013] The first-level outline material library is used to store the main chapters or outline titles extracted from the historical bidding documents;
[0014] The secondary outline material library is used to store each small directory or sub-chapter title under the primary outline;
[0015] The paragraph material library is used to store specific paragraph texts, and each paragraph text in the paragraph material library is correspondingly configured with a tag word.
[0016] In a specific implementation scheme, the large language model is a ChatGPT model trained and generated for the bidding field, and the training and generation of the large language model includes:
[0017] Collect bidding text data;
[0018] Clean and pre-process the bidding text data;
[0019] Select the GPT model based on OpenAI as the basic model;
[0020] Use the processed bidding text data to adaptively train the selected basic model;
[0021] After multiple iterations of training, the model parameters are adjusted to generate a large language model.
[0022] In a specific implementation scheme, the method for obtaining bid evaluation, based on the bid evaluation method, screening out the most matching first-level outline material in the first-level outline material library includes:
[0023] Obtaining the bid evaluation method from the bidding document, extracting the important words in the bid evaluation method and their corresponding importance weights, calculating the matching degree between each important word and each first-level outline material in the first-level outline material library, adding the matching degrees of all important words with each first-level outline material in the first-level outline material library to obtain the final matching score corresponding to each first-level outline material, and the first-level outline material with the highest final matching score is the most matching first-level outline material;
[0024] The calculation formula for the final matching score MatchScore is as follows:
[0025]
[0026] Among them, w i is the weight of the ith important word, n is the number of important words, k i is the ith important word, Sim(k i ,D j ) represents the matching degree between the ith important word and the jth first-level outline material in the first-level outline material library; Sim(k i ,D j ) is calculated as follows:
[0027]
[0028] Among them, cos(k i ,D j ) represents the cosine similarity between the ith important word and the jth first-level outline material in the first-level outline material library, Indicates the total frequency of occurrence of the ith important word in the primary outline material library. It represents the total occurrence frequency of the i-th important word in the j-th first-level outline material in the first-level outline material library, and ∈ is the smoothing parameter.
[0029] In a specific possible implementation scheme, the obtaining of the commercial bid and technical bid review requirements, and screening out the most matching secondary outline material in the secondary outline material library based on the commercial bid and technical bid review requirements includes:
[0030] Obtain the commercial and technical standards review requirements, extract the main text of the commercial and technical standards review requirements, and calculate the matching degree S between the main text of the commercial and technical standards review requirements and each secondary outline material in the secondary outline material library based on vocabulary and semantic information. The calculation formula is as follows:
[0031] S = α·CosSim(Q,E m )+(1-α)·SemanticSim(Q,E m )
[0032] Among them, Q is the vector representation of the main text of the commercial and technical standards review requirements, E m is the vector representation of the mth secondary outline material in the secondary outline material library, CosSim(Q,E m ) represents Q and E m The cosine similarity value of SemanticSim(Q,E m ) represents Q and E calculated using the existing word vector model m The semantic similarity of , α is the parameter that weighs the cosine similarity and semantic similarity;
[0033] The secondary outline material with the highest calculated matching degree is the most matching secondary outline material.
[0034] In a specific implementation scheme, the obtaining of keywords and selecting the most matching paragraph material in the paragraph material library based on the keywords includes:
[0035] By searching in the paragraph material library to determine whether there is a tag word that is the same as the keyword, if so, the paragraph material corresponding to the tag word is used as the most matching paragraph material; if not, the following formula is used to calculate the matching degree between the keyword and each paragraph material:
[0036] Matching degree = tag word matching degree + paragraph text matching degree
[0037] The calculation formula for label word matching is as follows:
[0038]
[0039] Among them, x is the number of keywords, y is the number of label words, sim(K q ,L u ) is the keyword K q and label word L u The cosine similarity value of ;
[0040] The formula for calculating paragraph text matching is as follows:
[0041]
[0042] Among them, z is the number of words in the paragraph text, sim(K q ,P r ) is the keyword K q and paragraph text word P r The cosine similarity value of ;
[0043] After calculating the matching degree between the keyword and each paragraph material, the paragraph material with the greatest matching degree is the most matching paragraph material.
[0044] In a specific possible implementation scheme, combining the primary outline, the secondary outline and the paragraph text of the bidding document to generate the final bidding text of the bidding document includes:
[0045] Obtain the first-level outline, second-level outline and paragraph text of the generated bidding text;
[0046] Insert each first-level outline as a main chapter title into the preset document in the order of the first-level outline;
[0047] Under each primary outline, insert the corresponding secondary outline in order;
[0048] Under each secondary outline, insert the corresponding paragraph text in sequence to generate the final bid text.
[0049] In the second aspect, the present application provides a bidding text generation system based on a large language model, which adopts the following technical solutions:
[0050] A bidding text generation system based on a large language model, comprising:
[0051] A material library construction module, used to collect historical bidding documents and construct a historical material library based on the historical bidding documents, wherein the historical material library includes a primary outline material library, a secondary outline material library and a paragraph material library;
[0052] A first-level outline generation module, used to obtain a bid evaluation method, select the most matching first-level outline material from the first-level outline material library based on the bid evaluation method, input the bid evaluation method and the first-level outline material into a preset large language model, and output a first-level outline of the bidding text;
[0053] A secondary outline generation module is used to obtain the commercial bid and technical bid review requirements, select the most matching secondary outline material from the secondary outline material library based on the commercial bid and technical bid review requirements, input the primary outline of the bidding text, the commercial bid and technical bid review requirements and the secondary outline material into the large language model, and output the secondary outline of the bidding text;
[0054] A paragraph text generation module is used to obtain keywords, select the most matching paragraph materials in the paragraph material library based on the keywords, input the secondary outline of the bidding text, the keywords and the paragraph materials into the large language model, and output the paragraph text of the bidding document;
[0055] The bidding text generation module is used to combine the primary outline, secondary outline and paragraph text of the bidding document to generate the final bidding text of the bidding document.
[0056] In a third aspect, the present application provides an electronic device, comprising a processor and a memory; the memory stores a program, and the program is loaded and executed by the processor to implement a bidding text generation method based on a large language model as described in the first aspect.
[0057] In a fourth aspect, the present application provides a computer-readable storage medium, wherein the storage medium stores a program, and when the program is executed by a processor, the program is used to implement a bidding text generation method based on a large language model as described in the first aspect.
[0058] By building a historical material library and combining the bid evaluation method and the requirements for the commercial and technical bid evaluation, the most matching materials are selected from the first-level outline, second-level outline and paragraph material library. Then, these materials are input into the large language model to gradually generate and combine the complete bidding text. By building the first-level outline, second-level outline and paragraph material library and using the large language model to automatically select and generate text, manual intervention is reduced, and the full process automation of bidding text generation is realized, thereby greatly improving the generation efficiency. Applying the preset large language model, according to the bid evaluation method and the requirements for the commercial and technical bid evaluation, the most matching materials are intelligently selected, and a bidding text that highly meets the project requirements and the bidder's intentions is generated, ensuring the professionalism and pertinence of the text. The keyword matching mechanism and preset templates are introduced to ensure that the generated text content is highly consistent with the key content provided by the bidder. By gradually generating the first-level outline, second-level outline and paragraph text, fine control of the generation process is achieved to ensure that the final generated bidding text accurately reflects the personalized needs of the bidder.
[0059] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application and implement it according to the contents of the specification, the following is a detailed description of the preferred embodiments of the present application in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0060] Figure 1 It is a flowchart of a bidding text generation method based on a large language model in an embodiment of the present application.
[0061] Figure 2 It is a flowchart of the training generation of a large language model in an embodiment of the present application.
[0062] Figure 3 It is a schematic diagram of the overall process of the bidding text generation method based on the large language model in the embodiment of the present application.
[0063] Figure 4It is a structural block diagram of a bidding text generation system based on a large language model in an embodiment of the present application.
[0064] Figure 5 It is a block diagram of an electronic device for generating bid text based on a large language model in an embodiment of the present application. DETAILED DESCRIPTION
[0065] The specific implementation methods of the present application are further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present application but are not intended to limit the scope of the present application.
[0066] Optionally, the present application takes the bidding text generation method based on the large language model provided in each embodiment as an example used in an electronic device, where the electronic device is a terminal or a server. The terminal may be a computer, a tablet computer, etc. This embodiment does not limit the type of electronic device.
[0067] Reference Figure 1 , is a flow chart of a method for generating bidding text based on a large language model provided by an embodiment of the present application, the method comprising at least the following steps:
[0068] Step S101: collect historical bidding documents, and build a historical material library based on the historical bidding documents. The historical material library includes a first-level outline material library, a second-level outline material library, and a paragraph material library.
[0069] Step S102, obtaining a bid evaluation method, selecting the most matching first-level outline material from the first-level outline material library based on the bid evaluation method, inputting the bid evaluation method and the first-level outline material into a preset large language model, and outputting the first-level outline of the bid text.
[0070] Step S103: Obtain the commercial bid and technical bid review requirements, select the most matching secondary outline material from the secondary outline material library based on the commercial bid and technical bid review requirements, input the primary outline of the bidding text, the commercial bid and technical bid review requirements, and the secondary outline material into the large language model, and output the secondary outline of the bidding text.
[0071] Step S104: Obtain keywords, select the most matching paragraph materials from the paragraph material library based on the keywords, input the secondary outline, keywords and paragraph materials of the bidding text into the large language model, and output the paragraph text of the bidding document.
[0072] Step S105: Combine the primary outline, secondary outline and paragraph text of the bidding document to generate the final bidding text of the bidding document.
[0073] In step S101, historical bidding documents are first collected. Optionally, historical bidding documents can be collected from historical bidding projects disclosed by the enterprise or other public bidding databases. This application does not limit the collection method of historical bidding documents. After collection, the historical bidding documents are pre-processed to remove irrelevant content, duplicate files, and files with irregular formats. Then, a historical material library is constructed based on the historical bidding documents. The historical material library includes a first-level outline material library, a second-level outline material library, and a paragraph material library. The first-level outline material library, the second-level outline material library, and the paragraph material library are as follows:
[0074] First-level outline material library: used to store the main chapters or outline titles extracted from historical bidding documents. Usually high-level topics such as "Project Overview", "Technical Solution", "Business Conditions", etc.
[0075] Secondary outline material library: used to store the sub-directories or sub-chapter titles under the primary outline. Usually it is a detailed breakdown of the primary outline content, such as "system architecture design" and "technical implementation method" under "technical solution".
[0076] Paragraph material library: used to store specific paragraph text. The text is a specific description and explanation of the secondary outline content, including specific technical details, business terms, etc.
[0077] In the implementation, firstly, natural language processing technology is used to automatically extract the main chapters or outline titles from the historical bidding documents, and several groups of keywords and phrases are defined, such as "project overview", "technical solution", "business conditions", etc. Several main chapters or outline titles are extracted from the historical bidding documents by searching and matching the defined keywords and phrases, and the extracted main chapters or outline titles are stored in the first-level outline material library. Secondly, the content under each first-level outline is analyzed in detail, the small directory or sub-chapter title is identified, and the hierarchical relationship and subordinate relationship of the second-level outline are automatically identified by using regular expressions, and the identified second-level outline is classified according to the first-level outline to which it belongs, and stored in the second-level outline material library. Finally, the paragraph segmentation algorithm is used to further refine the content under each second-level outline into specific paragraph texts, and the paragraph texts are stored in the paragraph material library. Through the above steps, the historical material library composed of historical bidding documents can be systematically divided into a first-level outline material library, a second-level outline material library and a paragraph material library. It should be noted that the natural language processing technology, regular expressions and paragraph segmentation algorithms in the above content are all applications of existing technologies, so they are not described in detail.
[0078] In addition, as a preference, each paragraph text in the paragraph material library is configured with a corresponding label word. For example, the paragraph text is "the sensor is set on the traction device, the sensor is connected to the module processor and the alarm in sequence, the alarm is fixed on the platform body, the load of the unloading platform is managed intelligently, and the early warning system is combined to provide timely early warning of overload conditions", and its label word corresponds to "sensor".
[0079] Optionally, this application uses a domain-fine-tuned and optimized summary model to automatically tag each paragraph text, and other existing technologies can also be used to implement the tagging operation. This application does not limit the tagging method.
[0080] In step S102, first obtain a detailed bid evaluation method from the bidding document, extract important words and their corresponding importance weights in the bid evaluation method according to the bid evaluation experts or natural language processing tool library, then calculate the matching degree between each important word and each first-level outline material in the first-level outline material library, and then add the matching degrees of all important words and each first-level outline material in the first-level outline material library to obtain the final matching score corresponding to each first-level outline material, and the one with the highest final matching score is the most matching first-level outline material. The calculation formula of the final matching score MatchScore is as follows:
[0081]
[0082] Among them, w i is the weight of the ith important word, indicating its importance in the bid evaluation method. It should be noted that the weights of all important words are added together to form one, n is the number of important words, and k is i is the ith important word, Sim(k i ,D j ) represents the matching degree between the ith important word and the jth first-level outline material in the first-level outline material library. Sim(k i ,D j ) is calculated as follows:
[0083]
[0084] Among them, cos(k i ,D j ) represents the cosine similarity between the ith important word and the jth first-level outline material in the first-level outline material library, Indicates the total frequency of occurrence of the ith important word in the primary outline material library. Represents the total frequency of occurrence of the i-th important word in the j-th first-level outline material in the first-level outline material library. ∈ is a smoothing parameter, which is usually set to 0.001. The weight in the bid evaluation method can accurately reflect the importance of each keyword in the whole. This ensures that the evaluation is not just a simple keyword frequency, but is based on the subjective judgment of the expert on the importance of the keyword. When calculating the matching degree, not only the frequency of occurrence and overall frequency of the keyword in the outline material library are considered, but also the semantic similarity between the keyword and the outline material. Such a comprehensive consideration is closer to the actual text semantic matching situation and avoids the limitations of being based solely on word frequency. The introduction of the smoothing parameter can effectively handle the situation where the keyword does not appear in some outline materials. This ensures that even if a keyword does not appear in a certain material, it will not lead to a zero overall matching degree, thereby improving the stability and accuracy of the evaluation. The final matching score combines the weights of all important words and their matching degree with each outline material. In this way, the degree of compliance of each outline material with the bid evaluation method can be comprehensively evaluated. This helps the tenderer to more accurately select the bidding document or plan that best matches the demand.
[0085] In one embodiment, the large language model in this application is a ChatGPT model trained and generated for the bidding field, referring to Figure 2 , is a flow chart of training and generating a large language model in an embodiment of the present application, and the method includes at least the following steps:
[0086] S1021. Collect bidding text data.
[0087] Specifically, the bidding text data includes historical bidding documents, tender documents, bid evaluation methods, commercial and technical bid evaluation requirements, and bid paragraph texts.
[0088] S1022. Clean and pre-process the bidding text data.
[0089] Specifically, the collected text data is cleaned, including removal of irrelevant information, formatting adjustment, tokenization (such as word segmentation, part-of-speech tagging), etc. Ensure the quality and consistency of the data to improve the effect and stability of model training.
[0090] S1023. Select the GPT model based on OpenAI as the basic model.
[0091] S1024. Use the processed bidding text data to perform adaptive training on the selected basic model.
[0092] S1025. After multiple iterations of training, the model parameters are adjusted to generate a large language model.
[0093] Optionally, the present application may also directly select the existing ChatGPT model as the large language model.
[0094] During implementation, after selecting the most matching first-level outline material from the first-level material library, the bid evaluation method and the first-level outline material are input into the large language model, and the large language model automatically outputs the first-level outline of the bidding text.
[0095] In step S103, the commercial bid technical bid review requirements are first obtained, and the main text of the commercial bid technical bid review requirements is extracted according to the bid evaluation experts or the natural language processing tool library. Then, the matching degree S between the main text of the commercial bid technical bid review requirements and each secondary outline material in the secondary outline material library is calculated based on vocabulary and semantic information. The calculation formula is as follows:
[0096] S = α·CosSim(Q,E m )+(1-α)·SemanticSim(Q,E m )
[0097] Among them, Q is the vector representation of the main text of the commercial and technical standards review requirements, E m is the vector representation of the mth secondary outline material in the secondary outline material library, CosSim(Q,E m ) represents Q and E m The cosine similarity value of SemanticSim(Q,E m ) represents Q and E calculated using the existing word vector model m The semantic similarity of the two models is usually the same, such as Word2Vec or BERT. α is a parameter that weighs cosine similarity and semantic similarity, and its value is between 0 and 1.
[0098] In implementation, the second-level outline material with the highest degree of match is calculated as the best matching second-level outline material. After selecting the best matching second-level outline material from the second-level outline material library, the first-level outline of the bidding text, the commercial and technical standards review requirements, and the second-level outline materials are input into the large language model, which refines the first-level outline and outputs the second-level outline of the bidding text that meets the commercial and technical standards review requirements.
[0099] Optionally, the method in step S102 may be used to filter out the best matching secondary outline material, or the method in step S103 may be used to filter out the best matching primary outline material, which will not be described here.
[0100] In step S104, keywords are obtained. Keywords are usually provided by bidders or extracted from bidding documents. Then, based on the keywords, the most matching paragraph materials are screened out in the paragraph material library. Specifically, firstly, it is determined whether there is a tag word with the same keyword in the paragraph material library. If there is, the paragraph material corresponding to the tag word is used as the most matching paragraph material. If not, the matching degree between the keyword and each paragraph material is calculated using the following formula:
[0101] Matching degree = tag word matching degree + paragraph text matching degree
[0102] The calculation formula for label word matching is as follows:
[0103]
[0104] Among them, x is the number of keywords, y is the number of label words, sim(K q ,L u ) is the keyword K q and label word L u The cosine similarity value of .
[0105] The formula for calculating paragraph text matching is as follows:
[0106]
[0107] Among them, z is the number of words in the paragraph text, sim(K q ,P r ) is the keyword K q and paragraph text word P r The cosine similarity value of . After calculating the matching degree between the keyword and each paragraph material, the paragraph material with the greatest matching degree is the most matching paragraph material. By considering both the tag word matching degree and the paragraph text matching degree, the matching degree between the keyword and the paragraph material can be measured more comprehensively. The tag word matching degree reflects the consistency between the keyword and the main theme of the paragraph, while the paragraph text matching degree measures the matching degree between the keyword and the specific details of the paragraph content. This dual matching degree calculation method can improve the matching accuracy and avoid relying solely on the single perspective brought by the tag word or text content, so as to find the most matching paragraph material more accurately.
[0108] During implementation, after selecting the most matching paragraph material from the paragraph material library, the secondary outline, keywords and paragraph materials of the bidding text are input into the large language model. The large language model refines the secondary outline and outputs the paragraph text of the bidding document that meets the keyword requirements.
[0109] In step S105, the first-level outline, second-level outline and paragraph text of the generated bidding text are obtained, and each first-level outline is inserted into the preset document as the main chapter title in the order of the first-level outline. Under each first-level outline, the corresponding second-level outline is inserted in sequence, and under each second-level outline, the corresponding paragraph text is inserted in sequence to generate the final bidding text. Finally, the generated bidding text is formally checked and corrected using a document checking tool.
[0110] In summary, refer to Figure 3 , the overall process diagram of the bidding text generation method based on the large language model in the embodiment of the present application, the present application proposes a method for intelligently generating bidding texts based on historical bidding documents and large language models, by constructing a historical material library, and combining the bid evaluation method and the commercial bid technical standard review requirements, the most matching materials are screened out in the first-level outline, second-level outline and paragraph material library. Then, these materials are input into the large language model, and the complete bidding text is gradually generated and combined. By constructing the first-level outline, second-level outline and paragraph material library, and using the large language model to automatically screen and generate texts, manual intervention is reduced, and the full process automation of bidding text generation is realized, thereby greatly improving the generation efficiency. Applying the preset large language model, according to the bid evaluation method and the commercial bid technical standard review requirements, the most matching materials are intelligently screened out, and a bidding text that highly meets the project requirements and the bidder's intentions is generated, ensuring the professionalism and pertinence of the text. The keyword matching mechanism and preset template are introduced to ensure that the generated text content is highly consistent with the key content provided by the bidder. By gradually generating the first-level outline, second-level outline and paragraph text, fine control of the generation process is achieved, ensuring that the final generated bidding text accurately reflects the personalized needs of the bidder.
[0111] Figure 4 This is a structural block diagram of a bidding text generation system based on a large language model provided by an embodiment of the present application. The device includes at least the following modules:
[0112] The material library construction module is used to collect historical bidding documents and build a historical material library based on the historical bidding documents. The historical material library includes a first-level outline material library, a second-level outline material library and a paragraph material library.
[0113] The first-level outline generation module is used to obtain the bid evaluation method, screen out the most matching first-level outline material in the first-level outline material library based on the bid evaluation method, input the bid evaluation method and the first-level outline material into the preset large language model, and output the first-level outline of the bidding text.
[0114] The secondary outline generation module is used to obtain the commercial bid and technical bid review requirements, screen out the most matching secondary outline materials in the secondary outline material library based on the commercial bid and technical bid review requirements, input the primary outline of the bidding text, the commercial bid and technical bid review requirements and the secondary outline materials into the large language model, and output the secondary outline of the bidding text.
[0115] The paragraph text generation module is used to obtain keywords, filter out the most matching paragraph materials in the paragraph material library based on the keywords, input the secondary outline, keywords and paragraph materials of the bidding text into the large language model, and output the paragraph text of the bidding document.
[0116] The bidding text generation module is used to combine the primary outline, secondary outline and paragraph text of the bidding document to generate the final bidding text of the bidding document.
[0117] For relevant details, refer to the above method embodiment.
[0118] Figure 5 5 is a block diagram of an electronic device provided by an embodiment of the present application. The device at least includes a processor 501 and a memory 502.
[0119] The processor 501 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 501 may be implemented in at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 501 may also include a main processor and a coprocessor. The main processor is a processor for processing data in the awake state, also known as a CPU (Central Processing Unit); the coprocessor is a low-power processor for processing data in the standby state. In some embodiments, the processor 501 may be integrated with a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 501 may also include an AI (Artificial Intelligence) processor, which is used to process computing operations related to machine learning.
[0120] The memory 502 may include one or more computer-readable storage media, which may be non-transitory. The memory 502 may also include a high-speed random access memory, and a non-volatile memory, such as one or more disk storage devices, flash memory storage devices. In some embodiments, the non-transitory computer-readable storage medium in the memory 502 is used to store at least one instruction, which is used to be executed by the processor 501 to implement the bidding text generation method based on a large language model provided in the method embodiment of the present application.
[0121] In some embodiments, the electronic device may further optionally include: a peripheral device interface and at least one peripheral device. The processor 501, the memory 502 and the peripheral device interface may be connected via a bus or a signal line. Each peripheral device may be connected to the peripheral device interface via a bus, a signal line or a circuit board. Schematically, the peripheral devices include but are not limited to: a radio frequency circuit, a touch display screen, an audio circuit, and a power supply.
[0122] Of course, the electronic device may also include fewer or more components, which is not limited in this embodiment.
[0123] Optionally, the present application also provides a computer-readable storage medium, in which a program is stored, and the program is loaded and executed by a processor to implement the bidding text generation method based on a large language model of the above method embodiment.
[0124] Optionally, the present application also provides a computer product, which includes a computer-readable storage medium, in which a program is stored, and the program is loaded and executed by a processor to implement the bidding text generation method based on a large language model of the above method embodiment.
[0125] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0126] The above embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.
Claims
1. A bidding text generation method based on a large language model, characterized in that: The method comprises: Collect historical bidding documents, and build a historical material library based on the historical bidding documents, wherein the historical material library includes a primary outline material library, a secondary outline material library, and a paragraph material library; Obtaining a bid evaluation method, selecting the most matching first-level outline material from the first-level outline material library based on the bid evaluation method, inputting the bid evaluation method and the first-level outline material into a preset large language model, and outputting a first-level outline of the bidding text; Obtaining the commercial bid and technical bid review requirements, selecting the most matching secondary outline material from the secondary outline material library based on the commercial bid and technical bid review requirements, inputting the primary outline of the bidding text, the commercial bid and technical bid review requirements, and the secondary outline material into the large language model, and outputting the secondary outline of the bidding text; Acquire keywords, select the most matching paragraph materials from the paragraph material library based on the keywords, input the secondary outline of the bidding text, the keywords and the paragraph materials into the large language model, and output the paragraph text of the bidding document; The first-level outline, the second-level outline and the paragraph text of the bidding document are combined to generate the final bidding text of the bidding document.
2. The bidding text generation method based on a large language model according to claim 1 is characterized in that: The collecting of historical bidding documents and building of a historical material library based on the historical bidding documents, wherein the historical material library includes a primary outline material library, a secondary outline material library and a paragraph material library, including: The first-level outline material library is used to store the main chapters or outline titles extracted from the historical bidding documents; The secondary outline material library is used to store each small directory or sub-chapter title under the primary outline; The paragraph material library is used to store specific paragraph texts, and each paragraph text in the paragraph material library is correspondingly configured with a tag word.
3. The bidding text generation method based on a large language model according to claim 1 is characterized in that: The large language model is a ChatGPT model trained and generated for the bidding field. The training and generation of the large language model includes: Collect bidding text data; Clean and pre-process the bidding text data; Select the GPT model based on OpenAI as the basic model; Use the processed bidding text data to adaptively train the selected basic model; After multiple iterations of training, the model parameters are adjusted to generate a large language model.
4. The bidding text generation method based on a large language model according to claim 2 is characterized in that: The method for obtaining bid evaluation, based on the bid evaluation method, selects the most matching first-level outline material from the first-level outline material library, including: Obtaining the bid evaluation method from the bidding document, extracting the important words in the bid evaluation method and their corresponding importance weights, calculating the matching degree between each important word and each first-level outline material in the first-level outline material library, adding the matching degrees of all important words with each first-level outline material in the first-level outline material library to obtain the final matching score corresponding to each first-level outline material, and the first-level outline material with the highest final matching score is the most matching first-level outline material; The calculation formula for the final matching score MatchScore is as follows: Among them, w i is the weight of the ith important word, n is the number of important words, k i is the ith important word, Sim(k i ,D j ) represents the matching degree between the ith important word and the jth first-level outline material in the first-level outline material library; Sim(k i ,D j ) is calculated as follows: Among them, cos(k i ,D j ) represents the cosine similarity between the ith important word and the jth first-level outline material in the first-level outline material library, Indicates the total frequency of occurrence of the ith important word in the primary outline material library. It represents the total occurrence frequency of the i-th important word in the j-th first-level outline material in the first-level outline material library, and ∈ is the smoothing parameter.
5. The bidding text generation method based on a large language model according to claim 2 is characterized in that: The obtaining of the commercial and technical standards review requirements and screening out the most matching secondary outline material in the secondary outline material library based on the commercial and technical standards review requirements include: Obtain the commercial and technical standards review requirements, extract the main text of the commercial and technical standards review requirements, and calculate the matching degree S between the main text of the commercial and technical standards review requirements and each secondary outline material in the secondary outline material library based on vocabulary and semantic information. The calculation formula is as follows: S=α·CosSim(Q,E m )+(1-α) SemanticSim(Q,E m ) Among them, Q is the vector representation of the main text of the commercial and technical standards review requirements, E m is the vector representation of the mth secondary outline material in the secondary outline material library, CosSim(Q,E m ) represents Q and E m The cosine similarity value of SemanticSim(Q,E m ) represents Q and E calculated using the existing word vector model m The semantic similarity of , α is the parameter that weighs the cosine similarity and semantic similarity; The secondary outline material with the highest calculated matching degree is the most matching secondary outline material.
6. The bidding text generation method based on a large language model according to claim 2 is characterized in that: The obtaining of keywords and selecting the most matching paragraph material in the paragraph material library based on the keywords includes: By searching in the paragraph material library to determine whether there is a tag word that is the same as the keyword, if so, the paragraph material corresponding to the tag word is used as the most matching paragraph material; if not, the following formula is used to calculate the matching degree between the keyword and each paragraph material: Matching degree = tag word matching degree + paragraph text matching degree The calculation formula for label word matching is as follows: Among them, x is the number of keywords, y is the number of label words, sim(K q ,L u ) is the keyword K q and label word L u The cosine similarity value of ; The formula for calculating paragraph text matching is as follows: Among them, z is the number of words in the paragraph text, sim(K q ,P r ) is the keyword K q and paragraph text word P r The cosine similarity value of ; After calculating the matching degree between the keyword and each paragraph material, the paragraph material with the greatest matching degree is the most matching paragraph material.
7. The bidding text generation method based on a large language model according to claim 1 is characterized in that: The step of combining the primary outline, the secondary outline and the paragraph text of the bidding document to generate the final bidding text of the bidding document comprises: Obtain the first-level outline, second-level outline and paragraph text of the generated bidding text; Insert each first-level outline as a main chapter title into the preset document in the order of the first-level outline; Under each primary outline, insert the corresponding secondary outline in order; Under each secondary outline, insert the corresponding paragraph text in sequence to generate the final bid text.
8. A bidding text generation system based on a large language model, characterized in that: include: A material library construction module, used to collect historical bidding documents and construct a historical material library based on the historical bidding documents, wherein the historical material library includes a primary outline material library, a secondary outline material library and a paragraph material library; A first-level outline generation module, used to obtain a bid evaluation method, select the most matching first-level outline material from the first-level outline material library based on the bid evaluation method, input the bid evaluation method and the first-level outline material into a preset large language model, and output a first-level outline of the bidding text; A secondary outline generation module is used to obtain the commercial bid and technical bid review requirements, select the most matching secondary outline material from the secondary outline material library based on the commercial bid and technical bid review requirements, input the primary outline of the bidding text, the commercial bid and technical bid review requirements and the secondary outline material into the large language model, and output the secondary outline of the bidding text; A paragraph text generation module is used to obtain keywords, select the most matching paragraph materials in the paragraph material library based on the keywords, input the secondary outline of the bidding text, the keywords and the paragraph materials into the large language model, and output the paragraph text of the bidding document; The bidding text generation module is used to combine the primary outline, secondary outline and paragraph text of the bidding document to generate the final bidding text of the bidding document.
9. An electronic device, characterized in that: The device includes a processor and a memory; a program is stored in the memory, and the program is loaded and executed by the processor to implement a bidding text generation method based on a large language model as described in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that: The storage medium stores a program, which, when executed by a processor, is used to implement a bidding text generation method based on a large language model as described in any one of claims 1 to 7.
Citation Information
Patent Citations
Bidding book generation method and device, equipment and storage medium
CN117150000A
Method and device for making, checking and pre-evaluating bidding scheme based on text recognition
CN117635283A
Method for creating text, device, apparatus, and storage medium
JP2021106017A
Rich-media document auxiliary generation apparatus
WO2024131054A1
Cited By
Bidding scheme generation method, bidding scheme generation system and computer equipment
CN119670717A
Official document generation method and system based on artificial intelligence
CN119886100A
Material salvage method based on modal fusion matching
CN120047186A
Bidding file generation method and device, equipment and storage medium
CN120068838A
Document outline generation method and system, terminal and medium
CN120911406A
