Template-based bidding document generation method, system and storage medium
Through the combination of vectorized matching and deep learning models, the fields to be filled in are automatically identified and backfilled in bidding documents, the problem of bidding documents reliance on manual operations in the existing technology is solved, and more efficient and accurate bidding documents are achieved.
Patent Information
- Application Number
- CN202411038044.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-31
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2044-07-31
AI Technical Summary
The existing bidding document generation method based on template templates relies on manual operations, which is time-consuming and error-prone, making it difficult to ensure accuracy.
The most matching template template is filtered through vector matching, identify the fields to be filled and classified, and use the BM25 search model and deep learning model to extract the fields directly filled, and summarize and expand the automatic backfill of the fields filled, improving the generation efficiency and accuracy.
It reduces the risk of human error, improves the accuracy and efficiency of tender document generation, reduces manual labor, and improves the automation level of text generation.
Smart Images

Figure CN118551748B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of electronic bidding and tendering, and in particular to a method, system and storage medium for generating bidding documents based on template examples. Background Art
[0002] As a core component of the bidding process, the quality of the bidding document can directly affect the execution effect and results of the project. In recent years, with the rapid development of information technology and natural language processing technology, the automatic generation of bidding documents has gradually become possible.
[0003] At present, there are two main forms of tender document generation: one is template-based tender document generation, and the other is template-free tender document generation. Template-based tender document generation refers to generating complete tender documents by filling in corresponding specific information into the template based on the existing template that meets the tender specifications. Template-free tender document generation needs to be written from scratch, which completely depends on the professional knowledge and experience of the writer. For template-based tender document generation, the existing method is based on a pre-established template, and the staff fills in specific fields and content according to the actual needs of the project to generate the tender document.
[0004] However, the existing method of generating bidding documents based on templates mainly relies on professional staff, which involves a lot of manual labor and meticulous attention. Staff need to have an in-depth understanding of the specific requirements of the project, contract terms, and technical specifications, which is not only time-consuming but also prone to errors. Therefore, how to develop an intelligent bidding document generation method based on templates to improve the accuracy of bidding text generation and reduce the risk of human errors has become a key direction of current technological development. Summary of the invention
[0005] The present application provides a method, system and storage medium for generating bidding documents based on templates, which can improve the accuracy of generating bidding documents and reduce the risk of human errors. The present application provides the following technical solutions:
[0006] In a first aspect, the present application provides a method for generating a bidding document based on a template, the method comprising:
[0007] Obtain the project name and project scale, perform vectorized matching in the pre-built historical template library, and select the most matching template;
[0008] Identify the fields to be filled in the template example, and classify the fields to be filled in into: direct filling fields, general filling fields, and extended filling fields;
[0009] Obtaining key fields and pre-bid documents, extracting directly filled-in fields from the pre-bid documents based on the key fields, and filling the directly filled-in fields back into corresponding positions of the template;
[0010] Obtaining regulations and summary points and inputting them into a preset deep learning model, the model outputs summary fill-in fields and fills them back into the corresponding positions of the template;
[0011] Based on the project name and engineering scale, the most similar bidding sections are screened out in the pre-built bidding material library, key fields are extracted from the bidding sections and input into the deep learning model for rewriting, and the model outputs the expanded filled-in fields and fills them back into the corresponding positions of the template example.
[0012] In a specific implementation scheme, the acquisition of the project name and engineering scale, performing vectorized matching in a pre-built historical template library, and screening out the most matching template templates include:
[0013] Obtain the project name and project scale of the current bidding project, which is input by the user or exported from the project management system;
[0014] The historical template database is a dynamically updated database containing templates of public bidding documents;
[0015] Use a vectorization model to convert the project name and project scale into vectorized representation, convert all the templates of the bidding documents in the historical template library into vectorized representation, and calculate the similarity between the project name and project scale and each template through cosine similarity. The template with the highest similarity is the most matching template.
[0016] In a specific implementation scheme, the fields to be filled in the identification template example are classified into: direct filling fields, general filling fields and extended filling fields, including:
[0017] Analyze the template template, use string matching technology to identify all the fields to be filled in through specific tags or placeholders, and use natural language processing technology or text classifier to classify the fields to be filled in into the following three categories:
[0018] Direct fill fields: fields that can be directly extracted from text;
[0019] Summary Fill Field: Field that needs to be summarized based on the entered regulations and key points;
[0020] Expanded filling fields: fields that need to be searched for similar sections and rewritten based on project name and engineering scale.
[0021] In a specific possible implementation scheme, the obtaining of key fields and pre-bid documents, extracting directly fill-in fields from the pre-bid documents based on the key fields, and filling the directly fill-in fields back into corresponding positions of the template example includes:
[0022] Key fields sort out 150 common key fields in the bidding field. After obtaining the key fields and pre-bid documents, determine whether the key fields are simple fields or complex fields;
[0023] If the key field is a simple field, use the BM25 search model to retrieve the pre-bid document segment with the highest correlation to the key field, locate the complete field where the key field is located in the pre-bid document segment with the highest correlation, extract the direct-fill field from the complete field, and fill the direct-fill field back to the corresponding position of the template;
[0024] If the key field is a complex field, the pre-bid document is divided into several paragraphs, and the coverage of the key field and each paragraph of the pre-bid document is calculated. and semantic similarity , use the following formula to calculate the relevance of the key fields to each paragraph of the pre-bid document :
[0025] ;
[0026] After calculating the correlation between the key field and each paragraph of the pre-bid document, locate the complete field where the key field is located in the paragraph of the pre-bid document with the highest correlation, extract the direct-fill field from the complete field, and fill the direct-fill field back into the corresponding position of the template example.
[0027] In a specific implementation scheme, the calculation of the coverage of the key fields and the paragraphs of each pre-bid document is as follows: and semantic similarity include:
[0028] Calculate the coverage of key fields and paragraphs of each pre-bid document. The calculation formula is as follows:
[0029] ;
[0030] in, Indicates the key field A paragraph in the pre-bid document The coverage, and Respectively represent the key fields A paragraph in the pre-bid document Length, Indicates the key field A paragraph in the pre-bid document The number of overlapping words, Indicates the key field A paragraph in the pre-bid document The number of overlapping characters;
[0031] Use the cosine similarity calculation method to calculate the semantic similarity between the key fields and the paragraphs of each pre-labeled document .
[0032] In a specific implementation scheme, the obtaining of regulations and summary points and inputting them into a preset deep learning model, the model outputting summary fill-in fields and filling them back into corresponding positions of the template example include:
[0033] Obtain regulations and summary points, where the regulations are those required by the bidding project and the summary points are provided by the tenderer. Each summary fill-in field corresponds to a summary point;
[0034] For each summary fill-in field, the regulations and the corresponding summary points are input into the preset deep learning model, and the deep learning model outputs the corresponding summary fill-in field, which is then filled back into the corresponding position of the template.
[0035] In a specific feasible implementation plan, the most similar bidding sections are selected from the pre-built bidding material library based on the project name and engineering scale, key fields are extracted from the bidding sections and input into the deep learning model for rewriting, and the model outputs the expanded filled fields and fills them back into the corresponding positions of the template template, including:
[0036] The pre-built bidding material library contains historical bidding documents and their corresponding bidding sections. The text similarity between the project name and project scale and each bidding section in the pre-built bidding material library is calculated. , the calculation formula is as follows:
[0037] ;
[0038] in, is the project name, For the project scale, For the first Sections, is the smoothing parameter, usually set to 0.01, Indicates the number of words contained in the project name and the bid section. Indicates the number of words that are included in the project scale and bid section. , and The total number of words representing the project name, project scale and bidding section respectively;
[0039] After calculating the text similarity between the project name and project scale and each bid section in the pre-built bidding material library, the bid section with the highest text similarity is the most similar bid section selected. Then, the information gain is used to calculate the key degree of each word in the bid section to extract the key fields. The information gain value It is possible to define the contribution of vocabulary to the overall information content in a section. The calculation formula is as follows:
[0040] ;
[0041] in, It is vocabulary The information gain score indicates the contribution of vocabulary to distinguishing different text paragraphs. The most similar segment of the word The word frequency in ; the information gain score is calculated as follows:
[0042] ;
[0043] in, is a collection of categories of text paragraphs, Yes Category The probability of Yes Category The entropy of In a given vocabulary In case of category The conditional entropy of
[0044] After calculating the information gain value of each word, select the word with the largest information gain value or the field where the top three words are located as the key field, extract the key field and input it into the deep learning model for rewriting, and the deep learning model outputs the expanded filled-in field and fills it back into the corresponding position of the template.
[0045] In the second aspect, the present application provides a system for generating bidding documents based on template examples, which adopts the following technical solutions:
[0046] A system for generating bidding documents based on template examples, comprising:
[0047] The template matching module is used to obtain the project name and project scale, perform vectorized matching in the pre-built historical template library, and select the most matching template;
[0048] A field identification module, used for identifying the fields to be filled in the template example, and classifying the fields to be filled in into: direct filling fields, general filling fields and extended filling fields;
[0049] A direct backfill module is used to obtain key fields and pre-bid documents, extract direct-fill fields from the pre-bid documents based on the key fields, and backfill the direct-fill fields to corresponding positions of the template;
[0050] A summary backfill module is used to obtain regulations and summary points and input them into a preset deep learning model, and the model outputs the summary fill-in field and backfills it into the corresponding position of the template;
[0051] The expansion and backfill module is used to select the most similar bidding sections in the pre-built bidding material library based on the project name and engineering scale, extract key fields from the bidding sections and input them into the deep learning model for rewriting, and the model outputs the expanded filled-in fields and backfills them into the corresponding positions of the template example.
[0052] 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 method for generating a bidding document based on a template example as described in the first aspect.
[0053] In a fourth aspect, the present application provides a computer-readable storage medium, in which a program is stored. When the program is executed by a processor, it is used to implement a method for generating a bidding document based on a template example as described in the first aspect.
[0054] By obtaining the project name and engineering scale and performing vectorized matching, the most matching template template is screened out. Then the fields to be filled in the template are identified and classified into direct-filled fields, summary-filled fields, and extended-filled fields. The BM25 retrieval model and comprehensive evaluation method are used to extract the direct-filled fields from the pre-bid documents and backfill them into the template. For the summary-filled fields, relevant regulations and key points are obtained, input into the deep learning model to generate and backfill. For the extended-filled fields, key fields are extracted through text similarity calculation and information gain, and then rewritten and generated by the deep learning model. This method solves the problems of manual reliance, time-consuming and error-proneness in the existing technology, and improves the accuracy and efficiency of bidding document generation.
[0055] 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
[0056] Figure 1 It is a flow chart of a method for generating bidding documents based on a template example in an embodiment of the present application.
[0057] Figure 2It is a schematic diagram of the overall process of the method for generating bidding documents based on a template example in an embodiment of the present application.
[0058] Figure 3 It is a structural block diagram of a bidding document generation system based on a template example in an embodiment of the present application.
[0059] Figure 4 It is a block diagram of an electronic device generated based on a bidding document of a template example in an embodiment of the present application. DETAILED DESCRIPTION
[0060] 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.
[0061] Optionally, the present application uses the template-based bidding document generation method provided in each embodiment as an example for use 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.
[0062] Reference Figure 1 , is a flow chart of a method for generating a bidding document based on a template provided by an embodiment of the present application, the method comprising at least the following steps:
[0063] Step S101: Obtain the project name and project scale, perform vectorized matching in a pre-built historical template library, and select the most matching template.
[0064] In step S101, the project name and engineering scale of the current bidding project are first obtained, which are input by the user or exported from the project management system. The historical template library is a dynamically updated database that contains a large number of template templates of public bidding documents. Then, the project name and engineering scale are vectorized and matched in the historical template library using an existing vectorization model, such as Word2Vec. Specifically, the project name and engineering scale are first converted into vectorized representations using a vectorization model, and then the template templates of all bidding documents in the historical template library are converted into vectorized representations. The similarity between the project name and engineering scale and each template template is calculated using cosine similarity. The template template with the highest similarity is the most matching template template.
[0065] Step S102, identifying the fields to be filled in the template example, and classifying the fields to be filled in into: direct filling fields, general filling fields and extended filling fields.
[0066] In step S102, the template is first analyzed, and all fields to be filled in are identified through specific tags or placeholders using existing string matching technology, and then the fields to be filled in are divided into the following three categories using existing natural language processing technology or text classifiers:
[0067] Direct fill field: A field that can be extracted directly from text.
[0068] Summary fill-in field: Field that needs to be summarized based on the entered regulations and key points.
[0069] Expanded filling fields: fields that need to be searched for similar sections and rewritten based on project name and engineering scale.
[0070] In addition, as a preference, manual review may be performed after the fields to be filled in are classified to ensure the accuracy and rationality of the classification as much as possible.
[0071] Step S103, obtaining key fields and pre-bid documents, extracting directly filled-in fields from the pre-bid documents based on the key fields, and filling the directly filled-in fields back into corresponding positions of the template example.
[0072] In step S103, the key fields are 150 common key fields sorted out in the bidding field. After obtaining the key fields and pre-bid documents, it is first determined whether the key fields are simple fields or complex fields. In this application, the judgment is made by determining whether the number of characters in the key fields exceeds the threshold. Other methods can also be used for judgment. This application does not limit the judgment method.
[0073] In implementation, if the key field is a simple field, first use the text data of the pre-bid document to build a BM25 retrieval model, input the key field as a query into the BM25 retrieval model, and the BM25 retrieval model retrieves the pre-bid document fragment with the highest correlation to the key field. The BM25 retrieval model is a retrieval model based on word frequency and inverse document frequency, which is used to calculate the correlation score between the query and the document. Then, locate the complete field where the key field is located in the pre-bid document fragment with the highest correlation, extract the direct fill-in field from the complete field, and fill the direct fill-in field back into the corresponding position of the template template. For example, the key field is the project name, and the fragments retrieved from the pre-bid document using the BM25 retrieval model are: Project Name: An Intelligent Building Management System; Project Location: Chaoyang District, Beijing; Project Budget: 50 million yuan. Among them, the complete field where the key field is located is: Project Name: An Intelligent Building Management System; the direct fill-in field extracted at the end is an intelligent building management system.
[0074] In implementation, if the key field is a complex field, the pre-bid document is divided into several paragraphs. First, the coverage of the key field and each paragraph of the pre-bid document is calculated. The calculation formula is as follows:
[0075] ;
[0076] in, Indicates the key field A paragraph in the pre-bid document The coverage, and Respectively represent the key fields A paragraph in the pre-bid document Length, Indicates the key field A paragraph in the pre-bid document The number of overlapping words, Indicates the key field A paragraph in the pre-bid document The above formula combines character-level overlap and word-level overlap to more comprehensively reflect the matching between key fields and paragraphs. This comprehensive evaluation helps reduce the bias or shortcomings that may be caused by a single measurement method. In addition, the denominator in the above formula is used to normalize the overlap, taking into account the length difference between key fields and paragraphs. This normalization can more fairly compare the similarities between texts of different lengths, making the evaluation results more universal.
[0077] Then the cosine similarity calculation method is used to calculate the semantic similarity between the key fields and the paragraphs of each pre-labeled document. Finally, the following formula is used to calculate the relevance between the key fields and each paragraph of the pre-bid document: :
[0078] ;
[0079] The above formula uses the product and weighted sum method to ensure that coverage and semantic similarity are properly weighed when calculating relevance. In addition, the product form design ensures that the range of relevance values can fully utilize the respective ranges of coverage and semantic similarity to a certain extent. This normalization method helps to avoid unreasonable results in extreme cases, such as when the overall relevance is not affected when a certain metric value is zero. The product form adds consideration for the nonlinear relationship between coverage and semantic similarity. This design can more flexibly capture the complex interaction between the two. For example, in some cases, even if one metric value is high, but the other metric value is low, a reasonable relevance assessment can be obtained.
[0080] In implementation, after calculating the relevance between the key field and each paragraph of the pre-bid document, locate the complete field where the key field is located in the paragraph of the pre-bid document with the highest relevance, extract the direct-fill field from the complete field, and fill the direct-fill field back into the corresponding position of the template. It should be noted that if the complete field where the key field is located cannot be located in the paragraph of the pre-bid document with the highest relevance, continue to locate it in the paragraph of the pre-bid document with the second highest relevance until it can be located successfully.
[0081] In step S103, simple fields usually have clear semantics and structures. The use of the BM25 retrieval model can quickly locate relevant information and extract fields for direct filling, avoiding complex calculation processes and improving extraction efficiency and accuracy. Complex fields may involve multiple paragraphs or multiple layers of information. The comprehensive evaluation of coverage and semantic similarity can better capture the complex structure and content of information, avoid missing key information, and improve the comprehensiveness and accuracy of extraction. By selecting a suitable extraction method, the system can maintain efficiency and stability when processing texts of various complexities and structures, enhance the adaptability and universality of the system, and make it suitable for a wider range of application scenarios. In summary, using different extraction methods for simple fields and complex fields can improve the accuracy and comprehensiveness of extraction while ensuring extraction efficiency, so that the system can operate more efficiently when processing large amounts of text information.
[0082] Step S104: Obtain the regulations and summary points and input them into the preset deep learning model. The model outputs the summary fill-in fields and fills them back into the corresponding positions of the template.
[0083] In step S104, firstly, the regulations and summary points are obtained, wherein the regulations are the regulations required by the bidding project, the summary points are provided by the tenderer, and each summary fill-in field corresponds to a summary point. Then, for each summary fill-in field, the regulations and the corresponding summary points are input into the preset deep learning model, the deep learning model outputs the corresponding summary fill-in field, and finally the summary fill-in field is filled back into the corresponding position of the template.
[0084] Optionally, the deep learning model in the present application can be an existing ChatGpt model, or a GPT model trained and generated for the bidding field.
[0085] Step S105: based on the project name and engineering scale, select the most similar bidding sections in the pre-built bidding material library, extract key fields from the bidding sections and input them into the deep learning model for rewriting, and the model outputs the expanded filled fields and fills them back into the corresponding positions of the template.
[0086] In step S105, the pre-built bidding material library contains a large number of historical bidding documents and their corresponding bidding sections, and the text similarity between the project name and engineering scale and each bidding section in the pre-built bidding material library is calculated. , the calculation formula is as follows:
[0087] ;
[0088] in, is the project name, For the project scale, For the first Sections, is the smoothing parameter, usually set to 0.01, Indicates the number of words contained in the project name and the bid section. Indicates the number of words that are included in the project scale and bid section. , and Represents the total number of words for project name, project scale and bid section respectively. In text similarity calculation, a common problem is that if two texts do not have any common words, the traditional similarity formula may result in a denominator of zero, making it impossible to compare effectively. To solve this problem, this application introduces normalization, that is, adding a small positive number to the formula This ensures that the denominator will not be zero and smooths the calculation of similarity. In addition, the length of the text can affect the calculation results of text similarity. Longer texts may have more words, which may lead to lower similarity calculation results with shorter texts. Therefore, the total number of words in each text is used in the formula to normalize the calculation, ensuring the fairness of length in similarity calculation.
[0089] Optionally, the existing cosine similarity calculation method can also be used to calculate the text similarity between the project name and engineering scale and each bid section in the pre-built bidding material library. This application does not limit the similarity calculation method.
[0090] In the implementation, after calculating the text similarity between the project name and project scale and each bid section in the pre-built bidding material library, the bid section with the highest text similarity is the most similar bid section selected. Then, the information gain is used to calculate the key degree of each word in the bid section to extract the key fields. The information gain value It is possible to define the contribution of vocabulary to the overall information content in a section. The calculation formula is as follows:
[0091] ;
[0092] in, It is vocabulary The information gain score indicates the contribution of vocabulary to distinguishing different text paragraphs (or sentences). The most similar segment of the word The information gain score is calculated as follows:
[0093] ;
[0094] in, is a set of categories of text paragraphs (or sentences), Yes Category The probability of Yes Category The entropy of , measures the uncertainty of the category. In a given vocabulary In case of category The conditional entropy of the vocabulary For Category The contribution of . It should be noted that entropy is a measure of the uncertainty of a random variable. In text classification or key field extraction, entropy can be understood as a quantification of the uncertainty of a category. For example, if there are multiple text paragraphs or sentences, the entropy of each category (such as different topics or categories) can be calculated to understand the amount of information in each category and the degree of variation within it. The specific calculation method of entropy can be described as follows: first, for a given set of categories, the frequency or probability of occurrence of each category is calculated, the probability of each category is substituted into the entropy calculation formula, and its uncertainty is measured according to its probability. A higher entropy value indicates that the variation within the category is larger or the probability distribution is more uniform. Conditional entropy is the uncertainty of another random variable when a condition (such as a certain word or feature) is known. In text processing, conditional entropy helps measure the contribution of a specific word to category distinction. The specific calculation method of conditional entropy can be described as follows: select a specific condition (such as a word) and then classify the text paragraph or sentence. Under the given condition, calculate the conditional probability of each category, that is, the probability of each category appearing under the given condition. The uncertainty of the category under the given condition is measured according to the conditional probability. The lower the conditional entropy, the more effective it is for a given condition to distinguish different categories. Both entropy and conditional entropy are basic concepts in information theory and belong to existing mathematical and probability theory concepts, so this application will not describe them in detail here.
[0095] In the above formula, the text similarity is first calculated to determine the section that is most similar to the given project name and engineering scale, ensuring that the subsequent information gain calculation is based on the real relevant sections. The most similar sections determined by text similarity, rather than simply selecting a fixed template, can more accurately reflect the specific needs and conditions of the current project. Secondly, the categories of multiple text paragraphs are considered. Through the probability weighting of the categories, the importance of words in different text paragraphs is comprehensively evaluated, which can more comprehensively capture the frequency and importance of words in different contexts, rather than simply counting the word frequency. The information gain calculation is intuitive and easy to understand, and can provide a quantitative importance indicator for each word, which helps decision makers understand and verify the selection process of key fields.
[0096] In implementation, after calculating the information gain value of each word, the word with the largest information gain value or the field where the top three words are located is selected as the key field for extraction, and the project name, project scale and extracted key fields are input into the deep learning model for rewriting. The deep learning model outputs the expanded filled-in fields and fills them back into the corresponding position of the template.
[0097] It should be noted that there is usually only one extended fill-in field. If there are multiple extended fill-in fields, the tenderer is required to design auxiliary prompt words for each extended fill-in field that needs to be generated. The auxiliary prompt words are used to guide the deep learning model to generate fields related to the extended fill-in fields, such as describing the overall goals and expected results of the project or providing a timetable and milestones for the implementation of the project. The project name, project scale, extracted key fields and the auxiliary prompt words corresponding to the extended fill-in field are input into the deep learning model for rewriting, and the deep learning model outputs the extended fill-in field.
[0098] In summary, refer to Figure 2 , is an overall flow chart of the method for generating bidding documents based on template examples in an embodiment of the present application. The present application obtains the project name and engineering scale and performs vectorized matching to screen out the most matching template example. Then identify the fields to be filled in the template and classify them into direct-filled fields, summary-filled fields, and extended-filled fields. Use the BM25 retrieval model and comprehensive evaluation method to extract the direct-filled fields from the pre-bid documents and backfill them into the template. For the summary-filled fields, obtain relevant laws and regulations and key points, input them into the deep learning model to generate and backfill. For the extended-filled fields, extract the key fields through text similarity calculation and information gain, and then rewrite and generate them through the deep learning model. This method solves the problems of manual reliance, time-consuming and error-proneness in the prior art, and improves the accuracy and efficiency of bidding document generation.
[0099] Reference Figure 3, is a structural block diagram of a system for generating bidding documents based on a template provided by an embodiment of the present application, which adopts the following technical solutions:
[0100] The template matching module is used to obtain the project name and project scale, perform vectorized matching in the pre-built historical template library, and select the most matching template;
[0101] A field identification module is used to identify the fields to be filled in the template example, and classify the fields to be filled in into: direct filling fields, general filling fields and extended filling fields;
[0102] The direct backfill module is used to obtain key fields and pre-bid documents, extract direct-fill fields in the pre-bid documents based on the key fields, and backfill the direct-fill fields to the corresponding positions of the template;
[0103] The summary backfill module is used to obtain regulations and summary points and input them into the preset deep learning model. The model outputs the summary fill-in fields and backfills them into the corresponding positions of the template;
[0104] The expansion backfill module is used to select the most similar bidding sections in the pre-built bidding material library based on the project name and engineering scale, extract key fields from the bidding sections and input them into the deep learning model for rewriting, and the model outputs the expanded filled fields and backfills them into the corresponding positions of the template example.
[0105] For relevant details, refer to the above method embodiment.
[0106] Figure 4 4 is a block diagram of an electronic device provided by an embodiment of the present application. The device at least includes a processor 401 and a memory 402.
[0107] The processor 401 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 401 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 401 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 401 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 401 may also include an AI (Artificial Intelligence) processor, which is used to process computing operations related to machine learning.
[0108] The memory 402 may include one or more computer-readable storage media, which may be non-transitory. The memory 402 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 402 is used to store at least one instruction, which is used to be executed by the processor 401 to implement the method for generating a bidding document based on a template template provided in the method embodiment of the present application.
[0109] In some embodiments, the electronic device may further optionally include: a peripheral device interface and at least one peripheral device. The processor 401, the memory 402 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.
[0110] Of course, the electronic device may also include fewer or more components, which is not limited in this embodiment.
[0111] 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 template-based bidding document generation method of the above method embodiment.
[0112] 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 template-based bidding document generation method of the above method embodiment.
[0113] 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.
[0114] 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 method for generating bidding documents based on a template, characterized in that: The method comprises: Obtain the project name and project scale, perform vectorized matching in the pre-built historical template library, and select the most matching template; Identify the fields to be filled in the template example, and classify the fields to be filled in into: direct filling fields, general filling fields, and extended filling fields; Obtaining key fields and pre-bid documents, extracting directly filled-in fields from the pre-bid documents based on the key fields, and filling the directly filled-in fields back into corresponding positions of the template; Obtaining regulations and summary points and inputting them into a preset deep learning model, the model outputs summary fill-in fields and fills them back into the corresponding positions of the template; Based on the project name and engineering scale, the most similar bidding sections are selected from the pre-built bidding material library, key fields are extracted from the bidding sections and input into the deep learning model for rewriting, and the model outputs the expanded filled fields and fills them back into the corresponding positions of the template; The most similar bidding sections are selected from the pre-built bidding material library based on the project name and engineering scale, key fields are extracted from the bidding sections and input into the deep learning model for rewriting, and the model outputs the expanded filled fields and fills them back into the corresponding positions of the template example, including: The pre-built bidding material library contains historical bidding documents and their corresponding bidding sections. The text similarity between the project name and project scale and each bidding section in the pre-built bidding material library is calculated. , the calculation formula is as follows: in, is the project name, For the project scale, For the first Sections, is the smoothing parameter, usually set to 0.
01. Indicates the number of words contained in the project name and the bid section. Indicates the number of words that are included in the project scale and bid section. , and The total number of words representing the project name, project scale and bidding section respectively; After calculating the text similarity between the project name and project scale and each bid section in the pre-built bidding material library, the bid section with the highest text similarity is the most similar bid section selected. Then, the information gain is used to calculate the key degree of each word in the bid section to extract the key fields. The information gain value It is possible to define the contribution of vocabulary to the overall information content in a section. The calculation formula is as follows: in, It is vocabulary The information gain score indicates the contribution of vocabulary to distinguishing different text paragraphs. The most similar segment of the word The word frequency in ; the information gain score is calculated as follows: in, is a collection of categories of text paragraphs, Yes Category The probability of Yes Category The entropy of In a given vocabulary In case of category The conditional entropy of After calculating the information gain value of each word, select the word with the largest information gain value or the field where the top three words are located as the key field, extract the key field and input it into the deep learning model for rewriting, and the deep learning model outputs the expanded filled-in field and fills it back into the corresponding position of the template.
2. The method for generating bidding documents based on templates according to claim 1, characterized in that: The project name and project scale are obtained, and vectorized matching is performed in the pre-built historical template library to screen out the most matching templates, including: Obtain the project name and project scale of the current bidding project, which is input by the user or exported from the project management system; The historical template database is a dynamically updated database containing templates of public bidding documents; Use a vectorization model to convert the project name and project scale into vectorized representation, convert all the templates of the bidding documents in the historical template library into vectorized representation, and calculate the similarity between the project name and project scale and each template through cosine similarity. The template with the highest similarity is the most matching template.
3. The method for generating bidding documents based on templates according to claim 1, characterized in that: The fields to be filled in the identification template are classified into: direct filling fields, general filling fields and extended filling fields, including: Analyze the template template, use string matching technology to identify all the fields to be filled in through specific tags or placeholders, and use natural language processing technology or text classifier to classify the fields to be filled in into the following three categories: Direct fill fields: fields that can be directly extracted from text; Summary Fill Field: Field that needs to be summarized based on the entered regulations and key points; Expanded filling fields: fields that need to be searched for similar sections and rewritten based on project name and engineering scale.
4. The method for generating bidding documents based on templates according to claim 1, characterized in that: The obtaining of key fields and pre-bid documents, extracting directly fill-in fields from the pre-bid documents based on the key fields, and filling the directly fill-in fields back into corresponding positions of the template example comprises: Key fields sort out 150 common key fields in the bidding field. After obtaining the key fields and pre-bid documents, determine whether the key fields are simple fields or complex fields; If the key field is a simple field, use the BM25 search model to retrieve the pre-bid document segment with the highest correlation to the key field, locate the complete field where the key field is located in the pre-bid document segment with the highest correlation, extract the direct-fill field from the complete field, and fill the direct-fill field back to the corresponding position of the template; If the key field is a complex field, the pre-bid document is divided into several paragraphs, and the coverage of the key field and each paragraph of the pre-bid document is calculated. and semantic similarity , use the following formula to calculate the relevance of the key fields to each paragraph of the pre-bid document : After calculating the correlation between the key field and each paragraph of the pre-bid document, locate the complete field where the key field is located in the paragraph of the pre-bid document with the highest correlation, extract the direct-fill field from the complete field, and fill the direct-fill field back into the corresponding position of the template example.
5. The method for generating bidding documents based on templates according to claim 4, characterized in that: The calculation focuses on the coverage of the key fields and paragraphs of each pre-bid document. and semantic similarity include: Calculate the coverage of key fields and paragraphs of each pre-bid document. The calculation formula is as follows: in, Indicates the key field A paragraph in the pre-bid document The coverage, and Respectively represent the key fields A paragraph in the pre-bid document Length, Indicates the key field A paragraph in the pre-bid document The number of overlapping words, Indicates the key field A paragraph in the pre-bid document The number of overlapping characters; Use the cosine similarity calculation method to calculate the semantic similarity between the key fields and the paragraphs of each pre-labeled document .
6. The method for generating bidding documents based on templates according to claim 1, characterized in that: The obtaining of regulations and summary points and inputting them into a preset deep learning model, the model outputting summary filled fields and filling them back into the corresponding positions of the template example include: Obtain regulations and summary points, where the regulations are those required by the bidding project and the summary points are provided by the tenderer. Each summary fill-in field corresponds to a summary point; For each summary fill-in field, the regulations and the corresponding summary points are input into the preset deep learning model, and the deep learning model outputs the corresponding summary fill-in field, which is then filled back into the corresponding position of the template.
7. A system for generating bidding documents based on templates, characterized in that: include: The template matching module is used to obtain the project name and project scale, perform vectorized matching in the pre-built historical template library, and select the most matching template; A field identification module, used for identifying the fields to be filled in the template example, and classifying the fields to be filled in into: direct filling fields, general filling fields and extended filling fields; A direct backfill module is used to obtain key fields and pre-bid documents, extract direct-fill fields from the pre-bid documents based on the key fields, and backfill the direct-fill fields to corresponding positions of the template; A summary backfill module is used to obtain regulations and summary points and input them into a preset deep learning model, and the model outputs the summary fill-in field and backfills it into the corresponding position of the template; An expansion and backfilling module is used to select the most similar bidding sections in the pre-built bidding material library based on the project name and engineering scale, extract key fields from the bidding sections and input them into the deep learning model for rewriting, and the model outputs the expanded filled fields and backfills them into the corresponding positions of the template; The most similar bidding sections are selected from the pre-built bidding material library based on the project name and engineering scale, key fields are extracted from the bidding sections and input into the deep learning model for rewriting, and the model outputs the expanded filled fields and fills them back into the corresponding positions of the template example, including: The pre-built bidding material library contains historical bidding documents and their corresponding bidding sections. The text similarity between the project name and project scale and each bidding section in the pre-built bidding material library is calculated. , the calculation formula is as follows: in, is the project name, For the project scale, For the first Sections, is the smoothing parameter, usually set to 0.
01. Indicates the number of words contained in the project name and the bid section. Indicates the number of words that are included in the project scale and bid section. , and The total number of words representing the project name, project scale and bidding section respectively; After calculating the text similarity between the project name and project scale and each bid section in the pre-built bidding material library, the bid section with the highest text similarity is the most similar bid section selected. Then, the information gain is used to calculate the key degree of each word in the bid section to extract the key fields. The information gain value It is possible to define the contribution of vocabulary to the overall information content in a section. The calculation formula is as follows: in, It is vocabulary The information gain score indicates the contribution of vocabulary to distinguishing different text paragraphs. The most similar segment of the word The word frequency in ; the information gain score is calculated as follows: in, is a collection of categories of text paragraphs, Yes Category The probability of Yes Category The entropy of In a given vocabulary In case of category The conditional entropy of After calculating the information gain value of each word, select the word with the largest information gain value or the field where the top three words are located as the key field, extract the key field and input it into the deep learning model for rewriting, and the deep learning model outputs the expanded filled-in field and fills it back into the corresponding position of the template.
8. 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 method for generating bidding documents based on a template example as described in any one of claims 1 to 6.
9. 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 method for generating a bidding document based on a template as described in any one of claims 1 to 6.
Citation Information
Patent Citations
Intelligent bid evaluation method and system, storage medium and electronic equipment
CN112632943A
Method and system for generating bid invitation file
CN114037493A
Technical method for automatically filling guarantee responsibility abstract applied to insurance clause analysis
CN114330283A
