Intelligent Generation Method and Computer System for Bidding Documents Based on Natural Language Processing

Abstract

The present application provides an intelligent generation method and computer system for tender documents based on natural language processing. When analyzing the tender possibility of a target tender document, the target tender document is divided into a set of target text sequences, and the first implicit representation set corresponding to the set of target text sequences is repeatedly iterated according to a feature accumulator to perform cyclic feature extraction on the set of target text sequences, obtaining a second implicit representation set; because the core tender semantic information can be propagated between text sequences based on the feature accumulator, the second implicit representation set can accurately reflect the core tender semantic information in the target tender document. When analyzing the tender possibility based on the second implicit representation set, the reliability of the tender possibility analysis can be increased, providing an accurate reference for generating a suitable tender document subsequently.

Description

Technical Field

[0001] This application relates to the field of electrical data processing. Specifically, it relates to an intelligent generation method and computer system for tender documents based on natural language processing. Background Art

[0002] In modern business activities, tendering and bidding is a widely used competition mechanism, which is of crucial significance for the implementation of various projects. During the tendering and bidding process, the quality of tender documents directly affects whether a bidder can stand out in the competition. Traditional tender document preparation mainly relies on manual operations, which involves a large amount of text reading, analysis, writing, and sorting work, consuming a great deal of manpower and time.

[0003] Tender documents usually have complex structures and semantic contents, and the tendering intentions contained therein may not always be intuitively obvious. When interpreted manually, due to individual differences in understanding and different levels of mastery of industry knowledge and project backgrounds, it is very easy to have inaccurate understandings of tendering intentions.

[0004] Before preparing tender documents, enterprises need to evaluate the feasibility of tendering, including whether their own capabilities meet the tender requirements, whether the project risks are controllable, and whether the expected returns are reasonable, etc. Traditional evaluation methods often rely on empirical judgments or simple cost-benefit analyses, lacking systematic and comprehensive evaluation methods.

[0005] With the development of information technology, natural language processing (NLP) technology has been widely applied in many fields, such as intelligent customer service, information retrieval, text automatic summarization, etc. However, in the field of tendering and bidding, the application of natural language processing technology is relatively less and its potential has not been fully exploited. Summary of the Invention

[0006] The object of the present invention is to provide an intelligent generation method and computer system for tender documents based on natural language processing. The present application is implemented as follows: In the first aspect, the present application provides an intelligent generation method for tender documents based on natural language processing, and the method includes: extracting implicit representations of a target text sequence set of a target tender document to obtain a first implicit representation set; performing a walk on the first implicit representation set, and performing the following operations on the x-th first implicit representation reached during the walk, where x is a positive integer less than X, and X is the number of target text sequences in the target text sequence set: iterating the y-th feature accumulator according to the x-th first implicit representation to obtain the x-th feature accumulator, where if x is 1, the y-th feature accumulator is the initial feature accumulator, and if x is greater than 1, the y-th feature accumulator is the feature accumulator of the y-th target text sequence, where y = x - 1; obtaining the x-th second implicit representation according to the x-th feature accumulator; obtaining a second implicit representation set according to the x second implicit representations obtained by performing a walk on the first implicit representation set; performing a tender possibility analysis according to the second implicit representation set to obtain a target tender support rate; and generating a tender document by invoking a matching preset tender document generation framework based on the target tender support rate.

[0007] In the second aspect, the present application provides a computer system, including: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the above-mentioned method is implemented.

[0008] The beneficial effects of the present application: When performing a tender possibility analysis on a target tender document in an embodiment of the present application, the target tender document is divided into a target text sequence set, and the first implicit representation set corresponding to the target text sequence set is repeatedly iterated according to the feature accumulator to perform cyclic feature extraction on the target text sequence set to obtain a second implicit representation set; because the core tender semantic information can be propagated between text sequences based on the feature accumulator, the second implicit representation set can accurately reflect the core tender semantic information in the target tender document, and when performing a tender possibility analysis according to the second implicit representation set, the reliability of the tender possibility analysis can be increased, providing an accurate reference for generating a suitable tender document subsequently. Description of the Drawings

[0009] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description in the embodiments of the present application.

[0010] Figure 1It is a flowchart of an intelligent tender document generation method based on natural language processing provided by an embodiment of the present application;

[0011] Figure 2 It is a schematic diagram of the composition of a computer system provided by an embodiment of the present application. Detailed implementation manners

[0012] In an embodiment of the present application, the execution subject of the intelligent tender document generation method based on natural language processing is a computer system, including but not limited to servers, personal computers, laptop computers, tablet computers, smart phones, etc. As Figure 1 shown, the method includes: Step S100: Extract implicit representations from the target text sequence set of the target tender document to obtain a first implicit representation set.

[0013] In step S100, the target tender document is a complete document content, which contains various detailed information about the tender project, such as the basic requirements, technical specifications, and business terms of the tender project. The computer system processes this tender document according to a certain logic or format and converts it into a target text sequence set. Here, the target text sequence set can be understood as a set of multiple text sequences obtained by dividing the text in the tender document according to specific rules. For example, taking a construction project tender as an example, it may be divided into different text sequences according to different chapters (such as project overview, construction requirements, quality standards, etc.).

[0014] Implicit representation extraction is a process of converting the semantic information in a text sequence into a vector representation that can be processed by a computer. The computer system can adopt various technical means to achieve implicit representation extraction. A common method is the word vector model, such as the Word2Vec model. Suppose we have a simple text sequence "The concrete strength needs to reach C30". In the Word2Vec model, each word (such as "concrete", "strength", "reach", "C30") will be mapped to a specific vector. The combination of these vectors or through further processing (such as weighted summation, averaging, etc.) can obtain the implicit representation of this text sequence. Specifically, if the word vectors are respectively , then a simple implicit representation formula for the text sequence can be , where I is the implicit representation of this text sequence.

[0015] The computer system performs such an implicit representation extraction operation on each text sequence in the target text sequence set. For example, for another text sequence "The construction period is 180 days" in the target text sequence set, its implicit representation is obtained in the same way as above (using the word vector model and related combination operations). By processing all the text sequences in the target text sequence set in this way, the computer system finally obtains a first implicit representation set. Each element in this first implicit representation set corresponds to the implicit representation of a text sequence in the target text sequence set, and it contains a digital representation of the semantic information of each text sequence. This representation can be used in subsequent operations of the computer system (such as the random walk operation in step S200, etc.) to further analyze and process the target tender document, such as performing bid possibility analysis and generating the final bid document, etc.

[0016] Step S200: Perform a random walk on the first implicit representation set and perform the following operations on the x-th first implicit representation reached, where x is a positive integer less than X, and X is the number of target text sequences in the target text sequence set: Iterate the y-th feature accumulator according to the x-th first implicit representation to obtain the x-th feature accumulator. If x is 1, then the y-th feature accumulator is the initial feature accumulator. If x is greater than 1, then the y-th feature accumulator is the feature accumulator of the y-th target text sequence, where y = x - 1.

[0017] In step S200, the computer system performs a random walk operation on the first implicit representation set obtained in step S100 and executes specific operations on the x-th first implicit representation reached. Among them, x is a positive integer less than X, and X represents the number of target text sequences in the target text sequence set.

[0018] The so-called random walk means that the computer system sequentially accesses the elements in the first implicit representation set in a certain order. For example, assume that there are 5 elements in the first implicit representation set (i.e., X = 5), then the computer system will start from the first element and gradually move to the subsequent elements. This process is like viewing each element in an ordered list in sequence.

[0019] For the x-th first implicit representation that has wandered, the computer system iterates the y-th feature accumulator (which can be understood as the Hidden State) based on it to obtain the x-th feature accumulator. Here, the feature accumulator is a structure for storing and transmitting information. When x is 1, the y-th feature accumulator is the initial feature accumulator, which is the starting state of the whole process. The initial feature accumulator can be initialized as a all-zero vector or a random vector, depending on the adopted technical solution. For example, if initialized with a all-zero vector, assuming the dimension of the feature accumulator is d-dimensional, then the initial feature accumulator H0 is a d-dimensional vector, where each element is 0, that is (a total of d zeros).

[0020] When x is greater than 1, the y-th feature accumulator is the feature accumulator of the y-th target text sequence, where y = x - 1. This means that during the wandering process, the computer system uses the information of the feature accumulator corresponding to the previous text sequence to update the feature accumulator corresponding to the current text sequence. To implement this iterative process, the computer system can adopt various technical means. A common way is to use a method that combines linear transformation and non-linear activation function. Assume the x-th first implicit representation is , and the y-th feature accumulator is . The computer system first performs a linear combination on them, for example , where and are learnable weight matrices, and their dimensions are determined according to and . Then, process H' through a non-linear activation function (such as the ReLU function) to obtain the x-th feature accumulator , that is .

[0021] This way of updating the current feature accumulator based on the previous feature accumulator enables the feature accumulator to transmit information between text sequences. Taking the bidding of an engineering project as an example, assume there are two different target text sequences regarding the project budget and the project quality requirements. When the computer system wanders from the text sequence of the project budget to the text sequence of the project quality requirements, the information contained in the feature accumulator corresponding to the project budget text sequence (such as the approximate range of the budget, the characteristics of the fund allocation, etc.) can be used to update the feature accumulator corresponding to the project quality requirements text sequence, so that the feature accumulator corresponding to the project quality requirements text sequence not only contains its own information (such as quality standards, inspection methods, etc.), but also integrates the relevant information of the previous project budget text sequence. Such an operation helps to integrate the information of different parts during the analysis process of the entire target bidding document, laying a foundation for subsequent accurate bid possibility analysis and bid document generation.

[0022] By performing such operations on each element in the first implicit representation set (i.e., each first implicit representation), the computer system can gradually construct feature accumulators corresponding to each target text sequence. These feature accumulators will play an important role in subsequent steps (such as obtaining the second implicit representation in step S300), which helps to deeply explore the semantic information and logical relationships in the target tender documents.

[0023] Step S300: Obtain the x-th second implicit representation based on the x-th feature accumulator.

[0024] In step S300, the x-th feature accumulator is the result obtained by the computer system through a series of operations in step S200. The feature accumulator contains information related to the previous text sequences. After integration and processing, this information becomes a state representation that can reflect some features of the target tender documents.

[0025] The computer system needs to obtain the second implicit representation from this feature accumulator, and this process can be achieved by various technical means. A possible way is to adopt the linear mapping technology. For example, assume that the x-th feature accumulator is a vector with a dimension of d . The computer system can define a weight matrix W with a dimension of e×d (where e is the target dimension of the second implicit representation). Then, through matrix multiplication, the x-th second implicit representation can be obtained, and the formula is .

[0026] Taking the tender of a software project as an example, in the target tender documents, there are different text sequences including project function requirements, technical architecture requirements, etc. When the computer system processes the x-th feature accumulator corresponding to the technical architecture requirements, this feature accumulator has already integrated the information related to the previous text sequences of project function requirements. Through the above linear mapping operation, the computer system converts this feature accumulator into the x-th second implicit representation. This second implicit representation can reflect the technical architecture requirements and the related project function requirement information in a new and more suitable form for subsequent analysis.

[0027] Another technical means can be based on the fully connected layer in the neural network. Taking the x-th feature accumulator as the input of the fully connected layer, after the calculation of the neurons in the fully connected layer, the x-th second implicit representation is output. Each neuron in the fully connected layer is connected to all neurons in the input layer (i.e., the feature accumulator), and each connection has a corresponding weight. The neuron performs a weighted sum based on the input values and these weights, and is processed through an activation function (such as the Sigmoid function or the Tanh function) to obtain the final output, which is the x-th second implicit representation. This way can process the information in the feature accumulator more flexibly, discover the complex relationships between different features, and thus generate a more representative second implicit representation, providing an effective data basis for subsequent operations such as bid possibility analysis based on the set of second implicit representations.

[0028] Step S400: Obtain a set of second implicit representations based on the x second implicit representations obtained by walking through the set of first implicit representations.

[0029] In the previous steps, the computer system performed a series of processes on the set of target text sequences of the target bidding documents. First, in step S100, a set of first implicit representations was extracted, then in step S200, a walk was performed on the set of first implicit representations and each corresponding feature accumulator was obtained through iterative feature accumulation, and then in step S300, one by one second implicit representations were obtained based on these feature accumulators. The second implicit representation here is a new representation form after complex processing of the semantic information of different parts of the target bidding documents.

[0030] When the computer system is walking through the set of first implicit representations, every time a second implicit representation is obtained, it is equivalent to re-encoding and integrating a part of the information in the target bidding documents. For example, in a construction project bidding, the target bidding documents include multiple aspects such as project scale, construction technical requirements, and construction period requirements. For the text sequence of the project scale part, a specific second implicit representation is obtained through the previous steps. This representation may be a new vector form that combines information such as the building area and building structure type in the project scale; for the text sequence of the construction technical requirements part, there is also its corresponding second implicit representation, which integrates information such as construction technology and material selection standards.

[0031] As the walking process progresses, the computer system obtains x such second implicit representations. To obtain a set of second implicit representations, the computer system can adopt a simple set construction technique, that is, combining these x second implicit representations together to form a set. From a mathematical perspective, if we denote the i-th second implicit representation as (where i = 1, 2,..., x), then the set of second implicit representations .

[0032] This way of combining multiple second implicit representations into a set enables the computer system to manage and subsequently process the information after the re - representation of each part of the target tender document as a whole. This set of second implicit representations contains rich semantic information of the target tender document, and this information has existed in a form more conducive to analysis after the previous processing, providing a comprehensive data basis for the subsequent computer system to conduct bid - possibility analysis (such as step S500), being able to reflect various requirements, conditions and other factors of the target tender document as a whole, and thus helping to accurately judge whether to bid and how to generate a suitable bid document.

[0033] Step S500: Conduct bid - possibility analysis based on the set of second implicit representations to obtain the target bid support rate.

[0034] In step S500, the set of second implicit representations is the result obtained after processing the target tender document through a series of previous steps. It contains the representation forms of various information in the target tender document after complex conversions. This information covers multiple aspects of the tender project, such as project requirements, technical specifications, commercial terms, etc.

[0035] Bid - possibility analysis is a complex process. The computer system mines information related to bid feasibility from the set of second implicit representations. A possible technical means is to construct an evaluation model. For example, a model based on probability statistics can be adopted. Assume that each element (i.e., each second implicit representation) in the set of second implicit representations is a vector (i = 1, 2, …, n, where n is the number of elements in the set), and the computer system can define a function to represent the bid - possibility score corresponding to each vector. This function can be predefined based on historical bid data. For example, for a specific type of tender project, if a certain dimension (corresponding to a specific tender requirement) in a certain vector meets specific conditions, a certain score is given.

[0036] Taking the tender for an information technology project as an example, a certain vector in the set of second implicit representations may represent the information of the technical requirements part of the project. If a certain element value in this vector indicates that the required technology highly matches the technology that the bidder is good at, then according to this function, a higher score will be given, indicating a higher possibility of bidding in terms of technology.

[0037] The computer system conducts such analysis on each element in the set of second implicit representations to obtain a series of scores . Then, in order to obtain the target bid support rate, the computer system can adopt the method of weighted average. Assume that each score The corresponding weight is , then the target bid support rate P can be calculated through the formula . These weights can be determined according to the importance of different tender requirements. For example, for the part corresponding to the core technical requirements, the corresponding weight may be higher, while for some auxiliary clauses, the corresponding weight may be lower.

[0038] Through such an analysis of the bid possibility, the target bid support rate can comprehensively reflect the feasibility degree of bidding based on the target tender document. This ratio will provide an important reference basis for whether to generate a bid document and how to generate it in the subsequent steps. If the target bid support rate is relatively high, it indicates a greater possibility of bid success, and then it is more likely to generate a bid document according to a suitable framework in the subsequent steps; on the contrary, if the target bid support rate is relatively low, it is necessary to re-evaluate the necessity of bidding or adjust the bidding strategy.

[0039] Step S600: Based on the target bid support rate, call the matching preset bid document generation framework to generate the bid document.

[0040] The target bid support rate is a quantitative index obtained through in-depth analysis of the relevant information of the target tender document in step S500, which reflects the degree of possibility of bid success.

[0041] The preset bid document generation framework is a series of pre-constructed templates or structures, which are designed for different bid possibilities or project types. For example, in a project tender for an engineering project, if it is a large-scale infrastructure construction project, there may be a preset bid document generation framework specifically for the scale of such a project and the complexity of technical requirements; while for a small building maintenance project, there is another suitable framework. These frameworks contain the structural and format requirements of each part that a bid document usually needs to cover, such as the project overview, solution, quotation strategy, project schedule plan, etc.

[0042] The computer system determines which preset bid document generation framework to call according to the numerical range of the target bid support rate. If the target bid support rate is relatively high, indicating a greater possibility of bid success, the computer system may call a more comprehensive, detailed and competitive preset bid document generation framework. For example, assuming that the target bid support rate exceeds 80%, the framework called by the computer system may elaborate on the advanced technologies adopted and unique project management methods in the solution part, and provide a more advantageous price structure in the quotation strategy part.

[0043] From a technical implementation perspective, the computer system can achieve this matching by establishing a mapping relationship table. This mapping relationship table records the corresponding relationships between different target bid support rate ranges and preset bid document generation frameworks. For example: 80% - 100% corresponds to the comprehensive competitiveness framework for large projects; 60% - 80% corresponds to the standard framework for medium-sized projects; 40% - 60% corresponds to the basic framework for small projects; 0% - 40% corresponds to the simple framework for special cases (such as only providing basic information for archiving, etc.).

[0044] After the computer system determines the matching preset bid document generation framework, it begins to generate the bid document. One possible technical means is the rule-based filling method. Taking the standard framework for medium-sized projects as an example, the project overview section in this framework may define rules for including information such as project name, tendering unit, and project background. The computer system obtains this content from the relevant information extracted during the previous processing of the target tender document and fills it into the corresponding positions of the bid document according to the specified format. For the solution part, the framework may stipulate that corresponding technical solutions, manpower arrangements, etc. should be listed according to the project requirements. The computer system generates the corresponding content according to its knowledge base or the analysis results of previous similar projects and fills it in according to the rules.

[0045] Another technical means can be the instance-based generation method. The computer system searches for instances similar to the current project in its stored historical bid document instance library (similarity can be judged based on multiple factors such as project type, scale, and technical requirements). For example, in a software project tender, if the current project is to develop an enterprise resource management system, the computer system will search for bid document instances of similar enterprise resource management system development projects in the instance library. Then, it adjusts and modifies the found instance according to the matching preset bid document generation framework to meet the specific requirements of the current project. Suppose the project schedule part in the found instance does not exactly match the time requirements of the current project, the computer system will adjust this part according to the construction period requirements of the current project.

[0046] As an implementation method, before step S100, which extracts the implicit representation of the target text sequence set of the target tender document to obtain the first implicit representation set, the method may further include:

[0047] Step S101: Divide the target tender document into text sequences to obtain a first text sequence set and a second text sequence set, and the text sequences in the first text sequence set and the second text sequence set are arranged in reverse order to each other;

[0048] Step S102: Perform a walk on the first text sequence set and the second text sequence set to obtain the target text sequence set.

[0049] Based on this, in step S500, a bid possibility analysis is performed according to the second implicit representation set to obtain the target bid support rate, including: step S500A: A bid possibility analysis is performed according to the second implicit representation set of the first text sequence set and the second implicit representation set of the second text sequence set to obtain the target bid support rate.

[0050] In step S101, when the computer system processes the target tender document, it first executes step S101. The target tender document is a document containing a lot of tender-related information, and this information exists in text form. The computer system performs text sequence division on it, aiming to organize the document content from different perspectives or orders.

[0051] For example, the content of the target tender document is: "The project requirements include building a three-story building with modern office facilities. The construction needs to be completed within 6 months, and the quality should meet the national building standards. In addition, the project budget is 5 million yuan." The computer system can divide it into text sequences according to certain rules. Assuming it is divided by sentence, for the first text sequence set, the text sequences may be formed in the original order of the sentences in the document. For example, "The project requirements include building a three-story building with modern office facilities." is a text sequence, "The construction needs to be completed within 6 months, and the quality should meet the national building standards." is another text sequence, "In addition, the project budget is 5 million yuan." is the third text sequence, etc.

[0052] The text sequences in the second text sequence set are arranged in reverse order with respect to the text sequences in the first text sequence set. This means that the first text sequence in the second text sequence set is the reverse arrangement of the last text sequence in the first text sequence set, the second text sequence is the reverse arrangement of the second last text sequence in the first text sequence set, and so on. For the above example, in the second text sequence set, the last text sequence "In addition, the project budget is 5 million yuan." will become the first text sequence, but in reverse order, such as "yuan wan 500 wei suan yu mu xiang, wai ling" (this is only for indicating the reverse arrangement, and in actual processing, appropriate operations will be carried out according to specific text processing technologies), and then so on.

[0053] The computer system can use a variety of technical means to achieve this division. A simple method is to perform sentence-level division based on punctuation marks, and then construct two sets in order and reverse order respectively. If the text of the target tender document is represented as T, the boundaries of sentences can be determined by identifying punctuation marks such as full stops and semicolons. Let (where t i(representing a single character), when a punctuation mark is encountered, a sentence is delimited as a text sequence.

[0054] In step S102, after the computer system obtains the first text sequence set and the second text sequence set, it executes step S102. Here, "wandering" means that the computer system sequentially accesses each text sequence element in the set according to a specific order.

[0055] For the first text sequence set, the computer system starts from the first text sequence and processes or analyzes each one in turn until the last text sequence. For example, for each text sequence in the first text sequence set of the above building project tender, the computer system will sequentially process text sequences regarding building construction requirements, construction period and quality requirements, project budget, etc. Similarly, for the second text sequence set, it also starts from the first (the first after reversal) text sequence and processes them in turn.

[0056] During the wandering process, the computer system can mark or number each text sequence for subsequent operations and management. For example, the text sequences in the first text sequence set can be marked as (where m is the number of text sequences in the first text sequence set), and the text sequences in the second text sequence set are marked as (where n is the number of text sequences in the second text sequence set). Through this wandering operation, the computer system recombines or filters the text sequences in the two sets to obtain the target text sequence set. This target text sequence set contains the sorting results of the content of the target tender document from different permutation order perspectives. For example, the target text sequence set may first contain some text sequences in the first text sequence set, and then contain some text sequences in the second text sequence set, arranged according to specific rules. This way of recombination helps to more comprehensively and deeply analyze the content of the target tender document in subsequent steps (such as a series of operations starting from step S100).

[0057] In step S500A, when reaching step S500A, the computer system has already obtained the second implicit representation sets corresponding to the first text sequence set and the second text sequence set respectively (this is the result obtained by processing the two sets separately in a series of previous steps, such as steps S100 - S400).

[0058] The second implicit representation set is a form of representation after complex feature extraction and transformation of the original text sequence, which contains semantic information and other content in the text sequence. For example, for each text sequence in the first text sequence set regarding construction projects (such as construction requirements, construction period requirements, etc.), after the previous processing, the elements in the corresponding second implicit representation set represent the relevant features of these text sequences in vector form.

[0059] The computer system conducts bid possibility analysis based on these two second implicit representation sets. First, for the second implicit representation set of the first text sequence set, the computer system can analyze the factors related to bid possibility reflected by each element (i.e., each second implicit representation) in it. Taking a construction project as an example, if a second implicit representation shows that the technical ability of the bidder highly matches the construction requirements of the project (analyzed from the second implicit representation obtained from the previous processing of the construction requirement text sequence), this will have a positive impact on the bid possibility.

[0060] Similarly, a similar analysis is also conducted on the second implicit representation set of the second text sequence set. For example, in the second text sequence set, there may be a second implicit representation obtained from the processing of the text sequence related to the project budget. If this representation shows that the cost control ability of the bidder can meet the project budget requirements, this is also a positive factor.

[0061] The computer system can conduct bid possibility analysis by constructing a comprehensive evaluation model. Let the second implicit representation set of the first text sequence set be (where p is the number of elements in the set), and the second implicit representation set of the second text sequence set be (where q is the number of elements in the set). A function F can be defined to comprehensively evaluate the bid possibility. For example, , where f is a function that analyzes a single second implicit representation to obtain a score related to the bid possibility, and are the weights assigned to the first text sequence set and the second text sequence set respectively, determined according to the actual situation of the project (for example, if the project pays more attention to the technical requirements in the forward-order text sequence, may be larger; if it pays more attention to factors such as the budget in the reverse-order text sequence, may be larger).

[0062] The result calculated by this comprehensive evaluation model is the target bid support rate. This ratio reflects the degree of the possibility of successful bidding after comprehensively considering the information contained in the two text sequence sets with different permutation orders, providing an important reference basis for whether to bid and how to generate a bid document in the follow-up (such as step S600).

[0063] As an implementation, in step S200, before obtaining the x-th feature accumulator by iterating the y-th feature accumulator according to the x-th first implicit representation, the method may further include: Step S201: Obtain discrete iteration variables corresponding to the continuous word embedding vectors according to the tokenization division step size.

[0064] Based on this, in step S200, obtaining the x-th feature accumulator by iterating the y-th feature accumulator according to the x-th first implicit representation may include: Step S210: Fuse the x-th first implicit representation and the y-th feature accumulator according to the discrete iteration variables to obtain the x-th feature accumulator, and the fusion of the x-th first implicit representation and the y-th feature accumulator is used to complete the iteration of the y-th feature accumulator.

[0065] In step S201, when the computer system executes step S201, it must first clarify the concept of the tokenization division step size. Tokenization is the process of splitting text into meaningful units (tokens). For example, when processing the sentence "We are conducting a construction project", tokenization may divide it into tokens such as "We", "are", "conducting", "a", "construction project", etc. The tokenization division step size is a measure that specifies how many characters or tokens to perform a specific operation in this division process.

[0066] Assume that the tokenization division step size is k. For a continuous sequence of word embedding vectors, the computer system processes them according to this step size. A word embedding vector is a representation that maps a token to a low-dimensional vector space. For example, the token "construction" may be mapped to an n-dimensional vector . A continuous sequence of word embedding vectors is a sequence formed by arranging the word embedding vectors corresponding to multiple tokens in a text in order.

[0067] Taking a simple text "This building is very tall and has many rooms" as an example, after tokenization, tokens such as "This", "building", "very tall", "has", "many", "rooms" are obtained. Assume that each token has a corresponding 3-dimensional word embedding vector (for example only). If the tokenization division step size k = 2, then the computer system will consider the word embedding vectors corresponding to "This" and "building" as a group.

[0068] The discrete iteration variable is a form of variable obtained according to the above tokenization division step size and continuous word embedding vectors. It can represent a certain relationship or feature between continuous word embedding vectors at a specific step size. The computer system can obtain the discrete iteration variable through statistical operations or mathematical transformations. For example, for each group of word embedding vectors selected according to the tokenization division step size, the sum of their Euclidean distances can be calculated as part of the discrete iteration variable. Let and are two consecutive word embedding vectors (selected according to the step size), and the Euclidean distance between them , if there are m groups of such consecutive word embedding vectors, then a part of the discrete iteration variable D can be . The computer system can also combine other statistical information, such as the mean and variance of the vectors, etc., to completely construct the discrete iteration variable.

[0069] When executing step S210, the computer system has obtained the discrete iteration variable and uses it to fuse the x-th first implicit representation and the y-th feature accumulator to obtain the x-th feature accumulator.

[0070] The x-th first implicit representation is a part of the result obtained after extracting the implicit representation of the target tender document in step S100. It contains the semantic information after the conversion of the text of a specific part in the target tender document. For example, when processing the tender document of a construction project, if the x-th first implicit representation corresponds to the text about the building structure requirements part, then this implicit representation summarizes the information related to the building structure requirements in the form of a vector.

[0071] The y-th feature accumulator is an information storage structure gradually constructed and updated during the traversal process. It accumulates the relevant information in the previous text sequence processing. When the computer system reaches the x-th first implicit representation, it uses the discrete iteration variable to fuse it with the y-th feature accumulator.

[0072] Assume the x-th first implicit representation is , the y-th feature accumulator is , and the discrete iteration variable is D. The computer system can use the weighted summation method for fusion. For example, first convert the discrete iteration variable D into a weight vector (where k is the dimension determined according to the specific composition of the discrete iteration variable). Then, the new x-th feature accumulator can be calculated by the following formula: .

[0073] This fusion method realizes the iteration of the y-th feature accumulator, so that the x-th feature accumulator contains both the new information in the current x-th first implicit representation and the historical information in the previous y-th feature accumulator. This helps to gradually integrate the information of different parts during the process of processing the target tender document, laying a foundation for subsequent accurate bid possibility analysis and bid document generation. For example, in a construction project, through this fusion, the information about the building structure requirements can be fused with the previously accumulated information such as construction technology and project budget, so that the feature accumulator can more comprehensively reflect the overall situation of the project.

[0074] As an implementation manner, in step S500, performing a bid possibility analysis based on the second implicit representation set to obtain a target bid support rate may include:

[0075] Step S510: Perform a walk on the second implicit representation set, and perform the following operations on the x-th second implicit representation reached during the walk:

[0076] Step S520: According to the x-th second implicit representation, obtain the confidence of the x-th target text sequence corresponding to the s-th normal model to obtain an appropriate confidence, where s is a positive integer less than or equal to S, and S is the number of normal models of the target bidding document;

[0077] Step S530: Perform a re-representation according to the S appropriate confidences of the x-th second implicit representation and the S first normal model coefficients of the S normal models to obtain an implicit representation of the text sequence to be analyzed;

[0078] Step S540: Perform a bid possibility analysis based on the x implicit representations of the text sequences to be analyzed obtained by walking on the second implicit representation set to obtain a target bid support rate.

[0079] In step S510, start the walk operation on the second implicit representation set. The second implicit representation set is obtained through a series of previous complex processes, and it contains the information representation after specific conversion of the target bidding document. For example, when dealing with the bidding of a construction project, each second implicit representation in this set may cover a certain comprehensive representation of information such as project scale, technical requirements, budget, etc.

[0080] The walk means that the computer system sequentially accesses each element in the second implicit representation set in a certain order, and each element here is the x-th second implicit representation (where x represents the current access serial number). This is similar to viewing each item in a list one by one, except that the items here are highly abstracted and converted second implicit representations. The computer system can implement this walk operation through indexing or pointers. For example, in programming, use a loop structure to start from the first element of the set and gradually move to the next element until all elements are accessed.

[0081] In step S520, for each x-th second implicit representation reached during the walk, the computer system needs to execute step S520. The normal model here is a pre-constructed model for analyzing the target bidding document, and each normal model may correspond to different characteristics or requirements in the target bidding document. For example, in the bidding of a construction project, there may be a normal model for analyzing the project cost budget and another for analyzing the project quality requirements, etc.

[0082] The computer system obtains the confidence corresponding to the s-th normal model according to the x-th second implicit representation. The confidence represents the degree of matching or conformity between the x-th second implicit representation and the s-th normal model. For example, assume that the x-th second implicit representation contains cost estimation information about a certain part of a construction project, and the s-th normal model is a model about cost budget. The computer system can obtain the confidence by calculating the distance or similarity between the two.

[0083] One possible technical means is to calculate the Mahalanobis distance. Let the x-th second implicit representation be the vector , and the parameters of the s-th normal model be the mean vector and the covariance matrix . Then the formula for calculating the Mahalanobis distance is . Then, based on this distance, the confidence is determined. For example, the confidence can be defined as , where is a pre-set maximum distance value.

[0084] Through such calculations, for each s-th normal model, the computer system can obtain a suitable confidence, which reflects the adaptation of the x-th target text sequence (indirectly related through the second implicit representation) to the s-th normal model.

[0085] In step S530, the computer system uses the S suitable confidences of the x-th second implicit representation obtained in step S520 and the first normal model coefficients of the S normal models for re-representation.

[0086] The first normal model coefficients (such as mean and variance) are important parameters of the normal model, which describe the distribution characteristics of the normal model. For example, for the normal model of cost budget, the mean may represent the expected average cost, and the variance represents the range of cost fluctuations.

[0087] Assume that the S suitable confidences of the x-th second implicit representation are , and the first normal model coefficients of the S normal models are respectively (here taking the mean and the variance as an example). The computer system can perform re-representation by weighted summation.

[0088] Let the implicit representation of the text sequence to be analyzed be , then , where represents an operation of adjusting the second implicit representation based on the normal model coefficients. For example, if it is the second implicit representation in vector form The operation can be 。

[0089] Through this re-representation, the implicitly represented text sequence to be analyzed incorporates the information of different normal models and the adaptation information of the x-th second implicit representation to these normal models, thus providing a new representation form for a more comprehensive and accurate analysis of the target tender document.

[0090] In step S540, when the computer system completes the random walk of the set of second implicit representations and obtains x implicitly represented text sequences to be analyzed, in step S540, these representations are used to analyze the bid possibility to obtain the target bid support rate.

[0091] Each implicitly represented text sequence to be analyzed contains comprehensive information about different aspects of the target tender document and the adaptation information to the normal model. The computer system can construct a comprehensive evaluation model to analyze the bid possibility.

[0092] For example, let the x implicitly represented text sequences to be analyzed be 。The computer system can define a function F to calculate the bid possibility score, , where f is a function that analyzes a single implicitly represented text sequence to be analyzed to obtain a score, is the weight corresponding to each implicitly represented text sequence to be analyzed, and this weight can be determined according to the importance of different parts in the target tender document.

[0093] Then, this score is converted into the target bid support rate. For example, it can be through a mapping function P = g(F), where P is the target bid support rate and g is a function that maps the score to the 0-1 interval (representing the support rate). For example, ,where and are the pre-determined minimum and maximum score values.

[0094] Through such a series of operations, the computer system starts from the set of second implicit representations, goes through steps such as random walk, obtaining confidence, and re-representation, and finally obtains the target bid support rate. This rate will provide an important reference basis for subsequent decisions on whether to bid and how to generate a bid document.

[0095] This way of gradually in-depth analysis starting from the set of second implicit representations makes full use of the information after the transformation of the target tender document, combines relevant knowledge such as normal models, and comprehensively considers all aspects of the target tender document, thus improving the accuracy and reliability of the bid possibility analysis.

[0096] As an implementation manner, in step S520, according to the x-th second implicit representation, obtaining the confidence of the x-th target text sequence corresponding to the s-th normal model to obtain an appropriate confidence may include:

[0097] Step S521: Obtain S corrected confidences between the x-th second implicit representation and the S first normal model coefficients respectively;

[0098] Step S522: Fuse the S corrected confidences to obtain a fused result of the corrected confidences;

[0099] Step S523: Among the S corrected confidences, obtain the target corrected confidence corresponding to the first normal model coefficient of the s-th normal model and the x-th second implicit representation;

[0100] Step S524: Use the weight of the target corrected confidence in the fused result of the corrected confidences as the appropriate confidence.

[0101] When the computer system executes step S521, the concepts of the x-th second implicit representation and the first normal model coefficient are clarified. The x-th second implicit representation is an abstract representation form of part of the content of the target tender document obtained through a series of previous processes, and it contains semantic information related to the target tender document. For example, when processing a tender document for a construction project, the x-th second implicit representation may cover a comprehensive vector representation of certain aspects or multiple aspects of the project progress, quality requirements, or cost budget of the construction project.

[0102] The first normal model coefficient is a parameter used to describe the characteristics of the normal model. In this context, each normal model corresponds to different aspects or requirements in the target tender document. Suppose in a construction project tender, there is a normal model used to describe the cost budget, and its first normal model coefficient may include the mean (representing the average cost budget) and the variance (representing the fluctuation range of the cost budget); another normal model is used to describe the engineering quality requirements, and its first normal model coefficient may be some statistics related to the quality standards.

[0103] In order for the computer system to obtain S corrected confidences between the x-th second implicit representation and the S first normal model coefficients respectively, specific technical means need to be adopted. A possible method is based on the calculation of the probability distribution function. Let the x-th second implicit representation be the vector , and the first normal model coefficient of the s-th normal model (s = 1, 2,..., S) be (taking the mean and the variance as an example).

[0104] For a univariate normal distribution, its probability density function is , where .

[0105] The computer system can regard each element in the x-th second implicit representation as an independent sample and calculate its probability density value under the s-th normal model . Then, through a certain comprehensive method, the corrected confidence level between the x-th second implicit representation and the s-th normal model is obtained . For example, the weighted summation method can be adopted. Let the weight vector be , then . Through such calculations, the computer system can obtain S such corrected confidence levels, and each corrected confidence level corresponds to the relationship between the x-th second implicit representation and a normal model.

[0106] In step S522, after obtaining the S corrected confidence levels , the computer system executes step S522 to fuse these corrected confidence levels.

[0107] The purpose of this fusion is to synthesize the relationships between the x-th second implicit representation and all normal models into a single result for subsequent further analysis. The computer system can adopt various fusion techniques, such as the simple average method, the weighted average method, or the fusion method based on other complex functions.

[0108] Taking the weighted average method as an example, let the weight vector be , then the fusion result of the corrected confidence levels can be calculated by the formula . The weights here can be determined according to the importance of different normal models in the entire analysis of the target bidding documents. For example, in the bidding for a construction project, if the normal model of the cost budget is relatively more important, then the weight of the corrected confidence level corresponding to the normal model of the cost budget (assuming s0 is the serial number corresponding to the normal model of the cost budget) may be set relatively high.

[0109] In step S523, the computer system, in step S523, among the already obtained S corrected confidence levels , finds the target corrected confidence level related to the s-th normal model.

[0110] This target correction confidence is specifically for the s-th normal model, which reflects the specific relationship between the x-th second implicit representation and the s-th normal model. For example, in the case of a construction project tender, if the s-th normal model is about the engineering quality requirements, then this target correction confidence represents a quantitative value of the matching degree or correlation degree between the x-th second implicit representation (which may contain some implicit information related to engineering quality) and the normal model of engineering quality requirements. The computer system directly obtains this target correction confidence corresponding to the s-th normal model from the S correction confidences through an indexing or matching mechanism. 。

[0111] In step S524, the computer system will use the target correction confidence obtained in step S523 and the fusion result of the correction confidences obtained in step S522 to determine the weight as the adaptive confidence.

[0112] The adaptive confidence represents the degree of adaptation of the x-th target text sequence (indirectly related through the x-th second implicit representation) to the s-th normal model. By calculating the weight of the target correction confidence in the fusion result, the computer system can more accurately measure this adaptability. For example, if the target correction confidence accounts for a large proportion in the fusion result , it indicates that the relationship between the x-th second implicit representation and the s-th normal model is more important or more matching in the overall relationship, resulting in a higher adaptive confidence; conversely, if the proportion is smaller, the adaptive confidence is lower.

[0113] Let the adaptive confidence be , then (assuming here ). Through such calculations, the computer system obtains the adaptive confidence, which will play an important role in subsequent steps (such as the re-representation operation in step S530), helping to further analyze the content of the target tender document in depth, and thus providing a more accurate basis for the final tender possibility analysis and tender document generation.

[0114] Through this series of steps S521 - S524, the computer system starts from the relationship between the x-th second implicit representation and each normal model, and gradually calculates the adaptive confidence. This process makes full use of the characteristics of the normal model and the information contained in the second implicit representation, laying a foundation for a more comprehensive and accurate analysis of the target tender document.

[0115] As an implementation, in step S520, according to the x-th second implicit representation, the confidence corresponding to the x-th target text sequence for the s-th normal model is obtained. After obtaining the adaptive confidence, the method may further include:

[0116] For the first normal model coefficient, the following processing is performed in a loop:

[0117] Step S520a: Obtain the second normal model coefficient of the s-th normal model based on the x adaptive confidence degrees of the s-th first normal model coefficient obtained by walking through the second implicit representation set and the second implicit representation set;

[0118] Step S520b: Iteratively adapt the confidence degrees based on the second normal model coefficient and the x-th second implicit representation to obtain the target adaptive confidence degree;

[0119] Step S520c: If the loop of the iterative processing stops, re-represent based on the S target adaptive confidence degrees of the x-th second implicit representation and the S second normal model coefficients of the S normal models to obtain the implicit representation of the text sequence to be analyzed.

[0120] The s-th first normal model coefficient is a parameter predefined to describe the characteristics of the s-th normal model, and has interacted with the x-th second implicit representation in the previous steps to generate adaptive confidence degrees. For example, in the context of a construction project tender, if the s-th normal model is about cost budget, the first normal model coefficient may include the average cost (mean) and the cost fluctuation range (variance), etc. The adaptive confidence degree reflects the adaptation degree of the x-th second implicit representation to the s-th normal model, and there is a corresponding adaptive confidence degree for each x-th second implicit representation (as walking through the second implicit representation set).

[0121] The computer system uses the x adaptive confidence degrees of the s-th first normal model coefficient obtained by walking through the second implicit representation set and the entire second implicit representation set to obtain the second normal model coefficient of the s-th normal model. A possible technical means is based on linear regression analysis.

[0122] Assume that the element in the second implicit representation set (i.e., the x-th second implicit representation) is , and the x adaptive confidence degrees corresponding to the s-th first normal model coefficient are . The computer system can construct a linear regression model with the adaptive confidence degree as the dependent variable and the respective elements of the second implicit representation as the independent variables.

[0123] For the linear regression model (where is the regression coefficient to be estimated, (which is the error term), the computer system estimates these regression coefficients by methods such as the least squares method. Then, based on the estimated regression coefficients, the second normal model coefficients of the s-th normal model are constructed. For example, the second normal model coefficients can be a certain transformation related to the regression coefficients, such as (where is the mean in the second implicit representation set) and (this is just an example construction method, and the actual situation may be more complex).

[0124] Taking the bidding of construction projects as an example, if the elements in the second implicit representation set contain information on different aspects of the project, such as project scale, construction difficulty, etc., through the above linear regression analysis, the computer system can adjust the coefficients of the normal model for cost budget (the s-th normal model) according to the relationship between this information and the appropriate confidence level, and obtain more realistic second normal model coefficients.

[0125] In step S520b, after obtaining the second normal model coefficients of the s-th normal model, the computer system executes step S520b. The goal here is to use the newly obtained second normal model coefficients and the x-th second implicit representation to iterate the previously obtained appropriate confidence level.

[0126] The purpose of iterating the appropriate confidence level is to update and optimize the original adaptation relationship based on the new model coefficient information. For example, continuing with the example of the bidding of construction projects, if the previously obtained appropriate confidence level was calculated based on the initial normal model coefficients (the first normal model coefficients), reflecting the degree of adaptation between the x-th second implicit representation and the s-th normal model (such as the cost budget model) under the initial model assumptions. Now, with the second normal model coefficients obtained based on more information (such as the overall information of the second implicit representation set), it is necessary to re-evaluate this degree of adaptation.

[0127] The computer system can adopt a method based on distance measurement to iterate the appropriate confidence level. Let the x-th second implicit representation be , and the second normal model coefficients of the s-th normal model be . The computer system can calculate a certain distance measurement between the x-th second implicit representation and the new normal model (defined by the second normal model coefficients). For example, the Mahalanobis distance (where is the vector composed of , is the covariance matrix related to ) is adopted.

[0128] Then, based on this distance measurement, the appropriate confidence level is iterated. For example, the target appropriate confidence level can be defined as (where is a pre-set maximum distance value). In this way, the target adaptation confidence comprehensively considers the new normal model coefficients and the information of the x-th second implicit representation, and more accurately reflects the adaptation relationship between them.

[0129] In step S520c, this loop stop may be based on pre-set stop conditions, such as reaching a certain number of iterations, convergence of the objective function, etc.

[0130] At this time, the computer system has S target adaptation confidences of the x-th second implicit representation and S second normal model coefficients of S normal models . The computer system uses this information for re-representation to obtain an implicit representation of the text sequence to be analyzed.

[0131] A possible re-representation method is based on weighted summation. Let the implicit representation of the text sequence to be analyzed be , then . Here represents an operation of adjusting the second implicit representation based on the normal model coefficients. For example, if , it can be defined as .

[0132] Taking the tender of a construction project as an example, if the S normal models respectively correspond to aspects such as the cost budget, engineering quality, and construction progress of the project, then through this re-representation, the implicit representation of the text sequence to be analyzed integrates the x-th second implicit representation with the adaptation information of each normal model after iterative optimization and the information of the new normal model coefficients. This fused representation can more comprehensively and accurately reflect the information related to the x-th second implicit representation in the target tender document, providing a more valuable data basis for subsequent operations such as bid possibility analysis.

[0133] As an implementation, step S520a, obtaining the second normal model coefficients of the s-th normal model based on the x adaptation confidences of the s-th first normal model coefficient obtained by walking through the set of second implicit representations, and the set of second implicit representations, may include:

[0134] Step S520a1: Fuse the x adaptation confidences of the s-th first normal model coefficient obtained by walking through the set of second implicit representations to obtain a fused result of the adaptation confidences;

[0135] Step S520a2: Fuse the x adaptation confidences and the set of second implicit representations to obtain a fused result of the confidence implicit representation;

[0136] Step S520a3: Obtain a second normal model coefficient that has a change trend consistent with the confidence implicit representation fusion result and a change trend opposite to the adaptive confidence fusion result.

[0137] When executing step S520a1, the sth first normal model coefficient is determined when the normal model is previously constructed for the target bidding document, and is used to describe the parameters of certain characteristics of the normal model. The adaptation confidence is a value that measures the degree of adaptation between the xth second implicit representation and the sth normal model. Here, each xth second implicit representation (as the second implicit representation set is wandered) has a corresponding adaptation confidence, and there are x such adaptation confidences in total.

[0138] For example, in a bidding analysis of a large infrastructure construction project, assuming that the sth normal model is about the project schedule, the first normal model coefficients may include the expected average schedule duration, the acceptable fluctuation range of the schedule, etc. For each second implicit representation (the xth second implicit representation) related to the normal model of the project schedule obtained from the second implicit representation set walk, an adaptation confidence is calculated.

[0139] The purpose of the fusion operation of the computer system is to integrate the x adaptive confidences associated with the sth first normal model coefficient into a single result for subsequent operations. A variety of technical means can be used to achieve fusion, such as weighted averaging.

[0140] Assume that the x adaptation confidences are , the corresponding weight is (These weights can be determined based on the relative importance of different second implicit representations in the overall analysis. For example, if the target text sequence corresponding to a second implicit representation contains more critical item information, then its corresponding adaptation confidence weight can be higher.) Adaptation confidence fusion results It can be calculated by the following formula: ; In this way, the computer system combines x adaptive confidences into one value , this fusion result can comprehensively reflect the overall adaptation of all second implicit representations to the sth first normal model coefficient.

[0141] In step S520a2, the computer system merges the x adaptive confidences with the second implicit representation set. The second implicit representation set includes information representations of the target tender document after a series of processing, and each element (the xth second implicit representation) reflects part of the semantic content of the tender document to a certain extent.

[0142] Continuing with the example of infrastructure construction project bidding, assume that the second implicit representation of the elements in the set (x = 1, 2, …), where each element may correspond to different aspects of the project, such as project scale, technical requirements, etc. And the fitness confidence reflects the degree of fitness of each second implicit representation with the s-th normal model.

[0143] The computer system can adopt a fusion method based on vector multiplication and summation. First, multiply each fitness confidence element-wise with the corresponding second implicit representation to obtain a new vector . Then, sum all these new vectors to obtain the fused result of the confidence implicit representation . ; This fusion method incorporates the information of the fitness confidence into the set of second implicit representations, making the fused result contain both the semantic information in the original set of second implicit representations and the information on the fitness relationship with the s-th normal model. This helps to more comprehensively mine the features related to the s-th normal model in the target bidding documents and provides a richer information basis for obtaining the coefficients of the second normal model subsequently.

[0144] In step S520a3, the computer system needs to obtain the coefficients of the second normal model based on the existing fused result. This involves the analysis of the changing trends of the fused result of the confidence implicit representation and the fused result of the fitness confidence .

[0145] The changing trend can be measured in various ways, such as calculating the first-order difference. For the fused result of the confidence implicit representation , its first-order difference can be expressed as ; For the fused result of the fitness confidence , assuming a series of values are obtained at different processing stages or for different data subsets, such as , then its first-order difference .

[0146] The coefficients of the second normal model sought by the computer system should satisfy the conditions of being consistent with the changing trend of the fused result of the confidence implicit representation and being opposite to the changing trend of the fused result of the fitness confidence at the same time. This means that if the fused result of the confidence implicit representation shows an upward trend in a certain dimension (e.g., ), then the coefficients of the second normal model should also show a change that matches this upward trend in the corresponding dimension; and if the fused result of the fitness confidence shows an upward trend at a certain stage (e.g., ), then the coefficients of the second normal model should show the opposite changing trend in the corresponding aspect.

[0147] For example, in the bidding for infrastructure construction projects, if the confidence level implicitly indicates that the dimension related to the project scale in the fusion result shows an increasing trend, this may mean that as the analysis of the bidding documents deepens, the relationship between the information related to the project scale and the s-th normal model (such as the project progress normal model) is strengthening. At this time, the part of the second normal model coefficient related to the project scale (for example, the coefficient of the impact of the project scale on the project progress) should match this increasing trend. At the same time, if the appropriate confidence level fusion result shows that the part related to the project budget rises at a certain stage, and the change in the project budget may have an inverse impact on the project progress, then the part of the second normal model coefficient related to the project budget should show the opposite change trend to reflect this complex relationship.

[0148] The computer system can determine the second normal model coefficient that meets this condition by establishing a mathematical model. For example, a model based on multiple linear regression can be constructed, taking the confidence level implicit representation fusion result and the appropriate confidence level fusion result as independent variables, and the second normal model coefficient as the dependent variable, and analyzing the relationship in the data to determine the specific coefficient value. Let be an n-dimensional vector, be a one-dimensional variable, and the second normal model coefficient be (m depends on the complexity of the normal model), then the model (where is the parameter to be estimated, is the error term) can be established, and the value of is estimated by methods such as the least squares method, so as to obtain the second normal model coefficient that meets the conditions.

[0149] Through steps S520a1 - S520a3, the computer system gradually fuses and analyzes the information related to the s-th first normal model coefficient, and finally obtains the second normal model coefficient. This process makes full use of the information in the appropriate confidence level and the second implicit representation set, provides a necessary basis for subsequent steps (such as the iteration of the appropriate confidence level in step S520b), helps to analyze the relationship between the target bidding documents and the normal model more deeply and accurately, and thus improves the accuracy of the bidding possibility analysis.

[0150] As an implementation, before step S500, when performing the bidding possibility analysis based on the second implicit representation set to obtain the target bidding support rate, the method may further include: step S501: performing an activation process based on the first implicit representation set to obtain an implicit representation significant coefficient set.

[0151] Based on this, step S500, when performing the bidding possibility analysis based on the second implicit representation set to obtain the target bidding support rate, may include:

[0152] Step S5001: Merge the second implicit representation set and the implicit representation significant coefficient set to obtain the implicit representation set to be fused;

[0153] Step S5002: Fuse the first implicit representation set and the implicit representation set to be fused to obtain the target implicit representation set;

[0154] Step S5003: Perform bid possibility analysis based on the target implicit representation set to obtain the target bid support rate.

[0155] When executing Step S501, operations are first performed on the first implicit representation set. The first implicit representation set is the result obtained by extracting implicit representations from the target text sequence set of the target bidding document in previous steps (such as Step S100). Each element (first implicit representation) in this set summarizes the semantic information of the corresponding text sequence in the target bidding document in the form of a vector.

[0156] Activation processing is an operation that performs non - linear transformation on the elements in the first implicit representation set, aiming to highlight some important features or information, thereby obtaining the implicit representation significant coefficient set. This set is essentially a weight set, used to represent the relative importance of each first implicit representation in subsequent analysis. For example, in the scenario of a construction project tender, the target bidding document contains text sequences in different aspects such as the project scale, construction technical requirements, and construction period requirements. After implicit representation extraction, the elements in the first implicit representation set respectively correspond to the information vectors of these different aspects. The computer system can use activation functions (such as the Sigmoid function, ReLU function, etc.) for activation processing.

[0157] Let a certain first implicit representation in the first implicit representation set be , if the Sigmoid function is used as the activation function, then the element j after activation processing is , and the activated vector obtained is . Each element in this activated vector is between 0 and 1, and can highlight some elements with larger values (indicating relatively more important information).

[0158] The computer system can further calculate the implicit representation significant coefficient set (weight set) based on the activated vector . For example, the weights can be determined by calculating the proportion of each element in the vector in the entire vector. Let the implicit representation significant coefficient set be , then In this way, the computer system obtains an implicitly represented set of significant coefficients, where each coefficient (weight) reflects the relative importance of the corresponding first implicit representation in the whole.

[0159] When performing step S5001, the second implicit representation set and the implicitly represented set of significant coefficients are already available. The second implicit representation set is another representation form of the target tender document information obtained through a series of processes (such as steps S100 - S400). It contains processing results different from the first implicit representation set, but also reflects various semantic information in the target tender document.

[0160] The merging operation combines each element in the second implicit representation set with the corresponding weight in the implicitly represented set of significant coefficients. For example, let the second implicit representation set be , and the implicitly represented set of significant coefficients be (here it is assumed that m = n, and the actual situation can be adjusted according to the specific processing method).

[0161] For each (j = 1, 2,..., m), the computer system combines it with the corresponding weight to obtain a new element . For example, if , , then .

[0162] By performing such an operation on each element in the second implicit representation set, the computer system obtains the set of implicitly represented sets to be fused . Each element in this set incorporates the weight information from the implicitly represented set of significant coefficients, enabling each element to be reasonably fused according to its importance in subsequent fusion operations (step S5002).

[0163] In step S5002, the computer system needs to fuse the first implicit representation set and the set of implicitly represented sets to be fused. The first implicit representation set contains the semantic information representation obtained from the early processing of the target tender document, while the set of implicitly represented sets to be fused is the result of fusing weight information based on the second implicit representation set.

[0164] The computer system can adopt various methods for fusion, such as the method of weighted summation. Let the first implicit representation set be , and the set of implicitly represented sets to be fused be (here it is assumed that q = m, and the actual situation can be adjusted according to the processing).

[0165] For each corresponding element and (k = 1, 2, …, q), the computer system calculates the fused elements , where and are pre-determined fusion weights (e.g., = 0.4, = 0.6). For example, if , , then .

[0166] By performing such fusion operations on all corresponding elements, the computer system obtains a set of target implicit representations . This set of target implicit representations fuses the information obtained from different processing stages of the target tender document, including both the original semantic information in the early first set of implicit representations and the information processed by the second set of implicit representations and fused with weights, thus providing a richer information basis for more comprehensive and accurate tendering probability analysis.

[0167] In step S5003, the computer system performs a tendering probability analysis based on the previously obtained set of target implicit representations to obtain a target tender support rate. Each element in the set of target implicit representations synthesizes information from multiple aspects of the target tender document.

[0168] The computer system can build a model based on machine learning or statistical analysis to perform the tendering probability analysis. For example, a logistic regression model is adopted. Let the set of target implicit representations be , and expand these vectors into feature vectors (r is the expanded feature dimension).

[0169] The form of the logistic regression model is , where P is the target tender support rate, are the parameters of the model. The computer system trains this logistic regression model with historical tender data (known tender document features and corresponding marks of tender success or failure) to determine appropriate parameter values.

[0170] For example, in the tendering of construction projects, if some features in the set of target implicit representations (such as the matching degree between the project scale and the bidder's ability, the matching degree between the construction technical requirements and the bidder's technical reserves, etc.) show strong correlations with tender success in historical data, then in the trained logistic regression model, the parameters corresponding to these features will reflect this correlation, thus giving appropriate weights when calculating the target tender support rate.

[0171] Through such bidding possibility analysis, the computer system obtains the target bidding support rate. This ratio can comprehensively reflect the feasibility of bidding based on the target bidding documents, and provide an important reference for subsequent bidding decisions (such as whether to generate bidding documents and how to generate bidding documents).

[0172] Through step S501 and a series of operations of steps S5001-S5003, the computer system starts from the first implicit representation set, undergoes activation processing, and multiple fusions with the second implicit representation set, and finally obtains the target bid support rate. This process makes full use of various information in the target bidding documents and improves the accuracy and reliability of the bid possibility analysis.

[0173] As an implementation mode, the target bid support rate is obtained based on the bid possibility analysis neural network processing, and the acquisition process of the bid possibility analysis neural network includes:

[0174] Step S10: Perform bidding possibility analysis on the training text of the bidding document by initializing the neural network to obtain the predicted bidding support rate;

[0175] Step S20: obtaining a target training cost according to an error between the predicted bid support rate and the bid support first verification mark of the training text of the bidding document;

[0176] Step S30: Initialize the neural network according to the target training cost training to obtain the bidding possibility analysis neural network.

[0177] When executing step S10, the initialized neural network is a neural network model with initial parameter settings (such as initial weights and biases), which is designed to process tasks related to bidding possibility analysis.

[0178] For the training texts of the bidding documents, these texts contain various information related to the project bidding, such as project requirements, business terms, technical specifications, etc. The computer system inputs these training texts into the initialized neural network.

[0179] The neural network contains multiple neuron layers, such as input layer, hidden layer and output layer. The input layer receives the vector representation of the training text of the bidding document after some encoding (such as word vector encoding). The hidden layer processes the data by linearly combining the input data (weighted summation) and applying activation functions (such as ReLU function: f(x)=max(0, x)) to mine the complex relationships in the data.

[0180] Assume that the training text of the bidding document is encoded into a vector , the weight matrix from the input layer to the hidden layer is , the bias vector of the hidden layer is , then the output of the hidden layer It can be calculated through the formula After being processed by multiple hidden layers, the predicted bid support rate is finally obtained at the output layer. For example, if there is only one neuron in the output layer, the output value y of this neuron is the predicted bid support rate, which represents the prediction of the success probability of bidding for the training text of this bidding document according to the currently initialized neural network. The value of y is between 0 and 1, and the closer it is to 1, the higher the success probability of bidding.

[0181] Taking the bidding of a construction project as an example, the training text contains information such as project scale, construction period requirements, budget, etc. The neural network analyzes and processes this information according to its initial parameter settings to obtain a predicted bid support rate.

[0182] In step S20, the computer system compares the difference between the predicted bid support rate and the prior label. The prior label is a known actual bid support rate corresponding to the training text of the bidding document or a label indicating whether the bid is successful (for example, a successful bid is marked as 1, and a failed bid is marked as 0).

[0183] There are various methods for error calculation, and the common one is the mean squared error (MSE). Let the predicted bid support rate be , and the prior label be y, then the mean squared error (where n is the number of training samples). This mean squared error is a form of representation of the target training cost.

[0184] In step S30, the computer system uses the target training cost obtained in step S20 to adjust the parameters of the initialized neural network, thereby obtaining a neural network for bid possibility analysis.

[0185] A common training method is to use the gradient descent algorithm. The gradient descent algorithm calculates the derivatives of the parameters (weights and biases) in the neural network based on the target training cost (such as the mean squared error) to find the parameter values that minimize the target training cost.

[0186] Assume that the parameters of the neural network are (including all weights and biases), and the target training cost is . According to the gradient descent algorithm, the update formula for the parameters is , where is the learning rate, which determines the step size of each parameter update, is the gradient of the target training cost with respect to the parameter .

[0187] For example, for an element in the weight matrix W, calculate its gradient , and then update it according to the learning rate ​ Value. By performing such update operations on all parameters in the neural network multiple times, the target training cost is continuously reduced until a preset stop condition is reached (such as the target training cost being less than a certain threshold or the maximum number of training rounds being reached).

[0188] After this training process, the parameters of the initialized neural network are optimized, thus becoming a neural network for bid possibility analysis that can perform bid possibility analysis more accurately. When this neural network conducts bid possibility analysis on new target tender documents subsequently, it can more reasonably analyze various information and output an accurate bid support rate based on the parameter patterns obtained from previous training.

[0189] As an implementation manner, step S30, training the initialized neural network according to the target training cost to obtain a neural network for bid possibility analysis, may include:

[0190] Step S31: Obtain the set of third normal model coefficients iterated from each of the R tender document training texts in the R tender document training texts, obtaining R sets of third normal model coefficients. The set of third normal model coefficients includes S third normal model coefficients corresponding to S normal models, where R ≥ 2;

[0191] Step S32: Fuse the R sets of third normal model coefficients to determine a fused set of normal model coefficients;

[0192] Step S33: Combine the initial set of normal model coefficients of the neural network for bid possibility analysis and the fused set of normal model coefficients to obtain an iterative set of normal model coefficients;

[0193] Step S34: Adjust the internal network configuration variables of the initialized neural network according to the target training cost to obtain updated internal network configuration variables;

[0194] Step S35: Combine the iterative set of normal model coefficients and the updated internal network configuration variables to obtain a neural network for bid possibility analysis.

[0195] When performing step S31, the tender document training text is the basic data for training the neural network for bid possibility analysis, and there are R such training texts here (R ≥ 2). The normal model is a mathematical model that is used to describe various characteristics or indicators related to tendering in this context. Each normal model has corresponding coefficients, which are called the third normal model coefficients here.

[0196] For each tender document training text, the computer system performs iterative operations to obtain a set of third normal model coefficients. For example, assume that in the scenario of a construction project tender, there are 3 (R = 3) tender document training texts. One is for a large commercial building project, another is for a small residential building project, and the other is for an industrial building project.

[0197] For these projects, assume there are 2 (S = 2) normal models, one for the cost budget normal model and the other for the project schedule normal model. For the tender document training text of the large commercial building project, during the processing by the computer system, for the cost budget normal model, a set of third normal model coefficients may be iteratively obtained, such as coefficients for the estimated average cost, the fluctuation range of the cost, etc.; for the project schedule normal model, coefficients such as the estimated average schedule duration and the acceptable fluctuation range of the schedule will also be iteratively obtained. Similarly, similar operations are performed on the tender document training texts of the small residential building project and the industrial building project, and finally 3 (R) sets of third normal model coefficients are obtained, and each set contains the third normal model coefficients corresponding to 2 (S) normal models.

[0198] The computer system can use statistical analysis and data mining techniques to perform such iterative operations. For example, for the cost budget normal model, based on information such as cost estimates and cost compositions for different project parts in the training text, the third normal model coefficients can be determined by calculating statistics such as the mean and variance. For the project schedule normal model, relevant coefficients can be determined by analyzing information such as the project duration requirements and milestone plans in the text.

[0199] In step S32, after the computer system obtains R sets of third normal model coefficients, it performs the fusion operation in step S32. The purpose of the fusion is to integrate the information in all tender document training texts to obtain a more comprehensive and representative set of normal model coefficients, that is, the normal model coefficient fusion set.

[0200] Continuing with the above example of the construction project tender, for the cost budget normal model, assume that from the R = 3 sets of third normal model coefficients for different projects (large commercial building, small residential building, industrial building), each set has a coefficient for the estimated average cost. The computer system can use the weighted average method for fusion.

[0201] Let the coefficient for the estimated average cost in the set of third normal model coefficients for the first project (large commercial building) be , and its weight be (The weight can be determined according to the scale, complexity or other relevant factors of the project. For example, a large commercial building project may be given a higher weight due to its large scale and high complexity); the corresponding coefficient of the second project (small residential building) is , and the weight is ; for the third project (industrial building), it is , and the weight is .

[0202] Then, in the set of fused coefficients of the normal model, the fusion coefficient of the average cost estimate of the normal model for cost budget can be calculated by the formula .

[0203] For the normal model of project progress and other possible normal models (S in total), they are all calculated according to a similar weighted average method to obtain the entire set of fused coefficients of the normal model. This fusion method can comprehensively consider the characteristics of different project types, making the fused coefficients better reflect the characteristics of the normal model in general cases.

[0204] In step S33, the initial set of normal model coefficients of the neural network for bid possibility analysis is combined with the set of fused coefficients of the normal model obtained previously. The neural network for bid possibility analysis has a set of initial normal model coefficients when initialized, and these coefficients are set in the network initialization stage, possibly based on some prior knowledge or simple estimates.

[0205] For example, when initializing the neural network, the initial coefficient of the average cost estimate set for the normal model of cost budget is , and the initial coefficient of the average progress duration set for the normal model of project progress is etc. (assuming there are a total of S initial coefficients corresponding to the normal models).

[0206] One way to combine the initial set of normal model coefficients and the set of fused coefficients of the normal model can be weighted summation. Let the fusion coefficient be , for the normal model of cost budget, the corresponding coefficient in the iterative set of normal model coefficients can be calculated by the formula , where is the coefficient of the normal model of cost budget in the set of fused coefficients of the normal model obtained in step S32.

[0207] Similarly, the coefficients corresponding to the other S-1 normal models are calculated in a similar manner to obtain an iterative set of normal model coefficients. This combination method not only considers the initial settings of the neural network but also incorporates the information mined from the training text of the tender documents, enabling the iterative set of coefficients to be optimized and adjusted on the original basis.

[0208] In step S34, the internal configuration variables of the initialized neural network are adjusted according to the target training cost. The target training cost is calculated in step S20 based on the error between the predicted bid support rate and the prior label, which reflects the deviation between the current prediction result of the neural network and the actual result.

[0209] The internal configuration variables of the network mainly include the connection weights between neurons and bias terms, etc. For example, for a simple neural network structure, assume that the input layer has n neurons, the hidden layer has m neurons, the weight matrix connecting the input layer and the hidden layer is W (W is an m×n matrix), and the bias vector of the hidden layer is (an m-dimensional vector).

[0210] The computer system can use the gradient descent algorithm to adjust these variables. Taking an element in the weight matrix W as an example, first, the partial derivative of the target training cost J with respect to needs to be calculated . The calculation of this partial derivative involves the forward propagation (calculating the prediction result from the input layer to the output layer) and backpropagation (calculating the influence of the error on the parameters of each layer from the output layer backward according to the error between the prediction result and the target) algorithms of the neural network.

[0211] Assume the learning rate is (a pre-set small positive number used to control the adjustment step size), then the update formula for is .

[0212] For the element in the bias vector , its partial derivative is also calculated in a similar manner, and then it is updated through .

[0213] By performing such update operations on all the connection weights and bias terms inside the network, the computer system obtains the updated internal configuration variables of the network. This process is repeated continuously to gradually improve the prediction ability of the network, with the goal of minimizing the target training cost.

[0214] In step S35, the computer system combines the iterative set of normal model coefficients and the updated internal configuration variables of the network to obtain the final neural network for bid possibility analysis.

[0215] The iterative set of normal model coefficients contains the coefficients related to S normal models, which reflect the information about the characteristics of the tender project (such as cost, schedule, etc.) comprehensively obtained from the training texts of tender documents. Updating the internal configuration variables of the network are the parameters such as the weights and biases of the neural network adjusted according to the target training cost in step S34.

[0216] For example, in the neural network for final tender possibility analysis, for the part that processes information related to cost budget, the coefficients of the normal model for cost budget in the iterative set of normal model coefficients and the updated weights, biases and other parameters will be used to construct the corresponding calculation logic. Similarly, for the processing of information in other aspects such as project schedule, the corresponding normal model coefficients and updated network parameters will also be comprehensively considered.

[0217] This joint operation enables the neural network for tender possibility analysis to have both the information related to project characteristics mined from a large number of tender document training texts (through the iterative set of normal model coefficients) and an optimized network structure (through updating the internal configuration variables of the network), so as to be able to more accurately analyze the tender possibility of new tender documents.

[0218] As an implementation manner, step S600, based on the target tender support rate, calling the matching preset tender document generation framework to generate a tender document may include:

[0219] Step S610: Call the pre-trained target tender document generation network;

[0220] Step S620: Input the general tender text and the matching preset tender document generation framework into the target tender document generation network, and obtain the tender document based on the target tender document generation network.

[0221] In step S610, it is necessary to call the pre-trained target tender document generation network. This target tender document generation network is a neural network structure specifically used to generate tender documents according to the input information. During the previous training process (related to steps S1 - S4), it has learned how to generate appropriate tender document content according to different inputs (such as tender training texts, example tender document generation frameworks, etc.).

[0222] For example, in the field of tender for construction projects, this network may have learned how to generate corresponding tender document content according to information such as the type of construction project (residential, commercial, industrial, etc.), scale, technical requirements, etc. This network contains multiple neuron layers, and the neurons are connected by different weights, which are determined during the training process and can reflect the relationship between different input features and tender document content.

[0223] In step S620, the computer system provides the general tender text and the matching preset tender document generation framework as inputs to the target tender document generation network. The general tender text contains some basic general information about the tender, such as the basic situation of the tenderer (company scale, qualifications, etc.), the basic understanding of the project, and other contents.

[0224] The preset tender document generation framework is a pre-defined structure or template that stipulates which parts the tender document should contain and the general structure of these parts. For example, a typical preset tender document generation framework may include parts such as project overview, solution, project schedule, and quotation strategy.

[0225] After the computer system passes these input information to the target tender document generation network, the neurons in the network process these inputs according to the pre-learned weights. For example, for the information about the tenderer's qualifications in the input general tender text, the neurons in the network will judge its importance in generating the content of different parts of the tender document according to the weights. If the qualification information is more important when generating the project overview part, the connection weights of the neurons related to qualifications will make the manifestation of this part of the information more prominent in the project overview.

[0226] Through layer-by-layer processing, the network gradually generates the content of each part of the tender document and finally combines them into a complete tender document. This process may involve complex calculations. For example, when generating the quotation strategy part, it may be necessary to generate a reasonable quotation according to the information such as the project scale and technical requirements in the input information, combined with the cost calculation model learned by the network.

[0227] Among them, the training process of the target tender document generation network includes:

[0228] Step S1: Obtain tender training texts and example tender document generation frameworks;

[0229] Step S2: In the tender document generation network to be trained, perform a greedy search on the tender training texts through the example tender document generation framework to obtain a reduced-dimensional representation vector of the training texts;

[0230] Step S3: In the tender document generation network to be trained, perform a beam search on the reduced-dimensional representation vector of the training texts through the example tender document generation framework to obtain an increased-dimensional representation vector of the training texts, and perform weighted fusion on the reduced-dimensional representation vector of the training texts and the increased-dimensional representation vector of the training texts through the initial fusion variable to obtain a training prediction representation vector;

[0231] Step S4: Generate a cost function based on the training prediction representation vector and the bid training text, and optimize the initial fusion variable based on the cost function until the representation vector fusion variable when the variable is stable is obtained. Determine the to-be-trained bid document generation network including the representation vector fusion variable as the target bid document generation network; the target bid document generation network is used to generate a bid document based on the bid document generation framework.

[0232] In step S2, the computer system operates on the to-be-trained bid document generation network. The guiding parameters in the example bid document generation framework are preset parameters used to guide the search process. For example, these parameters may specify the priority order when searching for important information in the bid training text.

[0233] Greedy search is a search strategy that selects the currently seemingly optimal option in each step of selection. The computer system performs greedy search on the bid training text through the guiding parameters in the example bid document generation framework. For example, during the search process, if the guiding parameters set the core technical requirements of the project as high priority, then the computer system will first look for relevant information in the bid training text about the core technologies of high-rise residential buildings (such as structural systems, seismic design, etc.).

[0234] During the search process, the computer system gradually transforms the information in the bid training text into a dimensionality-reduced representation vector. Assume that the information in the bid training text is high-dimensional (including numerous information elements such as words and sentences). Through the greedy search process, the computer system compresses and refines this information according to certain rules and weights to obtain a low-dimensional dimensionality-reduced representation vector of the training text. This vector can represent the key information in the bid training text in a more concise form, facilitating subsequent processing by the network.

[0235] In step S3, the computer system first uses the beam search method. Beam search is an improved search strategy that retains multiple (beam width number) optimal choices in each step of search, rather than only selecting one optimal choice as in greedy search. Through the example bid document generation framework, beam search is performed on the dimensionality-reduced representation vector of the training text in the to-be-trained bid document generation network. For example, in the scenario of generating a bid document for a high-rise residential building, beam search may consider multiple candidate solutions for the project solution (instead of just selecting one seemingly optimal solution), and these candidate solutions all meet the requirements of the example bid document generation framework to a certain extent.

[0236] In this process, the computer system generates the structural and parameter requirements in the framework based on multiple candidate solutions and sample tender documents, and expands the training text's low-dimensional representation vector into a high-dimensional representation vector. This high-dimensional representation vector contains more detailed information about different parts of the tender document. It is an enrichment and expansion of the low-dimensional representation vector considering multiple candidate solutions.

[0237] Then, the computer system performs weighted fusion on the training text's low-dimensional representation vector and the high-dimensional representation vector using the initial fusion variable. The initial fusion variable is a pre-set weight parameter used to control the relative importance of the two vectors during the fusion process. Let the training text's low-dimensional representation vector be , the training text's high-dimensional representation vector be , and the initial fusion variable be (0 < < 1), then the training prediction representation vector can be calculated by the formula . This training prediction representation vector synthesizes the information of the low-dimensional and high-dimensional representation vectors and is a preliminary prediction representation of the tender document content by the tender document generation network to be trained.

[0238] In step S4, the cost function is used to measure the degree of difference between the training prediction representation vector and the content of the real tender document. For example, the mean squared error (MSE) can be used as the cost function. Let the representation vector corresponding to the real tender document be , then (where n is the dimension of the vector).

[0239] Then, the computer system optimizes the initial fusion variable according to this cost function. The optimization process usually adopts an iterative method, such as using the gradient descent algorithm. In each iteration, according to the partial derivative of the cost function with respect to the initial fusion variable, the value of the initial fusion variable is adjusted to gradually reduce the value of the cost function. This process is repeated continuously until the initial fusion variable reaches a stable state, that is, further adjustment of it will not significantly improve the cost function.

[0240] When the representation vector fusion variable at the stable state of the variable is obtained, the tender document generation network to be trained containing this stable representation vector fusion variable is determined as the target tender document generation network. This target tender document generation network has been optimized and trained and can generate relatively accurate tender document content according to the input tender document generation framework. For example, when facing a new tender project for high-rise residential buildings, it can generate a tender document that meets the requirements based on the input relevant information, including parts such as a reasonable project overview, solution, and quotation strategy.

[0241] As an implementation manner, step S1, obtaining the bid training text and the example bid document generation framework may include:

[0242] Step S11: Obtain the first training objective, and obtain the second training objective corresponding to the first training objective; the second training objective includes the first training objective.

[0243] Step S12: Obtain the first bid document generation network corresponding to the second training objective, use the first bid document generation network as the basic bid document generation network for the first training objective, and obtain the training sample library corresponding to the first training objective and the basic bid document generation network; the basic bid document generation network is a trained bid document generation network.

[0244] Step S13: Obtain the bid training text with the features corresponding to the first training objective in the training sample library, and generate the example bid document generation framework based on the first training objective.

[0245] In step S11, the computer system first determines the first training objective. The first training objective is a relatively specific and targeted objective, which clarifies a specific direction for the training of the bid document generation network. For example, in the field of construction project bidding, the first training objective may be "generate a bid document for a large commercial complex construction project". Here, the "large commercial complex construction project" is a specific type, which has unique construction scale, functional requirements, commercial operation models, etc.

[0246] Then, the computer system obtains the second training objective corresponding to the first training objective. The second training objective is a broader concept, which includes the first training objective. Continuing with the construction project example, the second training objective may be "generate bid documents for all commercial construction projects". Commercial construction projects include various types such as large commercial complexes, shopping centers, office buildings, etc., so this second training objective covers the large commercial complex construction project in the first training objective.

[0247] The computer system can obtain these two objectives through a pre-established project type hierarchy or classification system. For example, in a construction project classification database, different types of construction projects are classified and stored in the order from specific to broad. The computer system can query this database and find the corresponding broader category (such as "commercial construction project") as the second training objective according to the keyword of the first training objective (such as "large commercial complex construction project").

[0248] In step S12, first, find the first bid document generation network corresponding to the second training objective. This first bid document generation network has been trained previously and is specifically used to handle the bid document generation task related to the second training objective. For example, in the context of generating bid documents for construction projects, if the second training objective is "generate bid documents for all commercial building projects", then the corresponding first bid document generation network has learned some general rules, patterns, and features for generating bid documents for commercial building projects.

[0249] The computer system uses the found first bid document generation network as the basic bid document generation network for the first training objective (such as "generate bid documents for large commercial complex building projects"). This is because although the first training objective is more specific, it is part of the second training objective, so this already trained and broader network can be used as a basis for further training.

[0250] Next, the computer system obtains a training sample library related to the first training objective and this basic bid document generation network. This training sample library contains a large amount of sample data related to the first training objective, and these data are crucial for training the bid document generation network. Taking large commercial complex building projects as an example, the training sample library may contain bid documents for numerous past large commercial complex building projects, relevant project requirement documents, descriptions of the bidders' solutions, and other materials. The computer system can obtain these data through the indexing and search functions in the data storage system. For example, if all bid-related data are stored in a large database, the computer system can retrieve the corresponding training sample library from the database based on the keywords of the first training objective (such as "large commercial complex building projects") and the identification information related to the first bid document generation network.

[0251] When the computer system executes step S13, first, it filters out bid training texts in the previously obtained training sample library that have the characteristics corresponding to the first training objective. In the example of large commercial complex building projects, these characteristics may include an extremely large building area, complex functional zoning (such as the integration of multiple functions like commerce, office, and entertainment), and high-standard building facilities (such as large parking lots, high-end fire protection and security systems, etc.). The computer system will analyze each sample in the training sample library to determine whether it contains these specific characteristics, thereby filtering out the bid training texts that meet the requirements.

[0252] For example, this screening process can be achieved through feature extraction and matching techniques. Assuming that each sample is stored in text form, the computer system can define a set of keywords or phrases related to the characteristics of large commercial complex construction projects (such as "construction area exceeding [X] square meters", "including commercial, office, and entertainment functional areas", etc.), and then perform keyword matching on each sample text, count the number of matches or the degree of importance of the matches. When a certain threshold is reached, the sample is determined to be a bid training text with the characteristics corresponding to the first training objective.

[0253] Based on these screened bid training texts, the computer system begins to construct an example bid document generation framework. This framework is generated according to the specific requirements of the first training objective, which stipulates which parts the bid document should contain and the general structure and content focus of these parts. For the bid document generation framework of large commercial complex construction projects, it may emphasize the scale of the project and the complexity of the functional combination in the project overview section; elaborate in detail in the solution section on how to meet various functional requirements while ensuring the overall coordination and safety of the building, etc.; consider the cost structure characteristics of large projects (such as land costs, large-scale construction costs, equipment and facility procurement costs, etc.) in the pricing strategy section.

[0254] The computer system can construct this framework by analyzing the content structure and common patterns in the bid training texts. For example, count the frequency of occurrence of content in different parts, the logical relationship between the content, etc., and then determine the structure and key content of the framework based on these statistical information. At the same time, some industry standards and best practices can also be referred to to improve this example bid document generation framework to ensure its rationality and effectiveness.

[0255] Through steps S11 - S13, the computer system prepares the necessary input data for the subsequent training of the bid document generation network, including specific bid training texts and an example bid document generation framework, which lays the foundation for accurately training the network to generate bid documents that meet the requirements.

[0256] As an implementation manner, the method may further include:

[0257] Step S14: If the first bid document generation network corresponding to the second training objective is not obtained, then use the basic network composed of the networks corresponding to the first training objective as the basic bid document generation network, and add an initial fusion variable to the basic bid document generation network;

[0258] Step S15: If the first bid document generation network corresponding to the second training objective is obtained, then perform the step of using the first bid document generation network as the basic bid document generation network for the first training objective.

[0259] Operations when the computer system fails to obtain the first tender document generation network corresponding to the second training objective. Here, the first tender document generation network refers to a network that has been trained and can be used for specific tender document generation tasks. For example, assume that the first training objective is to generate tender documents for "large-scale construction projects" (where "large-scale construction projects" is the first training objective), and the second training objective is to generate tender documents for "all construction projects" ( "all construction projects" includes "large-scale construction projects", which is the inclusion relationship between the second training objective and the first training objective). If the computer system fails to obtain the tender document generation network (the first tender document generation network) that has been trained for "all construction projects" (the second training objective), then the computer system will use the basic network composed of the networks corresponding to the first training objective as the basic tender document generation network. This basic network may be constructed from some basic modules related to the first training objective. For example, it may include a feature extraction module related to the basic elements of construction projects. At the same time, the computer system will add an initial fusion variable to this basic tender document generation network. This initial fusion variable may be a parameter used for weighted fusion of different representation vectors in subsequent steps. For example, it can be represented as α, and its initial value can be set to a certain value according to experience, such as α = 0.5.

[0260] Operations when the computer system obtains the first tender document generation network corresponding to the second training objective. Continuing with the above example, if the computer system has obtained the tender document generation network (the first tender document generation network) that has been trained for "all construction projects" (the second training objective), then the computer system will use this first tender document generation network as the basic tender document generation network for the first training objective ( "large-scale construction projects"). The advantage of doing this is that since "all construction projects" includes "large-scale construction projects", the tender document generation network for a more extensive project type may already contain some general features and functions applicable to a specific type (large-scale projects), thus providing a better basis for generating tender documents for a specific type ( "large-scale construction projects").

[0261] When a computer system determines whether it has obtained the first tender document generation network corresponding to the second training target, it can be achieved by querying a pre-established network repository. This repository stores information such as different types of tender document generation networks and their corresponding target types. When constructing a basic network as the basic tender document generation network, the computer system can combine the corresponding basic modules according to the feature definition of the first training target. For example, if the first training target is a "large-scale construction project", the basic modules may include a cost estimation module for large-scale projects, a construction period arrangement module for large-scale projects, etc. These modules can be combined into a basic network through specific interfaces and connection methods. For the addition of the initial fusion variable, it can be achieved by adding an adjustable parameter node in the network structure definition, and the value of this parameter node (such as ) will be adjusted according to factors such as the cost function during the subsequent optimization process.

[0262] As an implementation manner, step S4, optimizing the initial fusion variable based on the cost function until the representation vector fusion variable when the variable is stable is obtained, may include:

[0263] Step S41: If the basic tender document generation network is the first tender document generation network, optimize the initial fusion variable based on the cost function until the representation vector fusion variable when the variable is stable is obtained;

[0264] Step S42: If the basic tender document generation network is not the first tender document generation network, optimize the initial configuration variable and the initial fusion variable in the basic tender document generation network based on the cost function until the representation vector fusion variable and the network configuration variable when the variable is stable are obtained; the initial configuration variable is the configuration variable in the basic tender document generation network, the network configuration variable is the variable after the initial configuration variable is stable, and the representation vector fusion variable is the variable after the initial fusion variable is stable.

[0265] For step S41, when the computer system determines that the basic tender document generation network is the first tender document generation network, the computer system optimizes the initial fusion variable based on the cost function until the representation vector fusion variable when the variable is stable is obtained. Here, the basic tender document generation network being the first tender document generation network means that there already exists a network pre-trained for a specific training target (such as the target of generating a certain type of tender document mentioned before). For example, assume the first training target is to generate a tender document for an "electronic equipment procurement project", and the first tender document generation network is a network specifically trained to generate such tender documents. The initial fusion variable is a parameter used to perform weighted fusion on different representation vectors (such as the dimensionality-reduced representation vector and the dimensionality-increased representation vector of the training text) in the tender document generation network, denoted as . The cost function is a function used to measure the difference between the training prediction representation vector and the bid training text, and can be expressed as . The computer system minimizes the cost function by adjusting . The value of . A common technique is to use the gradient descent method, and the formula is , where is the current value, is the value for the next iteration, and is the learning rate. The computer system continuously iterates until the cost function converges. At this time, is the representation vector fusion variable when the variables are stable.

[0266] For step S42, when the computer system determines that the basic bid document generation network is not the first bid document generation network, the computer system optimizes the initial configuration variables and initial fusion variables in the basic bid document generation network based on the cost function until the representation vector fusion variable and network configuration variables when the variables are stable are obtained. Assume that the basic bid document generation network is a general network framework (not a network specifically for the current first training objective), for example, a basic bid document generation framework applicable to multiple types of projects. The initial configuration variables are the inherent parameters in this basic network framework, such as the number of network layers, the number of neurons in each layer, etc. Let the initial configuration variables be (here can be a vector representing multiple configuration parameters), and the initial fusion variable is still set as . The cost function simultaneously considers the influence of the initial fusion variable and the initial configuration variable on the difference between the training prediction representation vector and the bid training text. The computer system can also adopt a method similar to gradient descent, but this time it adjusts and simultaneously. For , the formula is similar to the above gradient descent formula ; for , , where and are the learning rates for and respectively. The computer system continuously iterates and until the cost function converges. At this time, is the representation vector fusion variable when the variables are stable, It is the network configuration variable. In this way, by processing different situations, the computer system can generate the network situation according to the basic tender document, reasonably optimize the relevant variables, and thus construct a target tender document generation network that is more suitable for the tender document generation task.

[0267] An embodiment of the present application provides a computer system, such as Figure 2 shown, the computer system 100 includes: a processor 101 and a memory 103. Among them, the processor 101 and the memory 103 are connected, such as connected through a bus 102. Optionally, the computer system 100 may further include a transceiver 104. It should be noted that in practical applications, the transceiver 104 is not limited to one, and the structure of the computer system 100 does not constitute a limitation to the embodiments of the present application.

[0268] An embodiment of the present application provides a computer system. The computer system in the embodiment of the present application includes: one or more processors; a memory; one or more computer programs, where one or more computer programs are stored in the memory and are configured to be executed by one or more processors. When the one or more programs are executed by the processor, the above method is implemented.

Claims

1. An intelligent generation method for tender documents based on natural language processing, characterized in that, The method includes: Performing implicit representation extraction on the target text sequence set of the target tender document to obtain a first implicit representation set; Performing a walk on the first implicit representation set, and performing the following operations on the x-th first implicit representation reached during the walk, where x is a positive integer less than X, and X is the number of target text sequences in the target text sequence set: obtaining discrete iterative variables corresponding to continuous word embedding vectors according to the tokenization division step size, and fusing the x-th first implicit representation and the y-th feature accumulator according to the discrete iterative variables to obtain the x-th feature accumulator. The fusion of the x-th first implicit representation and the y-th feature accumulator is used to complete the iteration of the y-th feature accumulator. If x is 1, the y-th feature accumulator is the initial feature accumulator. If x is greater than 1, the y-th feature accumulator is the feature accumulator of the y-th target text sequence, where y = x - 1; Obtaining a x-th second implicit representation according to the x-th feature accumulator; Obtaining a second implicit representation set according to the second implicit representations obtained by walking on the first implicit representation set; Performing a walk on the second implicit representation set, and performing the following operations on the x-th second implicit representation reached during the walk: obtaining the confidence of the x-th target text sequence corresponding to the s-th normal model according to the x-th second implicit representation to obtain an appropriate confidence, where s is a positive integer less than or equal to S, and S is the number of normal models of the target tender document; performing a re-representation according to the appropriate confidence of the x-th second implicit representation and the S first normal model coefficients of the S normal models to obtain an implicitly represented text sequence to be analyzed; performing a tender possibility analysis on the implicitly represented text sequence to be analyzed obtained by walking on the second implicit representation set to obtain a target tender support rate; Based on the target tender support rate, calling a matching preset tender document generation framework to generate a tender document.

2. The method according to claim 1, wherein Before performing implicit representation extraction on the target text sequence set of the target tender document to obtain a first implicit representation set, the method further includes: Performing text sequence division on the target tender document to obtain a first text sequence set and a second text sequence set, and the text sequences in the first text sequence set and the second text sequence set are arranged in reverse order to each other; Performing a walk on the first text sequence set and the second text sequence set to obtain the target text sequence set; The obtaining of the target tender support rate by performing a tender possibility analysis according to the second implicit representation set includes: Performing a tender possibility analysis according to the second implicit representation set of the first text sequence set and the second implicit representation set of the second text sequence set to obtain the target tender support rate.

3. The method according to claim 1, characterized in that, The obtaining of the confidence of the x-th target text sequence corresponding to the s-th normal model according to the x-th second implicit representation to obtain an appropriate confidence includes: Obtain S corrected confidence levels between the x-th second implicit representation and the S first normal model coefficients respectively; Fuse the S corrected confidence levels to obtain a fused result of the corrected confidence levels; Among the S corrected confidence levels, obtain the target corrected confidence level corresponding to the first normal model coefficient of the s-th normal model and the x-th second implicit representation; Take the weight of the target corrected confidence level in the fused result of the corrected confidence levels as the appropriate confidence level; After obtaining the appropriate confidence level by obtaining the confidence level of the x-th target text sequence corresponding to the s-th normal model according to the x-th second implicit representation, the method further includes: For the first normal model coefficient, perform the following processing in a loop: According to the appropriate confidence level obtained by walking through the second implicit representation set and the second implicit representation set, obtain the second normal model coefficient of the s-th normal model; Iterate the appropriate confidence level based on the second normal model coefficient and the x-th second implicit representation to obtain the target appropriate confidence level; If the loop of the iterative processing stops, perform re-representation according to the target appropriate confidence level of the x-th second implicit representation and the S second normal model coefficients of the S normal models to obtain the implicit representation of the text sequence to be analyzed; 4. The method according to claim 3, characterized in that The obtaining of the second normal model coefficient of the s-th normal model according to the appropriate confidence level obtained by walking through the second implicit representation set and the second implicit representation set includes: Fuse the appropriate confidence levels obtained by walking through the second implicit representation set to obtain a fused result of the appropriate confidence levels; Fuse the appropriate confidence levels and the second implicit representation set to obtain a fused result of the confidence level and the implicit representation; Obtain the second normal model coefficient that is consistent with the change trend of the fused result of the confidence level and the implicit representation and opposite to the change trend of the fused result of the appropriate confidence levels; 5. The method according to any one of claims 1 to 4, characterized in that, Before obtaining the target bid support rate based on the bid possibility analysis of the second implicit representation set, the method further includes: Perform activation processing on the first implicit representation set to obtain a set of implicit representation significant coefficients; The obtaining of the target bid support rate based on the bid possibility analysis of the second implicit representation set includes: Merge the second implicit representation set and the set of implicit representation significant coefficients to obtain a set of implicit representations to be fused; Fuse the first implicit representation set and the set of implicit representations to be fused to obtain a target set of implicit representations; Perform bid possibility analysis based on the target set of implicit representations to obtain the target bid support rate; 6. The method according to any one of claims 1 to 4, characterized in that, The target bid support rate is obtained through processing by a bid possibility analysis neural network. The obtaining process of the bid possibility analysis neural network includes: Perform bid possibility analysis on the training text of the bidding document through an initialized neural network to obtain a predicted bid support rate; Obtain the target training cost according to the error between the predicted bid support rate and the prior label of the bid support rate of the training text of the bidding document; Train the initialized neural network according to the target training cost to obtain the bid possibility analysis neural network.

7. The method according to claim 6, characterized in that, The training of the initialized neural network according to the target training cost to obtain the bid possibility analysis neural network includes: Obtain the third normal model coefficient sets iterated from each of the R tender document training texts in the R tender document training texts, to obtain R third normal model coefficient sets, where the third normal model coefficient set includes S third normal model coefficients corresponding to S normal models, and R≥2; Fuse the R third normal model coefficient sets to determine a fused normal model coefficient set; Combine the initial normal model coefficient set of the bid possibility analysis neural network and the fused normal model coefficient set to obtain an iterated normal model coefficient set; Adjust the internal network configuration variables of the initialized neural network according to the target training cost to obtain updated internal network configuration variables; Combine the iterated normal model coefficient set and the updated internal network configuration variables to obtain the bid possibility analysis neural network.

8. The method according to any one of claims 1 to 4, characterized in that Based on the target bid support rate, call a matching preset bid document generation framework to generate a bid document, including: Call a pre-trained target bid document generation network; Input the general bid text and the matching preset bid document generation framework into the target bid document generation network, and obtain a bid document based on the target bid document generation network; Among them, the training process of the target bid document generation network includes: Obtain bid training texts and an example bid document generation framework; In the bid document generation network to be trained, perform a greedy search on the bid training texts through the example bid document generation framework to obtain a reduced-dimensional representation vector of the training texts; In the bid document generation network to be trained, perform a beam search on the reduced-dimensional representation vector of the training texts through the example bid document generation framework to obtain an increased-dimensional representation vector of the training texts, and perform weighted fusion on the reduced-dimensional representation vector of the training texts and the increased-dimensional representation vector of the training texts through an initial fusion variable to obtain a training prediction representation vector; Based on the training prediction representation vector and the bid training text generation cost function, optimize the initial fusion variable based on the cost function until a representation vector fusion variable when the variable is stable is obtained, and determine the bid document generation network to be trained including the representation vector fusion variable as the target bid document generation network; the target bid document generation network is used to generate a bid document based on the bid document generation framework.

9. A computer system, characterized in that, Including: One or more processors; A memory; One or more computer programs; Among them, the one or more computer programs are stored in the memory and are configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processor, the method described in any one of claims 1 to 8 is implemented.

Images