Intelligent bid invitation agent whole-process service management method and system

Abstract

The invention belongs to the technical field of bid invitation management, and discloses an intelligent bid invitation agent full-process service management method and system, and the method comprises the steps: S1, obtaining a structure semantic vector based on a building information model, and obtaining a regulation term adaptation vector based on the structure semantic vector; s2, generating a bid invitation file draft based on the structure semantic vector and the regulation clause adaptation vector; s3, generating a structured question answering response set based on the questions proposed by the bidder and the bid invitation file draft; and S4, generating an audit report based on the bid invitation file draft and the structured question answering response set. According to the method, the industrial pain points that an existing system cannot process logic linkage between design information and regulation content and cannot guarantee consistency of question answering and document content are broken through, and service system reconstruction of'structure visualization, logic controllability, language review and regulation traceability 'of a bid invitation agent process is realized; the method is a key intelligent method path for proxy function remodeling in a whole-process engineering consultation mode.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bidding management, in particular to an intelligent bidding agent full-process service management method and system. BACKGROUND

[0002] With the continuous expansion of construction project scale in China, the increasing specialization of design, and the continuous improvement of supervision policy, the bidding agent work is shifting from traditional clerical affairs processing to a full-process service management mode based on data-driven. In this transition process, how to achieve information synchronization, semantic consistency and standard unification among different participants has become a core challenge. Under the existing mode, the agent usually processes the preparation of bidding documents, responds to questions and conducts normative review through fixed templates and manual experience. However, in the face of the actual situation of BIM model semantic complexity, changing regulations and highly specialized problems of bidders in medical, school and underground space projects, the traditional process often has problems such as clause adaptation error, content disconnection and response conflict, which leads to frequent complaints and even project flow. At the same time, although some information platforms introduce process node control functions, they generally lack the ability to understand the structure of "content itself" and make cross-stage correlation judgments, and cannot establish a complete closed-loop mechanism from design semantics to regulation reference control, to the consistency review of document content expression and response. In the existing system, the association between BIM model and bidding documents is broken, the activation of regulation clauses lacks traceability, and the response content cannot verify its consistency, which leads to the non-intelligent state of "compliance responsibility on people not on system" of the agent service. It is urgent to reconstruct the logic of the agent's behavior driven by the full-link semantic model to establish an intelligent management method from design semantics to document control, feedback through response generation and finally audit closed loop. SUMMARY

[0003] The purpose of the present application is to provide an intelligent bidding agent full-process service management method and system to solve the problems raised in the background.

[0004] In a first aspect, the present application provides an intelligent bidding agent full-process service management method, comprising:

[0005] S1, obtaining a structural semantic vector based on a building information model, and obtaining a regulation clause adaptation vector based on the structural semantic vector;

[0006] S2, generating a draft of a bidding document based on the structural semantic vector and the regulation clause adaptation vector;

[0007] S3, generating a structured response set based on the questions raised by the bidders and the draft of the bidding document;

[0008] S4, generating an audit report based on the draft of the bidding document and the structured response set.

[0009] Preferably, the structural semantic vector is obtained based on the building information model, comprising:

[0010] S10, obtaining original fields contained in each component object in the building information model respectively;

[0011] S11, converting the original fields into semantic labels based on a preset rule template;

[0012] S12, encoding all the semantic labels into a structural semantic vector.

[0013] Preferably, the regulation clause adaptation vector is obtained based on the structural semantic vector, comprising:

[0014] obtaining a regulation clause trigger matrix;

[0015] calculating a regulation clause adaptation vector based on the structural semantic vector and the regulation clause trigger matrix.

[0016] Preferably, S2 comprises:

[0017] S20, obtaining a paragraph template library;

[0018] S21, calculating a paragraph activation vector of each paragraph template in the paragraph template library based on the structural semantic vector and the regulation clause adaptation vector respectively;

[0019] S22, storing the number of the paragraph template corresponding to the component in the paragraph activation vector which is greater than a set threshold in a set ;

[0020] S23, screening out activated paragraph templates from the paragraph template library , ;

[0021] S24, performing field filling on the field slots in the paragraph templates;

[0022] S25, splicing all the filled paragraphs in module order to generate a draft of the tender document.

[0023] Preferably, S3 comprises:

[0024] S30, respectively constructing a semantic representation vector for each paragraph in the draft of the tender document ;

[0025] S31, encoding questions raised by the tenderers into vectors ;

[0026] S32, based on and calculating a matching score between the passage and the question;

[0027] S33, obtaining the passage corresponding to the highest matching score , extracting the passage associated structural semantic label and the activated set of regulation clause numbers ;

[0028] S34, generating a reply using a preset structured template;

[0029] S35, calculating a semantic consistency score based on the semantic encoding of the reply and the semantic representation vector of the passage ; ;

[0030] S36, respectively recording each reply as a structured five-tuple, and storing all structured five-tuples in a set of question-answering responses.

[0031] Preferably, S4 comprises:

[0032] S40, obtaining an unmatched set based on the set of question-answering responses;

[0033] S41, calculating a structural semantic coverage rate;

[0034] S42, obtaining an unreferenced set;

[0035] S43, calculating a legal clause reference completeness rate;

[0036] S44, obtaining a reply offset set;

[0037] S45, obtaining a term conflict set;

[0038] S46, generating an audit report based on the unmatched set, the structural semantic coverage rate, the unreferenced set, the legal clause reference completeness rate, the reply offset set, and the term conflict set.

[0039] Preferably, S40 comprises:

[0040] S400, extracting the structural semantic label set of all passages in the set of question-answering responses;

[0041] S401, comparing the structural semantic label set with the structural semantic vector to construct a structural semantic coverage matrix;

[0042] S402, respectively determining whether each dimension k satisfies a preset condition;

[0043] S403, taking the structural feature corresponding to the dimension that does not satisfy the preset condition as a structural semantic omission item,

[0044] S404, form a non-matching set by numbering all structural semantic missing items .

[0045] Preferably, S42 comprises:

[0046] S420, respectively acquire each paragraph corresponding clause reference subset;

[0047] S421, for each activated clause, determine whether it exists in any one clause reference subset, if not, store the number of the clause in the non-reference set.

[0048] Preferably, S44 comprises:

[0049] S440, respectively acquire the semantic consistency score of each reply from the question answering response set;

[0050] S441, store the index of the reply with a semantic consistency score less than a set score threshold in the reply offset set.

[0051] In a second aspect, the present application provides a smart bidding agent full-process service management system, comprising an acquisition module, a first generation module, a second generation module and a third generation module;

[0052] The acquisition module is used to acquire a structural semantic vector based on a building information model, and acquire a regulation clause adaptation vector based on the structural semantic vector;

[0053] The first generation module is used to generate a bidding document draft based on the structural semantic vector and the regulation clause adaptation vector;

[0054] The second generation module is used to generate a structured question answering response set based on the questions raised by the bidders and the bidding document draft;

[0055] The third generation module is used to generate an audit report based on the bidding document draft and the structured question answering response set.

[0056] Compared with the prior art, the present application has the following beneficial effects:

[0057] The present application is based on the BIM model output in the architectural design stage, a set of regulation clauses triggered by the actual conditions of the project is constructed by extracting the structural semantic vector and mapping the regulation knowledge graph, which is used as the logical starting point of the automatic generation of the bidding document; on this basis, the system constructs the traceable document structure content through the paragraph template and the clause reference matrix, and uses the content as the core semantic control source of the answer response generation, so as to realize the semantic closure of the answer content in the language and legal level; finally, the system takes the design semantics, regulation activation, document paragraph and answer content as the unified structure variable to participate in the whole-process consistency audit, automatically identifies the problems such as structural semantic omission, regulation reference missing, term expression conflict and answer semantic deviation, and outputs the complete and auditable content review report. The method takes the structural vector and the reference matrix as the main line, establishes the closed loop logic of information input-content generation-interaction control-output audit between stages, breaks through the industry pain points that the existing system cannot process the logical linkage between design information and regulation content, and cannot guarantee the consistency of the answer and the document content, realizes the reconstruction of the service system of the bidding agent process, that is, the "structural visibility, logical controllability, language reviewability and regulation traceability", and is the key intelligent method path for the remodeling of the agent function under the whole-process engineering consulting mode. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 FIG. 1 is a schematic diagram of the intelligent bidding agent whole-process service management method of the present application;

[0059] Figure 2 FIG. 2 is a schematic diagram of the intelligent bidding agent whole-process service management system of the present application. DETAILED DESCRIPTION

[0060] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0061] Please refer to Figure 1 The present application provides a technical solution: an intelligent bidding agent whole-process service management method, comprising:

[0062] S1, obtaining a structural semantic vector based on a building information model, and obtaining a regulation clause adaptive vector based on the structural semantic vector.

[0063] The task of this step is to convert the building information model provided by the architectural design unit into an intermediate representation with semantic structuring capability, and to drive the adaptation and selection of regulatory provisions, providing technical basis and compliance boundaries for the subsequent tender document generation process. The innovation of this process lies in that it is not simply reading the parameters in the building model, but according to the actual "design semantics" of the model's spatial configuration, component combination, building purpose, etc., a set of structural tag system is established, and through mapping method, it establishes corresponding relationship with the triggering conditions of regulations, so that the model content can automatically guide the selection of tender compliance provisions.

[0064] The input of this step is the building information model exported from the architectural design file, denoted as variable .

[0065] In specific operation, the model is usually exported as a.rvt file by the design unit after modeling in Autodesk Revit, or converted into a general IFC format (i.e. "Industry Foundation Classification Standard"), which is the mainstream open BIM data exchange format in the industry at present.

[0066] The model file will be transmitted into the semantic analysis module of the system for layer-by-layer component recognition and attribute extraction. The data collection method in the model is automatic field reading, which does not rely on manual reporting. Taking Revit file as an example, the system accesses the internal data structure of the model by calling its application programming interface (API) or using middleware (such as Dynamo, RevitPythonWrapper), and completes data extraction without opening the model interface.

[0067] During the extraction process, the system will obtain the following three types of core information:

[0068] (1) Component attributes, including component category (such as wall, column, beam, plate), size (length, width, height), material (such as reinforced concrete, brick masonry), whether load-bearing, etc.;

[0069] (2) Spatial configuration, including the functional type of each floor room (such as residential unit, corridor, equipment room, underground parking space), spatial position (above ground / underground), floor height and area, etc.;

[0070] (3) System identification information, such as whether the model contains air conditioning system, smoke exhaust system, electrical power distribution system, etc., which is usually identified by component category "air pipe", "cable bridge" or system naming containing keywords "HVAC" or "EPS" etc.

[0071] Further, based on the building information model, a structural semantic vector is obtained, including:

[0072] S10, respectively obtaining the original fields contained in each component object in the building information model.

[0073] The system first extracts the type, floor, space use, structure identification, etc. from each component object through model traversal operation.

[0074] S11, the original field is converted into semantic label based on the preset rule template.

[0075] The implementation method of the preset rule template is as follows:

[0076] For example, if the "use" attribute of a room is "Operating Room", it is automatically classified as "clean medical area"; if the component category is "basement floor" and the floor number is "-2", it is determined that the model has "two-story underground structure"; if "glass curtain wall" is found in the model and its projection area exceeds 100 square meters, the system marks the model as having "high-performance external envelope structure".

[0077] S12, all semantic labels are coded into structure semantic vector Each dimension represents a structure semantic label.

[0078] All dimensions are discrete Boolean values (1 represents existence, 0 represents nonexistence), or integer values mapped in coding intervals (such as building height: <24m is 0, 24-50m is 1, >50m is 2).

[0079] For example, a model is marked as "medical building", the structure system is "reinforced concrete frame shear", the number of floors is 12, the ground height is 55.8 meters, the total area is 18300 square meters, there is a three-story underground and garage system, so its "building use" dimension is medical (value = 1), "structure system" dimension is frame shear (value = 1), "number of floors" dimension is high-rise (value = 2), "underground structure" dimension is present (value = 1), and "seismic fortification grade" dimension is activated automatically.

[0080] The dimensions are fixed and controlled by system rules, each dimension represents whether a structure feature exists or its classification code, which is directly derived from the attribute analysis of .

[0081] Further, based on the structure semantic vector, a regulation clause adaptation vector is obtained, including:

[0082] Obtain the regulation clause trigger matrix.

[0083] The regulation clause trigger matrix has a size of Each row represents the structure semantic conditions required to meet a clause, which is defined by engineering experience templates.

[0084] The regulation clause adaptation vector is calculated based on the structural semantic vector and the regulation clause trigger matrix:

[0085] ;

[0086] Wherein:

[0087] : The regulation clause adaptation vector, with a length of , wherein is the total number of regulation clauses in the system. Each dimension represents whether the corresponding clause is activated by the current project conditions;

[0088] : The structural semantic vector is automatically parsed and constructed from the building model , containing fields such as building purpose, number of floors, underground structure, structure type, etc., with a dimension of ;

[0089] : The bias vector, with a length of , is used to represent whether the mandatory clause is unconditionally reserved.

[0090] For example, if the "structure type" dimension value in is "frame-shear", and the "building height" dimension is "high-rise", the 5th clause in the "Code for Seismic Design of Tall Buildings" has a product relationship with the above two dimensions, and the corresponding bit calculation value in will reach the activation threshold, so it is determined that the clause is triggered.

[0091] The dimension of is equal to the number of clauses in the regulation library, where each bit represents whether the corresponding clause is activated by the model semantic conditions, and its calculation is realized by the above matching matrix.

[0092] To maintain logical consistency in calculation, the system normalizes and encodes all model original fields before generating , ensuring that there is no logical conflict or mismatch in matrix operations for variables of different dimensions.

[0093] S2, based on the structural semantic vector and the regulation clause adaptation vector, generates a draft tender document.

[0094] The goal of this step is to generate a structured draft tender document , which is inputted from , respectively derived from the structural semantic extraction and regulation clause adaptation steps of the building model. is a vectorized representation, recording the core structural features in the building model, such as building purpose, structural system, building height interval, existence of underground structure, special function room type (such as operating room, clean area, teaching area), etc. Each dimension is derived from the component field or spatial identification field of the model .

[0095] Further, S2 includes:

[0096] S20, obtaining a paragraph template library:

[0097] Paragraph template library Each template contains the following fields:

[0098] (1) paragraph title;

[0099] (2) applicable semantic label (such as applicable to "medical building", "underground structure");

[0100] (3) fillable field slot (such as "

structure type

number of floors

clause number

[0101] S21, calculating the paragraph activation vector of each paragraph template in the paragraph template library based on the structure semantic vector and the regulation clause adaptation vector:

[0102] ;

[0103] Wherein is the paragraph activation vector, the length of which is the number of templates in the template library , each bit represents the activation strength of the template .

[0104] is a matrix constructed by the logical relationship between the template and the semantics, which is obtained by the system according to artificial rules or historical document frequency training. For example, the "green building design requirements" module template has high dependence on the "glass curtain wall" structural feature dimension and the green building clause number dimension.

[0105] is the coefficient of the influence degree of the control regulation on the paragraph generation; is the bias value of each paragraph, which identifies whether the paragraph is required content. For example, the "bidding scope description" paragraph should be generated even if there is no semantic label.

[0106] is a standard Sigmoid function, which normalizes the variable in the parentheses to Interval.

[0107] S22, the components in the paragraph activation vector greater than the set threshold value, the component corresponding to the paragraph template number is stored in the set ;

[0108] Set threshold value , such as 0.5, select greater than the threshold value The number of paragraph templates in the set .

[0109] S23, from the paragraph template library to filter out the activated paragraph template , ;

[0110] The filtering process is to traverse the paragraph template library , check if the number of each paragraph template is in If so, add the template to the activated paragraph template set as the object of subsequent field filling and splicing.

[0111] S24, field filling for field slots in paragraph templates:

[0112] Each paragraph template Usually contains 1-4 field slots that can be dynamically filled. For example, the template "structural design requirements" contains the sentence: "This project adopts

structure type

building height

clause number

[0113] For structural features in , such as "frame-shear wall structure" or "height 52 meters", directly replace the corresponding value in the slot;

[0114] For activated clause numbers in , such as Article 5.2 of "Code for Seismic Design of Tall Buildings", find the content, number and citation format in the regulations library for insertion;

[0115] If a slot cannot be derived from And , the system marks the paragraph as "to be supplemented" state, and inserts placeholder content such as "

field missing: roof drainage standard

[0116] For example, if the structure type in the model is "frame-shear", and the height is 55.8 meters, the field "

structure type

building height

Article Number

[0117] S25: Combine all the filled paragraphs in module order to generate a draft tender document.

[0118] For example, the sections should include "Project Overview" → "Design Basis" → "Technical Standards" → "Evaluation Methods" → "Contract Key Points," and record the sources cited in each paragraph. Which dimensions and Which clause values ​​are involved in the generation?

[0119] The module order here refers to the pre-set arrangement of the chapters in the tender document according to industry-standard or project-specific preparation specifications. For example, it is usually arranged in the following order: "Project Overview" first, "Design Basis" second, "Technical Standards" third, "Evaluation Method" and "Contract Key Points". When the system combines the completed paragraph templates into a draft tender document, it will strictly follow this predetermined chapter order to ensure that the generated document has a complete structure, clear logic and meets the specifications.

[0120] Final draft tender document The definition is as follows:

[0121] ;

[0122] in This represents the paragraph text after it has been filled with fields. Concat is an operation that concatenates the text in a preset order. This is a draft of the tender document.

[0123] While the document is being generated, the system constructs a triple index matrix of paragraphs, semantic tags, and legal clauses, and outputs an audit mapping table for subsequent steps to track consistency, locate questions, and review risks.

[0124] S3 generates a structured Q&A response set based on the questions raised by the bidders and the draft bidding documents.

[0125] This step aims to generate standardized responses to bidders' questions, ensuring complete content, proper wording, and semantic consistency with the aforementioned bidding documents. A structured consistency verification mechanism is introduced to guarantee that the response logic does not deviate from or conflict with the original text. In actual bidding agency processes, the Q&A session is the most direct information exchange between the bidding entity and bidders. The accuracy and reasonableness of the responses, and their consistency with existing content, applicable clauses, and terminology in the bidding documents, directly impact the direction of bid preparation, the evaluation criteria, and potential objections.

[0126] Further, S3 includes:

[0127] S30, respectively, the draft tender document Each paragraph Construct semantic representation vector :

[0128] This vector is generated by a two-level network:

[0129] The first layer is a context modeling layer based on word embedding, which uses a domain-specific language model to construct a sentence vector (for example, a Transformer encoder trained based on Chinese tender language);

[0130] The second layer is a sentence structure attention mechanism layer, which identifies the "regulatory semantic focus" and "structural semantic focus" in the paragraph, such as the sentence segment containing the keywords "frame-shear wall structure" and "three-level waterproofing", and increases its attention weight to enhance the model's positioning ability in the comparison process.

[0131] S31, the questions raised by the tenderer Encoded as a vector The method is consistent with paragraph encoding to ensure semantic space comparability.

[0132] S32, based on And Calculate the matching score between the paragraph and the question :

[0133] In the semantic comparison stage of the paragraph and the question, the system introduces a matching score function specially designed for this patent scenario to identify the best answer paragraph:

[0134] ;

[0135] Where:

[0136] is the basic semantic similarity calculation;

[0137] is the label overlap between the paragraph and the question in the structural semantic label space, for example, if involves "basement", and corresponds to "underground structure = 1" in the paragraph, it is considered to be overlapped;

[0138] is the "term ambiguity index" of the paragraph, indicating whether there are ambiguous expressions, incomplete fields, colloquial language residues, etc.

[0139] and These are the structural label enhancement factor and the fuzzy penalty factor, respectively, with empirical values ​​of 0.3 and 0.5.

[0140] The innovation of this function lies in the fact that it does not use pure text semantic similarity as the criterion, but instead incorporates structural consistency. Language risk penalties This better aligns with the actual needs of the bidding agency industry for "precise logic" and "conservative language." S33: Obtain the paragraph corresponding to the highest matching score. Extract paragraphs The associated structural semantic tags and the set of activated regulatory clause numbers ;

[0141] S34, generate a response using a preset structured template:

[0142] The response generation uses a preset structured template sentence format, such as:

[0143] Template sentence 1 (semantic direct response type): "Based on the content of the [module name] in the [paragraph number] section of the tender document, and in conjunction with the provisions of the [legal clauses], the reply is as follows: [standard statement]".

[0144] Template Sentence 2 (Legal Interpretation): "Based on the definition of [Term / Behavior] in [Legal Name] [Article Number], this item has already provided an explanation in [Module Name], specifically: [Paragraph Excerpt]."

[0145] The slots in the template are filled in the following way:

[0146] [Paragraph Numbering] originates from Index;

[0147] [Module Name] is by The paragraph structure marker field is obtained;

[0148] [Legal Provisions] The name and number of the clause with the highest number in the middle;

[0149] The standard sentences or paragraph excerpts are edited from the paragraph content, or are derived from... The mapping fields have been reorganized.

[0150] For example, when The system matched the paragraph as "Please clarify the roof drainage method". The cited clause is Article 7.2.3 of the "Standard for Design of Building Water Supply and Drainage", and If the "Roof Structure Type" is "Accessible Roof", the system will generate the following response:

[0151] According to the tender document No. 12, the roof structure drainage design specification, according to the building water supply and drainage design standard 7.2.3, the project adopts the bidirectional drainage ditch + external rainwater vertical pipe drainage system, which meets the 24-hour maximum rainfall intensity drainage capacity of the roof.

[0152] S35, based on , the corresponding semantic encoding of the answer and the semantic representation vector of the paragraph Calculate the semantic consistency score :

[0153] ;

[0154] Wherein:

[0155] is the semantic encoding of the answer statement;

[0156] is the redundancy score of the answer, indicating whether there are repeated references, term abuse or sentence length in the answer;

[0157] is the redundancy penalty coefficient, which is empirically set to 0.2;

[0158] is the final consistency score, and the system takes as the artificial review threshold.

[0159] , are the similarity weight coefficients of question and paragraph matching and answer and paragraph matching, respectively; .

[0160] The calculation method of the answer redundancy score is as follows: first, extract all the sentences in the answer text that refer to the tender document. If a sentence appears more than once in the answer, it is considered repeated reference, and the ratio of the number of repeated references to the total number of references is calculated to get ; Second, extract all the professional terms in the answer text. If a term appears more than once, it is considered redundant, and the ratio of the number of redundant terms to the total number of terms is calculated to get ; Third, count the total number of words in the answer text, set the word threshold (e.g. 200 words), if the total number of words exceeds , then the ratio of the number of words exceeding the threshold to the total number of words is taken as , otherwise ; Finally, add the three proportion values directly to get , which is the redundancy score of the answer, which is used The redundancy degree of the reply is reflected in the calculation formula.

[0161] S36, each reply record is recorded as a structured five tuple respectively, and all the structured five tuples are stored in the question answering response set .

[0162] Structured question answering response set , each of which contains: original question , system generated reply , quoted paragraph , activated regulatory clause number set and semantic consistency score . The data is stored in table structure or JSON structure, and the system provides question answering traceability, semantic risk prompt and question and answer version control function, which is suitable for standard question answering announcement, regulatory data archiving and subsequent compliance dispute traceability.

[0163] S4, generating an audit report based on the draft tender document and the structured question answering response set.

[0164] The goal of this step is to build a reviewable, traceable and quantifiable audit mechanism based on the completion of the tender document generation and the question answering response, and to generate the final audit report from four dimensions of structural semantics, regulatory coverage, question answering consistency and term expression.

[0165] Further, S4 includes:

[0166] S40, obtaining an unmatched set based on the question answering response set.

[0167] Further, S40 includes:

[0168] S400, extracting the structural semantic tag set of all paragraphs in the question answering response set , denoted as , representing the explicit reference structural semantic tag subset in the th paragraph (explicitly imported by the paragraph generation module and recorded).

[0169] S401, compare the structural semantic tag set with the structural semantic vector to build a structural semantic coverage matrix;

[0170] This matrix is the basic data source for calculating the structural semantic coverage . Each row in the matrix represents a dimension of global structural semantic tag , and each column represents a generated paragraph template instance , and the Boolean value of the matrix records whether the paragraph covers the semantic tag. The generation method of the matrix is as follows: first, from each paragraph template instance The middle reads its bound semantic tag set ; then traverse the global semantic tag vector of each dimension If the paragraph template covers the tag (i.e. ), assign 1 to the matrix at the row and the column, otherwise 0. The generated matrix has each row corresponding to a semantic tag and each column corresponding to a paragraph, 1 indicating coverage and 0 indicating no coverage.

[0171] S402, respectively, determine whether each dimension k satisfies the preset condition:

[0172] The system sequentially determines whether each dimension satisfies the following condition:

[0173] Under the premise that , whether there is a paragraph such that .

[0174] indicates that in the global structural semantic vector , the value of the dimension is 1, which means that in the design model (BIM data analysis result), the structural feature corresponding to the semantic tag exists and is activated, such as "two-story underground structure" and "frame-shear wall system".

[0175] indicates that in the semantic tag set of the paragraph template instance, the value of the dimension is 1, which means that the content of the paragraph template references or covers the structural feature corresponding to the semantic tag, for example, the paragraph writes about "two-story underground structure" or references its related technical provisions.

[0176] Therefore, is to confirm the existence of the feature from the perspective of design data, and is to confirm that the feature is used or expressed by a paragraph from the perspective of document generation. The combination of the two can be used to check whether the features existing in the design are all reflected in the document, thereby supporting the calculation of structural semantic coverage.

[0177] S403, the structural feature corresponding to the dimension that does not satisfy the preset condition is taken as a structural semantic omission item;

[0178] The structural feature here is the global structural semantic vector The semantic label corresponding to the dimension k in the model, so it is one-to-one correspondence with the aforementioned "dimension k".

[0179] If the condition is not met, it means that the structural feature exists in the model, but is not expressed by any paragraph, which is a structural semantic missing item.

[0180] S404, form the unmatched set of all structural semantic missing items , and according to the category of the label (such as "structure", "function", "equipment"), it is suggested that the module position should be inserted in the document.

[0181] For example, if the "underground structure" label is missing, it is suggested to be inserted into the "structure design description" part.

[0182] S41, calculate the structural semantic coverage :

[0183] ;

[0184] The closer the value is to 1, the higher the consistency between the model design semantics and the document expression. If , the system automatically triggers a structural risk prompt.

[0185] S42, get the un-referenced set.

[0186] Further, S42 includes:

[0187] S420, respectively get the clause reference subset corresponding to each paragraph.

[0188] Each paragraph records its clause reference subset, denoted as , which is injected by the corresponding number of when filling the step two field.

[0189] S421, for each activated clause, determine whether it exists in any clause reference subset, if not, store the number of the clause in the un-referenced set .

[0190] The system compares according to the following logic:

[0191] For each activated clause (i.e. ), determine whether it appears in any , if there is no match, it is determined that the clause is not referenced. All un-referenced clause numbers are recorded by the system as the un-referenced set , while extracting the original activation semantic label of these clauses (i.e. the semantic dimension in which activates the clause), display its trigger reason.

[0192] For example, the Building Fire Protection Code Article 6.7 is not mentioned in the document, but activated by the "underground parking structure", the system suggests adding a reference expression of this article in the "fire protection design" module.

[0193] S43, calculate the completeness of the legal article reference:

[0194] ;

[0195] When the value is less than 0.9, it is considered that there is a defect in the article coverage.

[0196] The kth regulation in the global regulation article vector is activated, applicable to the current project, The ith paragraph template references the definition of the kth regulation article.

[0197] S44, get the answer offset set.

[0198] Further, S44 includes:

[0199] S440, respectively get the semantic consistency score of each answer from the answer set;

[0200] The system scores the consistency of each record in the answer set The variable is the semantic consistency score defined in step three, indicating the degree of semantic and structural consistency between the answer and its reference paragraph The system will reuse the score results in this step.

[0201] S441, store the index of the answer whose semantic consistency score is less than the set score threshold (e.g. 0.65) in the answer offset set.

[0202] If , the answer is marked as an offset answer.

[0203] To enhance the interpretability of the audit structure, the system extracts the semantic tag subset of the reference paragraph and its reference article set of the answer, and compares them with the terms and regulation expressions in the content answered, to ensure that the references in the answer are not misused or misquoted.

[0204] All the indexes make up the answer offset set , and output the specific offset reason, paragraph number and suggested manual review label of each record.

[0205] S45, obtain the term conflict set.

[0206] To prevent the existence of module term conflicts within the document, the system compares paragraphs within the same module (such as "structural design" and "equipment description").

[0207] A term set is extracted from each paragraph, and a fuzzy similarity based on edit distance and an Euclidean distance between semantic vectors are combined to form a term conflict score .

[0208] The calculation process of the term conflict score is as follows:

[0209] First, the technical term set of each paragraph is extracted, and the terms are unified to lowercase, removed stop words, and then divided by words; for any two paragraphs and under the same module, the term sets are matched with each other, and the edit distance of the matched term pairs is calculated one by one, and the string similarity based on edit distance is obtained by formula ; At the same time, the matching terms are input into the pre-trained word vector model to obtain semantic vectors, and the Euclidean distance is calculated and converted into semantic similarity ; The average value of the two similarities is obtained as the comprehensive similarity of the term pair, and finally the average of the comprehensive similarity of all term pairs in the paragraph is taken as , and the term conflict score is used to represent the degree of term conflict of the paragraph pair, and the value closer to 1 indicates that the conflict is more serious.

[0210] When the term conflict score between any paragraph pair in a module is greater than 0.5, it is considered that the term expression is inconsistent, constituting an expressive conflict.

[0211] The system records the paragraph number, term conflict group and suggested revision item to form the term conflict set .

[0212] S46, based on the unmatched set, the structural semantic coverage rate, the unquoted set, the legal clause reference integrity, the reply offset set and the term conflict set to generate an audit report:

[0213] After completing the structural semantic comparison, the legal clause matching review, the answer response consistency judgment and the term expression conflict analysis, the intermediate results generated in all stages are structured and integrated to generate a complete audit report of information consistency and compliance:

[0214] ;

[0215] represents the audit report.

[0216] Each sub-item is generated based on the upstream and downstream paths of the structured variable, the semantics, source and role of all variables are clearly marked in the steps, and there is no variable ambiguity or naming conflict.

[0217] As Figure 2 The application also provides an intelligent bidding agent full-process service management system, comprising an acquisition module, a first generation module, a second generation module and a third generation module.

[0218] The acquisition module is used for acquiring a structural semantic vector based on a building information model and acquiring a regulation clause adaptive vector based on the structural semantic vector;

[0219] The first generation module is used for generating a bidding document draft based on the structural semantic vector and the regulation clause adaptive vector;

[0220] The second generation module is used for generating a structured question-answering response set based on questions raised by a bidder and the bidding document draft;

[0221] The third generation module is used for generating an audit report based on the bidding document draft and the structured question-answering response set.

[0222] The above embodiments are only for illustrating the technical concept and characteristics of the application, the purpose is to enable those skilled in the art to understand the content of the application and implement it, and cannot limit the protection scope of the application, any equivalent changes or modifications made according to the spirit and essence of the application should be covered within the protection scope of the application.

Claims

1.A method for intelligent bidding agent full-process service management, comprising: S1, obtaining a structural semantic vector based on a building information model, and obtaining a regulation clause adaptation vector based on the structural semantic vector; S2, generating a draft of a bidding document based on the structural semantic vector and the regulation clause adaptation vector; S3, generating a structured question-answering response set based on questions raised by bidders and the draft of the bidding document; S4, generating an audit report based on the draft of the bidding document and the structured question-answering response set. 2.The intelligent bidding agent whole-process service management method according to claim 1, characterized in that, The structural semantic vector is obtained based on the building information model, comprising: S10, obtaining original fields contained in each component object in the building information model respectively; S11, converting the original fields into semantic tags based on a preset rule template; S12, encoding all the semantic tags into a structural semantic vector. 3.The intelligent bidding agent whole-process service management method according to claim 1, characterized in that, The regulation clause adaptation vector is obtained based on the structural semantic vector, comprising: obtaining a regulation clause trigger matrix; calculating the regulation clause adaptation vector based on the structural semantic vector and the regulation clause trigger matrix. 4.The intelligent bidding agent whole-process service management method according to claim 1, wherein S2 Comprising: S20, obtaining a paragraph template library; S21, calculating a paragraph activation vector of each paragraph template in the paragraph template library based on the structural semantic vector and the regulation clause adaptation vector respectively; S22, components greater than a set threshold in the paragraph activation vector, the paragraph template number corresponding to the component is stored in the set ; S23, screening the activated paragraph template from the paragraph template library , ; S24, performing field filling on field slots in the paragraph template; S25, splicing all the filled paragraphs in module order to generate the draft of the bidding document. 5.The intelligent bidding agent whole-process service management method according to claim 1, wherein S3 Comprising: S30, for each paragraph in the draft tender document constructing a semantic representation vector ; S31, asking questions by the tenderer Encoding as a vector ; S32, based on and calculating a match score between the passage and the question; S33, obtaining the paragraph corresponding to the highest matching score , extracting the paragraph associated structural semantic label and activated regulation clause number set ; S34, generating a reply using a preset structured template; S35, based on , the corresponding semantic encoding and paragraph semantic representation vector calculates semantic consistency score ; S36, recording each reply as a structured five-tuple, and storing all the structured five-tuples into the question-answering response set. 6.The intelligent bidding agent whole-process service management method according to claim 1, characterized in that, S4 Comprising: S40, obtaining a non-matching set based on the question-answering response set; S41, calculating a structural semantic coverage rate; S42, obtaining a non-reference set; S43, calculating a legal clause reference completeness rate; S44, obtaining a reply offset set; S45, obtaining a term conflict set; S46, generating an audit report based on the non-matching set, the structural semantic coverage rate, the non-reference set, the legal clause reference completeness rate, the reply offset set, and the term conflict set. 7.The intelligent bidding agent whole-process service management method according to claim 6, characterized in that, S40 comprises: S400, extracting a structural semantic tag set from all the paragraphs in the question-answering response set; S401, comparing the structural semantic tag set with the structural semantic vector to construct a structural semantic coverage matrix; S402, judging whether each dimension k satisfies a preset condition respectively; S403, taking a structural feature corresponding to a dimension that does not satisfy the preset condition as a structural semantic omission item, S404, form a non-matching set with the numbers of all structural semantic missing items . 8.The intelligent bidding agent whole-process service management method according to claim 6, characterized in that, S42 comprises: S420, obtaining a clause reference subset corresponding to each paragraph respectively; S421, judging whether each activated clause exists in any of the clause reference subsets, and if not, storing the number of the clause into the non-reference set. 9.The intelligent bidding agent whole-process service management method according to claim 6, characterized in that, S44 comprises: S440, obtaining a semantic consistency score of each reply from the question-answering response set respectively; S441, storing an index of a reply whose semantic consistency score is less than a set score threshold into the reply offset set. 10.A system for intelligent bidding agent full-process service management, comprising an obtaining module, a first generating module, a second generating module, and a third generating module; The acquisition module is configured to acquire a structure semantic vector based on the building information model, and acquire a regulation clause adaptation vector based on the structure semantic vector; The first generation module is configured to generate a draft of the bidding document based on the structure semantic vector and the regulation clause adaptation vector; The second generation module is configured to generate a set of structured response to questions based on questions raised by the tenderer and the draft of the bidding document; The third generation module is configured to generate an audit report based on the draft of the bidding document and the set of structured response to questions.

Images