A bill of quantities analysis method and system based on contract and valuation specifications

Abstract

The application discloses a kind of based on contract and valuation specification's engineering quantity list analysis method and system, method includes list data acquisition, data preliminary processing, structured extraction model construction, list classification model construction and engineering quantity list analysis.The original data is obtained by data acquisition;Adopt text standardization, structure division, data purification alignment, numerical extraction and data set division data preliminary processing method;Improved BERT model is used as structured extraction model, through multi-granularity semantic capture, short phrase combination features and long sentence context logic are considered, and progressive logic association and path closure verification are used, to ensure that the mapping path has bidirectional traceability and self-consistency;Improved encoder model is used as list classification model, numerical features and text representation are jointly considered, and progressive judgment process is explicitly simulated, so that the decision path of the model naturally has explainability, and an uncertainty mechanism is introduced to reduce the risk of misjudgment.

Description

Technical Field

[0001] This invention relates to the field of intelligent bill of quantities analysis technology, specifically to a bill of quantities analysis method and system based on contracts and pricing standards. Background Technology

[0002] Bill of Quantities analysis is a process that utilizes advanced artificial intelligence technology and big data processing methods to automatically extract, classify, and analyze various data points within an engineering project. By transforming information from engineering drawings and related documents into analyzable structured data, it facilitates subsequent project management, cost estimation, and decision support, improving the efficiency and accuracy of engineering projects while reducing human error.

[0003] However, traditional bill of quantities analysis methods suffer from several drawbacks. Firstly, their text processing capabilities remain at the surface level, failing to effectively address diverse expressions and long-distance semantic dependencies within the same professional context. Secondly, they lack cross-verification of the consistency of the bill of quantities, contract, and specification ternary closed loop, easily leading to logically broken mapping paths. Thirdly, traditional bill of quantities analysis methods typically separate numerical rule-based judgments from textual condition-based judgments into two independent workflows, severing the coupling and constraint relationship between the two types of information. Furthermore, they lack a measure of the reliability of their own judgment results, making it difficult to effectively filter out high-risk misjudgment items. Summary of the Invention

[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a method and system for analyzing bill of quantities based on contracts and pricing standards. Traditional bill of quantities analysis methods suffer from limitations: their text processing capabilities are limited to the surface level, failing to effectively handle diverse expressions and long-distance semantic dependencies within the same professional meaning; and they lack cross-verification of the consistency of the bill of quantities-contract-standard ternary loop, easily leading to logically broken mapping paths. This solution creatively employs an improved BERT model as the structured extraction model. Through multi-granular semantic capture, it simultaneously considers the combined features of professional phrases and the contextual logic of long sentences. Furthermore, by utilizing progressive logical association and path closure verification, it ensures the established... While the mapping paths are statistically similar, they also possess bidirectional traceability and self-consistency. Addressing the technical problems of traditional bill of quantities analysis methods, which typically separate numerical rule-based judgments and textual condition-based judgments into two independent work lines, severing the coupling and constraint relationship between the two types of information and lacking a measurement output of the reliability of their own judgment results, thus failing to effectively screen out high-risk misjudged items, this solution creatively adopts an improved encoder model as the bill of quantities classification model. It places numerical features and textual representations in the same representation space for joint consideration and explicitly simulates the progressive judgment process in cost auditing, making the model's decision path inherently interpretable. The introduction of an uncertainty estimation mechanism reduces the risk of item misjudgment.

[0005] The technical solution adopted by this invention is as follows: This invention provides a bill of quantities analysis method based on contract and pricing specifications, which includes the following steps:

[0006] Step S1: Obtain list data;

[0007] Step S2: Preliminary data processing;

[0008] Step S3: Construction of the structured extraction model;

[0009] Step S4: Building the list classification model;

[0010] Step S5: Bill of Quantities Analysis.

[0011] Further, in step S1, the acquisition of the bill of quantities data is used to provide raw data for the training and application of the subsequent model. Specifically, it involves obtaining the raw dataset for bill of quantities analysis through data collection. The raw dataset for bill of quantities analysis includes a historical raw dataset for bill of quantities analysis and a current raw dataset for bill of quantities analysis. Both the historical raw dataset for bill of quantities analysis and the current raw dataset for bill of quantities analysis contain text data of the bill of quantities, text data of contract terms, and text data of pricing specification clauses. The historical raw dataset for bill of quantities analysis also contains classification tags for the bill of quantities.

[0012] Further, in step S2, the preliminary data processing is used to perform preliminary processing on the collected raw data, transforming the raw data into a regular format, specifically including the following steps:

[0013] Step S21: Text standardization is used to eliminate inconsistencies in encoding, font, and unit of measurement among the three types of data. Specifically, it involves unifying UTF-8 encoding, normalizing full-width and half-width characters, converting between simplified and traditional Chinese characters, and mapping synonymous units of measurement to standard words to obtain standardized text.

[0014] Step S22: Structural segmentation, used to break down continuous long text into independent processable units according to natural semantic boundaries. Specifically, the bill of quantities is segmented into item-level records by line, contract clauses are segmented into clause blocks by clause number regularization, and pricing specification clauses are segmented into clause blocks by chapter number and clause number.

[0015] Step S23: Data purification and alignment, used to remove irrelevant and interfering information and ensure that the structure of each segmentation unit is complete and consistent. Specifically, headers, footers and table of contents indexes are removed by regular expression matching and position rules, non-detailed data and list entries with missing item feature descriptions are filtered out, and a unique primary key is assigned to each segmentation unit.

[0016] Step S24: Numerical extraction, used to extract quantitative parameters from text, specifically extracting numerical fields from the bill of quantities and extracting numerical fields from the contract text using regular expressions and named entity recognition to form quantitative parameters. The numerical fields include, but are not limited to, tender quantity, change quantity, bid unit price, change rate threshold, and amount threshold.

[0017] Step S25: Dataset partitioning, used to obtain training data and validation data, specifically, splitting the original dataset of the historical list analysis to obtain the extracted training set and the extracted test set;

[0018] The current list analysis raw dataset is pre-processed through text standardization, structural segmentation, data purification and alignment, and numerical extraction to obtain the dataset to be extracted. The historical list analysis raw dataset is pre-processed through text standardization, structural segmentation, data purification and alignment, numerical extraction, and dataset partitioning to obtain the extraction training set and extraction test set.

[0019] Further, in step S3, the structured extraction model construction is used to construct the model required to extract structured data and establish the semantic mapping relationship between the feature description of the bill of quantities, contract terms, and pricing specification clauses. Specifically, it is to construct an improved BERT model as the structured extraction model. The improved BERT model outputs structured bill of quantities data and candidate triples with confidence scores by performing dual-granularity encoding, hierarchical attention matching, and triangulation calibration on the three types of text.

[0020] The construction of the structured extraction model specifically includes the following steps:

[0021] Step S31: Dual-granularity encoding, used to transform the original bill of quantities feature descriptions, contract terms, and pricing specification texts into vector representations containing both phrase-level and sentence-level semantic information. The steps include:

[0022] Step S311: Context encoding, used to generate an initial vector sequence containing context information for each text. Specifically, the pre-processed bill of quantities segmentation unit, contract clause segmentation unit, and pricing specification clause segmentation unit are fed into the pre-trained BERT model for processing. The output vector of each word is taken to form a sequence matrix, and the vector at the [CLS] position is extracted as the global representation vector.

[0023] Step S312: Dual-granularity convolution extraction, used to capture phrase-level semantic patterns and sentence-level semantic patterns in the text respectively. Specifically, two one-dimensional convolution kernels with sizes of 3 and 7 are used to perform parallel convolution on the sequence matrix. After activation by the ReLU activation function, mean pooling is performed, and granular fusion features are obtained through gated weighted fusion.

[0024] Step S313: Feature concatenation, which combines global sentence meaning and local multi-granular semantics into a more complete text representation. Specifically, it concatenates the global representation vector with the granular fusion features along the feature dimension to form the final text representation vector.

[0025] Step S32: Hierarchical attention mapping, used to construct mappings between three types of text through a three-level progressive attention mechanism. The steps include:

[0026] Step S321: The first coarse matching layer is used to establish a preliminary semantic relationship between the feature description of the bill of quantities and the contract terms, and to screen out the possible applicable contract terms. Specifically, it calculates the matching score between the final text representation vector of a bill of quantities segmentation unit and the final text representation vector of each contract term segmentation unit through bilinear attention. After normalization by the softmax function, the soft alignment weight is obtained. The soft alignment weight is then used to perform a weighted summation of the final text representation vectors of all contract term segmentation units to obtain the bill of quantities-aware contract aggregation features.

[0027] Step S322: The second layer of fine association is used to further associate the contract terms selected in the first layer with specific pricing normative clauses. Specifically, the final text representation vector of each contract term segmentation unit is repeatedly cross-attention calculated with the final text representation vectors of all pricing normative clause segmentation units to obtain the association weight of the contract terms with the pricing normative clauses. The association weight is then used to perform a weighted summation of the final text representation vectors of all pricing normative clause segmentation units to obtain the normative aggregation features of the terms.

[0028] Step S323: Triangulation verification calibration, used to check whether the mapping path formed by the two layers of transmission satisfies logical consistency. Specifically, it calculates the cosine similarity between the final text representation vector of the bill of quantities segmentation unit and the final text representation vector of the corresponding pricing specification segmentation unit as the direct semantic association value, and multiplies it with the weights of the first two layers as a adjustment factor to obtain the mapping path score.

[0029] Step S33: Mapping confidence evaluation, used to evaluate the quality of candidate mappings generated by the model, includes the following steps:

[0030] Step S331: Candidate extraction, used to filter out the top K mapping paths from the candidate space. Specifically, for the final text representation vector of each bill of quantities segmentation unit, take the final text representation vectors of the top K contract clause segmentation units with the highest soft alignment weight, and then take the final text representation vectors of the top K pricing specification clause segmentation units with the highest association weight for the final text representation vector of each contract clause segmentation unit, forming a preliminary candidate triplet set.

[0031] Step S332: Confidence scoring, which is used to assign a confidence index in the interval [0,1] to each candidate mapping path. Specifically, the mapping path score represented by each preliminary candidate triplet is mapped to a confidence score through a sigmoid function with learnable parameters. Only triplets in the preliminary candidate triplet set with confidence scores greater than a preset threshold are retained to obtain the candidate triplet set.

[0032] Step S34: Model construction and training, specifically, by integrating the dual-granularity encoding, the hierarchical attention mapping and the mapping confidence evaluation, an improved BERT model is constructed, and the model is trained and its performance is verified based on the extracted training set and extracted test set to obtain the improved BERT model as the structured extraction model;

[0033] The extracted training set and extracted test set are used as inputs to the structured extraction model to obtain the classification training set and classification test set.

[0034] Further, in step S4, the bill of quantities classification model construction is used to construct the model required for intelligent classification of the bill of quantities. Specifically, it is to construct an improved encoder model as the bill of quantities classification model. The improved encoder model outputs the bill of quantities classification label and uncertainty value by fusing numerical and semantic features and through three-stage causal reasoning.

[0035] The construction of the list classification model specifically includes the following steps:

[0036] Step S41: Conditional Aware Coding, used to fuse the numerical and textual features of the bill of quantities into a unified conditional representation vector. The steps include:

[0037] Step S411: Numerical feature construction, which is used to transform the original numerical data into a dimensionless ratio that reflects the degree of change. Specifically, it is to calculate the rate of change of the changed quantity of the project relative to the tendered quantity of the project and the ratio of the changed unit price to the tender unit price based on quantitative parameters, so as to obtain the numerical features.

[0038] Step S412: Text semantic representation, which integrates the feature description of the list itself and the contract terms and pricing specifications found through the mapping relationship into a semantically rich context vector. Specifically, it concatenates the three elements of the candidate triplet, feeds them into the pre-trained BERT model for processing, and extracts the vector at the [CLS] position as the text feature.

[0039] Step S413: Conditional vector concatenation, which combines numerical and semantic information into a complete conditional representation. Specifically, it concatenates numerical features and text features along the feature dimension to obtain a conditional representation vector.

[0040] Step S42: Design of the causal reasoning chain, used to design the syllogistic reasoning process, transforming the logical chain into a computable neural network structure. The steps include:

[0041] Step S421: Phase 1 Prerequisite Activation, used to determine whether there are applicable contract terms for the current list item. Specifically, the condition representation vector is processed through an independent gated loop unit, then activated by a linear transformation and a sigmoid activation function, outputting a scalar value to obtain the activation strength of the contract terms.

[0042] Step S422: Phase 2 minor premise check, used to seek supplementary basis from the pricing specification clause when the contract terms do not explicitly cover it. Specifically, the activation strength of the contract terms and the condition representation vector are concatenated and sent to another independent gated loop unit. After linear transformation and activation by the sigmoid activation function, a scalar value is output to obtain the pricing specification verification score.

[0043] Step S423: Stage 3 soft logic gating decision, used to comprehensively consider the activation strength of contract terms and the verification score of pricing specifications. Specifically, it simulates preset rules through a set of differentiable soft logic expressions. If the contract activation is strong and the change rate of the quantity of work does not exceed the preset contract threshold, it tends to settle according to the original contract price. If the contract activation is weak or the change rate of the quantity of work exceeds the preset contract threshold, and the verification score of the pricing specifications does not exceed the pre-designed price specification verification threshold, it tends to need to re-price. Otherwise, it tends to be on the abnormal list, resulting in a three-class probability vector.

[0044] Step S43: Uncertainty quantification, used to quantify the model's confidence in the current judgment result, specifically by calculating the negative log-likelihood expectation of the three label probabilities based on the three-class probability vector to obtain the uncertainty value;

[0045] Step S44: Construct and train the model. Specifically, by integrating the conditional-aware encoding, the causal inference chain design, and the uncertainty quantization, an improved encoder model is constructed. The model is trained and its performance is verified based on the classification training set and the classification test set to obtain the improved encoder model, which serves as the list classification model.

[0046] Further, in step S5, the bill of quantities analysis specifically involves processing the dataset to be extracted as input to a structured extraction model, outputting structured bill of quantities data and candidate triples with confidence scores, then processing the output of the structured extraction model as input to a bill of quantities classification model, outputting bill of quantities classification labels and uncertainty values, and finally using the structured bill of quantities data, candidate triples with confidence scores, bill of quantities classification labels, and uncertainty values ​​as the final bill of quantities analysis result.

[0047] The present invention provides a bill of quantities analysis system based on contract and pricing specifications, including a bill of quantities data acquisition module, a preliminary data processing module, a structured extraction model construction module, a bill of quantities classification model construction module, and a bill of quantities analysis module;

[0048] The list data acquisition module is used to collect raw data, obtain the list analysis raw dataset by collecting raw data, and send the list analysis raw dataset to the data preliminary processing module.

[0049] The data preliminary processing module is used for preliminary data processing. Through preliminary data processing, the dataset to be extracted, the training set to be extracted, and the test set to be extracted are obtained. The dataset to be extracted is sent to the bill of quantities analysis module, and the training set and the test set to be extracted are sent to the structured extraction model construction module.

[0050] The structured extraction model construction module is used to construct a structured extraction model. By constructing an improved BERT model, a structured extraction model is obtained. The extraction training set and extraction test set are used as inputs to the structured extraction model for processing to obtain a classification training set and a classification test set. The structured extraction model is then sent to the bill of quantities analysis module, and the classification training set and the classification test set are sent to the bill of quantities classification model construction module.

[0051] The bill of quantities classification model construction module is used to construct the bill of quantities classification model. By constructing an improved encoder model, the bill of quantities classification model is obtained, and the bill of quantities classification model is sent to the bill of quantities analysis module.

[0052] The bill of quantities analysis module is used for bill of quantities analysis. By combining the structured extraction model and the bill of quantities classification model, it performs real-time bill of quantities analysis to obtain the final bill of quantities analysis results.

[0053] The beneficial effects achieved by adopting the above solution are as follows:

[0054] (1) In view of the technical problems of traditional bill of quantities analysis methods, which have the ability to process text only at the surface level of words, cannot effectively deal with the diverse expressions and long-distance semantic dependencies under the same professional meaning, and lack cross-verification of the consistency of the three-element closed loop of bill of quantities-contract-standard, which easily leads to logically broken mapping paths, this solution creatively adopts an improved BERT model as a structured extraction model. Through multi-granular semantic capture, it can simultaneously take into account the combination features of professional phrases and the contextual logic of long sentences. At the same time, by using progressive logical association and path closure verification, it ensures that the established mapping path is statistically similar and has bidirectional traceability and self-consistency.

[0055] (2) In view of the technical problems of traditional bill of quantities analysis methods, which usually separate numerical rule judgment and text condition judgment into two independent work lines, severing the coupling and constraint relationship between the two types of information, and lacking the measurement output of the reliability of its own judgment results, making it impossible to effectively screen out high-risk misjudged items, this solution creatively adopts an improved encoder model as the bill of quantities classification model, puts numerical features and text representations in the same representation space for joint consideration, and explicitly simulates the step-by-step judgment process in cost audit, so that the decision path of the model is naturally interpretable, introduces an uncertainty estimation mechanism, and reduces the risk of item misjudgment. Attached Figure Description

[0056] Figure 1 A flowchart illustrating a bill of quantities analysis method based on contract and pricing standards provided by this invention;

[0057] Figure 2 A schematic diagram of a bill of quantities analysis system based on contract and pricing standards provided by the present invention;

[0058] Figure 3 This is a flowchart illustrating the preliminary data processing in step S2.

[0059] Figure 4 A flowchart illustrating the process of constructing the structured extraction model in step S3;

[0060] Figure 5 A flowchart illustrating the process of constructing the list classification model for step S4.

[0061] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0062] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0063] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0064] Example 1, see Figure 1 This invention provides a bill of quantities analysis method based on contract and pricing standards, which includes the following steps:

[0065] Step S1: Obtain list data;

[0066] Step S2: Preliminary data processing;

[0067] Step S3: Construction of the structured extraction model;

[0068] Step S4: Building the list classification model;

[0069] Step S5: Bill of Quantities Analysis.

[0070] Example 2, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S1, the acquisition of the bill of quantities data is used to provide raw data for the training and application of the subsequent model. Specifically, it is to obtain the original dataset for bill of quantities analysis through data collection. The original dataset for bill of quantities analysis specifically includes the historical original dataset for bill of quantities analysis and the current original dataset for bill of quantities analysis. Both the historical original dataset for bill of quantities analysis and the current original dataset for bill of quantities analysis contain bill of quantities text data, contract terms text data, and pricing specification clause text data. The historical original dataset for bill of quantities analysis also contains bill of quantities classification tags.

[0071] The bill of quantities text data is collected from the bidding documents and includes, but is not limited to, the bill of quantities code, project name, project characteristic description text, unit of measurement, tendered quantity, bid unit price, and revised quantity. The contract terms text data is collected from the construction contract documents and includes, but is not limited to, contract price method, risk range clause, change rate threshold, and amount threshold. The pricing specification text data is collected from the construction engineering bill of quantities pricing standard and local pricing specification clauses and includes, but is not limited to, the clauses of the sub-items of the project, mandatory clauses, pricing rules, and quota items. The bill of quantities classification labels specifically include the bill of quantities settled according to the original contract price, the bill of quantities requiring re-pricing, and the abnormal bill of quantities.

[0072] Example 3, see Figure 1 , Figure 2 and Figure 3 This embodiment is based on the above embodiment. In step S2, the preliminary data processing is used to perform preliminary processing on the collected raw data, transforming the raw data into a regular format. Specifically, it includes the following steps:

[0073] Step S21: Text standardization is used to eliminate inconsistencies in encoding, font, and unit of measurement among the three types of data. Specifically, it involves unifying UTF-8 encoding, normalizing full-width and half-width characters, converting between simplified and traditional Chinese characters, and mapping synonymous units of measurement to standard words to obtain standardized text.

[0074] Step S22: Structural segmentation, used to break down continuous long text into independent processable units according to natural semantic boundaries. Specifically, the bill of quantities is segmented into item-level records by line, contract clauses are segmented into clause blocks by clause number regularization, and pricing specification clauses are segmented into clause blocks by chapter number and clause number.

[0075] Step S23: Data purification and alignment, used to remove irrelevant and interfering information and ensure that the structure of each segmentation unit is complete and consistent. Specifically, headers, footers and table of contents indexes are removed by regular expression matching and position rules, non-detailed data and list entries with missing item feature descriptions are filtered out, and a unique primary key is assigned to each segmentation unit.

[0076] Step S24: Numerical extraction, used to extract quantitative parameters from text, specifically extracting numerical fields from the bill of quantities and extracting numerical fields from the contract text using regular expressions and named entity recognition to form quantitative parameters. The numerical fields include, but are not limited to, tender quantity, change quantity, bid unit price, change rate threshold, and amount threshold.

[0077] Step S25: Dataset partitioning, used to obtain training data and validation data, specifically, the original dataset of the historical list analysis is partitioned to obtain the extracted training set and the extracted test set;

[0078] The current list analysis raw dataset is pre-processed through text standardization, structural segmentation, data purification and alignment, and numerical extraction to obtain the dataset to be extracted. The historical list analysis raw dataset is pre-processed through text standardization, structural segmentation, data purification and alignment, numerical extraction, and dataset partitioning to obtain the extraction training set and extraction test set.

[0079] Example 4, see Figure 1 , Figure 2 and Figure 4 This embodiment is based on the above embodiment. In step S3, the structured extraction model is constructed to build the model required to extract structured data and establish the semantic mapping relationship between the feature description of the bill of quantities, contract terms and pricing specification clauses. Specifically, an improved BERT model is constructed as the structured extraction model. The improved BERT model outputs structured bill of quantities data and candidate triples with confidence scores by performing dual-granularity encoding, hierarchical attention matching and triangulation calibration on the three types of text.

[0080] The construction of the structured extraction model specifically includes the following steps:

[0081] Step S31: Dual-granularity encoding, used to transform the original bill of quantities feature descriptions, contract terms, and pricing specification texts into vector representations containing both phrase-level and sentence-level semantic information. The steps include:

[0082] Step S311: Context encoding, used to generate an initial vector sequence containing context information for each text. Specifically, the pre-processed bill of quantities segmentation unit, contract clause segmentation unit, and pricing specification clause segmentation unit are fed into the pre-trained BERT model for processing. The output vector of each word is taken to form a sequence matrix, and the vector at the [CLS] position is extracted as the global representation vector.

[0083] Step S312: Dual-granularity convolution extraction, used to capture phrase-level semantic patterns and sentence-level semantic patterns in the text respectively. Specifically, two one-dimensional convolution kernels of sizes 3 and 7 are used to perform parallel convolution on the sequence matrix. After activation by the ReLU activation function, mean pooling is performed, and then gated weighted fusion is used to obtain granular fusion features. The formula used is as follows:

[0084] ;

[0085] In the formula, 'a' represents the text category index, BOQ represents the bill of quantities, CTR represents the contract terms, and COV represents the pricing specification clauses. The first text category a represents the text category a. Phrasal-level features of each segmentation unit The first text category a represents the text category a. Sentence-level features of each segmentation unit, The first text category a represents the text category a. The sequence matrix of each segmentation unit, This represents the mean pooling function. This represents a one-dimensional convolution function of size 3. This represents a one-dimensional convolution function with size 7, and gw represents the granular fusion gate weights. Represents the gated mapping weight matrix. Indicates the gating mapping bias term. This indicates a feature concatenation operation. The first text category a represents the text category a. Granularity fusion features of each segmentation unit This represents element-wise multiplication.

[0086] Step S313: Feature concatenation, which combines global sentence meaning and local multi-granular semantics into a more complete text representation. Specifically, it concatenates the global representation vector with the granular fusion features along the feature dimension to form the final text representation vector.

[0087] Step S32: Hierarchical attention mapping, used to construct mappings between three types of text through a three-level progressive attention mechanism. The steps include:

[0088] Step S321: The first coarse matching layer is used to establish a preliminary semantic relationship between the feature description of the bill of quantities and the contract terms, and to screen out the possible applicable contract terms. Specifically, it calculates the matching score between the final text representation vector of a bill of quantities segmentation unit and the final text representation vector of each contract term segmentation unit through bilinear attention. After normalization by the softmax function, the soft alignment weight is obtained. The soft alignment weight is then used to perform a weighted summation of the final text representation vectors of all contract term segmentation units to obtain the bill of quantities-aware contract aggregation features.

[0089] Step S322: The second layer of fine association is used to further associate the contract terms selected in the first layer with specific pricing normative clauses. Specifically, the final text representation vector of each contract term segmentation unit is repeatedly cross-attention calculated with the final text representation vectors of all pricing normative clause segmentation units to obtain the association weight of the contract terms with the pricing normative clauses. The association weight is then used to perform a weighted summation of the final text representation vectors of all pricing normative clause segmentation units to obtain the normative aggregation features of the terms.

[0090] Step S323: Triangulation verification calibration, used to verify whether the mapping path formed through two layers satisfies logical consistency. Specifically, it calculates the cosine similarity between the final text representation vector of the bill of quantities segmentation unit and the final text representation vector of the corresponding pricing specification segmentation unit as the direct semantic association value, and multiplies it with the weights of the first two layers as a adjustment factor to obtain the mapping path score. The formula used is as follows:

[0091] ;

[0092] In the formula, Indicates the first The final text representation vector of each bill of quantities segmentation unit and the first The direct semantic association value of the final text representation vector of each pricing specification clause segmentation unit. This represents the function for calculating cosine similarity. Indicates the first The final text representation vector of each unit segmented from the bill of quantities. Indicates the first The final text representation vector of each pricing specification clause segmentation unit. Indicates the first The final text representation vector of each bill of quantities segment unit is then used to represent the first... The final text representation vector of the contract clause segmentation unit to the first... The mapping path score of the final text representation vector of each pricing specification clause segmentation unit. Indicates the first The final text representation vector of each bill of quantities segmentation unit and the first The soft alignment weights of the final text representation vectors of each contract clause segmentation unit. Indicates the first The soft alignment weight of the final text representation vector of the contract clause segmentation unit is related to the first... The association weights of the final text representation vectors of each pricing specification clause segmentation unit Indicates the adjustment of hyperparameters;

[0093] Step S33: Mapping confidence evaluation, used to evaluate the quality of candidate mappings generated by the model, includes the following steps:

[0094] Step S331: Candidate extraction, used to filter out the top K mapping paths from the candidate space. Specifically, for the final text representation vector of each bill of quantities segmentation unit, take the final text representation vectors of the top K contract clause segmentation units with the highest soft alignment weight, and then take the final text representation vectors of the top K pricing specification clause segmentation units with the highest association weight for the final text representation vector of each contract clause segmentation unit, forming a preliminary candidate triplet set.

[0095] Step S332: Confidence scoring, which is used to assign a confidence index in the interval [0,1] to each candidate mapping path. Specifically, the mapping path score represented by each preliminary candidate triplet is mapped to a confidence score through a sigmoid function with learnable parameters. Only triplets in the preliminary candidate triplet set with confidence scores greater than a preset threshold are retained to obtain the candidate triplet set.

[0096] Step S34: Model construction and training, specifically, by integrating the dual-granularity encoding, the hierarchical attention mapping and the mapping confidence evaluation, an improved BERT model is constructed, and the model is trained and its performance is verified based on the extracted training set and extracted test set to obtain the improved BERT model as the structured extraction model;

[0097] The extracted training set and extracted test set are used as inputs to the structured extraction model to obtain the classification training set and classification test set.

[0098] By performing the above operations, this solution addresses the technical problems of traditional bill of quantities analysis methods, which have text processing capabilities limited to the lexical surface level, are unable to effectively handle diverse expressions and long-distance semantic dependencies under the same professional meaning, and lack cross-verification of the consistency of the bill of quantities-contract-specification ternary closed loop, easily leading to logically broken mapping paths. This solution creatively adopts an improved BERT model as a structured extraction model. Through multi-granular semantic capture, it can simultaneously take into account the combination features of professional phrases and the contextual logic of long sentences. At the same time, by using progressive logical association and path closure verification, it ensures that the established mapping path is statistically similar and has bidirectional traceability and self-consistency.

[0099] Example 5, see Figure 1 , Figure 2 and Figure 5 This embodiment is based on the above embodiment. In step S4, the bill of quantities classification model is constructed to build the model required for intelligent classification of the bill of quantities. Specifically, an improved encoder model is constructed as the bill of quantities classification model. The improved encoder model outputs the bill of quantities classification label and uncertainty value by fusing numerical and semantic features and through three-stage causal reasoning.

[0100] The construction of the list classification model specifically includes the following steps:

[0101] Step S41: Conditional Aware Coding, used to fuse the numerical and textual features of the bill of quantities into a unified conditional representation vector. The steps include:

[0102] Step S411: Numerical feature construction, which is used to transform the original numerical data into a dimensionless ratio that reflects the degree of change. Specifically, it is to calculate the rate of change of the changed quantity of the project relative to the tendered quantity of the project and the ratio of the changed unit price to the tender unit price based on quantitative parameters, so as to obtain the numerical features.

[0103] Step S412: Text semantic representation, which integrates the feature description of the list itself and the contract terms and pricing specifications found through the mapping relationship into a semantically rich context vector. Specifically, it concatenates the three elements of the candidate triplet, feeds them into the pre-trained BERT model for processing, and extracts the vector at the [CLS] position as the text feature.

[0104] Step S413: Conditional vector concatenation, which combines numerical and semantic information into a complete conditional representation. Specifically, it concatenates numerical features and text features along the feature dimension to obtain a conditional representation vector.

[0105] Step S42: Design of the causal reasoning chain, used to design the syllogistic reasoning process, transforming the logical chain into a computable neural network structure. The steps include:

[0106] Step S421: Phase 1 Prerequisite Activation, used to determine whether there are applicable contract terms for the current list item. Specifically, the condition representation vector is processed through an independent gated loop unit, then activated by a linear transformation and a sigmoid activation function, outputting a scalar value to obtain the activation strength of the contract terms.

[0107] Step S422: Phase 2 minor premise check, used to seek supplementary basis from the pricing specification clause when the contract terms do not explicitly cover it. Specifically, the activation strength of the contract terms and the condition representation vector are concatenated and sent to another independent gated loop unit. After linear transformation and activation by the sigmoid activation function, a scalar value is output to obtain the pricing specification verification score.

[0108] Step S423: Stage Three Soft Logic Gating Decision, used to comprehensively consider the activation strength of contract terms and the verification score of pricing specifications. Specifically, it simulates preset rules through a set of differentiable soft logic expressions. If the contract activation is strong and the change rate of the work quantity does not exceed the preset contract threshold, it tends to settle according to the original contract price. If the contract activation is weak or the change rate of the work quantity exceeds the preset contract threshold, and the verification score of the pricing specifications does not exceed the pre-designed price specification verification threshold, it tends to require repricing. Otherwise, it tends to be on the exception list, resulting in a three-class probability vector. The formula used is as follows:

[0109] ;

[0110] In the formula, Indicates the first The original score for each unit of the bill of quantities is settled based on the original contract price. Indicates the first The original scores for each unit of the bill of quantities need to be repriced. Indicates the first The original score of the abnormal list for each bill of quantities segmentation unit. Indicates the first The activation strength of contract terms for each unit of the bill of quantities. Indicates the first Verification score of pricing specifications for each unit of the bill of quantities. Indicates the first Change rate of quantities for each unit of the bill of quantities This indicates a preset contract threshold. Indicates an indicator function, Indicates the first The three-class probability vector of each unit segmented by the bill of quantities;

[0111] Step S43: Uncertainty quantification, used to quantify the model's confidence in the current judgment result, specifically by calculating the negative log-likelihood expectation of the three label probabilities based on the three-class probability vector to obtain the uncertainty value;

[0112] Step S44: Construct and train the model. Specifically, by integrating the conditional-aware encoding, the causal inference chain design, and the uncertainty quantization, an improved encoder model is constructed. The model is trained and its performance is verified based on the classification training set and the classification test set to obtain the improved encoder model, which serves as the list classification model.

[0113] By performing the above operations, this solution addresses the technical problems of traditional bill of quantities analysis methods, which typically separate numerical rule judgment and textual condition judgment into two independent work lines, severing the coupling and constraint relationship between the two types of information, and lacking a measurement output of the reliability of its own judgment results, making it impossible to effectively screen out high-risk misjudged items. This solution creatively adopts an improved encoder model as the bill of quantities classification model, placing numerical features and textual representations in the same representation space for joint consideration, and explicitly simulating the progressive judgment process in cost auditing, making the model's decision path naturally interpretable, introducing an uncertainty estimation mechanism, and reducing the risk of item misjudgment.

[0114] Example 6, see Figure 1 and Figure 2This embodiment is based on the above embodiment. Specifically, the bill of quantities analysis involves processing the dataset to be extracted as input to a structured extraction model to output structured bill of quantities data and candidate triples with confidence scores. Then, the output of the structured extraction model is processed as input to a bill of quantities classification model to output bill of quantities classification labels and uncertainty values. Finally, the structured bill of quantities data, candidate triples with confidence scores, bill of quantities classification labels, and uncertainty values ​​are used as the final bill of quantities analysis results.

[0115] Example 7, see Figure 1 and Figure 2 Based on the above embodiments, this embodiment provides a bill of quantities analysis system based on contract and pricing specifications, including a bill of quantities data acquisition module, a preliminary data processing module, a structured extraction model construction module, a bill of quantities classification model construction module, and a bill of quantities analysis module.

[0116] The list data acquisition module is used to collect raw data, obtain the list analysis raw dataset by collecting raw data, and send the list analysis raw dataset to the data preliminary processing module.

[0117] The data preliminary processing module is used for preliminary data processing. Through preliminary data processing, the dataset to be extracted, the training set to be extracted, and the test set to be extracted are obtained. The dataset to be extracted is sent to the bill of quantities analysis module, and the training set and the test set to be extracted are sent to the structured extraction model construction module.

[0118] The structured extraction model construction module is used to construct a structured extraction model. By constructing an improved BERT model, a structured extraction model is obtained. The extraction training set and extraction test set are used as inputs to the structured extraction model for processing to obtain a classification training set and a classification test set. The structured extraction model is then sent to the bill of quantities analysis module, and the classification training set and the classification test set are sent to the bill of quantities classification model construction module.

[0119] The bill of quantities classification model construction module is used to construct the bill of quantities classification model. By constructing an improved encoder model, the bill of quantities classification model is obtained, and the bill of quantities classification model is sent to the bill of quantities analysis module.

[0120] The bill of quantities analysis module is used for bill of quantities analysis. By combining the structured extraction model and the bill of quantities classification model, it performs real-time bill of quantities analysis to obtain the final bill of quantities analysis results.

[0121] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0122] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0123] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for analyzing bill of quantities based on contracts and pricing standards, characterized in that: The method includes the following steps: S1: Inventory data acquisition. By collecting data, the raw dataset for inventory analysis is obtained. Specifically, the raw dataset for inventory analysis includes historical raw datasets for inventory analysis and the current raw dataset for inventory analysis. S2: Preliminary data processing: The collected raw data is preliminarily processed to obtain the dataset to be extracted, the training set to be extracted, and the test set to be extracted; S3: Structured extraction model construction, used to build the model required to extract structured data and establish the semantic mapping relationship between the feature description of the bill of quantities, contract terms and pricing specification clauses. Specifically, it is to build an improved BERT model as the structured extraction model. The improved BERT model outputs structured bill of quantities data and candidate triples with confidence scores by performing dual-granularity encoding, hierarchical attention matching and triangulation calibration on the three types of text. After the extraction training set and the extraction test set are processed by the structured extraction model, a classification training set and a classification test set are obtained. S4: Bill of Quantities Classification Model Construction, used to construct the model required for intelligent classification of the bill of quantities. Specifically, it involves constructing an improved encoder model as the bill of quantities classification model. The improved encoder model, through the fusion of numerical and semantic features and three-stage causal reasoning, outputs the bill of quantities classification label and uncertainty value. S5: Bill of Quantities Analysis. Specifically, the dataset to be extracted is processed as input to the structured extraction model, outputting structured bill of quantities data and candidate triples with confidence scores. Then, the output of the structured extraction model is processed as input to the bill of quantities classification model, outputting the bill of quantities classification label and uncertainty value. The structured bill of quantities data, candidate triples with confidence scores, bill of quantities classification label, and uncertainty value are used as the final bill of quantities analysis results.

2. The method for analyzing bill of quantities based on contract and pricing standards according to claim 1, characterized in that: The construction of the structured extraction model specifically includes the following steps: Step S31: Dual-granularity encoding, used to transform the original bill of quantities feature descriptions, contract terms, and pricing specification texts into vector representations containing both phrase-level and sentence-level semantic information. The steps include: Step S311: Context encoding, used to generate an initial vector sequence containing context information for each text. Specifically, the pre-processed bill of quantities segmentation unit, contract clause segmentation unit, and pricing specification clause segmentation unit are fed into the pre-trained BERT model for processing. The output vector of each word is taken to form a sequence matrix, and the vector at the [CLS] position is extracted as the global representation vector. Step S312: Dual-granularity convolution extraction, used to capture phrase-level semantic patterns and sentence-level semantic patterns in the text respectively. Specifically, two one-dimensional convolution kernels with sizes of 3 and 7 are used to perform parallel convolution on the sequence matrix. After activation by the ReLU activation function, mean pooling is performed, and granular fusion features are obtained through gated weighted fusion. Step S313: Feature concatenation, which combines global sentence meaning and local multi-granular semantics into a more complete text representation. Specifically, it concatenates the global representation vector with the granular fusion features along the feature dimension to form the final text representation vector. Step S32: Hierarchical attention mapping, used to construct mappings between three types of text through a three-level progressive attention mechanism. The steps include: Step S321: The first coarse matching layer is used to establish a preliminary semantic relationship between the feature description of the bill of quantities and the contract terms, and to screen out the possible applicable contract terms. Specifically, it calculates the matching score between the final text representation vector of a bill of quantities segmentation unit and the final text representation vector of each contract term segmentation unit through bilinear attention. After normalization by the softmax function, the soft alignment weight is obtained. The soft alignment weight is then used to perform a weighted summation of the final text representation vectors of all contract term segmentation units to obtain the bill of quantities-aware contract aggregation features. Step S322: The second layer of fine association is used to further associate the contract terms selected in the first layer with specific pricing normative clauses. Specifically, the final text representation vector of each contract term segmentation unit is repeatedly cross-attention calculated with the final text representation vectors of all pricing normative clause segmentation units to obtain the association weight of the contract terms with the pricing normative clauses. The association weight is then used to perform a weighted summation of the final text representation vectors of all pricing normative clause segmentation units to obtain the normative aggregation features of the terms. Step S323: Triangulation verification calibration, used to check whether the mapping path formed by the two layers of transmission satisfies logical consistency. Specifically, it calculates the cosine similarity between the final text representation vector of the bill of quantities segmentation unit and the final text representation vector of the corresponding pricing specification segmentation unit as the direct semantic association value, and multiplies it with the weights of the first two layers as a adjustment factor to obtain the mapping path score. Step S33: Mapping confidence evaluation, used to evaluate the quality of candidate mappings generated by the model, includes the following steps: Step S331: Candidate extraction, used to filter out the top K mapping paths from the candidate space. Specifically, for the final text representation vector of each bill of quantities segmentation unit, take the final text representation vectors of the top K contract clause segmentation units with the highest soft alignment weight, and then take the final text representation vectors of the top K pricing specification clause segmentation units with the highest association weight for the final text representation vector of each contract clause segmentation unit, forming a preliminary candidate triplet set. Step S332: Confidence scoring, which is used to assign a confidence index in the interval [0,1] to each candidate mapping path. Specifically, the mapping path score represented by each preliminary candidate triplet is mapped to a confidence score through a sigmoid function with learnable parameters. Only triplets in the preliminary candidate triplet set with confidence scores greater than a preset threshold are retained to obtain the candidate triplet set. Step S34: Model construction and training, specifically, by integrating the dual-granularity encoding, the hierarchical attention mapping and the mapping confidence evaluation, an improved BERT model is constructed, and the model is trained and its performance is verified based on the extracted training set and extracted test set to obtain the improved BERT model as the structured extraction model; The extracted training set and extracted test set are used as inputs to the structured extraction model to obtain the classification training set and classification test set.

3. The method for analyzing bill of quantities based on contract and pricing standards according to claim 1, characterized in that: The construction of the list classification model specifically includes the following steps: Step S41: Conditional Aware Coding, used to fuse the numerical and textual features of the bill of quantities into a unified conditional representation vector. The steps include: Step S411: Numerical feature construction, which is used to transform the original numerical data into a dimensionless ratio that reflects the degree of change. Specifically, it is to calculate the rate of change of the changed quantity of the project relative to the tendered quantity of the project and the ratio of the changed unit price to the tender unit price based on quantitative parameters, so as to obtain the numerical features. Step S412: Text semantic representation, which integrates the feature description of the list itself and the contract terms and pricing specifications found through the mapping relationship into a semantically rich context vector. Specifically, it concatenates the three elements of the candidate triplet, feeds them into the pre-trained BERT model for processing, and extracts the vector at the [CLS] position as the text feature. Step S413: Conditional vector concatenation, which combines numerical and semantic information into a complete conditional representation. Specifically, it concatenates numerical features and text features along the feature dimension to obtain a conditional representation vector. Step S42: Design of the causal reasoning chain, used to design the syllogistic reasoning process, transforming the logical chain into a computable neural network structure. The steps include: Step S421: Phase 1 Prerequisite Activation, used to determine whether there are applicable contract terms for the current list item. Specifically, the condition representation vector is processed through an independent gated loop unit, then activated by a linear transformation and a sigmoid activation function, outputting a scalar value to obtain the activation strength of the contract terms. Step S422: Phase 2 minor premise check, used to seek supplementary basis from the pricing specification clause when the contract terms do not explicitly cover it. Specifically, the activation strength of the contract terms and the condition representation vector are concatenated and sent to another independent gated loop unit. After linear transformation and activation by the sigmoid activation function, a scalar value is output to obtain the pricing specification verification score. Step S423: Stage 3 soft logic gating decision, used to comprehensively consider the activation strength of contract terms and the verification score of pricing specifications. Specifically, it simulates preset rules through a set of differentiable soft logic expressions. If the contract activation is strong and the change rate of the quantity of work does not exceed the preset contract threshold, it tends to settle according to the original contract price. If the contract activation is weak or the change rate of the quantity of work exceeds the preset contract threshold, and the verification score of the pricing specifications does not exceed the pre-designed price specification verification threshold, it tends to need to re-price. Otherwise, it tends to be on the abnormal list, resulting in a three-class probability vector. Step S43: Uncertainty quantification, used to quantify the model's confidence in the current judgment result, specifically by calculating the negative log-likelihood expectation of the three label probabilities based on the three-class probability vector to obtain the uncertainty value; Step S44: Construct and train the model. Specifically, by integrating the conditional-aware encoding, the causal inference chain design, and the uncertainty quantization, an improved encoder model is constructed. The model is trained and its performance is verified based on the classification training set and the classification test set to obtain the improved encoder model, which serves as the list classification model.

4. The method for analyzing bill of quantities based on contract and pricing standards according to claim 1, characterized in that: Both the historical bill of quantities analysis original dataset and the current bill of quantities analysis original dataset contain bill of quantities text data, contract terms text data, and pricing specification clause text data. The historical bill of quantities analysis original dataset also contains bill of quantities classification tags.

5. The method for analyzing bill of quantities based on contract and pricing standards according to claim 1, characterized in that: The preliminary data processing specifically includes the following steps: Step S21: Text standardization is used to eliminate inconsistencies in encoding, font, and unit of measurement among the three types of data. Specifically, it involves unifying UTF-8 encoding, normalizing full-width and half-width characters, converting between simplified and traditional Chinese characters, and mapping synonymous units of measurement to standard words to obtain standardized text. Step S22: Structural segmentation, used to break down continuous long text into independent processable units according to natural semantic boundaries. Specifically, the bill of quantities is segmented into item-level records by line, contract clauses are segmented into clause blocks by clause number regularization, and pricing specification clauses are segmented into clause blocks by chapter number and clause number. Step S23: Data purification and alignment, used to remove irrelevant and interfering information and ensure that the structure of each segmentation unit is complete and consistent. Specifically, headers, footers and table of contents indexes are removed by regular expression matching and position rules, non-detailed data and list entries with missing item feature descriptions are filtered out, and a unique primary key is assigned to each segmentation unit. Step S24: Numerical extraction, used to extract quantitative parameters from text, specifically extracting numerical fields from the bill of quantities and extracting numerical fields from the contract text using regular expressions and named entity recognition to form quantitative parameters. The numerical fields include, but are not limited to, tender quantity, change quantity, bid unit price, change rate threshold, and amount threshold. Step S25: Dataset partitioning, used to obtain training data and validation data, specifically, splitting the original dataset of the historical list analysis to obtain the extracted training set and the extracted test set; The current list analysis raw dataset is pre-processed through text standardization, structural segmentation, data purification and alignment, and numerical extraction to obtain the dataset to be extracted. The historical list analysis raw dataset is pre-processed through text standardization, structural segmentation, data purification and alignment, numerical extraction, and dataset partitioning to obtain the extraction training set and extraction test set.

6. A bill of quantities analysis system based on contract and pricing specifications, used to implement the bill of quantities analysis method based on contract and pricing specifications as described in any one of claims 1-5, characterized in that: It includes a bill of quantities data acquisition module, a preliminary data processing module, a structured extraction model construction module, a bill of quantities classification model construction module, and a bill of quantities analysis module.

7. The bill of quantities analysis system based on contract and pricing standards according to claim 6, characterized in that: The list data acquisition module is used to collect raw data, obtain the list analysis raw dataset by collecting raw data, and send the list analysis raw dataset to the data preliminary processing module. The data preliminary processing module is used for preliminary data processing. Through preliminary data processing, the dataset to be extracted, the training set to be extracted, and the test set to be extracted are obtained. The dataset to be extracted is sent to the bill of quantities analysis module, and the training set and the test set to be extracted are sent to the structured extraction model construction module. The structured extraction model construction module is used to construct a structured extraction model. By constructing an improved BERT model, a structured extraction model is obtained. The extraction training set and extraction test set are used as inputs to the structured extraction model for processing to obtain a classification training set and a classification test set. The structured extraction model is then sent to the bill of quantities analysis module, and the classification training set and the classification test set are sent to the bill of quantities classification model construction module. The bill of quantities classification model construction module is used to construct the bill of quantities classification model. By constructing an improved encoder model, the bill of quantities classification model is obtained, and the bill of quantities classification model is sent to the bill of quantities analysis module. The bill of quantities analysis module is used for bill of quantities analysis. By combining the structured extraction model and the bill of quantities classification model, it performs real-time bill of quantities analysis to obtain the final bill of quantities analysis results.

Images